pageserver: implement aligned io buffer

Signed-off-by: Yuchen Liang <yuchen@neon.tech>

.github/workflows/_benchmarking_preparation.yml

@@ -3,19 +3,23 @@ name: Prepare benchmarking databases by restoring dumps
 on:
   workflow_call:
     # no inputs needed
-    
+
 defaults:
   run:
     shell: bash -euxo pipefail {0}
 
 jobs:
   setup-databases:
+    permissions:
+      contents: write
+      statuses: write
+      id-token: write # aws-actions/configure-aws-credentials
     strategy:
       fail-fast: false
       matrix:
-        platform: [ aws-rds-postgres, aws-aurora-serverless-v2-postgres, neon ] 
+        platform: [ aws-rds-postgres, aws-aurora-serverless-v2-postgres, neon ]
         database: [ clickbench, tpch, userexample ]
-  
+
     env:
       LD_LIBRARY_PATH: /tmp/neon/pg_install/v16/lib
       PLATFORM: ${{ matrix.platform }}
@@ -23,7 +27,10 @@ jobs:
 
     runs-on: [ self-hosted, us-east-2, x64 ]
     container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:pinned
+      image: neondatabase/build-tools:pinned
+      credentials:
+        username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
+        password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
       options: --init
 
     steps:
@@ -32,13 +39,13 @@ jobs:
       run: |
         case "${PLATFORM}" in
           neon)
-            CONNSTR=${{ secrets.BENCHMARK_CAPTEST_CONNSTR }} 
+            CONNSTR=${{ secrets.BENCHMARK_CAPTEST_CONNSTR }}
             ;;
           aws-rds-postgres)
-            CONNSTR=${{ secrets.BENCHMARK_RDS_POSTGRES_CONNSTR }} 
+            CONNSTR=${{ secrets.BENCHMARK_RDS_POSTGRES_CONNSTR }}
             ;;
           aws-aurora-serverless-v2-postgres)
-            CONNSTR=${{ secrets.BENCHMARK_RDS_AURORA_CONNSTR }} 
+            CONNSTR=${{ secrets.BENCHMARK_RDS_AURORA_CONNSTR }}
             ;;
           *)
             echo >&2 "Unknown PLATFORM=${PLATFORM}"
@@ -46,10 +53,17 @@ jobs:
             ;;
         esac
 
-        echo "connstr=${CONNSTR}" >> $GITHUB_OUTPUT  
+        echo "connstr=${CONNSTR}" >> $GITHUB_OUTPUT
 
     - uses: actions/checkout@v4
 
+    - name: Configure AWS credentials
+      uses: aws-actions/configure-aws-credentials@v4
+      with:
+        aws-region: eu-central-1
+        role-to-assume: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+        role-duration-seconds: 18000 # 5 hours
+
     - name: Download Neon artifact
       uses: ./.github/actions/download
       with:
@@ -57,23 +71,23 @@ jobs:
         path: /tmp/neon/
         prefix: latest
 
-    # we create a table that has one row for each database that we want to restore with the status whether the restore is done    
+    # we create a table that has one row for each database that we want to restore with the status whether the restore is done
     - name: Create benchmark_restore_status table if it does not exist
       env:
         BENCHMARK_CONNSTR: ${{ steps.set-up-prep-connstr.outputs.connstr }}
         DATABASE_NAME: ${{ matrix.database }}
-      # to avoid a race condition of multiple jobs trying to create the table at the same time, 
+      # to avoid a race condition of multiple jobs trying to create the table at the same time,
       # we use an advisory lock
       run: |
         ${PG_BINARIES}/psql "${{ env.BENCHMARK_CONNSTR }}" -c "
-        SELECT pg_advisory_lock(4711);  
+        SELECT pg_advisory_lock(4711);
         CREATE TABLE IF NOT EXISTS benchmark_restore_status (
         databasename text primary key,
         restore_done boolean
         );
         SELECT pg_advisory_unlock(4711);
         "
-    
+
     - name: Check if restore is already done
       id: check-restore-done
       env:
@@ -107,7 +121,7 @@ jobs:
         DATABASE_NAME: ${{ matrix.database }}
       run: |
         mkdir -p /tmp/dumps
-        aws s3 cp s3://neon-github-dev/performance/pgdumps/$DATABASE_NAME/$DATABASE_NAME.pg_dump /tmp/dumps/ 
+        aws s3 cp s3://neon-github-dev/performance/pgdumps/$DATABASE_NAME/$DATABASE_NAME.pg_dump /tmp/dumps/
 
     - name: Replace database name in connection string
       if: steps.check-restore-done.outputs.skip != 'true'
@@ -126,17 +140,17 @@ jobs:
         else
           new_connstr="${base_connstr}/${DATABASE_NAME}"
         fi
-        echo "database_connstr=${new_connstr}" >> $GITHUB_OUTPUT  
+        echo "database_connstr=${new_connstr}" >> $GITHUB_OUTPUT
 
     - name: Restore dump
       if: steps.check-restore-done.outputs.skip != 'true'
       env:
         DATABASE_NAME: ${{ matrix.database }}
         DATABASE_CONNSTR: ${{ steps.replace-dbname.outputs.database_connstr }}
-        # the following works only with larger computes: 
+        # the following works only with larger computes:
         # PGOPTIONS: "-c maintenance_work_mem=8388608 -c max_parallel_maintenance_workers=7"
         # we add the || true because:
-        # the dumps were created with Neon and contain neon extensions that are not 
+        # the dumps were created with Neon and contain neon extensions that are not
         # available in RDS, so we will always report an error, but we can ignore it
       run: |
         ${PG_BINARIES}/pg_restore --clean --if-exists --no-owner --jobs=4 \

.github/workflows/_build-and-test-locally.yml

@@ -236,9 +236,7 @@ jobs:
 
           # run pageserver tests with different settings
           for io_engine in std-fs tokio-epoll-uring ; do
-            for io_buffer_alignment in 0 1 512 ; do
-              NEON_PAGESERVER_UNIT_TEST_VIRTUAL_FILE_IOENGINE=$io_engine NEON_PAGESERVER_UNIT_TEST_IO_BUFFER_ALIGNMENT=$io_buffer_alignment ${cov_prefix} cargo nextest run $CARGO_FLAGS $CARGO_FEATURES  -E 'package(pageserver)'
-            done
+            NEON_PAGESERVER_UNIT_TEST_VIRTUAL_FILE_IOENGINE=$io_engine ${cov_prefix} cargo nextest run $CARGO_FLAGS $CARGO_FEATURES  -E 'package(pageserver)'
           done
 
           # Run separate tests for real S3

.github/workflows/benchmarking.yml

@@ -12,7 +12,6 @@ on:
     #          │ │ │ ┌───────────── month (1 - 12 or JAN-DEC)
     #          │ │ │ │ ┌───────────── day of the week (0 - 6 or SUN-SAT)
     - cron:   '0 3 * * *' # run once a day, timezone is utc
-
   workflow_dispatch: # adds ability to run this manually
     inputs:
       region_id:
@@ -59,7 +58,7 @@ jobs:
     permissions:
       contents: write
       statuses: write
-      id-token: write # Required for OIDC authentication in azure runners
+      id-token: write # aws-actions/configure-aws-credentials
     strategy:
       fail-fast: false
       matrix:
@@ -68,12 +67,10 @@ jobs:
             PLATFORM: "neon-staging"
             region_id: ${{ github.event.inputs.region_id || 'aws-us-east-2' }}
             RUNNER: [ self-hosted, us-east-2, x64 ]
-            IMAGE: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:pinned
           - DEFAULT_PG_VERSION: 16
             PLATFORM: "azure-staging"
             region_id: 'azure-eastus2'
             RUNNER: [ self-hosted, eastus2, x64 ]
-            IMAGE: neondatabase/build-tools:pinned
     env:
       TEST_PG_BENCH_DURATIONS_MATRIX: "300"
       TEST_PG_BENCH_SCALES_MATRIX: "10,100"
@@ -86,7 +83,10 @@ jobs:
 
     runs-on: ${{ matrix.RUNNER }}
     container:
-      image: ${{ matrix.IMAGE }}
+      image: neondatabase/build-tools:pinned
+      credentials:
+        username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
+        password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
       options: --init
 
     steps:
@@ -164,6 +164,10 @@ jobs:
 
   replication-tests:
     if: ${{ github.event.inputs.run_only_pgvector_tests == 'false' || github.event.inputs.run_only_pgvector_tests == null }}
+    permissions:
+      contents: write
+      statuses: write
+      id-token: write # aws-actions/configure-aws-credentials
     env:
       POSTGRES_DISTRIB_DIR: /tmp/neon/pg_install
       DEFAULT_PG_VERSION: 16
@@ -174,12 +178,21 @@ jobs:
 
     runs-on: [ self-hosted, us-east-2, x64 ]
     container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:pinned
+      image: neondatabase/build-tools:pinned
+      credentials:
+        username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
+        password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
       options: --init
 
     steps:
     - uses: actions/checkout@v4
 
+    - name: Configure AWS credentials
+      uses: aws-actions/configure-aws-credentials@v4
+      with:
+        aws-region: eu-central-1
+        role-to-assume: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+        role-duration-seconds: 18000 # 5 hours
 
     - name: Download Neon artifact
       uses: ./.github/actions/download
@@ -267,7 +280,7 @@ jobs:
         region_id_default=${{ env.DEFAULT_REGION_ID }}
         runner_default='["self-hosted", "us-east-2", "x64"]'
         runner_azure='["self-hosted", "eastus2", "x64"]'
-        image_default="369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:pinned"
+        image_default="neondatabase/build-tools:pinned"
         matrix='{
           "pg_version" : [
             16
@@ -344,7 +357,7 @@ jobs:
     permissions:
       contents: write
       statuses: write
-      id-token: write # Required for OIDC authentication in azure runners
+      id-token: write # aws-actions/configure-aws-credentials
 
     strategy:
       fail-fast: false
@@ -371,7 +384,7 @@ jobs:
     steps:
     - uses: actions/checkout@v4
 
-    - name: Configure AWS credentials # necessary on Azure runners
+    - name: Configure AWS credentials
       uses: aws-actions/configure-aws-credentials@v4
       with:
         aws-region: eu-central-1
@@ -492,17 +505,15 @@ jobs:
     permissions:
       contents: write
       statuses: write
-      id-token: write # Required for OIDC authentication in azure runners
+      id-token: write # aws-actions/configure-aws-credentials
     strategy:
       fail-fast: false
       matrix:
         include:
           - PLATFORM: "neonvm-captest-pgvector"
             RUNNER: [ self-hosted, us-east-2, x64 ]
-            IMAGE: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:pinned
           - PLATFORM: "azure-captest-pgvector"
             RUNNER: [ self-hosted, eastus2, x64 ]
-            IMAGE: neondatabase/build-tools:pinned
 
     env:
       TEST_PG_BENCH_DURATIONS_MATRIX: "15m"
@@ -511,13 +522,16 @@ jobs:
       DEFAULT_PG_VERSION: 16
       TEST_OUTPUT: /tmp/test_output
       BUILD_TYPE: remote
-      LD_LIBRARY_PATH: /home/nonroot/pg/usr/lib/x86_64-linux-gnu
+
       SAVE_PERF_REPORT: ${{ github.event.inputs.save_perf_report || ( github.ref_name == 'main' ) }}
       PLATFORM: ${{ matrix.PLATFORM }}
 
     runs-on: ${{ matrix.RUNNER }}
     container:
-      image: ${{ matrix.IMAGE }}
+      image: neondatabase/build-tools:pinned
+      credentials:
+        username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
+        password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
       options: --init
 
     steps:
@@ -527,17 +541,26 @@ jobs:
     # instead of using Neon artifacts containing pgbench
     - name: Install postgresql-16 where pytest expects it
       run: |
+        # Just to make it easier to test things locally on macOS (with arm64)
+        arch=$(uname -m | sed 's/x86_64/amd64/g' | sed 's/aarch64/arm64/g')
+
         cd /home/nonroot
-        wget -q https://apt.postgresql.org/pub/repos/apt/pool/main/p/postgresql-16/libpq5_16.4-1.pgdg110%2B1_amd64.deb
-        wget -q https://apt.postgresql.org/pub/repos/apt/pool/main/p/postgresql-16/postgresql-client-16_16.4-1.pgdg110%2B1_amd64.deb
-        wget -q https://apt.postgresql.org/pub/repos/apt/pool/main/p/postgresql-16/postgresql-16_16.4-1.pgdg110%2B1_amd64.deb 
-        dpkg -x libpq5_16.4-1.pgdg110+1_amd64.deb pg
-        dpkg -x postgresql-client-16_16.4-1.pgdg110+1_amd64.deb pg
-        dpkg -x postgresql-16_16.4-1.pgdg110+1_amd64.deb pg
+        wget -q "https://apt.postgresql.org/pub/repos/apt/pool/main/p/postgresql-17/libpq5_17.0-1.pgdg110+1_${arch}.deb"
+        wget -q "https://apt.postgresql.org/pub/repos/apt/pool/main/p/postgresql-16/postgresql-client-16_16.4-1.pgdg110+2_${arch}.deb"
+        wget -q "https://apt.postgresql.org/pub/repos/apt/pool/main/p/postgresql-16/postgresql-16_16.4-1.pgdg110+2_${arch}.deb"
+        dpkg -x libpq5_17.0-1.pgdg110+1_${arch}.deb pg
+        dpkg -x postgresql-16_16.4-1.pgdg110+2_${arch}.deb pg
+        dpkg -x postgresql-client-16_16.4-1.pgdg110+2_${arch}.deb pg
+
         mkdir -p /tmp/neon/pg_install/v16/bin
-        ln -s /home/nonroot/pg/usr/lib/postgresql/16/bin/pgbench /tmp/neon/pg_install/v16/bin/pgbench  
-        ln -s /home/nonroot/pg/usr/lib/postgresql/16/bin/psql /tmp/neon/pg_install/v16/bin/psql  
-        ln -s /home/nonroot/pg/usr/lib/x86_64-linux-gnu /tmp/neon/pg_install/v16/lib 
+        ln -s /home/nonroot/pg/usr/lib/postgresql/16/bin/pgbench /tmp/neon/pg_install/v16/bin/pgbench
+        ln -s /home/nonroot/pg/usr/lib/postgresql/16/bin/psql    /tmp/neon/pg_install/v16/bin/psql
+        ln -s /home/nonroot/pg/usr/lib/$(uname -m)-linux-gnu     /tmp/neon/pg_install/v16/lib
+
+        LD_LIBRARY_PATH="/home/nonroot/pg/usr/lib/$(uname -m)-linux-gnu:${LD_LIBRARY_PATH:-}"
+        export LD_LIBRARY_PATH
+        echo "LD_LIBRARY_PATH=${LD_LIBRARY_PATH}" >> ${GITHUB_ENV}
+
         /tmp/neon/pg_install/v16/bin/pgbench --version
         /tmp/neon/pg_install/v16/bin/psql --version
 
@@ -559,7 +582,7 @@ jobs:
 
         echo "connstr=${CONNSTR}" >> $GITHUB_OUTPUT
 
-    - name: Configure AWS credentials # necessary on Azure runners to read/write from/to S3
+    - name: Configure AWS credentials
       uses: aws-actions/configure-aws-credentials@v4
       with:
         aws-region: eu-central-1
@@ -620,6 +643,10 @@ jobs:
     # *_CLICKBENCH_CONNSTR: Genuine ClickBench DB with ~100M rows
     # *_CLICKBENCH_10M_CONNSTR: DB with the first 10M rows of ClickBench DB
     if: ${{ !cancelled() && (github.event.inputs.run_only_pgvector_tests == 'false' || github.event.inputs.run_only_pgvector_tests == null) }}
+    permissions:
+      contents: write
+      statuses: write
+      id-token: write # aws-actions/configure-aws-credentials
     needs: [ generate-matrices, pgbench-compare, prepare_AWS_RDS_databases ]
 
     strategy:
@@ -638,12 +665,22 @@ jobs:
 
     runs-on: [ self-hosted, us-east-2, x64 ]
     container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:pinned
+      image: neondatabase/build-tools:pinned
+      credentials:
+        username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
+        password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
       options: --init
 
     steps:
     - uses: actions/checkout@v4
 
+    - name: Configure AWS credentials
+      uses: aws-actions/configure-aws-credentials@v4
+      with:
+        aws-region: eu-central-1
+        role-to-assume: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+        role-duration-seconds: 18000 # 5 hours
+
     - name: Download Neon artifact
       uses: ./.github/actions/download
       with:
@@ -714,6 +751,10 @@ jobs:
     #
     # *_TPCH_S10_CONNSTR: DB generated with scale factor 10 (~10 GB)
     if: ${{ !cancelled() && (github.event.inputs.run_only_pgvector_tests == 'false' || github.event.inputs.run_only_pgvector_tests == null) }}
+    permissions:
+      contents: write
+      statuses: write
+      id-token: write # aws-actions/configure-aws-credentials
     needs: [ generate-matrices, clickbench-compare, prepare_AWS_RDS_databases ]
 
     strategy:
@@ -731,12 +772,22 @@ jobs:
 
     runs-on: [ self-hosted, us-east-2, x64 ]
     container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:pinned
+      image: neondatabase/build-tools:pinned
+      credentials:
+        username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
+        password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
       options: --init
 
     steps:
     - uses: actions/checkout@v4
 
+    - name: Configure AWS credentials
+      uses: aws-actions/configure-aws-credentials@v4
+      with:
+        aws-region: eu-central-1
+        role-to-assume: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+        role-duration-seconds: 18000 # 5 hours
+
     - name: Download Neon artifact
       uses: ./.github/actions/download
       with:
@@ -806,6 +857,10 @@ jobs:
 
   user-examples-compare:
     if: ${{ !cancelled() && (github.event.inputs.run_only_pgvector_tests == 'false' || github.event.inputs.run_only_pgvector_tests == null) }}
+    permissions:
+      contents: write
+      statuses: write
+      id-token: write # aws-actions/configure-aws-credentials
     needs: [ generate-matrices, tpch-compare, prepare_AWS_RDS_databases ]
 
     strategy:
@@ -822,12 +877,22 @@ jobs:
 
     runs-on: [ self-hosted, us-east-2, x64 ]
     container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:pinned
+      image: neondatabase/build-tools:pinned
+      credentials:
+        username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
+        password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
       options: --init
 
     steps:
     - uses: actions/checkout@v4
 
+    - name: Configure AWS credentials
+      uses: aws-actions/configure-aws-credentials@v4
+      with:
+        aws-region: eu-central-1
+        role-to-assume: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+        role-duration-seconds: 18000 # 5 hours
+
     - name: Download Neon artifact
       uses: ./.github/actions/download
       with:

.github/workflows/build_and_test.yml

@@ -341,7 +341,7 @@ jobs:
           PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
           TEST_RESULT_CONNSTR: "${{ secrets.REGRESS_TEST_RESULT_CONNSTR_NEW }}"
           PAGESERVER_VIRTUAL_FILE_IO_ENGINE: tokio-epoll-uring
-          SYNC_AFTER_EACH_TEST: true
+          SYNC_BETWEEN_TESTS: true
       # XXX: no coverage data handling here, since benchmarks are run on release builds,
       # while coverage is currently collected for the debug ones
 
@@ -773,7 +773,7 @@ jobs:
       matrix:
         version: [ v14, v15, v16, v17 ]
     env:
-      VM_BUILDER_VERSION: v0.29.3
+      VM_BUILDER_VERSION: v0.35.0
 
     steps:
       - uses: actions/checkout@v4

.github/workflows/trigger-e2e-tests.yml

@@ -102,12 +102,17 @@ jobs:
           # Default set of platforms to run e2e tests on
           platforms='["docker", "k8s"]'
 
-          # If the PR changes vendor/, pgxn/ or libs/vm_monitor/ directories, or compute/Dockerfile.compute-node, add k8s-neonvm to the list of platforms.
+          # If a PR changes anything that affects computes, add k8s-neonvm to the list of platforms.
           # If the workflow run is not a pull request, add k8s-neonvm to the list.
           if [ "$GITHUB_EVENT_NAME" == "pull_request" ]; then
             for f in $(gh api "/repos/${GITHUB_REPOSITORY}/pulls/${PR_NUMBER}/files" --paginate --jq '.[].filename'); do
               case "$f" in
-                vendor/*|pgxn/*|libs/vm_monitor/*|compute/Dockerfile.compute-node)
+                # List of directories that contain code which affect compute images.
+                #
+                # This isn't exhaustive, just the paths that are most directly compute-related.
+                # For example, compute_ctl also depends on libs/utils, but we don't trigger
+                # an e2e run on that.
+                vendor/*|pgxn/*|compute_tools/*|libs/vm_monitor/*|compute/Dockerfile.compute-node)
                   platforms=$(echo "${platforms}" | jq --compact-output '. += ["k8s-neonvm"] | unique')
                   ;;
                 *)

Cargo.toml

@@ -53,15 +53,15 @@ azure_storage_blobs = { version = "0.19", default-features = false, features = [
 flate2 = "1.0.26"
 async-stream = "0.3"
 async-trait = "0.1"
-aws-config = { version = "1.3", default-features = false, features=["rustls"] }
-aws-sdk-s3 = "1.26"
-aws-sdk-iam = "1.15.0"
+aws-config = { version = "1.5", default-features = false, features=["rustls", "sso"] }
+aws-sdk-s3 = "1.52"
+aws-sdk-iam = "1.46.0"
 aws-smithy-async = { version = "1.2.1", default-features = false, features=["rt-tokio"] }
-aws-smithy-types = "1.1.9"
+aws-smithy-types = "1.2"
 aws-credential-types = "1.2.0"
-aws-sigv4 = { version = "1.2.1", features = ["sign-http"] }
-aws-types = "1.2.0"
-axum = { version = "0.6.20", features = ["ws"] }
+aws-sigv4 = { version = "1.2", features = ["sign-http"] }
+aws-types = "1.3"
+axum = { version = "0.7.5", features = ["ws"] }
 base64 = "0.13.0"
 bincode = "1.3"
 bindgen = "0.70"
@@ -96,10 +96,13 @@ hmac = "0.12.1"
 hostname = "0.4"
 http = {version = "1.1.0", features = ["std"]}
 http-types = { version = "2", default-features = false }
+http-body-util = "0.1.2"
 humantime = "2.1"
 humantime-serde = "1.1.1"
-hyper = "0.14"
-tokio-tungstenite = "0.20.0"
+hyper0 = { package = "hyper", version = "0.14" }
+hyper = "1.4"
+hyper-util = "0.1"
+tokio-tungstenite = "0.21.0"
 indexmap = "2"
 indoc = "2"
 ipnet = "2.9.0"
@@ -116,9 +119,10 @@ notify = "6.0.0"
 num_cpus = "1.15"
 num-traits = "0.2.15"
 once_cell = "1.13"
-opentelemetry = "0.20.0"
-opentelemetry-otlp = { version = "0.13.0", default-features=false, features = ["http-proto", "trace", "http", "reqwest-client"] }
-opentelemetry-semantic-conventions = "0.12.0"
+opentelemetry = "0.24"
+opentelemetry_sdk = "0.24"
+opentelemetry-otlp = { version = "0.17", default-features=false, features = ["http-proto", "trace", "http", "reqwest-client"] }
+opentelemetry-semantic-conventions = "0.16"
 parking_lot = "0.12"
 parquet = { version = "53", default-features = false, features = ["zstd"] }
 parquet_derive = "53"
@@ -126,12 +130,12 @@ pbkdf2 = { version = "0.12.1", features = ["simple", "std"] }
 pin-project-lite = "0.2"
 procfs = "0.16"
 prometheus = {version = "0.13", default-features=false, features = ["process"]} # removes protobuf dependency
-prost = "0.11"
+prost = "0.13"
 rand = "0.8"
 redis = { version = "0.25.2", features = ["tokio-rustls-comp", "keep-alive"] }
 regex = "1.10.2"
 reqwest = { version = "0.12", default-features = false, features = ["rustls-tls"] }
-reqwest-tracing = { version = "0.5", features = ["opentelemetry_0_20"] }
+reqwest-tracing = { version = "0.5", features = ["opentelemetry_0_24"] }
 reqwest-middleware = "0.3.0"
 reqwest-retry = "0.5"
 routerify = "3"
@@ -174,11 +178,11 @@ tokio-tar = "0.3"
 tokio-util = { version = "0.7.10", features = ["io", "rt"] }
 toml = "0.8"
 toml_edit = "0.22"
-tonic = {version = "0.9", features = ["tls", "tls-roots"]}
+tonic = {version = "0.12.3", features = ["tls", "tls-roots"]}
 tower-service = "0.3.2"
 tracing = "0.1"
-tracing-error = "0.2.0"
-tracing-opentelemetry = "0.21.0"
+tracing-error = "0.2"
+tracing-opentelemetry = "0.25"
 tracing-subscriber = { version = "0.3", default-features = false, features = ["smallvec", "fmt", "tracing-log", "std", "env-filter", "json"] }
 try-lock = "0.2.5"
 twox-hash = { version = "1.6.3", default-features = false }
@@ -242,7 +246,7 @@ criterion = "0.5.1"
 rcgen = "0.12"
 rstest = "0.18"
 camino-tempfile = "1.0.2"
-tonic-build = "0.9"
+tonic-build = "0.12"
 
 [patch.crates-io]
 

Dockerfile.build-tools

@@ -13,6 +13,9 @@ RUN useradd -ms /bin/bash nonroot -b /home
 SHELL ["/bin/bash", "-c"]
 
 # System deps
+#
+# 'gdb' is included so that we get backtraces of core dumps produced in
+# regression tests
 RUN set -e \
     && apt update \
     && apt install -y \
@@ -24,6 +27,7 @@ RUN set -e \
         cmake \
         curl \
         flex \
+        gdb \
         git \
         gnupg \
         gzip \

