pq bench: avoid repeated conversion to_i128

depends on https://github.com/neondatabase/rust-postgres/pull/25

.github/actionlint.yml

@@ -5,4 +5,6 @@ self-hosted-runner:
     - small
     - us-east-2
 config-variables:
+  - REMOTE_STORAGE_AZURE_CONTAINER
+  - REMOTE_STORAGE_AZURE_REGION
   - SLACK_UPCOMING_RELEASE_CHANNEL_ID

.github/actions/allure-report-generate/action.yml

@@ -203,6 +203,10 @@ runs:
         COMMIT_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
         BASE_S3_URL: ${{ steps.generate-report.outputs.base-s3-url }}
       run: |
+        if [ ! -d "${WORKDIR}/report/data/test-cases" ]; then
+          exit 0
+        fi
+
         export DATABASE_URL=${REGRESS_TEST_RESULT_CONNSTR_NEW}
 
         ./scripts/pysync

.github/workflows/build_and_test.yml

@@ -320,6 +320,9 @@ jobs:
       - name: Build neon extensions
         run: mold -run make neon-pg-ext -j$(nproc)
 
+      - name: Build walproposer-lib
+        run: mold -run make walproposer-lib -j$(nproc)
+
       - name: Run cargo build
         run: |
           ${cov_prefix} mold -run cargo build $CARGO_FLAGS $CARGO_FEATURES --bins --tests
@@ -335,6 +338,16 @@ jobs:
           # Avoid `$CARGO_FEATURES` since there's no `testing` feature in the e2e tests now
           ${cov_prefix} cargo test $CARGO_FLAGS --package remote_storage --test test_real_s3
 
+          # Run separate tests for real Azure Blob Storage
+          # XXX: replace region with `eu-central-1`-like region
+          export ENABLE_REAL_AZURE_REMOTE_STORAGE=y
+          export AZURE_STORAGE_ACCOUNT="${{ secrets.AZURE_STORAGE_ACCOUNT_DEV }}"
+          export AZURE_STORAGE_ACCESS_KEY="${{ secrets.AZURE_STORAGE_ACCESS_KEY_DEV }}"
+          export REMOTE_STORAGE_AZURE_CONTAINER="${{ vars.REMOTE_STORAGE_AZURE_CONTAINER }}"
+          export REMOTE_STORAGE_AZURE_REGION="${{ vars.REMOTE_STORAGE_AZURE_REGION }}"
+          # Avoid `$CARGO_FEATURES` since there's no `testing` feature in the e2e tests now
+          ${cov_prefix} cargo test $CARGO_FLAGS --package remote_storage --test test_real_azure
+
       - name: Install rust binaries
         run: |
           # Install target binaries
@@ -420,7 +433,7 @@ jobs:
           rerun_flaky: true
           pg_version: ${{ matrix.pg_version }}
         env:
-          TEST_RESULT_CONNSTR: ${{ secrets.REGRESS_TEST_RESULT_CONNSTR }}
+          TEST_RESULT_CONNSTR: ${{ secrets.REGRESS_TEST_RESULT_CONNSTR_NEW }}
           CHECK_ONDISK_DATA_COMPATIBILITY: nonempty
 
       - name: Merge and upload coverage data
@@ -455,7 +468,7 @@ jobs:
         env:
           VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
           PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
-          TEST_RESULT_CONNSTR: "${{ secrets.REGRESS_TEST_RESULT_CONNSTR }}"
+          TEST_RESULT_CONNSTR: "${{ secrets.REGRESS_TEST_RESULT_CONNSTR_NEW }}"
       # XXX: no coverage data handling here, since benchmarks are run on release builds,
       # while coverage is currently collected for the debug ones
 
@@ -834,7 +847,7 @@ jobs:
       run:
         shell: sh -eu {0}
     env:
-      VM_BUILDER_VERSION: v0.17.12
+      VM_BUILDER_VERSION: v0.18.2
 
     steps:
       - name: Checkout
@@ -1092,8 +1105,10 @@ jobs:
         run: |
           if [[ "$GITHUB_REF_NAME" == "main" ]]; then
             gh workflow --repo neondatabase/aws run deploy-dev.yml --ref main -f branch=main -f dockerTag=${{needs.tag.outputs.build-tag}} -f deployPreprodRegion=false
-          elif [[ "$GITHUB_REF_NAME" == "release" ]]; then
+
+            # TODO: move deployPreprodRegion to release (`"$GITHUB_REF_NAME" == "release"` block), once Staging support different compute tag prefixes for different regions
             gh workflow --repo neondatabase/aws run deploy-dev.yml --ref main -f branch=main -f dockerTag=${{needs.tag.outputs.build-tag}} -f deployPreprodRegion=true
+          elif [[ "$GITHUB_REF_NAME" == "release" ]]; then
             gh workflow --repo neondatabase/aws run deploy-prod.yml --ref main -f branch=main -f dockerTag=${{needs.tag.outputs.build-tag}} -f disclamerAcknowledged=true
           else
             echo "GITHUB_REF_NAME (value '$GITHUB_REF_NAME') is not set to either 'main' or 'release'"

.github/workflows/neon_extra_builds.yml

@@ -32,7 +32,7 @@ jobs:
 
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           submodules: true
           fetch-depth: 1
@@ -90,18 +90,21 @@ jobs:
 
       - name: Build postgres v14
         if: steps.cache_pg_14.outputs.cache-hit != 'true'
-        run: make postgres-v14 -j$(nproc)
+        run: make postgres-v14 -j$(sysctl -n hw.ncpu)
 
       - name: Build postgres v15
         if: steps.cache_pg_15.outputs.cache-hit != 'true'
-        run: make postgres-v15 -j$(nproc)
+        run: make postgres-v15 -j$(sysctl -n hw.ncpu)
 
       - name: Build postgres v16
         if: steps.cache_pg_16.outputs.cache-hit != 'true'
-        run: make postgres-v16 -j$(nproc)
+        run: make postgres-v16 -j$(sysctl -n hw.ncpu)
 
       - name: Build neon extensions
-        run: make neon-pg-ext -j$(nproc)
+        run: make neon-pg-ext -j$(sysctl -n hw.ncpu)
+
+      - name: Build walproposer-lib
+        run: make walproposer-lib -j$(sysctl -n hw.ncpu)
 
       - name: Run cargo build
         run: cargo build --all --release
@@ -126,7 +129,7 @@ jobs:
 
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           submodules: true
           fetch-depth: 1
@@ -135,6 +138,9 @@ jobs:
       - name: Get postgres headers
         run: make postgres-headers -j$(nproc)
 
+      - name: Build walproposer-lib
+        run: make walproposer-lib -j$(nproc)
+
       - name: Produce the build stats
         run: cargo build --all --release --timings
 

Cargo.toml

@@ -26,6 +26,7 @@ members = [
     "libs/tracing-utils",
     "libs/postgres_ffi/wal_craft",
     "libs/vm_monitor",
+    "libs/walproposer",
 ]
 
 [workspace.package]
@@ -36,6 +37,10 @@ license = "Apache-2.0"
 [workspace.dependencies]
 anyhow = { version = "1.0", features = ["backtrace"] }
 async-compression = { version = "0.4.0", features = ["tokio", "gzip"] }
+azure_core = "0.16"
+azure_identity = "0.16"
+azure_storage = "0.16"
+azure_storage_blobs = "0.16"
 flate2 = "1.0.26"
 async-stream = "0.3"
 async-trait = "0.1"
@@ -76,6 +81,7 @@ hex = "0.4"
 hex-literal = "0.4"
 hmac = "0.12.1"
 hostname = "0.3.1"
+http-types = "2"
 humantime = "2.1"
 humantime-serde = "1.1.1"
 hyper = "0.14"
@@ -155,11 +161,11 @@ env_logger = "0.10"
 log = "0.4"
 
 ## Libraries from neondatabase/ git forks, ideally with changes to be upstreamed
-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
-postgres-native-tls = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
-postgres-protocol = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
-postgres-types = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
-tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
+postgres = { git = "https://github.com/neondatabase/rust-postgres.git", branch="problame/copy-both-duplex-public" }
+postgres-native-tls = { git = "https://github.com/neondatabase/rust-postgres.git", branch="problame/copy-both-duplex-public" }
+postgres-protocol = { git = "https://github.com/neondatabase/rust-postgres.git", branch="problame/copy-both-duplex-public" }
+postgres-types = { git = "https://github.com/neondatabase/rust-postgres.git", branch="problame/copy-both-duplex-public" }
+tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", branch="problame/copy-both-duplex-public" }
 
 ## Other git libraries
 heapless = { default-features=false, features=[], git = "https://github.com/japaric/heapless.git", rev = "644653bf3b831c6bb4963be2de24804acf5e5001" } # upstream release pending
@@ -180,6 +186,7 @@ tenant_size_model = { version = "0.1", path = "./libs/tenant_size_model/" }
 tracing-utils = { version = "0.1", path = "./libs/tracing-utils/" }
 utils = { version = "0.1", path = "./libs/utils/" }
 vm_monitor = { version = "0.1", path = "./libs/vm_monitor/" }
+walproposer = { version = "0.1", path = "./libs/walproposer/" }
 
 ## Common library dependency
 workspace_hack = { version = "0.1", path = "./workspace_hack/" }
@@ -195,7 +202,7 @@ tonic-build = "0.9"
 
 # This is only needed for proxy's tests.
 # TODO: we should probably fork `tokio-postgres-rustls` instead.
-tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
+tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", branch="problame/copy-both-duplex-public" }
 
 ################# Binary contents sections
 

Dockerfile.compute-node

@@ -224,8 +224,8 @@ RUN wget https://github.com/df7cb/postgresql-unit/archive/refs/tags/7.7.tar.gz -
 FROM build-deps AS vector-pg-build
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
 
-RUN wget https://github.com/pgvector/pgvector/archive/refs/tags/v0.5.0.tar.gz -O pgvector.tar.gz && \
-    echo "d8aa3504b215467ca528525a6de12c3f85f9891b091ce0e5864dd8a9b757f77b pgvector.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/pgvector/pgvector/archive/refs/tags/v0.5.1.tar.gz -O pgvector.tar.gz && \
+    echo "cc7a8e034a96e30a819911ac79d32f6bc47bdd1aa2de4d7d4904e26b83209dc8 pgvector.tar.gz" | sha256sum --check && \
     mkdir pgvector-src && cd pgvector-src && tar xvzf ../pgvector.tar.gz --strip-components=1 -C . && \
     make -j $(getconf _NPROCESSORS_ONLN) PG_CONFIG=/usr/local/pgsql/bin/pg_config && \
     make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config && \
@@ -368,8 +368,8 @@ RUN wget https://github.com/citusdata/postgresql-hll/archive/refs/tags/v2.18.tar
 FROM build-deps AS plpgsql-check-pg-build
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
 
-RUN wget https://github.com/okbob/plpgsql_check/archive/refs/tags/v2.4.0.tar.gz -O plpgsql_check.tar.gz && \
-    echo "9ba58387a279b35a3bfa39ee611e5684e6cddb2ba046ddb2c5190b3bd2ca254a plpgsql_check.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/okbob/plpgsql_check/archive/refs/tags/v2.5.3.tar.gz -O plpgsql_check.tar.gz && \
+    echo "6631ec3e7fb3769eaaf56e3dfedb829aa761abf163d13dba354b4c218508e1c0 plpgsql_check.tar.gz" | sha256sum --check && \
     mkdir plpgsql_check-src && cd plpgsql_check-src && tar xvzf ../plpgsql_check.tar.gz --strip-components=1 -C . && \
     make -j $(getconf _NPROCESSORS_ONLN) PG_CONFIG=/usr/local/pgsql/bin/pg_config USE_PGXS=1 && \
     make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config USE_PGXS=1 && \

Makefile

@@ -62,7 +62,7 @@ all: neon postgres neon-pg-ext
 #
 # The 'postgres_ffi' depends on the Postgres headers.
 .PHONY: neon
-neon: postgres-headers
+neon: postgres-headers walproposer-lib
 	+@echo "Compiling Neon"
 	$(CARGO_CMD_PREFIX) cargo build $(CARGO_BUILD_FLAGS)
 
@@ -168,6 +168,42 @@ neon-pg-ext-clean-%:
 	-C $(POSTGRES_INSTALL_DIR)/build/neon-utils-$* \
 	-f $(ROOT_PROJECT_DIR)/pgxn/neon_utils/Makefile clean
 
+# Build walproposer as a static library. walproposer source code is located
+# in the pgxn/neon directory.
+# 
+# We also need to include libpgport.a and libpgcommon.a, because walproposer
+# uses some functions from those libraries.
+# 
+# Some object files are removed from libpgport.a and libpgcommon.a because
+# they depend on openssl and other libraries that are not included in our
+# Rust build.
+.PHONY: walproposer-lib
+walproposer-lib: neon-pg-ext-v16
+	+@echo "Compiling walproposer-lib"
+	mkdir -p $(POSTGRES_INSTALL_DIR)/build/walproposer-lib
+	$(MAKE) PG_CONFIG=$(POSTGRES_INSTALL_DIR)/v16/bin/pg_config CFLAGS='$(PG_CFLAGS) $(COPT)' \
+		-C $(POSTGRES_INSTALL_DIR)/build/walproposer-lib \
+		-f $(ROOT_PROJECT_DIR)/pgxn/neon/Makefile walproposer-lib
+	cp $(POSTGRES_INSTALL_DIR)/v16/lib/libpgport.a $(POSTGRES_INSTALL_DIR)/build/walproposer-lib
+	cp $(POSTGRES_INSTALL_DIR)/v16/lib/libpgcommon.a $(POSTGRES_INSTALL_DIR)/build/walproposer-lib
+ifeq ($(UNAME_S),Linux)
+	$(AR) d $(POSTGRES_INSTALL_DIR)/build/walproposer-lib/libpgport.a \
+		pg_strong_random.o
+	$(AR) d $(POSTGRES_INSTALL_DIR)/build/walproposer-lib/libpgcommon.a \
+		pg_crc32c.o \
+		hmac_openssl.o \
+		cryptohash_openssl.o \
+		scram-common.o \
+		md5_common.o \
+		checksum_helper.o
+endif
+
+.PHONY: walproposer-lib-clean
+walproposer-lib-clean:
+	$(MAKE) PG_CONFIG=$(POSTGRES_INSTALL_DIR)/v16/bin/pg_config \
+		-C $(POSTGRES_INSTALL_DIR)/build/walproposer-lib \
+		-f $(ROOT_PROJECT_DIR)/pgxn/neon/Makefile clean
+
 .PHONY: neon-pg-ext
 neon-pg-ext: \
 	neon-pg-ext-v14 \

NOTICE

@@ -1,5 +1,5 @@
 Neon
 Copyright 2022 Neon Inc.
 
-The PostgreSQL submodules in vendor/postgres-v14 and vendor/postgres-v15 are licensed under the
-PostgreSQL license. See vendor/postgres-v14/COPYRIGHT and vendor/postgres-v15/COPYRIGHT.
+The PostgreSQL submodules in vendor/ are licensed under the PostgreSQL license.
+See vendor/postgres-vX/COPYRIGHT for details.

compute_tools/src/compute.rs

@@ -252,7 +252,7 @@ fn create_neon_superuser(spec: &ComputeSpec, client: &mut Client) -> Result<()>
                     IF NOT EXISTS (
                         SELECT FROM pg_catalog.pg_roles WHERE rolname = 'neon_superuser')
                     THEN
-                        CREATE ROLE neon_superuser CREATEDB CREATEROLE NOLOGIN IN ROLE pg_read_all_data, pg_write_all_data;
+                        CREATE ROLE neon_superuser CREATEDB CREATEROLE NOLOGIN REPLICATION IN ROLE pg_read_all_data, pg_write_all_data;
                         IF array_length(roles, 1) IS NOT NULL THEN
                             EXECUTE format('GRANT neon_superuser TO %s',
                                            array_to_string(ARRAY(SELECT quote_ident(x) FROM unnest(roles) as x), ', '));
@@ -692,10 +692,11 @@ impl ComputeNode {
         // Proceed with post-startup configuration. Note, that order of operations is important.
         let spec = &compute_state.pspec.as_ref().expect("spec must be set").spec;
         create_neon_superuser(spec, &mut client)?;
+        cleanup_instance(&mut client)?;
         handle_roles(spec, &mut client)?;
         handle_databases(spec, &mut client)?;
         handle_role_deletions(spec, self.connstr.as_str(), &mut client)?;
-        handle_grants(spec, self.connstr.as_str())?;
+        handle_grants(spec, &mut client, self.connstr.as_str())?;
         handle_extensions(spec, &mut client)?;
         create_availability_check_data(&mut client)?;
 
@@ -731,10 +732,11 @@ impl ComputeNode {
         // Disable DDL forwarding because control plane already knows about these roles/databases.
         if spec.mode == ComputeMode::Primary {
             client.simple_query("SET neon.forward_ddl = false")?;
+            cleanup_instance(&mut client)?;
             handle_roles(&spec, &mut client)?;
             handle_databases(&spec, &mut client)?;
             handle_role_deletions(&spec, self.connstr.as_str(), &mut client)?;
-            handle_grants(&spec, self.connstr.as_str())?;
+            handle_grants(&spec, &mut client, self.connstr.as_str())?;
             handle_extensions(&spec, &mut client)?;
         }
 

compute_tools/src/pg_helpers.rs

@@ -1,3 +1,4 @@
+use std::collections::HashMap;
 use std::fmt::Write;
 use std::fs;
 use std::fs::File;
@@ -205,22 +206,37 @@ pub fn get_existing_roles(xact: &mut Transaction<'_>) -> Result<Vec<Role>> {
 }
 
 /// Build a list of existing Postgres databases
-pub fn get_existing_dbs(client: &mut Client) -> Result<Vec<Database>> {
-    let postgres_dbs = client
+pub fn get_existing_dbs(client: &mut Client) -> Result<HashMap<String, Database>> {
+    // `pg_database.datconnlimit = -2` means that the database is in the
+    // invalid state. See:
+    //   https://github.com/postgres/postgres/commit/a4b4cc1d60f7e8ccfcc8ff8cb80c28ee411ad9a9
+    let postgres_dbs: Vec<Database> = client
         .query(
-            "SELECT datname, datdba::regrole::text as owner
-               FROM pg_catalog.pg_database;",
+            "SELECT
+                datname AS name,
+                datdba::regrole::text AS owner,
+                NOT datallowconn AS restrict_conn,
+                datconnlimit = - 2 AS invalid
+            FROM
+                pg_catalog.pg_database;",
             &[],
         )?
         .iter()
         .map(|row| Database {
-            name: row.get("datname"),
+            name: row.get("name"),
             owner: row.get("owner"),
+            restrict_conn: row.get("restrict_conn"),
+            invalid: row.get("invalid"),
             options: None,
         })
         .collect();
 
-    Ok(postgres_dbs)
+    let dbs_map = postgres_dbs
+        .iter()
+        .map(|db| (db.name.clone(), db.clone()))
+        .collect::<HashMap<_, _>>();
+
+    Ok(dbs_map)
 }
 
 /// Wait for Postgres to become ready to accept connections. It's ready to

compute_tools/src/spec.rs

@@ -13,7 +13,7 @@ use crate::params::PG_HBA_ALL_MD5;
 use crate::pg_helpers::*;
 
 use compute_api::responses::{ControlPlaneComputeStatus, ControlPlaneSpecResponse};
-use compute_api::spec::{ComputeSpec, Database, PgIdent, Role};
+use compute_api::spec::{ComputeSpec, PgIdent, Role};
 
 // Do control plane request and return response if any. In case of error it
 // returns a bool flag indicating whether it makes sense to retry the request
@@ -161,6 +161,38 @@ pub fn add_standby_signal(pgdata_path: &Path) -> Result<()> {
     Ok(())
 }
 
+/// Compute could be unexpectedly shut down, for example, during the
+/// database dropping. This leaves the database in the invalid state,
+/// which prevents new db creation with the same name. This function
+/// will clean it up before proceeding with catalog updates. All
+/// possible future cleanup operations may go here too.
+#[instrument(skip_all)]
+pub fn cleanup_instance(client: &mut Client) -> Result<()> {
+    let existing_dbs = get_existing_dbs(client)?;
+
+    for (_, db) in existing_dbs {
+        if db.invalid {
+            // After recent commit in Postgres, interrupted DROP DATABASE
+            // leaves the database in the invalid state. According to the
+            // commit message, the only option for user is to drop it again.
+            // See:
+            //   https://github.com/postgres/postgres/commit/a4b4cc1d60f7e8ccfcc8ff8cb80c28ee411ad9a9
+            //
+            // Postgres Neon extension is done the way, that db is de-registered
+            // in the control plane metadata only after it is dropped. So there is
+            // a chance that it still thinks that db should exist. This means
+            // that it will be re-created by `handle_databases()`. Yet, it's fine
+            // as user can just repeat drop (in vanilla Postgres they would need
+            // to do the same, btw).
+            let query = format!("DROP DATABASE IF EXISTS {}", db.name.pg_quote());
+            info!("dropping invalid database {}", db.name);
+            client.execute(query.as_str(), &[])?;
+        }
+    }
+
+    Ok(())
+}
+
 /// Given a cluster spec json and open transaction it handles roles creation,
 /// deletion and update.
 #[instrument(skip_all)]
@@ -270,7 +302,7 @@ pub fn handle_roles(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
             }
             RoleAction::Create => {
                 let mut query: String = format!(
-                    "CREATE ROLE {} CREATEROLE CREATEDB BYPASSRLS IN ROLE neon_superuser",
+                    "CREATE ROLE {} CREATEROLE CREATEDB BYPASSRLS REPLICATION IN ROLE neon_superuser",
                     name.pg_quote()
                 );
                 info!("role create query: '{}'", &query);
@@ -379,13 +411,13 @@ fn reassign_owned_objects(spec: &ComputeSpec, connstr: &str, role_name: &PgIdent
 /// which together provide us idempotency.
 #[instrument(skip_all)]
 pub fn handle_databases(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
-    let existing_dbs: Vec<Database> = get_existing_dbs(client)?;
+    let existing_dbs = get_existing_dbs(client)?;
 
     // Print a list of existing Postgres databases (only in debug mode)
     if span_enabled!(Level::INFO) {
         info!("postgres databases:");
-        for r in &existing_dbs {
-            info!("    {}:{}", r.name, r.owner);
+        for (dbname, db) in &existing_dbs {
+            info!("    {}:{}", dbname, db.owner);
         }
     }
 
@@ -439,8 +471,7 @@ pub fn handle_databases(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
                 "rename_db" => {
                     let new_name = op.new_name.as_ref().unwrap();
 
-                    // XXX: with a limited number of roles it is fine, but consider making it a HashMap
-                    if existing_dbs.iter().any(|r| r.name == op.name) {
+                    if existing_dbs.get(&op.name).is_some() {
                         let query: String = format!(
                             "ALTER DATABASE {} RENAME TO {}",
                             op.name.pg_quote(),
@@ -457,14 +488,12 @@ pub fn handle_databases(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
     }
 
     // Refresh Postgres databases info to handle possible renames
-    let existing_dbs: Vec<Database> = get_existing_dbs(client)?;
+    let existing_dbs = get_existing_dbs(client)?;
 
     info!("cluster spec databases:");
     for db in &spec.cluster.databases {
         let name = &db.name;
-
-        // XXX: with a limited number of databases it is fine, but consider making it a HashMap
-        let pg_db = existing_dbs.iter().find(|r| r.name == *name);
+        let pg_db = existing_dbs.get(name);
 
         enum DatabaseAction {
             None,
@@ -530,13 +559,32 @@ pub fn handle_databases(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
 /// Grant CREATE ON DATABASE to the database owner and do some other alters and grants
 /// to allow users creating trusted extensions and re-creating `public` schema, for example.
 #[instrument(skip_all)]
-pub fn handle_grants(spec: &ComputeSpec, connstr: &str) -> Result<()> {
-    info!("cluster spec grants:");
+pub fn handle_grants(spec: &ComputeSpec, client: &mut Client, connstr: &str) -> Result<()> {
+    info!("modifying database permissions");
+    let existing_dbs = get_existing_dbs(client)?;
 
     // Do some per-database access adjustments. We'd better do this at db creation time,
     // but CREATE DATABASE isn't transactional. So we cannot create db + do some grants
     // atomically.
     for db in &spec.cluster.databases {
+        match existing_dbs.get(&db.name) {
+            Some(pg_db) => {
+                if pg_db.restrict_conn || pg_db.invalid {
+                    info!(
+                        "skipping grants for db {} (invalid: {}, connections not allowed: {})",
+                        db.name, pg_db.invalid, pg_db.restrict_conn
+                    );
+                    continue;
+                }
+            }
+            None => {
+                bail!(
+                    "database {} doesn't exist in Postgres after handle_databases()",
+                    db.name
+                );
+            }
+        }
+
         let mut conf = Config::from_str(connstr)?;
         conf.dbname(&db.name);
 
@@ -575,6 +623,11 @@ pub fn handle_grants(spec: &ComputeSpec, connstr: &str) -> Result<()> {
 
         // Explicitly grant CREATE ON SCHEMA PUBLIC to the web_access user.
         // This is needed because since postgres 15 this privilege is removed by default.
+        // TODO: web_access isn't created for almost 1 year. It could be that we have
+        // active users of 1 year old projects, but hopefully not, so check it and
+        // remove this code if possible. The worst thing that could happen is that
+        // user won't be able to use public schema in NEW databases created in the
+        // very OLD project.
         let grant_query = "DO $$\n\
                 BEGIN\n\
                     IF EXISTS(\n\

compute_tools/tests/pg_helpers_tests.rs

@@ -28,7 +28,7 @@ mod pg_helpers_tests {
         assert_eq!(
             spec.cluster.settings.as_pg_settings(),
             r#"fsync = off
-wal_level = replica
+wal_level = logical
 hot_standby = on
 neon.safekeepers = '127.0.0.1:6502,127.0.0.1:6503,127.0.0.1:6501'
 wal_log_hints = on

control_plane/src/background_process.rs

@@ -86,7 +86,7 @@ where
         .stdout(process_log_file)
         .stderr(same_file_for_stderr)
         .args(args);
-    let filled_cmd = fill_aws_secrets_vars(fill_rust_env_vars(background_command));
+    let filled_cmd = fill_remote_storage_secrets_vars(fill_rust_env_vars(background_command));
     filled_cmd.envs(envs);
 
     let pid_file_to_check = match initial_pid_file {
@@ -238,11 +238,13 @@ fn fill_rust_env_vars(cmd: &mut Command) -> &mut Command {
     filled_cmd
 }
 
-fn fill_aws_secrets_vars(mut cmd: &mut Command) -> &mut Command {
+fn fill_remote_storage_secrets_vars(mut cmd: &mut Command) -> &mut Command {
     for env_key in [
         "AWS_ACCESS_KEY_ID",
         "AWS_SECRET_ACCESS_KEY",
         "AWS_SESSION_TOKEN",
+        "AZURE_STORAGE_ACCOUNT",
+        "AZURE_STORAGE_ACCESS_KEY",
     ] {
         if let Ok(value) = std::env::var(env_key) {
             cmd = cmd.env(env_key, value);

control_plane/src/bin/attachment_service.rs

@@ -13,6 +13,7 @@ use serde::{Deserialize, Serialize};
 use std::path::{Path, PathBuf};
 use std::{collections::HashMap, sync::Arc};
 use utils::logging::{self, LogFormat};
+use utils::signals::{ShutdownSignals, Signal};
 
 use utils::{
     http::{
@@ -268,7 +269,16 @@ async fn main() -> anyhow::Result<()> {
     let server = hyper::Server::from_tcp(http_listener)?.serve(service);
 
     tracing::info!("Serving on {0}", args.listen);
-    server.await?;
+
+    tokio::task::spawn(server);
+
+    ShutdownSignals::handle(|signal| match signal {
+        Signal::Interrupt | Signal::Terminate | Signal::Quit => {
+            tracing::info!("Got {}. Terminating", signal.name());
+            // We're just a test helper: no graceful shutdown.
+            std::process::exit(0);
+        }
+    })?;
 
     Ok(())
 }

control_plane/src/bin/neon_local.rs

@@ -116,6 +116,7 @@ fn main() -> Result<()> {
             "attachment_service" => handle_attachment_service(sub_args, &env),
             "safekeeper" => handle_safekeeper(sub_args, &env),
             "endpoint" => handle_endpoint(sub_args, &env),
+            "mappings" => handle_mappings(sub_args, &mut env),
             "pg" => bail!("'pg' subcommand has been renamed to 'endpoint'"),
             _ => bail!("unexpected subcommand {sub_name}"),
         };
@@ -816,6 +817,38 @@ fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<(
     Ok(())
 }
 
+fn handle_mappings(sub_match: &ArgMatches, env: &mut local_env::LocalEnv) -> Result<()> {
+    let (sub_name, sub_args) = match sub_match.subcommand() {
+        Some(ep_subcommand_data) => ep_subcommand_data,
+        None => bail!("no mappings subcommand provided"),
+    };
+
+    match sub_name {
+        "map" => {
+            let branch_name = sub_args
+                .get_one::<String>("branch-name")
+                .expect("branch-name argument missing");
+
+            let tenant_id = sub_args
+                .get_one::<String>("tenant-id")
+                .map(|x| TenantId::from_str(x))
+                .expect("tenant-id argument missing")
+                .expect("malformed tenant-id arg");
+
+            let timeline_id = sub_args
+                .get_one::<String>("timeline-id")
+                .map(|x| TimelineId::from_str(x))
+                .expect("timeline-id argument missing")
+                .expect("malformed timeline-id arg");
+
+            env.register_branch_mapping(branch_name.to_owned(), tenant_id, timeline_id)?;
+
+            Ok(())
+        }
+        other => unimplemented!("mappings subcommand {other}"),
+    }
+}
+
 fn handle_pageserver(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<()> {
     fn get_pageserver(env: &local_env::LocalEnv, args: &ArgMatches) -> Result<PageServerNode> {
         let node_id = if let Some(id_str) = args.get_one::<String>("pageserver-id") {
@@ -1084,6 +1117,7 @@ fn cli() -> Command {
     // --id, when using a pageserver command
     let pageserver_id_arg = Arg::new("pageserver-id")
         .long("id")
+        .global(true)
         .help("pageserver id")
         .required(false);
     // --pageserver-id when using a non-pageserver command
@@ -1254,17 +1288,20 @@ fn cli() -> Command {
             Command::new("pageserver")
                 .arg_required_else_help(true)
                 .about("Manage pageserver")
+                .arg(pageserver_id_arg)
                 .subcommand(Command::new("status"))
-                .arg(pageserver_id_arg.clone())
-                .subcommand(Command::new("start").about("Start local pageserver")
-                .arg(pageserver_id_arg.clone())
-                .arg(pageserver_config_args.clone()))
-                .subcommand(Command::new("stop").about("Stop local pageserver")
-                .arg(pageserver_id_arg.clone())
-                            .arg(stop_mode_arg.clone()))
-                .subcommand(Command::new("restart").about("Restart local pageserver")
-                .arg(pageserver_id_arg.clone())
-                .arg(pageserver_config_args.clone()))
+                .subcommand(Command::new("start")
+                    .about("Start local pageserver")
+                    .arg(pageserver_config_args.clone())
+                )
+                .subcommand(Command::new("stop")
+                    .about("Stop local pageserver")
+                    .arg(stop_mode_arg.clone())
+                )
+                .subcommand(Command::new("restart")
+                    .about("Restart local pageserver")
+                    .arg(pageserver_config_args.clone())
+                )
         )
         .subcommand(
             Command::new("attachment_service")
@@ -1321,8 +1358,8 @@ fn cli() -> Command {
                     .about("Start postgres.\n If the endpoint doesn't exist yet, it is created.")
                     .arg(endpoint_id_arg.clone())
                     .arg(tenant_id_arg.clone())
-                    .arg(branch_name_arg)
-                    .arg(timeline_id_arg)
+                    .arg(branch_name_arg.clone())
+                    .arg(timeline_id_arg.clone())
                     .arg(lsn_arg)
                     .arg(pg_port_arg)
                     .arg(http_port_arg)
@@ -1335,7 +1372,7 @@ fn cli() -> Command {
                 .subcommand(
                     Command::new("stop")
                     .arg(endpoint_id_arg)
-                    .arg(tenant_id_arg)
+                    .arg(tenant_id_arg.clone())
                     .arg(
                         Arg::new("destroy")
                             .help("Also delete data directory (now optional, should be default in future)")
@@ -1346,6 +1383,18 @@ fn cli() -> Command {
                 )
 
         )
+        .subcommand(
+            Command::new("mappings")
+                .arg_required_else_help(true)
+                .about("Manage neon_local branch name mappings")
+                .subcommand(
+                    Command::new("map")
+                        .about("Create new mapping which cannot exist already")
+                        .arg(branch_name_arg.clone())
+                        .arg(tenant_id_arg.clone())
+                        .arg(timeline_id_arg.clone())
+                )
+        )
         // Obsolete old name for 'endpoint'. We now just print an error if it's used.
         .subcommand(
             Command::new("pg")

control_plane/src/endpoint.rs

@@ -253,7 +253,7 @@ impl Endpoint {
         conf.append("shared_buffers", "1MB");
         conf.append("fsync", "off");
         conf.append("max_connections", "100");
-        conf.append("wal_level", "replica");
+        conf.append("wal_level", "logical");
         // wal_sender_timeout is the maximum time to wait for WAL replication.
         // It also defines how often the walreciever will send a feedback message to the wal sender.
         conf.append("wal_sender_timeout", "5s");

docker-compose/compute_wrapper/var/db/postgres/specs/spec.json

@@ -25,7 +25,7 @@
             },
             {
                 "name": "wal_level",
-                "value": "replica",
+                "value": "logical",
                 "vartype": "enum"
             },
             {

docs/error-handling.md

@@ -188,11 +188,60 @@ that.
 
 ## Error message style
 
+### PostgreSQL extensions
+
 PostgreSQL has a style guide for writing error messages:
 
 https://www.postgresql.org/docs/current/error-style-guide.html
 
 Follow that guide when writing error messages in the PostgreSQL
-extension. We don't follow it strictly in the pageserver and
-safekeeper, but the advice in the PostgreSQL style guide is generally
-good, and you can't go wrong by following it.
+extensions.
+
+### Neon Rust code
+
+#### Anyhow Context
+
+When adding anyhow `context()`, use form `present-tense-verb+action`.
+
+Example:
+- Bad: `file.metadata().context("could not get file metadata")?;`
+- Good: `file.metadata().context("get file metadata")?;`
+
+#### Logging Errors
+
+When logging any error `e`, use `could not {e:#}` or `failed to {e:#}`.
+
+If `e` is an `anyhow` error and you want to log the backtrace that it contains,
+use `{e:?}` instead of `{e:#}`.
+
+#### Rationale
+
+The `{:#}` ("alternate Display") of an `anyhow` error chain is concatenation fo the contexts, using `: `.
+
+For example, the following Rust code will result in output
+```
+ERROR  failed to list users: load users from server: parse response: invalid json
+```
+
+This is more concise / less noisy than what happens if you do `.context("could not ...")?` at each level, i.e.:
+
+```
+ERROR  could not list users: could not load users from server: could not parse response: invalid json
+```
+
+
+```rust
+fn main() {
+  match list_users().context("list users") else {
+    Ok(_) => ...,
+    Err(e) => tracing::error!("failed to {e:#}"),
+  }
+}
+fn list_users() {
+  http_get_users().context("load users from server")?;
+}
+fn http_get_users() {
+  let response = client....?;
+  response.parse().context("parse response")?; // fails with serde error "invalid json"
+}
+```

docs/pageserver-services.md

@@ -96,6 +96,16 @@ prefix_in_bucket = '/test_prefix/'
 
 `AWS_SECRET_ACCESS_KEY` and `AWS_ACCESS_KEY_ID` env variables can be used to specify the S3 credentials if needed.
 
+or
+
+```toml
+[remote_storage]
+container_name = 'some-container-name'
+container_region = 'us-east'
+prefix_in_container = '/test-prefix/'
+```
+
+`AZURE_STORAGE_ACCOUNT` and `AZURE_STORAGE_ACCESS_KEY` env variables can be used to specify the azure credentials if needed.
 
 ## Repository background tasks
 

libs/compute_api/src/spec.rs

@@ -200,6 +200,12 @@ pub struct Database {
     pub name: PgIdent,
     pub owner: PgIdent,
     pub options: GenericOptions,
+    // These are derived flags, not present in the spec file.
+    // They are never set by the control plane.
+    #[serde(skip_deserializing, default)]
+    pub restrict_conn: bool,
+    #[serde(skip_deserializing, default)]
+    pub invalid: bool,
 }
 
 /// Common type representing both SQL statement params with or without value,

libs/compute_api/tests/cluster_spec.json

@@ -76,7 +76,7 @@
             },
             {
                 "name": "wal_level",
-                "value": "replica",
+                "value": "logical",
                 "vartype": "enum"
             },
             {

libs/metrics/src/wrappers.rs

@@ -1,6 +1,6 @@
 use std::io::{Read, Result, Write};
 
-/// A wrapper for an object implementing [Read](std::io::Read)
+/// A wrapper for an object implementing [Read]
 /// which allows a closure to observe the amount of bytes read.
 /// This is useful in conjunction with metrics (e.g. [IntCounter](crate::IntCounter)).
 ///
@@ -51,17 +51,17 @@ impl<'a, T> CountedReader<'a, T> {
         }
     }
 
-    /// Get an immutable reference to the underlying [Read](std::io::Read) implementor
+    /// Get an immutable reference to the underlying [Read] implementor
     pub fn inner(&self) -> &T {
         &self.reader
     }
 
-    /// Get a mutable reference to the underlying [Read](std::io::Read) implementor
+    /// Get a mutable reference to the underlying [Read] implementor
     pub fn inner_mut(&mut self) -> &mut T {
         &mut self.reader
     }
 
-    /// Consume the wrapper and return the underlying [Read](std::io::Read) implementor
+    /// Consume the wrapper and return the underlying [Read] implementor
     pub fn into_inner(self) -> T {
         self.reader
     }
@@ -75,7 +75,7 @@ impl<T: Read> Read for CountedReader<'_, T> {
     }
 }
 
-/// A wrapper for an object implementing [Write](std::io::Write)
+/// A wrapper for an object implementing [Write]
 /// which allows a closure to observe the amount of bytes written.
 /// This is useful in conjunction with metrics (e.g. [IntCounter](crate::IntCounter)).
 ///
@@ -122,17 +122,17 @@ impl<'a, T> CountedWriter<'a, T> {
         }
     }
 
-    /// Get an immutable reference to the underlying [Write](std::io::Write) implementor
+    /// Get an immutable reference to the underlying [Write] implementor
     pub fn inner(&self) -> &T {
         &self.writer
     }
 
-    /// Get a mutable reference to the underlying [Write](std::io::Write) implementor
+    /// Get a mutable reference to the underlying [Write] implementor
     pub fn inner_mut(&mut self) -> &mut T {
         &mut self.writer
     }
 
-    /// Consume the wrapper and return the underlying [Write](std::io::Write) implementor
+    /// Consume the wrapper and return the underlying [Write] implementor
     pub fn into_inner(self) -> T {
         self.writer
     }

libs/pageserver_api/src/models.rs

@@ -10,6 +10,7 @@ use serde_with::{serde_as, DisplayFromStr};
 use strum_macros;
 use utils::{
     completion,
+    generation::Generation,
     history_buffer::HistoryBufferWithDropCounter,
     id::{NodeId, TenantId, TimelineId},
     lsn::Lsn,
@@ -17,7 +18,7 @@ use utils::{
 
 use crate::reltag::RelTag;
 use anyhow::bail;
-use bytes::{BufMut, Bytes, BytesMut};
+use bytes::{Buf, BufMut, Bytes, BytesMut};
 
 /// The state of a tenant in this pageserver.
 ///
@@ -218,6 +219,8 @@ impl std::ops::Deref for TenantCreateRequest {
     }
 }
 
+/// An alternative representation of `pageserver::tenant::TenantConf` with
+/// simpler types.
 #[derive(Serialize, Deserialize, Debug, Default)]
 pub struct TenantConfig {
     pub checkpoint_distance: Option<u64>,
@@ -243,6 +246,39 @@ pub struct TenantConfig {
     pub gc_feedback: Option<bool>,
 }
 
+/// A flattened analog of a `pagesever::tenant::LocationMode`, which
+/// lists out all possible states (and the virtual "Detached" state)
+/// in a flat form rather than using rust-style enums.
+#[derive(Serialize, Deserialize, Debug)]
+pub enum LocationConfigMode {
+    AttachedSingle,
+    AttachedMulti,
+    AttachedStale,
+    Secondary,
+    Detached,
+}
+
+#[derive(Serialize, Deserialize, Debug)]
+pub struct LocationConfigSecondary {
+    pub warm: bool,
+}
+
+/// An alternative representation of `pageserver::tenant::LocationConf`,
+/// for use in external-facing APIs.
+#[derive(Serialize, Deserialize, Debug)]
+pub struct LocationConfig {
+    pub mode: LocationConfigMode,
+    /// If attaching, in what generation?
+    #[serde(default)]
+    pub generation: Option<Generation>,
+    #[serde(default)]
+    pub secondary_conf: Option<LocationConfigSecondary>,
+
+    // If requesting mode `Secondary`, configuration for that.
+    // Custom storage configuration for the tenant, if any
+    pub tenant_conf: TenantConfig,
+}
+
 #[serde_as]
 #[derive(Serialize, Deserialize)]
 #[serde(transparent)]
@@ -253,6 +289,16 @@ pub struct StatusResponse {
     pub id: NodeId,
 }
 
+#[serde_as]
+#[derive(Serialize, Deserialize, Debug)]
+#[serde(deny_unknown_fields)]
+pub struct TenantLocationConfigRequest {
+    #[serde_as(as = "DisplayFromStr")]
+    pub tenant_id: TenantId,
+    #[serde(flatten)]
+    pub config: LocationConfig, // as we have a flattened field, we should reject all unknown fields in it
+}
+
 #[serde_as]
 #[derive(Serialize, Deserialize, Debug)]
 #[serde(deny_unknown_fields)]
@@ -761,6 +807,36 @@ impl PagestreamBeMessage {
 
         bytes.into()
     }
+
+    pub fn deserialize(buf: Bytes) -> anyhow::Result<Self> {
+        let mut buf = buf.reader();
+        let msg_tag = buf.read_u8()?;
+        match msg_tag {
+            100 => todo!(),
+            101 => todo!(),
+            102 => {
+                let buf = buf.get_ref();
+                /* TODO use constant */
+                if buf.len() == 8192 {
+                    Ok(PagestreamBeMessage::GetPage(PagestreamGetPageResponse {
+                        page: buf.clone(),
+                    }))
+                } else {
+                    anyhow::bail!("invalid page size: {}", buf.len());
+                }
+            }
+            103 => {
+                let buf = buf.get_ref();
+                let cstr = std::ffi::CStr::from_bytes_until_nul(&buf)?;
+                let rust_str = cstr.to_str()?;
+                Ok(PagestreamBeMessage::Error(PagestreamErrorResponse {
+                    message: rust_str.to_owned(),
+                }))
+            }
+            104 => todo!(),
+            _ => bail!("unknown tag: {:?}", msg_tag),
+        }
+    }
 }
 
 #[cfg(test)]

libs/pageserver_api/src/reltag.rs

@@ -22,9 +22,9 @@ use postgres_ffi::Oid;
 /// [See more related comments here](https:///github.com/postgres/postgres/blob/99c5852e20a0987eca1c38ba0c09329d4076b6a0/src/include/storage/relfilenode.h#L57).
 ///
 // FIXME: should move 'forknum' as last field to keep this consistent with Postgres.
-// Then we could replace the custo Ord and PartialOrd implementations below with
-// deriving them.
-#[derive(Debug, PartialEq, Eq, Hash, Clone, Copy, Serialize, Deserialize)]
+// Then we could replace the custom Ord and PartialOrd implementations below with
+// deriving them. This will require changes in walredoproc.c.
+#[derive(Debug, PartialEq, Eq, Hash, Clone, Copy, Serialize)]
 pub struct RelTag {
     pub forknum: u8,
     pub spcnode: Oid,
@@ -40,21 +40,9 @@ impl PartialOrd for RelTag {
 
 impl Ord for RelTag {
     fn cmp(&self, other: &Self) -> Ordering {
-        let mut cmp = self.spcnode.cmp(&other.spcnode);
-        if cmp != Ordering::Equal {
-            return cmp;
-        }
-        cmp = self.dbnode.cmp(&other.dbnode);
-        if cmp != Ordering::Equal {
-            return cmp;
-        }
-        cmp = self.relnode.cmp(&other.relnode);
-        if cmp != Ordering::Equal {
-            return cmp;
-        }
-        cmp = self.forknum.cmp(&other.forknum);
-
-        cmp
+        // Custom ordering where we put forknum to the end of the list
+        let other_tup = (other.spcnode, other.dbnode, other.relnode, other.forknum);
+        (self.spcnode, self.dbnode, self.relnode, self.forknum).cmp(&other_tup)
     }
 }
 

libs/postgres_backend/src/lib.rs

@@ -19,8 +19,8 @@ use tracing::{debug, error, info, trace};
 
 use pq_proto::framed::{ConnectionError, Framed, FramedReader, FramedWriter};
 use pq_proto::{
-    BeMessage, FeMessage, FeStartupPacket, ProtocolError, SQLSTATE_INTERNAL_ERROR,
-    SQLSTATE_SUCCESSFUL_COMPLETION,
+    BeMessage, FeMessage, FeStartupPacket, ProtocolError, SQLSTATE_ADMIN_SHUTDOWN,
+    SQLSTATE_INTERNAL_ERROR, SQLSTATE_SUCCESSFUL_COMPLETION,
 };
 
 /// An error, occurred during query processing:
@@ -30,6 +30,9 @@ pub enum QueryError {
     /// The connection was lost while processing the query.
     #[error(transparent)]
     Disconnected(#[from] ConnectionError),
+    /// We were instructed to shutdown while processing the query
+    #[error("Shutting down")]
+    Shutdown,
     /// Some other error
     #[error(transparent)]
     Other(#[from] anyhow::Error),
@@ -44,7 +47,8 @@ impl From<io::Error> for QueryError {
 impl QueryError {
     pub fn pg_error_code(&self) -> &'static [u8; 5] {
         match self {
-            Self::Disconnected(_) => b"08006",         // connection failure
+            Self::Disconnected(_) => b"08006", // connection failure
+            Self::Shutdown => SQLSTATE_ADMIN_SHUTDOWN,
             Self::Other(_) => SQLSTATE_INTERNAL_ERROR, // internal error
         }
     }
@@ -396,7 +400,20 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
         // socket might be already closed, e.g. if previously received error,
         // so ignore result.
         self.framed.shutdown().await.ok();
-        ret
+        match ret {
+            Ok(()) => Ok(()),
+            Err(QueryError::Shutdown) => {
+                info!("Stopped due to shutdown");
+                Ok(())
+            }
+            Err(QueryError::Disconnected(e)) => {
+                info!("Disconnected ({e:#})");
+                // Disconnection is not an error: we just use it that way internally to drop
+                // out of loops.
+                Ok(())
+            }
+            e => e,
+        }
     }
 
     async fn run_message_loop<F, S>(
@@ -416,15 +433,11 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
             _ = shutdown_watcher() => {
                 // We were requested to shut down.
                 tracing::info!("shutdown request received during handshake");
-                return Ok(())
+                return Err(QueryError::Shutdown)
             },
 
-            result = self.handshake(handler) => {
-                // Handshake complete.
-                result?;
-                if self.state == ProtoState::Closed {
-                    return Ok(()); // EOF during handshake
-                }
+            handshake_r = self.handshake(handler) => {
+                handshake_r?;
             }
         );
 
@@ -435,17 +448,34 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
             _ = shutdown_watcher() => {
                 // We were requested to shut down.
                 tracing::info!("shutdown request received in run_message_loop");
-                Ok(None)
+                return Err(QueryError::Shutdown)
             },
             msg = self.read_message() => { msg },
         )? {
             trace!("got message {:?}", msg);
 
             let result = self.process_message(handler, msg, &mut query_string).await;
-            self.flush().await?;
+            tokio::select!(
+                biased;
+                _ = shutdown_watcher() => {
+                    // We were requested to shut down.
+                    tracing::info!("shutdown request received during response flush");
+
+                    // If we exited process_message with a shutdown error, there may be
+                    // some valid response content on in our transmit buffer: permit sending
+                    // this within a short timeout.  This is a best effort thing so we don't
+                    // care about the result.
+                    tokio::time::timeout(std::time::Duration::from_millis(500), self.flush()).await.ok();
+
+                    return Err(QueryError::Shutdown)
+                },
+                flush_r = self.flush() => {
+                    flush_r?;
+                }
+            );
+
             match result? {
                 ProcessMsgResult::Continue => {
-                    self.flush().await?;
                     continue;
                 }
                 ProcessMsgResult::Break => break,
@@ -550,7 +580,9 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
                         self.peer_addr
                     );
                     self.state = ProtoState::Closed;
-                    return Ok(());
+                    return Err(QueryError::Disconnected(ConnectionError::Protocol(
+                        ProtocolError::Protocol("EOF during handshake".to_string()),
+                    )));
                 }
             }
         }
@@ -589,7 +621,9 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
                         self.peer_addr
                     );
                     self.state = ProtoState::Closed;
-                    return Ok(());
+                    return Err(QueryError::Disconnected(ConnectionError::Protocol(
+                        ProtocolError::Protocol("EOF during auth".to_string()),
+                    )));
                 }
             }
         }
@@ -913,6 +947,7 @@ impl<'a, IO: AsyncRead + AsyncWrite + Unpin> AsyncWrite for CopyDataWriter<'a, I
 pub fn short_error(e: &QueryError) -> String {
     match e {
         QueryError::Disconnected(connection_error) => connection_error.to_string(),
+        QueryError::Shutdown => "shutdown".to_string(),
         QueryError::Other(e) => format!("{e:#}"),
     }
 }
@@ -929,6 +964,9 @@ fn log_query_error(query: &str, e: &QueryError) {
         QueryError::Disconnected(other_connection_error) => {
             error!("query handler for '{query}' failed with connection error: {other_connection_error:?}")
         }
+        QueryError::Shutdown => {
+            info!("query handler for '{query}' cancelled during tenant shutdown")
+        }
         QueryError::Other(e) => {
             error!("query handler for '{query}' failed: {e:?}");
         }

libs/postgres_ffi/src/lib.rs

@@ -131,6 +131,7 @@ pub const MAX_SEND_SIZE: usize = XLOG_BLCKSZ * 16;
 
 // Export some version independent functions that are used outside of this mod
 pub use v14::xlog_utils::encode_logical_message;
+pub use v14::xlog_utils::from_pg_timestamp;
 pub use v14::xlog_utils::get_current_timestamp;
 pub use v14::xlog_utils::to_pg_timestamp;
 pub use v14::xlog_utils::XLogFileName;

libs/postgres_ffi/src/pg_constants.rs

@@ -220,6 +220,10 @@ pub const XLOG_CHECKPOINT_ONLINE: u8 = 0x10;
 pub const XLP_FIRST_IS_CONTRECORD: u16 = 0x0001;
 pub const XLP_LONG_HEADER: u16 = 0x0002;
 
+/* From replication/slot.h */
+pub const REPL_SLOT_ON_DISK_OFFSETOF_RESTART_LSN: usize = 4*4  /* offset of `slotdata` in ReplicationSlotOnDisk  */
+   + 64 /* NameData */  + 4*4;
+
 /* From fsm_internals.h */
 const FSM_NODES_PER_PAGE: usize = BLCKSZ as usize - SIZEOF_PAGE_HEADER_DATA - 4;
 const FSM_NON_LEAF_NODES_PER_PAGE: usize = BLCKSZ as usize / 2 - 1;

libs/postgres_ffi/src/xlog_utils.rs

@@ -136,21 +136,42 @@ pub fn get_current_timestamp() -> TimestampTz {
     to_pg_timestamp(SystemTime::now())
 }
 
-pub fn to_pg_timestamp(time: SystemTime) -> TimestampTz {
-    const UNIX_EPOCH_JDATE: u64 = 2440588; /* == date2j(1970, 1, 1) */
-    const POSTGRES_EPOCH_JDATE: u64 = 2451545; /* == date2j(2000, 1, 1) */
+// Module to reduce the scope of the constants
+mod timestamp_conversions {
+    use std::time::Duration;
+
+    use super::*;
+
+    const UNIX_EPOCH_JDATE: u64 = 2440588; // == date2j(1970, 1, 1)
+    const POSTGRES_EPOCH_JDATE: u64 = 2451545; // == date2j(2000, 1, 1)
     const SECS_PER_DAY: u64 = 86400;
     const USECS_PER_SEC: u64 = 1000000;
-    match time.duration_since(SystemTime::UNIX_EPOCH) {
-        Ok(n) => {
-            ((n.as_secs() - ((POSTGRES_EPOCH_JDATE - UNIX_EPOCH_JDATE) * SECS_PER_DAY))
-                * USECS_PER_SEC
-                + n.subsec_micros() as u64) as i64
+    const SECS_DIFF_UNIX_TO_POSTGRES_EPOCH: u64 =
+        (POSTGRES_EPOCH_JDATE - UNIX_EPOCH_JDATE) * SECS_PER_DAY;
+
+    pub fn to_pg_timestamp(time: SystemTime) -> TimestampTz {
+        match time.duration_since(SystemTime::UNIX_EPOCH) {
+            Ok(n) => {
+                ((n.as_secs() - SECS_DIFF_UNIX_TO_POSTGRES_EPOCH) * USECS_PER_SEC
+                    + n.subsec_micros() as u64) as i64
+            }
+            Err(_) => panic!("SystemTime before UNIX EPOCH!"),
         }
-        Err(_) => panic!("SystemTime before UNIX EPOCH!"),
+    }
+
+    pub fn from_pg_timestamp(time: TimestampTz) -> SystemTime {
+        let time: u64 = time
+            .try_into()
+            .expect("timestamp before millenium (postgres epoch)");
+        let since_unix_epoch = time + SECS_DIFF_UNIX_TO_POSTGRES_EPOCH * USECS_PER_SEC;
+        SystemTime::UNIX_EPOCH
+            .checked_add(Duration::from_micros(since_unix_epoch))
+            .expect("SystemTime overflow")
     }
 }
 
+pub use timestamp_conversions::{from_pg_timestamp, to_pg_timestamp};
+
 // Returns (aligned) end_lsn of the last record in data_dir with WAL segments.
 // start_lsn must point to some previously known record boundary (beginning of
 // the next record). If no valid record after is found, start_lsn is returned
@@ -481,4 +502,24 @@ pub fn encode_logical_message(prefix: &str, message: &str) -> Vec<u8> {
     wal
 }
 
-// If you need to craft WAL and write tests for this module, put it at wal_craft crate.
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_ts_conversion() {
+        let now = SystemTime::now();
+        let round_trip = from_pg_timestamp(to_pg_timestamp(now));
+
+        let now_since = now.duration_since(SystemTime::UNIX_EPOCH).unwrap();
+        let round_trip_since = round_trip.duration_since(SystemTime::UNIX_EPOCH).unwrap();
+        assert_eq!(now_since.as_micros(), round_trip_since.as_micros());
+
+        let now_pg = get_current_timestamp();
+        let round_trip_pg = to_pg_timestamp(from_pg_timestamp(now_pg));
+
+        assert_eq!(now_pg, round_trip_pg);
+    }
+
+    // If you need to craft WAL and write tests for this module, put it at wal_craft crate.
+}

libs/pq_proto/src/lib.rs

@@ -670,6 +670,7 @@ pub fn read_cstr(buf: &mut Bytes) -> Result<Bytes, ProtocolError> {
 }
 
 pub const SQLSTATE_INTERNAL_ERROR: &[u8; 5] = b"XX000";
+pub const SQLSTATE_ADMIN_SHUTDOWN: &[u8; 5] = b"57P01";
 pub const SQLSTATE_SUCCESSFUL_COMPLETION: &[u8; 5] = b"00000";
 
 impl<'a> BeMessage<'a> {

libs/remote_storage/Cargo.toml

@@ -13,6 +13,7 @@ aws-types.workspace = true
 aws-config.workspace = true
 aws-sdk-s3.workspace = true
 aws-credential-types.workspace = true
+bytes.workspace = true
 camino.workspace = true
 hyper = { workspace = true, features = ["stream"] }
 serde.workspace = true
@@ -26,6 +27,13 @@ metrics.workspace = true
 utils.workspace = true
 pin-project-lite.workspace = true
 workspace_hack.workspace = true
+azure_core.workspace = true
+azure_identity.workspace = true
+azure_storage.workspace = true
+azure_storage_blobs.workspace = true
+futures-util.workspace = true
+http-types.workspace = true
+itertools.workspace = true
 
