pq bench: avoid repeated conversion to_i128

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

.github/workflows/build_and_test.yml

@@ -847,7 +847,7 @@ jobs:
       run:
         shell: sh -eu {0}
     env:
-      VM_BUILDER_VERSION: v0.18.5
+      VM_BUILDER_VERSION: v0.18.2
 
     steps:
       - name: Checkout

.github/workflows/release.yml

@@ -2,7 +2,7 @@ name: Create Release Branch
 
 on:
   schedule:
-    - cron: '0 7 * * 5'
+    - cron: '0 7 * * 2'
   workflow_dispatch:
 
 jobs:

Cargo.lock

@@ -1609,6 +1609,16 @@ dependencies = [
  "subtle",
 ]
 
+[[package]]
+name = "ctor"
+version = "0.1.26"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6d2301688392eb071b0bf1a37be05c469d3cc4dbbd95df672fe28ab021e6a096"
+dependencies = [
+ "quote",
+ "syn 1.0.109",
+]
+
 [[package]]
 name = "ctr"
 version = "0.6.0"
@@ -2704,10 +2714,11 @@ dependencies = [
 
 [[package]]
 name = "log"
-version = "0.4.20"
+version = "0.4.17"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b5e6163cb8c49088c2c36f57875e58ccd8c87c7427f7fbd50ea6710b2f3f2e8f"
+checksum = "abb12e687cfb44aa40f41fc3978ef76448f9b6038cad6aef4259d3c095a2382e"
 dependencies = [
+ "cfg-if",
  "value-bag",
 ]
 
@@ -2921,6 +2932,16 @@ dependencies = [
  "winapi",
 ]
 
+[[package]]
+name = "nu-ansi-term"
+version = "0.46.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "77a8165726e8236064dbb45459242600304b42a5ea24ee2948e18e023bf7ba84"
+dependencies = [
+ "overload",
+ "winapi",
+]
+
 [[package]]
 name = "num-bigint"
 version = "0.4.3"
@@ -3187,6 +3208,12 @@ version = "0.5.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "4030760ffd992bef45b0ae3f10ce1aba99e33464c90d14dd7c039884963ddc7a"
 
+[[package]]
+name = "overload"
+version = "0.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b15813163c1d831bf4a13c3610c05c0d03b39feb07f7e09fa234dac9b15aaf39"
+
 [[package]]
 name = "pagectl"
 version = "0.1.0"
@@ -3272,10 +3299,12 @@ dependencies = [
  "tokio",
  "tokio-io-timeout",
  "tokio-postgres",
+ "tokio-stream",
  "tokio-tar",
  "tokio-util",
  "toml_edit",
  "tracing",
+ "tracing-subscriber",
  "url",
  "utils",
  "walkdir",
@@ -3550,7 +3579,7 @@ dependencies = [
 [[package]]
 name = "postgres"
 version = "0.19.4"
-source = "git+https://github.com/neondatabase/rust-postgres.git?rev=ce7260db5998fe27167da42503905a12e7ad9048#ce7260db5998fe27167da42503905a12e7ad9048"
+source = "git+https://github.com/neondatabase/rust-postgres.git?branch=problame/copy-both-duplex-public#5c462bd3500e657c014ef087e4eef2c1a8f0ebda"
 dependencies = [
  "bytes",
  "fallible-iterator",
@@ -3563,7 +3592,7 @@ dependencies = [
 [[package]]
 name = "postgres-native-tls"
 version = "0.5.0"
-source = "git+https://github.com/neondatabase/rust-postgres.git?rev=ce7260db5998fe27167da42503905a12e7ad9048#ce7260db5998fe27167da42503905a12e7ad9048"
+source = "git+https://github.com/neondatabase/rust-postgres.git?branch=problame/copy-both-duplex-public#5c462bd3500e657c014ef087e4eef2c1a8f0ebda"
 dependencies = [
  "native-tls",
  "tokio",
@@ -3574,7 +3603,7 @@ dependencies = [
 [[package]]
 name = "postgres-protocol"
 version = "0.6.4"
-source = "git+https://github.com/neondatabase/rust-postgres.git?rev=ce7260db5998fe27167da42503905a12e7ad9048#ce7260db5998fe27167da42503905a12e7ad9048"
+source = "git+https://github.com/neondatabase/rust-postgres.git?branch=problame/copy-both-duplex-public#5c462bd3500e657c014ef087e4eef2c1a8f0ebda"
 dependencies = [
  "base64 0.20.0",
  "byteorder",
@@ -3592,7 +3621,7 @@ dependencies = [
 [[package]]
 name = "postgres-types"
 version = "0.2.4"
-source = "git+https://github.com/neondatabase/rust-postgres.git?rev=ce7260db5998fe27167da42503905a12e7ad9048#ce7260db5998fe27167da42503905a12e7ad9048"
+source = "git+https://github.com/neondatabase/rust-postgres.git?branch=problame/copy-both-duplex-public#5c462bd3500e657c014ef087e4eef2c1a8f0ebda"
 dependencies = [
  "bytes",
  "fallible-iterator",
@@ -4419,7 +4448,6 @@ dependencies = [
  "itertools",
  "pageserver",
  "rand 0.8.5",
- "remote_storage",
  "reqwest",
  "serde",
  "serde_json",
@@ -4478,7 +4506,6 @@ dependencies = [
  "tokio",
  "tokio-io-timeout",
  "tokio-postgres",
- "tokio-stream",
  "toml_edit",
  "tracing",
  "url",
@@ -4681,16 +4708,6 @@ dependencies = [
  "serde_derive",
 ]
 
-[[package]]
-name = "serde_assert"
-version = "0.5.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "eda563240c1288b044209be1f0d38bb4d15044fb3e00dc354fbc922ab4733e80"
-dependencies = [
- "hashbrown 0.13.2",
- "serde",
-]
-
 [[package]]
 name = "serde_derive"
 version = "1.0.183"
@@ -5408,7 +5425,7 @@ dependencies = [
 [[package]]
 name = "tokio-postgres"
 version = "0.7.7"
-source = "git+https://github.com/neondatabase/rust-postgres.git?rev=ce7260db5998fe27167da42503905a12e7ad9048#ce7260db5998fe27167da42503905a12e7ad9048"
+source = "git+https://github.com/neondatabase/rust-postgres.git?branch=problame/copy-both-duplex-public#5c462bd3500e657c014ef087e4eef2c1a8f0ebda"
 dependencies = [
  "async-trait",
  "byteorder",
@@ -5765,6 +5782,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "30a651bc37f915e81f087d86e62a18eec5f79550c7faff886f7090b4ea757c77"
 dependencies = [
  "matchers",
+ "nu-ansi-term",
  "once_cell",
  "regex",
  "serde",
@@ -5977,7 +5995,6 @@ dependencies = [
  "routerify",
  "sentry",
  "serde",
- "serde_assert",
  "serde_json",
  "serde_with",
  "signal-hook",
@@ -6013,9 +6030,13 @@ checksum = "830b7e5d4d90034032940e4ace0d9a9a057e7a45cd94e6c007832e39edb82f6d"
 
 [[package]]
 name = "value-bag"
-version = "1.4.2"
+version = "1.0.0-alpha.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4a72e1902dde2bd6441347de2b70b7f5d59bf157c6c62f0c44572607a1d55bbe"
+checksum = "2209b78d1249f7e6f3293657c9779fe31ced465df091bbd433a1cf88e916ec55"
+dependencies = [
+ "ctor",
+ "version_check",
+]
 
 [[package]]
 name = "vcpkg"

Cargo.toml

@@ -124,7 +124,6 @@ sentry = { version = "0.31", default-features = false, features = ["backtrace",
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1"
 serde_with = "2.0"
-serde_assert = "0.5.0"
 sha2 = "0.10.2"
 signal-hook = "0.3"
 smallvec = "1.11"
@@ -162,11 +161,11 @@ env_logger = "0.10"
 log = "0.4"
 
 ## Libraries from neondatabase/ git forks, ideally with changes to be upstreamed
-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="ce7260db5998fe27167da42503905a12e7ad9048" }
-postgres-native-tls = { git = "https://github.com/neondatabase/rust-postgres.git", rev="ce7260db5998fe27167da42503905a12e7ad9048" }
-postgres-protocol = { git = "https://github.com/neondatabase/rust-postgres.git", rev="ce7260db5998fe27167da42503905a12e7ad9048" }
-postgres-types = { git = "https://github.com/neondatabase/rust-postgres.git", rev="ce7260db5998fe27167da42503905a12e7ad9048" }
-tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="ce7260db5998fe27167da42503905a12e7ad9048" }
+postgres = { git = "https://github.com/neondatabase/rust-postgres.git", branch="problame/copy-both-duplex-public" }
+postgres-native-tls = { git = "https://github.com/neondatabase/rust-postgres.git", branch="problame/copy-both-duplex-public" }
+postgres-protocol = { git = "https://github.com/neondatabase/rust-postgres.git", branch="problame/copy-both-duplex-public" }
+postgres-types = { git = "https://github.com/neondatabase/rust-postgres.git", branch="problame/copy-both-duplex-public" }
+tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", branch="problame/copy-both-duplex-public" }
 
 ## Other git libraries
 heapless = { default-features=false, features=[], git = "https://github.com/japaric/heapless.git", rev = "644653bf3b831c6bb4963be2de24804acf5e5001" } # upstream release pending
@@ -203,7 +202,7 @@ tonic-build = "0.9"
 
 # This is only needed for proxy's tests.
 # TODO: we should probably fork `tokio-postgres-rustls` instead.
-tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="ce7260db5998fe27167da42503905a12e7ad9048" }
+tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", branch="problame/copy-both-duplex-public" }
 
 ################# Binary contents sections
 

compute_tools/src/bin/compute_ctl.rs

@@ -156,7 +156,6 @@ fn main() -> Result<()> {
                 let path = Path::new(sp);
                 let file = File::open(path)?;
                 spec = Some(serde_json::from_reader(file)?);
-                live_config_allowed = true;
             } else if let Some(id) = compute_id {
                 if let Some(cp_base) = control_plane_uri {
                     live_config_allowed = true;
@@ -278,26 +277,32 @@ fn main() -> Result<()> {
         if #[cfg(target_os = "linux")] {
             use std::env;
             use tokio_util::sync::CancellationToken;
-            let vm_monitor_addr = matches
-                .get_one::<String>("vm-monitor-addr")
-                .expect("--vm-monitor-addr should always be set because it has a default arg");
+            use tracing::warn;
+            let vm_monitor_addr = matches.get_one::<String>("vm-monitor-addr");
             let file_cache_connstr = matches.get_one::<String>("filecache-connstr");
             let cgroup = matches.get_one::<String>("cgroup");
+            let file_cache_on_disk = matches.get_flag("file-cache-on-disk");
 
             // Only make a runtime if we need to.
             // Note: it seems like you can make a runtime in an inner scope and
             // if you start a task in it it won't be dropped. However, make it
             // in the outermost scope just to be safe.
-            let rt = if env::var_os("AUTOSCALING").is_some() {
-                Some(
+            let rt = match (env::var_os("AUTOSCALING"), vm_monitor_addr) {
+                (None, None) => None,
+                (None, Some(_)) => {
+                    warn!("--vm-monitor-addr option set but AUTOSCALING env var not present");
+                    None
+                }
+                (Some(_), None) => {
+                    panic!("AUTOSCALING env var present but --vm-monitor-addr option not set")
+                }
+                (Some(_), Some(_)) => Some(
                     tokio::runtime::Builder::new_multi_thread()
                         .worker_threads(4)
                         .enable_all()
                         .build()
-                        .expect("failed to create tokio runtime for monitor")
-                )
-            } else {
-                None
+                        .expect("failed to create tokio runtime for monitor"),
+                ),
             };
 
             // This token is used internally by the monitor to clean up all threads
@@ -308,7 +313,8 @@ fn main() -> Result<()> {
                     Box::leak(Box::new(vm_monitor::Args {
                         cgroup: cgroup.cloned(),
                         pgconnstr: file_cache_connstr.cloned(),
-                        addr: vm_monitor_addr.clone(),
+                        addr: vm_monitor_addr.cloned().unwrap(),
+                        file_cache_on_disk,
                     })),
                     token.clone(),
                 ))
@@ -480,8 +486,6 @@ fn cli() -> clap::Command {
                 .value_name("FILECACHE_CONNSTR"),
         )
         .arg(
-            // DEPRECATED, NO LONGER DOES ANYTHING.
-            // See https://github.com/neondatabase/cloud/issues/7516
             Arg::new("file-cache-on-disk")
                 .long("file-cache-on-disk")
                 .action(clap::ArgAction::SetTrue),

compute_tools/src/pg_helpers.rs

@@ -193,16 +193,11 @@ impl Escaping for PgIdent {
 /// Build a list of existing Postgres roles
 pub fn get_existing_roles(xact: &mut Transaction<'_>) -> Result<Vec<Role>> {
     let postgres_roles = xact
-        .query(
-            "SELECT rolname, rolpassword, rolreplication, rolbypassrls FROM pg_catalog.pg_authid",
-            &[],
-        )?
+        .query("SELECT rolname, rolpassword FROM pg_catalog.pg_authid", &[])?
         .iter()
         .map(|row| Role {
             name: row.get("rolname"),
             encrypted_password: row.get("rolpassword"),
-            replication: Some(row.get("rolreplication")),
-            bypassrls: Some(row.get("rolbypassrls")),
             options: None,
         })
         .collect();

compute_tools/src/spec.rs

@@ -24,7 +24,7 @@ fn do_control_plane_request(
 ) -> Result<ControlPlaneSpecResponse, (bool, String)> {
     let resp = reqwest::blocking::Client::new()
         .get(uri)
-        .header("Authorization", format!("Bearer {}", jwt))
+        .header("Authorization", jwt)
         .send()
         .map_err(|e| {
             (
@@ -68,7 +68,7 @@ pub fn get_spec_from_control_plane(
     base_uri: &str,
     compute_id: &str,
 ) -> Result<Option<ComputeSpec>> {
-    let cp_uri = format!("{base_uri}/compute/api/v2/computes/{compute_id}/spec");
+    let cp_uri = format!("{base_uri}/management/api/v2/computes/{compute_id}/spec");
     let jwt: String = match std::env::var("NEON_CONTROL_PLANE_TOKEN") {
         Ok(v) => v,
         Err(_) => "".to_string(),
@@ -265,8 +265,6 @@ pub fn handle_roles(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
         let action = if let Some(r) = pg_role {
             if (r.encrypted_password.is_none() && role.encrypted_password.is_some())
                 || (r.encrypted_password.is_some() && role.encrypted_password.is_none())
-                || !r.bypassrls.unwrap_or(false)
-                || !r.replication.unwrap_or(false)
             {
                 RoleAction::Update
             } else if let Some(pg_pwd) = &r.encrypted_password {
@@ -298,8 +296,7 @@ pub fn handle_roles(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
         match action {
             RoleAction::None => {}
             RoleAction::Update => {
-                let mut query: String =
-                    format!("ALTER ROLE {} BYPASSRLS REPLICATION", name.pg_quote());
+                let mut query: String = format!("ALTER ROLE {} ", name.pg_quote());
                 query.push_str(&role.to_pg_options());
                 xact.execute(query.as_str(), &[])?;
             }

control_plane/src/attachment_service.rs

@@ -2,6 +2,7 @@ use crate::{background_process, local_env::LocalEnv};
 use anyhow::anyhow;
 use camino::Utf8PathBuf;
 use serde::{Deserialize, Serialize};
+use serde_with::{serde_as, DisplayFromStr};
 use std::{path::PathBuf, process::Child};
 use utils::id::{NodeId, TenantId};
 
@@ -13,10 +14,12 @@ pub struct AttachmentService {
 
 const COMMAND: &str = "attachment_service";
 
+#[serde_as]
 #[derive(Serialize, Deserialize)]
 pub struct AttachHookRequest {
+    #[serde_as(as = "DisplayFromStr")]
     pub tenant_id: TenantId,
-    pub node_id: Option<NodeId>,
+    pub pageserver_id: Option<NodeId>,
 }
 
 #[derive(Serialize, Deserialize)]
@@ -82,7 +85,7 @@ impl AttachmentService {
             .control_plane_api
             .clone()
             .unwrap()
-            .join("attach-hook")
+            .join("attach_hook")
             .unwrap();
         let client = reqwest::blocking::ClientBuilder::new()
             .build()
@@ -90,7 +93,7 @@ impl AttachmentService {
 
         let request = AttachHookRequest {
             tenant_id,
-            node_id: Some(pageserver_id),
+            pageserver_id: Some(pageserver_id),
         };
 
         let response = client.post(url).json(&request).send()?;

control_plane/src/bin/attachment_service.rs

@@ -12,7 +12,6 @@ use hyper::{Body, Request, Response};
 use serde::{Deserialize, Serialize};
 use std::path::{Path, PathBuf};
 use std::{collections::HashMap, sync::Arc};
-use utils::http::endpoint::request_span;
 use utils::logging::{self, LogFormat};
 use utils::signals::{ShutdownSignals, Signal};
 
@@ -172,7 +171,7 @@ async fn handle_re_attach(mut req: Request<Body>) -> Result<Response<Body>, ApiE
             state.generation += 1;
             response.tenants.push(ReAttachResponseTenant {
                 id: *t,
-                gen: state.generation,
+                generation: state.generation,
             });
         }
     }
@@ -218,31 +217,14 @@ async fn handle_attach_hook(mut req: Request<Body>) -> Result<Response<Body>, Ap
         .tenants
         .entry(attach_req.tenant_id)
         .or_insert_with(|| TenantState {
-            pageserver: attach_req.node_id,
+            pageserver: attach_req.pageserver_id,
             generation: 0,
         });
 
-    if let Some(attaching_pageserver) = attach_req.node_id.as_ref() {
+    if attach_req.pageserver_id.is_some() {
         tenant_state.generation += 1;
-        tracing::info!(
-            tenant_id = %attach_req.tenant_id,
-            ps_id = %attaching_pageserver,
-            generation = %tenant_state.generation,
-            "issuing",
-        );
-    } else if let Some(ps_id) = tenant_state.pageserver {
-        tracing::info!(
-            tenant_id = %attach_req.tenant_id,
-            %ps_id,
-            generation = %tenant_state.generation,
-            "dropping",
-        );
-    } else {
-        tracing::info!(
-            tenant_id = %attach_req.tenant_id,
-            "no-op: tenant already has no pageserver");
     }
-    tenant_state.pageserver = attach_req.node_id;
+    tenant_state.pageserver = attach_req.pageserver_id;
     let generation = tenant_state.generation;
 
     locked.save().await.map_err(ApiError::InternalServerError)?;
@@ -250,7 +232,7 @@ async fn handle_attach_hook(mut req: Request<Body>) -> Result<Response<Body>, Ap
     json_response(
         StatusCode::OK,
         AttachHookResponse {
-            gen: attach_req.node_id.map(|_| generation),
+            gen: attach_req.pageserver_id.map(|_| generation),
         },
     )
 }
@@ -258,9 +240,9 @@ async fn handle_attach_hook(mut req: Request<Body>) -> Result<Response<Body>, Ap
 fn make_router(persistent_state: PersistentState) -> RouterBuilder<hyper::Body, ApiError> {
     endpoint::make_router()
         .data(Arc::new(State::new(persistent_state)))
-        .post("/re-attach", |r| request_span(r, handle_re_attach))
-        .post("/validate", |r| request_span(r, handle_validate))
-        .post("/attach-hook", |r| request_span(r, handle_attach_hook))
+        .post("/re-attach", handle_re_attach)
+        .post("/validate", handle_validate)
+        .post("/attach_hook", handle_attach_hook)
 }
 
 #[tokio::main]

control_plane/src/bin/neon_local.rs

@@ -798,24 +798,6 @@ fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<(
                 ep.start(&auth_token, safekeepers, remote_ext_config)?;
             }
         }
-        "reconfigure" => {
-            let endpoint_id = sub_args
-                .get_one::<String>("endpoint_id")
-                .ok_or_else(|| anyhow!("No endpoint ID provided to reconfigure"))?;
-            let endpoint = cplane
-                .endpoints
-                .get(endpoint_id.as_str())
-                .with_context(|| format!("postgres endpoint {endpoint_id} is not found"))?;
-            let pageserver_id =
-                if let Some(id_str) = sub_args.get_one::<String>("endpoint-pageserver-id") {
-                    Some(NodeId(
-                        id_str.parse().context("while parsing pageserver id")?,
-                    ))
-                } else {
-                    None
-                };
-            endpoint.reconfigure(pageserver_id)?;
-        }
         "stop" => {
             let endpoint_id = sub_args
                 .get_one::<String>("endpoint_id")
@@ -1387,12 +1369,6 @@ fn cli() -> Command {
                     .arg(safekeepers_arg)
                     .arg(remote_ext_config_args)
                 )
-                .subcommand(Command::new("reconfigure")
-                            .about("Reconfigure the endpoint")
-                            .arg(endpoint_pageserver_id_arg)
-                            .arg(endpoint_id_arg.clone())
-                            .arg(tenant_id_arg.clone())
-                )
                 .subcommand(
                     Command::new("stop")
                     .arg(endpoint_id_arg)

control_plane/src/endpoint.rs

@@ -46,6 +46,7 @@ use std::time::Duration;
 
 use anyhow::{anyhow, bail, Context, Result};
 use serde::{Deserialize, Serialize};
+use serde_with::{serde_as, DisplayFromStr};
 use utils::id::{NodeId, TenantId, TimelineId};
 
 use crate::local_env::LocalEnv;
@@ -56,10 +57,13 @@ use compute_api::responses::{ComputeState, ComputeStatus};
 use compute_api::spec::{Cluster, ComputeMode, ComputeSpec};
 
 // contents of a endpoint.json file
+#[serde_as]
 #[derive(Serialize, Deserialize, PartialEq, Eq, Clone, Debug)]
 pub struct EndpointConf {
     endpoint_id: String,
+    #[serde_as(as = "DisplayFromStr")]
     tenant_id: TenantId,
+    #[serde_as(as = "DisplayFromStr")]
     timeline_id: TimelineId,
     mode: ComputeMode,
     pg_port: u16,
@@ -410,32 +414,16 @@ impl Endpoint {
             );
         }
 
-        Ok(())
-    }
-
-    fn wait_for_compute_ctl_to_exit(&self) -> Result<()> {
+        // Also wait for the compute_ctl process to die. It might have some cleanup
+        // work to do after postgres stops, like syncing safekeepers, etc.
+        //
         // TODO use background_process::stop_process instead
         let pidfile_path = self.endpoint_path().join("compute_ctl.pid");
         let pid: u32 = std::fs::read_to_string(pidfile_path)?.parse()?;
         let pid = nix::unistd::Pid::from_raw(pid as i32);
         crate::background_process::wait_until_stopped("compute_ctl", pid)?;
-        Ok(())
-    }
 
-    fn read_postgresql_conf(&self) -> Result<String> {
-        // Slurp the endpoints/<endpoint id>/postgresql.conf file into
-        // memory. We will include it in the spec file that we pass to
-        // `compute_ctl`, and `compute_ctl` will write it to the postgresql.conf
-        // in the data directory.
-        let postgresql_conf_path = self.endpoint_path().join("postgresql.conf");
-        match std::fs::read(&postgresql_conf_path) {
-            Ok(content) => Ok(String::from_utf8(content)?),
-            Err(e) if e.kind() == std::io::ErrorKind::NotFound => Ok("".to_string()),
-            Err(e) => Err(anyhow::Error::new(e).context(format!(
-                "failed to read config file in {}",
-                postgresql_conf_path.to_str().unwrap()
-            ))),
-        }
+        Ok(())
     }
 
     pub fn start(
@@ -448,7 +436,21 @@ impl Endpoint {
             anyhow::bail!("The endpoint is already running");
         }
 
-        let postgresql_conf = self.read_postgresql_conf()?;
+        // Slurp the endpoints/<endpoint id>/postgresql.conf file into
+        // memory. We will include it in the spec file that we pass to
+        // `compute_ctl`, and `compute_ctl` will write it to the postgresql.conf
+        // in the data directory.
+        let postgresql_conf_path = self.endpoint_path().join("postgresql.conf");
+        let postgresql_conf = match std::fs::read(&postgresql_conf_path) {
+            Ok(content) => String::from_utf8(content)?,
+            Err(e) if e.kind() == std::io::ErrorKind::NotFound => "".to_string(),
+            Err(e) => {
+                return Err(anyhow::Error::new(e).context(format!(
+                    "failed to read config file in {}",
+                    postgresql_conf_path.to_str().unwrap()
+                )))
+            }
+        };
 
         // We always start the compute node from scratch, so if the Postgres
         // data dir exists from a previous launch, remove it first.
@@ -619,61 +621,6 @@ impl Endpoint {
         }
     }
 
-    pub fn reconfigure(&self, pageserver_id: Option<NodeId>) -> Result<()> {
-        let mut spec: ComputeSpec = {
-            let spec_path = self.endpoint_path().join("spec.json");
-            let file = std::fs::File::open(spec_path)?;
-            serde_json::from_reader(file)?
-        };
-
-        let postgresql_conf = self.read_postgresql_conf()?;
-        spec.cluster.postgresql_conf = Some(postgresql_conf);
-
-        if let Some(pageserver_id) = pageserver_id {
-            let endpoint_config_path = self.endpoint_path().join("endpoint.json");
-            let mut endpoint_conf: EndpointConf = {
-                let file = std::fs::File::open(&endpoint_config_path)?;
-                serde_json::from_reader(file)?
-            };
-            endpoint_conf.pageserver_id = pageserver_id;
-            std::fs::write(
-                endpoint_config_path,
-                serde_json::to_string_pretty(&endpoint_conf)?,
-            )?;
-
-            let pageserver =
-                PageServerNode::from_env(&self.env, self.env.get_pageserver_conf(pageserver_id)?);
-            let ps_http_conf = &pageserver.pg_connection_config;
-            let (host, port) = (ps_http_conf.host(), ps_http_conf.port());
-            spec.pageserver_connstring = Some(format!("postgresql://no_user@{host}:{port}"));
-        }
-
-        let client = reqwest::blocking::Client::new();
-        let response = client
-            .post(format!(
-                "http://{}:{}/configure",
-                self.http_address.ip(),
-                self.http_address.port()
-            ))
-            .body(format!(
-                "{{\"spec\":{}}}",
-                serde_json::to_string_pretty(&spec)?
-            ))
-            .send()?;
-
-        let status = response.status();
-        if !(status.is_client_error() || status.is_server_error()) {
-            Ok(())
-        } else {
-            let url = response.url().to_owned();
-            let msg = match response.text() {
-                Ok(err_body) => format!("Error: {}", err_body),
-                Err(_) => format!("Http error ({}) at {}.", status.as_u16(), url),
-            };
-            Err(anyhow::anyhow!(msg))
-        }
-    }
-
     pub fn stop(&self, destroy: bool) -> Result<()> {
         // If we are going to destroy data directory,
         // use immediate shutdown mode, otherwise,
@@ -682,25 +629,15 @@ impl Endpoint {
         // Postgres is always started from scratch, so stop
         // without destroy only used for testing and debugging.
         //
-        self.pg_ctl(
-            if destroy {
-                &["-m", "immediate", "stop"]
-            } else {
-                &["stop"]
-            },
-            &None,
-        )?;
-
-        // Also wait for the compute_ctl process to die. It might have some cleanup
-        // work to do after postgres stops, like syncing safekeepers, etc.
-        //
-        self.wait_for_compute_ctl_to_exit()?;
         if destroy {
+            self.pg_ctl(&["-m", "immediate", "stop"], &None)?;
             println!(
                 "Destroying postgres data directory '{}'",
                 self.pgdata().to_str().unwrap()
             );
             std::fs::remove_dir_all(self.endpoint_path())?;
+        } else {
+            self.pg_ctl(&["stop"], &None)?;
         }
         Ok(())
     }

control_plane/src/local_env.rs

@@ -8,6 +8,7 @@ use anyhow::{bail, ensure, Context};
 use postgres_backend::AuthType;
 use reqwest::Url;
 use serde::{Deserialize, Serialize};
+use serde_with::{serde_as, DisplayFromStr};
 use std::collections::HashMap;
 use std::env;
 use std::fs;
@@ -32,6 +33,7 @@ pub const DEFAULT_PG_VERSION: u32 = 15;
 // to 'neon_local init --config=<path>' option. See control_plane/simple.conf for
 // an example.
 //
+#[serde_as]
 #[derive(Serialize, Deserialize, PartialEq, Eq, Clone, Debug)]
 pub struct LocalEnv {
     // Base directory for all the nodes (the pageserver, safekeepers and
@@ -57,6 +59,7 @@ pub struct LocalEnv {
     // Default tenant ID to use with the 'neon_local' command line utility, when
     // --tenant_id is not explicitly specified.
     #[serde(default)]
+    #[serde_as(as = "Option<DisplayFromStr>")]
     pub default_tenant_id: Option<TenantId>,
 
     // used to issue tokens during e.g pg start
@@ -81,6 +84,7 @@ pub struct LocalEnv {
     // 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".
+    #[serde_as(as = "HashMap<_, Vec<(DisplayFromStr, DisplayFromStr)>>")]
     branch_name_mappings: HashMap<String, Vec<(TenantId, TimelineId)>>,
 }
 

docs/updating-postgres.md

@@ -1,108 +0,0 @@
-# Updating Postgres
-
-## Minor Versions
-
-When upgrading to a new minor version of Postgres, please follow these steps:
-
-_Example: 15.4 is the new minor version to upgrade to from 15.3._
-
-1. Clone the Neon Postgres repository if you have not done so already.
-
-    ```shell
-    git clone git@github.com:neondatabase/postgres.git
-    ```
-
-1. Add the Postgres upstream remote.
-
-    ```shell
-    git remote add upstream https://git.postgresql.org/git/postgresql.git
-    ```
-
-1. Create a new branch based on the stable branch you are updating.
-
-    ```shell
-    git checkout -b my-branch REL_15_STABLE_neon
-    ```
-
-1. Tag the last commit on the stable branch you are updating.
-
-    ```shell
-    git tag REL_15_3_neon
-    ```
-
-1. Push the new tag to the Neon Postgres repository.
-
-    ```shell
-    git push origin REL_15_3_neon
-    ```
-
-1. Find the release tags you're looking for. They are of the form `REL_X_Y`.
-
-1. Rebase the branch you created on the tag and resolve any conflicts.
-
-    ```shell
-    git fetch upstream REL_15_4
-    git rebase REL_15_4
-    ```
-
-1. Run the Postgres test suite to make sure our commits have not affected
-Postgres in a negative way.
-
-    ```shell
-    make check
-    # OR
-    meson test -C builddir
-    ```
-
-1. Push your branch to the Neon Postgres repository.
-
-    ```shell
-    git push origin my-branch
-    ```
-
-1. Clone the Neon repository if you have not done so already.
-
-    ```shell
-    git clone git@github.com:neondatabase/neon.git
-    ```
-
-1. Create a new branch.
-
-1. Change the `revisions.json` file to point at the HEAD of your Postgres
-branch.
-
-1. Update the Git submodule.
-
-    ```shell
-    git submodule set-branch --branch my-branch vendor/postgres-v15
-    git submodule update --remote vendor/postgres-v15
-    ```
-
-1. Run the Neon test suite to make sure that Neon is still good to go on this
-minor Postgres release.
-
-    ```shell
-    ./scripts/poetry -k pg15
-    ```
-
-1. Commit your changes.
-
-1. Create a pull request, and wait for CI to go green.
-
-1. Force push the rebased Postgres branches into the Neon Postgres repository.
-
-    ```shell
-    git push --force origin my-branch:REL_15_STABLE_neon
-    ```
-
-    It may require disabling various branch protections.
-
-1. Update your Neon PR to point at the branches.
-
-    ```shell
-    git submodule set-branch --branch REL_15_STABLE_neon vendor/postgres-v15
-    git commit --amend --no-edit
-    git push --force origin
-    ```
-
-1. Merge the pull request after getting approval(s) and CI completion.

libs/compute_api/src/spec.rs

@@ -6,6 +6,7 @@
 use std::collections::HashMap;
 
 use serde::{Deserialize, Serialize};
+use serde_with::{serde_as, DisplayFromStr};
 use utils::id::{TenantId, TimelineId};
 use utils::lsn::Lsn;
 
@@ -18,6 +19,7 @@ pub type PgIdent = String;
 
 /// Cluster spec or configuration represented as an optional number of
 /// delta operations + final cluster state description.
+#[serde_as]
 #[derive(Clone, Debug, Default, Deserialize, Serialize)]
 pub struct ComputeSpec {
     pub format_version: f32,
@@ -48,12 +50,12 @@ pub struct ComputeSpec {
     // these, and instead set the "neon.tenant_id", "neon.timeline_id",
     // etc. GUCs in cluster.settings. TODO: Once the control plane has been
     // updated to fill these fields, we can make these non optional.
+    #[serde_as(as = "Option<DisplayFromStr>")]
     pub tenant_id: Option<TenantId>,
-
+    #[serde_as(as = "Option<DisplayFromStr>")]
     pub timeline_id: Option<TimelineId>,
-
+    #[serde_as(as = "Option<DisplayFromStr>")]
     pub pageserver_connstring: Option<String>,
-
     #[serde(default)]
     pub safekeeper_connstrings: Vec<String>,
 
@@ -138,13 +140,14 @@ impl RemoteExtSpec {
     }
 }
 
+#[serde_as]
 #[derive(Clone, Copy, Debug, Default, Eq, PartialEq, Deserialize, Serialize)]
 pub enum ComputeMode {
     /// A read-write node
     #[default]
     Primary,
     /// A read-only node, pinned at a particular LSN
-    Static(Lsn),
+    Static(#[serde_as(as = "DisplayFromStr")] Lsn),
     /// A read-only node that follows the tip of the branch in hot standby mode
     ///
     /// Future versions may want to distinguish between replicas with hot standby
@@ -187,8 +190,6 @@ pub struct DeltaOp {
 pub struct Role {
     pub name: PgIdent,
     pub encrypted_password: Option<String>,
-    pub replication: Option<bool>,
-    pub bypassrls: Option<bool>,
     pub options: GenericOptions,
 }
 

libs/metrics/src/lib.rs

@@ -89,14 +89,14 @@ pub const DISK_WRITE_SECONDS_BUCKETS: &[f64] = &[
     0.000_050, 0.000_100, 0.000_500, 0.001, 0.003, 0.005, 0.01, 0.05, 0.1, 0.3, 0.5,
 ];
 
-pub fn set_build_info_metric(revision: &str, build_tag: &str) {
+pub fn set_build_info_metric(revision: &str) {
     let metric = register_int_gauge_vec!(
         "libmetrics_build_info",
         "Build/version information",
-        &["revision", "build_tag"]
+        &["revision"]
     )
     .expect("Failed to register build info metric");
-    metric.with_label_values(&[revision, build_tag]).set(1);
+    metric.with_label_values(&[revision]).set(1);
 }
 
 // Records I/O stats in a "cross-platform" way.

libs/pageserver_api/src/control_api.rs

@@ -4,6 +4,7 @@
 //! See docs/rfcs/025-generation-numbers.md
 
 use serde::{Deserialize, Serialize};
+use serde_with::{serde_as, DisplayFromStr};
 use utils::id::{NodeId, TenantId};
 
 #[derive(Serialize, Deserialize)]
@@ -11,10 +12,12 @@ pub struct ReAttachRequest {
     pub node_id: NodeId,
 }
 
+#[serde_as]
 #[derive(Serialize, Deserialize)]
 pub struct ReAttachResponseTenant {
+    #[serde_as(as = "DisplayFromStr")]
     pub id: TenantId,
-    pub gen: u32,
+    pub generation: u32,
 }
 
 #[derive(Serialize, Deserialize)]
@@ -22,8 +25,10 @@ pub struct ReAttachResponse {
     pub tenants: Vec<ReAttachResponseTenant>,
 }
 
+#[serde_as]
 #[derive(Serialize, Deserialize)]
 pub struct ValidateRequestTenant {
+    #[serde_as(as = "DisplayFromStr")]
     pub id: TenantId,
     pub gen: u32,
 }
@@ -38,8 +43,10 @@ pub struct ValidateResponse {
     pub tenants: Vec<ValidateResponseTenant>,
 }
 
+#[serde_as]
 #[derive(Serialize, Deserialize)]
 pub struct ValidateResponseTenant {
+    #[serde_as(as = "DisplayFromStr")]
     pub id: TenantId,
     pub valid: bool,
 }

libs/pageserver_api/src/models.rs

@@ -6,7 +6,7 @@ use std::{
 
 use byteorder::{BigEndian, ReadBytesExt};
 use serde::{Deserialize, Serialize};
-use serde_with::serde_as;
+use serde_with::{serde_as, DisplayFromStr};
 use strum_macros;
 use utils::{
     completion,
@@ -18,7 +18,7 @@ use utils::{
 
 use crate::reltag::RelTag;
 use anyhow::bail;
-use bytes::{BufMut, Bytes, BytesMut};
+use bytes::{Buf, BufMut, Bytes, BytesMut};
 
 /// The state of a tenant in this pageserver.
 ///
@@ -110,6 +110,7 @@ impl TenantState {
             // So, return `Maybe` while Attaching, making Console wait for the attach task to finish.
             Self::Attaching | Self::Activating(ActivatingFrom::Attaching) => Maybe,
             // tenant mgr startup distinguishes attaching from loading via marker file.
+            // If it's loading, there is no attach marker file, i.e., attach had finished in the past.
             Self::Loading | Self::Activating(ActivatingFrom::Loading) => Attached,
             // We only reach Active after successful load / attach.
             // So, call atttachment status Attached.
@@ -174,19 +175,25 @@ pub enum TimelineState {
     Broken { reason: String, backtrace: String },
 }
 
+#[serde_as]
 #[derive(Serialize, Deserialize)]
 pub struct TimelineCreateRequest {
+    #[serde_as(as = "DisplayFromStr")]
     pub new_timeline_id: TimelineId,
     #[serde(default)]
+    #[serde_as(as = "Option<DisplayFromStr>")]
     pub ancestor_timeline_id: Option<TimelineId>,
     #[serde(default)]
+    #[serde_as(as = "Option<DisplayFromStr>")]
     pub ancestor_start_lsn: Option<Lsn>,
     pub pg_version: Option<u32>,
 }
 
+#[serde_as]
 #[derive(Serialize, Deserialize, Debug)]
 #[serde(deny_unknown_fields)]
 pub struct TenantCreateRequest {
+    #[serde_as(as = "DisplayFromStr")]
     pub new_tenant_id: TenantId,
     #[serde(default)]
     #[serde(skip_serializing_if = "Option::is_none")]
@@ -195,6 +202,7 @@ pub struct TenantCreateRequest {
     pub config: TenantConfig, // as we have a flattened field, we should reject all unknown fields in it
 }
 
+#[serde_as]
 #[derive(Deserialize, Debug)]
 #[serde(deny_unknown_fields)]
 pub struct TenantLoadRequest {
@@ -271,26 +279,31 @@ pub struct LocationConfig {
     pub tenant_conf: TenantConfig,
 }
 
+#[serde_as]
 #[derive(Serialize, Deserialize)]
 #[serde(transparent)]
-pub struct TenantCreateResponse(pub TenantId);
+pub struct TenantCreateResponse(#[serde_as(as = "DisplayFromStr")] pub TenantId);
 
 #[derive(Serialize)]
 pub struct StatusResponse {
     pub id: NodeId,
 }
 
+#[serde_as]
 #[derive(Serialize, Deserialize, Debug)]
 #[serde(deny_unknown_fields)]
 pub struct TenantLocationConfigRequest {
+    #[serde_as(as = "DisplayFromStr")]
     pub tenant_id: TenantId,
     #[serde(flatten)]
     pub config: LocationConfig, // as we have a flattened field, we should reject all unknown fields in it
 }
 
+#[serde_as]
 #[derive(Serialize, Deserialize, Debug)]
 #[serde(deny_unknown_fields)]
 pub struct TenantConfigRequest {
+    #[serde_as(as = "DisplayFromStr")]
     pub tenant_id: TenantId,
     #[serde(flatten)]
     pub config: TenantConfig, // as we have a flattened field, we should reject all unknown fields in it
@@ -362,8 +375,10 @@ pub enum TenantAttachmentStatus {
     Failed { reason: String },
 }
 
+#[serde_as]
 #[derive(Serialize, Deserialize, Clone)]
 pub struct TenantInfo {
+    #[serde_as(as = "DisplayFromStr")]
     pub id: TenantId,
     // NB: intentionally not part of OpenAPI, we don't want to commit to a specific set of TenantState's
     pub state: TenantState,
@@ -374,22 +389,33 @@ pub struct TenantInfo {
 }
 
 /// This represents the output of the "timeline_detail" and "timeline_list" API calls.
+#[serde_as]
 #[derive(Debug, Serialize, Deserialize, Clone)]
 pub struct TimelineInfo {
+    #[serde_as(as = "DisplayFromStr")]
     pub tenant_id: TenantId,
+    #[serde_as(as = "DisplayFromStr")]
     pub timeline_id: TimelineId,
 
+    #[serde_as(as = "Option<DisplayFromStr>")]
     pub ancestor_timeline_id: Option<TimelineId>,
+    #[serde_as(as = "Option<DisplayFromStr>")]
     pub ancestor_lsn: Option<Lsn>,
+    #[serde_as(as = "DisplayFromStr")]
     pub last_record_lsn: Lsn,
+    #[serde_as(as = "Option<DisplayFromStr>")]
     pub prev_record_lsn: Option<Lsn>,
+    #[serde_as(as = "DisplayFromStr")]
     pub latest_gc_cutoff_lsn: Lsn,
+    #[serde_as(as = "DisplayFromStr")]
     pub disk_consistent_lsn: Lsn,
 
     /// The LSN that we have succesfully uploaded to remote storage
+    #[serde_as(as = "DisplayFromStr")]
     pub remote_consistent_lsn: Lsn,
 
     /// The LSN that we are advertizing to safekeepers
+    #[serde_as(as = "DisplayFromStr")]
     pub remote_consistent_lsn_visible: Lsn,
 
     pub current_logical_size: Option<u64>, // is None when timeline is Unloaded
@@ -401,6 +427,7 @@ pub struct TimelineInfo {
     pub timeline_dir_layer_file_size_sum: Option<u64>,
 
     pub wal_source_connstr: Option<String>,
+    #[serde_as(as = "Option<DisplayFromStr>")]
     pub last_received_msg_lsn: Option<Lsn>,
     /// the timestamp (in microseconds) of the last received message
     pub last_received_msg_ts: Option<u128>,
@@ -497,13 +524,23 @@ pub struct LayerAccessStats {
     pub residence_events_history: HistoryBufferWithDropCounter<LayerResidenceEvent, 16>,
 }
 
+#[serde_as]
 #[derive(Debug, Clone, Serialize)]
 #[serde(tag = "kind")]
 pub enum InMemoryLayerInfo {
-    Open { lsn_start: Lsn },
-    Frozen { lsn_start: Lsn, lsn_end: Lsn },
+    Open {
+        #[serde_as(as = "DisplayFromStr")]
+        lsn_start: Lsn,
+    },
+    Frozen {
+        #[serde_as(as = "DisplayFromStr")]
+        lsn_start: Lsn,
+        #[serde_as(as = "DisplayFromStr")]
+        lsn_end: Lsn,
+    },
 }
 
+#[serde_as]
 #[derive(Debug, Clone, Serialize)]
 #[serde(tag = "kind")]
 pub enum HistoricLayerInfo {
@@ -511,7 +548,9 @@ pub enum HistoricLayerInfo {
         layer_file_name: String,
         layer_file_size: u64,
 
+        #[serde_as(as = "DisplayFromStr")]
         lsn_start: Lsn,
+        #[serde_as(as = "DisplayFromStr")]
         lsn_end: Lsn,
         remote: bool,
         access_stats: LayerAccessStats,
@@ -520,6 +559,7 @@ pub enum HistoricLayerInfo {
         layer_file_name: String,
         layer_file_size: u64,
 
+        #[serde_as(as = "DisplayFromStr")]
         lsn_start: Lsn,
         remote: bool,
         access_stats: LayerAccessStats,
@@ -767,6 +807,36 @@ impl PagestreamBeMessage {
 
         bytes.into()
     }
+
+    pub fn deserialize(buf: Bytes) -> anyhow::Result<Self> {
+        let mut buf = buf.reader();
+        let msg_tag = buf.read_u8()?;
+        match msg_tag {
+            100 => todo!(),
+            101 => todo!(),
+            102 => {
+                let buf = buf.get_ref();
+                /* TODO use constant */
+                if buf.len() == 8192 {
+                    Ok(PagestreamBeMessage::GetPage(PagestreamGetPageResponse {
+                        page: buf.clone(),
+                    }))
+                } else {
+                    anyhow::bail!("invalid page size: {}", buf.len());
+                }
+            }
+            103 => {
+                let buf = buf.get_ref();
+                let cstr = std::ffi::CStr::from_bytes_until_nul(&buf)?;
+                let rust_str = cstr.to_str()?;
+                Ok(PagestreamBeMessage::Error(PagestreamErrorResponse {
+                    message: rust_str.to_owned(),
+                }))
+            }
+            104 => todo!(),
+            _ => bail!("unknown tag: {:?}", msg_tag),
+        }
+    }
 }
 
 #[cfg(test)]

libs/postgres_backend/src/lib.rs

@@ -242,7 +242,6 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> MaybeWriteOnly<IO> {
         }
     }
 
-    /// Cancellation safe as long as the underlying IO is cancellation safe.
     async fn shutdown(&mut self) -> io::Result<()> {
         match self {
             MaybeWriteOnly::Full(framed) => framed.shutdown().await,
@@ -394,23 +393,13 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
         shutdown_watcher: F,
     ) -> Result<(), QueryError>
     where
-        F: Fn() -> S + Clone,
+        F: Fn() -> S,
         S: Future,
     {
-        let ret = self
-            .run_message_loop(handler, shutdown_watcher.clone())
-            .await;
-
-        tokio::select! {
-            _ = shutdown_watcher() => {
-                // do nothing; we most likely got already stopped by shutdown and will log it next.
-            }
-            _ = self.framed.shutdown() => {
-                // socket might be already closed, e.g. if previously received error,
-                // so ignore result.
-            },
-        }
-
+        let ret = self.run_message_loop(handler, shutdown_watcher).await;
+        // socket might be already closed, e.g. if previously received error,
+        // so ignore result.
+        self.framed.shutdown().await.ok();
         match ret {
             Ok(()) => Ok(()),
             Err(QueryError::Shutdown) => {
@@ -728,17 +717,12 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
 
                 trace!("got query {query_string:?}");
                 if let Err(e) = handler.process_query(self, query_string).await {
-                    match e {
-                        QueryError::Shutdown => return Ok(ProcessMsgResult::Break),
-                        e => {
-                            log_query_error(query_string, &e);
-                            let short_error = short_error(&e);
-                            self.write_message_noflush(&BeMessage::ErrorResponse(
-                                &short_error,
-                                Some(e.pg_error_code()),
-                            ))?;
-                        }
-                    }
+                    log_query_error(query_string, &e);
+                    let short_error = short_error(&e);
+                    self.write_message_noflush(&BeMessage::ErrorResponse(
+                        &short_error,
+                        Some(e.pg_error_code()),
+                    ))?;
                 }
                 self.write_message_noflush(&BeMessage::ReadyForQuery)?;
             }

libs/postgres_ffi/wal_craft/src/lib.rs

@@ -14,7 +14,6 @@ macro_rules! xlog_utils_test {
     ($version:ident) => {
         #[path = "."]
         mod $version {
-            #[allow(unused_imports)]
             pub use postgres_ffi::$version::wal_craft_test_export::*;
             #[allow(clippy::duplicate_mod)]
             #[cfg(test)]

libs/pq_proto/src/framed.rs

@@ -214,24 +214,27 @@ where
     }
 }
 
-/// Cancellation safe as long as the AsyncWrite is cancellation safe.
 async fn flush<S: AsyncWrite + Unpin>(
     stream: &mut S,
     write_buf: &mut BytesMut,
 ) -> Result<(), io::Error> {
     while write_buf.has_remaining() {
-        let bytes_written = stream.write_buf(write_buf).await?;
+        let bytes_written = stream.write(write_buf.chunk()).await?;
         if bytes_written == 0 {
             return Err(io::Error::new(
                 ErrorKind::WriteZero,
                 "failed to write message",
             ));
         }
+        // The advanced part will be garbage collected, likely during shifting
+        // data left on next attempt to write to buffer when free space is not
+        // enough.
+        write_buf.advance(bytes_written);
     }
+    write_buf.clear();
     stream.flush().await
 }
 