README.md

@@ -58,7 +58,7 @@ curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
 1. Install XCode and dependencies
 ```
 xcode-select --install
-brew install protobuf openssl flex bison icu4c pkg-config
+brew install protobuf openssl flex bison icu4c pkg-config m4
 
 # add openssl to PATH, required for ed25519 keys generation in neon_local
 echo 'export PATH="$(brew --prefix openssl)/bin:$PATH"' >> ~/.zshrc

compute/Dockerfile.compute-node

@@ -12,10 +12,25 @@ ARG DEBIAN_FLAVOR=bullseye-slim
 #########################################################################################
 FROM debian:$DEBIAN_FLAVOR AS build-deps
 ARG DEBIAN_FLAVOR
-RUN apt update &&  \
-    apt install -y git autoconf automake libtool build-essential bison flex libreadline-dev \
-    zlib1g-dev libxml2-dev libcurl4-openssl-dev libossp-uuid-dev wget pkg-config libssl-dev \
-    libicu-dev libxslt1-dev liblz4-dev libzstd-dev zstd
+
+RUN case $DEBIAN_FLAVOR in \
+      # Version-specific installs for Bullseye (PG14-PG16):
+      # The h3_pg extension needs a cmake 3.20+, but Debian bullseye has 3.18.
+      # Install newer version (3.25) from backports.
+      bullseye*) \
+        echo "deb http://deb.debian.org/debian bullseye-backports main" > /etc/apt/sources.list.d/bullseye-backports.list; \
+        VERSION_INSTALLS="cmake/bullseye-backports cmake-data/bullseye-backports"; \
+      ;; \
+      # Version-specific installs for Bookworm (PG17):
+      bookworm*) \
+        VERSION_INSTALLS="cmake"; \
+      ;; \
+    esac && \
+    apt update &&  \
+    apt install --no-install-recommends -y git autoconf automake libtool build-essential bison flex libreadline-dev \
+    zlib1g-dev libxml2-dev libcurl4-openssl-dev libossp-uuid-dev wget ca-certificates pkg-config libssl-dev \
+    libicu-dev libxslt1-dev liblz4-dev libzstd-dev zstd \
+    $VERSION_INSTALLS
 
 #########################################################################################
 #
@@ -89,7 +104,7 @@ FROM build-deps AS postgis-build
 ARG PG_VERSION
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
 RUN apt update && \
-    apt install -y cmake gdal-bin libboost-dev libboost-thread-dev libboost-filesystem-dev \
+    apt install --no-install-recommends -y gdal-bin libboost-dev libboost-thread-dev libboost-filesystem-dev \
     libboost-system-dev libboost-iostreams-dev libboost-program-options-dev libboost-timer-dev \
     libcgal-dev libgdal-dev libgmp-dev libmpfr-dev libopenscenegraph-dev libprotobuf-c-dev \
     protobuf-c-compiler xsltproc
@@ -167,7 +182,7 @@ RUN case "${PG_VERSION}" in "v17") \
     echo "v17 extensions are not supported yet. Quit" && exit 0;; \
     esac && \
     apt update && \
-    apt install -y ninja-build python3-dev libncurses5 binutils clang
+    apt install --no-install-recommends -y ninja-build python3-dev libncurses5 binutils clang
 
 RUN case "${PG_VERSION}" in "v17") \
     echo "v17 extensions are not supported yet. Quit" && exit 0;; \
@@ -200,27 +215,6 @@ FROM build-deps AS h3-pg-build
 ARG PG_VERSION
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
 
-RUN case "${PG_VERSION}" in "v17") \
-    echo "v17 extensions are not supported yet. Quit" && exit 0;; \
-    esac && \
-    case "$(uname -m)" in \
-      "x86_64") \
-        export CMAKE_CHECKSUM=739d372726cb23129d57a539ce1432453448816e345e1545f6127296926b6754 \
-        ;; \
-      "aarch64") \
-        export CMAKE_CHECKSUM=281b42627c9a1beed03e29706574d04c6c53fae4994472e90985ef018dd29c02 \
-        ;; \
-      *) \
-        echo "Unsupported architecture '$(uname -m)'. Supported are x86_64 and aarch64" && exit 1 \
-        ;; \
-    esac && \
-    wget https://github.com/Kitware/CMake/releases/download/v3.24.2/cmake-3.24.2-linux-$(uname -m).sh \
-      -q -O /tmp/cmake-install.sh \
-      && echo "${CMAKE_CHECKSUM} /tmp/cmake-install.sh" | sha256sum --check \
-      && chmod u+x /tmp/cmake-install.sh \
-      && /tmp/cmake-install.sh --skip-license --prefix=/usr/local/ \
-      && rm /tmp/cmake-install.sh
-
 RUN case "${PG_VERSION}" in "v17") \
         mkdir -p /h3/usr/ && \
     echo "v17 extensions are not supported yet. Quit" && exit 0;; \
@@ -506,8 +500,6 @@ RUN case "${PG_VERSION}" in "v17") \
         export TIMESCALEDB_CHECKSUM=584a351c7775f0e067eaa0e7277ea88cab9077cc4c455cbbf09a5d9723dce95d \
         ;; \
     esac && \
-    apt-get update && \
-    apt-get install -y cmake && \
     wget https://github.com/timescale/timescaledb/archive/refs/tags/${TIMESCALEDB_VERSION}.tar.gz -O timescaledb.tar.gz && \
     echo "${TIMESCALEDB_CHECKSUM} timescaledb.tar.gz" | sha256sum --check && \
     mkdir timescaledb-src && cd timescaledb-src && tar xzf ../timescaledb.tar.gz --strip-components=1 -C . && \
@@ -595,8 +587,7 @@ RUN case "${PG_VERSION}" in "v17") \
     echo "v17 extensions are not supported yet. Quit" && exit 0;; \
     esac && \
     apt-get update && \
-    apt-get install -y \
-        cmake \
+    apt-get install --no-install-recommends -y \
         libboost-iostreams1.74-dev \
         libboost-regex1.74-dev \
         libboost-serialization1.74-dev \
@@ -761,7 +752,7 @@ ARG PG_VERSION
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
 
 RUN apt-get update && \
-    apt-get install -y curl libclang-dev cmake && \
+    apt-get install --no-install-recommends -y curl libclang-dev && \
     useradd -ms /bin/bash nonroot -b /home
 
 ENV HOME=/home/nonroot
@@ -871,6 +862,25 @@ RUN case "${PG_VERSION}" in "v17") \
     cargo pgrx install --release && \
     echo "trusted = true" >> /usr/local/pgsql/share/extension/ulid.control
 
+#########################################################################################
+#
+# Layer "pg-session-jwt-build"
+# Compile "pg_session_jwt" extension
+#
+#########################################################################################
+
+FROM rust-extensions-build AS pg-session-jwt-build
+ARG PG_VERSION
+
+RUN case "${PG_VERSION}" in "v17") \
+    echo "pg_session_jwt does not yet have a release that supports pg17" && exit 0;; \
+    esac && \
+    wget https://github.com/neondatabase/pg_session_jwt/archive/ff0a72440e8ff584dab24b3f9b7c00c56c660b8e.tar.gz -O pg_session_jwt.tar.gz && \
+    echo "1fbb2b5a339263bcf6daa847fad8bccbc0b451cea6a62e6d3bf232b0087f05cb pg_session_jwt.tar.gz" | sha256sum --check && \
+    mkdir pg_session_jwt-src && cd pg_session_jwt-src && tar xzf ../pg_session_jwt.tar.gz --strip-components=1 -C . && \
+    sed -i 's/pgrx = "=0.11.3"/pgrx = { version = "=0.11.3", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
+    cargo pgrx install --release
+
 #########################################################################################
 #
 # Layer "wal2json-build"
@@ -967,6 +977,7 @@ COPY --from=timescaledb-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-hint-plan-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-cron-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-pgx-ulid-build /usr/local/pgsql/ /usr/local/pgsql/
+COPY --from=pg-session-jwt-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=rdkit-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-uuidv7-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-roaringbitmap-pg-build /usr/local/pgsql/ /usr/local/pgsql/
@@ -1044,9 +1055,12 @@ FROM debian:$DEBIAN_FLAVOR AS pgbouncer
 ARG DEBIAN_FLAVOR
 RUN set -e \
     && apt-get update \
-    && apt-get install -y \
+    && apt-get install --no-install-recommends -y \
         build-essential \
         git \
+        ca-certificates \
+        autoconf \
+        automake \
         libevent-dev \
         libtool \
         pkg-config
@@ -1061,6 +1075,20 @@ RUN set -e \
     && make -j $(nproc) dist_man_MANS= \
     && make install dist_man_MANS=
 
+#########################################################################################
+#
+# Compile the Neon-specific `local_proxy` binary
+#
+#########################################################################################
+FROM $REPOSITORY/$IMAGE:$TAG AS local_proxy
+ARG BUILD_TAG
+ENV BUILD_TAG=$BUILD_TAG
+
+USER nonroot
+# Copy entire project to get Cargo.* files with proper dependencies for the whole project
+COPY --chown=nonroot . .
+RUN mold -run cargo build --locked --profile release-line-debug-size-lto --bin local_proxy
+
 #########################################################################################
 #
 # Layers "postgres-exporter" and "sql-exporter"
@@ -1154,11 +1182,6 @@ RUN case "${PG_VERSION}" in "v17") \
     echo "v17 extensions are not supported yet. Quit" && exit 0;; \
     esac && \
     cd /ext-src/pgvector-src && patch -p1 <../pgvector.patch
-# cmake is required for the h3 test
-RUN case "${PG_VERSION}" in "v17") \
-    echo "v17 extensions are not supported yet. Quit" && exit 0;; \
-    esac && \
-    apt-get update && apt-get install -y cmake
 RUN case "${PG_VERSION}" in "v17") \
     echo "v17 extensions are not supported yet. Quit" && exit 0;; \
     esac && \
@@ -1185,7 +1208,6 @@ ENV PGDATABASE=postgres
 #########################################################################################
 FROM debian:$DEBIAN_FLAVOR
 ARG DEBIAN_FLAVOR
-ENV DEBIAN_FLAVOR=$DEBIAN_FLAVOR
 # Add user postgres
 RUN mkdir /var/db && useradd -m -d /var/db/postgres postgres && \
     echo "postgres:test_console_pass" | chpasswd && \
@@ -1205,6 +1227,10 @@ COPY --from=compute-tools --chown=postgres /home/nonroot/target/release-line-deb
 COPY --from=pgbouncer         /usr/local/pgbouncer/bin/pgbouncer /usr/local/bin/pgbouncer
 COPY --chmod=0666 --chown=postgres compute/etc/pgbouncer.ini /etc/pgbouncer.ini
 
+# local_proxy and its config
+COPY --from=local_proxy --chown=postgres /home/nonroot/target/release-line-debug-size-lto/local_proxy /usr/local/bin/local_proxy
+RUN mkdir -p /etc/local_proxy && chown postgres:postgres /etc/local_proxy
+
 # Metrics exporter binaries and  configuration files
 COPY --from=postgres-exporter /bin/postgres_exporter /bin/postgres_exporter
 COPY --from=sql-exporter      /bin/sql_exporter      /bin/sql_exporter
@@ -1258,7 +1284,7 @@ RUN apt update && \
         libxml2 \
         libxslt1.1 \
         libzstd1 \
-        libcurl4-openssl-dev \
+        libcurl4 \
         locales \
         procps \
         ca-certificates \

compute/etc/neon_collector.yml

@@ -94,6 +94,68 @@ metrics:
   query: |
     select sum(pg_database_size(datname)) as total from pg_database;
 
+- metric_name: getpage_wait_seconds_count
+  type: counter
+  help: 'Number of getpage requests'
+  values: [getpage_wait_seconds_count]
+  query_ref: neon_perf_counters
+
+- metric_name: getpage_wait_seconds_sum
+  type: counter
+  help: 'Time spent in getpage requests'
+  values: [getpage_wait_seconds_sum]
+  query_ref: neon_perf_counters
+
+- metric_name: getpage_prefetch_requests_total
+  type: counter
+  help: 'Number of getpage issued for prefetching'
+  values: [getpage_prefetch_requests_total]
+  query_ref: neon_perf_counters
+
+- metric_name: getpage_sync_requests_total
+  type: counter
+  help: 'Number of synchronous getpage issued'
+  values: [getpage_sync_requests_total]
+  query_ref: neon_perf_counters
+
+- metric_name: getpage_prefetch_misses_total
+  type: counter
+  help: 'Total number of readahead misses; consisting of either prefetches that don''t satisfy the LSN bounds once the prefetch got read by the backend, or cases where somehow no readahead was issued for the read'
+  values: [getpage_prefetch_misses_total]
+  query_ref: neon_perf_counters
+
+- metric_name: getpage_prefetch_discards_total
+  type: counter
+  help: 'Number of prefetch responses issued but not used'
+  values: [getpage_prefetch_discards_total]
+  query_ref: neon_perf_counters
+
+- metric_name: pageserver_requests_sent_total
+  type: counter
+  help: 'Number of all requests sent to the pageserver (not just GetPage requests)'
+  values: [pageserver_requests_sent_total]
+  query_ref: neon_perf_counters
+
+- metric_name: pageserver_disconnects_total
+  type: counter
+  help: 'Number of times that the connection to the pageserver was lost'
+  values: [pageserver_disconnects_total]
+  query_ref: neon_perf_counters
+
+- metric_name: pageserver_send_flushes_total
+  type: counter
+  help: 'Number of flushes to the pageserver connection'
+  values: [pageserver_send_flushes_total]
+  query_ref: neon_perf_counters
+
+- metric_name: getpage_wait_seconds_bucket
+  type: counter
+  help: 'Histogram buckets of getpage request latency'
+  key_labels:
+      - bucket_le
+  values: [value]
+  query_ref: getpage_wait_seconds_buckets
+
 # DEPRECATED
 - metric_name: lfc_approximate_working_set_size
   type: gauge
@@ -244,3 +306,26 @@ metrics:
     SELECT slot_name,
            CASE WHEN wal_status = 'lost' THEN 1 ELSE 0 END AS wal_is_lost
     FROM pg_replication_slots;
+
+queries:
+  - query_name: neon_perf_counters
+    query: |
+      WITH c AS (
+        SELECT pg_catalog.jsonb_object_agg(metric, value) jb FROM neon.neon_perf_counters
+      )
+      SELECT d.*
+      FROM pg_catalog.jsonb_to_record((select jb from c)) as d(
+          getpage_wait_seconds_count numeric,
+          getpage_wait_seconds_sum numeric,
+          getpage_prefetch_requests_total numeric,
+          getpage_sync_requests_total numeric,
+          getpage_prefetch_misses_total numeric,
+          getpage_prefetch_discards_total numeric,
+          pageserver_requests_sent_total numeric,
+          pageserver_disconnects_total numeric,
+          pageserver_send_flushes_total numeric
+      );
+
+  - query_name: getpage_wait_seconds_buckets
+    query: |
+      SELECT bucket_le, value FROM neon.neon_perf_counters WHERE metric = 'getpage_wait_seconds_bucket';

compute/vm-image-spec.yaml

@@ -11,10 +11,18 @@ commands:
     user: root
     sysvInitAction: sysinit
     shell: 'chmod 711 /neonvm/bin/resize-swap'
+  - name: chmod-set-disk-quota
+    user: root
+    sysvInitAction: sysinit
+    shell: 'chmod 711 /neonvm/bin/set-disk-quota'
   - name: pgbouncer
     user: postgres
     sysvInitAction: respawn
     shell: '/usr/local/bin/pgbouncer /etc/pgbouncer.ini'
+  - name: local_proxy
+    user: postgres
+    sysvInitAction: respawn
+    shell: '/usr/local/bin/local_proxy --config-path /etc/local_proxy/config.json --pid-path /etc/local_proxy/pid --http 0.0.0.0:10432'
   - name: postgres-exporter
     user: nobody
     sysvInitAction: respawn
@@ -30,11 +38,12 @@ commands:
 shutdownHook: |
   su -p postgres --session-command '/usr/local/bin/pg_ctl stop -D /var/db/postgres/compute/pgdata -m fast --wait -t 10'
 files:
-  - filename: compute_ctl-resize-swap
+  - filename: compute_ctl-sudoers
     content: |
       # Allow postgres user (which is what compute_ctl runs as) to run /neonvm/bin/resize-swap
-      # as root without requiring entering a password (NOPASSWD), regardless of hostname (ALL)
-      postgres ALL=(root) NOPASSWD: /neonvm/bin/resize-swap
+      # and /neonvm/bin/set-disk-quota as root without requiring entering a password (NOPASSWD),
+      # regardless of hostname (ALL)
+      postgres ALL=(root) NOPASSWD: /neonvm/bin/resize-swap, /neonvm/bin/set-disk-quota
   - filename: cgconfig.conf
     content: |
       # Configuration for cgroups in VM compute nodes
@@ -100,7 +109,7 @@ merge: |
       && apt install --no-install-recommends -y \
              sudo \
       && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
-  COPY compute_ctl-resize-swap /etc/sudoers.d/compute_ctl-resize-swap
+  COPY compute_ctl-sudoers /etc/sudoers.d/compute_ctl-sudoers
 
   COPY cgconfig.conf /etc/cgconfig.conf
 

compute_tools/Cargo.toml

@@ -11,16 +11,18 @@ testing = []
 
 [dependencies]
 anyhow.workspace = true
+camino.workspace = true
 chrono.workspace = true
 cfg-if.workspace = true
 clap.workspace = true
 flate2.workspace = true
 futures.workspace = true
-hyper = { workspace = true, features = ["full"] }
+hyper0 = { workspace = true, features = ["full"] }
 nix.workspace = true
 notify.workspace = true
 num_cpus.workspace = true
 opentelemetry.workspace = true
+opentelemetry_sdk.workspace = true
 postgres.workspace = true
 regex.workspace = true
 serde_json.workspace = true

compute_tools/src/bin/compute_ctl.rs

@@ -44,6 +44,7 @@ use std::{thread, time::Duration};
 use anyhow::{Context, Result};
 use chrono::Utc;
 use clap::Arg;
+use compute_tools::disk_quota::set_disk_quota;
 use compute_tools::lsn_lease::launch_lsn_lease_bg_task_for_static;
 use signal_hook::consts::{SIGQUIT, SIGTERM};
 use signal_hook::{consts::SIGINT, iterator::Signals};
@@ -151,6 +152,7 @@ fn process_cli(matches: &clap::ArgMatches) -> Result<ProcessCliResult> {
     let spec_json = matches.get_one::<String>("spec");
     let spec_path = matches.get_one::<String>("spec-path");
     let resize_swap_on_bind = matches.get_flag("resize-swap-on-bind");
+    let set_disk_quota_for_fs = matches.get_one::<String>("set-disk-quota-for-fs");
 
     Ok(ProcessCliResult {
         connstr,
@@ -161,6 +163,7 @@ fn process_cli(matches: &clap::ArgMatches) -> Result<ProcessCliResult> {
         spec_json,
         spec_path,
         resize_swap_on_bind,
+        set_disk_quota_for_fs,
     })
 }
 
@@ -173,6 +176,7 @@ struct ProcessCliResult<'clap> {
     spec_json: Option<&'clap String>,
     spec_path: Option<&'clap String>,
     resize_swap_on_bind: bool,
+    set_disk_quota_for_fs: Option<&'clap String>,
 }
 
 fn startup_context_from_env() -> Option<opentelemetry::ContextGuard> {
@@ -214,7 +218,7 @@ fn startup_context_from_env() -> Option<opentelemetry::ContextGuard> {
     }
     if !startup_tracing_carrier.is_empty() {
         use opentelemetry::propagation::TextMapPropagator;
-        use opentelemetry::sdk::propagation::TraceContextPropagator;
+        use opentelemetry_sdk::propagation::TraceContextPropagator;
         let guard = TraceContextPropagator::new()
             .extract(&startup_tracing_carrier)
             .attach();
@@ -293,6 +297,7 @@ fn wait_spec(
         pgbin,
         ext_remote_storage,
         resize_swap_on_bind,
+        set_disk_quota_for_fs,
         http_port,
         ..
     }: ProcessCliResult,
@@ -373,6 +378,7 @@ fn wait_spec(
         compute,
         http_port,
         resize_swap_on_bind,
+        set_disk_quota_for_fs: set_disk_quota_for_fs.cloned(),
     })
 }
 
@@ -381,6 +387,7 @@ struct WaitSpecResult {
     // passed through from ProcessCliResult
     http_port: u16,
     resize_swap_on_bind: bool,
+    set_disk_quota_for_fs: Option<String>,
 }
 
 fn start_postgres(
@@ -390,12 +397,12 @@ fn start_postgres(
         compute,
         http_port,
         resize_swap_on_bind,
+        set_disk_quota_for_fs,
     }: WaitSpecResult,
 ) -> Result<(Option<PostgresHandle>, StartPostgresResult)> {
     // We got all we need, update the state.
     let mut state = compute.state.lock().unwrap();
-    state.status = ComputeStatus::Init;
-    compute.state_changed.notify_all();
+    state.set_status(ComputeStatus::Init, &compute.state_changed);
 
     info!(
         "running compute with features: {:?}",
@@ -403,6 +410,7 @@ fn start_postgres(
     );
     // before we release the mutex, fetch the swap size (if any) for later.
     let swap_size_bytes = state.pspec.as_ref().unwrap().spec.swap_size_bytes;
+    let disk_quota_bytes = state.pspec.as_ref().unwrap().spec.disk_quota_bytes;
     drop(state);
 
     // Launch remaining service threads
@@ -422,8 +430,8 @@ fn start_postgres(
         // OOM-killed during startup because swap wasn't available yet.
         match resize_swap(size_bytes) {
             Ok(()) => {
-                let size_gib = size_bytes as f32 / (1 << 20) as f32; // just for more coherent display.
-                info!(%size_bytes, %size_gib, "resized swap");
+                let size_mib = size_bytes as f32 / (1 << 20) as f32; // just for more coherent display.
+                info!(%size_bytes, %size_mib, "resized swap");
             }
             Err(err) => {
                 let err = err.context("failed to resize swap");
@@ -432,10 +440,29 @@ fn start_postgres(
                 // Mark compute startup as failed; don't try to start postgres, and report this
                 // error to the control plane when it next asks.
                 prestartup_failed = true;
-                let mut state = compute.state.lock().unwrap();
-                state.error = Some(format!("{err:?}"));
-                state.status = ComputeStatus::Failed;
-                compute.state_changed.notify_all();
+                compute.set_failed_status(err);
+                delay_exit = true;
+            }
+        }
+    }
+
+    // Set disk quota if the compute spec says so
+    if let (Some(disk_quota_bytes), Some(disk_quota_fs_mountpoint)) =
+        (disk_quota_bytes, set_disk_quota_for_fs)
+    {
+        match set_disk_quota(disk_quota_bytes, &disk_quota_fs_mountpoint) {
+            Ok(()) => {
+                let size_mib = disk_quota_bytes as f32 / (1 << 20) as f32; // just for more coherent display.
+                info!(%disk_quota_bytes, %size_mib, "set disk quota");
+            }
+            Err(err) => {
+                let err = err.context("failed to set disk quota");
+                error!("{err:#}");
+
+                // Mark compute startup as failed; don't try to start postgres, and report this
+                // error to the control plane when it next asks.
+                prestartup_failed = true;
+                compute.set_failed_status(err);
                 delay_exit = true;
             }
         }
@@ -450,16 +477,7 @@ fn start_postgres(
             Ok(pg) => Some(pg),
             Err(err) => {
                 error!("could not start the compute node: {:#}", err);
-                let mut state = compute.state.lock().unwrap();
-                state.error = Some(format!("{:?}", err));
-                state.status = ComputeStatus::Failed;
-                // Notify others that Postgres failed to start. In case of configuring the
-                // empty compute, it's likely that API handler is still waiting for compute
-                // state change. With this we will notify it that compute is in Failed state,
-                // so control plane will know about it earlier and record proper error instead
-                // of timeout.
-                compute.state_changed.notify_all();
-                drop(state); // unlock
+                compute.set_failed_status(err);
                 delay_exit = true;
                 None
             }
@@ -750,6 +768,11 @@ fn cli() -> clap::Command {
                 .long("resize-swap-on-bind")
                 .action(clap::ArgAction::SetTrue),
         )
+        .arg(
+            Arg::new("set-disk-quota-for-fs")
+                .long("set-disk-quota-for-fs")
+                .value_name("SET_DISK_QUOTA_FOR_FS")
+        )
 }
 
 /// When compute_ctl is killed, send also termination signal to sync-safekeepers

compute_tools/src/compute.rs

@@ -10,6 +10,7 @@ use std::sync::atomic::AtomicU32;
 use std::sync::atomic::Ordering;
 use std::sync::{Condvar, Mutex, RwLock};
 use std::thread;
+use std::time::Duration;
 use std::time::Instant;
 
 use anyhow::{Context, Result};
@@ -33,6 +34,7 @@ use nix::sys::signal::{kill, Signal};
 use remote_storage::{DownloadError, RemotePath};
 
 use crate::checker::create_availability_check_data;
+use crate::local_proxy;
 use crate::logger::inlinify;
 use crate::pg_helpers::*;
 use crate::spec::*;
@@ -107,6 +109,18 @@ impl ComputeState {
             metrics: ComputeMetrics::default(),
         }
     }
+
+    pub fn set_status(&mut self, status: ComputeStatus, state_changed: &Condvar) {
+        let prev = self.status;
+        info!("Changing compute status from {} to {}", prev, status);
+        self.status = status;
+        state_changed.notify_all();
+    }
+
+    pub fn set_failed_status(&mut self, err: anyhow::Error, state_changed: &Condvar) {
+        self.error = Some(format!("{err:?}"));
+        self.set_status(ComputeStatus::Failed, state_changed);
+    }
 }
 
 impl Default for ComputeState {
@@ -301,8 +315,12 @@ impl ComputeNode {
 
     pub fn set_status(&self, status: ComputeStatus) {
         let mut state = self.state.lock().unwrap();
-        state.status = status;
-        self.state_changed.notify_all();
+        state.set_status(status, &self.state_changed);
+    }
+
+    pub fn set_failed_status(&self, err: anyhow::Error) {
+        let mut state = self.state.lock().unwrap();
+        state.set_failed_status(err, &self.state_changed);
     }
 
     pub fn get_status(&self) -> ComputeStatus {
@@ -710,7 +728,7 @@ impl ComputeNode {
         info!("running initdb");
         let initdb_bin = Path::new(&self.pgbin).parent().unwrap().join("initdb");
         Command::new(initdb_bin)
-            .args(["-D", pgdata])
+            .args(["--pgdata", pgdata])
             .output()
             .expect("cannot start initdb process");
 
@@ -878,6 +896,11 @@ impl ComputeNode {
         // 'Close' connection
         drop(client);
 
+        if let Some(ref local_proxy) = spec.local_proxy_config {
+            info!("configuring local_proxy");
+            local_proxy::configure(local_proxy).context("apply_config local_proxy")?;
+        }
+
         // Run migrations separately to not hold up cold starts
         thread::spawn(move || {
             let mut connstr = connstr.clone();
@@ -928,6 +951,19 @@ impl ComputeNode {
             });
         }
 
+        if let Some(ref local_proxy) = spec.local_proxy_config {
+            info!("configuring local_proxy");
+
+            // Spawn a thread to do the configuration,
+            // so that we don't block the main thread that starts Postgres.
+            let local_proxy = local_proxy.clone();
+            let _handle = Some(thread::spawn(move || {
+                if let Err(err) = local_proxy::configure(&local_proxy) {
+                    error!("error while configuring local_proxy: {err:?}");
+                }
+            }));
+        }
+
         // Write new config
         let pgdata_path = Path::new(&self.pgdata);
         let postgresql_conf_path = pgdata_path.join("postgresql.conf");
@@ -1015,6 +1051,19 @@ impl ComputeNode {
             });
         }
 
+        if let Some(local_proxy) = &pspec.spec.local_proxy_config {
+            info!("configuring local_proxy");
+
+            // Spawn a thread to do the configuration,
+            // so that we don't block the main thread that starts Postgres.
+            let local_proxy = local_proxy.clone();
+            let _handle = thread::spawn(move || {
+                if let Err(err) = local_proxy::configure(&local_proxy) {
+                    error!("error while configuring local_proxy: {err:?}");
+                }
+            });
+        }
+
         info!(
             "start_compute spec.remote_extensions {:?}",
             pspec.spec.remote_extensions
@@ -1052,19 +1101,26 @@ impl ComputeNode {
         let pg_process = self.start_postgres(pspec.storage_auth_token.clone())?;
 
         let config_time = Utc::now();
-        if pspec.spec.mode == ComputeMode::Primary && !pspec.spec.skip_pg_catalog_updates {
-            let pgdata_path = Path::new(&self.pgdata);
-            // temporarily reset max_cluster_size in config
-            // to avoid the possibility of hitting the limit, while we are applying config:
-            // creating new extensions, roles, etc...
-            config::with_compute_ctl_tmp_override(pgdata_path, "neon.max_cluster_size=-1", || {
+        if pspec.spec.mode == ComputeMode::Primary {
+            if !pspec.spec.skip_pg_catalog_updates {
+                let pgdata_path = Path::new(&self.pgdata);
+                // temporarily reset max_cluster_size in config
+                // to avoid the possibility of hitting the limit, while we are applying config:
+                // creating new extensions, roles, etc...
+                config::with_compute_ctl_tmp_override(
+                    pgdata_path,
+                    "neon.max_cluster_size=-1",
+                    || {
+                        self.pg_reload_conf()?;
+
+                        self.apply_config(&compute_state)?;
+
+                        Ok(())
+                    },
+                )?;
                 self.pg_reload_conf()?;
-
-                self.apply_config(&compute_state)?;
-
-                Ok(())
-            })?;
-            self.pg_reload_conf()?;
+            }
+            self.post_apply_config()?;
         }
 
         let startup_end_time = Utc::now();
@@ -1123,6 +1179,9 @@ impl ComputeNode {
     //
     // Use that as a default location and pattern, except macos where core dumps are written
     // to /cores/ directory by default.
+    //
+    // With default Linux settings, the core dump file is called just "core", so check for
+    // that too.
     pub fn check_for_core_dumps(&self) -> Result<()> {
         let core_dump_dir = match std::env::consts::OS {
             "macos" => Path::new("/cores/"),
@@ -1134,8 +1193,17 @@ impl ComputeNode {
         let files = fs::read_dir(core_dump_dir)?;
         let cores = files.filter_map(|entry| {
             let entry = entry.ok()?;
-            let _ = entry.file_name().to_str()?.strip_prefix("core.")?;
-            Some(entry.path())
+
+            let is_core_dump = match entry.file_name().to_str()? {
+                n if n.starts_with("core.") => true,
+                "core" => true,
+                _ => false,
+            };
+            if is_core_dump {
+                Some(entry.path())
+            } else {
+                None
+            }
         });
 
         // Print backtrace for each core dump
@@ -1386,6 +1454,36 @@ LIMIT 100",
         }
         Ok(remote_ext_metrics)
     }
+
+    /// Waits until current thread receives a state changed notification and
+    /// the pageserver connection strings has changed.
+    ///
+    /// The operation will time out after a specified duration.
+    pub fn wait_timeout_while_pageserver_connstr_unchanged(&self, duration: Duration) {
+        let state = self.state.lock().unwrap();
+        let old_pageserver_connstr = state
+            .pspec
+            .as_ref()
+            .expect("spec must be set")
+            .pageserver_connstr
+            .clone();
+        let mut unchanged = true;
+        let _ = self
+            .state_changed
+            .wait_timeout_while(state, duration, |s| {
+                let pageserver_connstr = &s
+                    .pspec
+                    .as_ref()
+                    .expect("spec must be set")
+                    .pageserver_connstr;
+                unchanged = pageserver_connstr == &old_pageserver_connstr;
+                unchanged
+            })
+            .unwrap();
+        if !unchanged {
+            info!("Pageserver config changed");
+        }
+    }
 }
 
 pub fn forward_termination_signal() {

compute_tools/src/configurator.rs

@@ -24,8 +24,7 @@ fn configurator_main_loop(compute: &Arc<ComputeNode>) {
         // Re-check the status after waking up
         if state.status == ComputeStatus::ConfigurationPending {
             info!("got configuration request");
-            state.status = ComputeStatus::Configuration;
-            compute.state_changed.notify_all();
+            state.set_status(ComputeStatus::Configuration, &compute.state_changed);
             drop(state);
 
             let mut new_status = ComputeStatus::Failed;

compute_tools/src/disk_quota.rs

@@ -0,0 +1,25 @@
+use anyhow::Context;
+
+pub const DISK_QUOTA_BIN: &str = "/neonvm/bin/set-disk-quota";
+
+/// If size_bytes is 0, it disables the quota. Otherwise, it sets filesystem quota to size_bytes.
+/// `fs_mountpoint` should point to the mountpoint of the filesystem where the quota should be set.
+pub fn set_disk_quota(size_bytes: u64, fs_mountpoint: &str) -> anyhow::Result<()> {
+    let size_kb = size_bytes / 1024;
+    // run `/neonvm/bin/set-disk-quota {size_kb} {mountpoint}`
+    let child_result = std::process::Command::new("/usr/bin/sudo")
+        .arg(DISK_QUOTA_BIN)
+        .arg(size_kb.to_string())
+        .arg(fs_mountpoint)
+        .spawn();
+
+    child_result
+        .context("spawn() failed")
+        .and_then(|mut child| child.wait().context("wait() failed"))
+        .and_then(|status| match status.success() {
+            true => Ok(()),
+            false => Err(anyhow::anyhow!("process exited with {status}")),
+        })
+        // wrap any prior error with the overall context that we couldn't run the command
+        .with_context(|| format!("could not run `/usr/bin/sudo {DISK_QUOTA_BIN}`"))
+}

compute_tools/src/http/api.rs

@@ -288,8 +288,7 @@ async fn handle_configure_request(
                 return Err((msg, StatusCode::PRECONDITION_FAILED));
             }
             state.pspec = Some(parsed_spec);
-            state.status = ComputeStatus::ConfigurationPending;
-            compute.state_changed.notify_all();
+            state.set_status(ComputeStatus::ConfigurationPending, &compute.state_changed);
             drop(state);
             info!("set new spec and notified waiters");
         }
@@ -362,15 +361,15 @@ async fn handle_terminate_request(compute: &Arc<ComputeNode>) -> Result<(), (Str
         }
         if state.status != ComputeStatus::Empty && state.status != ComputeStatus::Running {
             let msg = format!(
-                "invalid compute status for termination request: {:?}",
-                state.status.clone()
+                "invalid compute status for termination request: {}",
+                state.status
             );
             return Err((msg, StatusCode::PRECONDITION_FAILED));
         }
-        state.status = ComputeStatus::TerminationPending;
-        compute.state_changed.notify_all();
+        state.set_status(ComputeStatus::TerminationPending, &compute.state_changed);
         drop(state);
     }
+
     forward_termination_signal();
     info!("sent signal and notified waiters");
 
@@ -384,7 +383,8 @@ async fn handle_terminate_request(compute: &Arc<ComputeNode>) -> Result<(), (Str
         while state.status != ComputeStatus::Terminated {
             state = c.state_changed.wait(state).unwrap();
             info!(
-                "waiting for compute to become Terminated, current status: {:?}",
+                "waiting for compute to become {}, current status: {:?}",
+                ComputeStatus::Terminated,
                 state.status
             );
         }

compute_tools/src/lib.rs

@@ -2,6 +2,9 @@
 //! configuration.
 #![deny(unsafe_code)]
 #![deny(clippy::undocumented_unsafe_blocks)]
+
+extern crate hyper0 as hyper;
+
 pub mod checker;
 pub mod config;
 pub mod configurator;
@@ -10,7 +13,9 @@ pub mod http;
 pub mod logger;
 pub mod catalog;
 pub mod compute;
+pub mod disk_quota;
 pub mod extension_server;
+pub mod local_proxy;
 pub mod lsn_lease;
 mod migration;
 pub mod monitor;

compute_tools/src/local_proxy.rs

@@ -0,0 +1,56 @@
+//! Local Proxy is a feature of our BaaS Neon Authorize project.
+//!
+//! Local Proxy validates JWTs and manages the pg_session_jwt extension.
+//! It also maintains a connection pool to postgres.
+
+use anyhow::{Context, Result};
+use camino::Utf8Path;
+use compute_api::spec::LocalProxySpec;
+use nix::sys::signal::Signal;
+use utils::pid_file::{self, PidFileRead};
+
+pub fn configure(local_proxy: &LocalProxySpec) -> Result<()> {
+    write_local_proxy_conf("/etc/local_proxy/config.json".as_ref(), local_proxy)?;
+    notify_local_proxy("/etc/local_proxy/pid".as_ref())?;
+
+    Ok(())
+}
+
+/// Create or completely rewrite configuration file specified by `path`
+fn write_local_proxy_conf(path: &Utf8Path, local_proxy: &LocalProxySpec) -> Result<()> {
+    let config =
+        serde_json::to_string_pretty(local_proxy).context("serializing LocalProxySpec to json")?;
+    std::fs::write(path, config).with_context(|| format!("writing {path}"))?;
+
+    Ok(())
+}
+
+/// Notify local proxy about a new config file.
+fn notify_local_proxy(path: &Utf8Path) -> Result<()> {
+    match pid_file::read(path)? {
+        // if the file doesn't exist, or isn't locked, local_proxy isn't running
+        // and will naturally pick up our config later
+        PidFileRead::NotExist | PidFileRead::NotHeldByAnyProcess(_) => {}
+        PidFileRead::LockedByOtherProcess(pid) => {
+            // From the pid_file docs:
+            //
+            // > 1. The other process might exit at any time, turning the given PID stale.
+            // > 2. There is a small window in which `claim_for_current_process` has already
+            // >    locked the file but not yet updates its contents. [`read`] will return
+            // >    this variant here, but with the old file contents, i.e., a stale PID.
+            // >
+            // > The kernel is free to recycle PID once it has been `wait(2)`ed upon by
+            // > its creator. Thus, acting upon a stale PID, e.g., by issuing a `kill`
+            // > system call on it, bears the risk of killing an unrelated process.
+            // > This is an inherent limitation of using pidfiles.
+            // > The only race-free solution is to have a supervisor-process with a lifetime
+            // > that exceeds that of all of its child-processes (e.g., `runit`, `supervisord`).
+            //
+            // This is an ok risk as we only send a SIGHUP which likely won't actually
+            // kill the process, only reload config.
+            nix::sys::signal::kill(pid, Signal::SIGHUP).context("sending signal to local_proxy")?;
+        }
+    }
+
+    Ok(())
+}

compute_tools/src/logger.rs

@@ -1,4 +1,3 @@
-use tracing_opentelemetry::OpenTelemetryLayer;
 use tracing_subscriber::layer::SubscriberExt;
 use tracing_subscriber::prelude::*;
 
@@ -23,8 +22,7 @@ pub fn init_tracing_and_logging(default_log_level: &str) -> anyhow::Result<()> {
         .with_writer(std::io::stderr);
 
     // Initialize OpenTelemetry
-    let otlp_layer =
-        tracing_utils::init_tracing_without_runtime("compute_ctl").map(OpenTelemetryLayer::new);
+    let otlp_layer = tracing_utils::init_tracing_without_runtime("compute_ctl");
 
     // Put it all together
     tracing_subscriber::registry()

compute_tools/src/lsn_lease.rs

@@ -57,10 +57,10 @@ fn lsn_lease_bg_task(
             .max(valid_duration / 2);
 
         info!(
-            "Succeeded, sleeping for {} seconds",
+            "Request succeeded, sleeping for {} seconds",
             sleep_duration.as_secs()
         );
-        thread::sleep(sleep_duration);
+        compute.wait_timeout_while_pageserver_connstr_unchanged(sleep_duration);
     }
 }
 
@@ -89,10 +89,7 @@ fn acquire_lsn_lease_with_retry(
                 .map(|connstr| {
                     let mut config = postgres::Config::from_str(connstr).expect("Invalid connstr");
                     if let Some(storage_auth_token) = &spec.storage_auth_token {
-                        info!("Got storage auth token from spec file");
                         config.password(storage_auth_token.clone());
-                    } else {
-                        info!("Storage auth token not set");
                     }
                     config
                 })
@@ -108,9 +105,11 @@ fn acquire_lsn_lease_with_retry(
                 bail!("Permanent error: lease could not be obtained, LSN is behind the GC cutoff");
             }
             Err(e) => {
-                warn!("Failed to acquire lsn lease: {e} (attempt {attempts}");
+                warn!("Failed to acquire lsn lease: {e} (attempt {attempts})");
 
-                thread::sleep(Duration::from_millis(retry_period_ms as u64));
+                compute.wait_timeout_while_pageserver_connstr_unchanged(Duration::from_millis(
+                    retry_period_ms as u64,
+                ));
                 retry_period_ms *= 1.5;
                 retry_period_ms = retry_period_ms.min(MAX_RETRY_PERIOD_MS);
             }

compute_tools/src/spec.rs

@@ -1,3 +1,4 @@
+use std::collections::HashSet;
 use std::fs::File;
 use std::path::Path;
 use std::str::FromStr;
@@ -189,6 +190,15 @@ pub fn handle_roles(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