 [dev-dependencies]
 camino-tempfile.workspace = true

libs/remote_storage/src/azure_blob.rs

@@ -0,0 +1,356 @@
+//! Azure Blob Storage wrapper
+
+use std::env;
+use std::num::NonZeroU32;
+use std::sync::Arc;
+use std::{borrow::Cow, collections::HashMap, io::Cursor};
+
+use super::REMOTE_STORAGE_PREFIX_SEPARATOR;
+use anyhow::Result;
+use azure_core::request_options::{MaxResults, Metadata, Range};
+use azure_core::Header;
+use azure_identity::DefaultAzureCredential;
+use azure_storage::StorageCredentials;
+use azure_storage_blobs::prelude::ClientBuilder;
+use azure_storage_blobs::{
+    blob::operations::GetBlobBuilder,
+    prelude::{BlobClient, ContainerClient},
+};
+use futures_util::StreamExt;
+use http_types::StatusCode;
+use tokio::io::AsyncRead;
+use tracing::debug;
+
+use crate::s3_bucket::RequestKind;
+use crate::{
+    AzureConfig, ConcurrencyLimiter, Download, DownloadError, RemotePath, RemoteStorage,
+    StorageMetadata,
+};
+
+pub struct AzureBlobStorage {
+    client: ContainerClient,
+    prefix_in_container: Option<String>,
+    max_keys_per_list_response: Option<NonZeroU32>,
+    concurrency_limiter: ConcurrencyLimiter,
+}
+
+impl AzureBlobStorage {
+    pub fn new(azure_config: &AzureConfig) -> Result<Self> {
+        debug!(
+            "Creating azure remote storage for azure container {}",
+            azure_config.container_name
+        );
+
+        let account = env::var("AZURE_STORAGE_ACCOUNT").expect("missing AZURE_STORAGE_ACCOUNT");
+
+        // If the `AZURE_STORAGE_ACCESS_KEY` env var has an access key, use that,
+        // otherwise try the token based credentials.
+        let credentials = if let Ok(access_key) = env::var("AZURE_STORAGE_ACCESS_KEY") {
+            StorageCredentials::access_key(account.clone(), access_key)
+        } else {
+            let token_credential = DefaultAzureCredential::default();
+            StorageCredentials::token_credential(Arc::new(token_credential))
+        };
+
+        let builder = ClientBuilder::new(account, credentials);
+
+        let client = builder.container_client(azure_config.container_name.to_owned());
+
+        let max_keys_per_list_response =
+            if let Some(limit) = azure_config.max_keys_per_list_response {
+                Some(
+                    NonZeroU32::new(limit as u32)
+                        .ok_or_else(|| anyhow::anyhow!("max_keys_per_list_response can't be 0"))?,
+                )
+            } else {
+                None
+            };
+
+        Ok(AzureBlobStorage {
+            client,
+            prefix_in_container: azure_config.prefix_in_container.to_owned(),
+            max_keys_per_list_response,
+            concurrency_limiter: ConcurrencyLimiter::new(azure_config.concurrency_limit.get()),
+        })
+    }
+
+    pub fn relative_path_to_name(&self, path: &RemotePath) -> String {
+        assert_eq!(std::path::MAIN_SEPARATOR, REMOTE_STORAGE_PREFIX_SEPARATOR);
+        let path_string = path
+            .get_path()
+            .as_str()
+            .trim_end_matches(REMOTE_STORAGE_PREFIX_SEPARATOR);
+        match &self.prefix_in_container {
+            Some(prefix) => {
+                if prefix.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR) {
+                    prefix.clone() + path_string
+                } else {
+                    format!("{prefix}{REMOTE_STORAGE_PREFIX_SEPARATOR}{path_string}")
+                }
+            }
+            None => path_string.to_string(),
+        }
+    }
+
+    fn name_to_relative_path(&self, key: &str) -> RemotePath {
+        let relative_path =
+            match key.strip_prefix(self.prefix_in_container.as_deref().unwrap_or_default()) {
+                Some(stripped) => stripped,
+                // we rely on Azure to return properly prefixed paths
+                // for requests with a certain prefix
+                None => panic!(
+                    "Key {key} does not start with container prefix {:?}",
+                    self.prefix_in_container
+                ),
+            };
+        RemotePath(
+            relative_path
+                .split(REMOTE_STORAGE_PREFIX_SEPARATOR)
+                .collect(),
+        )
+    }
+
+    async fn download_for_builder(
+        &self,
+        metadata: StorageMetadata,
+        builder: GetBlobBuilder,
+    ) -> Result<Download, DownloadError> {
+        let mut response = builder.into_stream();
+
+        // TODO give proper streaming response instead of buffering into RAM
+        // https://github.com/neondatabase/neon/issues/5563
+        let mut buf = Vec::new();
+        while let Some(part) = response.next().await {
+            let part = part.map_err(to_download_error)?;
+            let data = part
+                .data
+                .collect()
+                .await
+                .map_err(|e| DownloadError::Other(e.into()))?;
+            buf.extend_from_slice(&data.slice(..));
+        }
+        Ok(Download {
+            download_stream: Box::pin(Cursor::new(buf)),
+            metadata: Some(metadata),
+        })
+    }
+    // TODO get rid of this function once we have metadata included in the response
+    // https://github.com/Azure/azure-sdk-for-rust/issues/1439
+    async fn get_metadata(
+        &self,
+        blob_client: &BlobClient,
+    ) -> Result<StorageMetadata, DownloadError> {
+        let builder = blob_client.get_metadata();
+
+        let response = builder.into_future().await.map_err(to_download_error)?;
+        let mut map = HashMap::new();
+
+        for md in response.metadata.iter() {
+            map.insert(
+                md.name().as_str().to_string(),
+                md.value().as_str().to_string(),
+            );
+        }
+        Ok(StorageMetadata(map))
+    }
+
+    async fn permit(&self, kind: RequestKind) -> tokio::sync::SemaphorePermit<'_> {
+        self.concurrency_limiter
+            .acquire(kind)
+            .await
+            .expect("semaphore is never closed")
+    }
+}
+
+fn to_azure_metadata(metadata: StorageMetadata) -> Metadata {
+    let mut res = Metadata::new();
+    for (k, v) in metadata.0.into_iter() {
+        res.insert(k, v);
+    }
+    res
+}
+
+fn to_download_error(error: azure_core::Error) -> DownloadError {
+    if let Some(http_err) = error.as_http_error() {
+        match http_err.status() {
+            StatusCode::NotFound => DownloadError::NotFound,
+            StatusCode::BadRequest => DownloadError::BadInput(anyhow::Error::new(error)),
+            _ => DownloadError::Other(anyhow::Error::new(error)),
+        }
+    } else {
+        DownloadError::Other(error.into())
+    }
+}
+
+#[async_trait::async_trait]
+impl RemoteStorage for AzureBlobStorage {
+    async fn list_prefixes(
+        &self,
+        prefix: Option<&RemotePath>,
+    ) -> Result<Vec<RemotePath>, DownloadError> {
+        // get the passed prefix or if it is not set use prefix_in_bucket value
+        let list_prefix = prefix
+            .map(|p| self.relative_path_to_name(p))
+            .or_else(|| self.prefix_in_container.clone())
+            .map(|mut p| {
+                // required to end with a separator
+                // otherwise request will return only the entry of a prefix
+                if !p.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR) {
+                    p.push(REMOTE_STORAGE_PREFIX_SEPARATOR);
+                }
+                p
+            });
+
+        let mut builder = self
+            .client
+            .list_blobs()
+            .delimiter(REMOTE_STORAGE_PREFIX_SEPARATOR.to_string());
+
+        if let Some(prefix) = list_prefix {
+            builder = builder.prefix(Cow::from(prefix.to_owned()));
+        }
+
+        if let Some(limit) = self.max_keys_per_list_response {
+            builder = builder.max_results(MaxResults::new(limit));
+        }
+
+        let mut response = builder.into_stream();
+        let mut res = Vec::new();
+        while let Some(entry) = response.next().await {
+            let entry = entry.map_err(to_download_error)?;
+            let name_iter = entry
+                .blobs
+                .prefixes()
+                .map(|prefix| self.name_to_relative_path(&prefix.name));
+            res.extend(name_iter);
+        }
+        Ok(res)
+    }
+
+    async fn list_files(&self, folder: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>> {
+        let folder_name = folder
+            .map(|p| self.relative_path_to_name(p))
+            .or_else(|| self.prefix_in_container.clone());
+
+        let mut builder = self.client.list_blobs();
+
+        if let Some(folder_name) = folder_name {
+            builder = builder.prefix(Cow::from(folder_name.to_owned()));
+        }
+
+        if let Some(limit) = self.max_keys_per_list_response {
+            builder = builder.max_results(MaxResults::new(limit));
+        }
+
+        let mut response = builder.into_stream();
+        let mut res = Vec::new();
+        while let Some(l) = response.next().await {
+            let entry = l.map_err(anyhow::Error::new)?;
+            let name_iter = entry
+                .blobs
+                .blobs()
+                .map(|bl| self.name_to_relative_path(&bl.name));
+            res.extend(name_iter);
+        }
+        Ok(res)
+    }
+
+    async fn upload(
+        &self,
+        mut from: impl AsyncRead + Unpin + Send + Sync + 'static,
+        data_size_bytes: usize,
+        to: &RemotePath,
+        metadata: Option<StorageMetadata>,
+    ) -> anyhow::Result<()> {
+        let _permit = self.permit(RequestKind::Put).await;
+        let blob_client = self.client.blob_client(self.relative_path_to_name(to));
+
+        // TODO FIX THIS UGLY HACK and don't buffer the entire object
+        // into RAM here, but use the streaming interface. For that,
+        // we'd have to change the interface though...
+        // https://github.com/neondatabase/neon/issues/5563
+        let mut buf = Vec::with_capacity(data_size_bytes);
+        tokio::io::copy(&mut from, &mut buf).await?;
+        let body = azure_core::Body::Bytes(buf.into());
+
+        let mut builder = blob_client.put_block_blob(body);
+
+        if let Some(metadata) = metadata {
+            builder = builder.metadata(to_azure_metadata(metadata));
+        }
+
+        let _response = builder.into_future().await?;
+
+        Ok(())
+    }
+
+    async fn download(&self, from: &RemotePath) -> Result<Download, DownloadError> {
+        let _permit = self.permit(RequestKind::Get).await;
+        let blob_client = self.client.blob_client(self.relative_path_to_name(from));
+
+        let metadata = self.get_metadata(&blob_client).await?;
+
+        let builder = blob_client.get();
+
+        self.download_for_builder(metadata, builder).await
+    }
+
+    async fn download_byte_range(
+        &self,
+        from: &RemotePath,
+        start_inclusive: u64,
+        end_exclusive: Option<u64>,
+    ) -> Result<Download, DownloadError> {
+        let _permit = self.permit(RequestKind::Get).await;
+        let blob_client = self.client.blob_client(self.relative_path_to_name(from));
+
+        let metadata = self.get_metadata(&blob_client).await?;
+
+        let mut builder = blob_client.get();
+
+        if let Some(end_exclusive) = end_exclusive {
+            builder = builder.range(Range::new(start_inclusive, end_exclusive));
+        } else {
+            // Open ranges are not supported by the SDK so we work around
+            // by setting the upper limit extremely high (but high enough
+            // to still be representable by signed 64 bit integers).
+            // TODO remove workaround once the SDK adds open range support
+            // https://github.com/Azure/azure-sdk-for-rust/issues/1438
+            let end_exclusive = u64::MAX / 4;
+            builder = builder.range(Range::new(start_inclusive, end_exclusive));
+        }
+
+        self.download_for_builder(metadata, builder).await
+    }
+
+    async fn delete(&self, path: &RemotePath) -> anyhow::Result<()> {
+        let _permit = self.permit(RequestKind::Delete).await;
+        let blob_client = self.client.blob_client(self.relative_path_to_name(path));
+
+        let builder = blob_client.delete();
+
+        match builder.into_future().await {
+            Ok(_response) => Ok(()),
+            Err(e) => {
+                if let Some(http_err) = e.as_http_error() {
+                    if http_err.status() == StatusCode::NotFound {
+                        return Ok(());
+                    }
+                }
+                Err(anyhow::Error::new(e))
+            }
+        }
+    }
+
+    async fn delete_objects<'a>(&self, paths: &'a [RemotePath]) -> anyhow::Result<()> {
+        // Permit is already obtained by inner delete function
+
+        // TODO batch requests are also not supported by the SDK
+        // https://github.com/Azure/azure-sdk-for-rust/issues/1068
+        // https://github.com/Azure/azure-sdk-for-rust/issues/1249
+        for path in paths {
+            self.delete(path).await?;
+        }
+        Ok(())
+    }
+}

libs/remote_storage/src/lib.rs

@@ -4,7 +4,10 @@
 //! [`RemoteStorage`] trait a CRUD-like generic abstraction to use for adapting external storages with a few implementations:
 //!   * [`local_fs`] allows to use local file system as an external storage
 //!   * [`s3_bucket`] uses AWS S3 bucket as an external storage
+//!   * [`azure_blob`] allows to use Azure Blob storage as an external storage
 //!
+
+mod azure_blob;
 mod local_fs;
 mod s3_bucket;
 mod simulate_failures;
@@ -21,11 +24,15 @@ use anyhow::{bail, Context};
 use camino::{Utf8Path, Utf8PathBuf};
 
 use serde::{Deserialize, Serialize};
-use tokio::io;
+use tokio::{io, sync::Semaphore};
 use toml_edit::Item;
 use tracing::info;
 
-pub use self::{local_fs::LocalFs, s3_bucket::S3Bucket, simulate_failures::UnreliableWrapper};
+pub use self::{
+    azure_blob::AzureBlobStorage, local_fs::LocalFs, s3_bucket::S3Bucket,
+    simulate_failures::UnreliableWrapper,
+};
+use s3_bucket::RequestKind;
 
 /// How many different timelines can be processed simultaneously when synchronizing layers with the remote storage.
 /// During regular work, pageserver produces one layer file per timeline checkpoint, with bursts of concurrency
@@ -39,6 +46,11 @@ pub const DEFAULT_REMOTE_STORAGE_MAX_SYNC_ERRORS: u32 = 10;
 /// ~3500 PUT/COPY/POST/DELETE or 5500 GET/HEAD S3 requests
 /// <https://aws.amazon.com/premiumsupport/knowledge-center/s3-request-limit-avoid-throttling/>
 pub const DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT: usize = 100;
+/// We set this a little bit low as we currently buffer the entire file into RAM
+///
+/// Here, a limit of max 20k concurrent connections was noted.
+/// <https://learn.microsoft.com/en-us/answers/questions/1301863/is-there-any-limitation-to-concurrent-connections>
+pub const DEFAULT_REMOTE_STORAGE_AZURE_CONCURRENCY_LIMIT: usize = 30;
 /// No limits on the client side, which currenltly means 1000 for AWS S3.
 /// <https://docs.aws.amazon.com/AmazonS3/latest/API/API_ListObjectsV2.html#API_ListObjectsV2_RequestSyntax>
 pub const DEFAULT_MAX_KEYS_PER_LIST_RESPONSE: Option<i32> = None;
@@ -100,7 +112,7 @@ impl RemotePath {
         self.0.file_name()
     }
 
-    pub fn join(&self, segment: &Utf8Path) -> Self {
+    pub fn join<P: AsRef<Utf8Path>>(&self, segment: P) -> Self {
         Self(self.0.join(segment))
     }
 
@@ -217,6 +229,7 @@ impl std::error::Error for DownloadError {}
 pub enum GenericRemoteStorage {
     LocalFs(LocalFs),
     AwsS3(Arc<S3Bucket>),
+    AzureBlob(Arc<AzureBlobStorage>),
     Unreliable(Arc<UnreliableWrapper>),
 }
 
@@ -228,6 +241,7 @@ impl GenericRemoteStorage {
         match self {
             Self::LocalFs(s) => s.list_files(folder).await,
             Self::AwsS3(s) => s.list_files(folder).await,
+            Self::AzureBlob(s) => s.list_files(folder).await,
             Self::Unreliable(s) => s.list_files(folder).await,
         }
     }
@@ -242,6 +256,7 @@ impl GenericRemoteStorage {
         match self {
             Self::LocalFs(s) => s.list_prefixes(prefix).await,
             Self::AwsS3(s) => s.list_prefixes(prefix).await,
+            Self::AzureBlob(s) => s.list_prefixes(prefix).await,
             Self::Unreliable(s) => s.list_prefixes(prefix).await,
         }
     }
@@ -256,6 +271,7 @@ impl GenericRemoteStorage {
         match self {
             Self::LocalFs(s) => s.upload(from, data_size_bytes, to, metadata).await,
             Self::AwsS3(s) => s.upload(from, data_size_bytes, to, metadata).await,
+            Self::AzureBlob(s) => s.upload(from, data_size_bytes, to, metadata).await,
             Self::Unreliable(s) => s.upload(from, data_size_bytes, to, metadata).await,
         }
     }
@@ -264,6 +280,7 @@ impl GenericRemoteStorage {
         match self {
             Self::LocalFs(s) => s.download(from).await,
             Self::AwsS3(s) => s.download(from).await,
+            Self::AzureBlob(s) => s.download(from).await,
             Self::Unreliable(s) => s.download(from).await,
         }
     }
@@ -283,6 +300,10 @@ impl GenericRemoteStorage {
                 s.download_byte_range(from, start_inclusive, end_exclusive)
                     .await
             }
+            Self::AzureBlob(s) => {
+                s.download_byte_range(from, start_inclusive, end_exclusive)
+                    .await
+            }
             Self::Unreliable(s) => {
                 s.download_byte_range(from, start_inclusive, end_exclusive)
                     .await
@@ -294,6 +315,7 @@ impl GenericRemoteStorage {
         match self {
             Self::LocalFs(s) => s.delete(path).await,
             Self::AwsS3(s) => s.delete(path).await,
+            Self::AzureBlob(s) => s.delete(path).await,
             Self::Unreliable(s) => s.delete(path).await,
         }
     }
@@ -302,6 +324,7 @@ impl GenericRemoteStorage {
         match self {
             Self::LocalFs(s) => s.delete_objects(paths).await,
             Self::AwsS3(s) => s.delete_objects(paths).await,
+            Self::AzureBlob(s) => s.delete_objects(paths).await,
             Self::Unreliable(s) => s.delete_objects(paths).await,
         }
     }
@@ -319,6 +342,11 @@ impl GenericRemoteStorage {
                       s3_config.bucket_name, s3_config.bucket_region, s3_config.prefix_in_bucket, s3_config.endpoint);
                 Self::AwsS3(Arc::new(S3Bucket::new(s3_config)?))
             }
+            RemoteStorageKind::AzureContainer(azure_config) => {
+                info!("Using azure container '{}' in region '{}' as a remote storage, prefix in container: '{:?}'",
+                      azure_config.container_name, azure_config.container_region, azure_config.prefix_in_container);
+                Self::AzureBlob(Arc::new(AzureBlobStorage::new(azure_config)?))
+            }
         })
     }
 
@@ -383,6 +411,9 @@ pub enum RemoteStorageKind {
     /// AWS S3 based storage, storing all files in the S3 bucket
     /// specified by the config
     AwsS3(S3Config),
+    /// Azure Blob based storage, storing all files in the container
+    /// specified by the config
+    AzureContainer(AzureConfig),
 }
 
 /// AWS S3 bucket coordinates and access credentials to manage the bucket contents (read and write).
@@ -422,11 +453,45 @@ impl Debug for S3Config {
     }
 }
 
+/// Azure  bucket coordinates and access credentials to manage the bucket contents (read and write).
+#[derive(Clone, PartialEq, Eq)]
+pub struct AzureConfig {
+    /// Name of the container to connect to.
+    pub container_name: String,
+    /// The region where the bucket is located at.
+    pub container_region: String,
+    /// A "subfolder" in the container, to use the same container separately by multiple remote storage users at once.
+    pub prefix_in_container: Option<String>,
+    /// Azure has various limits on its API calls, we need not to exceed those.
+    /// See [`DEFAULT_REMOTE_STORAGE_AZURE_CONCURRENCY_LIMIT`] for more details.
+    pub concurrency_limit: NonZeroUsize,
+    pub max_keys_per_list_response: Option<i32>,
+}
+
+impl Debug for AzureConfig {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("AzureConfig")
+            .field("bucket_name", &self.container_name)
+            .field("bucket_region", &self.container_region)
+            .field("prefix_in_bucket", &self.prefix_in_container)
+            .field("concurrency_limit", &self.concurrency_limit)
+            .field(
+                "max_keys_per_list_response",
+                &self.max_keys_per_list_response,
+            )
+            .finish()
+    }
+}
+
 impl RemoteStorageConfig {
     pub fn from_toml(toml: &toml_edit::Item) -> anyhow::Result<Option<RemoteStorageConfig>> {
         let local_path = toml.get("local_path");
         let bucket_name = toml.get("bucket_name");
         let bucket_region = toml.get("bucket_region");
+        let container_name = toml.get("container_name");
+        let container_region = toml.get("container_region");
+
+        let use_azure = container_name.is_some() && container_region.is_some();
 
         let max_concurrent_syncs = NonZeroUsize::new(
             parse_optional_integer("max_concurrent_syncs", toml)?
@@ -440,9 +505,13 @@ impl RemoteStorageConfig {
         )
         .context("Failed to parse 'max_sync_errors' as a positive integer")?;
 
+        let default_concurrency_limit = if use_azure {
+            DEFAULT_REMOTE_STORAGE_AZURE_CONCURRENCY_LIMIT
+        } else {
+            DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT
+        };
         let concurrency_limit = NonZeroUsize::new(
-            parse_optional_integer("concurrency_limit", toml)?
-                .unwrap_or(DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT),
+            parse_optional_integer("concurrency_limit", toml)?.unwrap_or(default_concurrency_limit),
         )
         .context("Failed to parse 'concurrency_limit' as a positive integer")?;
 
@@ -451,33 +520,70 @@ impl RemoteStorageConfig {
                 .context("Failed to parse 'max_keys_per_list_response' as a positive integer")?
                 .or(DEFAULT_MAX_KEYS_PER_LIST_RESPONSE);
 
-        let storage = match (local_path, bucket_name, bucket_region) {
+        let endpoint = toml
+            .get("endpoint")
+            .map(|endpoint| parse_toml_string("endpoint", endpoint))
+            .transpose()?;
+
+        let storage = match (
+            local_path,
+            bucket_name,
+            bucket_region,
+            container_name,
+            container_region,
+        ) {
             // no 'local_path' nor 'bucket_name' options are provided, consider this remote storage disabled
-            (None, None, None) => return Ok(None),
-            (_, Some(_), None) => {
+            (None, None, None, None, None) => return Ok(None),
+            (_, Some(_), None, ..) => {
                 bail!("'bucket_region' option is mandatory if 'bucket_name' is given ")
             }
-            (_, None, Some(_)) => {
+            (_, None, Some(_), ..) => {
                 bail!("'bucket_name' option is mandatory if 'bucket_region' is given ")
             }
-            (None, Some(bucket_name), Some(bucket_region)) => RemoteStorageKind::AwsS3(S3Config {
-                bucket_name: parse_toml_string("bucket_name", bucket_name)?,
-                bucket_region: parse_toml_string("bucket_region", bucket_region)?,
-                prefix_in_bucket: toml
-                    .get("prefix_in_bucket")
-                    .map(|prefix_in_bucket| parse_toml_string("prefix_in_bucket", prefix_in_bucket))
-                    .transpose()?,
-                endpoint: toml
-                    .get("endpoint")
-                    .map(|endpoint| parse_toml_string("endpoint", endpoint))
-                    .transpose()?,
-                concurrency_limit,
-                max_keys_per_list_response,
-            }),
-            (Some(local_path), None, None) => RemoteStorageKind::LocalFs(Utf8PathBuf::from(
-                parse_toml_string("local_path", local_path)?,
-            )),
-            (Some(_), Some(_), _) => bail!("local_path and bucket_name are mutually exclusive"),
+            (None, Some(bucket_name), Some(bucket_region), ..) => {
+                RemoteStorageKind::AwsS3(S3Config {
+                    bucket_name: parse_toml_string("bucket_name", bucket_name)?,
+                    bucket_region: parse_toml_string("bucket_region", bucket_region)?,
+                    prefix_in_bucket: toml
+                        .get("prefix_in_bucket")
+                        .map(|prefix_in_bucket| {
+                            parse_toml_string("prefix_in_bucket", prefix_in_bucket)
+                        })
+                        .transpose()?,
+                    endpoint,
+                    concurrency_limit,
+                    max_keys_per_list_response,
+                })
+            }
+            (_, _, _, Some(_), None) => {
+                bail!("'container_name' option is mandatory if 'container_region' is given ")
+            }
+            (_, _, _, None, Some(_)) => {
+                bail!("'container_name' option is mandatory if 'container_region' is given ")
+            }
+            (None, None, None, Some(container_name), Some(container_region)) => {
+                RemoteStorageKind::AzureContainer(AzureConfig {
+                    container_name: parse_toml_string("container_name", container_name)?,
+                    container_region: parse_toml_string("container_region", container_region)?,
+                    prefix_in_container: toml
+                        .get("prefix_in_container")
+                        .map(|prefix_in_container| {
+                            parse_toml_string("prefix_in_container", prefix_in_container)
+                        })
+                        .transpose()?,
+                    concurrency_limit,
+                    max_keys_per_list_response,
+                })
+            }
+            (Some(local_path), None, None, None, None) => RemoteStorageKind::LocalFs(
+                Utf8PathBuf::from(parse_toml_string("local_path", local_path)?),
+            ),
+            (Some(_), Some(_), ..) => {
+                bail!("'local_path' and 'bucket_name' are mutually exclusive")
+            }
+            (Some(_), _, _, Some(_), Some(_)) => {
+                bail!("local_path and 'container_name' are mutually exclusive")
+            }
         };
 
         Ok(Some(RemoteStorageConfig {
@@ -513,6 +619,46 @@ fn parse_toml_string(name: &str, item: &Item) -> anyhow::Result<String> {
     Ok(s.to_string())
 }
 
+struct ConcurrencyLimiter {
+    // Every request to S3 can be throttled or cancelled, if a certain number of requests per second is exceeded.
+    // Same goes to IAM, which is queried before every S3 request, if enabled. IAM has even lower RPS threshold.
+    // The helps to ensure we don't exceed the thresholds.
+    write: Arc<Semaphore>,
+    read: Arc<Semaphore>,
+}
+
+impl ConcurrencyLimiter {
+    fn for_kind(&self, kind: RequestKind) -> &Arc<Semaphore> {
+        match kind {
+            RequestKind::Get => &self.read,
+            RequestKind::Put => &self.write,
+            RequestKind::List => &self.read,
+            RequestKind::Delete => &self.write,
+        }
+    }
+
+    async fn acquire(
+        &self,
+        kind: RequestKind,
+    ) -> Result<tokio::sync::SemaphorePermit<'_>, tokio::sync::AcquireError> {
+        self.for_kind(kind).acquire().await
+    }
+
+    async fn acquire_owned(
+        &self,
+        kind: RequestKind,
+    ) -> Result<tokio::sync::OwnedSemaphorePermit, tokio::sync::AcquireError> {
+        Arc::clone(self.for_kind(kind)).acquire_owned().await
+    }
+
+    fn new(limit: usize) -> ConcurrencyLimiter {
+        Self {
+            read: Arc::new(Semaphore::new(limit)),
+            write: Arc::new(Semaphore::new(limit)),
+        }
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;

libs/remote_storage/src/s3_bucket.rs

@@ -4,7 +4,7 @@
 //! allowing multiple api users to independently work with the same S3 bucket, if
 //! their bucket prefixes are both specified and different.
 
-use std::sync::Arc;
+use std::borrow::Cow;
 
 use anyhow::Context;
 use aws_config::{
@@ -24,22 +24,20 @@ use aws_sdk_s3::{
 use aws_smithy_http::body::SdkBody;
 use hyper::Body;
 use scopeguard::ScopeGuard;
-use tokio::{
-    io::{self, AsyncRead},
-    sync::Semaphore,
-};
+use tokio::io::{self, AsyncRead};
 use tokio_util::io::ReaderStream;
 use tracing::debug;
 
 use super::StorageMetadata;
 use crate::{
-    Download, DownloadError, RemotePath, RemoteStorage, S3Config, MAX_KEYS_PER_DELETE,
-    REMOTE_STORAGE_PREFIX_SEPARATOR,
+    ConcurrencyLimiter, Download, DownloadError, RemotePath, RemoteStorage, S3Config,
+    MAX_KEYS_PER_DELETE, REMOTE_STORAGE_PREFIX_SEPARATOR,
 };
 
 pub(super) mod metrics;
 
-use self::metrics::{AttemptOutcome, RequestKind};
+use self::metrics::AttemptOutcome;
+pub(super) use self::metrics::RequestKind;
 
 /// AWS S3 storage.
 pub struct S3Bucket {
@@ -50,46 +48,6 @@ pub struct S3Bucket {
     concurrency_limiter: ConcurrencyLimiter,
 }
 
-struct ConcurrencyLimiter {
-    // Every request to S3 can be throttled or cancelled, if a certain number of requests per second is exceeded.
-    // Same goes to IAM, which is queried before every S3 request, if enabled. IAM has even lower RPS threshold.
-    // The helps to ensure we don't exceed the thresholds.
-    write: Arc<Semaphore>,
-    read: Arc<Semaphore>,
-}
-
-impl ConcurrencyLimiter {
-    fn for_kind(&self, kind: RequestKind) -> &Arc<Semaphore> {
-        match kind {
-            RequestKind::Get => &self.read,
-            RequestKind::Put => &self.write,
-            RequestKind::List => &self.read,
-            RequestKind::Delete => &self.write,
-        }
-    }
-
-    async fn acquire(
-        &self,
-        kind: RequestKind,
-    ) -> Result<tokio::sync::SemaphorePermit<'_>, tokio::sync::AcquireError> {
-        self.for_kind(kind).acquire().await
-    }
-
-    async fn acquire_owned(
-        &self,
-        kind: RequestKind,
-    ) -> Result<tokio::sync::OwnedSemaphorePermit, tokio::sync::AcquireError> {
-        Arc::clone(self.for_kind(kind)).acquire_owned().await
-    }
-
-    fn new(limit: usize) -> ConcurrencyLimiter {
-        Self {
-            read: Arc::new(Semaphore::new(limit)),
-            write: Arc::new(Semaphore::new(limit)),
-        }
-    }
-}
-
 #[derive(Default)]
 struct GetObjectRequest {
     bucket: String,
@@ -556,6 +514,20 @@ impl RemoteStorage for S3Bucket {
                         .deleted_objects_total
                         .inc_by(chunk.len() as u64);
                     if let Some(errors) = resp.errors {
+                        // Log a bounded number of the errors within the response:
+                        // these requests can carry 1000 keys so logging each one
+                        // would be too verbose, especially as errors may lead us
+                        // to retry repeatedly.
+                        const LOG_UP_TO_N_ERRORS: usize = 10;
+                        for e in errors.iter().take(LOG_UP_TO_N_ERRORS) {
+                            tracing::warn!(
+                                "DeleteObjects key {} failed: {}: {}",
+                                e.key.as_ref().map(Cow::from).unwrap_or("".into()),
+                                e.code.as_ref().map(Cow::from).unwrap_or("".into()),
+                                e.message.as_ref().map(Cow::from).unwrap_or("".into())
+                            );
+                        }
+
                         return Err(anyhow::format_err!(
                             "Failed to delete {} objects",
                             errors.len()

libs/remote_storage/src/s3_bucket/metrics.rs

@@ -6,7 +6,7 @@ use once_cell::sync::Lazy;
 pub(super) static BUCKET_METRICS: Lazy<BucketMetrics> = Lazy::new(Default::default);
 
 #[derive(Clone, Copy, Debug)]
-pub(super) enum RequestKind {
+pub(crate) enum RequestKind {
     Get = 0,
     Put = 1,
     Delete = 2,

libs/remote_storage/tests/test_real_azure.rs

@@ -0,0 +1,625 @@
+use std::collections::HashSet;
+use std::env;
+use std::num::{NonZeroU32, NonZeroUsize};
+use std::ops::ControlFlow;
+use std::path::PathBuf;
+use std::sync::Arc;
+use std::time::UNIX_EPOCH;
+
+use anyhow::Context;
+use camino::Utf8Path;
+use once_cell::sync::OnceCell;
+use remote_storage::{
+    AzureConfig, Download, GenericRemoteStorage, RemotePath, RemoteStorageConfig, RemoteStorageKind,
+};
+use test_context::{test_context, AsyncTestContext};
+use tokio::task::JoinSet;
+use tracing::{debug, error, info};
+
+static LOGGING_DONE: OnceCell<()> = OnceCell::new();
+
+const ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME: &str = "ENABLE_REAL_AZURE_REMOTE_STORAGE";
+
+const BASE_PREFIX: &str = "test";
+
+/// Tests that the Azure client can list all prefixes, even if the response comes paginated and requires multiple HTTP queries.
+/// Uses real Azure and requires [`ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME`] and related Azure cred env vars specified.
+/// See the client creation in [`create_azure_client`] for details on the required env vars.
+/// If real Azure tests are disabled, the test passes, skipping any real test run: currently, there's no way to mark the test ignored in runtime with the
+/// deafult test framework, see https://github.com/rust-lang/rust/issues/68007 for details.
+///
+/// First, the test creates a set of Azure blobs with keys `/${random_prefix_part}/${base_prefix_str}/sub_prefix_${i}/blob_${i}` in [`upload_azure_data`]
+/// where
+/// * `random_prefix_part` is set for the entire Azure client during the Azure client creation in [`create_azure_client`], to avoid multiple test runs interference
+/// * `base_prefix_str` is a common prefix to use in the client requests: we would want to ensure that the client is able to list nested prefixes inside the bucket
+///
+/// Then, verifies that the client does return correct prefixes when queried:
+/// * with no prefix, it lists everything after its `${random_prefix_part}/` — that should be `${base_prefix_str}` value only
+/// * with `${base_prefix_str}/` prefix, it lists every `sub_prefix_${i}`
+///
+/// With the real Azure enabled and `#[cfg(test)]` Rust configuration used, the Azure client test adds a `max-keys` param to limit the response keys.
+/// This way, we are able to test the pagination implicitly, by ensuring all results are returned from the remote storage and avoid uploading too many blobs to Azure.
+///
+/// Lastly, the test attempts to clean up and remove all uploaded Azure files.
+/// If any errors appear during the clean up, they get logged, but the test is not failed or stopped until clean up is finished.
+#[test_context(MaybeEnabledAzureWithTestBlobs)]
+#[tokio::test]
+async fn azure_pagination_should_work(
+    ctx: &mut MaybeEnabledAzureWithTestBlobs,
+) -> anyhow::Result<()> {
+    let ctx = match ctx {
+        MaybeEnabledAzureWithTestBlobs::Enabled(ctx) => ctx,
+        MaybeEnabledAzureWithTestBlobs::Disabled => return Ok(()),
+        MaybeEnabledAzureWithTestBlobs::UploadsFailed(e, _) => {
+            anyhow::bail!("Azure init failed: {e:?}")
+        }
+    };
+
+    let test_client = Arc::clone(&ctx.enabled.client);
+    let expected_remote_prefixes = ctx.remote_prefixes.clone();
+
+    let base_prefix = RemotePath::new(Utf8Path::new(ctx.enabled.base_prefix))
+        .context("common_prefix construction")?;
+    let root_remote_prefixes = test_client
+        .list_prefixes(None)
+        .await
+        .context("client list root prefixes failure")?
+        .into_iter()
+        .collect::<HashSet<_>>();
+    assert_eq!(
+        root_remote_prefixes, HashSet::from([base_prefix.clone()]),
+        "remote storage root prefixes list mismatches with the uploads. Returned prefixes: {root_remote_prefixes:?}"
+    );
+
+    let nested_remote_prefixes = test_client
+        .list_prefixes(Some(&base_prefix))
+        .await
+        .context("client list nested prefixes failure")?
+        .into_iter()
+        .collect::<HashSet<_>>();
+    let remote_only_prefixes = nested_remote_prefixes
+        .difference(&expected_remote_prefixes)
+        .collect::<HashSet<_>>();
+    let missing_uploaded_prefixes = expected_remote_prefixes
+        .difference(&nested_remote_prefixes)
+        .collect::<HashSet<_>>();
+    assert_eq!(
+        remote_only_prefixes.len() + missing_uploaded_prefixes.len(), 0,
+        "remote storage nested prefixes list mismatches with the uploads. Remote only prefixes: {remote_only_prefixes:?}, missing uploaded prefixes: {missing_uploaded_prefixes:?}",
+    );
+
+    Ok(())
+}
+
+/// Tests that Azure client can list all files in a folder, even if the response comes paginated and requirees multiple Azure queries.
+/// Uses real Azure and requires [`ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME`] and related Azure cred env vars specified. Test will skip real code and pass if env vars not set.
+/// See `Azure_pagination_should_work` for more information.
+///
+/// First, create a set of Azure objects with keys `random_prefix/folder{j}/blob_{i}.txt` in [`upload_azure_data`]
+/// Then performs the following queries:
+///    1. `list_files(None)`. This should return all files `random_prefix/folder{j}/blob_{i}.txt`
+///    2. `list_files("folder1")`.  This  should return all files `random_prefix/folder1/blob_{i}.txt`
+#[test_context(MaybeEnabledAzureWithSimpleTestBlobs)]
+#[tokio::test]
+async fn azure_list_files_works(
+    ctx: &mut MaybeEnabledAzureWithSimpleTestBlobs,
+) -> anyhow::Result<()> {
+    let ctx = match ctx {
+        MaybeEnabledAzureWithSimpleTestBlobs::Enabled(ctx) => ctx,
+        MaybeEnabledAzureWithSimpleTestBlobs::Disabled => return Ok(()),
+        MaybeEnabledAzureWithSimpleTestBlobs::UploadsFailed(e, _) => {
+            anyhow::bail!("Azure init failed: {e:?}")
+        }
+    };
+    let test_client = Arc::clone(&ctx.enabled.client);
+    let base_prefix =
+        RemotePath::new(Utf8Path::new("folder1")).context("common_prefix construction")?;
+    let root_files = test_client
+        .list_files(None)
+        .await
+        .context("client list root files failure")?
+        .into_iter()
+        .collect::<HashSet<_>>();
+    assert_eq!(
+        root_files,
+        ctx.remote_blobs.clone(),
+        "remote storage list_files on root mismatches with the uploads."
+    );
+    let nested_remote_files = test_client
+        .list_files(Some(&base_prefix))
+        .await
+        .context("client list nested files failure")?
+        .into_iter()
+        .collect::<HashSet<_>>();
+    let trim_remote_blobs: HashSet<_> = ctx
+        .remote_blobs
+        .iter()
+        .map(|x| x.get_path())
+        .filter(|x| x.starts_with("folder1"))
+        .map(|x| RemotePath::new(x).expect("must be valid path"))
+        .collect();
+    assert_eq!(
+        nested_remote_files, trim_remote_blobs,
+        "remote storage list_files on subdirrectory mismatches with the uploads."
+    );
+    Ok(())
+}
+
+#[test_context(MaybeEnabledAzure)]
+#[tokio::test]
+async fn azure_delete_non_exising_works(ctx: &mut MaybeEnabledAzure) -> anyhow::Result<()> {
+    let ctx = match ctx {
+        MaybeEnabledAzure::Enabled(ctx) => ctx,
+        MaybeEnabledAzure::Disabled => return Ok(()),
+    };
+
+    let path = RemotePath::new(Utf8Path::new(
+        format!("{}/for_sure_there_is_nothing_there_really", ctx.base_prefix).as_str(),
+    ))
+    .with_context(|| "RemotePath conversion")?;
+
+    ctx.client.delete(&path).await.expect("should succeed");
+
+    Ok(())
+}
+
+#[test_context(MaybeEnabledAzure)]
+#[tokio::test]
+async fn azure_delete_objects_works(ctx: &mut MaybeEnabledAzure) -> anyhow::Result<()> {
+    let ctx = match ctx {
+        MaybeEnabledAzure::Enabled(ctx) => ctx,
+        MaybeEnabledAzure::Disabled => return Ok(()),
+    };
+
+    let path1 = RemotePath::new(Utf8Path::new(format!("{}/path1", ctx.base_prefix).as_str()))
+        .with_context(|| "RemotePath conversion")?;
+
+    let path2 = RemotePath::new(Utf8Path::new(format!("{}/path2", ctx.base_prefix).as_str()))
+        .with_context(|| "RemotePath conversion")?;
+
+    let path3 = RemotePath::new(Utf8Path::new(format!("{}/path3", ctx.base_prefix).as_str()))
+        .with_context(|| "RemotePath conversion")?;
+
+    let data1 = "remote blob data1".as_bytes();
+    let data1_len = data1.len();
+    let data2 = "remote blob data2".as_bytes();
+    let data2_len = data2.len();
+    let data3 = "remote blob data3".as_bytes();
+    let data3_len = data3.len();
+    ctx.client
+        .upload(std::io::Cursor::new(data1), data1_len, &path1, None)
+        .await?;
+
+    ctx.client
+        .upload(std::io::Cursor::new(data2), data2_len, &path2, None)
+        .await?;
+
+    ctx.client
+        .upload(std::io::Cursor::new(data3), data3_len, &path3, None)
+        .await?;
+
+    ctx.client.delete_objects(&[path1, path2]).await?;
+
+    let prefixes = ctx.client.list_prefixes(None).await?;
+
+    assert_eq!(prefixes.len(), 1);
+
+    ctx.client.delete_objects(&[path3]).await?;
+
+    Ok(())
+}
+
+#[test_context(MaybeEnabledAzure)]
+#[tokio::test]
+async fn azure_upload_download_works(ctx: &mut MaybeEnabledAzure) -> anyhow::Result<()> {
+    let MaybeEnabledAzure::Enabled(ctx) = ctx else {
+        return Ok(());
+    };
+
+    let path = RemotePath::new(Utf8Path::new(format!("{}/file", ctx.base_prefix).as_str()))
+        .with_context(|| "RemotePath conversion")?;
+
+    let data = "remote blob data here".as_bytes();
+    let data_len = data.len() as u64;
+
+    ctx.client
+        .upload(std::io::Cursor::new(data), data.len(), &path, None)
+        .await?;
+
+    async fn download_and_compare(mut dl: Download) -> anyhow::Result<Vec<u8>> {
+        let mut buf = Vec::new();
+        tokio::io::copy(&mut dl.download_stream, &mut buf).await?;
+        Ok(buf)
+    }
+    // Normal download request
+    let dl = ctx.client.download(&path).await?;
+    let buf = download_and_compare(dl).await?;
+    assert_eq!(buf, data);
+
+    // Full range (end specified)
+    let dl = ctx
+        .client
+        .download_byte_range(&path, 0, Some(data_len))
+        .await?;
+    let buf = download_and_compare(dl).await?;
+    assert_eq!(buf, data);
+
+    // partial range (end specified)
+    let dl = ctx.client.download_byte_range(&path, 4, Some(10)).await?;
+    let buf = download_and_compare(dl).await?;
+    assert_eq!(buf, data[4..10]);
+
+    // partial range (end beyond real end)
+    let dl = ctx
+        .client
+        .download_byte_range(&path, 8, Some(data_len * 100))
+        .await?;
+    let buf = download_and_compare(dl).await?;
+    assert_eq!(buf, data[8..]);
+
+    // Partial range (end unspecified)
+    let dl = ctx.client.download_byte_range(&path, 4, None).await?;
+    let buf = download_and_compare(dl).await?;
+    assert_eq!(buf, data[4..]);
+
+    // Full range (end unspecified)
+    let dl = ctx.client.download_byte_range(&path, 0, None).await?;
+    let buf = download_and_compare(dl).await?;
+    assert_eq!(buf, data);
+
+    debug!("Cleanup: deleting file at path {path:?}");
+    ctx.client
+        .delete(&path)
+        .await
+        .with_context(|| format!("{path:?} removal"))?;
+
+    Ok(())
+}
+
+fn ensure_logging_ready() {
+    LOGGING_DONE.get_or_init(|| {
+        utils::logging::init(
+            utils::logging::LogFormat::Test,
+            utils::logging::TracingErrorLayerEnablement::Disabled,
+        )
+        .expect("logging init failed");
+    });
+}
+
+struct EnabledAzure {
+    client: Arc<GenericRemoteStorage>,
+    base_prefix: &'static str,
+}
+
+impl EnabledAzure {
+    async fn setup(max_keys_in_list_response: Option<i32>) -> Self {
+        let client = create_azure_client(max_keys_in_list_response)
+            .context("Azure client creation")
+            .expect("Azure client creation failed");
+
+        EnabledAzure {
+            client,
+            base_prefix: BASE_PREFIX,
+        }
+    }
+}
+
+enum MaybeEnabledAzure {
+    Enabled(EnabledAzure),
+    Disabled,
+}
+
+#[async_trait::async_trait]
+impl AsyncTestContext for MaybeEnabledAzure {
+    async fn setup() -> Self {
+        ensure_logging_ready();
+
+        if env::var(ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME).is_err() {
+            info!(
+                "`{}` env variable is not set, skipping the test",
+                ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME
+            );
+            return Self::Disabled;
+        }
+
+        Self::Enabled(EnabledAzure::setup(None).await)
+    }
+}
+
+enum MaybeEnabledAzureWithTestBlobs {
+    Enabled(AzureWithTestBlobs),
+    Disabled,
+    UploadsFailed(anyhow::Error, AzureWithTestBlobs),
+}
+
+struct AzureWithTestBlobs {
+    enabled: EnabledAzure,
+    remote_prefixes: HashSet<RemotePath>,
+    remote_blobs: HashSet<RemotePath>,
+}
+
+#[async_trait::async_trait]
+impl AsyncTestContext for MaybeEnabledAzureWithTestBlobs {
+    async fn setup() -> Self {
+        ensure_logging_ready();
+        if env::var(ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME).is_err() {
+            info!(
+                "`{}` env variable is not set, skipping the test",
+                ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME
+            );
+            return Self::Disabled;
+        }
+
+        let max_keys_in_list_response = 10;
+        let upload_tasks_count = 1 + (2 * usize::try_from(max_keys_in_list_response).unwrap());
+
+        let enabled = EnabledAzure::setup(Some(max_keys_in_list_response)).await;
+
+        match upload_azure_data(&enabled.client, enabled.base_prefix, upload_tasks_count).await {
+            ControlFlow::Continue(uploads) => {
+                info!("Remote objects created successfully");
+
+                Self::Enabled(AzureWithTestBlobs {
+                    enabled,
+                    remote_prefixes: uploads.prefixes,
+                    remote_blobs: uploads.blobs,
+                })
+            }
+            ControlFlow::Break(uploads) => Self::UploadsFailed(
+                anyhow::anyhow!("One or multiple blobs failed to upload to Azure"),
+                AzureWithTestBlobs {
+                    enabled,
+                    remote_prefixes: uploads.prefixes,
+                    remote_blobs: uploads.blobs,
+                },
+            ),
+        }
+    }
+
+    async fn teardown(self) {
+        match self {
+            Self::Disabled => {}
+            Self::Enabled(ctx) | Self::UploadsFailed(_, ctx) => {
+                cleanup(&ctx.enabled.client, ctx.remote_blobs).await;
+            }
+        }
+    }
+}
+
+// NOTE: the setups for the list_prefixes test and the list_files test are very similar
+// However, they are not idential. The list_prefixes function is concerned with listing prefixes,
+// whereas the list_files function is concerned with listing files.
+// See `RemoteStorage::list_files` documentation for more details
+enum MaybeEnabledAzureWithSimpleTestBlobs {
+    Enabled(AzureWithSimpleTestBlobs),
+    Disabled,
+    UploadsFailed(anyhow::Error, AzureWithSimpleTestBlobs),
+}
+struct AzureWithSimpleTestBlobs {
+    enabled: EnabledAzure,
+    remote_blobs: HashSet<RemotePath>,
+}
+
+#[async_trait::async_trait]
+impl AsyncTestContext for MaybeEnabledAzureWithSimpleTestBlobs {
+    async fn setup() -> Self {
+        ensure_logging_ready();
+        if env::var(ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME).is_err() {
+            info!(
+                "`{}` env variable is not set, skipping the test",
+                ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME
+            );
+            return Self::Disabled;
+        }
+
+        let max_keys_in_list_response = 10;
+        let upload_tasks_count = 1 + (2 * usize::try_from(max_keys_in_list_response).unwrap());
+
+        let enabled = EnabledAzure::setup(Some(max_keys_in_list_response)).await;
+
+        match upload_simple_azure_data(&enabled.client, upload_tasks_count).await {
+            ControlFlow::Continue(uploads) => {
+                info!("Remote objects created successfully");
+
+                Self::Enabled(AzureWithSimpleTestBlobs {
+                    enabled,
+                    remote_blobs: uploads,
+                })
+            }
+            ControlFlow::Break(uploads) => Self::UploadsFailed(
+                anyhow::anyhow!("One or multiple blobs failed to upload to Azure"),
+                AzureWithSimpleTestBlobs {
+                    enabled,
+                    remote_blobs: uploads,
+                },
+            ),
+        }
+    }
+
+    async fn teardown(self) {
+        match self {
+            Self::Disabled => {}
+            Self::Enabled(ctx) | Self::UploadsFailed(_, ctx) => {
+                cleanup(&ctx.enabled.client, ctx.remote_blobs).await;
+            }
+        }
+    }
+}
+
+fn create_azure_client(
+    max_keys_per_list_response: Option<i32>,
+) -> anyhow::Result<Arc<GenericRemoteStorage>> {
+    use rand::Rng;
+
+    let remote_storage_azure_container = env::var("REMOTE_STORAGE_AZURE_CONTAINER").context(
+        "`REMOTE_STORAGE_AZURE_CONTAINER` env var is not set, but real Azure tests are enabled",
+    )?;
+    let remote_storage_azure_region = env::var("REMOTE_STORAGE_AZURE_REGION").context(
+        "`REMOTE_STORAGE_AZURE_REGION` env var is not set, but real Azure tests are enabled",
+    )?;
+
+    // due to how time works, we've had test runners use the same nanos as bucket prefixes.
+    // millis is just a debugging aid for easier finding the prefix later.
+    let millis = std::time::SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .context("random Azure test prefix part calculation")?
+        .as_millis();
+
+    // because nanos can be the same for two threads so can millis, add randomness
+    let random = rand::thread_rng().gen::<u32>();
+
+    let remote_storage_config = RemoteStorageConfig {
+        max_concurrent_syncs: NonZeroUsize::new(100).unwrap(),
+        max_sync_errors: NonZeroU32::new(5).unwrap(),
+        storage: RemoteStorageKind::AzureContainer(AzureConfig {
+            container_name: remote_storage_azure_container,
+            container_region: remote_storage_azure_region,
+            prefix_in_container: Some(format!("test_{millis}_{random:08x}/")),
+            concurrency_limit: NonZeroUsize::new(100).unwrap(),
+            max_keys_per_list_response,
+        }),
+    };
+    Ok(Arc::new(
+        GenericRemoteStorage::from_config(&remote_storage_config).context("remote storage init")?,
+    ))
+}
+
+struct Uploads {
+    prefixes: HashSet<RemotePath>,
+    blobs: HashSet<RemotePath>,
+}
+
+async fn upload_azure_data(
+    client: &Arc<GenericRemoteStorage>,
+    base_prefix_str: &'static str,
+    upload_tasks_count: usize,
+) -> ControlFlow<Uploads, Uploads> {
+    info!("Creating {upload_tasks_count} Azure files");
+    let mut upload_tasks = JoinSet::new();
+    for i in 1..upload_tasks_count + 1 {
+        let task_client = Arc::clone(client);
+        upload_tasks.spawn(async move {
+            let prefix = format!("{base_prefix_str}/sub_prefix_{i}/");
+            let blob_prefix = RemotePath::new(Utf8Path::new(&prefix))
+                .with_context(|| format!("{prefix:?} to RemotePath conversion"))?;
+            let blob_path = blob_prefix.join(Utf8Path::new(&format!("blob_{i}")));
+            debug!("Creating remote item {i} at path {blob_path:?}");
+
+            let data = format!("remote blob data {i}").into_bytes();
+            let data_len = data.len();
+            task_client
+                .upload(std::io::Cursor::new(data), data_len, &blob_path, None)
+                .await?;
+
+            Ok::<_, anyhow::Error>((blob_prefix, blob_path))
+        });
+    }
+
+    let mut upload_tasks_failed = false;
+    let mut uploaded_prefixes = HashSet::with_capacity(upload_tasks_count);
+    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
+    while let Some(task_run_result) = upload_tasks.join_next().await {
+        match task_run_result
+            .context("task join failed")
+            .and_then(|task_result| task_result.context("upload task failed"))
+        {
+            Ok((upload_prefix, upload_path)) => {
+                uploaded_prefixes.insert(upload_prefix);
+                uploaded_blobs.insert(upload_path);
+            }
+            Err(e) => {
+                error!("Upload task failed: {e:?}");
+                upload_tasks_failed = true;
+            }
+        }
+    }
+
+    let uploads = Uploads {
+        prefixes: uploaded_prefixes,
+        blobs: uploaded_blobs,
+    };
+    if upload_tasks_failed {
+        ControlFlow::Break(uploads)
+    } else {
+        ControlFlow::Continue(uploads)
+    }
+}
+
+async fn cleanup(client: &Arc<GenericRemoteStorage>, objects_to_delete: HashSet<RemotePath>) {
+    info!(
+        "Removing {} objects from the remote storage during cleanup",
+        objects_to_delete.len()
+    );
+    let mut delete_tasks = JoinSet::new();
+    for object_to_delete in objects_to_delete {
+        let task_client = Arc::clone(client);
+        delete_tasks.spawn(async move {
+            debug!("Deleting remote item at path {object_to_delete:?}");
+            task_client
+                .delete(&object_to_delete)
+                .await
+                .with_context(|| format!("{object_to_delete:?} removal"))
+        });
+    }
+
+    while let Some(task_run_result) = delete_tasks.join_next().await {
+        match task_run_result {
+            Ok(task_result) => match task_result {
+                Ok(()) => {}
+                Err(e) => error!("Delete task failed: {e:?}"),
+            },
+            Err(join_err) => error!("Delete task did not finish correctly: {join_err}"),
+        }
+    }
+}
+
+// Uploads files `folder{j}/blob{i}.txt`. See test description for more details.
+async fn upload_simple_azure_data(
+    client: &Arc<GenericRemoteStorage>,
+    upload_tasks_count: usize,
+) -> ControlFlow<HashSet<RemotePath>, HashSet<RemotePath>> {
+    info!("Creating {upload_tasks_count} Azure files");
+    let mut upload_tasks = JoinSet::new();
+    for i in 1..upload_tasks_count + 1 {
+        let task_client = Arc::clone(client);
+        upload_tasks.spawn(async move {
+            let blob_path = PathBuf::from(format!("folder{}/blob_{}.txt", i / 7, i));
+            let blob_path = RemotePath::new(
+                Utf8Path::from_path(blob_path.as_path()).expect("must be valid blob path"),
+            )
+            .with_context(|| format!("{blob_path:?} to RemotePath conversion"))?;
+            debug!("Creating remote item {i} at path {blob_path:?}");
+
+            let data = format!("remote blob data {i}").into_bytes();
+            let data_len = data.len();
+            task_client
+                .upload(std::io::Cursor::new(data), data_len, &blob_path, None)
+                .await?;
+
+            Ok::<_, anyhow::Error>(blob_path)
+        });
+    }
+
+    let mut upload_tasks_failed = false;
+    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
+    while let Some(task_run_result) = upload_tasks.join_next().await {
+        match task_run_result
+            .context("task join failed")
+            .and_then(|task_result| task_result.context("upload task failed"))
+        {
+            Ok(upload_path) => {
+                uploaded_blobs.insert(upload_path);
+            }
+            Err(e) => {
+                error!("Upload task failed: {e:?}");
+                upload_tasks_failed = true;
+            }
+        }
+    }
+
+    if upload_tasks_failed {
+        ControlFlow::Break(uploaded_blobs)
+    } else {
+        ControlFlow::Continue(uploaded_blobs)
+    }
+}

libs/utils/src/http/error.rs

@@ -1,8 +1,9 @@
 use hyper::{header, Body, Response, StatusCode};
 use serde::{Deserialize, Serialize};
+use std::borrow::Cow;
 use std::error::Error as StdError;
 use thiserror::Error;
-use tracing::error;
+use tracing::{error, info};
 
 #[derive(Debug, Error)]
 pub enum ApiError {
@@ -25,7 +26,7 @@ pub enum ApiError {
     PreconditionFailed(Box<str>),
 
     #[error("Resource temporarily unavailable: {0}")]
-    ResourceUnavailable(String),
+    ResourceUnavailable(Cow<'static, str>),
 
     #[error("Shutting down")]
     ShuttingDown,
@@ -115,10 +116,12 @@ pub async fn route_error_handler(err: routerify::RouteError) -> Response<Body> {
 
 pub fn api_error_handler(api_error: ApiError) -> Response<Body> {
     // Print a stack trace for Internal Server errors
-    if let ApiError::InternalServerError(_) = api_error {
-        error!("Error processing HTTP request: {api_error:?}");
-    } else {
-        error!("Error processing HTTP request: {api_error:#}");
+
+    match api_error {
+        ApiError::ResourceUnavailable(_) => info!("Error processing HTTP request: {api_error:#}"),
+        ApiError::NotFound(_) => info!("Error processing HTTP request: {api_error:#}"),
+        ApiError::InternalServerError(_) => error!("Error processing HTTP request: {api_error:?}"),
+        _ => error!("Error processing HTTP request: {api_error:#}"),
     }
 
     api_error.into_response()

libs/utils/src/seqwait.rs

@@ -58,7 +58,7 @@ where
 // to get that.
 impl<T: Ord> PartialOrd for Waiter<T> {
     fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
-        other.wake_num.partial_cmp(&self.wake_num)
+        Some(self.cmp(other))
     }
 }
 

libs/vm_monitor/README.md

@@ -27,8 +27,8 @@ and old one if it exists.
 * the filecache: a struct that allows communication with the Postgres file cache.
 On startup, we connect to the filecache and hold on to the connection for the
 entire monitor lifetime.
-* the cgroup watcher: the `CgroupWatcher` manages the `neon-postgres` cgroup by
-listening for `memory.high` events and setting its `memory.{high,max}` values.
+* the cgroup watcher: the `CgroupWatcher` polls the `neon-postgres` cgroup's memory
+usage and sends rolling aggregates to the runner.
 * the runner: the runner marries the filecache and cgroup watcher together,
 communicating with the agent throught the `Dispatcher`, and then calling filecache
 and cgroup watcher functions as needed to upscale and downscale

libs/vm_monitor/src/cgroup.rs

@@ -1,161 +1,38 @@
-use std::{
-    fmt::{Debug, Display},
-    fs,
-    pin::pin,
-    sync::atomic::{AtomicU64, Ordering},
-};
+use std::fmt::{self, Debug, Formatter};
+use std::time::{Duration, Instant};
 
-use anyhow::{anyhow, bail, Context};
+use anyhow::{anyhow, Context};
 use cgroups_rs::{
-    freezer::FreezerController,
-    hierarchies::{self, is_cgroup2_unified_mode, UNIFIED_MOUNTPOINT},
+    hierarchies::{self, is_cgroup2_unified_mode},
     memory::MemController,
-    MaxValue,
-    Subsystem::{Freezer, Mem},
+    Subsystem,
 };
-use inotify::{EventStream, Inotify, WatchMask};
-use tokio::sync::mpsc::{self, error::TryRecvError};
-use tokio::time::{Duration, Instant};
-use tokio_stream::{Stream, StreamExt};
+use tokio::sync::watch;
 use tracing::{info, warn};
 
-use crate::protocol::Resources;
-use crate::MiB;
-
-/// Monotonically increasing counter of the number of memory.high events
-/// the cgroup has experienced.
-///
-/// We use this to determine if a modification to the `memory.events` file actually
-/// changed the `high` field. If not, we don't care about the change. When we
-/// read the file, we check the `high` field in the file against `MEMORY_EVENT_COUNT`
-/// to see if it changed since last time.
-pub static MEMORY_EVENT_COUNT: AtomicU64 = AtomicU64::new(0);
-
-/// Monotonically increasing counter that gives each cgroup event a unique id.
-///
-/// This allows us to answer questions like "did this upscale arrive before this
-/// memory.high?". This static is also used by the `Sequenced` type to "tag" values
-/// with a sequence number. As such, prefer to used the `Sequenced` type rather
-/// than this static directly.
-static EVENT_SEQUENCE_NUMBER: AtomicU64 = AtomicU64::new(0);
-
-/// A memory event type reported in memory.events.
-#[derive(Debug, Eq, PartialEq, Copy, Clone)]
-pub enum MemoryEvent {
-    Low,
-    High,
-    Max,
-    Oom,
-    OomKill,
-    OomGroupKill,
-}
-
-impl MemoryEvent {
-    fn as_str(&self) -> &str {
-        match self {
-            MemoryEvent::Low => "low",
-            MemoryEvent::High => "high",
-            MemoryEvent::Max => "max",
-            MemoryEvent::Oom => "oom",
-            MemoryEvent::OomKill => "oom_kill",
-            MemoryEvent::OomGroupKill => "oom_group_kill",
-        }
-    }
-}
-
-impl Display for MemoryEvent {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        f.write_str(self.as_str())
-    }
-}
-
 /// Configuration for a `CgroupWatcher`
 #[derive(Debug, Clone)]
 pub struct Config {
-    // The target difference between the total memory reserved for the cgroup
-    // and the value of the cgroup's memory.high.
-    //
-    // In other words, memory.high + oom_buffer_bytes will equal the total memory that the cgroup may
-    // use (equal to system memory, minus whatever's taken out for the file cache).
-    oom_buffer_bytes: u64,
+    /// Interval at which we should be fetching memory statistics
+    memory_poll_interval: Duration,
 
-    // The amount of memory, in bytes, below a proposed new value for
-    // memory.high that the cgroup's memory usage must be for us to downscale
-    //
-    // In other words, we can downscale only when:
-    //
-    //   memory.current + memory_high_buffer_bytes < (proposed) memory.high
-    //
-    // TODO: there's some minor issues with this approach -- in particular, that we might have
-    // memory in use by the kernel's page cache that we're actually ok with getting rid of.
-    pub(crate) memory_high_buffer_bytes: u64,
-
-    // The maximum duration, in milliseconds, that we're allowed to pause
-    // the cgroup for while waiting for the autoscaler-agent to upscale us
-    max_upscale_wait: Duration,
-
-    // The required minimum time, in milliseconds, that we must wait before re-freezing
-    // the cgroup while waiting for the autoscaler-agent to upscale us.
-    do_not_freeze_more_often_than: Duration,
-
-    // The amount of memory, in bytes, that we should periodically increase memory.high
-    // by while waiting for the autoscaler-agent to upscale us.
-    //
-    // This exists to avoid the excessive throttling that happens when a cgroup is above its
-    // memory.high for too long. See more here:
-    // https://github.com/neondatabase/autoscaling/issues/44#issuecomment-1522487217
-    memory_high_increase_by_bytes: u64,
-
-    // The period, in milliseconds, at which we should repeatedly increase the value
-    // of the cgroup's memory.high while we're waiting on upscaling and memory.high
-    // is still being hit.
-    //
-    // Technically speaking, this actually serves as a rate limit to moderate responding to
-    // memory.high events, but these are roughly equivalent if the process is still allocating
-    // memory.
-    memory_high_increase_every: Duration,
-}
-
-impl Config {
-    /// Calculate the new value for the cgroups memory.high based on system memory
-    pub fn calculate_memory_high_value(&self, total_system_mem: u64) -> u64 {
-        total_system_mem.saturating_sub(self.oom_buffer_bytes)
-    }
+    /// The number of samples used in constructing aggregated memory statistics
+    memory_history_len: usize,
+    /// The number of most recent samples that will be periodically logged.
+    ///
+    /// Each sample is logged exactly once. Increasing this value means that recent samples will be
+    /// logged less frequently, and vice versa.
+    ///
+    /// For simplicity, this value must be greater than or equal to `memory_history_len`.
+    memory_history_log_interval: usize,
 }
 
 impl Default for Config {
     fn default() -> Self {
         Self {
-            oom_buffer_bytes: 100 * MiB,
-            memory_high_buffer_bytes: 100 * MiB,
-            // while waiting for upscale, don't freeze for more than 20ms every 1s
-            max_upscale_wait: Duration::from_millis(20),
-            do_not_freeze_more_often_than: Duration::from_millis(1000),
-            // while waiting for upscale, increase memory.high by 10MiB every 25ms
-            memory_high_increase_by_bytes: 10 * MiB,
-            memory_high_increase_every: Duration::from_millis(25),
-        }
-    }
-}
-
-/// Used to represent data that is associated with a certain point in time, such
-/// as an upscale request or memory.high event.
-///
-/// Internally, creating a `Sequenced` uses a static atomic counter to obtain
-/// a unique sequence number. Sequence numbers are monotonically increasing,
-/// allowing us to answer questions like "did this upscale happen after this
-/// memory.high event?" by comparing the sequence numbers of the two events.
-#[derive(Debug, Clone)]
-pub struct Sequenced<T> {
-    seqnum: u64,
-    data: T,
-}
-
-impl<T> Sequenced<T> {
-    pub fn new(data: T) -> Self {
-        Self {
-            seqnum: EVENT_SEQUENCE_NUMBER.fetch_add(1, Ordering::AcqRel),
-            data,
+            memory_poll_interval: Duration::from_millis(100),
+            memory_history_len: 5, // use 500ms of history for decision-making
+            memory_history_log_interval: 20, // but only log every ~2s (otherwise it's spammy)
         }
     }
 }
@@ -170,74 +47,14 @@ impl<T> Sequenced<T> {
 pub struct CgroupWatcher {
     pub config: Config,
 
-    /// The sequence number of the last upscale.
-    ///
-    /// If we receive a memory.high event that has a _lower_ sequence number than
-    /// `last_upscale_seqnum`, then we know it occured before the upscale, and we
-    /// can safely ignore it.
-    ///
-    /// Note: Like the `events` field, this doesn't _need_ interior mutability but we
-    /// use it anyways so that methods take `&self`, not `&mut self`.
-    last_upscale_seqnum: AtomicU64,
-
-    /// A channel on which we send messages to request upscale from the dispatcher.
-    upscale_requester: mpsc::Sender<()>,
-
     /// The actual cgroup we are watching and managing.
     cgroup: cgroups_rs::Cgroup,
 }
 