-/// Cancellation safe as long as the AsyncWrite is cancellation safe.
 async fn shutdown<S: AsyncWrite + Unpin>(
     stream: &mut S,
     write_buf: &mut BytesMut,

libs/remote_storage/src/azure_blob.rs

@@ -1,18 +1,21 @@
 //! Azure Blob Storage wrapper
 
-use std::collections::HashMap;
 use std::env;
 use std::num::NonZeroU32;
 use std::sync::Arc;
-use std::{borrow::Cow, io::Cursor};
+use std::{borrow::Cow, collections::HashMap, io::Cursor};
 
 use super::REMOTE_STORAGE_PREFIX_SEPARATOR;
 use anyhow::Result;
 use azure_core::request_options::{MaxResults, Metadata, Range};
+use azure_core::Header;
 use azure_identity::DefaultAzureCredential;
 use azure_storage::StorageCredentials;
 use azure_storage_blobs::prelude::ClientBuilder;
-use azure_storage_blobs::{blob::operations::GetBlobBuilder, prelude::ContainerClient};
+use azure_storage_blobs::{
+    blob::operations::GetBlobBuilder,
+    prelude::{BlobClient, ContainerClient},
+};
 use futures_util::StreamExt;
 use http_types::StatusCode;
 use tokio::io::AsyncRead;
@@ -20,8 +23,8 @@ use tracing::debug;
 
 use crate::s3_bucket::RequestKind;
 use crate::{
-    AzureConfig, ConcurrencyLimiter, Download, DownloadError, Listing, ListingMode, RemotePath,
-    RemoteStorage, StorageMetadata,
+    AzureConfig, ConcurrencyLimiter, Download, DownloadError, RemotePath, RemoteStorage,
+    StorageMetadata,
 };
 
 pub struct AzureBlobStorage {
@@ -109,19 +112,16 @@ impl AzureBlobStorage {
 
     async fn download_for_builder(
         &self,
+        metadata: StorageMetadata,
         builder: GetBlobBuilder,
     ) -> Result<Download, DownloadError> {
         let mut response = builder.into_stream();
 
-        let mut metadata = HashMap::new();
         // TODO give proper streaming response instead of buffering into RAM
         // https://github.com/neondatabase/neon/issues/5563
         let mut buf = Vec::new();
         while let Some(part) = response.next().await {
             let part = part.map_err(to_download_error)?;
-            if let Some(blob_meta) = part.blob.metadata {
-                metadata.extend(blob_meta.iter().map(|(k, v)| (k.to_owned(), v.to_owned())));
-            }
             let data = part
                 .data
                 .collect()
@@ -131,9 +131,28 @@ impl AzureBlobStorage {
         }
         Ok(Download {
             download_stream: Box::pin(Cursor::new(buf)),
-            metadata: Some(StorageMetadata(metadata)),
+            metadata: Some(metadata),
         })
     }
+    // TODO get rid of this function once we have metadata included in the response
+    // https://github.com/Azure/azure-sdk-for-rust/issues/1439
+    async fn get_metadata(
+        &self,
+        blob_client: &BlobClient,
+    ) -> Result<StorageMetadata, DownloadError> {
+        let builder = blob_client.get_metadata();
+
+        let response = builder.into_future().await.map_err(to_download_error)?;
+        let mut map = HashMap::new();
+
+        for md in response.metadata.iter() {
+            map.insert(
+                md.name().as_str().to_string(),
+                md.value().as_str().to_string(),
+            );
+        }
+        Ok(StorageMetadata(map))
+    }
 
     async fn permit(&self, kind: RequestKind) -> tokio::sync::SemaphorePermit<'_> {
         self.concurrency_limiter
@@ -165,11 +184,10 @@ fn to_download_error(error: azure_core::Error) -> DownloadError {
 
 #[async_trait::async_trait]
 impl RemoteStorage for AzureBlobStorage {
-    async fn list(
+    async fn list_prefixes(
         &self,
         prefix: Option<&RemotePath>,
-        mode: ListingMode,
-    ) -> anyhow::Result<Listing, DownloadError> {
+    ) -> Result<Vec<RemotePath>, DownloadError> {
         // get the passed prefix or if it is not set use prefix_in_bucket value
         let list_prefix = prefix
             .map(|p| self.relative_path_to_name(p))
@@ -177,19 +195,16 @@ impl RemoteStorage for AzureBlobStorage {
             .map(|mut p| {
                 // required to end with a separator
                 // otherwise request will return only the entry of a prefix
-                if matches!(mode, ListingMode::WithDelimiter)
-                    && !p.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR)
-                {
+                if !p.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR) {
                     p.push(REMOTE_STORAGE_PREFIX_SEPARATOR);
                 }
                 p
             });
 
-        let mut builder = self.client.list_blobs();
-
-        if let ListingMode::WithDelimiter = mode {
-            builder = builder.delimiter(REMOTE_STORAGE_PREFIX_SEPARATOR.to_string());
-        }
+        let mut builder = self
+            .client
+            .list_blobs()
+            .delimiter(REMOTE_STORAGE_PREFIX_SEPARATOR.to_string());
 
         if let Some(prefix) = list_prefix {
             builder = builder.prefix(Cow::from(prefix.to_owned()));
@@ -200,23 +215,46 @@ impl RemoteStorage for AzureBlobStorage {
         }
 
         let mut response = builder.into_stream();
-        let mut res = Listing::default();
-        while let Some(l) = response.next().await {
-            let entry = l.map_err(to_download_error)?;
-            let prefix_iter = entry
+        let mut res = Vec::new();
+        while let Some(entry) = response.next().await {
+            let entry = entry.map_err(to_download_error)?;
+            let name_iter = entry
                 .blobs
                 .prefixes()
                 .map(|prefix| self.name_to_relative_path(&prefix.name));
-            res.prefixes.extend(prefix_iter);
-
-            let blob_iter = entry
-                .blobs
-                .blobs()
-                .map(|k| self.name_to_relative_path(&k.name));
-            res.keys.extend(blob_iter);
+            res.extend(name_iter);
         }
         Ok(res)
     }
+
+    async fn list_files(&self, folder: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>> {
+        let folder_name = folder
+            .map(|p| self.relative_path_to_name(p))
+            .or_else(|| self.prefix_in_container.clone());
+
+        let mut builder = self.client.list_blobs();
+
+        if let Some(folder_name) = folder_name {
+            builder = builder.prefix(Cow::from(folder_name.to_owned()));
+        }
+
+        if let Some(limit) = self.max_keys_per_list_response {
+            builder = builder.max_results(MaxResults::new(limit));
+        }
+
+        let mut response = builder.into_stream();
+        let mut res = Vec::new();
+        while let Some(l) = response.next().await {
+            let entry = l.map_err(anyhow::Error::new)?;
+            let name_iter = entry
+                .blobs
+                .blobs()
+                .map(|bl| self.name_to_relative_path(&bl.name));
+            res.extend(name_iter);
+        }
+        Ok(res)
+    }
+
     async fn upload(
         &self,
         mut from: impl AsyncRead + Unpin + Send + Sync + 'static,
@@ -250,9 +288,11 @@ impl RemoteStorage for AzureBlobStorage {
         let _permit = self.permit(RequestKind::Get).await;
         let blob_client = self.client.blob_client(self.relative_path_to_name(from));
 
+        let metadata = self.get_metadata(&blob_client).await?;
+
         let builder = blob_client.get();
 
-        self.download_for_builder(builder).await
+        self.download_for_builder(metadata, builder).await
     }
 
     async fn download_byte_range(
@@ -264,6 +304,8 @@ impl RemoteStorage for AzureBlobStorage {
         let _permit = self.permit(RequestKind::Get).await;
         let blob_client = self.client.blob_client(self.relative_path_to_name(from));
 
+        let metadata = self.get_metadata(&blob_client).await?;
+
         let mut builder = blob_client.get();
 
         if let Some(end_exclusive) = end_exclusive {
@@ -278,7 +320,7 @@ impl RemoteStorage for AzureBlobStorage {
             builder = builder.range(Range::new(start_inclusive, end_exclusive));
         }
 
-        self.download_for_builder(builder).await
+        self.download_for_builder(metadata, builder).await
     }
 
     async fn delete(&self, path: &RemotePath) -> anyhow::Result<()> {

libs/remote_storage/src/lib.rs

@@ -112,7 +112,7 @@ impl RemotePath {
         self.0.file_name()
     }
 
-    pub fn join(&self, segment: &Utf8Path) -> Self {
+    pub fn join<P: AsRef<Utf8Path>>(&self, segment: P) -> Self {
         Self(self.0.join(segment))
     }
 
@@ -129,22 +129,6 @@ impl RemotePath {
     }
 }
 
-/// We don't need callers to be able to pass arbitrary delimiters: just control
-/// whether listings will use a '/' separator or not.
-///
-/// The WithDelimiter mode will populate `prefixes` and `keys` in the result.  The
-/// NoDelimiter mode will only populate `keys`.
-pub enum ListingMode {
-    WithDelimiter,
-    NoDelimiter,
-}
-
-#[derive(Default)]
-pub struct Listing {
-    pub prefixes: Vec<RemotePath>,
-    pub keys: Vec<RemotePath>,
-}
-
 /// 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.
@@ -157,13 +141,8 @@ pub trait RemoteStorage: Send + Sync + 'static {
     async fn list_prefixes(
         &self,
         prefix: Option<&RemotePath>,
-    ) -> Result<Vec<RemotePath>, DownloadError> {
-        let result = self
-            .list(prefix, ListingMode::WithDelimiter)
-            .await?
-            .prefixes;
-        Ok(result)
-    }
+    ) -> Result<Vec<RemotePath>, DownloadError>;
+
     /// Lists all files in directory "recursively"
     /// (not really recursively, because AWS has a flat namespace)
     /// Note: This is subtely different than list_prefixes,
@@ -175,16 +154,7 @@ pub trait RemoteStorage: Send + Sync + 'static {
     /// whereas,
     /// list_prefixes("foo/bar/") = ["cat", "dog"]
     /// See `test_real_s3.rs` for more details.
-    async fn list_files(&self, prefix: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>> {
-        let result = self.list(prefix, ListingMode::NoDelimiter).await?.keys;
-        Ok(result)
-    }
-
-    async fn list(
-        &self,
-        prefix: Option<&RemotePath>,
-        _mode: ListingMode,
-    ) -> anyhow::Result<Listing, DownloadError>;
+    async fn list_files(&self, folder: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>>;
 
     /// Streams the local file contents into remote into the remote storage entry.
     async fn upload(
@@ -235,9 +205,6 @@ pub enum DownloadError {
     BadInput(anyhow::Error),
     /// The file was not found in the remote storage.
     NotFound,
-    /// A cancellation token aborted the download, typically during
-    /// tenant detach or process shutdown.
-    Cancelled,
     /// The file was found in the remote storage, but the download failed.
     Other(anyhow::Error),
 }
@@ -248,7 +215,6 @@ impl std::fmt::Display for DownloadError {
             DownloadError::BadInput(e) => {
                 write!(f, "Failed to download a remote file due to user input: {e}")
             }
-            DownloadError::Cancelled => write!(f, "Cancelled, shutting down"),
             DownloadError::NotFound => write!(f, "No file found for the remote object id given"),
             DownloadError::Other(e) => write!(f, "Failed to download a remote file: {e:?}"),
         }
@@ -268,19 +234,6 @@ pub enum GenericRemoteStorage {
 }
 
 impl GenericRemoteStorage {
-    pub async fn list(
-        &self,
-        prefix: Option<&RemotePath>,
-        mode: ListingMode,
-    ) -> anyhow::Result<Listing, DownloadError> {
-        match self {
-            Self::LocalFs(s) => s.list(prefix, mode).await,
-            Self::AwsS3(s) => s.list(prefix, mode).await,
-            Self::AzureBlob(s) => s.list(prefix, mode).await,
-            Self::Unreliable(s) => s.list(prefix, mode).await,
-        }
-    }
-
     // A function for listing all the files in a "directory"
     // Example:
     // list_files("foo/bar") = ["foo/bar/a.txt", "foo/bar/b.txt"]

libs/remote_storage/src/local_fs.rs

@@ -15,7 +15,7 @@ use tokio::{
 use tracing::*;
 use utils::{crashsafe::path_with_suffix_extension, fs_ext::is_directory_empty};
 
-use crate::{Download, DownloadError, Listing, ListingMode, RemotePath};
+use crate::{Download, DownloadError, RemotePath};
 
 use super::{RemoteStorage, StorageMetadata};
 
@@ -75,7 +75,7 @@ impl LocalFs {
     }
 
     #[cfg(test)]
-    async fn list_all(&self) -> anyhow::Result<Vec<RemotePath>> {
+    async fn list(&self) -> anyhow::Result<Vec<RemotePath>> {
         Ok(get_all_files(&self.storage_root, true)
             .await?
             .into_iter()
@@ -89,10 +89,52 @@ impl LocalFs {
             })
             .collect())
     }
+}
+
+#[async_trait::async_trait]
+impl RemoteStorage for LocalFs {
+    async fn list_prefixes(
+        &self,
+        prefix: Option<&RemotePath>,
+    ) -> Result<Vec<RemotePath>, DownloadError> {
+        let path = match prefix {
+            Some(prefix) => Cow::Owned(prefix.with_base(&self.storage_root)),
+            None => Cow::Borrowed(&self.storage_root),
+        };
+
+        let prefixes_to_filter = get_all_files(path.as_ref(), false)
+            .await
+            .map_err(DownloadError::Other)?;
+
+        let mut prefixes = Vec::with_capacity(prefixes_to_filter.len());
+
+        // filter out empty directories to mirror s3 behavior.
+        for prefix in prefixes_to_filter {
+            if prefix.is_dir()
+                && is_directory_empty(&prefix)
+                    .await
+                    .map_err(DownloadError::Other)?
+            {
+                continue;
+            }
+
+            prefixes.push(
+                prefix
+                    .strip_prefix(&self.storage_root)
+                    .context("Failed to strip prefix")
+                    .and_then(RemotePath::new)
+                    .expect(
+                        "We list files for storage root, hence should be able to remote the prefix",
+                    ),
+            )
+        }
+
+        Ok(prefixes)
+    }
 
     // recursively lists all files in a directory,
     // mirroring the `list_files` for `s3_bucket`
-    async fn list_recursive(&self, folder: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>> {
+    async fn list_files(&self, folder: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>> {
         let full_path = match folder {
             Some(folder) => folder.with_base(&self.storage_root),
             None => self.storage_root.clone(),
@@ -144,70 +186,6 @@ impl LocalFs {
 
         Ok(files)
     }
-}
-
-#[async_trait::async_trait]
-impl RemoteStorage for LocalFs {
-    async fn list(
-        &self,
-        prefix: Option<&RemotePath>,
-        mode: ListingMode,
-    ) -> Result<Listing, DownloadError> {
-        let mut result = Listing::default();
-
-        if let ListingMode::NoDelimiter = mode {
-            let keys = self
-                .list_recursive(prefix)
-                .await
-                .map_err(DownloadError::Other)?;
-
-            result.keys = keys
-                .into_iter()
-                .filter(|k| {
-                    let path = k.with_base(&self.storage_root);
-                    !path.is_dir()
-                })
-                .collect();
-
-            return Ok(result);
-        }
-
-        let path = match prefix {
-            Some(prefix) => Cow::Owned(prefix.with_base(&self.storage_root)),
-            None => Cow::Borrowed(&self.storage_root),
-        };
-
-        let prefixes_to_filter = get_all_files(path.as_ref(), false)
-            .await
-            .map_err(DownloadError::Other)?;
-
-        // filter out empty directories to mirror s3 behavior.
-        for prefix in prefixes_to_filter {
-            if prefix.is_dir()
-                && is_directory_empty(&prefix)
-                    .await
-                    .map_err(DownloadError::Other)?
-            {
-                continue;
-            }
-
-            let stripped = prefix
-                .strip_prefix(&self.storage_root)
-                .context("Failed to strip prefix")
-                .and_then(RemotePath::new)
-                .expect(
-                    "We list files for storage root, hence should be able to remote the prefix",
-                );
-
-            if prefix.is_dir() {
-                result.prefixes.push(stripped);
-            } else {
-                result.keys.push(stripped);
-            }
-        }
-
-        Ok(result)
-    }
 
     async fn upload(
         &self,
@@ -501,7 +479,7 @@ mod fs_tests {
 
         let target_path_1 = upload_dummy_file(&storage, "upload_1", None).await?;
         assert_eq!(
-            storage.list_all().await?,
+            storage.list().await?,
             vec![target_path_1.clone()],
             "Should list a single file after first upload"
         );
@@ -689,7 +667,7 @@ mod fs_tests {
         let upload_target = upload_dummy_file(&storage, upload_name, None).await?;
 
         storage.delete(&upload_target).await?;
-        assert!(storage.list_all().await?.is_empty());
+        assert!(storage.list().await?.is_empty());
 
         storage
             .delete(&upload_target)
@@ -747,43 +725,6 @@ mod fs_tests {
         Ok(())
     }
 
-    #[tokio::test]
-    async fn list() -> anyhow::Result<()> {
-        // No delimiter: should recursively list everything
-        let storage = create_storage()?;
-        let child = upload_dummy_file(&storage, "grandparent/parent/child", None).await?;
-        let uncle = upload_dummy_file(&storage, "grandparent/uncle", None).await?;
-
-        let listing = storage.list(None, ListingMode::NoDelimiter).await?;
-        assert!(listing.prefixes.is_empty());
-        assert_eq!(listing.keys, [uncle.clone(), child.clone()].to_vec());
-
-        // Delimiter: should only go one deep
-        let listing = storage.list(None, ListingMode::WithDelimiter).await?;
-
-        assert_eq!(
-            listing.prefixes,
-            [RemotePath::from_string("timelines").unwrap()].to_vec()
-        );
-        assert!(listing.keys.is_empty());
-
-        // Delimiter & prefix
-        let listing = storage
-            .list(
-                Some(&RemotePath::from_string("timelines/some_timeline/grandparent").unwrap()),
-                ListingMode::WithDelimiter,
-            )
-            .await?;
-        assert_eq!(
-            listing.prefixes,
-            [RemotePath::from_string("timelines/some_timeline/grandparent/parent").unwrap()]
-                .to_vec()
-        );
-        assert_eq!(listing.keys, [uncle.clone()].to_vec());
-
-        Ok(())
-    }
-
     async fn upload_dummy_file(
         storage: &LocalFs,
         name: &str,
@@ -836,7 +777,7 @@ mod fs_tests {
     }
 
     async fn list_files_sorted(storage: &LocalFs) -> anyhow::Result<Vec<RemotePath>> {
-        let mut files = storage.list_all().await?;
+        let mut files = storage.list().await?;
         files.sort_by(|a, b| a.0.cmp(&b.0));
         Ok(files)
     }

libs/remote_storage/src/s3_bucket.rs

@@ -30,8 +30,8 @@ use tracing::debug;
 
 use super::StorageMetadata;
 use crate::{
-    ConcurrencyLimiter, Download, DownloadError, Listing, ListingMode, RemotePath, RemoteStorage,
-    S3Config, MAX_KEYS_PER_DELETE, REMOTE_STORAGE_PREFIX_SEPARATOR,
+    ConcurrencyLimiter, Download, DownloadError, RemotePath, RemoteStorage, S3Config,
+    MAX_KEYS_PER_DELETE, REMOTE_STORAGE_PREFIX_SEPARATOR,
 };
 
 pub(super) mod metrics;
@@ -299,13 +299,13 @@ impl<S: AsyncRead> AsyncRead for TimedDownload<S> {
 
 #[async_trait::async_trait]
 impl RemoteStorage for S3Bucket {
-    async fn list(
+    /// See the doc for `RemoteStorage::list_prefixes`
+    /// Note: it wont include empty "directories"
+    async fn list_prefixes(
         &self,
         prefix: Option<&RemotePath>,
-        mode: ListingMode,
-    ) -> Result<Listing, DownloadError> {
+    ) -> Result<Vec<RemotePath>, DownloadError> {
         let kind = RequestKind::List;
-        let mut result = Listing::default();
 
         // get the passed prefix or if it is not set use prefix_in_bucket value
         let list_prefix = prefix
@@ -314,33 +314,28 @@ impl RemoteStorage for S3Bucket {
             .map(|mut p| {
                 // required to end with a separator
                 // otherwise request will return only the entry of a prefix
-                if matches!(mode, ListingMode::WithDelimiter)
-                    && !p.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR)
-                {
+                if !p.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR) {
                     p.push(REMOTE_STORAGE_PREFIX_SEPARATOR);
                 }
                 p
             });
 
+        let mut document_keys = Vec::new();
+
         let mut continuation_token = None;
 
         loop {
             let _guard = self.permit(kind).await;
             let started_at = start_measuring_requests(kind);
 
-            let mut request = self
+            let fetch_response = self
                 .client
                 .list_objects_v2()
                 .bucket(self.bucket_name.clone())
                 .set_prefix(list_prefix.clone())
                 .set_continuation_token(continuation_token)
-                .set_max_keys(self.max_keys_per_list_response);
-
-            if let ListingMode::WithDelimiter = mode {
-                request = request.delimiter(REMOTE_STORAGE_PREFIX_SEPARATOR.to_string());
-            }
-
-            let response = request
+                .delimiter(REMOTE_STORAGE_PREFIX_SEPARATOR.to_string())
+                .set_max_keys(self.max_keys_per_list_response)
                 .send()
                 .await
                 .context("Failed to list S3 prefixes")
@@ -350,35 +345,71 @@ impl RemoteStorage for S3Bucket {
 
             metrics::BUCKET_METRICS
                 .req_seconds
-                .observe_elapsed(kind, &response, started_at);
+                .observe_elapsed(kind, &fetch_response, started_at);
 
-            let response = response?;
+            let fetch_response = fetch_response?;
 
-            let keys = response.contents().unwrap_or_default();
-            let empty = Vec::new();
-            let prefixes = response.common_prefixes.as_ref().unwrap_or(&empty);
-
-            tracing::info!("list: {} prefixes, {} keys", prefixes.len(), keys.len());
-
-            for object in keys {
-                let object_path = object.key().expect("response does not contain a key");
-                let remote_path = self.s3_object_to_relative_path(object_path);
-                result.keys.push(remote_path);
-            }
-
-            result.prefixes.extend(
-                prefixes
-                    .iter()
+            document_keys.extend(
+                fetch_response
+                    .common_prefixes
+                    .unwrap_or_default()
+                    .into_iter()
                     .filter_map(|o| Some(self.s3_object_to_relative_path(o.prefix()?))),
             );
 
-            continuation_token = match response.next_continuation_token {
+            continuation_token = match fetch_response.next_continuation_token {
                 Some(new_token) => Some(new_token),
                 None => break,
             };
         }
 
-        Ok(result)
+        Ok(document_keys)
+    }
+
+    /// See the doc for `RemoteStorage::list_files`
+    async fn list_files(&self, folder: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>> {
+        let kind = RequestKind::List;
+
+        let folder_name = folder
+            .map(|p| self.relative_path_to_s3_object(p))
+            .or_else(|| self.prefix_in_bucket.clone());
+
+        // AWS may need to break the response into several parts
+        let mut continuation_token = None;
+        let mut all_files = vec![];
+        loop {
+            let _guard = self.permit(kind).await;
+            let started_at = start_measuring_requests(kind);
+
+            let response = self
+                .client
+                .list_objects_v2()
+                .bucket(self.bucket_name.clone())
+                .set_prefix(folder_name.clone())
+                .set_continuation_token(continuation_token)
+                .set_max_keys(self.max_keys_per_list_response)
+                .send()
+                .await
+                .context("Failed to list files in S3 bucket");
+
+            let started_at = ScopeGuard::into_inner(started_at);
+            metrics::BUCKET_METRICS
+                .req_seconds
+                .observe_elapsed(kind, &response, started_at);
+
+            let response = response?;
+
+            for object in response.contents().unwrap_or_default() {
+                let object_path = object.key().expect("response does not contain a key");
+                let remote_path = self.s3_object_to_relative_path(object_path);
+                all_files.push(remote_path);
+            }
+            match response.next_continuation_token {
+                Some(new_token) => continuation_token = Some(new_token),
+                None => break,
+            }
+        }
+        Ok(all_files)
     }
 
     async fn upload(

libs/remote_storage/src/simulate_failures.rs

@@ -5,9 +5,7 @@ use std::collections::hash_map::Entry;
 use std::collections::HashMap;
 use std::sync::Mutex;
 
-use crate::{
-    Download, DownloadError, Listing, ListingMode, RemotePath, RemoteStorage, StorageMetadata,
-};
+use crate::{Download, DownloadError, RemotePath, RemoteStorage, StorageMetadata};
 
 pub struct UnreliableWrapper {
     inner: crate::GenericRemoteStorage,
@@ -97,15 +95,6 @@ impl RemoteStorage for UnreliableWrapper {
         self.inner.list_files(folder).await
     }
 
-    async fn list(
-        &self,
-        prefix: Option<&RemotePath>,
-        mode: ListingMode,
-    ) -> Result<Listing, DownloadError> {
-        self.attempt(RemoteOp::ListPrefixes(prefix.cloned()))?;
-        self.inner.list(prefix, mode).await
-    }
-
     async fn upload(
         &self,
         data: impl tokio::io::AsyncRead + Unpin + Send + Sync + 'static,

libs/safekeeper_api/src/models.rs

@@ -1,18 +1,23 @@
 use serde::{Deserialize, Serialize};
+use serde_with::{serde_as, DisplayFromStr};
 
 use utils::{
     id::{NodeId, TenantId, TimelineId},
     lsn::Lsn,
 };
 
+#[serde_as]
 #[derive(Serialize, Deserialize)]
 pub struct TimelineCreateRequest {
+    #[serde_as(as = "DisplayFromStr")]
     pub tenant_id: TenantId,
+    #[serde_as(as = "DisplayFromStr")]
     pub timeline_id: TimelineId,
     pub peer_ids: Option<Vec<NodeId>>,
     pub pg_version: u32,
     pub system_id: Option<u64>,
     pub wal_seg_size: Option<u32>,
+    #[serde_as(as = "DisplayFromStr")]
     pub commit_lsn: Lsn,
     // If not passed, it is assigned to the beginning of commit_lsn segment.
     pub local_start_lsn: Option<Lsn>,
@@ -23,6 +28,7 @@ fn lsn_invalid() -> Lsn {
 }
 
 /// Data about safekeeper's timeline, mirrors broker.proto.
+#[serde_as]
 #[derive(Debug, Clone, Deserialize, Serialize)]
 pub struct SkTimelineInfo {
     /// Term.
@@ -30,19 +36,25 @@ pub struct SkTimelineInfo {
     /// Term of the last entry.
     pub last_log_term: Option<u64>,
     /// LSN of the last record.
+    #[serde_as(as = "DisplayFromStr")]
     #[serde(default = "lsn_invalid")]
     pub flush_lsn: Lsn,
     /// Up to which LSN safekeeper regards its WAL as committed.
+    #[serde_as(as = "DisplayFromStr")]
     #[serde(default = "lsn_invalid")]
     pub commit_lsn: Lsn,
     /// LSN up to which safekeeper has backed WAL.
+    #[serde_as(as = "DisplayFromStr")]
     #[serde(default = "lsn_invalid")]
     pub backup_lsn: Lsn,
     /// LSN of last checkpoint uploaded by pageserver.
+    #[serde_as(as = "DisplayFromStr")]
     #[serde(default = "lsn_invalid")]
     pub remote_consistent_lsn: Lsn,
+    #[serde_as(as = "DisplayFromStr")]
     #[serde(default = "lsn_invalid")]
     pub peer_horizon_lsn: Lsn,
+    #[serde_as(as = "DisplayFromStr")]
     #[serde(default = "lsn_invalid")]
     pub local_start_lsn: Lsn,
     /// A connection string to use for WAL receiving.

libs/utils/Cargo.toml

@@ -55,7 +55,6 @@ bytes.workspace = true
 criterion.workspace = true
 hex-literal.workspace = true
 camino-tempfile.workspace = true
-serde_assert.workspace = true
 
 [[bench]]
 name = "benchmarks"

libs/utils/src/auth.rs

@@ -9,6 +9,7 @@ use jsonwebtoken::{
     decode, encode, Algorithm, DecodingKey, EncodingKey, Header, TokenData, Validation,
 };
 use serde::{Deserialize, Serialize};
+use serde_with::{serde_as, DisplayFromStr};
 
 use crate::id::TenantId;
 
@@ -31,9 +32,11 @@ pub enum Scope {
 }
 
 /// JWT payload. See docs/authentication.md for the format
+#[serde_as]
 #[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
 pub struct Claims {
     #[serde(default)]
+    #[serde_as(as = "Option<DisplayFromStr>")]
     pub tenant_id: Option<TenantId>,
     pub scope: Scope,
 }

libs/utils/src/generation.rs

@@ -7,7 +7,7 @@ use serde::{Deserialize, Serialize};
 ///
 /// See docs/rfcs/025-generation-numbers.md for detail on how generation
 /// numbers are used.
-#[derive(Copy, Clone, Eq, PartialEq, PartialOrd, Ord, Hash)]
+#[derive(Copy, Clone, Eq, PartialEq, PartialOrd, Ord)]
 pub enum Generation {
     // Generations with this magic value will not add a suffix to S3 keys, and will not
     // be included in persisted index_part.json.  This value is only to be used

libs/utils/src/hex.rs

@@ -1,41 +0,0 @@
-/// Useful type for asserting that expected bytes match reporting the bytes more readable
-/// array-syntax compatible hex bytes.
-///
-/// # Usage
-///
-/// ```
-/// use utils::Hex;
-///
-/// let actual = serialize_something();
-/// let expected = [0x68, 0x65, 0x6c, 0x6c, 0x6f, 0x20, 0x77, 0x6f, 0x72, 0x6c, 0x64];
-///
-/// // the type implements PartialEq and on mismatch, both sides are printed in 16 wide multiline
-/// // output suffixed with an array style length for easier comparisons.
-/// assert_eq!(Hex(&actual), Hex(&expected));
-///
-/// // with `let expected = [0x68];` the error would had been:
-/// // assertion `left == right` failed
-/// //  left: [0x68, 0x65, 0x6c, 0x6c, 0x6f, 0x20, 0x77, 0x6f, 0x72, 0x6c, 0x64; 11]
-/// // right: [0x68; 1]
-/// # fn serialize_something() -> Vec<u8> { "hello world".as_bytes().to_vec() }
-/// ```
-#[derive(PartialEq)]
-pub struct Hex<'a>(pub &'a [u8]);
-
-impl std::fmt::Debug for Hex<'_> {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        write!(f, "[")?;
-        for (i, c) in self.0.chunks(16).enumerate() {
-            if i > 0 && !c.is_empty() {
-                writeln!(f, ", ")?;
-            }
-            for (j, b) in c.iter().enumerate() {
-                if j > 0 {
-                    write!(f, ", ")?;
-                }
-                write!(f, "0x{b:02x}")?;
-            }
-        }
-        write!(f, "; {}]", self.0.len())
-    }
-}

libs/utils/src/http/endpoint.rs

@@ -14,11 +14,6 @@ use tracing::{self, debug, info, info_span, warn, Instrument};
 use std::future::Future;
 use std::str::FromStr;
 
-use bytes::{Bytes, BytesMut};
-use std::io::Write as _;
-use tokio::sync::mpsc;
-use tokio_stream::wrappers::ReceiverStream;
-
 static SERVE_METRICS_COUNT: Lazy<IntCounter> = Lazy::new(|| {
     register_int_counter!(
         "libmetrics_metric_handler_requests_total",
@@ -151,89 +146,94 @@ impl Drop for RequestCancelled {
     }
 }
 
-/// An [`std::io::Write`] implementation on top of a channel sending [`bytes::Bytes`] chunks.
-pub struct ChannelWriter {
-    buffer: BytesMut,
-    pub tx: mpsc::Sender<std::io::Result<Bytes>>,
-    written: usize,
-}
-
-impl ChannelWriter {
-    pub fn new(buf_len: usize, tx: mpsc::Sender<std::io::Result<Bytes>>) -> Self {
-        assert_ne!(buf_len, 0);
-        ChannelWriter {
-            // split about half off the buffer from the start, because we flush depending on
-            // capacity. first flush will come sooner than without this, but now resizes will
-            // have better chance of picking up the "other" half. not guaranteed of course.
-            buffer: BytesMut::with_capacity(buf_len).split_off(buf_len / 2),
-            tx,
-            written: 0,
-        }
-    }
-
-    pub fn flush0(&mut self) -> std::io::Result<usize> {
-        let n = self.buffer.len();
-        if n == 0 {
-            return Ok(0);
-        }
-
-        tracing::trace!(n, "flushing");
-        let ready = self.buffer.split().freeze();
-
-        // not ideal to call from blocking code to block_on, but we are sure that this
-        // operation does not spawn_blocking other tasks
-        let res: Result<(), ()> = tokio::runtime::Handle::current().block_on(async {
-            self.tx.send(Ok(ready)).await.map_err(|_| ())?;
-
-            // throttle sending to allow reuse of our buffer in `write`.
-            self.tx.reserve().await.map_err(|_| ())?;
-
-            // now the response task has picked up the buffer and hopefully started
-            // sending it to the client.
-            Ok(())
-        });
-        if res.is_err() {
-            return Err(std::io::ErrorKind::BrokenPipe.into());
-        }
-        self.written += n;
-        Ok(n)
-    }
-
-    pub fn flushed_bytes(&self) -> usize {
-        self.written
-    }
-}
-
-impl std::io::Write for ChannelWriter {
-    fn write(&mut self, mut buf: &[u8]) -> std::io::Result<usize> {
-        let remaining = self.buffer.capacity() - self.buffer.len();
-
-        let out_of_space = remaining < buf.len();
-
-        let original_len = buf.len();
-
-        if out_of_space {
-            let can_still_fit = buf.len() - remaining;
-            self.buffer.extend_from_slice(&buf[..can_still_fit]);
-            buf = &buf[can_still_fit..];
-            self.flush0()?;
-        }
-
-        // assume that this will often under normal operation just move the pointer back to the
-        // beginning of allocation, because previous split off parts are already sent and
-        // dropped.
-        self.buffer.extend_from_slice(buf);
-        Ok(original_len)
-    }
-
-    fn flush(&mut self) -> std::io::Result<()> {
-        self.flush0().map(|_| ())
-    }
-}
-
 async fn prometheus_metrics_handler(_req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    use bytes::{Bytes, BytesMut};
+    use std::io::Write as _;
+    use tokio::sync::mpsc;
+    use tokio_stream::wrappers::ReceiverStream;
+
     SERVE_METRICS_COUNT.inc();
 
+    /// An [`std::io::Write`] implementation on top of a channel sending [`bytes::Bytes`] chunks.
+    struct ChannelWriter {
+        buffer: BytesMut,
+        tx: mpsc::Sender<std::io::Result<Bytes>>,
+        written: usize,
+    }
+
+    impl ChannelWriter {
+        fn new(buf_len: usize, tx: mpsc::Sender<std::io::Result<Bytes>>) -> Self {
+            assert_ne!(buf_len, 0);
+            ChannelWriter {
+                // split about half off the buffer from the start, because we flush depending on
+                // capacity. first flush will come sooner than without this, but now resizes will
+                // have better chance of picking up the "other" half. not guaranteed of course.
+                buffer: BytesMut::with_capacity(buf_len).split_off(buf_len / 2),
+                tx,
+                written: 0,
+            }
+        }
+
+        fn flush0(&mut self) -> std::io::Result<usize> {
+            let n = self.buffer.len();
+            if n == 0 {
+                return Ok(0);
+            }
+
+            tracing::trace!(n, "flushing");
+            let ready = self.buffer.split().freeze();
+
+            // not ideal to call from blocking code to block_on, but we are sure that this
+            // operation does not spawn_blocking other tasks
+            let res: Result<(), ()> = tokio::runtime::Handle::current().block_on(async {
+                self.tx.send(Ok(ready)).await.map_err(|_| ())?;
+
+                // throttle sending to allow reuse of our buffer in `write`.
+                self.tx.reserve().await.map_err(|_| ())?;
+
+                // now the response task has picked up the buffer and hopefully started
+                // sending it to the client.
+                Ok(())
+            });
+            if res.is_err() {
+                return Err(std::io::ErrorKind::BrokenPipe.into());
+            }
+            self.written += n;
+            Ok(n)
+        }
+
+        fn flushed_bytes(&self) -> usize {
+            self.written
+        }
+    }
+
+    impl std::io::Write for ChannelWriter {
+        fn write(&mut self, mut buf: &[u8]) -> std::io::Result<usize> {
+            let remaining = self.buffer.capacity() - self.buffer.len();
+
+            let out_of_space = remaining < buf.len();
+
+            let original_len = buf.len();
+
+            if out_of_space {
+                let can_still_fit = buf.len() - remaining;
+                self.buffer.extend_from_slice(&buf[..can_still_fit]);
+                buf = &buf[can_still_fit..];
+                self.flush0()?;
+            }
+
+            // assume that this will often under normal operation just move the pointer back to the
+            // beginning of allocation, because previous split off parts are already sent and
+            // dropped.
+            self.buffer.extend_from_slice(buf);
+            Ok(original_len)
+        }
+
+        fn flush(&mut self) -> std::io::Result<()> {
+            self.flush0().map(|_| ())
+        }
+    }
+
     let started_at = std::time::Instant::now();
 
     let (tx, rx) = mpsc::channel(1);

libs/utils/src/id.rs

@@ -3,7 +3,6 @@ use std::{fmt, str::FromStr};
 use anyhow::Context;
 use hex::FromHex;
 use rand::Rng;
-use serde::de::Visitor;
 use serde::{Deserialize, Serialize};
 use thiserror::Error;
 
@@ -18,74 +17,12 @@ pub enum IdError {
 ///
 /// NOTE: It (de)serializes as an array of hex bytes, so the string representation would look
 /// like `[173,80,132,115,129,226,72,254,170,201,135,108,199,26,228,24]`.
-#[derive(Clone, Copy, PartialEq, Eq, Hash, PartialOrd, Ord)]
+///
+/// Use `#[serde_as(as = "DisplayFromStr")]` to (de)serialize it as hex string instead: `ad50847381e248feaac9876cc71ae418`.
+/// Check the `serde_with::serde_as` documentation for options for more complex types.
+#[derive(Clone, Copy, PartialEq, Eq, Hash, Serialize, Deserialize, PartialOrd, Ord)]
 struct Id([u8; 16]);
 
-impl Serialize for Id {
-    fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
-    where
-        S: serde::Serializer,
-    {
-        if serializer.is_human_readable() {
-            serializer.collect_str(self)
-        } else {
-            self.0.serialize(serializer)
-        }
-    }
-}
-
-impl<'de> Deserialize<'de> for Id {
-    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
-    where
-        D: serde::Deserializer<'de>,
-    {
-        struct IdVisitor {
-            is_human_readable_deserializer: bool,
-        }
-
-        impl<'de> Visitor<'de> for IdVisitor {
-            type Value = Id;
-
-            fn expecting(&self, formatter: &mut fmt::Formatter) -> fmt::Result {
-                if self.is_human_readable_deserializer {
-                    formatter.write_str("value in form of hex string")
-                } else {
-                    formatter.write_str("value in form of integer array([u8; 16])")
-                }
-            }
-
-            fn visit_seq<A>(self, seq: A) -> Result<Self::Value, A::Error>
-            where
-                A: serde::de::SeqAccess<'de>,
-            {
-                let s = serde::de::value::SeqAccessDeserializer::new(seq);
-                let id: [u8; 16] = Deserialize::deserialize(s)?;
-                Ok(Id::from(id))
-            }
-
-            fn visit_str<E>(self, v: &str) -> Result<Self::Value, E>
-            where
-                E: serde::de::Error,
-            {
-                Id::from_str(v).map_err(E::custom)
-            }
-        }
-
-        if deserializer.is_human_readable() {
-            deserializer.deserialize_str(IdVisitor {
-                is_human_readable_deserializer: true,
-            })
-        } else {
-            deserializer.deserialize_tuple(
-                16,
-                IdVisitor {
-                    is_human_readable_deserializer: false,
-                },
-            )
-        }
-    }
-}
-
 impl Id {
     pub fn get_from_buf(buf: &mut impl bytes::Buf) -> Id {
         let mut arr = [0u8; 16];
@@ -371,112 +308,3 @@ impl fmt::Display for NodeId {
         write!(f, "{}", self.0)
     }
 }
-
-#[cfg(test)]
-mod tests {
-    use serde_assert::{Deserializer, Serializer, Token, Tokens};
-
-    use crate::bin_ser::BeSer;
-
-    use super::*;
-
-    #[test]
-    fn test_id_serde_non_human_readable() {
-        let original_id = Id([
-            173, 80, 132, 115, 129, 226, 72, 254, 170, 201, 135, 108, 199, 26, 228, 24,
-        ]);
-        let expected_tokens = Tokens(vec![
-            Token::Tuple { len: 16 },
-            Token::U8(173),
-            Token::U8(80),
-            Token::U8(132),
-            Token::U8(115),
-            Token::U8(129),
-            Token::U8(226),
-            Token::U8(72),
-            Token::U8(254),
-            Token::U8(170),
-            Token::U8(201),
-            Token::U8(135),
-            Token::U8(108),
-            Token::U8(199),
-            Token::U8(26),
-            Token::U8(228),
-            Token::U8(24),
-            Token::TupleEnd,
-        ]);
-
-        let serializer = Serializer::builder().is_human_readable(false).build();
-        let serialized_tokens = original_id.serialize(&serializer).unwrap();
-        assert_eq!(serialized_tokens, expected_tokens);
-
-        let mut deserializer = Deserializer::builder()
-            .is_human_readable(false)
-            .tokens(serialized_tokens)
-            .build();
-        let deserialized_id = Id::deserialize(&mut deserializer).unwrap();
-        assert_eq!(deserialized_id, original_id);
-    }
-
-    #[test]
-    fn test_id_serde_human_readable() {
-        let original_id = Id([
-            173, 80, 132, 115, 129, 226, 72, 254, 170, 201, 135, 108, 199, 26, 228, 24,
-        ]);
-        let expected_tokens = Tokens(vec![Token::Str(String::from(
-            "ad50847381e248feaac9876cc71ae418",
-        ))]);
-
-        let serializer = Serializer::builder().is_human_readable(true).build();
-        let serialized_tokens = original_id.serialize(&serializer).unwrap();
-        assert_eq!(serialized_tokens, expected_tokens);
-
-        let mut deserializer = Deserializer::builder()
-            .is_human_readable(true)
-            .tokens(Tokens(vec![Token::Str(String::from(
-                "ad50847381e248feaac9876cc71ae418",
-            ))]))
-            .build();
-        assert_eq!(Id::deserialize(&mut deserializer).unwrap(), original_id);
-    }
-
-    macro_rules! roundtrip_type {
-        ($type:ty, $expected_bytes:expr) => {{
-            let expected_bytes: [u8; 16] = $expected_bytes;
-            let original_id = <$type>::from(expected_bytes);
-
-            let ser_bytes = original_id.ser().unwrap();
-            assert_eq!(ser_bytes, expected_bytes);
-
-            let des_id = <$type>::des(&ser_bytes).unwrap();
-            assert_eq!(des_id, original_id);
-        }};
-    }
-
-    #[test]
-    fn test_id_bincode_serde() {
-        let expected_bytes = [
-            173, 80, 132, 115, 129, 226, 72, 254, 170, 201, 135, 108, 199, 26, 228, 24,
-        ];
-
-        roundtrip_type!(Id, expected_bytes);
-    }
-
-    #[test]
-    fn test_tenant_id_bincode_serde() {
-        let expected_bytes = [
-            173, 80, 132, 115, 129, 226, 72, 254, 170, 201, 135, 108, 199, 26, 228, 24,
-        ];
-
-        roundtrip_type!(TenantId, expected_bytes);
-    }
-
-    #[test]
-    fn test_timeline_id_bincode_serde() {
-        let expected_bytes = [
-            173, 80, 132, 115, 129, 226, 72, 254, 170, 201, 135, 108, 199, 26, 228, 24,
-        ];
-
-        roundtrip_type!(TimelineId, expected_bytes);
-    }
-}

libs/utils/src/lib.rs

@@ -24,10 +24,6 @@ pub mod auth;
 
 // utility functions and helper traits for unified unique id generation/serialization etc.
 pub mod id;
-
-mod hex;
-pub use hex::Hex;
-
 // http endpoint utils
 pub mod http;
 
@@ -77,11 +73,6 @@ pub mod completion;
 /// Reporting utilities
 pub mod error;
 
-/// async timeout helper
-pub mod timeout;
-
-pub mod sync;
-
 /// This is a shortcut to embed git sha into binaries and avoid copying the same build script to all packages
 ///
 /// we have several cases:
@@ -137,21 +128,6 @@ macro_rules! project_git_version {
     };
 }
 
-/// This is a shortcut to embed build tag into binaries and avoid copying the same build script to all packages
-#[macro_export]
-macro_rules! project_build_tag {
-    ($const_identifier:ident) => {
-        const $const_identifier: &::core::primitive::str = {
-            const __ARG: &[&::core::primitive::str; 2] = &match ::core::option_env!("BUILD_TAG") {
-                ::core::option::Option::Some(x) => ["build_tag-env:", x],
-                ::core::option::Option::None => ["build_tag:", ""],
-            };
-
-            $crate::__const_format::concatcp!(__ARG[0], __ARG[1])
-        };
-    };
-}
-
 /// Re-export for `project_git_version` macro
 #[doc(hidden)]
 pub use const_format as __const_format;

libs/utils/src/lsn.rs

@@ -1,7 +1,7 @@
 #![warn(missing_docs)]
 
 use camino::Utf8Path;
-use serde::{de::Visitor, Deserialize, Serialize};
+use serde::{Deserialize, Serialize};
 use std::fmt;
 use std::ops::{Add, AddAssign};
 use std::str::FromStr;
@@ -13,114 +13,10 @@ use crate::seqwait::MonotonicCounter;
 pub const XLOG_BLCKSZ: u32 = 8192;
 
