Line data Source code
1 : //! This module acts as a switchboard to access different repositories managed by this
2 : //! page server.
3 :
4 : use camino::{Utf8DirEntry, Utf8Path, Utf8PathBuf};
5 : use futures::StreamExt;
6 : use itertools::Itertools;
7 : use pageserver_api::key::Key;
8 : use pageserver_api::models::LocationConfigMode;
9 : use pageserver_api::shard::{
10 : ShardCount, ShardIdentity, ShardIndex, ShardNumber, ShardStripeSize, TenantShardId,
11 : };
12 : use pageserver_api::upcall_api::ReAttachResponseTenant;
13 : use rand::{distributions::Alphanumeric, Rng};
14 : use std::borrow::Cow;
15 : use std::cmp::Ordering;
16 : use std::collections::{BTreeMap, HashMap};
17 : use std::ops::Deref;
18 : use std::sync::Arc;
19 : use std::time::Duration;
20 : use sysinfo::SystemExt;
21 : use tokio::fs;
22 :
23 : use anyhow::Context;
24 : use once_cell::sync::Lazy;
25 : use tokio::task::JoinSet;
26 : use tokio_util::sync::CancellationToken;
27 : use tracing::*;
28 :
29 : use utils::{backoff, completion, crashsafe};
30 :
31 : use crate::config::PageServerConf;
32 : use crate::context::{DownloadBehavior, RequestContext};
33 : use crate::control_plane_client::{
34 : ControlPlaneClient, ControlPlaneGenerationsApi, RetryForeverError,
35 : };
36 : use crate::deletion_queue::DeletionQueueClient;
37 : use crate::http::routes::ACTIVE_TENANT_TIMEOUT;
38 : use crate::metrics::{TENANT, TENANT_MANAGER as METRICS};
39 : use crate::task_mgr::{TaskKind, BACKGROUND_RUNTIME};
40 : use crate::tenant::config::{
41 : AttachedLocationConfig, AttachmentMode, LocationConf, LocationMode, SecondaryLocationConfig,
42 : };
43 : use crate::tenant::span::debug_assert_current_span_has_tenant_id;
44 : use crate::tenant::storage_layer::inmemory_layer;
45 : use crate::tenant::timeline::ShutdownMode;
46 : use crate::tenant::{AttachedTenantConf, GcError, LoadConfigError, SpawnMode, Tenant, TenantState};
47 : use crate::virtual_file::MaybeFatalIo;
48 : use crate::{InitializationOrder, TEMP_FILE_SUFFIX};
49 :
50 : use utils::crashsafe::path_with_suffix_extension;
51 : use utils::fs_ext::PathExt;
52 : use utils::generation::Generation;
53 : use utils::id::{TenantId, TimelineId};
54 :
55 : use super::remote_timeline_client::remote_tenant_path;
56 : use super::secondary::SecondaryTenant;
57 : use super::timeline::detach_ancestor::PreparedTimelineDetach;
58 : use super::{GlobalShutDown, TenantSharedResources};
59 :
60 : /// For a tenant that appears in TenantsMap, it may either be
61 : /// - `Attached`: has a full Tenant object, is elegible to service
62 : /// reads and ingest WAL.
63 : /// - `Secondary`: is only keeping a local cache warm.
64 : ///
65 : /// Secondary is a totally distinct state rather than being a mode of a `Tenant`, because
66 : /// that way we avoid having to carefully switch a tenant's ingestion etc on and off during
67 : /// its lifetime, and we can preserve some important safety invariants like `Tenant` always
68 : /// having a properly acquired generation (Secondary doesn't need a generation)
69 : #[derive(Clone)]
70 : pub(crate) enum TenantSlot {
71 : Attached(Arc<Tenant>),
72 : Secondary(Arc<SecondaryTenant>),
73 : /// In this state, other administrative operations acting on the TenantId should
74 : /// block, or return a retry indicator equivalent to HTTP 503.