-/// Read memory.events for the desired event type.
-///
-/// `path` specifies the path to the desired `memory.events` file.
-/// For more info, see the `memory.events` section of the [kernel docs]
-/// <https://docs.kernel.org/admin-guide/cgroup-v2.html#memory-interface-files>
-fn get_event_count(path: &str, event: MemoryEvent) -> anyhow::Result<u64> {
-    let contents = fs::read_to_string(path)
-        .with_context(|| format!("failed to read memory.events from {path}"))?;
-
-    // Then contents of the file look like:
-    // low 42
-    // high 101
-    // ...
-    contents
-        .lines()
-        .filter_map(|s| s.split_once(' '))
-        .find(|(e, _)| *e == event.as_str())
-        .ok_or_else(|| anyhow!("failed to find entry for memory.{event} events in {path}"))
-        .and_then(|(_, count)| {
-            count
-                .parse::<u64>()
-                .with_context(|| format!("failed to parse memory.{event} as u64"))
-        })
-}
-
-/// Create an event stream that produces events whenever the file at the provided
-/// path is modified.
-fn create_file_watcher(path: &str) -> anyhow::Result<EventStream<[u8; 1024]>> {
-    info!("creating file watcher for {path}");
-    let inotify = Inotify::init().context("failed to initialize file watcher")?;
-    inotify
-        .watches()
-        .add(path, WatchMask::MODIFY)
-        .with_context(|| format!("failed to start watching {path}"))?;
-    inotify
-        // The inotify docs use [0u8; 1024] so we'll just copy them. We only need
-        // to store one event at a time - if the event gets written over, that's
-        // ok. We still see that there is an event. For more information, see:
-        // https://man7.org/linux/man-pages/man7/inotify.7.html
-        .into_event_stream([0u8; 1024])
-        .context("failed to start inotify event stream")
-}
-
 impl CgroupWatcher {
     /// Create a new `CgroupWatcher`.
     #[tracing::instrument(skip_all, fields(%name))]
-    pub fn new(
-        name: String,
-        // A channel on which to send upscale requests
-        upscale_requester: mpsc::Sender<()>,
-    ) -> anyhow::Result<(Self, impl Stream<Item = Sequenced<u64>>)> {
+    pub fn new(name: String) -> anyhow::Result<Self> {
         // TODO: clarify exactly why we need v2
         // Make sure cgroups v2 (aka unified) are supported
         if !is_cgroup2_unified_mode() {
@@ -245,410 +62,203 @@ impl CgroupWatcher {
         }
         let cgroup = cgroups_rs::Cgroup::load(hierarchies::auto(), &name);
 
-        // Start monitoring the cgroup for memory events. In general, for
-        // cgroups v2 (aka unified), metrics are reported in files like
-        // > `/sys/fs/cgroup/{name}/{metric}`
-        // We are looking for `memory.high` events, which are stored in the
-        // file `memory.events`. For more info, see the `memory.events` section
-        // of https://docs.kernel.org/admin-guide/cgroup-v2.html#memory-interface-files
-        let path = format!("{}/{}/memory.events", UNIFIED_MOUNTPOINT, &name);
-        let memory_events = create_file_watcher(&path)
-            .with_context(|| format!("failed to create event watcher for {path}"))?
-            // This would be nice with with .inspect_err followed by .ok
-            .filter_map(move |_| match get_event_count(&path, MemoryEvent::High) {
-                Ok(high) => Some(high),
-                Err(error) => {
-                    // TODO: Might want to just panic here
-                    warn!(?error, "failed to read high events count from {}", &path);
-                    None
-                }
-            })
-            // Only report the event if the memory.high count increased
-            .filter_map(|high| {
-                if MEMORY_EVENT_COUNT.fetch_max(high, Ordering::AcqRel) < high {
-                    Some(high)
-                } else {
-                    None
-                }
-            })
-            .map(Sequenced::new);
-
-        let initial_count = get_event_count(
-            &format!("{}/{}/memory.events", UNIFIED_MOUNTPOINT, &name),
-            MemoryEvent::High,
-        )?;
-
-        info!(initial_count, "initial memory.high event count");
-
-        // Hard update `MEMORY_EVENT_COUNT` since there could have been processes
-        // running in the cgroup before that caused it to be non-zero.
-        MEMORY_EVENT_COUNT.fetch_max(initial_count, Ordering::AcqRel);
-
-        Ok((
-            Self {
-                cgroup,
-                upscale_requester,
-                last_upscale_seqnum: AtomicU64::new(0),
-                config: Default::default(),
-            },
-            memory_events,
-        ))
+        Ok(Self {
+            cgroup,
+            config: Default::default(),
+        })
     }
 
     /// The entrypoint for the `CgroupWatcher`.
     #[tracing::instrument(skip_all)]
-    pub async fn watch<E>(
+    pub async fn watch(
         &self,
-        // These are ~dependency injected~ (fancy, I know) because this function
-        // should never return.
-        // -> therefore: when we tokio::spawn it, we don't await the JoinHandle.
-        // -> therefore: if we want to stick it in an Arc so many threads can access
-        //    it, methods can never take mutable access.
-        //     - note: we use the Arc strategy so that a) we can call this function
-        //             right here and b) the runner can call the set/get_memory methods
-        // -> since calling recv() on a tokio::sync::mpsc::Receiver takes &mut self,
-        //    we just pass them in here instead of holding them in fields, as that
-        //    would require this method to take &mut self.
-        mut upscales: mpsc::Receiver<Sequenced<Resources>>,
-        events: E,
-    ) -> anyhow::Result<()>
-    where
-        E: Stream<Item = Sequenced<u64>>,
-    {
-        let mut wait_to_freeze = pin!(tokio::time::sleep(Duration::ZERO));
-        let mut last_memory_high_increase_at: Option<Instant> = None;
-        let mut events = pin!(events);
-
-        // Are we waiting to be upscaled? Could be true if we request upscale due
-        // to a memory.high event and it does not arrive in time.
-        let mut waiting_on_upscale = false;
-
-        loop {
-            tokio::select! {
-                upscale = upscales.recv() => {
-                    let Sequenced { seqnum, data } = upscale
-                        .context("failed to listen on upscale notification channel")?;
-                    waiting_on_upscale = false;
-                    last_memory_high_increase_at = None;
-                    self.last_upscale_seqnum.store(seqnum, Ordering::Release);
-                    info!(cpu = data.cpu, mem_bytes = data.mem, "received upscale");
-                }
-                event = events.next() => {
-                    let Some(Sequenced { seqnum, .. }) = event else {
-                        bail!("failed to listen for memory.high events")
-                    };
-                    // The memory.high came before our last upscale, so we consider
-                    // it resolved
-                    if self.last_upscale_seqnum.fetch_max(seqnum, Ordering::AcqRel) > seqnum {
-                        info!(
-                            "received memory.high event, but it came before our last upscale -> ignoring it"
-                        );
-                        continue;
-                    }
-
-                    // The memory.high came after our latest upscale. We don't
-                    // want to do anything yet, so peek the next event in hopes
-                    // that it's an upscale.
-                    if let Some(upscale_num) = self
-                        .upscaled(&mut upscales)
-                        .context("failed to check if we were upscaled")?
-                    {
-                        if upscale_num > seqnum {
-                            info!(
-                                "received memory.high event, but it came before our last upscale -> ignoring it"
-                            );
-                            continue;
-                        }
-                    }
-
-                    // If it's been long enough since we last froze, freeze the
-                    // cgroup and request upscale
-                    if wait_to_freeze.is_elapsed() {
-                        info!("received memory.high event -> requesting upscale");
-                        waiting_on_upscale = self
-                            .handle_memory_high_event(&mut upscales)
-                            .await
-                            .context("failed to handle upscale")?;
-                        wait_to_freeze
-                            .as_mut()
-                            .reset(Instant::now() + self.config.do_not_freeze_more_often_than);
-                        continue;
-                    }
-
-                    // Ok, we can't freeze, just request upscale
-                    if !waiting_on_upscale {
-                        info!("received memory.high event, but too soon to refreeze -> requesting upscale");
-
-                        // Make check to make sure we haven't been upscaled in the
-                        // meantine (can happen if the agent independently decides
-                        // to upscale us again)
-                        if self
-                            .upscaled(&mut upscales)
-                            .context("failed to check if we were upscaled")?
-                            .is_some()
-                        {
-                            info!("no need to request upscaling because we got upscaled");
-                            continue;
-                        }
-                        self.upscale_requester
-                            .send(())
-                            .await
-                            .context("failed to request upscale")?;
-                        waiting_on_upscale = true;
-                        continue;
-                    }
-
-                    // Shoot, we can't freeze or and we're still waiting on upscale,
-                    // increase memory.high to reduce throttling
-                    let can_increase_memory_high = match last_memory_high_increase_at {
-                        None => true,
-                        Some(t) => t.elapsed() > self.config.memory_high_increase_every,
-                    };
-                    if can_increase_memory_high {
-                        info!(
-                            "received memory.high event, \
-                            but too soon to refreeze and already requested upscale \
-                            -> increasing memory.high"
-                        );
-
-                        // Make check to make sure we haven't been upscaled in the
-                        // meantine (can happen if the agent independently decides
-                        // to upscale us again)
-                        if self
-                            .upscaled(&mut upscales)
-                            .context("failed to check if we were upscaled")?
-                            .is_some()
-                        {
-                            info!("no need to increase memory.high because got upscaled");
-                            continue;
-                        }
-
-                        // Request upscale anyways (the agent will handle deduplicating
-                        // requests)
-                        self.upscale_requester
-                            .send(())
-                            .await
-                            .context("failed to request upscale")?;
-
-                        let memory_high =
-                            self.get_memory_high_bytes().context("failed to get memory.high")?;
-                        let new_high = memory_high + self.config.memory_high_increase_by_bytes;
-                        info!(
-                            current_high_bytes = memory_high,
-                            new_high_bytes = new_high,
-                            "updating memory.high"
-                        );
-                        self.set_memory_high_bytes(new_high)
-                            .context("failed to set memory.high")?;
-                        last_memory_high_increase_at = Some(Instant::now());
-                        continue;
-                    }
-
-                    info!("received memory.high event, but can't do anything");
-                }
-            };
-        }
-    }
-
-    /// Handle a `memory.high`, returning whether we are still waiting on upscale
-    /// by the time the function returns.
-    ///
-    /// The general plan for handling a `memory.high` event is as follows:
-    /// 1. Freeze the cgroup
-    /// 2. Start a timer for `self.config.max_upscale_wait`
-    /// 3. Request upscale
-    /// 4. After the timer elapses or we receive upscale, thaw the cgroup.
-    /// 5. Return whether or not we are still waiting for upscale. If we are,
-    ///    we'll increase the cgroups memory.high to avoid getting oom killed
-    #[tracing::instrument(skip_all)]
-    async fn handle_memory_high_event(
-        &self,
-        upscales: &mut mpsc::Receiver<Sequenced<Resources>>,
-    ) -> anyhow::Result<bool> {
-        // Immediately freeze the cgroup before doing anything else.
-        info!("received memory.high event -> freezing cgroup");
-        self.freeze().context("failed to freeze cgroup")?;
-
-        // We'll use this for logging durations
-        let start_time = Instant::now();
-
-        // Await the upscale until we have to unfreeze
-        let timed =
-            tokio::time::timeout(self.config.max_upscale_wait, self.await_upscale(upscales));
-
-        // Request the upscale
-        info!(
-            wait = ?self.config.max_upscale_wait,
-            "sending request for immediate upscaling",
-        );
-        self.upscale_requester
-            .send(())
-            .await
-            .context("failed to request upscale")?;
-
-        let waiting_on_upscale = match timed.await {
-            Ok(Ok(())) => {
-                info!(elapsed = ?start_time.elapsed(), "received upscale in time");
-                false
-            }
-            // **important**: unfreeze the cgroup before ?-reporting the error
-            Ok(Err(e)) => {
-                info!("error waiting for upscale -> thawing cgroup");
-                self.thaw()
-                    .context("failed to thaw cgroup after errored waiting for upscale")?;
-                Err(e.context("failed to await upscale"))?
-            }
-            Err(_) => {
-                info!(elapsed = ?self.config.max_upscale_wait, "timed out waiting for upscale");
-                true
-            }
-        };
-
-        info!("thawing cgroup");
-        self.thaw().context("failed to thaw cgroup")?;
-
-        Ok(waiting_on_upscale)
-    }
-
-    /// Checks whether we were just upscaled, returning the upscale's sequence
-    /// number if so.
-    #[tracing::instrument(skip_all)]
-    fn upscaled(
-        &self,
-        upscales: &mut mpsc::Receiver<Sequenced<Resources>>,
-    ) -> anyhow::Result<Option<u64>> {
-        let Sequenced { seqnum, data } = match upscales.try_recv() {
-            Ok(upscale) => upscale,
-            Err(TryRecvError::Empty) => return Ok(None),
-            Err(TryRecvError::Disconnected) => {
-                bail!("upscale notification channel was disconnected")
-            }
-        };
-
-        // Make sure to update the last upscale sequence number
-        self.last_upscale_seqnum.store(seqnum, Ordering::Release);
-        info!(cpu = data.cpu, mem_bytes = data.mem, "received upscale");
-        Ok(Some(seqnum))
-    }
-
-    /// Await an upscale event, discarding any `memory.high` events received in
-    /// the process.
-    ///
-    /// This is used in `handle_memory_high_event`, where we need to listen
-    /// for upscales in particular so we know if we can thaw the cgroup early.
-    #[tracing::instrument(skip_all)]
-    async fn await_upscale(
-        &self,
-        upscales: &mut mpsc::Receiver<Sequenced<Resources>>,
+        updates: watch::Sender<(Instant, MemoryHistory)>,
     ) -> anyhow::Result<()> {
-        let Sequenced { seqnum, .. } = upscales
-            .recv()
-            .await
-            .context("error listening for upscales")?;
+        // this requirement makes the code a bit easier to work with; see the config for more.
+        assert!(self.config.memory_history_len <= self.config.memory_history_log_interval);
 
-        self.last_upscale_seqnum.store(seqnum, Ordering::Release);
-        Ok(())
-    }
+        let mut ticker = tokio::time::interval(self.config.memory_poll_interval);
+        ticker.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Skip);
+        // ticker.reset_immediately(); // FIXME: enable this once updating to tokio >= 1.30.0
 
-    /// Get the cgroup's name.
-    pub fn path(&self) -> &str {
-        self.cgroup.path()
-    }
-}
+        let mem_controller = self.memory()?;
 
-// Methods for manipulating the actual cgroup
-impl CgroupWatcher {
-    /// Get a handle on the freezer subsystem.
-    fn freezer(&self) -> anyhow::Result<&FreezerController> {
-        if let Some(Freezer(freezer)) = self
-            .cgroup
-            .subsystems()
-            .iter()
-            .find(|sub| matches!(sub, Freezer(_)))
-        {
-            Ok(freezer)
-        } else {
-            anyhow::bail!("could not find freezer subsystem")
+        // buffer for samples that will be logged. once full, it remains so.
+        let history_log_len = self.config.memory_history_log_interval;
+        let mut history_log_buf = vec![MemoryStatus::zeroed(); history_log_len];
+
+        for t in 0_u64.. {
+            ticker.tick().await;
+
+            let now = Instant::now();
+            let mem = Self::memory_usage(mem_controller);
+
+            let i = t as usize % history_log_len;
+            history_log_buf[i] = mem;
+
+            // We're taking *at most* memory_history_len values; we may be bounded by the total
+            // number of samples that have come in so far.
+            let samples_count = (t + 1).min(self.config.memory_history_len as u64) as usize;
+            // NB: in `ring_buf_recent_values_iter`, `i` is *inclusive*, which matches the fact
+            // that we just inserted a value there, so the end of the iterator will *include* the
+            // value at i, rather than stopping just short of it.
+            let samples = ring_buf_recent_values_iter(&history_log_buf, i, samples_count);
+
+            let summary = MemoryHistory {
+                avg_non_reclaimable: samples.map(|h| h.non_reclaimable).sum::<u64>()
+                    / samples_count as u64,
+                samples_count,
+                samples_span: self.config.memory_poll_interval * (samples_count - 1) as u32,
+            };
+
+            // Log the current history if it's time to do so. Because `history_log_buf` has length
+            // equal to the logging interval, we can just log the entire buffer every time we set
+            // the last entry, which also means that for this log line, we can ignore that it's a
+            // ring buffer (because all the entries are in order of increasing time).
+            if i == history_log_len - 1 {
+                info!(
+                    history = ?MemoryStatus::debug_slice(&history_log_buf),
+                    summary = ?summary,
+                    "Recent cgroup memory statistics history"
+                );
+            }
+
+            updates
+                .send((now, summary))
+                .context("failed to send MemoryHistory")?;
         }
-    }
 
-    /// Attempt to freeze the cgroup.
-    pub fn freeze(&self) -> anyhow::Result<()> {
-        self.freezer()
-            .context("failed to get freezer subsystem")?
-            .freeze()
-            .context("failed to freeze")
-    }
-
-    /// Attempt to thaw the cgroup.
-    pub fn thaw(&self) -> anyhow::Result<()> {
-        self.freezer()
-            .context("failed to get freezer subsystem")?
-            .thaw()
-            .context("failed to thaw")
+        unreachable!()
     }
 
     /// Get a handle on the memory subsystem.
-    ///
-    /// Note: this method does not require `self.memory_update_lock` because
-    /// getting a handle to the subsystem does not access any of the files we
-    /// care about, such as memory.high and memory.events
     fn memory(&self) -> anyhow::Result<&MemController> {
-        if let Some(Mem(memory)) = self
-            .cgroup
+        self.cgroup
             .subsystems()
             .iter()
-            .find(|sub| matches!(sub, Mem(_)))
-        {
-            Ok(memory)
-        } else {
-            anyhow::bail!("could not find memory subsystem")
-        }
-    }
-
-    /// Get cgroup current memory usage.
-    pub fn current_memory_usage(&self) -> anyhow::Result<u64> {
-        Ok(self
-            .memory()
-            .context("failed to get memory subsystem")?
-            .memory_stat()
-            .usage_in_bytes)
-    }
-
-    /// Set cgroup memory.high threshold.
-    pub fn set_memory_high_bytes(&self, bytes: u64) -> anyhow::Result<()> {
-        self.set_memory_high_internal(MaxValue::Value(u64::min(bytes, i64::MAX as u64) as i64))
-    }
-
-    /// Set the cgroup's memory.high to 'max', disabling it.
-    pub fn unset_memory_high(&self) -> anyhow::Result<()> {
-        self.set_memory_high_internal(MaxValue::Max)
-    }
-
-    fn set_memory_high_internal(&self, value: MaxValue) -> anyhow::Result<()> {
-        self.memory()
-            .context("failed to get memory subsystem")?
-            .set_mem(cgroups_rs::memory::SetMemory {
-                low: None,
-                high: Some(value),
-                min: None,
-                max: None,
+            .find_map(|sub| match sub {
+                Subsystem::Mem(c) => Some(c),
+                _ => None,
             })
-            .map_err(anyhow::Error::from)
+            .ok_or_else(|| anyhow!("could not find memory subsystem"))
     }
 
-    /// Get memory.high threshold.
-    pub fn get_memory_high_bytes(&self) -> anyhow::Result<u64> {
-        let high = self
-            .memory()
-            .context("failed to get memory subsystem while getting memory statistics")?
-            .get_mem()
-            .map(|mem| mem.high)
-            .context("failed to get memory statistics from subsystem")?;
-        match high {
-            Some(MaxValue::Max) => Ok(i64::MAX as u64),
-            Some(MaxValue::Value(high)) => Ok(high as u64),
-            None => anyhow::bail!("failed to read memory.high from memory subsystem"),
+    /// Given a handle on the memory subsystem, returns the current memory information
+    fn memory_usage(mem_controller: &MemController) -> MemoryStatus {
+        let stat = mem_controller.memory_stat().stat;
+        MemoryStatus {
+            non_reclaimable: stat.active_anon + stat.inactive_anon,
         }
     }
 }
+
+// Helper function for `CgroupWatcher::watch`
+fn ring_buf_recent_values_iter<T>(
+    buf: &[T],
+    last_value_idx: usize,
+    count: usize,
+) -> impl '_ + Iterator<Item = &T> {
+    // Assertion carried over from `CgroupWatcher::watch`, to make the logic in this function
+    // easier (we only have to add `buf.len()` once, rather than a dynamic number of times).
+    assert!(count <= buf.len());
+
+    buf.iter()
+        // 'cycle' because the values could wrap around
+        .cycle()
+        // with 'cycle', this skip is more like 'offset', and functionally this is
+        // offsettting by 'last_value_idx - count (mod buf.len())', but we have to be
+        // careful to avoid underflow, so we pre-add buf.len().
+        // The '+ 1' is because `last_value_idx` is inclusive, rather than exclusive.
+        .skip((buf.len() + last_value_idx + 1 - count) % buf.len())
+        .take(count)
+}
+
+/// Summary of recent memory usage
+#[derive(Debug, Copy, Clone)]
+pub struct MemoryHistory {
+    /// Rolling average of non-reclaimable memory usage samples over the last `history_period`
+    pub avg_non_reclaimable: u64,
+
+    /// The number of samples used to construct this summary
+    pub samples_count: usize,
+    /// Total timespan between the first and last sample used for this summary
+    pub samples_span: Duration,
+}
+
+#[derive(Debug, Copy, Clone)]
+pub struct MemoryStatus {
+    non_reclaimable: u64,
+}
+
+impl MemoryStatus {
+    fn zeroed() -> Self {
+        MemoryStatus { non_reclaimable: 0 }
+    }
+
+    fn debug_slice(slice: &[Self]) -> impl '_ + Debug {
+        struct DS<'a>(&'a [MemoryStatus]);
+
+        impl<'a> Debug for DS<'a> {
+            fn fmt(&self, f: &mut Formatter) -> fmt::Result {
+                f.debug_struct("[MemoryStatus]")
+                    .field(
+                        "non_reclaimable[..]",
+                        &Fields(self.0, |stat: &MemoryStatus| {
+                            BytesToGB(stat.non_reclaimable)
+                        }),
+                    )
+                    .finish()
+            }
+        }
+
+        struct Fields<'a, F>(&'a [MemoryStatus], F);
+
+        impl<'a, F: Fn(&MemoryStatus) -> T, T: Debug> Debug for Fields<'a, F> {
+            fn fmt(&self, f: &mut Formatter) -> fmt::Result {
+                f.debug_list().entries(self.0.iter().map(&self.1)).finish()
+            }
+        }
+
+        struct BytesToGB(u64);
+
+        impl Debug for BytesToGB {
+            fn fmt(&self, f: &mut Formatter) -> fmt::Result {
+                f.write_fmt(format_args!(
+                    "{:.3}Gi",
+                    self.0 as f64 / (1_u64 << 30) as f64
+                ))
+            }
+        }
+
+        DS(slice)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #[test]
+    fn ring_buf_iter() {
+        let buf = vec![0_i32, 1, 2, 3, 4, 5, 6, 7, 8, 9];
+
+        let values = |offset, count| {
+            super::ring_buf_recent_values_iter(&buf, offset, count)
+                .copied()
+                .collect::<Vec<i32>>()
+        };
+
+        // Boundary conditions: start, end, and entire thing:
+        assert_eq!(values(0, 1), [0]);
+        assert_eq!(values(3, 4), [0, 1, 2, 3]);
+        assert_eq!(values(9, 4), [6, 7, 8, 9]);
+        assert_eq!(values(9, 10), [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]);
+
+        // "normal" operation: no wraparound
+        assert_eq!(values(7, 4), [4, 5, 6, 7]);
+
+        // wraparound:
+        assert_eq!(values(0, 4), [7, 8, 9, 0]);
+        assert_eq!(values(1, 4), [8, 9, 0, 1]);
+        assert_eq!(values(2, 4), [9, 0, 1, 2]);
+        assert_eq!(values(2, 10), [3, 4, 5, 6, 7, 8, 9, 0, 1, 2]);
+    }
+}

libs/vm_monitor/src/dispatcher.rs

@@ -12,12 +12,10 @@ use futures::{
     stream::{SplitSink, SplitStream},
     SinkExt, StreamExt,
 };
-use tokio::sync::mpsc;
 use tracing::info;
 
-use crate::cgroup::Sequenced;
 use crate::protocol::{
-    OutboundMsg, ProtocolRange, ProtocolResponse, ProtocolVersion, Resources, PROTOCOL_MAX_VERSION,
+    OutboundMsg, ProtocolRange, ProtocolResponse, ProtocolVersion, PROTOCOL_MAX_VERSION,
     PROTOCOL_MIN_VERSION,
 };
 
@@ -36,13 +34,6 @@ pub struct Dispatcher {
     /// We send messages to the agent through `sink`
     sink: SplitSink<WebSocket, Message>,
 
-    /// Used to notify the cgroup when we are upscaled.
-    pub(crate) notify_upscale_events: mpsc::Sender<Sequenced<Resources>>,
-
-    /// When the cgroup requests upscale it will send on this channel. In response
-    /// we send an `UpscaleRequst` to the agent.
-    pub(crate) request_upscale_events: mpsc::Receiver<()>,
-
     /// The protocol version we have agreed to use with the agent. This is negotiated
     /// during the creation of the dispatcher, and should be the highest shared protocol
     /// version.
@@ -61,11 +52,7 @@ impl Dispatcher {
     /// 1. Wait for the agent to sent the range of protocols it supports.
     /// 2. Send a protocol version that works for us as well, or an error if there
     ///    is no compatible version.
-    pub async fn new(
-        stream: WebSocket,
-        notify_upscale_events: mpsc::Sender<Sequenced<Resources>>,
-        request_upscale_events: mpsc::Receiver<()>,
-    ) -> anyhow::Result<Self> {
+    pub async fn new(stream: WebSocket) -> anyhow::Result<Self> {
         let (mut sink, mut source) = stream.split();
 
         // Figure out the highest protocol version we both support
@@ -119,22 +106,10 @@ impl Dispatcher {
         Ok(Self {
             sink,
             source,
-            notify_upscale_events,
-            request_upscale_events,
             proto_version: highest_shared_version,
         })
     }
 
-    /// Notify the cgroup manager that we have received upscale and wait for
-    /// the acknowledgement.
-    #[tracing::instrument(skip_all, fields(?resources))]
-    pub async fn notify_upscale(&self, resources: Sequenced<Resources>) -> anyhow::Result<()> {
-        self.notify_upscale_events
-            .send(resources)
-            .await
-            .context("failed to send resources and oneshot sender across channel")
-    }
-
     /// Send a message to the agent.
     ///
     /// Although this function is small, it has one major benefit: it is the only

libs/vm_monitor/src/runner.rs

@@ -5,18 +5,16 @@
 //! all functionality.
 
 use std::fmt::Debug;
-use std::sync::Arc;
 use std::time::{Duration, Instant};
 
 use anyhow::{bail, Context};
 use axum::extract::ws::{Message, WebSocket};
 use futures::StreamExt;
-use tokio::sync::broadcast;
-use tokio::sync::mpsc;
+use tokio::sync::{broadcast, watch};
 use tokio_util::sync::CancellationToken;
 use tracing::{error, info, warn};
 
-use crate::cgroup::{CgroupWatcher, Sequenced};
+use crate::cgroup::{self, CgroupWatcher};
 use crate::dispatcher::Dispatcher;
 use crate::filecache::{FileCacheConfig, FileCacheState};
 use crate::protocol::{InboundMsg, InboundMsgKind, OutboundMsg, OutboundMsgKind, Resources};
@@ -28,7 +26,7 @@ use crate::{bytes_to_mebibytes, get_total_system_memory, spawn_with_cancel, Args
 pub struct Runner {
     config: Config,
     filecache: Option<FileCacheState>,
-    cgroup: Option<Arc<CgroupWatcher>>,
+    cgroup: Option<CgroupState>,
     dispatcher: Dispatcher,
 
     /// We "mint" new message ids by incrementing this counter and taking the value.
@@ -45,6 +43,14 @@ pub struct Runner {
     kill: broadcast::Receiver<()>,
 }
 
+#[derive(Debug)]
+struct CgroupState {
+    watcher: watch::Receiver<(Instant, cgroup::MemoryHistory)>,
+    /// If [`cgroup::MemoryHistory::avg_non_reclaimable`] exceeds `threshold`, we send upscale
+    /// requests.
+    threshold: u64,
+}
+
 /// Configuration for a `Runner`
 #[derive(Debug)]
 pub struct Config {
@@ -62,16 +68,56 @@ pub struct Config {
     /// upscale resource amounts (because we might not *actually* have been upscaled yet). This field
     /// should be removed once we have a better solution there.
     sys_buffer_bytes: u64,
+
+    /// Minimum fraction of total system memory reserved *before* the the cgroup threshold; in
+    /// other words, providing a ceiling for the highest value of the threshold by enforcing that
+    /// there's at least `cgroup_min_overhead_fraction` of the total memory remaining beyond the
+    /// threshold.
+    ///
+    /// For example, a value of `0.1` means that 10% of total memory must remain after exceeding
+    /// the threshold, so the value of the cgroup threshold would always be capped at 90% of total
+    /// memory.
+    ///
+    /// The default value of `0.15` means that we *guarantee* sending upscale requests if the
+    /// cgroup is using more than 85% of total memory (even if we're *not* separately reserving
+    /// memory for the file cache).
+    cgroup_min_overhead_fraction: f64,
+
+    cgroup_downscale_threshold_buffer_bytes: u64,
 }
 
 impl Default for Config {
     fn default() -> Self {
         Self {
             sys_buffer_bytes: 100 * MiB,
+            cgroup_min_overhead_fraction: 0.15,
+            cgroup_downscale_threshold_buffer_bytes: 100 * MiB,
         }
     }
 }
 
+impl Config {
+    fn cgroup_threshold(&self, total_mem: u64, file_cache_disk_size: u64) -> u64 {
+        // If the file cache is in tmpfs, then it will count towards shmem usage of the cgroup,
+        // and thus be non-reclaimable, so we should allow for additional memory usage.
+        //
+        // If the file cache sits on disk, our desired stable system state is for it to be fully
+        // page cached (its contents should only be paged to/from disk in situations where we can't
+        // upscale fast enough). Page-cached memory is reclaimable, so we need to lower the
+        // threshold for non-reclaimable memory so we scale up *before* the kernel starts paging
+        // out the file cache.
+        let memory_remaining_for_cgroup = total_mem.saturating_sub(file_cache_disk_size);
+
+        // Even if we're not separately making room for the file cache (if it's in tmpfs), we still
+        // want our threshold to be met gracefully instead of letting postgres get OOM-killed.
+        // So we guarantee that there's at least `cgroup_min_overhead_fraction` of total memory
+        // remaining above the threshold.
+        let max_threshold = (total_mem as f64 * (1.0 - self.cgroup_min_overhead_fraction)) as u64;
+
+        memory_remaining_for_cgroup.min(max_threshold)
+    }
+}
+
 impl Runner {
     /// Create a new monitor.
     #[tracing::instrument(skip_all, fields(?config, ?args))]
@@ -87,12 +133,7 @@ impl Runner {
             "invalid monitor Config: sys_buffer_bytes cannot be 0"
         );
 
-        // *NOTE*: the dispatcher and cgroup manager talk through these channels
-        // so make sure they each get the correct half, nothing is droppped, etc.
-        let (notified_send, notified_recv) = mpsc::channel(1);
-        let (requesting_send, requesting_recv) = mpsc::channel(1);
-
-        let dispatcher = Dispatcher::new(ws, notified_send, requesting_recv)
+        let dispatcher = Dispatcher::new(ws)
             .await
             .context("error creating new dispatcher")?;
 
@@ -106,46 +147,10 @@ impl Runner {
             kill,
         };
 
-        // If we have both the cgroup and file cache integrations enabled, it's possible for
-        // temporary failures to result in cgroup throttling (from memory.high), that in turn makes
-        // it near-impossible to connect to the file cache (because it times out). Unfortunately,
-        // we *do* still want to determine the file cache size before setting the cgroup's
-        // memory.high, so it's not as simple as just swapping the order.
-        //
-        // Instead, the resolution here is that on vm-monitor startup (note: happens on each
-        // connection from autoscaler-agent, possibly multiple times per compute_ctl lifecycle), we
-        // temporarily unset memory.high, to allow any existing throttling to dissipate. It's a bit
-        // of a hacky solution, but helps with reliability.
-        if let Some(name) = &args.cgroup {
-            // Best not to set up cgroup stuff more than once, so we'll initialize cgroup state
-            // now, and then set limits later.
-            info!("initializing cgroup");
-
-            let (cgroup, cgroup_event_stream) = CgroupWatcher::new(name.clone(), requesting_send)
-                .context("failed to create cgroup manager")?;
-
-            info!("temporarily unsetting memory.high");
-
-            // Temporarily un-set cgroup memory.high; see above.
-            cgroup
-                .unset_memory_high()
-                .context("failed to unset memory.high")?;
-
-            let cgroup = Arc::new(cgroup);
-
-            let cgroup_clone = Arc::clone(&cgroup);
-            spawn_with_cancel(
-                token.clone(),
-                |_| error!("cgroup watcher terminated"),
-                async move { cgroup_clone.watch(notified_recv, cgroup_event_stream).await },
-            );
-
-            state.cgroup = Some(cgroup);
-        }
-
-        let mut file_cache_reserved_bytes = 0;
         let mem = get_total_system_memory();
 
+        let mut file_cache_disk_size = 0;
+
         // We need to process file cache initialization before cgroup initialization, so that the memory
         // allocated to the file cache is appropriately taken into account when we decide the cgroup's
         // memory limits.
@@ -156,7 +161,7 @@ impl Runner {
                 false => FileCacheConfig::default_in_memory(),
             };
 
-            let mut file_cache = FileCacheState::new(connstr, config, token)
+            let mut file_cache = FileCacheState::new(connstr, config, token.clone())
                 .await
                 .context("failed to create file cache")?;
 
@@ -181,23 +186,40 @@ impl Runner {
             if actual_size != new_size {
                 info!("file cache size actually got set to {actual_size}")
             }
-            // Mark the resources given to the file cache as reserved, but only if it's in memory.
-            if !args.file_cache_on_disk {
-                file_cache_reserved_bytes = actual_size;
+
+            if args.file_cache_on_disk {
+                file_cache_disk_size = actual_size;
             }
 
             state.filecache = Some(file_cache);
         }
 
-        if let Some(cgroup) = &state.cgroup {
-            let available = mem - file_cache_reserved_bytes;
-            let value = cgroup.config.calculate_memory_high_value(available);
+        if let Some(name) = &args.cgroup {
+            // Best not to set up cgroup stuff more than once, so we'll initialize cgroup state
+            // now, and then set limits later.
+            info!("initializing cgroup");
 
-            info!(value, "setting memory.high");
+            let cgroup =
+                CgroupWatcher::new(name.clone()).context("failed to create cgroup manager")?;
 
-            cgroup
-                .set_memory_high_bytes(value)
-                .context("failed to set cgroup memory.high")?;
+            let init_value = cgroup::MemoryHistory {
+                avg_non_reclaimable: 0,
+                samples_count: 0,
+                samples_span: Duration::ZERO,
+            };
+            let (hist_tx, hist_rx) = watch::channel((Instant::now(), init_value));
+
+            spawn_with_cancel(token, |_| error!("cgroup watcher terminated"), async move {
+                cgroup.watch(hist_tx).await
+            });
+
+            let threshold = state.config.cgroup_threshold(mem, file_cache_disk_size);
+            info!(threshold, "set initial cgroup threshold",);
+
+            state.cgroup = Some(CgroupState {
+                watcher: hist_rx,
+                threshold,
+            });
         }
 
         Ok(state)
@@ -217,28 +239,51 @@ impl Runner {
 
         let requested_mem = target.mem;
         let usable_system_memory = requested_mem.saturating_sub(self.config.sys_buffer_bytes);
-        let expected_file_cache_mem_usage = self
+        let (expected_file_cache_size, expected_file_cache_disk_size) = self
             .filecache
             .as_ref()
-            .map(|file_cache| file_cache.config.calculate_cache_size(usable_system_memory))
-            .unwrap_or(0);
-        let mut new_cgroup_mem_high = 0;
+            .map(|file_cache| {
+                let size = file_cache.config.calculate_cache_size(usable_system_memory);
+                match file_cache.config.in_memory {
+                    true => (size, 0),
+                    false => (size, size),
+                }
+            })
+            .unwrap_or((0, 0));
         if let Some(cgroup) = &self.cgroup {
-            new_cgroup_mem_high = cgroup
+            let (last_time, last_history) = *cgroup.watcher.borrow();
+
+            // NB: The ordering of these conditions is intentional. During startup, we should deny
+            // downscaling until we have enough information to determine that it's safe to do so
+            // (i.e. enough samples have come in). But if it's been a while and we *still* haven't
+            // received any information, we should *fail* instead of just denying downscaling.
+            //
+            // `last_time` is set to `Instant::now()` on startup, so checking `last_time.elapsed()`
+            // serves double-duty: it trips if we haven't received *any* metrics for long enough,
+            // OR if we haven't received metrics *recently enough*.
+            //
+            // TODO: make the duration here configurable.
+            if last_time.elapsed() > Duration::from_secs(5) {
+                bail!("haven't gotten cgroup memory stats recently enough to determine downscaling information");
+            } else if last_history.samples_count <= 1 {
+                let status = "haven't received enough cgroup memory stats yet";
+                info!(status, "discontinuing downscale");
+                return Ok((false, status.to_owned()));
+            }
+
+            let new_threshold = self
                 .config
-                .calculate_memory_high_value(usable_system_memory - expected_file_cache_mem_usage);
+                .cgroup_threshold(usable_system_memory, expected_file_cache_disk_size);
 
-            let current = cgroup
-                .current_memory_usage()
-                .context("failed to fetch cgroup memory")?;
+            let current = last_history.avg_non_reclaimable;
 
-            if new_cgroup_mem_high < current + cgroup.config.memory_high_buffer_bytes {
+            if new_threshold < current + self.config.cgroup_downscale_threshold_buffer_bytes {
                 let status = format!(
-                    "{}: {} MiB (new high) < {} (current usage) + {} (buffer)",
-                    "calculated memory.high too low",
-                    bytes_to_mebibytes(new_cgroup_mem_high),
+                    "{}: {} MiB (new threshold) < {} (current usage) + {} (downscale buffer)",
+                    "calculated memory threshold too low",
+                    bytes_to_mebibytes(new_threshold),
                     bytes_to_mebibytes(current),
-                    bytes_to_mebibytes(cgroup.config.memory_high_buffer_bytes)
+                    bytes_to_mebibytes(self.config.cgroup_downscale_threshold_buffer_bytes)
                 );
 
                 info!(status, "discontinuing downscale");
@@ -249,14 +294,14 @@ impl Runner {
 
         // The downscaling has been approved. Downscale the file cache, then the cgroup.
         let mut status = vec![];
-        let mut file_cache_mem_usage = 0;
+        let mut file_cache_disk_size = 0;
         if let Some(file_cache) = &mut self.filecache {
             let actual_usage = file_cache
-                .set_file_cache_size(expected_file_cache_mem_usage)
+                .set_file_cache_size(expected_file_cache_size)
                 .await
                 .context("failed to set file cache size")?;
-            if file_cache.config.in_memory {
-                file_cache_mem_usage = actual_usage;
+            if !file_cache.config.in_memory {
+                file_cache_disk_size = actual_usage;
             }
             let message = format!(
                 "set file cache size to {} MiB (in memory = {})",
@@ -267,24 +312,18 @@ impl Runner {
             status.push(message);
         }
 
-        if let Some(cgroup) = &self.cgroup {
-            let available_memory = usable_system_memory - file_cache_mem_usage;
-
-            if file_cache_mem_usage != expected_file_cache_mem_usage {
-                new_cgroup_mem_high = cgroup.config.calculate_memory_high_value(available_memory);
-            }
-
-            // new_cgroup_mem_high is initialized to 0 but it is guaranteed to not be here
-            // since it is properly initialized in the previous cgroup if let block
-            cgroup
-                .set_memory_high_bytes(new_cgroup_mem_high)
-                .context("failed to set cgroup memory.high")?;
+        if let Some(cgroup) = &mut self.cgroup {
+            let new_threshold = self
+                .config
+                .cgroup_threshold(usable_system_memory, file_cache_disk_size);
 
             let message = format!(
-                "set cgroup memory.high to {} MiB, of new max {} MiB",
-                bytes_to_mebibytes(new_cgroup_mem_high),
-                bytes_to_mebibytes(available_memory)
+                "set cgroup memory threshold from {} MiB to {} MiB, of new total {} MiB",
+                bytes_to_mebibytes(cgroup.threshold),
+                bytes_to_mebibytes(new_threshold),
+                bytes_to_mebibytes(usable_system_memory)
             );
+            cgroup.threshold = new_threshold;
             info!("downscale: {message}");
             status.push(message);
         }
@@ -305,8 +344,7 @@ impl Runner {
         let new_mem = resources.mem;
         let usable_system_memory = new_mem.saturating_sub(self.config.sys_buffer_bytes);
 
-        // Get the file cache's expected contribution to the memory usage
-        let mut file_cache_mem_usage = 0;
+        let mut file_cache_disk_size = 0;
         if let Some(file_cache) = &mut self.filecache {
             let expected_usage = file_cache.config.calculate_cache_size(usable_system_memory);
             info!(
@@ -319,8 +357,8 @@ impl Runner {
                 .set_file_cache_size(expected_usage)
                 .await
                 .context("failed to set file cache size")?;
-            if file_cache.config.in_memory {
-                file_cache_mem_usage = actual_usage;
+            if !file_cache.config.in_memory {
+                file_cache_disk_size = actual_usage;
             }
 
             if actual_usage != expected_usage {
@@ -332,18 +370,18 @@ impl Runner {
             }
         }
 
-        if let Some(cgroup) = &self.cgroup {
-            let available_memory = usable_system_memory - file_cache_mem_usage;
-            let new_cgroup_mem_high = cgroup.config.calculate_memory_high_value(available_memory);
+        if let Some(cgroup) = &mut self.cgroup {
+            let new_threshold = self
+                .config
+                .cgroup_threshold(usable_system_memory, file_cache_disk_size);
+
             info!(
-                target = bytes_to_mebibytes(new_cgroup_mem_high),
-                total = bytes_to_mebibytes(new_mem),
-                name = cgroup.path(),
-                "updating cgroup memory.high",
+                "set cgroup memory threshold from {} MiB to {} MiB of new total {} MiB",
+                bytes_to_mebibytes(cgroup.threshold),
+                bytes_to_mebibytes(new_threshold),
+                bytes_to_mebibytes(usable_system_memory)
             );
-            cgroup
-                .set_memory_high_bytes(new_cgroup_mem_high)
-                .context("failed to set cgroup memory.high")?;
+            cgroup.threshold = new_threshold;
         }
 
         Ok(())
@@ -361,10 +399,6 @@ impl Runner {
                 self.handle_upscale(granted)
                     .await
                     .context("failed to handle upscale")?;
-                self.dispatcher
-                    .notify_upscale(Sequenced::new(granted))
-                    .await
-                    .context("failed to notify notify cgroup of upscale")?;
                 Ok(Some(OutboundMsg::new(
                     OutboundMsgKind::UpscaleConfirmation {},
                     id,
@@ -408,33 +442,53 @@ impl Runner {
                         Err(e) => bail!("failed to receive kill signal: {e}")
                     }
                 }
-                // we need to propagate an upscale request
-                request = self.dispatcher.request_upscale_events.recv(), if self.cgroup.is_some() => {
-                    if request.is_none() {
-                        bail!("failed to listen for upscale event from cgroup")
+
+                // New memory stats from the cgroup, *may* need to request upscaling, if we've
+                // exceeded the threshold
+                result = self.cgroup.as_mut().unwrap().watcher.changed(), if self.cgroup.is_some() => {
+                    result.context("failed to receive from cgroup memory stats watcher")?;
+
+                    let cgroup = self.cgroup.as_ref().unwrap();
+
+                    let (_time, cgroup_mem_stat) = *cgroup.watcher.borrow();
+
+                    // If we haven't exceeded the threshold, then we're all ok
+                    if cgroup_mem_stat.avg_non_reclaimable < cgroup.threshold {
+                        continue;
                     }
 
-                    // If it's been less than 1 second since the last time we requested upscaling,
-                    // ignore the event, to avoid spamming the agent (otherwise, this can happen
-                    // ~1k times per second).
+                    // Otherwise, we generally want upscaling. But, if it's been less than 1 second
+                    // since the last time we requested upscaling, ignore the event, to avoid
+                    // spamming the agent.
                     if let Some(t) = self.last_upscale_request_at {
                         let elapsed = t.elapsed();
                         if elapsed < Duration::from_secs(1) {
-                            info!(elapsed_millis = elapsed.as_millis(), "cgroup asked for upscale but too soon to forward the request, ignoring");
+                            info!(
+                                elapsed_millis = elapsed.as_millis(),
+                                avg_non_reclaimable = bytes_to_mebibytes(cgroup_mem_stat.avg_non_reclaimable),
+                                threshold = bytes_to_mebibytes(cgroup.threshold),
+                                "cgroup memory stats are high enough to upscale but too soon to forward the request, ignoring",
+                            );
                             continue;
                         }
                     }
 
                     self.last_upscale_request_at = Some(Instant::now());
 
-                    info!("cgroup asking for upscale; forwarding request");
+                    info!(
+                        avg_non_reclaimable = bytes_to_mebibytes(cgroup_mem_stat.avg_non_reclaimable),
+                        threshold = bytes_to_mebibytes(cgroup.threshold),
+                        "cgroup memory stats are high enough to upscale, requesting upscale",
+                    );
+
                     self.counter += 2; // Increment, preserving parity (i.e. keep the
                                        // counter odd). See the field comment for more.
                     self.dispatcher
                         .send(OutboundMsg::new(OutboundMsgKind::UpscaleRequest {}, self.counter))
                         .await
                         .context("failed to send message")?;
-                }
+                },
+
                 // there is a message from the agent
                 msg = self.dispatcher.source.next() => {
                     if let Some(msg) = msg {
@@ -462,11 +516,14 @@ impl Runner {
                                     Ok(Some(out)) => out,
                                     Ok(None) => continue,
                                     Err(e) => {
-                                        let error = e.to_string();
-                                        warn!(?error, "error handling message");
+                                        // use {:#} for our logging because the display impl only
+                                        // gives the outermost cause, and the debug impl
+                                        // pretty-prints the error, whereas {:#} contains all the
+                                        // causes, but is compact (no newlines).
+                                        warn!(error = format!("{e:#}"), "error handling message");
                                         OutboundMsg::new(
                                             OutboundMsgKind::InternalError {
-                                                error
+                                                error: e.to_string(),
                                             },
                                             message.id
                                         )

libs/walproposer/Cargo.toml

@@ -0,0 +1,16 @@
+[package]
+name = "walproposer"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+
+[dependencies]
+anyhow.workspace = true
+utils.workspace = true
+postgres_ffi.workspace = true
+
+workspace_hack.workspace = true
+
+[build-dependencies]
+anyhow.workspace = true
+bindgen.workspace = true

libs/walproposer/bindgen_deps.h

@@ -0,0 +1 @@
+#include "walproposer.h"

libs/walproposer/build.rs

@@ -0,0 +1,113 @@
+use std::{env, path::PathBuf, process::Command};
+
+use anyhow::{anyhow, Context};
+use bindgen::CargoCallbacks;
+
+fn main() -> anyhow::Result<()> {
+    // Tell cargo to invalidate the built crate whenever the wrapper changes
+    println!("cargo:rerun-if-changed=bindgen_deps.h");
+
+    // Finding the location of built libraries and Postgres C headers:
+    // - if POSTGRES_INSTALL_DIR is set look into it, otherwise look into `<project_root>/pg_install`
+    // - if there's a `bin/pg_config` file use it for getting include server, otherwise use `<project_root>/pg_install/{PG_MAJORVERSION}/include/postgresql/server`
+    let pg_install_dir = if let Some(postgres_install_dir) = env::var_os("POSTGRES_INSTALL_DIR") {
+        postgres_install_dir.into()
+    } else {
+        PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("../../pg_install")
+    };
+
+    let pg_install_abs = std::fs::canonicalize(pg_install_dir)?;
+    let walproposer_lib_dir = pg_install_abs.join("build/walproposer-lib");
+    let walproposer_lib_search_str = walproposer_lib_dir
+        .to_str()
+        .ok_or(anyhow!("Bad non-UTF path"))?;
+
+    let pgxn_neon = PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("../../pgxn/neon");
+    let pgxn_neon = std::fs::canonicalize(pgxn_neon)?;
+    let pgxn_neon = pgxn_neon.to_str().ok_or(anyhow!("Bad non-UTF path"))?;
+
+    println!("cargo:rustc-link-lib=static=pgport");
+    println!("cargo:rustc-link-lib=static=pgcommon");
+    println!("cargo:rustc-link-lib=static=walproposer");
+    println!("cargo:rustc-link-search={walproposer_lib_search_str}");
+
+    let pg_config_bin = pg_install_abs.join("v16").join("bin").join("pg_config");
+    let inc_server_path: String = if pg_config_bin.exists() {
+        let output = Command::new(pg_config_bin)
+            .arg("--includedir-server")
+            .output()
+            .context("failed to execute `pg_config --includedir-server`")?;
+
+        if !output.status.success() {
+            panic!("`pg_config --includedir-server` failed")
+        }
+
+        String::from_utf8(output.stdout)
+            .context("pg_config output is not UTF-8")?
+            .trim_end()
+            .into()
+    } else {
+        let server_path = pg_install_abs
+            .join("v16")
+            .join("include")
+            .join("postgresql")
+            .join("server")
+            .into_os_string();
+        server_path
+            .into_string()
+            .map_err(|s| anyhow!("Bad postgres server path {s:?}"))?
+    };
+
+    // The bindgen::Builder is the main entry point
+    // to bindgen, and lets you build up options for
+    // the resulting bindings.
+    let bindings = bindgen::Builder::default()
+        // The input header we would like to generate
+        // bindings for.
+        .header("bindgen_deps.h")
+        // Tell cargo to invalidate the built crate whenever any of the
+        // included header files changed.
+        .parse_callbacks(Box::new(CargoCallbacks))
+        .allowlist_type("WalProposer")
+        .allowlist_type("WalProposerConfig")
+        .allowlist_type("walproposer_api")
+        .allowlist_function("WalProposerCreate")
+        .allowlist_function("WalProposerStart")
+        .allowlist_function("WalProposerBroadcast")
+        .allowlist_function("WalProposerPoll")
+        .allowlist_function("WalProposerFree")
+        .allowlist_var("DEBUG5")
+        .allowlist_var("DEBUG4")
+        .allowlist_var("DEBUG3")
+        .allowlist_var("DEBUG2")
+        .allowlist_var("DEBUG1")
+        .allowlist_var("LOG")
+        .allowlist_var("INFO")
+        .allowlist_var("NOTICE")
+        .allowlist_var("WARNING")
+        .allowlist_var("ERROR")
+        .allowlist_var("FATAL")
+        .allowlist_var("PANIC")
+        .allowlist_var("WPEVENT")
+        .allowlist_var("WL_LATCH_SET")
+        .allowlist_var("WL_SOCKET_READABLE")
+        .allowlist_var("WL_SOCKET_WRITEABLE")
+        .allowlist_var("WL_TIMEOUT")
+        .allowlist_var("WL_SOCKET_CLOSED")
+        .allowlist_var("WL_SOCKET_MASK")
+        .clang_arg("-DWALPROPOSER_LIB")
+        .clang_arg(format!("-I{pgxn_neon}"))
+        .clang_arg(format!("-I{inc_server_path}"))
+        // Finish the builder and generate the bindings.
+        .generate()
+        // Unwrap the Result and panic on failure.
+        .expect("Unable to generate bindings");
+
+    // Write the bindings to the $OUT_DIR/bindings.rs file.
+    let out_path = PathBuf::from(env::var("OUT_DIR").unwrap()).join("bindings.rs");
+    bindings
+        .write_to_file(out_path)
+        .expect("Couldn't write bindings!");
+
+    Ok(())
+}

libs/walproposer/src/api_bindings.rs

@@ -0,0 +1,455 @@
+#![allow(dead_code)]
+
+use std::ffi::CStr;
+use std::ffi::CString;
+
+use crate::bindings::uint32;
+use crate::bindings::walproposer_api;
+use crate::bindings::PGAsyncReadResult;
+use crate::bindings::PGAsyncWriteResult;
+use crate::bindings::Safekeeper;
+use crate::bindings::Size;
+use crate::bindings::StringInfoData;
+use crate::bindings::TimeLineID;
+use crate::bindings::TimestampTz;
+use crate::bindings::WalProposer;
+use crate::bindings::WalProposerConnStatusType;
+use crate::bindings::WalProposerConnectPollStatusType;
+use crate::bindings::WalProposerExecStatusType;
+use crate::bindings::WalproposerShmemState;
+use crate::bindings::XLogRecPtr;
+use crate::walproposer::ApiImpl;
+use crate::walproposer::WaitResult;
+
+extern "C" fn get_shmem_state(wp: *mut WalProposer) -> *mut WalproposerShmemState {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).get_shmem_state()
+    }
+}
+
+extern "C" fn start_streaming(wp: *mut WalProposer, startpos: XLogRecPtr) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).start_streaming(startpos)
+    }
+}
+
+extern "C" fn get_flush_rec_ptr(wp: *mut WalProposer) -> XLogRecPtr {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).get_flush_rec_ptr()
+    }
+}
+
+extern "C" fn get_current_timestamp(wp: *mut WalProposer) -> TimestampTz {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).get_current_timestamp()
+    }
+}
+
+extern "C" fn conn_error_message(sk: *mut Safekeeper) -> *mut ::std::os::raw::c_char {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        let msg = (*api).conn_error_message(&mut (*sk));
+        let msg = CString::new(msg).unwrap();
+        // TODO: fix leaking error message
+        msg.into_raw()
+    }
+}
+
+extern "C" fn conn_status(sk: *mut Safekeeper) -> WalProposerConnStatusType {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_status(&mut (*sk))
+    }
+}
+
+extern "C" fn conn_connect_start(sk: *mut Safekeeper) {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_connect_start(&mut (*sk))
+    }
+}
+
+extern "C" fn conn_connect_poll(sk: *mut Safekeeper) -> WalProposerConnectPollStatusType {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_connect_poll(&mut (*sk))
+    }
+}
+
+extern "C" fn conn_send_query(sk: *mut Safekeeper, query: *mut ::std::os::raw::c_char) -> bool {
+    let query = unsafe { CStr::from_ptr(query) };
+    let query = query.to_str().unwrap();
+
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_send_query(&mut (*sk), query)
+    }
+}
+
+extern "C" fn conn_get_query_result(sk: *mut Safekeeper) -> WalProposerExecStatusType {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_get_query_result(&mut (*sk))
+    }
+}
+
+extern "C" fn conn_flush(sk: *mut Safekeeper) -> ::std::os::raw::c_int {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_flush(&mut (*sk))
+    }
+}
+
+extern "C" fn conn_finish(sk: *mut Safekeeper) {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_finish(&mut (*sk))
+    }
+}
+
+extern "C" fn conn_async_read(
+    sk: *mut Safekeeper,
+    buf: *mut *mut ::std::os::raw::c_char,
+    amount: *mut ::std::os::raw::c_int,
+) -> PGAsyncReadResult {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        let (res, result) = (*api).conn_async_read(&mut (*sk));
+
+        // This function has guarantee that returned buf will be valid until
+        // the next call. So we can store a Vec in each Safekeeper and reuse
+        // it on the next call.
+        let mut inbuf = take_vec_u8(&mut (*sk).inbuf).unwrap_or_default();
+
+        inbuf.clear();
+        inbuf.extend_from_slice(res);
+
+        // Put a Vec back to sk->inbuf and return data ptr.
+        *buf = store_vec_u8(&mut (*sk).inbuf, inbuf);
+        *amount = res.len() as i32;
+
+        result
+    }
+}
+
+extern "C" fn conn_async_write(
+    sk: *mut Safekeeper,
+    buf: *const ::std::os::raw::c_void,
+    size: usize,
+) -> PGAsyncWriteResult {
+    unsafe {
+        let buf = std::slice::from_raw_parts(buf as *const u8, size);
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_async_write(&mut (*sk), buf)
+    }
+}
+
+extern "C" fn conn_blocking_write(
+    sk: *mut Safekeeper,
+    buf: *const ::std::os::raw::c_void,
+    size: usize,
+) -> bool {
+    unsafe {
+        let buf = std::slice::from_raw_parts(buf as *const u8, size);
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_blocking_write(&mut (*sk), buf)
+    }
+}
+
+extern "C" fn recovery_download(
+    sk: *mut Safekeeper,
+    _timeline: TimeLineID,
+    startpos: XLogRecPtr,
+    endpos: XLogRecPtr,
+) -> bool {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).recovery_download(&mut (*sk), startpos, endpos)
+    }
+}
+
+extern "C" fn wal_read(
+    sk: *mut Safekeeper,
+    buf: *mut ::std::os::raw::c_char,
+    startptr: XLogRecPtr,
+    count: Size,
+) {
+    unsafe {
+        let buf = std::slice::from_raw_parts_mut(buf as *mut u8, count);
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).wal_read(&mut (*sk), buf, startptr)
+    }
+}
+
+extern "C" fn wal_reader_allocate(sk: *mut Safekeeper) {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).wal_reader_allocate(&mut (*sk));
+    }
+}
+
+extern "C" fn free_event_set(wp: *mut WalProposer) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).free_event_set(&mut (*wp));
+    }
+}
+
+extern "C" fn init_event_set(wp: *mut WalProposer) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).init_event_set(&mut (*wp));
+    }
+}
+
+extern "C" fn update_event_set(sk: *mut Safekeeper, events: uint32) {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).update_event_set(&mut (*sk), events);
+    }
+}
+
+extern "C" fn add_safekeeper_event_set(sk: *mut Safekeeper, events: uint32) {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).add_safekeeper_event_set(&mut (*sk), events);
+    }
+}
+
+extern "C" fn wait_event_set(
+    wp: *mut WalProposer,
+    timeout: ::std::os::raw::c_long,
+    event_sk: *mut *mut Safekeeper,
+    events: *mut uint32,
+) -> ::std::os::raw::c_int {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        let result = (*api).wait_event_set(&mut (*wp), timeout);
+        match result {
+            WaitResult::Latch => {
+                *event_sk = std::ptr::null_mut();
+                *events = crate::bindings::WL_LATCH_SET;
+                1
+            }
+            WaitResult::Timeout => {
+                *event_sk = std::ptr::null_mut();
+                *events = crate::bindings::WL_TIMEOUT;
+                0
+            }
+            WaitResult::Network(sk, event_mask) => {
+                *event_sk = sk;
+                *events = event_mask;
+                1
+            }
+        }
+    }
+}
+
+extern "C" fn strong_random(
+    wp: *mut WalProposer,
+    buf: *mut ::std::os::raw::c_void,
+    len: usize,
+) -> bool {
+    unsafe {
+        let buf = std::slice::from_raw_parts_mut(buf as *mut u8, len);
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).strong_random(buf)
+    }
+}
+
+extern "C" fn get_redo_start_lsn(wp: *mut WalProposer) -> XLogRecPtr {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).get_redo_start_lsn()
+    }
+}
+
+extern "C" fn finish_sync_safekeepers(wp: *mut WalProposer, lsn: XLogRecPtr) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).finish_sync_safekeepers(lsn)
+    }
+}
+
+extern "C" fn process_safekeeper_feedback(wp: *mut WalProposer, commit_lsn: XLogRecPtr) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).process_safekeeper_feedback(&mut (*wp), commit_lsn)
+    }
+}
+
+extern "C" fn confirm_wal_streamed(wp: *mut WalProposer, lsn: XLogRecPtr) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).confirm_wal_streamed(&mut (*wp), lsn)
+    }
+}
+
+extern "C" fn log_internal(
+    wp: *mut WalProposer,
+    level: ::std::os::raw::c_int,
+    line: *const ::std::os::raw::c_char,
+) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        let line = CStr::from_ptr(line);
+        let line = line.to_str().unwrap();
+        (*api).log_internal(&mut (*wp), Level::from(level as u32), line)
+    }
+}
+
+extern "C" fn after_election(wp: *mut WalProposer) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).after_election(&mut (*wp))
+    }
+}
+
+#[derive(Debug)]
+pub enum Level {
+    Debug5,
+    Debug4,
+    Debug3,
+    Debug2,
+    Debug1,
+    Log,
+    Info,
+    Notice,
+    Warning,
+    Error,
+    Fatal,
+    Panic,
+    WPEvent,
+}
+
+impl Level {
+    pub fn from(elevel: u32) -> Level {
+        use crate::bindings::*;
+
+        match elevel {
+            DEBUG5 => Level::Debug5,
+            DEBUG4 => Level::Debug4,
+            DEBUG3 => Level::Debug3,
+            DEBUG2 => Level::Debug2,
+            DEBUG1 => Level::Debug1,
+            LOG => Level::Log,
+            INFO => Level::Info,
+            NOTICE => Level::Notice,
+            WARNING => Level::Warning,
+            ERROR => Level::Error,
+            FATAL => Level::Fatal,
+            PANIC => Level::Panic,
+            WPEVENT => Level::WPEvent,
+            _ => panic!("unknown log level {}", elevel),
+        }
+    }
+}
+
+pub(crate) fn create_api() -> walproposer_api {
+    walproposer_api {
+        get_shmem_state: Some(get_shmem_state),
+        start_streaming: Some(start_streaming),
+        get_flush_rec_ptr: Some(get_flush_rec_ptr),
+        get_current_timestamp: Some(get_current_timestamp),
+        conn_error_message: Some(conn_error_message),
+        conn_status: Some(conn_status),
+        conn_connect_start: Some(conn_connect_start),
+        conn_connect_poll: Some(conn_connect_poll),
+        conn_send_query: Some(conn_send_query),
+        conn_get_query_result: Some(conn_get_query_result),
+        conn_flush: Some(conn_flush),
+        conn_finish: Some(conn_finish),
+        conn_async_read: Some(conn_async_read),
+        conn_async_write: Some(conn_async_write),
+        conn_blocking_write: Some(conn_blocking_write),
+        recovery_download: Some(recovery_download),
+        wal_read: Some(wal_read),
+        wal_reader_allocate: Some(wal_reader_allocate),
+        free_event_set: Some(free_event_set),
+        init_event_set: Some(init_event_set),
+        update_event_set: Some(update_event_set),
+        add_safekeeper_event_set: Some(add_safekeeper_event_set),
+        wait_event_set: Some(wait_event_set),
+        strong_random: Some(strong_random),
+        get_redo_start_lsn: Some(get_redo_start_lsn),
+        finish_sync_safekeepers: Some(finish_sync_safekeepers),
+        process_safekeeper_feedback: Some(process_safekeeper_feedback),
+        confirm_wal_streamed: Some(confirm_wal_streamed),
+        log_internal: Some(log_internal),
+        after_election: Some(after_election),
+    }
+}
+
+impl std::fmt::Display for Level {
+    fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
+        write!(f, "{:?}", self)
+    }
+}
+
+/// Take ownership of `Vec<u8>` from StringInfoData.
+pub(crate) fn take_vec_u8(pg: &mut StringInfoData) -> Option<Vec<u8>> {
+    if pg.data.is_null() {
+        return None;
+    }
+
+    let ptr = pg.data as *mut u8;
+    let length = pg.len as usize;
+    let capacity = pg.maxlen as usize;
+
+    pg.data = std::ptr::null_mut();
+    pg.len = 0;
+    pg.maxlen = 0;
+
+    unsafe { Some(Vec::from_raw_parts(ptr, length, capacity)) }
+}
+
+/// Store `Vec<u8>` in StringInfoData.
+fn store_vec_u8(pg: &mut StringInfoData, vec: Vec<u8>) -> *mut ::std::os::raw::c_char {
+    let ptr = vec.as_ptr() as *mut ::std::os::raw::c_char;
+    let length = vec.len();
+    let capacity = vec.capacity();
+
+    assert!(pg.data.is_null());
+
+    pg.data = ptr;
+    pg.len = length as i32;
+    pg.maxlen = capacity as i32;
+
+    std::mem::forget(vec);
+
+    ptr
+}