 /// A Postgres LSN (Log Sequence Number), also known as an XLogRecPtr
-#[derive(Clone, Copy, Eq, Ord, PartialEq, PartialOrd, Hash)]
+#[derive(Clone, Copy, Eq, Ord, PartialEq, PartialOrd, Hash, Serialize, Deserialize)]
+#[serde(transparent)]
 pub struct Lsn(pub u64);
 
-impl Serialize for Lsn {
-    fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
-    where
-        S: serde::Serializer,
-    {
-        if serializer.is_human_readable() {
-            serializer.collect_str(self)
-        } else {
-            self.0.serialize(serializer)
-        }
-    }
-}
-
-impl<'de> Deserialize<'de> for Lsn {
-    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
-    where
-        D: serde::Deserializer<'de>,
-    {
-        struct LsnVisitor {
-            is_human_readable_deserializer: bool,
-        }
-
-        impl<'de> Visitor<'de> for LsnVisitor {
-            type Value = Lsn;
-
-            fn expecting(&self, formatter: &mut fmt::Formatter) -> fmt::Result {
-                if self.is_human_readable_deserializer {
-                    formatter.write_str(
-                        "value in form of hex string({upper_u32_hex}/{lower_u32_hex}) representing u64 integer",
-                    )
-                } else {
-                    formatter.write_str("value in form of integer(u64)")
-                }
-            }
-
-            fn visit_u64<E>(self, v: u64) -> Result<Self::Value, E>
-            where
-                E: serde::de::Error,
-            {
-                Ok(Lsn(v))
-            }
-
-            fn visit_str<E>(self, v: &str) -> Result<Self::Value, E>
-            where
-                E: serde::de::Error,
-            {
-                Lsn::from_str(v).map_err(|e| E::custom(e))
-            }
-        }
-
-        if deserializer.is_human_readable() {
-            deserializer.deserialize_str(LsnVisitor {
-                is_human_readable_deserializer: true,
-            })
-        } else {
-            deserializer.deserialize_u64(LsnVisitor {
-                is_human_readable_deserializer: false,
-            })
-        }
-    }
-}
-
-/// Allows (de)serialization of an `Lsn` always as `u64`.
-///
-/// ### Example
-///
-/// ```rust
-/// # use serde::{Serialize, Deserialize};
-/// use utils::lsn::Lsn;
-///
-/// #[derive(PartialEq, Serialize, Deserialize, Debug)]
-/// struct Foo {
-///   #[serde(with = "utils::lsn::serde_as_u64")]
-///   always_u64: Lsn,
-/// }
-///
-/// let orig = Foo { always_u64: Lsn(1234) };
-///
-/// let res = serde_json::to_string(&orig).unwrap();
-/// assert_eq!(res, r#"{"always_u64":1234}"#);
-///
-/// let foo = serde_json::from_str::<Foo>(&res).unwrap();
-/// assert_eq!(foo, orig);
-/// ```
-///
-pub mod serde_as_u64 {
-    use super::Lsn;
-
-    /// Serializes the Lsn as u64 disregarding the human readability of the format.
-    ///
-    /// Meant to be used via `#[serde(with = "...")]` or `#[serde(serialize_with = "...")]`.
-    pub fn serialize<S: serde::Serializer>(lsn: &Lsn, serializer: S) -> Result<S::Ok, S::Error> {
-        use serde::Serialize;
-        lsn.0.serialize(serializer)
-    }
-
-    /// Deserializes the Lsn as u64 disregarding the human readability of the format.
-    ///
-    /// Meant to be used via `#[serde(with = "...")]` or `#[serde(deserialize_with = "...")]`.
-    pub fn deserialize<'de, D: serde::Deserializer<'de>>(deserializer: D) -> Result<Lsn, D::Error> {
-        use serde::Deserialize;
-        u64::deserialize(deserializer).map(Lsn)
-    }
-}
-
 /// We tried to parse an LSN from a string, but failed
 #[derive(Debug, PartialEq, Eq, thiserror::Error)]
 #[error("LsnParseError")]
@@ -368,13 +264,8 @@ impl MonotonicCounter<Lsn> for RecordLsn {
 
 #[cfg(test)]
 mod tests {
-    use crate::bin_ser::BeSer;
-
     use super::*;
 
-    use serde::ser::Serialize;
-    use serde_assert::{Deserializer, Serializer, Token, Tokens};
-
     #[test]
     fn test_lsn_strings() {
         assert_eq!("12345678/AAAA5555".parse(), Ok(Lsn(0x12345678AAAA5555)));
@@ -450,95 +341,4 @@ mod tests {
         assert_eq!(lsn.fetch_max(Lsn(6000)), Lsn(5678));
         assert_eq!(lsn.fetch_max(Lsn(5000)), Lsn(6000));
     }
-
-    #[test]
-    fn test_lsn_serde() {
-        let original_lsn = Lsn(0x0123456789abcdef);
-        let expected_readable_tokens = Tokens(vec![Token::U64(0x0123456789abcdef)]);
-        let expected_non_readable_tokens =
-            Tokens(vec![Token::Str(String::from("1234567/89ABCDEF"))]);
-
-        // Testing human_readable ser/de
-        let serializer = Serializer::builder().is_human_readable(false).build();
-        let readable_ser_tokens = original_lsn.serialize(&serializer).unwrap();
-        assert_eq!(readable_ser_tokens, expected_readable_tokens);
-
-        let mut deserializer = Deserializer::builder()
-            .is_human_readable(false)
-            .tokens(readable_ser_tokens)
-            .build();
-        let des_lsn = Lsn::deserialize(&mut deserializer).unwrap();
-        assert_eq!(des_lsn, original_lsn);
-
-        // Testing NON human_readable ser/de
-        let serializer = Serializer::builder().is_human_readable(true).build();
-        let non_readable_ser_tokens = original_lsn.serialize(&serializer).unwrap();
-        assert_eq!(non_readable_ser_tokens, expected_non_readable_tokens);
-
-        let mut deserializer = Deserializer::builder()
-            .is_human_readable(true)
-            .tokens(non_readable_ser_tokens)
-            .build();
-        let des_lsn = Lsn::deserialize(&mut deserializer).unwrap();
-        assert_eq!(des_lsn, original_lsn);
-
-        // Testing mismatching ser/de
-        let serializer = Serializer::builder().is_human_readable(false).build();
-        let non_readable_ser_tokens = original_lsn.serialize(&serializer).unwrap();
-
-        let mut deserializer = Deserializer::builder()
-            .is_human_readable(true)
-            .tokens(non_readable_ser_tokens)
-            .build();
-        Lsn::deserialize(&mut deserializer).unwrap_err();
-
-        let serializer = Serializer::builder().is_human_readable(true).build();
-        let readable_ser_tokens = original_lsn.serialize(&serializer).unwrap();
-
-        let mut deserializer = Deserializer::builder()
-            .is_human_readable(false)
-            .tokens(readable_ser_tokens)
-            .build();
-        Lsn::deserialize(&mut deserializer).unwrap_err();
-    }
-
-    #[test]
-    fn test_lsn_ensure_roundtrip() {
-        let original_lsn = Lsn(0xaaaabbbb);
-
-        let serializer = Serializer::builder().is_human_readable(false).build();
-        let ser_tokens = original_lsn.serialize(&serializer).unwrap();
-
-        let mut deserializer = Deserializer::builder()
-            .is_human_readable(false)
-            .tokens(ser_tokens)
-            .build();
-
-        let des_lsn = Lsn::deserialize(&mut deserializer).unwrap();
-        assert_eq!(des_lsn, original_lsn);
-    }
-
-    #[test]
-    fn test_lsn_bincode_serde() {
-        let lsn = Lsn(0x0123456789abcdef);
-        let expected_bytes = [0x01, 0x23, 0x45, 0x67, 0x89, 0xab, 0xcd, 0xef];
-
-        let ser_bytes = lsn.ser().unwrap();
-        assert_eq!(ser_bytes, expected_bytes);
-
-        let des_lsn = Lsn::des(&ser_bytes).unwrap();
-        assert_eq!(des_lsn, lsn);
-    }
-
-    #[test]
-    fn test_lsn_bincode_ensure_roundtrip() {
-        let original_lsn = Lsn(0x01_02_03_04_05_06_07_08);
-        let expected_bytes = vec![0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08];
-
-        let ser_bytes = original_lsn.ser().unwrap();
-        assert_eq!(ser_bytes, expected_bytes);
-
-        let des_lsn = Lsn::des(&ser_bytes).unwrap();
-        assert_eq!(des_lsn, original_lsn);
-    }
 }

libs/utils/src/pageserver_feedback.rs

@@ -3,6 +3,7 @@ use std::time::{Duration, SystemTime};
 use bytes::{Buf, BufMut, Bytes, BytesMut};
 use pq_proto::{read_cstr, PG_EPOCH};
 use serde::{Deserialize, Serialize};
+use serde_with::{serde_as, DisplayFromStr};
 use tracing::{trace, warn};
 
 use crate::lsn::Lsn;
@@ -14,17 +15,21 @@ use crate::lsn::Lsn;
 ///
 /// serde Serialize is used only for human readable dump to json (e.g. in
 /// safekeepers debug_dump).
+#[serde_as]
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
 pub struct PageserverFeedback {
     /// Last known size of the timeline. Used to enforce timeline size limit.
     pub current_timeline_size: u64,
     /// LSN last received and ingested by the pageserver. Controls backpressure.
+    #[serde_as(as = "DisplayFromStr")]
     pub last_received_lsn: Lsn,
     /// LSN up to which data is persisted by the pageserver to its local disc.
     /// Controls backpressure.
+    #[serde_as(as = "DisplayFromStr")]
     pub disk_consistent_lsn: Lsn,
     /// LSN up to which data is persisted by the pageserver on s3; safekeepers
     /// consider WAL before it can be removed.
+    #[serde_as(as = "DisplayFromStr")]
     pub remote_consistent_lsn: Lsn,
     // Serialize with RFC3339 format.
     #[serde(with = "serde_systemtime")]

libs/utils/src/sync.rs

@@ -1,3 +0,0 @@
-pub mod heavier_once_cell;
-
-pub mod gate;

libs/utils/src/sync/gate.rs

@@ -1,151 +0,0 @@
-use std::{sync::Arc, time::Duration};
-
-/// Gates are a concurrency helper, primarily used for implementing safe shutdown.
-///
-/// Users of a resource call `enter()` to acquire a GateGuard, and the owner of
-/// the resource calls `close()` when they want to ensure that all holders of guards
-/// have released them, and that no future guards will be issued.
-pub struct Gate {
-    /// Each caller of enter() takes one unit from the semaphore. In close(), we
-    /// take all the units to ensure all GateGuards are destroyed.
-    sem: Arc<tokio::sync::Semaphore>,
-
-    /// For observability only: a name that will be used to log warnings if a particular
-    /// gate is holding up shutdown
-    name: String,
-}
-
-/// RAII guard for a [`Gate`]: as long as this exists, calls to [`Gate::close`] will
-/// not complete.
-#[derive(Debug)]
-pub struct GateGuard(tokio::sync::OwnedSemaphorePermit);
-
-/// Observability helper: every `warn_period`, emit a log warning that we're still waiting on this gate
-async fn warn_if_stuck<Fut: std::future::Future>(
-    fut: Fut,
-    name: &str,
-    warn_period: std::time::Duration,
-) -> <Fut as std::future::Future>::Output {
-    let started = std::time::Instant::now();
-
-    let mut fut = std::pin::pin!(fut);
-
-    loop {
-        match tokio::time::timeout(warn_period, &mut fut).await {
-            Ok(ret) => return ret,
-            Err(_) => {
-                tracing::warn!(
-                    gate = name,
-                    elapsed_ms = started.elapsed().as_millis(),
-                    "still waiting, taking longer than expected..."
-                );
-            }
-        }
-    }
-}
-
-#[derive(Debug)]
-pub enum GateError {
-    GateClosed,
-}
-
-impl Gate {
-    const MAX_UNITS: u32 = u32::MAX;
-
-    pub fn new(name: String) -> Self {
-        Self {
-            sem: Arc::new(tokio::sync::Semaphore::new(Self::MAX_UNITS as usize)),
-            name,
-        }
-    }
-
-    /// Acquire a guard that will prevent close() calls from completing. If close()
-    /// was already called, this will return an error which should be interpreted
-    /// as "shutting down".
-    ///
-    /// This function would typically be used from e.g. request handlers. While holding
-    /// the guard returned from this function, it is important to respect a CancellationToken
-    /// to avoid blocking close() indefinitely: typically types that contain a Gate will
-    /// also contain a CancellationToken.
-    pub fn enter(&self) -> Result<GateGuard, GateError> {
-        self.sem
-            .clone()
-            .try_acquire_owned()
-            .map(GateGuard)
-            .map_err(|_| GateError::GateClosed)
-    }
-
-    /// Types with a shutdown() method and a gate should call this method at the
-    /// end of shutdown, to ensure that all GateGuard holders are done.
-    ///
-    /// This will wait for all guards to be destroyed.  For this to complete promptly, it is
-    /// important that the holders of such guards are respecting a CancellationToken which has
-    /// been cancelled before entering this function.
-    pub async fn close(&self) {
-        warn_if_stuck(self.do_close(), &self.name, Duration::from_millis(1000)).await
-    }
-
-    async fn do_close(&self) {
-        tracing::debug!(gate = self.name, "Closing Gate...");
-        match self.sem.acquire_many(Self::MAX_UNITS).await {
-            Ok(_units) => {
-                // While holding all units, close the semaphore.  All subsequent calls to enter() will fail.
-                self.sem.close();
-            }
-            Err(_) => {
-                // Semaphore closed: we are the only function that can do this, so it indicates a double-call.
-                // This is legal.  Timeline::shutdown for example is not protected from being called more than
-                // once.
-                tracing::debug!(gate = self.name, "Double close")
-            }
-        }
-        tracing::debug!(gate = self.name, "Closed Gate.")
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use futures::FutureExt;
-
-    use super::*;
-
-    #[tokio::test]
-    async fn test_idle_gate() {
-        // Having taken no gates, we should not be blocked in close
-        let gate = Gate::new("test".to_string());
-        gate.close().await;
-
-        // If a guard is dropped before entering, close should not be blocked
-        let gate = Gate::new("test".to_string());
-        let guard = gate.enter().unwrap();
-        drop(guard);
-        gate.close().await;
-
-        // Entering a closed guard fails
-        gate.enter().expect_err("enter should fail after close");
-    }
-
-    #[tokio::test]
-    async fn test_busy_gate() {
-        let gate = Gate::new("test".to_string());
-
-        let guard = gate.enter().unwrap();
-
-        let mut close_fut = std::pin::pin!(gate.close());
-
-        // Close should be blocked
-        assert!(close_fut.as_mut().now_or_never().is_none());
-
-        // Attempting to enter() should fail, even though close isn't done yet.
-        gate.enter()
-            .expect_err("enter should fail after entering close");
-
-        drop(guard);
-
-        // Guard is gone, close should finish
-        assert!(close_fut.as_mut().now_or_never().is_some());
-
-        // Attempting to enter() is still forbidden
-        gate.enter().expect_err("enter should fail finishing close");
-    }
-}

libs/utils/src/sync/heavier_once_cell.rs

@@ -1,383 +0,0 @@
-use std::sync::{
-    atomic::{AtomicUsize, Ordering},
-    Arc, Mutex, MutexGuard,
-};
-use tokio::sync::Semaphore;
-
-/// Custom design like [`tokio::sync::OnceCell`] but using [`OwnedSemaphorePermit`] instead of
-/// `SemaphorePermit`, allowing use of `take` which does not require holding an outer mutex guard
-/// for the duration of initialization.
-///
-/// Has no unsafe, builds upon [`tokio::sync::Semaphore`] and [`std::sync::Mutex`].
-///
-/// [`OwnedSemaphorePermit`]: tokio::sync::OwnedSemaphorePermit
-pub struct OnceCell<T> {
-    inner: Mutex<Inner<T>>,
-    initializers: AtomicUsize,
-}
-
-impl<T> Default for OnceCell<T> {
-    /// Create new uninitialized [`OnceCell`].
-    fn default() -> Self {
-        Self {
-            inner: Default::default(),
-            initializers: AtomicUsize::new(0),
-        }
-    }
-}
-
-/// Semaphore is the current state:
-/// - open semaphore means the value is `None`, not yet initialized
-/// - closed semaphore means the value has been initialized
-#[derive(Debug)]
-struct Inner<T> {
-    init_semaphore: Arc<Semaphore>,
-    value: Option<T>,
-}
-
-impl<T> Default for Inner<T> {
-    fn default() -> Self {
-        Self {
-            init_semaphore: Arc::new(Semaphore::new(1)),
-            value: None,
-        }
-    }
-}
-
-impl<T> OnceCell<T> {
-    /// Creates an already initialized `OnceCell` with the given value.
-    pub fn new(value: T) -> Self {
-        let sem = Semaphore::new(1);
-        sem.close();
-        Self {
-            inner: Mutex::new(Inner {
-                init_semaphore: Arc::new(sem),
-                value: Some(value),
-            }),
-            initializers: AtomicUsize::new(0),
-        }
-    }
-
-    /// Returns a guard to an existing initialized value, or uniquely initializes the value before
-    /// returning the guard.
-    ///
-    /// Initializing might wait on any existing [`Guard::take_and_deinit`] deinitialization.
-    ///
-    /// Initialization is panic-safe and cancellation-safe.
-    pub async fn get_or_init<F, Fut, E>(&self, factory: F) -> Result<Guard<'_, T>, E>
-    where
-        F: FnOnce(InitPermit) -> Fut,
-        Fut: std::future::Future<Output = Result<(T, InitPermit), E>>,
-    {
-        let sem = {
-            let guard = self.inner.lock().unwrap();
-            if guard.value.is_some() {
-                return Ok(Guard(guard));
-            }
-            guard.init_semaphore.clone()
-        };
-
-        let permit = {
-            // increment the count for the duration of queued
-            let _guard = CountWaitingInitializers::start(self);
-            sem.acquire_owned().await
-        };
-
-        match permit {
-            Ok(permit) => {
-                let permit = InitPermit(permit);
-                let (value, _permit) = factory(permit).await?;
-
-                let guard = self.inner.lock().unwrap();
-
-                Ok(Self::set0(value, guard))
-            }
-            Err(_closed) => {
-                let guard = self.inner.lock().unwrap();
-                assert!(
-                    guard.value.is_some(),
-                    "semaphore got closed, must be initialized"
-                );
-                return Ok(Guard(guard));
-            }
-        }
-    }
-
-    /// Assuming a permit is held after previous call to [`Guard::take_and_deinit`], it can be used
-    /// to complete initializing the inner value.
-    ///
-    /// # Panics
-    ///
-    /// If the inner has already been initialized.
-    pub fn set(&self, value: T, _permit: InitPermit) -> Guard<'_, T> {
-        let guard = self.inner.lock().unwrap();
-
-        // cannot assert that this permit is for self.inner.semaphore, but we can assert it cannot
-        // give more permits right now.
-        if guard.init_semaphore.try_acquire().is_ok() {
-            drop(guard);
-            panic!("permit is of wrong origin");
-        }
-
-        Self::set0(value, guard)
-    }
-
-    fn set0(value: T, mut guard: std::sync::MutexGuard<'_, Inner<T>>) -> Guard<'_, T> {
-        if guard.value.is_some() {
-            drop(guard);
-            unreachable!("we won permit, must not be initialized");
-        }
-        guard.value = Some(value);
-        guard.init_semaphore.close();
-        Guard(guard)
-    }
-
-    /// Returns a guard to an existing initialized value, if any.
-    pub fn get(&self) -> Option<Guard<'_, T>> {
-        let guard = self.inner.lock().unwrap();
-        if guard.value.is_some() {
-            Some(Guard(guard))
-        } else {
-            None
-        }
-    }
-
-    /// Return the number of [`Self::get_or_init`] calls waiting for initialization to complete.
-    pub fn initializer_count(&self) -> usize {
-        self.initializers.load(Ordering::Relaxed)
-    }
-}
-
-/// DropGuard counter for queued tasks waiting to initialize, mainly accessible for the
-/// initializing task for example at the end of initialization.
-struct CountWaitingInitializers<'a, T>(&'a OnceCell<T>);
-
-impl<'a, T> CountWaitingInitializers<'a, T> {
-    fn start(target: &'a OnceCell<T>) -> Self {
-        target.initializers.fetch_add(1, Ordering::Relaxed);
-        CountWaitingInitializers(target)
-    }
-}
-
-impl<'a, T> Drop for CountWaitingInitializers<'a, T> {
-    fn drop(&mut self) {
-        self.0.initializers.fetch_sub(1, Ordering::Relaxed);
-    }
-}
-
-/// Uninteresting guard object to allow short-lived access to inspect or clone the held,
-/// initialized value.
-#[derive(Debug)]
-pub struct Guard<'a, T>(MutexGuard<'a, Inner<T>>);
-
-impl<T> std::ops::Deref for Guard<'_, T> {
-    type Target = T;
-
-    fn deref(&self) -> &Self::Target {
-        self.0
-            .value
-            .as_ref()
-            .expect("guard is not created unless value has been initialized")
-    }
-}
-
-impl<T> std::ops::DerefMut for Guard<'_, T> {
-    fn deref_mut(&mut self) -> &mut Self::Target {
-        self.0
-            .value
-            .as_mut()
-            .expect("guard is not created unless value has been initialized")
-    }
-}
-
-impl<'a, T> Guard<'a, T> {
-    /// Take the current value, and a new permit for it's deinitialization.
-    ///
-    /// The permit will be on a semaphore part of the new internal value, and any following
-    /// [`OnceCell::get_or_init`] will wait on it to complete.
-    pub fn take_and_deinit(&mut self) -> (T, InitPermit) {
-        let mut swapped = Inner::default();
-        let permit = swapped
-            .init_semaphore
-            .clone()
-            .try_acquire_owned()
-            .expect("we just created this");
-        std::mem::swap(&mut *self.0, &mut swapped);
-        swapped
-            .value
-            .map(|v| (v, InitPermit(permit)))
-            .expect("guard is not created unless value has been initialized")
-    }
-}
-
-/// Type held by OnceCell (de)initializing task.
-pub struct InitPermit(tokio::sync::OwnedSemaphorePermit);
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-    use std::{
-        convert::Infallible,
-        sync::atomic::{AtomicUsize, Ordering},
-        time::Duration,
-    };
-
-    #[tokio::test]
-    async fn many_initializers() {
-        #[derive(Default, Debug)]
-        struct Counters {
-            factory_got_to_run: AtomicUsize,
-            future_polled: AtomicUsize,
-            winners: AtomicUsize,
-        }
-
-        let initializers = 100;
-
-        let cell = Arc::new(OnceCell::default());
-        let counters = Arc::new(Counters::default());
-        let barrier = Arc::new(tokio::sync::Barrier::new(initializers + 1));
-
-        let mut js = tokio::task::JoinSet::new();
-        for i in 0..initializers {
-            js.spawn({
-                let cell = cell.clone();
-                let counters = counters.clone();
-                let barrier = barrier.clone();
-
-                async move {
-                    barrier.wait().await;
-                    let won = {
-                        let g = cell
-                            .get_or_init(|permit| {
-                                counters.factory_got_to_run.fetch_add(1, Ordering::Relaxed);
-                                async {
-                                    counters.future_polled.fetch_add(1, Ordering::Relaxed);
-                                    Ok::<_, Infallible>((i, permit))
-                                }
-                            })
-                            .await
-                            .unwrap();
-
-                        *g == i
-                    };
-
-                    if won {
-                        counters.winners.fetch_add(1, Ordering::Relaxed);
-                    }
-                }
-            });
-        }
-
-        barrier.wait().await;
-
-        while let Some(next) = js.join_next().await {
-            next.expect("no panics expected");
-        }
-
-        let mut counters = Arc::try_unwrap(counters).unwrap();
-
-        assert_eq!(*counters.factory_got_to_run.get_mut(), 1);
-        assert_eq!(*counters.future_polled.get_mut(), 1);
-        assert_eq!(*counters.winners.get_mut(), 1);
-    }
-
-    #[tokio::test(start_paused = true)]
-    async fn reinit_waits_for_deinit() {
-        // with the tokio::time paused, we will "sleep" for 1s while holding the reinitialization
-        let sleep_for = Duration::from_secs(1);
-        let initial = 42;
-        let reinit = 1;
-        let cell = Arc::new(OnceCell::new(initial));
-
-        let deinitialization_started = Arc::new(tokio::sync::Barrier::new(2));
-
-        let jh = tokio::spawn({
-            let cell = cell.clone();
-            let deinitialization_started = deinitialization_started.clone();
-            async move {
-                let (answer, _permit) = cell.get().expect("initialized to value").take_and_deinit();
-                assert_eq!(answer, initial);
-
-                deinitialization_started.wait().await;
-                tokio::time::sleep(sleep_for).await;
-            }
-        });
-
-        deinitialization_started.wait().await;
-
-        let started_at = tokio::time::Instant::now();
-        cell.get_or_init(|permit| async { Ok::<_, Infallible>((reinit, permit)) })
-            .await
-            .unwrap();
-
-        let elapsed = started_at.elapsed();
-        assert!(
-            elapsed >= sleep_for,
-            "initialization should had taken at least the time time slept with permit"
-        );
-
-        jh.await.unwrap();
-
-        assert_eq!(*cell.get().unwrap(), reinit);
-    }
-
-    #[test]
-    fn reinit_with_deinit_permit() {
-        let cell = Arc::new(OnceCell::new(42));
-
-        let (mol, permit) = cell.get().unwrap().take_and_deinit();
-        cell.set(5, permit);
-        assert_eq!(*cell.get().unwrap(), 5);
-
-        let (five, permit) = cell.get().unwrap().take_and_deinit();
-        assert_eq!(5, five);
-        cell.set(mol, permit);
-        assert_eq!(*cell.get().unwrap(), 42);
-    }
-
-    #[tokio::test]
-    async fn initialization_attemptable_until_ok() {
-        let cell = OnceCell::default();
-
-        for _ in 0..10 {
-            cell.get_or_init(|_permit| async { Err("whatever error") })
-                .await
-                .unwrap_err();
-        }
-
-        let g = cell
-            .get_or_init(|permit| async { Ok::<_, Infallible>(("finally success", permit)) })
-            .await
-            .unwrap();
-        assert_eq!(*g, "finally success");
-    }
-
-    #[tokio::test]
-    async fn initialization_is_cancellation_safe() {
-        let cell = OnceCell::default();
-
-        let barrier = tokio::sync::Barrier::new(2);
-
-        let initializer = cell.get_or_init(|permit| async {
-            barrier.wait().await;
-            futures::future::pending::<()>().await;
-
-            Ok::<_, Infallible>(("never reached", permit))
-        });
-
-        tokio::select! {
-            _ = initializer => { unreachable!("cannot complete; stuck in pending().await") },
-            _ = barrier.wait() => {}
-        };
-
-        // now initializer is dropped
-
-        assert!(cell.get().is_none());
-
-        let g = cell
-            .get_or_init(|permit| async { Ok::<_, Infallible>(("now initialized", permit)) })
-            .await
-            .unwrap();
-        assert_eq!(*g, "now initialized");
-    }
-}

libs/utils/src/timeout.rs

@@ -1,37 +0,0 @@
-use std::time::Duration;
-
-use tokio_util::sync::CancellationToken;
-
-pub enum TimeoutCancellableError {
-    Timeout,
-    Cancelled,
-}
-
-/// Wrap [`tokio::time::timeout`] with a CancellationToken.
-///
-/// This wrapper is appropriate for any long running operation in a task
-/// that ought to respect a CancellationToken (which means most tasks).
-///
-/// The only time you should use a bare tokio::timeout is when the future `F`
-/// itself respects a CancellationToken: otherwise, always use this wrapper
-/// with your CancellationToken to ensure that your task does not hold up
-/// graceful shutdown.
-pub async fn timeout_cancellable<F>(
-    duration: Duration,
-    cancel: &CancellationToken,
-    future: F,
-) -> Result<F::Output, TimeoutCancellableError>
-where
-    F: std::future::Future,
-{
-    tokio::select!(
-        r = tokio::time::timeout(duration, future) => {
-            r.map_err(|_| TimeoutCancellableError::Timeout)
-
-        },
-        _ = cancel.cancelled() => {
-            Err(TimeoutCancellableError::Cancelled)
-
-        }
-    )
-}

libs/vm_monitor/src/filecache.rs

@@ -21,6 +21,11 @@ pub struct FileCacheState {
 
 #[derive(Debug)]
 pub struct FileCacheConfig {
+    /// Whether the file cache is *actually* stored in memory (e.g. by writing to
+    /// a tmpfs or shmem file). If true, the size of the file cache will be counted against the
+    /// memory available for the cgroup.
+    pub(crate) in_memory: bool,
+
     /// The size of the file cache, in terms of the size of the resource it consumes
     /// (currently: only memory)
     ///
@@ -54,9 +59,22 @@ pub struct FileCacheConfig {
     spread_factor: f64,
 }
 
-impl Default for FileCacheConfig {
-    fn default() -> Self {
+impl FileCacheConfig {
+    pub fn default_in_memory() -> Self {
         Self {
+            in_memory: true,
+            // 75 %
+            resource_multiplier: 0.75,
+            // 640 MiB; (512 + 128)
+            min_remaining_after_cache: NonZeroU64::new(640 * MiB).unwrap(),
+            // ensure any increase in file cache size is split 90-10 with 10% to other memory
+            spread_factor: 0.1,
+        }
+    }
+
+    pub fn default_on_disk() -> Self {
+        Self {
+            in_memory: false,
             resource_multiplier: 0.75,
             // 256 MiB - lower than when in memory because overcommitting is safe; if we don't have
             // memory, the kernel will just evict from its page cache, rather than e.g. killing
@@ -65,9 +83,7 @@ impl Default for FileCacheConfig {
             spread_factor: 0.1,
         }
     }
-}
 
-impl FileCacheConfig {
     /// Make sure fields of the config are consistent.
     pub fn validate(&self) -> anyhow::Result<()> {
         // Single field validity

libs/vm_monitor/src/lib.rs

@@ -39,6 +39,16 @@ pub struct Args {
     #[arg(short, long)]
     pub pgconnstr: Option<String>,
 
+    /// Flag to signal that the Postgres file cache is on disk (i.e. not in memory aside from the
+    /// kernel's page cache), and therefore should not count against available memory.
+    //
+    // NB: Ideally this flag would directly refer to whether the file cache is in memory (rather
+    // than a roundabout way, via whether it's on disk), but in order to be backwards compatible
+    // during the switch away from an in-memory file cache, we had to default to the previous
+    // behavior.
+    #[arg(long)]
+    pub file_cache_on_disk: bool,
+
     /// The address we should listen on for connection requests. For the
     /// agent, this is 0.0.0.0:10301. For the informant, this is 127.0.0.1:10369.
     #[arg(short, long)]

libs/vm_monitor/src/runner.rs

@@ -156,7 +156,10 @@ impl Runner {
         // memory limits.
         if let Some(connstr) = &args.pgconnstr {
             info!("initializing file cache");
-            let config = FileCacheConfig::default();
+            let config = match args.file_cache_on_disk {
+                true => FileCacheConfig::default_on_disk(),
+                false => FileCacheConfig::default_in_memory(),
+            };
 
             let mut file_cache = FileCacheState::new(connstr, config, token.clone())
                 .await
@@ -184,7 +187,10 @@ impl Runner {
                 info!("file cache size actually got set to {actual_size}")
             }
 
-            file_cache_disk_size = actual_size;
+            if args.file_cache_on_disk {
+                file_cache_disk_size = actual_size;
+            }
+
             state.filecache = Some(file_cache);
         }
 
@@ -233,11 +239,17 @@ impl Runner {
 
         let requested_mem = target.mem;
         let usable_system_memory = requested_mem.saturating_sub(self.config.sys_buffer_bytes);
-        let expected_file_cache_size = self
+        let (expected_file_cache_size, expected_file_cache_disk_size) = self
             .filecache
             .as_ref()
-            .map(|file_cache| file_cache.config.calculate_cache_size(usable_system_memory))
-            .unwrap_or(0);
+            .map(|file_cache| {
+                let size = file_cache.config.calculate_cache_size(usable_system_memory);
+                match file_cache.config.in_memory {
+                    true => (size, 0),
+                    false => (size, size),
+                }
+            })
+            .unwrap_or((0, 0));
         if let Some(cgroup) = &self.cgroup {
             let (last_time, last_history) = *cgroup.watcher.borrow();
 
@@ -261,7 +273,7 @@ impl Runner {
 
             let new_threshold = self
                 .config
-                .cgroup_threshold(usable_system_memory, expected_file_cache_size);
+                .cgroup_threshold(usable_system_memory, expected_file_cache_disk_size);
 
             let current = last_history.avg_non_reclaimable;
 
@@ -288,10 +300,13 @@ impl Runner {
                 .set_file_cache_size(expected_file_cache_size)
                 .await
                 .context("failed to set file cache size")?;
-            file_cache_disk_size = actual_usage;
+            if !file_cache.config.in_memory {
+                file_cache_disk_size = actual_usage;
+            }
             let message = format!(
-                "set file cache size to {} MiB",
+                "set file cache size to {} MiB (in memory = {})",
                 bytes_to_mebibytes(actual_usage),
+                file_cache.config.in_memory,
             );
             info!("downscale: {message}");
             status.push(message);
@@ -342,7 +357,9 @@ impl Runner {
                 .set_file_cache_size(expected_usage)
                 .await
                 .context("failed to set file cache size")?;
-            file_cache_disk_size = actual_usage;
+            if !file_cache.config.in_memory {
+                file_cache_disk_size = actual_usage;
+            }
 
             if actual_usage != expected_usage {
                 warn!(

pageserver/Cargo.toml

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

pageserver/src/bin/getpage_bench_http.rs

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

pageserver/src/bin/getpage_bench_libpq.rs

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

pageserver/src/bin/pageserver.rs

@@ -34,12 +34,11 @@ use postgres_backend::AuthType;
 use utils::logging::TracingErrorLayerEnablement;
 use utils::signals::ShutdownSignals;
 use utils::{
-    auth::JwtAuth, logging, project_build_tag, project_git_version, sentry_init::init_sentry,
-    signals::Signal, tcp_listener,
+    auth::JwtAuth, logging, project_git_version, sentry_init::init_sentry, signals::Signal,
+    tcp_listener,
 };
 
 project_git_version!(GIT_VERSION);
-project_build_tag!(BUILD_TAG);
 
 const PID_FILE_NAME: &str = "pageserver.pid";
 
@@ -259,12 +258,11 @@ fn start_pageserver(
     // A changed version string indicates changed software.
     // A changed launch timestamp indicates a pageserver restart.
     info!(
-        "version: {} launch_timestamp: {} build_tag: {}",
+        "version: {} launch_timestamp: {}",
         version(),
-        launch_ts.to_string(),
-        BUILD_TAG,
+        launch_ts.to_string()
     );
-    set_build_info_metric(GIT_VERSION, BUILD_TAG);
+    set_build_info_metric(GIT_VERSION);
     set_launch_timestamp_metric(launch_ts);
     pageserver::preinitialize_metrics();
 

pageserver/src/config.rs

@@ -33,7 +33,8 @@ use crate::disk_usage_eviction_task::DiskUsageEvictionTaskConfig;
 use crate::tenant::config::TenantConf;
 use crate::tenant::config::TenantConfOpt;
 use crate::tenant::{
-    TENANTS_SEGMENT_NAME, TENANT_DELETED_MARKER_FILE_NAME, TIMELINES_SEGMENT_NAME,
+    TENANTS_SEGMENT_NAME, TENANT_ATTACHING_MARKER_FILENAME, TENANT_DELETED_MARKER_FILE_NAME,
+    TIMELINES_SEGMENT_NAME,
 };
 use crate::{
     IGNORED_TENANT_FILE_NAME, METADATA_FILE_NAME, TENANT_CONFIG_NAME, TENANT_LOCATION_CONFIG_NAME,
@@ -632,6 +633,11 @@ impl PageServerConf {
         self.tenants_path().join(tenant_id.to_string())
     }
 
+    pub fn tenant_attaching_mark_file_path(&self, tenant_id: &TenantId) -> Utf8PathBuf {
+        self.tenant_path(tenant_id)
+            .join(TENANT_ATTACHING_MARKER_FILENAME)
+    }
+
     pub fn tenant_ignore_mark_file_path(&self, tenant_id: &TenantId) -> Utf8PathBuf {
         self.tenant_path(tenant_id).join(IGNORED_TENANT_FILE_NAME)
     }

pageserver/src/consumption_metrics.rs

@@ -266,7 +266,7 @@ async fn calculate_synthetic_size_worker(
                 continue;
             }
 
-            if let Ok(tenant) = mgr::get_tenant(tenant_id, true) {
+            if let Ok(tenant) = mgr::get_tenant(tenant_id, true).await {
                 // TODO should we use concurrent_background_tasks_rate_limit() here, like the other background tasks?
                 // We can put in some prioritization for consumption metrics.
                 // Same for the loop that fetches computed metrics.

pageserver/src/consumption_metrics/metrics.rs

@@ -3,6 +3,7 @@ use anyhow::Context;
 use chrono::{DateTime, Utc};
 use consumption_metrics::EventType;
 use futures::stream::StreamExt;
+use serde_with::serde_as;
 use std::{sync::Arc, time::SystemTime};
 use utils::{
     id::{TenantId, TimelineId},
@@ -41,10 +42,13 @@ pub(super) enum Name {
 ///
 /// This is a denormalization done at the MetricsKey const methods; these should not be constructed
 /// elsewhere.
+#[serde_with::serde_as]
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, serde::Serialize, serde::Deserialize)]
 pub(crate) struct MetricsKey {
+    #[serde_as(as = "serde_with::DisplayFromStr")]
     pub(super) tenant_id: TenantId,
 
+    #[serde_as(as = "Option<serde_with::DisplayFromStr>")]
     #[serde(skip_serializing_if = "Option::is_none")]
     pub(super) timeline_id: Option<TimelineId>,
 
@@ -202,6 +206,7 @@ pub(super) async fn collect_all_metrics(
             None
         } else {
             crate::tenant::mgr::get_tenant(id, true)
+                .await
                 .ok()
                 .map(|tenant| (id, tenant))
         }

pageserver/src/consumption_metrics/upload.rs

@@ -1,4 +1,5 @@
 use consumption_metrics::{Event, EventChunk, IdempotencyKey, CHUNK_SIZE};
+use serde_with::serde_as;
 use tokio_util::sync::CancellationToken;
 use tracing::Instrument;
 
@@ -6,9 +7,12 @@ use super::{metrics::Name, Cache, MetricsKey, RawMetric};
 use utils::id::{TenantId, TimelineId};
 
 /// How the metrics from pageserver are identified.
+#[serde_with::serde_as]
 #[derive(serde::Serialize, serde::Deserialize, Debug, Clone, Copy, PartialEq)]
 struct Ids {
+    #[serde_as(as = "serde_with::DisplayFromStr")]
     pub(super) tenant_id: TenantId,
+    #[serde_as(as = "Option<serde_with::DisplayFromStr>")]
     #[serde(skip_serializing_if = "Option::is_none")]
     pub(super) timeline_id: Option<TimelineId>,
 }

pageserver/src/control_plane_client.rs

@@ -57,10 +57,7 @@ impl ControlPlaneClient {
 
         if let Some(jwt) = &conf.control_plane_api_token {
             let mut headers = hyper::HeaderMap::new();
-            headers.insert(
-                "Authorization",
-                format!("Bearer {}", jwt.get_contents()).parse().unwrap(),
-            );
+            headers.insert("Authorization", jwt.get_contents().parse().unwrap());
             client = client.default_headers(headers);
         }
 
@@ -147,7 +144,7 @@ impl ControlPlaneGenerationsApi for ControlPlaneClient {
         Ok(response
             .tenants
             .into_iter()
-            .map(|t| (t.id, Generation::new(t.gen)))
+            .map(|t| (t.id, Generation::new(t.generation)))
             .collect::<HashMap<_, _>>())
     }
 

pageserver/src/deletion_queue.rs

@@ -10,7 +10,6 @@ use crate::control_plane_client::ControlPlaneGenerationsApi;
 use crate::metrics;
 use crate::tenant::remote_timeline_client::remote_layer_path;
 use crate::tenant::remote_timeline_client::remote_timeline_path;
-use crate::virtual_file::MaybeFatalIo;
 use crate::virtual_file::VirtualFile;
 use anyhow::Context;
 use camino::Utf8PathBuf;
@@ -18,6 +17,7 @@ use hex::FromHex;
 use remote_storage::{GenericRemoteStorage, RemotePath};
 use serde::Deserialize;
 use serde::Serialize;
+use serde_with::serde_as;
 use thiserror::Error;
 use tokio;
 use tokio_util::sync::CancellationToken;
@@ -214,6 +214,7 @@ where
 /// during recovery as startup.
 const TEMP_SUFFIX: &str = "tmp";
 
+#[serde_as]
 #[derive(Debug, Serialize, Deserialize)]
 struct DeletionList {
     /// Serialization version, for future use
@@ -242,6 +243,7 @@ struct DeletionList {
     validated: bool,
 }
 
+#[serde_as]
 #[derive(Debug, Serialize, Deserialize)]
 struct DeletionHeader {
     /// Serialization version, for future use
@@ -269,9 +271,7 @@ impl DeletionHeader {
         let temp_path = path_with_suffix_extension(&header_path, TEMP_SUFFIX);
         VirtualFile::crashsafe_overwrite(&header_path, &temp_path, &header_bytes)
             .await
-            .maybe_fatal_err("save deletion header")?;
-
-        Ok(())
+            .map_err(Into::into)
     }
 }
 
@@ -360,7 +360,6 @@ impl DeletionList {
         let bytes = serde_json::to_vec(self).expect("Failed to serialize deletion list");
         VirtualFile::crashsafe_overwrite(&path, &temp_path, &bytes)
             .await
-            .maybe_fatal_err("save deletion list")
             .map_err(Into::into)
     }
 }

pageserver/src/deletion_queue/list_writer.rs

@@ -34,8 +34,6 @@ use crate::deletion_queue::TEMP_SUFFIX;
 use crate::metrics;
 use crate::tenant::remote_timeline_client::remote_layer_path;
 use crate::tenant::storage_layer::LayerFileName;
-use crate::virtual_file::on_fatal_io_error;
-use crate::virtual_file::MaybeFatalIo;
 
 // The number of keys in a DeletionList before we will proactively persist it
 // (without reaching a flush deadline).  This aims to deliver objects of the order
@@ -197,7 +195,7 @@ impl ListWriter {
                     debug!("Deletion header {header_path} not found, first start?");
                     Ok(None)
                 } else {
-                    on_fatal_io_error(&e, "reading deletion header");
+                    Err(anyhow::anyhow!(e))
                 }
             }
         }
@@ -218,9 +216,16 @@ impl ListWriter {
         self.pending.sequence = validated_sequence + 1;
 
         let deletion_directory = self.conf.deletion_prefix();
-        let mut dir = tokio::fs::read_dir(&deletion_directory)
-            .await
-            .fatal_err("read deletion directory");
+        let mut dir = match tokio::fs::read_dir(&deletion_directory).await {
+            Ok(d) => d,
+            Err(e) => {
+                warn!("Failed to open deletion list directory {deletion_directory}: {e:#}");
+
+                // Give up: if we can't read the deletion list directory, we probably can't
+                // write lists into it later, so the queue won't work.
+                return Err(e.into());
+            }
+        };
 
         let list_name_pattern =
             Regex::new("(?<sequence>[a-zA-Z0-9]{16})-(?<version>[a-zA-Z0-9]{2}).list").unwrap();
@@ -228,7 +233,7 @@ impl ListWriter {
         let temp_extension = format!(".{TEMP_SUFFIX}");
         let header_path = self.conf.deletion_header_path();
         let mut seqs: Vec<u64> = Vec::new();
-        while let Some(dentry) = dir.next_entry().await.fatal_err("read deletion dentry") {
+        while let Some(dentry) = dir.next_entry().await? {
             let file_name = dentry.file_name();
             let dentry_str = file_name.to_string_lossy();
 
@@ -241,9 +246,11 @@ impl ListWriter {
                 info!("Cleaning up temporary file {dentry_str}");
                 let absolute_path =
                     deletion_directory.join(dentry.file_name().to_str().expect("non-Unicode path"));
-                tokio::fs::remove_file(&absolute_path)
-                    .await
-                    .fatal_err("delete temp file");
+                if let Err(e) = tokio::fs::remove_file(&absolute_path).await {
+                    // Non-fatal error: we will just leave the file behind but not
+                    // try and load it.
+                    warn!("Failed to clean up temporary file {absolute_path}: {e:#}");
+                }
 
                 continue;
             }
@@ -283,9 +290,7 @@ impl ListWriter {
         for s in seqs {
             let list_path = self.conf.deletion_list_path(s);
 
-            let list_bytes = tokio::fs::read(&list_path)
-                .await
-                .fatal_err("read deletion list");
+            let list_bytes = tokio::fs::read(&list_path).await?;
 
             let mut deletion_list = match serde_json::from_slice::<DeletionList>(&list_bytes) {
                 Ok(l) => l,
@@ -344,7 +349,7 @@ impl ListWriter {
         info!("Started deletion frontend worker");
 
         // Synchronous, but we only do it once per process lifetime so it's tolerable
-        if let Err(e) = create_dir_all(self.conf.deletion_prefix()) {
+        if let Err(e) = create_dir_all(&self.conf.deletion_prefix()) {
             tracing::error!(
                 "Failed to create deletion list directory {}, deletions will not be executed ({e})",
                 self.conf.deletion_prefix(),

pageserver/src/deletion_queue/validator.rs

@@ -28,7 +28,6 @@ use crate::config::PageServerConf;
 use crate::control_plane_client::ControlPlaneGenerationsApi;
 use crate::control_plane_client::RetryForeverError;
 use crate::metrics;
-use crate::virtual_file::MaybeFatalIo;
 
 use super::deleter::DeleterMessage;
 use super::DeletionHeader;
@@ -288,9 +287,16 @@ where
     async fn cleanup_lists(&mut self, list_paths: Vec<Utf8PathBuf>) {
         for list_path in list_paths {
             debug!("Removing deletion list {list_path}");
-            tokio::fs::remove_file(&list_path)
-                .await
-                .fatal_err("remove deletion list");
+
+            if let Err(e) = tokio::fs::remove_file(&list_path).await {
+                // Unexpected: we should have permissions and nothing else should
+                // be touching these files.  We will leave the file behind.  Subsequent
+                // pageservers will try and load it again: hopefully whatever storage
+                // issue (probably permissions) has been fixed by then.
+                tracing::error!("Failed to delete {list_path}: {e:#}");
+                metrics::DELETION_QUEUE.unexpected_errors.inc();
+                break;
+            }
         }
     }
 

pageserver/src/disk_usage_eviction_task.rs

@@ -60,11 +60,7 @@ use utils::serde_percent::Percent;
 use crate::{
     config::PageServerConf,
     task_mgr::{self, TaskKind, BACKGROUND_RUNTIME},
-    tenant::{
-        self,
-        storage_layer::{AsLayerDesc, EvictionError, Layer},
-        Timeline,
-    },
+    tenant::{self, storage_layer::PersistentLayer, timeline::EvictionError, Timeline},
 };
 