75 : InProgress(utils::completion::Barrier),
76 : }
77 :
78 : impl std::fmt::Debug for TenantSlot {
79 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
80 0 : match self {
81 0 : Self::Attached(tenant) => write!(f, "Attached({})", tenant.current_state()),
82 0 : Self::Secondary(_) => write!(f, "Secondary"),
83 0 : Self::InProgress(_) => write!(f, "InProgress"),
84 : }
85 0 : }
86 : }
87 :
88 : impl TenantSlot {
89 : /// Return the `Tenant` in this slot if attached, else None
90 0 : fn get_attached(&self) -> Option<&Arc<Tenant>> {
91 0 : match self {
92 0 : Self::Attached(t) => Some(t),
93 0 : Self::Secondary(_) => None,
94 0 : Self::InProgress(_) => None,
95 : }
96 0 : }
97 : }
98 :
99 : /// The tenants known to the pageserver.
100 : /// The enum variants are used to distinguish the different states that the pageserver can be in.
101 : pub(crate) enum TenantsMap {
102 : /// [`init_tenant_mgr`] is not done yet.
103 : Initializing,
104 : /// [`init_tenant_mgr`] is done, all on-disk tenants have been loaded.
105 : /// New tenants can be added using [`tenant_map_acquire_slot`].
106 : Open(BTreeMap<TenantShardId, TenantSlot>),
107 : /// The pageserver has entered shutdown mode via [`TenantManager::shutdown`].
108 : /// Existing tenants are still accessible, but no new tenants can be created.
109 : ShuttingDown(BTreeMap<TenantShardId, TenantSlot>),
110 : }
111 :
112 : /// When resolving a TenantId to a shard, we may be looking for the 0th
113 : /// shard, or we might be looking for whichever shard holds a particular page.
114 : #[derive(Copy, Clone)]
115 : pub(crate) enum ShardSelector {
116 : /// Only return the 0th shard, if it is present. If a non-0th shard is present,
117 : /// ignore it.
118 : Zero,
119 : /// Pick the shard that holds this key
120 : Page(Key),
121 : /// The shard ID is known: pick the given shard
122 : Known(ShardIndex),
123 : }
124 :
125 : /// A convenience for use with the re_attach ControlPlaneClient function: rather
126 : /// than the serializable struct, we build this enum that encapsulates
127 : /// the invariant that attached tenants always have generations.
128 : ///
129 : /// This represents the subset of a LocationConfig that we receive during re-attach.
130 : pub(crate) enum TenantStartupMode {
131 : Attached((AttachmentMode, Generation)),
132 : Secondary,
133 : }
134 :
135 : impl TenantStartupMode {
136 : /// Return the generation & mode that should be used when starting
137 : /// this tenant.
138 : ///
139 : /// If this returns None, the re-attach struct is in an invalid state and
140 : /// should be ignored in the response.
141 0 : fn from_reattach_tenant(rart: ReAttachResponseTenant) -> Option<Self> {
142 0 : match (rart.mode, rart.gen) {
143 0 : (LocationConfigMode::Detached, _) => None,
144 0 : (LocationConfigMode::Secondary, _) => Some(Self::Secondary),
145 0 : (LocationConfigMode::AttachedMulti, Some(g)) => {
146 0 : Some(Self::Attached((AttachmentMode::Multi, Generation::new(g))))
147 : }
148 0 : (LocationConfigMode::AttachedSingle, Some(g)) => {
149 0 : Some(Self::Attached((AttachmentMode::Single, Generation::new(g))))
150 : }
151 0 : (LocationConfigMode::AttachedStale, Some(g)) => {
152 0 : Some(Self::Attached((AttachmentMode::Stale, Generation::new(g))))
153 : }
154 : _ => {
155 0 : tracing::warn!(
156 0 : "Received invalid re-attach state for tenant {}: {rart:?}",
157 : rart.id
158 : );
159 0 : None
160 : }
161 : }
162 0 : }
163 : }
164 :
165 : /// Result type for looking up a TenantId to a specific shard
166 : pub(crate) enum ShardResolveResult {
167 : NotFound,
168 : Found(Arc<Tenant>),
169 : // Wait for this barrrier, then query again
170 : InProgress(utils::completion::Barrier),
171 : }
172 :
173 : impl TenantsMap {
174 : /// Convenience function for typical usage, where we want to get a `Tenant` object, for
175 : /// working with attached tenants. If the TenantId is in the map but in Secondary state,
176 : /// None is returned.