libs/walproposer/src/lib.rs

@@ -0,0 +1,14 @@
+pub mod bindings {
+    #![allow(non_upper_case_globals)]
+    #![allow(non_camel_case_types)]
+    #![allow(non_snake_case)]
+    // bindgen creates some unsafe code with no doc comments.
+    #![allow(clippy::missing_safety_doc)]
+    // noted at 1.63 that in many cases there's a u32 -> u32 transmutes in bindgen code.
+    #![allow(clippy::useless_transmute)]
+
+    include!(concat!(env!("OUT_DIR"), "/bindings.rs"));
+}
+
+pub mod api_bindings;
+pub mod walproposer;

libs/walproposer/src/walproposer.rs

@@ -0,0 +1,485 @@
+use std::ffi::CString;
+
+use postgres_ffi::WAL_SEGMENT_SIZE;
+use utils::id::TenantTimelineId;
+
+use crate::{
+    api_bindings::{create_api, take_vec_u8, Level},
+    bindings::{
+        Safekeeper, WalProposer, WalProposerConfig, WalProposerCreate, WalProposerFree,
+        WalProposerStart,
+    },
+};
+
+/// Rust high-level wrapper for C walproposer API. Many methods are not required
+/// for simple cases, hence todo!() in default implementations.
+///
+/// Refer to `pgxn/neon/walproposer.h` for documentation.
+pub trait ApiImpl {
+    fn get_shmem_state(&self) -> &mut crate::bindings::WalproposerShmemState {
+        todo!()
+    }
+
+    fn start_streaming(&self, _startpos: u64) {
+        todo!()
+    }
+
+    fn get_flush_rec_ptr(&self) -> u64 {
+        todo!()
+    }
+
+    fn get_current_timestamp(&self) -> i64 {
+        todo!()
+    }
+
+    fn conn_error_message(&self, _sk: &mut Safekeeper) -> String {
+        todo!()
+    }
+
+    fn conn_status(&self, _sk: &mut Safekeeper) -> crate::bindings::WalProposerConnStatusType {
+        todo!()
+    }
+
+    fn conn_connect_start(&self, _sk: &mut Safekeeper) {
+        todo!()
+    }
+
+    fn conn_connect_poll(
+        &self,
+        _sk: &mut Safekeeper,
+    ) -> crate::bindings::WalProposerConnectPollStatusType {
+        todo!()
+    }
+
+    fn conn_send_query(&self, _sk: &mut Safekeeper, _query: &str) -> bool {
+        todo!()
+    }
+
+    fn conn_get_query_result(
+        &self,
+        _sk: &mut Safekeeper,
+    ) -> crate::bindings::WalProposerExecStatusType {
+        todo!()
+    }
+
+    fn conn_flush(&self, _sk: &mut Safekeeper) -> i32 {
+        todo!()
+    }
+
+    fn conn_finish(&self, _sk: &mut Safekeeper) {
+        todo!()
+    }
+
+    fn conn_async_read(&self, _sk: &mut Safekeeper) -> (&[u8], crate::bindings::PGAsyncReadResult) {
+        todo!()
+    }
+
+    fn conn_async_write(
+        &self,
+        _sk: &mut Safekeeper,
+        _buf: &[u8],
+    ) -> crate::bindings::PGAsyncWriteResult {
+        todo!()
+    }
+
+    fn conn_blocking_write(&self, _sk: &mut Safekeeper, _buf: &[u8]) -> bool {
+        todo!()
+    }
+
+    fn recovery_download(&self, _sk: &mut Safekeeper, _startpos: u64, _endpos: u64) -> bool {
+        todo!()
+    }
+
+    fn wal_read(&self, _sk: &mut Safekeeper, _buf: &mut [u8], _startpos: u64) {
+        todo!()
+    }
+
+    fn wal_reader_allocate(&self, _sk: &mut Safekeeper) {
+        todo!()
+    }
+
+    fn free_event_set(&self, _wp: &mut WalProposer) {
+        todo!()
+    }
+
+    fn init_event_set(&self, _wp: &mut WalProposer) {
+        todo!()
+    }
+
+    fn update_event_set(&self, _sk: &mut Safekeeper, _events_mask: u32) {
+        todo!()
+    }
+
+    fn add_safekeeper_event_set(&self, _sk: &mut Safekeeper, _events_mask: u32) {
+        todo!()
+    }
+
+    fn wait_event_set(&self, _wp: &mut WalProposer, _timeout_millis: i64) -> WaitResult {
+        todo!()
+    }
+
+    fn strong_random(&self, _buf: &mut [u8]) -> bool {
+        todo!()
+    }
+
+    fn get_redo_start_lsn(&self) -> u64 {
+        todo!()
+    }
+
+    fn finish_sync_safekeepers(&self, _lsn: u64) {
+        todo!()
+    }
+
+    fn process_safekeeper_feedback(&self, _wp: &mut WalProposer, _commit_lsn: u64) {
+        todo!()
+    }
+
+    fn confirm_wal_streamed(&self, _wp: &mut WalProposer, _lsn: u64) {
+        todo!()
+    }
+
+    fn log_internal(&self, _wp: &mut WalProposer, _level: Level, _msg: &str) {
+        todo!()
+    }
+
+    fn after_election(&self, _wp: &mut WalProposer) {
+        todo!()
+    }
+}
+
+pub enum WaitResult {
+    Latch,
+    Timeout,
+    Network(*mut Safekeeper, u32),
+}
+
+pub struct Config {
+    /// Tenant and timeline id
+    pub ttid: TenantTimelineId,
+    /// List of safekeepers in format `host:port`
+    pub safekeepers_list: Vec<String>,
+    /// Safekeeper reconnect timeout in milliseconds
+    pub safekeeper_reconnect_timeout: i32,
+    /// Safekeeper connection timeout in milliseconds
+    pub safekeeper_connection_timeout: i32,
+    /// walproposer mode, finish when all safekeepers are synced or subscribe
+    /// to WAL streaming
+    pub sync_safekeepers: bool,
+}
+
+/// WalProposer main struct. C methods are reexported as Rust functions.
+pub struct Wrapper {
+    wp: *mut WalProposer,
+    _safekeepers_list_vec: Vec<u8>,
+}
+
+impl Wrapper {
+    pub fn new(api: Box<dyn ApiImpl>, config: Config) -> Wrapper {
+        let neon_tenant = CString::new(config.ttid.tenant_id.to_string())
+            .unwrap()
+            .into_raw();
+        let neon_timeline = CString::new(config.ttid.timeline_id.to_string())
+            .unwrap()
+            .into_raw();
+
+        let mut safekeepers_list_vec = CString::new(config.safekeepers_list.join(","))
+            .unwrap()
+            .into_bytes_with_nul();
+        assert!(safekeepers_list_vec.len() == safekeepers_list_vec.capacity());
+        let safekeepers_list = safekeepers_list_vec.as_mut_ptr() as *mut i8;
+
+        let callback_data = Box::into_raw(Box::new(api)) as *mut ::std::os::raw::c_void;
+
+        let c_config = WalProposerConfig {
+            neon_tenant,
+            neon_timeline,
+            safekeepers_list,
+            safekeeper_reconnect_timeout: config.safekeeper_reconnect_timeout,
+            safekeeper_connection_timeout: config.safekeeper_connection_timeout,
+            wal_segment_size: WAL_SEGMENT_SIZE as i32, // default 16MB
+            syncSafekeepers: config.sync_safekeepers,
+            systemId: 0,
+            pgTimeline: 1,
+            callback_data,
+        };
+        let c_config = Box::into_raw(Box::new(c_config));
+
+        let api = create_api();
+        let wp = unsafe { WalProposerCreate(c_config, api) };
+        Wrapper {
+            wp,
+            _safekeepers_list_vec: safekeepers_list_vec,
+        }
+    }
+
+    pub fn start(&self) {
+        unsafe { WalProposerStart(self.wp) }
+    }
+}
+
+impl Drop for Wrapper {
+    fn drop(&mut self) {
+        unsafe {
+            let config = (*self.wp).config;
+            drop(Box::from_raw(
+                (*config).callback_data as *mut Box<dyn ApiImpl>,
+            ));
+            drop(CString::from_raw((*config).neon_tenant));
+            drop(CString::from_raw((*config).neon_timeline));
+            drop(Box::from_raw(config));
+
+            for i in 0..(*self.wp).n_safekeepers {
+                let sk = &mut (*self.wp).safekeeper[i as usize];
+                take_vec_u8(&mut sk.inbuf);
+            }
+
+            WalProposerFree(self.wp);
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::{
+        cell::Cell,
+        sync::{atomic::AtomicUsize, mpsc::sync_channel},
+    };
+
+    use utils::id::TenantTimelineId;
+
+    use crate::{api_bindings::Level, walproposer::Wrapper};
+
+    use super::ApiImpl;
+
+    #[derive(Clone, Copy, Debug)]
+    struct WaitEventsData {
+        sk: *mut crate::bindings::Safekeeper,
+        event_mask: u32,
+    }
+
+    struct MockImpl {
+        // data to return from wait_event_set
+        wait_events: Cell<WaitEventsData>,
+        // walproposer->safekeeper messages
+        expected_messages: Vec<Vec<u8>>,
+        expected_ptr: AtomicUsize,
+        // safekeeper->walproposer messages
+        safekeeper_replies: Vec<Vec<u8>>,
+        replies_ptr: AtomicUsize,
+        // channel to send LSN to the main thread
+        sync_channel: std::sync::mpsc::SyncSender<u64>,
+    }
+
+    impl MockImpl {
+        fn check_walproposer_msg(&self, msg: &[u8]) {
+            let ptr = self
+                .expected_ptr
+                .fetch_add(1, std::sync::atomic::Ordering::SeqCst);
+
+            if ptr >= self.expected_messages.len() {
+                panic!("unexpected message from walproposer");
+            }
+
+            let expected_msg = &self.expected_messages[ptr];
+            assert_eq!(msg, expected_msg.as_slice());
+        }
+
+        fn next_safekeeper_reply(&self) -> &[u8] {
+            let ptr = self
+                .replies_ptr
+                .fetch_add(1, std::sync::atomic::Ordering::SeqCst);
+
+            if ptr >= self.safekeeper_replies.len() {
+                panic!("no more safekeeper replies");
+            }
+
+            &self.safekeeper_replies[ptr]
+        }
+    }
+
+    impl ApiImpl for MockImpl {
+        fn get_current_timestamp(&self) -> i64 {
+            println!("get_current_timestamp");
+            0
+        }
+
+        fn conn_status(
+            &self,
+            _: &mut crate::bindings::Safekeeper,
+        ) -> crate::bindings::WalProposerConnStatusType {
+            println!("conn_status");
+            crate::bindings::WalProposerConnStatusType_WP_CONNECTION_OK
+        }
+
+        fn conn_connect_start(&self, _: &mut crate::bindings::Safekeeper) {
+            println!("conn_connect_start");
+        }
+
+        fn conn_connect_poll(
+            &self,
+            _: &mut crate::bindings::Safekeeper,
+        ) -> crate::bindings::WalProposerConnectPollStatusType {
+            println!("conn_connect_poll");
+            crate::bindings::WalProposerConnectPollStatusType_WP_CONN_POLLING_OK
+        }
+
+        fn conn_send_query(&self, _: &mut crate::bindings::Safekeeper, query: &str) -> bool {
+            println!("conn_send_query: {}", query);
+            true
+        }
+
+        fn conn_get_query_result(
+            &self,
+            _: &mut crate::bindings::Safekeeper,
+        ) -> crate::bindings::WalProposerExecStatusType {
+            println!("conn_get_query_result");
+            crate::bindings::WalProposerExecStatusType_WP_EXEC_SUCCESS_COPYBOTH
+        }
+
+        fn conn_async_read(
+            &self,
+            _: &mut crate::bindings::Safekeeper,
+        ) -> (&[u8], crate::bindings::PGAsyncReadResult) {
+            println!("conn_async_read");
+            let reply = self.next_safekeeper_reply();
+            println!("conn_async_read result: {:?}", reply);
+            (
+                reply,
+                crate::bindings::PGAsyncReadResult_PG_ASYNC_READ_SUCCESS,
+            )
+        }
+
+        fn conn_blocking_write(&self, _: &mut crate::bindings::Safekeeper, buf: &[u8]) -> bool {
+            println!("conn_blocking_write: {:?}", buf);
+            self.check_walproposer_msg(buf);
+            true
+        }
+
+        fn wal_reader_allocate(&self, _: &mut crate::bindings::Safekeeper) {
+            println!("wal_reader_allocate")
+        }
+
+        fn free_event_set(&self, _: &mut crate::bindings::WalProposer) {
+            println!("free_event_set")
+        }
+
+        fn init_event_set(&self, _: &mut crate::bindings::WalProposer) {
+            println!("init_event_set")
+        }
+
+        fn update_event_set(&self, sk: &mut crate::bindings::Safekeeper, event_mask: u32) {
+            println!(
+                "update_event_set, sk={:?}, events_mask={:#b}",
+                sk as *mut crate::bindings::Safekeeper, event_mask
+            );
+            self.wait_events.set(WaitEventsData { sk, event_mask });
+        }
+
+        fn add_safekeeper_event_set(&self, sk: &mut crate::bindings::Safekeeper, event_mask: u32) {
+            println!(
+                "add_safekeeper_event_set, sk={:?}, events_mask={:#b}",
+                sk as *mut crate::bindings::Safekeeper, event_mask
+            );
+            self.wait_events.set(WaitEventsData { sk, event_mask });
+        }
+
+        fn wait_event_set(
+            &self,
+            _: &mut crate::bindings::WalProposer,
+            timeout_millis: i64,
+        ) -> super::WaitResult {
+            let data = self.wait_events.get();
+            println!(
+                "wait_event_set, timeout_millis={}, res={:?}",
+                timeout_millis, data
+            );
+            super::WaitResult::Network(data.sk, data.event_mask)
+        }
+
+        fn strong_random(&self, buf: &mut [u8]) -> bool {
+            println!("strong_random");
+            buf.fill(0);
+            true
+        }
+
+        fn finish_sync_safekeepers(&self, lsn: u64) {
+            self.sync_channel.send(lsn).unwrap();
+            panic!("sync safekeepers finished at lsn={}", lsn);
+        }
+
+        fn log_internal(&self, _wp: &mut crate::bindings::WalProposer, level: Level, msg: &str) {
+            println!("walprop_log[{}] {}", level, msg);
+        }
+
+        fn after_election(&self, _wp: &mut crate::bindings::WalProposer) {
+            println!("after_election");
+        }
+    }
+
+    /// Test that walproposer can successfully connect to safekeeper and finish
+    /// sync_safekeepers. API is mocked in MockImpl.
+    ///
+    /// Run this test with valgrind to detect leaks:
+    /// `valgrind --leak-check=full target/debug/deps/walproposer-<build>`
+    #[test]
+    fn test_simple_sync_safekeepers() -> anyhow::Result<()> {
+        let ttid = TenantTimelineId::new(
+            "9e4c8f36063c6c6e93bc20d65a820f3d".parse()?,
+            "9e4c8f36063c6c6e93bc20d65a820f3d".parse()?,
+        );
+
+        let (sender, receiver) = sync_channel(1);
+
+        let my_impl: Box<dyn ApiImpl> = Box::new(MockImpl {
+            wait_events: Cell::new(WaitEventsData {
+                sk: std::ptr::null_mut(),
+                event_mask: 0,
+            }),
+            expected_messages: vec![
+                // Greeting(ProposerGreeting { protocol_version: 2, pg_version: 160000, proposer_id: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], system_id: 0, timeline_id: 9e4c8f36063c6c6e93bc20d65a820f3d, tenant_id: 9e4c8f36063c6c6e93bc20d65a820f3d, tli: 1, wal_seg_size: 16777216 })
+                vec![
+                    103, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 113, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 158, 76, 143, 54, 6, 60, 108, 110,
+                    147, 188, 32, 214, 90, 130, 15, 61, 158, 76, 143, 54, 6, 60, 108, 110, 147,
+                    188, 32, 214, 90, 130, 15, 61, 1, 0, 0, 0, 0, 0, 0, 1,
+                ],
+                // VoteRequest(VoteRequest { term: 3 })
+                vec![
+                    118, 0, 0, 0, 0, 0, 0, 0, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                    0, 0, 0, 0, 0, 0,
+                ],
+            ],
+            expected_ptr: AtomicUsize::new(0),
+            safekeeper_replies: vec![
+                // Greeting(AcceptorGreeting { term: 2, node_id: NodeId(1) })
+                vec![
+                    103, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0,
+                ],
+                // VoteResponse(VoteResponse { term: 3, vote_given: 1, flush_lsn: 0/539, truncate_lsn: 0/539, term_history: [(2, 0/539)], timeline_start_lsn: 0/539 })
+                vec![
+                    118, 0, 0, 0, 0, 0, 0, 0, 3, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 57,
+                    5, 0, 0, 0, 0, 0, 0, 57, 5, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0,
+                    0, 57, 5, 0, 0, 0, 0, 0, 0, 57, 5, 0, 0, 0, 0, 0, 0,
+                ],
+            ],
+            replies_ptr: AtomicUsize::new(0),
+            sync_channel: sender,
+        });
+        let config = crate::walproposer::Config {
+            ttid,
+            safekeepers_list: vec!["localhost:5000".to_string()],
+            safekeeper_reconnect_timeout: 1000,
+            safekeeper_connection_timeout: 10000,
+            sync_safekeepers: true,
+        };
+
+        let wp = Wrapper::new(my_impl, config);
+
+        // walproposer will panic when it finishes sync_safekeepers
+        std::panic::catch_unwind(|| wp.start()).unwrap_err();
+        // validate the resulting LSN
+        assert_eq!(receiver.recv()?, 1337);
+        Ok(())
+        // drop() will free up resources here
+    }
+}

pageserver/Cargo.toml

@@ -82,6 +82,8 @@ enum-map.workspace = true
 enumset.workspace = true
 strum.workspace = true
 strum_macros.workspace = true
+tokio-stream.workspace = true
+tracing-subscriber = { version = "0.3.17", features = ["env-filter"] }
 
 [dev-dependencies]
 criterion.workspace = true

pageserver/benches/bench_walredo.rs

@@ -11,10 +11,7 @@ use std::sync::{Arc, Barrier};
 
 use bytes::{Buf, Bytes};
 use pageserver::{
-    config::PageServerConf,
-    repository::Key,
-    walrecord::NeonWalRecord,
-    walredo::{PostgresRedoManager, WalRedoError},
+    config::PageServerConf, repository::Key, walrecord::NeonWalRecord, walredo::PostgresRedoManager,
 };
 use utils::{id::TenantId, lsn::Lsn};
 
@@ -35,9 +32,15 @@ fn redo_scenarios(c: &mut Criterion) {
 
     let manager = Arc::new(manager);
 
-    tracing::info!("executing first");
-    short().execute(&manager).unwrap();
-    tracing::info!("first executed");
+    {
+        let rt = tokio::runtime::Builder::new_current_thread()
+            .enable_all()
+            .build()
+            .unwrap();
+        tracing::info!("executing first");
+        short().execute(rt.handle(), &manager).unwrap();
+        tracing::info!("first executed");
+    }
 
     let thread_counts = [1, 2, 4, 8, 16];
 
@@ -80,9 +83,14 @@ fn add_multithreaded_walredo_requesters(
     assert_ne!(threads, 0);
 
     if threads == 1 {
+        let rt = tokio::runtime::Builder::new_current_thread()
+            .enable_all()
+            .build()
+            .unwrap();
+        let handle = rt.handle();
         b.iter_batched_ref(
             || Some(input_factory()),
-            |input| execute_all(input.take(), manager),
+            |input| execute_all(input.take(), handle, manager),
             criterion::BatchSize::PerIteration,
         );
     } else {
@@ -98,19 +106,26 @@ fn add_multithreaded_walredo_requesters(
                     let manager = manager.clone();
                     let barrier = barrier.clone();
                     let work_rx = work_rx.clone();
-                    move || loop {
-                        // queue up and wait if we want to go another round
-                        if work_rx.lock().unwrap().recv().is_err() {
-                            break;
+                    move || {
+                        let rt = tokio::runtime::Builder::new_current_thread()
+                            .enable_all()
+                            .build()
+                            .unwrap();
+                        let handle = rt.handle();
+                        loop {
+                            // queue up and wait if we want to go another round
+                            if work_rx.lock().unwrap().recv().is_err() {
+                                break;
+                            }
+
+                            let input = Some(input_factory());
+
+                            barrier.wait();
+
+                            execute_all(input, handle, &manager).unwrap();
+
+                            barrier.wait();
                         }
-
-                        let input = Some(input_factory());
-
-                        barrier.wait();
-
-                        execute_all(input, &manager).unwrap();
-
-                        barrier.wait();
                     }
                 })
             })
@@ -152,15 +167,19 @@ impl Drop for JoinOnDrop {
     }
 }
 
-fn execute_all<I>(input: I, manager: &PostgresRedoManager) -> Result<(), WalRedoError>
+fn execute_all<I>(
+    input: I,
+    handle: &tokio::runtime::Handle,
+    manager: &PostgresRedoManager,
+) -> anyhow::Result<()>
 where
     I: IntoIterator<Item = Request>,
 {
     // just fire all requests as fast as possible
     input.into_iter().try_for_each(|req| {
-        let page = req.execute(manager)?;
+        let page = req.execute(handle, manager)?;
         assert_eq!(page.remaining(), 8192);
-        Ok::<_, WalRedoError>(())
+        anyhow::Ok(())
     })
 }
 
@@ -473,9 +492,11 @@ struct Request {
 }
 
 impl Request {
-    fn execute(self, manager: &PostgresRedoManager) -> Result<Bytes, WalRedoError> {
-        use pageserver::walredo::WalRedoManager;
-
+    fn execute(
+        self,
+        rt: &tokio::runtime::Handle,
+        manager: &PostgresRedoManager,
+    ) -> anyhow::Result<Bytes> {
         let Request {
             key,
             lsn,
@@ -484,6 +505,6 @@ impl Request {
             pg_version,
         } = self;
 
-        manager.request_redo(key, lsn, base_img, records, pg_version)
+        rt.block_on(manager.request_redo(key, lsn, base_img, records, pg_version))
     }
 }

pageserver/src/basebackup.rs

@@ -13,6 +13,7 @@
 use anyhow::{anyhow, bail, ensure, Context};
 use bytes::{BufMut, BytesMut};
 use fail::fail_point;
+use postgres_ffi::pg_constants;
 use std::fmt::Write as FmtWrite;
 use std::time::SystemTime;
 use tokio::io;
@@ -180,6 +181,7 @@ where
             }
         }
 
+        let mut min_restart_lsn: Lsn = Lsn::MAX;
         // Create tablespace directories
         for ((spcnode, dbnode), has_relmap_file) in
             self.timeline.list_dbdirs(self.lsn, self.ctx).await?
@@ -213,6 +215,34 @@ where
                     self.add_rel(rel, rel).await?;
                 }
             }
+
+            for (path, content) in self.timeline.list_aux_files(self.lsn, self.ctx).await? {
+                if path.starts_with("pg_replslot") {
+                    let offs = pg_constants::REPL_SLOT_ON_DISK_OFFSETOF_RESTART_LSN;
+                    let restart_lsn = Lsn(u64::from_le_bytes(
+                        content[offs..offs + 8].try_into().unwrap(),
+                    ));
+                    info!("Replication slot {} restart LSN={}", path, restart_lsn);
+                    min_restart_lsn = Lsn::min(min_restart_lsn, restart_lsn);
+                }
+                let header = new_tar_header(&path, content.len() as u64)?;
+                self.ar
+                    .append(&header, &*content)
+                    .await
+                    .context("could not add aux file to basebackup tarball")?;
+            }
+        }
+        if min_restart_lsn != Lsn::MAX {
+            info!(
+                "Min restart LSN for logical replication is {}",
+                min_restart_lsn
+            );
+            let data = min_restart_lsn.0.to_le_bytes();
+            let header = new_tar_header("restart.lsn", data.len() as u64)?;
+            self.ar
+                .append(&header, &data[..])
+                .await
+                .context("could not add restart.lsn file to basebackup tarball")?;
         }
         for xid in self
             .timeline

pageserver/src/bin/getpage_bench_http.rs

@@ -0,0 +1,245 @@
+use clap::Parser;
+use hyper::client::conn::Parts;
+use hyper::client::HttpConnector;
+use hyper::{Body, Client, Uri};
+use pageserver::{repository, tenant};
+use rand::prelude::*;
+use std::env::args;
+use std::future::Future;
+use std::str::FromStr;
+use std::sync::atomic::{AtomicU64, Ordering};
+use std::sync::{Arc, Mutex};
+use std::thread;
+use tokio::sync::mpsc::{channel, Sender};
+use tokio::sync::Mutex as AsyncMutex;
+use tokio::task::JoinHandle;
+
+struct Key(repository::Key);
+
+impl std::str::FromStr for Key {
+    type Err = anyhow::Error;
+
+    fn from_str(s: &str) -> std::result::Result<Self, Self::Err> {
+        repository::Key::from_hex(s).map(Key)
+    }
+}
+
+struct KeyRange {
+    start: Key,
+    end: Key,
+}
+
+impl KeyRange {
+    fn len(&self) -> i128 {
+        self.end.0.to_i128() - self.start.0.to_i128()
+    }
+}
+
+#[derive(clap::Parser)]
+struct Args {
+    #[clap(long, default_value = "http://localhost:9898")]
+    ps_endpoint: String,
+    // tenant_id: String,
+    // timeline_id: String,
+    num_tasks: usize,
+    num_requests: usize,
+    tenants: Option<Vec<String>>,
+    #[clap(long)]
+    pick_n_tenants: Option<usize>,
+}
+
+#[derive(Debug, Default)]
+struct Stats {
+    completed_requests: AtomicU64,
+}
+
+impl Stats {
+    fn inc(&self) {
+        self.completed_requests.fetch_add(1, Ordering::Relaxed);
+    }
+}
+
+#[tokio::main]
+async fn main() {
+    let args: &'static Args = Box::leak(Box::new(Args::parse()));
+
+    let client = Client::new();
+
+    let tenants = if let Some(tenants) = &args.tenants {
+        tenants.clone()
+    } else {
+        // let tenant_id = "b97965931096047b2d54958756baee7b";
+        // let timeline_id = "2868f84a8d166779e4c651b116c45059";
+
+        let resp = client
+            .get(Uri::try_from(&format!("{}/v1/tenant", args.ps_endpoint)).unwrap())
+            .await
+            .unwrap();
+
+        let body = hyper::body::to_bytes(resp).await.unwrap();
+        let tenants: serde_json::Value = serde_json::from_slice(&body).unwrap();
+        let mut out = Vec::new();
+        for t in tenants.as_array().unwrap() {
+            if let Some(limit) = args.pick_n_tenants {
+                if out.len() >= limit {
+                    break;
+                }
+            }
+            out.push(t.get("id").unwrap().as_str().unwrap().to_owned());
+        }
+        if let Some(limit) = args.pick_n_tenants {
+            assert_eq!(out.len(), limit);
+        }
+        out
+    };
+
+    let mut tenant_timelines = Vec::new();
+    for tenant_id in tenants {
+        let resp = client
+            .get(
+                Uri::try_from(&format!(
+                    "{}/v1/tenant/{}/timeline",
+                    args.ps_endpoint, tenant_id
+                ))
+                .unwrap(),
+            )
+            .await
+            .unwrap();
+
+        let body = hyper::body::to_bytes(resp).await.unwrap();
+        let timelines: serde_json::Value = serde_json::from_slice(&body).unwrap();
+        for t in timelines.as_array().unwrap() {
+            let timeline_id = t.get("timeline_id").unwrap().as_str().unwrap().to_owned();
+            tenant_timelines.push((tenant_id.clone(), timeline_id));
+        }
+    }
+    println!("tenant_timelines:\n{:?}", tenant_timelines);
+
+    let mut stats = Arc::new(Stats::default());
+
+    tokio::spawn({
+        let stats = Arc::clone(&stats);
+        async move {
+            loop {
+                let start = std::time::Instant::now();
+                tokio::time::sleep(std::time::Duration::from_secs(1)).await;
+                let completed_requests = stats.completed_requests.swap(0, Ordering::Relaxed);
+                let elapsed = start.elapsed();
+                println!(
+                    "RPS: {:.0}",
+                    completed_requests as f64 / elapsed.as_secs_f64()
+                );
+            }
+        }
+    });
+
+    let mut tasks = Vec::new();
+    for (tenant_id, timeline_id) in tenant_timelines {
+        let t = tokio::spawn(timeline(
+            args,
+            client.clone(),
+            tenant_id,
+            timeline_id,
+            Arc::clone(&stats),
+        ));
+        tasks.push(t);
+    }
+
+    for t in tasks {
+        t.await.unwrap();
+    }
+}
+
+fn timeline(
+    args: &'static Args,
+    client: Client<HttpConnector, Body>,
+    tenant_id: String,
+    timeline_id: String,
+    stats: Arc<Stats>,
+) -> impl Future<Output = ()> {
+    async move {
+        let mut resp = client
+            .get(
+                Uri::try_from(&format!(
+                    "{}/v1/tenant/{}/timeline/{}/keyspace",
+                    args.ps_endpoint, tenant_id, timeline_id
+                ))
+                .unwrap(),
+            )
+            .await
+            .unwrap();
+        if !resp.status().is_success() {
+            panic!("Failed to get keyspace: {resp:?}");
+        }
+        let body = hyper::body::to_bytes(resp).await.unwrap();
+        let keyspace: serde_json::Value = serde_json::from_slice(&body).unwrap();
+
+        let lsn = Arc::new(keyspace["at_lsn"].as_str().unwrap().to_owned());
+
+        let ranges = keyspace["keys"]
+            .as_array()
+            .unwrap()
+            .iter()
+            .map(|r| {
+                let r = r.as_array().unwrap();
+                assert_eq!(r.len(), 2);
+                let start = Key::from_str(r[0].as_str().unwrap()).unwrap();
+                let end = Key::from_str(r[1].as_str().unwrap()).unwrap();
+                KeyRange { start, end }
+            })
+            .collect::<Vec<_>>();
+
+        // weighted ranges
+        let weights = ranges.iter().map(|r| r.len()).collect::<Vec<_>>();
+
+        let ranges = Arc::new(ranges);
+        let weights = Arc::new(weights);
+
+        let (tx, mut rx) = channel::<i32>(1000);
+        let tx = Arc::new(AsyncMutex::new(tx));
+
+        let mut tasks = Vec::<JoinHandle<()>>::new();
+
+        let start = std::time::Instant::now();
+
+        for i in 0..args.num_tasks {
+            let ranges = ranges.clone();
+            let weights = weights.clone();
+            let lsn = lsn.clone();
+            let client = client.clone();
+            let tenant_id = tenant_id.clone();
+            let timeline_id = timeline_id.clone();
+            let stats = Arc::clone(&stats);
+            let task = tokio::spawn(async move {
+                for i in 0..args.num_requests {
+                    let key = {
+                        let mut rng = rand::thread_rng();
+                        let r = ranges.choose_weighted(&mut rng, |r| r.len()).unwrap();
+                        let key = rng.gen_range((r.start.0.to_i128()..r.end.0.to_i128()));
+                        key
+                    };
+                    let url = format!(
+                        "{}/v1/tenant/{}/timeline/{}/getpage?key={:036x}&lsn={}",
+                        args.ps_endpoint, tenant_id, timeline_id, key, lsn
+                    );
+                    let uri = url.parse::<Uri>().unwrap();
+                    let resp = client.get(uri).await.unwrap();
+                    stats.inc();
+                }
+            });
+            tasks.push(task);
+        }
+
+        drop(tx);
+
+        for task in tasks {
+            task.await.unwrap();
+        }
+
+        let elapsed = start.elapsed();
+        println!(
+            "RPS: {:.0}",
+            (args.num_requests * args.num_tasks) as f64 / elapsed.as_secs_f64()
+        );
+    }
+}

pageserver/src/bin/getpage_bench_libpq.rs

@@ -0,0 +1,373 @@
+use anyhow::Context;
+use clap::Parser;
+use futures::{SinkExt, TryStreamExt};
+use hyper::client::conn::Parts;
+use hyper::client::HttpConnector;
+use hyper::{Client, Uri};
+use pageserver::page_cache::PAGE_SZ;
+use pageserver::pgdatadir_mapping::{is_rel_block_key, key_to_rel_block};
+use pageserver::{repository, tenant};
+use pageserver_api::models::{
+    PagestreamBeMessage, PagestreamFeMessage, PagestreamGetPageRequest, PagestreamGetPageResponse,
+};
+use pageserver_api::reltag::RelTag;
+use rand::prelude::*;
+use scopeguard::defer;
+use std::env::args;
+use std::future::Future;
+use std::str::FromStr;
+use std::sync::atomic::{AtomicU64, Ordering};
+use std::sync::{Arc, Mutex};
+use std::thread;
+use tokio::sync::mpsc::{channel, Sender};
+use tokio::sync::Mutex as AsyncMutex;
+use tokio::task::JoinHandle;
+use tokio_stream::{Stream, StreamExt};
+use utils::completion;
+use utils::lsn::Lsn;
+
+struct Key(repository::Key);
+
+impl std::str::FromStr for Key {
+    type Err = anyhow::Error;
+
+    fn from_str(s: &str) -> std::result::Result<Self, Self::Err> {
+        repository::Key::from_hex(s).map(Key)
+    }
+}
+
+struct KeyRange {
+    start: i128,
+    end: i128,
+}
+
+impl KeyRange {
+    fn len(&self) -> i128 {
+        self.end - self.start
+    }
+}
+
+struct RelTagBlockNo {
+    rel_tag: RelTag,
+    block_no: u32,
+}
+
+#[derive(clap::Parser)]
+struct Args {
+    #[clap(long, default_value = "http://localhost:9898")]
+    ps_endpoint: String,
+    // tenant_id: String,
+    // timeline_id: String,
+    num_tasks: usize,
+    num_requests: usize,
+    tenants: Option<Vec<String>>,
+    #[clap(long)]
+    pick_n_tenants: Option<usize>,
+}
+
+#[derive(Debug, Default)]
+struct Stats {
+    completed_requests: AtomicU64,
+}
+
+impl Stats {
+    fn inc(&self) {
+        self.completed_requests.fetch_add(1, Ordering::Relaxed);
+    }
+}
+
+#[tokio::main]
+async fn main() {
+    let args: &'static Args = Box::leak(Box::new(Args::parse()));
+
+    // std::env::set_var("RUST_LOG", "info,tokio_postgres=trace");
+    // tracing_subscriber::fmt::init();
+
+    let client = Client::new();
+
+    let tenants = if let Some(tenants) = &args.tenants {
+        tenants.clone()
+    } else {
+        // let tenant_id = "b97965931096047b2d54958756baee7b";
+        // let timeline_id = "2868f84a8d166779e4c651b116c45059";
+
+        let resp = client
+            .get(Uri::try_from(&format!("{}/v1/tenant", args.ps_endpoint)).unwrap())
+            .await
+            .unwrap();
+
+        let body = hyper::body::to_bytes(resp).await.unwrap();
+        let tenants: serde_json::Value = serde_json::from_slice(&body).unwrap();
+        let mut out = Vec::new();
+        for t in tenants.as_array().unwrap() {
+            if let Some(limit) = args.pick_n_tenants {
+                if out.len() >= limit {
+                    break;
+                }
+            }
+            out.push(t.get("id").unwrap().as_str().unwrap().to_owned());
+        }
+        if let Some(limit) = args.pick_n_tenants {
+            assert_eq!(out.len(), limit);
+        }
+        out
+    };
+
+    let mut tenant_timelines = Vec::new();
+    for tenant_id in tenants {
+        let resp = client
+            .get(
+                Uri::try_from(&format!(
+                    "{}/v1/tenant/{}/timeline",
+                    args.ps_endpoint, tenant_id
+                ))
+                .unwrap(),
+            )
+            .await
+            .unwrap();
+
+        let body = hyper::body::to_bytes(resp).await.unwrap();
+        let timelines: serde_json::Value = serde_json::from_slice(&body).unwrap();
+        for t in timelines.as_array().unwrap() {
+            let timeline_id = t.get("timeline_id").unwrap().as_str().unwrap().to_owned();
+            tenant_timelines.push((tenant_id.clone(), timeline_id));
+        }
+    }
+    println!("tenant_timelines:\n{:?}", tenant_timelines);
+
+    let mut stats = Arc::new(Stats::default());
+
+    tokio::spawn({
+        let stats = Arc::clone(&stats);
+        async move {
+            loop {
+                let start = std::time::Instant::now();
+                tokio::time::sleep(std::time::Duration::from_secs(1)).await;
+                let completed_requests = stats.completed_requests.swap(0, Ordering::Relaxed);
+                let elapsed = start.elapsed();
+                println!(
+                    "RPS: {:.0}",
+                    completed_requests as f64 / elapsed.as_secs_f64()
+                );
+            }
+        }
+    });
+
+    let mut tasks = Vec::new();
+    for (tenant_id, timeline_id) in tenant_timelines {
+        let stats = Arc::clone(&stats);
+        let t = tokio::spawn(timeline(
+            args,
+            client.clone(),
+            tenant_id,
+            timeline_id,
+            stats,
+        ));
+        tasks.push(t);
+    }
+
+    for t in tasks {
+        t.await.unwrap();
+    }
+}
+
+fn timeline(
+    args: &'static Args,
+    http_client: Client<HttpConnector, hyper::Body>,
+    tenant_id: String,
+    timeline_id: String,
+    stats: Arc<Stats>,
+) -> impl Future<Output = ()> + Send + Sync {
+    async move {
+        let mut resp = http_client
+            .get(
+                Uri::try_from(&format!(
+                    "{}/v1/tenant/{}/timeline/{}/keyspace",
+                    args.ps_endpoint, tenant_id, timeline_id
+                ))
+                .unwrap(),
+            )
+            .await
+            .unwrap();
+        if !resp.status().is_success() {
+            panic!("Failed to get keyspace: {resp:?}");
+        }
+        let body = hyper::body::to_bytes(resp).await.unwrap();
+        let keyspace: serde_json::Value = serde_json::from_slice(&body).unwrap();
+        let lsn: Lsn = keyspace["at_lsn"].as_str().unwrap().parse().unwrap();
+
+        let ranges = keyspace["keys"]
+            .as_array()
+            .unwrap()
+            .iter()
+            .filter_map(|r| {
+                let r = r.as_array().unwrap();
+                assert_eq!(r.len(), 2);
+                let start = Key::from_str(r[0].as_str().unwrap()).unwrap();
+                let end = Key::from_str(r[1].as_str().unwrap()).unwrap();
+                // filter out non-relblock keys
+                match (is_rel_block_key(start.0), is_rel_block_key(end.0)) {
+                    (true, true) => Some(KeyRange {
+                        start: start.0.to_i128(),
+                        end: end.0.to_i128(),
+                    }),
+                    (true, false) | (false, true) => {
+                        unimplemented!("split up range")
+                    }
+                    (false, false) => None,
+                }
+            })
+            .collect::<Vec<_>>();
+
+        // weighted ranges
+        let weights = ranges.iter().map(|r| r.len()).collect::<Vec<_>>();
+
+        let ranges = Arc::new(ranges);
+        let weights = Arc::new(weights);
+
+        let mut tasks = Vec::<JoinHandle<()>>::new();
+
+        let start = std::time::Instant::now();
+
+        for i in 0..args.num_tasks {
+            let ranges = ranges.clone();
+            let weights = weights.clone();
+            let client = http_client.clone();
+            let tenant_id = tenant_id.clone();
+            let timeline_id = timeline_id.clone();
+            let task = tokio::spawn({
+                let stats = Arc::clone(&stats);
+                async move {
+                    let mut client =
+                        getpage_client::Client::new(tenant_id.clone(), timeline_id.clone())
+                            .await
+                            .unwrap();
+                    for i in 0..args.num_requests {
+                        let key = {
+                            let mut rng = rand::thread_rng();
+                            let r = ranges.choose_weighted(&mut rng, |r| r.len()).unwrap();
+                            let key: i128 = rng.gen_range((r.start..r.end));
+                            let key = repository::Key::from_i128(key);
+                            // XXX filter these out when we iterate the keyspace
+                            assert!(
+                                is_rel_block_key(key),
+                                "we filter non-relblock keys out above"
+                            );
+                            let (rel_tag, block_no) =
+                                key_to_rel_block(key).expect("we just checked");
+                            RelTagBlockNo { rel_tag, block_no }
+                        };
+                        client
+                            .getpage(key, lsn)
+                            .await
+                            .with_context(|| {
+                                format!("getpage for tenant {} timeline {}", tenant_id, timeline_id)
+                            })
+                            .unwrap();
+                        stats.inc();
+                    }
+                    client.shutdown().await;
+                }
+            });
+            tasks.push(task);
+        }
+
+        for task in tasks {
+            task.await.unwrap();
+        }
+    }
+}
+
+mod getpage_client {
+    use std::pin::Pin;
+
+    use futures::SinkExt;
+    use pageserver_api::models::{
+        PagestreamBeMessage, PagestreamFeMessage, PagestreamGetPageRequest,
+        PagestreamGetPageResponse,
+    };
+    use tokio::task::JoinHandle;
+    use tokio_stream::StreamExt;
+    use tokio_util::sync::CancellationToken;
+    use utils::lsn::Lsn;
+
+    use crate::RelTagBlockNo;
+
+    pub(crate) struct Client {
+        copy_both: Pin<Box<tokio_postgres::CopyBothDuplex<bytes::Bytes>>>,
+        cancel_on_client_drop: Option<tokio_util::sync::DropGuard>,
+        conn_task: JoinHandle<()>,
+    }
+
+    impl Client {
+        pub fn new(
+            tenant_id: String,
+            timeline_id: String,
+        ) -> impl std::future::Future<Output = anyhow::Result<Self>> + Send {
+            async move {
+                let (client, connection) =
+                    tokio_postgres::connect("host=localhost port=64000", postgres::NoTls).await?;
+
+                let conn_task_cancel = CancellationToken::new();
+                let conn_task = tokio::spawn({
+                    let conn_task_cancel = conn_task_cancel.clone();
+                    async move {
+                        tokio::select! {
+                            _ = conn_task_cancel.cancelled() => {
+                                return;
+                            }
+                            res = connection => {
+                                res.unwrap();
+                            }
+                        }
+                    }
+                });
+
+                let copy_both: tokio_postgres::CopyBothDuplex<bytes::Bytes> = client
+                    .copy_both_simple(&format!("pagestream {tenant_id} {timeline_id}"))
+                    .await?;
+
+                Ok(Self {
+                    copy_both: Box::pin(copy_both),
+                    conn_task,
+                    cancel_on_client_drop: Some(conn_task_cancel.drop_guard()),
+                })
+            }
+        }
+
+        pub async fn shutdown(mut self) {
+            let _ = self.cancel_on_client_drop.take();
+            self.conn_task.await.unwrap();
+        }
+
+        pub async fn getpage(
+            &mut self,
+            key: RelTagBlockNo,
+            lsn: Lsn,
+        ) -> anyhow::Result<PagestreamGetPageResponse> {
+            let req = PagestreamGetPageRequest {
+                latest: false,
+                rel: key.rel_tag,
+                blkno: key.block_no,
+                lsn,
+            };
+            let req = PagestreamFeMessage::GetPage(req);
+            let req: bytes::Bytes = req.serialize();
+            // let mut req = tokio_util::io::ReaderStream::new(&req);
+            let mut req = tokio_stream::once(Ok(req));
+
+            self.copy_both.send_all(&mut req).await?;
+
+            let next: Option<Result<bytes::Bytes, _>> = self.copy_both.next().await;
+            let next = next.unwrap().unwrap();
+
+            match PagestreamBeMessage::deserialize(next)? {
+                PagestreamBeMessage::Exists(_) => todo!(),
+                PagestreamBeMessage::Nblocks(_) => todo!(),
+                PagestreamBeMessage::GetPage(p) => Ok(p),
+                PagestreamBeMessage::Error(e) => anyhow::bail!("Error: {:?}", e),
+                PagestreamBeMessage::DbSize(_) => todo!(),
+            }
+        }
+    }
+}

pageserver/src/bin/pageserver.rs

@@ -2,6 +2,7 @@
 
 use std::env::{var, VarError};
 use std::sync::Arc;
+use std::time::Duration;
 use std::{env, ops::ControlFlow, str::FromStr};
 
 use anyhow::{anyhow, Context};
@@ -200,6 +201,51 @@ fn initialize_config(
     })
 }
 
+struct WaitForPhaseResult<F: std::future::Future + Unpin> {
+    timeout_remaining: Duration,
+    skipped: Option<F>,
+}
+
+/// During startup, we apply a timeout to our waits for readiness, to avoid
+/// stalling the whole service if one Tenant experiences some problem.  Each
+/// phase may consume some of the timeout: this function returns the updated
+/// timeout for use in the next call.
+async fn wait_for_phase<F>(phase: &str, mut fut: F, timeout: Duration) -> WaitForPhaseResult<F>
+where
+    F: std::future::Future + Unpin,
+{
+    let initial_t = Instant::now();
+    let skipped = match tokio::time::timeout(timeout, &mut fut).await {
+        Ok(_) => None,
+        Err(_) => {
+            tracing::info!(
+                timeout_millis = timeout.as_millis(),
+                %phase,
+                "Startup phase timed out, proceeding anyway"
+            );
+            Some(fut)
+        }
+    };
+
+    WaitForPhaseResult {
+        timeout_remaining: timeout
+            .checked_sub(Instant::now().duration_since(initial_t))
+            .unwrap_or(Duration::ZERO),
+        skipped,
+    }
+}
+
+fn startup_checkpoint(started_at: Instant, phase: &str, human_phase: &str) {
+    let elapsed = started_at.elapsed();
+    let secs = elapsed.as_secs_f64();
+    STARTUP_DURATION.with_label_values(&[phase]).set(secs);
+
+    info!(
+        elapsed_ms = elapsed.as_millis(),
+        "{human_phase} ({secs:.3}s since start)"
+    )
+}
+
 fn start_pageserver(
     launch_ts: &'static LaunchTimestamp,
     conf: &'static PageServerConf,
@@ -207,16 +253,6 @@ fn start_pageserver(
     // Monotonic time for later calculating startup duration
     let started_startup_at = Instant::now();
 
-    let startup_checkpoint = move |phase: &str, human_phase: &str| {
-        let elapsed = started_startup_at.elapsed();
-        let secs = elapsed.as_secs_f64();
-        STARTUP_DURATION.with_label_values(&[phase]).set(secs);
-        info!(
-            elapsed_ms = elapsed.as_millis(),
-            "{human_phase} ({secs:.3}s since start)"
-        )
-    };
-
     // Print version and launch timestamp to the log,
     // and expose them as prometheus metrics.
     // A changed version string indicates changed software.
@@ -341,7 +377,7 @@ fn start_pageserver(
 
     // Up to this point no significant I/O has been done: this should have been fast.  Record
     // duration prior to starting I/O intensive phase of startup.
-    startup_checkpoint("initial", "Starting loading tenants");
+    startup_checkpoint(started_startup_at, "initial", "Starting loading tenants");
     STARTUP_IS_LOADING.set(1);
 
     // Startup staging or optimizing:
@@ -355,6 +391,7 @@ fn start_pageserver(
     // consumer side) will be dropped once we can start the background jobs. Currently it is behind
     // completing all initial logical size calculations (init_logical_size_done_rx) and a timeout
     // (background_task_maximum_delay).
+    let (init_remote_done_tx, init_remote_done_rx) = utils::completion::channel();
     let (init_done_tx, init_done_rx) = utils::completion::channel();
 
     let (init_logical_size_done_tx, init_logical_size_done_rx) = utils::completion::channel();
@@ -362,7 +399,8 @@ fn start_pageserver(
     let (background_jobs_can_start, background_jobs_barrier) = utils::completion::channel();
 
     let order = pageserver::InitializationOrder {
-        initial_tenant_load: Some(init_done_tx),
+        initial_tenant_load_remote: Some(init_done_tx),
+        initial_tenant_load: Some(init_remote_done_tx),
         initial_logical_size_can_start: init_done_rx.clone(),
         initial_logical_size_attempt: Some(init_logical_size_done_tx),
         background_jobs_can_start: background_jobs_barrier.clone(),
@@ -386,55 +424,93 @@ fn start_pageserver(
         let shutdown_pageserver = shutdown_pageserver.clone();
         let drive_init = async move {
             // NOTE: unlike many futures in pageserver, this one is cancellation-safe
-            let guard = scopeguard::guard_on_success((), |_| tracing::info!("Cancelled before initial load completed"));
+            let guard = scopeguard::guard_on_success((), |_| {
+                tracing::info!("Cancelled before initial load completed")
+            });
 
-            init_done_rx.wait().await;
-            startup_checkpoint("initial_tenant_load", "Initial load completed");
-            STARTUP_IS_LOADING.set(0);
+            let timeout = conf.background_task_maximum_delay;
+
+            let init_remote_done = std::pin::pin!(async {
+                init_remote_done_rx.wait().await;
+                startup_checkpoint(
+                    started_startup_at,
+                    "initial_tenant_load_remote",
+                    "Remote part of initial load completed",
+                );
+            });
+
+            let WaitForPhaseResult {
+                timeout_remaining: timeout,
+                skipped: init_remote_skipped,
+            } = wait_for_phase("initial_tenant_load_remote", init_remote_done, timeout).await;
+
+            let init_load_done = std::pin::pin!(async {
+                init_done_rx.wait().await;
+                startup_checkpoint(
+                    started_startup_at,
+                    "initial_tenant_load",
+                    "Initial load completed",
+                );
+                STARTUP_IS_LOADING.set(0);
+            });
+
+            let WaitForPhaseResult {
+                timeout_remaining: timeout,
+                skipped: init_load_skipped,
+            } = wait_for_phase("initial_tenant_load", init_load_done, timeout).await;
 
             // initial logical sizes can now start, as they were waiting on init_done_rx.
 
             scopeguard::ScopeGuard::into_inner(guard);
 
-            let mut init_sizes_done = std::pin::pin!(init_logical_size_done_rx.wait());
+            let guard = scopeguard::guard_on_success((), |_| {
+                tracing::info!("Cancelled before initial logical sizes completed")
+            });
 
-            let timeout = conf.background_task_maximum_delay;
+            let logical_sizes_done = std::pin::pin!(async {
+                init_logical_size_done_rx.wait().await;
+                startup_checkpoint(
+                    started_startup_at,
+                    "initial_logical_sizes",
+                    "Initial logical sizes completed",
+                );
+            });
 
-            let guard = scopeguard::guard_on_success((), |_| tracing::info!("Cancelled before initial logical sizes completed"));
-
-            let init_sizes_done = match tokio::time::timeout(timeout, &mut init_sizes_done).await {
-                Ok(_) => {
-                    startup_checkpoint("initial_logical_sizes", "Initial logical sizes completed");
-                    None
-                }
-                Err(_) => {
-                    tracing::info!(
-                        timeout_millis = timeout.as_millis(),
-                        "Initial logical size timeout elapsed; starting background jobs"
-                    );
-                    Some(init_sizes_done)
-                }
-            };
+            let WaitForPhaseResult {
+                timeout_remaining: _,
+                skipped: logical_sizes_skipped,
+            } = wait_for_phase("initial_logical_sizes", logical_sizes_done, timeout).await;
 
             scopeguard::ScopeGuard::into_inner(guard);
 
-            // allow background jobs to start
+            // allow background jobs to start: we either completed prior stages, or they reached timeout
+            // and were skipped.  It is important that we do not let them block background jobs indefinitely,
+            // because things like consumption metrics for billing are blocked by this barrier.
             drop(background_jobs_can_start);
-            startup_checkpoint("background_jobs_can_start", "Starting background jobs");
-
-            if let Some(init_sizes_done) = init_sizes_done {
-                // ending up here is not a bug; at the latest logical sizes will be queried by
-                // consumption metrics.
-                let guard = scopeguard::guard_on_success((), |_| tracing::info!("Cancelled before initial logical sizes completed"));
-                init_sizes_done.await;
-
-                scopeguard::ScopeGuard::into_inner(guard);
-
-                startup_checkpoint("initial_logical_sizes", "Initial logical sizes completed after timeout (background jobs already started)");
+            startup_checkpoint(
+                started_startup_at,
+                "background_jobs_can_start",
+                "Starting background jobs",
+            );
 
+            // We are done. If we skipped any phases due to timeout, run them to completion here so that
+            // they will eventually update their startup_checkpoint, and so that we do not declare the
+            // 'complete' stage until all the other stages are really done.
+            let guard = scopeguard::guard_on_success((), |_| {
+                tracing::info!("Cancelled before waiting for skipped phases done")
+            });
+            if let Some(f) = init_remote_skipped {
+                f.await;
             }
+            if let Some(f) = init_load_skipped {
+                f.await;
+            }
+            if let Some(f) = logical_sizes_skipped {
+                f.await;
+            }
+            scopeguard::ScopeGuard::into_inner(guard);
 
-            startup_checkpoint("complete", "Startup complete");
+            startup_checkpoint(started_startup_at, "complete", "Startup complete");
         };
 
         async move {
@@ -574,6 +650,7 @@ fn start_pageserver(
                     pageserver_listener,
                     conf.pg_auth_type,
                     libpq_ctx,
+                    task_mgr::shutdown_token(),
                 )
                 .await
             },

pageserver/src/config.rs

@@ -37,8 +37,8 @@ use crate::tenant::{
     TIMELINES_SEGMENT_NAME,
 };
 use crate::{
-    IGNORED_TENANT_FILE_NAME, METADATA_FILE_NAME, TENANT_CONFIG_NAME, TIMELINE_DELETE_MARK_SUFFIX,
-    TIMELINE_UNINIT_MARK_SUFFIX,
+    IGNORED_TENANT_FILE_NAME, METADATA_FILE_NAME, TENANT_CONFIG_NAME, TENANT_LOCATION_CONFIG_NAME,
+    TIMELINE_DELETE_MARK_SUFFIX, TIMELINE_UNINIT_MARK_SUFFIX,
 };
 
 pub mod defaults {
@@ -211,6 +211,10 @@ pub struct PageServerConf {
 
     /// JWT token for use with the control plane API.
     pub control_plane_api_token: Option<SecretString>,
+
+    /// If true, pageserver will make best-effort to operate without a control plane: only
+    /// for use in major incidents.
+    pub control_plane_emergency_mode: bool,
 }
 
 /// We do not want to store this in a PageServerConf because the latter may be logged
@@ -288,6 +292,7 @@ struct PageServerConfigBuilder {
 
     control_plane_api: BuilderValue<Option<Url>>,
     control_plane_api_token: BuilderValue<Option<SecretString>>,
+    control_plane_emergency_mode: BuilderValue<bool>,
 }
 
 impl Default for PageServerConfigBuilder {
@@ -355,6 +360,7 @@ impl Default for PageServerConfigBuilder {
 
             control_plane_api: Set(None),
             control_plane_api_token: Set(None),
+            control_plane_emergency_mode: Set(false),
         }
     }
 }
@@ -491,6 +497,10 @@ impl PageServerConfigBuilder {
         self.control_plane_api_token = BuilderValue::Set(token)
     }
 
+    pub fn control_plane_emergency_mode(&mut self, enabled: bool) {
+        self.control_plane_emergency_mode = BuilderValue::Set(enabled)
+    }
+
     pub fn build(self) -> anyhow::Result<PageServerConf> {
         let concurrent_tenant_size_logical_size_queries = self
             .concurrent_tenant_size_logical_size_queries
@@ -582,6 +592,9 @@ impl PageServerConfigBuilder {
             control_plane_api_token: self
                 .control_plane_api_token
                 .ok_or(anyhow!("missing control_plane_api_token"))?,
+            control_plane_emergency_mode: self
+                .control_plane_emergency_mode
+                .ok_or(anyhow!("missing control_plane_emergency_mode"))?,
         })
     }
 }
@@ -631,10 +644,18 @@ impl PageServerConf {
 
     /// Points to a place in pageserver's local directory,
     /// where certain tenant's tenantconf file should be located.
+    ///
+    /// Legacy: superseded by tenant_location_config_path.  Eventually
+    /// remove this function.
     pub fn tenant_config_path(&self, tenant_id: &TenantId) -> Utf8PathBuf {
         self.tenant_path(tenant_id).join(TENANT_CONFIG_NAME)
     }
 
+    pub fn tenant_location_config_path(&self, tenant_id: &TenantId) -> Utf8PathBuf {
+        self.tenant_path(tenant_id)
+            .join(TENANT_LOCATION_CONFIG_NAME)
+    }
+
     pub fn timelines_path(&self, tenant_id: &TenantId) -> Utf8PathBuf {
         self.tenant_path(tenant_id).join(TIMELINES_SEGMENT_NAME)
     }
@@ -799,6 +820,10 @@ impl PageServerConf {
                         builder.control_plane_api_token(Some(parsed.into()))
                     }
                 },
+                "control_plane_emergency_mode" => {
+                    builder.control_plane_emergency_mode(parse_toml_bool(key, item)?)
+
+                },
                 _ => bail!("unrecognized pageserver option '{key}'"),
             }
         }
@@ -968,6 +993,7 @@ impl PageServerConf {
             background_task_maximum_delay: Duration::ZERO,
             control_plane_api: None,
             control_plane_api_token: None,
+            control_plane_emergency_mode: false,
         }
     }
 }
@@ -1191,7 +1217,8 @@ background_task_maximum_delay = '334 s'
                     defaults::DEFAULT_BACKGROUND_TASK_MAXIMUM_DELAY
                 )?,
                 control_plane_api: None,
-                control_plane_api_token: None
+                control_plane_api_token: None,
+                control_plane_emergency_mode: false
             },
             "Correct defaults should be used when no config values are provided"
         );
@@ -1247,7 +1274,8 @@ background_task_maximum_delay = '334 s'
                 ondemand_download_behavior_treat_error_as_warn: false,
                 background_task_maximum_delay: Duration::from_secs(334),
                 control_plane_api: None,
-                control_plane_api_token: None
+                control_plane_api_token: None,
+                control_plane_emergency_mode: false
             },
             "Should be able to parse all basic config values correctly"
         );

pageserver/src/consumption_metrics.rs

@@ -2,6 +2,7 @@
 //! and push them to a HTTP endpoint.
 use crate::context::{DownloadBehavior, RequestContext};
 use crate::task_mgr::{self, TaskKind, BACKGROUND_RUNTIME};
+use crate::tenant::tasks::BackgroundLoopKind;
 use crate::tenant::{mgr, LogicalSizeCalculationCause};
 use camino::Utf8PathBuf;
 use consumption_metrics::EventType;
@@ -10,6 +11,7 @@ use reqwest::Url;
 use std::collections::HashMap;
 use std::sync::Arc;
 use std::time::{Duration, SystemTime};
+use tokio::time::Instant;
 use tracing::*;
 use utils::id::NodeId;
 