     let mut xact = client.transaction()?;
     let existing_roles: Vec<Role> = get_existing_roles(&mut xact)?;
 
+    let mut jwks_roles = HashSet::new();
+    if let Some(local_proxy) = &spec.local_proxy_config {
+        for jwks_setting in local_proxy.jwks.iter().flatten() {
+            for role_name in &jwks_setting.role_names {
+                jwks_roles.insert(role_name.clone());
+            }
+        }
+    }
+
     // Print a list of existing Postgres roles (only in debug mode)
     if span_enabled!(Level::INFO) {
         let mut vec = Vec::new();
@@ -308,6 +318,9 @@ pub fn handle_roles(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
                     "CREATE ROLE {} INHERIT CREATEROLE CREATEDB BYPASSRLS REPLICATION IN ROLE neon_superuser",
                     name.pg_quote()
                 );
+                if jwks_roles.contains(name.as_str()) {
+                    query = format!("CREATE ROLE {}", name.pg_quote());
+                }
                 info!("running role create query: '{}'", &query);
                 query.push_str(&role.to_pg_options());
                 xact.execute(query.as_str(), &[])?;

control_plane/Cargo.toml

@@ -9,11 +9,12 @@ anyhow.workspace = true
 camino.workspace = true
 clap.workspace = true
 comfy-table.workspace = true
+futures.workspace = true
 humantime.workspace = true
 nix.workspace = true
 once_cell.workspace = true
 humantime-serde.workspace = true
-hyper.workspace = true
+hyper0.workspace = true
 regex.workspace = true
 reqwest = { workspace = true, features = ["blocking", "json"] }
 scopeguard.workspace = true

control_plane/src/branch_mappings.rs

@@ -0,0 +1,94 @@
+//! Branch mappings for convenience
+
+use std::collections::HashMap;
+use std::fs;
+use std::path::Path;
+
+use anyhow::{bail, Context};
+use serde::{Deserialize, Serialize};
+
+use utils::id::{TenantId, TenantTimelineId, TimelineId};
+
+/// Keep human-readable aliases in memory (and persist them to config XXX), to hide tenant/timeline hex strings from the user.
+#[derive(PartialEq, Eq, Clone, Debug, Default, Serialize, Deserialize)]
+#[serde(default, deny_unknown_fields)]
+pub struct BranchMappings {
+    /// Default tenant ID to use with the 'neon_local' command line utility, when
+    /// --tenant_id is not explicitly specified. This comes from the branches.
+    pub default_tenant_id: Option<TenantId>,
+
+    // A `HashMap<String, HashMap<TenantId, TimelineId>>` would be more appropriate here,
+    // but deserialization into a generic toml object as `toml::Value::try_from` fails with an error.
+    // https://toml.io/en/v1.0.0 does not contain a concept of "a table inside another table".
+    pub mappings: HashMap<String, Vec<(TenantId, TimelineId)>>,
+}
+
+impl BranchMappings {
+    pub fn register_branch_mapping(
+        &mut self,
+        branch_name: String,
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+    ) -> anyhow::Result<()> {
+        let existing_values = self.mappings.entry(branch_name.clone()).or_default();
+
+        let existing_ids = existing_values
+            .iter()
+            .find(|(existing_tenant_id, _)| existing_tenant_id == &tenant_id);
+
+        if let Some((_, old_timeline_id)) = existing_ids {
+            if old_timeline_id == &timeline_id {
+                Ok(())
+            } else {
+                bail!("branch '{branch_name}' is already mapped to timeline {old_timeline_id}, cannot map to another timeline {timeline_id}");
+            }
+        } else {
+            existing_values.push((tenant_id, timeline_id));
+            Ok(())
+        }
+    }
+
+    pub fn get_branch_timeline_id(
+        &self,
+        branch_name: &str,
+        tenant_id: TenantId,
+    ) -> Option<TimelineId> {
+        // If it looks like a timeline ID, return it as it is
+        if let Ok(timeline_id) = branch_name.parse::<TimelineId>() {
+            return Some(timeline_id);
+        }
+
+        self.mappings
+            .get(branch_name)?
+            .iter()
+            .find(|(mapped_tenant_id, _)| mapped_tenant_id == &tenant_id)
+            .map(|&(_, timeline_id)| timeline_id)
+            .map(TimelineId::from)
+    }
+
+    pub fn timeline_name_mappings(&self) -> HashMap<TenantTimelineId, String> {
+        self.mappings
+            .iter()
+            .flat_map(|(name, tenant_timelines)| {
+                tenant_timelines.iter().map(|&(tenant_id, timeline_id)| {
+                    (TenantTimelineId::new(tenant_id, timeline_id), name.clone())
+                })
+            })
+            .collect()
+    }
+
+    pub fn persist(&self, path: &Path) -> anyhow::Result<()> {
+        let content = &toml::to_string_pretty(self)?;
+        fs::write(path, content).with_context(|| {
+            format!(
+                "Failed to write branch information into path '{}'",
+                path.display()
+            )
+        })
+    }
+
+    pub fn load(path: &Path) -> anyhow::Result<BranchMappings> {
+        let branches_file_contents = fs::read_to_string(path)?;
+        Ok(toml::from_str(branches_file_contents.as_str())?)
+    }
+}

control_plane/src/endpoint.rs

@@ -561,6 +561,7 @@ impl Endpoint {
             operation_uuid: None,
             features: self.features.clone(),
             swap_size_bytes: None,
+            disk_quota_bytes: None,
             cluster: Cluster {
                 cluster_id: None, // project ID: not used
                 name: None,       // project name: not used
@@ -598,6 +599,7 @@ impl Endpoint {
             remote_extensions,
             pgbouncer_settings: None,
             shard_stripe_size: Some(shard_stripe_size),
+            local_proxy_config: None,
         };
         let spec_path = self.endpoint_path().join("spec.json");
         std::fs::write(spec_path, serde_json::to_string_pretty(&spec)?)?;

control_plane/src/local_env.rs

@@ -168,6 +168,9 @@ pub struct NeonStorageControllerConf {
 
     #[serde(with = "humantime_serde")]
     pub heartbeat_interval: Duration,
+
+    #[serde(with = "humantime_serde")]
+    pub long_reconcile_threshold: Option<Duration>,
 }
 
 impl NeonStorageControllerConf {
@@ -190,6 +193,7 @@ impl Default for NeonStorageControllerConf {
             split_threshold: None,
             max_secondary_lag_bytes: None,
             heartbeat_interval: Self::DEFAULT_HEARTBEAT_INTERVAL,
+            long_reconcile_threshold: None,
         }
     }
 }

control_plane/src/safekeeper.rs

@@ -113,7 +113,7 @@ impl SafekeeperNode {
 
     pub async fn start(
         &self,
-        extra_opts: Vec<String>,
+        extra_opts: &[String],
         retry_timeout: &Duration,
     ) -> anyhow::Result<()> {
         print!(
@@ -196,7 +196,7 @@ impl SafekeeperNode {
             ]);
         }
 
-        args.extend(extra_opts);
+        args.extend_from_slice(extra_opts);
 
         background_process::start_process(
             &format!("safekeeper-{id}"),

control_plane/src/storage_controller.rs

@@ -3,7 +3,7 @@ use crate::{
     local_env::{LocalEnv, NeonStorageControllerConf},
 };
 use camino::{Utf8Path, Utf8PathBuf};
-use hyper::Uri;
+use hyper0::Uri;
 use nix::unistd::Pid;
 use pageserver_api::{
     controller_api::{
@@ -347,7 +347,7 @@ impl StorageController {
 
             if !tokio::fs::try_exists(&pg_data_path).await? {
                 let initdb_args = [
-                    "-D",
+                    "--pgdata",
                     pg_data_path.as_ref(),
                     "--username",
                     &username(),
@@ -517,6 +517,13 @@ impl StorageController {
             args.push(format!("--max-secondary-lag-bytes={lag}"))
         }
 
+        if let Some(threshold) = self.config.long_reconcile_threshold {
+            args.push(format!(
+                "--long-reconcile-threshold={}",
+                humantime::Duration::from(threshold)
+            ))
+        }
+
         args.push(format!(
             "--neon-local-repo-dir={}",
             self.env.base_data_dir.display()

docs/docker.md

@@ -5,7 +5,7 @@
 Currently we build two main images:
 
 - [neondatabase/neon](https://hub.docker.com/repository/docker/neondatabase/neon) — image with pre-built `pageserver`, `safekeeper` and `proxy` binaries and all the required runtime dependencies. Built from [/Dockerfile](/Dockerfile).
-- [neondatabase/compute-node-v16](https://hub.docker.com/repository/docker/neondatabase/compute-node-v16) — compute node image with pre-built Postgres binaries from [neondatabase/postgres](https://github.com/neondatabase/postgres). Similar images exist for v15 and v14.
+- [neondatabase/compute-node-v16](https://hub.docker.com/repository/docker/neondatabase/compute-node-v16) — compute node image with pre-built Postgres binaries from [neondatabase/postgres](https://github.com/neondatabase/postgres). Similar images exist for v15 and v14. Built from [/compute-node/Dockerfile](/compute/Dockerfile.compute-node).
 
 And additional intermediate image:
 

docs/rfcs/038-aux-file-v2.md

@@ -0,0 +1,112 @@
+# AUX file v2
+
+## Summary
+
+This is a retrospective RFC describing a new storage strategy for AUX files.
+
+## Motivation
+
+The original aux file storage strategy stores everything in a single `AUX_FILES_KEY`.
+Every time the compute node streams a `neon-file` record to the pageserver, it will
+update the aux file hash map, and then write the serialized hash map into the key.
+This creates serious space bloat. There was a fix to log delta records (i.e., update
+a key in the hash map) to the aux file key. In this way, the pageserver only stores
+the deltas at each of the LSNs. However, this improved v1 storage strategy still
+requires us to store everything in an aux file cache in memory, because we cannot
+fetch a single key (or file) from the compound `AUX_FILES_KEY`.
+
+### Prior art
+
+For storing large amount of small files, we can use a key-value store where the key
+is the filename and the value is the file content.
+
+## Requirements
+
+- No space bloat, fixed space amplification.
+- No write bloat, fixed write amplification.
+
+## Impacted Components
+
+pageserver
+
+## Sparse Keyspace
+
+In pageserver, we had assumed the keyspaces are always contiguous. For example, if the keyspace 0x0000-0xFFFF
+exists in the pageserver, every single key in the key range would exist in the storage. Based on the prior
+assumption, there are code that traverses the keyspace by iterating every single key.
+
+```rust
+loop {
+    // do something
+    key = key.next();
+}
+```
+
+If a keyspace is very large, for example, containing `2^64` keys, this loop will take infinite time to run.
+Therefore, we introduce the concept of sparse keyspace in this RFC. For a sparse keyspace, not every key would
+exist in the key range. Developers should not attempt to iterate every single key in the keyspace. Instead,
+they should fetch all the layer files in the key range, and then do a merge of them.
+
+In aux file v2, we store aux files within the sparse keyspace of the prefix `AUX_KEY_PREFIX`.
+
+## AUX v2 Keyspace and Key Mapping
+
+Pageserver uses fixed-size keys. The key is 128b. In order to store files of arbitrary filenames into the
+keyspace, we assign a predetermined prefix based on the directory storing the aux file, and use the FNV hash
+of the filename for the rest bits of the key. The encoding scheme is defined in `encode_aux_file_key`.
+
+For example, `pg_logical/mappings/test1` will be encoded as:
+
+```
+62 0000 01 01 7F8B83D94F7081693471ABF91C
+^ aux prefix
+        ^ assigned prefix of pg_logical/
+           ^ assigned prefix of mappings/
+              ^ 13B FNV hash of test1
+   ^ not used due to key representation
+```
+
+The prefixes of the directories should be assigned every time we add a new type of aux file into the storage within `aux_file.rs`. For all directories without an assigned prefix, it will be put into the `0xFFFF` keyspace.
+
+Note that inside pageserver, there are two representations of the keys: the 18B full key representation
+and the 16B compact key representation. For the 18B representation, some fields have restricted ranges
+of values. Therefore, the aux keys only use the 16B compact portion of the full key.
+
+It is possible that two files get mapped to the same key due to hash collision. Therefore, the value of
+each of the aux key is an array that contains all filenames and file content that should be stored in
+this key.
+
+We use `Value::Image` to store the aux keys. Therefore, page reconstruction works in the same way as before,
+and we do not need addition code to support reconstructing the value. We simply get the latest image from
+the storage.
+
+## Inbound Logical Replication Key Mapping
+
+For inbound logical replication, Postgres needs the `replorigin_checkpoint` file to store the data.
+This file not directly stored in the pageserver using the aux v2 mechanism. It is constructed during
+generating the basebackup by scanning the `REPL_ORIGIN_KEY_PREFIX` keyspace.
+
+## Sparse Keyspace Read Path
+
+There are two places we need to read the aux files from the pageserver:
+
+* On the write path, when the compute node adds an aux file to the pageserver, we will retrieve the key from the storage, append the file to the hashed key, and write it back. The current `get` API already supports that.
+*  We use the vectored get API to retrieve all aux files during generating the basebackup. Because we need to scan a sparse keyspace, we slightly modified the vectored get path. The vectorized API will attempt to retrieve every single key within the requested key range, and therefore, we modified it in a way that keys within `NON_INHERITED_SPARSE_RANGE` will not trigger missing key error.
+
+## Compaction and Image Layer Generation
+
+With the add of sparse keyspaces, we also modified the compaction code to accommodate the fact that sparse keyspaces do not have every single key stored in the storage.
+
+* L0 compaction: we modified the hole computation code so that it can handle sparse keyspaces when computing holes.
+* Image layer creation: instead of calling `key.next()` and getting/reconstructing images for every single key, we use the vectored get API to scan all keys in the keyspace at a given LSN. Image layers are only created if there are too many delta layers between the latest LSN and the last image layer we generated for sparse keyspaces. The created image layer always cover the full aux key range for now, and could be optimized later.
+
+## Migration
+
+We decided not to make the new aux storage strategy (v1) compatible with the original one (v1). One feasible way of doing a seamless migration is to store new data in aux v2 while old data in aux v1, but this complicates file deletions. We want all users to start with a clean state with no aux files in the storage, and therefore, we need to do manual migrations for users using aux v1 by using the [migration script](https://github.com/neondatabase/aux_v2_migration).
+
+During the period of migration, we store the aux policy in the `index_part.json` file. When a tenant is attached
+with no policy set, the pageserver will scan the aux file keyspaces to identify the current aux policy being used (v1 or v2).
+
+If a timeline has aux v1 files stored, it will use aux file policy v1 unless we do a manual migration for them. Otherwise, the default aux file policy for new timelines is aux v2. Users enrolled in logical replication before we set aux v2 as default use aux v1 policy. Users who tried setting up inbound replication (which was not supported at that time) may also create some file entries in aux v1 store, even if they did not enroll in the logical replication testing program.
+
+The code for aux v2 migration is in https://github.com/neondatabase/aux_v2_migration. The toolkit scans all projects with logical replication enabled. For all these projects, it put the computes into maintenance mode (suspend all of then), call the migration API to switch the aux file policy on the pageserver (which drops all replication states), and restart all the computes.

libs/compute_api/src/responses.rs

@@ -1,5 +1,7 @@
 //! Structs representing the JSON formats used in the compute_ctl's HTTP API.
 
+use std::fmt::Display;
+
 use chrono::{DateTime, Utc};
 use serde::{Deserialize, Serialize, Serializer};
 
@@ -58,6 +60,21 @@ pub enum ComputeStatus {
     Terminated,
 }
 
+impl Display for ComputeStatus {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            ComputeStatus::Empty => f.write_str("empty"),
+            ComputeStatus::ConfigurationPending => f.write_str("configuration-pending"),
+            ComputeStatus::Init => f.write_str("init"),
+            ComputeStatus::Running => f.write_str("running"),
+            ComputeStatus::Configuration => f.write_str("configuration"),
+            ComputeStatus::Failed => f.write_str("failed"),
+            ComputeStatus::TerminationPending => f.write_str("termination-pending"),
+            ComputeStatus::Terminated => f.write_str("terminated"),
+        }
+    }
+}
+
 fn rfc3339_serialize<S>(x: &Option<DateTime<Utc>>, s: S) -> Result<S::Ok, S::Error>
 where
     S: Serializer,

libs/compute_api/src/spec.rs

@@ -50,6 +50,16 @@ pub struct ComputeSpec {
     #[serde(default)]
     pub swap_size_bytes: Option<u64>,
 
+    /// If compute_ctl was passed `--set-disk-quota-for-fs`, a value of `Some(_)` instructs
+    /// compute_ctl to run `/neonvm/bin/set-disk-quota` with the given size and fs, when the
+    /// spec is first received.
+    ///
+    /// Both this field and `--set-disk-quota-for-fs` are required, so that the control plane's
+    /// spec generation doesn't need to be aware of the actual compute it's running on, while
+    /// guaranteeing gradual rollout of disk quota.
+    #[serde(default)]
+    pub disk_quota_bytes: Option<u64>,
+
     /// Expected cluster state at the end of transition process.
     pub cluster: Cluster,
     pub delta_operations: Option<Vec<DeltaOp>>,
@@ -96,6 +106,10 @@ pub struct ComputeSpec {
     // Stripe size for pageserver sharding, in pages
     #[serde(default)]
     pub shard_stripe_size: Option<usize>,
+
+    /// Local Proxy configuration used for JWT authentication
+    #[serde(default)]
+    pub local_proxy_config: Option<LocalProxySpec>,
 }
 
 /// Feature flag to signal `compute_ctl` to enable certain experimental functionality.
@@ -268,11 +282,13 @@ pub struct GenericOption {
 /// declare a `trait` on it.
 pub type GenericOptions = Option<Vec<GenericOption>>;
 
-/// Configured the local-proxy application with the relevant JWKS and roles it should
+/// Configured the local_proxy application with the relevant JWKS and roles it should
 /// use for authorizing connect requests using JWT.
 #[derive(Clone, Debug, Deserialize, Serialize)]
 pub struct LocalProxySpec {
-    pub jwks: Vec<JwksSettings>,
+    #[serde(default)]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub jwks: Option<Vec<JwksSettings>>,
 }
 
 #[derive(Clone, Debug, Deserialize, Serialize)]

libs/pageserver_api/src/config.rs

@@ -104,8 +104,7 @@ pub struct ConfigToml {
     pub image_compression: ImageCompressionAlgorithm,
     pub ephemeral_bytes_per_memory_kb: usize,
     pub l0_flush: Option<crate::models::L0FlushConfig>,
-    pub virtual_file_direct_io: crate::models::virtual_file::DirectIoMode,
-    pub io_buffer_alignment: usize,
+    pub virtual_file_io_mode: Option<crate::models::virtual_file::IoMode>,
 }
 
 #[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
@@ -296,7 +295,14 @@ pub mod defaults {
 
     pub const DEFAULT_INGEST_BATCH_SIZE: u64 = 100;
 
-    pub const DEFAULT_MAX_VECTORED_READ_BYTES: usize = 128 * 1024; // 128 KiB
+    /// Soft limit for the maximum size of a vectored read.
+    ///
+    /// This is determined by the largest NeonWalRecord that can exist (minus dbdir and reldir keys
+    /// which are bounded by the blob io limits only). As of this writing, that is a `NeonWalRecord::ClogSetCommitted` record,
+    /// with 32k xids. That's the max number of XIDS on a single CLOG page. The size of such a record
+    /// is `sizeof(Transactionid) * 32768 + (some fixed overhead from 'timestamp`, the Vec length and whatever extra serde serialization adds)`.
+    /// That is, slightly above 128 kB.
+    pub const DEFAULT_MAX_VECTORED_READ_BYTES: usize = 130 * 1024; // 130 KiB
 
     pub const DEFAULT_IMAGE_COMPRESSION: ImageCompressionAlgorithm =
         ImageCompressionAlgorithm::Zstd { level: Some(1) };
@@ -381,10 +387,7 @@ impl Default for ConfigToml {
             image_compression: (DEFAULT_IMAGE_COMPRESSION),
             ephemeral_bytes_per_memory_kb: (DEFAULT_EPHEMERAL_BYTES_PER_MEMORY_KB),
             l0_flush: None,
-            virtual_file_direct_io: crate::models::virtual_file::DirectIoMode::default(),
-
-            io_buffer_alignment: DEFAULT_IO_BUFFER_ALIGNMENT,
-
+            virtual_file_io_mode: None,
             tenant_config: TenantConfigToml::default(),
         }
     }

libs/pageserver_api/src/key.rs

@@ -748,6 +748,16 @@ impl Key {
         self.field1 == 0x00 && self.field4 != 0 && self.field6 != 0xffffffff
     }
 
+    #[inline(always)]
+    pub fn is_rel_dir_key(&self) -> bool {
+        self.field1 == 0x00
+            && self.field2 != 0
+            && self.field3 != 0
+            && self.field4 == 0
+            && self.field5 == 0
+            && self.field6 == 1
+    }
+
     /// Guaranteed to return `Ok()` if [`Self::is_rel_block_key`] returns `true` for `key`.
     #[inline(always)]
     pub fn to_rel_block(self) -> anyhow::Result<(RelTag, BlockNumber)> {

libs/pageserver_api/src/models.rs

@@ -972,8 +972,6 @@ pub struct TopTenantShardsResponse {
 }
 
 pub mod virtual_file {
-    use std::path::PathBuf;
-
     #[derive(
         Copy,
         Clone,
@@ -994,50 +992,45 @@ pub mod virtual_file {
     }
 
     /// Direct IO modes for a pageserver.
-    #[derive(Debug, PartialEq, Eq, Clone, serde::Deserialize, serde::Serialize, Default)]
-    #[serde(tag = "mode", rename_all = "kebab-case", deny_unknown_fields)]
-    pub enum DirectIoMode {
-        /// Direct IO disabled (uses usual buffered IO).
-        #[default]
-        Disabled,
-        /// Direct IO disabled (performs checks and perf simulations).
-        Evaluate {
-            /// Alignment check level
-            alignment_check: DirectIoAlignmentCheckLevel,
-            /// Latency padded for performance simulation.
-            latency_padding: DirectIoLatencyPadding,
-        },
-        /// Direct IO enabled.
-        Enabled {
-            /// Actions to perform on alignment error.
-            on_alignment_error: DirectIoOnAlignmentErrorAction,
-        },
+    #[derive(
+        Copy,
+        Clone,
+        PartialEq,
+        Eq,
+        Hash,
+        strum_macros::EnumString,
+        strum_macros::Display,
+        serde_with::DeserializeFromStr,
+        serde_with::SerializeDisplay,
+        Debug,
+    )]
+    #[strum(serialize_all = "kebab-case")]
+    #[repr(u8)]
+    pub enum IoMode {
+        /// Uses buffered IO.
+        Buffered,
+        /// Uses direct IO, error out if the operation fails.
+        #[cfg(target_os = "linux")]
+        Direct,
     }
 
-    #[derive(Debug, PartialEq, Eq, Clone, serde::Deserialize, serde::Serialize, Default)]
-    #[serde(rename_all = "kebab-case")]
-    pub enum DirectIoAlignmentCheckLevel {
-        #[default]
-        Error,
-        Log,
-        None,
+    impl IoMode {
+        pub const fn preferred() -> Self {
+            Self::Buffered
+        }
     }
 
-    #[derive(Debug, PartialEq, Eq, Clone, serde::Deserialize, serde::Serialize, Default)]
-    #[serde(rename_all = "kebab-case")]
-    pub enum DirectIoOnAlignmentErrorAction {
-        Error,
-        #[default]
-        FallbackToBuffered,
-    }
+    impl TryFrom<u8> for IoMode {
+        type Error = u8;
 
-    #[derive(Debug, PartialEq, Eq, Clone, serde::Deserialize, serde::Serialize, Default)]
-    #[serde(tag = "type", rename_all = "kebab-case")]
-    pub enum DirectIoLatencyPadding {
-        /// Pad virtual file operations with IO to a fake file.
-        FakeFileRW { path: PathBuf },
-        #[default]
-        None,
+        fn try_from(value: u8) -> Result<Self, Self::Error> {
+            Ok(match value {
+                v if v == (IoMode::Buffered as u8) => IoMode::Buffered,
+                #[cfg(target_os = "linux")]
+                v if v == (IoMode::Direct as u8) => IoMode::Direct,
+                x => return Err(x),
+            })
+        }
     }
 }
 

libs/postgres_ffi/wal_craft/src/lib.rs

@@ -93,9 +93,9 @@ impl Conf {
         );
         let output = self
             .new_pg_command("initdb")?
-            .arg("-D")
+            .arg("--pgdata")
             .arg(&self.datadir)
-            .args(["-U", "postgres", "--no-instructions", "--no-sync"])
+            .args(["--username", "postgres", "--no-instructions", "--no-sync"])
             .output()?;
         debug!("initdb output: {:?}", output);
         ensure!(

libs/remote_storage/Cargo.toml

@@ -16,7 +16,7 @@ aws-sdk-s3.workspace = true
 bytes.workspace = true
 camino = { workspace = true, features = ["serde1"] }
 humantime-serde.workspace = true
-hyper = { workspace = true, features = ["stream"] }
+hyper0 = { workspace = true, features = ["stream"] }
 futures.workspace = true
 serde.workspace = true
 serde_json.workspace = true

libs/remote_storage/src/azure_blob.rs

@@ -14,7 +14,7 @@ use std::time::SystemTime;
 
 use super::REMOTE_STORAGE_PREFIX_SEPARATOR;
 use anyhow::Result;
-use azure_core::request_options::{MaxResults, Metadata, Range};
+use azure_core::request_options::{IfMatchCondition, MaxResults, Metadata, Range};
 use azure_core::{Continuable, RetryOptions};
 use azure_identity::DefaultAzureCredential;
 use azure_storage::StorageCredentials;
@@ -33,10 +33,10 @@ use tracing::debug;
 use utils::backoff;
 
 use crate::metrics::{start_measuring_requests, AttemptOutcome, RequestKind};
-use crate::ListingObject;
 use crate::{
-    config::AzureConfig, error::Cancelled, ConcurrencyLimiter, Download, DownloadError, Listing,
-    ListingMode, RemotePath, RemoteStorage, StorageMetadata, TimeTravelError, TimeoutOrCancel,
+    config::AzureConfig, error::Cancelled, ConcurrencyLimiter, Download, DownloadError,
+    DownloadOpts, Listing, ListingMode, ListingObject, RemotePath, RemoteStorage, StorageMetadata,
+    TimeTravelError, TimeoutOrCancel,
 };
 
 pub struct AzureBlobStorage {
@@ -259,6 +259,7 @@ 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::NotModified => DownloadError::Unmodified,
             StatusCode::BadRequest => DownloadError::BadInput(anyhow::Error::new(error)),
             _ => DownloadError::Other(anyhow::Error::new(error)),
         }
@@ -484,11 +485,16 @@ impl RemoteStorage for AzureBlobStorage {
     async fn download(
         &self,
         from: &RemotePath,
+        opts: &DownloadOpts,
         cancel: &CancellationToken,
     ) -> Result<Download, DownloadError> {
         let blob_client = self.client.blob_client(self.relative_path_to_name(from));
 
-        let builder = blob_client.get();
+        let mut builder = blob_client.get();
+
+        if let Some(ref etag) = opts.etag {
+            builder = builder.if_match(IfMatchCondition::NotMatch(etag.to_string()))
+        }
 
         self.download_for_builder(builder, cancel).await
     }

libs/remote_storage/src/error.rs

@@ -5,6 +5,8 @@ pub enum DownloadError {
     BadInput(anyhow::Error),
     /// The file was not found in the remote storage.
     NotFound,
+    /// The caller provided an ETag, and the file was not modified.
+    Unmodified,
     /// A cancellation token aborted the download, typically during
     /// tenant detach or process shutdown.
     Cancelled,
@@ -24,6 +26,7 @@ impl std::fmt::Display for DownloadError {
                 write!(f, "Failed to download a remote file due to user input: {e}")
             }
             DownloadError::NotFound => write!(f, "No file found for the remote object id given"),
+            DownloadError::Unmodified => write!(f, "File was not modified"),
             DownloadError::Cancelled => write!(f, "Cancelled, shutting down"),
             DownloadError::Timeout => write!(f, "timeout"),
             DownloadError::Other(e) => write!(f, "Failed to download a remote file: {e:?}"),
@@ -38,7 +41,7 @@ impl DownloadError {
     pub fn is_permanent(&self) -> bool {
         use DownloadError::*;
         match self {
-            BadInput(_) | NotFound | Cancelled => true,
+            BadInput(_) | NotFound | Unmodified | Cancelled => true,
             Timeout | Other(_) => false,
         }
     }

libs/remote_storage/src/lib.rs

@@ -161,6 +161,14 @@ pub struct Listing {
     pub keys: Vec<ListingObject>,
 }
 
+/// Options for downloads. The default value is a plain GET.
+#[derive(Default)]
+pub struct DownloadOpts {
+    /// If given, returns [`DownloadError::Unmodified`] if the object still has
+    /// the same ETag (using If-None-Match).
+    pub etag: Option<Etag>,
+}
+
 /// Storage (potentially remote) API to manage its state.
 /// This storage tries to be unaware of any layered repository context,
 /// providing basic CRUD operations for storage files.
@@ -245,6 +253,7 @@ pub trait RemoteStorage: Send + Sync + 'static {
     async fn download(
         &self,
         from: &RemotePath,
+        opts: &DownloadOpts,
         cancel: &CancellationToken,
     ) -> Result<Download, DownloadError>;
 
@@ -401,16 +410,18 @@ impl<Other: RemoteStorage> GenericRemoteStorage<Arc<Other>> {
         }
     }
 
+    /// See [`RemoteStorage::download`]
     pub async fn download(
         &self,
         from: &RemotePath,
+        opts: &DownloadOpts,
         cancel: &CancellationToken,
     ) -> Result<Download, DownloadError> {
         match self {
-            Self::LocalFs(s) => s.download(from, cancel).await,
-            Self::AwsS3(s) => s.download(from, cancel).await,
-            Self::AzureBlob(s) => s.download(from, cancel).await,
-            Self::Unreliable(s) => s.download(from, cancel).await,
+            Self::LocalFs(s) => s.download(from, opts, cancel).await,
+            Self::AwsS3(s) => s.download(from, opts, cancel).await,
+            Self::AzureBlob(s) => s.download(from, opts, cancel).await,
+            Self::Unreliable(s) => s.download(from, opts, cancel).await,
         }
     }
 
@@ -572,7 +583,7 @@ impl GenericRemoteStorage {
     ) -> Result<Download, DownloadError> {
         match byte_range {
             Some((start, end)) => self.download_byte_range(from, start, end, cancel).await,
-            None => self.download(from, cancel).await,
+            None => self.download(from, &DownloadOpts::default(), cancel).await,
         }
     }
 

libs/remote_storage/src/local_fs.rs

@@ -23,8 +23,8 @@ use tokio_util::{io::ReaderStream, sync::CancellationToken};
 use utils::crashsafe::path_with_suffix_extension;
 
 use crate::{
-    Download, DownloadError, Listing, ListingMode, ListingObject, RemotePath, TimeTravelError,
-    TimeoutOrCancel, REMOTE_STORAGE_PREFIX_SEPARATOR,
+    Download, DownloadError, DownloadOpts, Listing, ListingMode, ListingObject, RemotePath,
+    TimeTravelError, TimeoutOrCancel, REMOTE_STORAGE_PREFIX_SEPARATOR,
 };
 
 use super::{RemoteStorage, StorageMetadata};
@@ -494,11 +494,17 @@ impl RemoteStorage for LocalFs {
     async fn download(
         &self,
         from: &RemotePath,
+        opts: &DownloadOpts,
         cancel: &CancellationToken,
     ) -> Result<Download, DownloadError> {
         let target_path = from.with_base(&self.storage_root);
 
         let file_metadata = file_metadata(&target_path).await?;
+        let etag = mock_etag(&file_metadata);
+
+        if opts.etag.as_ref() == Some(&etag) {
+            return Err(DownloadError::Unmodified);
+        }
 
         let source = ReaderStream::new(
             fs::OpenOptions::new()
@@ -519,7 +525,6 @@ impl RemoteStorage for LocalFs {
         let cancel_or_timeout = crate::support::cancel_or_timeout(self.timeout, cancel.clone());
         let source = crate::support::DownloadStream::new(cancel_or_timeout, source);
 
-        let etag = mock_etag(&file_metadata);
         Ok(Download {
             metadata,
             last_modified: file_metadata
@@ -692,7 +697,7 @@ mod fs_tests {
     ) -> anyhow::Result<String> {
         let cancel = CancellationToken::new();
         let download = storage
-            .download(remote_storage_path, &cancel)
+            .download(remote_storage_path, &DownloadOpts::default(), &cancel)
             .await
             .map_err(|e| anyhow::anyhow!("Download failed: {e}"))?;
         ensure!(
@@ -773,8 +778,8 @@ mod fs_tests {
             "We should upload and download the same contents"
         );
 
-        let non_existing_path = "somewhere/else";
-        match storage.download(&RemotePath::new(Utf8Path::new(non_existing_path))?, &cancel).await {
+        let non_existing_path = RemotePath::new(Utf8Path::new("somewhere/else"))?;
+        match storage.download(&non_existing_path, &DownloadOpts::default(), &cancel).await {
             Err(DownloadError::NotFound) => {} // Should get NotFound for non existing keys
             other => panic!("Should get a NotFound error when downloading non-existing storage files, but got: {other:?}"),
         }
@@ -1101,7 +1106,13 @@ mod fs_tests {
             storage.upload(body, len, &path, None, &cancel).await?;
         }
 
-        let read = aggregate(storage.download(&path, &cancel).await?.download_stream).await?;
+        let read = aggregate(
+            storage
+                .download(&path, &DownloadOpts::default(), &cancel)
+                .await?
+                .download_stream,
+        )
+        .await?;
         assert_eq!(body, read);
 
         let shorter = Bytes::from_static(b"shorter body");
@@ -1112,7 +1123,13 @@ mod fs_tests {
             storage.upload(body, len, &path, None, &cancel).await?;
         }
 
-        let read = aggregate(storage.download(&path, &cancel).await?.download_stream).await?;
+        let read = aggregate(
+            storage
+                .download(&path, &DownloadOpts::default(), &cancel)
+                .await?
+                .download_stream,
+        )
+        .await?;
         assert_eq!(shorter, read);
         Ok(())
     }
@@ -1145,7 +1162,13 @@ mod fs_tests {
             storage.upload(body, len, &path, None, &cancel).await?;
         }
 
-        let read = aggregate(storage.download(&path, &cancel).await?.download_stream).await?;
+        let read = aggregate(
+            storage
+                .download(&path, &DownloadOpts::default(), &cancel)
+                .await?
+                .download_stream,
+        )
+        .await?;
         assert_eq!(body, read);
 
         Ok(())

libs/remote_storage/src/s3_bucket.rs

@@ -28,12 +28,13 @@ use aws_sdk_s3::{
     Client,
 };
 use aws_smithy_async::rt::sleep::TokioSleep;
+use http_types::StatusCode;
 
 use aws_smithy_types::{body::SdkBody, DateTime};
 use aws_smithy_types::{byte_stream::ByteStream, date_time::ConversionError};
 use bytes::Bytes;
 use futures::stream::Stream;
-use hyper::Body;
+use hyper0::Body;
 use scopeguard::ScopeGuard;
 use tokio_util::sync::CancellationToken;
 use utils::backoff;
@@ -44,8 +45,8 @@ use crate::{
     error::Cancelled,
     metrics::{start_counting_cancelled_wait, start_measuring_requests},
     support::PermitCarrying,
-    ConcurrencyLimiter, Download, DownloadError, Listing, ListingMode, ListingObject, RemotePath,
-    RemoteStorage, TimeTravelError, TimeoutOrCancel, MAX_KEYS_PER_DELETE,
+    ConcurrencyLimiter, Download, DownloadError, DownloadOpts, Listing, ListingMode, ListingObject,
+    RemotePath, RemoteStorage, TimeTravelError, TimeoutOrCancel, MAX_KEYS_PER_DELETE,
     REMOTE_STORAGE_PREFIX_SEPARATOR,
 };
 
@@ -67,6 +68,7 @@ pub struct S3Bucket {
 struct GetObjectRequest {
     bucket: String,
     key: String,
+    etag: Option<String>,
     range: Option<String>,
 }
 impl S3Bucket {
@@ -248,13 +250,18 @@ impl S3Bucket {
 
         let started_at = start_measuring_requests(kind);
 
-        let get_object = self
+        let mut builder = self
             .client
             .get_object()
             .bucket(request.bucket)
             .key(request.key)
-            .set_range(request.range)
-            .send();
+            .set_range(request.range);
+
+        if let Some(etag) = request.etag {
+            builder = builder.if_none_match(etag);
+        }
+
+        let get_object = builder.send();
 
         let get_object = tokio::select! {
             res = get_object => res,
@@ -277,6 +284,20 @@ impl S3Bucket {
                 );
                 return Err(DownloadError::NotFound);
             }
+            Err(SdkError::ServiceError(e))
+                // aws_smithy_runtime_api::http::response::StatusCode isn't
+                // re-exported by any aws crates, so just check the numeric
+                // status against http_types::StatusCode instead of pulling it.
+                if e.raw().status().as_u16() == StatusCode::NotModified =>
+            {
+                // Count an unmodified file as a success.
+                crate::metrics::BUCKET_METRICS.req_seconds.observe_elapsed(
+                    kind,
+                    AttemptOutcome::Ok,
+                    started_at,
+                );
+                return Err(DownloadError::Unmodified);
+            }
             Err(e) => {
                 crate::metrics::BUCKET_METRICS.req_seconds.observe_elapsed(
                     kind,
@@ -773,6 +794,7 @@ impl RemoteStorage for S3Bucket {
     async fn download(
         &self,
         from: &RemotePath,
+        opts: &DownloadOpts,
         cancel: &CancellationToken,
     ) -> Result<Download, DownloadError> {
         // if prefix is not none then download file `prefix/from`
@@ -781,6 +803,7 @@ impl RemoteStorage for S3Bucket {
             GetObjectRequest {
                 bucket: self.bucket_name.clone(),
                 key: self.relative_path_to_s3_object(from),
+                etag: opts.etag.as_ref().map(|e| e.to_string()),
                 range: None,
             },
             cancel,
@@ -807,6 +830,7 @@ impl RemoteStorage for S3Bucket {
             GetObjectRequest {
                 bucket: self.bucket_name.clone(),
                 key: self.relative_path_to_s3_object(from),
+                etag: None,
                 range,
             },
             cancel,

libs/remote_storage/src/simulate_failures.rs

@@ -12,8 +12,8 @@ use std::{collections::hash_map::Entry, sync::Arc};
 use tokio_util::sync::CancellationToken;
 
 use crate::{
-    Download, DownloadError, GenericRemoteStorage, Listing, ListingMode, RemotePath, RemoteStorage,
-    StorageMetadata, TimeTravelError,
+    Download, DownloadError, DownloadOpts, GenericRemoteStorage, Listing, ListingMode, RemotePath,
+    RemoteStorage, StorageMetadata, TimeTravelError,
 };
 
 pub struct UnreliableWrapper {
@@ -167,11 +167,12 @@ impl RemoteStorage for UnreliableWrapper {
     async fn download(
         &self,
         from: &RemotePath,
+        opts: &DownloadOpts,
         cancel: &CancellationToken,
     ) -> Result<Download, DownloadError> {
         self.attempt(RemoteOp::Download(from.clone()))
             .map_err(DownloadError::Other)?;
-        self.inner.download(from, cancel).await
+        self.inner.download(from, opts, cancel).await
     }
 
     async fn download_byte_range(

libs/remote_storage/tests/common/tests.rs

@@ -1,8 +1,7 @@
 use anyhow::Context;
 use camino::Utf8Path;
 use futures::StreamExt;
-use remote_storage::ListingMode;
-use remote_storage::RemotePath;
+use remote_storage::{DownloadError, DownloadOpts, ListingMode, ListingObject, RemotePath};
 use std::sync::Arc;
 use std::{collections::HashSet, num::NonZeroU32};
 use test_context::test_context;
@@ -284,7 +283,10 @@ async fn upload_download_works(ctx: &mut MaybeEnabledStorage) -> anyhow::Result<
     ctx.client.upload(data, len, &path, None, &cancel).await?;
 
     // Normal download request
-    let dl = ctx.client.download(&path, &cancel).await?;
+    let dl = ctx
