Fix test

pageserver/src/bin/pageserver.rs

@@ -263,6 +263,8 @@ fn start_pageserver(conf: &'static PageServerConf, daemonize: bool) -> Result<()
     // start profiler (if enabled)
     let profiler_guard = profiling::init_profiler(conf);
 
+    pageserver::tenant_tasks::init_tenant_task_pool()?;
+
     // initialize authentication for incoming connections
     let auth = match &conf.auth_type {
         AuthType::Trust | AuthType::MD5 => None,

pageserver/src/layered_repository.rs

@@ -158,6 +158,13 @@ pub struct LayeredRepository {
     // Global pageserver config parameters
     pub conf: &'static PageServerConf,
 
+    // Allows us to gracefully cancel operations that edit the directory
+    // that backs this layered repository. Usage:
+    //
+    // Use `let _guard = file_lock.try_read()` while writing any files.
+    // Use `let _guard = file_lock.write().unwrap()` to wait for all writes to finish.
+    pub file_lock: RwLock<()>,
+
     // Overridden tenant-specific config parameters.
     // We keep TenantConfOpt sturct here to preserve the information
     // about parameters that are not set.
@@ -685,6 +692,7 @@ impl LayeredRepository {
     ) -> LayeredRepository {
         LayeredRepository {
             tenant_id,
+            file_lock: RwLock::new(()),
             conf,
             tenant_conf: Arc::new(RwLock::new(tenant_conf)),
             timelines: Mutex::new(HashMap::new()),

pageserver/src/lib.rs

@@ -13,7 +13,7 @@ pub mod repository;
 pub mod storage_sync;
 pub mod tenant_config;
 pub mod tenant_mgr;
-pub mod tenant_threads;
+pub mod tenant_tasks;
 pub mod thread_mgr;
 pub mod timelines;
 pub mod virtual_file;

pageserver/src/tenant_mgr.rs

@@ -230,8 +230,6 @@ pub fn shutdown_all_tenants() {
     drop(m);
 
     thread_mgr::shutdown_threads(Some(ThreadKind::WalReceiverManager), None, None);
-    thread_mgr::shutdown_threads(Some(ThreadKind::GarbageCollector), None, None);
-    thread_mgr::shutdown_threads(Some(ThreadKind::Compactor), None, None);
 
     // Ok, no background threads running anymore. Flush any remaining data in
     // memory to disk.
@@ -330,44 +328,12 @@ pub fn set_tenant_state(tenant_id: ZTenantId, new_state: TenantState) -> anyhow:
         }
         (TenantState::Idle, TenantState::Active) => {
             info!("activating tenant {tenant_id}");
-            let compactor_spawn_result = thread_mgr::spawn(
-                ThreadKind::Compactor,
-                Some(tenant_id),
-                None,
-                "Compactor thread",
-                false,
-                move || crate::tenant_threads::compact_loop(tenant_id),
-            );
-            if compactor_spawn_result.is_err() {
-                let mut m = tenants_state::write_tenants();
-                m.get_mut(&tenant_id)
-                    .with_context(|| format!("Tenant not found for id {tenant_id}"))?
-                    .state = old_state;
-                drop(m);
-            }
-            compactor_spawn_result?;
 
-            let gc_spawn_result = thread_mgr::spawn(
-                ThreadKind::GarbageCollector,
-                Some(tenant_id),
-                None,
-                "GC thread",
-                false,
-                move || crate::tenant_threads::gc_loop(tenant_id),
-            )
-            .map(|_thread_id| ()) // update the `Result::Ok` type to match the outer function's return signature
-            .with_context(|| format!("Failed to launch GC thread for tenant {tenant_id}"));
-
-            if let Err(e) = &gc_spawn_result {
-                let mut m = tenants_state::write_tenants();
-                m.get_mut(&tenant_id)
-                    .with_context(|| format!("Tenant not found for id {tenant_id}"))?
-                    .state = old_state;
-                drop(m);
-                error!("Failed to start GC thread for tenant {tenant_id}, stopping its checkpointer thread: {e:?}");
-                thread_mgr::shutdown_threads(Some(ThreadKind::Compactor), Some(tenant_id), None);
-                return gc_spawn_result;
-            }
+            // Spawn gc and compaction loops. The loops will shut themselves
+            // down when they notice that the tenant is inactive.
+            // TODO maybe use tokio::sync::watch instead?
+            crate::tenant_tasks::start_compaction_loop(tenant_id)?;
+            crate::tenant_tasks::start_gc_loop(tenant_id)?;
         }
         (TenantState::Idle, TenantState::Stopping) => {
             info!("stopping idle tenant {tenant_id}");
@@ -379,8 +345,11 @@ pub fn set_tenant_state(tenant_id: ZTenantId, new_state: TenantState) -> anyhow:
                 Some(tenant_id),
                 None,
             );
-            thread_mgr::shutdown_threads(Some(ThreadKind::GarbageCollector), Some(tenant_id), None);
-            thread_mgr::shutdown_threads(Some(ThreadKind::Compactor), Some(tenant_id), None);
+
+            // Wait until all gc/compaction tasks finish
+            // TODO send cancellation signal too, or make the state a watch
+            let repo = get_repository_for_tenant(tenant_id)?;
+            let _guard = repo.file_lock.write().unwrap();
         }
     }
 