@@ -87,22 +89,12 @@ pub async fn collect_metrics(
 
     let node_id = node_id.to_string();
 
-    // reminder: ticker is ready immediatedly
-    let mut ticker = tokio::time::interval(metric_collection_interval);
-
     loop {
-        let tick_at = tokio::select! {
-            _ = cancel.cancelled() => return Ok(()),
-            tick_at = ticker.tick() => tick_at,
-        };
+        let started_at = Instant::now();
 
         // these are point in time, with variable "now"
         let metrics = metrics::collect_all_metrics(&cached_metrics, &ctx).await;
 
-        if metrics.is_empty() {
-            continue;
-        }
-
         let metrics = Arc::new(metrics);
 
         // why not race cancellation here? because we are one of the last tasks, and if we are
@@ -141,10 +133,19 @@ pub async fn collect_metrics(
         let (_, _) = tokio::join!(flush, upload);
 
         crate::tenant::tasks::warn_when_period_overrun(
-            tick_at.elapsed(),
+            started_at.elapsed(),
             metric_collection_interval,
-            "consumption_metrics_collect_metrics",
+            BackgroundLoopKind::ConsumptionMetricsCollectMetrics,
         );
+
+        let res = tokio::time::timeout_at(
+            started_at + metric_collection_interval,
+            task_mgr::shutdown_token().cancelled(),
+        )
+        .await;
+        if res.is_ok() {
+            return Ok(());
+        }
     }
 }
 
@@ -243,16 +244,14 @@ async fn calculate_synthetic_size_worker(
     ctx: &RequestContext,
 ) -> anyhow::Result<()> {
     info!("starting calculate_synthetic_size_worker");
+    scopeguard::defer! {
+        info!("calculate_synthetic_size_worker stopped");
+    };
 
-    // reminder: ticker is ready immediatedly
-    let mut ticker = tokio::time::interval(synthetic_size_calculation_interval);
     let cause = LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize;
 
     loop {
-        let tick_at = tokio::select! {
-            _ = task_mgr::shutdown_watcher() => return Ok(()),
-            tick_at = ticker.tick() => tick_at,
-        };
+        let started_at = Instant::now();
 
         let tenants = match mgr::list_tenants().await {
             Ok(tenants) => tenants,
@@ -268,6 +267,11 @@ async fn calculate_synthetic_size_worker(
             }
 
             if let Ok(tenant) = mgr::get_tenant(tenant_id, true).await {
+                // TODO should we use concurrent_background_tasks_rate_limit() here, like the other background tasks?
+                // We can put in some prioritization for consumption metrics.
+                // Same for the loop that fetches computed metrics.
+                // By using the same limiter, we centralize metrics collection for "start" and "finished" counters,
+                // which turns out is really handy to understand the system.
                 if let Err(e) = tenant.calculate_synthetic_size(cause, ctx).await {
                     error!("failed to calculate synthetic size for tenant {tenant_id}: {e:#}");
                 }
@@ -275,9 +279,18 @@ async fn calculate_synthetic_size_worker(
         }
 
         crate::tenant::tasks::warn_when_period_overrun(
-            tick_at.elapsed(),
+            started_at.elapsed(),
             synthetic_size_calculation_interval,
-            "consumption_metrics_synthetic_size_worker",
+            BackgroundLoopKind::ConsumptionMetricsSyntheticSizeWorker,
         );
+
+        let res = tokio::time::timeout_at(
+            started_at + synthetic_size_calculation_interval,
+            task_mgr::shutdown_token().cancelled(),
+        )
+        .await;
+        if res.is_ok() {
+            return Ok(());
+        }
     }
 }

pageserver/src/control_plane_client.rs

@@ -133,6 +133,8 @@ impl ControlPlaneGenerationsApi for ControlPlaneClient {
             node_id: self.node_id,
         };
 
+        fail::fail_point!("control-plane-client-re-attach");
+
         let response: ReAttachResponse = self.retry_http_forever(&re_attach_path, request).await?;
         tracing::info!(
             "Received re-attach response with {} tenants",
@@ -168,6 +170,8 @@ impl ControlPlaneGenerationsApi for ControlPlaneClient {
                 .collect(),
         };
 
+        fail::fail_point!("control-plane-client-validate");
+
         let response: ValidateResponse = self.retry_http_forever(&re_attach_path, request).await?;
 
         Ok(response

pageserver/src/deletion_queue.rs

@@ -40,7 +40,6 @@ use validator::ValidatorQueueMessage;
 
 use crate::{config::PageServerConf, tenant::storage_layer::LayerFileName};
 
-// TODO: adminstrative "panic button" config property to disable all deletions
 // TODO: configurable for how long to wait before executing deletions
 
 /// We aggregate object deletions from many tenants in one place, for several reasons:
@@ -154,7 +153,7 @@ impl FlushOp {
 
 #[derive(Clone, Debug)]
 pub struct DeletionQueueClient {
-    tx: tokio::sync::mpsc::Sender<ListWriterQueueMessage>,
+    tx: tokio::sync::mpsc::UnboundedSender<ListWriterQueueMessage>,
     executor_tx: tokio::sync::mpsc::Sender<DeleterMessage>,
 
     lsn_table: Arc<std::sync::RwLock<VisibleLsnUpdates>>,
@@ -186,7 +185,7 @@ where
     V: Serialize,
     I: AsRef<[u8]>,
 {
-    let transformed = input.iter().map(|(k, v)| (hex::encode(k), v.clone()));
+    let transformed = input.iter().map(|(k, v)| (hex::encode(k), v));
 
     transformed
         .collect::<HashMap<String, &V>>()
@@ -213,7 +212,7 @@ where
 
 /// Files ending with this suffix will be ignored and erased
 /// during recovery as startup.
-const TEMP_SUFFIX: &str = ".tmp";
+const TEMP_SUFFIX: &str = "tmp";
 
 #[serde_as]
 #[derive(Debug, Serialize, Deserialize)]
@@ -325,10 +324,7 @@ impl DeletionList {
             return false;
         }
 
-        let timeline_entry = tenant_entry
-            .timelines
-            .entry(*timeline)
-            .or_insert_with(Vec::new);
+        let timeline_entry = tenant_entry.timelines.entry(*timeline).or_default();
 
         let timeline_remote_path = remote_timeline_path(tenant, timeline);
 
@@ -420,7 +416,7 @@ pub enum DeletionQueueError {
 impl DeletionQueueClient {
     pub(crate) fn broken() -> Self {
         // Channels whose receivers are immediately dropped.
-        let (tx, _rx) = tokio::sync::mpsc::channel(1);
+        let (tx, _rx) = tokio::sync::mpsc::unbounded_channel();
         let (executor_tx, _executor_rx) = tokio::sync::mpsc::channel(1);
         Self {
             tx,
@@ -432,12 +428,12 @@ impl DeletionQueueClient {
     /// This is cancel-safe.  If you drop the future before it completes, the message
     /// is not pushed, although in the context of the deletion queue it doesn't matter: once
     /// we decide to do a deletion the decision is always final.
-    async fn do_push<T>(
+    fn do_push<T>(
         &self,
-        queue: &tokio::sync::mpsc::Sender<T>,
+        queue: &tokio::sync::mpsc::UnboundedSender<T>,
         msg: T,
     ) -> Result<(), DeletionQueueError> {
-        match queue.send(msg).await {
+        match queue.send(msg) {
             Ok(_) => Ok(()),
             Err(e) => {
                 // This shouldn't happen, we should shut down all tenants before
@@ -449,7 +445,7 @@ impl DeletionQueueClient {
         }
     }
 
-    pub(crate) async fn recover(
+    pub(crate) fn recover(
         &self,
         attached_tenants: HashMap<TenantId, Generation>,
     ) -> Result<(), DeletionQueueError> {
@@ -457,7 +453,6 @@ impl DeletionQueueClient {
             &self.tx,
             ListWriterQueueMessage::Recover(RecoverOp { attached_tenants }),
         )
-        .await
     }
 
     /// When a Timeline wishes to update the remote_consistent_lsn that it exposes to the outside
@@ -530,6 +525,21 @@ impl DeletionQueueClient {
             return self.flush_immediate().await;
         }
 
+        self.push_layers_sync(tenant_id, timeline_id, current_generation, layers)
+    }
+
+    /// When a Tenant has a generation, push_layers is always synchronous because
+    /// the ListValidator channel is an unbounded channel.
+    ///
+    /// This can be merged into push_layers when we remove the Generation-less mode
+    /// support (`<https://github.com/neondatabase/neon/issues/5395>`)
+    pub(crate) fn push_layers_sync(
+        &self,
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+        current_generation: Generation,
+        layers: Vec<(LayerFileName, Generation)>,
+    ) -> Result<(), DeletionQueueError> {
         metrics::DELETION_QUEUE
             .keys_submitted
             .inc_by(layers.len() as u64);
@@ -543,17 +553,16 @@ impl DeletionQueueClient {
                 objects: Vec::new(),
             }),
         )
-        .await
     }
 
     /// This is cancel-safe.  If you drop the future the flush may still happen in the background.
     async fn do_flush<T>(
         &self,
-        queue: &tokio::sync::mpsc::Sender<T>,
+        queue: &tokio::sync::mpsc::UnboundedSender<T>,
         msg: T,
         rx: tokio::sync::oneshot::Receiver<()>,
     ) -> Result<(), DeletionQueueError> {
-        self.do_push(queue, msg).await?;
+        self.do_push(queue, msg)?;
         if rx.await.is_err() {
             // This shouldn't happen if tenants are shut down before deletion queue.  If we
             // encounter a bug like this, then a flusher will incorrectly believe it has flushed
@@ -574,6 +583,18 @@ impl DeletionQueueClient {
             .await
     }
 
+    /// Issue a flush without waiting for it to complete.  This is useful on advisory flushes where
+    /// the caller wants to avoid the risk of waiting for lots of enqueued work, such as on tenant
+    /// detach where flushing is nice but not necessary.
+    ///
+    /// This function provides no guarantees of work being done.
+    pub fn flush_advisory(&self) {
+        let (flush_op, _) = FlushOp::new();
+
+        // Transmit the flush message, ignoring any result (such as a closed channel during shutdown).
+        drop(self.tx.send(ListWriterQueueMessage::FlushExecute(flush_op)));
+    }
+
     // Wait until all previous deletions are executed
     pub(crate) async fn flush_execute(&self) -> Result<(), DeletionQueueError> {
         debug!("flush_execute: flushing to deletion lists...");
@@ -590,9 +611,7 @@ impl DeletionQueueClient {
         // Flush any immediate-mode deletions (the above backend flush will only flush
         // the executor if deletions had flowed through the backend)
         debug!("flush_execute: flushing execution...");
-        let (flush_op, rx) = FlushOp::new();
-        self.do_flush(&self.executor_tx, DeleterMessage::Flush(flush_op), rx)
-            .await?;
+        self.flush_immediate().await?;
         debug!("flush_execute: finished flushing execution...");
         Ok(())
     }
@@ -647,8 +666,10 @@ impl DeletionQueue {
     where
         C: ControlPlaneGenerationsApi + Send + Sync,
     {
-        // Deep channel: it consumes deletions from all timelines and we do not want to block them
-        let (tx, rx) = tokio::sync::mpsc::channel(16384);
+        // Unbounded channel: enables non-async functions to submit deletions.  The actual length is
+        // constrained by how promptly the ListWriter wakes up and drains it, which should be frequent
+        // enough to avoid this taking pathologically large amount of memory.
+        let (tx, rx) = tokio::sync::mpsc::unbounded_channel();
 
         // Shallow channel: it carries DeletionLists which each contain up to thousands of deletions
         let (backend_tx, backend_rx) = tokio::sync::mpsc::channel(16);
@@ -961,7 +982,7 @@ mod test {
         // Basic test that the deletion queue processes the deletions we pass into it
         let ctx = setup("deletion_queue_smoke").expect("Failed test setup");
         let client = ctx.deletion_queue.new_client();
-        client.recover(HashMap::new()).await?;
+        client.recover(HashMap::new())?;
 
         let layer_file_name_1: LayerFileName = "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap();
         let tenant_id = ctx.harness.tenant_id;
@@ -1029,7 +1050,7 @@ mod test {
     async fn deletion_queue_validation() -> anyhow::Result<()> {
         let ctx = setup("deletion_queue_validation").expect("Failed test setup");
         let client = ctx.deletion_queue.new_client();
-        client.recover(HashMap::new()).await?;
+        client.recover(HashMap::new())?;
 
         // Generation that the control plane thinks is current
         let latest_generation = Generation::new(0xdeadbeef);
@@ -1086,7 +1107,7 @@ mod test {
         // Basic test that the deletion queue processes the deletions we pass into it
         let mut ctx = setup("deletion_queue_recovery").expect("Failed test setup");
         let client = ctx.deletion_queue.new_client();
-        client.recover(HashMap::new()).await?;
+        client.recover(HashMap::new())?;
 
         let tenant_id = ctx.harness.tenant_id;
 
@@ -1149,9 +1170,7 @@ mod test {
         drop(client);
         ctx.restart().await;
         let client = ctx.deletion_queue.new_client();
-        client
-            .recover(HashMap::from([(tenant_id, now_generation)]))
-            .await?;
+        client.recover(HashMap::from([(tenant_id, now_generation)]))?;
 
         info!("Flush-executing");
         client.flush_execute().await?;
@@ -1177,7 +1196,7 @@ pub(crate) mod mock {
     };
 
     pub struct ConsumerState {
-        rx: tokio::sync::mpsc::Receiver<ListWriterQueueMessage>,
+        rx: tokio::sync::mpsc::UnboundedReceiver<ListWriterQueueMessage>,
         executor_rx: tokio::sync::mpsc::Receiver<DeleterMessage>,
     }
 
@@ -1254,7 +1273,7 @@ pub(crate) mod mock {
     }
 
     pub struct MockDeletionQueue {
-        tx: tokio::sync::mpsc::Sender<ListWriterQueueMessage>,
+        tx: tokio::sync::mpsc::UnboundedSender<ListWriterQueueMessage>,
         executor_tx: tokio::sync::mpsc::Sender<DeleterMessage>,
         executed: Arc<AtomicUsize>,
         remote_storage: Option<GenericRemoteStorage>,
@@ -1264,7 +1283,7 @@ pub(crate) mod mock {
 
     impl MockDeletionQueue {
         pub fn new(remote_storage: Option<GenericRemoteStorage>) -> Self {
-            let (tx, rx) = tokio::sync::mpsc::channel(16384);
+            let (tx, rx) = tokio::sync::mpsc::unbounded_channel();
             let (executor_tx, executor_rx) = tokio::sync::mpsc::channel(16384);
 
             let executed = Arc::new(AtomicUsize::new(0));
@@ -1279,10 +1298,6 @@ pub(crate) mod mock {
             }
         }
 
-        pub fn get_executed(&self) -> usize {
-            self.executed.load(Ordering::Relaxed)
-        }
-
         #[allow(clippy::await_holding_lock)]
         pub async fn pump(&self) {
             if let Some(remote_storage) = &self.remote_storage {

pageserver/src/deletion_queue/deleter.rs

@@ -13,6 +13,7 @@ use std::time::Duration;
 use tokio_util::sync::CancellationToken;
 use tracing::info;
 use tracing::warn;
+use utils::backoff;
 
 use crate::metrics;
 
@@ -63,7 +64,19 @@ impl Deleter {
             Err(anyhow::anyhow!("failpoint hit"))
         });
 
-        self.remote_storage.delete_objects(&self.accumulator).await
+        // A backoff::retry is used here for two reasons:
+        // - To provide a backoff rather than busy-polling the API on errors
+        // - To absorb transient 429/503 conditions without hitting our error
+        //   logging path for issues deleting objects.
+        backoff::retry(
+            || async { self.remote_storage.delete_objects(&self.accumulator).await },
+            |_| false,
+            3,
+            10,
+            "executing deletion batch",
+            backoff::Cancel::new(self.cancel.clone(), || anyhow::anyhow!("Shutting down")),
+        )
+        .await
     }
 
     /// Block until everything in accumulator has been executed
@@ -88,7 +101,10 @@ impl Deleter {
                     self.accumulator.clear();
                 }
                 Err(e) => {
-                    warn!("DeleteObjects request failed: {e:#}, will retry");
+                    if self.cancel.is_cancelled() {
+                        return Err(DeletionQueueError::ShuttingDown);
+                    }
+                    warn!("DeleteObjects request failed: {e:#}, will continue trying");
                     metrics::DELETION_QUEUE
                         .remote_errors
                         .with_label_values(&["execute"])

pageserver/src/deletion_queue/list_writer.rs

@@ -34,8 +34,6 @@ use crate::deletion_queue::TEMP_SUFFIX;
 use crate::metrics;
 use crate::tenant::remote_timeline_client::remote_layer_path;
 use crate::tenant::storage_layer::LayerFileName;
-use crate::virtual_file;
-use crate::virtual_file::on_fatal_io_error;
 
 // The number of keys in a DeletionList before we will proactively persist it
 // (without reaching a flush deadline).  This aims to deliver objects of the order
@@ -87,7 +85,7 @@ pub(super) struct ListWriter {
     conf: &'static PageServerConf,
 
     // Incoming frontend requests to delete some keys
-    rx: tokio::sync::mpsc::Receiver<ListWriterQueueMessage>,
+    rx: tokio::sync::mpsc::UnboundedReceiver<ListWriterQueueMessage>,
 
     // Outbound requests to the backend to execute deletion lists we have composed.
     tx: tokio::sync::mpsc::Sender<ValidatorQueueMessage>,
@@ -113,7 +111,7 @@ impl ListWriter {
 
     pub(super) fn new(
         conf: &'static PageServerConf,
-        rx: tokio::sync::mpsc::Receiver<ListWriterQueueMessage>,
+        rx: tokio::sync::mpsc::UnboundedReceiver<ListWriterQueueMessage>,
         tx: tokio::sync::mpsc::Sender<ValidatorQueueMessage>,
         cancel: CancellationToken,
     ) -> Self {
@@ -197,7 +195,7 @@ impl ListWriter {
                     debug!("Deletion header {header_path} not found, first start?");
                     Ok(None)
                 } else {
-                    on_fatal_io_error(&e);
+                    Err(anyhow::anyhow!(e))
                 }
             }
         }
@@ -223,15 +221,16 @@ impl ListWriter {
             Err(e) => {
                 warn!("Failed to open deletion list directory {deletion_directory}: {e:#}");
 
-                // This is fatal: any failure to read this local directory indicates a
-                // storage problem or configuration problem of the node.
-                virtual_file::on_fatal_io_error(&e);
+                // Give up: if we can't read the deletion list directory, we probably can't
+                // write lists into it later, so the queue won't work.
+                return Err(e.into());
             }
         };
 
         let list_name_pattern =
             Regex::new("(?<sequence>[a-zA-Z0-9]{16})-(?<version>[a-zA-Z0-9]{2}).list").unwrap();
 
+        let temp_extension = format!(".{TEMP_SUFFIX}");
         let header_path = self.conf.deletion_header_path();
         let mut seqs: Vec<u64> = Vec::new();
         while let Some(dentry) = dir.next_entry().await? {
@@ -243,7 +242,7 @@ impl ListWriter {
                 continue;
             }
 
-            if dentry_str.ends_with(TEMP_SUFFIX) {
+            if dentry_str.ends_with(&temp_extension) {
                 info!("Cleaning up temporary file {dentry_str}");
                 let absolute_path =
                     deletion_directory.join(dentry.file_name().to_str().expect("non-Unicode path"));
@@ -251,8 +250,6 @@ impl ListWriter {
                     // Non-fatal error: we will just leave the file behind but not
                     // try and load it.
                     warn!("Failed to clean up temporary file {absolute_path}: {e:#}");
-
-                    virtual_file::on_fatal_io_error(&e);
                 }
 
                 continue;
@@ -265,7 +262,7 @@ impl ListWriter {
                     .expect("Non optional group should be present")
                     .as_str()
             } else {
-                warn!("Unexpected filename in deletion queue: {basename}");
+                warn!("Unexpected key in deletion queue: {basename}");
                 metrics::DELETION_QUEUE.unexpected_errors.inc();
                 continue;
             };
@@ -293,12 +290,7 @@ impl ListWriter {
         for s in seqs {
             let list_path = self.conf.deletion_list_path(s);
 
-            let list_bytes = match tokio::fs::read(&list_path).await {
-                Ok(b) => b,
-                Err(e) => {
-                    virtual_file::on_fatal_io_error(&e);
-                }
-            };
+            let list_bytes = tokio::fs::read(&list_path).await?;
 
             let mut deletion_list = match serde_json::from_slice::<DeletionList>(&list_bytes) {
                 Ok(l) => l,

pageserver/src/deletion_queue/validator.rs

@@ -28,7 +28,6 @@ use crate::config::PageServerConf;
 use crate::control_plane_client::ControlPlaneGenerationsApi;
 use crate::control_plane_client::RetryForeverError;
 use crate::metrics;
-use crate::virtual_file::on_fatal_io_error;
 
 use super::deleter::DeleterMessage;
 use super::DeletionHeader;
@@ -117,11 +116,6 @@ where
     /// Valid LSN updates propagate back to Timelines immediately, valid DeletionLists
     /// go into the queue of ready-to-execute lists.
     async fn validate(&mut self) -> Result<(), DeletionQueueError> {
-        // Figure out for each tenant which generation number to validate.
-        //
-        // It is sufficient to validate the max generation number of each tenant because only the
-        // highest generation number can possibly be valid. Hence this map will collect the
-        // highest generation pending validation for each tenant.
         let mut tenant_generations = HashMap::new();
         for list in &self.pending_lists {
             for (tenant_id, tenant_list) in &list.tenants {
@@ -226,6 +220,8 @@ where
                     warn!("Dropping stale deletions for tenant {tenant_id} in generation {:?}, objects may be leaked", tenant.generation);
                     metrics::DELETION_QUEUE.keys_dropped.inc_by(tenant.len() as u64);
                     mutated = true;
+                } else {
+                    metrics::DELETION_QUEUE.keys_validated.inc_by(tenant.len() as u64);
                 }
                 this_list_valid
             });
@@ -252,11 +248,6 @@ where
                 }
             }
 
-            // Assert monotonicity of the list sequence numbers we are processing
-            if let Some(validated) = validated_sequence {
-                assert!(list.sequence >= validated)
-            }
-
             validated_sequence = Some(list.sequence);
         }
 
@@ -304,8 +295,7 @@ where
                 // issue (probably permissions) has been fixed by then.
                 tracing::error!("Failed to delete {list_path}: {e:#}");
                 metrics::DELETION_QUEUE.unexpected_errors.inc();
-
-                on_fatal_io_error(&e);
+                break;
             }
         }
     }

pageserver/src/disk_usage_eviction_task.rs

@@ -411,6 +411,11 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
                                 evictions_failed.file_sizes += file_size;
                                 evictions_failed.count += 1;
                             }
+                            Some(Err(EvictionError::MetadataInconsistency(detail))) => {
+                                warn!(%layer, "failed to evict layer: {detail}");
+                                evictions_failed.file_sizes += file_size;
+                                evictions_failed.count += 1;
+                            }
                             None => {
                                 assert!(cancel.is_cancelled());
                                 return;

pageserver/src/http/openapi_spec.yml

@@ -93,9 +93,16 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
     delete:
       description: |
-        Attempts to delete specified tenant. 500 and 409 errors should be retried until 404 is retrieved.
+        Attempts to delete specified tenant. 500, 503 and 409 errors should be retried until 404 is retrieved.
         404 means that deletion successfully finished"
       responses:
         "400":
@@ -134,6 +141,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
 
   /v1/tenant/{tenant_id}/timeline:
     parameters:
@@ -178,6 +192,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
 
   /v1/tenant/{tenant_id}/timeline/{timeline_id}:
     parameters:
@@ -226,6 +247,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
     delete:
       description: "Attempts to delete specified timeline. 500 and 409 errors should be retried"
       responses:
@@ -265,7 +293,74 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/PreconditionFailedError"
+        "500":
+          description: Generic operation error
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
 
+  /v1/tenant/{tenant_id}/timeline/{timeline_id}/get_timestamp_of_lsn:
+    parameters:
+      - name: tenant_id
+        in: path
+        required: true
+        schema:
+          type: string
+          format: hex
+      - name: timeline_id
+        in: path
+        required: true
+        schema:
+          type: string
+          format: hex
+    get:
+      description: Get timestamp for a given LSN
+      parameters:
+        - name: lsn
+          in: query
+          required: true
+          schema:
+            type: integer
+          description: A LSN to get the timestamp
+      responses:
+        "200":
+          description: OK
+          content:
+            application/json:
+              schema:
+                type: string
+                format: date-time
+        "400":
+          description: Error when no tenant id found in path, no timeline id or invalid timestamp
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/Error"
+        "401":
+          description: Unauthorized Error
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/UnauthorizedError"
+        "403":
+          description: Forbidden Error
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ForbiddenError"
+        "404":
+          description: Timeline not found, or there is no timestamp information for the given lsn
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/NotFoundError"
         "500":
           description: Generic operation error
           content:
@@ -297,13 +392,19 @@ paths:
             type: string
             format: date-time
           description: A timestamp to get the LSN
+        - name: version
+          in: query
+          required: false
+          schema:
+            type: integer
+          description: The version of the endpoint to use
       responses:
         "200":
           description: OK
           content:
             application/json:
               schema:
-                type: string
+                $ref: "#/components/schemas/LsnByTimestampResponse"
         "400":
           description: Error when no tenant id found in path, no timeline id or invalid timestamp
           content:
@@ -328,6 +429,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
   /v1/tenant/{tenant_id}/timeline/{timeline_id}/do_gc:
     parameters:
       - name: tenant_id
@@ -375,6 +483,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
   /v1/tenant/{tenant_id}/attach:
     parameters:
       - name: tenant_id
@@ -465,6 +580,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
 
   /v1/tenant/{tenant_id}/detach:
     parameters:
@@ -518,6 +640,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
 
   /v1/tenant/{tenant_id}/ignore:
     parameters:
@@ -560,6 +689,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
 
   /v1/tenant/{tenant_id}/load:
     parameters:
@@ -604,6 +740,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
 
   /v1/tenant/{tenant_id}/synthetic_size:
     parameters:
@@ -641,6 +784,12 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
 
   /v1/tenant/{tenant_id}/size:
     parameters:
@@ -704,6 +853,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
 
   /v1/tenant/{tenant_id}/timeline/:
     parameters:
@@ -780,6 +936,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
   /v1/tenant/:
     get:
       description: Get tenants list
@@ -810,6 +973,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
     post:
       description: |
         Create a tenant. Returns new tenant id on success.
@@ -860,6 +1030,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
 
   /v1/tenant/config:
     put:
@@ -905,6 +1082,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
   /v1/tenant/{tenant_id}/config/:
     parameters:
       - name: tenant_id
@@ -954,6 +1138,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
 components:
   securitySchemes:
     JWT:
@@ -1199,6 +1390,19 @@ components:
           type: string
           format: hex
 
+    LsnByTimestampResponse:
+      type: object
+      required:
+        - lsn
+        - kind
+      properties:
+        lsn:
+          type: string
+          format: hex
+        kind:
+          type: string
+          enum: [past, present, future, nodata]
+
     Error:
       type: object
       required:
@@ -1220,6 +1424,13 @@ components:
       properties:
         msg:
           type: string
+    ServiceUnavailableError:
+      type: object
+      required:
+        - msg
+      properties:
+        msg:
+          type: string
     NotFoundError:
       type: object
       required:

pageserver/src/http/routes.rs

@@ -2,17 +2,22 @@
 //! Management HTTP API
 //!
 use std::collections::HashMap;
+use std::str::FromStr;
 use std::sync::Arc;
 
 use anyhow::{anyhow, Context, Result};
 use futures::TryFutureExt;
+use humantime::format_rfc3339;
+use hyper::header;
 use hyper::StatusCode;
 use hyper::{Body, Request, Response, Uri};
 use metrics::launch_timestamp::LaunchTimestamp;
 use pageserver_api::models::{
-    DownloadRemoteLayersTaskSpawnRequest, TenantAttachRequest, TenantLoadRequest,
+    DownloadRemoteLayersTaskSpawnRequest, LocationConfigMode, TenantAttachRequest,
+    TenantLoadRequest, TenantLocationConfigRequest,
 };
 use remote_storage::GenericRemoteStorage;
+use serde_with::{serde_as, DisplayFromStr};
 use tenant_size_model::{SizeResult, StorageModel};
 use tokio_util::sync::CancellationToken;
 use tracing::*;
@@ -29,7 +34,7 @@ use crate::deletion_queue::DeletionQueueClient;
 use crate::metrics::{StorageTimeOperation, STORAGE_TIME_GLOBAL};
 use crate::pgdatadir_mapping::LsnForTimestamp;
 use crate::task_mgr::TaskKind;
-use crate::tenant::config::TenantConfOpt;
+use crate::tenant::config::{LocationConf, TenantConfOpt};
 use crate::tenant::mgr::{
     GetTenantError, SetNewTenantConfigError, TenantMapInsertError, TenantStateError,
 };
@@ -75,7 +80,7 @@ impl State {
         disk_usage_eviction_state: Arc<disk_usage_eviction_task::State>,
         deletion_queue_client: DeletionQueueClient,
     ) -> anyhow::Result<Self> {
-        let allowlist_routes = ["/v1/status", "/v1/doc", "/swagger.yml"]
+        let allowlist_routes = ["/v1/status", "/v1/doc", "/swagger.yml", "/metrics"]
             .iter()
             .map(|v| v.parse().unwrap())
             .collect::<Vec<_>>();
@@ -132,11 +137,9 @@ impl From<PageReconstructError> for ApiError {
                 ApiError::InternalServerError(anyhow::anyhow!("request was cancelled"))
             }
             PageReconstructError::AncestorStopping(_) => {
-                ApiError::ResourceUnavailable(format!("{pre}"))
-            }
-            PageReconstructError::WalRedo(pre) => {
-                ApiError::InternalServerError(anyhow::Error::new(pre))
+                ApiError::ResourceUnavailable(format!("{pre}").into())
             }
+            PageReconstructError::WalRedo(pre) => ApiError::InternalServerError(pre),
         }
     }
 }
@@ -145,12 +148,15 @@ impl From<TenantMapInsertError> for ApiError {
     fn from(tmie: TenantMapInsertError) -> ApiError {
         match tmie {
             TenantMapInsertError::StillInitializing | TenantMapInsertError::ShuttingDown => {
-                ApiError::ResourceUnavailable(format!("{tmie}"))
+                ApiError::ResourceUnavailable(format!("{tmie}").into())
             }
             TenantMapInsertError::TenantAlreadyExists(id, state) => {
                 ApiError::Conflict(format!("tenant {id} already exists, state: {state:?}"))
             }
-            TenantMapInsertError::Closure(e) => ApiError::InternalServerError(e),
+            TenantMapInsertError::TenantExistsSecondary(id) => {
+                ApiError::Conflict(format!("tenant {id} already exists as secondary"))
+            }
+            TenantMapInsertError::Other(e) => ApiError::InternalServerError(e),
         }
     }
 }
@@ -159,9 +165,6 @@ impl From<TenantStateError> for ApiError {
     fn from(tse: TenantStateError) -> ApiError {
         match tse {
             TenantStateError::NotFound(tid) => ApiError::NotFound(anyhow!("tenant {}", tid).into()),
-            TenantStateError::NotActive(_) => {
-                ApiError::ResourceUnavailable("Tenant not yet active".into())
-            }
             TenantStateError::IsStopping(_) => {
                 ApiError::ResourceUnavailable("Tenant is stopping".into())
             }
@@ -391,6 +394,9 @@ async fn timeline_create_handler(
                     format!("{err:#}")
                 ))
             }
+            Err(e @ tenant::CreateTimelineError::AncestorNotActive) => {
+                json_response(StatusCode::SERVICE_UNAVAILABLE, HttpErrorBody::from_msg(e.to_string()))
+            }
             Err(tenant::CreateTimelineError::Other(err)) => Err(ApiError::InternalServerError(err)),
         }
     }
@@ -479,6 +485,8 @@ async fn get_lsn_by_timestamp_handler(
     let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
     check_permission(&request, Some(tenant_id))?;
 
+    let version: Option<u8> = parse_query_param(&request, "version")?;
+
     let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
     let timestamp_raw = must_get_query_param(&request, "timestamp")?;
     let timestamp = humantime::parse_rfc3339(&timestamp_raw)
@@ -490,13 +498,59 @@ async fn get_lsn_by_timestamp_handler(
     let timeline = active_timeline_of_active_tenant(tenant_id, timeline_id).await?;
     let result = timeline.find_lsn_for_timestamp(timestamp_pg, &ctx).await?;
 
-    let result = match result {
-        LsnForTimestamp::Present(lsn) => format!("{lsn}"),
-        LsnForTimestamp::Future(_lsn) => "future".into(),
-        LsnForTimestamp::Past(_lsn) => "past".into(),
-        LsnForTimestamp::NoData(_lsn) => "nodata".into(),
-    };
-    json_response(StatusCode::OK, result)
+    if version.unwrap_or(0) > 1 {
+        #[serde_as]
+        #[derive(serde::Serialize)]
+        struct Result {
+            #[serde_as(as = "DisplayFromStr")]
+            lsn: Lsn,
+            kind: &'static str,
+        }
+        let (lsn, kind) = match result {
+            LsnForTimestamp::Present(lsn) => (lsn, "present"),
+            LsnForTimestamp::Future(lsn) => (lsn, "future"),
+            LsnForTimestamp::Past(lsn) => (lsn, "past"),
+            LsnForTimestamp::NoData(lsn) => (lsn, "nodata"),
+        };
+        json_response(StatusCode::OK, Result { lsn, kind })
+    } else {
+        // FIXME: this is a temporary crutch not to break backwards compatibility
+        // See https://github.com/neondatabase/neon/pull/5608
+        let result = match result {
+            LsnForTimestamp::Present(lsn) => format!("{lsn}"),
+            LsnForTimestamp::Future(_lsn) => "future".into(),
+            LsnForTimestamp::Past(_lsn) => "past".into(),
+            LsnForTimestamp::NoData(_lsn) => "nodata".into(),
+        };
+        json_response(StatusCode::OK, result)
+    }
+}
+
+async fn get_timestamp_of_lsn_handler(
+    request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
+    check_permission(&request, Some(tenant_id))?;
+
+    let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
+
+    let lsn_str = must_get_query_param(&request, "lsn")?;
+    let lsn = Lsn::from_str(&lsn_str)
+        .with_context(|| format!("Invalid LSN: {lsn_str:?}"))
+        .map_err(ApiError::BadRequest)?;
+
+    let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
+    let timeline = active_timeline_of_active_tenant(tenant_id, timeline_id).await?;
+    let result = timeline.get_timestamp_for_lsn(lsn, &ctx).await?;
+
+    match result {
+        Some(time) => {
+            let time = format_rfc3339(postgres_ffi::from_pg_timestamp(time)).to_string();
+            json_response(StatusCode::OK, time)
+        }
+        None => json_response(StatusCode::NOT_FOUND, ()),
+    }
 }
 
 async fn tenant_attach_handler(
@@ -567,9 +621,14 @@ async fn tenant_detach_handler(
 
     let state = get_state(&request);
     let conf = state.conf;
-    mgr::detach_tenant(conf, tenant_id, detach_ignored.unwrap_or(false))
-        .instrument(info_span!("tenant_detach", %tenant_id))
-        .await?;
+    mgr::detach_tenant(
+        conf,
+        tenant_id,
+        detach_ignored.unwrap_or(false),
+        &state.deletion_queue_client,
+    )
+    .instrument(info_span!("tenant_detach", %tenant_id))
+    .await?;
 
     json_response(StatusCode::OK, ())
 }
@@ -622,6 +681,45 @@ async fn tenant_ignore_handler(
     json_response(StatusCode::OK, ())
 }
 
+async fn tenant_duplicate_handler(
+    mut request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let src_tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
+
+    let request_data: TenantCreateRequest = json_request(&mut request).await?;
+    let new_tenant_id = request_data.new_tenant_id;
+    check_permission(&request, None)?;
+
+    let _timer = STORAGE_TIME_GLOBAL
+        .get_metric_with_label_values(&[StorageTimeOperation::DuplicateTenant.into()])
+        .expect("bug")
+        .start_timer();
+
+    let tenant_conf =
+        TenantConfOpt::try_from(&request_data.config).map_err(ApiError::BadRequest)?;
+
+    let state = get_state(&request);
+
+    let generation = get_request_generation(state, request_data.generation)?;
+
+    let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Warn);
+
+    mgr::duplicate_tenant(
+        state.conf,
+        tenant_conf,
+        src_tenant_id,
+        new_tenant_id,
+        generation,
+        state.tenant_resources(),
+        &ctx,
+    )
+    .instrument(info_span!("tenant_duplicate", %src_tenant_id, tenant_id = %new_tenant_id))
+    .await?;
+
+    json_response(StatusCode::CREATED, TenantCreateResponse(new_tenant_id))
+}
+
 async fn tenant_list_handler(
     request: Request<Body>,
     _cancel: CancellationToken,
@@ -632,7 +730,7 @@ async fn tenant_list_handler(
         .instrument(info_span!("tenant_list"))
         .await
         .map_err(|_| {
-            ApiError::ResourceUnavailable("Tenant map is initializing or shutting down".to_string())
+            ApiError::ResourceUnavailable("Tenant map is initializing or shutting down".into())
         })?
         .iter()
         .map(|(id, state)| TenantInfo {
@@ -730,6 +828,10 @@ async fn tenant_size_handler(
         .map_err(ApiError::InternalServerError)?;
 
     let mut sizes = None;
+    let accepts_html = headers
+        .get(header::ACCEPT)
+        .map(|v| v == "text/html")
+        .unwrap_or_default();
     if !inputs_only.unwrap_or(false) {
         let storage_model = inputs
             .calculate_model()
@@ -737,11 +839,11 @@ async fn tenant_size_handler(
         let size = storage_model.calculate();
 
         // If request header expects html, return html
-        if headers["Accept"] == "text/html" {
+        if accepts_html {
             return synthetic_size_html_response(inputs, storage_model, size);
         }
         sizes = Some(size);
-    } else if headers["Accept"] == "text/html" {
+    } else if accepts_html {
         return Err(ApiError::BadRequest(anyhow!(
             "inputs_only parameter is incompatible with html output request"
         )));
@@ -892,7 +994,7 @@ fn synthetic_size_html_response(
 pub fn html_response(status: StatusCode, data: String) -> Result<Response<Body>, ApiError> {
     let response = Response::builder()
         .status(status)
-        .header(hyper::header::CONTENT_TYPE, "text/html")
+        .header(header::CONTENT_TYPE, "text/html")
         .body(Body::from(data.as_bytes().to_vec()))
         .map_err(|e| ApiError::InternalServerError(e.into()))?;
     Ok(response)
@@ -1011,6 +1113,56 @@ async fn update_tenant_config_handler(
     json_response(StatusCode::OK, ())
 }
 
+async fn put_tenant_location_config_handler(
+    mut request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let request_data: TenantLocationConfigRequest = json_request(&mut request).await?;
+    let tenant_id = request_data.tenant_id;
+    check_permission(&request, Some(tenant_id))?;
+
+    let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Warn);
+    let state = get_state(&request);
+    let conf = state.conf;
+
+    // The `Detached` state is special, it doesn't upsert a tenant, it removes
+    // its local disk content and drops it from memory.
+    if let LocationConfigMode::Detached = request_data.config.mode {
+        if let Err(e) = mgr::detach_tenant(conf, tenant_id, true, &state.deletion_queue_client)
+            .instrument(info_span!("tenant_detach", %tenant_id))
+            .await
+        {
+            match e {
+                TenantStateError::NotFound(_) => {
+                    // This API is idempotent: a NotFound on a detach is fine.
+                }
+                _ => return Err(e.into()),
+            }
+        }
+        return json_response(StatusCode::OK, ());
+    }
+
+    let location_conf =
+        LocationConf::try_from(&request_data.config).map_err(ApiError::BadRequest)?;
+
+    mgr::upsert_location(
+        state.conf,
+        tenant_id,
+        location_conf,
+        state.broker_client.clone(),
+        state.remote_storage.clone(),
+        state.deletion_queue_client.clone(),
+        &ctx,
+    )
+    .await
+    // TODO: badrequest assumes the caller was asking for something unreasonable, but in
+    // principle we might have hit something like concurrent API calls to the same tenant,
+    // which is not a 400 but a 409.
+    .map_err(ApiError::BadRequest)?;
+
+    json_response(StatusCode::OK, ())
+}
+
 /// Testing helper to transition a tenant to [`crate::tenant::TenantState::Broken`].
 async fn handle_tenant_break(
     r: Request<Body>,
@@ -1190,6 +1342,136 @@ async fn deletion_queue_flush(
     }
 }
 
+/// Try if `GetPage@Lsn` is successful, useful for manual debugging.
+async fn getpage_at_lsn_handler(
+    request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
+    let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
+    check_permission(&request, Some(tenant_id))?;
+
+    struct Key(crate::repository::Key);
+
+    impl std::str::FromStr for Key {
+        type Err = anyhow::Error;
+
+        fn from_str(s: &str) -> std::result::Result<Self, Self::Err> {
+            crate::repository::Key::from_hex(s).map(Key)
+        }
+    }
+
+    let key: Key = parse_query_param(&request, "key")?
+        .ok_or_else(|| ApiError::BadRequest(anyhow!("missing 'key' query parameter")))?;
+    let lsn: Lsn = parse_query_param(&request, "lsn")?
+        .ok_or_else(|| ApiError::BadRequest(anyhow!("missing 'lsn' query parameter")))?;
+
+    async {
+        let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
+        let timeline = active_timeline_of_active_tenant(tenant_id, timeline_id).await?;
+
+        let page = timeline.get(key.0, lsn, &ctx).await?;
+
+        Result::<_, ApiError>::Ok(
+            Response::builder()
+                .status(StatusCode::OK)
+                .header(header::CONTENT_TYPE, "application/octet-stream")
+                .body(hyper::Body::from(page))
+                .unwrap(),
+        )
+    }
+    .instrument(info_span!("timeline_get", %tenant_id, %timeline_id))
+    .await
+}
+
+async fn timeline_collect_keyspace(
+    request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
+    let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
+    check_permission(&request, Some(tenant_id))?;
+
+    struct Partitioning {
+        keys: crate::keyspace::KeySpace,
+
+        at_lsn: Lsn,
+    }
+
+    impl serde::Serialize for Partitioning {
+        fn serialize<S>(&self, serializer: S) -> std::result::Result<S::Ok, S::Error>
+        where
+            S: serde::Serializer,
+        {
+            use serde::ser::SerializeMap;
+            let mut map = serializer.serialize_map(Some(2))?;
+            map.serialize_key("keys")?;
+            map.serialize_value(&KeySpace(&self.keys))?;
+            map.serialize_key("at_lsn")?;
+            map.serialize_value(&WithDisplay(&self.at_lsn))?;
+            map.end()
+        }
+    }
+
+    struct WithDisplay<'a, T>(&'a T);
+
+    impl<'a, T: std::fmt::Display> serde::Serialize for WithDisplay<'a, T> {
+        fn serialize<S>(&self, serializer: S) -> std::result::Result<S::Ok, S::Error>
+        where
+            S: serde::Serializer,
+        {
+            serializer.collect_str(&self.0)
+        }
+    }
+
+    struct KeySpace<'a>(&'a crate::keyspace::KeySpace);
+
+    impl<'a> serde::Serialize for KeySpace<'a> {
+        fn serialize<S>(&self, serializer: S) -> std::result::Result<S::Ok, S::Error>
+        where
+            S: serde::Serializer,
+        {
+            use serde::ser::SerializeSeq;
+            let mut seq = serializer.serialize_seq(Some(self.0.ranges.len()))?;
+            for kr in &self.0.ranges {
+                seq.serialize_element(&KeyRange(kr))?;
+            }
+            seq.end()
+        }
+    }
+
+    struct KeyRange<'a>(&'a std::ops::Range<crate::repository::Key>);
+
+    impl<'a> serde::Serialize for KeyRange<'a> {
+        fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
+        where
+            S: serde::Serializer,
+        {
+            use serde::ser::SerializeTuple;
+            let mut t = serializer.serialize_tuple(2)?;
+            t.serialize_element(&WithDisplay(&self.0.start))?;
+            t.serialize_element(&WithDisplay(&self.0.end))?;
+            t.end()
+        }
+    }
+
+    let at_lsn: Option<Lsn> = parse_query_param(&request, "at_lsn")?;
+
+    async {
+        let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
+        let timeline = active_timeline_of_active_tenant(tenant_id, timeline_id).await?;
+        let at_lsn = at_lsn.unwrap_or_else(|| timeline.get_last_record_lsn());
+        let keys = timeline
+            .collect_keyspace(at_lsn, &ctx)
+            .await
+            .map_err(ApiError::InternalServerError)?;
+
+        json_response(StatusCode::OK, Partitioning { keys, at_lsn })
+    }
+    .instrument(info_span!("timeline_collect_keyspace", %tenant_id, %timeline_id))
+    .await
+}
+
 async fn active_timeline_of_active_tenant(
     tenant_id: TenantId,
     timeline_id: TimelineId,
@@ -1464,6 +1746,9 @@ pub fn make_router(
         .get("/v1/tenant/:tenant_id/config", |r| {
             api_handler(r, get_tenant_config_handler)
         })
+        .put("/v1/tenant/:tenant_id/location_config", |r| {
+            api_handler(r, put_tenant_location_config_handler)
+        })
         .get("/v1/tenant/:tenant_id/timeline", |r| {
             api_handler(r, timeline_list_handler)
         })
@@ -1482,6 +1767,9 @@ pub fn make_router(
         .post("/v1/tenant/:tenant_id/ignore", |r| {
             api_handler(r, tenant_ignore_handler)
         })
+        .post("/v1/tenant/:tenant_id/duplicate", |r| {
+            api_handler(r, tenant_duplicate_handler)
+        })
         .get("/v1/tenant/:tenant_id/timeline/:timeline_id", |r| {
             api_handler(r, timeline_detail_handler)
         })
@@ -1489,6 +1777,10 @@ pub fn make_router(
             "/v1/tenant/:tenant_id/timeline/:timeline_id/get_lsn_by_timestamp",
             |r| api_handler(r, get_lsn_by_timestamp_handler),
         )
+        .get(
+            "/v1/tenant/:tenant_id/timeline/:timeline_id/get_timestamp_of_lsn",
+            |r| api_handler(r, get_timestamp_of_lsn_handler),
+        )
         .put("/v1/tenant/:tenant_id/timeline/:timeline_id/do_gc", |r| {
             api_handler(r, timeline_gc_handler)
         })
@@ -1534,5 +1826,12 @@ pub fn make_router(
         .post("/v1/tracing/event", |r| {
             testing_api_handler("emit a tracing event", r, post_tracing_event_handler)
         })
+        .get("/v1/tenant/:tenant_id/timeline/:timeline_id/getpage", |r| {
+            testing_api_handler("getpage@lsn", r, getpage_at_lsn_handler)
+        })
+        .get(
+            "/v1/tenant/:tenant_id/timeline/:timeline_id/keyspace",
+            |r| testing_api_handler("read out the keyspace", r, timeline_collect_keyspace),
+        )
         .any(handler_404))
 }

pageserver/src/lib.rs

@@ -112,6 +112,10 @@ pub const METADATA_FILE_NAME: &str = "metadata";
 /// Full path: `tenants/<tenant_id>/config`.
 pub const TENANT_CONFIG_NAME: &str = "config";
 
+/// Per-tenant configuration file.
+/// Full path: `tenants/<tenant_id>/config`.
+pub const TENANT_LOCATION_CONFIG_NAME: &str = "config-v1";
+
 /// A suffix used for various temporary files. Any temporary files found in the
 /// data directory at pageserver startup can be automatically removed.
 pub const TEMP_FILE_SUFFIX: &str = "___temp";
@@ -145,6 +149,10 @@ fn ends_with_suffix(path: &Utf8Path, suffix: &str) -> bool {
     }
 }
 
+// FIXME: DO NOT ADD new query methods like this, which will have a next step of parsing timelineid
+// from the directory name. Instead create type "UninitMark(TimelineId)" and only parse it once
+// from the name.
+
 pub fn is_uninit_mark(path: &Utf8Path) -> bool {
     ends_with_suffix(path, TIMELINE_UNINIT_MARK_SUFFIX)
 }
@@ -169,6 +177,9 @@ fn is_walkdir_io_not_found(e: &walkdir::Error) -> bool {
 /// delaying is needed.
 #[derive(Clone)]
 pub struct InitializationOrder {
+    /// Each initial tenant load task carries this until it is done loading timelines from remote storage
+    pub initial_tenant_load_remote: Option<utils::completion::Completion>,
+
     /// Each initial tenant load task carries this until completion.
     pub initial_tenant_load: Option<utils::completion::Completion>,
 

pageserver/src/metrics.rs

@@ -51,6 +51,9 @@ pub enum StorageTimeOperation {
 
     #[strum(serialize = "create tenant")]
     CreateTenant,
+
+    #[strum(serialize = "duplicate tenant")]
+    DuplicateTenant,
 }
 
 pub static STORAGE_TIME_SUM_PER_TIMELINE: Lazy<CounterVec> = Lazy::new(|| {
@@ -691,10 +694,9 @@ impl StorageIoTime {
         .expect("failed to define a metric");
         let metrics = std::array::from_fn(|i| {
             let op = StorageIoOperation::from_repr(i).unwrap();
-            let metric = storage_io_histogram_vec
+            storage_io_histogram_vec
                 .get_metric_with_label_values(&[op.as_str()])
-                .unwrap();
-            metric
+                .unwrap()
         });
         Self { metrics }
     }
@@ -967,6 +969,7 @@ pub(crate) struct DeletionQueueMetrics {
     pub(crate) keys_submitted: IntCounter,
     pub(crate) keys_dropped: IntCounter,
     pub(crate) keys_executed: IntCounter,
+    pub(crate) keys_validated: IntCounter,
     pub(crate) dropped_lsn_updates: IntCounter,
     pub(crate) unexpected_errors: IntCounter,
     pub(crate) remote_errors: IntCounterVec,
@@ -988,7 +991,13 @@ pub(crate) static DELETION_QUEUE: Lazy<DeletionQueueMetrics> = Lazy::new(|| {
 
     keys_executed: register_int_counter!(
         "pageserver_deletion_queue_executed_total",
-        "Number of objects deleted. Only includes objects that we actually deleted, sum with pageserver_deletion_queue_dropped_total for the total number of keys processed."
+        "Number of objects deleted. Only includes objects that we actually deleted, sum with pageserver_deletion_queue_dropped_total for the total number of keys processed to completion"
+    )
+    .expect("failed to define a metric"),
+
+    keys_validated: register_int_counter!(
+        "pageserver_deletion_queue_validated_total",
+        "Number of keys validated for deletion.  Sum with pageserver_deletion_queue_dropped_total for the total number of keys that have passed through the validation stage."
     )
     .expect("failed to define a metric"),
 
@@ -1061,6 +1070,26 @@ pub(crate) static TENANT_TASK_EVENTS: Lazy<IntCounterVec> = Lazy::new(|| {
     .expect("Failed to register tenant_task_events metric")
 });
 
+pub(crate) static BACKGROUND_LOOP_SEMAPHORE_WAIT_START_COUNT: Lazy<IntCounterVec> =
+    Lazy::new(|| {
+        register_int_counter_vec!(
+            "pageserver_background_loop_semaphore_wait_start_count",
+            "Counter for background loop concurrency-limiting semaphore acquire calls started",
+            &["task"],
+        )
+        .unwrap()
+    });
+
+pub(crate) static BACKGROUND_LOOP_SEMAPHORE_WAIT_FINISH_COUNT: Lazy<IntCounterVec> =
+    Lazy::new(|| {
+        register_int_counter_vec!(
+            "pageserver_background_loop_semaphore_wait_finish_count",
+            "Counter for background loop concurrency-limiting semaphore acquire calls finished",
+            &["task"],
+        )
+        .unwrap()
+    });
+
 pub(crate) static BACKGROUND_LOOP_PERIOD_OVERRUN_COUNT: Lazy<IntCounterVec> = Lazy::new(|| {
     register_int_counter_vec!(
         "pageserver_background_loop_period_overrun_count",
@@ -1362,27 +1391,22 @@ impl TimelineMetrics {
         }
     }
 
-    pub fn record_new_file_metrics(&self, sz: u64) {
+    pub(crate) fn record_new_file_metrics(&self, sz: u64) {
         self.resident_physical_size_add(sz);
         self.num_persistent_files_created.inc_by(1);
         self.persistent_bytes_written.inc_by(sz);
     }
 
-    pub fn resident_physical_size_sub(&self, sz: u64) {
+    pub(crate) fn resident_physical_size_sub(&self, sz: u64) {
         self.resident_physical_size_gauge.sub(sz);
         crate::metrics::RESIDENT_PHYSICAL_SIZE_GLOBAL.sub(sz);
     }
 
-    pub fn resident_physical_size_add(&self, sz: u64) {
+    pub(crate) fn resident_physical_size_add(&self, sz: u64) {
         self.resident_physical_size_gauge.add(sz);
         crate::metrics::RESIDENT_PHYSICAL_SIZE_GLOBAL.add(sz);
     }
 
-    pub fn resident_physical_size_set(&self, sz: u64) {
-        self.resident_physical_size_gauge.set(sz);
-        crate::metrics::RESIDENT_PHYSICAL_SIZE_GLOBAL.set(sz);
-    }
-
     pub fn resident_physical_size_get(&self) -> u64 {
         self.resident_physical_size_gauge.get()
     }

pageserver/src/page_cache.rs

@@ -66,8 +66,7 @@
 //! inserted to the mapping, but you must hold the write-lock on the slot until
 //! the contents are valid. If you need to release the lock without initializing
 //! the contents, you must remove the mapping first. We make that easy for the
-//! callers with PageWriteGuard: when lock_for_write() returns an uninitialized
-//! page, the caller must explicitly call guard.mark_valid() after it has
+//! callers with PageWriteGuard: the caller must explicitly call guard.mark_valid() after it has
 //! initialized it. If the guard is dropped without calling mark_valid(), the
 //! mapping is automatically removed and the slot is marked free.
 //!
@@ -286,23 +285,25 @@ impl AsRef<[u8; PAGE_SZ]> for PageReadGuard<'_> {
 ///
 /// Counterintuitively, this is used even for a read, if the requested page is not
 /// currently found in the page cache. In that case, the caller of lock_for_read()
-/// is expected to fill in the page contents and call mark_valid(). Similarly
-/// lock_for_write() can return an invalid buffer that the caller is expected to
-/// to initialize.
-///
+/// is expected to fill in the page contents and call mark_valid().
 pub struct PageWriteGuard<'i> {
-    inner: tokio::sync::RwLockWriteGuard<'i, SlotInner>,
+    state: PageWriteGuardState<'i>,
+}
 
-    _permit: PinnedSlotsPermit,
-
-    // Are the page contents currently valid?
-    // Used to mark pages as invalid that are assigned but not yet filled with data.
-    valid: bool,
+enum PageWriteGuardState<'i> {
+    Invalid {
+        inner: tokio::sync::RwLockWriteGuard<'i, SlotInner>,
+        _permit: PinnedSlotsPermit,
+    },
+    Downgraded,
 }
 
 impl std::ops::DerefMut for PageWriteGuard<'_> {
     fn deref_mut(&mut self) -> &mut Self::Target {
-        self.inner.buf
+        match &mut self.state {
+            PageWriteGuardState::Invalid { inner, _permit } => inner.buf,
+            PageWriteGuardState::Downgraded => unreachable!(),
+        }
     }
 }
 
@@ -310,25 +311,28 @@ impl std::ops::Deref for PageWriteGuard<'_> {
     type Target = [u8; PAGE_SZ];
 
     fn deref(&self) -> &Self::Target {
-        self.inner.buf
+        match &self.state {
+            PageWriteGuardState::Invalid { inner, _permit } => inner.buf,
+            PageWriteGuardState::Downgraded => unreachable!(),
+        }
     }
 }
 
-impl AsMut<[u8; PAGE_SZ]> for PageWriteGuard<'_> {
-    fn as_mut(&mut self) -> &mut [u8; PAGE_SZ] {
-        self.inner.buf
-    }
-}
-
-impl PageWriteGuard<'_> {
+impl<'a> PageWriteGuard<'a> {
     /// Mark that the buffer contents are now valid.
-    pub fn mark_valid(&mut self) {
-        assert!(self.inner.key.is_some());
-        assert!(
-            !self.valid,
-            "mark_valid called on a buffer that was already valid"
-        );
-        self.valid = true;
+    #[must_use]
+    pub fn mark_valid(mut self) -> PageReadGuard<'a> {
+        let prev = std::mem::replace(&mut self.state, PageWriteGuardState::Downgraded);
+        match prev {
+            PageWriteGuardState::Invalid { inner, _permit } => {
+                assert!(inner.key.is_some());
+                PageReadGuard {
+                    _permit: Arc::new(_permit),
+                    slot_guard: inner.downgrade(),
+                }
+            }
+            PageWriteGuardState::Downgraded => unreachable!(),
+        }
     }
 }
 
@@ -339,11 +343,14 @@ impl Drop for PageWriteGuard<'_> {
     /// initializing it, remove the mapping from the page cache.
     ///
     fn drop(&mut self) {
-        assert!(self.inner.key.is_some());
-        if !self.valid {
-            let self_key = self.inner.key.as_ref().unwrap();
-            PAGE_CACHE.get().unwrap().remove_mapping(self_key);
-            self.inner.key = None;
+        match &mut self.state {
+            PageWriteGuardState::Invalid { inner, _permit } => {
+                assert!(inner.key.is_some());
+                let self_key = inner.key.as_ref().unwrap();
+                PAGE_CACHE.get().unwrap().remove_mapping(self_key);
+                inner.key = None;
+            }
+            PageWriteGuardState::Downgraded => {}
         }
     }
 }
@@ -354,12 +361,6 @@ pub enum ReadBufResult<'a> {
     NotFound(PageWriteGuard<'a>),
 }
 
-/// lock_for_write() return value
-pub enum WriteBufResult<'a> {
-    Found(PageWriteGuard<'a>),
-    NotFound(PageWriteGuard<'a>),
-}
-
 impl PageCache {
     //
     // Section 1.1: Public interface functions for looking up and memorizing materialized page
@@ -446,20 +447,77 @@ impl PageCache {
             lsn,
         };
 
-        match self.lock_for_write(&cache_key).await? {
-            WriteBufResult::Found(write_guard) => {
-                // We already had it in cache. Another thread must've put it there
-                // concurrently. Check that it had the same contents that we
-                // replayed.
-                assert!(*write_guard == img);
+        let mut permit = Some(self.try_get_pinned_slot_permit().await?);
+        loop {
+            // First check if the key already exists in the cache.
+            if let Some(slot_idx) = self.search_mapping_exact(&cache_key) {
+                // The page was found in the mapping. Lock the slot, and re-check
+                // that it's still what we expected (because we don't released the mapping
+                // lock already, another thread could have evicted the page)
+                let slot = &self.slots[slot_idx];
+                let inner = slot.inner.write().await;
+                if inner.key.as_ref() == Some(&cache_key) {
+                    slot.inc_usage_count();
+                    debug_assert!(
+                        {
+                            let guard = inner.permit.lock().unwrap();
+                            guard.upgrade().is_none()
+                        },
+                        "we hold a write lock, so, no one else should have a permit"
+                    );
+                    debug_assert_eq!(inner.buf.len(), img.len());
+                    // We already had it in cache. Another thread must've put it there
+                    // concurrently. Check that it had the same contents that we
+                    // replayed.
+                    assert!(inner.buf == img);
+                    return Ok(());
+                }
             }
-            WriteBufResult::NotFound(mut write_guard) => {
-                write_guard.copy_from_slice(img);
-                write_guard.mark_valid();
-            }
-        }
+            debug_assert!(permit.is_some());
 
-        Ok(())
+            // Not found. Find a victim buffer
+            let (slot_idx, mut inner) = self
+                .find_victim(permit.as_ref().unwrap())
+                .await
+                .context("Failed to find evict victim")?;
+
+            // Insert mapping for this. At this point, we may find that another
+            // thread did the same thing concurrently. In that case, we evicted
+            // our victim buffer unnecessarily. Put it into the free list and
+            // continue with the slot that the other thread chose.
+            if let Some(_existing_slot_idx) = self.try_insert_mapping(&cache_key, slot_idx) {
+                // TODO: put to free list
+
+                // We now just loop back to start from beginning. This is not
+                // optimal, we'll perform the lookup in the mapping again, which
+                // is not really necessary because we already got
+                // 'existing_slot_idx'.  But this shouldn't happen often enough
+                // to matter much.
+                continue;
+            }
+
+            // Make the slot ready
+            let slot = &self.slots[slot_idx];
+            inner.key = Some(cache_key.clone());
+            slot.set_usage_count(1);
+            // Create a write guard for the slot so we go through the expected motions.
+            debug_assert!(
+                {
+                    let guard = inner.permit.lock().unwrap();
+                    guard.upgrade().is_none()
+                },
+                "we hold a write lock, so, no one else should have a permit"
+            );
+            let mut write_guard = PageWriteGuard {
+                state: PageWriteGuardState::Invalid {
+                    _permit: permit.take().unwrap(),
+                    inner,
+                },
+            };
+            write_guard.copy_from_slice(img);
+            let _ = write_guard.mark_valid();
+            return Ok(());
+        }
     }
 