177 0 : pub(crate) fn get(&self, tenant_shard_id: &TenantShardId) -> Option<&Arc<Tenant>> {
178 0 : match self {
179 0 : TenantsMap::Initializing => None,
180 0 : TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => {
181 0 : m.get(tenant_shard_id).and_then(|slot| slot.get_attached())
182 : }
183 : }
184 0 : }
185 :
186 : #[cfg(all(debug_assertions, not(test)))]
187 0 : pub(crate) fn len(&self) -> usize {
188 0 : match self {
189 0 : TenantsMap::Initializing => 0,
190 0 : TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => m.len(),
191 : }
192 0 : }
193 : }
194 :
195 : /// Precursor to deletion of a tenant dir: we do a fast rename to a tmp path, and then
196 : /// the slower actual deletion in the background.
197 : ///
198 : /// This is "safe" in that that it won't leave behind a partially deleted directory
199 : /// at the original path, because we rename with TEMP_FILE_SUFFIX before starting deleting
200 : /// the contents.
201 : ///
202 : /// This is pageserver-specific, as it relies on future processes after a crash to check
203 : /// for TEMP_FILE_SUFFIX when loading things.
204 0 : async fn safe_rename_tenant_dir(path: impl AsRef<Utf8Path>) -> std::io::Result<Utf8PathBuf> {
205 0 : let parent = path
206 0 : .as_ref()
207 0 : .parent()
208 0 : // It is invalid to call this function with a relative path. Tenant directories
209 0 : // should always have a parent.
210 0 : .ok_or(std::io::Error::new(
211 0 : std::io::ErrorKind::InvalidInput,
212 0 : "Path must be absolute",
213 0 : ))?;
214 0 : let rand_suffix = rand::thread_rng()
215 0 : .sample_iter(&Alphanumeric)
216 0 : .take(8)
217 0 : .map(char::from)
218 0 : .collect::<String>()
219 0 : + TEMP_FILE_SUFFIX;
220 0 : let tmp_path = path_with_suffix_extension(&path, &rand_suffix);
221 0 : fs::rename(path.as_ref(), &tmp_path).await?;
222 0 : fs::File::open(parent).await?.sync_all().await?;
223 0 : Ok(tmp_path)
224 0 : }
225 :
226 : /// See [`Self::spawn`].
227 : #[derive(Clone)]
228 : pub struct BackgroundPurges(Arc<std::sync::Mutex<BackgroundPurgesInner>>);
229 : enum BackgroundPurgesInner {
230 : Open(tokio::task::JoinSet<()>),
231 : // we use the async mutex for coalescing
232 : ShuttingDown(Arc<tokio::sync::Mutex<tokio::task::JoinSet<()>>>),
233 : }
234 :
235 : impl Default for BackgroundPurges {
236 0 : fn default() -> Self {
237 0 : Self(Arc::new(std::sync::Mutex::new(
238 0 : BackgroundPurgesInner::Open(JoinSet::new()),
239 0 : )))
240 0 : }
241 : }
242 :
243 : impl BackgroundPurges {
244 : /// When we have moved a tenant's content to a temporary directory, we may delete it lazily in
245 : /// the background, and thereby avoid blocking any API requests on this deletion completing.