pageserver/src/tenant_tasks.rs

@@ -0,0 +1,289 @@
+//! This module contains functions to serve per-tenant background processes,
+//! such as compaction and GC
+
+use std::collections::HashMap;
+use std::ops::ControlFlow;
+use std::time::Duration;
+
+use crate::repository::Repository;
+use crate::tenant_mgr::TenantState;
+use crate::thread_mgr::ThreadKind;
+use crate::{tenant_mgr, thread_mgr};
+use anyhow::{self, Context};
+use futures::stream::FuturesUnordered;
+use futures::StreamExt;
+use metrics::{register_int_counter_vec, IntCounterVec};
+use once_cell::sync::{Lazy, OnceCell};
+use tokio::sync::mpsc;
+use tokio::sync::watch;
+use tracing::*;
+use utils::zid::ZTenantId;
+
+static TENANT_TASK_EVENTS: Lazy<IntCounterVec> = Lazy::new(|| {
+    register_int_counter_vec!(
+        "pageserver_tenant_task_events",
+        "Number of task start/stop/fail events.",
+        &["event"],
+    )
+    .expect("Failed to register tenant_task_events metric")
+});
+
+///
+/// Compaction task's main loop
+///
+async fn compaction_loop(tenantid: ZTenantId, mut cancel: watch::Receiver<()>) {
+    loop {
+        trace!("waking up");
+
+        // Run blocking part of the task
+        let period: Result<Result<_, anyhow::Error>, _> = tokio::task::spawn_blocking(move || {
+            // Break if tenant is not active
+            if tenant_mgr::get_tenant_state(tenantid) != Some(TenantState::Active) {
+                return Ok(ControlFlow::Break(()));
+            }
+
+            // Break if we're not allowed to write to disk
+            let repo = tenant_mgr::get_repository_for_tenant(tenantid)?;
+            // TODO do this inside repo.compaction_iteration instead.
+            let _guard = match repo.file_lock.try_read() {
+                Ok(g) => g,
+                Err(_) => return Ok(ControlFlow::Break(())),
+            };
+
+            // Run compaction
+            let compaction_period = repo.get_compaction_period();
+            repo.compaction_iteration()?;
+            Ok(ControlFlow::Continue(compaction_period))
+        })
+        .await;
+
+        // Decide whether to sleep or break
+        let sleep_duration = match period {
+            Ok(Ok(ControlFlow::Continue(period))) => period,
+            Ok(Ok(ControlFlow::Break(()))) => break,
+            Ok(Err(e)) => {
+                error!("Compaction failed, retrying: {}", e);
+                Duration::from_secs(2)
+            }
+            Err(e) => {
+                error!("Compaction join error, retrying: {}", e);
+                Duration::from_secs(2)
+            }
+        };
+
+        // Sleep
+        tokio::select! {
+            _ = cancel.changed() => {
+                trace!("received cancellation request");
+                break;
+            },
+            _ = tokio::time::sleep(sleep_duration) => {},
+        }
+    }
+
+    trace!(
+        "compaction loop stopped. State is {:?}",
+        tenant_mgr::get_tenant_state(tenantid)
+    );
+}
+
+static START_GC_LOOP: OnceCell<mpsc::Sender<ZTenantId>> = OnceCell::new();
+static START_COMPACTION_LOOP: OnceCell<mpsc::Sender<ZTenantId>> = OnceCell::new();
+
+/// Spawn a task that will periodically schedule garbage collection until
+/// the tenant becomes inactive. This should be called on tenant
+/// activation.
+pub fn start_gc_loop(tenantid: ZTenantId) -> anyhow::Result<()> {
+    START_GC_LOOP
+        .get()
+        .context("Failed to get START_GC_LOOP")?
+        .blocking_send(tenantid)
+        .context("Failed to send to START_GC_LOOP channel")?;
+    Ok(())
+}
+
+/// Spawn a task that will periodically schedule compaction until
+/// the tenant becomes inactive. This should be called on tenant
+/// activation.
+pub fn start_compaction_loop(tenantid: ZTenantId) -> anyhow::Result<()> {
+    START_COMPACTION_LOOP
+        .get()
+        .context("failed to get START_COMPACTION_LOOP")?
+        .blocking_send(tenantid)
+        .context("failed to send to START_COMPACTION_LOOP")?;
+    Ok(())
+}
+
+/// Spawn the TenantTaskManager
+/// This needs to be called before start_gc_loop or start_compaction_loop
+pub fn init_tenant_task_pool() -> anyhow::Result<()> {
+    let runtime = tokio::runtime::Builder::new_multi_thread()
+        .thread_name("tenant-task-worker")
+        .worker_threads(40) // Way more than necessary
+        .max_blocking_threads(100) // Way more than necessary
+        .enable_all()
+        .build()?;
+
+    let (gc_send, mut gc_recv) = mpsc::channel::<ZTenantId>(100);
+    START_GC_LOOP
+        .set(gc_send)
+        .expect("Failed to set START_GC_LOOP");
+
+    let (compaction_send, mut compaction_recv) = mpsc::channel::<ZTenantId>(100);
+    START_COMPACTION_LOOP
+        .set(compaction_send)
+        .expect("Failed to set START_COMPACTION_LOOP");
+
+    // TODO this is getting repetitive
+    let mut gc_loops = HashMap::<ZTenantId, watch::Sender<()>>::new();
+    let mut compaction_loops = HashMap::<ZTenantId, watch::Sender<()>>::new();
+
+    thread_mgr::spawn(
+        ThreadKind::TenantTaskManager,
+        None,
+        None,
+        "Tenant task manager main thread",
+        true,
+        move || {
+            runtime.block_on(async move {
+                let mut futures = FuturesUnordered::new();
+                loop {
+                    tokio::select! {
+                        _ = thread_mgr::shutdown_watcher() => {
+                            // Send cancellation to all tasks
+                            for (_, cancel) in gc_loops.drain() {
+                                cancel.send(()).ok();
+                            }
+                            for (_, cancel) in compaction_loops.drain() {