+        .client
+        .download(&path, &DownloadOpts::default(), &cancel)
+        .await?;
     let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig);
 
@@ -337,6 +339,54 @@ async fn upload_download_works(ctx: &mut MaybeEnabledStorage) -> anyhow::Result<
     Ok(())
 }
 
+/// Tests that conditional downloads work properly, by returning
+/// DownloadError::Unmodified when the object ETag matches the given ETag.
+#[test_context(MaybeEnabledStorage)]
+#[tokio::test]
+async fn download_conditional(ctx: &mut MaybeEnabledStorage) -> anyhow::Result<()> {
+    let MaybeEnabledStorage::Enabled(ctx) = ctx else {
+        return Ok(());
+    };
+    let cancel = CancellationToken::new();
+
+    // Create a file.
+    let path = RemotePath::new(Utf8Path::new(format!("{}/file", ctx.base_prefix).as_str()))?;
+    let data = bytes::Bytes::from_static("foo".as_bytes());
+    let (stream, len) = wrap_stream(data);
+    ctx.client.upload(stream, len, &path, None, &cancel).await?;
+
+    // Download it to obtain its etag.
+    let mut opts = DownloadOpts::default();
+    let download = ctx.client.download(&path, &opts, &cancel).await?;
+
+    // Download with the etag yields DownloadError::Unmodified.
+    opts.etag = Some(download.etag);
+    let result = ctx.client.download(&path, &opts, &cancel).await;
+    assert!(
+        matches!(result, Err(DownloadError::Unmodified)),
+        "expected DownloadError::Unmodified, got {result:?}"
+    );
+
+    // Replace the file contents.
+    let data = bytes::Bytes::from_static("bar".as_bytes());
+    let (stream, len) = wrap_stream(data);
+    ctx.client.upload(stream, len, &path, None, &cancel).await?;
+
+    // A download with the old etag should yield the new file.
+    let download = ctx.client.download(&path, &opts, &cancel).await?;
+    assert_ne!(download.etag, opts.etag.unwrap(), "ETag did not change");
+
+    // A download with the new etag should yield Unmodified again.
+    opts.etag = Some(download.etag);
+    let result = ctx.client.download(&path, &opts, &cancel).await;
+    assert!(
+        matches!(result, Err(DownloadError::Unmodified)),
+        "expected DownloadError::Unmodified, got {result:?}"
+    );
+
+    Ok(())
+}
+
 #[test_context(MaybeEnabledStorage)]
 #[tokio::test]
 async fn copy_works(ctx: &mut MaybeEnabledStorage) -> anyhow::Result<()> {
@@ -364,7 +414,10 @@ async fn copy_works(ctx: &mut MaybeEnabledStorage) -> anyhow::Result<()> {
     // Normal download request
     ctx.client.copy_object(&path, &path_dest, &cancel).await?;
 
-    let dl = ctx.client.download(&path_dest, &cancel).await?;
+    let dl = ctx
+        .client
+        .download(&path_dest, &DownloadOpts::default(), &cancel)
+        .await?;
     let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig);
 
@@ -376,3 +429,56 @@ async fn copy_works(ctx: &mut MaybeEnabledStorage) -> anyhow::Result<()> {
 
     Ok(())
 }
+
+/// Tests that head_object works properly.
+#[test_context(MaybeEnabledStorage)]
+#[tokio::test]
+async fn head_object(ctx: &mut MaybeEnabledStorage) -> anyhow::Result<()> {
+    let MaybeEnabledStorage::Enabled(ctx) = ctx else {
+        return Ok(());
+    };
+    let cancel = CancellationToken::new();
+
+    let path = RemotePath::new(Utf8Path::new(format!("{}/file", ctx.base_prefix).as_str()))?;
+
+    // Errors on missing file.
+    let result = ctx.client.head_object(&path, &cancel).await;
+    assert!(
+        matches!(result, Err(DownloadError::NotFound)),
+        "expected NotFound, got {result:?}"
+    );
+
+    // Create the file.
+    let data = bytes::Bytes::from_static("foo".as_bytes());
+    let (stream, len) = wrap_stream(data);
+    ctx.client.upload(stream, len, &path, None, &cancel).await?;
+
+    // Fetch the head metadata.
+    let object = ctx.client.head_object(&path, &cancel).await?;
+    assert_eq!(
+        object,
+        ListingObject {
+            key: path.clone(),
+            last_modified: object.last_modified, // ignore
+            size: 3
+        }
+    );
+
+    // Wait for a couple of seconds, and then update the file to check the last
+    // modified timestamp.
+    tokio::time::sleep(std::time::Duration::from_secs(2)).await;
+
+    let data = bytes::Bytes::from_static("bar".as_bytes());
+    let (stream, len) = wrap_stream(data);
+    ctx.client.upload(stream, len, &path, None, &cancel).await?;
+    let new = ctx.client.head_object(&path, &cancel).await?;
+
+    assert!(
+        !new.last_modified
+            .duration_since(object.last_modified)?
+            .is_zero(),
+        "last_modified did not advance"
+    );
+
+    Ok(())
+}

libs/remote_storage/tests/test_real_s3.rs

@@ -12,8 +12,8 @@ use anyhow::Context;
 use camino::Utf8Path;
 use futures_util::StreamExt;
 use remote_storage::{
-    DownloadError, GenericRemoteStorage, ListingMode, RemotePath, RemoteStorageConfig,
-    RemoteStorageKind, S3Config,
+    DownloadError, DownloadOpts, GenericRemoteStorage, ListingMode, RemotePath,
+    RemoteStorageConfig, RemoteStorageKind, S3Config,
 };
 use test_context::test_context;
 use test_context::AsyncTestContext;
@@ -121,7 +121,8 @@ async fn s3_time_travel_recovery_works(ctx: &mut MaybeEnabledStorage) -> anyhow:
 
     // A little check to ensure that our clock is not too far off from the S3 clock
     {
-        let dl = retry(|| ctx.client.download(&path2, &cancel)).await?;
+        let opts = DownloadOpts::default();
+        let dl = retry(|| ctx.client.download(&path2, &opts, &cancel)).await?;
         let last_modified = dl.last_modified;
         let half_wt = WAIT_TIME.mul_f32(0.5);
         let t0_hwt = t0 + half_wt;
@@ -159,7 +160,12 @@ async fn s3_time_travel_recovery_works(ctx: &mut MaybeEnabledStorage) -> anyhow:
     let t2_files_recovered = list_files(&ctx.client, &cancel).await?;
     println!("after recovery to t2: {t2_files_recovered:?}");
     assert_eq!(t2_files, t2_files_recovered);
-    let path2_recovered_t2 = download_to_vec(ctx.client.download(&path2, &cancel).await?).await?;
+    let path2_recovered_t2 = download_to_vec(
+        ctx.client
+            .download(&path2, &DownloadOpts::default(), &cancel)
+            .await?,
+    )
+    .await?;
     assert_eq!(path2_recovered_t2, new_data.as_bytes());
 
     // after recovery to t1: path1 is back, path2 has the old content
@@ -170,7 +176,12 @@ async fn s3_time_travel_recovery_works(ctx: &mut MaybeEnabledStorage) -> anyhow:
     let t1_files_recovered = list_files(&ctx.client, &cancel).await?;
     println!("after recovery to t1: {t1_files_recovered:?}");
     assert_eq!(t1_files, t1_files_recovered);
-    let path2_recovered_t1 = download_to_vec(ctx.client.download(&path2, &cancel).await?).await?;
+    let path2_recovered_t1 = download_to_vec(
+        ctx.client
+            .download(&path2, &DownloadOpts::default(), &cancel)
+            .await?,
+    )
+    .await?;
     assert_eq!(path2_recovered_t1, old_data.as_bytes());
 
     // after recovery to t0: everything is gone except for path1
@@ -416,7 +427,7 @@ async fn download_is_timeouted(ctx: &mut MaybeEnabledStorage) {
     let started_at = std::time::Instant::now();
     let mut stream = ctx
         .client
-        .download(&path, &cancel)
+        .download(&path, &DownloadOpts::default(), &cancel)
         .await
         .expect("download succeeds")
         .download_stream;
@@ -491,7 +502,7 @@ async fn download_is_cancelled(ctx: &mut MaybeEnabledStorage) {
     {
         let stream = ctx
             .client
-            .download(&path, &cancel)
+            .download(&path, &DownloadOpts::default(), &cancel)
             .await
             .expect("download succeeds")
             .download_stream;

libs/tracing-utils/Cargo.toml

@@ -5,13 +5,15 @@ edition.workspace = true
 license.workspace = true
 
 [dependencies]
-hyper.workspace = true
-opentelemetry = { workspace = true, features=["rt-tokio"] }
-opentelemetry-otlp = { workspace = true, default-features=false, features = ["http-proto", "trace", "http", "reqwest-client"] }
+hyper0.workspace = true
+opentelemetry = { workspace = true, features = ["trace"] }
+opentelemetry_sdk = { workspace = true, features = ["rt-tokio"] }
+opentelemetry-otlp = { workspace = true, default-features = false, features = ["http-proto", "trace", "http", "reqwest-client"] }
 opentelemetry-semantic-conventions.workspace = true
 tokio = { workspace = true, features = ["rt", "rt-multi-thread"] }
 tracing.workspace = true
 tracing-opentelemetry.workspace = true
+tracing-subscriber.workspace = true
 
 [dev-dependencies]
 tracing-subscriber.workspace = true    # For examples in docs

libs/tracing-utils/src/http.rs

@@ -1,7 +1,7 @@
 //! Tracing wrapper for Hyper HTTP server
 
-use hyper::HeaderMap;
-use hyper::{Body, Request, Response};
+use hyper0::HeaderMap;
+use hyper0::{Body, Request, Response};
 use std::future::Future;
 use tracing::Instrument;
 use tracing_opentelemetry::OpenTelemetrySpanExt;

libs/tracing-utils/src/lib.rs

@@ -10,7 +10,6 @@
 //!
 //! ```rust,no_run
 //! use tracing_subscriber::prelude::*;
-//! use tracing_opentelemetry::OpenTelemetryLayer;
 //!
 //! #[tokio::main]
 //! async fn main() {
@@ -22,7 +21,7 @@
 //!         .with_writer(std::io::stderr);
 //!
 //!     // Initialize OpenTelemetry. Exports tracing spans as OpenTelemetry traces
-//!     let otlp_layer = tracing_utils::init_tracing("my_application").await.map(OpenTelemetryLayer::new);
+//!     let otlp_layer = tracing_utils::init_tracing("my_application").await;
 //!
 //!     // Put it all together
 //!     tracing_subscriber::registry()
@@ -35,15 +34,15 @@
 #![deny(unsafe_code)]
 #![deny(clippy::undocumented_unsafe_blocks)]
 
-use opentelemetry::sdk::Resource;
-use opentelemetry::KeyValue;
-use opentelemetry_otlp::WithExportConfig;
-use opentelemetry_otlp::{OTEL_EXPORTER_OTLP_ENDPOINT, OTEL_EXPORTER_OTLP_TRACES_ENDPOINT};
-
-pub use tracing_opentelemetry::OpenTelemetryLayer;
-
 pub mod http;
 
+use opentelemetry::trace::TracerProvider;
+use opentelemetry::KeyValue;
+use opentelemetry_sdk::Resource;
+use tracing::Subscriber;
+use tracing_subscriber::registry::LookupSpan;
+use tracing_subscriber::Layer;
+
 /// Set up OpenTelemetry exporter, using configuration from environment variables.
 ///
 /// `service_name` is set as the OpenTelemetry 'service.name' resource (see
@@ -71,7 +70,10 @@ pub mod http;
 ///
 /// This doesn't block, but is marked as 'async' to hint that this must be called in
 /// asynchronous execution context.
-pub async fn init_tracing(service_name: &str) -> Option<opentelemetry::sdk::trace::Tracer> {
+pub async fn init_tracing<S>(service_name: &str) -> Option<impl Layer<S>>
+where
+    S: Subscriber + for<'span> LookupSpan<'span>,
+{
     if std::env::var("OTEL_SDK_DISABLED") == Ok("true".to_string()) {
         return None;
     };
@@ -80,9 +82,10 @@ pub async fn init_tracing(service_name: &str) -> Option<opentelemetry::sdk::trac
 
 /// Like `init_tracing`, but creates a separate tokio Runtime for the tracing
 /// tasks.
-pub fn init_tracing_without_runtime(
-    service_name: &str,
-) -> Option<opentelemetry::sdk::trace::Tracer> {
+pub fn init_tracing_without_runtime<S>(service_name: &str) -> Option<impl Layer<S>>
+where
+    S: Subscriber + for<'span> LookupSpan<'span>,
+{
     if std::env::var("OTEL_SDK_DISABLED") == Ok("true".to_string()) {
         return None;
     };
@@ -113,54 +116,36 @@ pub fn init_tracing_without_runtime(
     Some(init_tracing_internal(service_name.to_string()))
 }
 
-fn init_tracing_internal(service_name: String) -> opentelemetry::sdk::trace::Tracer {
-    // Set up exporter from the OTEL_EXPORTER_* environment variables
-    let mut exporter = opentelemetry_otlp::new_exporter().http().with_env();
+fn init_tracing_internal<S>(service_name: String) -> impl Layer<S>
+where
+    S: Subscriber + for<'span> LookupSpan<'span>,
+{
+    // Sets up exporter from the OTEL_EXPORTER_* environment variables.
+    let exporter = opentelemetry_otlp::new_exporter().http();
 
-    // XXX opentelemetry-otlp v0.18.0 has a bug in how it uses the
-    // OTEL_EXPORTER_OTLP_ENDPOINT env variable. According to the
-    // OpenTelemetry spec at
-    // <https://github.com/open-telemetry/opentelemetry-specification/blob/main/specification/protocol/exporter.md#endpoint-urls-for-otlphttp>,
-    // the full exporter URL is formed by appending "/v1/traces" to the value
-    // of OTEL_EXPORTER_OTLP_ENDPOINT. However, opentelemetry-otlp only does
-    // that with the grpc-tonic exporter. Other exporters, like the HTTP
-    // exporter, use the URL from OTEL_EXPORTER_OTLP_ENDPOINT as is, without
-    // appending "/v1/traces".
-    //
-    // See https://github.com/open-telemetry/opentelemetry-rust/pull/950
-    //
-    // Work around that by checking OTEL_EXPORTER_OTLP_ENDPOINT, and setting
-    // the endpoint url with the "/v1/traces" path ourselves. If the bug is
-    // fixed in a later version, we can remove this code. But if we don't
-    // remember to remove this, it won't do any harm either, as the crate will
-    // just ignore the OTEL_EXPORTER_OTLP_ENDPOINT setting when the endpoint
-    // is set directly with `with_endpoint`.
-    if std::env::var(OTEL_EXPORTER_OTLP_TRACES_ENDPOINT).is_err() {
-        if let Ok(mut endpoint) = std::env::var(OTEL_EXPORTER_OTLP_ENDPOINT) {
-            if !endpoint.ends_with('/') {
-                endpoint.push('/');
-            }
-            endpoint.push_str("v1/traces");
-            exporter = exporter.with_endpoint(endpoint);
-        }
-    }
+    // TODO: opentelemetry::global::set_error_handler() with custom handler that
+    //       bypasses default tracing layers, but logs regular looking log
+    //       messages.
 
     // Propagate trace information in the standard W3C TraceContext format.
     opentelemetry::global::set_text_map_propagator(
-        opentelemetry::sdk::propagation::TraceContextPropagator::new(),
+        opentelemetry_sdk::propagation::TraceContextPropagator::new(),
     );
 
-    opentelemetry_otlp::new_pipeline()
+    let tracer = opentelemetry_otlp::new_pipeline()
         .tracing()
         .with_exporter(exporter)
-        .with_trace_config(
-            opentelemetry::sdk::trace::config().with_resource(Resource::new(vec![KeyValue::new(
+        .with_trace_config(opentelemetry_sdk::trace::Config::default().with_resource(
+            Resource::new(vec![KeyValue::new(
                 opentelemetry_semantic_conventions::resource::SERVICE_NAME,
                 service_name,
-            )])),
-        )
-        .install_batch(opentelemetry::runtime::Tokio)
+            )]),
+        ))
+        .install_batch(opentelemetry_sdk::runtime::Tokio)
         .expect("could not initialize opentelemetry exporter")
+        .tracer("global");
+
+    tracing_opentelemetry::layer().with_tracer(tracer)
 }
 
 // Shutdown trace pipeline gracefully, so that it has a chance to send any

libs/utils/Cargo.toml

@@ -22,7 +22,7 @@ chrono.workspace = true
 git-version.workspace = true
 hex = { workspace = true, features = ["serde"] }
 humantime.workspace = true
-hyper = { workspace = true, features = ["full"] }
+hyper0 = { workspace = true, features = ["full"] }
 fail.workspace = true
 futures = { workspace = true}
 jsonwebtoken.workspace = true

libs/utils/src/lib.rs

@@ -2,6 +2,8 @@
 //! between other crates in this repository.
 #![deny(clippy::undocumented_unsafe_blocks)]
 
+extern crate hyper0 as hyper;
+
 pub mod backoff;
 
 /// `Lsn` type implements common tasks on Log Sequence Numbers

libs/vm_monitor/src/lib.rs

@@ -7,11 +7,13 @@ use axum::{
     extract::{ws::WebSocket, State, WebSocketUpgrade},
     response::Response,
 };
-use axum::{routing::get, Router, Server};
+use axum::{routing::get, Router};
 use clap::Parser;
 use futures::Future;
+use std::net::SocketAddr;
 use std::{fmt::Debug, time::Duration};
 use sysinfo::{RefreshKind, System, SystemExt};
+use tokio::net::TcpListener;
 use tokio::{sync::broadcast, task::JoinHandle};
 use tokio_util::sync::CancellationToken;
 use tracing::{error, info};
@@ -132,14 +134,14 @@ pub async fn start(args: &'static Args, token: CancellationToken) -> anyhow::Res
             args,
         });
 
-    let addr = args.addr();
-    let bound = Server::try_bind(&addr.parse().expect("parsing address should not fail"))
+    let addr_str = args.addr();
+    let addr: SocketAddr = addr_str.parse().expect("parsing address should not fail");
+
+    let listener = TcpListener::bind(&addr)
+        .await
         .with_context(|| format!("failed to bind to {addr}"))?;
-
-    info!(addr, "server bound");
-
-    bound
-        .serve(app.into_make_service())
+    info!(addr_str, "server bound");
+    axum::serve(listener, app.into_make_service())
         .await
         .context("server exited")?;
 

pageserver/Cargo.toml

@@ -30,7 +30,7 @@ futures.workspace = true
 hex.workspace = true
 humantime.workspace = true
 humantime-serde.workspace = true
-hyper.workspace = true
+hyper0.workspace = true
 itertools.workspace = true
 md5.workspace = true
 nix.workspace = true

pageserver/benches/bench_ingest.rs

@@ -164,11 +164,7 @@ fn criterion_benchmark(c: &mut Criterion) {
     let conf: &'static PageServerConf = Box::leak(Box::new(
         pageserver::config::PageServerConf::dummy_conf(temp_dir.path().to_path_buf()),
     ));
-    virtual_file::init(
-        16384,
-        virtual_file::io_engine_for_bench(),
-        pageserver_api::config::defaults::DEFAULT_IO_BUFFER_ALIGNMENT,
-    );
+    virtual_file::init(16384, virtual_file::io_engine_for_bench());
     page_cache::init(conf.page_cache_size);
 
     {

pageserver/client/src/mgmt_api.rs

@@ -540,10 +540,13 @@ impl Client {
             .map_err(Error::ReceiveBody)
     }
 
-    /// Configs io buffer alignment at runtime.
-    pub async fn put_io_alignment(&self, align: usize) -> Result<()> {
-        let uri = format!("{}/v1/io_alignment", self.mgmt_api_endpoint);
-        self.request(Method::PUT, uri, align)
+    /// Configs io mode at runtime.
+    pub async fn put_io_mode(
+        &self,
+        mode: &pageserver_api::models::virtual_file::IoMode,
+    ) -> Result<()> {
+        let uri = format!("{}/v1/io_mode", self.mgmt_api_endpoint);
+        self.request(Method::PUT, uri, mode)
             .await?
             .json()
             .await
@@ -736,4 +739,22 @@ impl Client {
             .await
             .map_err(Error::ReceiveBody)
     }
+
+    pub async fn timeline_init_lsn_lease(
+        &self,
+        tenant_shard_id: TenantShardId,
+        timeline_id: TimelineId,
+        lsn: Lsn,
+    ) -> Result<LsnLease> {
+        let uri = format!(
+            "{}/v1/tenant/{tenant_shard_id}/timeline/{timeline_id}/lsn_lease",
+            self.mgmt_api_endpoint,
+        );
+
+        self.request(Method::POST, &uri, LsnLeaseRequest { lsn })
+            .await?
+            .json()
+            .await
+            .map_err(Error::ReceiveBody)
+    }
 }

pageserver/ctl/src/layer_map_analyzer.rs

@@ -152,11 +152,7 @@ pub(crate) async fn main(cmd: &AnalyzeLayerMapCmd) -> Result<()> {
     let ctx = RequestContext::new(TaskKind::DebugTool, DownloadBehavior::Error);
 
     // Initialize virtual_file (file desriptor cache) and page cache which are needed to access layer persistent B-Tree.
-    pageserver::virtual_file::init(
-        10,
-        virtual_file::api::IoEngineKind::StdFs,
-        pageserver_api::config::defaults::DEFAULT_IO_BUFFER_ALIGNMENT,
-    );
+    pageserver::virtual_file::init(10, virtual_file::api::IoEngineKind::StdFs);
     pageserver::page_cache::init(100);
 
     let mut total_delta_layers = 0usize;

pageserver/ctl/src/layers.rs

@@ -59,7 +59,7 @@ pub(crate) enum LayerCmd {
 
 async fn read_delta_file(path: impl AsRef<Path>, ctx: &RequestContext) -> Result<()> {
     let path = Utf8Path::from_path(path.as_ref()).expect("non-Unicode path");
-    virtual_file::init(10, virtual_file::api::IoEngineKind::StdFs, 1);
+    virtual_file::init(10, virtual_file::api::IoEngineKind::StdFs);
     page_cache::init(100);
     let file = VirtualFile::open(path, ctx).await?;
     let file_id = page_cache::next_file_id();
@@ -190,11 +190,7 @@ pub(crate) async fn main(cmd: &LayerCmd) -> Result<()> {
             new_tenant_id,
             new_timeline_id,
         } => {
-            pageserver::virtual_file::init(
-                10,
-                virtual_file::api::IoEngineKind::StdFs,
-                pageserver_api::config::defaults::DEFAULT_IO_BUFFER_ALIGNMENT,
-            );
+            pageserver::virtual_file::init(10, virtual_file::api::IoEngineKind::StdFs);
             pageserver::page_cache::init(100);
 
             let ctx = RequestContext::new(TaskKind::DebugTool, DownloadBehavior::Error);

pageserver/ctl/src/main.rs

@@ -26,7 +26,7 @@ use pageserver::{
     tenant::{dump_layerfile_from_path, metadata::TimelineMetadata},
     virtual_file,
 };
-use pageserver_api::{config::defaults::DEFAULT_IO_BUFFER_ALIGNMENT, shard::TenantShardId};
+use pageserver_api::shard::TenantShardId;
 use postgres_ffi::ControlFileData;
 use remote_storage::{RemotePath, RemoteStorageConfig};
 use tokio_util::sync::CancellationToken;
@@ -205,11 +205,7 @@ fn read_pg_control_file(control_file_path: &Utf8Path) -> anyhow::Result<()> {
 
 async fn print_layerfile(path: &Utf8Path) -> anyhow::Result<()> {
     // Basic initialization of things that don't change after startup
-    virtual_file::init(
-        10,
-        virtual_file::api::IoEngineKind::StdFs,
-        DEFAULT_IO_BUFFER_ALIGNMENT,
-    );
+    virtual_file::init(10, virtual_file::api::IoEngineKind::StdFs);
     page_cache::init(100);
     let ctx = RequestContext::new(TaskKind::DebugTool, DownloadBehavior::Error);
     dump_layerfile_from_path(path, true, &ctx).await

pageserver/pagebench/src/cmd/getpage_latest_lsn.rs

@@ -59,9 +59,9 @@ pub(crate) struct Args {
     #[clap(long)]
     set_io_engine: Option<pageserver_api::models::virtual_file::IoEngineKind>,
 
-    /// Before starting the benchmark, live-reconfigure the pageserver to use specified alignment for io buffers.
+    /// Before starting the benchmark, live-reconfigure the pageserver to use specified io mode (buffered vs. direct).
     #[clap(long)]
-    set_io_alignment: Option<usize>,
+    set_io_mode: Option<pageserver_api::models::virtual_file::IoMode>,
 
     targets: Option<Vec<TenantTimelineId>>,
 }
@@ -129,8 +129,8 @@ async fn main_impl(
         mgmt_api_client.put_io_engine(engine_str).await?;
     }
 
-    if let Some(align) = args.set_io_alignment {
-        mgmt_api_client.put_io_alignment(align).await?;
+    if let Some(mode) = &args.set_io_mode {
+        mgmt_api_client.put_io_mode(mode).await?;
     }
 
     // discover targets

pageserver/src/bin/pageserver.rs

@@ -125,8 +125,7 @@ fn main() -> anyhow::Result<()> {
 
     // after setting up logging, log the effective IO engine choice and read path implementations
     info!(?conf.virtual_file_io_engine, "starting with virtual_file IO engine");
-    info!(?conf.virtual_file_direct_io, "starting with virtual_file Direct IO settings");
-    info!(?conf.io_buffer_alignment, "starting with setting for IO buffer alignment");
+    info!(?conf.virtual_file_io_mode, "starting with virtual_file IO mode");
 
     // The tenants directory contains all the pageserver local disk state.
     // Create if not exists and make sure all the contents are durable before proceeding.
@@ -168,11 +167,7 @@ fn main() -> anyhow::Result<()> {
     let scenario = failpoint_support::init();
 
     // Basic initialization of things that don't change after startup
-    virtual_file::init(
-        conf.max_file_descriptors,
-        conf.virtual_file_io_engine,
-        conf.io_buffer_alignment,
-    );
+    virtual_file::init(conf.max_file_descriptors, conf.virtual_file_io_engine);
     page_cache::init(conf.page_cache_size);
 
     start_pageserver(launch_ts, conf).context("Failed to start pageserver")?;
@@ -575,7 +570,7 @@ fn start_pageserver(
             .build()
             .map_err(|err| anyhow!(err))?;
         let service = utils::http::RouterService::new(router).unwrap();
-        let server = hyper::Server::from_tcp(http_listener)?
+        let server = hyper0::Server::from_tcp(http_listener)?
             .serve(service)
             .with_graceful_shutdown({
                 let cancel = cancel.clone();

pageserver/src/config.rs

@@ -174,9 +174,7 @@ pub struct PageServerConf {
     pub l0_flush: crate::l0_flush::L0FlushConfig,
 
     /// Direct IO settings
-    pub virtual_file_direct_io: virtual_file::DirectIoMode,
-
-    pub io_buffer_alignment: usize,
+    pub virtual_file_io_mode: virtual_file::IoMode,
 }
 
 /// Token for authentication to safekeepers
@@ -325,11 +323,10 @@ impl PageServerConf {
             image_compression,
             ephemeral_bytes_per_memory_kb,
             l0_flush,
-            virtual_file_direct_io,
+            virtual_file_io_mode,
             concurrent_tenant_warmup,
             concurrent_tenant_size_logical_size_queries,
             virtual_file_io_engine,
-            io_buffer_alignment,
             tenant_config,
         } = config_toml;
 
@@ -368,8 +365,6 @@ impl PageServerConf {
             max_vectored_read_bytes,
             image_compression,
             ephemeral_bytes_per_memory_kb,
-            virtual_file_direct_io,
-            io_buffer_alignment,
 
             // ------------------------------------------------------------
             // fields that require additional validation or custom handling
@@ -408,6 +403,7 @@ impl PageServerConf {
             l0_flush: l0_flush
                 .map(crate::l0_flush::L0FlushConfig::from)
                 .unwrap_or_default(),
+            virtual_file_io_mode: virtual_file_io_mode.unwrap_or(virtual_file::IoMode::preferred()),
         };
 
         // ------------------------------------------------------------

pageserver/src/http/routes.rs

@@ -17,6 +17,7 @@ use hyper::header;
 use hyper::StatusCode;
 use hyper::{Body, Request, Response, Uri};
 use metrics::launch_timestamp::LaunchTimestamp;
+use pageserver_api::models::virtual_file::IoMode;
 use pageserver_api::models::AuxFilePolicy;
 use pageserver_api::models::DownloadRemoteLayersTaskSpawnRequest;
 use pageserver_api::models::IngestAuxFilesRequest;
@@ -56,6 +57,7 @@ use utils::http::endpoint::request_span;
 use utils::http::request::must_parse_query_param;
 use utils::http::request::{get_request_param, must_get_query_param, parse_query_param};
 
+use crate::config::PageServerConf;
 use crate::context::{DownloadBehavior, RequestContext};
 use crate::deletion_queue::DeletionQueueClient;
 use crate::pgdatadir_mapping::LsnForTimestamp;
@@ -80,7 +82,6 @@ use crate::tenant::timeline::CompactionError;
 use crate::tenant::timeline::Timeline;
 use crate::tenant::GetTimelineError;
 use crate::tenant::{LogicalSizeCalculationCause, PageReconstructError};
-use crate::{config::PageServerConf, tenant::mgr};
 use crate::{disk_usage_eviction_task, tenant};
 use pageserver_api::models::{
     StatusResponse, TenantConfigRequest, TenantInfo, TimelineCreateRequest, TimelineGcRequest,
@@ -824,7 +825,7 @@ async fn get_lsn_by_timestamp_handler(
 
     let lease = if with_lease {
         timeline
-            .make_lsn_lease(lsn, timeline.get_lsn_lease_length_for_ts(), &ctx)
+            .init_lsn_lease(lsn, timeline.get_lsn_lease_length_for_ts(), &ctx)
             .inspect_err(|_| {
                 warn!("fail to grant a lease to {}", lsn);
             })
@@ -1692,9 +1693,18 @@ async fn lsn_lease_handler(
     let timeline =
         active_timeline_of_active_tenant(&state.tenant_manager, tenant_shard_id, timeline_id)
             .await?;
-    let result = timeline
-        .make_lsn_lease(lsn, timeline.get_lsn_lease_length(), &ctx)
-        .map_err(|e| ApiError::InternalServerError(e.context("lsn lease http handler")))?;
+
+    let result = async {
+        timeline
+            .init_lsn_lease(lsn, timeline.get_lsn_lease_length(), &ctx)
+            .map_err(|e| {
+                ApiError::InternalServerError(
+                    e.context(format!("invalid lsn lease request at {lsn}")),
+                )
+            })
+    }
+    .instrument(info_span!("init_lsn_lease", tenant_id = %tenant_shard_id.tenant_id, shard_id = %tenant_shard_id.shard_slug(), %timeline_id))
+    .await?;
 
     json_response(StatusCode::OK, result)
 }
@@ -1710,8 +1720,13 @@ async fn timeline_gc_handler(
 
     let gc_req: TimelineGcRequest = json_request(&mut request).await?;
 
+    let state = get_state(&request);
+
     let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
-    let gc_result = mgr::immediate_gc(tenant_shard_id, timeline_id, gc_req, cancel, &ctx).await?;
+    let gc_result = state
+        .tenant_manager
+        .immediate_gc(tenant_shard_id, timeline_id, gc_req, cancel, &ctx)
+        .await?;
 
     json_response(StatusCode::OK, gc_result)
 }
@@ -1728,6 +1743,10 @@ async fn timeline_compact_handler(
     let state = get_state(&request);
 
     let mut flags = EnumSet::empty();
+
+    if Some(true) == parse_query_param::<_, bool>(&request, "force_l0_compaction")? {
+        flags |= CompactFlags::ForceL0Compaction;
+    }
     if Some(true) == parse_query_param::<_, bool>(&request, "force_repartition")? {
         flags |= CompactFlags::ForceRepartition;
     }
@@ -1774,6 +1793,9 @@ async fn timeline_checkpoint_handler(
     let state = get_state(&request);
 
     let mut flags = EnumSet::empty();
+    if Some(true) == parse_query_param::<_, bool>(&request, "force_l0_compaction")? {
+        flags |= CompactFlags::ForceL0Compaction;
+    }
     if Some(true) == parse_query_param::<_, bool>(&request, "force_repartition")? {
         flags |= CompactFlags::ForceRepartition;
     }
@@ -2358,17 +2380,13 @@ async fn put_io_engine_handler(
     json_response(StatusCode::OK, ())
 }
 
-async fn put_io_alignment_handler(
+async fn put_io_mode_handler(
     mut r: Request<Body>,
     _cancel: CancellationToken,
 ) -> Result<Response<Body>, ApiError> {
     check_permission(&r, None)?;
-    let align: usize = json_request(&mut r).await?;
-    crate::virtual_file::set_io_buffer_alignment(align).map_err(|align| {
-        ApiError::PreconditionFailed(
-            format!("Requested io alignment ({align}) is not a power of two").into(),
-        )
-    })?;
+    let mode: IoMode = json_request(&mut r).await?;
+    crate::virtual_file::set_io_mode(mode);
     json_response(StatusCode::OK, ())
 }
 
@@ -3059,9 +3077,7 @@ pub fn make_router(
             |r| api_handler(r, timeline_collect_keyspace),
         )
         .put("/v1/io_engine", |r| api_handler(r, put_io_engine_handler))
-        .put("/v1/io_alignment", |r| {
-            api_handler(r, put_io_alignment_handler)
-        })
+        .put("/v1/io_mode", |r| api_handler(r, put_io_mode_handler))
         .put(
             "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/force_aux_policy_switch",
             |r| api_handler(r, force_aux_policy_switch_handler),

pageserver/src/lib.rs

@@ -13,6 +13,8 @@ pub mod http;
 pub mod import_datadir;
 pub mod l0_flush;
 
+extern crate hyper0 as hyper;
+
 use futures::{stream::FuturesUnordered, StreamExt};
 pub use pageserver_api::keyspace;
 use tokio_util::sync::CancellationToken;

pageserver/src/page_service.rs

@@ -273,10 +273,20 @@ async fn page_service_conn_main(
                 info!("Postgres client disconnected ({io_error})");
                 Ok(())
             } else {
-                Err(io_error).context("Postgres connection error")
+                let tenant_id = conn_handler.timeline_handles.tenant_id();
+                Err(io_error).context(format!(
+                    "Postgres connection error for tenant_id={:?} client at peer_addr={}",
+                    tenant_id, peer_addr
+                ))
             }
         }
-        other => other.context("Postgres query error"),
+        other => {
+            let tenant_id = conn_handler.timeline_handles.tenant_id();
+            other.context(format!(
+                "Postgres query error for tenant_id={:?} client peer_addr={}",
+                tenant_id, peer_addr
+            ))
+        }
     }
 }
 
@@ -340,6 +350,10 @@ impl TimelineHandles {
                 }
             })
     }
+
+    fn tenant_id(&self) -> Option<TenantId> {
+        self.wrapper.tenant_id.get().copied()
+    }
 }
 
 pub(crate) struct TenantManagerWrapper {
@@ -819,7 +833,7 @@ impl PageServerHandler {
         set_tracing_field_shard_id(&timeline);
 
         let lease = timeline
-            .make_lsn_lease(lsn, timeline.get_lsn_lease_length(), ctx)
+            .renew_lsn_lease(lsn, timeline.get_lsn_lease_length(), ctx)
             .inspect_err(|e| {
                 warn!("{e}");
             })

pageserver/src/tenant.rs

@@ -21,6 +21,7 @@ use futures::stream::FuturesUnordered;
 use futures::StreamExt;
 use pageserver_api::models;
 use pageserver_api::models::AuxFilePolicy;
+use pageserver_api::models::LsnLease;
 use pageserver_api::models::TimelineArchivalState;
 use pageserver_api::models::TimelineState;
 use pageserver_api::models::TopTenantShardItem;
@@ -96,6 +97,7 @@ use crate::tenant::remote_timeline_client::MaybeDeletedIndexPart;
 use crate::tenant::remote_timeline_client::INITDB_PATH;
 use crate::tenant::storage_layer::DeltaLayer;
 use crate::tenant::storage_layer::ImageLayer;
+use crate::walingest::WalLagCooldown;
 use crate::walredo;
 use crate::InitializationOrder;
 use std::collections::hash_map::Entry;
@@ -182,27 +184,54 @@ pub struct TenantSharedResources {
 pub(super) struct AttachedTenantConf {
     tenant_conf: TenantConfOpt,
     location: AttachedLocationConfig,
+    /// The deadline before which we are blocked from GC so that
+    /// leases have a chance to be renewed.
+    lsn_lease_deadline: Option<tokio::time::Instant>,
 }
 
 impl AttachedTenantConf {
     fn new(tenant_conf: TenantConfOpt, location: AttachedLocationConfig) -> Self {
+        // Sets a deadline before which we cannot proceed to GC due to lsn lease.
+        //
+        // We do this as the leases mapping are not persisted to disk. By delaying GC by lease
+        // length, we guarantee that all the leases we granted before will have a chance to renew
+        // when we run GC for the first time after restart / transition from AttachedMulti to AttachedSingle.
+        let lsn_lease_deadline = if location.attach_mode == AttachmentMode::Single {
+            Some(
+                tokio::time::Instant::now()
+                    + tenant_conf
+                        .lsn_lease_length
+                        .unwrap_or(LsnLease::DEFAULT_LENGTH),
+            )
+        } else {
+            // We don't use `lsn_lease_deadline` to delay GC in AttachedMulti and AttachedStale
+            // because we don't do GC in these modes.
+            None
+        };
+
         Self {
             tenant_conf,
             location,
+            lsn_lease_deadline,
         }
     }
 
     fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
         match &location_conf.mode {
-            LocationMode::Attached(attach_conf) => Ok(Self {
-                tenant_conf: location_conf.tenant_conf,
-                location: *attach_conf,
-            }),
+            LocationMode::Attached(attach_conf) => {
+                Ok(Self::new(location_conf.tenant_conf, *attach_conf))
+            }
             LocationMode::Secondary(_) => {
                 anyhow::bail!("Attempted to construct AttachedTenantConf from a LocationConf in secondary mode")
             }
         }
     }
+
+    fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
+        self.lsn_lease_deadline
+            .map(|d| tokio::time::Instant::now() < d)
+            .unwrap_or(false)
+    }
 }
 struct TimelinePreload {
     timeline_id: TimelineId,
@@ -291,6 +320,9 @@ pub struct Tenant {
     /// background warmup.
     pub(crate) activate_now_sem: tokio::sync::Semaphore,
 
+    /// Time it took for the tenant to activate. Zero if not active yet.
+    attach_wal_lag_cooldown: Arc<std::sync::OnceLock<WalLagCooldown>>,
+
     // Cancellation token fires when we have entered shutdown().  This is a parent of
     // Timelines' cancellation token.
     pub(crate) cancel: CancellationToken,
@@ -972,11 +1004,15 @@ impl Tenant {
                 // Remote preload is complete.
                 drop(remote_load_completion);
 
+
                 // We will time the duration of the attach phase unless this is a creation (attach will do no work)
+                let attach_start = std::time::Instant::now();
                 let attached = {
                     let _attach_timer = Some(TENANT.attach.start_timer());
                     tenant_clone.attach(preload, &ctx).await
                 };
+                let attach_duration = attach_start.elapsed();
+                _ = tenant_clone.attach_wal_lag_cooldown.set(WalLagCooldown::new(attach_start, attach_duration));
 
                 match attached {
                     Ok(()) => {
@@ -1822,6 +1858,11 @@ impl Tenant {
                 info!("Skipping GC in location state {:?}", conf.location);
                 return Ok(GcResult::default());
             }
+
+            if conf.is_gc_blocked_by_lsn_lease_deadline() {
+                info!("Skipping GC because lsn lease deadline is not reached");
+                return Ok(GcResult::default());
+            }
         }
 
         let _guard = match self.gc_block.start().await {
@@ -2630,6 +2671,8 @@ impl Tenant {
             Arc::new(AttachedTenantConf {
                 tenant_conf: new_tenant_conf.clone(),
                 location: inner.location,
+                // Attached location is not changed, no need to update lsn lease deadline.
+                lsn_lease_deadline: inner.lsn_lease_deadline,
             })
         });
 
@@ -2719,6 +2762,7 @@ impl Tenant {
             pg_version,
             state,
             last_aux_file_policy,
+            self.attach_wal_lag_cooldown.clone(),
             self.cancel.child_token(),
         );
 
@@ -2825,6 +2869,7 @@ impl Tenant {
                 Some(Duration::from_secs(3600 * 24)),
             )),
             activate_now_sem: tokio::sync::Semaphore::new(0),
+            attach_wal_lag_cooldown: Arc::new(std::sync::OnceLock::new()),
             cancel: CancellationToken::default(),
             gate: Gate::default(),
             timeline_get_throttle: Arc::new(throttle::Throttle::new(
@@ -3887,9 +3932,9 @@ async fn run_initdb(
     let _permit = INIT_DB_SEMAPHORE.acquire().await;
 
     let initdb_command = tokio::process::Command::new(&initdb_bin_path)
-        .args(["-D", initdb_target_dir.as_ref()])
-        .args(["-U", &conf.superuser])
-        .args(["-E", "utf8"])
+        .args(["--pgdata", initdb_target_dir.as_ref()])
+        .args(["--username", &conf.superuser])
+        .args(["--encoding", "utf8"])
         .arg("--no-instructions")
         .arg("--no-sync")
         .env_clear()
@@ -4461,13 +4506,17 @@ mod tests {
         tline.freeze_and_flush().await.map_err(|e| e.into())
     }
 
-    #[tokio::test]
+    #[tokio::test(start_paused = true)]
     async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
         let (tenant, ctx) =
             TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
                 .await?
                 .load()
                 .await;
+        // Advance to the lsn lease deadline so that GC is not blocked by
+        // initial transition into AttachedSingle.
+        tokio::time::advance(tenant.get_lsn_lease_length()).await;
+        tokio::time::resume();
         let tline = tenant
             .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
             .await?;
@@ -7244,9 +7293,17 @@ mod tests {
         Ok(())
     }
 
-    #[tokio::test]
+    #[tokio::test(start_paused = true)]
     async fn test_lsn_lease() -> anyhow::Result<()> {
-        let (tenant, ctx) = TenantHarness::create("test_lsn_lease").await?.load().await;
+        let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
+            .await
+            .unwrap()
+            .load()
+            .await;