 #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
@@ -112,7 +108,7 @@ pub fn launch_disk_usage_global_eviction_task(
                 _ = background_jobs_barrier.wait() => { }
             };
 
-            disk_usage_eviction_task(&state, task_config, &storage, &conf.tenants_path(), cancel)
+            disk_usage_eviction_task(&state, task_config, storage, &conf.tenants_path(), cancel)
                 .await;
             Ok(())
         },
@@ -125,7 +121,7 @@ pub fn launch_disk_usage_global_eviction_task(
 async fn disk_usage_eviction_task(
     state: &State,
     task_config: &DiskUsageEvictionTaskConfig,
-    _storage: &GenericRemoteStorage,
+    storage: GenericRemoteStorage,
     tenants_dir: &Utf8Path,
     cancel: CancellationToken,
 ) {
@@ -149,8 +145,14 @@ async fn disk_usage_eviction_task(
         let start = Instant::now();
 
         async {
-            let res =
-                disk_usage_eviction_task_iteration(state, task_config, tenants_dir, &cancel).await;
+            let res = disk_usage_eviction_task_iteration(
+                state,
+                task_config,
+                &storage,
+                tenants_dir,
+                &cancel,
+            )
+            .await;
 
             match res {
                 Ok(()) => {}
@@ -181,12 +183,13 @@ pub trait Usage: Clone + Copy + std::fmt::Debug {
 async fn disk_usage_eviction_task_iteration(
     state: &State,
     task_config: &DiskUsageEvictionTaskConfig,
+    storage: &GenericRemoteStorage,
     tenants_dir: &Utf8Path,
     cancel: &CancellationToken,
 ) -> anyhow::Result<()> {
     let usage_pre = filesystem_level_usage::get(tenants_dir, task_config)
         .context("get filesystem-level disk usage before evictions")?;
-    let res = disk_usage_eviction_task_iteration_impl(state, usage_pre, cancel).await;
+    let res = disk_usage_eviction_task_iteration_impl(state, storage, usage_pre, cancel).await;
     match res {
         Ok(outcome) => {
             debug!(?outcome, "disk_usage_eviction_iteration finished");
@@ -270,6 +273,7 @@ struct LayerCount {
 
 pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
     state: &State,
+    storage: &GenericRemoteStorage,
     usage_pre: U,
     cancel: &CancellationToken,
 ) -> anyhow::Result<IterationOutcome<U>> {
@@ -326,10 +330,9 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
     // If we get far enough in the list that we start to evict layers that are below
     // the tenant's min-resident-size threshold, print a warning, and memorize the disk
     // usage at that point, in 'usage_planned_min_resident_size_respecting'.
-    let mut batched: HashMap<_, Vec<_>> = HashMap::new();
+    let mut batched: HashMap<_, Vec<Arc<dyn PersistentLayer>>> = HashMap::new();
     let mut warned = None;
     let mut usage_planned = usage_pre;
-    let mut max_batch_size = 0;
     for (i, (partition, candidate)) in candidates.into_iter().enumerate() {
         if !usage_planned.has_pressure() {
             debug!(
@@ -346,18 +349,10 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
 
         usage_planned.add_available_bytes(candidate.layer.layer_desc().file_size);
 
-        // FIXME: batching makes no sense anymore because of no layermap locking, should just spawn
-        // tasks to evict all seen layers until we have evicted enough
-
-        let batch = batched.entry(TimelineKey(candidate.timeline)).or_default();
-
-        // semaphore will later be used to limit eviction concurrency, and we can express at
-        // most u32 number of permits. unlikely we would have u32::MAX layers to be evicted,
-        // but fail gracefully by not making batches larger.
-        if batch.len() < u32::MAX as usize {
-            batch.push(candidate.layer);
-            max_batch_size = max_batch_size.max(batch.len());
-        }
+        batched
+            .entry(TimelineKey(candidate.timeline))
+            .or_default()
+            .push(candidate.layer);
     }
 
     let usage_planned = match warned {
@@ -374,101 +369,69 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
 
     // phase2: evict victims batched by timeline
 
-    let mut js = tokio::task::JoinSet::new();
-
-    // ratelimit to 1k files or any higher max batch size
-    let limit = Arc::new(tokio::sync::Semaphore::new(1000.max(max_batch_size)));
-
+    // After the loop, `usage_assumed` is the post-eviction usage,
+    // according to internal accounting.
+    let mut usage_assumed = usage_pre;
+    let mut evictions_failed = LayerCount::default();
     for (timeline, batch) in batched {
         let tenant_id = timeline.tenant_id;
         let timeline_id = timeline.timeline_id;
-        let batch_size =
-            u32::try_from(batch.len()).expect("batch size limited to u32::MAX during partitioning");
-
-        // I dislike naming of `available_permits` but it means current total amount of permits
-        // because permits can be added
-        assert!(batch_size as usize <= limit.available_permits());
+        let batch_size = batch.len();
 
         debug!(%timeline_id, "evicting batch for timeline");
 
-        let evict = {
-            let limit = limit.clone();
-            let cancel = cancel.clone();
-            async move {
-                let mut evicted_bytes = 0;
-                let mut evictions_failed = LayerCount::default();
+        async {
+            let results = timeline.evict_layers(storage, &batch, cancel.clone()).await;
 
-                let Ok(_permit) = limit.acquire_many_owned(batch_size).await else {
-                    // semaphore closing means cancelled
-                    return (evicted_bytes, evictions_failed);
-                };
-
-                let results = timeline.evict_layers(&batch).await;
-
-                match results {
-                    Ok(results) => {
-                        assert_eq!(results.len(), batch.len());
-                        for (result, layer) in results.into_iter().zip(batch.iter()) {
-                            let file_size = layer.layer_desc().file_size;
-                            match result {
-                                Some(Ok(())) => {
-                                    evicted_bytes += file_size;
-                                }
-                                Some(Err(EvictionError::NotFound | EvictionError::Downloaded)) => {
-                                    evictions_failed.file_sizes += file_size;
-                                    evictions_failed.count += 1;
-                                }
-                                None => {
-                                    assert!(cancel.is_cancelled());
-                                }
+            match results {
+                Err(e) => {
+                    warn!("failed to evict batch: {:#}", e);
+                }
+                Ok(results) => {
+                    assert_eq!(results.len(), batch.len());
+                    for (result, layer) in results.into_iter().zip(batch.iter()) {
+                        let file_size = layer.layer_desc().file_size;
+                        match result {
+                            Some(Ok(())) => {
+                                usage_assumed.add_available_bytes(file_size);
+                            }
+                            Some(Err(EvictionError::CannotEvictRemoteLayer)) => {
+                                unreachable!("get_local_layers_for_disk_usage_eviction finds only local layers")
+                            }
+                            Some(Err(EvictionError::FileNotFound)) => {
+                                evictions_failed.file_sizes += file_size;
+                                evictions_failed.count += 1;
+                            }
+                            Some(Err(
+                                e @ EvictionError::LayerNotFound(_)
+                                | e @ EvictionError::StatFailed(_),
+                            )) => {
+                                let e = utils::error::report_compact_sources(&e);
+                                warn!(%layer, "failed to evict layer: {e}");
+                                evictions_failed.file_sizes += file_size;
+                                evictions_failed.count += 1;
+                            }
+                            Some(Err(EvictionError::MetadataInconsistency(detail))) => {
+                                warn!(%layer, "failed to evict layer: {detail}");
+                                evictions_failed.file_sizes += file_size;
+                                evictions_failed.count += 1;
+                            }
+                            None => {
+                                assert!(cancel.is_cancelled());
+                                return;
                             }
                         }
                     }
-                    Err(e) => {
-                        warn!("failed to evict batch: {:#}", e);
-                    }
                 }
-                (evicted_bytes, evictions_failed)
             }
         }
-        .instrument(tracing::info_span!("evict_batch", %tenant_id, %timeline_id, batch_size));
+        .instrument(tracing::info_span!("evict_batch", %tenant_id, %timeline_id, batch_size))
+        .await;
 
-        js.spawn(evict);
-
-        // spwaning multiple thousands of these is essentially blocking, so give already spawned a
-        // chance of making progress
-        tokio::task::yield_now().await;
-    }
-
-    let join_all = async move {
-        // After the evictions, `usage_assumed` is the post-eviction usage,
-        // according to internal accounting.
-        let mut usage_assumed = usage_pre;
-        let mut evictions_failed = LayerCount::default();
-
-        while let Some(res) = js.join_next().await {
-            match res {
-                Ok((evicted_bytes, failed)) => {
-                    usage_assumed.add_available_bytes(evicted_bytes);
-                    evictions_failed.file_sizes += failed.file_sizes;
-                    evictions_failed.count += failed.count;
-                }
-                Err(je) if je.is_cancelled() => unreachable!("not used"),
-                Err(je) if je.is_panic() => { /* already logged */ }
-                Err(je) => tracing::error!("unknown JoinError: {je:?}"),
-            }
-        }
-        (usage_assumed, evictions_failed)
-    };
-
-    let (usage_assumed, evictions_failed) = tokio::select! {
-        tuple = join_all => { tuple },
-        _ = cancel.cancelled() => {
-            // close the semaphore to stop any pending acquires
-            limit.close();
+        if cancel.is_cancelled() {
             return Ok(IterationOutcome::Cancelled);
         }
-    };
+    }
 
     Ok(IterationOutcome::Finished(IterationOutcomeFinished {
         before: usage_pre,
@@ -483,7 +446,7 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
 #[derive(Clone)]
 struct EvictionCandidate {
     timeline: Arc<Timeline>,
-    layer: Layer,
+    layer: Arc<dyn PersistentLayer>,
     last_activity_ts: SystemTime,
 }
 
@@ -545,7 +508,7 @@ async fn collect_eviction_candidates(
         if cancel.is_cancelled() {
             return Ok(EvictionCandidates::Cancelled);
         }
-        let tenant = match tenant::mgr::get_tenant(*tenant_id, true) {
+        let tenant = match tenant::mgr::get_tenant(*tenant_id, true).await {
             Ok(tenant) => tenant,
             Err(e) => {
                 // this can happen if tenant has lifecycle transition after we fetched it
@@ -554,11 +517,6 @@ async fn collect_eviction_candidates(
             }
         };
 
-        if tenant.cancel.is_cancelled() {
-            info!(%tenant_id, "Skipping tenant for eviction, it is shutting down");
-            continue;
-        }
-
         // collect layers from all timelines in this tenant
         //
         // If one of the timelines becomes `!is_active()` during the iteration,

pageserver/src/http/openapi_spec.yml

@@ -569,17 +569,7 @@ paths:
               schema:
                 $ref: "#/components/schemas/NotFoundError"
         "409":
-          description: |
-            The tenant is already known to Pageserver in some way,
-            and hence this `/attach` call has been rejected.
-
-            Some examples of how this can happen:
-            - tenant was created on this pageserver
-            - tenant attachment was started by an earlier call to `/attach`.
-
-            Callers should poll the tenant status's `attachment_status` field,
-            like for status 202. See the longer description for `POST /attach`
-            for details.
+          description: Tenant download is already in progress
           content:
             application/json:
               schema:
@@ -723,12 +713,6 @@ paths:
 
         Errors if the tenant is absent on disk, already present in memory or fails to schedule its load.
         Scheduling a load does not mean that the tenant would load successfully, check tenant status to ensure load correctness.
-      requestBody:
-        required: false
-        content:
-          application/json:
-            schema:
-              $ref: "#/components/schemas/TenantLoadRequest"
       responses:
         "202":
           description: Tenant scheduled to load successfully
@@ -1219,15 +1203,6 @@ components:
             new_tenant_id:
               type: string
               format: hex
-            generation:
-              type: integer
-              description: Attachment generation number.
-    TenantLoadRequest:
-      type: object
-      properties:
-        generation:
-          type: integer
-          description: Attachment generation number.
     TenantAttachRequest:
       type: object
       required:

pageserver/src/http/routes.rs

@@ -17,6 +17,7 @@ use pageserver_api::models::{
     TenantLoadRequest, TenantLocationConfigRequest,
 };
 use remote_storage::GenericRemoteStorage;
+use serde_with::{serde_as, DisplayFromStr};
 use tenant_size_model::{SizeResult, StorageModel};
 use tokio_util::sync::CancellationToken;
 use tracing::*;
@@ -35,8 +36,7 @@ use crate::pgdatadir_mapping::LsnForTimestamp;
 use crate::task_mgr::TaskKind;
 use crate::tenant::config::{LocationConf, TenantConfOpt};
 use crate::tenant::mgr::{
-    GetTenantError, SetNewTenantConfigError, TenantMapError, TenantMapInsertError, TenantSlotError,
-    TenantSlotUpsertError, TenantStateError,
+    GetTenantError, SetNewTenantConfigError, TenantMapInsertError, TenantStateError,
 };
 use crate::tenant::size::ModelInputs;
 use crate::tenant::storage_layer::LayerAccessStatsReset;
@@ -147,59 +147,28 @@ impl From<PageReconstructError> for ApiError {
 impl From<TenantMapInsertError> for ApiError {
     fn from(tmie: TenantMapInsertError) -> ApiError {
         match tmie {
-            TenantMapInsertError::SlotError(e) => e.into(),
-            TenantMapInsertError::SlotUpsertError(e) => e.into(),
+            TenantMapInsertError::StillInitializing | TenantMapInsertError::ShuttingDown => {
+                ApiError::ResourceUnavailable(format!("{tmie}").into())
+            }
+            TenantMapInsertError::TenantAlreadyExists(id, state) => {
+                ApiError::Conflict(format!("tenant {id} already exists, state: {state:?}"))
+            }
+            TenantMapInsertError::TenantExistsSecondary(id) => {
+                ApiError::Conflict(format!("tenant {id} already exists as secondary"))
+            }
             TenantMapInsertError::Other(e) => ApiError::InternalServerError(e),
         }
     }
 }
 
-impl From<TenantSlotError> for ApiError {
-    fn from(e: TenantSlotError) -> ApiError {
-        use TenantSlotError::*;
-        match e {
-            NotFound(tenant_id) => {
-                ApiError::NotFound(anyhow::anyhow!("NotFound: tenant {tenant_id}").into())
-            }
-            e @ (AlreadyExists(_, _) | Conflict(_)) => ApiError::Conflict(format!("{e}")),
-            InProgress => {
-                ApiError::ResourceUnavailable("Tenant is being modified concurrently".into())
-            }
-            MapState(e) => e.into(),
-        }
-    }
-}
-
-impl From<TenantSlotUpsertError> for ApiError {
-    fn from(e: TenantSlotUpsertError) -> ApiError {
-        use TenantSlotUpsertError::*;
-        match e {
-            InternalError(e) => ApiError::InternalServerError(anyhow::anyhow!("{e}")),
-            MapState(e) => e.into(),
-        }
-    }
-}
-
-impl From<TenantMapError> for ApiError {
-    fn from(e: TenantMapError) -> ApiError {
-        use TenantMapError::*;
-        match e {
-            StillInitializing | ShuttingDown => {
-                ApiError::ResourceUnavailable(format!("{e}").into())
-            }
-        }
-    }
-}
-
 impl From<TenantStateError> for ApiError {
     fn from(tse: TenantStateError) -> ApiError {
         match tse {
+            TenantStateError::NotFound(tid) => ApiError::NotFound(anyhow!("tenant {}", tid).into()),
             TenantStateError::IsStopping(_) => {
                 ApiError::ResourceUnavailable("Tenant is stopping".into())
             }
-            TenantStateError::SlotError(e) => e.into(),
-            TenantStateError::SlotUpsertError(e) => e.into(),
-            TenantStateError::Other(e) => ApiError::InternalServerError(anyhow!(e)),
+            _ => ApiError::InternalServerError(anyhow::Error::new(tse)),
         }
     }
 }
@@ -220,7 +189,6 @@ impl From<GetTenantError> for ApiError {
                 // (We can produce this variant only in `mgr::get_tenant(..., active=true)` calls).
                 ApiError::ResourceUnavailable("Tenant not yet active".into())
             }
-            GetTenantError::MapState(e) => ApiError::ResourceUnavailable(format!("{e}").into()),
         }
     }
 }
@@ -275,9 +243,6 @@ impl From<crate::tenant::delete::DeleteTenantError> for ApiError {
             Get(g) => ApiError::from(g),
             e @ AlreadyInProgress => ApiError::Conflict(e.to_string()),
             Timeline(t) => ApiError::from(t),
-            NotAttached => ApiError::NotFound(anyhow::anyhow!("Tenant is not attached").into()),
-            SlotError(e) => e.into(),
-            SlotUpsertError(e) => e.into(),
             Other(o) => ApiError::InternalServerError(o),
             e @ InvalidState(_) => ApiError::PreconditionFailed(e.to_string().into_boxed_str()),
         }
@@ -404,7 +369,7 @@ async fn timeline_create_handler(
     let state = get_state(&request);
 
     async {
-        let tenant = mgr::get_tenant(tenant_id, true)?;
+        let tenant = mgr::get_tenant(tenant_id, true).await?;
         match tenant.create_timeline(
             new_timeline_id,
             request_data.ancestor_timeline_id.map(TimelineId::from),
@@ -432,9 +397,6 @@ async fn timeline_create_handler(
             Err(e @ tenant::CreateTimelineError::AncestorNotActive) => {
                 json_response(StatusCode::SERVICE_UNAVAILABLE, HttpErrorBody::from_msg(e.to_string()))
             }
-            Err(tenant::CreateTimelineError::ShuttingDown) => {
-                json_response(StatusCode::SERVICE_UNAVAILABLE,HttpErrorBody::from_msg("tenant shutting down".to_string()))
-            }
             Err(tenant::CreateTimelineError::Other(err)) => Err(ApiError::InternalServerError(err)),
         }
     }
@@ -454,7 +416,7 @@ async fn timeline_list_handler(
     let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
 
     let response_data = async {
-        let tenant = mgr::get_tenant(tenant_id, true)?;
+        let tenant = mgr::get_tenant(tenant_id, true).await?;
         let timelines = tenant.list_timelines();
 
         let mut response_data = Vec::with_capacity(timelines.len());
@@ -493,7 +455,7 @@ async fn timeline_detail_handler(
     let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
 
     let timeline_info = async {
-        let tenant = mgr::get_tenant(tenant_id, true)?;
+        let tenant = mgr::get_tenant(tenant_id, true).await?;
 
         let timeline = tenant
             .get_timeline(timeline_id, false)
@@ -537,8 +499,10 @@ async fn get_lsn_by_timestamp_handler(
     let result = timeline.find_lsn_for_timestamp(timestamp_pg, &ctx).await?;
 
     if version.unwrap_or(0) > 1 {
+        #[serde_as]
         #[derive(serde::Serialize)]
         struct Result {
+            #[serde_as(as = "DisplayFromStr")]
             lsn: Lsn,
             kind: &'static str,
         }
@@ -717,6 +681,45 @@ async fn tenant_ignore_handler(
     json_response(StatusCode::OK, ())
 }
 
+async fn tenant_duplicate_handler(
+    mut request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let src_tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
+
+    let request_data: TenantCreateRequest = json_request(&mut request).await?;
+    let new_tenant_id = request_data.new_tenant_id;
+    check_permission(&request, None)?;
+
+    let _timer = STORAGE_TIME_GLOBAL
+        .get_metric_with_label_values(&[StorageTimeOperation::DuplicateTenant.into()])
+        .expect("bug")
+        .start_timer();
+
+    let tenant_conf =
+        TenantConfOpt::try_from(&request_data.config).map_err(ApiError::BadRequest)?;
+
+    let state = get_state(&request);
+
+    let generation = get_request_generation(state, request_data.generation)?;
+
+    let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Warn);
+
+    mgr::duplicate_tenant(
+        state.conf,
+        tenant_conf,
+        src_tenant_id,
+        new_tenant_id,
+        generation,
+        state.tenant_resources(),
+        &ctx,
+    )
+    .instrument(info_span!("tenant_duplicate", %src_tenant_id, tenant_id = %new_tenant_id))
+    .await?;
+
+    json_response(StatusCode::CREATED, TenantCreateResponse(new_tenant_id))
+}
+
 async fn tenant_list_handler(
     request: Request<Body>,
     _cancel: CancellationToken,
@@ -749,7 +752,7 @@ async fn tenant_status(
     check_permission(&request, Some(tenant_id))?;
 
     let tenant_info = async {
-        let tenant = mgr::get_tenant(tenant_id, false)?;
+        let tenant = mgr::get_tenant(tenant_id, false).await?;
 
         // Calculate total physical size of all timelines
         let mut current_physical_size = 0;
@@ -812,7 +815,7 @@ async fn tenant_size_handler(
     let headers = request.headers();
 
     let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
-    let tenant = mgr::get_tenant(tenant_id, true)?;
+    let tenant = mgr::get_tenant(tenant_id, true).await?;
 
     // this can be long operation
     let inputs = tenant
@@ -847,8 +850,10 @@ async fn tenant_size_handler(
     }
 
     /// The type resides in the pageserver not to expose `ModelInputs`.
+    #[serde_with::serde_as]
     #[derive(serde::Serialize)]
     struct TenantHistorySize {
+        #[serde_as(as = "serde_with::DisplayFromStr")]
         id: TenantId,
         /// Size is a mixture of WAL and logical size, so the unit is bytes.
         ///
@@ -1069,7 +1074,7 @@ async fn get_tenant_config_handler(
     let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
     check_permission(&request, Some(tenant_id))?;
 
-    let tenant = mgr::get_tenant(tenant_id, false)?;
+    let tenant = mgr::get_tenant(tenant_id, false).await?;
 
     let response = HashMap::from([
         (
@@ -1128,7 +1133,7 @@ async fn put_tenant_location_config_handler(
             .await
         {
             match e {
-                TenantStateError::SlotError(TenantSlotError::NotFound(_)) => {
+                TenantStateError::NotFound(_) => {
                     // This API is idempotent: a NotFound on a detach is fine.
                 }
                 _ => return Err(e.into()),
@@ -1166,6 +1171,7 @@ async fn handle_tenant_break(
     let tenant_id: TenantId = parse_request_param(&r, "tenant_id")?;
 
     let tenant = crate::tenant::mgr::get_tenant(tenant_id, true)
+        .await
         .map_err(|_| ApiError::Conflict(String::from("no active tenant found")))?;
 
     tenant.set_broken("broken from test".to_owned()).await;
@@ -1238,7 +1244,7 @@ async fn timeline_compact_handler(
         timeline
             .compact(&cancel, &ctx)
             .await
-            .map_err(|e| ApiError::InternalServerError(e.into()))?;
+            .map_err(ApiError::InternalServerError)?;
         json_response(StatusCode::OK, ())
     }
     .instrument(info_span!("manual_compaction", %tenant_id, %timeline_id))
@@ -1263,7 +1269,7 @@ async fn timeline_checkpoint_handler(
         timeline
             .compact(&cancel, &ctx)
             .await
-            .map_err(|e| ApiError::InternalServerError(e.into()))?;
+            .map_err(ApiError::InternalServerError)?;
 
         json_response(StatusCode::OK, ())
     }
@@ -1470,7 +1476,7 @@ async fn active_timeline_of_active_tenant(
     tenant_id: TenantId,
     timeline_id: TimelineId,
 ) -> Result<Arc<Timeline>, ApiError> {
-    let tenant = mgr::get_tenant(tenant_id, true)?;
+    let tenant = mgr::get_tenant(tenant_id, true).await?;
     tenant
         .get_timeline(timeline_id, true)
         .map_err(|e| ApiError::NotFound(e.into()))
@@ -1533,11 +1539,11 @@ async fn disk_usage_eviction_run(
 
     let state = get_state(&r);
 
-    if state.remote_storage.as_ref().is_none() {
+    let Some(storage) = state.remote_storage.clone() else {
         return Err(ApiError::InternalServerError(anyhow::anyhow!(
             "remote storage not configured, cannot run eviction iteration"
         )));
-    }
+    };
 
     let state = state.disk_usage_eviction_state.clone();
 
@@ -1555,6 +1561,7 @@ async fn disk_usage_eviction_run(
         async move {
             let res = crate::disk_usage_eviction_task::disk_usage_eviction_task_iteration_impl(
                 &state,
+                &storage,
                 usage,
                 &child_cancel,
             )
@@ -1760,6 +1767,9 @@ pub fn make_router(
         .post("/v1/tenant/:tenant_id/ignore", |r| {
             api_handler(r, tenant_ignore_handler)
         })
+        .post("/v1/tenant/:tenant_id/duplicate", |r| {
+            api_handler(r, tenant_duplicate_handler)
+        })
         .get("/v1/tenant/:tenant_id/timeline/:timeline_id", |r| {
             api_handler(r, timeline_detail_handler)
         })

pageserver/src/lib.rs

@@ -61,6 +61,14 @@ pub async fn shutdown_pageserver(deletion_queue: Option<DeletionQueue>, exit_cod
     )
     .await;
 
+    // Shut down any page service tasks.
+    timed(
+        task_mgr::shutdown_tasks(Some(TaskKind::PageRequestHandler), None, None),
+        "shutdown PageRequestHandlers",
+        Duration::from_secs(1),
+    )
+    .await;
+
     // Shut down all the tenants. This flushes everything to disk and kills
     // the checkpoint and GC tasks.
     timed(
@@ -70,15 +78,6 @@ pub async fn shutdown_pageserver(deletion_queue: Option<DeletionQueue>, exit_cod
     )
     .await;
 
-    // Shut down any page service tasks: any in-progress work for particular timelines or tenants
-    // should already have been canclled via mgr::shutdown_all_tenants
-    timed(
-        task_mgr::shutdown_tasks(Some(TaskKind::PageRequestHandler), None, None),
-        "shutdown PageRequestHandlers",
-        Duration::from_secs(1),
-    )
-    .await;
-
     // Best effort to persist any outstanding deletions, to avoid leaking objects
     if let Some(mut deletion_queue) = deletion_queue {
         deletion_queue.shutdown(Duration::from_secs(5)).await;

pageserver/src/metrics.rs

@@ -51,6 +51,9 @@ pub enum StorageTimeOperation {
 
     #[strum(serialize = "create tenant")]
     CreateTenant,
+
+    #[strum(serialize = "duplicate tenant")]
+    DuplicateTenant,
 }
 
 pub static STORAGE_TIME_SUM_PER_TIMELINE: Lazy<CounterVec> = Lazy::new(|| {
@@ -962,32 +965,6 @@ static REMOTE_TIMELINE_CLIENT_BYTES_FINISHED_COUNTER: Lazy<IntCounterVec> = Lazy
     .expect("failed to define a metric")
 });
 
-pub(crate) struct TenantManagerMetrics {
-    pub(crate) tenant_slots: UIntGauge,
-    pub(crate) tenant_slot_writes: IntCounter,
-    pub(crate) unexpected_errors: IntCounter,
-}
-
-pub(crate) static TENANT_MANAGER: Lazy<TenantManagerMetrics> = Lazy::new(|| {
-    TenantManagerMetrics {
-    tenant_slots: register_uint_gauge!(
-        "pageserver_tenant_manager_slots",
-        "How many slots currently exist, including all attached, secondary and in-progress operations",
-    )
-    .expect("failed to define a metric"),
-    tenant_slot_writes: register_int_counter!(
-        "pageserver_tenant_manager_slot_writes",
-        "Writes to a tenant slot, including all of create/attach/detach/delete"
-    )
-    .expect("failed to define a metric"),
-    unexpected_errors: register_int_counter!(
-        "pageserver_tenant_manager_unexpected_errors_total",
-        "Number of unexpected conditions encountered: nonzero value indicates a non-fatal bug."
-    )
-    .expect("failed to define a metric"),
-}
-});
-
 pub(crate) struct DeletionQueueMetrics {
     pub(crate) keys_submitted: IntCounter,
     pub(crate) keys_dropped: IntCounter,
@@ -1430,7 +1407,7 @@ impl TimelineMetrics {
         crate::metrics::RESIDENT_PHYSICAL_SIZE_GLOBAL.add(sz);
     }
 
-    pub(crate) fn resident_physical_size_get(&self) -> u64 {
+    pub fn resident_physical_size_get(&self) -> u64 {
         self.resident_physical_size_gauge.get()
     }
 }
@@ -1910,9 +1887,6 @@ pub fn preinitialize_metrics() {
     // Deletion queue stats
     Lazy::force(&DELETION_QUEUE);
 
-    // Tenant manager stats
-    Lazy::force(&TENANT_MANAGER);
-
     // countervecs
     [&BACKGROUND_LOOP_PERIOD_OVERRUN_COUNT]
         .into_iter()

pageserver/src/page_service.rs

@@ -14,6 +14,7 @@ use async_compression::tokio::write::GzipEncoder;
 use bytes::Buf;
 use bytes::Bytes;
 use futures::Stream;
+use pageserver_api::models::TenantState;
 use pageserver_api::models::{
     PagestreamBeMessage, PagestreamDbSizeRequest, PagestreamDbSizeResponse,
     PagestreamErrorResponse, PagestreamExistsRequest, PagestreamExistsResponse,
@@ -54,20 +55,16 @@ use crate::metrics;
 use crate::metrics::LIVE_CONNECTIONS_COUNT;
 use crate::task_mgr;
 use crate::task_mgr::TaskKind;
+use crate::tenant;
 use crate::tenant::debug_assert_current_span_has_tenant_and_timeline_id;
 use crate::tenant::mgr;
-use crate::tenant::mgr::get_active_tenant_with_timeout;
-use crate::tenant::mgr::GetActiveTenantError;
-use crate::tenant::Timeline;
+use crate::tenant::mgr::GetTenantError;
+use crate::tenant::{Tenant, Timeline};
 use crate::trace::Tracer;
 
 use postgres_ffi::pg_constants::DEFAULTTABLESPACE_OID;
 use postgres_ffi::BLCKSZ;
 
-// How long we may block waiting for a [`TenantSlot::InProgress`]` and/or a [`Tenant`] which
-// is not yet in state [`TenantState::Active`].
-const ACTIVE_TENANT_TIMEOUT: Duration = Duration::from_millis(5000);
-
 /// Read the end of a tar archive.
 ///
 /// A tar archive normally ends with two consecutive blocks of zeros, 512 bytes each.
@@ -226,7 +223,13 @@ async fn page_service_conn_main(
     // and create a child per-query context when it invokes process_query.
     // But it's in a shared crate, so, we store connection_ctx inside PageServerHandler
     // and create the per-query context in process_query ourselves.
-    let mut conn_handler = PageServerHandler::new(conf, broker_client, auth, connection_ctx);
+    let mut conn_handler = PageServerHandler::new(
+        conf,
+        broker_client,
+        auth,
+        connection_ctx,
+        task_mgr::shutdown_token(),
+    );
     let pgbackend = PostgresBackend::new_from_io(socket, peer_addr, auth_type, None)?;
 
     match pgbackend
@@ -260,6 +263,10 @@ struct PageServerHandler {
     /// For each query received over the connection,
     /// `process_query` creates a child context from this one.
     connection_ctx: RequestContext,
+
+    /// A token that should fire when the tenant transitions from
+    /// attached state, or when the pageserver is shutting down.
+    cancel: CancellationToken,
 }
 
 impl PageServerHandler {
@@ -268,6 +275,7 @@ impl PageServerHandler {
         broker_client: storage_broker::BrokerClientChannel,
         auth: Option<Arc<JwtAuth>>,
         connection_ctx: RequestContext,
+        cancel: CancellationToken,
     ) -> Self {
         PageServerHandler {
             _conf: conf,
@@ -275,6 +283,7 @@ impl PageServerHandler {
             auth,
             claims: None,
             connection_ctx,
+            cancel,
         }
     }
 
@@ -282,11 +291,7 @@ impl PageServerHandler {
     /// this rather than naked flush() in order to shut down promptly.  Without this, we would
     /// block shutdown of a tenant if a postgres client was failing to consume bytes we send
     /// in the flush.
-    async fn flush_cancellable<IO>(
-        &self,
-        pgb: &mut PostgresBackend<IO>,
-        cancel: &CancellationToken,
-    ) -> Result<(), QueryError>
+    async fn flush_cancellable<IO>(&self, pgb: &mut PostgresBackend<IO>) -> Result<(), QueryError>
     where
         IO: AsyncRead + AsyncWrite + Send + Sync + Unpin,
     {
@@ -294,7 +299,7 @@ impl PageServerHandler {
             flush_r = pgb.flush() => {
                 Ok(flush_r?)
             },
-            _ = cancel.cancelled() => {
+            _ = self.cancel.cancelled() => {
                 Err(QueryError::Shutdown)
             }
         )
@@ -303,7 +308,6 @@ impl PageServerHandler {
     fn copyin_stream<'a, IO>(
         &'a self,
         pgb: &'a mut PostgresBackend<IO>,
-        cancel: &'a CancellationToken,
     ) -> impl Stream<Item = io::Result<Bytes>> + 'a
     where
         IO: AsyncRead + AsyncWrite + Send + Sync + Unpin,
@@ -313,7 +317,7 @@ impl PageServerHandler {
                 let msg = tokio::select! {
                     biased;
 
-                    _ = cancel.cancelled() => {
+                    _ = self.cancel.cancelled() => {
                         // We were requested to shut down.
                         let msg = "pageserver is shutting down";
                         let _ = pgb.write_message_noflush(&BeMessage::ErrorResponse(msg, None));
@@ -353,7 +357,7 @@ impl PageServerHandler {
                         let query_error = QueryError::Disconnected(ConnectionError::Io(io::Error::new(io::ErrorKind::ConnectionReset, msg)));
                         // error can't happen here, ErrorResponse serialization should be always ok
                         pgb.write_message_noflush(&BeMessage::ErrorResponse(msg, Some(query_error.pg_error_code()))).map_err(|e| e.into_io_error())?;
-                        self.flush_cancellable(pgb, cancel).await.map_err(|e| io::Error::new(io::ErrorKind::Other, e.to_string()))?;
+                        self.flush_cancellable(pgb).await.map_err(|e| io::Error::new(io::ErrorKind::Other, e.to_string()))?;
                         Err(io::Error::new(io::ErrorKind::ConnectionReset, msg))?;
                     }
                     Err(QueryError::Disconnected(ConnectionError::Io(io_error))) => {
@@ -380,13 +384,12 @@ impl PageServerHandler {
     {
         debug_assert_current_span_has_tenant_and_timeline_id();
 
+        // NOTE: pagerequests handler exits when connection is closed,
+        //       so there is no need to reset the association
+        task_mgr::associate_with(Some(tenant_id), Some(timeline_id));
+
         // Make request tracer if needed
-        let tenant = mgr::get_active_tenant_with_timeout(
-            tenant_id,
-            ACTIVE_TENANT_TIMEOUT,
-            &task_mgr::shutdown_token(),
-        )
-        .await?;
+        let tenant = get_active_tenant_with_timeout(tenant_id, &ctx).await?;
         let mut tracer = if tenant.get_trace_read_requests() {
             let connection_id = ConnectionId::generate();
             let path = tenant
@@ -402,14 +405,9 @@ impl PageServerHandler {
             .get_timeline(timeline_id, true)
             .map_err(|e| anyhow::anyhow!(e))?;
 
-        // Avoid starting new requests if the timeline has already started shutting down,
-        // and block timeline shutdown until this request is complete, or drops out due
-        // to cancellation.
-        let _timeline_guard = timeline.gate.enter().map_err(|_| QueryError::Shutdown)?;
-
         // switch client to COPYBOTH
         pgb.write_message_noflush(&BeMessage::CopyBothResponse)?;
-        self.flush_cancellable(pgb, &timeline.cancel).await?;
+        self.flush_cancellable(pgb).await?;
 
         let metrics = metrics::SmgrQueryTimePerTimeline::new(&tenant_id, &timeline_id);
 
@@ -417,7 +415,7 @@ impl PageServerHandler {
             let msg = tokio::select! {
                 biased;
 
-                _ = timeline.cancel.cancelled() => {
+                _ = self.cancel.cancelled() => {
                     // We were requested to shut down.
                     info!("shutdown request received in page handler");
                     return Err(QueryError::Shutdown)
@@ -492,20 +490,9 @@ impl PageServerHandler {
                 }
             };
 
-            if let Err(e) = &response {
-                if timeline.cancel.is_cancelled() {
-                    // If we fail to fulfil a request during shutdown, which may be _because_ of
-                    // shutdown, then do not send the error to the client.  Instead just drop the
-                    // connection.
-                    span.in_scope(|| info!("dropped response during shutdown: {e:#}"));
-                    return Err(QueryError::Shutdown);
-                }
-            }
-
             let response = response.unwrap_or_else(|e| {
                 // print the all details to the log with {:#}, but for the client the
-                // error message is enough.  Do not log if shutting down, as the anyhow::Error
-                // here includes cancellation which is not an error.
+                // error message is enough
                 span.in_scope(|| error!("error reading relation or page version: {:#}", e));
                 PagestreamBeMessage::Error(PagestreamErrorResponse {
                     message: e.to_string(),
@@ -513,7 +500,7 @@ impl PageServerHandler {
             });
 
             pgb.write_message_noflush(&BeMessage::CopyData(&response.serialize()))?;
-            self.flush_cancellable(pgb, &timeline.cancel).await?;
+            self.flush_cancellable(pgb).await?;
         }
         Ok(())
     }
@@ -535,14 +522,10 @@ impl PageServerHandler {
     {
         debug_assert_current_span_has_tenant_and_timeline_id();
 
+        task_mgr::associate_with(Some(tenant_id), Some(timeline_id));
         // Create empty timeline
         info!("creating new timeline");
-        let tenant = get_active_tenant_with_timeout(
-            tenant_id,
-            ACTIVE_TENANT_TIMEOUT,
-            &task_mgr::shutdown_token(),
-        )
-        .await?;
+        let tenant = get_active_tenant_with_timeout(tenant_id, &ctx).await?;
         let timeline = tenant
             .create_empty_timeline(timeline_id, base_lsn, pg_version, &ctx)
             .await?;
@@ -560,9 +543,9 @@ impl PageServerHandler {
         // Import basebackup provided via CopyData
         info!("importing basebackup");
         pgb.write_message_noflush(&BeMessage::CopyInResponse)?;
-        self.flush_cancellable(pgb, &tenant.cancel).await?;
+        self.flush_cancellable(pgb).await?;
 
-        let mut copyin_reader = pin!(StreamReader::new(self.copyin_stream(pgb, &tenant.cancel)));
+        let mut copyin_reader = pin!(StreamReader::new(self.copyin_stream(pgb)));
         timeline
             .import_basebackup_from_tar(
                 &mut copyin_reader,
@@ -599,10 +582,9 @@ impl PageServerHandler {
         IO: AsyncRead + AsyncWrite + Send + Sync + Unpin,
     {
         debug_assert_current_span_has_tenant_and_timeline_id();
+        task_mgr::associate_with(Some(tenant_id), Some(timeline_id));
 
-        let timeline = self
-            .get_active_tenant_timeline(tenant_id, timeline_id)
-            .await?;
+        let timeline = get_active_tenant_timeline(tenant_id, timeline_id, &ctx).await?;
         let last_record_lsn = timeline.get_last_record_lsn();
         if last_record_lsn != start_lsn {
             return Err(QueryError::Other(
@@ -616,8 +598,8 @@ impl PageServerHandler {
         // Import wal provided via CopyData
         info!("importing wal");
         pgb.write_message_noflush(&BeMessage::CopyInResponse)?;
-        self.flush_cancellable(pgb, &timeline.cancel).await?;
-        let mut copyin_reader = pin!(StreamReader::new(self.copyin_stream(pgb, &timeline.cancel)));
+        self.flush_cancellable(pgb).await?;
+        let mut copyin_reader = pin!(StreamReader::new(self.copyin_stream(pgb)));
         import_wal_from_tar(&timeline, &mut copyin_reader, start_lsn, end_lsn, &ctx).await?;
         info!("wal import complete");
 
@@ -810,9 +792,7 @@ impl PageServerHandler {
         let started = std::time::Instant::now();
 
         // check that the timeline exists
-        let timeline = self
-            .get_active_tenant_timeline(tenant_id, timeline_id)
-            .await?;
+        let timeline = get_active_tenant_timeline(tenant_id, timeline_id, &ctx).await?;
         let latest_gc_cutoff_lsn = timeline.get_latest_gc_cutoff_lsn();
         if let Some(lsn) = lsn {
             // Backup was requested at a particular LSN. Wait for it to arrive.
@@ -827,7 +807,7 @@ impl PageServerHandler {
 
         // switch client to COPYOUT
         pgb.write_message_noflush(&BeMessage::CopyOutResponse)?;
-        self.flush_cancellable(pgb, &timeline.cancel).await?;
+        self.flush_cancellable(pgb).await?;
 
         // Send a tarball of the latest layer on the timeline. Compress if not
         // fullbackup. TODO Compress in that case too (tests need to be updated)
@@ -879,7 +859,7 @@ impl PageServerHandler {
         }
 
         pgb.write_message_noflush(&BeMessage::CopyDone)?;
-        self.flush_cancellable(pgb, &timeline.cancel).await?;
+        self.flush_cancellable(pgb).await?;
 
         let basebackup_after = started
             .elapsed()
@@ -911,25 +891,6 @@ impl PageServerHandler {
             .expect("claims presence already checked");
         check_permission(claims, tenant_id)
     }
-
-    /// Shorthand for getting a reference to a Timeline of an Active tenant.
-    async fn get_active_tenant_timeline(
-        &self,
-        tenant_id: TenantId,
-        timeline_id: TimelineId,
-    ) -> Result<Arc<Timeline>, GetActiveTimelineError> {
-        let tenant = get_active_tenant_with_timeout(
-            tenant_id,
-            ACTIVE_TENANT_TIMEOUT,
-            &task_mgr::shutdown_token(),
-        )
-        .await
-        .map_err(GetActiveTimelineError::Tenant)?;
-        let timeline = tenant
-            .get_timeline(timeline_id, true)
-            .map_err(|e| GetActiveTimelineError::Timeline(anyhow::anyhow!(e)))?;
-        Ok(timeline)
-    }
 }
 
 #[async_trait::async_trait]
@@ -1087,9 +1048,7 @@ where
                 .record("timeline_id", field::display(timeline_id));
 
             self.check_permission(Some(tenant_id))?;
-            let timeline = self
-                .get_active_tenant_timeline(tenant_id, timeline_id)
-                .await?;
+            let timeline = get_active_tenant_timeline(tenant_id, timeline_id, &ctx).await?;
 
             let end_of_timeline = timeline.get_last_record_rlsn();
 
@@ -1273,12 +1232,7 @@ where
 
             self.check_permission(Some(tenant_id))?;
 
-            let tenant = get_active_tenant_with_timeout(
-                tenant_id,
-                ACTIVE_TENANT_TIMEOUT,
-                &task_mgr::shutdown_token(),
-            )
-            .await?;
+            let tenant = get_active_tenant_with_timeout(tenant_id, &ctx).await?;
             pgb.write_message_noflush(&BeMessage::RowDescription(&[
                 RowDescriptor::int8_col(b"checkpoint_distance"),
                 RowDescriptor::int8_col(b"checkpoint_timeout"),
@@ -1324,13 +1278,67 @@ where
     }
 }
 
+#[derive(thiserror::Error, Debug)]
+enum GetActiveTenantError {
+    #[error(
+        "Timed out waiting {wait_time:?} for tenant active state. Latest state: {latest_state:?}"
+    )]
+    WaitForActiveTimeout {
+        latest_state: TenantState,
+        wait_time: Duration,
+    },
+    #[error(transparent)]
+    NotFound(GetTenantError),
+    #[error(transparent)]
+    WaitTenantActive(tenant::WaitToBecomeActiveError),
+}
+
 impl From<GetActiveTenantError> for QueryError {
     fn from(e: GetActiveTenantError) -> Self {
         match e {
             GetActiveTenantError::WaitForActiveTimeout { .. } => QueryError::Disconnected(
                 ConnectionError::Io(io::Error::new(io::ErrorKind::TimedOut, e.to_string())),
             ),
-            e => QueryError::Other(anyhow::anyhow!(e)),
+            GetActiveTenantError::WaitTenantActive(e) => QueryError::Other(anyhow::Error::new(e)),
+            GetActiveTenantError::NotFound(e) => QueryError::Other(anyhow::Error::new(e)),
+        }
+    }
+}
+
+/// Get active tenant.
+///
+/// If the tenant is Loading, waits for it to become Active, for up to 30 s. That
+/// ensures that queries don't fail immediately after pageserver startup, because
+/// all tenants are still loading.
+async fn get_active_tenant_with_timeout(
+    tenant_id: TenantId,
+    _ctx: &RequestContext, /* require get a context to support cancellation in the future */
+) -> Result<Arc<Tenant>, GetActiveTenantError> {
+    let tenant = match mgr::get_tenant(tenant_id, false).await {
+        Ok(tenant) => tenant,
+        Err(e @ GetTenantError::NotFound(_)) => return Err(GetActiveTenantError::NotFound(e)),
+        Err(GetTenantError::NotActive(_)) => {
+            unreachable!("we're calling get_tenant with active_only=false")
+        }
+        Err(GetTenantError::Broken(_)) => {
+            unreachable!("we're calling get_tenant with active_only=false")
+        }
+    };
+    let wait_time = Duration::from_secs(30);
+    match tokio::time::timeout(wait_time, tenant.wait_to_become_active()).await {
+        Ok(Ok(())) => Ok(tenant),
+        // no .context(), the error message is good enough and some tests depend on it
+        Ok(Err(e)) => Err(GetActiveTenantError::WaitTenantActive(e)),
+        Err(_) => {
+            let latest_state = tenant.current_state();
+            if latest_state == TenantState::Active {
+                Ok(tenant)
+            } else {
+                Err(GetActiveTenantError::WaitForActiveTimeout {
+                    latest_state,
+                    wait_time,
+                })
+            }
         }
     }
 }
@@ -1351,3 +1359,18 @@ impl From<GetActiveTimelineError> for QueryError {
         }
     }
 }
+
+/// Shorthand for getting a reference to a Timeline of an Active tenant.
+async fn get_active_tenant_timeline(
+    tenant_id: TenantId,
+    timeline_id: TimelineId,
+    ctx: &RequestContext,
+) -> Result<Arc<Timeline>, GetActiveTimelineError> {
+    let tenant = get_active_tenant_with_timeout(tenant_id, ctx)
+        .await
+        .map_err(GetActiveTimelineError::Tenant)?;
+    let timeline = tenant
+        .get_timeline(timeline_id, true)
+        .map_err(|e| GetActiveTimelineError::Timeline(anyhow::anyhow!(e)))?;
+    Ok(timeline)
+}

pageserver/src/pgdatadir_mapping.rs

@@ -44,17 +44,6 @@ pub enum CalculateLogicalSizeError {
     Other(#[from] anyhow::Error),
 }
 
-impl From<PageReconstructError> for CalculateLogicalSizeError {
-    fn from(pre: PageReconstructError) -> Self {
-        match pre {
-            PageReconstructError::AncestorStopping(_) | PageReconstructError::Cancelled => {
-                Self::Cancelled
-            }
-            _ => Self::Other(pre.into()),
-        }
-    }
-}
-
 #[derive(Debug, thiserror::Error)]
 pub enum RelationError {
     #[error("Relation Already Exists")]
@@ -563,8 +552,7 @@ impl Timeline {
                 Err(e) => Err(PageReconstructError::from(e)),
             },
             Err(e) => {
-                // This is expected: historical databases do not have the key.
-                debug!("Failed to get info about AUX files: {}", e);
+                warn!("Failed to get info about AUX files: {}", e);
                 Ok(HashMap::new())
             }
         }
@@ -584,7 +572,7 @@ impl Timeline {
         crate::tenant::debug_assert_current_span_has_tenant_and_timeline_id();
 
         // Fetch list of database dirs and iterate them
-        let buf = self.get(DBDIR_KEY, lsn, ctx).await?;
+        let buf = self.get(DBDIR_KEY, lsn, ctx).await.context("read dbdir")?;
         let dbdir = DbDirectory::des(&buf).context("deserialize db directory")?;
 
         let mut total_size: u64 = 0;
@@ -598,7 +586,10 @@ impl Timeline {
                     return Err(CalculateLogicalSizeError::Cancelled);
                 }
                 let relsize_key = rel_size_to_key(rel);
-                let mut buf = self.get(relsize_key, lsn, ctx).await?;
+                let mut buf = self
+                    .get(relsize_key, lsn, ctx)
+                    .await
+                    .with_context(|| format!("read relation size of {rel:?}"))?;
                 let relsize = buf.get_u32_le();
 
                 total_size += relsize as u64;
@@ -1211,8 +1202,7 @@ impl<'a> DatadirModification<'a> {
         let mut dir = match self.get(AUX_FILES_KEY, ctx).await {
             Ok(buf) => AuxFilesDirectory::des(&buf)?,
             Err(e) => {
-                // This is expected: historical databases do not have the key.
-                debug!("Failed to get info about AUX files: {}", e);
+                warn!("Failed to get info about AUX files: {}", e);
                 AuxFilesDirectory {
                     files: HashMap::new(),
                 }
@@ -1704,6 +1694,7 @@ const AUX_FILES_KEY: Key = Key {
 // Reverse mappings for a few Keys.
 // These are needed by WAL redo manager.
 