     // Section 1.2: Public interface functions for working with immutable file pages.
@@ -638,99 +696,10 @@ impl PageCache {
             );
 
             return Ok(ReadBufResult::NotFound(PageWriteGuard {
-                _permit: permit.take().unwrap(),
-                inner,
-                valid: false,
-            }));
-        }
-    }
-
-    /// Look up a page in the cache and lock it in write mode. If it's not
-    /// found, returns None.
-    ///
-    /// When locking a page for writing, the search criteria is always "exact".
-    async fn try_lock_for_write(
-        &self,
-        cache_key: &CacheKey,
-        permit: &mut Option<PinnedSlotsPermit>,
-    ) -> Option<PageWriteGuard> {
-        if let Some(slot_idx) = self.search_mapping_for_write(cache_key) {
-            // The page was found in the mapping. Lock the slot, and re-check
-            // that it's still what we expected (because we don't released the mapping
-            // lock already, another thread could have evicted the page)
-            let slot = &self.slots[slot_idx];
-            let inner = slot.inner.write().await;
-            if inner.key.as_ref() == Some(cache_key) {
-                slot.inc_usage_count();
-                debug_assert!(
-                    {
-                        let guard = inner.permit.lock().unwrap();
-                        guard.upgrade().is_none()
-                    },
-                    "we hold a write lock, so, no one else should have a permit"
-                );
-                return Some(PageWriteGuard {
+                state: PageWriteGuardState::Invalid {
                     _permit: permit.take().unwrap(),
                     inner,
-                    valid: true,
-                });
-            }
-        }
-        None
-    }
-
-    /// Return a write-locked buffer for given block.
-    ///
-    /// Similar to lock_for_read(), but the returned buffer is write-locked and
-    /// may be modified by the caller even if it's already found in the cache.
-    async fn lock_for_write(&self, cache_key: &CacheKey) -> anyhow::Result<WriteBufResult> {
-        let mut permit = Some(self.try_get_pinned_slot_permit().await?);
-        loop {
-            // First check if the key already exists in the cache.
-            if let Some(write_guard) = self.try_lock_for_write(cache_key, &mut permit).await {
-                debug_assert!(permit.is_none());
-                return Ok(WriteBufResult::Found(write_guard));
-            }
-            debug_assert!(permit.is_some());
-
-            // Not found. Find a victim buffer
-            let (slot_idx, mut inner) = self
-                .find_victim(permit.as_ref().unwrap())
-                .await
-                .context("Failed to find evict victim")?;
-
-            // Insert mapping for this. At this point, we may find that another
-            // thread did the same thing concurrently. In that case, we evicted
-            // our victim buffer unnecessarily. Put it into the free list and
-            // continue with the slot that the other thread chose.
-            if let Some(_existing_slot_idx) = self.try_insert_mapping(cache_key, slot_idx) {
-                // TODO: put to free list
-
-                // We now just loop back to start from beginning. This is not
-                // optimal, we'll perform the lookup in the mapping again, which
-                // is not really necessary because we already got
-                // 'existing_slot_idx'.  But this shouldn't happen often enough
-                // to matter much.
-                continue;
-            }
-
-            // Make the slot ready
-            let slot = &self.slots[slot_idx];
-            inner.key = Some(cache_key.clone());
-            slot.set_usage_count(1);
-
-            debug_assert!(
-                {
-                    let guard = inner.permit.lock().unwrap();
-                    guard.upgrade().is_none()
                 },
-                "we hold a write lock, so, no one else should have a permit"
-            );
-
-            return Ok(WriteBufResult::NotFound(PageWriteGuard {
-                _permit: permit.take().unwrap(),
-                inner,
-                valid: false,
             }));
         }
     }
@@ -775,7 +744,7 @@ impl PageCache {
     ///
     /// Like 'search_mapping, but performs an "exact" search. Used for
     /// allocating a new buffer.
-    fn search_mapping_for_write(&self, key: &CacheKey) -> Option<usize> {
+    fn search_mapping_exact(&self, key: &CacheKey) -> Option<usize> {
         match key {
             CacheKey::MaterializedPage { hash_key, lsn } => {
                 let map = self.materialized_page_map.read().unwrap();

pageserver/src/page_service.rs

@@ -35,6 +35,7 @@ use std::time::Duration;
 use tokio::io::AsyncWriteExt;
 use tokio::io::{AsyncRead, AsyncWrite};
 use tokio_util::io::StreamReader;
+use tokio_util::sync::CancellationToken;
 use tracing::field;
 use tracing::*;
 use utils::id::ConnectionId;
@@ -64,69 +65,6 @@ use crate::trace::Tracer;
 use postgres_ffi::pg_constants::DEFAULTTABLESPACE_OID;
 use postgres_ffi::BLCKSZ;
 
-fn copyin_stream<IO>(pgb: &mut PostgresBackend<IO>) -> impl Stream<Item = io::Result<Bytes>> + '_
-where
-    IO: AsyncRead + AsyncWrite + Unpin,
-{
-    async_stream::try_stream! {
-        loop {
-            let msg = tokio::select! {
-                biased;
-
-                _ = task_mgr::shutdown_watcher() => {
-                    // We were requested to shut down.
-                    let msg = "pageserver is shutting down";
-                    let _ = pgb.write_message_noflush(&BeMessage::ErrorResponse(msg, None));
-                    Err(QueryError::Other(anyhow::anyhow!(msg)))
-                }
-
-                msg = pgb.read_message() => { msg.map_err(QueryError::from)}
-            };
-
-            match msg {
-                Ok(Some(message)) => {
-                    let copy_data_bytes = match message {
-                        FeMessage::CopyData(bytes) => bytes,
-                        FeMessage::CopyDone => { break },
-                        FeMessage::Sync => continue,
-                        FeMessage::Terminate => {
-                            let msg = "client terminated connection with Terminate message during COPY";
-                            let query_error = QueryError::Disconnected(ConnectionError::Io(io::Error::new(io::ErrorKind::ConnectionReset, msg)));
-                            // error can't happen here, ErrorResponse serialization should be always ok
-                            pgb.write_message_noflush(&BeMessage::ErrorResponse(msg, Some(query_error.pg_error_code()))).map_err(|e| e.into_io_error())?;
-                            Err(io::Error::new(io::ErrorKind::ConnectionReset, msg))?;
-                            break;
-                        }
-                        m => {
-                            let msg = format!("unexpected message {m:?}");
-                            // error can't happen here, ErrorResponse serialization should be always ok
-                            pgb.write_message_noflush(&BeMessage::ErrorResponse(&msg, None)).map_err(|e| e.into_io_error())?;
-                            Err(io::Error::new(io::ErrorKind::Other, msg))?;
-                            break;
-                        }
-                    };
-
-                    yield copy_data_bytes;
-                }
-                Ok(None) => {
-                    let msg = "client closed connection during COPY";
-                    let query_error = QueryError::Disconnected(ConnectionError::Io(io::Error::new(io::ErrorKind::ConnectionReset, msg)));
-                    // error can't happen here, ErrorResponse serialization should be always ok
-                    pgb.write_message_noflush(&BeMessage::ErrorResponse(msg, Some(query_error.pg_error_code()))).map_err(|e| e.into_io_error())?;
-                    pgb.flush().await?;
-                    Err(io::Error::new(io::ErrorKind::ConnectionReset, msg))?;
-                }
-                Err(QueryError::Disconnected(ConnectionError::Io(io_error))) => {
-                    Err(io_error)?;
-                }
-                Err(other) => {
-                    Err(io::Error::new(io::ErrorKind::Other, other.to_string()))?;
-                }
-            };
-        }
-    }
-}
-
 /// Read the end of a tar archive.
 ///
 /// A tar archive normally ends with two consecutive blocks of zeros, 512 bytes each.
@@ -184,6 +122,7 @@ pub async fn libpq_listener_main(
     listener: TcpListener,
     auth_type: AuthType,
     listener_ctx: RequestContext,
+    cancel: CancellationToken,
 ) -> anyhow::Result<()> {
     listener.set_nonblocking(true)?;
     let tokio_listener = tokio::net::TcpListener::from_std(listener)?;
@@ -192,7 +131,7 @@ pub async fn libpq_listener_main(
     while let Some(res) = tokio::select! {
         biased;
 
-        _ = task_mgr::shutdown_watcher() => {
+        _ = cancel.cancelled() => {
             // We were requested to shut down.
             None
         }
@@ -284,7 +223,13 @@ async fn page_service_conn_main(
     // and create a child per-query context when it invokes process_query.
     // But it's in a shared crate, so, we store connection_ctx inside PageServerHandler
     // and create the per-query context in process_query ourselves.
-    let mut conn_handler = PageServerHandler::new(conf, broker_client, auth, connection_ctx);
+    let mut conn_handler = PageServerHandler::new(
+        conf,
+        broker_client,
+        auth,
+        connection_ctx,
+        task_mgr::shutdown_token(),
+    );
     let pgbackend = PostgresBackend::new_from_io(socket, peer_addr, auth_type, None)?;
 
     match pgbackend
@@ -318,6 +263,10 @@ struct PageServerHandler {
     /// For each query received over the connection,
     /// `process_query` creates a child context from this one.
     connection_ctx: RequestContext,
+
+    /// A token that should fire when the tenant transitions from
+    /// attached state, or when the pageserver is shutting down.
+    cancel: CancellationToken,
 }
 
 impl PageServerHandler {
@@ -326,6 +275,7 @@ impl PageServerHandler {
         broker_client: storage_broker::BrokerClientChannel,
         auth: Option<Arc<JwtAuth>>,
         connection_ctx: RequestContext,
+        cancel: CancellationToken,
     ) -> Self {
         PageServerHandler {
             _conf: conf,
@@ -333,6 +283,91 @@ impl PageServerHandler {
             auth,
             claims: None,
             connection_ctx,
+            cancel,
+        }
+    }
+
+    /// Wrap PostgresBackend::flush to respect our CancellationToken: it is important to use
+    /// this rather than naked flush() in order to shut down promptly.  Without this, we would
+    /// block shutdown of a tenant if a postgres client was failing to consume bytes we send
+    /// in the flush.
+    async fn flush_cancellable<IO>(&self, pgb: &mut PostgresBackend<IO>) -> Result<(), QueryError>
+    where
+        IO: AsyncRead + AsyncWrite + Send + Sync + Unpin,
+    {
+        tokio::select!(
+            flush_r = pgb.flush() => {
+                Ok(flush_r?)
+            },
+            _ = self.cancel.cancelled() => {
+                Err(QueryError::Shutdown)
+            }
+        )
+    }
+
+    fn copyin_stream<'a, IO>(
+        &'a self,
+        pgb: &'a mut PostgresBackend<IO>,
+    ) -> impl Stream<Item = io::Result<Bytes>> + 'a
+    where
+        IO: AsyncRead + AsyncWrite + Send + Sync + Unpin,
+    {
+        async_stream::try_stream! {
+            loop {
+                let msg = tokio::select! {
+                    biased;
+
+                    _ = self.cancel.cancelled() => {
+                        // We were requested to shut down.
+                        let msg = "pageserver is shutting down";
+                        let _ = pgb.write_message_noflush(&BeMessage::ErrorResponse(msg, None));
+                        Err(QueryError::Shutdown)
+                    }
+
+                    msg = pgb.read_message() => { msg.map_err(QueryError::from)}
+                };
+
+                match msg {
+                    Ok(Some(message)) => {
+                        let copy_data_bytes = match message {
+                            FeMessage::CopyData(bytes) => bytes,
+                            FeMessage::CopyDone => { break },
+                            FeMessage::Sync => continue,
+                            FeMessage::Terminate => {
+                                let msg = "client terminated connection with Terminate message during COPY";
+                                let query_error = QueryError::Disconnected(ConnectionError::Io(io::Error::new(io::ErrorKind::ConnectionReset, msg)));
+                                // error can't happen here, ErrorResponse serialization should be always ok
+                                pgb.write_message_noflush(&BeMessage::ErrorResponse(msg, Some(query_error.pg_error_code()))).map_err(|e| e.into_io_error())?;
+                                Err(io::Error::new(io::ErrorKind::ConnectionReset, msg))?;
+                                break;
+                            }
+                            m => {
+                                let msg = format!("unexpected message {m:?}");
+                                // error can't happen here, ErrorResponse serialization should be always ok
+                                pgb.write_message_noflush(&BeMessage::ErrorResponse(&msg, None)).map_err(|e| e.into_io_error())?;
+                                Err(io::Error::new(io::ErrorKind::Other, msg))?;
+                                break;
+                            }
+                        };
+
+                        yield copy_data_bytes;
+                    }
+                    Ok(None) => {
+                        let msg = "client closed connection during COPY";
+                        let query_error = QueryError::Disconnected(ConnectionError::Io(io::Error::new(io::ErrorKind::ConnectionReset, msg)));
+                        // error can't happen here, ErrorResponse serialization should be always ok
+                        pgb.write_message_noflush(&BeMessage::ErrorResponse(msg, Some(query_error.pg_error_code()))).map_err(|e| e.into_io_error())?;
+                        self.flush_cancellable(pgb).await.map_err(|e| io::Error::new(io::ErrorKind::Other, e.to_string()))?;
+                        Err(io::Error::new(io::ErrorKind::ConnectionReset, msg))?;
+                    }
+                    Err(QueryError::Disconnected(ConnectionError::Io(io_error))) => {
+                        Err(io_error)?;
+                    }
+                    Err(other) => {
+                        Err(io::Error::new(io::ErrorKind::Other, other.to_string()))?;
+                    }
+                };
+            }
         }
     }
 
@@ -372,7 +407,7 @@ impl PageServerHandler {
 
         // switch client to COPYBOTH
         pgb.write_message_noflush(&BeMessage::CopyBothResponse)?;
-        pgb.flush().await?;
+        self.flush_cancellable(pgb).await?;
 
         let metrics = metrics::SmgrQueryTimePerTimeline::new(&tenant_id, &timeline_id);
 
@@ -380,10 +415,10 @@ impl PageServerHandler {
             let msg = tokio::select! {
                 biased;
 
-                _ = task_mgr::shutdown_watcher() => {
+                _ = self.cancel.cancelled() => {
                     // We were requested to shut down.
                     info!("shutdown request received in page handler");
-                    break;
+                    return Err(QueryError::Shutdown)
                 }
 
                 msg = pgb.read_message() => { msg }
@@ -465,7 +500,7 @@ impl PageServerHandler {
             });
 
             pgb.write_message_noflush(&BeMessage::CopyData(&response.serialize()))?;
-            pgb.flush().await?;
+            self.flush_cancellable(pgb).await?;
         }
         Ok(())
     }
@@ -508,9 +543,9 @@ impl PageServerHandler {
         // Import basebackup provided via CopyData
         info!("importing basebackup");
         pgb.write_message_noflush(&BeMessage::CopyInResponse)?;
-        pgb.flush().await?;
+        self.flush_cancellable(pgb).await?;
 
-        let mut copyin_reader = pin!(StreamReader::new(copyin_stream(pgb)));
+        let mut copyin_reader = pin!(StreamReader::new(self.copyin_stream(pgb)));
         timeline
             .import_basebackup_from_tar(
                 &mut copyin_reader,
@@ -563,8 +598,8 @@ impl PageServerHandler {
         // Import wal provided via CopyData
         info!("importing wal");
         pgb.write_message_noflush(&BeMessage::CopyInResponse)?;
-        pgb.flush().await?;
-        let mut copyin_reader = pin!(StreamReader::new(copyin_stream(pgb)));
+        self.flush_cancellable(pgb).await?;
+        let mut copyin_reader = pin!(StreamReader::new(self.copyin_stream(pgb)));
         import_wal_from_tar(&timeline, &mut copyin_reader, start_lsn, end_lsn, &ctx).await?;
         info!("wal import complete");
 
@@ -772,7 +807,7 @@ impl PageServerHandler {
 
         // switch client to COPYOUT
         pgb.write_message_noflush(&BeMessage::CopyOutResponse)?;
-        pgb.flush().await?;
+        self.flush_cancellable(pgb).await?;
 
         // Send a tarball of the latest layer on the timeline. Compress if not
         // fullbackup. TODO Compress in that case too (tests need to be updated)
@@ -824,7 +859,7 @@ impl PageServerHandler {
         }
 
         pgb.write_message_noflush(&BeMessage::CopyDone)?;
-        pgb.flush().await?;
+        self.flush_cancellable(pgb).await?;
 
         let basebackup_after = started
             .elapsed()

pageserver/src/pgdatadir_mapping.rs

@@ -19,6 +19,7 @@ use postgres_ffi::BLCKSZ;
 use postgres_ffi::{Oid, TimestampTz, TransactionId};
 use serde::{Deserialize, Serialize};
 use std::collections::{hash_map, HashMap, HashSet};
+use std::ops::ControlFlow;
 use std::ops::Range;
 use tokio_util::sync::CancellationToken;
 use tracing::{debug, trace, warn};
@@ -370,7 +371,6 @@ impl Timeline {
         }
     }
 
-    ///
     /// Subroutine of find_lsn_for_timestamp(). Returns true, if there are any
     /// commits that committed after 'search_timestamp', at LSN 'probe_lsn'.
     ///
@@ -385,6 +385,50 @@ impl Timeline {
         found_larger: &mut bool,
         ctx: &RequestContext,
     ) -> Result<bool, PageReconstructError> {
+        self.map_all_timestamps(probe_lsn, ctx, |timestamp| {
+            if timestamp >= search_timestamp {
+                *found_larger = true;
+                return ControlFlow::Break(true);
+            } else {
+                *found_smaller = true;
+            }
+            ControlFlow::Continue(())
+        })
+        .await
+    }
+
+    /// Obtain the possible timestamp range for the given lsn.
+    ///
+    /// If the lsn has no timestamps, returns None. returns `(min, max, median)` if it has timestamps.
+    pub async fn get_timestamp_for_lsn(
+        &self,
+        probe_lsn: Lsn,
+        ctx: &RequestContext,
+    ) -> Result<Option<TimestampTz>, PageReconstructError> {
+        let mut max: Option<TimestampTz> = None;
+        self.map_all_timestamps(probe_lsn, ctx, |timestamp| {
+            if let Some(max_prev) = max {
+                max = Some(max_prev.max(timestamp));
+            } else {
+                max = Some(timestamp);
+            }
+            ControlFlow::Continue(())
+        })
+        .await?;
+
+        Ok(max)
+    }
+
+    /// Runs the given function on all the timestamps for a given lsn
+    ///
+    /// The return value is either given by the closure, or set to the `Default`
+    /// impl's output.
+    async fn map_all_timestamps<T: Default>(
+        &self,
+        probe_lsn: Lsn,
+        ctx: &RequestContext,
+        mut f: impl FnMut(TimestampTz) -> ControlFlow<T>,
+    ) -> Result<T, PageReconstructError> {
         for segno in self
             .list_slru_segments(SlruKind::Clog, probe_lsn, ctx)
             .await?
@@ -402,16 +446,14 @@ impl Timeline {
                     timestamp_bytes.copy_from_slice(&clog_page[BLCKSZ as usize..]);
                     let timestamp = TimestampTz::from_be_bytes(timestamp_bytes);
 
-                    if timestamp >= search_timestamp {
-                        *found_larger = true;
-                        return Ok(true);
-                    } else {
-                        *found_smaller = true;
+                    match f(timestamp) {
+                        ControlFlow::Break(b) => return Ok(b),
+                        ControlFlow::Continue(()) => (),
                     }
                 }
             }
         }
-        Ok(false)
+        Ok(Default::default())
     }
 
     /// Get a list of SLRU segments
@@ -499,6 +541,23 @@ impl Timeline {
         self.get(CHECKPOINT_KEY, lsn, ctx).await
     }
 
+    pub async fn list_aux_files(
+        &self,
+        lsn: Lsn,
+        ctx: &RequestContext,
+    ) -> Result<HashMap<String, Bytes>, PageReconstructError> {
+        match self.get(AUX_FILES_KEY, lsn, ctx).await {
+            Ok(buf) => match AuxFilesDirectory::des(&buf).context("deserialization failure") {
+                Ok(dir) => Ok(dir.files),
+                Err(e) => Err(PageReconstructError::from(e)),
+            },
+            Err(e) => {
+                warn!("Failed to get info about AUX files: {}", e);
+                Ok(HashMap::new())
+            }
+        }
+    }
+
     /// Does the same as get_current_logical_size but counted on demand.
     /// Used to initialize the logical size tracking on startup.
     ///
@@ -616,7 +675,9 @@ impl Timeline {
 
         result.add_key(CONTROLFILE_KEY);
         result.add_key(CHECKPOINT_KEY);
-
+        if self.get(AUX_FILES_KEY, lsn, ctx).await.is_ok() {
+            result.add_key(AUX_FILES_KEY);
+        }
         Ok(result.to_keyspace())
     }
 
@@ -692,6 +753,12 @@ impl<'a> DatadirModification<'a> {
         })?;
         self.put(DBDIR_KEY, Value::Image(buf.into()));
 
+        // Create AuxFilesDirectory
+        let buf = AuxFilesDirectory::ser(&AuxFilesDirectory {
+            files: HashMap::new(),
+        })?;
+        self.put(AUX_FILES_KEY, Value::Image(Bytes::from(buf)));
+
         let buf = TwoPhaseDirectory::ser(&TwoPhaseDirectory {
             xids: HashSet::new(),
         })?;
@@ -796,6 +863,12 @@ impl<'a> DatadirModification<'a> {
             // 'true', now write the updated 'dbdirs' map back.
             let buf = DbDirectory::ser(&dbdir)?;
             self.put(DBDIR_KEY, Value::Image(buf.into()));
+
+            // Create AuxFilesDirectory as well
+            let buf = AuxFilesDirectory::ser(&AuxFilesDirectory {
+                files: HashMap::new(),
+            })?;
+            self.put(AUX_FILES_KEY, Value::Image(Bytes::from(buf)));
         }
         if r.is_none() {
             // Create RelDirectory
@@ -1120,6 +1193,36 @@ impl<'a> DatadirModification<'a> {
         Ok(())
     }
 
+    pub async fn put_file(
+        &mut self,
+        path: &str,
+        content: &[u8],
+        ctx: &RequestContext,
+    ) -> anyhow::Result<()> {
+        let mut dir = match self.get(AUX_FILES_KEY, ctx).await {
+            Ok(buf) => AuxFilesDirectory::des(&buf)?,
+            Err(e) => {
+                warn!("Failed to get info about AUX files: {}", e);
+                AuxFilesDirectory {
+                    files: HashMap::new(),
+                }
+            }
+        };
+        let path = path.to_string();
+        if content.is_empty() {
+            dir.files.remove(&path);
+        } else {
+            dir.files.insert(path, Bytes::copy_from_slice(content));
+        }
+        self.put(
+            AUX_FILES_KEY,
+            Value::Image(Bytes::from(
+                AuxFilesDirectory::ser(&dir).context("serialize")?,
+            )),
+        );
+        Ok(())
+    }
+
     ///
     /// Flush changes accumulated so far to the underlying repository.
     ///
@@ -1255,6 +1358,11 @@ struct RelDirectory {
     rels: HashSet<(Oid, u8)>,
 }
 
+#[derive(Debug, Serialize, Deserialize, Default)]
+struct AuxFilesDirectory {
+    files: HashMap<String, Bytes>,
+}
+
 #[derive(Debug, Serialize, Deserialize)]
 struct RelSizeEntry {
     nblocks: u32,
@@ -1303,10 +1411,12 @@ static ZERO_PAGE: Bytes = Bytes::from_static(&[0u8; BLCKSZ as usize]);
 // 02 pg_twophase
 //
 // 03 misc
-//    controlfile
+//    Controlfile
 //    checkpoint
 //    pg_version
 //
+// 04 aux files
+//
 // Below is a full list of the keyspace allocation:
 //
 // DbDir:
@@ -1344,6 +1454,11 @@ static ZERO_PAGE: Bytes = Bytes::from_static(&[0u8; BLCKSZ as usize]);
 //
 // Checkpoint:
 // 03 00000000 00000000 00000000 00   00000001
+//
+// AuxFiles:
+// 03 00000000 00000000 00000000 00   00000002
+//
+
 //-- Section 01: relation data and metadata
 
 const DBDIR_KEY: Key = Key {
@@ -1567,9 +1682,19 @@ const CHECKPOINT_KEY: Key = Key {
     field6: 1,
 };
 
+const AUX_FILES_KEY: Key = Key {
+    field1: 0x03,
+    field2: 0,
+    field3: 0,
+    field4: 0,
+    field5: 0,
+    field6: 2,
+};
+
 // Reverse mappings for a few Keys.
 // These are needed by WAL redo manager.
 
+/// Guaranteed to return `Ok()` if [[is_rel_block_key]] returns `true` for `key`.
 pub fn key_to_rel_block(key: Key) -> anyhow::Result<(RelTag, BlockNumber)> {
     Ok(match key.field1 {
         0x00 => (
@@ -1585,7 +1710,8 @@ pub fn key_to_rel_block(key: Key) -> anyhow::Result<(RelTag, BlockNumber)> {
     })
 }
 
-fn is_rel_block_key(key: Key) -> bool {
+/// See [[key_to_rel_block]].
+pub fn is_rel_block_key(key: Key) -> bool {
     key.field1 == 0x00 && key.field4 != 0
 }
 

pageserver/src/tenant/blob_io.rs

@@ -234,10 +234,7 @@ impl BlobWriter<false> {
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::{
-        context::DownloadBehavior, task_mgr::TaskKind, tenant::block_io::BlockReaderRef,
-        virtual_file::Error,
-    };
+    use crate::{context::DownloadBehavior, task_mgr::TaskKind, tenant::block_io::BlockReaderRef};
     use rand::{Rng, SeedableRng};
 
     async fn round_trip_test<const BUFFERED: bool>(blobs: &[Vec<u8>]) -> Result<(), Error> {

pageserver/src/tenant/block_io.rs

@@ -6,7 +6,7 @@ use super::ephemeral_file::EphemeralFile;
 use super::storage_layer::delta_layer::{Adapter, DeltaLayerInner};
 use crate::context::RequestContext;
 use crate::page_cache::{self, PageReadGuard, ReadBufResult, PAGE_SZ};
-use crate::virtual_file::{self, VirtualFile};
+use crate::virtual_file::VirtualFile;
 use bytes::Bytes;
 use std::ops::{Deref, DerefMut};
 
@@ -96,7 +96,7 @@ impl<'a> BlockReaderRef<'a> {
             #[cfg(test)]
             TestDisk(r) => r.read_blk(blknum),
             #[cfg(test)]
-            VirtualFile(r) => r.read_blk(blknum).await.map_err(virtual_file::Error::into),
+            VirtualFile(r) => r.read_blk(blknum).await,
         }
     }
 }
@@ -174,7 +174,6 @@ impl FileBlockReader {
         self.file
             .read_exact_at(buf, blkno as u64 * PAGE_SZ as u64)
             .await
-            .map_err(virtual_file::Error::into)
     }
     /// Read a block.
     ///
@@ -187,26 +186,21 @@ impl FileBlockReader {
         ctx: &RequestContext,
     ) -> Result<BlockLease, std::io::Error> {
         let cache = page_cache::get();
-        loop {
-            match cache
-                .read_immutable_buf(self.file_id, blknum, ctx)
-                .await
-                .map_err(|e| {
-                    std::io::Error::new(
-                        std::io::ErrorKind::Other,
-                        format!("Failed to read immutable buf: {e:#}"),
-                    )
-                })? {
-                ReadBufResult::Found(guard) => break Ok(guard.into()),
-                ReadBufResult::NotFound(mut write_guard) => {
-                    // Read the page from disk into the buffer
-                    self.fill_buffer(write_guard.deref_mut(), blknum).await?;
-                    write_guard.mark_valid();
-
-                    // Swap for read lock
-                    continue;
-                }
-            };
+        match cache
+            .read_immutable_buf(self.file_id, blknum, ctx)
+            .await
+            .map_err(|e| {
+                std::io::Error::new(
+                    std::io::ErrorKind::Other,
+                    format!("Failed to read immutable buf: {e:#}"),
+                )
+            })? {
+            ReadBufResult::Found(guard) => Ok(guard.into()),
+            ReadBufResult::NotFound(mut write_guard) => {
+                // Read the page from disk into the buffer
+                self.fill_buffer(write_guard.deref_mut(), blknum).await?;
+                Ok(write_guard.mark_valid().into())
+            }
         }
     }
 }

pageserver/src/tenant/config.rs

@@ -13,6 +13,7 @@ use pageserver_api::models;
 use serde::{Deserialize, Serialize};
 use std::num::NonZeroU64;
 use std::time::Duration;
+use utils::generation::Generation;
 
 pub mod defaults {
     // FIXME: This current value is very low. I would imagine something like 1 GB or 10 GB
@@ -44,7 +45,211 @@ pub mod defaults {
     pub const DEFAULT_EVICTIONS_LOW_RESIDENCE_DURATION_METRIC_THRESHOLD: &str = "24 hour";
 }
 
-/// Per-tenant configuration options
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
+pub(crate) enum AttachmentMode {
+    /// Our generation is current as far as we know, and as far as we know we are the only attached
+    /// pageserver.  This is the "normal" attachment mode.
+    Single,
+    /// Our generation number is current as far as we know, but we are advised that another
+    /// pageserver is still attached, and therefore to avoid executing deletions.   This is
+    /// the attachment mode of a pagesever that is the destination of a migration.
+    Multi,
+    /// Our generation number is superseded, or about to be superseded.  We are advised
+    /// to avoid remote storage writes if possible, and to avoid sending billing data.  This
+    /// is the attachment mode of a pageserver that is the origin of a migration.
+    Stale,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
+pub(crate) struct AttachedLocationConfig {
+    pub(crate) generation: Generation,
+    pub(crate) attach_mode: AttachmentMode,
+    // TODO: add a flag to override AttachmentMode's policies under
+    // disk pressure (i.e. unblock uploads under disk pressure in Stale
+    // state, unblock deletions after timeout in Multi state)
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
+pub(crate) struct SecondaryLocationConfig {
+    /// If true, keep the local cache warm by polling remote storage
+    pub(crate) warm: bool,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
+pub(crate) enum LocationMode {
+    Attached(AttachedLocationConfig),
+    Secondary(SecondaryLocationConfig),
+}
+
+/// Per-tenant, per-pageserver configuration.  All pageservers use the same TenantConf,
+/// but have distinct LocationConf.
+#[derive(Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub(crate) struct LocationConf {
+    /// The location-specific part of the configuration, describes the operating
+    /// mode of this pageserver for this tenant.
+    pub(crate) mode: LocationMode,
+    /// The pan-cluster tenant configuration, the same on all locations
+    pub(crate) tenant_conf: TenantConfOpt,
+}
+
+impl std::fmt::Debug for LocationConf {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match &self.mode {
+            LocationMode::Attached(conf) => {
+                write!(
+                    f,
+                    "Attached {:?}, gen={:?}",
+                    conf.attach_mode, conf.generation
+                )
+            }
+            LocationMode::Secondary(conf) => {
+                write!(f, "Secondary, warm={}", conf.warm)
+            }
+        }
+    }
+}
+
+impl AttachedLocationConfig {
+    /// Consult attachment mode to determine whether we are currently permitted
+    /// to delete layers.  This is only advisory, not required for data safety.
+    /// See [`AttachmentMode`] for more context.
+    pub(crate) fn may_delete_layers_hint(&self) -> bool {
+        // TODO: add an override for disk pressure in AttachedLocationConfig,
+        // and respect it here.
+        match &self.attach_mode {
+            AttachmentMode::Single => true,
+            AttachmentMode::Multi | AttachmentMode::Stale => {
+                // In Multi mode we avoid doing deletions because some other
+                // attached pageserver might get 404 while trying to read
+                // a layer we delete which is still referenced in their metadata.
+                //
+                // In Stale mode, we avoid doing deletions because we expect
+                // that they would ultimately fail validation in the deletion
+                // queue due to our stale generation.
+                false
+            }
+        }
+    }
+
+    /// Whether we are currently hinted that it is worthwhile to upload layers.
+    /// This is only advisory, not required for data safety.
+    /// See [`AttachmentMode`] for more context.
+    pub(crate) fn may_upload_layers_hint(&self) -> bool {
+        // TODO: add an override for disk pressure in AttachedLocationConfig,
+        // and respect it here.
+        match &self.attach_mode {
+            AttachmentMode::Single | AttachmentMode::Multi => true,
+            AttachmentMode::Stale => {
+                // In Stale mode, we avoid doing uploads because we expect that
+                // our replacement pageserver will already have started its own
+                // IndexPart that will never reference layers we upload: it is
+                // wasteful.
+                false
+            }
+        }
+    }
+}
+
+impl LocationConf {
+    /// For use when loading from a legacy configuration: presence of a tenant
+    /// implies it is in AttachmentMode::Single, which used to be the only
+    /// possible state.  This function should eventually be removed.
+    pub(crate) fn attached_single(tenant_conf: TenantConfOpt, generation: Generation) -> Self {
+        Self {
+            mode: LocationMode::Attached(AttachedLocationConfig {
+                generation,
+                attach_mode: AttachmentMode::Single,
+            }),
+            tenant_conf,
+        }
+    }
+
+    /// For use when attaching/re-attaching: update the generation stored in this
+    /// structure.  If we were in a secondary state, promote to attached (posession
+    /// of a fresh generation implies this).
+    pub(crate) fn attach_in_generation(&mut self, generation: Generation) {
+        match &mut self.mode {
+            LocationMode::Attached(attach_conf) => {
+                attach_conf.generation = generation;
+            }
+            LocationMode::Secondary(_) => {
+                // We are promoted to attached by the control plane's re-attach response
+                self.mode = LocationMode::Attached(AttachedLocationConfig {
+                    generation,
+                    attach_mode: AttachmentMode::Single,
+                })
+            }
+        }
+    }
+
+    pub(crate) fn try_from(conf: &'_ models::LocationConfig) -> anyhow::Result<Self> {
+        let tenant_conf = TenantConfOpt::try_from(&conf.tenant_conf)?;
+
+        fn get_generation(conf: &'_ models::LocationConfig) -> Result<Generation, anyhow::Error> {
+            conf.generation
+                .ok_or_else(|| anyhow::anyhow!("Generation must be set when attaching"))
+        }
+
+        let mode = match &conf.mode {
+            models::LocationConfigMode::AttachedMulti => {
+                LocationMode::Attached(AttachedLocationConfig {
+                    generation: get_generation(conf)?,
+                    attach_mode: AttachmentMode::Multi,
+                })
+            }
+            models::LocationConfigMode::AttachedSingle => {
+                LocationMode::Attached(AttachedLocationConfig {
+                    generation: get_generation(conf)?,
+                    attach_mode: AttachmentMode::Single,
+                })
+            }
+            models::LocationConfigMode::AttachedStale => {
+                LocationMode::Attached(AttachedLocationConfig {
+                    generation: get_generation(conf)?,
+                    attach_mode: AttachmentMode::Stale,
+                })
+            }
+            models::LocationConfigMode::Secondary => {
+                anyhow::ensure!(conf.generation.is_none());
+
+                let warm = conf
+                    .secondary_conf
+                    .as_ref()
+                    .map(|c| c.warm)
+                    .unwrap_or(false);
+                LocationMode::Secondary(SecondaryLocationConfig { warm })
+            }
+            models::LocationConfigMode::Detached => {
+                // Should not have been called: API code should translate this mode
+                // into a detach rather than trying to decode it as a LocationConf
+                return Err(anyhow::anyhow!("Cannot decode a Detached configuration"));
+            }
+        };
+
+        Ok(Self { mode, tenant_conf })
+    }
+}
+
+impl Default for LocationConf {
+    // TODO: this should be removed once tenant loading can guarantee that we are never
+    // loading from a directory without a configuration.
+    // => tech debt since https://github.com/neondatabase/neon/issues/1555
+    fn default() -> Self {
+        Self {
+            mode: LocationMode::Attached(AttachedLocationConfig {
+                generation: Generation::none(),
+                attach_mode: AttachmentMode::Single,
+            }),
+            tenant_conf: TenantConfOpt::default(),
+        }
+    }
+}
+
+/// A tenant's calcuated configuration, which is the result of merging a
+/// tenant's TenantConfOpt with the global TenantConf from PageServerConf.
+///
+/// For storing and transmitting individual tenant's configuration, see
+/// TenantConfOpt.
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
 pub struct TenantConf {
     // Flush out an inmemory layer, if it's holding WAL older than this

pageserver/src/tenant/delete.rs

@@ -31,7 +31,7 @@ use super::{
 const SHOULD_RESUME_DELETION_FETCH_MARK_ATTEMPTS: u32 = 3;
 
 #[derive(Debug, thiserror::Error)]
-pub enum DeleteTenantError {
+pub(crate) enum DeleteTenantError {
     #[error("GetTenant {0}")]
     Get(#[from] GetTenantError),
 
@@ -197,6 +197,7 @@ async fn cleanup_remaining_fs_traces(
     };
 
     rm(conf.tenant_config_path(tenant_id), false).await?;
+    rm(conf.tenant_location_config_path(tenant_id), false).await?;
 
     fail::fail_point!("tenant-delete-before-remove-timelines-dir", |_| {
         Err(anyhow::anyhow!(
@@ -375,7 +376,7 @@ impl DeleteTenantFlow {
         Ok(())
     }
 
-    pub async fn should_resume_deletion(
+    pub(crate) async fn should_resume_deletion(
         conf: &'static PageServerConf,
         remote_storage: Option<&GenericRemoteStorage>,
         tenant: &Tenant,
@@ -431,7 +432,7 @@ impl DeleteTenantFlow {
         // Tenant may not be loadable if we fail late in cleanup_remaining_fs_traces (e g remove timelines dir)
         let timelines_path = tenant.conf.timelines_path(&tenant.tenant_id);
         if timelines_path.exists() {
-            tenant.load(init_order, ctx).await.context("load")?;
+            tenant.load(init_order, None, ctx).await.context("load")?;
         }
 
         Self::background(
@@ -457,7 +458,10 @@ impl DeleteTenantFlow {
             .await
             .expect("cant be stopping or broken");
 
-        tenant.attach(ctx).await.context("attach")?;
+        tenant
+            .attach(ctx, super::AttachMarkerMode::Expect)
+            .await
+            .context("attach")?;
 
         Self::background(
             guard,

pageserver/src/tenant/ephemeral_file.rs

@@ -72,36 +72,32 @@ impl EphemeralFile {
         let flushed_blknums = 0..self.len / PAGE_SZ as u64;
         if flushed_blknums.contains(&(blknum as u64)) {
             let cache = page_cache::get();
-            loop {
-                match cache
-                    .read_immutable_buf(self.page_cache_file_id, blknum, ctx)
-                    .await
-                    .map_err(|e| {
-                        std::io::Error::new(
-                            std::io::ErrorKind::Other,
-                            // order path before error because error is anyhow::Error => might have many contexts
-                            format!(
-                                "ephemeral file: read immutable page #{}: {}: {:#}",
-                                blknum, self.file.path, e,
-                            ),
-                        )
-                    })? {
-                    page_cache::ReadBufResult::Found(guard) => {
-                        return Ok(BlockLease::PageReadGuard(guard))
-                    }
-                    page_cache::ReadBufResult::NotFound(mut write_guard) => {
-                        let buf: &mut [u8] = write_guard.deref_mut();
-                        debug_assert_eq!(buf.len(), PAGE_SZ);
-                        self.file
-                            .read_exact_at(&mut buf[..], blknum as u64 * PAGE_SZ as u64)
-                            .await?;
-                        write_guard.mark_valid();
-
-                        // Swap for read lock
-                        continue;
-                    }
-                };
-            }
+            match cache
+                .read_immutable_buf(self.page_cache_file_id, blknum, ctx)
+                .await
+                .map_err(|e| {
+                    std::io::Error::new(
+                        std::io::ErrorKind::Other,
+                        // order path before error because error is anyhow::Error => might have many contexts
+                        format!(
+                            "ephemeral file: read immutable page #{}: {}: {:#}",
+                            blknum, self.file.path, e,
+                        ),
+                    )
+                })? {
+                page_cache::ReadBufResult::Found(guard) => {
+                    return Ok(BlockLease::PageReadGuard(guard))
+                }
+                page_cache::ReadBufResult::NotFound(mut write_guard) => {
+                    let buf: &mut [u8] = write_guard.deref_mut();
+                    debug_assert_eq!(buf.len(), PAGE_SZ);
+                    self.file
+                        .read_exact_at(&mut buf[..], blknum as u64 * PAGE_SZ as u64)
+                        .await?;
+                    let read_guard = write_guard.mark_valid();
+                    return Ok(BlockLease::PageReadGuard(read_guard));
+                }
+            };
         } else {
             debug_assert_eq!(blknum as u64, self.len / PAGE_SZ as u64);
             Ok(BlockLease::EphemeralFileMutableTail(&self.mutable_tail))
@@ -171,7 +167,7 @@ impl EphemeralFile {
                                         let buf: &mut [u8] = write_guard.deref_mut();
                                         debug_assert_eq!(buf.len(), PAGE_SZ);
                                         buf.copy_from_slice(&self.ephemeral_file.mutable_tail);
-                                        write_guard.mark_valid();
+                                        let _ = write_guard.mark_valid();
                                         // pre-warm successful
                                     }
                                     Err(e) => {
@@ -358,8 +354,7 @@ mod tests {
         }
 
         // Test a large blob that spans multiple pages
-        let mut large_data = Vec::new();
-        large_data.resize(20000, 0);
+        let mut large_data = vec![0; 20000];
         thread_rng().fill_bytes(&mut large_data);
         let pos_large = file.write_blob(&large_data, &ctx).await?;
         let result = file.block_cursor().read_blob(pos_large, &ctx).await?;

pageserver/src/tenant/remote_timeline_client.rs

@@ -627,7 +627,7 @@ impl RemoteTimelineClient {
     ///
     /// Launch an upload operation in the background.
     ///
-    pub fn schedule_layer_file_upload(
+    pub(crate) fn schedule_layer_file_upload(
         self: &Arc<Self>,
         layer_file_name: &LayerFileName,
         layer_metadata: &LayerFileMetadata,
@@ -635,6 +635,17 @@ impl RemoteTimelineClient {
         let mut guard = self.upload_queue.lock().unwrap();
         let upload_queue = guard.initialized_mut()?;
 
+        self.schedule_layer_file_upload0(upload_queue, layer_file_name, layer_metadata);
+        self.launch_queued_tasks(upload_queue);
+        Ok(())
+    }
+
+    fn schedule_layer_file_upload0(
+        self: &Arc<Self>,
+        upload_queue: &mut UploadQueueInitialized,
+        layer_file_name: &LayerFileName,
+        layer_metadata: &LayerFileMetadata,
+    ) {
         upload_queue
             .latest_files
             .insert(layer_file_name.clone(), layer_metadata.clone());
@@ -643,21 +654,15 @@ impl RemoteTimelineClient {
         let op = UploadOp::UploadLayer(layer_file_name.clone(), layer_metadata.clone());
         self.calls_unfinished_metric_begin(&op);
         upload_queue.queued_operations.push_back(op);
-
         info!("scheduled layer file upload {layer_file_name}");
-
-        // Launch the task immediately, if possible
-        self.launch_queued_tasks(upload_queue);
-        Ok(())
     }
 
     /// Launch a delete operation in the background.
     ///
-    /// The operation does not modify local state but assumes the local files have already been
-    /// deleted, and is used to mirror those changes to remote.
+    /// The operation does not modify local filesystem state.
     ///
     /// Note: This schedules an index file upload before the deletions.  The
-    /// deletion won't actually be performed, until any previously scheduled
+    /// deletion won't actually be performed, until all previously scheduled
     /// upload operations, and the index file upload, have completed
     /// successfully.
     pub fn schedule_layer_file_deletion(
@@ -667,61 +672,133 @@ impl RemoteTimelineClient {
         let mut guard = self.upload_queue.lock().unwrap();
         let upload_queue = guard.initialized_mut()?;
 
+        let with_generations =
+            self.schedule_unlinking_of_layers_from_index_part0(upload_queue, &names);
+
+        self.schedule_deletion_of_unlinked0(upload_queue, with_generations);
+
+        // Launch the tasks immediately, if possible
+        self.launch_queued_tasks(upload_queue);
+        Ok(())
+    }
+
+    /// Unlinks the layer files from `index_part.json` but does not yet schedule deletion for the
+    /// layer files, leaving them dangling.
+    ///
+    /// The files will be leaked in remote storage unless [`Self::schedule_deletion_of_unlinked`]
+    /// is invoked on them.
+    #[allow(unused)] // will be used by PR#4938
+    pub(crate) fn schedule_unlinking_of_layers_from_index_part(
+        self: &Arc<Self>,
+        names: Vec<LayerFileName>,
+    ) -> anyhow::Result<()> {
+        let mut guard = self.upload_queue.lock().unwrap();
+        let upload_queue = guard.initialized_mut()?;
+
+        // just forget the return value; after uploading the next index_part.json, we can consider
+        // the layer files as "dangling". this is fine however.
+        self.schedule_unlinking_of_layers_from_index_part0(upload_queue, &names);
+
+        self.launch_queued_tasks(upload_queue);
+
+        Ok(())
+    }
+
+    /// Update the remote index file, removing the to-be-deleted files from the index,
+    /// allowing scheduling of actual deletions later.
+    fn schedule_unlinking_of_layers_from_index_part0(
+        self: &Arc<Self>,
+        upload_queue: &mut UploadQueueInitialized,
+        names: &[LayerFileName],
+    ) -> Vec<(LayerFileName, Generation)> {
         // Deleting layers doesn't affect the values stored in TimelineMetadata,
         // so we don't need update it. Just serialize it.
         let metadata = upload_queue.latest_metadata.clone();
 
-        // Update the remote index file, removing the to-be-deleted files from the index,
-        // before deleting the actual files.
-        //
-        // Once we start removing files from upload_queue.latest_files, there's
-        // no going back! Otherwise, some of the files would already be removed
-        // from latest_files, but not yet scheduled for deletion. Use a closure
-        // to syntactically forbid ? or bail! calls here.
-        let no_bail_here = || {
-            // Decorate our list of names with each name's generation, dropping
-            // makes that are unexpectedly missing from our metadata.
-            let with_generations: Vec<_> = names
-                .into_iter()
-                .filter_map(|name| {
-                    // Remove from latest_files, learning the file's remote generation in the process
-                    let meta = upload_queue.latest_files.remove(&name);
+        // Decorate our list of names with each name's generation, dropping
+        // makes that are unexpectedly missing from our metadata.
+        let with_generations: Vec<_> = names
+            .iter()
+            .filter_map(|name| {
+                // Remove from latest_files, learning the file's remote generation in the process
+                let meta = upload_queue.latest_files.remove(name);
 
-                    if let Some(meta) = meta {
-                        upload_queue.latest_files_changes_since_metadata_upload_scheduled += 1;
-                        Some((name, meta.generation))
-                    } else {
-                        // This can only happen if we forgot to to schedule the file upload
-                        // before scheduling the delete. Log it because it is a rare/strange
-                        // situation, and in case something is misbehaving, we'd like to know which
-                        // layers experienced this.
-                        info!(
-                            "Deleting layer {name} not found in latest_files list, never uploaded?"
-                        );
-                        None
-                    }
-                })
-                .collect();
+                if let Some(meta) = meta {
+                    upload_queue.latest_files_changes_since_metadata_upload_scheduled += 1;
+                    Some((name.to_owned(), meta.generation))
+                } else {
+                    // This can only happen if we forgot to to schedule the file upload
+                    // before scheduling the delete. Log it because it is a rare/strange
+                    // situation, and in case something is misbehaving, we'd like to know which
+                    // layers experienced this.
+                    info!("Deleting layer {name} not found in latest_files list, never uploaded?");
+                    None
+                }
+            })
+            .collect();
 
-            if upload_queue.latest_files_changes_since_metadata_upload_scheduled > 0 {
-                self.schedule_index_upload(upload_queue, metadata);
-            }
+        // after unlinking files from the upload_queue.latest_files we must always schedule an
+        // index_part update, because that needs to be uploaded before we can actually delete the
+        // files.
+        if upload_queue.latest_files_changes_since_metadata_upload_scheduled > 0 {
+            self.schedule_index_upload(upload_queue, metadata);
+        }
 
-            for (name, gen) in &with_generations {
-                info!("scheduling deletion of layer {}{}", name, gen.get_suffix());
-            }
+        with_generations
+    }
 
-            // schedule the actual deletions
-            let op = UploadOp::Delete(Delete {
-                layers: with_generations,
-            });
-            self.calls_unfinished_metric_begin(&op);
-            upload_queue.queued_operations.push_back(op);
+    /// Schedules deletion for layer files which have previously been unlinked from the
+    /// `index_part.json` with [`Self::schedule_unlinking_of_layers_from_index_part`].
+    #[allow(unused)] // will be used by Layer::drop in PR#4938
+    pub(crate) fn schedule_deletion_of_unlinked(
+        self: &Arc<Self>,
+        layers: Vec<(LayerFileName, Generation)>,
+    ) -> anyhow::Result<()> {
+        let mut guard = self.upload_queue.lock().unwrap();
+        let upload_queue = guard.initialized_mut()?;
+
+        self.schedule_deletion_of_unlinked0(upload_queue, layers);
+        self.launch_queued_tasks(upload_queue);
+        Ok(())
+    }
+
+    fn schedule_deletion_of_unlinked0(
+        self: &Arc<Self>,
+        upload_queue: &mut UploadQueueInitialized,
+        with_generations: Vec<(LayerFileName, Generation)>,
+    ) {
+        for (name, gen) in &with_generations {
+            info!("scheduling deletion of layer {}{}", name, gen.get_suffix());
+        }
+
+        // schedule the actual deletions
+        let op = UploadOp::Delete(Delete {
+            layers: with_generations,
+        });
+        self.calls_unfinished_metric_begin(&op);
+        upload_queue.queued_operations.push_back(op);
+    }
+
+    /// Schedules a compaction update to the remote `index_part.json`.
+    ///
+    /// `compacted_from` represent the L0 names which have been `compacted_to` L1 layers.
+    pub(crate) fn schedule_compaction_update(
+        self: &Arc<Self>,
+        compacted_from: &[LayerFileName],
+        compacted_to: &[(LayerFileName, LayerFileMetadata)],
+    ) -> anyhow::Result<()> {
+        let mut guard = self.upload_queue.lock().unwrap();
+        let upload_queue = guard.initialized_mut()?;
+
+        for (name, m) in compacted_to {
+            self.schedule_layer_file_upload0(upload_queue, name, m);
+        }
+
+        let with_generations =
+            self.schedule_unlinking_of_layers_from_index_part0(upload_queue, compacted_from);
+        self.schedule_deletion_of_unlinked0(upload_queue, with_generations);
+        self.launch_queued_tasks(upload_queue);
 
-            // Launch the tasks immediately, if possible
-            self.launch_queued_tasks(upload_queue);
-        };
-        no_bail_here();
         Ok(())
     }
 
@@ -901,9 +978,27 @@ impl RemoteTimelineClient {
         .await
         .context("list prefixes")?;
 
-        let remaining: Vec<RemotePath> = remaining
+        // We will delete the current index_part object last, since it acts as a deletion
+        // marker via its deleted_at attribute
+        let latest_index = remaining
+            .iter()
+            .filter(|p| {
+                p.object_name()
+                    .map(|n| n.starts_with(IndexPart::FILE_NAME))
+                    .unwrap_or(false)
+            })
+            .filter_map(|path| parse_remote_index_path(path.clone()).map(|gen| (path, gen)))
+            .max_by_key(|i| i.1)
+            .map(|i| i.0.clone())
+            .unwrap_or(
+                // No generation-suffixed indices, assume we are dealing with
+                // a legacy index.
+                remote_index_path(&self.tenant_id, &self.timeline_id, Generation::none()),
+            );
+
+        let remaining_layers: Vec<RemotePath> = remaining
             .into_iter()
-            .filter(|p| p.object_name() != Some(IndexPart::FILE_NAME))
+            .filter(|p| p!= &latest_index)
             .inspect(|path| {
                 if let Some(name) = path.object_name() {
                     info!(%name, "deleting a file not referenced from index_part.json");
@@ -913,9 +1008,11 @@ impl RemoteTimelineClient {
             })
             .collect();
 
-        let not_referenced_count = remaining.len();
-        if !remaining.is_empty() {
-            self.deletion_queue_client.push_immediate(remaining).await?;
+        let not_referenced_count = remaining_layers.len();
+        if !remaining_layers.is_empty() {
+            self.deletion_queue_client
+                .push_immediate(remaining_layers)
+                .await?;
         }
 
         fail::fail_point!("timeline-delete-before-index-delete", |_| {
@@ -924,11 +1021,9 @@ impl RemoteTimelineClient {
             ))?
         });
 
-        let index_file_path = timeline_storage_path.join(Utf8Path::new(IndexPart::FILE_NAME));
-
         debug!("enqueuing index part deletion");
         self.deletion_queue_client
-            .push_immediate([index_file_path].to_vec())
+            .push_immediate([latest_index].to_vec())
             .await?;
 
         // Timeline deletion is rare and we have probably emitted a reasonably number of objects: wait
@@ -1401,6 +1496,13 @@ impl RemoteTimelineClient {
             }
         }
     }
+
+    pub(crate) fn get_layer_metadata(
+        &self,
+        name: &LayerFileName,
+    ) -> anyhow::Result<Option<LayerFileMetadata>> {
+        self.upload_queue.lock().unwrap().get_layer_metadata(name)
+    }
 }
 
 pub fn remote_timelines_path(tenant_id: &TenantId) -> RemotePath {

pageserver/src/tenant/remote_timeline_client/download.rs

@@ -18,7 +18,7 @@ use crate::config::PageServerConf;
 use crate::tenant::remote_timeline_client::{remote_layer_path, remote_timelines_path};
 use crate::tenant::storage_layer::LayerFileName;
 use crate::tenant::timeline::span::debug_assert_current_span_has_tenant_and_timeline_id;
-use crate::tenant::Generation;
+use crate::tenant::{Generation, TENANT_DELETED_MARKER_FILE_NAME};
 use remote_storage::{DownloadError, GenericRemoteStorage};
 use utils::crashsafe::path_with_suffix_extension;
 use utils::id::{TenantId, TimelineId};
@@ -190,6 +190,12 @@ pub async fn list_remote_timelines(
     let mut timeline_ids = HashSet::new();
 
     for timeline_remote_storage_key in timelines {
+        if timeline_remote_storage_key.object_name() == Some(TENANT_DELETED_MARKER_FILE_NAME) {
+            // A `deleted` key within `timelines/` is a marker file, not a timeline.  Ignore it.
+            // This code will be removed in https://github.com/neondatabase/neon/pull/5580
+            continue;
+        }
+
         let object_name = timeline_remote_storage_key.object_name().ok_or_else(|| {
             anyhow::anyhow!("failed to get timeline id for remote tenant {tenant_id}")
         })?;

pageserver/src/tenant/remote_timeline_client/upload.rs

@@ -31,6 +31,7 @@ pub(super) async fn upload_index_part<'a>(
     fail_point!("before-upload-index", |_| {
         bail!("failpoint before-upload-index")
     });
+    pausable_failpoint!("before-upload-index-pausable");
 
     let index_part_bytes =
         serde_json::to_vec(&index_part).context("serialize index part file into bytes")?;
@@ -59,6 +60,8 @@ pub(super) async fn upload_timeline_layer<'a>(
         bail!("failpoint before-upload-layer")
     });
 
+    pausable_failpoint!("before-upload-layer-pausable");
+
     let storage_path = remote_path(conf, source_path, generation)?;
     let source_file_res = fs::File::open(&source_path).await;
     let source_file = match source_file_res {

pageserver/src/tenant/storage_layer/delta_layer.rs

@@ -511,8 +511,7 @@ impl DeltaLayer {
     ///
     /// This variant is only used for debugging purposes, by the 'pagectl' binary.
     pub fn new_for_path(path: &Utf8Path, file: File) -> Result<Self> {
-        let mut summary_buf = Vec::new();
-        summary_buf.resize(PAGE_SZ, 0);
+        let mut summary_buf = vec![0; PAGE_SZ];
         file.read_exact_at(&mut summary_buf, 0)?;
         let summary = Summary::des_prefix(&summary_buf)?;
 
@@ -845,6 +844,49 @@ impl Drop for DeltaLayerWriter {
     }
 }
 
+impl DeltaLayer {
+    /// Assume the file at `path` is corrupt if this function returns with an error.
+    pub(crate) async fn rewrite_tenant_timeline(
+        path: &Utf8Path,
+        new_tenant: TenantId,
+        new_timeline: TimelineId,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<()> {
+        let file = VirtualFile::open_with_options(
+            path,
+            &*std::fs::OpenOptions::new().read(true).write(true),
+        )
+        .await
+        .with_context(|| format!("Failed to open file '{}'", path))?;
+        let file = FileBlockReader::new(file);
+        let summary_blk = file.read_blk(0, ctx).await?;
+        let actual_summary = Summary::des_prefix(summary_blk.as_ref())?;
+        let mut file = file.file;
+        if actual_summary.magic != DELTA_FILE_MAGIC {
+            bail!("File '{}' is not a delta layer", path);
+        }
+        let new_summary = Summary {
+            tenant_id: new_tenant,
+            timeline_id: new_timeline,
+            ..actual_summary
+        };
+
+        let mut buf = smallvec::SmallVec::<[u8; PAGE_SZ]>::new();
+        Summary::ser_into(&new_summary, &mut buf)?;
+        if buf.spilled() {
+            // The code in ImageLayerWriterInner just warn!()s for this.
+            // It should probably error out as well.
+            anyhow::bail!(
+                "Used more than one page size for summary buffer: {}",
+                buf.len()
+            );
+        }
+        file.seek(SeekFrom::Start(0)).await?;
+        file.write_all(&buf).await?;
+        Ok(())
+    }
+}
+
 impl DeltaLayerInner {
     pub(super) async fn load(
         path: &Utf8Path,

pageserver/src/tenant/storage_layer/filename.rs

@@ -226,6 +226,14 @@ impl LayerFileName {
             _ => false,
         }
     }
+
+    pub(crate) fn kind(&self) -> &'static str {
+        use LayerFileName::*;
+        match self {
+            Delta(_) => "delta",
+            Image(_) => "image",
+        }
+    }
 }
 
 impl fmt::Display for LayerFileName {

pageserver/src/tenant/storage_layer/image_layer.rs

@@ -400,8 +400,7 @@ impl ImageLayer {
     ///
     /// This variant is only used for debugging purposes, by the 'pagectl' binary.
     pub fn new_for_path(path: &Utf8Path, file: File) -> Result<ImageLayer> {
-        let mut summary_buf = Vec::new();
-        summary_buf.resize(PAGE_SZ, 0);
+        let mut summary_buf = vec![0; PAGE_SZ];
         file.read_exact_at(&mut summary_buf, 0)?;
         let summary = Summary::des_prefix(&summary_buf)?;
         let metadata = file
@@ -437,6 +436,49 @@ impl ImageLayer {
     }
 }
 
+impl ImageLayer {
+    /// Assume the file at `path` is corrupt if this function returns with an error.
+    pub(crate) async fn rewrite_tenant_timeline(
+        path: &Utf8Path,
+        new_tenant: TenantId,
+        new_timeline: TimelineId,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<()> {
+        let file = VirtualFile::open_with_options(
+            path,
+            &*std::fs::OpenOptions::new().read(true).write(true),
+        )
+        .await
+        .with_context(|| format!("Failed to open file '{}'", path))?;
+        let file = FileBlockReader::new(file);
+        let summary_blk = file.read_blk(0, ctx).await?;
+        let actual_summary = Summary::des_prefix(summary_blk.as_ref())?;
+        let mut file = file.file;
+        if actual_summary.magic != IMAGE_FILE_MAGIC {
+            bail!("File '{}' is not a delta layer", path);
+        }
+        let new_summary = Summary {
+            tenant_id: new_tenant,
+            timeline_id: new_timeline,
+            ..actual_summary
+        };
+
+        let mut buf = smallvec::SmallVec::<[u8; PAGE_SZ]>::new();
+        Summary::ser_into(&new_summary, &mut buf)?;
+        if buf.spilled() {
+            // The code in ImageLayerWriterInner just warn!()s for this.
+            // It should probably error out as well.
+            anyhow::bail!(
+                "Used more than one page size for summary buffer: {}",
+                buf.len()
+            );
+        }
+        file.seek(SeekFrom::Start(0)).await?;
+        file.write_all(&buf).await?;
+        Ok(())
+    }
+}
+
 impl ImageLayerInner {
     pub(super) async fn load(
         path: &Utf8Path,

pageserver/src/tenant/storage_layer/remote_layer.rs

@@ -25,7 +25,7 @@ use super::{
 };
 
 /// RemoteLayer is a not yet downloaded [`ImageLayer`] or
-/// [`DeltaLayer`](super::DeltaLayer).
+/// [`DeltaLayer`].
 ///
 /// RemoteLayer might be downloaded on-demand during operations which are
 /// allowed download remote layers and during which, it gets replaced with a

pageserver/src/tenant/tasks.rs

@@ -14,6 +14,73 @@ use tokio_util::sync::CancellationToken;
 use tracing::*;
 use utils::completion;
 