+        // Advance to the lsn lease deadline so that GC is not blocked by
+        // initial transition into AttachedSingle.
+        tokio::time::advance(tenant.get_lsn_lease_length()).await;
+        tokio::time::resume();
         let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
 
         let end_lsn = Lsn(0x100);
@@ -7274,24 +7331,33 @@ mod tests {
 
         let leased_lsns = [0x30, 0x50, 0x70];
         let mut leases = Vec::new();
-        let _: anyhow::Result<_> = leased_lsns.iter().try_for_each(|n| {
-            leases.push(timeline.make_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)?);
-            Ok(())
+        leased_lsns.iter().for_each(|n| {
+            leases.push(
+                timeline
+                    .init_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)
+                    .expect("lease request should succeed"),
+            );
         });
 
-        // Renewing with shorter lease should not change the lease.
-        let updated_lease_0 =
-            timeline.make_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)?;
-        assert_eq!(updated_lease_0.valid_until, leases[0].valid_until);
+        let updated_lease_0 = timeline
+            .renew_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)
+            .expect("lease renewal should succeed");
+        assert_eq!(
+            updated_lease_0.valid_until, leases[0].valid_until,
+            " Renewing with shorter lease should not change the lease."
+        );
 
-        // Renewing with a long lease should renew lease with later expiration time.
-        let updated_lease_1 = timeline.make_lsn_lease(
-            Lsn(leased_lsns[1]),
-            timeline.get_lsn_lease_length() * 2,
-            &ctx,
-        )?;
-
-        assert!(updated_lease_1.valid_until > leases[1].valid_until);
+        let updated_lease_1 = timeline
+            .renew_lsn_lease(
+                Lsn(leased_lsns[1]),
+                timeline.get_lsn_lease_length() * 2,
+                &ctx,
+            )
+            .expect("lease renewal should succeed");
+        assert!(
+            updated_lease_1.valid_until > leases[1].valid_until,
+            "Renewing with a long lease should renew lease with later expiration time."
+        );
 
         // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
         info!(
@@ -7308,7 +7374,8 @@ mod tests {
                 &CancellationToken::new(),
                 &ctx,
             )
-            .await?;
+            .await
+            .unwrap();
 
         // Keeping everything <= Lsn(0x80) b/c leases:
         // 0/10: initdb layer
@@ -7322,13 +7389,16 @@ mod tests {
         // Make lease on a already GC-ed LSN.
         // 0/80 does not have a valid lease + is below latest_gc_cutoff
         assert!(Lsn(0x80) < *timeline.get_latest_gc_cutoff_lsn());
-        let res = timeline.make_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx);
-        assert!(res.is_err());
+        timeline
+            .init_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx)
+            .expect_err("lease request on GC-ed LSN should fail");
 
         // Should still be able to renew a currently valid lease
         // Assumption: original lease to is still valid for 0/50.
-        let _ =
-            timeline.make_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)?;
+        // (use `Timeline::init_lsn_lease` for testing so it always does validation)
+        timeline
+            .init_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)
+            .expect("lease renewal with validation should succeed");
 
         Ok(())
     }

pageserver/src/tenant/config.rs

@@ -8,7 +8,6 @@
 //! We cannot use global or default config instead, because wrong settings
 //! may lead to a data loss.
 //!
-use anyhow::bail;
 pub(crate) use pageserver_api::config::TenantConfigToml as TenantConf;
 use pageserver_api::models::AuxFilePolicy;
 use pageserver_api::models::CompactionAlgorithmSettings;
@@ -441,29 +440,6 @@ impl TryFrom<&'_ models::TenantConfig> for TenantConfOpt {
     }
 }
 
-impl TryFrom<toml_edit::Item> for TenantConfOpt {
-    type Error = anyhow::Error;
-
-    fn try_from(item: toml_edit::Item) -> Result<Self, Self::Error> {
-        match item {
-            toml_edit::Item::Value(value) => {
-                let d = value.into_deserializer();
-                return serde_path_to_error::deserialize(d)
-                    .map_err(|e| anyhow::anyhow!("{}: {}", e.path(), e.inner().message()));
-            }
-            toml_edit::Item::Table(table) => {
-                let deserializer =
-                    toml_edit::de::Deserializer::from(toml_edit::DocumentMut::from(table));
-                return serde_path_to_error::deserialize(deserializer)
-                    .map_err(|e| anyhow::anyhow!("{}: {}", e.path(), e.inner().message()));
-            }
-            _ => {
-                bail!("expected non-inline table but found {item}")
-            }
-        }
-    }
-}
-
 /// This is a conversion from our internal tenant config object to the one used
 /// in external APIs.
 impl From<TenantConfOpt> for models::TenantConfig {

pageserver/src/tenant/disk_btree.rs

@@ -260,11 +260,11 @@ where
         Ok(result)
     }
 
-    pub fn iter<'a>(self, start_key: &'a [u8; L], ctx: &'a RequestContext) -> DiskBtreeIter<'a>
+    pub fn iter<'a>(self, start_key: &'a [u8; L], ctx: &'a RequestContext) -> DiskBtreeIterator<'a>
     where
         R: 'a + Send,
     {
-        DiskBtreeIter {
+        DiskBtreeIterator {
             stream: Box::pin(self.into_stream(start_key, ctx)),
         }
     }
@@ -525,128 +525,19 @@ where
     }
 }
 
-pub trait DiskBtreeIterator {
-    type Item;
-    async fn next(&mut self) -> Option<std::result::Result<Self::Item, DiskBtreeError>>;
-    fn map<B, F>(self, f: F) -> MapDiskBtreeIter<Self, F>
-    where
-        Self: Sized,
-        F: FnMut(Self::Item) -> B,
-    {
-        MapDiskBtreeIter::new(self, f)
-    }
-
-    fn filter<P>(self, predicate: P) -> FilterDiskBtreeIter<Self, P>
-    where
-        Self: Sized,
-        P: FnMut(&Self::Item) -> bool,
-    {
-        FilterDiskBtreeIter::new(self, predicate)
-    }
-}
-
-pub struct DiskBtreeIter<'a> {
+pub struct DiskBtreeIterator<'a> {
     #[allow(clippy::type_complexity)]
     stream: std::pin::Pin<
         Box<dyn Stream<Item = std::result::Result<(Vec<u8>, u64), DiskBtreeError>> + 'a + Send>,
     >,
 }
 
-impl<'a> DiskBtreeIterator for DiskBtreeIter<'a> {
-    type Item = (Vec<u8>, u64);
-
-    async fn next(&mut self) -> Option<std::result::Result<Self::Item, DiskBtreeError>> {
+impl<'a> DiskBtreeIterator<'a> {
+    pub async fn next(&mut self) -> Option<std::result::Result<(Vec<u8>, u64), DiskBtreeError>> {
         self.stream.next().await
     }
 }
 
-pub struct MapDiskBtreeIter<I, F> {
-    iter: I,
-    f: F,
-}
-
-impl<I, F> MapDiskBtreeIter<I, F> {
-    pub fn new(iter: I, f: F) -> Self {
-        MapDiskBtreeIter { iter, f }
-    }
-}
-
-impl<B, I: DiskBtreeIterator, F> DiskBtreeIterator for MapDiskBtreeIter<I, F>
-where
-    F: FnMut(I::Item) -> B,
-{
-    type Item = B;
-
-    async fn next(&mut self) -> Option<std::result::Result<Self::Item, DiskBtreeError>> {
-        self.iter.next().await.map(|res| res.map(|x| (self.f)(x)))
-    }
-}
-
-pub struct FilterDiskBtreeIter<I, P> {
-    iter: I,
-    predicate: P,
-}
-
-impl<I, P> FilterDiskBtreeIter<I, P> {
-    pub fn new(iter: I, predicate: P) -> Self {
-        FilterDiskBtreeIter { iter, predicate }
-    }
-}
-
-impl<I: DiskBtreeIterator, P> DiskBtreeIterator for FilterDiskBtreeIter<I, P>
-where
-    P: FnMut(&I::Item) -> bool,
-{
-    type Item = I::Item;
-
-    async fn next(&mut self) -> Option<std::result::Result<Self::Item, DiskBtreeError>> {
-        self.iter.next().await.and_then(|res| match res {
-            Ok(x) => {
-                if (self.predicate)(&x) {
-                    Some(Ok(x))
-                } else {
-                    None
-                }
-            }
-            Err(e) => Some(Err(e)),
-        })
-    }
-}
-
-// pub struct DeltaLayerDiskBtreeIterator<'a, F> {
-//     inner: DiskBtreeIterator<'a, F, >,
-//     extract: F,
-// }
-
-// impl<'a, F> DeltaLayerDiskBtreeIterator<'a, F>
-// where
-//     F: FnMut(Vec<u8>, u64) -> (Key, Lsn, u64),
-// {
-//     pub async fn next(&mut self) -> Option<std::result::Result<(Key, Lsn, u64), DiskBtreeError>> {
-//         self.inner
-//             .next()
-//             .await
-//             .map(|res| res.map(|(raw_key, value)| (self.extract)(raw_key, value)))
-//     }
-// }
-
-// pub struct ImageLayerDiskBtreeIterator<'a, F> {
-//     inner: DiskBtreeIterator<'a>,
-//     extract: F,
-// }
-
-// impl<'a, F> ImageLayerDiskBtreeIterator<'a, F>
-// where
-//     F: FnMut(Vec<u8>, u64) -> (Key, Lsn, u64),
-// {
-//     pub async fn next(&mut self) -> Option<std::result::Result<(Key, Lsn, u64), DiskBtreeError>> {
-//         self.inner
-//             .next()
-//             .await
-//             .map(|res| res.map(|(raw_key, value)| (self.extract)(raw_key, value)))
-//     }
-// }
-
 ///
 /// Public builder object, for creating a new tree.
 ///

pageserver/src/tenant/ephemeral_file.rs

@@ -84,7 +84,7 @@ impl Drop for EphemeralFile {
     fn drop(&mut self) {
         // unlink the file
         // we are clear to do this, because we have entered a gate
-        let path = &self.buffered_writer.as_inner().as_inner().path;
+        let path = self.buffered_writer.as_inner().as_inner().path();
         let res = std::fs::remove_file(path);
         if let Err(e) = res {
             if e.kind() != std::io::ErrorKind::NotFound {
@@ -356,7 +356,7 @@ mod tests {
         }
 
         let file_contents =
-            std::fs::read(&file.buffered_writer.as_inner().as_inner().path).unwrap();
+            std::fs::read(file.buffered_writer.as_inner().as_inner().path()).unwrap();
         assert_eq!(file_contents, &content[0..cap]);
 
         let buffer_contents = file.buffered_writer.inspect_buffer();
@@ -392,7 +392,7 @@ mod tests {
             .buffered_writer
             .as_inner()
             .as_inner()
-            .path
+            .path()
             .metadata()
             .unwrap();
         assert_eq!(

pageserver/src/tenant/gc_block.rs

@@ -1,29 +1,12 @@
-use std::{collections::HashMap, time::Duration};
+use std::collections::HashMap;
 
-use super::remote_timeline_client::index::GcBlockingReason;
-use tokio::time::Instant;
 use utils::id::TimelineId;
 
-type TimelinesBlocked = HashMap<TimelineId, enumset::EnumSet<GcBlockingReason>>;
+use super::remote_timeline_client::index::GcBlockingReason;
 
-#[derive(Default)]
-struct Storage {
-    timelines_blocked: TimelinesBlocked,
-    /// The deadline before which we are blocked from GC so that
-    /// leases have a chance to be renewed.
-    lsn_lease_deadline: Option<Instant>,
-}
+type Storage = HashMap<TimelineId, enumset::EnumSet<GcBlockingReason>>;
 
-impl Storage {
-    fn is_blocked_by_lsn_lease_deadline(&self) -> bool {
-        self.lsn_lease_deadline
-            .map(|d| Instant::now() < d)
-            .unwrap_or(false)
-    }
-}
-
-/// GcBlock provides persistent (per-timeline) gc blocking and facilitates transient time based gc
-/// blocking.
+/// GcBlock provides persistent (per-timeline) gc blocking.
 #[derive(Default)]
 pub(crate) struct GcBlock {
     /// The timelines which have current reasons to block gc.
@@ -66,17 +49,6 @@ impl GcBlock {
         }
     }
 
-    /// Sets a deadline before which we cannot proceed to GC due to lsn lease.
-    ///
-    /// We do this as the leases mapping are not persisted to disk. By delaying GC by lease
-    /// length, we guarantee that all the leases we granted before will have a chance to renew
-    /// when we run GC for the first time after restart / transition from AttachedMulti to AttachedSingle.
-    pub(super) fn set_lsn_lease_deadline(&self, lsn_lease_length: Duration) {
-        let deadline = Instant::now() + lsn_lease_length;
-        let mut g = self.reasons.lock().unwrap();
-        g.lsn_lease_deadline = Some(deadline);
-    }
-
     /// Describe the current gc blocking reasons.
     ///
     /// TODO: make this json serializable.
@@ -102,7 +74,7 @@ impl GcBlock {
     ) -> anyhow::Result<bool> {
         let (added, uploaded) = {
             let mut g = self.reasons.lock().unwrap();
-            let set = g.timelines_blocked.entry(timeline.timeline_id).or_default();
+            let set = g.entry(timeline.timeline_id).or_default();
             let added = set.insert(reason);
 
             // LOCK ORDER: intentionally hold the lock, see self.reasons.
@@ -133,7 +105,7 @@ impl GcBlock {
 
         let (remaining_blocks, uploaded) = {
             let mut g = self.reasons.lock().unwrap();
-            match g.timelines_blocked.entry(timeline.timeline_id) {
+            match g.entry(timeline.timeline_id) {
                 Entry::Occupied(mut oe) => {
                     let set = oe.get_mut();
                     set.remove(reason);
@@ -147,7 +119,7 @@ impl GcBlock {
                 }
             }
 
-            let remaining_blocks = g.timelines_blocked.len();
+            let remaining_blocks = g.len();
 
             // LOCK ORDER: intentionally hold the lock while scheduling; see self.reasons
             let uploaded = timeline
@@ -172,11 +144,11 @@ impl GcBlock {
     pub(crate) fn before_delete(&self, timeline: &super::Timeline) {
         let unblocked = {
             let mut g = self.reasons.lock().unwrap();
-            if g.timelines_blocked.is_empty() {
+            if g.is_empty() {
                 return;
             }
 
-            g.timelines_blocked.remove(&timeline.timeline_id);
+            g.remove(&timeline.timeline_id);
 
             BlockingReasons::clean_and_summarize(g).is_none()
         };
@@ -187,11 +159,10 @@ impl GcBlock {
     }
 
     /// Initialize with the non-deleted timelines of this tenant.
-    pub(crate) fn set_scanned(&self, scanned: TimelinesBlocked) {
+    pub(crate) fn set_scanned(&self, scanned: Storage) {
         let mut g = self.reasons.lock().unwrap();
-        assert!(g.timelines_blocked.is_empty());
-        g.timelines_blocked
-            .extend(scanned.into_iter().filter(|(_, v)| !v.is_empty()));
+        assert!(g.is_empty());
+        g.extend(scanned.into_iter().filter(|(_, v)| !v.is_empty()));
 
         if let Some(reasons) = BlockingReasons::clean_and_summarize(g) {
             tracing::info!(summary=?reasons, "initialized with gc blocked");
@@ -205,7 +176,6 @@ pub(super) struct Guard<'a> {
 
 #[derive(Debug)]
 pub(crate) struct BlockingReasons {
-    tenant_blocked_by_lsn_lease_deadline: bool,
     timelines: usize,
     reasons: enumset::EnumSet<GcBlockingReason>,
 }
@@ -214,8 +184,8 @@ impl std::fmt::Display for BlockingReasons {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
         write!(
             f,
-            "tenant_blocked_by_lsn_lease_deadline: {}, {} timelines block for {:?}",
-            self.tenant_blocked_by_lsn_lease_deadline, self.timelines, self.reasons
+            "{} timelines block for {:?}",
+            self.timelines, self.reasons
         )
     }
 }
@@ -223,15 +193,13 @@ impl std::fmt::Display for BlockingReasons {
 impl BlockingReasons {
     fn clean_and_summarize(mut g: std::sync::MutexGuard<'_, Storage>) -> Option<Self> {
         let mut reasons = enumset::EnumSet::empty();
-        g.timelines_blocked.retain(|_key, value| {
+        g.retain(|_key, value| {
             reasons = reasons.union(*value);
             !value.is_empty()
         });
-        let blocked_by_lsn_lease_deadline = g.is_blocked_by_lsn_lease_deadline();
-        if !g.timelines_blocked.is_empty() || blocked_by_lsn_lease_deadline {
+        if !g.is_empty() {
             Some(BlockingReasons {
-                tenant_blocked_by_lsn_lease_deadline: blocked_by_lsn_lease_deadline,
-                timelines: g.timelines_blocked.len(),
+                timelines: g.len(),
                 reasons,
             })
         } else {
@@ -240,17 +208,14 @@ impl BlockingReasons {
     }
 
     fn summarize(g: &std::sync::MutexGuard<'_, Storage>) -> Option<Self> {
-        let blocked_by_lsn_lease_deadline = g.is_blocked_by_lsn_lease_deadline();
-        if g.timelines_blocked.is_empty() && !blocked_by_lsn_lease_deadline {
+        if g.is_empty() {
             None
         } else {
             let reasons = g
-                .timelines_blocked
                 .values()
                 .fold(enumset::EnumSet::empty(), |acc, next| acc.union(*next));
             Some(BlockingReasons {
-                tenant_blocked_by_lsn_lease_deadline: blocked_by_lsn_lease_deadline,
-                timelines: g.timelines_blocked.len(),
+                timelines: g.len(),
                 reasons,
             })
         }

pageserver/src/tenant/mgr.rs

@@ -219,7 +219,11 @@ async fn safe_rename_tenant_dir(path: impl AsRef<Utf8Path>) -> std::io::Result<U
         + TEMP_FILE_SUFFIX;
     let tmp_path = path_with_suffix_extension(&path, &rand_suffix);
     fs::rename(path.as_ref(), &tmp_path).await?;
-    fs::File::open(parent).await?.sync_all().await?;
+    fs::File::open(parent)
+        .await?
+        .sync_all()
+        .await
+        .maybe_fatal_err("safe_rename_tenant_dir")?;
     Ok(tmp_path)
 }
 
@@ -949,12 +953,6 @@ impl TenantManager {
                 (LocationMode::Attached(attach_conf), Some(TenantSlot::Attached(tenant))) => {
                     match attach_conf.generation.cmp(&tenant.generation) {
                         Ordering::Equal => {
-                            if attach_conf.attach_mode == AttachmentMode::Single {
-                                tenant
-                                    .gc_block
-                                    .set_lsn_lease_deadline(tenant.get_lsn_lease_length());
-                            }
-
                             // A transition from Attached to Attached in the same generation, we may
                             // take our fast path and just provide the updated configuration
                             // to the tenant.
@@ -2199,6 +2197,82 @@ impl TenantManager {
 
         Ok((wanted_bytes, shard_count as u32))
     }
+
+    #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), %timeline_id))]
+    pub(crate) async fn immediate_gc(
+        &self,
+        tenant_shard_id: TenantShardId,
+        timeline_id: TimelineId,
+        gc_req: TimelineGcRequest,
+        cancel: CancellationToken,
+        ctx: &RequestContext,
+    ) -> Result<GcResult, ApiError> {
+        let tenant = {
+            let guard = self.tenants.read().unwrap();
+            guard
+                .get(&tenant_shard_id)
+                .cloned()
+                .with_context(|| format!("tenant {tenant_shard_id}"))
+                .map_err(|e| ApiError::NotFound(e.into()))?
+        };
+
+        let gc_horizon = gc_req.gc_horizon.unwrap_or_else(|| tenant.get_gc_horizon());
+        // Use tenant's pitr setting
+        let pitr = tenant.get_pitr_interval();
+
+        tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
+
+        // Run in task_mgr to avoid race with tenant_detach operation
+        let ctx: RequestContext =
+            ctx.detached_child(TaskKind::GarbageCollector, DownloadBehavior::Download);
+
+        let _gate_guard = tenant.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
+
+        fail::fail_point!("immediate_gc_task_pre");
+
+        #[allow(unused_mut)]
+        let mut result = tenant
+            .gc_iteration(Some(timeline_id), gc_horizon, pitr, &cancel, &ctx)
+            .await;
+        // FIXME: `gc_iteration` can return an error for multiple reasons; we should handle it
+        // better once the types support it.
+
+        #[cfg(feature = "testing")]
+        {
+            // we need to synchronize with drop completion for python tests without polling for
+            // log messages
+            if let Ok(result) = result.as_mut() {
+                let mut js = tokio::task::JoinSet::new();
+                for layer in std::mem::take(&mut result.doomed_layers) {
+                    js.spawn(layer.wait_drop());
+                }
+                tracing::info!(
+                    total = js.len(),
+                    "starting to wait for the gc'd layers to be dropped"
+                );
+                while let Some(res) = js.join_next().await {
+                    res.expect("wait_drop should not panic");
+                }
+            }
+
+            let timeline = tenant.get_timeline(timeline_id, false).ok();
+            let rtc = timeline.as_ref().map(|x| &x.remote_client);
+
+            if let Some(rtc) = rtc {
+                // layer drops schedule actions on remote timeline client to actually do the
+                // deletions; don't care about the shutdown error, just exit fast
+                drop(rtc.wait_completion().await);
+            }
+        }
+
+        result.map_err(|e| match e {
+            GcError::TenantCancelled | GcError::TimelineCancelled => ApiError::ShuttingDown,
+            GcError::TimelineNotFound => {
+                ApiError::NotFound(anyhow::anyhow!("Timeline not found").into())
+            }
+            other => ApiError::InternalServerError(anyhow::anyhow!(other)),
+        })
+    }
 }
 
 #[derive(Debug, thiserror::Error)]
@@ -2343,7 +2417,7 @@ enum TenantSlotDropError {
 /// Errors that can happen any time we are walking the tenant map to try and acquire
 /// the TenantSlot for a particular tenant.
 #[derive(Debug, thiserror::Error)]
-pub enum TenantMapError {
+pub(crate) enum TenantMapError {
     // Tried to read while initializing
     #[error("tenant map is still initializing")]
     StillInitializing,
@@ -2373,7 +2447,7 @@ pub enum TenantMapError {
 /// The `old_value` may be dropped before the SlotGuard is dropped, by calling
 /// `drop_old_value`.  It is an error to call this without shutting down
 /// the conents of `old_value`.
-pub struct SlotGuard {
+pub(crate) struct SlotGuard {
     tenant_shard_id: TenantShardId,
     old_value: Option<TenantSlot>,
     upserted: bool,
@@ -2766,81 +2840,6 @@ use {
     utils::http::error::ApiError,
 };
 
-#[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), %timeline_id))]
-pub(crate) async fn immediate_gc(
-    tenant_shard_id: TenantShardId,
-    timeline_id: TimelineId,
-    gc_req: TimelineGcRequest,
-    cancel: CancellationToken,
-    ctx: &RequestContext,
-) -> Result<GcResult, ApiError> {
-    let tenant = {
-        let guard = TENANTS.read().unwrap();
-        guard
-            .get(&tenant_shard_id)
-            .cloned()
-            .with_context(|| format!("tenant {tenant_shard_id}"))
-            .map_err(|e| ApiError::NotFound(e.into()))?
-    };
-
-    let gc_horizon = gc_req.gc_horizon.unwrap_or_else(|| tenant.get_gc_horizon());
-    // Use tenant's pitr setting
-    let pitr = tenant.get_pitr_interval();
-
-    tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
-
-    // Run in task_mgr to avoid race with tenant_detach operation
-    let ctx: RequestContext =
-        ctx.detached_child(TaskKind::GarbageCollector, DownloadBehavior::Download);
-
-    let _gate_guard = tenant.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
-
-    fail::fail_point!("immediate_gc_task_pre");
-
-    #[allow(unused_mut)]
-    let mut result = tenant
-        .gc_iteration(Some(timeline_id), gc_horizon, pitr, &cancel, &ctx)
-        .await;
-    // FIXME: `gc_iteration` can return an error for multiple reasons; we should handle it
-    // better once the types support it.
-
-    #[cfg(feature = "testing")]
-    {
-        // we need to synchronize with drop completion for python tests without polling for
-        // log messages
-        if let Ok(result) = result.as_mut() {
-            let mut js = tokio::task::JoinSet::new();
-            for layer in std::mem::take(&mut result.doomed_layers) {
-                js.spawn(layer.wait_drop());
-            }
-            tracing::info!(
-                total = js.len(),
-                "starting to wait for the gc'd layers to be dropped"
-            );
-            while let Some(res) = js.join_next().await {
-                res.expect("wait_drop should not panic");
-            }
-        }
-
-        let timeline = tenant.get_timeline(timeline_id, false).ok();
-        let rtc = timeline.as_ref().map(|x| &x.remote_client);
-
-        if let Some(rtc) = rtc {
-            // layer drops schedule actions on remote timeline client to actually do the
-            // deletions; don't care about the shutdown error, just exit fast
-            drop(rtc.wait_completion().await);
-        }
-    }
-
-    result.map_err(|e| match e {
-        GcError::TenantCancelled | GcError::TimelineCancelled => ApiError::ShuttingDown,
-        GcError::TimelineNotFound => {
-            ApiError::NotFound(anyhow::anyhow!("Timeline not found").into())
-        }
-        other => ApiError::InternalServerError(anyhow::anyhow!(other)),
-    })
-}
-
 #[cfg(test)]
 mod tests {
     use std::collections::BTreeMap;

pageserver/src/tenant/remote_timeline_client/download.rs

@@ -27,7 +27,7 @@ use crate::tenant::Generation;
 use crate::virtual_file::owned_buffers_io::io_buf_ext::IoBufExt;
 use crate::virtual_file::{on_fatal_io_error, MaybeFatalIo, VirtualFile};
 use crate::TEMP_FILE_SUFFIX;
-use remote_storage::{DownloadError, GenericRemoteStorage, ListingMode, RemotePath};
+use remote_storage::{DownloadError, DownloadOpts, GenericRemoteStorage, ListingMode, RemotePath};
 use utils::crashsafe::path_with_suffix_extension;
 use utils::id::{TenantId, TimelineId};
 use utils::pausable_failpoint;
@@ -153,7 +153,9 @@ async fn download_object<'a>(
                     .with_context(|| format!("create a destination file for layer '{dst_path}'"))
                     .map_err(DownloadError::Other)?;
 
-                let download = storage.download(src_path, cancel).await?;
+                let download = storage
+                    .download(src_path, &DownloadOpts::default(), cancel)
+                    .await?;
 
                 pausable_failpoint!("before-downloading-layer-stream-pausable");
 
@@ -178,6 +180,7 @@ async fn download_object<'a>(
                 destination_file
                     .flush()
                     .await
+                    .maybe_fatal_err("download_object sync_all")
                     .with_context(|| format!("flush source file at {dst_path}"))
                     .map_err(DownloadError::Other)?;
 
@@ -185,6 +188,7 @@ async fn download_object<'a>(
                 destination_file
                     .sync_all()
                     .await
+                    .maybe_fatal_err("download_object sync_all")
                     .with_context(|| format!("failed to fsync source file at {dst_path}"))
                     .map_err(DownloadError::Other)?;
 
@@ -202,7 +206,9 @@ async fn download_object<'a>(
                     .with_context(|| format!("create a destination file for layer '{dst_path}'"))
                     .map_err(DownloadError::Other)?;
 
-                let mut download = storage.download(src_path, cancel).await?;
+                let mut download = storage
+                    .download(src_path, &DownloadOpts::default(), cancel)
+                    .await?;
 
                 pausable_failpoint!("before-downloading-layer-stream-pausable");
 
@@ -232,6 +238,7 @@ async fn download_object<'a>(
                 destination_file
                     .sync_all()
                     .await
+                    .maybe_fatal_err("download_object sync_all")
                     .with_context(|| format!("failed to fsync source file at {dst_path}"))
                     .map_err(DownloadError::Other)?;
 
@@ -341,7 +348,9 @@ async fn do_download_index_part(
 
     let index_part_bytes = download_retry_forever(
         || async {
-            let download = storage.download(&remote_path, cancel).await?;
+            let download = storage
+                .download(&remote_path, &DownloadOpts::default(), cancel)
+                .await?;
 
             let mut bytes = Vec::new();
 
@@ -523,10 +532,15 @@ pub(crate) async fn download_initdb_tar_zst(
                 .with_context(|| format!("tempfile creation {temp_path}"))
                 .map_err(DownloadError::Other)?;
 
-            let download = match storage.download(&remote_path, cancel).await {
+            let download = match storage
+                .download(&remote_path, &DownloadOpts::default(), cancel)
+                .await
+            {
                 Ok(dl) => dl,
                 Err(DownloadError::NotFound) => {
-                    storage.download(&remote_preserved_path, cancel).await?
+                    storage
+                        .download(&remote_preserved_path, &DownloadOpts::default(), cancel)
+                        .await?
                 }
                 Err(other) => Err(other)?,
             };

pageserver/src/tenant/secondary/downloader.rs

@@ -49,7 +49,7 @@ use futures::Future;
 use metrics::UIntGauge;
 use pageserver_api::models::SecondaryProgress;
 use pageserver_api::shard::TenantShardId;
-use remote_storage::{DownloadError, Etag, GenericRemoteStorage};
+use remote_storage::{DownloadError, DownloadOpts, Etag, GenericRemoteStorage};
 
 use tokio_util::sync::CancellationToken;
 use tracing::{info_span, instrument, warn, Instrument};
@@ -944,36 +944,34 @@ impl<'a> TenantDownloader<'a> {
     ) -> Result<HeatMapDownload, UpdateError> {
         debug_assert_current_span_has_tenant_id();
         let tenant_shard_id = self.secondary_state.get_tenant_shard_id();
-        // TODO: pull up etag check into the request, to do a conditional GET rather than
-        // issuing a GET and then maybe ignoring the response body
-        // (https://github.com/neondatabase/neon/issues/6199)
         tracing::debug!("Downloading heatmap for secondary tenant",);
 
         let heatmap_path = remote_heatmap_path(tenant_shard_id);
         let cancel = &self.secondary_state.cancel;
+        let opts = DownloadOpts {
+            etag: prev_etag.cloned(),
+        };
 
         backoff::retry(
             || async {
-                let download = self
+                let download = match self
                     .remote_storage
-                    .download(&heatmap_path, cancel)
+                    .download(&heatmap_path, &opts, cancel)
                     .await
-                    .map_err(UpdateError::from)?;
+                {
+                    Ok(download) => download,
+                    Err(DownloadError::Unmodified) => return Ok(HeatMapDownload::Unmodified),
+                    Err(err) => return Err(err.into()),
+                };
 
-                SECONDARY_MODE.download_heatmap.inc();
-
-                if Some(&download.etag) == prev_etag {
-                    Ok(HeatMapDownload::Unmodified)
-                } else {
-                    let mut heatmap_bytes = Vec::new();
-                    let mut body = tokio_util::io::StreamReader::new(download.download_stream);
-                    let _size = tokio::io::copy_buf(&mut body, &mut heatmap_bytes).await?;
-                    Ok(HeatMapDownload::Modified(HeatMapModified {
-                        etag: download.etag,
-                        last_modified: download.last_modified,
-                        bytes: heatmap_bytes,
-                    }))
-                }
+                let mut heatmap_bytes = Vec::new();
+                let mut body = tokio_util::io::StreamReader::new(download.download_stream);
+                let _size = tokio::io::copy_buf(&mut body, &mut heatmap_bytes).await?;
+                Ok(HeatMapDownload::Modified(HeatMapModified {
+                    etag: download.etag,
+                    last_modified: download.last_modified,
+                    bytes: heatmap_bytes,
+                }))
             },
             |e| matches!(e, UpdateError::NoData | UpdateError::Cancelled),
             FAILED_DOWNLOAD_WARN_THRESHOLD,
@@ -984,6 +982,7 @@ impl<'a> TenantDownloader<'a> {
         .await
         .ok_or_else(|| UpdateError::Cancelled)
         .and_then(|x| x)
+        .inspect(|_| SECONDARY_MODE.download_heatmap.inc())
     }
 
     /// Download heatmap layers that are not present on local disk, or update their

pageserver/src/tenant/storage_layer/delta_layer.rs

@@ -34,17 +34,17 @@ use crate::repository::{Key, Value, KEY_SIZE};
 use crate::tenant::blob_io::BlobWriter;
 use crate::tenant::block_io::{BlockBuf, BlockCursor, BlockLease, BlockReader, FileBlockReader};
 use crate::tenant::disk_btree::{
-    DiskBtreeBuilder, DiskBtreeIter, DiskBtreeIterator, DiskBtreeReader, VisitDirection,
+    DiskBtreeBuilder, DiskBtreeIterator, DiskBtreeReader, VisitDirection,
 };
 use crate::tenant::storage_layer::layer::S3_UPLOAD_LIMIT;
 use crate::tenant::timeline::GetVectoredError;
 use crate::tenant::vectored_blob_io::{
     BlobFlag, BufView, StreamingVectoredReadPlanner, VectoredBlobReader, VectoredRead,
-    VectoredReadCoalesceMode, VectoredReadPlanner,
+    VectoredReadPlanner,
 };
 use crate::tenant::PageReconstructError;
 use crate::virtual_file::owned_buffers_io::io_buf_ext::{FullSlice, IoBufExt};
-use crate::virtual_file::{self, VirtualFile};
+use crate::virtual_file::{self, MaybeFatalIo, VirtualFile};
 use crate::{walrecord, TEMP_FILE_SUFFIX};
 use crate::{DELTA_FILE_MAGIC, STORAGE_FORMAT_VERSION};
 use anyhow::{anyhow, bail, ensure, Context, Result};
@@ -53,6 +53,7 @@ use camino::{Utf8Path, Utf8PathBuf};
 use futures::StreamExt;
 use itertools::Itertools;
 use pageserver_api::config::MaxVectoredReadBytes;
+use pageserver_api::key::DBDIR_KEY;
 use pageserver_api::keyspace::KeySpace;
 use pageserver_api::models::ImageCompressionAlgorithm;
 use pageserver_api::shard::TenantShardId;
@@ -572,7 +573,7 @@ impl DeltaLayerWriterInner {
         ensure!(
             metadata.len() <= S3_UPLOAD_LIMIT,
             "Created delta layer file at {} of size {} above limit {S3_UPLOAD_LIMIT}!",
-            file.path,
+            file.path(),
             metadata.len()
         );
 
@@ -589,7 +590,9 @@ impl DeltaLayerWriterInner {
         );
 
         // fsync the file
-        file.sync_all().await?;
+        file.sync_all()
+            .await
+            .maybe_fatal_err("delta_layer sync_all")?;
 
         trace!("created delta layer {}", self.path);
 
@@ -788,7 +791,7 @@ impl DeltaLayerInner {
         max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
         ctx: &RequestContext,
     ) -> anyhow::Result<Self> {
-        let file = VirtualFile::open(path, ctx)
+        let file = VirtualFile::open_v2(path, ctx)
             .await
             .context("open layer file")?;
 
@@ -961,14 +964,25 @@ impl DeltaLayerInner {
                 .blobs_at
                 .as_slice()
                 .iter()
-                .map(|(_, blob_meta)| format!("{}@{}", blob_meta.key, blob_meta.lsn))
+                .filter_map(|(_, blob_meta)| {
+                    if blob_meta.key.is_rel_dir_key() || blob_meta.key == DBDIR_KEY {
+                        // The size of values for these keys is unbounded and can
+                        // grow very large in pathological cases.
+                        None
+                    } else {
+                        Some(format!("{}@{}", blob_meta.key, blob_meta.lsn))
+                    }
+                })
                 .join(", ");
-            tracing::warn!(
-                "Oversized vectored read ({} > {}) for keys {}",
-                largest_read_size,
-                read_size_soft_max,
-                offenders
-            );
+
+            if !offenders.is_empty() {
+                tracing::warn!(
+                    "Oversized vectored read ({} > {}) for keys {}",
+                    largest_read_size,
+                    read_size_soft_max,
+                    offenders
+                );
+            }
         }
 
         largest_read_size
@@ -1008,7 +1022,7 @@ impl DeltaLayerInner {
                             blob_meta.key,
                             PageReconstructError::Other(anyhow!(
                                 "Failed to read blobs from virtual file {}: {}",
-                                self.file.path,
+                                self.file.path(),
                                 kind
                             )),
                         );
@@ -1034,7 +1048,7 @@ impl DeltaLayerInner {
                             meta.meta.key,
                             PageReconstructError::Other(anyhow!(e).context(format!(
                                 "Failed to decompress blob from virtual file {}",
-                                self.file.path,
+                                self.file.path(),
                             ))),
                         );
 
@@ -1052,7 +1066,7 @@ impl DeltaLayerInner {
                             meta.meta.key,
                             PageReconstructError::Other(anyhow!(e).context(format!(
                                 "Failed to deserialize blob from virtual file {}",
-                                self.file.path,
+                                self.file.path(),
                             ))),
                         );
 
@@ -1133,7 +1147,7 @@ impl DeltaLayerInner {
         ctx: &RequestContext,
     ) -> anyhow::Result<usize> {
         use crate::tenant::vectored_blob_io::{
-            BlobMeta, VectoredReadBuilder, VectoredReadExtended,
+            BlobMeta, ChunkedVectoredReadBuilder, VectoredReadExtended,
         };
         use futures::stream::TryStreamExt;
 
@@ -1183,8 +1197,7 @@ impl DeltaLayerInner {
 
         let mut prev: Option<(Key, Lsn, BlobRef)> = None;
 
-        let mut read_builder: Option<VectoredReadBuilder> = None;
-        let read_mode = VectoredReadCoalesceMode::get();
+        let mut read_builder: Option<ChunkedVectoredReadBuilder> = None;
 
         let max_read_size = self
             .max_vectored_read_bytes
@@ -1228,12 +1241,11 @@ impl DeltaLayerInner {
                 {
                     None
                 } else {
-                    read_builder.replace(VectoredReadBuilder::new(
+                    read_builder.replace(ChunkedVectoredReadBuilder::new(
                         offsets.start.pos(),
                         offsets.end.pos(),
                         meta,
                         max_read_size,
-                        read_mode,
                     ))
                 }
             } else {
@@ -1426,17 +1438,14 @@ impl DeltaLayerInner {
         offset
     }
 
-    pub(crate) fn iter<'a>(
-        &'a self,
-        ctx: &'a RequestContext,
-    ) -> DeltaLayerIterator<'a, DeltaLayerDiskBtreeIter<'a>> {
+    pub(crate) fn iter<'a>(&'a self, ctx: &'a RequestContext) -> DeltaLayerIterator<'a> {
         let block_reader = FileBlockReader::new(&self.file, self.file_id);
         let tree_reader =
             DiskBtreeReader::new(self.index_start_blk, self.index_root_blk, block_reader);
         DeltaLayerIterator {
             delta_layer: self,
             ctx,
-            index_iter: tree_reader.iter(&[0; DELTA_KEY_SIZE], ctx).map_delta(),
+            index_iter: tree_reader.iter(&[0; DELTA_KEY_SIZE], ctx),
             key_values_batch: std::collections::VecDeque::new(),
             is_end: false,
             planner: StreamingVectoredReadPlanner::new(
@@ -1511,19 +1520,16 @@ impl<'a> pageserver_compaction::interface::CompactionDeltaEntry<'a, Key> for Del
     }
 }
 
-pub struct DeltaLayerIterator<'a, I = DeltaLayerDiskBtreeIter<'a>> {
+pub struct DeltaLayerIterator<'a> {
     delta_layer: &'a DeltaLayerInner,
     ctx: &'a RequestContext,
     planner: StreamingVectoredReadPlanner,
-    index_iter: I,
+    index_iter: DiskBtreeIterator<'a>,
     key_values_batch: VecDeque<(Key, Lsn, Value)>,
     is_end: bool,
 }
 
-impl<'a, I> DeltaLayerIterator<'a, I>
-where
-    I: DiskBtreeIterator<Item = (Key, Lsn, u64)>,
-{
+impl<'a> DeltaLayerIterator<'a> {
     pub(crate) fn layer_dbg_info(&self) -> String {