+/// Guaranteed to return `Ok()` if [[is_rel_block_key]] returns `true` for `key`.
 pub fn key_to_rel_block(key: Key) -> anyhow::Result<(RelTag, BlockNumber)> {
     Ok(match key.field1 {
         0x00 => (
@@ -1719,7 +1710,8 @@ pub fn key_to_rel_block(key: Key) -> anyhow::Result<(RelTag, BlockNumber)> {
     })
 }
 
-fn is_rel_block_key(key: Key) -> bool {
+/// See [[key_to_rel_block]].
+pub fn is_rel_block_key(key: Key) -> bool {
     key.field1 == 0x00 && key.field4 != 0
 }
 

pageserver/src/task_mgr.rs

@@ -299,6 +299,10 @@ pub enum TaskKind {
 
 #[derive(Default)]
 struct MutableTaskState {
+    /// Tenant and timeline that this task is associated with.
+    tenant_id: Option<TenantId>,
+    timeline_id: Option<TimelineId>,
+
     /// Handle for waiting for the task to exit. It can be None, if the
     /// the task has already exited.
     join_handle: Option<JoinHandle<()>>,
@@ -315,11 +319,6 @@ struct PageServerTask {
     // To request task shutdown, just cancel this token.
     cancel: CancellationToken,
 
-    /// Tasks may optionally be launched for a particular tenant/timeline, enabling
-    /// later cancelling tasks for that tenant/timeline in [`shutdown_tasks`]
-    tenant_id: Option<TenantId>,
-    timeline_id: Option<TimelineId>,
-
     mutable: Mutex<MutableTaskState>,
 }
 
@@ -345,9 +344,11 @@ where
         kind,
         name: name.to_string(),
         cancel: cancel.clone(),
-        tenant_id,
-        timeline_id,
-        mutable: Mutex::new(MutableTaskState { join_handle: None }),
+        mutable: Mutex::new(MutableTaskState {
+            tenant_id,
+            timeline_id,
+            join_handle: None,
+        }),
     });
 
     TASKS.lock().unwrap().insert(task_id, Arc::clone(&task));
@@ -417,6 +418,8 @@ async fn task_finish(
 
     let mut shutdown_process = false;
     {
+        let task_mut = task.mutable.lock().unwrap();
+
         match result {
             Ok(Ok(())) => {
                 debug!("Task '{}' exited normally", task_name);
@@ -425,13 +428,13 @@ async fn task_finish(
                 if shutdown_process_on_error {
                     error!(
                         "Shutting down: task '{}' tenant_id: {:?}, timeline_id: {:?} exited with error: {:?}",
-                        task_name, task.tenant_id, task.timeline_id, err
+                        task_name, task_mut.tenant_id, task_mut.timeline_id, err
                     );
                     shutdown_process = true;
                 } else {
                     error!(
                         "Task '{}' tenant_id: {:?}, timeline_id: {:?} exited with error: {:?}",
-                        task_name, task.tenant_id, task.timeline_id, err
+                        task_name, task_mut.tenant_id, task_mut.timeline_id, err
                     );
                 }
             }
@@ -439,13 +442,13 @@ async fn task_finish(
                 if shutdown_process_on_error {
                     error!(
                         "Shutting down: task '{}' tenant_id: {:?}, timeline_id: {:?} panicked: {:?}",
-                        task_name, task.tenant_id, task.timeline_id, err
+                        task_name, task_mut.tenant_id, task_mut.timeline_id, err
                     );
                     shutdown_process = true;
                 } else {
                     error!(
                         "Task '{}' tenant_id: {:?}, timeline_id: {:?} panicked: {:?}",
-                        task_name, task.tenant_id, task.timeline_id, err
+                        task_name, task_mut.tenant_id, task_mut.timeline_id, err
                     );
                 }
             }
@@ -457,6 +460,17 @@ async fn task_finish(
     }
 }
 
+// expected to be called from the task of the given id.
+pub fn associate_with(tenant_id: Option<TenantId>, timeline_id: Option<TimelineId>) {
+    CURRENT_TASK.with(|ct| {
+        let mut task_mut = ct.mutable.lock().unwrap();
+        task_mut.tenant_id = tenant_id;
+        task_mut.timeline_id = timeline_id;
+    });
+}
+
+/// Is there a task running that matches the criteria
+
 /// Signal and wait for tasks to shut down.
 ///
 ///
@@ -479,16 +493,17 @@ pub async fn shutdown_tasks(
     {
         let tasks = TASKS.lock().unwrap();
         for task in tasks.values() {
+            let task_mut = task.mutable.lock().unwrap();
             if (kind.is_none() || Some(task.kind) == kind)
-                && (tenant_id.is_none() || task.tenant_id == tenant_id)
-                && (timeline_id.is_none() || task.timeline_id == timeline_id)
+                && (tenant_id.is_none() || task_mut.tenant_id == tenant_id)
+                && (timeline_id.is_none() || task_mut.timeline_id == timeline_id)
             {
                 task.cancel.cancel();
                 victim_tasks.push((
                     Arc::clone(task),
                     task.kind,
-                    task.tenant_id,
-                    task.timeline_id,
+                    task_mut.tenant_id,
+                    task_mut.timeline_id,
                 ));
             }
         }

pageserver/src/tenant/delete.rs

@@ -3,10 +3,10 @@ use std::sync::Arc;
 use anyhow::Context;
 use camino::{Utf8Path, Utf8PathBuf};
 use pageserver_api::models::TenantState;
-use remote_storage::{GenericRemoteStorage, RemotePath};
+use remote_storage::{DownloadError, GenericRemoteStorage, RemotePath};
 use tokio::sync::OwnedMutexGuard;
 use tokio_util::sync::CancellationToken;
-use tracing::{error, instrument, warn, Instrument, Span};
+use tracing::{error, info, instrument, warn, Instrument, Span};
 
 use utils::{
     backoff, completion, crashsafe, fs_ext,
@@ -21,33 +21,26 @@ use crate::{
 };
 
 use super::{
-    mgr::{GetTenantError, TenantSlotError, TenantSlotUpsertError, TenantsMap},
+    mgr::{GetTenantError, TenantsMap},
     remote_timeline_client::{FAILED_REMOTE_OP_RETRIES, FAILED_UPLOAD_WARN_THRESHOLD},
     span,
     timeline::delete::DeleteTimelineFlow,
-    tree_sort_timelines, DeleteTimelineError, Tenant, TenantPreload,
+    tree_sort_timelines, DeleteTimelineError, Tenant,
 };
 
+const SHOULD_RESUME_DELETION_FETCH_MARK_ATTEMPTS: u32 = 3;
+
 #[derive(Debug, thiserror::Error)]
 pub(crate) enum DeleteTenantError {
     #[error("GetTenant {0}")]
     Get(#[from] GetTenantError),
 
-    #[error("Tenant not attached")]
-    NotAttached,
-
     #[error("Invalid state {0}. Expected Active or Broken")]
     InvalidState(TenantState),
 
     #[error("Tenant deletion is already in progress")]
     AlreadyInProgress,
 
-    #[error("Tenant map slot error {0}")]
-    SlotError(#[from] TenantSlotError),
-
-    #[error("Tenant map slot upsert error {0}")]
-    SlotUpsertError(#[from] TenantSlotUpsertError),
-
     #[error("Timeline {0}")]
     Timeline(#[from] DeleteTimelineError),
 
@@ -67,7 +60,7 @@ fn remote_tenant_delete_mark_path(
         .context("Failed to strip workdir prefix")
         .and_then(RemotePath::new)
         .context("tenant path")?;
-    Ok(tenant_remote_path.join(Utf8Path::new("timelines/deleted")))
+    Ok(tenant_remote_path.join(Utf8Path::new("deleted")))
 }
 
 async fn create_remote_delete_mark(
@@ -157,8 +150,7 @@ async fn ensure_timelines_dir_empty(timelines_path: &Utf8Path) -> Result<(), Del
     // Assert timelines dir is empty.
     if !fs_ext::is_directory_empty(timelines_path).await? {
         // Display first 10 items in directory
-        let list = fs_ext::list_dir(timelines_path).await.context("list_dir")?;
-        let list = &list.into_iter().take(10).collect::<Vec<_>>();
+        let list = &fs_ext::list_dir(timelines_path).await.context("list_dir")?[..10];
         return Err(DeleteTenantError::Other(anyhow::anyhow!(
             "Timelines directory is not empty after all timelines deletion: {list:?}"
         )));
@@ -247,6 +239,32 @@ async fn cleanup_remaining_fs_traces(
     Ok(())
 }
 
+pub(crate) async fn remote_delete_mark_exists(
+    conf: &PageServerConf,
+    tenant_id: &TenantId,
+    remote_storage: &GenericRemoteStorage,
+) -> anyhow::Result<bool> {
+    // If remote storage is there we rely on it
+    let remote_mark_path = remote_tenant_delete_mark_path(conf, tenant_id).context("path")?;
+
+    let result = backoff::retry(
+        || async { remote_storage.download(&remote_mark_path).await },
+        |e| matches!(e, DownloadError::NotFound),
+        SHOULD_RESUME_DELETION_FETCH_MARK_ATTEMPTS,
+        SHOULD_RESUME_DELETION_FETCH_MARK_ATTEMPTS,
+        "fetch_tenant_deletion_mark",
+        // TODO: use a cancellation token (https://github.com/neondatabase/neon/issues/5066)
+        backoff::Cancel::new(CancellationToken::new(), || unreachable!()),
+    )
+    .await;
+
+    match result {
+        Ok(_) => Ok(true),
+        Err(DownloadError::NotFound) => Ok(false),
+        Err(e) => Err(anyhow::anyhow!(e)).context("remote_delete_mark_exists")?,
+    }
+}
+
 /// Orchestrates tenant shut down of all tasks, removes its in-memory structures,
 /// and deletes its data from both disk and s3.
 /// The sequence of steps:
@@ -258,9 +276,10 @@ async fn cleanup_remaining_fs_traces(
 /// 6. Remove remote mark
 /// 7. Cleanup remaining fs traces, tenant dir, config, timelines dir, local delete mark
 /// It is resumable from any step in case a crash/restart occurs.
-/// There are two entrypoints to the process:
+/// There are three entrypoints to the process:
 /// 1. [`DeleteTenantFlow::run`] this is the main one called by a management api handler.
-/// 2. [`DeleteTenantFlow::resume_from_attach`] is called when deletion is resumed tenant is found to be deleted during attach process.
+/// 2. [`DeleteTenantFlow::resume_from_load`] is called during restarts when local or remote deletion marks are still there.
+/// 3. [`DeleteTenantFlow::resume_from_attach`] is called when deletion is resumed tenant is found to be deleted during attach process.
 ///  Note the only other place that messes around timeline delete mark is the `Tenant::spawn_load` function.
 #[derive(Default)]
 pub enum DeleteTenantFlow {
@@ -282,12 +301,12 @@ impl DeleteTenantFlow {
     pub(crate) async fn run(
         conf: &'static PageServerConf,
         remote_storage: Option<GenericRemoteStorage>,
-        tenants: &'static std::sync::RwLock<TenantsMap>,
-        tenant: Arc<Tenant>,
+        tenants: &'static tokio::sync::RwLock<TenantsMap>,
+        tenant_id: TenantId,
     ) -> Result<(), DeleteTenantError> {
         span::debug_assert_current_span_has_tenant_id();
 
-        let mut guard = Self::prepare(&tenant).await?;
+        let (tenant, mut guard) = Self::prepare(tenants, tenant_id).await?;
 
         if let Err(e) = Self::run_inner(&mut guard, conf, remote_storage.as_ref(), &tenant).await {
             tenant.set_broken(format!("{e:#}")).await;
@@ -359,7 +378,7 @@ impl DeleteTenantFlow {
 
     pub(crate) async fn should_resume_deletion(
         conf: &'static PageServerConf,
-        remote_mark_exists: bool,
+        remote_storage: Option<&GenericRemoteStorage>,
         tenant: &Tenant,
     ) -> Result<Option<DeletionGuard>, DeleteTenantError> {
         let acquire = |t: &Tenant| {
@@ -370,25 +389,66 @@ impl DeleteTenantFlow {
             )
         };
 
-        if remote_mark_exists {
-            return Ok(acquire(tenant));
-        }
-
         let tenant_id = tenant.tenant_id;
         // Check local mark first, if its there there is no need to go to s3 to check whether remote one exists.
         if conf.tenant_deleted_mark_file_path(&tenant_id).exists() {
+            return Ok(acquire(tenant));
+        }
+
+        let remote_storage = match remote_storage {
+            Some(remote_storage) => remote_storage,
+            None => return Ok(None),
+        };
+
+        if remote_delete_mark_exists(conf, &tenant_id, remote_storage).await? {
             Ok(acquire(tenant))
         } else {
             Ok(None)
         }
     }
 
+    pub(crate) async fn resume_from_load(
+        guard: DeletionGuard,
+        tenant: &Arc<Tenant>,
+        init_order: Option<&InitializationOrder>,
+        tenants: &'static tokio::sync::RwLock<TenantsMap>,
+        ctx: &RequestContext,
+    ) -> Result<(), DeleteTenantError> {
+        let (_, progress) = completion::channel();
+
+        tenant
+            .set_stopping(progress, true, false)
+            .await
+            .expect("cant be stopping or broken");
+
+        // Do not consume valuable resources during the load phase, continue deletion once init phase is complete.
+        let background_jobs_can_start = init_order.as_ref().map(|x| &x.background_jobs_can_start);
+        if let Some(background) = background_jobs_can_start {
+            info!("waiting for backgound jobs barrier");
+            background.clone().wait().await;
+            info!("ready for backgound jobs barrier");
+        }
+
+        // Tenant may not be loadable if we fail late in cleanup_remaining_fs_traces (e g remove timelines dir)
+        let timelines_path = tenant.conf.timelines_path(&tenant.tenant_id);
+        if timelines_path.exists() {
+            tenant.load(init_order, None, ctx).await.context("load")?;
+        }
+
+        Self::background(
+            guard,
+            tenant.conf,
+            tenant.remote_storage.clone(),
+            tenants,
+            tenant,
+        )
+        .await
+    }
+
     pub(crate) async fn resume_from_attach(
         guard: DeletionGuard,
         tenant: &Arc<Tenant>,
-        preload: Option<TenantPreload>,
-        tenants: &'static std::sync::RwLock<TenantsMap>,
-        init_order: Option<InitializationOrder>,
+        tenants: &'static tokio::sync::RwLock<TenantsMap>,
         ctx: &RequestContext,
     ) -> Result<(), DeleteTenantError> {
         let (_, progress) = completion::channel();
@@ -399,7 +459,7 @@ impl DeleteTenantFlow {
             .expect("cant be stopping or broken");
 
         tenant
-            .attach(init_order, preload, ctx)
+            .attach(ctx, super::AttachMarkerMode::Expect)
             .await
             .context("attach")?;
 
@@ -414,8 +474,15 @@ impl DeleteTenantFlow {
     }
 
     async fn prepare(
-        tenant: &Arc<Tenant>,
-    ) -> Result<tokio::sync::OwnedMutexGuard<Self>, DeleteTenantError> {
+        tenants: &tokio::sync::RwLock<TenantsMap>,
+        tenant_id: TenantId,
+    ) -> Result<(Arc<Tenant>, tokio::sync::OwnedMutexGuard<Self>), DeleteTenantError> {
+        let m = tenants.read().await;
+
+        let tenant = m
+            .get(&tenant_id)
+            .ok_or(GetTenantError::NotFound(tenant_id))?;
+
         // FIXME: unsure about active only. Our init jobs may not be cancellable properly,
         // so at least for now allow deletions only for active tenants. TODO recheck
         // Broken and Stopping is needed for retries.
@@ -449,14 +516,14 @@ impl DeleteTenantFlow {
             )));
         }
 
-        Ok(guard)
+        Ok((Arc::clone(tenant), guard))
     }
 
     fn schedule_background(
         guard: OwnedMutexGuard<Self>,
         conf: &'static PageServerConf,
         remote_storage: Option<GenericRemoteStorage>,
-        tenants: &'static std::sync::RwLock<TenantsMap>,
+        tenants: &'static tokio::sync::RwLock<TenantsMap>,
         tenant: Arc<Tenant>,
     ) {
         let tenant_id = tenant.tenant_id;
@@ -489,7 +556,7 @@ impl DeleteTenantFlow {
         mut guard: OwnedMutexGuard<Self>,
         conf: &PageServerConf,
         remote_storage: Option<GenericRemoteStorage>,
-        tenants: &'static std::sync::RwLock<TenantsMap>,
+        tenants: &'static tokio::sync::RwLock<TenantsMap>,
         tenant: &Arc<Tenant>,
     ) -> Result<(), DeleteTenantError> {
         // Tree sort timelines, schedule delete for them. Mention retries from the console side.
@@ -537,18 +604,10 @@ impl DeleteTenantFlow {
             .await
             .context("cleanup_remaining_fs_traces")?;
 
-        {
-            let mut locked = tenants.write().unwrap();
-            if locked.remove(&tenant.tenant_id).is_none() {
-                warn!("Tenant got removed from tenants map during deletion");
-            };
-
-            // FIXME: we should not be modifying this from outside of mgr.rs.
-            // This will go away when we simplify deletion (https://github.com/neondatabase/neon/issues/5080)
-            crate::metrics::TENANT_MANAGER
-                .tenant_slots
-                .set(locked.len() as u64);
-        }
+        let mut locked = tenants.write().await;
+        if locked.remove(&tenant.tenant_id).is_none() {
+            warn!("Tenant got removed from tenants map during deletion");
+        };
 
         *guard = Self::Finished;
 

pageserver/src/tenant/disk_btree.rs

@@ -573,10 +573,10 @@ impl<const L: usize> BuildNode<L> {
         BuildNode {
             num_children: 0,
             level,
-            prefix: Vec::with_capacity(16),
+            prefix: Vec::new(),
             suffix_len: 0,
-            keys: Vec::with_capacity(5024),
-            values: Vec::with_capacity(3140),
+            keys: Vec::new(),
+            values: Vec::new(),
             size: NODE_HDR_SIZE,
         }
     }

pageserver/src/tenant/layer_map.rs

@@ -639,10 +639,147 @@ impl LayerMap {
         }
 
         println!("historic_layers:");
-        for desc in self.iter_historic_layers() {
-            desc.dump();
+        for layer in self.iter_historic_layers() {
+            layer.dump(verbose, ctx)?;
         }
         println!("End dump LayerMap");
         Ok(())
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::LayerMap;
+    use crate::tenant::storage_layer::LayerFileName;
+    use std::str::FromStr;
+    use std::sync::Arc;
+
+    mod l0_delta_layers_updated {
+
+        use crate::tenant::{
+            storage_layer::{AsLayerDesc, PersistentLayerDesc},
+            timeline::layer_manager::LayerFileManager,
+        };
+
+        use super::*;
+
+        struct LayerObject(PersistentLayerDesc);
+
+        impl AsLayerDesc for LayerObject {
+            fn layer_desc(&self) -> &PersistentLayerDesc {
+                &self.0
+            }
+        }
+
+        impl LayerObject {
+            fn new(desc: PersistentLayerDesc) -> Self {
+                LayerObject(desc)
+            }
+        }
+
+        type TestLayerFileManager = LayerFileManager<LayerObject>;
+
+        #[test]
+        fn for_full_range_delta() {
+            // l0_delta_layers are used by compaction, and should observe all buffered updates
+            l0_delta_layers_updated_scenario(
+                 "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000053423C21-0000000053424D69",
+                 true
+             )
+        }
+
+        #[test]
+        fn for_non_full_range_delta() {
+            // has minimal uncovered areas compared to l0_delta_layers_updated_on_insert_replace_remove_for_full_range_delta
+            l0_delta_layers_updated_scenario(
+                 "000000000000000000000000000000000001-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFE__0000000053423C21-0000000053424D69",
+                 // because not full range
+                 false
+             )
+        }
+
+        #[test]
+        fn for_image() {
+            l0_delta_layers_updated_scenario(
+                 "000000000000000000000000000000000000-000000000000000000000000000000010000__0000000053424D69",
+                 // code only checks if it is a full range layer, doesn't care about images, which must
+                 // mean we should in practice never have full range images
+                 false
+             )
+        }
+
+        #[test]
+        fn replacing_missing_l0_is_notfound() {
+            // original impl had an oversight, and L0 was an anyhow::Error. anyhow::Error should
+            // however only happen for precondition failures.
+
+            let layer = "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000053423C21-0000000053424D69";
+            let layer = LayerFileName::from_str(layer).unwrap();
+            let layer = PersistentLayerDesc::from(layer);
+
+            // same skeletan construction; see scenario below
+            let not_found = Arc::new(LayerObject::new(layer.clone()));
+            let new_version = Arc::new(LayerObject::new(layer));
+
+            // after the immutable storage state refactor, the replace operation
+            // will not use layer map any more. We keep it here for consistency in test cases
+            // and can remove it in the future.
+            let _map = LayerMap::default();
+
+            let mut mapping = TestLayerFileManager::new();
+
+            mapping
+                .replace_and_verify(not_found, new_version)
+                .unwrap_err();
+        }
+
+        fn l0_delta_layers_updated_scenario(layer_name: &str, expected_l0: bool) {
+            let name = LayerFileName::from_str(layer_name).unwrap();
+            let skeleton = PersistentLayerDesc::from(name);
+
+            let remote = Arc::new(LayerObject::new(skeleton.clone()));
+            let downloaded = Arc::new(LayerObject::new(skeleton));
+
+            let mut map = LayerMap::default();
+            let mut mapping = LayerFileManager::new();
+
+            // two disjoint Arcs in different lifecycle phases. even if it seems they must be the
+            // same layer, we use LayerMap::compare_arced_layers as the identity of layers.
+            assert_eq!(remote.layer_desc(), downloaded.layer_desc());
+
+            let expected_in_counts = (1, usize::from(expected_l0));
+
+            map.batch_update()
+                .insert_historic(remote.layer_desc().clone());
+            mapping.insert(remote.clone());
+            assert_eq!(
+                count_layer_in(&map, remote.layer_desc()),
+                expected_in_counts
+            );
+
+            mapping
+                .replace_and_verify(remote, downloaded.clone())
+                .expect("name derived attributes are the same");
+            assert_eq!(
+                count_layer_in(&map, downloaded.layer_desc()),
+                expected_in_counts
+            );
+
+            map.batch_update().remove_historic(downloaded.layer_desc());
+            assert_eq!(count_layer_in(&map, downloaded.layer_desc()), (0, 0));
+        }
+
+        fn count_layer_in(map: &LayerMap, layer: &PersistentLayerDesc) -> (usize, usize) {
+            let historic = map
+                .iter_historic_layers()
+                .filter(|x| x.key() == layer.key())
+                .count();
+            let l0s = map
+                .get_level0_deltas()
+                .expect("why does this return a result");
+            let l0 = l0s.iter().filter(|x| x.key() == layer.key()).count();
+
+            (historic, l0)
+        }
+    }
+}

pageserver/src/tenant/metadata.rs

@@ -406,123 +406,4 @@ mod tests {
             METADATA_OLD_FORMAT_VERSION, METADATA_FORMAT_VERSION
         );
     }
-
-    #[test]
-    fn test_metadata_bincode_serde() {
-        let original_metadata = TimelineMetadata::new(
-            Lsn(0x200),
-            Some(Lsn(0x100)),
-            Some(TIMELINE_ID),
-            Lsn(0),
-            Lsn(0),
-            Lsn(0),
-            // Any version will do here, so use the default
-            crate::DEFAULT_PG_VERSION,
-        );
-        let metadata_bytes = original_metadata
-            .to_bytes()
-            .expect("Cannot create bytes array from metadata");
-
-        let metadata_bincode_be_bytes = original_metadata
-            .ser()
-            .expect("Cannot serialize the metadata");
-
-        // 8 bytes for the length of the vector
-        assert_eq!(metadata_bincode_be_bytes.len(), 8 + metadata_bytes.len());
-
-        let expected_bincode_bytes = {
-            let mut temp = vec![];
-            let len_bytes = metadata_bytes.len().to_be_bytes();
-            temp.extend_from_slice(&len_bytes);
-            temp.extend_from_slice(&metadata_bytes);
-            temp
-        };
-        assert_eq!(metadata_bincode_be_bytes, expected_bincode_bytes);
-
-        let deserialized_metadata = TimelineMetadata::des(&metadata_bincode_be_bytes).unwrap();
-        // Deserialized metadata has the metadata header, which is different from the serialized one.
-        //   Reference: TimelineMetaData::to_bytes()
-        let expected_metadata = {
-            let mut temp_metadata = original_metadata;
-            let body_bytes = temp_metadata
-                .body
-                .ser()
-                .expect("Cannot serialize the metadata body");
-            let metadata_size = METADATA_HDR_SIZE + body_bytes.len();
-            let hdr = TimelineMetadataHeader {
-                size: metadata_size as u16,
-                format_version: METADATA_FORMAT_VERSION,
-                checksum: crc32c::crc32c(&body_bytes),
-            };
-            temp_metadata.hdr = hdr;
-            temp_metadata
-        };
-        assert_eq!(deserialized_metadata, expected_metadata);
-    }
-
-    #[test]
-    fn test_metadata_bincode_serde_ensure_roundtrip() {
-        let original_metadata = TimelineMetadata::new(
-            Lsn(0x200),
-            Some(Lsn(0x100)),
-            Some(TIMELINE_ID),
-            Lsn(0),
-            Lsn(0),
-            Lsn(0),
-            // Any version will do here, so use the default
-            crate::DEFAULT_PG_VERSION,
-        );
-        let expected_bytes = vec![
-            /* bincode length encoding bytes */
-            0, 0, 0, 0, 0, 0, 2, 0, // 8 bytes for the length of the serialized vector
-            /* TimelineMetadataHeader */
-            4, 37, 101, 34, 0, 70, 0, 4, // checksum, size, format_version (4 + 2 + 2)
-            /* TimelineMetadataBodyV2 */
-            0, 0, 0, 0, 0, 0, 2, 0, // disk_consistent_lsn (8 bytes)
-            1, 0, 0, 0, 0, 0, 0, 1, 0, // prev_record_lsn (9 bytes)
-            1, 17, 34, 51, 68, 85, 102, 119, 136, 17, 34, 51, 68, 85, 102, 119,
-            136, // ancestor_timeline (17 bytes)
-            0, 0, 0, 0, 0, 0, 0, 0, // ancestor_lsn (8 bytes)
-            0, 0, 0, 0, 0, 0, 0, 0, // latest_gc_cutoff_lsn (8 bytes)
-            0, 0, 0, 0, 0, 0, 0, 0, // initdb_lsn (8 bytes)
-            0, 0, 0, 15, // pg_version (4 bytes)
-            /* padding bytes */
-            0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-            0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-            0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-            0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-            0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-            0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-            0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-            0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-            0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-            0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-            0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-            0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-            0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-            0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-            0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-            0, 0, 0, 0, 0, 0, 0,
-        ];
-        let metadata_ser_bytes = original_metadata.ser().unwrap();
-        assert_eq!(metadata_ser_bytes, expected_bytes);
-
-        let expected_metadata = {
-            let mut temp_metadata = original_metadata;
-            let body_bytes = temp_metadata
-                .body
-                .ser()
-                .expect("Cannot serialize the metadata body");
-            let metadata_size = METADATA_HDR_SIZE + body_bytes.len();
-            let hdr = TimelineMetadataHeader {
-                size: metadata_size as u16,
-                format_version: METADATA_FORMAT_VERSION,
-                checksum: crc32c::crc32c(&body_bytes),
-            };
-            temp_metadata.hdr = hdr;
-            temp_metadata
-        };
-        let des_metadata = TimelineMetadata::des(&metadata_ser_bytes).unwrap();
-        assert_eq!(des_metadata, expected_metadata);
-    }
 }

pageserver/src/tenant/par_fsync.rs

@@ -57,7 +57,8 @@ pub fn par_fsync(paths: &[Utf8PathBuf]) -> io::Result<()> {
     fsync_in_thread_pool(paths)
 }
 
-/// Parallel fsync asynchronously.
+/// Parallel fsync asynchronously. If number of files are less than PARALLEL_PATH_THRESHOLD, fsync is done in the current
+/// execution thread. Otherwise, we will spawn_blocking and run it in tokio.
 pub async fn par_fsync_async(paths: &[Utf8PathBuf]) -> io::Result<()> {
     const MAX_CONCURRENT_FSYNC: usize = 64;
     let mut next = paths.iter().peekable();

pageserver/src/tenant/remote_timeline_client.rs

@@ -167,15 +167,39 @@
 //!   - download their remote [`IndexPart`]s
 //!   - create `Timeline` struct and a `RemoteTimelineClient`
 //!   - initialize the client's upload queue with its `IndexPart`
+//!   - create [`RemoteLayer`](super::storage_layer::RemoteLayer) instances
+//!     for layers that are referenced by `IndexPart` but not present locally
 //!   - schedule uploads for layers that are only present locally.
+//!   - if the remote `IndexPart`'s metadata was newer than the metadata in
+//!     the local filesystem, write the remote metadata to the local filesystem
 //! - After the above is done for each timeline, open the tenant for business by
 //!   transitioning it from `TenantState::Attaching` to `TenantState::Active` state.
 //!   This starts the timelines' WAL-receivers and the tenant's GC & Compaction loops.
 //!
+//! We keep track of the fact that a client is in `Attaching` state in a marker
+//! file on the local disk. This is critical because, when we restart the pageserver,
+//! we do not want to do the `List timelines` step for each tenant that has already
+//! been successfully attached (for performance & cost reasons).
+//! Instead, for a tenant without the attach marker file, we assume that the
+//! local state is in sync or ahead of the remote state. This includes the list
+//! of all of the tenant's timelines, which is particularly critical to be up-to-date:
+//! if there's a timeline on the remote that the pageserver doesn't know about,
+//! the GC will not consider its branch point, leading to data loss.
+//! So, for a tenant with the attach marker file, we know that we do not yet have
+//! persisted all the remote timeline's metadata files locally. To exclude the
+//! risk above, we re-run the procedure for such tenants
+//!
 //! # Operating Without Remote Storage
 //!
 //! If no remote storage configuration is provided, the [`RemoteTimelineClient`] is
 //! not created and the uploads are skipped.
+//! Theoretically, it should be ok to remove and re-add remote storage configuration to
+//! the pageserver config at any time, since it doesn't make a difference to
+//! [`Timeline::load_layer_map`].
+//! Of course, the remote timeline dir must not change while we have de-configured
+//! remote storage, i.e., the pageserver must remain the owner of the given prefix
+//! in remote storage.
+//! But note that we don't test any of this right now.
 //!
 //! [`Tenant::timeline_init_and_sync`]: super::Tenant::timeline_init_and_sync
 //! [`Timeline::load_layer_map`]: super::Timeline::load_layer_map
@@ -187,7 +211,8 @@ mod upload;
 use anyhow::Context;
 use camino::Utf8Path;
 use chrono::{NaiveDateTime, Utc};
-
+// re-export these
+pub use download::{is_temp_download_file, list_remote_timelines};
 use scopeguard::ScopeGuard;
 use tokio_util::sync::CancellationToken;
 use utils::backoff::{
@@ -212,7 +237,7 @@ use crate::metrics::{
 };
 use crate::task_mgr::shutdown_token;
 use crate::tenant::debug_assert_current_span_has_tenant_and_timeline_id;
-use crate::tenant::storage_layer::AsLayerDesc;
+use crate::tenant::remote_timeline_client::index::LayerFileMetadata;
 use crate::tenant::upload_queue::Delete;
 use crate::tenant::TIMELINES_SEGMENT_NAME;
 use crate::{
@@ -230,13 +255,10 @@ use utils::id::{TenantId, TimelineId};
 
 use self::index::IndexPart;
 
-use super::storage_layer::{Layer, LayerFileName, ResidentLayer};
+use super::storage_layer::LayerFileName;
 use super::upload_queue::SetDeletedFlagProgress;
 use super::Generation;
 
-pub(crate) use download::{is_temp_download_file, list_remote_timelines};
-pub(crate) use index::LayerFileMetadata;
-
 // Occasional network issues and such can cause remote operations to fail, and
 // that's expected. If a download fails, we log it at info-level, and retry.
 // But after FAILED_DOWNLOAD_WARN_THRESHOLD retries, we start to log it at WARN
@@ -446,10 +468,7 @@ impl RemoteTimelineClient {
     //
 
     /// Download index file
-    pub async fn download_index_file(
-        &self,
-        cancel: CancellationToken,
-    ) -> Result<MaybeDeletedIndexPart, DownloadError> {
+    pub async fn download_index_file(&self) -> Result<MaybeDeletedIndexPart, DownloadError> {
         let _unfinished_gauge_guard = self.metrics.call_begin(
             &RemoteOpFileKind::Index,
             &RemoteOpKind::Download,
@@ -463,7 +482,6 @@ impl RemoteTimelineClient {
             &self.tenant_id,
             &self.timeline_id,
             self.generation,
-            cancel,
         )
         .measure_remote_op(
             self.tenant_id,
@@ -611,12 +629,13 @@ impl RemoteTimelineClient {
     ///
     pub(crate) fn schedule_layer_file_upload(
         self: &Arc<Self>,
-        layer: ResidentLayer,
+        layer_file_name: &LayerFileName,
+        layer_metadata: &LayerFileMetadata,
     ) -> anyhow::Result<()> {
         let mut guard = self.upload_queue.lock().unwrap();
         let upload_queue = guard.initialized_mut()?;
 
-        self.schedule_layer_file_upload0(upload_queue, layer);
+        self.schedule_layer_file_upload0(upload_queue, layer_file_name, layer_metadata);
         self.launch_queued_tasks(upload_queue);
         Ok(())
     }
@@ -624,19 +643,18 @@ impl RemoteTimelineClient {
     fn schedule_layer_file_upload0(
         self: &Arc<Self>,
         upload_queue: &mut UploadQueueInitialized,
-        layer: ResidentLayer,
+        layer_file_name: &LayerFileName,
+        layer_metadata: &LayerFileMetadata,
     ) {
-        let metadata = layer.metadata();
-
         upload_queue
             .latest_files
-            .insert(layer.layer_desc().filename(), metadata.clone());
+            .insert(layer_file_name.clone(), layer_metadata.clone());
         upload_queue.latest_files_changes_since_metadata_upload_scheduled += 1;
 
-        info!("scheduled layer file upload {layer}");
-        let op = UploadOp::UploadLayer(layer, metadata);
+        let op = UploadOp::UploadLayer(layer_file_name.clone(), layer_metadata.clone());
         self.calls_unfinished_metric_begin(&op);
         upload_queue.queued_operations.push_back(op);
+        info!("scheduled layer file upload {layer_file_name}");
     }
 
     /// Launch a delete operation in the background.
@@ -649,13 +667,13 @@ impl RemoteTimelineClient {
     /// successfully.
     pub fn schedule_layer_file_deletion(
         self: &Arc<Self>,
-        names: &[LayerFileName],
+        names: Vec<LayerFileName>,
     ) -> anyhow::Result<()> {
         let mut guard = self.upload_queue.lock().unwrap();
         let upload_queue = guard.initialized_mut()?;
 
         let with_generations =
-            self.schedule_unlinking_of_layers_from_index_part0(upload_queue, names.iter().cloned());
+            self.schedule_unlinking_of_layers_from_index_part0(upload_queue, &names);
 
         self.schedule_deletion_of_unlinked0(upload_queue, with_generations);
 
@@ -669,17 +687,17 @@ impl RemoteTimelineClient {
     ///
     /// The files will be leaked in remote storage unless [`Self::schedule_deletion_of_unlinked`]
     /// is invoked on them.
-    pub(crate) fn schedule_gc_update(self: &Arc<Self>, gc_layers: &[Layer]) -> anyhow::Result<()> {
+    #[allow(unused)] // will be used by PR#4938
+    pub(crate) fn schedule_unlinking_of_layers_from_index_part(
+        self: &Arc<Self>,
+        names: Vec<LayerFileName>,
+    ) -> anyhow::Result<()> {
         let mut guard = self.upload_queue.lock().unwrap();
         let upload_queue = guard.initialized_mut()?;
 
         // just forget the return value; after uploading the next index_part.json, we can consider
-        // the layer files as "dangling". this is fine, at worst case we create work for the
-        // scrubber.
-
-        let names = gc_layers.iter().map(|x| x.layer_desc().filename());
-
-        self.schedule_unlinking_of_layers_from_index_part0(upload_queue, names);
+        // the layer files as "dangling". this is fine however.
+        self.schedule_unlinking_of_layers_from_index_part0(upload_queue, &names);
 
         self.launch_queued_tasks(upload_queue);
 
@@ -688,28 +706,26 @@ impl RemoteTimelineClient {
 
     /// Update the remote index file, removing the to-be-deleted files from the index,
     /// allowing scheduling of actual deletions later.
-    fn schedule_unlinking_of_layers_from_index_part0<I>(
+    fn schedule_unlinking_of_layers_from_index_part0(
         self: &Arc<Self>,
         upload_queue: &mut UploadQueueInitialized,
-        names: I,
-    ) -> Vec<(LayerFileName, Generation)>
-    where
-        I: IntoIterator<Item = LayerFileName>,
-    {
+        names: &[LayerFileName],
+    ) -> Vec<(LayerFileName, Generation)> {
         // Deleting layers doesn't affect the values stored in TimelineMetadata,
         // so we don't need update it. Just serialize it.
         let metadata = upload_queue.latest_metadata.clone();
 
         // Decorate our list of names with each name's generation, dropping
-        // names that are unexpectedly missing from our metadata.
+        // makes that are unexpectedly missing from our metadata.
         let with_generations: Vec<_> = names
-            .into_iter()
+            .iter()
             .filter_map(|name| {
-                let meta = upload_queue.latest_files.remove(&name);
+                // Remove from latest_files, learning the file's remote generation in the process
+                let meta = upload_queue.latest_files.remove(name);
 
                 if let Some(meta) = meta {
                     upload_queue.latest_files_changes_since_metadata_upload_scheduled += 1;
-                    Some((name, meta.generation))
+                    Some((name.to_owned(), meta.generation))
                 } else {
                     // This can only happen if we forgot to to schedule the file upload
                     // before scheduling the delete. Log it because it is a rare/strange
@@ -721,17 +737,6 @@ impl RemoteTimelineClient {
             })
             .collect();
 
-        #[cfg(feature = "testing")]
-        for (name, gen) in &with_generations {
-            if let Some(unexpected) = upload_queue.dangling_files.insert(name.to_owned(), *gen) {
-                if &unexpected == gen {
-                    tracing::error!("{name} was unlinked twice with same generation");
-                } else {
-                    tracing::error!("{name} was unlinked twice with different generations {gen:?} and {unexpected:?}");
-                }
-            }
-        }
-
         // after unlinking files from the upload_queue.latest_files we must always schedule an
         // index_part update, because that needs to be uploaded before we can actually delete the
         // files.
@@ -743,7 +748,8 @@ impl RemoteTimelineClient {
     }
 
     /// Schedules deletion for layer files which have previously been unlinked from the
-    /// `index_part.json` with [`Self::schedule_gc_update`] or [`Self::schedule_compaction_update`].
+    /// `index_part.json` with [`Self::schedule_unlinking_of_layers_from_index_part`].
+    #[allow(unused)] // will be used by Layer::drop in PR#4938
     pub(crate) fn schedule_deletion_of_unlinked(
         self: &Arc<Self>,
         layers: Vec<(LayerFileName, Generation)>,
@@ -765,19 +771,6 @@ impl RemoteTimelineClient {
             info!("scheduling deletion of layer {}{}", name, gen.get_suffix());
         }
 
-        #[cfg(feature = "testing")]
-        for (name, gen) in &with_generations {
-            match upload_queue.dangling_files.remove(name) {
-                Some(same) if &same == gen => { /* expected */ }
-                Some(other) => {
-                    tracing::error!("{name} was unlinked with {other:?} but deleted with {gen:?}");
-                }
-                None => {
-                    tracing::error!("{name} was unlinked but was not dangling");
-                }
-            }
-        }
-
         // schedule the actual deletions
         let op = UploadOp::Delete(Delete {
             layers: with_generations,
@@ -791,19 +784,19 @@ impl RemoteTimelineClient {
     /// `compacted_from` represent the L0 names which have been `compacted_to` L1 layers.
     pub(crate) fn schedule_compaction_update(
         self: &Arc<Self>,
-        compacted_from: &[Layer],
-        compacted_to: &[ResidentLayer],
+        compacted_from: &[LayerFileName],
+        compacted_to: &[(LayerFileName, LayerFileMetadata)],
     ) -> anyhow::Result<()> {
         let mut guard = self.upload_queue.lock().unwrap();
         let upload_queue = guard.initialized_mut()?;
 
-        for layer in compacted_to {
-            self.schedule_layer_file_upload0(upload_queue, layer.clone());
+        for (name, m) in compacted_to {
+            self.schedule_layer_file_upload0(upload_queue, name, m);
         }
 
-        let names = compacted_from.iter().map(|x| x.layer_desc().filename());
-
-        self.schedule_unlinking_of_layers_from_index_part0(upload_queue, names);
+        let with_generations =
+            self.schedule_unlinking_of_layers_from_index_part0(upload_queue, compacted_from);
+        self.schedule_deletion_of_unlinked0(upload_queue, with_generations);
         self.launch_queued_tasks(upload_queue);
 
         Ok(())
@@ -1177,12 +1170,16 @@ impl RemoteTimelineClient {
             }
 
             let upload_result: anyhow::Result<()> = match &task.op {
-                UploadOp::UploadLayer(ref layer, ref layer_metadata) => {
-                    let path = layer.local_path();
+                UploadOp::UploadLayer(ref layer_file_name, ref layer_metadata) => {
+                    let path = self
+                        .conf
+                        .timeline_path(&self.tenant_id, &self.timeline_id)
+                        .join(layer_file_name.file_name());
+
                     upload::upload_timeline_layer(
                         self.conf,
                         &self.storage_impl,
-                        path,
+                        &path,
                         layer_metadata,
                         self.generation,
                     )
@@ -1456,8 +1453,6 @@ impl RemoteTimelineClient {
                         num_inprogress_deletions: 0,
                         inprogress_tasks: HashMap::default(),
                         queued_operations: VecDeque::default(),
-                        #[cfg(feature = "testing")]
-                        dangling_files: HashMap::default(),
                     };
 
                     let upload_queue = std::mem::replace(
@@ -1501,6 +1496,13 @@ impl RemoteTimelineClient {
             }
         }
     }
+
+    pub(crate) fn get_layer_metadata(
+        &self,
+        name: &LayerFileName,
+    ) -> anyhow::Result<Option<LayerFileMetadata>> {
+        self.upload_queue.lock().unwrap().get_layer_metadata(name)
+    }
 }
 
 pub fn remote_timelines_path(tenant_id: &TenantId) -> RemotePath {
@@ -1542,7 +1544,7 @@ pub fn remote_index_path(
 }
 
 /// Given the key of an index, parse out the generation part of the name
-pub fn parse_remote_index_path(path: RemotePath) -> Option<Generation> {
+pub(crate) fn parse_remote_index_path(path: RemotePath) -> Option<Generation> {
     let file_name = match path.get_path().file_name() {
         Some(f) => f,
         None => {
@@ -1588,7 +1590,6 @@ mod tests {
         context::RequestContext,
         tenant::{
             harness::{TenantHarness, TIMELINE_ID},
-            storage_layer::Layer,
             Generation, Tenant, Timeline,
         },
         DEFAULT_PG_VERSION,
@@ -1731,11 +1732,7 @@ mod tests {
         let client = timeline.remote_client.as_ref().unwrap();
 
         // Download back the index.json, and check that the list of files is correct
-        let initial_index_part = match client
-            .download_index_file(CancellationToken::new())
-            .await
-            .unwrap()
-        {
+        let initial_index_part = match client.download_index_file().await.unwrap() {
             MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
             MaybeDeletedIndexPart::Deleted(_) => panic!("unexpectedly got deleted index part"),
         };
@@ -1761,29 +1758,32 @@ mod tests {
         let generation = harness.generation;
 
         // Create a couple of dummy files,  schedule upload for them
+        let layer_file_name_1: LayerFileName = "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap();
+        let layer_file_name_2: LayerFileName = "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D9-00000000016B5A52".parse().unwrap();
+        let layer_file_name_3: LayerFileName = "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59DA-00000000016B5A53".parse().unwrap();
+        let content_1 = dummy_contents("foo");
+        let content_2 = dummy_contents("bar");
+        let content_3 = dummy_contents("baz");
 