+static CONCURRENT_BACKGROUND_TASKS: once_cell::sync::Lazy<tokio::sync::Semaphore> =
+    once_cell::sync::Lazy::new(|| {
+        let total_threads = *task_mgr::BACKGROUND_RUNTIME_WORKER_THREADS;
+        let permits = usize::max(
+            1,
+            // while a lot of the work is done on spawn_blocking, we still do
+            // repartitioning in the async context. this should give leave us some workers
+            // unblocked to be blocked on other work, hopefully easing any outside visible
+            // effects of restarts.
+            //
+            // 6/8 is a guess; previously we ran with unlimited 8 and more from
+            // spawn_blocking.
+            (total_threads * 3).checked_div(4).unwrap_or(0),
+        );
+        assert_ne!(permits, 0, "we will not be adding in permits later");
+        assert!(
+            permits < total_threads,
+            "need threads avail for shorter work"
+        );
+        tokio::sync::Semaphore::new(permits)
+    });
+
+#[derive(Debug, PartialEq, Eq, Clone, Copy, strum_macros::IntoStaticStr)]
+#[strum(serialize_all = "snake_case")]
+pub(crate) enum BackgroundLoopKind {
+    Compaction,
+    Gc,
+    Eviction,
+    ConsumptionMetricsCollectMetrics,
+    ConsumptionMetricsSyntheticSizeWorker,
+}
+
+impl BackgroundLoopKind {
+    fn as_static_str(&self) -> &'static str {
+        let s: &'static str = self.into();
+        s
+    }
+}
+
+pub(crate) enum RateLimitError {
+    Cancelled,
+}
+
+pub(crate) async fn concurrent_background_tasks_rate_limit(
+    loop_kind: BackgroundLoopKind,
+    _ctx: &RequestContext,
+    cancel: &CancellationToken,
+) -> Result<impl Drop, RateLimitError> {
+    crate::metrics::BACKGROUND_LOOP_SEMAPHORE_WAIT_START_COUNT
+        .with_label_values(&[loop_kind.as_static_str()])
+        .inc();
+    scopeguard::defer!(
+        crate::metrics::BACKGROUND_LOOP_SEMAPHORE_WAIT_FINISH_COUNT.with_label_values(&[loop_kind.as_static_str()]).inc();
+    );
+    tokio::select! {
+        permit = CONCURRENT_BACKGROUND_TASKS.acquire() => {
+            match permit {
+                Ok(permit) => Ok(permit),
+                Err(_closed) => unreachable!("we never close the semaphore"),
+            }
+        },
+        _ = cancel.cancelled() => {
+            Err(RateLimitError::Cancelled)
+        }
+    }
+}
+
 /// Start per tenant background loops: compaction and gc.
 pub fn start_background_loops(
     tenant: &Arc<Tenant>,
@@ -92,11 +159,8 @@ async fn compaction_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
 
             // TODO: we shouldn't need to await to find tenant and this could be moved outside of
             // loop, #3501. There are also additional "allowed_errors" in tests.
-            if first {
-                first = false;
-                if random_init_delay(period, &cancel).await.is_err() {
-                    break;
-                }
+            if first && random_init_delay(period, &cancel).await.is_err() {
+                break;
             }
 
             let started_at = Instant::now();
@@ -116,7 +180,16 @@ async fn compaction_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
                 }
             };
 
-            warn_when_period_overrun(started_at.elapsed(), period, "compaction");
+            if !first {
+                // The first iteration is typically much slower, because all tenants compete for the
+                // compaction sempahore to run, and because of concurrent startup work like initializing
+                // logical sizes.  To avoid routinely spamming warnings, we suppress this log on first iteration.
+                warn_when_period_overrun(
+                    started_at.elapsed(),
+                    period,
+                    BackgroundLoopKind::Compaction,
+                );
+            }
 
             // Sleep
             if tokio::time::timeout(sleep_duration, cancel.cancelled())
@@ -125,6 +198,8 @@ async fn compaction_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
             {
                 break;
             }
+
+            first = false;
         }
     }
     .await;
@@ -156,11 +231,8 @@ async fn gc_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
 
             let period = tenant.get_gc_period();
 
-            if first {
-                first = false;
-                if random_init_delay(period, &cancel).await.is_err() {
-                    break;
-                }
+            if first && random_init_delay(period, &cancel).await.is_err() {
+                break;
             }
 
             let started_at = Instant::now();
@@ -184,7 +256,12 @@ async fn gc_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
                 }
             };
 
-            warn_when_period_overrun(started_at.elapsed(), period, "gc");
+            if !first {
+                // The first iteration is typically much slower, because all tenants compete for the
+                // compaction sempahore to run, and because of concurrent startup work like initializing
+                // logical sizes.  To avoid routinely spamming warnings, we suppress this log on first iteration.
+                warn_when_period_overrun(started_at.elapsed(), period, BackgroundLoopKind::Gc);
+            }
 
             // Sleep
             if tokio::time::timeout(sleep_duration, cancel.cancelled())
@@ -193,6 +270,8 @@ async fn gc_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
             {
                 break;
             }
+
+            first = false;
         }
     }
     .await;
@@ -258,7 +337,11 @@ pub(crate) async fn random_init_delay(
 }
 
 /// Attention: the `task` and `period` beocme labels of a pageserver-wide prometheus metric.
-pub(crate) fn warn_when_period_overrun(elapsed: Duration, period: Duration, task: &str) {
+pub(crate) fn warn_when_period_overrun(
+    elapsed: Duration,
+    period: Duration,
+    task: BackgroundLoopKind,
+) {
     // Duration::ZERO will happen because it's the "disable [bgtask]" value.
     if elapsed >= period && period != Duration::ZERO {
         // humantime does no significant digits clamping whereas Duration's debug is a bit more
@@ -267,11 +350,11 @@ pub(crate) fn warn_when_period_overrun(elapsed: Duration, period: Duration, task
         warn!(
             ?elapsed,
             period = %humantime::format_duration(period),
-            task,
+            ?task,
             "task iteration took longer than the configured period"
         );
         crate::metrics::BACKGROUND_LOOP_PERIOD_OVERRUN_COUNT
-            .with_label_values(&[task, &format!("{}", period.as_secs())])
+            .with_label_values(&[task.as_static_str(), &format!("{}", period.as_secs())])
             .inc();
     }
 }

pageserver/src/tenant/timeline.rs

@@ -44,6 +44,7 @@ use crate::tenant::storage_layer::delta_layer::DeltaEntry;
 use crate::tenant::storage_layer::{
     DeltaLayerWriter, ImageLayerWriter, InMemoryLayer, LayerAccessStats, LayerFileName, RemoteLayer,
 };
+use crate::tenant::tasks::{BackgroundLoopKind, RateLimitError};
 use crate::tenant::timeline::logical_size::CurrentLogicalSize;
 use crate::tenant::{
     layer_map::{LayerMap, SearchResult},
@@ -80,7 +81,6 @@ use crate::repository::GcResult;
 use crate::repository::{Key, Value};
 use crate::task_mgr;
 use crate::task_mgr::TaskKind;
-use crate::walredo::WalRedoManager;
 use crate::ZERO_PAGE;
 
 use self::delete::DeleteTimelineFlow;
@@ -91,12 +91,12 @@ use self::logical_size::LogicalSize;
 use self::walreceiver::{WalReceiver, WalReceiverConf};
 
 use super::config::TenantConf;
-use super::debug_assert_current_span_has_tenant_and_timeline_id;
 use super::remote_timeline_client::index::IndexPart;
 use super::remote_timeline_client::RemoteTimelineClient;
 use super::storage_layer::{
     AsLayerDesc, DeltaLayer, ImageLayer, LayerAccessStatsReset, PersistentLayerDesc,
 };
+use super::{debug_assert_current_span_has_tenant_and_timeline_id, AttachedTenantConf};
 
 #[derive(Debug, PartialEq, Eq, Clone, Copy)]
 pub(super) enum FlushLoopState {
@@ -149,7 +149,7 @@ pub struct TimelineResources {
 
 pub struct Timeline {
     conf: &'static PageServerConf,
-    tenant_conf: Arc<RwLock<TenantConfOpt>>,
+    tenant_conf: Arc<RwLock<AttachedTenantConf>>,
 
     myself: Weak<Self>,
 
@@ -158,6 +158,9 @@ pub struct Timeline {
 
     /// The generation of the tenant that instantiated us: this is used for safety when writing remote objects.
     /// Never changes for the lifetime of this [`Timeline`] object.
+    ///
+    /// This duplicates the generation stored in LocationConf, but that structure is mutable:
+    /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
     generation: Generation,
 
     pub pg_version: u32,
@@ -197,7 +200,7 @@ pub struct Timeline {
     last_freeze_ts: RwLock<Instant>,
 
     // WAL redo manager
-    walredo_mgr: Arc<dyn WalRedoManager + Sync + Send>,
+    walredo_mgr: Arc<super::WalRedoManager>,
 
     /// Remote storage client.
     /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
@@ -367,7 +370,7 @@ pub enum PageReconstructError {
 
     /// An error happened replaying WAL records
     #[error(transparent)]
-    WalRedo(#[from] crate::walredo::WalRedoError),
+    WalRedo(anyhow::Error),
 }
 
 impl std::fmt::Debug for PageReconstructError {
@@ -681,37 +684,17 @@ impl Timeline {
     ) -> anyhow::Result<()> {
         const ROUNDS: usize = 2;
 
-        static CONCURRENT_COMPACTIONS: once_cell::sync::Lazy<tokio::sync::Semaphore> =
-            once_cell::sync::Lazy::new(|| {
-                let total_threads = *task_mgr::BACKGROUND_RUNTIME_WORKER_THREADS;
-                let permits = usize::max(
-                    1,
-                    // while a lot of the work is done on spawn_blocking, we still do
-                    // repartitioning in the async context. this should give leave us some workers
-                    // unblocked to be blocked on other work, hopefully easing any outside visible
-                    // effects of restarts.
-                    //
-                    // 6/8 is a guess; previously we ran with unlimited 8 and more from
-                    // spawn_blocking.
-                    (total_threads * 3).checked_div(4).unwrap_or(0),
-                );
-                assert_ne!(permits, 0, "we will not be adding in permits later");
-                assert!(
-                    permits < total_threads,
-                    "need threads avail for shorter work"
-                );
-                tokio::sync::Semaphore::new(permits)
-            });
-
         // this wait probably never needs any "long time spent" logging, because we already nag if
         // compaction task goes over it's period (20s) which is quite often in production.
-        let _permit = tokio::select! {
-            permit = CONCURRENT_COMPACTIONS.acquire() => {
-                permit
-            },
-            _ = cancel.cancelled() => {
-                return Ok(());
-            }
+        let _permit = match super::tasks::concurrent_background_tasks_rate_limit(
+            BackgroundLoopKind::Compaction,
+            ctx,
+            cancel,
+        )
+        .await
+        {
+            Ok(permit) => permit,
+            Err(RateLimitError::Cancelled) => return Ok(()),
         };
 
         let last_record_lsn = self.get_last_record_lsn();
@@ -1291,7 +1274,23 @@ impl Timeline {
                 Ok(delta) => Some(delta),
             };
 
-        let layer_metadata = LayerFileMetadata::new(layer_file_size, self.generation);
+        // RemoteTimelineClient holds the metadata on layers' remote generations, so
+        // query it to construct a RemoteLayer.
+        let layer_metadata = self
+            .remote_client
+            .as_ref()
+            .expect("Eviction is not called without remote storage")
+            .get_layer_metadata(&local_layer.filename())
+            .map_err(EvictionError::LayerNotFound)?
+            .ok_or_else(|| {
+                EvictionError::LayerNotFound(anyhow::anyhow!("Layer not in remote metadata"))
+            })?;
+        if layer_metadata.file_size() != layer_file_size {
+            return Err(EvictionError::MetadataInconsistency(format!(
+                "Layer size {layer_file_size} doesn't match remote metadata file size {}",
+                layer_metadata.file_size()
+            )));
+        }
 
         let new_remote_layer = Arc::new(match local_layer.filename() {
             LayerFileName::Image(image_name) => RemoteLayer::new_img(
@@ -1370,6 +1369,10 @@ pub(crate) enum EvictionError {
     /// different objects in memory.
     #[error("layer was no longer part of LayerMap")]
     LayerNotFound(#[source] anyhow::Error),
+
+    /// This should never happen
+    #[error("Metadata inconsistency")]
+    MetadataInconsistency(String),
 }
 
 /// Number of times we will compute partition within a checkpoint distance.
@@ -1378,42 +1381,42 @@ const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
 // Private functions
 impl Timeline {
     fn get_checkpoint_distance(&self) -> u64 {
-        let tenant_conf = self.tenant_conf.read().unwrap();
+        let tenant_conf = self.tenant_conf.read().unwrap().tenant_conf;
         tenant_conf
             .checkpoint_distance
             .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
     }
 
     fn get_checkpoint_timeout(&self) -> Duration {
-        let tenant_conf = self.tenant_conf.read().unwrap();
+        let tenant_conf = self.tenant_conf.read().unwrap().tenant_conf;
         tenant_conf
             .checkpoint_timeout
             .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
     }
 
     fn get_compaction_target_size(&self) -> u64 {
-        let tenant_conf = self.tenant_conf.read().unwrap();
+        let tenant_conf = self.tenant_conf.read().unwrap().tenant_conf;
         tenant_conf
             .compaction_target_size
             .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
     }
 
     fn get_compaction_threshold(&self) -> usize {
-        let tenant_conf = self.tenant_conf.read().unwrap();
+        let tenant_conf = self.tenant_conf.read().unwrap().tenant_conf;
         tenant_conf
             .compaction_threshold
             .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
     }
 
     fn get_image_creation_threshold(&self) -> usize {
-        let tenant_conf = self.tenant_conf.read().unwrap();
+        let tenant_conf = self.tenant_conf.read().unwrap().tenant_conf;
         tenant_conf
             .image_creation_threshold
             .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
     }
 
     fn get_eviction_policy(&self) -> EvictionPolicy {
-        let tenant_conf = self.tenant_conf.read().unwrap();
+        let tenant_conf = self.tenant_conf.read().unwrap().tenant_conf;
         tenant_conf
             .eviction_policy
             .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
@@ -1429,7 +1432,7 @@ impl Timeline {
     }
 
     fn get_gc_feedback(&self) -> bool {
-        let tenant_conf = self.tenant_conf.read().unwrap();
+        let tenant_conf = &self.tenant_conf.read().unwrap().tenant_conf;
         tenant_conf
             .gc_feedback
             .unwrap_or(self.conf.default_tenant_conf.gc_feedback)
@@ -1442,7 +1445,7 @@ impl Timeline {
         // The threshold is embedded in the metric. So, we need to update it.
         {
             let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
-                &self.tenant_conf.read().unwrap(),
+                &self.tenant_conf.read().unwrap().tenant_conf,
                 &self.conf.default_tenant_conf,
             );
             let tenant_id_str = self.tenant_id.to_string();
@@ -1461,13 +1464,13 @@ impl Timeline {
     #[allow(clippy::too_many_arguments)]
     pub(super) fn new(
         conf: &'static PageServerConf,
-        tenant_conf: Arc<RwLock<TenantConfOpt>>,
+        tenant_conf: Arc<RwLock<AttachedTenantConf>>,
         metadata: &TimelineMetadata,
         ancestor: Option<Arc<Timeline>>,
         timeline_id: TimelineId,
         tenant_id: TenantId,
         generation: Generation,
-        walredo_mgr: Arc<dyn WalRedoManager + Send + Sync>,
+        walredo_mgr: Arc<super::WalRedoManager>,
         resources: TimelineResources,
         pg_version: u32,
         initial_logical_size_can_start: Option<completion::Barrier>,
@@ -1484,7 +1487,7 @@ impl Timeline {
 
         let evictions_low_residence_duration_metric_threshold =
             Self::get_evictions_low_residence_duration_metric_threshold(
-                &tenant_conf_guard,
+                &tenant_conf_guard.tenant_conf,
                 &conf.default_tenant_conf,
             );
         drop(tenant_conf_guard);
@@ -1649,12 +1652,15 @@ impl Timeline {
 
         let tenant_conf_guard = self.tenant_conf.read().unwrap();
         let wal_connect_timeout = tenant_conf_guard
+            .tenant_conf
             .walreceiver_connect_timeout
             .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
         let lagging_wal_timeout = tenant_conf_guard
+            .tenant_conf
             .lagging_wal_timeout
             .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
         let max_lsn_wal_lag = tenant_conf_guard
+            .tenant_conf
             .max_lsn_wal_lag
             .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
         drop(tenant_conf_guard);
@@ -1693,7 +1699,7 @@ impl Timeline {
         disk_consistent_lsn: Lsn,
         index_part: Option<IndexPart>,
     ) -> anyhow::Result<()> {
-        use init::{Decision::*, Discovered, FutureLayer};
+        use init::{Decision::*, Discovered, DismissedLayer};
         use LayerFileName::*;
 
         let mut guard = self.layers.write().await;
@@ -1709,7 +1715,7 @@ impl Timeline {
         // Copy to move into the task we're about to spawn
         let generation = self.generation;
 
-        let (loaded_layers, to_sync, total_physical_size) = tokio::task::spawn_blocking({
+        let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
             move || {
                 let _g = span.entered();
                 let discovered = init::scan_timeline_dir(&timeline_path)?;
@@ -1758,7 +1764,6 @@ impl Timeline {
                 );
 
                 let mut loaded_layers = Vec::new();
-                let mut needs_upload = Vec::new();
                 let mut needs_cleanup = Vec::new();
                 let mut total_physical_size = 0;
 
@@ -1779,7 +1784,7 @@ impl Timeline {
                             }
                         }
                         Ok(decision) => decision,
-                        Err(FutureLayer { local }) => {
+                        Err(DismissedLayer::Future { local }) => {
                             if local.is_some() {
                                 path.push(name.file_name());
                                 init::cleanup_future_layer(&path, &name, disk_consistent_lsn)?;
@@ -1788,6 +1793,13 @@ impl Timeline {
                             needs_cleanup.push(name);
                             continue;
                         }
+                        Err(DismissedLayer::LocalOnly(local)) => {
+                            path.push(name.file_name());
+                            init::cleanup_local_only_file(&path, &name, &local)?;
+                            path.pop();
+                            // this file never existed remotely, we will have to do rework
+                            continue;
+                        }
                     };
 
                     match &name {
@@ -1796,14 +1808,16 @@ impl Timeline {
                     }
 
                     let status = match &decision {
-                        UseLocal(_) | NeedsUpload(_) => LayerResidenceStatus::Resident,
+                        UseLocal(_) => LayerResidenceStatus::Resident,
                         Evicted(_) | UseRemote { .. } => LayerResidenceStatus::Evicted,
                     };
 
+                    tracing::debug!(layer=%name, ?decision, ?status, "applied");
+
                     let stats = LayerAccessStats::for_loading_layer(status);
 
                     let layer: Arc<dyn PersistentLayer> = match (name, &decision) {
-                        (Delta(d), UseLocal(m) | NeedsUpload(m)) => {
+                        (Delta(d), UseLocal(m)) => {
                             total_physical_size += m.file_size();
                             Arc::new(DeltaLayer::new(
                                 conf,
@@ -1814,7 +1828,7 @@ impl Timeline {
                                 stats,
                             ))
                         }
-                        (Image(i), UseLocal(m) | NeedsUpload(m)) => {
+                        (Image(i), UseLocal(m)) => {
                             total_physical_size += m.file_size();
                             Arc::new(ImageLayer::new(
                                 conf,
@@ -1833,17 +1847,9 @@ impl Timeline {
                         ),
                     };
 
-                    if let NeedsUpload(m) = decision {
-                        needs_upload.push((layer.clone(), m));
-                    }
-
                     loaded_layers.push(layer);
                 }
-                Ok((
-                    loaded_layers,
-                    (needs_upload, needs_cleanup),
-                    total_physical_size,
-                ))
+                Ok((loaded_layers, needs_cleanup, total_physical_size))
             }
         })
         .await
@@ -1855,10 +1861,6 @@ impl Timeline {
         guard.initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
 
         if let Some(rtc) = self.remote_client.as_ref() {
-            let (needs_upload, needs_cleanup) = to_sync;
-            for (layer, m) in needs_upload {
-                rtc.schedule_layer_file_upload(&layer.layer_desc().filename(), &m)?;
-            }
             rtc.schedule_layer_file_deletion(needs_cleanup)?;
             rtc.schedule_index_upload_for_file_changes()?;
             // Tenant::create_timeline will wait for these uploads to happen before returning, or
@@ -1869,7 +1871,7 @@ impl Timeline {
             "loaded layer map with {} layers at {}, total physical size: {}",
             num_layers, disk_consistent_lsn, total_physical_size
         );
-        self.metrics.resident_physical_size_set(total_physical_size);
+        self.metrics.resident_physical_size_add(total_physical_size);
 
         timer.stop_and_record();
         Ok(())
@@ -2357,7 +2359,7 @@ impl Timeline {
                 // during branch creation.
                 match ancestor.wait_to_become_active(ctx).await {
                     Ok(()) => {}
-                    Err(state) if state == TimelineState::Stopping => {
+                    Err(TimelineState::Stopping) => {
                         return Err(PageReconstructError::AncestorStopping(ancestor.timeline_id));
                     }
                     Err(state) => {
@@ -2791,10 +2793,13 @@ impl Timeline {
                 )
             };
 
+        let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
+        let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
+
         // The new on-disk layers are now in the layer map. We can remove the
         // in-memory layer from the map now. The flushed layer is stored in
         // the mapping in `create_delta_layer`.
-        {
+        let metadata = {
             let mut guard = self.layers.write().await;
 
             if let Some(ref l) = delta_layer_to_add {
@@ -2810,8 +2815,17 @@ impl Timeline {
             }
 
             guard.finish_flush_l0_layer(delta_layer_to_add, &frozen_layer);
+            if disk_consistent_lsn != old_disk_consistent_lsn {
+                assert!(disk_consistent_lsn > old_disk_consistent_lsn);
+                self.disk_consistent_lsn.store(disk_consistent_lsn);
+
+                // Schedule remote uploads that will reflect our new disk_consistent_lsn
+                Some(self.schedule_uploads(disk_consistent_lsn, layer_paths_to_upload)?)
+            } else {
+                None
+            }
             // release lock on 'layers'
-        }
+        };
 
         // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
         // a compaction can delete the file and then it won't be available for uploads any more.
@@ -2827,28 +2841,22 @@ impl Timeline {
         //
         // TODO: This perhaps should be done in 'flush_frozen_layers', after flushing
         // *all* the layers, to avoid fsyncing the file multiple times.
-        let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
-        let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
 
-        // If we were able to advance 'disk_consistent_lsn', save it the metadata file.
-        // After crash, we will restart WAL streaming and processing from that point.
-        if disk_consistent_lsn != old_disk_consistent_lsn {
-            assert!(disk_consistent_lsn > old_disk_consistent_lsn);
-            self.update_metadata_file(disk_consistent_lsn, layer_paths_to_upload)
+        // If we updated our disk_consistent_lsn, persist the updated metadata to local disk.
+        if let Some(metadata) = metadata {
+            save_metadata(self.conf, &self.tenant_id, &self.timeline_id, &metadata)
                 .await
-                .context("update_metadata_file")?;
-            // Also update the in-memory copy
-            self.disk_consistent_lsn.store(disk_consistent_lsn);
+                .context("save_metadata")?;
         }
         Ok(())
     }
 
     /// Update metadata file
-    async fn update_metadata_file(
+    fn schedule_uploads(
         &self,
         disk_consistent_lsn: Lsn,
         layer_paths_to_upload: HashMap<LayerFileName, LayerFileMetadata>,
-    ) -> anyhow::Result<()> {
+    ) -> anyhow::Result<TimelineMetadata> {
         // We can only save a valid 'prev_record_lsn' value on disk if we
         // flushed *all* in-memory changes to disk. We only track
         // 'prev_record_lsn' in memory for the latest processed record, so we
@@ -2885,10 +2893,6 @@ impl Timeline {
             x.unwrap()
         ));
 
-        save_metadata(self.conf, &self.tenant_id, &self.timeline_id, &metadata)
-            .await
-            .context("save_metadata")?;
-
         if let Some(remote_client) = &self.remote_client {
             for (path, layer_metadata) in layer_paths_to_upload {
                 remote_client.schedule_layer_file_upload(&path, &layer_metadata)?;
@@ -2896,6 +2900,20 @@ impl Timeline {
             remote_client.schedule_index_upload_for_metadata_update(&metadata)?;
         }
 
+        Ok(metadata)
+    }
+
+    async fn update_metadata_file(
+        &self,
+        disk_consistent_lsn: Lsn,
+        layer_paths_to_upload: HashMap<LayerFileName, LayerFileMetadata>,
+    ) -> anyhow::Result<()> {
+        let metadata = self.schedule_uploads(disk_consistent_lsn, layer_paths_to_upload)?;
+
+        save_metadata(self.conf, &self.tenant_id, &self.timeline_id, &metadata)
+            .await
+            .context("save_metadata")?;
+
         Ok(())
     }
 
@@ -3061,6 +3079,7 @@ impl Timeline {
         Ok(false)
     }
 
+    #[tracing::instrument(skip_all, fields(%lsn, %force))]
     async fn create_image_layers(
         &self,
         partitioning: &KeyPartitioning,
@@ -3851,22 +3870,21 @@ impl Timeline {
         // now, we just skip the file to avoid unintentional modification to files on the disk and in the layer map.
         let mut duplicated_layers = HashSet::new();
 
+        let mut uploaded_layers = Vec::with_capacity(new_layers.len());
         let mut insert_layers = Vec::new();
         let mut remove_layers = Vec::new();
 
-        for l in new_layers {
+        for l in &new_layers {
             let new_delta_path = l.path();
 
             let metadata = new_delta_path.metadata().with_context(|| {
                 format!("read file metadata for new created layer {new_delta_path}")
             })?;
 
-            if let Some(remote_client) = &self.remote_client {
-                remote_client.schedule_layer_file_upload(
-                    &l.filename(),
-                    &LayerFileMetadata::new(metadata.len(), self.generation),
-                )?;
-            }
+            uploaded_layers.push((
+                l.filename(),
+                LayerFileMetadata::new(metadata.len(), self.generation),
+            ));
 
             // update metrics, including the timeline's physical size
             self.metrics.record_new_file_metrics(metadata.len());
@@ -3879,7 +3897,7 @@ impl Timeline {
                 LayerResidenceStatus::Resident,
                 LayerResidenceEventReason::LayerCreate,
             );
-            let l = l as Arc<dyn PersistentLayer>;
+            let l = l.to_owned() as Arc<dyn PersistentLayer>;
             if guard.contains(&l) {
                 tracing::error!(layer=%l, "duplicated L1 layer");
                 duplicated_layers.insert(l.layer_desc().key());
@@ -3911,13 +3929,12 @@ impl Timeline {
             &self.metrics,
         )?;
 
-        drop_wlock(guard);
-
-        // Also schedule the deletions in remote storage
-        if let Some(remote_client) = &self.remote_client {
-            remote_client.schedule_layer_file_deletion(layer_names_to_delete)?;
+        if let Some(remote_client) = self.remote_client.as_ref() {
+            remote_client.schedule_compaction_update(&layer_names_to_delete, &uploaded_layers)?;
         }
 
+        drop_wlock(guard);
+
         Ok(())
     }
 
@@ -4321,6 +4338,7 @@ impl Timeline {
                 let img = match self
                     .walredo_mgr
                     .request_redo(key, request_lsn, data.img, data.records, self.pg_version)
+                    .await
                     .context("Failed to reconstruct a page image:")
                 {
                     Ok(img) => img,

pageserver/src/tenant/timeline/eviction_task.rs

@@ -30,6 +30,7 @@ use crate::{
     tenant::{
         config::{EvictionPolicy, EvictionPolicyLayerAccessThreshold},
         storage_layer::PersistentLayer,
+        tasks::{BackgroundLoopKind, RateLimitError},
         timeline::EvictionError,
         LogicalSizeCalculationCause, Tenant,
     },
@@ -129,7 +130,11 @@ impl Timeline {
                     ControlFlow::Continue(()) => (),
                 }
                 let elapsed = start.elapsed();
-                crate::tenant::tasks::warn_when_period_overrun(elapsed, p.period, "eviction");
+                crate::tenant::tasks::warn_when_period_overrun(
+                    elapsed,
+                    p.period,
+                    BackgroundLoopKind::Eviction,
+                );
                 crate::metrics::EVICTION_ITERATION_DURATION
                     .get_metric_with_label_values(&[
                         &format!("{}", p.period.as_secs()),
@@ -150,6 +155,17 @@ impl Timeline {
     ) -> ControlFlow<()> {
         let now = SystemTime::now();
 
+        let _permit = match crate::tenant::tasks::concurrent_background_tasks_rate_limit(
+            BackgroundLoopKind::Eviction,
+            ctx,
+            cancel,
+        )
+        .await
+        {
+            Ok(permit) => permit,
+            Err(RateLimitError::Cancelled) => return ControlFlow::Break(()),
+        };
+
         // If we evict layers but keep cached values derived from those layers, then
         // we face a storm of on-demand downloads after pageserver restart.
         // The reason is that the restart empties the caches, and so, the values
@@ -285,6 +301,10 @@ impl Timeline {
                     warn!(layer = %l, "failed to evict layer: {e}");
                     stats.not_evictable += 1;
                 }
+                Some(Err(EvictionError::MetadataInconsistency(detail))) => {
+                    warn!(layer = %l, "failed to evict layer: {detail}");
+                    stats.not_evictable += 1;
+                }
             }
         }
         if stats.candidates == stats.not_evictable {

pageserver/src/tenant/timeline/init.rs

@@ -72,7 +72,7 @@ pub(super) fn scan_timeline_dir(path: &Utf8Path) -> anyhow::Result<Vec<Discovere
 }
 
 /// Decision on what to do with a layer file after considering its local and remote metadata.
-#[derive(Clone)]
+#[derive(Clone, Debug)]
 pub(super) enum Decision {
     /// The layer is not present locally.
     Evicted(LayerFileMetadata),
@@ -84,27 +84,30 @@ pub(super) enum Decision {
     },
     /// The layer is present locally, and metadata matches.
     UseLocal(LayerFileMetadata),
-    /// The layer is only known locally, it needs to be uploaded.
-    NeedsUpload(LayerFileMetadata),
 }
 
-/// The related layer is is in future compared to disk_consistent_lsn, it must not be loaded.
+/// A layer needs to be left out of the layer map.
 #[derive(Debug)]
-pub(super) struct FutureLayer {
-    /// The local metadata. `None` if the layer is only known through [`IndexPart`].
-    pub(super) local: Option<LayerFileMetadata>,
+pub(super) enum DismissedLayer {
+    /// The related layer is is in future compared to disk_consistent_lsn, it must not be loaded.
+    Future {
+        /// The local metadata. `None` if the layer is only known through [`IndexPart`].
+        local: Option<LayerFileMetadata>,
+    },
+    /// The layer only exists locally.
+    ///
+    /// In order to make crash safe updates to layer map, we must dismiss layers which are only
+    /// found locally or not yet included in the remote `index_part.json`.
+    LocalOnly(LayerFileMetadata),
 }
 
 /// Merges local discoveries and remote [`IndexPart`] to a collection of decisions.
-///
-/// This function should not gain additional reasons to fail than [`FutureLayer`], consider adding
-/// the checks earlier to [`scan_timeline_dir`].
 pub(super) fn reconcile(
     discovered: Vec<(LayerFileName, u64)>,
     index_part: Option<&IndexPart>,
     disk_consistent_lsn: Lsn,
     generation: Generation,
-) -> Vec<(LayerFileName, Result<Decision, FutureLayer>)> {
+) -> Vec<(LayerFileName, Result<Decision, DismissedLayer>)> {
     use Decision::*;
 
     // name => (local, remote)
@@ -142,17 +145,19 @@ pub(super) fn reconcile(
         .into_iter()
         .map(|(name, (local, remote))| {
             let decision = if name.is_in_future(disk_consistent_lsn) {
-                Err(FutureLayer { local })
+                Err(DismissedLayer::Future { local })
             } else {
-                Ok(match (local, remote) {
-                    (Some(local), Some(remote)) if local != remote => UseRemote { local, remote },
-                    (Some(x), Some(_)) => UseLocal(x),
-                    (None, Some(x)) => Evicted(x),
-                    (Some(x), None) => NeedsUpload(x),
+                match (local, remote) {
+                    (Some(local), Some(remote)) if local != remote => {
+                        Ok(UseRemote { local, remote })
+                    }
+                    (Some(x), Some(_)) => Ok(UseLocal(x)),
+                    (None, Some(x)) => Ok(Evicted(x)),
+                    (Some(x), None) => Err(DismissedLayer::LocalOnly(x)),
                     (None, None) => {
                         unreachable!("there must not be any non-local non-remote files")
                     }
-                })
+                }
             };
 
             (name, decision)
@@ -192,14 +197,21 @@ pub(super) fn cleanup_future_layer(
     name: &LayerFileName,
     disk_consistent_lsn: Lsn,
 ) -> anyhow::Result<()> {
-    use LayerFileName::*;
-    let kind = match name {
-        Delta(_) => "delta",
-        Image(_) => "image",
-    };
     // future image layers are allowed to be produced always for not yet flushed to disk
     // lsns stored in InMemoryLayer.
+    let kind = name.kind();
     tracing::info!("found future {kind} layer {name} disk_consistent_lsn is {disk_consistent_lsn}");
-    crate::tenant::timeline::rename_to_backup(path)?;
+    std::fs::remove_file(path)?;
+    Ok(())
+}
+
+pub(super) fn cleanup_local_only_file(
+    path: &Utf8Path,
+    name: &LayerFileName,
+    local: &LayerFileMetadata,
+) -> anyhow::Result<()> {
+    let kind = name.kind();
+    tracing::info!("found local-only {kind} layer {name}, metadata {local:?}");
+    std::fs::remove_file(path)?;
     Ok(())
 }

pageserver/src/tenant/timeline/walreceiver/connection_manager.rs

@@ -26,8 +26,7 @@ use storage_broker::proto::subscribe_safekeeper_info_request::SubscriptionKey;
 use storage_broker::proto::SafekeeperTimelineInfo;
 use storage_broker::proto::SubscribeSafekeeperInfoRequest;
 use storage_broker::proto::TenantTimelineId as ProtoTenantTimelineId;
-use storage_broker::BrokerClientChannel;
-use storage_broker::Streaming;
+use storage_broker::{BrokerClientChannel, Code, Streaming};
 use tokio::select;
 use tracing::*;
 
@@ -137,8 +136,17 @@ pub(super) async fn connection_manager_loop_step(
             broker_update = broker_subscription.message() => {
                 match broker_update {
                     Ok(Some(broker_update)) => connection_manager_state.register_timeline_update(broker_update),
-                    Err(e) => {
-                        error!("broker subscription failed: {e}");
+                    Err(status) => {
+                        match status.code() {
+                            Code::Unknown if status.message().contains("stream closed because of a broken pipe") => {
+                                // tonic's error handling doesn't provide a clear code for disconnections: we get
+                                // "h2 protocol error: error reading a body from connection: stream closed because of a broken pipe"
+                                info!("broker disconnected: {status}");
+                            },
+                            _ => {
+                                warn!("broker subscription failed: {status}");
+                            }
+                        }
                         return ControlFlow::Continue(());
                     }
                     Ok(None) => {

pageserver/src/tenant/upload_queue.rs

@@ -203,6 +203,18 @@ impl UploadQueue {
             UploadQueue::Stopped(stopped) => Ok(stopped),
         }
     }
+
+    pub(crate) fn get_layer_metadata(
+        &self,
+        name: &LayerFileName,
+    ) -> anyhow::Result<Option<LayerFileMetadata>> {
+        match self {
+            UploadQueue::Stopped(_) | UploadQueue::Uninitialized => {
+                anyhow::bail!("queue is in state {}", self.as_str())
+            }
+            UploadQueue::Initialized(inner) => Ok(inner.latest_files.get(name).cloned()),
+        }
+    }
 }
 
 /// An in-progress upload or delete task.

pageserver/src/virtual_file.rs

@@ -15,10 +15,11 @@ use crate::tenant::TENANTS_SEGMENT_NAME;
 use camino::{Utf8Path, Utf8PathBuf};
 use once_cell::sync::OnceCell;
 use std::fs::{self, File, OpenOptions};
-use std::io::{ErrorKind, Seek, SeekFrom};
+use std::io::{Error, ErrorKind, Seek, SeekFrom};
 use std::os::unix::fs::FileExt;
 use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
-use std::sync::{RwLock, RwLockWriteGuard};
+use tokio::sync::{RwLock, RwLockReadGuard, RwLockWriteGuard};
+use tokio::time::Instant;
 
 ///
 /// A virtual file descriptor. You can use this just like std::fs::File, but internally
@@ -110,7 +111,7 @@ impl OpenFiles {
     ///
     /// On return, we hold a lock on the slot, and its 'tag' has been updated
     /// recently_used has been set. It's all ready for reuse.
-    fn find_victim_slot(&self) -> (SlotHandle, RwLockWriteGuard<SlotInner>) {
+    async fn find_victim_slot(&self) -> (SlotHandle, RwLockWriteGuard<SlotInner>) {
         //
         // Run the clock algorithm to find a slot to replace.
         //
@@ -142,7 +143,7 @@ impl OpenFiles {
                 }
                 retries += 1;
             } else {
-                slot_guard = slot.inner.write().unwrap();
+                slot_guard = slot.inner.write().await;
                 index = next;
                 break;
             }
@@ -153,7 +154,7 @@ impl OpenFiles {
         // old file.
         //
         if let Some(old_file) = slot_guard.file.take() {
-            // the normal path of dropping VirtualFile uses "close", use "close-by-replace" here to
+            // the normal path of dropping VirtualFile uses `Close`, use `CloseByReplace` here to
             // distinguish the two.
             STORAGE_IO_TIME_METRIC
                 .get(StorageIoOperation::CloseByReplace)
@@ -173,148 +174,78 @@ impl OpenFiles {
     }
 }
 
-/// Call this when the local filesystem gives us an error with an external
-/// cause: this includes EIO, EROFS, and EACCESS: all these indicate either
-/// bad storage or bad configuration, and we can't fix that from inside
-/// a running process.
-pub(crate) fn on_fatal_io_error(e: &std::io::Error) -> ! {
-    tracing::error!("Fatal I/O error: {}", &e);
-    std::process::abort();
+#[derive(Debug, thiserror::Error)]
+pub enum CrashsafeOverwriteError {
+    #[error("final path has no parent dir")]
+    FinalPathHasNoParentDir,
+    #[error("remove tempfile")]
+    RemovePreviousTempfile(#[source] std::io::Error),
+    #[error("create tempfile")]
+    CreateTempfile(#[source] std::io::Error),
+    #[error("write tempfile")]
+    WriteContents(#[source] std::io::Error),
+    #[error("sync tempfile")]
+    SyncTempfile(#[source] std::io::Error),
+    #[error("rename tempfile to final path")]
+    RenameTempfileToFinalPath(#[source] std::io::Error),
+    #[error("open final path parent dir")]
+    OpenFinalPathParentDir(#[source] std::io::Error),
+    #[error("sync final path parent dir")]
+    SyncFinalPathParentDir(#[source] std::io::Error),
 }
-
-/// Identify error types that should alwways terminate the process.  Other
-/// error types may be elegible for retry.
-pub(crate) fn is_fatal_io_error(e: &std::io::Error) -> bool {
-    use nix::errno::Errno::*;
-    match e.raw_os_error().map(nix::errno::from_i32) {
-        Some(EIO) => {
-            // Terminate on EIO because we no longer trust the device to store
-            // data safely, or to uphold persistence guarantees on fsync.
-            true
-        }
-        Some(EROFS) => {
-            // Terminate on EROFS because a filesystem is usually remounted
-            // readonly when it has experienced some critical issue, so the same
-            // logic as EIO applies.
-            true
-        }
-        Some(EACCES) => {
-            // Terminate on EACCESS because we should always have permissions
-            // for our own data dir: if we don't, then we can't do our job and
-            // need administrative intervention to fix permissions.  Terminating
-            // is the best way to make sure we stop cleanly rather than going
-            // into infinite retry loops, and will make it clear to the outside
-            // world that we need help.
-            true
-        }
-        _ => {
-            // Treat all other local file I/O errors are retryable.  This includes:
-            // - ENOSPC: we stay up and wait for eviction to free some space
-            // - EINVAL, EBADF, EBADFD: this is a code bug, not a filesystem/hardware issue
-            // - WriteZero, Interrupted: these are used internally VirtualFile
-            false
+impl CrashsafeOverwriteError {
+    /// Returns true iff the new contents are durably stored.
+    pub fn are_new_contents_durable(&self) -> bool {
+        match self {
+            Self::FinalPathHasNoParentDir => false,
+            Self::RemovePreviousTempfile(_) => false,
+            Self::CreateTempfile(_) => false,
+            Self::WriteContents(_) => false,
+            Self::SyncTempfile(_) => false,
+            Self::RenameTempfileToFinalPath(_) => false,
+            Self::OpenFinalPathParentDir(_) => false,
+            Self::SyncFinalPathParentDir(_) => true,
         }
     }
 }
 
-/// Wrap std::io::Error with a behavior where we will terminate the process
-/// on most I/O errors from local storage.  The rational for terminating is:
-/// - EIO means we can't trust the drive any more
-/// - EROFS means the local filesystem or drive is damaged, we shouldn't use it any more
-/// - EACCESS means something is fatally misconfigured about the pageserver, such
-///   as running the process as the wrong user, or the filesystem having the wrong
-///   ownership or permission bits.  We terminate so that it's obvious to
-///   the operator why the pageserver isn't working, and they can restart it when
-///   they've fixed the problem.
-#[derive(thiserror::Error, Debug)]
-pub struct Error {
-    inner: std::io::Error,
-    context: Option<String>,
+/// Observe duration for the given storage I/O operation
+///
+/// Unlike `observe_closure_duration`, this supports async,
+/// where "support" means that we measure wall clock time.
+macro_rules! observe_duration {
+    ($op:expr, $($body:tt)*) => {{
+        let instant = Instant::now();
+        let result = $($body)*;
+        let elapsed = instant.elapsed().as_secs_f64();
+        STORAGE_IO_TIME_METRIC
+            .get($op)
+            .observe(elapsed);
+        result
+    }}
 }
 
-impl Error {
-    /// Wrap a io::Error with some context & terminate
-    /// the process if the io::Error matches our policy for termination
-    fn new_with_context(e: std::io::Error, context: &str) -> Self {
-        Self::build(e, Some(context.to_string()))
-    }
-
-    fn context(e: Self, context: &str) -> Self {
-        Self {
-            inner: e.inner,
-            context: Some(context.to_string()),
-        }
-    }
-
-    fn new(e: std::io::Error) -> Self {
-        Self::build(e, None)
-    }
-
-    fn invalid(reason: &str) -> Self {
-        Self::new(std::io::Error::new(ErrorKind::InvalidInput, reason))
-    }
-
-    fn build(e: std::io::Error, context: Option<String>) -> Self {
-        // Construct instance early so that we have it for
-        // using Display in termination message.
-        let instance = Self { inner: e, context };
-
-        // Maybe terminate: this violates the usual expectation that callers
-        // should make their own decisions about how to handle an Error, but
-        // it's worthwhile to avoid every single user of the local filesystem
-        // having to apply the same "terminate on errors" behavior.
-        if is_fatal_io_error(&instance.inner) {
-            on_fatal_io_error(&instance.inner);
-        }
-
-        instance
-    }
-
-    fn kind(&self) -> ErrorKind {
-        self.inner.kind()
-    }
-}
-
-impl From<std::io::Error> for Error {
-    fn from(e: std::io::Error) -> Self {
-        Self::build(e, None)
-    }
-}
-
-impl From<Error> for std::io::Error {
-    fn from(e: Error) -> std::io::Error {
-        e.inner
-    }
-}
-
-impl std::fmt::Display for Error {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        match &self.context {
-            Some(context) => {
-                write!(f, "{}: {}", context, self.inner)
-            }
-            None => self.inner.fmt(f),
-        }
-    }
+macro_rules! with_file {
+    ($this:expr, $op:expr, | $ident:ident | $($body:tt)*) => {{
+        let $ident = $this.lock_file().await?;
+        observe_duration!($op, $($body)*)
+    }};
 }
 
 impl VirtualFile {
     /// Open a file in read-only mode. Like File::open.
-    pub async fn open(path: &Utf8Path) -> Result<VirtualFile, Error> {
-        Self::open_with_options(path, OpenOptions::new().read(true))
-            .await
-            .map_err(Error::from)
+    pub async fn open(path: &Utf8Path) -> Result<VirtualFile, std::io::Error> {
+        Self::open_with_options(path, OpenOptions::new().read(true)).await
     }
 
     /// Create a new file for writing. If the file exists, it will be truncated.
     /// Like File::create.
-    pub async fn create(path: &Utf8Path) -> Result<VirtualFile, Error> {
+    pub async fn create(path: &Utf8Path) -> Result<VirtualFile, std::io::Error> {
         Self::open_with_options(
             path,
             OpenOptions::new().write(true).create(true).truncate(true),
         )
         .await
-        .map_err(Error::from)
     }
 
     /// Open a file with given options.
@@ -325,7 +256,7 @@ impl VirtualFile {
     pub async fn open_with_options(
         path: &Utf8Path,
         open_options: &OpenOptions,
-    ) -> Result<VirtualFile, Error> {
+    ) -> Result<VirtualFile, std::io::Error> {
         let path_str = path.to_string();
         let parts = path_str.split('/').collect::<Vec<&str>>();
         let tenant_id;
@@ -337,11 +268,9 @@ impl VirtualFile {
             tenant_id = "*".to_string();
             timeline_id = "*".to_string();
         }
-        let (handle, mut slot_guard) = get_open_files().find_victim_slot();
+        let (handle, mut slot_guard) = get_open_files().find_victim_slot().await;
 
-        let file = STORAGE_IO_TIME_METRIC
-            .get(StorageIoOperation::Open)
-            .observe_closure_duration(|| open_options.open(path))?;
+        let file = observe_duration!(StorageIoOperation::Open, open_options.open(path))?;
 
         // Strip all options other than read and write.
         //
@@ -377,16 +306,14 @@ impl VirtualFile {
         final_path: &Utf8Path,
         tmp_path: &Utf8Path,
         content: &[u8],
-    ) -> Result<(), Error> {
-        let final_path_parent = final_path.parent().ok_or(std::io::Error::new(
-            ErrorKind::InvalidInput,
-            "Path must be absolute",
-        ))?;
-
+    ) -> Result<(), CrashsafeOverwriteError> {
+        let Some(final_path_parent) = final_path.parent() else {
+            return Err(CrashsafeOverwriteError::FinalPathHasNoParentDir);
+        };
         match std::fs::remove_file(tmp_path) {
             Ok(()) => {}
             Err(e) if e.kind() == std::io::ErrorKind::NotFound => {}
-            Err(e) => return Err(Error::new_with_context(e, "removing tempfile")),
+            Err(e) => return Err(CrashsafeOverwriteError::RemovePreviousTempfile(e)),
         }
         let mut file = Self::open_with_options(
             tmp_path,
@@ -397,17 +324,17 @@ impl VirtualFile {
                 .create_new(true),
         )
         .await
-        .map_err(|e| Error::context(e, "create tempfile"))?;
+        .map_err(CrashsafeOverwriteError::CreateTempfile)?;
         file.write_all(content)
             .await
-            .map_err(|e| Error::context(e, "write contents"))?;
+            .map_err(CrashsafeOverwriteError::WriteContents)?;
         file.sync_all()
             .await
-            .map_err(|e| Error::context(e, "sync tempfile"))?;
+            .map_err(CrashsafeOverwriteError::SyncTempfile)?;
         drop(file); // before the rename, that's important!
                     // renames are atomic
         std::fs::rename(tmp_path, final_path)
-            .map_err(|e| Error::new_with_context(e, "rename tempfile to final path"))?;
+            .map_err(CrashsafeOverwriteError::RenameTempfileToFinalPath)?;
         // Only open final path parent dirfd now, so that this operation only
         // ever holds one VirtualFile fd at a time.  That's important because
         // the current `find_victim_slot` impl might pick the same slot for both
@@ -416,34 +343,34 @@ impl VirtualFile {
         let final_parent_dirfd =
             Self::open_with_options(final_path_parent, OpenOptions::new().read(true))
                 .await
-                .map_err(|e| Error::context(e, "open final path parent"))?;
+                .map_err(CrashsafeOverwriteError::OpenFinalPathParentDir)?;
         final_parent_dirfd
             .sync_all()
             .await
-            .map_err(|e| Error::context(e, "sync final path parent"))?;
+            .map_err(CrashsafeOverwriteError::SyncFinalPathParentDir)?;
         Ok(())
     }
 
     /// Call File::sync_all() on the underlying File.
     pub async fn sync_all(&self) -> Result<(), Error> {
-        self.with_file(StorageIoOperation::Fsync, |file| file.sync_all())
-            .await?
-            .map_err(Error::new)
+        with_file!(self, StorageIoOperation::Fsync, |file| file
+            .as_ref()
+            .sync_all())
     }
 
     pub async fn metadata(&self) -> Result<fs::Metadata, Error> {
-        self.with_file(StorageIoOperation::Metadata, |file| file.metadata())
-            .await?
-            .map_err(Error::new)
+        with_file!(self, StorageIoOperation::Metadata, |file| file
+            .as_ref()
+            .metadata())
     }
 
-    /// Helper function that looks up the underlying File for this VirtualFile,
-    /// opening it and evicting some other File if necessary. It calls 'func'
-    /// with the physical File.
-    async fn with_file<F, R>(&self, op: StorageIoOperation, mut func: F) -> Result<R, Error>
-    where
-        F: FnMut(&File) -> R,
-    {
+    /// Helper function internal to `VirtualFile` that looks up the underlying File,
+    /// opens it and evicts some other File if necessary. The passed parameter is
+    /// assumed to be a function available for the physical `File`.
+    ///
+    /// We are doing it via a macro as Rust doesn't support async closures that
+    /// take on parameters with lifetimes.
+    async fn lock_file(&self) -> Result<FileGuard<'_>, Error> {
         let open_files = get_open_files();
 
         let mut handle_guard = {
@@ -453,27 +380,23 @@ impl VirtualFile {
             // We only need to hold the handle lock while we read the current handle. If
             // another thread closes the file and recycles the slot for a different file,
             // we will notice that the handle we read is no longer valid and retry.
-            let mut handle = *self.handle.read().unwrap();
+            let mut handle = *self.handle.read().await;
             loop {
                 // Check if the slot contains our File
                 {
                     let slot = &open_files.slots[handle.index];
-                    let slot_guard = slot.inner.read().unwrap();
-                    if slot_guard.tag == handle.tag {
-                        if let Some(file) = &slot_guard.file {
-                            // Found a cached file descriptor.
-                            slot.recently_used.store(true, Ordering::Relaxed);
-                            return Ok(STORAGE_IO_TIME_METRIC
-                                .get(op)
-                                .observe_closure_duration(|| func(file)));
-                        }
+                    let slot_guard = slot.inner.read().await;
+                    if slot_guard.tag == handle.tag && slot_guard.file.is_some() {
+                        // Found a cached file descriptor.
+                        slot.recently_used.store(true, Ordering::Relaxed);
+                        return Ok(FileGuard { slot_guard });
                     }
                 }
 
                 // The slot didn't contain our File. We will have to open it ourselves,
                 // but before that, grab a write lock on handle in the VirtualFile, so
                 // that no other thread will try to concurrently open the same file.
-                let handle_guard = self.handle.write().unwrap();
+                let handle_guard = self.handle.write().await;
 
                 // If another thread changed the handle while we were not holding the lock,
                 // then the handle might now be valid again. Loop back to retry.
@@ -487,17 +410,10 @@ impl VirtualFile {
 
         // We need to open the file ourselves. The handle in the VirtualFile is
         // now locked in write-mode. Find a free slot to put it in.
-        let (handle, mut slot_guard) = open_files.find_victim_slot();
+        let (handle, mut slot_guard) = open_files.find_victim_slot().await;
 
         // Open the physical file
-        let file = STORAGE_IO_TIME_METRIC
-            .get(StorageIoOperation::Open)
-            .observe_closure_duration(|| self.open_options.open(&self.path))?;
-
-        // Perform the requested operation on it
-        let result = STORAGE_IO_TIME_METRIC
-            .get(op)
-            .observe_closure_duration(|| func(&file));
+        let file = observe_duration!(StorageIoOperation::Open, self.open_options.open(&self.path))?;
 
         // Store the File in the slot and update the handle in the VirtualFile
         // to point to it.
@@ -505,7 +421,9 @@ impl VirtualFile {
 
         *handle_guard = handle;
 
-        Ok(result)
+        return Ok(FileGuard {
+            slot_guard: slot_guard.downgrade(),
+        });
     }
 
     pub fn remove(self) {
@@ -520,19 +438,20 @@ impl VirtualFile {
                 self.pos = offset;
             }
             SeekFrom::End(offset) => {
-                self.pos = self
-                    .with_file(StorageIoOperation::Seek, |mut file| {
-                        file.seek(SeekFrom::End(offset))
-                    })
-                    .await??
+                self.pos = with_file!(self, StorageIoOperation::Seek, |file| file
+                    .as_ref()
+                    .seek(SeekFrom::End(offset)))?
             }
             SeekFrom::Current(offset) => {
                 let pos = self.pos as i128 + offset as i128;
                 if pos < 0 {
-                    return Err(Error::invalid("offset would be negative"));
+                    return Err(Error::new(
+                        ErrorKind::InvalidInput,
+                        "offset would be negative",
+                    ));
                 }
                 if pos > u64::MAX as i128 {
-                    return Err(Error::invalid("offset overflow"));
+                    return Err(Error::new(ErrorKind::InvalidInput, "offset overflow"));
                 }
                 self.pos = pos as u64;
             }
@@ -545,11 +464,10 @@ impl VirtualFile {
         while !buf.is_empty() {
             match self.read_at(buf, offset).await {
                 Ok(0) => {
-                    return Err(std::io::Error::new(
+                    return Err(Error::new(
                         std::io::ErrorKind::UnexpectedEof,
                         "failed to fill whole buffer",
-                    )
-                    .into())
+                    ))
                 }
                 Ok(n) => {
                     buf = &mut buf[n..];
@@ -567,11 +485,10 @@ impl VirtualFile {
         while !buf.is_empty() {
             match self.write_at(buf, offset).await {
                 Ok(0) => {
-                    return Err(std::io::Error::new(
+                    return Err(Error::new(
                         std::io::ErrorKind::WriteZero,
                         "failed to write whole buffer",
-                    )
-                    .into());
+                    ));
                 }
                 Ok(n) => {
                     buf = &buf[n..];
@@ -588,11 +505,10 @@ impl VirtualFile {
         while !buf.is_empty() {
             match self.write(buf).await {
                 Ok(0) => {
-                    return Err(std::io::Error::new(
+                    return Err(Error::new(
                         std::io::ErrorKind::WriteZero,
                         "failed to write whole buffer",
-                    )
-                    .into());
+                    ));
                 }
                 Ok(n) => {
                     buf = &buf[n..];
@@ -604,7 +520,7 @@ impl VirtualFile {
         Ok(())
     }
 
-    async fn write(&mut self, buf: &[u8]) -> Result<usize, Error> {
+    async fn write(&mut self, buf: &[u8]) -> Result<usize, std::io::Error> {
         let pos = self.pos;
         let n = self.write_at(buf, pos).await?;
         self.pos += n as u64;
@@ -612,27 +528,39 @@ impl VirtualFile {
     }
 
     pub async fn read_at(&self, buf: &mut [u8], offset: u64) -> Result<usize, Error> {
-        let result = self
-            .with_file(StorageIoOperation::Read, |file| file.read_at(buf, offset))
-            .await?;
+        let result = with_file!(self, StorageIoOperation::Read, |file| file
+            .as_ref()
+            .read_at(buf, offset));
         if let Ok(size) = result {
             STORAGE_IO_SIZE
                 .with_label_values(&["read", &self.tenant_id, &self.timeline_id])
                 .add(size as i64);
         }
-        result.map_err(Error::new)
+        result
     }
 
     async fn write_at(&self, buf: &[u8], offset: u64) -> Result<usize, Error> {
-        let result = self
-            .with_file(StorageIoOperation::Write, |file| file.write_at(buf, offset))
-            .await?;
+        let result = with_file!(self, StorageIoOperation::Write, |file| file
+            .as_ref()
+            .write_at(buf, offset));
         if let Ok(size) = result {
             STORAGE_IO_SIZE
                 .with_label_values(&["write", &self.tenant_id, &self.timeline_id])
                 .add(size as i64);
         }
-        result.map_err(Error::new)
+        result
+    }
+}
+
+struct FileGuard<'a> {
+    slot_guard: RwLockReadGuard<'a, SlotInner>,
+}
+
+impl<'a> AsRef<File> for FileGuard<'a> {
+    fn as_ref(&self) -> &File {
+        // This unwrap is safe because we only create `FileGuard`s
+        // if we know that the file is Some.
+        self.slot_guard.file.as_ref().unwrap()
     }
 }
 
@@ -641,7 +569,7 @@ impl VirtualFile {
     pub(crate) async fn read_blk(
         &self,
         blknum: u32,
-    ) -> Result<crate::tenant::block_io::BlockLease<'_>, Error> {
+    ) -> Result<crate::tenant::block_io::BlockLease<'_>, std::io::Error> {
         use crate::page_cache::PAGE_SZ;
         let mut buf = [0; PAGE_SZ];
         self.read_exact_at(&mut buf, blknum as u64 * (PAGE_SZ as u64))
@@ -668,20 +596,39 @@ impl VirtualFile {
 impl Drop for VirtualFile {
     /// If a VirtualFile is dropped, close the underlying file if it was open.
     fn drop(&mut self) {
-        let handle = self.handle.get_mut().unwrap();
+        let handle = self.handle.get_mut();
 
-        // We could check with a read-lock first, to avoid waiting on an
-        // unrelated I/O.
-        let slot = &get_open_files().slots[handle.index];
-        let mut slot_guard = slot.inner.write().unwrap();
-        if slot_guard.tag == handle.tag {
-            slot.recently_used.store(false, Ordering::Relaxed);
-            // there is also operation "close-by-replace" for closes done on eviction for
-            // comparison.
-            STORAGE_IO_TIME_METRIC
-                .get(StorageIoOperation::Close)
-                .observe_closure_duration(|| drop(slot_guard.file.take()));
+        fn clean_slot(slot: &Slot, mut slot_guard: RwLockWriteGuard<'_, SlotInner>, tag: u64) {
+            if slot_guard.tag == tag {
+                slot.recently_used.store(false, Ordering::Relaxed);
+                // there is also the `CloseByReplace` operation for closes done on eviction for
+                // comparison.
+                STORAGE_IO_TIME_METRIC
+                    .get(StorageIoOperation::Close)
+                    .observe_closure_duration(|| drop(slot_guard.file.take()));
+            }
         }
+
+        // We don't have async drop so we cannot directly await the lock here.
+        // Instead, first do a best-effort attempt at closing the underlying
+        // file descriptor by using `try_write`, and if that fails, spawn
+        // a tokio task to do it asynchronously: we just want it to be
+        // cleaned up eventually.
+        // Most of the time, the `try_lock` should succeed though,
+        // as we have `&mut self` access. In other words, if the slot
+        // is still occupied by our file, there should be no access from
+        // other I/O operations; the only other possible place to lock
+        // the slot is the lock algorithm looking for free slots.
+        let slot = &get_open_files().slots[handle.index];
+        if let Ok(slot_guard) = slot.inner.try_write() {
+            clean_slot(slot, slot_guard, handle.tag);
+        } else {
+            let tag = handle.tag;
+            tokio::spawn(async move {
+                let slot_guard = slot.inner.write().await;
+                clean_slot(slot, slot_guard, tag);
+            });
+        };
     }
 }
 
@@ -757,25 +704,25 @@ mod tests {
         async fn read_exact_at(&self, buf: &mut [u8], offset: u64) -> Result<(), Error> {
             match self {
                 MaybeVirtualFile::VirtualFile(file) => file.read_exact_at(buf, offset).await,
-                MaybeVirtualFile::File(file) => file.read_exact_at(buf, offset).map_err(Error::new),
+                MaybeVirtualFile::File(file) => file.read_exact_at(buf, offset),
             }
         }
         async fn write_all_at(&self, buf: &[u8], offset: u64) -> Result<(), Error> {
             match self {
                 MaybeVirtualFile::VirtualFile(file) => file.write_all_at(buf, offset).await,
-                MaybeVirtualFile::File(file) => file.write_all_at(buf, offset).map_err(Error::new),
+                MaybeVirtualFile::File(file) => file.write_all_at(buf, offset),
             }
         }
         async fn seek(&mut self, pos: SeekFrom) -> Result<u64, Error> {
             match self {
                 MaybeVirtualFile::VirtualFile(file) => file.seek(pos).await,
-                MaybeVirtualFile::File(file) => file.seek(pos).map_err(Error::new),
+                MaybeVirtualFile::File(file) => file.seek(pos),
             }
         }
         async fn write_all(&mut self, buf: &[u8]) -> Result<(), Error> {
             match self {
                 MaybeVirtualFile::VirtualFile(file) => file.write_all(buf).await,
-                MaybeVirtualFile::File(file) => file.write_all(buf).map_err(Error::new),
+                MaybeVirtualFile::File(file) => file.write_all(buf),
             }
         }
 