+                                cancel.send(()).ok();
+                            }
+
+                            // Exit after all tasks finish
+                            while let Some(result) = futures.next().await {
+                                match result {
+                                    Ok(()) => {
+                                        TENANT_TASK_EVENTS.with_label_values(&["stop"]).inc();
+                                    },
+                                    Err(e) => {
+                                        TENANT_TASK_EVENTS.with_label_values(&["panic"]).inc();
+                                        error!("loop join error {}", e)
+                                    },
+                                }
+                            }
+                            break;
+                        },
+                        tenantid = gc_recv.recv() => {
+                            let tenantid = tenantid.expect("Gc task channel closed unexpectedly");
+
+                            // Spawn new task, request cancellation of the old one if exists
+                            let (cancel_send, cancel_recv) = watch::channel(());
+                            let handle = tokio::spawn(gc_loop(tenantid, cancel_recv)
+                                .instrument(trace_span!("gc loop", tenant = %tenantid)));
+                            if let Some(old_cancel_send) = gc_loops.insert(tenantid, cancel_send) {
+                                old_cancel_send.send(()).ok();
+                            }
+
+                            // Update metrics, remember handle
+                            TENANT_TASK_EVENTS.with_label_values(&["start"]).inc();
+                            futures.push(handle);
+                        },
+                        tenantid = compaction_recv.recv() => {
+                            let tenantid = tenantid.expect("Compaction task channel closed unexpectedly");
+
+                            // Spawn new task, request cancellation of the old one if exists
+                            let (cancel_send, cancel_recv) = watch::channel(());
+                            // TODO this instrument doesn't work
+                            let handle = tokio::spawn(compaction_loop(tenantid, cancel_recv)
+                                .instrument(trace_span!("compaction loop", tenant = %tenantid)));
+                            if let Some(old_cancel_send) = compaction_loops.insert(tenantid, cancel_send) {
+                                old_cancel_send.send(()).ok();
+                            }
+
+                            // Update metrics, remember handle
+                            TENANT_TASK_EVENTS.with_label_values(&["start"]).inc();
+                            futures.push(handle);
+                        },
+                        result = futures.next() => {
+                            // Log and count any unhandled panics
+                            match result {
+                                Some(Ok(())) => {
+                                    TENANT_TASK_EVENTS.with_label_values(&["stop"]).inc();
+                                },
+                                Some(Err(e)) => {
+                                    TENANT_TASK_EVENTS.with_label_values(&["panic"]).inc();
+                                    error!("loop join error {}", e)
+                                },
+                                None => {},
+                            };
+                        },
+                    }
+                }
+            });
+            Ok(())
+        },
+    )?;
+
+    Ok(())
+}
+
+///
+/// GC task's main loop
+///
+async fn gc_loop(tenantid: ZTenantId, mut cancel: watch::Receiver<()>) {
+    loop {
+        trace!("waking up");
+
+        // Run blocking part of the task
+        let period: Result<Result<_, anyhow::Error>, _> = tokio::task::spawn_blocking(move || {
+            // Break if tenant is not active
+            if tenant_mgr::get_tenant_state(tenantid) != Some(TenantState::Active) {
+                return Ok(ControlFlow::Break(()));
+            }
+
+            // Break if we're not allowed to write to disk
+            let repo = tenant_mgr::get_repository_for_tenant(tenantid)?;
+            // TODO do this inside repo.gc_iteration instead.
+            let _guard = match repo.file_lock.try_read() {
+                Ok(g) => g,
+                Err(_) => return Ok(ControlFlow::Break(())),
+            };
+
+            // Run gc
+            let gc_period = repo.get_gc_period();
+            let gc_horizon = repo.get_gc_horizon();
+            if gc_horizon > 0 {
+                repo.gc_iteration(None, gc_horizon, repo.get_pitr_interval(), false)?;
+            }
+
+            Ok(ControlFlow::Continue(gc_period))
+        })
+        .await;
+
+        // Decide whether to sleep or break
+        let sleep_duration = match period {
+            Ok(Ok(ControlFlow::Continue(period))) => period,
+            Ok(Ok(ControlFlow::Break(()))) => break,
+            Ok(Err(e)) => {
+                error!("Gc failed, retrying: {}", e);
+                Duration::from_secs(2)
+            }
+            Err(e) => {
+                error!("Gc join error, retrying: {}", e);
+                Duration::from_secs(2)
+            }
+        };
+
+        // Sleep
+        tokio::select! {
+            _ = cancel.changed() => {
+                trace!("received cancellation request");
+                break;
+            },
+            _ = tokio::time::sleep(sleep_duration) => {},
+        }
+    }
+    trace!(
+        "GC loop stopped. State is {:?}",
+        tenant_mgr::get_tenant_state(tenantid)
+    );
+}