-        let layers = [
-            ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), dummy_contents("foo")),
-            ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D9-00000000016B5A52".parse().unwrap(), dummy_contents("bar")),
-            ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59DA-00000000016B5A53".parse().unwrap(), dummy_contents("baz"))
-        ]
-        .into_iter()
-        .map(|(name, contents): (LayerFileName, Vec<u8>)| {
-            std::fs::write(timeline_path.join(name.file_name()), &contents).unwrap();
-
-            Layer::for_resident(
-                harness.conf,
-                &timeline,
-                name,
-                LayerFileMetadata::new(contents.len() as u64, generation),
-            )
-        }).collect::<Vec<_>>();
+        for (filename, content) in [
+            (&layer_file_name_1, &content_1),
+            (&layer_file_name_2, &content_2),
+            (&layer_file_name_3, &content_3),
+        ] {
+            std::fs::write(timeline_path.join(filename.file_name()), content).unwrap();
+        }
 
         client
-            .schedule_layer_file_upload(layers[0].clone())
+            .schedule_layer_file_upload(
+                &layer_file_name_1,
+                &LayerFileMetadata::new(content_1.len() as u64, generation),
+            )
             .unwrap();
         client
-            .schedule_layer_file_upload(layers[1].clone())
+            .schedule_layer_file_upload(
+                &layer_file_name_2,
+                &LayerFileMetadata::new(content_2.len() as u64, generation),
+            )
             .unwrap();
 
         // Check that they are started immediately, not queued
@@ -1824,11 +1824,7 @@ mod tests {
         }
 
         // Download back the index.json, and check that the list of files is correct
-        let index_part = match client
-            .download_index_file(CancellationToken::new())
-            .await
-            .unwrap()
-        {
+        let index_part = match client.download_index_file().await.unwrap() {
             MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
             MaybeDeletedIndexPart::Deleted(_) => panic!("unexpectedly got deleted index part"),
         };
@@ -1841,42 +1837,38 @@ mod tests {
                 .collect(),
             &[
                 &initial_layer.file_name(),
-                &layers[0].layer_desc().filename().file_name(),
-                &layers[1].layer_desc().filename().file_name(),
+                &layer_file_name_1.file_name(),
+                &layer_file_name_2.file_name(),
             ],
         );
         assert_eq!(index_part.metadata, metadata);
 
         // Schedule upload and then a deletion. Check that the deletion is queued
         client
-            .schedule_layer_file_upload(layers[2].clone())
+            .schedule_layer_file_upload(
+                &layer_file_name_3,
+                &LayerFileMetadata::new(content_3.len() as u64, generation),
+            )
             .unwrap();
-
-        // this is no longer consistent with how deletion works with Layer::drop, but in this test
-        // keep using schedule_layer_file_deletion because we don't have a way to wait for the
-        // spawn_blocking started by the drop.
         client
-            .schedule_layer_file_deletion(&[layers[0].layer_desc().filename()])
+            .schedule_layer_file_deletion([layer_file_name_1.clone()].to_vec())
             .unwrap();
         {
             let mut guard = client.upload_queue.lock().unwrap();
             let upload_queue = guard.initialized_mut().unwrap();
 
             // Deletion schedules upload of the index file, and the file deletion itself
-            assert_eq!(upload_queue.queued_operations.len(), 2);
-            assert_eq!(upload_queue.inprogress_tasks.len(), 1);
-            assert_eq!(upload_queue.num_inprogress_layer_uploads, 1);
-            assert_eq!(upload_queue.num_inprogress_deletions, 0);
-            assert_eq!(
-                upload_queue.latest_files_changes_since_metadata_upload_scheduled,
-                0
-            );
+            assert!(upload_queue.queued_operations.len() == 2);
+            assert!(upload_queue.inprogress_tasks.len() == 1);
+            assert!(upload_queue.num_inprogress_layer_uploads == 1);
+            assert!(upload_queue.num_inprogress_deletions == 0);
+            assert!(upload_queue.latest_files_changes_since_metadata_upload_scheduled == 0);
         }
         assert_remote_files(
             &[
                 &initial_layer.file_name(),
-                &layers[0].layer_desc().filename().file_name(),
-                &layers[1].layer_desc().filename().file_name(),
+                &layer_file_name_1.file_name(),
+                &layer_file_name_2.file_name(),
                 "index_part.json",
             ],
             &remote_timeline_dir,
@@ -1890,8 +1882,8 @@ mod tests {
         assert_remote_files(
             &[
                 &initial_layer.file_name(),
-                &layers[1].layer_desc().filename().file_name(),
-                &layers[2].layer_desc().filename().file_name(),
+                &layer_file_name_2.file_name(),
+                &layer_file_name_3.file_name(),
                 "index_part.json",
             ],
             &remote_timeline_dir,
@@ -1920,13 +1912,6 @@ mod tests {
         )
         .unwrap();
 
-        let layer_file_1 = Layer::for_resident(
-            harness.conf,
-            &timeline,
-            layer_file_name_1.clone(),
-            LayerFileMetadata::new(content_1.len() as u64, harness.generation),
-        );
-
         #[derive(Debug, PartialEq, Clone, Copy)]
         struct BytesStartedFinished {
             started: Option<usize>,
@@ -1962,7 +1947,10 @@ mod tests {
         let actual_a = get_bytes_started_stopped();
 
         client
-            .schedule_layer_file_upload(layer_file_1.clone())
+            .schedule_layer_file_upload(
+                &layer_file_name_1,
+                &LayerFileMetadata::new(content_1.len() as u64, harness.generation),
+            )
             .unwrap();
 
         let actual_b = get_bytes_started_stopped();
@@ -2027,7 +2015,7 @@ mod tests {
         let client = test_state.build_client(get_generation);
 
         let download_r = client
-            .download_index_file(CancellationToken::new())
+            .download_index_file()
             .await
             .expect("download should always succeed");
         assert!(matches!(download_r, MaybeDeletedIndexPart::IndexPart(_)));

pageserver/src/tenant/remote_timeline_client/download.rs

@@ -18,8 +18,8 @@ use crate::config::PageServerConf;
 use crate::tenant::remote_timeline_client::{remote_layer_path, remote_timelines_path};
 use crate::tenant::storage_layer::LayerFileName;
 use crate::tenant::timeline::span::debug_assert_current_span_has_tenant_and_timeline_id;
-use crate::tenant::Generation;
-use remote_storage::{DownloadError, GenericRemoteStorage, ListingMode};
+use crate::tenant::{Generation, TENANT_DELETED_MARKER_FILE_NAME};
+use remote_storage::{DownloadError, GenericRemoteStorage};
 use utils::crashsafe::path_with_suffix_extension;
 use utils::id::{TenantId, TimelineId};
 
@@ -170,43 +170,53 @@ pub fn is_temp_download_file(path: &Utf8Path) -> bool {
 pub async fn list_remote_timelines(
     storage: &GenericRemoteStorage,
     tenant_id: TenantId,
-    cancel: CancellationToken,
-) -> anyhow::Result<(HashSet<TimelineId>, HashSet<String>)> {
+) -> anyhow::Result<HashSet<TimelineId>> {
     let remote_path = remote_timelines_path(&tenant_id);
 
     fail::fail_point!("storage-sync-list-remote-timelines", |_| {
         anyhow::bail!("storage-sync-list-remote-timelines");
     });
 
-    let listing = download_retry_forever(
-        || storage.list(Some(&remote_path), ListingMode::WithDelimiter),
-        &format!("list timelines for {tenant_id}"),
-        cancel,
+    let timelines = download_retry(
+        || storage.list_prefixes(Some(&remote_path)),
+        &format!("list prefixes for {tenant_id}"),
     )
     .await?;
 
-    let mut timeline_ids = HashSet::new();
-    let mut other_prefixes = HashSet::new();
+    if timelines.is_empty() {
+        anyhow::bail!("no timelines found on the remote storage")
+    }
+
+    let mut timeline_ids = HashSet::new();
+
+    for timeline_remote_storage_key in timelines {
+        if timeline_remote_storage_key.object_name() == Some(TENANT_DELETED_MARKER_FILE_NAME) {
+            // A `deleted` key within `timelines/` is a marker file, not a timeline.  Ignore it.
+            // This code will be removed in https://github.com/neondatabase/neon/pull/5580
+            continue;
+        }
 
-    for timeline_remote_storage_key in listing.prefixes {
         let object_name = timeline_remote_storage_key.object_name().ok_or_else(|| {
             anyhow::anyhow!("failed to get timeline id for remote tenant {tenant_id}")
         })?;
 
-        match object_name.parse::<TimelineId>() {
-            Ok(t) => timeline_ids.insert(t),
-            Err(_) => other_prefixes.insert(object_name.to_string()),
-        };
+        let timeline_id: TimelineId = object_name
+            .parse()
+            .with_context(|| format!("parse object name into timeline id '{object_name}'"))?;
+
+        // list_prefixes is assumed to return unique names. Ensure this here.
+        // NB: it's safer to bail out than warn-log this because the pageserver
+        //     needs to absolutely know about _all_ timelines that exist, so that
+        //     GC knows all the branchpoints. If we skipped over a timeline instead,
+        //     GC could delete a layer that's still needed by that timeline.
+        anyhow::ensure!(
+            !timeline_ids.contains(&timeline_id),
+            "list_prefixes contains duplicate timeline id {timeline_id}"
+        );
+        timeline_ids.insert(timeline_id);
     }
 
-    for key in listing.keys {
-        let object_name = key
-            .object_name()
-            .ok_or_else(|| anyhow::anyhow!("object name for key {key}"))?;
-        other_prefixes.insert(object_name.to_string());
-    }
-
-    Ok((timeline_ids, other_prefixes))
+    Ok(timeline_ids)
 }
 
 async fn do_download_index_part(
@@ -214,11 +224,10 @@ async fn do_download_index_part(
     tenant_id: &TenantId,
     timeline_id: &TimelineId,
     index_generation: Generation,
-    cancel: CancellationToken,
 ) -> Result<IndexPart, DownloadError> {
     let remote_path = remote_index_path(tenant_id, timeline_id, index_generation);
 
-    let index_part_bytes = download_retry_forever(
+    let index_part_bytes = download_retry(
         || async {
             let mut index_part_download = storage.download(&remote_path).await?;
 
@@ -233,7 +242,6 @@ async fn do_download_index_part(
             Ok(index_part_bytes)
         },
         &format!("download {remote_path:?}"),
-        cancel,
     )
     .await?;
 
@@ -255,28 +263,19 @@ pub(super) async fn download_index_part(
     tenant_id: &TenantId,
     timeline_id: &TimelineId,
     my_generation: Generation,
-    cancel: CancellationToken,
 ) -> Result<IndexPart, DownloadError> {
     debug_assert_current_span_has_tenant_and_timeline_id();
 
     if my_generation.is_none() {
         // Operating without generations: just fetch the generation-less path
-        return do_download_index_part(storage, tenant_id, timeline_id, my_generation, cancel)
-            .await;
+        return do_download_index_part(storage, tenant_id, timeline_id, my_generation).await;
     }
 
     // Stale case: If we were intentionally attached in a stale generation, there may already be a remote
     // index in our generation.
     //
     // This is an optimization to avoid doing the listing for the general case below.
-    let res = do_download_index_part(
-        storage,
-        tenant_id,
-        timeline_id,
-        my_generation,
-        cancel.clone(),
-    )
-    .await;
+    let res = do_download_index_part(storage, tenant_id, timeline_id, my_generation).await;
     match res {
         Ok(index_part) => {
             tracing::debug!(
@@ -296,14 +295,8 @@ pub(super) async fn download_index_part(
     //    we want to find the most recent index from a previous generation.
     //
     // This is an optimization to avoid doing the listing for the general case below.
-    let res = do_download_index_part(
-        storage,
-        tenant_id,
-        timeline_id,
-        my_generation.previous(),
-        cancel.clone(),
-    )
-    .await;
+    let res =
+        do_download_index_part(storage, tenant_id, timeline_id, my_generation.previous()).await;
     match res {
         Ok(index_part) => {
             tracing::debug!("Found index_part from previous generation");
@@ -347,14 +340,13 @@ pub(super) async fn download_index_part(
     match max_previous_generation {
         Some(g) => {
             tracing::debug!("Found index_part in generation {g:?}");
-            do_download_index_part(storage, tenant_id, timeline_id, g, cancel).await
+            do_download_index_part(storage, tenant_id, timeline_id, g).await
         }
         None => {
             // Migration from legacy pre-generation state: we have a generation but no prior
             // attached pageservers did.  Try to load from a no-generation path.
             tracing::info!("No index_part.json* found");
-            do_download_index_part(storage, tenant_id, timeline_id, Generation::none(), cancel)
-                .await
+            do_download_index_part(storage, tenant_id, timeline_id, Generation::none()).await
         }
     }
 }
@@ -384,23 +376,3 @@ where
     )
     .await
 }
-
-async fn download_retry_forever<T, O, F>(
-    op: O,
-    description: &str,
-    cancel: CancellationToken,
-) -> Result<T, DownloadError>
-where
-    O: FnMut() -> F,
-    F: Future<Output = Result<T, DownloadError>>,
-{
-    backoff::retry(
-        op,
-        |e| matches!(e, DownloadError::BadInput(_) | DownloadError::NotFound),
-        FAILED_DOWNLOAD_WARN_THRESHOLD,
-        u32::MAX,
-        description,
-        backoff::Cancel::new(cancel, || DownloadError::Cancelled),
-    )
-    .await
-}

pageserver/src/tenant/remote_timeline_client/index.rs

@@ -6,6 +6,7 @@ use std::collections::HashMap;
 
 use chrono::NaiveDateTime;
 use serde::{Deserialize, Serialize};
+use serde_with::{serde_as, DisplayFromStr};
 use utils::bin_ser::SerializeError;
 
 use crate::tenant::metadata::TimelineMetadata;
@@ -57,6 +58,7 @@ impl LayerFileMetadata {
 ///
 /// This type needs to be backwards and forwards compatible. When changing the fields,
 /// remember to add a test case for the changed version.
+#[serde_as]
 #[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize)]
 pub struct IndexPart {
     /// Debugging aid describing the version of this type.
@@ -76,6 +78,7 @@ pub struct IndexPart {
     // 'disk_consistent_lsn' is a copy of the 'disk_consistent_lsn' in the metadata.
     // It's duplicated for convenience when reading the serialized structure, but is
     // private because internally we would read from metadata instead.
+    #[serde_as(as = "DisplayFromStr")]
     disk_consistent_lsn: Lsn,
 
     #[serde(rename = "metadata_bytes")]
@@ -95,7 +98,7 @@ impl IndexPart {
     const LATEST_VERSION: usize = 4;
 
     // Versions we may see when reading from a bucket.
-    pub const KNOWN_VERSIONS: &'static [usize] = &[1, 2, 3, 4];
+    pub const KNOWN_VERSIONS: &[usize] = &[1, 2, 3, 4];
 
     pub const FILE_NAME: &'static str = "index_part.json";
 
@@ -152,7 +155,7 @@ pub struct IndexLayerMetadata {
 
     #[serde(default = "Generation::none")]
     #[serde(skip_serializing_if = "Generation::is_none")]
-    pub generation: Generation,
+    pub(super) generation: Generation,
 }
 
 impl From<LayerFileMetadata> for IndexLayerMetadata {

pageserver/src/tenant/remote_timeline_client/upload.rs

@@ -72,8 +72,6 @@ pub(super) async fn upload_timeline_layer<'a>(
             // upload. However, a nonexistent file can also be indicative of
             // something worse, like when a file is scheduled for upload before
             // it has been written to disk yet.
-            //
-            // This is tested against `test_compaction_delete_before_upload`
             info!(path = %source_path, "File to upload doesn't exist. Likely the file has been deleted and an upload is not required any more.");
             return Ok(());
         }

pageserver/src/tenant/size.rs

@@ -29,6 +29,7 @@ use tenant_size_model::{Segment, StorageModel};
 /// needs. We will convert this into a StorageModel when it's time to perform
 /// the calculation.
 ///
+#[serde_with::serde_as]
 #[derive(Debug, serde::Serialize, serde::Deserialize)]
 pub struct ModelInputs {
     pub segments: Vec<SegmentMeta>,
@@ -36,9 +37,11 @@ pub struct ModelInputs {
 }
 
 /// A [`Segment`], with some extra information for display purposes
+#[serde_with::serde_as]
 #[derive(Debug, serde::Serialize, serde::Deserialize)]
 pub struct SegmentMeta {
     pub segment: Segment,
+    #[serde_as(as = "serde_with::DisplayFromStr")]
     pub timeline_id: TimelineId,
     pub kind: LsnKind,
 }
@@ -74,22 +77,32 @@ pub enum LsnKind {
 
 /// Collect all relevant LSNs to the inputs. These will only be helpful in the serialized form as
 /// part of [`ModelInputs`] from the HTTP api, explaining the inputs.
+#[serde_with::serde_as]
 #[derive(Debug, serde::Serialize, serde::Deserialize)]
 pub struct TimelineInputs {
+    #[serde_as(as = "serde_with::DisplayFromStr")]
     pub timeline_id: TimelineId,
 
+    #[serde_as(as = "Option<serde_with::DisplayFromStr>")]
     pub ancestor_id: Option<TimelineId>,
 
+    #[serde_as(as = "serde_with::DisplayFromStr")]
     ancestor_lsn: Lsn,
+    #[serde_as(as = "serde_with::DisplayFromStr")]
     last_record: Lsn,
+    #[serde_as(as = "serde_with::DisplayFromStr")]
     latest_gc_cutoff: Lsn,
+    #[serde_as(as = "serde_with::DisplayFromStr")]
     horizon_cutoff: Lsn,
+    #[serde_as(as = "serde_with::DisplayFromStr")]
     pitr_cutoff: Lsn,
 
     /// Cutoff point based on GC settings
+    #[serde_as(as = "serde_with::DisplayFromStr")]
     next_gc_cutoff: Lsn,
 
     /// Cutoff point calculated from the user-supplied 'max_retention_period'
+    #[serde_as(as = "Option<serde_with::DisplayFromStr>")]
     retention_param_cutoff: Option<Lsn>,
 }
 
@@ -406,12 +419,10 @@ async fn fill_logical_sizes(
                 have_any_error = true;
             }
             Ok(Ok(TimelineAtLsnSizeResult(timeline, lsn, Err(error)))) => {
-                if !matches!(error, CalculateLogicalSizeError::Cancelled) {
-                    warn!(
-                        timeline_id=%timeline.timeline_id,
-                        "failed to calculate logical size at {lsn}: {error:#}"
-                    );
-                }
+                warn!(
+                    timeline_id=%timeline.timeline_id,
+                    "failed to calculate logical size at {lsn}: {error:#}"
+                );
                 have_any_error = true;
             }
             Ok(Ok(TimelineAtLsnSizeResult(timeline, lsn, Ok(size)))) => {

pageserver/src/tenant/storage_layer.rs

@@ -4,21 +4,26 @@ pub mod delta_layer;
 mod filename;
 mod image_layer;
 mod inmemory_layer;
-mod layer;
 mod layer_desc;
+mod remote_layer;
 
+use crate::config::PageServerConf;
 use crate::context::{AccessStatsBehavior, RequestContext};
+use crate::repository::Key;
 use crate::task_mgr::TaskKind;
 use crate::walrecord::NeonWalRecord;
+use anyhow::Result;
 use bytes::Bytes;
+use camino::Utf8PathBuf;
 use enum_map::EnumMap;
 use enumset::EnumSet;
 use once_cell::sync::Lazy;
+use pageserver_api::models::LayerAccessKind;
 use pageserver_api::models::{
-    LayerAccessKind, LayerResidenceEvent, LayerResidenceEventReason, LayerResidenceStatus,
+    HistoricLayerInfo, LayerResidenceEvent, LayerResidenceEventReason, LayerResidenceStatus,
 };
 use std::ops::Range;
-use std::sync::Mutex;
+use std::sync::{Arc, Mutex};
 use std::time::{Duration, SystemTime, UNIX_EPOCH};
 use tracing::warn;
 use utils::history_buffer::HistoryBufferWithDropCounter;
@@ -34,8 +39,7 @@ pub use filename::{DeltaFileName, ImageFileName, LayerFileName};
 pub use image_layer::{ImageLayer, ImageLayerWriter};
 pub use inmemory_layer::InMemoryLayer;
 pub use layer_desc::{PersistentLayerDesc, PersistentLayerKey};
-
-pub(crate) use layer::{EvictionError, Layer, ResidentLayer};
+pub use remote_layer::RemoteLayer;
 
 pub fn range_overlaps<T>(a: &Range<T>, b: &Range<T>) -> bool
 where
@@ -70,7 +74,7 @@ pub struct ValueReconstructState {
     pub img: Option<(Lsn, Bytes)>,
 }
 
-/// Return value from [`Layer::get_value_reconstruct_data`]
+/// Return value from Layer::get_page_reconstruct_data
 #[derive(Clone, Copy, Debug)]
 pub enum ValueReconstructResult {
     /// Got all the data needed to reconstruct the requested page
@@ -175,6 +179,26 @@ impl LayerAccessStats {
         new
     }
 
+    /// Creates a clone of `self` and records `new_status` in the clone.
+    ///
+    /// The `new_status` is not recorded in `self`.
+    ///
+    /// See [`record_residence_event`] for why you need to do this while holding the layer map lock.
+    ///
+    /// [`record_residence_event`]: Self::record_residence_event
+    pub(crate) fn clone_for_residence_change(
+        &self,
+        new_status: LayerResidenceStatus,
+    ) -> LayerAccessStats {
+        let clone = {
+            let inner = self.0.lock().unwrap();
+            inner.clone()
+        };
+        let new = LayerAccessStats(Mutex::new(clone));
+        new.record_residence_event(new_status, LayerResidenceEventReason::ResidenceChange);
+        new
+    }
+
     /// Record a change in layer residency.
     ///
     /// Recording the event must happen while holding the layer map lock to
@@ -297,12 +321,95 @@ impl LayerAccessStats {
     }
 }
 
+/// Supertrait of the [`Layer`] trait that captures the bare minimum interface
+/// required by [`LayerMap`](super::layer_map::LayerMap).
+///
+/// All layers should implement a minimal `std::fmt::Debug` without tenant or
+/// timeline names, because those are known in the context of which the layers
+/// are used in (timeline).
+#[async_trait::async_trait]
+pub trait Layer: std::fmt::Debug + std::fmt::Display + Send + Sync + 'static {
+    ///
+    /// Return data needed to reconstruct given page at LSN.
+    ///
+    /// It is up to the caller to collect more data from previous layer and
+    /// perform WAL redo, if necessary.
+    ///
+    /// See PageReconstructResult for possible return values. The collected data
+    /// is appended to reconstruct_data; the caller should pass an empty struct
+    /// on first call, or a struct with a cached older image of the page if one
+    /// is available. If this returns ValueReconstructResult::Continue, look up
+    /// the predecessor layer and call again with the same 'reconstruct_data' to
+    /// collect more data.
+    async fn get_value_reconstruct_data(
+        &self,
+        key: Key,
+        lsn_range: Range<Lsn>,
+        reconstruct_data: &mut ValueReconstructState,
+        ctx: &RequestContext,
+    ) -> Result<ValueReconstructResult>;
+}
+
 /// Get a layer descriptor from a layer.
 pub trait AsLayerDesc {
     /// Get the layer descriptor.
     fn layer_desc(&self) -> &PersistentLayerDesc;
 }
 
+/// A Layer contains all data in a "rectangle" consisting of a range of keys and
+/// range of LSNs.
+///
+/// There are two kinds of layers, in-memory and on-disk layers. In-memory
+/// layers are used to ingest incoming WAL, and provide fast access to the
+/// recent page versions. On-disk layers are stored as files on disk, and are
+/// immutable. This trait presents the common functionality of in-memory and
+/// on-disk layers.
+///
+/// Furthermore, there are two kinds of on-disk layers: delta and image layers.
+/// A delta layer contains all modifications within a range of LSNs and keys.
+/// An image layer is a snapshot of all the data in a key-range, at a single
+/// LSN.
+pub trait PersistentLayer: Layer + AsLayerDesc {
+    /// File name used for this layer, both in the pageserver's local filesystem
+    /// state as well as in the remote storage.
+    fn filename(&self) -> LayerFileName {
+        self.layer_desc().filename()
+    }
+
+    // Path to the layer file in the local filesystem.
+    // `None` for `RemoteLayer`.
+    fn local_path(&self) -> Option<Utf8PathBuf>;
+
+    /// Permanently remove this layer from disk.
+    fn delete_resident_layer_file(&self) -> Result<()>;
+
+    fn downcast_remote_layer(self: Arc<Self>) -> Option<std::sync::Arc<RemoteLayer>> {
+        None
+    }
+
+    fn downcast_delta_layer(self: Arc<Self>) -> Option<std::sync::Arc<DeltaLayer>> {
+        None
+    }
+
+    fn is_remote_layer(&self) -> bool {
+        false
+    }
+
+    fn info(&self, reset: LayerAccessStatsReset) -> HistoricLayerInfo;
+
+    fn access_stats(&self) -> &LayerAccessStats;
+}
+
+pub fn downcast_remote_layer(
+    layer: &Arc<dyn PersistentLayer>,
+) -> Option<std::sync::Arc<RemoteLayer>> {
+    if layer.is_remote_layer() {
+        Arc::clone(layer).downcast_remote_layer()
+    } else {
+        None
+    }
+}
+
 pub mod tests {
     use super::*;
 
@@ -340,6 +447,19 @@ pub mod tests {
     }
 }
 
+/// Helper enum to hold a PageServerConf, or a path
+///
+/// This is used by DeltaLayer and ImageLayer. Normally, this holds a reference to the
+/// global config, and paths to layer files are constructed using the tenant/timeline
+/// path from the config. But in the 'pagectl' binary, we need to construct a Layer
+/// struct for a file on disk, without having a page server running, so that we have no
+/// config. In that case, we use the Path variant to hold the full path to the file on
+/// disk.
+enum PathOrConf {
+    Path(Utf8PathBuf),
+    Conf(&'static PageServerConf),
+}
+
 /// Range wrapping newtype, which uses display to render Debug.
 ///
 /// Useful with `Key`, which has too verbose `{:?}` for printing multiple layers.

pageserver/src/tenant/storage_layer/delta_layer.rs

@@ -34,17 +34,18 @@ 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, DiskBtreeReader, VisitDirection};
-use crate::tenant::storage_layer::{Layer, ValueReconstructResult, ValueReconstructState};
-use crate::tenant::Timeline;
+use crate::tenant::storage_layer::{
+    PersistentLayer, ValueReconstructResult, ValueReconstructState,
+};
 use crate::virtual_file::VirtualFile;
 use crate::{walrecord, TEMP_FILE_SUFFIX};
 use crate::{DELTA_FILE_MAGIC, STORAGE_FORMAT_VERSION};
 use anyhow::{bail, ensure, Context, Result};
 use camino::{Utf8Path, Utf8PathBuf};
-use pageserver_api::models::LayerAccessKind;
+use pageserver_api::models::{HistoricLayerInfo, LayerAccessKind};
 use rand::{distributions::Alphanumeric, Rng};
 use serde::{Deserialize, Serialize};
-use std::fs::File;
+use std::fs::{self, File};
 use std::io::SeekFrom;
 use std::ops::Range;
 use std::os::unix::fs::FileExt;
@@ -58,7 +59,10 @@ use utils::{
     lsn::Lsn,
 };
 
-use super::{AsLayerDesc, LayerAccessStats, PersistentLayerDesc, ResidentLayer};
+use super::{
+    AsLayerDesc, DeltaFileName, Layer, LayerAccessStats, LayerAccessStatsReset, PathOrConf,
+    PersistentLayerDesc,
+};
 
 ///
 /// Header stored in the beginning of the file
@@ -178,12 +182,20 @@ impl DeltaKey {
     }
 }
 
-/// This is used only from `pagectl`. Within pageserver, all layers are
-/// [`crate::tenant::storage_layer::Layer`], which can hold a [`DeltaLayerInner`].
+/// DeltaLayer is the in-memory data structure associated with an on-disk delta
+/// file.
+///
+/// We keep a DeltaLayer in memory for each file, in the LayerMap. If a layer
+/// is in "loaded" state, we have a copy of the index in memory, in 'inner'.
+/// Otherwise the struct is just a placeholder for a file that exists on disk,
+/// and it needs to be loaded before using it in queries.
 pub struct DeltaLayer {
-    path: Utf8PathBuf,
+    path_or_conf: PathOrConf,
+
     pub desc: PersistentLayerDesc,
+
     access_stats: LayerAccessStats,
+
     inner: OnceCell<Arc<DeltaLayerInner>>,
 }
 
@@ -200,8 +212,6 @@ impl std::fmt::Debug for DeltaLayer {
     }
 }
 
-/// `DeltaLayerInner` is the in-memory data structure associated with an on-disk delta
-/// file.
 pub struct DeltaLayerInner {
     // values copied from summary
     index_start_blk: u32,
@@ -211,6 +221,12 @@ pub struct DeltaLayerInner {
     file: FileBlockReader,
 }
 
+impl AsRef<DeltaLayerInner> for DeltaLayerInner {
+    fn as_ref(&self) -> &DeltaLayerInner {
+        self
+    }
+}
+
 impl std::fmt::Debug for DeltaLayerInner {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
         f.debug_struct("DeltaLayerInner")
@@ -220,6 +236,19 @@ impl std::fmt::Debug for DeltaLayerInner {
     }
 }
 
+#[async_trait::async_trait]
+impl Layer for DeltaLayer {
+    async fn get_value_reconstruct_data(
+        &self,
+        key: Key,
+        lsn_range: Range<Lsn>,
+        reconstruct_state: &mut ValueReconstructState,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<ValueReconstructResult> {
+        self.get_value_reconstruct_data(key, lsn_range, reconstruct_state, ctx)
+            .await
+    }
+}
 /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
 impl std::fmt::Display for DeltaLayer {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
@@ -233,9 +262,40 @@ impl AsLayerDesc for DeltaLayer {
     }
 }
 
+impl PersistentLayer for DeltaLayer {
+    fn downcast_delta_layer(self: Arc<Self>) -> Option<std::sync::Arc<DeltaLayer>> {
+        Some(self)
+    }
+
+    fn local_path(&self) -> Option<Utf8PathBuf> {
+        self.local_path()
+    }
+
+    fn delete_resident_layer_file(&self) -> Result<()> {
+        self.delete_resident_layer_file()
+    }
+
+    fn info(&self, reset: LayerAccessStatsReset) -> HistoricLayerInfo {
+        self.info(reset)
+    }
+
+    fn access_stats(&self) -> &LayerAccessStats {
+        self.access_stats()
+    }
+}
+
 impl DeltaLayer {
     pub(crate) async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
-        self.desc.dump();
+        println!(
+            "----- delta layer for ten {} tli {} keys {}-{} lsn {}-{} size {} ----",
+            self.desc.tenant_id,
+            self.desc.timeline_id,
+            self.desc.key_range.start,
+            self.desc.key_range.end,
+            self.desc.lsn_range.start,
+            self.desc.lsn_range.end,
+            self.desc.file_size,
+        );
 
         if !verbose {
             return Ok(());
@@ -243,7 +303,119 @@ impl DeltaLayer {
 
         let inner = self.load(LayerAccessKind::Dump, ctx).await?;
 
-        inner.dump(ctx).await
+        println!(
+            "index_start_blk: {}, root {}",
+            inner.index_start_blk, inner.index_root_blk
+        );
+
+        let file = &inner.file;
+        let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
+            inner.index_start_blk,
+            inner.index_root_blk,
+            file,
+        );
+
+        tree_reader.dump().await?;
+
+        let keys = DeltaLayerInner::load_keys(&inner, ctx).await?;
+
+        // A subroutine to dump a single blob
+        async fn dump_blob(val: ValueRef<'_>, ctx: &RequestContext) -> Result<String> {
+            let buf = val.reader.read_blob(val.blob_ref.pos(), ctx).await?;
+            let val = Value::des(&buf)?;
+            let desc = match val {
+                Value::Image(img) => {
+                    format!(" img {} bytes", img.len())
+                }
+                Value::WalRecord(rec) => {
+                    let wal_desc = walrecord::describe_wal_record(&rec)?;
+                    format!(
+                        " rec {} bytes will_init: {} {}",
+                        buf.len(),
+                        rec.will_init(),
+                        wal_desc
+                    )
+                }
+            };
+            Ok(desc)
+        }
+
+        for entry in keys {
+            let DeltaEntry { key, lsn, val, .. } = entry;
+            let desc = match dump_blob(val, ctx).await {
+                Ok(desc) => desc,
+                Err(err) => {
+                    let err: anyhow::Error = err;
+                    format!("ERROR: {err}")
+                }
+            };
+            println!("  key {key} at {lsn}: {desc}");
+        }
+
+        Ok(())
+    }
+
+    pub(crate) async fn get_value_reconstruct_data(
+        &self,
+        key: Key,
+        lsn_range: Range<Lsn>,
+        reconstruct_state: &mut ValueReconstructState,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<ValueReconstructResult> {
+        ensure!(lsn_range.start >= self.desc.lsn_range.start);
+
+        ensure!(self.desc.key_range.contains(&key));
+
+        let inner = self
+            .load(LayerAccessKind::GetValueReconstructData, ctx)
+            .await?;
+        inner
+            .get_value_reconstruct_data(key, lsn_range, reconstruct_state, ctx)
+            .await
+    }
+
+    pub(crate) fn local_path(&self) -> Option<Utf8PathBuf> {
+        Some(self.path())
+    }
+
+    pub(crate) fn delete_resident_layer_file(&self) -> Result<()> {
+        // delete underlying file
+        fs::remove_file(self.path())?;
+        Ok(())
+    }
+
+    pub(crate) fn info(&self, reset: LayerAccessStatsReset) -> HistoricLayerInfo {
+        let layer_file_name = self.layer_desc().filename().file_name();
+        let lsn_range = self.layer_desc().lsn_range.clone();
+
+        let access_stats = self.access_stats.as_api_model(reset);
+
+        HistoricLayerInfo::Delta {
+            layer_file_name,
+            layer_file_size: self.desc.file_size,
+            lsn_start: lsn_range.start,
+            lsn_end: lsn_range.end,
+            remote: false,
+            access_stats,
+        }
+    }
+
+    pub(crate) fn access_stats(&self) -> &LayerAccessStats {
+        &self.access_stats
+    }
+
+    fn path_for(
+        path_or_conf: &PathOrConf,
+        tenant_id: &TenantId,
+        timeline_id: &TimelineId,
+        fname: &DeltaFileName,
+    ) -> Utf8PathBuf {
+        match path_or_conf {
+            PathOrConf::Path(path) => path.clone(),
+            PathOrConf::Conf(conf) => conf
+                .timeline_path(tenant_id, timeline_id)
+                .join(fname.to_string()),
+        }
     }
 
     fn temp_path_for(
@@ -289,21 +461,52 @@ impl DeltaLayer {
     async fn load_inner(&self, ctx: &RequestContext) -> Result<Arc<DeltaLayerInner>> {
         let path = self.path();
 
-        let loaded = DeltaLayerInner::load(&path, None, ctx).await?;
+        let summary = match &self.path_or_conf {
+            PathOrConf::Conf(_) => Some(Summary::from(self)),
+            PathOrConf::Path(_) => None,
+        };
 
-        // not production code
-        let actual_filename = path.file_name().unwrap().to_owned();
-        let expected_filename = self.layer_desc().filename().file_name();
+        let loaded = DeltaLayerInner::load(&path, summary, ctx).await?;
 
-        if actual_filename != expected_filename {
-            println!("warning: filename does not match what is expected from in-file summary");
-            println!("actual: {:?}", actual_filename);
-            println!("expected: {:?}", expected_filename);
+        if let PathOrConf::Path(ref path) = self.path_or_conf {
+            // not production code
+
+            let actual_filename = path.file_name().unwrap().to_owned();
+            let expected_filename = self.filename().file_name();
+
+            if actual_filename != expected_filename {
+                println!("warning: filename does not match what is expected from in-file summary");
+                println!("actual: {:?}", actual_filename);
+                println!("expected: {:?}", expected_filename);
+            }
         }
 
         Ok(Arc::new(loaded))
     }
 
+    /// Create a DeltaLayer struct representing an existing file on disk.
+    pub fn new(
+        conf: &'static PageServerConf,
+        timeline_id: TimelineId,
+        tenant_id: TenantId,
+        filename: &DeltaFileName,
+        file_size: u64,
+        access_stats: LayerAccessStats,
+    ) -> DeltaLayer {
+        DeltaLayer {
+            path_or_conf: PathOrConf::Conf(conf),
+            desc: PersistentLayerDesc::new_delta(
+                tenant_id,
+                timeline_id,
+                filename.key_range.clone(),
+                filename.lsn_range.clone(),
+                file_size,
+            ),
+            access_stats,
+            inner: OnceCell::new(),
+        }
+    }
+
     /// Create a DeltaLayer struct representing an existing file on disk.
     ///
     /// This variant is only used for debugging purposes, by the 'pagectl' binary.
@@ -317,7 +520,7 @@ impl DeltaLayer {
             .context("get file metadata to determine size")?;
 
         Ok(DeltaLayer {
-            path: path.to_path_buf(),
+            path_or_conf: PathOrConf::Path(path.to_path_buf()),
             desc: PersistentLayerDesc::new_delta(
                 summary.tenant_id,
                 summary.timeline_id,
@@ -330,9 +533,29 @@ impl DeltaLayer {
         })
     }
 
+    fn layer_name(&self) -> DeltaFileName {
+        self.desc.delta_file_name()
+    }
     /// Path to the layer file in pageserver workdir.
-    fn path(&self) -> Utf8PathBuf {
-        self.path.clone()
+    pub fn path(&self) -> Utf8PathBuf {
+        Self::path_for(
+            &self.path_or_conf,
+            &self.desc.tenant_id,
+            &self.desc.timeline_id,
+            &self.layer_name(),
+        )
+    }
+    /// Loads all keys stored in the layer. Returns key, lsn, value size and value reference.
+    ///
+    /// The value can be obtained via the [`ValueRef::load`] function.
+    pub(crate) async fn load_keys(&self, ctx: &RequestContext) -> Result<Vec<DeltaEntry<'_>>> {
+        let inner = self
+            .load(LayerAccessKind::KeyIter, ctx)
+            .await
+            .context("load delta layer keys")?;
+        DeltaLayerInner::load_keys(inner, ctx)
+            .await
+            .context("Layer index is corrupted")
     }
 }
 
@@ -437,7 +660,7 @@ impl DeltaLayerWriterInner {
     ///
     /// Finish writing the delta layer.
     ///
-    async fn finish(self, key_end: Key, timeline: &Arc<Timeline>) -> anyhow::Result<ResidentLayer> {
+    async fn finish(self, key_end: Key) -> anyhow::Result<DeltaLayer> {
         let index_start_blk =
             ((self.blob_writer.size() + PAGE_SZ as u64 - 1) / PAGE_SZ as u64) as u32;
 
@@ -494,21 +717,37 @@ impl DeltaLayerWriterInner {
         // Note: Because we opened the file in write-only mode, we cannot
         // reuse the same VirtualFile for reading later. That's why we don't
         // set inner.file here. The first read will have to re-open it.
-
-        let desc = PersistentLayerDesc::new_delta(
-            self.tenant_id,
-            self.timeline_id,
-            self.key_start..key_end,
-            self.lsn_range.clone(),
-            metadata.len(),
-        );
+        let layer = DeltaLayer {
+            path_or_conf: PathOrConf::Conf(self.conf),
+            desc: PersistentLayerDesc::new_delta(
+                self.tenant_id,
+                self.timeline_id,
+                self.key_start..key_end,
+                self.lsn_range.clone(),
+                metadata.len(),
+            ),
+            access_stats: LayerAccessStats::empty_will_record_residence_event_later(),
+            inner: OnceCell::new(),
+        };
 
         // fsync the file
         file.sync_all().await?;
+        // Rename the file to its final name
+        //
+        // Note: This overwrites any existing file. There shouldn't be any.
+        // FIXME: throw an error instead?
+        let final_path = DeltaLayer::path_for(
+            &PathOrConf::Conf(self.conf),
+            &self.tenant_id,
+            &self.timeline_id,
+            &DeltaFileName {
+                key_range: self.key_start..key_end,
+                lsn_range: self.lsn_range,
+            },
+        );
+        std::fs::rename(self.path, &final_path)?;
 
-        let layer = Layer::finish_creating(self.conf, timeline, desc, &self.path)?;
-
-        trace!("created delta layer {}", layer.local_path());
+        trace!("created delta layer {final_path}");
 
         Ok(layer)
     }
@@ -589,12 +828,8 @@ impl DeltaLayerWriter {
     ///
     /// Finish writing the delta layer.
     ///
-    pub(crate) async fn finish(
-        mut self,
-        key_end: Key,
-        timeline: &Arc<Timeline>,
-    ) -> anyhow::Result<ResidentLayer> {
-        self.inner.take().unwrap().finish(key_end, timeline).await
+    pub async fn finish(mut self, key_end: Key) -> anyhow::Result<DeltaLayer> {
+        self.inner.take().unwrap().finish(key_end).await
     }
 }
 
@@ -609,6 +844,49 @@ impl Drop for DeltaLayerWriter {
     }
 }
 
+impl DeltaLayer {
+    /// Assume the file at `path` is corrupt if this function returns with an error.
+    pub(crate) async fn rewrite_tenant_timeline(
+        path: &Utf8Path,
+        new_tenant: TenantId,
+        new_timeline: TimelineId,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<()> {
+        let file = VirtualFile::open_with_options(
+            path,
+            &*std::fs::OpenOptions::new().read(true).write(true),
+        )
+        .await
+        .with_context(|| format!("Failed to open file '{}'", path))?;
+        let file = FileBlockReader::new(file);
+        let summary_blk = file.read_blk(0, ctx).await?;
+        let actual_summary = Summary::des_prefix(summary_blk.as_ref())?;
+        let mut file = file.file;
+        if actual_summary.magic != DELTA_FILE_MAGIC {
+            bail!("File '{}' is not a delta layer", path);
+        }
+        let new_summary = Summary {
+            tenant_id: new_tenant,
+            timeline_id: new_timeline,
+            ..actual_summary
+        };
+
+        let mut buf = smallvec::SmallVec::<[u8; PAGE_SZ]>::new();
+        Summary::ser_into(&new_summary, &mut buf)?;
+        if buf.spilled() {
+            // The code in ImageLayerWriterInner just warn!()s for this.
+            // It should probably error out as well.
+            anyhow::bail!(
+                "Used more than one page size for summary buffer: {}",
+                buf.len()
+            );
+        }
+        file.seek(SeekFrom::Start(0)).await?;
+        file.write_all(&buf).await?;
+        Ok(())
+    }
+}
+
 impl DeltaLayerInner {
     pub(super) async fn load(
         path: &Utf8Path,
@@ -732,17 +1010,15 @@ impl DeltaLayerInner {
         }
     }
 
-    pub(super) async fn load_keys<'a>(
-        &'a self,
-        ctx: &RequestContext,
+    pub(super) async fn load_keys<'a, 'b, T: AsRef<DeltaLayerInner> + Clone>(
+        this: &'a T,
+        ctx: &'b RequestContext,
     ) -> Result<Vec<DeltaEntry<'a>>> {
-        let file = &self.file;
+        let dl = this.as_ref();
+        let file = &dl.file;
 
-        let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
-            self.index_start_blk,
-            self.index_root_blk,
-            file,
-        );
+        let tree_reader =
+            DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(dl.index_start_blk, dl.index_root_blk, file);
 
         let mut all_keys: Vec<DeltaEntry<'_>> = Vec::new();
 
@@ -755,7 +1031,7 @@ impl DeltaLayerInner {
                     let val_ref = ValueRef {
                         blob_ref: BlobRef(value),
                         reader: BlockCursor::new(crate::tenant::block_io::BlockReaderRef::Adapter(
-                            Adapter(self),
+                            Adapter(dl),
                         )),
                     };
                     let pos = BlobRef(value).pos();
@@ -782,61 +1058,10 @@ impl DeltaLayerInner {
         if let Some(last) = all_keys.last_mut() {
             // Last key occupies all space till end of value storage,
             // which corresponds to beginning of the index
-            last.size = self.index_start_blk as u64 * PAGE_SZ as u64 - last.size;
+            last.size = dl.index_start_blk as u64 * PAGE_SZ as u64 - last.size;
         }
         Ok(all_keys)
     }
-
-    pub(super) async fn dump(&self, ctx: &RequestContext) -> anyhow::Result<()> {
-        println!(
-            "index_start_blk: {}, root {}",
-            self.index_start_blk, self.index_root_blk
-        );
-
-        let file = &self.file;
-        let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
-            self.index_start_blk,
-            self.index_root_blk,
-            file,
-        );
-
-        tree_reader.dump().await?;
-
-        let keys = self.load_keys(ctx).await?;
-
-        async fn dump_blob(val: ValueRef<'_>, ctx: &RequestContext) -> anyhow::Result<String> {
-            let buf = val.reader.read_blob(val.blob_ref.pos(), ctx).await?;
-            let val = Value::des(&buf)?;
-            let desc = match val {
-                Value::Image(img) => {
-                    format!(" img {} bytes", img.len())
-                }
-                Value::WalRecord(rec) => {
-                    let wal_desc = walrecord::describe_wal_record(&rec)?;
-                    format!(
-                        " rec {} bytes will_init: {} {}",
-                        buf.len(),
-                        rec.will_init(),
-                        wal_desc
-                    )
-                }
-            };
-            Ok(desc)
-        }
-
-        for entry in keys {
-            let DeltaEntry { key, lsn, val, .. } = entry;
-            let desc = match dump_blob(val, ctx).await {
-                Ok(desc) => desc,
-                Err(err) => {
-                    format!("ERROR: {err}")
-                }
-            };
-            println!("  key {key} at {lsn}: {desc}");
-        }
-
-        Ok(())
-    }
 }
 
 /// A set of data associated with a delta layer key and its value
@@ -876,9 +1101,3 @@ impl<T: AsRef<DeltaLayerInner>> Adapter<T> {
         self.0.as_ref().file.read_blk(blknum, ctx).await
     }
 }
-
-impl AsRef<DeltaLayerInner> for DeltaLayerInner {
-    fn as_ref(&self) -> &DeltaLayerInner {
-        self
-    }
-}

pageserver/src/tenant/storage_layer/image_layer.rs

@@ -31,23 +31,21 @@ use crate::tenant::blob_io::BlobWriter;
 use crate::tenant::block_io::{BlockBuf, BlockReader, FileBlockReader};
 use crate::tenant::disk_btree::{DiskBtreeBuilder, DiskBtreeReader, VisitDirection};
 use crate::tenant::storage_layer::{
-    LayerAccessStats, ValueReconstructResult, ValueReconstructState,
+    LayerAccessStats, PersistentLayer, ValueReconstructResult, ValueReconstructState,
 };
-use crate::tenant::Timeline;
 use crate::virtual_file::VirtualFile;
 use crate::{IMAGE_FILE_MAGIC, STORAGE_FORMAT_VERSION, TEMP_FILE_SUFFIX};
 use anyhow::{bail, ensure, Context, Result};
 use bytes::Bytes;
 use camino::{Utf8Path, Utf8PathBuf};
 use hex;
-use pageserver_api::models::LayerAccessKind;
+use pageserver_api::models::{HistoricLayerInfo, LayerAccessKind};
 use rand::{distributions::Alphanumeric, Rng};
 use serde::{Deserialize, Serialize};
-use std::fs::File;
+use std::fs::{self, File};
 use std::io::SeekFrom;
 use std::ops::Range;
 use std::os::unix::prelude::FileExt;
-use std::sync::Arc;
 use tokio::sync::OnceCell;
 use tracing::*;
 