246 : ///
247 : /// Although we are cleaning up the tenant, this task is not meant to be bound by the lifetime of the tenant in memory.
248 : /// Thus the [`BackgroundPurges`] type to keep track of these tasks.
249 0 : pub fn spawn(&self, tmp_path: Utf8PathBuf) {
250 0 : let mut guard = self.0.lock().unwrap();
251 0 : let jset = match &mut *guard {
252 0 : BackgroundPurgesInner::Open(ref mut jset) => jset,
253 : BackgroundPurgesInner::ShuttingDown(_) => {
254 0 : warn!("trying to spawn background purge during shutdown, ignoring");
255 0 : return;
256 : }
257 : };
258 0 : jset.spawn_on(
259 0 : async move {
260 0 : if let Err(error) = fs::remove_dir_all(tmp_path.as_path()).await {
261 : // should we fatal_io_error here?
262 0 : warn!(%error, path=%tmp_path, "failed to purge tenant directory");
263 0 : }
264 0 : }
265 0 : .instrument(info_span!(parent: None, "background_purge")),
266 0 : BACKGROUND_RUNTIME.handle(),
267 : );
268 0 : }
269 :
270 : /// When this future completes, all background purges have completed.
271 : /// The first poll of the future will already lock out new background purges spawned via [`Self::spawn`].
272 : ///
273 : /// Concurrent calls will coalesce.
274 : ///
275 : /// # Cancellation-Safety
276 : ///
277 : /// If this future is dropped before polled to completion, concurrent and subsequent
278 : /// instances of this future will continue to be correct.
279 0 : #[instrument(skip_all)]
280 : pub async fn shutdown(&self) {
281 : let jset = {
282 : let mut guard = self.0.lock().unwrap();
283 : match &mut *guard {
284 : BackgroundPurgesInner::Open(jset) => {
285 : *guard = BackgroundPurgesInner::ShuttingDown(Arc::new(tokio::sync::Mutex::new(
286 : std::mem::take(jset),
287 : )))
288 : }
289 : BackgroundPurgesInner::ShuttingDown(_) => {
290 : // calling shutdown multiple times is most likely a bug in pageserver shutdown code
291 : warn!("already shutting down");
292 : }
293 : };
294 : match &mut *guard {
295 : BackgroundPurgesInner::ShuttingDown(ref mut jset) => jset.clone(),
296 : BackgroundPurgesInner::Open(_) => {
297 : unreachable!("above code transitions into shut down state");
298 : }
299 : }
300 : };
301 : let mut jset = jset.lock().await; // concurrent callers coalesce here
302 : while let Some(res) = jset.join_next().await {
303 : match res {
304 : Ok(()) => {}
305 : Err(e) if e.is_panic() => {
306 : // If it panicked, the error is already logged by the panic hook.
307 : }
308 : Err(e) if e.is_cancelled() => {
309 : unreachable!("we don't cancel the joinset or runtime")
310 : }
311 : Err(e) => {
312 : // No idea when this can happen, but let's log it.
313 : warn!(%e, "background purge task failed or panicked");
314 : }
315 : }
316 : }
317 : }
318 : }
319 :
320 : static TENANTS: Lazy<std::sync::RwLock<TenantsMap>> =
321 2 : Lazy::new(|| std::sync::RwLock::new(TenantsMap::Initializing));
322 :
323 : /// The TenantManager is responsible for storing and mutating the collection of all tenants
324 : /// that this pageserver process has state for. Every Tenant and SecondaryTenant instance
325 : /// lives inside the TenantManager.
326 : ///
327 : /// The most important role of the TenantManager is to prevent conflicts: e.g. trying to attach
328 : /// the same tenant twice concurrently, or trying to configure the same tenant into secondary
329 : /// and attached modes concurrently.
330 : pub struct TenantManager {
331 : conf: &'static PageServerConf,
332 : // TODO: currently this is a &'static pointing to TENANTs. When we finish refactoring
333 : // out of that static variable, the TenantManager can own this.