pageserver/src/tenant_threads.rs

@@ -1,79 +0,0 @@
-//! This module contains functions to serve per-tenant background processes,
-//! such as compaction and GC
-use crate::repository::Repository;
-use crate::tenant_mgr;
-use crate::tenant_mgr::TenantState;
-use anyhow::Result;
-use std::time::Duration;
-use tracing::*;
-use utils::zid::ZTenantId;
-
-///
-/// Compaction thread's main loop
-///
-pub fn compact_loop(tenantid: ZTenantId) -> Result<()> {
-    if let Err(err) = compact_loop_ext(tenantid) {
-        error!("compact loop terminated with error: {:?}", err);
-        Err(err)
-    } else {
-        Ok(())
-    }
-}
-
-fn compact_loop_ext(tenantid: ZTenantId) -> Result<()> {
-    loop {
-        if tenant_mgr::get_tenant_state(tenantid) != Some(TenantState::Active) {
-            break;
-        }
-        let repo = tenant_mgr::get_repository_for_tenant(tenantid)?;
-        let compaction_period = repo.get_compaction_period();
-
-        std::thread::sleep(compaction_period);
-        trace!("compaction thread for tenant {} waking up", tenantid);
-
-        // Compact timelines
-        let repo = tenant_mgr::get_repository_for_tenant(tenantid)?;
-        repo.compaction_iteration()?;
-    }
-
-    trace!(
-        "compaction thread stopped for tenant {} state is {:?}",
-        tenantid,
-        tenant_mgr::get_tenant_state(tenantid)
-    );
-    Ok(())
-}
-
-///
-/// GC thread's main loop
-///
-pub fn gc_loop(tenantid: ZTenantId) -> Result<()> {
-    loop {
-        if tenant_mgr::get_tenant_state(tenantid) != Some(TenantState::Active) {
-            break;
-        }
-
-        trace!("gc thread for tenant {} waking up", tenantid);
-        let repo = tenant_mgr::get_repository_for_tenant(tenantid)?;
-        let gc_horizon = repo.get_gc_horizon();
-        // Garbage collect old files that are not needed for PITR anymore
-        if gc_horizon > 0 {
-            repo.gc_iteration(None, gc_horizon, repo.get_pitr_interval(), false)?;
-        }
-
-        // TODO Write it in more adequate way using
-        // condvar.wait_timeout() or something
-        let mut sleep_time = repo.get_gc_period().as_secs();
-        while sleep_time > 0 && tenant_mgr::get_tenant_state(tenantid) == Some(TenantState::Active)
-        {
-            sleep_time -= 1;
-            std::thread::sleep(Duration::from_secs(1));
-        }
-    }
-    trace!(
-        "GC thread stopped for tenant {} state is {:?}",
-        tenantid,
-        tenant_mgr::get_tenant_state(tenantid)
-    );
-    Ok(())
-}