@@ -58,7 +56,7 @@ use utils::{
 };
 
 use super::filename::ImageFileName;
-use super::{AsLayerDesc, Layer, PersistentLayerDesc, ResidentLayer};
+use super::{AsLayerDesc, Layer, LayerAccessStatsReset, PathOrConf, PersistentLayerDesc};
 
 ///
 /// Header stored in the beginning of the file
@@ -116,14 +114,22 @@ impl Summary {
     }
 }
 
-/// This is used only from `pagectl`. Within pageserver, all layers are
-/// [`crate::tenant::storage_layer::Layer`], which can hold an [`ImageLayerInner`].
+/// ImageLayer is the in-memory data structure associated with an on-disk image
+/// file.
+///
+/// We keep an ImageLayer in memory for each file, in the LayerMap. If a layer
+/// is in "loaded" state, we have a copy of the index in memory, in 'inner'.
+/// Otherwise the struct is just a placeholder for a file that exists on disk,
+/// and it needs to be loaded before using it in queries.
 pub struct ImageLayer {
-    path: Utf8PathBuf,
+    path_or_conf: PathOrConf,
+
     pub desc: PersistentLayerDesc,
     // This entry contains an image of all pages as of this LSN, should be the same as desc.lsn
     pub lsn: Lsn,
+
     access_stats: LayerAccessStats,
+
     inner: OnceCell<ImageLayerInner>,
 }
 
@@ -140,8 +146,6 @@ impl std::fmt::Debug for ImageLayer {
     }
 }
 
-/// ImageLayer is the in-memory data structure associated with an on-disk image
-/// file.
 pub struct ImageLayerInner {
     // values copied from summary
     index_start_blk: u32,
@@ -162,11 +166,73 @@ impl std::fmt::Debug for ImageLayerInner {
     }
 }
 
-impl ImageLayerInner {
-    pub(super) async fn dump(&self, ctx: &RequestContext) -> anyhow::Result<()> {
-        let file = &self.file;
+#[async_trait::async_trait]
+impl Layer for ImageLayer {
+    /// Look up given page in the file
+    async fn get_value_reconstruct_data(
+        &self,
+        key: Key,
+        lsn_range: Range<Lsn>,
+        reconstruct_state: &mut ValueReconstructState,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<ValueReconstructResult> {
+        self.get_value_reconstruct_data(key, lsn_range, reconstruct_state, ctx)
+            .await
+    }
+}
+
+/// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
+impl std::fmt::Display for ImageLayer {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{}", self.layer_desc().short_id())
+    }
+}
+
+impl AsLayerDesc for ImageLayer {
+    fn layer_desc(&self) -> &PersistentLayerDesc {
+        &self.desc
+    }
+}
+
+impl PersistentLayer for ImageLayer {
+    fn local_path(&self) -> Option<Utf8PathBuf> {
+        self.local_path()
+    }
+
+    fn delete_resident_layer_file(&self) -> Result<()> {
+        self.delete_resident_layer_file()
+    }
+
+    fn info(&self, reset: LayerAccessStatsReset) -> HistoricLayerInfo {
+        self.info(reset)
+    }
+
+    fn access_stats(&self) -> &LayerAccessStats {
+        self.access_stats()
+    }
+}
+
+impl ImageLayer {
+    pub(crate) async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
+        println!(
+            "----- image layer for ten {} tli {} key {}-{} at {} is_incremental {} size {} ----",
+            self.desc.tenant_id,
+            self.desc.timeline_id,
+            self.desc.key_range.start,
+            self.desc.key_range.end,
+            self.lsn,
+            self.desc.is_incremental(),
+            self.desc.file_size
+        );
+
+        if !verbose {
+            return Ok(());
+        }
+
+        let inner = self.load(LayerAccessKind::Dump, ctx).await?;
+        let file = &inner.file;
         let tree_reader =
-            DiskBtreeReader::<_, KEY_SIZE>::new(self.index_start_blk, self.index_root_blk, file);
+            DiskBtreeReader::<_, KEY_SIZE>::new(inner.index_start_blk, inner.index_root_blk, file);
 
         tree_reader.dump().await?;
 
@@ -184,36 +250,69 @@ impl ImageLayerInner {
 
         Ok(())
     }
-}
 
-/// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
-impl std::fmt::Display for ImageLayer {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        write!(f, "{}", self.layer_desc().short_id())
+    pub(crate) async fn get_value_reconstruct_data(
+        &self,
+        key: Key,
+        lsn_range: Range<Lsn>,
+        reconstruct_state: &mut ValueReconstructState,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<ValueReconstructResult> {
+        assert!(self.desc.key_range.contains(&key));
+        assert!(lsn_range.start >= self.lsn);
+        assert!(lsn_range.end >= self.lsn);
+
+        let inner = self
+            .load(LayerAccessKind::GetValueReconstructData, ctx)
+            .await?;
+        inner
+            .get_value_reconstruct_data(key, reconstruct_state, ctx)
+            .await
+            // FIXME: makes no sense to dump paths
+            .with_context(|| format!("read {}", self.path()))
     }
-}
 
-impl AsLayerDesc for ImageLayer {
-    fn layer_desc(&self) -> &PersistentLayerDesc {
-        &self.desc
+    pub(crate) fn local_path(&self) -> Option<Utf8PathBuf> {
+        Some(self.path())
     }
-}
-
-impl ImageLayer {
-    pub(crate) async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
-        self.desc.dump();
-
-        if !verbose {
-            return Ok(());
-        }
-
-        let inner = self.load(LayerAccessKind::Dump, ctx).await?;
-
-        inner.dump(ctx).await?;
 
+    pub(crate) fn delete_resident_layer_file(&self) -> Result<()> {
+        // delete underlying file
+        fs::remove_file(self.path())?;
         Ok(())
     }
 
+    pub(crate) fn info(&self, reset: LayerAccessStatsReset) -> HistoricLayerInfo {
+        let layer_file_name = self.layer_desc().filename().file_name();
+        let lsn_start = self.layer_desc().image_layer_lsn();
+
+        HistoricLayerInfo::Image {
+            layer_file_name,
+            layer_file_size: self.desc.file_size,
+            lsn_start,
+            remote: false,
+            access_stats: self.access_stats.as_api_model(reset),
+        }
+    }
+
+    pub(crate) fn access_stats(&self) -> &LayerAccessStats {
+        &self.access_stats
+    }
+
+    fn path_for(
+        path_or_conf: &PathOrConf,
+        timeline_id: TimelineId,
+        tenant_id: TenantId,
+        fname: &ImageFileName,
+    ) -> Utf8PathBuf {
+        match path_or_conf {
+            PathOrConf::Path(path) => path.to_path_buf(),
+            PathOrConf::Conf(conf) => conf
+                .timeline_path(&tenant_id, &timeline_id)
+                .join(fname.to_string()),
+        }
+    }
+
     fn temp_path_for(
         conf: &PageServerConf,
         timeline_id: TimelineId,
@@ -249,21 +348,54 @@ impl ImageLayer {
     async fn load_inner(&self, ctx: &RequestContext) -> Result<ImageLayerInner> {
         let path = self.path();
 
-        let loaded = ImageLayerInner::load(&path, self.desc.image_layer_lsn(), None, ctx).await?;
+        let expected_summary = match &self.path_or_conf {
+            PathOrConf::Conf(_) => Some(Summary::from(self)),
+            PathOrConf::Path(_) => None,
+        };
 
-        // not production code
-        let actual_filename = path.file_name().unwrap().to_owned();
-        let expected_filename = self.layer_desc().filename().file_name();
+        let loaded =
+            ImageLayerInner::load(&path, self.desc.image_layer_lsn(), expected_summary, ctx)
+                .await?;
 
-        if actual_filename != expected_filename {
-            println!("warning: filename does not match what is expected from in-file summary");
-            println!("actual: {:?}", actual_filename);
-            println!("expected: {:?}", expected_filename);
+        if let PathOrConf::Path(ref path) = self.path_or_conf {
+            // not production code
+            let actual_filename = path.file_name().unwrap().to_owned();
+            let expected_filename = self.filename().file_name();
+
+            if actual_filename != expected_filename {
+                println!("warning: filename does not match what is expected from in-file summary");
+                println!("actual: {:?}", actual_filename);
+                println!("expected: {:?}", expected_filename);
+            }
         }
 
         Ok(loaded)
     }
 
+    /// Create an ImageLayer struct representing an existing file on disk
+    pub fn new(
+        conf: &'static PageServerConf,
+        timeline_id: TimelineId,
+        tenant_id: TenantId,
+        filename: &ImageFileName,
+        file_size: u64,
+        access_stats: LayerAccessStats,
+    ) -> ImageLayer {
+        ImageLayer {
+            path_or_conf: PathOrConf::Conf(conf),
+            desc: PersistentLayerDesc::new_img(
+                tenant_id,
+                timeline_id,
+                filename.key_range.clone(),
+                filename.lsn,
+                file_size,
+            ), // Now we assume image layer ALWAYS covers the full range. This may change in the future.
+            lsn: filename.lsn,
+            access_stats,
+            inner: OnceCell::new(),
+        }
+    }
+
     /// Create an ImageLayer struct representing an existing file on disk.
     ///
     /// This variant is only used for debugging purposes, by the 'pagectl' binary.
@@ -275,7 +407,7 @@ impl ImageLayer {
             .metadata()
             .context("get file metadata to determine size")?;
         Ok(ImageLayer {
-            path: path.to_path_buf(),
+            path_or_conf: PathOrConf::Path(path.to_path_buf()),
             desc: PersistentLayerDesc::new_img(
                 summary.tenant_id,
                 summary.timeline_id,
@@ -289,8 +421,61 @@ impl ImageLayer {
         })
     }
 
-    fn path(&self) -> Utf8PathBuf {
-        self.path.clone()
+    fn layer_name(&self) -> ImageFileName {
+        self.desc.image_file_name()
+    }
+
+    /// Path to the layer file in pageserver workdir.
+    pub fn path(&self) -> Utf8PathBuf {
+        Self::path_for(
+            &self.path_or_conf,
+            self.desc.timeline_id,
+            self.desc.tenant_id,
+            &self.layer_name(),
+        )
+    }
+}
+
+impl ImageLayer {
+    /// Assume the file at `path` is corrupt if this function returns with an error.
+    pub(crate) async fn rewrite_tenant_timeline(
+        path: &Utf8Path,
+        new_tenant: TenantId,
+        new_timeline: TimelineId,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<()> {
+        let file = VirtualFile::open_with_options(
+            path,
+            &*std::fs::OpenOptions::new().read(true).write(true),
+        )
+        .await
+        .with_context(|| format!("Failed to open file '{}'", path))?;
+        let file = FileBlockReader::new(file);
+        let summary_blk = file.read_blk(0, ctx).await?;
+        let actual_summary = Summary::des_prefix(summary_blk.as_ref())?;
+        let mut file = file.file;
+        if actual_summary.magic != IMAGE_FILE_MAGIC {
+            bail!("File '{}' is not a delta layer", path);
+        }
+        let new_summary = Summary {
+            tenant_id: new_tenant,
+            timeline_id: new_timeline,
+            ..actual_summary
+        };
+
+        let mut buf = smallvec::SmallVec::<[u8; PAGE_SZ]>::new();
+        Summary::ser_into(&new_summary, &mut buf)?;
+        if buf.spilled() {
+            // The code in ImageLayerWriterInner just warn!()s for this.
+            // It should probably error out as well.
+            anyhow::bail!(
+                "Used more than one page size for summary buffer: {}",
+                buf.len()
+            );
+        }
+        file.seek(SeekFrom::Start(0)).await?;
+        file.write_all(&buf).await?;
+        Ok(())
     }
 }
 
@@ -462,7 +647,7 @@ impl ImageLayerWriterInner {
     ///
     /// Finish writing the image layer.
     ///
-    async fn finish(self, timeline: &Arc<Timeline>) -> anyhow::Result<ResidentLayer> {
+    async fn finish(self) -> anyhow::Result<ImageLayer> {
         let index_start_blk =
             ((self.blob_writer.size() + PAGE_SZ as u64 - 1) / PAGE_SZ as u64) as u32;
 
@@ -516,14 +701,33 @@ impl ImageLayerWriterInner {
         // Note: Because we open the file in write-only mode, we cannot
         // reuse the same VirtualFile for reading later. That's why we don't
         // set inner.file here. The first read will have to re-open it.
+        let layer = ImageLayer {
+            path_or_conf: PathOrConf::Conf(self.conf),
+            desc,
+            lsn: self.lsn,
+            access_stats: LayerAccessStats::empty_will_record_residence_event_later(),
+            inner: OnceCell::new(),
+        };
 
         // fsync the file
         file.sync_all().await?;
 
-        // FIXME: why not carry the virtualfile here, it supports renaming?
-        let layer = Layer::finish_creating(self.conf, timeline, desc, &self.path)?;
+        // Rename the file to its final name
+        //
+        // Note: This overwrites any existing file. There shouldn't be any.
+        // FIXME: throw an error instead?
+        let final_path = ImageLayer::path_for(
+            &PathOrConf::Conf(self.conf),
+            self.timeline_id,
+            self.tenant_id,
+            &ImageFileName {
+                key_range: self.key_range.clone(),
+                lsn: self.lsn,
+            },
+        );
+        std::fs::rename(self.path, final_path)?;
 
-        trace!("created image layer {}", layer.local_path());
+        trace!("created image layer {}", layer.path());
 
         Ok(layer)
     }
@@ -585,11 +789,8 @@ impl ImageLayerWriter {
     ///
     /// Finish writing the image layer.
     ///
-    pub(crate) async fn finish(
-        mut self,
-        timeline: &Arc<Timeline>,
-    ) -> anyhow::Result<super::ResidentLayer> {
-        self.inner.take().unwrap().finish(timeline).await
+    pub async fn finish(mut self) -> anyhow::Result<ImageLayer> {
+        self.inner.take().unwrap().finish().await
     }
 }
 

pageserver/src/tenant/storage_layer/inmemory_layer.rs

@@ -10,12 +10,11 @@ use crate::repository::{Key, Value};
 use crate::tenant::block_io::BlockReader;
 use crate::tenant::ephemeral_file::EphemeralFile;
 use crate::tenant::storage_layer::{ValueReconstructResult, ValueReconstructState};
-use crate::tenant::Timeline;
 use crate::walrecord;
 use anyhow::{ensure, Result};
 use pageserver_api::models::InMemoryLayerInfo;
 use std::collections::HashMap;
-use std::sync::{Arc, OnceLock};
+use std::sync::OnceLock;
 use tracing::*;
 use utils::{
     bin_ser::BeSer,
@@ -29,7 +28,7 @@ use std::fmt::Write as _;
 use std::ops::Range;
 use tokio::sync::RwLock;
 
-use super::{DeltaLayerWriter, ResidentLayer};
+use super::{DeltaLayer, DeltaLayerWriter, Layer};
 
 pub struct InMemoryLayer {
     conf: &'static PageServerConf,
@@ -208,6 +207,20 @@ impl InMemoryLayer {
     }
 }
 
+#[async_trait::async_trait]
+impl Layer for InMemoryLayer {
+    async fn get_value_reconstruct_data(
+        &self,
+        key: Key,
+        lsn_range: Range<Lsn>,
+        reconstruct_data: &mut ValueReconstructState,
+        ctx: &RequestContext,
+    ) -> Result<ValueReconstructResult> {
+        self.get_value_reconstruct_data(key, lsn_range, reconstruct_data, ctx)
+            .await
+    }
+}
+
 impl std::fmt::Display for InMemoryLayer {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
         let end_lsn = self.end_lsn_or_max();
@@ -216,13 +229,17 @@ impl std::fmt::Display for InMemoryLayer {
 }
 
 impl InMemoryLayer {
+    ///
     /// Get layer size.
+    ///
     pub async fn size(&self) -> Result<u64> {
         let inner = self.inner.read().await;
         Ok(inner.file.len())
     }
 
+    ///
     /// Create a new, empty, in-memory layer
+    ///
     pub async fn create(
         conf: &'static PageServerConf,
         timeline_id: TimelineId,
@@ -314,11 +331,7 @@ impl InMemoryLayer {
     /// Write this frozen in-memory layer to disk.
     ///
     /// Returns a new delta layer with all the same data as this in-memory layer
-    pub(crate) async fn write_to_disk(
-        &self,
-        timeline: &Arc<Timeline>,
-        ctx: &RequestContext,
-    ) -> Result<ResidentLayer> {
+    pub(crate) async fn write_to_disk(&self, ctx: &RequestContext) -> Result<DeltaLayer> {
         // Grab the lock in read-mode. We hold it over the I/O, but because this
         // layer is not writeable anymore, no one should be trying to acquire the
         // write lock on it, so we shouldn't block anyone. There's one exception
@@ -345,19 +358,14 @@ impl InMemoryLayer {
 
         let cursor = inner.file.block_cursor();
 
-        // Sort the keys because delta layer writer expects them sorted.
-        //
-        // NOTE: this sort can take up significant time if the layer has millions of
-        //       keys. To speed up all the comparisons we convert the key to i128 and
-        //       keep the value as a reference.
-        let mut keys: Vec<_> = inner.index.iter().map(|(k, m)| (k.to_i128(), m)).collect();
-        keys.sort_unstable_by_key(|k| k.0);
+        let mut keys: Vec<(&Key, &VecMap<Lsn, u64>)> = inner.index.iter().collect();
+        keys.sort_by_key(|k| k.0);
 
         let ctx = RequestContextBuilder::extend(ctx)
             .page_content_kind(PageContentKind::InMemoryLayer)
             .build();
         for (key, vec_map) in keys.iter() {
-            let key = Key::from_i128(*key);
+            let key = **key;
             // Write all page versions
             for (lsn, pos) in vec_map.as_slice() {
                 cursor.read_blob_into_buf(*pos, &mut buf, &ctx).await?;
@@ -368,8 +376,7 @@ impl InMemoryLayer {
             }
         }
 
-        // MAX is used here because we identify L0 layers by full key range
-        let delta_layer = delta_layer_writer.finish(Key::MAX, timeline).await?;
+        let delta_layer = delta_layer_writer.finish(Key::MAX).await?;
         Ok(delta_layer)
     }
 }

pageserver/src/tenant/storage_layer/layer_desc.rs

@@ -1,3 +1,4 @@
+use anyhow::Result;
 use core::fmt::Display;
 use std::ops::Range;
 use utils::{
@@ -5,7 +6,7 @@ use utils::{
     lsn::Lsn,
 };
 
-use crate::repository::Key;
+use crate::{context::RequestContext, repository::Key};
 
 use super::{DeltaFileName, ImageFileName, LayerFileName};
 
@@ -99,22 +100,6 @@ impl PersistentLayerDesc {
         }
     }
 
-    pub fn from_filename(
-        tenant_id: TenantId,
-        timeline_id: TimelineId,
-        filename: LayerFileName,
-        file_size: u64,
-    ) -> Self {
-        match filename {
-            LayerFileName::Image(i) => {
-                Self::new_img(tenant_id, timeline_id, i.key_range, i.lsn, file_size)
-            }
-            LayerFileName::Delta(d) => {
-                Self::new_delta(tenant_id, timeline_id, d.key_range, d.lsn_range, file_size)
-            }
-        }
-    }
-
     /// Get the LSN that the image layer covers.
     pub fn image_layer_lsn(&self) -> Lsn {
         assert!(!self.is_delta);
@@ -188,31 +173,21 @@ impl PersistentLayerDesc {
         self.is_delta
     }
 
-    pub fn dump(&self) {
-        if self.is_delta {
-            println!(
-                "----- delta layer for ten {} tli {} keys {}-{} lsn {}-{} is_incremental {} size {} ----",
-                self.tenant_id,
-                self.timeline_id,
-                self.key_range.start,
-                self.key_range.end,
-                self.lsn_range.start,
-                self.lsn_range.end,
-                self.is_incremental(),
-                self.file_size,
-            );
-        } else {
-            println!(
-                "----- image layer for ten {} tli {} key {}-{} at {} is_incremental {} size {} ----",
-                self.tenant_id,
-                self.timeline_id,
-                self.key_range.start,
-                self.key_range.end,
-                self.image_layer_lsn(),
-                self.is_incremental(),
-                self.file_size
-            );
-        }
+    pub fn dump(&self, _verbose: bool, _ctx: &RequestContext) -> Result<()> {
+        println!(
+            "----- layer for ten {} tli {} keys {}-{} lsn {}-{} is_delta {} is_incremental {} size {} ----",
+            self.tenant_id,
+            self.timeline_id,
+            self.key_range.start,
+            self.key_range.end,
+            self.lsn_range.start,
+            self.lsn_range.end,
+            self.is_delta,
+            self.is_incremental(),
+            self.file_size,
+        );
+
+        Ok(())
     }
 
     pub fn file_size(&self) -> u64 {

pageserver/src/tenant/storage_layer/remote_layer.rs

@@ -0,0 +1,216 @@
+//! A RemoteLayer is an in-memory placeholder for a layer file that exists
+//! in remote storage.
+//!
+use crate::config::PageServerConf;
+use crate::context::RequestContext;
+use crate::repository::Key;
+use crate::tenant::remote_timeline_client::index::LayerFileMetadata;
+use crate::tenant::storage_layer::{Layer, ValueReconstructResult, ValueReconstructState};
+use crate::tenant::timeline::layer_manager::LayerManager;
+use anyhow::{bail, Result};
+use camino::Utf8PathBuf;
+use pageserver_api::models::HistoricLayerInfo;
+use std::ops::Range;
+use std::sync::Arc;
+
+use utils::{
+    id::{TenantId, TimelineId},
+    lsn::Lsn,
+};
+
+use super::filename::{DeltaFileName, ImageFileName};
+use super::{
+    AsLayerDesc, DeltaLayer, ImageLayer, LayerAccessStats, LayerAccessStatsReset,
+    LayerResidenceStatus, PersistentLayer, PersistentLayerDesc,
+};
+
+/// RemoteLayer is a not yet downloaded [`ImageLayer`] or
+/// [`DeltaLayer`].
+///
+/// RemoteLayer might be downloaded on-demand during operations which are
+/// allowed download remote layers and during which, it gets replaced with a
+/// concrete `DeltaLayer` or `ImageLayer`.
+///
+/// See: [`crate::context::RequestContext`] for authorization to download
+pub struct RemoteLayer {
+    pub desc: PersistentLayerDesc,
+
+    pub layer_metadata: LayerFileMetadata,
+
+    access_stats: LayerAccessStats,
+
+    pub(crate) ongoing_download: Arc<tokio::sync::Semaphore>,
+
+    /// Has `LayerMap::replace` failed for this (true) or not (false).
+    ///
+    /// Used together with [`ongoing_download`] semaphore in `Timeline::download_remote_layer`.
+    /// The field is used to mark a RemoteLayer permanently (until restart or ignore+load)
+    /// unprocessable, because a LayerMap::replace failed.
+    ///
+    /// It is very unlikely to accumulate these in the Timeline's LayerMap, but having this avoids
+    /// a possible fast loop between `Timeline::get_reconstruct_data` and
+    /// `Timeline::download_remote_layer`, which also logs.
+    ///
+    /// [`ongoing_download`]: Self::ongoing_download
+    pub(crate) download_replacement_failure: std::sync::atomic::AtomicBool,
+}
+
+impl std::fmt::Debug for RemoteLayer {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("RemoteLayer")
+            .field("file_name", &self.desc.filename())
+            .field("layer_metadata", &self.layer_metadata)
+            .field("is_incremental", &self.desc.is_incremental())
+            .finish()
+    }
+}
+
+#[async_trait::async_trait]
+impl Layer for RemoteLayer {
+    async fn get_value_reconstruct_data(
+        &self,
+        _key: Key,
+        _lsn_range: Range<Lsn>,
+        _reconstruct_state: &mut ValueReconstructState,
+        _ctx: &RequestContext,
+    ) -> Result<ValueReconstructResult> {
+        Err(anyhow::anyhow!("layer {self} needs to be downloaded"))
+    }
+}
+
+/// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
+impl std::fmt::Display for RemoteLayer {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{}", self.layer_desc().short_id())
+    }
+}
+
+impl AsLayerDesc for RemoteLayer {
+    fn layer_desc(&self) -> &PersistentLayerDesc {
+        &self.desc
+    }
+}
+
+impl PersistentLayer for RemoteLayer {
+    fn local_path(&self) -> Option<Utf8PathBuf> {
+        None
+    }
+
+    fn delete_resident_layer_file(&self) -> Result<()> {
+        bail!("remote layer has no layer file");
+    }
+
+    fn downcast_remote_layer<'a>(self: Arc<Self>) -> Option<std::sync::Arc<RemoteLayer>> {
+        Some(self)
+    }
+
+    fn is_remote_layer(&self) -> bool {
+        true
+    }
+
+    fn info(&self, reset: LayerAccessStatsReset) -> HistoricLayerInfo {
+        let layer_file_name = self.layer_desc().filename().file_name();
+        let lsn_range = self.layer_desc().lsn_range.clone();
+
+        if self.desc.is_delta {
+            HistoricLayerInfo::Delta {
+                layer_file_name,
+                layer_file_size: self.layer_metadata.file_size(),
+                lsn_start: lsn_range.start,
+                lsn_end: lsn_range.end,
+                remote: true,
+                access_stats: self.access_stats.as_api_model(reset),
+            }
+        } else {
+            HistoricLayerInfo::Image {
+                layer_file_name,
+                layer_file_size: self.layer_metadata.file_size(),
+                lsn_start: lsn_range.start,
+                remote: true,
+                access_stats: self.access_stats.as_api_model(reset),
+            }
+        }
+    }
+
+    fn access_stats(&self) -> &LayerAccessStats {
+        &self.access_stats
+    }
+}
+
+impl RemoteLayer {
+    pub fn new_img(
+        tenantid: TenantId,
+        timelineid: TimelineId,
+        fname: &ImageFileName,
+        layer_metadata: &LayerFileMetadata,
+        access_stats: LayerAccessStats,
+    ) -> RemoteLayer {
+        RemoteLayer {
+            desc: PersistentLayerDesc::new_img(
+                tenantid,
+                timelineid,
+                fname.key_range.clone(),
+                fname.lsn,
+                layer_metadata.file_size(),
+            ),
+            layer_metadata: layer_metadata.clone(),
+            ongoing_download: Arc::new(tokio::sync::Semaphore::new(1)),
+            download_replacement_failure: std::sync::atomic::AtomicBool::default(),
+            access_stats,
+        }
+    }
+
+    pub fn new_delta(
+        tenantid: TenantId,
+        timelineid: TimelineId,
+        fname: &DeltaFileName,
+        layer_metadata: &LayerFileMetadata,
+        access_stats: LayerAccessStats,
+    ) -> RemoteLayer {
+        RemoteLayer {
+            desc: PersistentLayerDesc::new_delta(
+                tenantid,
+                timelineid,
+                fname.key_range.clone(),
+                fname.lsn_range.clone(),
+                layer_metadata.file_size(),
+            ),
+            layer_metadata: layer_metadata.clone(),
+            ongoing_download: Arc::new(tokio::sync::Semaphore::new(1)),
+            download_replacement_failure: std::sync::atomic::AtomicBool::default(),
+            access_stats,
+        }
+    }
+
+    /// Create a Layer struct representing this layer, after it has been downloaded.
+    pub(crate) fn create_downloaded_layer(
+        &self,
+        _layer_map_lock_held_witness: &LayerManager,
+        conf: &'static PageServerConf,
+        file_size: u64,
+    ) -> Arc<dyn PersistentLayer> {
+        if self.desc.is_delta {
+            let fname = self.desc.delta_file_name();
+            Arc::new(DeltaLayer::new(
+                conf,
+                self.desc.timeline_id,
+                self.desc.tenant_id,
+                &fname,
+                file_size,
+                self.access_stats
+                    .clone_for_residence_change(LayerResidenceStatus::Resident),
+            ))
+        } else {
+            let fname = self.desc.image_file_name();
+            Arc::new(ImageLayer::new(
+                conf,
+                self.desc.timeline_id,
+                self.desc.tenant_id,
+                &fname,
+                file_size,
+                self.access_stats
+                    .clone_for_residence_change(LayerResidenceStatus::Resident),
+            ))
+        }
+    }
+}

pageserver/src/tenant/tasks.rs

@@ -12,7 +12,7 @@ use crate::task_mgr::{TaskKind, BACKGROUND_RUNTIME};
 use crate::tenant::{Tenant, TenantState};
 use tokio_util::sync::CancellationToken;
 use tracing::*;
-use utils::{backoff, completion};
+use utils::completion;
 
 static CONCURRENT_BACKGROUND_TASKS: once_cell::sync::Lazy<tokio::sync::Semaphore> =
     once_cell::sync::Lazy::new(|| {
@@ -139,10 +139,7 @@ pub fn start_background_loops(
 /// Compaction task's main loop
 ///
 async fn compaction_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
-    const MAX_BACKOFF_SECS: f64 = 300.0;
-    // How many errors we have seen consequtively
-    let mut error_run_count = 0;
-
+    let wait_duration = Duration::from_secs(2);
     TENANT_TASK_EVENTS.with_label_values(&["start"]).inc();
     async {
         let ctx = RequestContext::todo_child(TaskKind::Compaction, DownloadBehavior::Download);
@@ -162,11 +159,8 @@ async fn compaction_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
 
             // TODO: we shouldn't need to await to find tenant and this could be moved outside of
             // loop, #3501. There are also additional "allowed_errors" in tests.
-            if first {
-                first = false;
-                if random_init_delay(period, &cancel).await.is_err() {
-                    break;
-                }
+            if first && random_init_delay(period, &cancel).await.is_err() {
+                break;
             }
 
             let started_at = Instant::now();
@@ -179,24 +173,23 @@ async fn compaction_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
             } else {
                 // Run compaction
                 if let Err(e) = tenant.compaction_iteration(&cancel, &ctx).await {
-                    let wait_duration = backoff::exponential_backoff_duration_seconds(
-                        error_run_count,
-                        1.0,
-                        MAX_BACKOFF_SECS,
-                    );
-                    error_run_count += 1;
-                    error!(
-                        "Compaction failed {error_run_count} times, retrying in {:?}: {e:?}",
-                        wait_duration
-                    );
-                    Duration::from_secs_f64(wait_duration)
+                    error!("Compaction failed, retrying in {:?}: {e:?}", wait_duration);
+                    wait_duration
                 } else {
-                    error_run_count = 0;
                     period
                 }
             };
 
-            warn_when_period_overrun(started_at.elapsed(), period, BackgroundLoopKind::Compaction);
+            if !first {
+                // The first iteration is typically much slower, because all tenants compete for the
+                // compaction sempahore to run, and because of concurrent startup work like initializing
+                // logical sizes.  To avoid routinely spamming warnings, we suppress this log on first iteration.
+                warn_when_period_overrun(
+                    started_at.elapsed(),
+                    period,
+                    BackgroundLoopKind::Compaction,
+                );
+            }
 
             // Sleep
             if tokio::time::timeout(sleep_duration, cancel.cancelled())
@@ -205,6 +198,8 @@ async fn compaction_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
             {
                 break;
             }
+
+            first = false;
         }
     }
     .await;
@@ -215,10 +210,7 @@ async fn compaction_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
 /// GC task's main loop
 ///
 async fn gc_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
-    const MAX_BACKOFF_SECS: f64 = 300.0;
-    // How many errors we have seen consequtively
-    let mut error_run_count = 0;
-
+    let wait_duration = Duration::from_secs(2);
     TENANT_TASK_EVENTS.with_label_values(&["start"]).inc();
     async {
         // GC might require downloading, to find the cutoff LSN that corresponds to the
@@ -239,11 +231,8 @@ async fn gc_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
 
             let period = tenant.get_gc_period();
 
-            if first {
-                first = false;
-                if random_init_delay(period, &cancel).await.is_err() {
-                    break;
-                }
+            if first && random_init_delay(period, &cancel).await.is_err() {
+                break;
             }
 
             let started_at = Instant::now();
@@ -260,24 +249,19 @@ async fn gc_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
                     .gc_iteration(None, gc_horizon, tenant.get_pitr_interval(), &ctx)
                     .await;
                 if let Err(e) = res {
-                    let wait_duration = backoff::exponential_backoff_duration_seconds(
-                        error_run_count,
-                        1.0,
-                        MAX_BACKOFF_SECS,
-                    );
-                    error_run_count += 1;
-                    error!(
-                        "Gc failed {error_run_count} times, retrying in {:?}: {e:?}",
-                        wait_duration
-                    );
-                    Duration::from_secs_f64(wait_duration)
+                    error!("Gc failed, retrying in {:?}: {e:?}", wait_duration);
+                    wait_duration
                 } else {
-                    error_run_count = 0;
                     period
                 }
             };
 
-            warn_when_period_overrun(started_at.elapsed(), period, BackgroundLoopKind::Gc);
+            if !first {
+                // The first iteration is typically much slower, because all tenants compete for the
+                // compaction sempahore to run, and because of concurrent startup work like initializing
+                // logical sizes.  To avoid routinely spamming warnings, we suppress this log on first iteration.
+                warn_when_period_overrun(started_at.elapsed(), period, BackgroundLoopKind::Gc);
+            }
 
             // Sleep
             if tokio::time::timeout(sleep_duration, cancel.cancelled())
@@ -286,6 +270,8 @@ async fn gc_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
             {
                 break;
             }
+
+            first = false;
         }
     }
     .await;
@@ -361,7 +347,7 @@ pub(crate) fn warn_when_period_overrun(
         // humantime does no significant digits clamping whereas Duration's debug is a bit more
         // intelligent. however it makes sense to keep the "configuration format" for period, even
         // though there's no way to output the actual config value.
-        info!(
+        warn!(
             ?elapsed,
             period = %humantime::format_duration(period),
             ?task,

pageserver/src/tenant/timeline/delete.rs

@@ -17,7 +17,6 @@ use crate::{
     deletion_queue::DeletionQueueClient,
     task_mgr::{self, TaskKind},
     tenant::{
-        debug_assert_current_span_has_tenant_and_timeline_id,
         metadata::TimelineMetadata,
         remote_timeline_client::{
             self, PersistIndexPartWithDeletedFlagError, RemoteTimelineClient,
@@ -31,11 +30,6 @@ use super::{Timeline, TimelineResources};
 
 /// Now that the Timeline is in Stopping state, request all the related tasks to shut down.
 async fn stop_tasks(timeline: &Timeline) -> Result<(), DeleteTimelineError> {
-    debug_assert_current_span_has_tenant_and_timeline_id();
-    // Notify any timeline work to drop out of loops/requests
-    tracing::debug!("Cancelling CancellationToken");
-    timeline.cancel.cancel();
-
     // Stop the walreceiver first.
     debug!("waiting for wal receiver to shutdown");
     let maybe_started_walreceiver = { timeline.walreceiver.lock().unwrap().take() };
@@ -44,14 +38,6 @@ async fn stop_tasks(timeline: &Timeline) -> Result<(), DeleteTimelineError> {
     }
     debug!("wal receiver shutdown confirmed");
 
-    // Shut down the layer flush task before the remote client, as one depends on the other
-    task_mgr::shutdown_tasks(
-        Some(TaskKind::LayerFlushTask),
-        Some(timeline.tenant_id),
-        Some(timeline.timeline_id),
-    )
-    .await;
-
     // Prevent new uploads from starting.
     if let Some(remote_client) = timeline.remote_client.as_ref() {
         let res = remote_client.stop();
@@ -80,11 +66,6 @@ async fn stop_tasks(timeline: &Timeline) -> Result<(), DeleteTimelineError> {
             "failpoint: timeline-delete-before-index-deleted-at"
         ))?
     });
-
-    tracing::debug!("Waiting for gate...");
-    timeline.gate.close().await;
-    tracing::debug!("Shutdown complete");
-
     Ok(())
 }
 
@@ -313,7 +294,6 @@ async fn cleanup_remaining_timeline_fs_traces(
     // Remove delete mark
     tokio::fs::remove_file(conf.timeline_delete_mark_file_path(tenant_id, timeline_id))
         .await
-        .or_else(fs_ext::ignore_not_found)
         .context("remove delete mark")
 }
 

pageserver/src/tenant/timeline/eviction_task.rs

@@ -29,6 +29,7 @@ use crate::{
     task_mgr::{self, TaskKind, BACKGROUND_RUNTIME},
     tenant::{
         config::{EvictionPolicy, EvictionPolicyLayerAccessThreshold},
+        storage_layer::PersistentLayer,
         tasks::{BackgroundLoopKind, RateLimitError},
         timeline::EvictionError,
         LogicalSizeCalculationCause, Tenant,
@@ -209,26 +210,15 @@ impl Timeline {
         // NB: all the checks can be invalidated as soon as we release the layer map lock.
         // We don't want to hold the layer map lock during eviction.
         // So, we just need to deal with this.
-        let candidates: Vec<_> = {
+        let candidates: Vec<Arc<dyn PersistentLayer>> = {
             let guard = self.layers.read().await;
             let layers = guard.layer_map();
             let mut candidates = Vec::new();
             for hist_layer in layers.iter_historic_layers() {
                 let hist_layer = guard.get_from_desc(&hist_layer);
-
-                // guard against eviction while we inspect it; it might be that eviction_task and
-                // disk_usage_eviction_task both select the same layers to be evicted, and
-                // seemingly free up double the space. both succeeding is of no consequence.
-                let guard = match hist_layer.keep_resident().await {
-                    Ok(Some(l)) => l,
-                    Ok(None) => continue,
-                    Err(e) => {
-                        // these should not happen, but we cannot make them statically impossible right
-                        // now.
-                        tracing::warn!(layer=%hist_layer, "failed to keep the layer resident: {e:#}");
-                        continue;
-                    }
-                };
+                if hist_layer.is_remote_layer() {
+                    continue;
+                }
 
                 let last_activity_ts = hist_layer.access_stats().latest_activity().unwrap_or_else(|| {
                     // We only use this fallback if there's an implementation error.
@@ -259,7 +249,7 @@ impl Timeline {
                     }
                 };
                 if no_activity_for > p.threshold {
-                    candidates.push(guard.drop_eviction_guard())
+                    candidates.push(hist_layer)
                 }
             }
             candidates
@@ -277,7 +267,10 @@ impl Timeline {
             Some(c) => c,
         };
 
-        let results = match self.evict_layer_batch(remote_client, &candidates).await {
+        let results = match self
+            .evict_layer_batch(remote_client, &candidates[..], cancel.clone())
+            .await
+        {
             Err(pre_err) => {
                 stats.errors += candidates.len();
                 error!("could not do any evictions: {pre_err:#}");
@@ -286,7 +279,7 @@ impl Timeline {
             Ok(results) => results,
         };
         assert_eq!(results.len(), candidates.len());
-        for result in results {
+        for (l, result) in candidates.iter().zip(results) {
             match result {
                 None => {
                     stats.skipped_for_shutdown += 1;
@@ -294,10 +287,24 @@ impl Timeline {
                 Some(Ok(())) => {
                     stats.evicted += 1;
                 }
-                Some(Err(EvictionError::NotFound | EvictionError::Downloaded)) => {
+                Some(Err(EvictionError::CannotEvictRemoteLayer)) => {
+                    stats.not_evictable += 1;
+                }
+                Some(Err(EvictionError::FileNotFound)) => {
                     // compaction/gc removed the file while we were waiting on layer_removal_cs
                     stats.not_evictable += 1;
                 }
+                Some(Err(
+                    e @ EvictionError::LayerNotFound(_) | e @ EvictionError::StatFailed(_),
+                )) => {
+                    let e = utils::error::report_compact_sources(&e);
+                    warn!(layer = %l, "failed to evict layer: {e}");
+                    stats.not_evictable += 1;
+                }
+                Some(Err(EvictionError::MetadataInconsistency(detail))) => {
+                    warn!(layer = %l, "failed to evict layer: {detail}");
+                    stats.not_evictable += 1;
+                }
             }
         }
         if stats.candidates == stats.not_evictable {
@@ -341,7 +348,20 @@ impl Timeline {
         // Make one of the tenant's timelines draw the short straw and run the calculation.
         // The others wait until the calculation is done so that they take into account the
         // imitated accesses that the winner made.
-        let tenant = match crate::tenant::mgr::get_tenant(self.tenant_id, true) {
+        //
+        // It is critical we are responsive to cancellation here. Otherwise, we deadlock with
+        // tenant deletion (holds TENANTS in read mode) any other task that attempts to
+        // acquire TENANTS in write mode before we here call get_tenant.
+        // See https://github.com/neondatabase/neon/issues/5284.
+        let res = tokio::select! {
+            _ = cancel.cancelled() => {
+                return ControlFlow::Break(());
+            }
+            res = crate::tenant::mgr::get_tenant(self.tenant_id, true) => {
+                res
+            }
+        };
+        let tenant = match res {
             Ok(t) => t,
             Err(_) => {
                 return ControlFlow::Break(());

pageserver/src/tenant/timeline/layer_manager.rs

@@ -12,16 +12,27 @@ use crate::{
     tenant::{
         layer_map::{BatchedUpdates, LayerMap},
         storage_layer::{
-            AsLayerDesc, InMemoryLayer, Layer, PersistentLayerDesc, PersistentLayerKey,
-            ResidentLayer,
+            AsLayerDesc, DeltaLayer, ImageLayer, InMemoryLayer, PersistentLayer,
+            PersistentLayerDesc, PersistentLayerKey,
         },
+        timeline::compare_arced_layers,
     },
 };
 
 /// Provides semantic APIs to manipulate the layer map.
 pub(crate) struct LayerManager {
     layer_map: LayerMap,
-    layer_fmgr: LayerFileManager<Layer>,
+    layer_fmgr: LayerFileManager,
+}
+
+/// After GC, the layer map changes will not be applied immediately. Users should manually apply the changes after
+/// scheduling deletes in remote client.
+pub(crate) struct ApplyGcResultGuard<'a>(BatchedUpdates<'a>);
+
+impl ApplyGcResultGuard<'_> {
+    pub(crate) fn flush(self) {
+        self.0.flush();
+    }
 }
 
 impl LayerManager {
@@ -32,7 +43,7 @@ impl LayerManager {
         }
     }
 
-    pub(crate) fn get_from_desc(&self, desc: &PersistentLayerDesc) -> Layer {
+    pub(crate) fn get_from_desc(&self, desc: &PersistentLayerDesc) -> Arc<dyn PersistentLayer> {
         self.layer_fmgr.get_from_desc(desc)
     }
 
@@ -44,12 +55,21 @@ impl LayerManager {
         &self.layer_map
     }
 
+    /// Replace layers in the layer file manager, used in evictions and layer downloads.
+    pub(crate) fn replace_and_verify(
+        &mut self,
+        expected: Arc<dyn PersistentLayer>,
+        new: Arc<dyn PersistentLayer>,
+    ) -> Result<()> {
+        self.layer_fmgr.replace_and_verify(expected, new)
+    }
+
     /// Called from `load_layer_map`. Initialize the layer manager with:
     /// 1. all on-disk layers
     /// 2. next open layer (with disk disk_consistent_lsn LSN)
     pub(crate) fn initialize_local_layers(
         &mut self,
-        on_disk_layers: Vec<Layer>,
+        on_disk_layers: Vec<Arc<dyn PersistentLayer>>,
         next_open_layer_at: Lsn,
     ) {
         let mut updates = self.layer_map.batch_update();
@@ -144,19 +164,10 @@ impl LayerManager {
     }
 
     /// Add image layers to the layer map, called from `create_image_layers`.
-    pub(crate) fn track_new_image_layers(
-        &mut self,
-        image_layers: &[ResidentLayer],
-        metrics: &TimelineMetrics,
-    ) {
+    pub(crate) fn track_new_image_layers(&mut self, image_layers: Vec<ImageLayer>) {
         let mut updates = self.layer_map.batch_update();
         for layer in image_layers {
-            Self::insert_historic_layer(layer.as_ref().clone(), &mut updates, &mut self.layer_fmgr);
-
-            // record these here instead of Layer::finish_creating because otherwise partial
-            // failure with create_image_layers would balloon up the physical size gauge. downside
-            // is that all layers need to be created before metrics are updated.
-            metrics.record_new_file_metrics(layer.layer_desc().file_size);
+            Self::insert_historic_layer(Arc::new(layer), &mut updates, &mut self.layer_fmgr);
         }
         updates.flush();
     }
@@ -164,71 +175,76 @@ impl LayerManager {
     /// Flush a frozen layer and add the written delta layer to the layer map.
     pub(crate) fn finish_flush_l0_layer(
         &mut self,
-        delta_layer: Option<&ResidentLayer>,
+        delta_layer: Option<DeltaLayer>,
         frozen_layer_for_check: &Arc<InMemoryLayer>,
-        metrics: &TimelineMetrics,
     ) {
-        let inmem = self
-            .layer_map
-            .frozen_layers
-            .pop_front()
-            .expect("there must be a inmem layer to flush");
+        let l = self.layer_map.frozen_layers.pop_front();
+        let mut updates = self.layer_map.batch_update();
 
-        // Only one task may call this function at a time (for this
-        // timeline). If two tasks tried to flush the same frozen
+        // Only one thread may call this function at a time (for this
+        // timeline). If two threads tried to flush the same frozen
         // layer to disk at the same time, that would not work.
-        assert_eq!(Arc::as_ptr(&inmem), Arc::as_ptr(frozen_layer_for_check));
+        assert!(compare_arced_layers(&l.unwrap(), frozen_layer_for_check));
 
-        if let Some(l) = delta_layer {
-            let mut updates = self.layer_map.batch_update();
-            Self::insert_historic_layer(l.as_ref().clone(), &mut updates, &mut self.layer_fmgr);
-            metrics.record_new_file_metrics(l.layer_desc().file_size);
-            updates.flush();
+        if let Some(delta_layer) = delta_layer {
+            Self::insert_historic_layer(Arc::new(delta_layer), &mut updates, &mut self.layer_fmgr);
         }
+        updates.flush();
     }
 
     /// Called when compaction is completed.
     pub(crate) fn finish_compact_l0(
         &mut self,
-        layer_removal_cs: &Arc<tokio::sync::OwnedMutexGuard<()>>,
-        compact_from: &[Layer],
-        compact_to: &[ResidentLayer],
+        layer_removal_cs: Arc<tokio::sync::OwnedMutexGuard<()>>,
+        compact_from: Vec<Arc<dyn PersistentLayer>>,
+        compact_to: Vec<Arc<dyn PersistentLayer>>,
         metrics: &TimelineMetrics,
-    ) {
+    ) -> Result<()> {
         let mut updates = self.layer_map.batch_update();
         for l in compact_to {
-            Self::insert_historic_layer(l.as_ref().clone(), &mut updates, &mut self.layer_fmgr);
-            metrics.record_new_file_metrics(l.layer_desc().file_size);
+            Self::insert_historic_layer(l, &mut updates, &mut self.layer_fmgr);
         }
         for l in compact_from {
-            Self::delete_historic_layer(layer_removal_cs, l, &mut updates, &mut self.layer_fmgr);
+            // NB: the layer file identified by descriptor `l` is guaranteed to be present
+            // in the LayerFileManager because compaction kept holding `layer_removal_cs` the entire
+            // time, even though we dropped `Timeline::layers` inbetween.
+            Self::delete_historic_layer(
+                layer_removal_cs.clone(),
+                l,
+                &mut updates,
+                metrics,
+                &mut self.layer_fmgr,
+            )?;
         }
         updates.flush();
+        Ok(())
     }
 