         self.delta_layer.layer_dbg_info()
     }
@@ -1535,7 +1541,11 @@ where
 
         let plan = loop {
             if let Some(res) = self.index_iter.next().await {
-                let (key, lsn, offset) = res?;
+                let (raw_key, value) = res?;
+                let key = Key::from_slice(&raw_key[..KEY_SIZE]);
+                let lsn = DeltaKey::extract_lsn_from_buf(&raw_key);
+                let blob_ref = BlobRef(value);
+                let offset = blob_ref.pos();
                 if let Some(batch_plan) = self.planner.handle(key, lsn, offset) {
                     break batch_plan;
                 }
@@ -1581,51 +1591,6 @@ where
                 .expect("should not be empty"),
         ))
     }
-
-    pub async fn filter_key<P>(
-        self,
-        predicate: P,
-    ) -> DeltaLayerIterator<'a, impl DiskBtreeIterator<Item = (Key, Lsn, u64)>>
-    where
-        P: FnMut(&(Key, Lsn, u64)) -> bool,
-    {
-        DeltaLayerIterator {
-            delta_layer: self.delta_layer,
-            ctx: self.ctx,
-            planner: self.planner,
-            index_iter: self.index_iter.filter(predicate),
-            key_values_batch: self.key_values_batch,
-            is_end: self.is_end,
-        }
-    }
-}
-
-pub struct DeltaLayerDiskBtreeIter<'a> {
-    iter: DiskBtreeIter<'a>,
-}
-
-impl<'a> DiskBtreeIter<'a> {
-    pub(crate) fn map_delta(self) -> DeltaLayerDiskBtreeIter<'a> {
-        DeltaLayerDiskBtreeIter { iter: self }
-    }
-}
-
-impl<'a> DiskBtreeIterator for DeltaLayerDiskBtreeIter<'a> {
-    type Item = (Key, Lsn, u64);
-
-    async fn next(
-        &mut self,
-    ) -> Option<std::result::Result<Self::Item, crate::tenant::disk_btree::DiskBtreeError>> {
-        self.iter.next().await.map(|res| {
-            res.map(|(raw_key, value)| {
-                let key = Key::from_slice(&raw_key[..KEY_SIZE]);
-                let lsn = DeltaKey::extract_lsn_from_buf(&raw_key);
-                let blob_ref = BlobRef(value);
-                let offset = blob_ref.pos();
-                (key, lsn, offset)
-            })
-        })
-    }
 }
 
 #[cfg(test)]

pageserver/src/tenant/storage_layer/image_layer.rs

@@ -32,7 +32,7 @@ use crate::repository::{Key, Value, KEY_SIZE};
 use crate::tenant::blob_io::BlobWriter;
 use crate::tenant::block_io::{BlockBuf, FileBlockReader};
 use crate::tenant::disk_btree::{
-    DiskBtreeBuilder, DiskBtreeIter, DiskBtreeIterator, DiskBtreeReader, VisitDirection,
+    DiskBtreeBuilder, DiskBtreeIterator, DiskBtreeReader, VisitDirection,
 };
 use crate::tenant::timeline::GetVectoredError;
 use crate::tenant::vectored_blob_io::{
@@ -41,7 +41,7 @@ use crate::tenant::vectored_blob_io::{
 };
 use crate::tenant::PageReconstructError;
 use crate::virtual_file::owned_buffers_io::io_buf_ext::IoBufExt;
-use crate::virtual_file::{self, VirtualFile};
+use crate::virtual_file::{self, MaybeFatalIo, VirtualFile};
 use crate::{IMAGE_FILE_MAGIC, STORAGE_FORMAT_VERSION, TEMP_FILE_SUFFIX};
 use anyhow::{anyhow, bail, ensure, Context, Result};
 use bytes::{Bytes, BytesMut};
@@ -49,6 +49,7 @@ use camino::{Utf8Path, Utf8PathBuf};
 use hex;
 use itertools::Itertools;
 use pageserver_api::config::MaxVectoredReadBytes;
+use pageserver_api::key::DBDIR_KEY;
 use pageserver_api::keyspace::KeySpace;
 use pageserver_api::shard::{ShardIdentity, TenantShardId};
 use rand::{distributions::Alphanumeric, Rng};
@@ -388,7 +389,7 @@ impl ImageLayerInner {
         max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
         ctx: &RequestContext,
     ) -> anyhow::Result<Self> {
-        let file = VirtualFile::open(path, ctx)
+        let file = VirtualFile::open_v2(path, ctx)
             .await
             .context("open layer file")?;
         let file_id = page_cache::next_file_id();
@@ -587,14 +588,25 @@ impl ImageLayerInner {
                     .blobs_at
                     .as_slice()
                     .iter()
-                    .map(|(_, blob_meta)| format!("{}@{}", blob_meta.key, blob_meta.lsn))
+                    .filter_map(|(_, blob_meta)| {
+                        if blob_meta.key.is_rel_dir_key() || blob_meta.key == DBDIR_KEY {
+                            // The size of values for these keys is unbounded and can
+                            // grow very large in pathological cases.
+                            None
+                        } else {
+                            Some(format!("{}@{}", blob_meta.key, blob_meta.lsn))
+                        }
+                    })
                     .join(", ");
-                tracing::warn!(
-                    "Oversized vectored read ({} > {}) for keys {}",
-                    buf_size,
-                    max_vectored_read_bytes,
-                    offenders
-                );
+
+                if !offenders.is_empty() {
+                    tracing::warn!(
+                        "Oversized vectored read ({} > {}) for keys {}",
+                        buf_size,
+                        max_vectored_read_bytes,
+                        offenders
+                    );
+                }
             }
 
             let buf = BytesMut::with_capacity(buf_size);
@@ -614,7 +626,7 @@ impl ImageLayerInner {
                                     meta.meta.key,
                                     PageReconstructError::Other(anyhow!(e).context(format!(
                                         "Failed to decompress blob from virtual file {}",
-                                        self.file.path,
+                                        self.file.path(),
                                     ))),
                                 );
 
@@ -635,7 +647,7 @@ impl ImageLayerInner {
                             blob_meta.key,
                             PageReconstructError::from(anyhow!(
                                 "Failed to read blobs from virtual file {}: {}",
-                                self.file.path,
+                                self.file.path(),
                                 kind
                             )),
                         );
@@ -652,7 +664,7 @@ impl ImageLayerInner {
         ImageLayerIterator {
             image_layer: self,
             ctx,
-            index_iter: tree_reader.iter(&[0; KEY_SIZE], ctx).map_image(self.lsn),
+            index_iter: tree_reader.iter(&[0; KEY_SIZE], ctx),
             key_values_batch: VecDeque::new(),
             is_end: false,
             planner: StreamingVectoredReadPlanner::new(
@@ -889,7 +901,9 @@ impl ImageLayerWriterInner {
         // set inner.file here. The first read will have to re-open it.
 
         // fsync the file
-        file.sync_all().await?;
+        file.sync_all()
+            .await
+            .maybe_fatal_err("image_layer sync_all")?;
 
         trace!("created image layer {}", self.path);
 
@@ -999,37 +1013,11 @@ pub struct ImageLayerIterator<'a> {
     image_layer: &'a ImageLayerInner,
     ctx: &'a RequestContext,
     planner: StreamingVectoredReadPlanner,
-    index_iter: ImageLayerDiskBtreeIter<'a>,
+    index_iter: DiskBtreeIterator<'a>,
     key_values_batch: VecDeque<(Key, Lsn, Value)>,
     is_end: bool,
 }
 
-pub struct ImageLayerDiskBtreeIter<'a> {
-    iter: DiskBtreeIter<'a>,
-    lsn: Lsn,
-}
-
-impl<'a> DiskBtreeIter<'a> {
-    pub(crate) fn map_image(self, lsn: Lsn) -> ImageLayerDiskBtreeIter<'a> {
-        ImageLayerDiskBtreeIter { iter: self, lsn }
-    }
-}
-
-impl<'a> DiskBtreeIterator for ImageLayerDiskBtreeIter<'a> {
-    type Item = (Key, Lsn, u64);
-
-    async fn next(
-        &mut self,
-    ) -> Option<std::result::Result<Self::Item, crate::tenant::disk_btree::DiskBtreeError>> {
-        self.iter.next().await.map(|res| {
-            res.map(|(raw_key, offset)| {
-                let key = Key::from_slice(&raw_key[..KEY_SIZE]);
-                (key, self.lsn, offset)
-            })
-        })
-    }
-}
-
 impl<'a> ImageLayerIterator<'a> {
     pub(crate) fn layer_dbg_info(&self) -> String {
         self.image_layer.layer_dbg_info()
@@ -1042,8 +1030,12 @@ impl<'a> ImageLayerIterator<'a> {
 
         let plan = loop {
             if let Some(res) = self.index_iter.next().await {
-                let (key, lsn, offset) = res?;
-                if let Some(batch_plan) = self.planner.handle(key, lsn, offset) {
+                let (raw_key, offset) = res?;
+                if let Some(batch_plan) = self.planner.handle(
+                    Key::from_slice(&raw_key[..KEY_SIZE]),
+                    self.image_layer.lsn,
+                    offset,
+                ) {
                     break batch_plan;
                 }
             } else {

pageserver/src/tenant/storage_layer/layer.rs

@@ -442,11 +442,13 @@ impl Layer {
             // Visibility was modified to Visible: maybe log about this
             match ctx.task_kind() {
                 TaskKind::CalculateSyntheticSize
+                | TaskKind::OndemandLogicalSizeCalculation
                 | TaskKind::GarbageCollector
                 | TaskKind::MgmtRequest => {
                     // This situation is expected in code paths do binary searches of the LSN space to resolve
                     // an LSN to a timestamp, which happens during GC, during GC cutoff calculations in synthetic size,
-                    // and on-demand for certain HTTP API requests.
+                    // and on-demand for certain HTTP API requests. On-demand logical size calculation is also included
+                    // because it is run as a sub-task of synthetic size.
                 }
                 _ => {
                     // In all other contexts, it is unusual to do I/O involving layers which are not visible at
@@ -456,8 +458,8 @@ impl Layer {
                     // This case is legal in brief time windows: for example an in-flight getpage request can hold on to a layer object
                     // which was covered by a concurrent compaction.
                     tracing::info!(
-                        "Layer {} became visible as a result of access",
-                        self.0.desc.key()
+                        layer=%self,
+                        "became visible as a result of access",
                     );
                 }
             }
@@ -686,7 +688,9 @@ impl Drop for LayerInner {
             // and we could be delaying shutdown for nothing.
         }
 
-        if let Some(timeline) = self.timeline.upgrade() {
+        let timeline = self.timeline.upgrade();
+
+        if let Some(timeline) = timeline.as_ref() {
             // Only need to decrement metrics if the timeline still exists: otherwise
             // it will have already de-registered these metrics via TimelineMetrics::shutdown
             if self.desc.is_delta() {
@@ -717,7 +721,6 @@ impl Drop for LayerInner {
         let path = std::mem::take(&mut self.path);
         let file_name = self.layer_desc().layer_name();
         let file_size = self.layer_desc().file_size;
-        let timeline = self.timeline.clone();
         let meta = self.metadata();
         let status = self.status.take();
 
@@ -727,7 +730,7 @@ impl Drop for LayerInner {
             // carry this until we are finished for [`Layer::wait_drop`] support
             let _status = status;
 
-            let Some(timeline) = timeline.upgrade() else {
+            let Some(timeline) = timeline else {
                 // no need to nag that timeline is gone: under normal situation on
                 // task_mgr::remove_tenant_from_memory the timeline is gone before we get dropped.
                 LAYER_IMPL_METRICS.inc_deletes_failed(DeleteFailed::TimelineGone);

pageserver/src/tenant/tasks.rs

@@ -330,7 +330,6 @@ async fn gc_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
             RequestContext::todo_child(TaskKind::GarbageCollector, DownloadBehavior::Download);
 
         let mut first = true;
-        tenant.gc_block.set_lsn_lease_deadline(tenant.get_lsn_lease_length());
         loop {
             tokio::select! {
                 _ = cancel.cancelled() => {

pageserver/src/tenant/timeline.rs

@@ -48,7 +48,6 @@ use utils::{
     sync::gate::{Gate, GateGuard},
 };
 
-use std::pin::pin;
 use std::sync::atomic::Ordering as AtomicOrdering;
 use std::sync::{Arc, Mutex, RwLock, Weak};
 use std::time::{Duration, Instant, SystemTime};
@@ -62,14 +61,17 @@ use std::{
     collections::btree_map::Entry,
     ops::{Deref, Range},
 };
+use std::{pin::pin, sync::OnceLock};
 
 use crate::{
     aux_file::AuxFileSizeEstimator,
     tenant::{
+        config::AttachmentMode,
         layer_map::{LayerMap, SearchResult},
         metadata::TimelineMetadata,
         storage_layer::{inmemory_layer::IndexEntry, PersistentLayerDesc},
     },
+    walingest::WalLagCooldown,
     walredo,
 };
 use crate::{
@@ -428,6 +430,8 @@ pub struct Timeline {
     pub(crate) l0_flush_global_state: L0FlushGlobalState,
 
     pub(crate) handles: handle::PerTimelineState<crate::page_service::TenantManagerTypes>,
+
+    pub(crate) attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
 }
 
 pub struct WalReceiverInfo {
@@ -736,6 +740,7 @@ pub enum GetLogicalSizePriority {
 pub(crate) enum CompactFlags {
     ForceRepartition,
     ForceImageLayerCreation,
+    ForceL0Compaction,
     EnhancedGcBottomMostCompaction,
     DryRun,
 }
@@ -1324,16 +1329,38 @@ impl Timeline {
         Ok(())
     }
 
-    /// Obtains a temporary lease blocking garbage collection for the given LSN.
-    ///
-    /// This function will error if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is also
-    /// no existing lease to renew. If there is an existing lease in the map, the lease will be renewed only if
-    /// the request extends the lease. The returned lease is therefore the maximum between the existing lease and
-    /// the requesting lease.
-    pub(crate) fn make_lsn_lease(
+    /// Initializes an LSN lease. The function will return an error if the requested LSN is less than the `latest_gc_cutoff_lsn`.
+    pub(crate) fn init_lsn_lease(
         &self,
         lsn: Lsn,
         length: Duration,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<LsnLease> {
+        self.make_lsn_lease(lsn, length, true, ctx)
+    }
+
+    /// Renews a lease at a particular LSN. The requested LSN is not validated against the `latest_gc_cutoff_lsn` when we are in the grace period.
+    pub(crate) fn renew_lsn_lease(
+        &self,
+        lsn: Lsn,
+        length: Duration,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<LsnLease> {
+        self.make_lsn_lease(lsn, length, false, ctx)
+    }
+
+    /// Obtains a temporary lease blocking garbage collection for the given LSN.
+    ///
+    /// If we are in `AttachedSingle` mode and is not blocked by the lsn lease deadline, this function will error
+    /// if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is no existing request present.
+    ///
+    /// If there is an existing lease in the map, the lease will be renewed only if the request extends the lease.
+    /// The returned lease is therefore the maximum between the existing lease and the requesting lease.
+    fn make_lsn_lease(
+        &self,
+        lsn: Lsn,
+        length: Duration,
+        init: bool,
         _ctx: &RequestContext,
     ) -> anyhow::Result<LsnLease> {
         let lease = {
@@ -1347,8 +1374,8 @@ impl Timeline {
 
             let entry = gc_info.leases.entry(lsn);
 
-            let lease = {
-                if let Entry::Occupied(mut occupied) = entry {
+            match entry {
+                Entry::Occupied(mut occupied) => {
                     let existing_lease = occupied.get_mut();
                     if valid_until > existing_lease.valid_until {
                         existing_lease.valid_until = valid_until;
@@ -1360,20 +1387,28 @@ impl Timeline {
                     }
 
                     existing_lease.clone()
-                } else {
-                    // Reject already GC-ed LSN (lsn < latest_gc_cutoff)
-                    let latest_gc_cutoff_lsn = self.get_latest_gc_cutoff_lsn();
-                    if lsn < *latest_gc_cutoff_lsn {
-                        bail!("tried to request a page version that was garbage collected. requested at {} gc cutoff {}", lsn, *latest_gc_cutoff_lsn);
+                }
+                Entry::Vacant(vacant) => {
+                    // Reject already GC-ed LSN (lsn < latest_gc_cutoff) if we are in AttachedSingle and
+                    // not blocked by the lsn lease deadline.
+                    let validate = {
+                        let conf = self.tenant_conf.load();
+                        conf.location.attach_mode == AttachmentMode::Single
+                            && !conf.is_gc_blocked_by_lsn_lease_deadline()
+                    };
+
+                    if init || validate {
+                        let latest_gc_cutoff_lsn = self.get_latest_gc_cutoff_lsn();
+                        if lsn < *latest_gc_cutoff_lsn {
+                            bail!("tried to request a page version that was garbage collected. requested at {} gc cutoff {}", lsn, *latest_gc_cutoff_lsn);
+                        }
                     }
 
                     let dt: DateTime<Utc> = valid_until.into();
                     info!("lease created, valid until {}", dt);
-                    entry.or_insert(LsnLease { valid_until }).clone()
+                    vacant.insert(LsnLease { valid_until }).clone()
                 }
-            };
-
-            lease
+            }
         };
 
         Ok(lease)
@@ -1950,8 +1985,6 @@ impl Timeline {
             .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
     }
 
-    // TODO(yuchen): remove unused flag after implementing https://github.com/neondatabase/neon/issues/8072
-    #[allow(unused)]
     pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
         let tenant_conf = self.tenant_conf.load();
         tenant_conf
@@ -2101,6 +2134,7 @@ impl Timeline {
         pg_version: u32,
         state: TimelineState,
         aux_file_policy: Option<AuxFilePolicy>,
+        attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
         cancel: CancellationToken,
     ) -> Arc<Self> {
         let disk_consistent_lsn = metadata.disk_consistent_lsn();
@@ -2242,6 +2276,8 @@ impl Timeline {
                 l0_flush_global_state: resources.l0_flush_global_state,
 
                 handles: Default::default(),
+
+                attach_wal_lag_cooldown,
             };
 
             if aux_file_policy == Some(AuxFilePolicy::V1) {

pageserver/src/tenant/timeline/analysis.rs

@@ -11,6 +11,7 @@ pub(crate) struct RangeAnalysis {
     has_image: bool,
     num_of_deltas_above_image: usize,
     total_num_of_deltas: usize,
+    num_of_l0: usize,
 }
 
 impl Timeline {
@@ -20,8 +21,10 @@ impl Timeline {
         let mut delta_ranges = Vec::new();
         let mut image_ranges = Vec::new();
 
+        let num_of_l0;
         let all_layer_files = {
             let guard = self.layers.read().await;
+            num_of_l0 = guard.layer_map().unwrap().level0_deltas().len();
             guard.all_persistent_layers()
         };
         let lsn = self.get_last_record_lsn();
@@ -82,6 +85,7 @@ impl Timeline {
                 has_image: image_layer.is_some(),
                 num_of_deltas_above_image: maybe_delta_layers.len(),
                 total_num_of_deltas: pitr_delta_layers.len(),
+                num_of_l0,
             });
         }
 

pageserver/src/tenant/timeline/compaction.rs

@@ -353,7 +353,13 @@ impl Timeline {
 
                 // 2. Compact
                 let timer = self.metrics.compact_time_histo.start_timer();
-                let fully_compacted = self.compact_level0(target_file_size, ctx).await?;
+                let fully_compacted = self
+                    .compact_level0(
+                        target_file_size,
+                        flags.contains(CompactFlags::ForceL0Compaction),
+                        ctx,
+                    )
+                    .await?;
                 timer.stop_and_record();
 
                 let mut partitioning = dense_partitioning;
@@ -658,6 +664,7 @@ impl Timeline {
     async fn compact_level0(
         self: &Arc<Self>,
         target_file_size: u64,
+        force_compaction_ignore_threshold: bool,
         ctx: &RequestContext,
     ) -> Result<bool, CompactionError> {
         let CompactLevel0Phase1Result {
@@ -679,9 +686,15 @@ impl Timeline {
             let now = tokio::time::Instant::now();
             stats.read_lock_acquisition_micros =
                 DurationRecorder::Recorded(RecordedDuration(now - begin), now);
-            self.compact_level0_phase1(phase1_layers_locked, stats, target_file_size, &ctx)
-                .instrument(phase1_span)
-                .await?
+            self.compact_level0_phase1(
+                phase1_layers_locked,
+                stats,
+                target_file_size,
+                force_compaction_ignore_threshold,
+                &ctx,
+            )
+            .instrument(phase1_span)
+            .await?
         };
 
         if new_layers.is_empty() && deltas_to_compact.is_empty() {
@@ -700,6 +713,7 @@ impl Timeline {
         guard: tokio::sync::RwLockReadGuard<'a, LayerManager>,
         mut stats: CompactLevel0Phase1StatsBuilder,
         target_file_size: u64,
+        force_compaction_ignore_threshold: bool,
         ctx: &RequestContext,
     ) -> Result<CompactLevel0Phase1Result, CompactionError> {
         stats.read_lock_held_spawn_blocking_startup_micros =
@@ -711,11 +725,26 @@ impl Timeline {
         // Only compact if enough layers have accumulated.
         let threshold = self.get_compaction_threshold();
         if level0_deltas.is_empty() || level0_deltas.len() < threshold {
-            debug!(
-                level0_deltas = level0_deltas.len(),
-                threshold, "too few deltas to compact"
-            );
-            return Ok(CompactLevel0Phase1Result::default());
+            if force_compaction_ignore_threshold {
+                if !level0_deltas.is_empty() {
+                    info!(
+                        level0_deltas = level0_deltas.len(),
+                        threshold, "too few deltas to compact, but forcing compaction"
+                    );
+                } else {
+                    info!(
+                        level0_deltas = level0_deltas.len(),
+                        threshold, "too few deltas to compact, cannot force compaction"
+                    );
+                    return Ok(CompactLevel0Phase1Result::default());
+                }
+            } else {
+                debug!(
+                    level0_deltas = level0_deltas.len(),
+                    threshold, "too few deltas to compact"
+                );
+                return Ok(CompactLevel0Phase1Result::default());
+            }
         }
 
         let mut level0_deltas = level0_deltas

pageserver/src/tenant/vectored_blob_io.rs

@@ -185,171 +185,7 @@ pub(crate) enum VectoredReadExtended {
     No,
 }
 
-#[derive(Copy, Clone, Debug, PartialEq, Eq)]
-pub enum VectoredReadCoalesceMode {
-    /// Only coalesce exactly adjacent reads.
-    AdjacentOnly,
-    /// In addition to adjacent reads, also consider reads whose corresponding
-    /// `end` and `start` offsets reside at the same chunk.
-    Chunked(usize),
-}
-
-impl VectoredReadCoalesceMode {
-    /// [`AdjacentVectoredReadBuilder`] is used if alignment requirement is 0,
-    /// whereas [`ChunkedVectoredReadBuilder`] is used for alignment requirement 1 and higher.
-    pub(crate) fn get() -> Self {
-        let align = virtual_file::get_io_buffer_alignment_raw();
-        if align == 0 {
-            VectoredReadCoalesceMode::AdjacentOnly
-        } else {
-            VectoredReadCoalesceMode::Chunked(align)
-        }
-    }
-}
-
-pub(crate) enum VectoredReadBuilder {
-    Adjacent(AdjacentVectoredReadBuilder),
-    Chunked(ChunkedVectoredReadBuilder),
-}
-
-impl VectoredReadBuilder {
-    fn new_impl(
-        start_offset: u64,
-        end_offset: u64,
-        meta: BlobMeta,
-        max_read_size: Option<usize>,
-        mode: VectoredReadCoalesceMode,
-    ) -> Self {
-        match mode {
-            VectoredReadCoalesceMode::AdjacentOnly => Self::Adjacent(
-                AdjacentVectoredReadBuilder::new(start_offset, end_offset, meta, max_read_size),
-            ),
-            VectoredReadCoalesceMode::Chunked(chunk_size) => {
-                Self::Chunked(ChunkedVectoredReadBuilder::new(
-                    start_offset,
-                    end_offset,
-                    meta,
-                    max_read_size,
-                    chunk_size,
-                ))
-            }
-        }
-    }
-
-    pub(crate) fn new(
-        start_offset: u64,
-        end_offset: u64,
-        meta: BlobMeta,
-        max_read_size: usize,
-        mode: VectoredReadCoalesceMode,
-    ) -> Self {
-        Self::new_impl(start_offset, end_offset, meta, Some(max_read_size), mode)
-    }
-
-    pub(crate) fn new_streaming(
-        start_offset: u64,
-        end_offset: u64,
-        meta: BlobMeta,
-        mode: VectoredReadCoalesceMode,
-    ) -> Self {
-        Self::new_impl(start_offset, end_offset, meta, None, mode)
-    }
-
-    pub(crate) fn extend(&mut self, start: u64, end: u64, meta: BlobMeta) -> VectoredReadExtended {
-        match self {
-            VectoredReadBuilder::Adjacent(builder) => builder.extend(start, end, meta),
-            VectoredReadBuilder::Chunked(builder) => builder.extend(start, end, meta),
-        }
-    }
-
-    pub(crate) fn build(self) -> VectoredRead {
-        match self {
-            VectoredReadBuilder::Adjacent(builder) => builder.build(),
-            VectoredReadBuilder::Chunked(builder) => builder.build(),
-        }
-    }
-
-    pub(crate) fn size(&self) -> usize {
-        match self {
-            VectoredReadBuilder::Adjacent(builder) => builder.size(),
-            VectoredReadBuilder::Chunked(builder) => builder.size(),
-        }
-    }
-}
-
-pub(crate) struct AdjacentVectoredReadBuilder {
-    /// Start offset of the read.
-    start: u64,
-    // End offset of the read.
-    end: u64,
-    /// Start offset and metadata for each blob in this read
-    blobs_at: VecMap<u64, BlobMeta>,
-    max_read_size: Option<usize>,
-}
-
-impl AdjacentVectoredReadBuilder {
-    /// Start building a new vectored read.
-    ///
-    /// Note that by design, this does not check against reading more than `max_read_size` to
-    /// support reading larger blobs than the configuration value. The builder will be single use
-    /// however after that.
-    pub(crate) fn new(
-        start_offset: u64,
-        end_offset: u64,
-        meta: BlobMeta,
-        max_read_size: Option<usize>,
-    ) -> Self {
-        let mut blobs_at = VecMap::default();
-        blobs_at
-            .append(start_offset, meta)
-            .expect("First insertion always succeeds");
-
-        Self {
-            start: start_offset,
-            end: end_offset,
-            blobs_at,
-            max_read_size,
-        }
-    }
-    /// Attempt to extend the current read with a new blob if the start
-    /// offset matches with the current end of the vectored read
-    /// and the resuting size is below the max read size
-    pub(crate) fn extend(&mut self, start: u64, end: u64, meta: BlobMeta) -> VectoredReadExtended {
-        tracing::trace!(start, end, "trying to extend");
-        let size = (end - start) as usize;
-        let not_limited_by_max_read_size = {
-            if let Some(max_read_size) = self.max_read_size {
-                self.size() + size <= max_read_size
-            } else {
-                true
-            }
-        };
-
-        if self.end == start && not_limited_by_max_read_size {
-            self.end = end;
-            self.blobs_at
-                .append(start, meta)
-                .expect("LSNs are ordered within vectored reads");
-
-            return VectoredReadExtended::Yes;
-        }
-
-        VectoredReadExtended::No
-    }
-
-    pub(crate) fn size(&self) -> usize {
-        (self.end - self.start) as usize
-    }
-
-    pub(crate) fn build(self) -> VectoredRead {
-        VectoredRead {
-            start: self.start,
-            end: self.end,
-            blobs_at: self.blobs_at,
-        }
-    }
-}
-
+/// A vectored read builder that tries to coalesce all reads that fits in a chunk.
 pub(crate) struct ChunkedVectoredReadBuilder {
     /// Start block number
     start_blk_no: usize,
@@ -358,8 +194,6 @@ pub(crate) struct ChunkedVectoredReadBuilder {
     /// Start offset and metadata for each blob in this read
     blobs_at: VecMap<u64, BlobMeta>,
     max_read_size: Option<usize>,
-    /// Chunk size reads are coalesced into.
-    chunk_size: usize,
 }
 
 /// Computes x / d rounded up.
@@ -368,45 +202,57 @@ fn div_round_up(x: usize, d: usize) -> usize {
 }
 
 impl ChunkedVectoredReadBuilder {
+    const CHUNK_SIZE: usize = virtual_file::get_io_buffer_alignment();
     /// Start building a new vectored read.
     ///
     /// Note that by design, this does not check against reading more than `max_read_size` to
     /// support reading larger blobs than the configuration value. The builder will be single use
     /// however after that.
-    pub(crate) fn new(
+    fn new_impl(
         start_offset: u64,
         end_offset: u64,
         meta: BlobMeta,
         max_read_size: Option<usize>,
-        chunk_size: usize,
     ) -> Self {
         let mut blobs_at = VecMap::default();
         blobs_at
             .append(start_offset, meta)
             .expect("First insertion always succeeds");
 
-        let start_blk_no = start_offset as usize / chunk_size;
-        let end_blk_no = div_round_up(end_offset as usize, chunk_size);
+        let start_blk_no = start_offset as usize / Self::CHUNK_SIZE;
+        let end_blk_no = div_round_up(end_offset as usize, Self::CHUNK_SIZE);
         Self {
             start_blk_no,
             end_blk_no,
             blobs_at,
             max_read_size,
-            chunk_size,
         }
     }
 
+    pub(crate) fn new(
+        start_offset: u64,
+        end_offset: u64,
+        meta: BlobMeta,
+        max_read_size: usize,
+    ) -> Self {
+        Self::new_impl(start_offset, end_offset, meta, Some(max_read_size))
+    }
+
+    pub(crate) fn new_streaming(start_offset: u64, end_offset: u64, meta: BlobMeta) -> Self {
+        Self::new_impl(start_offset, end_offset, meta, None)
+    }
+
     /// Attempts to extend the current read with a new blob if the new blob resides in the same or the immediate next chunk.
     ///
     /// The resulting size also must be below the max read size.
     pub(crate) fn extend(&mut self, start: u64, end: u64, meta: BlobMeta) -> VectoredReadExtended {
         tracing::trace!(start, end, "trying to extend");
-        let start_blk_no = start as usize / self.chunk_size;
-        let end_blk_no = div_round_up(end as usize, self.chunk_size);
+        let start_blk_no = start as usize / Self::CHUNK_SIZE;
+        let end_blk_no = div_round_up(end as usize, Self::CHUNK_SIZE);
 
         let not_limited_by_max_read_size = {
             if let Some(max_read_size) = self.max_read_size {
-                let coalesced_size = (end_blk_no - self.start_blk_no) * self.chunk_size;
+                let coalesced_size = (end_blk_no - self.start_blk_no) * Self::CHUNK_SIZE;
                 coalesced_size <= max_read_size
             } else {
                 true
@@ -437,12 +283,12 @@ impl ChunkedVectoredReadBuilder {
     }
 
     pub(crate) fn size(&self) -> usize {
-        (self.end_blk_no - self.start_blk_no) * self.chunk_size
+        (self.end_blk_no - self.start_blk_no) * Self::CHUNK_SIZE
     }
 
     pub(crate) fn build(self) -> VectoredRead {
-        let start = (self.start_blk_no * self.chunk_size) as u64;
-        let end = (self.end_blk_no * self.chunk_size) as u64;
+        let start = (self.start_blk_no * Self::CHUNK_SIZE) as u64;
+        let end = (self.end_blk_no * Self::CHUNK_SIZE) as u64;
         VectoredRead {
             start,
             end,
@@ -473,18 +319,14 @@ pub struct VectoredReadPlanner {
     prev: Option<(Key, Lsn, u64, BlobFlag)>,
 
     max_read_size: usize,
-
-    mode: VectoredReadCoalesceMode,
 }
 
 impl VectoredReadPlanner {
     pub fn new(max_read_size: usize) -> Self {
-        let mode = VectoredReadCoalesceMode::get();
         Self {
             blobs: BTreeMap::new(),
             prev: None,
             max_read_size,
-            mode,
         }
     }
 
@@ -545,7 +387,7 @@ impl VectoredReadPlanner {
     }
 
     pub fn finish(self) -> Vec<VectoredRead> {
-        let mut current_read_builder: Option<VectoredReadBuilder> = None;
+        let mut current_read_builder: Option<ChunkedVectoredReadBuilder> = None;
         let mut reads = Vec::new();
 
         for (key, blobs_for_key) in self.blobs {
@@ -558,12 +400,11 @@ impl VectoredReadPlanner {
                 };
 
                 if extended == VectoredReadExtended::No {
-                    let next_read_builder = VectoredReadBuilder::new(
+                    let next_read_builder = ChunkedVectoredReadBuilder::new(
                         start_offset,
                         end_offset,
                         BlobMeta { key, lsn },
                         self.max_read_size,
-                        self.mode,
                     );
 
                     let prev_read_builder = current_read_builder.replace(next_read_builder);
@@ -617,13 +458,13 @@ impl<'a> VectoredBlobReader<'a> {
         );
 
         if cfg!(debug_assertions) {
-            let align = virtual_file::get_io_buffer_alignment() as u64;
+            const ALIGN: u64 = virtual_file::get_io_buffer_alignment() as u64;
             debug_assert_eq!(
-                read.start % align,
+                read.start % ALIGN,
                 0,
                 "Read start at {} does not satisfy the required io buffer alignment ({} bytes)",
                 read.start,
-                align
+                ALIGN
             );
         }
 
@@ -688,7 +529,7 @@ impl<'a> VectoredBlobReader<'a> {
 /// `handle` gets called and when the current key would just exceed the read_size and
 /// max_cnt constraints.
 pub struct StreamingVectoredReadPlanner {
-    read_builder: Option<VectoredReadBuilder>,
+    read_builder: Option<ChunkedVectoredReadBuilder>,
     // Arguments for previous blob passed into [`StreamingVectoredReadPlanner::handle`]
     prev: Option<(Key, Lsn, u64)>,
     /// Max read size per batch. This is not a strict limit. If there are [0, 100) and [100, 200), while the `max_read_size` is 150,
@@ -698,22 +539,18 @@ pub struct StreamingVectoredReadPlanner {
     max_cnt: usize,
     /// Size of the current batch
     cnt: usize,
-
-    mode: VectoredReadCoalesceMode,
 }
 
 impl StreamingVectoredReadPlanner {
     pub fn new(max_read_size: u64, max_cnt: usize) -> Self {
         assert!(max_cnt > 0);
         assert!(max_read_size > 0);
-        let mode = VectoredReadCoalesceMode::get();
         Self {
             read_builder: None,
             prev: None,
             max_cnt,
             max_read_size,
             cnt: 0,
-            mode,
         }
     }
 
@@ -762,11 +599,10 @@ impl StreamingVectoredReadPlanner {
             }
             None => {
                 self.read_builder = {
-                    Some(VectoredReadBuilder::new_streaming(
+                    Some(ChunkedVectoredReadBuilder::new_streaming(
                         start_offset,
                         end_offset,
                         BlobMeta { key, lsn },
-                        self.mode,
                     ))
                 };
             }
@@ -801,9 +637,9 @@ mod tests {
     use super::*;
 
     fn validate_read(read: &VectoredRead, offset_range: &[(Key, Lsn, u64, BlobFlag)]) {
-        let align = virtual_file::get_io_buffer_alignment() as u64;
-        assert_eq!(read.start % align, 0);
-        assert_eq!(read.start / align, offset_range.first().unwrap().2 / align);
+        const ALIGN: u64 = virtual_file::get_io_buffer_alignment() as u64;
+        assert_eq!(read.start % ALIGN, 0);
+        assert_eq!(read.start / ALIGN, offset_range.first().unwrap().2 / ALIGN);
 
         let expected_offsets_in_read: Vec<_> = offset_range.iter().map(|o| o.2).collect();
 
@@ -821,32 +657,27 @@ mod tests {
     fn planner_chunked_coalesce_all_test() {
         use crate::virtual_file;
 
-        let chunk_size = virtual_file::get_io_buffer_alignment() as u64;
+        const CHUNK_SIZE: u64 = virtual_file::get_io_buffer_alignment() as u64;
 
-        // The test explicitly does not check chunk size < 512
-        if chunk_size < 512 {
-            return;
-        }
-
-        let max_read_size = chunk_size as usize * 8;
+        let max_read_size = CHUNK_SIZE as usize * 8;
         let key = Key::MIN;
         let lsn = Lsn(0);
 
         let blob_descriptions = [
-            (key, lsn, chunk_size / 8, BlobFlag::None), // Read 1 BEGIN
-            (key, lsn, chunk_size / 4, BlobFlag::Ignore), // Gap
-            (key, lsn, chunk_size / 2, BlobFlag::None),
-            (key, lsn, chunk_size - 2, BlobFlag::Ignore), // Gap
-            (key, lsn, chunk_size, BlobFlag::None),
-            (key, lsn, chunk_size * 2 - 1, BlobFlag::None),
-            (key, lsn, chunk_size * 2 + 1, BlobFlag::Ignore), // Gap
-            (key, lsn, chunk_size * 3 + 1, BlobFlag::None),
-            (key, lsn, chunk_size * 5 + 1, BlobFlag::None),
-            (key, lsn, chunk_size * 6 + 1, BlobFlag::Ignore), // skipped chunk size, but not a chunk: should coalesce.
-            (key, lsn, chunk_size * 7 + 1, BlobFlag::None),
-            (key, lsn, chunk_size * 8, BlobFlag::None), // Read 2 BEGIN (b/c max_read_size)
-            (key, lsn, chunk_size * 9, BlobFlag::Ignore), // ==== skipped a chunk
-            (key, lsn, chunk_size * 10, BlobFlag::None), // Read 3 BEGIN (cannot coalesce)
+            (key, lsn, CHUNK_SIZE / 8, BlobFlag::None), // Read 1 BEGIN
+            (key, lsn, CHUNK_SIZE / 4, BlobFlag::Ignore), // Gap
+            (key, lsn, CHUNK_SIZE / 2, BlobFlag::None),
+            (key, lsn, CHUNK_SIZE - 2, BlobFlag::Ignore), // Gap
+            (key, lsn, CHUNK_SIZE, BlobFlag::None),
+            (key, lsn, CHUNK_SIZE * 2 - 1, BlobFlag::None),
+            (key, lsn, CHUNK_SIZE * 2 + 1, BlobFlag::Ignore), // Gap
+            (key, lsn, CHUNK_SIZE * 3 + 1, BlobFlag::None),
+            (key, lsn, CHUNK_SIZE * 5 + 1, BlobFlag::None),
+            (key, lsn, CHUNK_SIZE * 6 + 1, BlobFlag::Ignore), // skipped chunk size, but not a chunk: should coalesce.
+            (key, lsn, CHUNK_SIZE * 7 + 1, BlobFlag::None),
+            (key, lsn, CHUNK_SIZE * 8, BlobFlag::None), // Read 2 BEGIN (b/c max_read_size)
+            (key, lsn, CHUNK_SIZE * 9, BlobFlag::Ignore), // ==== skipped a chunk
+            (key, lsn, CHUNK_SIZE * 10, BlobFlag::None), // Read 3 BEGIN (cannot coalesce)
         ];
 
         let ranges = [
@@ -925,19 +756,19 @@ mod tests {
 
     #[test]
     fn planner_replacement_test() {
-        let chunk_size = virtual_file::get_io_buffer_alignment() as u64;
-        let max_read_size = 128 * chunk_size as usize;
+        const CHUNK_SIZE: u64 = virtual_file::get_io_buffer_alignment() as u64;
+        let max_read_size = 128 * CHUNK_SIZE as usize;
         let first_key = Key::MIN;
         let second_key = first_key.next();
         let lsn = Lsn(0);
 
         let blob_descriptions = vec![
             (first_key, lsn, 0, BlobFlag::None),          // First in read 1
-            (first_key, lsn, chunk_size, BlobFlag::None), // Last in read 1
-            (second_key, lsn, 2 * chunk_size, BlobFlag::ReplaceAll),
-            (second_key, lsn, 3 * chunk_size, BlobFlag::None),
-            (second_key, lsn, 4 * chunk_size, BlobFlag::ReplaceAll), // First in read 2
-            (second_key, lsn, 5 * chunk_size, BlobFlag::None),       // Last in read 2
+            (first_key, lsn, CHUNK_SIZE, BlobFlag::None), // Last in read 1