334 : // See https://github.com/neondatabase/neon/issues/5796
335 : tenants: &'static std::sync::RwLock<TenantsMap>,
336 : resources: TenantSharedResources,
337 :
338 : // Long-running operations that happen outside of a [`Tenant`] lifetime should respect this token.
339 : // This is for edge cases like tenant deletion. In normal cases (within a Tenant lifetime),
340 : // tenants have their own cancellation tokens, which we fire individually in [`Self::shutdown`], or
341 : // when the tenant detaches.
342 : cancel: CancellationToken,
343 :
344 : background_purges: BackgroundPurges,
345 : }
346 :
347 0 : fn emergency_generations(
348 0 : tenant_confs: &HashMap<TenantShardId, Result<LocationConf, LoadConfigError>>,
349 0 : ) -> HashMap<TenantShardId, TenantStartupMode> {
350 0 : tenant_confs
351 0 : .iter()
352 0 : .filter_map(|(tid, lc)| {
353 0 : let lc = match lc {
354 0 : Ok(lc) => lc,
355 0 : Err(_) => return None,
356 : };
357 : Some((
358 0 : *tid,
359 0 : match &lc.mode {
360 0 : LocationMode::Attached(alc) => {
361 0 : TenantStartupMode::Attached((alc.attach_mode, alc.generation))
362 : }
363 0 : LocationMode::Secondary(_) => TenantStartupMode::Secondary,
364 : },
365 : ))
366 0 : })
367 0 : .collect()
368 0 : }
369 :
370 0 : async fn init_load_generations(
371 0 : conf: &'static PageServerConf,
372 0 : tenant_confs: &HashMap<TenantShardId, Result<LocationConf, LoadConfigError>>,
373 0 : resources: &TenantSharedResources,
374 0 : cancel: &CancellationToken,
375 0 : ) -> anyhow::Result<Option<HashMap<TenantShardId, TenantStartupMode>>> {
376 0 : let generations = if conf.control_plane_emergency_mode {
377 0 : error!(
378 0 : "Emergency mode! Tenants will be attached unsafely using their last known generation"
379 : );
380 0 : emergency_generations(tenant_confs)
381 0 : } else if let Some(client) = ControlPlaneClient::new(conf, cancel) {
382 0 : info!("Calling control plane API to re-attach tenants");
383 : // If we are configured to use the control plane API, then it is the source of truth for what tenants to load.
384 0 : match client.re_attach(conf).await {
385 0 : Ok(tenants) => tenants
386 0 : .into_iter()
387 0 : .flat_map(|(id, rart)| {
388 0 : TenantStartupMode::from_reattach_tenant(rart).map(|tsm| (id, tsm))
389 0 : })
390 0 : .collect(),
391 : Err(RetryForeverError::ShuttingDown) => {
392 0 : anyhow::bail!("Shut down while waiting for control plane re-attach response")
393 : }
394 : }
395 : } else {
396 0 : info!("Control plane API not configured, tenant generations are disabled");
397 0 : return Ok(None);
398 : };
399 :
400 : // The deletion queue needs to know about the startup attachment state to decide which (if any) stored
401 : // deletion list entries may still be valid. We provide that by pushing a recovery operation into
402 : // the queue. Sequential processing of te queue ensures that recovery is done before any new tenant deletions
403 : // are processed, even though we don't block on recovery completing here.
404 0 : let attached_tenants = generations
405 0 : .iter()
406 0 : .flat_map(|(id, start_mode)| {
407 0 : match start_mode {
408 0 : TenantStartupMode::Attached((_mode, generation)) => Some(generation),
409 0 : TenantStartupMode::Secondary => None,
410 : }
411 0 : .map(|gen| (*id, *gen))
412 0 : })
413 0 : .collect();
414 0 : resources.deletion_queue_client.recover(attached_tenants)?;
415 :
416 0 : Ok(Some(generations))
417 0 : }
418 :
419 : /// Given a directory discovered in the pageserver's tenants/ directory, attempt
420 : /// to load a tenant config from it.