     /// Called when garbage collect the timeline. Returns a guard that will apply the updates to the layer map.
     pub(crate) fn finish_gc_timeline(
         &mut self,
-        layer_removal_cs: &Arc<tokio::sync::OwnedMutexGuard<()>>,
-        gc_layers: Vec<Layer>,
-    ) {
+        layer_removal_cs: Arc<tokio::sync::OwnedMutexGuard<()>>,
+        gc_layers: Vec<Arc<dyn PersistentLayer>>,
+        metrics: &TimelineMetrics,
+    ) -> Result<ApplyGcResultGuard> {
         let mut updates = self.layer_map.batch_update();
         for doomed_layer in gc_layers {
             Self::delete_historic_layer(
-                layer_removal_cs,
-                &doomed_layer,
+                layer_removal_cs.clone(),
+                doomed_layer,
                 &mut updates,
+                metrics,
                 &mut self.layer_fmgr,
-            );
+            )?; // FIXME: schedule succeeded deletions in timeline.rs `gc_timeline` instead of in batch?
         }
-        updates.flush()
+        Ok(ApplyGcResultGuard(updates))
     }
 
     /// Helper function to insert a layer into the layer map and file manager.
     fn insert_historic_layer(
-        layer: Layer,
+        layer: Arc<dyn PersistentLayer>,
         updates: &mut BatchedUpdates<'_>,
-        mapping: &mut LayerFileManager<Layer>,
+        mapping: &mut LayerFileManager,
     ) {
         updates.insert_historic(layer.layer_desc().clone());
         mapping.insert(layer);
@@ -238,12 +254,17 @@ impl LayerManager {
     /// Remote storage is not affected by this operation.
     fn delete_historic_layer(
         // we cannot remove layers otherwise, since gc and compaction will race
-        _layer_removal_cs: &Arc<tokio::sync::OwnedMutexGuard<()>>,
-        layer: &Layer,
+        _layer_removal_cs: Arc<tokio::sync::OwnedMutexGuard<()>>,
+        layer: Arc<dyn PersistentLayer>,
         updates: &mut BatchedUpdates<'_>,
-        mapping: &mut LayerFileManager<Layer>,
-    ) {
+        metrics: &TimelineMetrics,
+        mapping: &mut LayerFileManager,
+    ) -> anyhow::Result<()> {
         let desc = layer.layer_desc();
+        if !layer.is_remote_layer() {
+            layer.delete_resident_layer_file()?;
+            metrics.resident_physical_size_sub(desc.file_size);
+        }
 
         // TODO Removing from the bottom of the layer map is expensive.
         //      Maybe instead discard all layer map historic versions that
@@ -252,18 +273,21 @@ impl LayerManager {
         //      map index without actually rebuilding the index.
         updates.remove_historic(desc);
         mapping.remove(layer);
-        layer.garbage_collect_on_drop();
+
+        Ok(())
     }
 
-    pub(crate) fn contains(&self, layer: &Layer) -> bool {
+    pub(crate) fn contains(&self, layer: &Arc<dyn PersistentLayer>) -> bool {
         self.layer_fmgr.contains(layer)
     }
 }
 
-pub(crate) struct LayerFileManager<T>(HashMap<PersistentLayerKey, T>);
+pub(crate) struct LayerFileManager<T: AsLayerDesc + ?Sized = dyn PersistentLayer>(
+    HashMap<PersistentLayerKey, Arc<T>>,
+);
 
-impl<T: AsLayerDesc + Clone> LayerFileManager<T> {
-    fn get_from_desc(&self, desc: &PersistentLayerDesc) -> T {
+impl<T: AsLayerDesc + ?Sized> LayerFileManager<T> {
+    fn get_from_desc(&self, desc: &PersistentLayerDesc) -> Arc<T> {
         // The assumption for the `expect()` is that all code maintains the following invariant:
         // A layer's descriptor is present in the LayerMap => the LayerFileManager contains a layer for the descriptor.
         self.0
@@ -273,14 +297,14 @@ impl<T: AsLayerDesc + Clone> LayerFileManager<T> {
             .clone()
     }
 
-    pub(crate) fn insert(&mut self, layer: T) {
+    pub(crate) fn insert(&mut self, layer: Arc<T>) {
         let present = self.0.insert(layer.layer_desc().key(), layer.clone());
         if present.is_some() && cfg!(debug_assertions) {
             panic!("overwriting a layer: {:?}", layer.layer_desc())
         }
     }
 
-    pub(crate) fn contains(&self, layer: &T) -> bool {
+    pub(crate) fn contains(&self, layer: &Arc<T>) -> bool {
         self.0.contains_key(&layer.layer_desc().key())
     }
 
@@ -288,7 +312,7 @@ impl<T: AsLayerDesc + Clone> LayerFileManager<T> {
         Self(HashMap::new())
     }
 
-    pub(crate) fn remove(&mut self, layer: &T) {
+    pub(crate) fn remove(&mut self, layer: Arc<T>) {
         let present = self.0.remove(&layer.layer_desc().key());
         if present.is_none() && cfg!(debug_assertions) {
             panic!(
@@ -297,4 +321,39 @@ impl<T: AsLayerDesc + Clone> LayerFileManager<T> {
             )
         }
     }
+
+    pub(crate) fn replace_and_verify(&mut self, expected: Arc<T>, new: Arc<T>) -> Result<()> {
+        let key = expected.layer_desc().key();
+        let other = new.layer_desc().key();
+
+        let expected_l0 = LayerMap::is_l0(expected.layer_desc());
+        let new_l0 = LayerMap::is_l0(new.layer_desc());
+
+        fail::fail_point!("layermap-replace-notfound", |_| anyhow::bail!(
+            "layermap-replace-notfound"
+        ));
+
+        anyhow::ensure!(
+            key == other,
+            "expected and new layer have different keys: {key:?} != {other:?}"
+        );
+
+        anyhow::ensure!(
+            expected_l0 == new_l0,
+            "one layer is l0 while the other is not: {expected_l0} != {new_l0}"
+        );
+
+        if let Some(layer) = self.0.get_mut(&key) {
+            anyhow::ensure!(
+                compare_arced_layers(&expected, layer),
+                "another layer was found instead of expected, expected={expected:?}, new={new:?}",
+                expected = Arc::as_ptr(&expected),
+                new = Arc::as_ptr(layer),
+            );
+            *layer = new;
+            Ok(())
+        } else {
+            anyhow::bail!("layer was not found");
+        }
+    }
 }

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

@@ -426,7 +426,7 @@ impl ConnectionManagerState {
                     timeline,
                     new_sk.wal_source_connconf,
                     events_sender,
-                    cancellation.clone(),
+                    cancellation,
                     connect_timeout,
                     ctx,
                     node_id,
@@ -447,14 +447,7 @@ impl ConnectionManagerState {
                             }
                             WalReceiverError::Other(e) => {
                                 // give out an error to have task_mgr give it a really verbose logging
-                                if cancellation.is_cancelled() {
-                                    // Ideally we would learn about this via some path other than Other, but
-                                    // that requires refactoring all the intermediate layers of ingest code
-                                    // that only emit anyhow::Error
-                                    Ok(())
-                                } else {
-                                    Err(e).context("walreceiver connection handling failure")
-                                }
+                                Err(e).context("walreceiver connection handling failure")
                             }
                         }
                     }

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

@@ -122,7 +122,7 @@ pub(super) async fn handle_walreceiver_connection(
     // Connect to the database in replication mode.
     info!("connecting to {wal_source_connconf:?}");
 
-    let (replication_client, connection) = {
+    let (mut replication_client, connection) = {
         let mut config = wal_source_connconf.to_tokio_postgres_config();
         config.application_name("pageserver");
         config.replication_mode(tokio_postgres::config::ReplicationMode::Physical);
@@ -205,7 +205,7 @@ pub(super) async fn handle_walreceiver_connection(
         gauge.dec();
     }
 
-    let identify = identify_system(&replication_client).await?;
+    let identify = identify_system(&mut replication_client).await?;
     info!("{identify:?}");
 
     let end_of_wal = Lsn::from(u64::from(identify.xlogpos));
@@ -444,7 +444,7 @@ struct IdentifySystem {
 struct IdentifyError;
 
 /// Run the postgres `IDENTIFY_SYSTEM` command
-async fn identify_system(client: &Client) -> anyhow::Result<IdentifySystem> {
+async fn identify_system(client: &mut Client) -> anyhow::Result<IdentifySystem> {
     let query_str = "IDENTIFY_SYSTEM";
     let response = client.simple_query(query_str).await?;
 
@@ -459,7 +459,7 @@ async fn identify_system(client: &Client) -> anyhow::Result<IdentifySystem> {
 
     // extract the row contents into an IdentifySystem struct.
     // written as a closure so I can use ? for Option here.
-    if let Some(SimpleQueryMessage::Row(first_row)) = response.first() {
+    if let Some(SimpleQueryMessage::Row(first_row)) = response.get(0) {
         Ok(IdentifySystem {
             systemid: get_parse(first_row, 0)?,
             timeline: get_parse(first_row, 1)?,

pageserver/src/tenant/upload_queue.rs

@@ -1,5 +1,4 @@
 use super::storage_layer::LayerFileName;
-use super::storage_layer::ResidentLayer;
 use super::Generation;
 use crate::tenant::metadata::TimelineMetadata;
 use crate::tenant::remote_timeline_client::index::IndexPart;
@@ -80,14 +79,6 @@ pub(crate) struct UploadQueueInitialized {
     /// tasks to finish. For example, metadata upload cannot be performed before all
     /// preceding layer file uploads have completed.
     pub(crate) queued_operations: VecDeque<UploadOp>,
-
-    /// Files which have been unlinked but not yet had scheduled a deletion for. Only kept around
-    /// for error logging.
-    ///
-    /// Putting this behind a testing feature to catch problems in tests, but assuming we could have a
-    /// bug causing leaks, then it's better to not leave this enabled for production builds.
-    #[cfg(feature = "testing")]
-    pub(crate) dangling_files: HashMap<LayerFileName, Generation>,
 }
 
 impl UploadQueueInitialized {
@@ -144,8 +135,6 @@ impl UploadQueue {
             num_inprogress_deletions: 0,
             inprogress_tasks: HashMap::new(),
             queued_operations: VecDeque::new(),
-            #[cfg(feature = "testing")]
-            dangling_files: HashMap::new(),
         };
 
         *self = UploadQueue::Initialized(state);
@@ -191,8 +180,6 @@ impl UploadQueue {
             num_inprogress_deletions: 0,
             inprogress_tasks: HashMap::new(),
             queued_operations: VecDeque::new(),
-            #[cfg(feature = "testing")]
-            dangling_files: HashMap::new(),
         };
 
         *self = UploadQueue::Initialized(state);
@@ -216,6 +203,18 @@ impl UploadQueue {
             UploadQueue::Stopped(stopped) => Ok(stopped),
         }
     }
+
+    pub(crate) fn get_layer_metadata(
+        &self,
+        name: &LayerFileName,
+    ) -> anyhow::Result<Option<LayerFileMetadata>> {
+        match self {
+            UploadQueue::Stopped(_) | UploadQueue::Uninitialized => {
+                anyhow::bail!("queue is in state {}", self.as_str())
+            }
+            UploadQueue::Initialized(inner) => Ok(inner.latest_files.get(name).cloned()),
+        }
+    }
 }
 
 /// An in-progress upload or delete task.
@@ -238,7 +237,7 @@ pub(crate) struct Delete {
 #[derive(Debug)]
 pub(crate) enum UploadOp {
     /// Upload a layer file
-    UploadLayer(ResidentLayer, LayerFileMetadata),
+    UploadLayer(LayerFileName, LayerFileMetadata),
 
     /// Upload the metadata file
     UploadMetadata(IndexPart, Lsn),
@@ -253,13 +252,13 @@ pub(crate) enum UploadOp {
 impl std::fmt::Display for UploadOp {
     fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
         match self {
-            UploadOp::UploadLayer(layer, metadata) => {
+            UploadOp::UploadLayer(path, metadata) => {
                 write!(
                     f,
                     "UploadLayer({}, size={:?}, gen={:?})",
-                    layer,
+                    path.file_name(),
                     metadata.file_size(),
-                    metadata.generation
+                    metadata.generation,
                 )
             }
             UploadOp::UploadMetadata(_, lsn) => {

pageserver/src/virtual_file.rs

@@ -18,8 +18,8 @@ use std::fs::{self, File, OpenOptions};
 use std::io::{Error, ErrorKind, Seek, SeekFrom};
 use std::os::unix::fs::FileExt;
 use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
-use std::sync::{RwLock, RwLockWriteGuard};
-use utils::fs_ext;
+use tokio::sync::{RwLock, RwLockReadGuard, RwLockWriteGuard};
+use tokio::time::Instant;
 
 ///
 /// A virtual file descriptor. You can use this just like std::fs::File, but internally
@@ -111,7 +111,7 @@ impl OpenFiles {
     ///
     /// On return, we hold a lock on the slot, and its 'tag' has been updated
     /// recently_used has been set. It's all ready for reuse.
-    fn find_victim_slot(&self) -> (SlotHandle, RwLockWriteGuard<SlotInner>) {
+    async fn find_victim_slot(&self) -> (SlotHandle, RwLockWriteGuard<SlotInner>) {
         //
         // Run the clock algorithm to find a slot to replace.
         //
@@ -143,7 +143,7 @@ impl OpenFiles {
                 }
                 retries += 1;
             } else {
-                slot_guard = slot.inner.write().unwrap();
+                slot_guard = slot.inner.write().await;
                 index = next;
                 break;
             }
@@ -154,7 +154,7 @@ impl OpenFiles {
         // old file.
         //
         if let Some(old_file) = slot_guard.file.take() {
-            // the normal path of dropping VirtualFile uses "close", use "close-by-replace" here to
+            // the normal path of dropping VirtualFile uses `Close`, use `CloseByReplace` here to
             // distinguish the two.
             STORAGE_IO_TIME_METRIC
                 .get(StorageIoOperation::CloseByReplace)
@@ -174,82 +174,64 @@ impl OpenFiles {
     }
 }
 
-/// Identify error types that should alwways terminate the process.  Other
-/// error types may be elegible for retry.
-pub(crate) fn is_fatal_io_error(e: &std::io::Error) -> bool {
-    use nix::errno::Errno::*;
-    match e.raw_os_error().map(nix::errno::from_i32) {
-        Some(EIO) => {
-            // Terminate on EIO because we no longer trust the device to store
-            // data safely, or to uphold persistence guarantees on fsync.
-            true
-        }
-        Some(EROFS) => {
-            // Terminate on EROFS because a filesystem is usually remounted
-            // readonly when it has experienced some critical issue, so the same
-            // logic as EIO applies.
-            true
-        }
-        Some(EACCES) => {
-            // Terminate on EACCESS because we should always have permissions
-            // for our own data dir: if we don't, then we can't do our job and
-            // need administrative intervention to fix permissions.  Terminating
-            // is the best way to make sure we stop cleanly rather than going
-            // into infinite retry loops, and will make it clear to the outside
-            // world that we need help.
-            true
-        }
-        _ => {
-            // Treat all other local file I/O errors are retryable.  This includes:
-            // - ENOSPC: we stay up and wait for eviction to free some space
-            // - EINVAL, EBADF, EBADFD: this is a code bug, not a filesystem/hardware issue
-            // - WriteZero, Interrupted: these are used internally VirtualFile
-            false
-        }
-    }
+#[derive(Debug, thiserror::Error)]
+pub enum CrashsafeOverwriteError {
+    #[error("final path has no parent dir")]
+    FinalPathHasNoParentDir,
+    #[error("remove tempfile")]
+    RemovePreviousTempfile(#[source] std::io::Error),
+    #[error("create tempfile")]
+    CreateTempfile(#[source] std::io::Error),
+    #[error("write tempfile")]
+    WriteContents(#[source] std::io::Error),
+    #[error("sync tempfile")]
+    SyncTempfile(#[source] std::io::Error),
+    #[error("rename tempfile to final path")]
+    RenameTempfileToFinalPath(#[source] std::io::Error),
+    #[error("open final path parent dir")]
+    OpenFinalPathParentDir(#[source] std::io::Error),
+    #[error("sync final path parent dir")]
+    SyncFinalPathParentDir(#[source] std::io::Error),
 }
-
-/// Call this when the local filesystem gives us an error with an external
-/// cause: this includes EIO, EROFS, and EACCESS: all these indicate either
-/// bad storage or bad configuration, and we can't fix that from inside
-/// a running process.
-pub(crate) fn on_fatal_io_error(e: &std::io::Error, context: &str) -> ! {
-    tracing::error!("Fatal I/O error: {e}: {context})");
-    std::process::abort();
-}
-
-pub(crate) trait MaybeFatalIo<T> {
-    fn maybe_fatal_err(self, context: &str) -> std::io::Result<T>;
-    fn fatal_err(self, context: &str) -> T;
-}
-
-impl<T> MaybeFatalIo<T> for std::io::Result<T> {
-    /// Terminate the process if the result is an error of a fatal type, else pass it through
-    ///
-    /// This is appropriate for writes, where we typically want to die on EIO/ACCES etc, but
-    /// not on ENOSPC.
-    fn maybe_fatal_err(self, context: &str) -> std::io::Result<T> {
-        if let Err(e) = &self {
-            if is_fatal_io_error(e) {
-                on_fatal_io_error(e, context);
-            }
-        }
-        self
-    }
-
-    /// Terminate the process on any I/O error.
-    ///
-    /// This is appropriate for reads on files that we know exist: they should always work.
-    fn fatal_err(self, context: &str) -> T {
+impl CrashsafeOverwriteError {
+    /// Returns true iff the new contents are durably stored.
+    pub fn are_new_contents_durable(&self) -> bool {
         match self {
-            Ok(v) => v,
-            Err(e) => {
-                on_fatal_io_error(&e, context);
-            }
+            Self::FinalPathHasNoParentDir => false,
+            Self::RemovePreviousTempfile(_) => false,
+            Self::CreateTempfile(_) => false,
+            Self::WriteContents(_) => false,
+            Self::SyncTempfile(_) => false,
+            Self::RenameTempfileToFinalPath(_) => false,
+            Self::OpenFinalPathParentDir(_) => false,
+            Self::SyncFinalPathParentDir(_) => true,
         }
     }
 }
 
+/// Observe duration for the given storage I/O operation
+///
+/// Unlike `observe_closure_duration`, this supports async,
+/// where "support" means that we measure wall clock time.
+macro_rules! observe_duration {
+    ($op:expr, $($body:tt)*) => {{
+        let instant = Instant::now();
+        let result = $($body)*;
+        let elapsed = instant.elapsed().as_secs_f64();
+        STORAGE_IO_TIME_METRIC
+            .get($op)
+            .observe(elapsed);
+        result
+    }}
+}
+
+macro_rules! with_file {
+    ($this:expr, $op:expr, | $ident:ident | $($body:tt)*) => {{
+        let $ident = $this.lock_file().await?;
+        observe_duration!($op, $($body)*)
+    }};
+}
+
 impl VirtualFile {
     /// Open a file in read-only mode. Like File::open.
     pub async fn open(path: &Utf8Path) -> Result<VirtualFile, std::io::Error> {
@@ -286,11 +268,9 @@ impl VirtualFile {
             tenant_id = "*".to_string();
             timeline_id = "*".to_string();
         }
-        let (handle, mut slot_guard) = get_open_files().find_victim_slot();
+        let (handle, mut slot_guard) = get_open_files().find_victim_slot().await;
 
-        let file = STORAGE_IO_TIME_METRIC
-            .get(StorageIoOperation::Open)
-            .observe_closure_duration(|| open_options.open(path))?;
+        let file = observe_duration!(StorageIoOperation::Open, open_options.open(path))?;
 
         // Strip all options other than read and write.
         //
@@ -326,13 +306,15 @@ impl VirtualFile {
         final_path: &Utf8Path,
         tmp_path: &Utf8Path,
         content: &[u8],
-    ) -> std::io::Result<()> {
+    ) -> Result<(), CrashsafeOverwriteError> {
         let Some(final_path_parent) = final_path.parent() else {
-            return Err(std::io::Error::from_raw_os_error(
-                nix::errno::Errno::EINVAL as i32,
-            ));
+            return Err(CrashsafeOverwriteError::FinalPathHasNoParentDir);
         };
-        std::fs::remove_file(tmp_path).or_else(fs_ext::ignore_not_found)?;
+        match std::fs::remove_file(tmp_path) {
+            Ok(()) => {}
+            Err(e) if e.kind() == std::io::ErrorKind::NotFound => {}
+            Err(e) => return Err(CrashsafeOverwriteError::RemovePreviousTempfile(e)),
+        }
         let mut file = Self::open_with_options(
             tmp_path,
             OpenOptions::new()
@@ -341,41 +323,54 @@ impl VirtualFile {
                 // we bail out instead of causing damage.
                 .create_new(true),
         )
-        .await?;
-        file.write_all(content).await?;
-        file.sync_all().await?;
+        .await
+        .map_err(CrashsafeOverwriteError::CreateTempfile)?;
+        file.write_all(content)
+            .await
+            .map_err(CrashsafeOverwriteError::WriteContents)?;
+        file.sync_all()
+            .await
+            .map_err(CrashsafeOverwriteError::SyncTempfile)?;
         drop(file); // before the rename, that's important!
                     // renames are atomic
-        std::fs::rename(tmp_path, final_path)?;
+        std::fs::rename(tmp_path, final_path)
+            .map_err(CrashsafeOverwriteError::RenameTempfileToFinalPath)?;
         // Only open final path parent dirfd now, so that this operation only
         // ever holds one VirtualFile fd at a time.  That's important because
         // the current `find_victim_slot` impl might pick the same slot for both
         // VirtualFile., and it eventually does a blocking write lock instead of
         // try_lock.
         let final_parent_dirfd =
-            Self::open_with_options(final_path_parent, OpenOptions::new().read(true)).await?;
-        final_parent_dirfd.sync_all().await?;
+            Self::open_with_options(final_path_parent, OpenOptions::new().read(true))
+                .await
+                .map_err(CrashsafeOverwriteError::OpenFinalPathParentDir)?;
+        final_parent_dirfd
+            .sync_all()
+            .await
+            .map_err(CrashsafeOverwriteError::SyncFinalPathParentDir)?;
         Ok(())
     }
 
     /// Call File::sync_all() on the underlying File.
     pub async fn sync_all(&self) -> Result<(), Error> {
-        self.with_file(StorageIoOperation::Fsync, |file| file.sync_all())
-            .await?
+        with_file!(self, StorageIoOperation::Fsync, |file| file
+            .as_ref()
+            .sync_all())
     }
 
     pub async fn metadata(&self) -> Result<fs::Metadata, Error> {
-        self.with_file(StorageIoOperation::Metadata, |file| file.metadata())
-            .await?
+        with_file!(self, StorageIoOperation::Metadata, |file| file
+            .as_ref()
+            .metadata())
     }
 
-    /// Helper function that looks up the underlying File for this VirtualFile,
-    /// opening it and evicting some other File if necessary. It calls 'func'
-    /// with the physical File.
-    async fn with_file<F, R>(&self, op: StorageIoOperation, mut func: F) -> Result<R, Error>
-    where
-        F: FnMut(&File) -> R,
-    {
+    /// Helper function internal to `VirtualFile` that looks up the underlying File,
+    /// opens it and evicts some other File if necessary. The passed parameter is
+    /// assumed to be a function available for the physical `File`.
+    ///
+    /// We are doing it via a macro as Rust doesn't support async closures that
+    /// take on parameters with lifetimes.
+    async fn lock_file(&self) -> Result<FileGuard<'_>, Error> {
         let open_files = get_open_files();
 
         let mut handle_guard = {
@@ -385,27 +380,23 @@ impl VirtualFile {
             // We only need to hold the handle lock while we read the current handle. If
             // another thread closes the file and recycles the slot for a different file,
             // we will notice that the handle we read is no longer valid and retry.
-            let mut handle = *self.handle.read().unwrap();
+            let mut handle = *self.handle.read().await;
             loop {
                 // Check if the slot contains our File
                 {
                     let slot = &open_files.slots[handle.index];
-                    let slot_guard = slot.inner.read().unwrap();
-                    if slot_guard.tag == handle.tag {
-                        if let Some(file) = &slot_guard.file {
-                            // Found a cached file descriptor.
-                            slot.recently_used.store(true, Ordering::Relaxed);
-                            return Ok(STORAGE_IO_TIME_METRIC
-                                .get(op)
-                                .observe_closure_duration(|| func(file)));
-                        }
+                    let slot_guard = slot.inner.read().await;
+                    if slot_guard.tag == handle.tag && slot_guard.file.is_some() {
+                        // Found a cached file descriptor.
+                        slot.recently_used.store(true, Ordering::Relaxed);
+                        return Ok(FileGuard { slot_guard });
                     }
                 }
 
                 // The slot didn't contain our File. We will have to open it ourselves,
                 // but before that, grab a write lock on handle in the VirtualFile, so
                 // that no other thread will try to concurrently open the same file.
-                let handle_guard = self.handle.write().unwrap();
+                let handle_guard = self.handle.write().await;
 
                 // If another thread changed the handle while we were not holding the lock,
                 // then the handle might now be valid again. Loop back to retry.
@@ -419,17 +410,10 @@ impl VirtualFile {
 
         // We need to open the file ourselves. The handle in the VirtualFile is
         // now locked in write-mode. Find a free slot to put it in.
-        let (handle, mut slot_guard) = open_files.find_victim_slot();
+        let (handle, mut slot_guard) = open_files.find_victim_slot().await;
 
         // Open the physical file
-        let file = STORAGE_IO_TIME_METRIC
-            .get(StorageIoOperation::Open)
-            .observe_closure_duration(|| self.open_options.open(&self.path))?;
-
-        // Perform the requested operation on it
-        let result = STORAGE_IO_TIME_METRIC
-            .get(op)
-            .observe_closure_duration(|| func(&file));
+        let file = observe_duration!(StorageIoOperation::Open, self.open_options.open(&self.path))?;
 
         // Store the File in the slot and update the handle in the VirtualFile
         // to point to it.
@@ -437,7 +421,9 @@ impl VirtualFile {
 
         *handle_guard = handle;
 
-        Ok(result)
+        return Ok(FileGuard {
+            slot_guard: slot_guard.downgrade(),
+        });
     }
 
     pub fn remove(self) {
@@ -452,11 +438,9 @@ impl VirtualFile {
                 self.pos = offset;
             }
             SeekFrom::End(offset) => {
-                self.pos = self
-                    .with_file(StorageIoOperation::Seek, |mut file| {
-                        file.seek(SeekFrom::End(offset))
-                    })
-                    .await??
+                self.pos = with_file!(self, StorageIoOperation::Seek, |file| file
+                    .as_ref()
+                    .seek(SeekFrom::End(offset)))?
             }
             SeekFrom::Current(offset) => {
                 let pos = self.pos as i128 + offset as i128;
@@ -544,9 +528,9 @@ impl VirtualFile {
     }
 
     pub async fn read_at(&self, buf: &mut [u8], offset: u64) -> Result<usize, Error> {
-        let result = self
-            .with_file(StorageIoOperation::Read, |file| file.read_at(buf, offset))
-            .await?;
+        let result = with_file!(self, StorageIoOperation::Read, |file| file
+            .as_ref()
+            .read_at(buf, offset));
         if let Ok(size) = result {
             STORAGE_IO_SIZE
                 .with_label_values(&["read", &self.tenant_id, &self.timeline_id])
@@ -556,9 +540,9 @@ impl VirtualFile {
     }
 
     async fn write_at(&self, buf: &[u8], offset: u64) -> Result<usize, Error> {
-        let result = self
-            .with_file(StorageIoOperation::Write, |file| file.write_at(buf, offset))
-            .await?;
+        let result = with_file!(self, StorageIoOperation::Write, |file| file
+            .as_ref()
+            .write_at(buf, offset));
         if let Ok(size) = result {
             STORAGE_IO_SIZE
                 .with_label_values(&["write", &self.tenant_id, &self.timeline_id])
@@ -568,6 +552,18 @@ impl VirtualFile {
     }
 }
 
+struct FileGuard<'a> {
+    slot_guard: RwLockReadGuard<'a, SlotInner>,
+}
+
+impl<'a> AsRef<File> for FileGuard<'a> {
+    fn as_ref(&self) -> &File {
+        // This unwrap is safe because we only create `FileGuard`s
+        // if we know that the file is Some.
+        self.slot_guard.file.as_ref().unwrap()
+    }
+}
+
 #[cfg(test)]
 impl VirtualFile {
     pub(crate) async fn read_blk(
@@ -600,20 +596,39 @@ impl VirtualFile {
 impl Drop for VirtualFile {
     /// If a VirtualFile is dropped, close the underlying file if it was open.
     fn drop(&mut self) {
-        let handle = self.handle.get_mut().unwrap();
+        let handle = self.handle.get_mut();
 
-        // We could check with a read-lock first, to avoid waiting on an
-        // unrelated I/O.
-        let slot = &get_open_files().slots[handle.index];
-        let mut slot_guard = slot.inner.write().unwrap();
-        if slot_guard.tag == handle.tag {
-            slot.recently_used.store(false, Ordering::Relaxed);
-            // there is also operation "close-by-replace" for closes done on eviction for
-            // comparison.
-            STORAGE_IO_TIME_METRIC
-                .get(StorageIoOperation::Close)
-                .observe_closure_duration(|| drop(slot_guard.file.take()));
+        fn clean_slot(slot: &Slot, mut slot_guard: RwLockWriteGuard<'_, SlotInner>, tag: u64) {
+            if slot_guard.tag == tag {
+                slot.recently_used.store(false, Ordering::Relaxed);
+                // there is also the `CloseByReplace` operation for closes done on eviction for
+                // comparison.
+                STORAGE_IO_TIME_METRIC
+                    .get(StorageIoOperation::Close)
+                    .observe_closure_duration(|| drop(slot_guard.file.take()));
+            }
         }
+
+        // We don't have async drop so we cannot directly await the lock here.
+        // Instead, first do a best-effort attempt at closing the underlying
+        // file descriptor by using `try_write`, and if that fails, spawn
+        // a tokio task to do it asynchronously: we just want it to be
+        // cleaned up eventually.
+        // Most of the time, the `try_lock` should succeed though,
+        // as we have `&mut self` access. In other words, if the slot
+        // is still occupied by our file, there should be no access from
+        // other I/O operations; the only other possible place to lock
+        // the slot is the lock algorithm looking for free slots.
+        let slot = &get_open_files().slots[handle.index];
+        if let Ok(slot_guard) = slot.inner.try_write() {
+            clean_slot(slot, slot_guard, handle.tag);
+        } else {
+            let tag = handle.tag;
+            tokio::spawn(async move {
+                let slot_guard = slot.inner.write().await;
+                clean_slot(slot, slot_guard, tag);
+            });
+        };
     }
 }
 

pageserver/src/walingest.rs

@@ -443,7 +443,7 @@ impl<'a> WalIngest<'a> {
         &mut self,
         buf: &mut Bytes,
         modification: &mut DatadirModification<'_>,
-        decoded: &DecodedWALRecord,
+        decoded: &mut DecodedWALRecord,
         ctx: &RequestContext,
     ) -> anyhow::Result<()> {
         // Handle VM bit updates that are implicitly part of heap records.
@@ -749,7 +749,7 @@ impl<'a> WalIngest<'a> {
         &mut self,
         buf: &mut Bytes,
         modification: &mut DatadirModification<'_>,
-        decoded: &DecodedWALRecord,
+        decoded: &mut DecodedWALRecord,
         ctx: &RequestContext,
     ) -> anyhow::Result<()> {
         // Handle VM bit updates that are implicitly part of heap records.

pageserver/src/walredo.rs

@@ -857,8 +857,7 @@ impl WalRedoProcess {
             let in_revents = stdin_pollfds[0].revents().unwrap();
             if in_revents & (PollFlags::POLLERR | PollFlags::POLLOUT) != PollFlags::empty() {
                 nwrite += proc.stdin.write(&writebuf[nwrite..])?;
-            }
-            if in_revents.contains(PollFlags::POLLHUP) {
+            } else if in_revents.contains(PollFlags::POLLHUP) {
                 // We still have more data to write, but the process closed the pipe.
                 anyhow::bail!("WAL redo process closed its stdin unexpectedly");
             }
@@ -908,8 +907,7 @@ impl WalRedoProcess {
                 let out_revents = stdout_pollfds[0].revents().unwrap();
                 if out_revents & (PollFlags::POLLERR | PollFlags::POLLIN) != PollFlags::empty() {
                     nresult += output.stdout.read(&mut resultbuf[nresult..])?;
-                }
-                if out_revents.contains(PollFlags::POLLHUP) {
+                } else if out_revents.contains(PollFlags::POLLHUP) {
                     anyhow::bail!("WAL redo process closed its stdout unexpectedly");
                 }
             }

pgxn/neon/libpagestore.c

@@ -19,7 +19,6 @@
 #include "access/xlog.h"
 #include "access/xlogutils.h"
 #include "storage/buf_internals.h"
-#include "c.h"
 
 #include "libpq-fe.h"
 #include "libpq/pqformat.h"
@@ -64,21 +63,6 @@ int			max_reconnect_attempts = 60;
 bool	(*old_redo_read_buffer_filter) (XLogReaderState *record, uint8 block_id) = NULL;
 
 static bool pageserver_flush(void);
-static void pageserver_disconnect(void);
-
-
-static pqsigfunc	 prev_signal_handler;
-
-static void
-pageserver_sighup_handler(SIGNAL_ARGS)
-{
-	if (prev_signal_handler)
-	{
-        	prev_signal_handler(postgres_signal_arg);
-	}
-	neon_log(LOG, "Received SIGHUP, disconnecting pageserver. New pageserver connstring is %s", page_server_connstring);
-	pageserver_disconnect();
-}
 
 static bool
 pageserver_connect(int elevel)
@@ -416,7 +400,7 @@ pg_init_libpagestore(void)
 							   NULL,
 							   &page_server_connstring,
 							   "",
-							   PGC_SIGHUP,
+							   PGC_POSTMASTER,
 							   0,	/* no flags required */
 							   NULL, NULL, NULL);
 
@@ -498,8 +482,5 @@ pg_init_libpagestore(void)
 		old_redo_read_buffer_filter = redo_read_buffer_filter;
 		redo_read_buffer_filter = neon_redo_read_buffer_filter;
 	}
-
-        prev_signal_handler = pqsignal(SIGHUP, pageserver_sighup_handler);
-
 	lfc_init();
 }

pgxn/neon/walproposer_pg.c

@@ -88,7 +88,7 @@ static void StartProposerReplication(WalProposer *wp, StartReplicationCmd *cmd);
 static void WalSndLoop(WalProposer *wp);
 static void XLogBroadcastWalProposer(WalProposer *wp);
 
-static void XLogWalPropWrite(WalProposer *wp, char *buf, Size nbytes, XLogRecPtr recptr);
+static void XLogWalPropWrite(char *buf, Size nbytes, XLogRecPtr recptr);
 static void XLogWalPropClose(XLogRecPtr recptr);
 
 static void
@@ -1241,7 +1241,7 @@ WalProposerRecovery(Safekeeper *sk, TimeLineID timeline, XLogRecPtr startpos, XL
 				rec_end_lsn = rec_start_lsn + len - XLOG_HDR_SIZE;
 
 				/* write WAL to disk */
-				XLogWalPropWrite(sk->wp, &buf[XLOG_HDR_SIZE], len - XLOG_HDR_SIZE, rec_start_lsn);
+				XLogWalPropWrite(&buf[XLOG_HDR_SIZE], len - XLOG_HDR_SIZE, rec_start_lsn);
 
 				ereport(DEBUG1,
 						(errmsg("Recover message %X/%X length %d",
@@ -1283,24 +1283,11 @@ static XLogSegNo walpropSegNo = 0;
  * Write XLOG data to disk.
  */
 static void
-XLogWalPropWrite(WalProposer *wp, char *buf, Size nbytes, XLogRecPtr recptr)
+XLogWalPropWrite(char *buf, Size nbytes, XLogRecPtr recptr)
 {
 	int			startoff;
 	int			byteswritten;
 
-	/*
-	 * Apart from walproposer, basebackup LSN page is also written out by
-	 * postgres itself which writes WAL only in pages, and in basebackup it is
-	 * inherently dummy (only safekeepers have historic WAL). Update WAL buffers
-	 * here to avoid dummy page overwriting correct one we download here. Ugly,
-	 * but alternatives are about the same ugly. We won't need that if we switch
-	 * to on-demand WAL download from safekeepers, without writing to disk.
-	 *
-	 * https://github.com/neondatabase/neon/issues/5749
-	 */
-	if (!wp->config->syncSafekeepers)
-		XLogUpdateWalBuffers(buf, recptr, nbytes);
-
 	while (nbytes > 0)
 	{
 		int			segbytes;

pgxn/neon_walredo/walredoproc.c

@@ -201,16 +201,6 @@ WalRedoMain(int argc, char *argv[])
 #endif
 
 	am_wal_redo_postgres = true;
-	/*
-	 * Pageserver treats any output to stderr as an ERROR, so we must
-	 * set the log level as early as possible to only log FATAL and 
-	 * above during WAL redo (note that loglevel ERROR also logs LOG,
-	 * which is super strange but that's not something we can solve
-	 * for here. ¯\_(-_-)_/¯
-	 */
-	SetConfigOption("log_min_messages", "FATAL", PGC_SUSET, PGC_S_OVERRIDE);
-	SetConfigOption("client_min_messages", "ERROR", PGC_SUSET,
-					PGC_S_OVERRIDE);
 
 	/*
 	 * WAL redo does not need a large number of buffers. And speed of
@@ -895,12 +885,7 @@ apply_error_callback(void *arg)
 	StringInfoData buf;
 
 	initStringInfo(&buf);
-#if PG_VERSION_NUM >= 150000
-	if (record->record)
-#else
-	if (record->decoded_record)
-#endif
-		xlog_outdesc(&buf, record);
+	xlog_outdesc(&buf, record);
 
 	/* translator: %s is a WAL record description */
 	errcontext("WAL redo at %X/%X for %s",

proxy/src/auth/backend.rs

@@ -5,7 +5,7 @@ mod link;
 pub use link::LinkAuthError;
 use tokio_postgres::config::AuthKeys;
 
-use crate::proxy::{handle_try_wake, retry_after, LatencyTimer};
+use crate::proxy::{handle_try_wake, retry_after};
 use crate::{
     auth::{self, ClientCredentials},
     config::AuthenticationConfig,
@@ -134,14 +134,13 @@ async fn auth_quirks_creds(
     client: &mut stream::PqStream<impl AsyncRead + AsyncWrite + Unpin>,
     allow_cleartext: bool,
     config: &'static AuthenticationConfig,
-    latency_timer: &mut LatencyTimer,
 ) -> auth::Result<AuthSuccess<ComputeCredentials>> {
     // If there's no project so far, that entails that client doesn't
     // support SNI or other means of passing the endpoint (project) name.
     // We now expect to see a very specific payload in the place of password.
     if creds.project.is_none() {
         // Password will be checked by the compute node later.
-        return hacks::password_hack(creds, client, latency_timer).await;
+        return hacks::password_hack(creds, client).await;
     }
 
     // Password hack should set the project name.
@@ -152,11 +151,11 @@ async fn auth_quirks_creds(
     // Currently, we use it for websocket connections (latency).
     if allow_cleartext {
         // Password will be checked by the compute node later.
-        return hacks::cleartext_hack(client, latency_timer).await;
+        return hacks::cleartext_hack(client).await;
     }
 
     // Finally, proceed with the main auth flow (SCRAM-based).
-    classic::authenticate(api, extra, creds, client, config, latency_timer).await
+    classic::authenticate(api, extra, creds, client, config).await
 }
 
 /// True to its name, this function encapsulates our current auth trade-offs.
@@ -168,18 +167,8 @@ async fn auth_quirks(
     client: &mut stream::PqStream<impl AsyncRead + AsyncWrite + Unpin>,
     allow_cleartext: bool,
     config: &'static AuthenticationConfig,
-    latency_timer: &mut LatencyTimer,
 ) -> auth::Result<AuthSuccess<CachedNodeInfo>> {
-    let auth_stuff = auth_quirks_creds(
-        api,
-        extra,
-        creds,
-        client,
-        allow_cleartext,
-        config,
-        latency_timer,
-    )
-    .await?;
+    let auth_stuff = auth_quirks_creds(api, extra, creds, client, allow_cleartext, config).await?;
 
     let mut num_retries = 0;
     let mut node = loop {
@@ -244,7 +233,6 @@ impl BackendType<'_, ClientCredentials<'_>> {
         client: &mut stream::PqStream<impl AsyncRead + AsyncWrite + Unpin>,
         allow_cleartext: bool,
         config: &'static AuthenticationConfig,
-        latency_timer: &mut LatencyTimer,
     ) -> auth::Result<AuthSuccess<CachedNodeInfo>> {
         use BackendType::*;
 
@@ -257,16 +245,7 @@ impl BackendType<'_, ClientCredentials<'_>> {
                 );
 
                 let api = api.as_ref();
-                auth_quirks(
-                    api,
-                    extra,
-                    creds,
-                    client,
-                    allow_cleartext,
-                    config,
-                    latency_timer,
-                )
-                .await?
+                auth_quirks(api, extra, creds, client, allow_cleartext, config).await?
             }
             Postgres(api, creds) => {
                 info!(
@@ -276,16 +255,7 @@ impl BackendType<'_, ClientCredentials<'_>> {
                 );
 
                 let api = api.as_ref();
-                auth_quirks(
-                    api,
-                    extra,
-                    creds,
-                    client,
-                    allow_cleartext,
-                    config,
-                    latency_timer,
-                )
-                .await?
+                auth_quirks(api, extra, creds, client, allow_cleartext, config).await?
             }
             // NOTE: this auth backend doesn't use client credentials.
             Link(url) => {

proxy/src/auth/backend/classic.rs

@@ -4,7 +4,6 @@ use crate::{
     compute,
     config::AuthenticationConfig,
     console::{self, AuthInfo, ConsoleReqExtra},
-    proxy::LatencyTimer,
     sasl, scram,
     stream::PqStream,
 };
@@ -17,7 +16,6 @@ pub(super) async fn authenticate(
     creds: &ClientCredentials<'_>,
     client: &mut PqStream<impl AsyncRead + AsyncWrite + Unpin>,
     config: &'static AuthenticationConfig,
-    latency_timer: &mut LatencyTimer,
 ) -> auth::Result<AuthSuccess<ComputeCredentials>> {
     info!("fetching user's authentication info");
     let info = api.get_auth_info(extra, creds).await?.unwrap_or_else(|| {
@@ -38,26 +36,24 @@ pub(super) async fn authenticate(
             info!("auth endpoint chooses SCRAM");
             let scram = auth::Scram(&secret);
 
+            let auth_flow = flow.begin(scram).await.map_err(|error| {
+                warn!(?error, "error sending scram acknowledgement");
+                error
+            })?;
+
             let auth_outcome = tokio::time::timeout(
                 config.scram_protocol_timeout,
-                async {
-                    // pause the timer while we communicate with the client
-                    let _paused = latency_timer.pause();
-
-                    flow.begin(scram).await.map_err(|error| {
-                        warn!(?error, "error sending scram acknowledgement");
-                        error
-                    })?.authenticate().await.map_err(|error| {
-                        warn!(?error, "error processing scram messages");
-                        error
-                    })
-                }
+                auth_flow.authenticate(),
             )
             .await
             .map_err(|error| {
                 warn!("error processing scram messages error = authentication timed out, execution time exeeded {} seconds", config.scram_protocol_timeout.as_secs());
                 auth::io::Error::new(auth::io::ErrorKind::TimedOut, error)
-            })??;
+            })?
+            .map_err(|error| {
+                warn!(?error, "error processing scram messages");
+                error
+            })?;
 
             let client_key = match auth_outcome {
                 sasl::Outcome::Success(key) => key,

proxy/src/auth/backend/hacks.rs

@@ -1,7 +1,6 @@
 use super::{AuthSuccess, ComputeCredentials};
 use crate::{
     auth::{self, AuthFlow, ClientCredentials},
-    proxy::LatencyTimer,
     stream,
 };
 use tokio::io::{AsyncRead, AsyncWrite};
@@ -13,13 +12,8 @@ use tracing::{info, warn};
 /// use this mechanism for websocket connections.
 pub async fn cleartext_hack(
     client: &mut stream::PqStream<impl AsyncRead + AsyncWrite + Unpin>,
-    latency_timer: &mut LatencyTimer,
 ) -> auth::Result<AuthSuccess<ComputeCredentials>> {
     warn!("cleartext auth flow override is enabled, proceeding");
-
-    // pause the timer while we communicate with the client
-    let _paused = latency_timer.pause();
-
     let password = AuthFlow::new(client)
         .begin(auth::CleartextPassword)
         .await?
@@ -38,13 +32,8 @@ pub async fn cleartext_hack(
 pub async fn password_hack(
     creds: &mut ClientCredentials<'_>,
     client: &mut stream::PqStream<impl AsyncRead + AsyncWrite + Unpin>,
-    latency_timer: &mut LatencyTimer,
 ) -> auth::Result<AuthSuccess<ComputeCredentials>> {
     warn!("project not specified, resorting to the password hack auth flow");
-
-    // pause the timer while we communicate with the client
-    let _paused = latency_timer.pause();
-
     let payload = AuthFlow::new(client)
         .begin(auth::PasswordHack)
         .await?

proxy/src/bin/proxy.rs

@@ -16,10 +16,9 @@ use tokio::task::JoinSet;
 use tokio_util::sync::CancellationToken;
 use tracing::info;
 use tracing::warn;
-use utils::{project_build_tag, project_git_version, sentry_init::init_sentry};
+use utils::{project_git_version, sentry_init::init_sentry};
 
 project_git_version!(GIT_VERSION);
-project_build_tag!(BUILD_TAG);
 
 use clap::{Parser, ValueEnum};
 
@@ -101,8 +100,7 @@ async fn main() -> anyhow::Result<()> {
     let _sentry_guard = init_sentry(Some(GIT_VERSION.into()), &[]);
 
     info!("Version: {GIT_VERSION}");
-    info!("Build_tag: {BUILD_TAG}");
-    ::metrics::set_build_info_metric(GIT_VERSION, BUILD_TAG);
+    ::metrics::set_build_info_metric(GIT_VERSION);
 
     let args = ProxyCliArgs::parse();
     let config = build_config(&args)?;