+            (second_key, lsn, 2 * CHUNK_SIZE, BlobFlag::ReplaceAll),
+            (second_key, lsn, 3 * CHUNK_SIZE, BlobFlag::None),
+            (second_key, lsn, 4 * CHUNK_SIZE, BlobFlag::ReplaceAll), // First in read 2
+            (second_key, lsn, 5 * CHUNK_SIZE, BlobFlag::None),       // Last in read 2
         ];
 
         let ranges = [&blob_descriptions[0..2], &blob_descriptions[4..]];
@@ -947,7 +778,7 @@ mod tests {
             planner.handle(key, lsn, offset, flag);
         }
 
-        planner.handle_range_end(6 * chunk_size);
+        planner.handle_range_end(6 * CHUNK_SIZE);
 
         let reads = planner.finish();
         assert_eq!(reads.len(), 2);
@@ -1092,7 +923,6 @@ mod tests {
         let reserved_bytes = blobs.iter().map(|bl| bl.len()).max().unwrap() * 2 + 16;
         let mut buf = BytesMut::with_capacity(reserved_bytes);
 
-        let mode = VectoredReadCoalesceMode::get();
         let vectored_blob_reader = VectoredBlobReader::new(&file);
         let meta = BlobMeta {
             key: Key::MIN,
@@ -1104,7 +934,7 @@ mod tests {
             if idx + 1 == offsets.len() {
                 continue;
             }
-            let read_builder = VectoredReadBuilder::new(*offset, *end, meta, 16 * 4096, mode);
+            let read_builder = ChunkedVectoredReadBuilder::new(*offset, *end, meta, 16 * 4096);
             let read = read_builder.build();
             let result = vectored_blob_reader.read_blobs(&read, buf, &ctx).await?;
             assert_eq!(result.blobs.len(), 1);

pageserver/src/virtual_file.rs

@@ -23,10 +23,12 @@ use pageserver_api::config::defaults::DEFAULT_IO_BUFFER_ALIGNMENT;
 use pageserver_api::shard::TenantShardId;
 use std::fs::File;
 use std::io::{Error, ErrorKind, Seek, SeekFrom};
+#[cfg(target_os = "linux")]
+use std::os::unix::fs::OpenOptionsExt;
 use tokio_epoll_uring::{BoundedBuf, IoBuf, IoBufMut, Slice};
 
 use std::os::fd::{AsRawFd, FromRawFd, IntoRawFd, OwnedFd, RawFd};
-use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
+use std::sync::atomic::{AtomicBool, AtomicU8, AtomicUsize, Ordering};
 use tokio::sync::{RwLock, RwLockReadGuard, RwLockWriteGuard};
 use tokio::time::Instant;
 
@@ -38,10 +40,11 @@ pub use io_engine::FeatureTestResult as IoEngineFeatureTestResult;
 mod metadata;
 mod open_options;
 use self::owned_buffers_io::write::OwnedAsyncWriter;
-pub(crate) use api::DirectIoMode;
+pub(crate) use api::IoMode;
 pub(crate) use io_engine::IoEngineKind;
 pub(crate) use metadata::Metadata;
 pub(crate) use open_options::*;
+pub(crate) mod dio;
 
 pub(crate) mod owned_buffers_io {
     //! Abstractions for IO with owned buffers.
@@ -53,6 +56,7 @@ pub(crate) mod owned_buffers_io {
     //! but for the time being we're proving out the primitives in the neon.git repo
     //! for faster iteration.
 
+    pub(crate) mod io_buf_aligned;
     pub(crate) mod io_buf_ext;
     pub(crate) mod slice;
     pub(crate) mod write;
@@ -61,6 +65,176 @@ pub(crate) mod owned_buffers_io {
     }
 }
 
+#[derive(Debug)]
+pub enum VirtualFile {
+    Buffered(VirtualFileInner),
+    Direct(VirtualFileInner),
+}
+
+impl VirtualFile {
+    fn inner(&self) -> &VirtualFileInner {
+        match self {
+            Self::Buffered(file) => file,
+            Self::Direct(file) => file,
+        }
+    }
+
+    fn inner_mut(&mut self) -> &mut VirtualFileInner {
+        match self {
+            Self::Buffered(file) => file,
+            Self::Direct(file) => file,
+        }
+    }
+
+    fn into_inner(self) -> VirtualFileInner {
+        match self {
+            Self::Buffered(file) => file,
+            Self::Direct(file) => file,
+        }
+    }
+    /// Open a file in read-only mode. Like File::open.
+    pub async fn open<P: AsRef<Utf8Path>>(
+        path: P,
+        ctx: &RequestContext,
+    ) -> Result<Self, std::io::Error> {
+        let file = VirtualFileInner::open(path, ctx).await?;
+        Ok(Self::Buffered(file))
+    }
+
+    /// Open a file in read-only mode. Like File::open.
+    ///
+    /// `O_DIRECT` will be enabled base on `virtual_file_io_mode`.
+    pub async fn open_v2<P: AsRef<Utf8Path>>(
+        path: P,
+        ctx: &RequestContext,
+    ) -> Result<Self, std::io::Error> {
+        Self::open_with_options_v2(path.as_ref(), OpenOptions::new().read(true), ctx).await
+    }
+
+    pub async fn create<P: AsRef<Utf8Path>>(
+        path: P,
+        ctx: &RequestContext,
+    ) -> Result<Self, std::io::Error> {
+        let file = VirtualFileInner::create(path, ctx).await?;
+        Ok(Self::Buffered(file))
+    }
+
+    pub async fn create_v2<P: AsRef<Utf8Path>>(
+        path: P,
+        ctx: &RequestContext,
+    ) -> Result<Self, std::io::Error> {
+        VirtualFile::open_with_options_v2(
+            path.as_ref(),
+            OpenOptions::new().write(true).create(true).truncate(true),
+            ctx,
+        )
+        .await
+    }
+
+    pub async fn open_with_options<P: AsRef<Utf8Path>>(
+        path: P,
+        open_options: &OpenOptions,
+        ctx: &RequestContext, /* TODO: carry a pointer to the metrics in the RequestContext instead of the parsing https://github.com/neondatabase/neon/issues/6107 */
+    ) -> Result<Self, std::io::Error> {
+        let file = VirtualFileInner::open_with_options(path, open_options, ctx).await?;
+        Ok(Self::Buffered(file))
+    }
+
+    pub async fn open_with_options_v2<P: AsRef<Utf8Path>>(
+        path: P,
+        open_options: &mut OpenOptions, // Uses `&mut` here to add `O_DIRECT`.
+        ctx: &RequestContext, /* TODO: carry a pointer to the metrics in the RequestContext instead of the parsing https://github.com/neondatabase/neon/issues/6107 */
+    ) -> Result<Self, std::io::Error> {
+        let file = match get_io_mode() {
+            IoMode::Buffered => {
+                let file = VirtualFileInner::open_with_options(path, open_options, ctx).await?;
+                Self::Buffered(file)
+            }
+            #[cfg(target_os = "linux")]
+            IoMode::Direct => {
+                let file = VirtualFileInner::open_with_options(
+                    path,
+                    open_options.custom_flags(nix::libc::O_DIRECT),
+                    ctx,
+                )
+                .await?;
+                Self::Direct(file)
+            }
+        };
+        Ok(file)
+    }
+
+    pub fn path(&self) -> &Utf8Path {
+        self.inner().path.as_path()
+    }
+
+    pub async fn crashsafe_overwrite<B: BoundedBuf<Buf = Buf> + Send, Buf: IoBuf + Send>(
+        final_path: Utf8PathBuf,
+        tmp_path: Utf8PathBuf,
+        content: B,
+    ) -> std::io::Result<()> {
+        VirtualFileInner::crashsafe_overwrite(final_path, tmp_path, content).await
+    }
+
+    pub async fn sync_all(&self) -> Result<(), Error> {
+        self.inner().sync_all().await
+    }
+
+    pub async fn sync_data(&self) -> Result<(), Error> {
+        self.inner().sync_data().await
+    }
+
+    pub async fn metadata(&self) -> Result<Metadata, Error> {
+        self.inner().metadata().await
+    }
+
+    pub fn remove(self) {
+        self.into_inner().remove();
+    }
+
+    pub async fn seek(&mut self, pos: SeekFrom) -> Result<u64, Error> {
+        self.inner_mut().seek(pos).await
+    }
+
+    pub async fn read_exact_at<Buf>(
+        &self,
+        slice: Slice<Buf>,
+        offset: u64,
+        ctx: &RequestContext,
+    ) -> Result<Slice<Buf>, Error>
+    where
+        Buf: IoBufMut + Send,
+    {
+        self.inner().read_exact_at(slice, offset, ctx).await
+    }
+
+    pub async fn read_exact_at_page(
+        &self,
+        page: PageWriteGuard<'static>,
+        offset: u64,
+        ctx: &RequestContext,
+    ) -> Result<PageWriteGuard<'static>, Error> {
+        self.inner().read_exact_at_page(page, offset, ctx).await
+    }
+
+    pub async fn write_all_at<Buf: IoBuf + Send>(
+        &self,
+        buf: FullSlice<Buf>,
+        offset: u64,
+        ctx: &RequestContext,
+    ) -> (FullSlice<Buf>, Result<(), Error>) {
+        self.inner().write_all_at(buf, offset, ctx).await
+    }
+
+    pub async fn write_all<Buf: IoBuf + Send>(
+        &mut self,
+        buf: FullSlice<Buf>,
+        ctx: &RequestContext,
+    ) -> (FullSlice<Buf>, Result<usize, Error>) {
+        self.inner_mut().write_all(buf, ctx).await
+    }
+}
+
 ///
 /// A virtual file descriptor. You can use this just like std::fs::File, but internally
 /// the underlying file is closed if the system is low on file descriptors,
@@ -77,7 +251,7 @@ pub(crate) mod owned_buffers_io {
 /// 'tag' field is used to detect whether the handle still is valid or not.
 ///
 #[derive(Debug)]
-pub struct VirtualFile {
+pub struct VirtualFileInner {
     /// Lazy handle to the global file descriptor cache. The slot that this points to
     /// might contain our File, or it may be empty, or it may contain a File that
     /// belongs to a different VirtualFile.
@@ -350,12 +524,12 @@ macro_rules! with_file {
     }};
 }
 
-impl VirtualFile {
+impl VirtualFileInner {
     /// Open a file in read-only mode. Like File::open.
     pub async fn open<P: AsRef<Utf8Path>>(
         path: P,
         ctx: &RequestContext,
-    ) -> Result<VirtualFile, std::io::Error> {
+    ) -> Result<VirtualFileInner, std::io::Error> {
         Self::open_with_options(path.as_ref(), OpenOptions::new().read(true), ctx).await
     }
 
@@ -364,7 +538,7 @@ impl VirtualFile {
     pub async fn create<P: AsRef<Utf8Path>>(
         path: P,
         ctx: &RequestContext,
-    ) -> Result<VirtualFile, std::io::Error> {
+    ) -> Result<VirtualFileInner, std::io::Error> {
         Self::open_with_options(
             path.as_ref(),
             OpenOptions::new().write(true).create(true).truncate(true),
@@ -382,7 +556,7 @@ impl VirtualFile {
         path: P,
         open_options: &OpenOptions,
         _ctx: &RequestContext, /* TODO: carry a pointer to the metrics in the RequestContext instead of the parsing https://github.com/neondatabase/neon/issues/6107 */
-    ) -> Result<VirtualFile, std::io::Error> {
+    ) -> Result<VirtualFileInner, std::io::Error> {
         let path_ref = path.as_ref();
         let path_str = path_ref.to_string();
         let parts = path_str.split('/').collect::<Vec<&str>>();
@@ -423,7 +597,7 @@ impl VirtualFile {
         reopen_options.create_new(false);
         reopen_options.truncate(false);
 
-        let vfile = VirtualFile {
+        let vfile = VirtualFileInner {
             handle: RwLock::new(handle),
             pos: 0,
             path: path_ref.to_path_buf(),
@@ -466,6 +640,7 @@ impl VirtualFile {
                 &[]
             };
             utils::crashsafe::overwrite(&final_path, &tmp_path, content)
+                .maybe_fatal_err("crashsafe_overwrite")
         })
         .await
         .expect("blocking task is never aborted")
@@ -475,7 +650,7 @@ impl VirtualFile {
     pub async fn sync_all(&self) -> Result<(), Error> {
         with_file!(self, StorageIoOperation::Fsync, |file_guard| {
             let (_file_guard, res) = io_engine::get().sync_all(file_guard).await;
-            res
+            res.maybe_fatal_err("sync_all")
         })
     }
 
@@ -483,7 +658,7 @@ impl VirtualFile {
     pub async fn sync_data(&self) -> Result<(), Error> {
         with_file!(self, StorageIoOperation::Fsync, |file_guard| {
             let (_file_guard, res) = io_engine::get().sync_data(file_guard).await;
-            res
+            res.maybe_fatal_err("sync_data")
         })
     }
 
@@ -1033,6 +1208,21 @@ impl tokio_epoll_uring::IoFd for FileGuard {
 
 #[cfg(test)]
 impl VirtualFile {
+    pub(crate) async fn read_blk(
+        &self,
+        blknum: u32,
+        ctx: &RequestContext,
+    ) -> Result<crate::tenant::block_io::BlockLease<'_>, std::io::Error> {
+        self.inner().read_blk(blknum, ctx).await
+    }
+
+    async fn read_to_end(&mut self, buf: &mut Vec<u8>, ctx: &RequestContext) -> Result<(), Error> {
+        self.inner_mut().read_to_end(buf, ctx).await
+    }
+}
+
+#[cfg(test)]
+impl VirtualFileInner {
     pub(crate) async fn read_blk(
         &self,
         blknum: u32,
@@ -1066,7 +1256,7 @@ impl VirtualFile {
     }
 }
 
-impl Drop for VirtualFile {
+impl Drop for VirtualFileInner {
     /// If a VirtualFile is dropped, close the underlying file if it was open.
     fn drop(&mut self) {
         let handle = self.handle.get_mut();
@@ -1142,13 +1332,10 @@ impl OpenFiles {
 /// server startup.
 ///
 #[cfg(not(test))]
-pub fn init(num_slots: usize, engine: IoEngineKind, io_buffer_alignment: usize) {
+pub fn init(num_slots: usize, engine: IoEngineKind) {
     if OPEN_FILES.set(OpenFiles::new(num_slots)).is_err() {
         panic!("virtual_file::init called twice");
     }
-    if set_io_buffer_alignment(io_buffer_alignment).is_err() {
-        panic!("IO buffer alignment ({io_buffer_alignment}) is not a power of two");
-    }
     io_engine::init(engine);
     crate::metrics::virtual_file_descriptor_cache::SIZE_MAX.set(num_slots as u64);
 }
@@ -1172,53 +1359,22 @@ fn get_open_files() -> &'static OpenFiles {
     }
 }
 
-static IO_BUFFER_ALIGNMENT: AtomicUsize = AtomicUsize::new(DEFAULT_IO_BUFFER_ALIGNMENT);
-
-/// Returns true if `x` is zero or a power of two.
-fn is_zero_or_power_of_two(x: usize) -> bool {
-    (x == 0) || ((x & (x - 1)) == 0)
+/// Gets the io buffer alignment.
+pub(crate) const fn get_io_buffer_alignment() -> usize {
+    DEFAULT_IO_BUFFER_ALIGNMENT
 }
 
-#[allow(unused)]
-pub(crate) fn set_io_buffer_alignment(align: usize) -> Result<(), usize> {
-    if is_zero_or_power_of_two(align) {
-        IO_BUFFER_ALIGNMENT.store(align, std::sync::atomic::Ordering::Relaxed);
-        Ok(())
-    } else {
-        Err(align)
-    }
+pub(crate) type IoBufferMut = dio::AlignedBufferMut<{ get_io_buffer_alignment() }>;
+
+static IO_MODE: AtomicU8 = AtomicU8::new(IoMode::preferred() as u8);
+
+pub(crate) fn set_io_mode(mode: IoMode) {
+    IO_MODE.store(mode as u8, std::sync::atomic::Ordering::Relaxed);
 }
 
-/// Gets the io buffer alignment requirement. Returns 0 if there is no requirement specified.
-///
-/// This function should be used to check the raw config value.
-pub(crate) fn get_io_buffer_alignment_raw() -> usize {
-    let align = IO_BUFFER_ALIGNMENT.load(std::sync::atomic::Ordering::Relaxed);
-
-    if cfg!(test) {
-        let env_var_name = "NEON_PAGESERVER_UNIT_TEST_IO_BUFFER_ALIGNMENT";
-        if let Some(test_align) = utils::env::var(env_var_name) {
-            if is_zero_or_power_of_two(test_align) {
-                test_align
-            } else {
-                panic!("IO buffer alignment ({test_align}) is not a power of two");
-            }
-        } else {
-            align
-        }
-    } else {
-        align
-    }
+pub(crate) fn get_io_mode() -> IoMode {
+    IoMode::try_from(IO_MODE.load(Ordering::Relaxed)).unwrap()
 }
-
-/// Gets the io buffer alignment requirement. Returns 1 if the alignment config is set to zero.
-///
-/// This function should be used for getting the actual alignment value to use.
-pub(crate) fn get_io_buffer_alignment() -> usize {
-    let align = get_io_buffer_alignment_raw();
-    align.max(1)
-}
-
 #[cfg(test)]
 mod tests {
     use crate::context::DownloadBehavior;
@@ -1527,7 +1683,7 @@ mod tests {
         // Open the file many times.
         let mut files = Vec::new();
         for _ in 0..VIRTUAL_FILES {
-            let f = VirtualFile::open_with_options(
+            let f = VirtualFileInner::open_with_options(
                 &test_file_path,
                 OpenOptions::new().read(true),
                 &ctx,
@@ -1579,7 +1735,7 @@ mod tests {
         let path = testdir.join("myfile");
         let tmp_path = testdir.join("myfile.tmp");
 
-        VirtualFile::crashsafe_overwrite(path.clone(), tmp_path.clone(), b"foo".to_vec())
+        VirtualFileInner::crashsafe_overwrite(path.clone(), tmp_path.clone(), b"foo".to_vec())
             .await
             .unwrap();
         let mut file = MaybeVirtualFile::from(VirtualFile::open(&path, &ctx).await.unwrap());
@@ -1588,7 +1744,7 @@ mod tests {
         assert!(!tmp_path.exists());
         drop(file);
 
-        VirtualFile::crashsafe_overwrite(path.clone(), tmp_path.clone(), b"bar".to_vec())
+        VirtualFileInner::crashsafe_overwrite(path.clone(), tmp_path.clone(), b"bar".to_vec())
             .await
             .unwrap();
         let mut file = MaybeVirtualFile::from(VirtualFile::open(&path, &ctx).await.unwrap());
@@ -1611,7 +1767,7 @@ mod tests {
         std::fs::write(&tmp_path, "some preexisting junk that should be removed").unwrap();
         assert!(tmp_path.exists());
 
-        VirtualFile::crashsafe_overwrite(path.clone(), tmp_path.clone(), b"foo".to_vec())
+        VirtualFileInner::crashsafe_overwrite(path.clone(), tmp_path.clone(), b"foo".to_vec())
             .await
             .unwrap();
 

pageserver/src/virtual_file/dio.rs

@@ -0,0 +1,405 @@
+#![allow(unused)]
+
+use core::slice;
+use std::{
+    alloc::{self, Layout},
+    cmp,
+    mem::{ManuallyDrop, MaybeUninit},
+    ops::{Deref, DerefMut},
+    ptr::{addr_of_mut, NonNull},
+};
+
+use bytes::buf::UninitSlice;
+
+struct IoBufferPtr(*mut u8);
+
+// SAFETY: We gurantees no one besides `IoBufferPtr` itself has the raw pointer.
+unsafe impl Send for IoBufferPtr {}
+
+/// An aligned buffer type used for I/O.
+pub struct AlignedBufferMut<const ALIGN: usize> {
+    ptr: IoBufferPtr,
+    capacity: usize,
+    len: usize,
+}
+
+impl<const ALIGN: usize> AlignedBufferMut<ALIGN> {
+    /// Constructs a new, empty `IoBufferMut` with at least the specified capacity and alignment.
+    ///
+    /// The buffer will be able to hold at most `capacity` elements and will never resize.
+    ///
+    ///
+    /// # Panics
+    ///
+    /// Panics if the new capacity exceeds `isize::MAX` _bytes_, or if the following alignment requirement is not met:
+    /// * `align` must not be zero,
+    ///
+    /// * `align` must be a power of two,
+    ///
+    /// * `capacity`, when rounded up to the nearest multiple of `align`,
+    ///    must not overflow isize (i.e., the rounded value must be
+    ///    less than or equal to `isize::MAX`).
+    pub fn with_capacity(capacity: usize) -> Self {
+        let layout = Layout::from_size_align(capacity, ALIGN).expect("Invalid layout");
+
+        // SAFETY:  Making an allocation with a sized and aligned layout. The memory is manually freed with the same layout.
+        let ptr = unsafe {
+            let ptr = alloc::alloc(layout);
+            if ptr.is_null() {
+                alloc::handle_alloc_error(layout);
+            }
+            IoBufferPtr(ptr)
+        };
+
+        AlignedBufferMut {
+            ptr,
+            capacity,
+            len: 0,
+        }
+    }
+
+    /// Constructs a new `IoBufferMut` with at least the specified capacity and alignment, filled with zeros.
+    pub fn with_capacity_zeroed(capacity: usize) -> Self {
+        use bytes::BufMut;
+        let mut buf = Self::with_capacity(capacity);
+        buf.put_bytes(0, capacity);
+        buf.len = capacity;
+        buf
+    }
+
+    /// Returns the total number of bytes the buffer can hold.
+    #[inline]
+    pub fn capacity(&self) -> usize {
+        self.capacity
+    }
+
+    /// Returns the alignment of the buffer.
+    #[inline]
+    pub const fn align(&self) -> usize {
+        ALIGN
+    }
+
+    /// Returns the number of bytes in the buffer, also referred to as its 'length'.
+    #[inline]
+    pub fn len(&self) -> usize {
+        self.len
+    }
+
+    /// Force the length of the buffer to `new_len`.
+    #[inline]
+    unsafe fn set_len(&mut self, new_len: usize) {
+        debug_assert!(new_len <= self.capacity());
+        self.len = new_len;
+    }
+
+    #[inline]
+    fn as_ptr(&self) -> *const u8 {
+        self.ptr.0
+    }
+
+    #[inline]
+    fn as_mut_ptr(&mut self) -> *mut u8 {
+        self.ptr.0
+    }
+
+    /// Extracts a slice containing the entire buffer.
+    ///
+    /// Equivalent to `&s[..]`.
+    #[inline]
+    fn as_slice(&self) -> &[u8] {
+        // SAFETY: The pointer is valid and `len` bytes are initialized.
+        unsafe { slice::from_raw_parts(self.as_ptr(), self.len) }
+    }
+
+    /// Extracts a mutable slice of the entire buffer.
+    ///
+    /// Equivalent to `&mut s[..]`.
+    fn as_mut_slice(&mut self) -> &mut [u8] {
+        // SAFETY: The pointer is valid and `len` bytes are initialized.
+        unsafe { slice::from_raw_parts_mut(self.as_mut_ptr(), self.len) }
+    }
+
+    /// Drops the all the contents of the buffer, setting its length to `0`.
+    #[inline]
+    pub fn clear(&mut self) {
+        self.len = 0;
+    }
+
+    /// Reserves capacity for at least `additional` more bytes to be inserted
+    /// in the given `IoBufferMut`. The collection may reserve more space to
+    /// speculatively avoid frequent reallocations. After calling `reserve`,
+    /// capacity will be greater than or equal to `self.len() + additional`.
+    /// Does nothing if capacity is already sufficient.
+    ///
+    /// # Panics
+    ///
+    /// Panics if the new capacity exceeds `isize::MAX` _bytes_.
+    pub fn reserve(&mut self, additional: usize) {
+        if additional > self.capacity() - self.len() {
+            self.reserve_inner(additional);
+        }
+    }
+
+    fn reserve_inner(&mut self, additional: usize) {
+        let Some(required_cap) = self.len().checked_add(additional) else {
+            capacity_overflow()
+        };
+
+        let old_capacity = self.capacity();
+        let align = self.align();
+        // This guarantees exponential growth. The doubling cannot overflow
+        // because `cap <= isize::MAX` and the type of `cap` is `usize`.
+        let cap = cmp::max(old_capacity * 2, required_cap);
+
+        if !is_valid_alloc(cap) {
+            capacity_overflow()
+        }
+        let new_layout = Layout::from_size_align(cap, self.align()).expect("Invalid layout");
+
+        let old_ptr = self.as_mut_ptr();
+
+        // SAFETY: old allocation was allocated with std::alloc::alloc with the same layout,
+        // and we panics on null pointer.
+        let (ptr, cap) = unsafe {
+            let old_layout = Layout::from_size_align_unchecked(old_capacity, align);
+            let ptr = alloc::realloc(old_ptr, old_layout, new_layout.size());
+            if ptr.is_null() {
+                alloc::handle_alloc_error(new_layout);
+            }
+            (IoBufferPtr(ptr), cap)
+        };
+
+        self.ptr = ptr;
+        self.capacity = cap;
+    }
+
+    /// Consumes and leaks the `IoBufferMut`, returning a mutable reference to the contents, &'a mut [u8].
+    pub fn leak<'a>(self) -> &'a mut [u8] {
+        let mut buf = ManuallyDrop::new(self);
+        // SAFETY: leaking the buffer as intended.
+        unsafe { slice::from_raw_parts_mut(buf.as_mut_ptr(), buf.len) }
+    }
+}
+
+fn capacity_overflow() -> ! {
+    panic!("capacity overflow")
+}
+
+// We need to guarantee the following:
+// * We don't ever allocate `> isize::MAX` byte-size objects.
+// * We don't overflow `usize::MAX` and actually allocate too little.
+//
+// On 64-bit we just need to check for overflow since trying to allocate
+// `> isize::MAX` bytes will surely fail. On 32-bit and 16-bit we need to add
+// an extra guard for this in case we're running on a platform which can use
+// all 4GB in user-space, e.g., PAE or x32.
+#[inline]
+fn is_valid_alloc(alloc_size: usize) -> bool {
+    !(usize::BITS < 64 && alloc_size > isize::MAX as usize)
+}
+
+impl<const ALIGN: usize> Drop for AlignedBufferMut<ALIGN> {
+    fn drop(&mut self) {
+        // SAFETY: memory was allocated with std::alloc::alloc with the same layout.
+        unsafe {
+            alloc::dealloc(
+                self.as_mut_ptr(),
+                Layout::from_size_align_unchecked(self.capacity, ALIGN),
+            )
+        }
+    }
+}
+
+impl<const ALIGN: usize> Deref for AlignedBufferMut<ALIGN> {
+    type Target = [u8];
+
+    fn deref(&self) -> &Self::Target {
+        self.as_slice()
+    }
+}
+
+impl<const ALIGN: usize> DerefMut for AlignedBufferMut<ALIGN> {
+    fn deref_mut(&mut self) -> &mut Self::Target {
+        self.as_mut_slice()
+    }
+}
+
+/// SAFETY: When advancing the internal cursor, the caller needs to make sure the bytes advcanced past have been initialized.
+unsafe impl<const ALIGN: usize> bytes::BufMut for AlignedBufferMut<ALIGN> {
+    #[inline]
+    fn remaining_mut(&self) -> usize {
+        // Although a `Vec` can have at most isize::MAX bytes, we never want to grow `IoBufferMut`.
+        // Thus, it can have at most `self.capacity` bytes.
+        self.capacity() - self.len()
+    }
+
+    // SAFETY: Caller needs to make sure the bytes being advanced past have been initialized.
+    #[inline]
+    unsafe fn advance_mut(&mut self, cnt: usize) {
+        let len = self.len();
+        let remaining = self.remaining_mut();
+
+        if remaining < cnt {
+            panic_advance(cnt, remaining);
+        }
+
+        // Addition will not overflow since the sum is at most the capacity.
+        self.set_len(len + cnt);
+    }
+
+    #[inline]
+    fn chunk_mut(&mut self) -> &mut bytes::buf::UninitSlice {
+        let cap = self.capacity();
+        let len = self.len();
+
+        // SAFETY: Since `self.ptr` is valid for `cap` bytes, `self.ptr.add(len)` must be
+        // valid for `cap - len` bytes. The subtraction will not underflow since
+        // `len <= cap`.
+        unsafe { UninitSlice::from_raw_parts_mut(self.as_mut_ptr().add(len), cap - len) }
+    }
+}
+
+/// Panic with a nice error message.
+#[cold]
+fn panic_advance(idx: usize, len: usize) -> ! {
+    panic!(
+        "advance out of bounds: the len is {} but advancing by {}",
+        len, idx
+    );
+}
+
+/// Safety: [`IoBufferMut`] has exclusive ownership of the io buffer,
+/// and the location remains stable even if [`Self`] is moved.
+unsafe impl<const ALIGN: usize> tokio_epoll_uring::IoBuf for AlignedBufferMut<ALIGN> {
+    fn stable_ptr(&self) -> *const u8 {
+        self.as_ptr()
+    }
+
+    fn bytes_init(&self) -> usize {
+        self.len()
+    }
+
+    fn bytes_total(&self) -> usize {
+        self.capacity()
+    }
+}
+
+// SAFETY: See above.
+unsafe impl<const ALIGN: usize> tokio_epoll_uring::IoBufMut for AlignedBufferMut<ALIGN> {
+    fn stable_mut_ptr(&mut self) -> *mut u8 {
+        self.as_mut_ptr()
+    }
+
+    unsafe fn set_init(&mut self, init_len: usize) {
+        if self.len() < init_len {
+            self.set_len(init_len);
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+
+    use super::*;
+
+    const ALIGN: usize = 4 * 1024;
+    type TestIoBufferMut = AlignedBufferMut<ALIGN>;
+
+    #[test]
+    fn test_with_capacity() {
+        let v = TestIoBufferMut::with_capacity(ALIGN * 4);
+        assert_eq!(v.len(), 0);
+        assert_eq!(v.capacity(), ALIGN * 4);
+        assert_eq!(v.align(), ALIGN);
+        assert_eq!(v.as_ptr().align_offset(ALIGN), 0);
+
+        let v = TestIoBufferMut::with_capacity(ALIGN / 2);
+        assert_eq!(v.len(), 0);
+        assert_eq!(v.capacity(), ALIGN / 2);
+        assert_eq!(v.align(), ALIGN);
+        assert_eq!(v.as_ptr().align_offset(ALIGN), 0);
+    }
+
+    #[test]
+    fn test_with_capacity_zeroed() {
+        let v = TestIoBufferMut::with_capacity_zeroed(ALIGN);
+        assert_eq!(v.len(), ALIGN);
+        assert_eq!(v.capacity(), ALIGN);
+        assert_eq!(v.align(), ALIGN);
+        assert_eq!(v.as_ptr().align_offset(ALIGN), 0);
+        assert_eq!(&v[..], &[0; ALIGN])
+    }
+
+    #[test]
+    fn test_reserve() {
+        use bytes::BufMut;
+        let mut v = TestIoBufferMut::with_capacity(ALIGN);
+        let capacity = v.capacity();
+        v.reserve(capacity);
+        assert_eq!(v.capacity(), capacity);
+        let data = [b'a'; ALIGN];
+        v.put(&data[..]);
+        v.reserve(capacity);
+        assert!(v.capacity() >= capacity * 2);
+        assert_eq!(&v[..], &data[..]);
+        let capacity = v.capacity();
+        v.clear();
+        v.reserve(capacity);
+        assert_eq!(capacity, v.capacity());
+    }
+
+    #[test]
+    fn test_bytes_put() {
+        use bytes::BufMut;
+        let mut v = TestIoBufferMut::with_capacity(ALIGN * 4);
+        let x = [b'a'; ALIGN];
+
+        for _ in 0..2 {
+            for _ in 0..4 {
+                v.put(&x[..]);
+            }
+            assert_eq!(v.len(), ALIGN * 4);
+            assert_eq!(v.capacity(), ALIGN * 4);
+            assert_eq!(v.align(), ALIGN);
+            assert_eq!(v.as_ptr().align_offset(ALIGN), 0);
+            v.clear()
+        }
+        assert_eq!(v.len(), 0);
+        assert_eq!(v.capacity(), ALIGN * 4);
+        assert_eq!(v.align(), ALIGN);
+        assert_eq!(v.as_ptr().align_offset(ALIGN), 0);
+    }
+
+    #[test]
+    #[should_panic]
+    fn test_bytes_put_panic() {
+        use bytes::BufMut;
+        const ALIGN: usize = 4 * 1024;
+        let mut v = TestIoBufferMut::with_capacity(ALIGN * 4);
+        let x = [b'a'; ALIGN];
+        for _ in 0..5 {
+            v.put_slice(&x[..]);
+        }
+    }
+
+    #[test]
+    fn test_io_buf_put_slice() {
+        use tokio_epoll_uring::BoundedBufMut;
+        const ALIGN: usize = 4 * 1024;
+        let mut v = TestIoBufferMut::with_capacity(ALIGN);
+        let x = [b'a'; ALIGN];
+
+        for _ in 0..2 {
+            v.put_slice(&x[..]);
+            assert_eq!(v.len(), ALIGN);
+            assert_eq!(v.capacity(), ALIGN);
+            assert_eq!(v.align(), ALIGN);
+            assert_eq!(v.as_ptr().align_offset(ALIGN), 0);
+            v.clear()
+        }
+        assert_eq!(v.len(), 0);
+        assert_eq!(v.capacity(), ALIGN);
+        assert_eq!(v.align(), ALIGN);
+        assert_eq!(v.as_ptr().align_offset(ALIGN), 0);
+    }
+}

pageserver/src/virtual_file/owned_buffers_io/io_buf_aligned.rs

@@ -0,0 +1,9 @@
+#![allow(unused)]
+
+use tokio_epoll_uring::IoBufMut;
+
+use crate::virtual_file::IoBufferMut;
+
+pub(crate) trait IoBufAlignedMut: IoBufMut {}
+
+impl IoBufAlignedMut for IoBufferMut {}

pageserver/src/virtual_file/owned_buffers_io/io_buf_ext.rs

@@ -1,5 +1,6 @@
 //! See [`FullSlice`].
 
+use crate::virtual_file::IoBufferMut;
 use bytes::{Bytes, BytesMut};
 use std::ops::{Deref, Range};
 use tokio_epoll_uring::{BoundedBuf, IoBuf, Slice};
@@ -76,3 +77,4 @@ macro_rules! impl_io_buf_ext {
 impl_io_buf_ext!(Bytes);
 impl_io_buf_ext!(BytesMut);
 impl_io_buf_ext!(Vec<u8>);
+impl_io_buf_ext!(IoBufferMut);

pageserver/src/walingest.rs

@@ -21,7 +21,10 @@
 //! redo Postgres process, but some records it can handle directly with
 //! bespoken Rust code.
 
+use std::sync::Arc;
+use std::sync::OnceLock;
 use std::time::Duration;
+use std::time::Instant;
 use std::time::SystemTime;
 
 use pageserver_api::shard::ShardIdentity;
@@ -69,7 +72,29 @@ impl CheckPoint {
     }
 }
 
+/// Temporary limitation of WAL lag warnings after attach
+///
+/// After tenant attach, we want to limit WAL lag warnings because
+/// we don't look at the WAL until the attach is complete, which
+/// might take a while.
+pub struct WalLagCooldown {
+    /// Until when should this limitation apply at all
+    active_until: std::time::Instant,
+    /// The maximum lag to suppress. Lags above this limit get reported anyways.
+    max_lag: Duration,
+}
+
+impl WalLagCooldown {
+    pub fn new(attach_start: Instant, attach_duration: Duration) -> Self {
+        Self {
+            active_until: attach_start + attach_duration * 3 + Duration::from_secs(120),
+            max_lag: attach_duration * 2 + Duration::from_secs(60),
+        }
+    }
+}
+
 pub struct WalIngest {
+    attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
     shard: ShardIdentity,
     checkpoint: CheckPoint,
     checkpoint_modified: bool,
@@ -103,6 +128,7 @@ impl WalIngest {
             shard: *timeline.get_shard_identity(),
             checkpoint,
             checkpoint_modified: false,
+            attach_wal_lag_cooldown: timeline.attach_wal_lag_cooldown.clone(),
             warn_ingest_lag: WarnIngestLag {
                 lag_msg_ratelimit: RateLimit::new(std::time::Duration::from_secs(10)),
                 future_lsn_msg_ratelimit: RateLimit::new(std::time::Duration::from_secs(10)),
@@ -1429,6 +1455,13 @@ impl WalIngest {
                     Ok(lag) => {
                         if lag > conf.wait_lsn_timeout {
                             rate_limits.lag_msg_ratelimit.call2(|rate_limit_stats| {
+                                if let Some(cooldown) = self.attach_wal_lag_cooldown.get() {
+                                    if std::time::Instant::now() < cooldown.active_until && lag <= cooldown.max_lag {
+                                        return;
+                                    }
+                                } else {
+                                    // Still loading? We shouldn't be here
+                                }
                                 let lag = humantime::format_duration(lag);
                                 warn!(%rate_limit_stats, %lag, "ingesting record with timestamp lagging more than wait_lsn_timeout");
                             })

pgxn/neon/file_cache.c

@@ -42,6 +42,7 @@
 
 #include "hll.h"
 #include "bitmap.h"
+#include "neon.h"
 
 #define CriticalAssert(cond) do if (!(cond)) elog(PANIC, "Assertion %s failed at %s:%d: ", #cond, __FILE__, __LINE__); while (0)
 
@@ -173,7 +174,9 @@ lfc_disable(char const *op)
 			 * If the reason of error is ENOSPC, then truncation of file may
 			 * help to reclaim some space
 			 */
+			pgstat_report_wait_start(WAIT_EVENT_NEON_LFC_TRUNCATE);
 			int			rc = ftruncate(lfc_desc, 0);
+			pgstat_report_wait_end();
 
 			if (rc < 0)
 				elog(WARNING, "Failed to truncate local file cache %s: %m", lfc_path);
@@ -769,8 +772,10 @@ lfc_readv_select(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
 
 		if (iteration_hits != 0)
 		{
+			pgstat_report_wait_start(WAIT_EVENT_NEON_LFC_READ);
 			rc = preadv(lfc_desc, iov, blocks_in_chunk,
 						((off_t) entry_offset * BLOCKS_PER_CHUNK + chunk_offs) * BLCKSZ);
+			pgstat_report_wait_end();
 
 			if (rc != (BLCKSZ * blocks_in_chunk))
 			{
@@ -944,8 +949,11 @@ lfc_writev(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
 		lfc_ctl->writes += blocks_in_chunk;
 		LWLockRelease(lfc_lock);
 
+		pgstat_report_wait_start(WAIT_EVENT_NEON_LFC_WRITE);
 		rc = pwritev(lfc_desc, iov, blocks_in_chunk,
 					 ((off_t) entry_offset * BLOCKS_PER_CHUNK + chunk_offs) * BLCKSZ);
+		pgstat_report_wait_end();
+
 		if (rc != BLCKSZ * blocks_in_chunk)
 		{
 			lfc_disable("write");

pgxn/neon/libpagestore.c

@@ -490,7 +490,7 @@ pageserver_connect(shardno_t shard_no, int elevel)
 											   WL_EXIT_ON_PM_DEATH | WL_LATCH_SET | WL_SOCKET_READABLE,
 											   PQsocket(shard->conn),
 											   0,
-											   PG_WAIT_EXTENSION);
+											   WAIT_EVENT_NEON_PS_STARTING);
 					elog(DEBUG5, "PGRES_POLLING_READING=>%d", rc);
 					if (rc & WL_LATCH_SET)
 					{
@@ -512,7 +512,7 @@ pageserver_connect(shardno_t shard_no, int elevel)
 											   WL_EXIT_ON_PM_DEATH | WL_LATCH_SET | WL_SOCKET_WRITEABLE,
 											   PQsocket(shard->conn),
 											   0,
-											   PG_WAIT_EXTENSION);
+											   WAIT_EVENT_NEON_PS_STARTING);
 					elog(DEBUG5, "PGRES_POLLING_WRITING=>%d", rc);
 					if (rc & WL_LATCH_SET)
 					{
@@ -608,7 +608,8 @@ pageserver_connect(shardno_t shard_no, int elevel)
 			WaitEvent	event;
 
 			/* Sleep until there's something to do */
-			(void) WaitEventSetWait(shard->wes_read, -1L, &event, 1, PG_WAIT_EXTENSION);
+			(void) WaitEventSetWait(shard->wes_read, -1L, &event, 1,
+									WAIT_EVENT_NEON_PS_CONFIGURING);
 			ResetLatch(MyLatch);
 
 			CHECK_FOR_INTERRUPTS();
@@ -656,7 +657,8 @@ static int
 call_PQgetCopyData(shardno_t shard_no, char **buffer)
 {
 	int			ret;
-	PGconn	   *pageserver_conn = page_servers[shard_no].conn;
+	PageServer *shard = &page_servers[shard_no];
+	PGconn	   *pageserver_conn = shard->conn;
 
 retry:
 	ret = PQgetCopyData(pageserver_conn, buffer, 1 /* async */ );
@@ -666,7 +668,8 @@ retry:
 		WaitEvent	event;
 
 		/* Sleep until there's something to do */
-		(void) WaitEventSetWait(page_servers[shard_no].wes_read, -1L, &event, 1, PG_WAIT_EXTENSION);
+		(void) WaitEventSetWait(shard->wes_read, -1L, &event, 1,
+								WAIT_EVENT_NEON_PS_READ);
 		ResetLatch(MyLatch);
 
 		CHECK_FOR_INTERRUPTS();
@@ -937,7 +940,7 @@ PagestoreShmemInit(void)
 
 	LWLockAcquire(AddinShmemInitLock, LW_EXCLUSIVE);
 	pagestore_shared = ShmemInitStruct("libpagestore shared state",
-									   PagestoreShmemSize(),
+									   sizeof(PagestoreShmemState),
 									   &found);
 	if (!found)
 	{

pgxn/neon/neon.c

@@ -41,6 +41,9 @@
 #include "pagestore_client.h"
 #include "control_plane_connector.h"
 #include "walsender_hooks.h"
+#if PG_MAJORVERSION_NUM >= 16
+#include "storage/ipc.h"
+#endif
 
 PG_MODULE_MAGIC;
 void		_PG_init(void);
@@ -49,6 +52,23 @@ static int	logical_replication_max_snap_files = 300;
 
 static int  running_xacts_overflow_policy;
 
+#if PG_MAJORVERSION_NUM >= 16
+static shmem_startup_hook_type prev_shmem_startup_hook;
+
+static void neon_shmem_startup_hook(void);
+#endif
+#if PG_MAJORVERSION_NUM >= 17
+uint32		WAIT_EVENT_NEON_LFC_MAINTENANCE;
+uint32		WAIT_EVENT_NEON_LFC_READ;
+uint32		WAIT_EVENT_NEON_LFC_TRUNCATE;
+uint32		WAIT_EVENT_NEON_LFC_WRITE;
+uint32		WAIT_EVENT_NEON_PS_STARTING;
+uint32		WAIT_EVENT_NEON_PS_CONFIGURING;
+uint32		WAIT_EVENT_NEON_PS_SEND;
+uint32		WAIT_EVENT_NEON_PS_READ;
+uint32		WAIT_EVENT_NEON_WAL_DL;
+#endif
+
 enum RunningXactsOverflowPolicies {
 	OP_IGNORE,
 	OP_SKIP,
@@ -635,6 +655,9 @@ _PG_init(void)
 	 */
 #if PG_VERSION_NUM >= 160000
 	load_file("$libdir/neon_rmgr", false);
+
+	prev_shmem_startup_hook = shmem_startup_hook;
+	shmem_startup_hook = neon_shmem_startup_hook;
 #endif
 
 	pg_init_libpagestore();
@@ -721,3 +744,25 @@ backpressure_throttling_time(PG_FUNCTION_ARGS)
 {
 	PG_RETURN_UINT64(BackpressureThrottlingTime());
 }
+
+#if PG_MAJORVERSION_NUM >= 16
+static void
+neon_shmem_startup_hook(void)
+{
+	/* Initialize */
+	if (prev_shmem_startup_hook)
+		prev_shmem_startup_hook();
+
+#if PG_PG_MAJORVERSION_NUM >= 17
+	WAIT_EVENT_NEON_LFC_MAINTENANCE = WaitEventExtensionNew("Neon/FileCache_Maintenance");
+	WAIT_EVENT_NEON_LFC_READ = WaitEventExtensionNew("Neon/FileCache_Read");
+	WAIT_EVENT_NEON_LFC_TRUNCATE = WaitEventExtensionNew("Neon/FileCache_Truncate");
+	WAIT_EVENT_NEON_LFC_WRITE = WaitEventExtensionNew("Neon/FileCache_Write");
+	WAIT_EVENT_NEON_PS_STARTING = WaitEventExtensionNew("Neon/PS_Starting");
+	WAIT_EVENT_NEON_PS_CONFIGURING = WaitEventExtensionNew("Neon/PS_Configuring");
+	WAIT_EVENT_NEON_PS_SEND = WaitEventExtensionNew("Neon/PS_SendIO");
+	WAIT_EVENT_NEON_PS_READ = WaitEventExtensionNew("Neon/PS_ReadIO");
+	WAIT_EVENT_NEON_WAL_DL = WaitEventExtensionNew("Neon/WAL_Download");
+#endif
+}
+#endif

pgxn/neon/neon.control

@@ -1,8 +1,6 @@
 # neon extension
 comment = 'cloud storage for PostgreSQL'
-# TODO: bump default version to 1.5, after we are certain that we don't
-# need to rollback the compute image
-default_version = '1.4'
+default_version = '1.5'
 module_pathname = '$libdir/neon'
 relocatable = true
 trusted = true

pgxn/neon/neon.h

@@ -12,6 +12,7 @@
 #ifndef NEON_H
 #define NEON_H
 #include "access/xlogreader.h"
+#include "utils/wait_event.h"
 
 /* GUCs */
 extern char *neon_auth_token;
@@ -22,6 +23,28 @@ extern char *wal_acceptors_list;
 extern int	wal_acceptor_reconnect_timeout;
 extern int	wal_acceptor_connection_timeout;
 
+#if PG_MAJORVERSION_NUM >= 17
+extern uint32		WAIT_EVENT_NEON_LFC_MAINTENANCE;
+extern uint32		WAIT_EVENT_NEON_LFC_READ;
+extern uint32		WAIT_EVENT_NEON_LFC_TRUNCATE;
+extern uint32		WAIT_EVENT_NEON_LFC_WRITE;
+extern uint32		WAIT_EVENT_NEON_PS_STARTING;
+extern uint32		WAIT_EVENT_NEON_PS_CONFIGURING;
+extern uint32		WAIT_EVENT_NEON_PS_SEND;
+extern uint32		WAIT_EVENT_NEON_PS_READ;
+extern uint32		WAIT_EVENT_NEON_WAL_DL;
+#else
+#define WAIT_EVENT_NEON_LFC_MAINTENANCE	PG_WAIT_EXTENSION
+#define WAIT_EVENT_NEON_LFC_READ		WAIT_EVENT_BUFFILE_READ
+#define WAIT_EVENT_NEON_LFC_TRUNCATE	WAIT_EVENT_BUFFILE_TRUNCATE