pageserver/src/thread_mgr.rs

@@ -94,11 +94,8 @@ pub enum ThreadKind {
     // Main walreceiver manager thread that ensures that every timeline spawns a connection to safekeeper, to fetch WAL.
     WalReceiverManager,
 
-    // Thread that handles compaction of all timelines for a tenant.
-    Compactor,
-
-    // Thread that handles GC of a tenant
-    GarbageCollector,
+    // Thread that schedules new compaction and gc jobs
+    TenantTaskManager,
 
     // Thread that flushes frozen in-memory layers to disk
     LayerFlushThread,

test_runner/batch_others/test_tenant_tasks.py

@@ -0,0 +1,66 @@
+from fixtures.neon_fixtures import NeonEnvBuilder
+from uuid import UUID
+
+
+def get_only_element(l):
+    assert len(l) == 1
+    return l[0]
+
+
+def test_tenant_tasks(neon_env_builder: NeonEnvBuilder):
+    name = "test_tenant_tasks"
+    env = neon_env_builder.init_start()
+    client = env.pageserver.http_client()
+
+    def get_state(tenant):
+        all_states = client.tenant_list()
+        matching = [t for t in all_states if t["id"] == tenant.hex]
+        return get_only_element(matching)["state"]
+
+    def get_metric_value(name):
+        metrics = client.get_metrics()
+        relevant = [line for line in metrics.splitlines() if line.startswith(name)]
+        if len(relevant) == 0:
+            return 0
+        line = get_only_element(relevant)
+        value = line.lstrip(name).strip()
+        return int(value)
+
+    def detach_all_timelines(tenant):
+        timelines = [UUID(t["timeline_id"]) for t in client.timeline_list(tenant)]
+        for t in timelines:
+            client.timeline_detach(tenant, t)
+
+    # Create tenant, start compute
+    tenant, _ = env.neon_cli.create_tenant()
+    timeline = env.neon_cli.create_timeline(name, tenant_id=tenant)
+    pg = env.postgres.create_start(name, tenant_id=tenant)
+    assert(get_state(tenant) == "Active")
+
+    # Stop compute, detach timelines
+    # TODO tenant should go idle even if we don't explicitly detach
+    pg.stop()
+    detach_all_timelines(tenant)
+
+    import time; time.sleep(1)
+    assert(get_state(tenant) == "Idle")
+
+    # Detach all tenants
+    for tenant_info in client.tenant_list():
+        tenant_id = UUID(tenant_info["id"])
+        detach_all_timelines(tenant_id)
+
+        # TODO poll wait until idle instead
+        import time; time.sleep(1)
+        assert get_state(tenant_id) == "Idle"
+
+    # Read metrics
+    import time; time.sleep(1)
+    tasks_started = get_metric_value('pageserver_tenant_task_events{event="start"}')
+    tasks_ended = get_metric_value('pageserver_tenant_task_events{event="stop"}')
+    tasks_panicked = get_metric_value('pageserver_tenant_task_events{event="panic"}')
+
+    # TODO this fails because the "active -> idle" transition only waits for gc to
+    #      finish without cancelling it, and gc waits for 100 seconds.
+    assert tasks_started == tasks_ended
+    assert tasks_panicked == 0