@@ -984,7 +931,7 @@ mod tests {
             hdls.push(hdl);
         }
         for hdl in hdls {
-            hdl.await.expect("joining")
+            hdl.await?;
         }
         std::mem::forget(rt);
 

pageserver/src/walingest.rs

@@ -338,11 +338,20 @@ impl<'a> WalIngest<'a> {
         } else if decoded.xl_rmid == pg_constants::RM_LOGICALMSG_ID {
             let info = decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK;
             if info == pg_constants::XLOG_LOGICAL_MESSAGE {
-                // This is a convenient way to make the WAL ingestion pause at
-                // particular point in the WAL. For more fine-grained control,
-                // we could peek into the message and only pause if it contains
-                // a particular string, for example, but this is enough for now.
-                crate::failpoint_support::sleep_millis_async!("wal-ingest-logical-message-sleep");
+                let xlrec = XlLogicalMessage::decode(&mut buf);
+                let prefix = std::str::from_utf8(&buf[0..xlrec.prefix_size - 1])?;
+                let message = &buf[xlrec.prefix_size..xlrec.prefix_size + xlrec.message_size];
+                if prefix == "neon-test" {
+                    // This is a convenient way to make the WAL ingestion pause at
+                    // particular point in the WAL. For more fine-grained control,
+                    // we could peek into the message and only pause if it contains
+                    // a particular string, for example, but this is enough for now.
+                    crate::failpoint_support::sleep_millis_async!(
+                        "wal-ingest-logical-message-sleep"
+                    );
+                } else if let Some(path) = prefix.strip_prefix("neon-file:") {
+                    modification.put_file(path, message, ctx).await?;
+                }
             }
         }
 
@@ -459,7 +468,6 @@ impl<'a> WalIngest<'a> {
                         }
                     } else if info == pg_constants::XLOG_HEAP_DELETE {
                         let xlrec = v14::XlHeapDelete::decode(buf);
-                        assert_eq!(0, buf.remaining());
                         if (xlrec.flags & pg_constants::XLH_DELETE_ALL_VISIBLE_CLEARED) != 0 {
                             new_heap_blkno = Some(decoded.blocks[0].blkno);
                         }
@@ -527,7 +535,6 @@ impl<'a> WalIngest<'a> {
                         }
                     } else if info == pg_constants::XLOG_HEAP_DELETE {
                         let xlrec = v15::XlHeapDelete::decode(buf);
-                        assert_eq!(0, buf.remaining());
                         if (xlrec.flags & pg_constants::XLH_DELETE_ALL_VISIBLE_CLEARED) != 0 {
                             new_heap_blkno = Some(decoded.blocks[0].blkno);
                         }
@@ -595,7 +602,6 @@ impl<'a> WalIngest<'a> {
                         }
                     } else if info == pg_constants::XLOG_HEAP_DELETE {
                         let xlrec = v16::XlHeapDelete::decode(buf);
-                        assert_eq!(0, buf.remaining());
                         if (xlrec.flags & pg_constants::XLH_DELETE_ALL_VISIBLE_CLEARED) != 0 {
                             new_heap_blkno = Some(decoded.blocks[0].blkno);
                         }
@@ -771,7 +777,6 @@ impl<'a> WalIngest<'a> {
                     }
                     pg_constants::XLOG_NEON_HEAP_DELETE => {
                         let xlrec = v16::rm_neon::XlNeonHeapDelete::decode(buf);
-                        assert_eq!(0, buf.remaining());
                         if (xlrec.flags & pg_constants::XLH_DELETE_ALL_VISIBLE_CLEARED) != 0 {
                             new_heap_blkno = Some(decoded.blocks[0].blkno);
                         }

pageserver/src/walrecord.rs

@@ -748,6 +748,26 @@ impl XlMultiXactTruncate {
     }
 }
 
+#[repr(C)]
+#[derive(Debug)]
+pub struct XlLogicalMessage {
+    pub db_id: Oid,
+    pub transactional: bool,
+    pub prefix_size: usize,
+    pub message_size: usize,
+}
+
+impl XlLogicalMessage {
+    pub fn decode(buf: &mut Bytes) -> XlLogicalMessage {
+        XlLogicalMessage {
+            db_id: buf.get_u32_le(),
+            transactional: buf.get_u32_le() != 0, // 4-bytes alignment
+            prefix_size: buf.get_u64_le() as usize,
+            message_size: buf.get_u64_le() as usize,
+        }
+    }
+}
+
 /// Main routine to decode a WAL record and figure out which blocks are modified
 //
 // See xlogrecord.h for details

pageserver/src/walredo.rs

@@ -18,38 +18,37 @@
 //! any WAL records, so that even if an attacker hijacks the Postgres
 //! process, he cannot escape out of it.
 //!
+use anyhow::Context;
 use byteorder::{ByteOrder, LittleEndian};
 use bytes::{BufMut, Bytes, BytesMut};
 use nix::poll::*;
 use serde::Serialize;
 use std::collections::VecDeque;
+use std::io;
 use std::io::prelude::*;
-use std::io::{Error, ErrorKind};
 use std::ops::{Deref, DerefMut};
-use std::os::unix::io::{AsRawFd, RawFd};
+use std::os::unix::io::AsRawFd;
 use std::os::unix::prelude::CommandExt;
 use std::process::Stdio;
-use std::process::{Child, ChildStderr, ChildStdin, ChildStdout, Command};
-use std::sync::{Mutex, MutexGuard};
+use std::process::{Child, ChildStdin, ChildStdout, Command};
+use std::sync::{Arc, Mutex, MutexGuard, RwLock};
 use std::time::Duration;
 use std::time::Instant;
-use std::{fs, io};
+use tokio_util::sync::CancellationToken;
 use tracing::*;
-use utils::crashsafe::path_with_suffix_extension;
 use utils::{bin_ser::BeSer, id::TenantId, lsn::Lsn, nonblock::set_nonblock};
 
 #[cfg(feature = "testing")]
 use std::sync::atomic::{AtomicUsize, Ordering};
 
+use crate::config::PageServerConf;
 use crate::metrics::{
     WAL_REDO_BYTES_HISTOGRAM, WAL_REDO_RECORDS_HISTOGRAM, WAL_REDO_RECORD_COUNTER, WAL_REDO_TIME,
     WAL_REDO_WAIT_TIME,
 };
 use crate::pgdatadir_mapping::{key_to_rel_block, key_to_slru_block};
 use crate::repository::Key;
-use crate::task_mgr::BACKGROUND_RUNTIME;
 use crate::walrecord::NeonWalRecord;
-use crate::{config::PageServerConf, TEMP_FILE_SUFFIX};
 use pageserver_api::reltag::{RelTag, SlruKind};
 use postgres_ffi::pg_constants;
 use postgres_ffi::relfile_utils::VISIBILITYMAP_FORKNUM;
@@ -66,37 +65,13 @@ use postgres_ffi::BLCKSZ;
 /// [See more related comments here](https://github.com/postgres/postgres/blob/99c5852e20a0987eca1c38ba0c09329d4076b6a0/src/include/storage/buf_internals.h#L91).
 ///
 #[derive(Debug, PartialEq, Eq, PartialOrd, Ord, Clone, Copy, Serialize)]
-pub struct BufferTag {
+pub(crate) struct BufferTag {
     pub rel: RelTag,
     pub blknum: u32,
 }
 
-///
-/// WAL Redo Manager is responsible for replaying WAL records.
-///
-/// Callers use the WAL redo manager through this abstract interface,
-/// which makes it easy to mock it in tests.
-pub trait WalRedoManager: Send + Sync {
-    /// Apply some WAL records.
-    ///
-    /// The caller passes an old page image, and WAL records that should be
-    /// applied over it. The return value is a new page image, after applying
-    /// the reords.
-    fn request_redo(
-        &self,
-        key: Key,
-        lsn: Lsn,
-        base_img: Option<(Lsn, Bytes)>,
-        records: Vec<(Lsn, NeonWalRecord)>,
-        pg_version: u32,
-    ) -> Result<Bytes, WalRedoError>;
-}
-
 struct ProcessInput {
-    child: NoLeakChild,
     stdin: ChildStdin,
-    stderr_fd: RawFd,
-    stdout_fd: RawFd,
     n_requests: usize,
 }
 
@@ -116,13 +91,7 @@ struct ProcessOutput {
 pub struct PostgresRedoManager {
     tenant_id: TenantId,
     conf: &'static PageServerConf,
-    /// Counter to separate same sized walredo inputs failing at the same millisecond.
-    #[cfg(feature = "testing")]
-    dump_sequence: AtomicUsize,
-
-    stdout: Mutex<Option<ProcessOutput>>,
-    stdin: Mutex<Option<ProcessInput>>,
-    stderr: Mutex<Option<ChildStderr>>,
+    redo_process: RwLock<Option<Arc<WalRedoProcess>>>,
 }
 
 /// Can this request be served by neon redo functions
@@ -140,41 +109,27 @@ fn can_apply_in_neon(rec: &NeonWalRecord) -> bool {
     }
 }
 
-/// An error happened in WAL redo
-#[derive(Debug, thiserror::Error)]
-pub enum WalRedoError {
-    #[error(transparent)]
-    IoError(#[from] std::io::Error),
-
-    #[error("cannot perform WAL redo now")]
-    InvalidState,
-    #[error("cannot perform WAL redo for this request")]
-    InvalidRequest,
-    #[error("cannot perform WAL redo for this record")]
-    InvalidRecord,
-}
-
 ///
 /// Public interface of WAL redo manager
 ///
-impl WalRedoManager for PostgresRedoManager {
+impl PostgresRedoManager {
     ///
     /// Request the WAL redo manager to apply some WAL records
     ///
     /// The WAL redo is handled by a separate thread, so this just sends a request
     /// to the thread and waits for response.
     ///
-    fn request_redo(
+    /// CANCEL SAFETY: NOT CANCEL SAFE.
+    pub async fn request_redo(
         &self,
         key: Key,
         lsn: Lsn,
         base_img: Option<(Lsn, Bytes)>,
         records: Vec<(Lsn, NeonWalRecord)>,
         pg_version: u32,
-    ) -> Result<Bytes, WalRedoError> {
+    ) -> anyhow::Result<Bytes> {
         if records.is_empty() {
-            error!("invalid WAL redo request with no records");
-            return Err(WalRedoError::InvalidRequest);
+            anyhow::bail!("invalid WAL redo request with no records");
         }
 
         let base_img_lsn = base_img.as_ref().map(|p| p.0).unwrap_or(Lsn::INVALID);
@@ -197,6 +152,7 @@ impl WalRedoManager for PostgresRedoManager {
                         self.conf.wal_redo_timeout,
                         pg_version,
                     )
+                    .await
                 };
                 img = Some(result?);
 
@@ -217,6 +173,7 @@ impl WalRedoManager for PostgresRedoManager {
                 self.conf.wal_redo_timeout,
                 pg_version,
             )
+            .await
         }
     }
 }
@@ -230,28 +187,15 @@ impl PostgresRedoManager {
         PostgresRedoManager {
             tenant_id,
             conf,
-            #[cfg(feature = "testing")]
-            dump_sequence: AtomicUsize::default(),
-            stdin: Mutex::new(None),
-            stdout: Mutex::new(None),
-            stderr: Mutex::new(None),
+            redo_process: RwLock::new(None),
         }
     }
 
-    /// Launch process pre-emptively. Should not be needed except for benchmarking.
-    pub fn launch_process(&self, pg_version: u32) -> anyhow::Result<()> {
-        let mut proc = self.stdin.lock().unwrap();
-        if proc.is_none() {
-            self.launch(&mut proc, pg_version)?;
-        }
-        Ok(())
-    }
-
     ///
     /// Process one request for WAL redo using wal-redo postgres
     ///
     #[allow(clippy::too_many_arguments)]
-    fn apply_batch_postgres(
+    async fn apply_batch_postgres(
         &self,
         key: Key,
         lsn: Lsn,
@@ -260,26 +204,45 @@ impl PostgresRedoManager {
         records: &[(Lsn, NeonWalRecord)],
         wal_redo_timeout: Duration,
         pg_version: u32,
-    ) -> Result<Bytes, WalRedoError> {
-        let (rel, blknum) = key_to_rel_block(key).or(Err(WalRedoError::InvalidRecord))?;
+    ) -> anyhow::Result<Bytes> {
+        let (rel, blknum) = key_to_rel_block(key).context("invalid record")?;
         const MAX_RETRY_ATTEMPTS: u32 = 1;
         let start_time = Instant::now();
         let mut n_attempts = 0u32;
         loop {
-            let mut proc = self.stdin.lock().unwrap();
             let lock_time = Instant::now();
 
             // launch the WAL redo process on first use
-            if proc.is_none() {
-                self.launch(&mut proc, pg_version)?;
-            }
+            let proc: Arc<WalRedoProcess> = {
+                let proc_guard = self.redo_process.read().unwrap();
+                match &*proc_guard {
+                    None => {
+                        // "upgrade" to write lock to launch the process
+                        drop(proc_guard);
+                        let mut proc_guard = self.redo_process.write().unwrap();
+                        match &*proc_guard {
+                            None => {
+                                let proc = Arc::new(
+                                    WalRedoProcess::launch(self.conf, self.tenant_id, pg_version)
+                                        .context("launch walredo process")?,
+                                );
+                                *proc_guard = Some(Arc::clone(&proc));
+                                proc
+                            }
+                            Some(proc) => Arc::clone(proc),
+                        }
+                    }
+                    Some(proc) => Arc::clone(proc),
+                }
+            };
+
             WAL_REDO_WAIT_TIME.observe(lock_time.duration_since(start_time).as_secs_f64());
 
             // Relational WAL records are applied using wal-redo-postgres
             let buf_tag = BufferTag { rel, blknum };
-            let result = self
-                .apply_wal_records(proc, buf_tag, &base_img, records, wal_redo_timeout)
-                .map_err(WalRedoError::IoError);
+            let result = proc
+                .apply_wal_records(buf_tag, &base_img, records, wal_redo_timeout)
+                .context("apply_wal_records");
 
             let end_time = Instant::now();
             let duration = end_time.duration_since(lock_time);
@@ -309,32 +272,50 @@ impl PostgresRedoManager {
             // next request will launch a new one.
             if let Err(e) = result.as_ref() {
                 error!(
-                    n_attempts,
-                    "error applying {} WAL records {}..{} ({} bytes) to base image with LSN {} to reconstruct page image at LSN {}: {}",
+                    "error applying {} WAL records {}..{} ({} bytes) to base image with LSN {} to reconstruct page image at LSN {} n_attempts={}: {:?}",
                     records.len(),
                     records.first().map(|p| p.0).unwrap_or(Lsn(0)),
                     records.last().map(|p| p.0).unwrap_or(Lsn(0)),
                     nbytes,
                     base_img_lsn,
                     lsn,
-                    utils::error::report_compact_sources(e),
+                    n_attempts,
+                    e,
                 );
-                // self.stdin only holds stdin & stderr as_raw_fd().
-                // Dropping it as part of take() doesn't close them.
-                // The owning objects (ChildStdout and ChildStderr) are stored in
-                // self.stdout and self.stderr, respsectively.
-                // We intentionally keep them open here to avoid a race between
-                // currently running `apply_wal_records()` and a `launch()` call
-                // after we return here.
-                // The currently running `apply_wal_records()` must not read from
-                // the newly launched process.
-                // By keeping self.stdout and self.stderr open here, `launch()` will
-                // get other file descriptors for the new child's stdout and stderr,
-                // and hence the current `apply_wal_records()` calls will observe
-                //  `output.stdout.as_raw_fd() != stdout_fd` .
-                if let Some(proc) = self.stdin.lock().unwrap().take() {
-                    proc.child.kill_and_wait();
+                // Avoid concurrent callers hitting the same issue.
+                // We can't prevent it from happening because we want to enable parallelism.
+                {
+                    let mut guard = self.redo_process.write().unwrap();
+                    match &*guard {
+                        Some(current_field_value) => {
+                            if Arc::ptr_eq(current_field_value, &proc) {
+                                // We're the first to observe an error from `proc`, it's our job to take it out of rotation.
+                                *guard = None;
+                            }
+                        }
+                        None => {
+                            // Another thread was faster to observe the error, and already took the process out of rotation.
+                        }
+                    }
                 }
+                // NB: there may still be other concurrent threads using `proc`.
+                // The last one will send SIGKILL when the underlying Arc reaches refcount 0.
+                // NB: it's important to drop(proc) after drop(guard). Otherwise we'd keep
+                // holding the lock while waiting for the process to exit.
+                // NB: the drop impl blocks the current threads with a wait() system call for
+                // the child process. We dropped the `guard` above so that other threads aren't
+                // affected. But, it's good that the current thread _does_ block to wait.
+                // If we instead deferred the waiting into the background / to tokio, it could
+                // happen that if walredo always fails immediately, we spawn processes faster
+                // than we can SIGKILL & `wait` for them to exit. By doing it the way we do here,
+                // we limit this risk of run-away to at most $num_runtimes * $num_executor_threads.
+                // This probably needs revisiting at some later point.
+                let mut wait_done = proc.stderr_logger_task_done.clone();
+                drop(proc);
+                wait_done
+                    .wait_for(|v| *v)
+                    .await
+                    .expect("we use scopeguard to ensure we always send `true` to the channel before dropping the sender");
             } else if n_attempts != 0 {
                 info!(n_attempts, "retried walredo succeeded");
             }
@@ -354,7 +335,7 @@ impl PostgresRedoManager {
         lsn: Lsn,
         base_img: Option<Bytes>,
         records: &[(Lsn, NeonWalRecord)],
-    ) -> Result<Bytes, WalRedoError> {
+    ) -> anyhow::Result<Bytes> {
         let start_time = Instant::now();
 
         let mut page = BytesMut::new();
@@ -363,8 +344,7 @@ impl PostgresRedoManager {
             page.extend_from_slice(&fpi[..]);
         } else {
             // All the current WAL record types that we can handle require a base image.
-            error!("invalid neon WAL redo request with no base image");
-            return Err(WalRedoError::InvalidRequest);
+            anyhow::bail!("invalid neon WAL redo request with no base image");
         }
 
         // Apply all the WAL records in the batch
@@ -392,14 +372,13 @@ impl PostgresRedoManager {
         page: &mut BytesMut,
         _record_lsn: Lsn,
         record: &NeonWalRecord,
-    ) -> Result<(), WalRedoError> {
+    ) -> anyhow::Result<()> {
         match record {
             NeonWalRecord::Postgres {
                 will_init: _,
                 rec: _,
             } => {
-                error!("tried to pass postgres wal record to neon WAL redo");
-                return Err(WalRedoError::InvalidRequest);
+                anyhow::bail!("tried to pass postgres wal record to neon WAL redo");
             }
             NeonWalRecord::ClearVisibilityMapFlags {
                 new_heap_blkno,
@@ -407,7 +386,7 @@ impl PostgresRedoManager {
                 flags,
             } => {
                 // sanity check that this is modifying the correct relation
-                let (rel, blknum) = key_to_rel_block(key).or(Err(WalRedoError::InvalidRecord))?;
+                let (rel, blknum) = key_to_rel_block(key).context("invalid record")?;
                 assert!(
                     rel.forknum == VISIBILITYMAP_FORKNUM,
                     "ClearVisibilityMapFlags record on unexpected rel {}",
@@ -445,7 +424,7 @@ impl PostgresRedoManager {
             // same effects as the corresponding Postgres WAL redo function.
             NeonWalRecord::ClogSetCommitted { xids, timestamp } => {
                 let (slru_kind, segno, blknum) =
-                    key_to_slru_block(key).or(Err(WalRedoError::InvalidRecord))?;
+                    key_to_slru_block(key).context("invalid record")?;
                 assert_eq!(
                     slru_kind,
                     SlruKind::Clog,
@@ -495,7 +474,7 @@ impl PostgresRedoManager {
             }
             NeonWalRecord::ClogSetAborted { xids } => {
                 let (slru_kind, segno, blknum) =
-                    key_to_slru_block(key).or(Err(WalRedoError::InvalidRecord))?;
+                    key_to_slru_block(key).context("invalid record")?;
                 assert_eq!(
                     slru_kind,
                     SlruKind::Clog,
@@ -526,7 +505,7 @@ impl PostgresRedoManager {
             }
             NeonWalRecord::MultixactOffsetCreate { mid, moff } => {
                 let (slru_kind, segno, blknum) =
-                    key_to_slru_block(key).or(Err(WalRedoError::InvalidRecord))?;
+                    key_to_slru_block(key).context("invalid record")?;
                 assert_eq!(
                     slru_kind,
                     SlruKind::MultiXactOffsets,
@@ -559,7 +538,7 @@ impl PostgresRedoManager {
             }
             NeonWalRecord::MultixactMembersCreate { moff, members } => {
                 let (slru_kind, segno, blknum) =
-                    key_to_slru_block(key).or(Err(WalRedoError::InvalidRecord))?;
+                    key_to_slru_block(key).context("invalid record")?;
                 assert_eq!(
                     slru_kind,
                     SlruKind::MultiXactMembers,
@@ -639,44 +618,33 @@ impl<C: CommandExt> CloseFileDescriptors for C {
     }
 }
 
-impl PostgresRedoManager {
+struct WalRedoProcess {
+    #[allow(dead_code)]
+    conf: &'static PageServerConf,
+    tenant_id: TenantId,
+    // Some() on construction, only becomes None on Drop.
+    child: Option<NoLeakChild>,
+    stdout: Mutex<ProcessOutput>,
+    stdin: Mutex<ProcessInput>,
+    stderr_logger_cancel: CancellationToken,
+    stderr_logger_task_done: tokio::sync::watch::Receiver<bool>,
+    /// Counter to separate same sized walredo inputs failing at the same millisecond.
+    #[cfg(feature = "testing")]
+    dump_sequence: AtomicUsize,
+}
+
+impl WalRedoProcess {
     //
     // Start postgres binary in special WAL redo mode.
     //
-    #[instrument(skip_all,fields(tenant_id=%self.tenant_id, pg_version=pg_version))]
+    #[instrument(skip_all,fields(tenant_id=%tenant_id, pg_version=pg_version))]
     fn launch(
-        &self,
-        input: &mut MutexGuard<Option<ProcessInput>>,
+        conf: &'static PageServerConf,
+        tenant_id: TenantId,
         pg_version: u32,
-    ) -> Result<(), Error> {
-        // Previous versions of wal-redo required data directory and that directories
-        // occupied some space on disk. Remove it if we face it.
-        //
-        // This code could be dropped after one release cycle.
-        let legacy_datadir = path_with_suffix_extension(
-            self.conf
-                .tenant_path(&self.tenant_id)
-                .join("wal-redo-datadir"),
-            TEMP_FILE_SUFFIX,
-        );
-        if legacy_datadir.exists() {
-            info!("legacy wal-redo datadir {legacy_datadir:?} exists, removing");
-            fs::remove_dir_all(&legacy_datadir).map_err(|e| {
-                Error::new(
-                    e.kind(),
-                    format!("legacy wal-redo datadir {legacy_datadir:?} removal failure: {e}"),
-                )
-            })?;
-        }
-
-        let pg_bin_dir_path = self
-            .conf
-            .pg_bin_dir(pg_version)
-            .map_err(|e| Error::new(ErrorKind::Other, format!("incorrect pg_bin_dir path: {e}")))?;
-        let pg_lib_dir_path = self
-            .conf
-            .pg_lib_dir(pg_version)
-            .map_err(|e| Error::new(ErrorKind::Other, format!("incorrect pg_lib_dir path: {e}")))?;
+    ) -> anyhow::Result<Self> {
+        let pg_bin_dir_path = conf.pg_bin_dir(pg_version).context("pg_bin_dir")?; // TODO these should be infallible.
+        let pg_lib_dir_path = conf.pg_lib_dir(pg_version).context("pg_lib_dir")?;
 
         // Start postgres itself
         let child = Command::new(pg_bin_dir_path.join("postgres"))
@@ -697,13 +665,8 @@ impl PostgresRedoManager {
             // as close-on-exec by default, but that's not enough, since we use
             // libraries that directly call libc open without setting that flag.
             .close_fds()
-            .spawn_no_leak_child(self.tenant_id)
-            .map_err(|e| {
-                Error::new(
-                    e.kind(),
-                    format!("postgres --wal-redo command failed to start: {}", e),
-                )
-            })?;
+            .spawn_no_leak_child(tenant_id)
+            .context("spawn process")?;
 
         let mut child = scopeguard::guard(child, |child| {
             error!("killing wal-redo-postgres process due to a problem during launch");
@@ -713,7 +676,6 @@ impl PostgresRedoManager {
         let stdin = child.stdin.take().unwrap();
         let stdout = child.stdout.take().unwrap();
         let stderr = child.stderr.take().unwrap();
-
         macro_rules! set_nonblock_or_log_err {
             ($file:ident) => {{
                 let res = set_nonblock($file.as_raw_fd());
@@ -727,39 +689,108 @@ impl PostgresRedoManager {
         set_nonblock_or_log_err!(stdout)?;
         set_nonblock_or_log_err!(stderr)?;
 
+        let mut stderr = tokio::io::unix::AsyncFd::new(stderr).context("AsyncFd::with_interest")?;
+
         // all fallible operations post-spawn are complete, so get rid of the guard
         let child = scopeguard::ScopeGuard::into_inner(child);
 
-        **input = Some(ProcessInput {
-            child,
-            stdout_fd: stdout.as_raw_fd(),
-            stderr_fd: stderr.as_raw_fd(),
-            stdin,
-            n_requests: 0,
+        let stderr_logger_cancel = CancellationToken::new();
+        let (stderr_logger_task_done_tx, stderr_logger_task_done_rx) =
+            tokio::sync::watch::channel(false);
+        tokio::spawn({
+            let stderr_logger_cancel = stderr_logger_cancel.clone();
+            async move {
+                scopeguard::defer! {
+                    debug!("wal-redo-postgres stderr_logger_task finished");
+                    let _ = stderr_logger_task_done_tx.send(true);
+                }
+                debug!("wal-redo-postgres stderr_logger_task started");
+                loop {
+                    // NB: we purposefully don't do a select! for the cancellation here.
+                    // The cancellation would likely cause us to miss stderr messages.
+                    // We can rely on this to return from .await because when we SIGKILL
+                    // the child, the writing end of the stderr pipe gets closed.
+                    match stderr.readable_mut().await {
+                        Ok(mut guard) => {
+                            let mut errbuf = [0; 16384];
+                            let res = guard.try_io(|fd| {
+                                use std::io::Read;
+                                fd.get_mut().read(&mut errbuf)
+                            });
+                            match res {
+                                Ok(Ok(0)) => {
+                                    // it closed the stderr pipe
+                                    break;
+                                }
+                                Ok(Ok(n)) => {
+                                    // The message might not be split correctly into lines here. But this is
+                                    // good enough, the important thing is to get the message to the log.
+                                    let output = String::from_utf8_lossy(&errbuf[0..n]).to_string();
+                                    error!(output, "received output");
+                                },
+                                Ok(Err(e)) => {
+                                    error!(error = ?e, "read() error, waiting for cancellation");
+                                    stderr_logger_cancel.cancelled().await;
+                                    error!(error = ?e, "read() error, cancellation complete");
+                                    break;
+                                }
+                                Err(e) => {
+                                    let _e: tokio::io::unix::TryIoError = e;
+                                    // the read() returned WouldBlock, that's expected
+                                }
+                            }
+                        }
+                        Err(e) => {
+                            error!(error = ?e, "read() error, waiting for cancellation");
+                            stderr_logger_cancel.cancelled().await;
+                            error!(error = ?e, "read() error, cancellation complete");
+                            break;
+                        }
+                    }
+                }
+            }.instrument(tracing::info_span!(parent: None, "wal-redo-postgres-stderr", pid = child.id(), tenant_id = %tenant_id, %pg_version))
         });
 
-        *self.stdout.lock().unwrap() = Some(ProcessOutput {
-            stdout,
-            pending_responses: VecDeque::new(),
-            n_processed_responses: 0,
-        });
-        *self.stderr.lock().unwrap() = Some(stderr);
+        Ok(Self {
+            conf,
+            tenant_id,
+            child: Some(child),
+            stdin: Mutex::new(ProcessInput {
+                stdin,
+                n_requests: 0,
+            }),
+            stdout: Mutex::new(ProcessOutput {
+                stdout,
+                pending_responses: VecDeque::new(),
+                n_processed_responses: 0,
+            }),
+            stderr_logger_cancel,
+            stderr_logger_task_done: stderr_logger_task_done_rx,
+            #[cfg(feature = "testing")]
+            dump_sequence: AtomicUsize::default(),
+        })
+    }
 
-        Ok(())
+    fn id(&self) -> u32 {
+        self.child
+            .as_ref()
+            .expect("must not call this during Drop")
+            .id()
     }
 
     // Apply given WAL records ('records') over an old page image. Returns
     // new page image.
     //
-    #[instrument(skip_all, fields(tenant_id=%self.tenant_id, pid=%input.as_ref().unwrap().child.id()))]
+    #[instrument(skip_all, fields(tenant_id=%self.tenant_id, pid=%self.id()))]
     fn apply_wal_records(
         &self,
-        input: MutexGuard<Option<ProcessInput>>,
         tag: BufferTag,
         base_img: &Option<Bytes>,
         records: &[(Lsn, NeonWalRecord)],
         wal_redo_timeout: Duration,
-    ) -> Result<Bytes, std::io::Error> {
+    ) -> anyhow::Result<Bytes> {
+        let input = self.stdin.lock().unwrap();
+
         // Serialize all the messages to send the WAL redo process first.
         //
         // This could be problematic if there are millions of records to replay,
@@ -782,10 +813,7 @@ impl PostgresRedoManager {
             {
                 build_apply_record_msg(*lsn, postgres_rec, &mut writebuf);
             } else {
-                return Err(Error::new(
-                    ErrorKind::Other,
-                    "tried to pass neon wal record to postgres WAL redo",
-                ));
+                anyhow::bail!("tried to pass neon wal record to postgres WAL redo");
             }
         }
         build_get_page_msg(tag, &mut writebuf);
@@ -805,77 +833,38 @@ impl PostgresRedoManager {
     fn apply_wal_records0(
         &self,
         writebuf: &[u8],
-        mut input: MutexGuard<Option<ProcessInput>>,
+        input: MutexGuard<ProcessInput>,
         wal_redo_timeout: Duration,
-    ) -> Result<Bytes, std::io::Error> {
-        let proc = input.as_mut().unwrap();
+    ) -> anyhow::Result<Bytes> {
+        let mut proc = { input }; // TODO: remove this legacy rename, but this keep the patch small.
         let mut nwrite = 0usize;
-        let stdout_fd = proc.stdout_fd;
 
-        // Prepare for calling poll()
-        let mut pollfds = [
-            PollFd::new(proc.stdin.as_raw_fd(), PollFlags::POLLOUT),
-            PollFd::new(proc.stderr_fd, PollFlags::POLLIN),
-            PollFd::new(stdout_fd, PollFlags::POLLIN),
-        ];
+        let mut stdin_pollfds = [PollFd::new(proc.stdin.as_raw_fd(), PollFlags::POLLOUT)];
 
-        // We do two things simultaneously: send the old base image and WAL records to
-        // the child process's stdin and forward any logging
-        // information that the child writes to its stderr to the page server's log.
         while nwrite < writebuf.len() {
             let n = loop {
-                match nix::poll::poll(&mut pollfds[0..2], wal_redo_timeout.as_millis() as i32) {
-                    Err(e) if e == nix::errno::Errno::EINTR => continue,
+                match nix::poll::poll(&mut stdin_pollfds[..], wal_redo_timeout.as_millis() as i32) {
+                    Err(nix::errno::Errno::EINTR) => continue,
                     res => break res,
                 }
             }?;
 
             if n == 0 {
-                return Err(Error::new(ErrorKind::Other, "WAL redo timed out"));
-            }
-
-            // If we have some messages in stderr, forward them to the log.
-            let err_revents = pollfds[1].revents().unwrap();
-            if err_revents & (PollFlags::POLLERR | PollFlags::POLLIN) != PollFlags::empty() {
-                let mut errbuf: [u8; 16384] = [0; 16384];
-                let mut stderr_guard = self.stderr.lock().unwrap();
-                let stderr = stderr_guard.as_mut().unwrap();
-                let len = stderr.read(&mut errbuf)?;
-
-                // The message might not be split correctly into lines here. But this is
-                // good enough, the important thing is to get the message to the log.
-                if len > 0 {
-                    error!(
-                        "wal-redo-postgres: {}",
-                        String::from_utf8_lossy(&errbuf[0..len])
-                    );
-
-                    // To make sure we capture all log from the process if it fails, keep
-                    // reading from the stderr, before checking the stdout.
-                    continue;
-                }
-            } else if err_revents.contains(PollFlags::POLLHUP) {
-                return Err(Error::new(
-                    ErrorKind::BrokenPipe,
-                    "WAL redo process closed its stderr unexpectedly",
-                ));
+                anyhow::bail!("WAL redo timed out");
             }
 
             // If 'stdin' is writeable, do write.
-            let in_revents = pollfds[0].revents().unwrap();
+            let in_revents = stdin_pollfds[0].revents().unwrap();
             if in_revents & (PollFlags::POLLERR | PollFlags::POLLOUT) != PollFlags::empty() {
                 nwrite += proc.stdin.write(&writebuf[nwrite..])?;
             } else if in_revents.contains(PollFlags::POLLHUP) {
                 // We still have more data to write, but the process closed the pipe.
-                return Err(Error::new(
-                    ErrorKind::BrokenPipe,
-                    "WAL redo process closed its stdin unexpectedly",
-                ));
+                anyhow::bail!("WAL redo process closed its stdin unexpectedly");
             }
         }
         let request_no = proc.n_requests;
         proc.n_requests += 1;
-        drop(input);
+        drop(proc);
 
         // To improve walredo performance we separate sending requests and receiving
         // responses. Them are protected by different mutexes (output and input).
@@ -889,23 +878,8 @@ impl PostgresRedoManager {
         // pending responses ring buffer and truncate all empty elements from the front,
         // advancing processed responses number.
 
-        let mut output_guard = self.stdout.lock().unwrap();
-        let output = output_guard.as_mut().unwrap();
-        if output.stdout.as_raw_fd() != stdout_fd {
-            // If stdout file descriptor is changed then it means that walredo process is crashed and restarted.
-            // As far as ProcessInput and ProcessOutout are protected by different mutexes,
-            // it can happen that we send request to one process and waiting response from another.
-            // To prevent such situation we compare stdout file descriptors.
-            // As far as old stdout pipe is destroyed only after new one is created,
-            // it can not reuse the same file descriptor, so this check is safe.
-            //
-            // Cross-read this with the comment in apply_batch_postgres if result.is_err().
-            // That's where we kill the child process.
-            return Err(Error::new(
-                ErrorKind::BrokenPipe,
-                "WAL redo process closed its stdout unexpectedly",
-            ));
-        }
+        let mut output = self.stdout.lock().unwrap();
+        let mut stdout_pollfds = [PollFd::new(output.stdout.as_raw_fd(), PollFlags::POLLIN)];
         let n_processed_responses = output.n_processed_responses;
         while n_processed_responses + output.pending_responses.len() <= request_no {
             // We expect the WAL redo process to respond with an 8k page image. We read it
@@ -916,52 +890,25 @@ impl PostgresRedoManager {
                 // We do two things simultaneously: reading response from stdout
                 // and forward any logging information that the child writes to its stderr to the page server's log.
                 let n = loop {
-                    match nix::poll::poll(&mut pollfds[1..3], wal_redo_timeout.as_millis() as i32) {
-                        Err(e) if e == nix::errno::Errno::EINTR => continue,
+                    match nix::poll::poll(
+                        &mut stdout_pollfds[..],
+                        wal_redo_timeout.as_millis() as i32,
+                    ) {
+                        Err(nix::errno::Errno::EINTR) => continue,
                         res => break res,
                     }
                 }?;
 
                 if n == 0 {
-                    return Err(Error::new(ErrorKind::Other, "WAL redo timed out"));
-                }
-
-                // If we have some messages in stderr, forward them to the log.
-                let err_revents = pollfds[1].revents().unwrap();
-                if err_revents & (PollFlags::POLLERR | PollFlags::POLLIN) != PollFlags::empty() {
-                    let mut errbuf: [u8; 16384] = [0; 16384];
-                    let mut stderr_guard = self.stderr.lock().unwrap();
-                    let stderr = stderr_guard.as_mut().unwrap();
-                    let len = stderr.read(&mut errbuf)?;
-
-                    // The message might not be split correctly into lines here. But this is
-                    // good enough, the important thing is to get the message to the log.
-                    if len > 0 {
-                        error!(
-                            "wal-redo-postgres: {}",
-                            String::from_utf8_lossy(&errbuf[0..len])
-                        );
-
-                        // To make sure we capture all log from the process if it fails, keep
-                        // reading from the stderr, before checking the stdout.
-                        continue;
-                    }
-                } else if err_revents.contains(PollFlags::POLLHUP) {
-                    return Err(Error::new(
-                        ErrorKind::BrokenPipe,
-                        "WAL redo process closed its stderr unexpectedly",
-                    ));
+                    anyhow::bail!("WAL redo timed out");
                 }
 
                 // If we have some data in stdout, read it to the result buffer.
-                let out_revents = pollfds[2].revents().unwrap();
+                let out_revents = stdout_pollfds[0].revents().unwrap();
                 if out_revents & (PollFlags::POLLERR | PollFlags::POLLIN) != PollFlags::empty() {
                     nresult += output.stdout.read(&mut resultbuf[nresult..])?;
                 } else if out_revents.contains(PollFlags::POLLHUP) {
-                    return Err(Error::new(
-                        ErrorKind::BrokenPipe,
-                        "WAL redo process closed its stdout unexpectedly",
-                    ));
+                    anyhow::bail!("WAL redo process closed its stdout unexpectedly");
                 }
             }
             output
@@ -1047,6 +994,17 @@ impl PostgresRedoManager {
     fn record_and_log(&self, _: &[u8]) {}
 }
 
+impl Drop for WalRedoProcess {
+    fn drop(&mut self) {
+        self.child
+            .take()
+            .expect("we only do this once")
+            .kill_and_wait();
+        self.stderr_logger_cancel.cancel();
+        // no way to wait for stderr_logger_task from Drop because that is async only
+    }
+}
+
 /// Wrapper type around `std::process::Child` which guarantees that the child
 /// will be killed and waited-for by this process before being dropped.
 struct NoLeakChild {
@@ -1125,7 +1083,7 @@ impl Drop for NoLeakChild {
         // Offload the kill+wait of the child process into the background.
         // If someone stops the runtime, we'll leak the child process.
         // We can ignore that case because we only stop the runtime on pageserver exit.
-        BACKGROUND_RUNTIME.spawn(async move {
+        tokio::runtime::Handle::current().spawn(async move {
             tokio::task::spawn_blocking(move || {
                 // Intentionally don't inherit the tracing context from whoever is dropping us.
                 // This thread here is going to outlive of our dropper.
@@ -1194,15 +1152,15 @@ fn build_get_page_msg(tag: BufferTag, buf: &mut Vec<u8>) {
 
 #[cfg(test)]
 mod tests {
-    use super::{PostgresRedoManager, WalRedoManager};
+    use super::PostgresRedoManager;
     use crate::repository::Key;
     use crate::{config::PageServerConf, walrecord::NeonWalRecord};
     use bytes::Bytes;
     use std::str::FromStr;
     use utils::{id::TenantId, lsn::Lsn};
 
-    #[test]
-    fn short_v14_redo() {
+    #[tokio::test]
+    async fn short_v14_redo() {
         let expected = std::fs::read("fixtures/short_v14_redo.page").unwrap();
 
         let h = RedoHarness::new().unwrap();
@@ -1223,13 +1181,14 @@ mod tests {
                 short_records(),
                 14,
             )
+            .await
             .unwrap();
 
         assert_eq!(&expected, &*page);
     }
 
-    #[test]
-    fn short_v14_fails_for_wrong_key_but_returns_zero_page() {
+    #[tokio::test]
+    async fn short_v14_fails_for_wrong_key_but_returns_zero_page() {
         let h = RedoHarness::new().unwrap();
 
         let page = h
@@ -1249,6 +1208,7 @@ mod tests {
                 short_records(),
                 14,
             )
+            .await
             .unwrap();
 
         // TODO: there will be some stderr printout, which is forwarded to tracing that could
@@ -1256,6 +1216,22 @@ mod tests {
         assert_eq!(page, crate::ZERO_PAGE);
     }
 
+    #[tokio::test]
+    async fn test_stderr() {
+        let h = RedoHarness::new().unwrap();
+        h
+            .manager
+            .request_redo(
+                Key::from_i128(0),
+                Lsn::INVALID,
+                None,
+                short_records(),
+                16, /* 16 currently produces stderr output on startup, which adds a nice extra edge */
+            )
+            .await
+            .unwrap_err();
+    }
+
     #[allow(clippy::octal_escapes)]
     fn short_records() -> Vec<(Lsn, NeonWalRecord)> {
         vec![
@@ -1284,6 +1260,8 @@ mod tests {
 
     impl RedoHarness {
         fn new() -> anyhow::Result<Self> {
+            crate::tenant::harness::setup_logging();
+
             let repo_dir = camino_tempfile::tempdir()?;
             let conf = PageServerConf::dummy_conf(repo_dir.path().to_path_buf());
             let conf = Box::leak(Box::new(conf));

pgxn/neon/Makefile

@@ -7,12 +7,12 @@ OBJS = \
 	extension_server.o \
 	file_cache.o \
 	libpagestore.o \
-	libpqwalproposer.o \
 	neon.o \
+	neon_utils.o \
 	pagestore_smgr.o \
 	relsize_cache.o \
 	walproposer.o \
-	walproposer_utils.o \
+	walproposer_pg.o \
 	control_plane_connector.o
 
 PG_CPPFLAGS = -I$(libpq_srcdir)
@@ -23,6 +23,23 @@ EXTENSION = neon
 DATA = neon--1.0.sql
 PGFILEDESC = "neon - cloud storage for PostgreSQL"
 
+EXTRA_CLEAN = \
+	libwalproposer.a
+
+WALPROP_OBJS = \
+	$(WIN32RES) \
+	walproposer.o \
+	neon_utils.o \
+	walproposer_compat.o
+
+.PHONY: walproposer-lib
+walproposer-lib: CPPFLAGS += -DWALPROPOSER_LIB
+walproposer-lib: libwalproposer.a;
+
+.PHONY: libwalproposer.a
+libwalproposer.a: $(WALPROP_OBJS)
+	rm -f $@
+	$(AR) $(AROPT) $@ $^
 
 PG_CONFIG = pg_config
 PGXS := $(shell $(PG_CONFIG) --pgxs)

pgxn/neon/control_plane_connector.c

@@ -741,13 +741,6 @@ NeonProcessUtility(
 			break;
 		case T_DropdbStmt:
 			HandleDropDb(castNode(DropdbStmt, parseTree));
-			/*
-			 * We do this here to hack around the fact that Postgres performs the drop
-			 * INSIDE of standard_ProcessUtility, which means that if we try to
-			 * abort the drop normally it'll be too late. DROP DATABASE can't be inside
-			 * of a transaction block anyway, so this should be fine to do.
-			 */
-			NeonXactCallback(XACT_EVENT_PRE_COMMIT, NULL);
 			break;
 		case T_CreateRoleStmt:
 			HandleCreateRole(castNode(CreateRoleStmt, parseTree));

pgxn/neon/libpagestore.c

@@ -30,7 +30,7 @@
 
 #include "neon.h"
 #include "walproposer.h"
-#include "walproposer_utils.h"
+#include "neon_utils.h"
 
 #define PageStoreTrace DEBUG5
 

pgxn/neon/libpqwalproposer.c

@@ -1,424 +0,0 @@
-#include "postgres.h"
-
-#include "libpq-fe.h"
-#include "neon.h"
-#include "walproposer.h"
-
-/* Header in walproposer.h -- Wrapper struct to abstract away the libpq connection */
-struct WalProposerConn
-{
-	PGconn	   *pg_conn;
-	bool		is_nonblocking; /* whether the connection is non-blocking */
-	char	   *recvbuf;		/* last received data from
-								 * walprop_async_read */
-};
-
-/* Helper function */
-static bool
-ensure_nonblocking_status(WalProposerConn *conn, bool is_nonblocking)
-{
-	/* If we're already correctly blocking or nonblocking, all good */
-	if (is_nonblocking == conn->is_nonblocking)
-		return true;
-
-	/* Otherwise, set it appropriately */
-	if (PQsetnonblocking(conn->pg_conn, is_nonblocking) == -1)
-		return false;
-
-	conn->is_nonblocking = is_nonblocking;
-	return true;
-}
-
-/* Exported function definitions */
-char *
-walprop_error_message(WalProposerConn *conn)
-{
-	return PQerrorMessage(conn->pg_conn);
-}
-
-WalProposerConnStatusType
-walprop_status(WalProposerConn *conn)
-{
-	switch (PQstatus(conn->pg_conn))
-	{
-		case CONNECTION_OK:
-			return WP_CONNECTION_OK;
-		case CONNECTION_BAD:
-			return WP_CONNECTION_BAD;
-		default:
-			return WP_CONNECTION_IN_PROGRESS;
-	}
-}
-
-WalProposerConn *
-walprop_connect_start(char *conninfo, char *password)
-{
-	WalProposerConn *conn;
-	PGconn	   *pg_conn;
-	const char *keywords[3];
-	const char *values[3];
-	int			n;
-
-	/*
-	 * Connect using the given connection string. If the
-	 * NEON_AUTH_TOKEN environment variable was set, use that as
-	 * the password.
-	 *
-	 * The connection options are parsed in the order they're given, so
-	 * when we set the password before the connection string, the
-	 * connection string can override the password from the env variable.
-	 * Seems useful, although we don't currently use that capability
-	 * anywhere.
-	 */
-	n = 0;
-	if (password)
-	{
-		keywords[n] = "password";
-		values[n] = password;
-		n++;
-	}
-	keywords[n] = "dbname";
-	values[n] = conninfo;
-	n++;
-	keywords[n] = NULL;
-	values[n] = NULL;
-	n++;
-	pg_conn = PQconnectStartParams(keywords, values, 1);
-
-	/*
-	 * Allocation of a PQconn can fail, and will return NULL. We want to fully
-	 * replicate the behavior of PQconnectStart here.
-	 */
-	if (!pg_conn)
-		return NULL;
-
-	/*
-	 * And in theory this allocation can fail as well, but it's incredibly
-	 * unlikely if we just successfully allocated a PGconn.
-	 *
-	 * palloc will exit on failure though, so there's not much we could do if
-	 * it *did* fail.
-	 */
-	conn = palloc(sizeof(WalProposerConn));
-	conn->pg_conn = pg_conn;
-	conn->is_nonblocking = false;	/* connections always start in blocking
-									 * mode */
-	conn->recvbuf = NULL;
-	return conn;
-}
-
-WalProposerConnectPollStatusType
-walprop_connect_poll(WalProposerConn *conn)
-{
-	WalProposerConnectPollStatusType return_val;
-
-	switch (PQconnectPoll(conn->pg_conn))
-	{
-		case PGRES_POLLING_FAILED:
-			return_val = WP_CONN_POLLING_FAILED;
-			break;
-		case PGRES_POLLING_READING:
-			return_val = WP_CONN_POLLING_READING;
-			break;
-		case PGRES_POLLING_WRITING:
-			return_val = WP_CONN_POLLING_WRITING;
-			break;
-		case PGRES_POLLING_OK:
-			return_val = WP_CONN_POLLING_OK;
-			break;
-
-			/*
-			 * There's a comment at its source about this constant being
-			 * unused. We'll expect it's never returned.
-			 */
-		case PGRES_POLLING_ACTIVE:
-			elog(FATAL, "Unexpected PGRES_POLLING_ACTIVE returned from PQconnectPoll");
-
-			/*
-			 * This return is never actually reached, but it's here to make
-			 * the compiler happy
-			 */
-			return WP_CONN_POLLING_FAILED;
-
-		default:
-			Assert(false);
-			return_val = WP_CONN_POLLING_FAILED;	/* keep the compiler quiet */
-	}
-
-	return return_val;
-}
-
-bool
-walprop_send_query(WalProposerConn *conn, char *query)
-{
-	/*
-	 * We need to be in blocking mode for sending the query to run without
-	 * requiring a call to PQflush
-	 */
-	if (!ensure_nonblocking_status(conn, false))
-		return false;
-
-	/* PQsendQuery returns 1 on success, 0 on failure */
-	if (!PQsendQuery(conn->pg_conn, query))
-		return false;
-
-	return true;
-}
-
-WalProposerExecStatusType
-walprop_get_query_result(WalProposerConn *conn)
-{
-	PGresult   *result;
-	WalProposerExecStatusType return_val;
-
-	/* Marker variable if we need to log an unexpected success result */
-	char	   *unexpected_success = NULL;
-
-	/* Consume any input that we might be missing */
-	if (!PQconsumeInput(conn->pg_conn))
-		return WP_EXEC_FAILED;
-
-	if (PQisBusy(conn->pg_conn))
-		return WP_EXEC_NEEDS_INPUT;
-
-
-	result = PQgetResult(conn->pg_conn);
-
-	/*
-	 * PQgetResult returns NULL only if getting the result was successful &
-	 * there's no more of the result to get.
-	 */
-	if (!result)
-	{
-		elog(WARNING, "[libpqwalproposer] Unexpected successful end of command results");
-		return WP_EXEC_UNEXPECTED_SUCCESS;
-	}
-
-	/* Helper macro to reduce boilerplate */
-#define UNEXPECTED_SUCCESS(msg) \
-		return_val = WP_EXEC_UNEXPECTED_SUCCESS; \
-		unexpected_success = msg; \
-		break;
-
-
-	switch (PQresultStatus(result))
-	{
-			/* "true" success case */
-		case PGRES_COPY_BOTH:
-			return_val = WP_EXEC_SUCCESS_COPYBOTH;
-			break;
-
-			/* Unexpected success case */
-		case PGRES_EMPTY_QUERY:
-			UNEXPECTED_SUCCESS("empty query return");
-		case PGRES_COMMAND_OK:
-			UNEXPECTED_SUCCESS("data-less command end");
-		case PGRES_TUPLES_OK:
-			UNEXPECTED_SUCCESS("tuples return");
-		case PGRES_COPY_OUT:
-			UNEXPECTED_SUCCESS("'Copy Out' response");
-		case PGRES_COPY_IN:
-			UNEXPECTED_SUCCESS("'Copy In' response");
-		case PGRES_SINGLE_TUPLE:
-			UNEXPECTED_SUCCESS("single tuple return");
-		case PGRES_PIPELINE_SYNC:
-			UNEXPECTED_SUCCESS("pipeline sync point");
-
-			/* Failure cases */
-		case PGRES_BAD_RESPONSE:
-		case PGRES_NONFATAL_ERROR:
-		case PGRES_FATAL_ERROR:
-		case PGRES_PIPELINE_ABORTED:
-			return_val = WP_EXEC_FAILED;
-			break;
-
-		default:
-			Assert(false);
-			return_val = WP_EXEC_FAILED;	/* keep the compiler quiet */
-	}
-
-	if (unexpected_success)
-		elog(WARNING, "[libpqwalproposer] Unexpected successful %s", unexpected_success);
-
-	return return_val;
-}
-
-pgsocket
-walprop_socket(WalProposerConn *conn)
-{
-	return PQsocket(conn->pg_conn);
-}
-
-int
-walprop_flush(WalProposerConn *conn)
-{
-	return (PQflush(conn->pg_conn));
-}
-
-void
-walprop_finish(WalProposerConn *conn)
-{
-	if (conn->recvbuf != NULL)
-		PQfreemem(conn->recvbuf);
-	PQfinish(conn->pg_conn);
-	pfree(conn);
-}
-
-/*
- * Receive a message from the safekeeper.
- *
- * On success, the data is placed in *buf. It is valid until the next call
- * to this function.
- */
-PGAsyncReadResult
-walprop_async_read(WalProposerConn *conn, char **buf, int *amount)
-{
-	int			result;
-
-	if (conn->recvbuf != NULL)
-	{
-		PQfreemem(conn->recvbuf);
-		conn->recvbuf = NULL;
-	}
-
-	/* Call PQconsumeInput so that we have the data we need */
-	if (!PQconsumeInput(conn->pg_conn))
-	{
-		*amount = 0;
-		*buf = NULL;
-		return PG_ASYNC_READ_FAIL;
-	}
-
-	/*
-	 * The docs for PQgetCopyData list the return values as: 0 if the copy is
-	 * still in progress, but no "complete row" is available -1 if the copy is
-	 * done -2 if an error occurred (> 0) if it was successful; that value is
-	 * the amount transferred.
-	 *
-	 * The protocol we use between walproposer and safekeeper means that we
-	 * *usually* wouldn't expect to see that the copy is done, but this can
-	 * sometimes be triggered by the server returning an ErrorResponse (which
-	 * also happens to have the effect that the copy is done).
-	 */
-	switch (result = PQgetCopyData(conn->pg_conn, &conn->recvbuf, true))
-	{
-		case 0:
-			*amount = 0;
-			*buf = NULL;
-			return PG_ASYNC_READ_TRY_AGAIN;
-		case -1:
-			{
-				/*
-				 * If we get -1, it's probably because of a server error; the
-				 * safekeeper won't normally send a CopyDone message.
-				 *
-				 * We can check PQgetResult to make sure that the server
-				 * failed; it'll always result in PGRES_FATAL_ERROR
-				 */
-				ExecStatusType status = PQresultStatus(PQgetResult(conn->pg_conn));
-
-				if (status != PGRES_FATAL_ERROR)
-					elog(FATAL, "unexpected result status %d after failed PQgetCopyData", status);
-
-				/*
-				 * If there was actually an error, it'll be properly reported
-				 * by calls to PQerrorMessage -- we don't have to do anything
-				 * else
-				 */
-				*amount = 0;
-				*buf = NULL;
-				return PG_ASYNC_READ_FAIL;
-			}
-		case -2:
-			*amount = 0;
-			*buf = NULL;
-			return PG_ASYNC_READ_FAIL;
-		default:
-			/* Positive values indicate the size of the returned result */
-			*amount = result;
-			*buf = conn->recvbuf;
-			return PG_ASYNC_READ_SUCCESS;
-	}
-}
-
-PGAsyncWriteResult
-walprop_async_write(WalProposerConn *conn, void const *buf, size_t size)
-{
-	int			result;
-
-	/* If we aren't in non-blocking mode, switch to it. */
-	if (!ensure_nonblocking_status(conn, true))
-		return PG_ASYNC_WRITE_FAIL;
-
-	/*
-	 * The docs for PQputcopyData list the return values as: 1 if the data was
-	 * queued, 0 if it was not queued because of full buffers, or -1 if an
-	 * error occurred
-	 */
-	result = PQputCopyData(conn->pg_conn, buf, size);
-
-	/*
-	 * We won't get a result of zero because walproposer always empties the
-	 * connection's buffers before sending more
-	 */
-	Assert(result != 0);
-
-	switch (result)
-	{
-		case 1:
-			/* good -- continue */
-			break;
-		case -1:
-			return PG_ASYNC_WRITE_FAIL;
-		default:
-			elog(FATAL, "invalid return %d from PQputCopyData", result);
-	}
-
-	/*
-	 * After queueing the data, we still need to flush to get it to send. This
-	 * might take multiple tries, but we don't want to wait around until it's
-	 * done.
-	 *
-	 * PQflush has the following returns (directly quoting the docs): 0 if
-	 * sucessful, 1 if it was unable to send all the data in the send queue
-	 * yet -1 if it failed for some reason
-	 */
-	switch (result = PQflush(conn->pg_conn))
-	{
-		case 0:
-			return PG_ASYNC_WRITE_SUCCESS;
-		case 1:
-			return PG_ASYNC_WRITE_TRY_FLUSH;
-		case -1:
-			return PG_ASYNC_WRITE_FAIL;
-		default:
-			elog(FATAL, "invalid return %d from PQflush", result);
-	}
-}
-
-/*
- * This function is very similar to walprop_async_write. For more
- * information, refer to the comments there.
- */
-bool
-walprop_blocking_write(WalProposerConn *conn, void const *buf, size_t size)
-{
-	int			result;
-
-	/* If we are in non-blocking mode, switch out of it. */
-	if (!ensure_nonblocking_status(conn, false))
-		return false;
-
-	if ((result = PQputCopyData(conn->pg_conn, buf, size)) == -1)
-		return false;
-
-	Assert(result == 1);
-
-	/* Because the connection is non-blocking, flushing returns 0 or -1 */
-
-	if ((result = PQflush(conn->pg_conn)) == -1)
-		return false;
-
-	Assert(result == 0);
-	return true;
-}

pgxn/neon/neon.h

@@ -18,6 +18,10 @@ extern char *neon_auth_token;
 extern char *neon_timeline;
 extern char *neon_tenant;
 
+extern char *wal_acceptors_list;
+extern int	wal_acceptor_reconnect_timeout;
+extern int	wal_acceptor_connection_timeout;
+
 extern void pg_init_libpagestore(void);
 extern void pg_init_walproposer(void);
 
@@ -30,4 +34,10 @@ extern void pg_init_extension_server(void);
 extern bool neon_redo_read_buffer_filter(XLogReaderState *record, uint8 block_id);
 extern bool	(*old_redo_read_buffer_filter) (XLogReaderState *record, uint8 block_id);
 
+extern uint64 BackpressureThrottlingTime(void);
+extern void replication_feedback_get_lsns(XLogRecPtr *writeLsn, XLogRecPtr *flushLsn, XLogRecPtr *applyLsn);
+
+extern void PGDLLEXPORT WalProposerSync(int argc, char *argv[]);
+extern void PGDLLEXPORT WalProposerMain(Datum main_arg);
+
 #endif							/* NEON_H */

pgxn/neon/neon_utils.c

@@ -0,0 +1,116 @@
+#include "postgres.h"
+
+#include "access/timeline.h"
+#include "access/xlogutils.h"
+#include "common/logging.h"
+#include "common/ip.h"
+#include "funcapi.h"
+#include "libpq/libpq.h"
+#include "libpq/pqformat.h"
+#include "miscadmin.h"
+#include "postmaster/interrupt.h"
+#include "replication/slot.h"
+#include "replication/walsender_private.h"
+
+#include "storage/ipc.h"
+#include "utils/builtins.h"
+#include "utils/ps_status.h"
+
+#include "libpq-fe.h"
+#include <netinet/tcp.h>
+#include <unistd.h>
+
+#if PG_VERSION_NUM >= 150000
+#include "access/xlogutils.h"
+#include "access/xlogrecovery.h"
+#endif
+#if PG_MAJORVERSION_NUM >= 16
+#include "utils/guc.h"
+#endif
+
+/*
+ * Convert a character which represents a hexadecimal digit to an integer.
+ *
+ * Returns -1 if the character is not a hexadecimal digit.
+ */
+int
+HexDecodeChar(char c)
+{
+	if (c >= '0' && c <= '9')
+		return c - '0';
+	if (c >= 'a' && c <= 'f')
+		return c - 'a' + 10;
+	if (c >= 'A' && c <= 'F')
+		return c - 'A' + 10;
+
+	return -1;
+}
+
+/*
+ * Decode a hex string into a byte string, 2 hex chars per byte.
+ *
+ * Returns false if invalid characters are encountered; otherwise true.
+ */
+bool
+HexDecodeString(uint8 *result, char *input, int nbytes)
+{
+	int			i;
+
+	for (i = 0; i < nbytes; ++i)
+	{
+		int			n1 = HexDecodeChar(input[i * 2]);
+		int			n2 = HexDecodeChar(input[i * 2 + 1]);
+
+		if (n1 < 0 || n2 < 0)
+			return false;
+		result[i] = n1 * 16 + n2;
+	}
+
+	return true;
+}
+
+/* --------------------------------
+ *		pq_getmsgint32_le	- get a binary 4-byte int from a message buffer in native (LE) order
+ * --------------------------------
+ */
+uint32
+pq_getmsgint32_le(StringInfo msg)
+{
+	uint32		n32;
+
+	pq_copymsgbytes(msg, (char *) &n32, sizeof(n32));
+
+	return n32;
+}
+
+/* --------------------------------
+ *		pq_getmsgint64	- get a binary 8-byte int from a message buffer in native (LE) order
+ * --------------------------------
+ */
+uint64
+pq_getmsgint64_le(StringInfo msg)
+{
+	uint64		n64;
+
+	pq_copymsgbytes(msg, (char *) &n64, sizeof(n64));
+
+	return n64;
+}
+
+/* append a binary [u]int32 to a StringInfo buffer in native (LE) order */
+void
+pq_sendint32_le(StringInfo buf, uint32 i)
+{
+	enlargeStringInfo(buf, sizeof(uint32));
+	memcpy(buf->data + buf->len, &i, sizeof(uint32));
+	buf->len += sizeof(uint32);
+}
+
+/* append a binary [u]int64 to a StringInfo buffer in native (LE) order */
+void
+pq_sendint64_le(StringInfo buf, uint64 i)
+{
+	enlargeStringInfo(buf, sizeof(uint64));
+	memcpy(buf->data + buf->len, &i, sizeof(uint64));
+	buf->len += sizeof(uint64);
+}

pgxn/neon/neon_utils.h

@@ -0,0 +1,12 @@
+#ifndef __NEON_UTILS_H__
+#define __NEON_UTILS_H__
+
+#include "postgres.h"
+
+bool		HexDecodeString(uint8 *result, char *input, int nbytes);
+uint32		pq_getmsgint32_le(StringInfo msg);
+uint64		pq_getmsgint64_le(StringInfo msg);
+void		pq_sendint32_le(StringInfo buf, uint32 i);
+void		pq_sendint64_le(StringInfo buf, uint64 i);
+
+#endif							/* __NEON_UTILS_H__ */