+#define WAIT_EVENT_NEON_LFC_WRITE		WAIT_EVENT_BUFFILE_WRITE
+#define WAIT_EVENT_NEON_PS_STARTING		PG_WAIT_EXTENSION
+#define WAIT_EVENT_NEON_PS_CONFIGURING	PG_WAIT_EXTENSION
+#define WAIT_EVENT_NEON_PS_SEND			PG_WAIT_EXTENSION
+#define WAIT_EVENT_NEON_PS_READ			PG_WAIT_EXTENSION
+#define WAIT_EVENT_NEON_WAL_DL			WAIT_EVENT_WAL_READ
+#endif
+
 extern void pg_init_libpagestore(void);
 extern void pg_init_walproposer(void);
 

pgxn/neon/neon_perf_counters.c

@@ -27,7 +27,8 @@ NeonPerfCountersShmemSize(void)
 {
 	Size		size = 0;
 
-	size = add_size(size, mul_size(MaxBackends, sizeof(neon_per_backend_counters)));
+	size = add_size(size, mul_size(NUM_NEON_PERF_COUNTER_SLOTS,
+								   sizeof(neon_per_backend_counters)));
 
 	return size;
 }
@@ -39,7 +40,7 @@ NeonPerfCountersShmemInit(void)
 
 	neon_per_backend_counters_shared =
 		ShmemInitStruct("Neon perf counters",
-						mul_size(MaxBackends,
+						mul_size(NUM_NEON_PERF_COUNTER_SLOTS,
 								 sizeof(neon_per_backend_counters)),
 						&found);
 	Assert(found == IsUnderPostmaster);
@@ -137,7 +138,7 @@ neon_perf_counters_to_metrics(neon_per_backend_counters *counters)
 	metrics[i].is_bucket = false;
 	metrics[i].value = (double) counters->pageserver_requests_sent_total;
 	i++;
-	metrics[i].name = "pageserver_requests_disconnects_total";
+	metrics[i].name = "pageserver_disconnects_total";
 	metrics[i].is_bucket = false;
 	metrics[i].value = (double) counters->pageserver_disconnects_total;
 	i++;
@@ -192,7 +193,7 @@ neon_get_backend_perf_counters(PG_FUNCTION_ARGS)
 	/* We put all the tuples into a tuplestore in one go. */
 	InitMaterializedSRF(fcinfo, 0);
 
-	for (int procno = 0; procno < MaxBackends; procno++)
+	for (int procno = 0; procno < NUM_NEON_PERF_COUNTER_SLOTS; procno++)
 	{
 		PGPROC	   *proc = GetPGProcByNumber(procno);
 		int			pid = proc->pid;
@@ -231,7 +232,7 @@ neon_get_perf_counters(PG_FUNCTION_ARGS)
 	InitMaterializedSRF(fcinfo, 0);
 
 	/* Aggregate the counters across all backends */
-	for (int procno = 0; procno < MaxBackends; procno++)
+	for (int procno = 0; procno < NUM_NEON_PERF_COUNTER_SLOTS; procno++)
 	{
 		neon_per_backend_counters *counters = &neon_per_backend_counters_shared[procno];
 

pgxn/neon/neon_perf_counters.h

@@ -96,6 +96,14 @@ typedef struct
 /* Pointer to the shared memory array of neon_per_backend_counters structs */
 extern neon_per_backend_counters *neon_per_backend_counters_shared;
 
+/*
+ * Size of the perf counters array in shared memory. One slot for each backend
+ * and aux process. IOW one for each PGPROC slot, except for slots reserved
+ * for prepared transactions, because they're not real processes and cannot do
+ * I/O.
+ */
+#define NUM_NEON_PERF_COUNTER_SLOTS (MaxBackends + NUM_AUXILIARY_PROCS)
+
 #if PG_VERSION_NUM >= 170000
 #define MyNeonCounters (&neon_per_backend_counters_shared[MyProcNumber])
 #else

pgxn/neon/neon_pgversioncompat.h

@@ -7,6 +7,7 @@
 #define NEON_PGVERSIONCOMPAT_H
 
 #include "fmgr.h"
+#include "storage/buf_internals.h"
 
 #if PG_MAJORVERSION_NUM < 17
 #define NRelFileInfoBackendIsTemp(rinfo) (rinfo.backend != InvalidBackendId)
@@ -20,11 +21,24 @@
 	NInfoGetRelNumber(a) == NInfoGetRelNumber(b) \
 )
 
-/* buftag population & RelFileNode/RelFileLocator rework */
+/* These macros were turned into static inline functions in v16 */
 #if PG_MAJORVERSION_NUM < 16
+static inline bool
+BufferTagsEqual(const BufferTag *tag1, const BufferTag *tag2)
+{
+	return BUFFERTAGS_EQUAL(*tag1, *tag2);
+}
 
-#define InitBufferTag(tag, rfn, fn, bn) INIT_BUFFERTAG(*tag, *rfn, fn, bn)
+static inline void
+InitBufferTag(BufferTag *tag, const RelFileNode *rnode,
+			  ForkNumber forkNum, BlockNumber blockNum)
+{
+	INIT_BUFFERTAG(*tag, *rnode, forkNum, blockNum);
+}
+#endif
 
+/* RelFileNode -> RelFileLocator rework */
+#if PG_MAJORVERSION_NUM < 16
 #define USE_RELFILENODE
 
 #define RELFILEINFO_HDR "storage/relfilenode.h"
@@ -73,8 +87,6 @@
 
 #define USE_RELFILELOCATOR
 
-#define BUFFERTAGS_EQUAL(a, b) BufferTagsEqual(&(a), &(b))
-
 #define RELFILEINFO_HDR "storage/relfilelocator.h"
 
 #define NRelFileInfo RelFileLocator

pgxn/neon/pagestore_smgr.c

@@ -215,7 +215,7 @@ typedef struct PrfHashEntry
 	sizeof(BufferTag) \
 )
 
-#define SH_EQUAL(tb, a, b)	(BUFFERTAGS_EQUAL((a)->buftag, (b)->buftag))
+#define SH_EQUAL(tb, a, b)	(BufferTagsEqual(&(a)->buftag, &(b)->buftag))
 #define SH_SCOPE			static inline
 #define SH_DEFINE
 #define SH_DECLARE
@@ -803,15 +803,19 @@ prefetch_register_bufferv(BufferTag tag, neon_request_lsns *frlsns,
 						  bool is_prefetch)
 {
 	uint64		min_ring_index;
-	PrefetchRequest req;
+	PrefetchRequest hashkey;
 #if USE_ASSERT_CHECKING
 	bool		any_hits = false;
 #endif
 	/* We will never read further ahead than our buffer can store. */
 	nblocks = Max(1, Min(nblocks, readahead_buffer_size));
 
-	/* use an intermediate PrefetchRequest struct to ensure correct alignment */
-	req.buftag = tag;
+	/*
+	 * Use an intermediate PrefetchRequest struct as the hash key to ensure
+	 * correct alignment and that the padding bytes are cleared.
+	 */
+	memset(&hashkey.buftag, 0, sizeof(BufferTag));
+	hashkey.buftag = tag;
 
 Retry:
 	min_ring_index = UINT64_MAX;
@@ -837,8 +841,8 @@ Retry:
 		slot = NULL;
 		entry = NULL;
 
-		req.buftag.blockNum = tag.blockNum + i;
-		entry = prfh_lookup(MyPState->prf_hash, (PrefetchRequest *) &req);
+		hashkey.buftag.blockNum = tag.blockNum + i;
+		entry = prfh_lookup(MyPState->prf_hash, &hashkey);
 
 		if (entry != NULL)
 		{
@@ -849,7 +853,7 @@ Retry:
 			Assert(slot->status != PRFS_UNUSED);
 			Assert(MyPState->ring_last <= ring_index &&
 				   ring_index < MyPState->ring_unused);
-			Assert(BUFFERTAGS_EQUAL(slot->buftag, req.buftag));
+			Assert(BufferTagsEqual(&slot->buftag, &hashkey.buftag));
 
 			/*
 			 * If the caller specified a request LSN to use, only accept
@@ -886,12 +890,19 @@ Retry:
 				{
 					min_ring_index = Min(min_ring_index, ring_index);
 					/* The buffered request is good enough, return that index */
-					pgBufferUsage.prefetch.duplicates++;
+					if (is_prefetch)
+						pgBufferUsage.prefetch.duplicates++;
+					else
+						pgBufferUsage.prefetch.hits++;
 					continue;
 				}
 			}
 		}
-
+		else if (!is_prefetch)
+		{
+			pgBufferUsage.prefetch.misses += 1;
+			MyNeonCounters->getpage_prefetch_misses_total++;
+		}
 		/*
 		 * We can only leave the block above by finding that there's
 		 * no entry that can satisfy this request, either because there
@@ -974,7 +985,7 @@ Retry:
 		 * We must update the slot data before insertion, because the hash
 		 * function reads the buffer tag from the slot.
 		 */
-		slot->buftag = req.buftag;
+		slot->buftag = hashkey.buftag;
 		slot->shard_no = get_shard_number(&tag);
 		slot->my_ring_index = ring_index;
 
@@ -1773,6 +1784,20 @@ neon_init(void)
 	if (MyPState != NULL)
 		return;
 
+	/*
+	 * Sanity check that theperf counters array is sized correctly. We got
+	 * this wrong once, and the formula for max number of backends and aux
+	 * processes might well change in the future, so better safe than sorry.
+	 * This is a very cheap check so we do it even without assertions.  On
+	 * v14, this gets called before initializing MyProc, so we cannot perform
+	 * the check here. That's OK, we don't expect the logic to change in old
+	 * releases.
+	 */
+#if PG_VERSION_NUM>=150000
+	if (MyNeonCounters >= &neon_per_backend_counters_shared[NUM_NEON_PERF_COUNTER_SLOTS])
+		elog(ERROR, "MyNeonCounters points past end of array");
+#endif
+
 	prfs_size = offsetof(PrefetchState, prf_buffer) +
 		sizeof(PrefetchRequest) * readahead_buffer_size;
 
@@ -2728,14 +2753,19 @@ neon_read_at_lsnv(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber base_block
 	uint64		ring_index;
 	PrfHashEntry *entry;
 	PrefetchRequest *slot;
-	BufferTag	buftag = {0};
+	PrefetchRequest hashkey;
 
 	Assert(PointerIsValid(request_lsns));
 	Assert(nblocks >= 1);
 
-	CopyNRelFileInfoToBufTag(buftag, rinfo);
-	buftag.forkNum = forkNum;
-	buftag.blockNum = base_blockno;
+	/*
+	 * Use an intermediate PrefetchRequest struct as the hash key to ensure
+	 * correct alignment and that the padding bytes are cleared.
+	 */
+	memset(&hashkey.buftag, 0, sizeof(BufferTag));
+	CopyNRelFileInfoToBufTag(hashkey.buftag, rinfo);
+	hashkey.buftag.forkNum = forkNum;
+	hashkey.buftag.blockNum = base_blockno;
 
 	/*
 	 * The redo process does not lock pages that it needs to replay but are
@@ -2753,7 +2783,7 @@ neon_read_at_lsnv(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber base_block
 	 * weren't for the behaviour of the LwLsn cache that uses the highest
 	 * value of the LwLsn cache when the entry is not found.
 	 */
-	prefetch_register_bufferv(buftag, request_lsns, nblocks, mask, false);
+	prefetch_register_bufferv(hashkey.buftag, request_lsns, nblocks, mask, false);
 
 	for (int i = 0; i < nblocks; i++)
 	{
@@ -2774,8 +2804,8 @@ neon_read_at_lsnv(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber base_block
 		 * Try to find prefetched page in the list of received pages.
 		 */
 Retry:
-		buftag.blockNum = blockno;
-		entry = prfh_lookup(MyPState->prf_hash, (PrefetchRequest *) &buftag);
+		hashkey.buftag.blockNum = blockno;
+		entry = prfh_lookup(MyPState->prf_hash, &hashkey);
 
 		if (entry != NULL)
 		{
@@ -2783,7 +2813,6 @@ Retry:
 			if (neon_prefetch_response_usable(reqlsns, slot))
 			{
 				ring_index = slot->my_ring_index;
-				pgBufferUsage.prefetch.hits += 1;
 			}
 			else
 			{
@@ -2813,10 +2842,7 @@ Retry:
 		{
 			if (entry == NULL)
 			{
-				pgBufferUsage.prefetch.misses += 1;
-				MyNeonCounters->getpage_prefetch_misses_total++;
-
-				ring_index = prefetch_register_bufferv(buftag, reqlsns, 1, NULL, false);
+				ring_index = prefetch_register_bufferv(hashkey.buftag, reqlsns, 1, NULL, false);
 				Assert(ring_index != UINT64_MAX);
 				slot = GetPrfSlot(ring_index);
 			}
@@ -2841,8 +2867,8 @@ Retry:
 		} while (!prefetch_wait_for(ring_index));
 
 		Assert(slot->status == PRFS_RECEIVED);
-		Assert(memcmp(&buftag, &slot->buftag, sizeof(BufferTag)) == 0);
-		Assert(buftag.blockNum == base_blockno + i);
+		Assert(memcmp(&hashkey.buftag, &slot->buftag, sizeof(BufferTag)) == 0);
+		Assert(hashkey.buftag.blockNum == base_blockno + i);
 
 		resp = slot->response;
 
@@ -3045,6 +3071,9 @@ neon_readv(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
 	lfc_result = lfc_readv_select(InfoFromSMgrRel(reln), forknum, blocknum, buffers,
 								  nblocks, read);
 
+	if (lfc_result > 0)
+		MyNeonCounters->file_cache_hits_total += lfc_result;
+
 	/* Read all blocks from LFC, so we're done */
 	if (lfc_result == nblocks)
 		return;

pgxn/neon/walproposer.c

@@ -213,7 +213,7 @@ WalProposerPoll(WalProposer *wp)
 		rc = wp->api.wait_event_set(wp, timeout, &sk, &events);
 
 		/* Exit loop if latch is set (we got new WAL) */
-		if ((rc == 1 && events & WL_LATCH_SET))
+		if (rc == 1 && (events & WL_LATCH_SET))
 			break;
 
 		/*

pgxn/neon/walproposer_pg.c

@@ -422,6 +422,9 @@ backpressure_throttling_impl(void)
 	TimestampTz start,
 				stop;
 	bool		retry = false;
+	char	   *new_status = NULL;
+	const char *old_status;
+	int			len;
 
 	if (PointerIsValid(PrevProcessInterruptsCallback))
 		retry = PrevProcessInterruptsCallback();
@@ -442,14 +445,24 @@ backpressure_throttling_impl(void)
 	if (lag == 0)
 		return retry;
 
-	/* Suspend writers until replicas catch up */
-	set_ps_display("backpressure throttling");
+
+	old_status = get_ps_display(&len);
+	new_status = (char *) palloc(len + 64 + 1);
+	memcpy(new_status, old_status, len);
+	snprintf(new_status + len, 64, "backpressure throttling: lag %lu", lag);
+	set_ps_display(new_status);
+	new_status[len] = '\0'; /* truncate off " backpressure ..." to later reset the ps */
 
 	elog(DEBUG2, "backpressure throttling: lag %lu", lag);
 	start = GetCurrentTimestamp();
 	pg_usleep(BACK_PRESSURE_DELAY);
 	stop = GetCurrentTimestamp();
 	pg_atomic_add_fetch_u64(&walprop_shared->backpressureThrottlingTime, stop - start);
+
+	/* Reset ps display */
+	set_ps_display(new_status);
+	pfree(new_status);
+
 	return true;
 }
 
@@ -1473,11 +1486,33 @@ walprop_pg_wal_read(Safekeeper *sk, char *buf, XLogRecPtr startptr, Size count,
 {
 	NeonWALReadResult res;
 
-	res = NeonWALRead(sk->xlogreader,
-					  buf,
-					  startptr,
-					  count,
-					  walprop_pg_get_timeline_id());
+#if PG_MAJORVERSION_NUM >= 17
+	if (!sk->wp->config->syncSafekeepers)
+	{
+		Size	rbytes;
+		rbytes = WALReadFromBuffers(buf, startptr, count,
+									walprop_pg_get_timeline_id());
+
+		startptr += rbytes;
+		count -= rbytes;
+	}
+#endif
+
+	if (count == 0)
+	{
+		res = NEON_WALREAD_SUCCESS;
+	}
+	else
+	{
+		Assert(count > 0);
+
+		/* Now read the remaining WAL from the WAL file */
+		res = NeonWALRead(sk->xlogreader,
+						  buf,
+						  startptr,
+						  count,
+						  walprop_pg_get_timeline_id());
+	}
 
 	if (res == NEON_WALREAD_SUCCESS)
 	{
@@ -1779,7 +1814,7 @@ walprop_pg_wait_event_set(WalProposer *wp, long timeout, Safekeeper **sk, uint32
 	 * If wait is terminated by latch set (walsenders' latch is set on each
 	 * wal flush). (no need for pm death check due to WL_EXIT_ON_PM_DEATH)
 	 */
-	if ((rc == 1 && event.events & WL_LATCH_SET) || late_cv_trigger)
+	if ((rc == 1 && (event.events & WL_LATCH_SET)) || late_cv_trigger)
 	{
 		/* Reset our latch */
 		ResetLatch(MyLatch);
@@ -1791,7 +1826,7 @@ walprop_pg_wait_event_set(WalProposer *wp, long timeout, Safekeeper **sk, uint32
 	 * If the event contains something about the socket, it means we got an
 	 * event from a safekeeper socket.
 	 */
-	if (rc == 1 && (event.events & (WL_SOCKET_MASK)))
+	if (rc == 1 && (event.events & WL_SOCKET_MASK))
 	{
 		*sk = (Safekeeper *) event.user_data;
 		*events = event.events;

pgxn/neon/walsender_hooks.c

@@ -160,7 +160,7 @@ NeonWALPageRead(
 							  WL_LATCH_SET | WL_EXIT_ON_PM_DEATH | reader_events,
 							  sock,
 							  timeout_ms,
-							  WAIT_EVENT_WAL_SENDER_MAIN);
+							  WAIT_EVENT_NEON_WAL_DL);
 		}
 	}
 }
@@ -191,13 +191,14 @@ NeonOnDemandXLogReaderRoutines(XLogReaderRoutine *xlr)
 
 	if (!wal_reader)
 	{
-		XLogRecPtr	epochStartLsn = pg_atomic_read_u64(&GetWalpropShmemState()->propEpochStartLsn);
+		XLogRecPtr	basebackupLsn = GetRedoStartLsn();
 
-		if (epochStartLsn == 0)
+		/* should never happen */
+		if (basebackupLsn == 0)
 		{
-			elog(ERROR, "Unable to start walsender when propEpochStartLsn is 0!");
+			elog(ERROR, "unable to start walsender when basebackupLsn is 0");
 		}
-		wal_reader = NeonWALReaderAllocate(wal_segment_size, epochStartLsn, "[walsender] ");
+		wal_reader = NeonWALReaderAllocate(wal_segment_size, basebackupLsn, "[walsender] ");
 	}
 	xlr->page_read = NeonWALPageRead;
 	xlr->segment_open = NeonWALReadSegmentOpen;

pgxn/neon_walredo/walredoproc.c

@@ -992,7 +992,7 @@ redo_block_filter(XLogReaderState *record, uint8 block_id)
 	 * If this block isn't one we are currently restoring, then return 'true'
 	 * so that this gets ignored
 	 */
-	return !BUFFERTAGS_EQUAL(target_tag, target_redo_tag);
+	return !BufferTagsEqual(&target_tag, &target_redo_tag);
 }
 
 /*

poetry.lock

@@ -2064,73 +2064,80 @@ test = ["enum34", "ipaddress", "mock", "pywin32", "wmi"]
 
 [[package]]
 name = "psycopg2-binary"
-version = "2.9.6"
+version = "2.9.9"
 description = "psycopg2 - Python-PostgreSQL Database Adapter"
 optional = false
-python-versions = ">=3.6"
+python-versions = ">=3.7"
 files = [
-    {file = "psycopg2-binary-2.9.6.tar.gz", hash = "sha256:1f64dcfb8f6e0c014c7f55e51c9759f024f70ea572fbdef123f85318c297947c"},
-    {file = "psycopg2_binary-2.9.6-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:d26e0342183c762de3276cca7a530d574d4e25121ca7d6e4a98e4f05cb8e4df7"},
-    {file = "psycopg2_binary-2.9.6-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:c48d8f2db17f27d41fb0e2ecd703ea41984ee19362cbce52c097963b3a1b4365"},
-    {file = "psycopg2_binary-2.9.6-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ffe9dc0a884a8848075e576c1de0290d85a533a9f6e9c4e564f19adf8f6e54a7"},
-    {file = "psycopg2_binary-2.9.6-cp310-cp310-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:8a76e027f87753f9bd1ab5f7c9cb8c7628d1077ef927f5e2446477153a602f2c"},
-    {file = "psycopg2_binary-2.9.6-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:6460c7a99fc939b849431f1e73e013d54aa54293f30f1109019c56a0b2b2ec2f"},
-    {file = "psycopg2_binary-2.9.6-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ae102a98c547ee2288637af07393dd33f440c25e5cd79556b04e3fca13325e5f"},
-    {file = "psycopg2_binary-2.9.6-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:9972aad21f965599ed0106f65334230ce826e5ae69fda7cbd688d24fa922415e"},
-    {file = "psycopg2_binary-2.9.6-cp310-cp310-musllinux_1_1_i686.whl", hash = "sha256:7a40c00dbe17c0af5bdd55aafd6ff6679f94a9be9513a4c7e071baf3d7d22a70"},
-    {file = "psycopg2_binary-2.9.6-cp310-cp310-musllinux_1_1_ppc64le.whl", hash = "sha256:cacbdc5839bdff804dfebc058fe25684cae322987f7a38b0168bc1b2df703fb1"},
-    {file = "psycopg2_binary-2.9.6-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:7f0438fa20fb6c7e202863e0d5ab02c246d35efb1d164e052f2f3bfe2b152bd0"},
-    {file = "psycopg2_binary-2.9.6-cp310-cp310-win32.whl", hash = "sha256:b6c8288bb8a84b47e07013bb4850f50538aa913d487579e1921724631d02ea1b"},
-    {file = "psycopg2_binary-2.9.6-cp310-cp310-win_amd64.whl", hash = "sha256:61b047a0537bbc3afae10f134dc6393823882eb263088c271331602b672e52e9"},
-    {file = "psycopg2_binary-2.9.6-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:964b4dfb7c1c1965ac4c1978b0f755cc4bd698e8aa2b7667c575fb5f04ebe06b"},
-    {file = "psycopg2_binary-2.9.6-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:afe64e9b8ea66866a771996f6ff14447e8082ea26e675a295ad3bdbffdd72afb"},
-    {file = "psycopg2_binary-2.9.6-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:15e2ee79e7cf29582ef770de7dab3d286431b01c3bb598f8e05e09601b890081"},
-    {file = "psycopg2_binary-2.9.6-cp311-cp311-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:dfa74c903a3c1f0d9b1c7e7b53ed2d929a4910e272add6700c38f365a6002820"},
-    {file = "psycopg2_binary-2.9.6-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:b83456c2d4979e08ff56180a76429263ea254c3f6552cd14ada95cff1dec9bb8"},
-    {file = "psycopg2_binary-2.9.6-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0645376d399bfd64da57148694d78e1f431b1e1ee1054872a5713125681cf1be"},
-    {file = "psycopg2_binary-2.9.6-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:e99e34c82309dd78959ba3c1590975b5d3c862d6f279f843d47d26ff89d7d7e1"},
-    {file = "psycopg2_binary-2.9.6-cp311-cp311-musllinux_1_1_i686.whl", hash = "sha256:4ea29fc3ad9d91162c52b578f211ff1c931d8a38e1f58e684c45aa470adf19e2"},
-    {file = "psycopg2_binary-2.9.6-cp311-cp311-musllinux_1_1_ppc64le.whl", hash = "sha256:4ac30da8b4f57187dbf449294d23b808f8f53cad6b1fc3623fa8a6c11d176dd0"},
-    {file = "psycopg2_binary-2.9.6-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:e78e6e2a00c223e164c417628572a90093c031ed724492c763721c2e0bc2a8df"},
-    {file = "psycopg2_binary-2.9.6-cp311-cp311-win32.whl", hash = "sha256:1876843d8e31c89c399e31b97d4b9725a3575bb9c2af92038464231ec40f9edb"},
-    {file = "psycopg2_binary-2.9.6-cp311-cp311-win_amd64.whl", hash = "sha256:b4b24f75d16a89cc6b4cdff0eb6a910a966ecd476d1e73f7ce5985ff1328e9a6"},
-    {file = "psycopg2_binary-2.9.6-cp36-cp36m-win32.whl", hash = "sha256:498807b927ca2510baea1b05cc91d7da4718a0f53cb766c154c417a39f1820a0"},
-    {file = "psycopg2_binary-2.9.6-cp36-cp36m-win_amd64.whl", hash = "sha256:0d236c2825fa656a2d98bbb0e52370a2e852e5a0ec45fc4f402977313329174d"},
-    {file = "psycopg2_binary-2.9.6-cp37-cp37m-macosx_10_9_x86_64.whl", hash = "sha256:34b9ccdf210cbbb1303c7c4db2905fa0319391bd5904d32689e6dd5c963d2ea8"},
-    {file = "psycopg2_binary-2.9.6-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:84d2222e61f313c4848ff05353653bf5f5cf6ce34df540e4274516880d9c3763"},
-    {file = "psycopg2_binary-2.9.6-cp37-cp37m-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:30637a20623e2a2eacc420059be11527f4458ef54352d870b8181a4c3020ae6b"},
-    {file = "psycopg2_binary-2.9.6-cp37-cp37m-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:8122cfc7cae0da9a3077216528b8bb3629c43b25053284cc868744bfe71eb141"},
-    {file = "psycopg2_binary-2.9.6-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:38601cbbfe600362c43714482f43b7c110b20cb0f8172422c616b09b85a750c5"},
-    {file = "psycopg2_binary-2.9.6-cp37-cp37m-musllinux_1_1_aarch64.whl", hash = "sha256:c7e62ab8b332147a7593a385d4f368874d5fe4ad4e341770d4983442d89603e3"},
-    {file = "psycopg2_binary-2.9.6-cp37-cp37m-musllinux_1_1_i686.whl", hash = "sha256:2ab652e729ff4ad76d400df2624d223d6e265ef81bb8aa17fbd63607878ecbee"},
-    {file = "psycopg2_binary-2.9.6-cp37-cp37m-musllinux_1_1_ppc64le.whl", hash = "sha256:c83a74b68270028dc8ee74d38ecfaf9c90eed23c8959fca95bd703d25b82c88e"},
-    {file = "psycopg2_binary-2.9.6-cp37-cp37m-musllinux_1_1_x86_64.whl", hash = "sha256:d4e6036decf4b72d6425d5b29bbd3e8f0ff1059cda7ac7b96d6ac5ed34ffbacd"},
-    {file = "psycopg2_binary-2.9.6-cp37-cp37m-win32.whl", hash = "sha256:a8c28fd40a4226b4a84bdf2d2b5b37d2c7bd49486b5adcc200e8c7ec991dfa7e"},
-    {file = "psycopg2_binary-2.9.6-cp37-cp37m-win_amd64.whl", hash = "sha256:51537e3d299be0db9137b321dfb6a5022caaab275775680e0c3d281feefaca6b"},
-    {file = "psycopg2_binary-2.9.6-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:cf4499e0a83b7b7edcb8dabecbd8501d0d3a5ef66457200f77bde3d210d5debb"},
-    {file = "psycopg2_binary-2.9.6-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:7e13a5a2c01151f1208d5207e42f33ba86d561b7a89fca67c700b9486a06d0e2"},
-    {file = "psycopg2_binary-2.9.6-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:0e0f754d27fddcfd74006455b6e04e6705d6c31a612ec69ddc040a5468e44b4e"},
-    {file = "psycopg2_binary-2.9.6-cp38-cp38-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:d57c3fd55d9058645d26ae37d76e61156a27722097229d32a9e73ed54819982a"},
-    {file = "psycopg2_binary-2.9.6-cp38-cp38-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:71f14375d6f73b62800530b581aed3ada394039877818b2d5f7fc77e3bb6894d"},
-    {file = "psycopg2_binary-2.9.6-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:441cc2f8869a4f0f4bb408475e5ae0ee1f3b55b33f350406150277f7f35384fc"},
-    {file = "psycopg2_binary-2.9.6-cp38-cp38-musllinux_1_1_aarch64.whl", hash = "sha256:65bee1e49fa6f9cf327ce0e01c4c10f39165ee76d35c846ade7cb0ec6683e303"},
-    {file = "psycopg2_binary-2.9.6-cp38-cp38-musllinux_1_1_i686.whl", hash = "sha256:af335bac6b666cc6aea16f11d486c3b794029d9df029967f9938a4bed59b6a19"},
-    {file = "psycopg2_binary-2.9.6-cp38-cp38-musllinux_1_1_ppc64le.whl", hash = "sha256:cfec476887aa231b8548ece2e06d28edc87c1397ebd83922299af2e051cf2827"},
-    {file = "psycopg2_binary-2.9.6-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:65c07febd1936d63bfde78948b76cd4c2a411572a44ac50719ead41947d0f26b"},
-    {file = "psycopg2_binary-2.9.6-cp38-cp38-win32.whl", hash = "sha256:4dfb4be774c4436a4526d0c554af0cc2e02082c38303852a36f6456ece7b3503"},
-    {file = "psycopg2_binary-2.9.6-cp38-cp38-win_amd64.whl", hash = "sha256:02c6e3cf3439e213e4ee930308dc122d6fb4d4bea9aef4a12535fbd605d1a2fe"},
-    {file = "psycopg2_binary-2.9.6-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:e9182eb20f41417ea1dd8e8f7888c4d7c6e805f8a7c98c1081778a3da2bee3e4"},
-    {file = "psycopg2_binary-2.9.6-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:8a6979cf527e2603d349a91060f428bcb135aea2be3201dff794813256c274f1"},
-    {file = "psycopg2_binary-2.9.6-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8338a271cb71d8da40b023a35d9c1e919eba6cbd8fa20a54b748a332c355d896"},
-    {file = "psycopg2_binary-2.9.6-cp39-cp39-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:e3ed340d2b858d6e6fb5083f87c09996506af483227735de6964a6100b4e6a54"},
-    {file = "psycopg2_binary-2.9.6-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:f81e65376e52f03422e1fb475c9514185669943798ed019ac50410fb4c4df232"},
-    {file = "psycopg2_binary-2.9.6-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:bfb13af3c5dd3a9588000910178de17010ebcccd37b4f9794b00595e3a8ddad3"},
-    {file = "psycopg2_binary-2.9.6-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:4c727b597c6444a16e9119386b59388f8a424223302d0c06c676ec8b4bc1f963"},
-    {file = "psycopg2_binary-2.9.6-cp39-cp39-musllinux_1_1_i686.whl", hash = "sha256:4d67fbdaf177da06374473ef6f7ed8cc0a9dc640b01abfe9e8a2ccb1b1402c1f"},
-    {file = "psycopg2_binary-2.9.6-cp39-cp39-musllinux_1_1_ppc64le.whl", hash = "sha256:0892ef645c2fabb0c75ec32d79f4252542d0caec1d5d949630e7d242ca4681a3"},
-    {file = "psycopg2_binary-2.9.6-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:02c0f3757a4300cf379eb49f543fb7ac527fb00144d39246ee40e1df684ab514"},
-    {file = "psycopg2_binary-2.9.6-cp39-cp39-win32.whl", hash = "sha256:c3dba7dab16709a33a847e5cd756767271697041fbe3fe97c215b1fc1f5c9848"},
-    {file = "psycopg2_binary-2.9.6-cp39-cp39-win_amd64.whl", hash = "sha256:f6a88f384335bb27812293fdb11ac6aee2ca3f51d3c7820fe03de0a304ab6249"},
+    {file = "psycopg2-binary-2.9.9.tar.gz", hash = "sha256:7f01846810177d829c7692f1f5ada8096762d9172af1b1a28d4ab5b77c923c1c"},
+    {file = "psycopg2_binary-2.9.9-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:c2470da5418b76232f02a2fcd2229537bb2d5a7096674ce61859c3229f2eb202"},
+    {file = "psycopg2_binary-2.9.9-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:c6af2a6d4b7ee9615cbb162b0738f6e1fd1f5c3eda7e5da17861eacf4c717ea7"},
+    {file = "psycopg2_binary-2.9.9-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:75723c3c0fbbf34350b46a3199eb50638ab22a0228f93fb472ef4d9becc2382b"},
+    {file = "psycopg2_binary-2.9.9-cp310-cp310-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:83791a65b51ad6ee6cf0845634859d69a038ea9b03d7b26e703f94c7e93dbcf9"},
+    {file = "psycopg2_binary-2.9.9-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:0ef4854e82c09e84cc63084a9e4ccd6d9b154f1dbdd283efb92ecd0b5e2b8c84"},
+    {file = "psycopg2_binary-2.9.9-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ed1184ab8f113e8d660ce49a56390ca181f2981066acc27cf637d5c1e10ce46e"},
+    {file = "psycopg2_binary-2.9.9-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:d2997c458c690ec2bc6b0b7ecbafd02b029b7b4283078d3b32a852a7ce3ddd98"},
+    {file = "psycopg2_binary-2.9.9-cp310-cp310-musllinux_1_1_i686.whl", hash = "sha256:b58b4710c7f4161b5e9dcbe73bb7c62d65670a87df7bcce9e1faaad43e715245"},
+    {file = "psycopg2_binary-2.9.9-cp310-cp310-musllinux_1_1_ppc64le.whl", hash = "sha256:0c009475ee389757e6e34611d75f6e4f05f0cf5ebb76c6037508318e1a1e0d7e"},
+    {file = "psycopg2_binary-2.9.9-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:8dbf6d1bc73f1d04ec1734bae3b4fb0ee3cb2a493d35ede9badbeb901fb40f6f"},
+    {file = "psycopg2_binary-2.9.9-cp310-cp310-win32.whl", hash = "sha256:3f78fd71c4f43a13d342be74ebbc0666fe1f555b8837eb113cb7416856c79682"},
+    {file = "psycopg2_binary-2.9.9-cp310-cp310-win_amd64.whl", hash = "sha256:876801744b0dee379e4e3c38b76fc89f88834bb15bf92ee07d94acd06ec890a0"},
+    {file = "psycopg2_binary-2.9.9-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:ee825e70b1a209475622f7f7b776785bd68f34af6e7a46e2e42f27b659b5bc26"},
+    {file = "psycopg2_binary-2.9.9-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:1ea665f8ce695bcc37a90ee52de7a7980be5161375d42a0b6c6abedbf0d81f0f"},
+    {file = "psycopg2_binary-2.9.9-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:143072318f793f53819048fdfe30c321890af0c3ec7cb1dfc9cc87aa88241de2"},
+    {file = "psycopg2_binary-2.9.9-cp311-cp311-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:c332c8d69fb64979ebf76613c66b985414927a40f8defa16cf1bc028b7b0a7b0"},
+    {file = "psycopg2_binary-2.9.9-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:f7fc5a5acafb7d6ccca13bfa8c90f8c51f13d8fb87d95656d3950f0158d3ce53"},
+    {file = "psycopg2_binary-2.9.9-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:977646e05232579d2e7b9c59e21dbe5261f403a88417f6a6512e70d3f8a046be"},
+    {file = "psycopg2_binary-2.9.9-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:b6356793b84728d9d50ead16ab43c187673831e9d4019013f1402c41b1db9b27"},
+    {file = "psycopg2_binary-2.9.9-cp311-cp311-musllinux_1_1_i686.whl", hash = "sha256:bc7bb56d04601d443f24094e9e31ae6deec9ccb23581f75343feebaf30423359"},
+    {file = "psycopg2_binary-2.9.9-cp311-cp311-musllinux_1_1_ppc64le.whl", hash = "sha256:77853062a2c45be16fd6b8d6de2a99278ee1d985a7bd8b103e97e41c034006d2"},
+    {file = "psycopg2_binary-2.9.9-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:78151aa3ec21dccd5cdef6c74c3e73386dcdfaf19bced944169697d7ac7482fc"},
+    {file = "psycopg2_binary-2.9.9-cp311-cp311-win32.whl", hash = "sha256:dc4926288b2a3e9fd7b50dc6a1909a13bbdadfc67d93f3374d984e56f885579d"},
+    {file = "psycopg2_binary-2.9.9-cp311-cp311-win_amd64.whl", hash = "sha256:b76bedd166805480ab069612119ea636f5ab8f8771e640ae103e05a4aae3e417"},
+    {file = "psycopg2_binary-2.9.9-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:8532fd6e6e2dc57bcb3bc90b079c60de896d2128c5d9d6f24a63875a95a088cf"},
+    {file = "psycopg2_binary-2.9.9-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8f8544b092a29a6ddd72f3556a9fcf249ec412e10ad28be6a0c0d948924f2212"},
+    {file = "psycopg2_binary-2.9.9-cp312-cp312-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:2d423c8d8a3c82d08fe8af900ad5b613ce3632a1249fd6a223941d0735fce493"},
+    {file = "psycopg2_binary-2.9.9-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:2e5afae772c00980525f6d6ecf7cbca55676296b580c0e6abb407f15f3706996"},
+    {file = "psycopg2_binary-2.9.9-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:6e6f98446430fdf41bd36d4faa6cb409f5140c1c2cf58ce0bbdaf16af7d3f119"},
+    {file = "psycopg2_binary-2.9.9-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:c77e3d1862452565875eb31bdb45ac62502feabbd53429fdc39a1cc341d681ba"},
+    {file = "psycopg2_binary-2.9.9-cp312-cp312-musllinux_1_1_i686.whl", hash = "sha256:cb16c65dcb648d0a43a2521f2f0a2300f40639f6f8c1ecbc662141e4e3e1ee07"},
+    {file = "psycopg2_binary-2.9.9-cp312-cp312-musllinux_1_1_ppc64le.whl", hash = "sha256:911dda9c487075abd54e644ccdf5e5c16773470a6a5d3826fda76699410066fb"},
+    {file = "psycopg2_binary-2.9.9-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:57fede879f08d23c85140a360c6a77709113efd1c993923c59fde17aa27599fe"},
+    {file = "psycopg2_binary-2.9.9-cp37-cp37m-macosx_10_9_x86_64.whl", hash = "sha256:2293b001e319ab0d869d660a704942c9e2cce19745262a8aba2115ef41a0a42a"},
+    {file = "psycopg2_binary-2.9.9-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:03ef7df18daf2c4c07e2695e8cfd5ee7f748a1d54d802330985a78d2a5a6dca9"},
+    {file = "psycopg2_binary-2.9.9-cp37-cp37m-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:0a602ea5aff39bb9fac6308e9c9d82b9a35c2bf288e184a816002c9fae930b77"},
+    {file = "psycopg2_binary-2.9.9-cp37-cp37m-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:8359bf4791968c5a78c56103702000105501adb557f3cf772b2c207284273984"},
+    {file = "psycopg2_binary-2.9.9-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:275ff571376626195ab95a746e6a04c7df8ea34638b99fc11160de91f2fef503"},
+    {file = "psycopg2_binary-2.9.9-cp37-cp37m-musllinux_1_1_aarch64.whl", hash = "sha256:f9b5571d33660d5009a8b3c25dc1db560206e2d2f89d3df1cb32d72c0d117d52"},
+    {file = "psycopg2_binary-2.9.9-cp37-cp37m-musllinux_1_1_i686.whl", hash = "sha256:420f9bbf47a02616e8554e825208cb947969451978dceb77f95ad09c37791dae"},
+    {file = "psycopg2_binary-2.9.9-cp37-cp37m-musllinux_1_1_ppc64le.whl", hash = "sha256:4154ad09dac630a0f13f37b583eae260c6aa885d67dfbccb5b02c33f31a6d420"},
+    {file = "psycopg2_binary-2.9.9-cp37-cp37m-musllinux_1_1_x86_64.whl", hash = "sha256:a148c5d507bb9b4f2030a2025c545fccb0e1ef317393eaba42e7eabd28eb6041"},
+    {file = "psycopg2_binary-2.9.9-cp37-cp37m-win32.whl", hash = "sha256:68fc1f1ba168724771e38bee37d940d2865cb0f562380a1fb1ffb428b75cb692"},
+    {file = "psycopg2_binary-2.9.9-cp37-cp37m-win_amd64.whl", hash = "sha256:281309265596e388ef483250db3640e5f414168c5a67e9c665cafce9492eda2f"},
+    {file = "psycopg2_binary-2.9.9-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:60989127da422b74a04345096c10d416c2b41bd7bf2a380eb541059e4e999980"},
+    {file = "psycopg2_binary-2.9.9-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:246b123cc54bb5361588acc54218c8c9fb73068bf227a4a531d8ed56fa3ca7d6"},
+    {file = "psycopg2_binary-2.9.9-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:34eccd14566f8fe14b2b95bb13b11572f7c7d5c36da61caf414d23b91fcc5d94"},
+    {file = "psycopg2_binary-2.9.9-cp38-cp38-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:18d0ef97766055fec15b5de2c06dd8e7654705ce3e5e5eed3b6651a1d2a9a152"},
+    {file = "psycopg2_binary-2.9.9-cp38-cp38-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:d3f82c171b4ccd83bbaf35aa05e44e690113bd4f3b7b6cc54d2219b132f3ae55"},
+    {file = "psycopg2_binary-2.9.9-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ead20f7913a9c1e894aebe47cccf9dc834e1618b7aa96155d2091a626e59c972"},
+    {file = "psycopg2_binary-2.9.9-cp38-cp38-musllinux_1_1_aarch64.whl", hash = "sha256:ca49a8119c6cbd77375ae303b0cfd8c11f011abbbd64601167ecca18a87e7cdd"},
+    {file = "psycopg2_binary-2.9.9-cp38-cp38-musllinux_1_1_i686.whl", hash = "sha256:323ba25b92454adb36fa425dc5cf6f8f19f78948cbad2e7bc6cdf7b0d7982e59"},
+    {file = "psycopg2_binary-2.9.9-cp38-cp38-musllinux_1_1_ppc64le.whl", hash = "sha256:1236ed0952fbd919c100bc839eaa4a39ebc397ed1c08a97fc45fee2a595aa1b3"},
+    {file = "psycopg2_binary-2.9.9-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:729177eaf0aefca0994ce4cffe96ad3c75e377c7b6f4efa59ebf003b6d398716"},
+    {file = "psycopg2_binary-2.9.9-cp38-cp38-win32.whl", hash = "sha256:804d99b24ad523a1fe18cc707bf741670332f7c7412e9d49cb5eab67e886b9b5"},
+    {file = "psycopg2_binary-2.9.9-cp38-cp38-win_amd64.whl", hash = "sha256:a6cdcc3ede532f4a4b96000b6362099591ab4a3e913d70bcbac2b56c872446f7"},
+    {file = "psycopg2_binary-2.9.9-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:72dffbd8b4194858d0941062a9766f8297e8868e1dd07a7b36212aaa90f49472"},
+    {file = "psycopg2_binary-2.9.9-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:30dcc86377618a4c8f3b72418df92e77be4254d8f89f14b8e8f57d6d43603c0f"},
+    {file = "psycopg2_binary-2.9.9-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:31a34c508c003a4347d389a9e6fcc2307cc2150eb516462a7a17512130de109e"},
+    {file = "psycopg2_binary-2.9.9-cp39-cp39-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:15208be1c50b99203fe88d15695f22a5bed95ab3f84354c494bcb1d08557df67"},
+    {file = "psycopg2_binary-2.9.9-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:1873aade94b74715be2246321c8650cabf5a0d098a95bab81145ffffa4c13876"},
+    {file = "psycopg2_binary-2.9.9-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:3a58c98a7e9c021f357348867f537017057c2ed7f77337fd914d0bedb35dace7"},
+    {file = "psycopg2_binary-2.9.9-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:4686818798f9194d03c9129a4d9a702d9e113a89cb03bffe08c6cf799e053291"},
+    {file = "psycopg2_binary-2.9.9-cp39-cp39-musllinux_1_1_i686.whl", hash = "sha256:ebdc36bea43063116f0486869652cb2ed7032dbc59fbcb4445c4862b5c1ecf7f"},
+    {file = "psycopg2_binary-2.9.9-cp39-cp39-musllinux_1_1_ppc64le.whl", hash = "sha256:ca08decd2697fdea0aea364b370b1249d47336aec935f87b8bbfd7da5b2ee9c1"},
+    {file = "psycopg2_binary-2.9.9-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:ac05fb791acf5e1a3e39402641827780fe44d27e72567a000412c648a85ba860"},
+    {file = "psycopg2_binary-2.9.9-cp39-cp39-win32.whl", hash = "sha256:9dba73be7305b399924709b91682299794887cbbd88e38226ed9f6712eabee90"},
+    {file = "psycopg2_binary-2.9.9-cp39-cp39-win_amd64.whl", hash = "sha256:f7ae5d65ccfbebdfa761585228eb4d0df3a8b15cfb53bd953e713e09fbb12957"},
 ]
 
 [[package]]
@@ -2577,7 +2584,6 @@ files = [
     {file = "PyYAML-6.0.1-cp311-cp311-win_amd64.whl", hash = "sha256:bf07ee2fef7014951eeb99f56f39c9bb4af143d8aa3c21b1677805985307da34"},
     {file = "PyYAML-6.0.1-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:855fb52b0dc35af121542a76b9a84f8d1cd886ea97c84703eaa6d88e37a2ad28"},
     {file = "PyYAML-6.0.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:40df9b996c2b73138957fe23a16a4f0ba614f4c0efce1e9406a184b6d07fa3a9"},
-    {file = "PyYAML-6.0.1-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a08c6f0fe150303c1c6b71ebcd7213c2858041a7e01975da3a99aed1e7a378ef"},
     {file = "PyYAML-6.0.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:6c22bec3fbe2524cde73d7ada88f6566758a8f7227bfbf93a408a9d86bcc12a0"},
     {file = "PyYAML-6.0.1-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:8d4e9c88387b0f5c7d5f281e55304de64cf7f9c0021a3525bd3b1c542da3b0e4"},
     {file = "PyYAML-6.0.1-cp312-cp312-win32.whl", hash = "sha256:d483d2cdf104e7c9fa60c544d92981f12ad66a457afae824d146093b8c294c54"},
@@ -2702,13 +2708,13 @@ files = [
 
 [[package]]
 name = "requests"
-version = "2.32.0"
+version = "2.32.3"
 description = "Python HTTP for Humans."
 optional = false
 python-versions = ">=3.8"
 files = [
-    {file = "requests-2.32.0-py3-none-any.whl", hash = "sha256:f2c3881dddb70d056c5bd7600a4fae312b2a300e39be6a118d30b90bd27262b5"},
-    {file = "requests-2.32.0.tar.gz", hash = "sha256:fa5490319474c82ef1d2c9bc459d3652e3ae4ef4c4ebdd18a21145a47ca4b6b8"},
+    {file = "requests-2.32.3-py3-none-any.whl", hash = "sha256:70761cfe03c773ceb22aa2f671b4757976145175cdfca038c02654d061d6dcc6"},
+    {file = "requests-2.32.3.tar.gz", hash = "sha256:55365417734eb18255590a9ff9eb97e9e1da868d4ccd6402399eaf68af20a760"},
 ]
 
 [package.dependencies]
@@ -3131,16 +3137,6 @@ files = [
     {file = "wrapt-1.14.1-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:8ad85f7f4e20964db4daadcab70b47ab05c7c1cf2a7c1e51087bfaa83831854c"},
     {file = "wrapt-1.14.1-cp310-cp310-win32.whl", hash = "sha256:a9a52172be0b5aae932bef82a79ec0a0ce87288c7d132946d645eba03f0ad8a8"},
     {file = "wrapt-1.14.1-cp310-cp310-win_amd64.whl", hash = "sha256:6d323e1554b3d22cfc03cd3243b5bb815a51f5249fdcbb86fda4bf62bab9e164"},
-    {file = "wrapt-1.14.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:ecee4132c6cd2ce5308e21672015ddfed1ff975ad0ac8d27168ea82e71413f55"},
-    {file = "wrapt-1.14.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:2020f391008ef874c6d9e208b24f28e31bcb85ccff4f335f15a3251d222b92d9"},
-    {file = "wrapt-1.14.1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:2feecf86e1f7a86517cab34ae6c2f081fd2d0dac860cb0c0ded96d799d20b335"},
-    {file = "wrapt-1.14.1-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:240b1686f38ae665d1b15475966fe0472f78e71b1b4903c143a842659c8e4cb9"},
-    {file = "wrapt-1.14.1-cp311-cp311-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a9008dad07d71f68487c91e96579c8567c98ca4c3881b9b113bc7b33e9fd78b8"},
-    {file = "wrapt-1.14.1-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:6447e9f3ba72f8e2b985a1da758767698efa72723d5b59accefd716e9e8272bf"},
-    {file = "wrapt-1.14.1-cp311-cp311-musllinux_1_1_i686.whl", hash = "sha256:acae32e13a4153809db37405f5eba5bac5fbe2e2ba61ab227926a22901051c0a"},
-    {file = "wrapt-1.14.1-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:49ef582b7a1152ae2766557f0550a9fcbf7bbd76f43fbdc94dd3bf07cc7168be"},
-    {file = "wrapt-1.14.1-cp311-cp311-win32.whl", hash = "sha256:358fe87cc899c6bb0ddc185bf3dbfa4ba646f05b1b0b9b5a27c2cb92c2cea204"},
-    {file = "wrapt-1.14.1-cp311-cp311-win_amd64.whl", hash = "sha256:26046cd03936ae745a502abf44dac702a5e6880b2b01c29aea8ddf3353b68224"},
     {file = "wrapt-1.14.1-cp35-cp35m-manylinux1_i686.whl", hash = "sha256:43ca3bbbe97af00f49efb06e352eae40434ca9d915906f77def219b88e85d907"},
     {file = "wrapt-1.14.1-cp35-cp35m-manylinux1_x86_64.whl", hash = "sha256:6b1a564e6cb69922c7fe3a678b9f9a3c54e72b469875aa8018f18b4d1dd1adf3"},
     {file = "wrapt-1.14.1-cp35-cp35m-manylinux2010_i686.whl", hash = "sha256:00b6d4ea20a906c0ca56d84f93065b398ab74b927a7a3dbd470f6fc503f95dc3"},
@@ -3378,4 +3374,4 @@ cffi = ["cffi (>=1.11)"]
 [metadata]
 lock-version = "2.0"
 python-versions = "^3.9"
-content-hash = "c09bcb333ab550958b33dbf4fec968c500d8e701fd4c96402cddbd9bb8048055"
+content-hash = "9055b73352f1534f664cd8af6ebf8d93cf3bf857f115756f312ff2e3ae1bbbc1"