421 : ///
422 : /// If we cleaned up something expected (like an empty dir or a temp dir), return None.
423 0 : fn load_tenant_config(
424 0 : conf: &'static PageServerConf,
425 0 : tenant_shard_id: TenantShardId,
426 0 : dentry: Utf8DirEntry,
427 0 : ) -> Option<Result<LocationConf, LoadConfigError>> {
428 0 : let tenant_dir_path = dentry.path().to_path_buf();
429 0 : if crate::is_temporary(&tenant_dir_path) {
430 0 : info!("Found temporary tenant directory, removing: {tenant_dir_path}");
431 : // No need to use safe_remove_tenant_dir_all because this is already
432 : // a temporary path
433 0 : std::fs::remove_dir_all(&tenant_dir_path).fatal_err("delete temporary tenant dir");
434 0 : return None;
435 0 : }
436 0 :
437 0 : // This case happens if we crash during attachment before writing a config into the dir
438 0 : let is_empty = tenant_dir_path
439 0 : .is_empty_dir()
440 0 : .fatal_err("Checking for empty tenant dir");
441 0 : if is_empty {
442 0 : info!("removing empty tenant directory {tenant_dir_path:?}");
443 0 : std::fs::remove_dir(&tenant_dir_path).fatal_err("delete empty tenant dir");
444 0 : return None;
445 0 : }
446 0 :
447 0 : Some(Tenant::load_tenant_config(conf, &tenant_shard_id))
448 0 : }
449 :
450 : /// Initial stage of load: walk the local tenants directory, clean up any temp files,
451 : /// and load configurations for the tenants we found.
452 : ///
453 : /// Do this in parallel, because we expect 10k+ tenants, so serial execution can take
454 : /// seconds even on reasonably fast drives.
455 0 : async fn init_load_tenant_configs(
456 0 : conf: &'static PageServerConf,
457 0 : ) -> HashMap<TenantShardId, Result<LocationConf, LoadConfigError>> {
458 0 : let tenants_dir = conf.tenants_path();
459 :
460 0 : let dentries = tokio::task::spawn_blocking(move || -> Vec<Utf8DirEntry> {
461 0 : let context = format!("read tenants dir {tenants_dir}");
462 0 : let dir_entries = tenants_dir.read_dir_utf8().fatal_err(&context);
463 0 :
464 0 : dir_entries
465 0 : .collect::<Result<Vec<_>, std::io::Error>>()
466 0 : .fatal_err(&context)
467 0 : })
468 0 : .await
469 0 : .expect("Config load task panicked");
470 0 :
471 0 : let mut configs = HashMap::new();
472 0 :
473 0 : let mut join_set = JoinSet::new();
474 0 : for dentry in dentries {
475 0 : let tenant_shard_id = match dentry.file_name().parse::<TenantShardId>() {
476 0 : Ok(id) => id,
477 : Err(_) => {
478 0 : warn!(
479 0 : "Invalid tenant path (garbage in our repo directory?): '{}'",
480 0 : dentry.file_name()
481 : );
482 0 : continue;
483 : }
484 : };
485 :
486 0 : join_set.spawn_blocking(move || {
487 0 : (
488 0 : tenant_shard_id,
489 0 : load_tenant_config(conf, tenant_shard_id, dentry),
490 0 : )
491 0 : });
492 0 : }
493 :
494 0 : while let Some(r) = join_set.join_next().await {
495 0 : let (tenant_shard_id, tenant_config) = r.expect("Panic in config load task");
496 0 : if let Some(tenant_config) = tenant_config {
497 0 : configs.insert(tenant_shard_id, tenant_config);
498 0 : }
499 : }
500 :
501 0 : configs
502 0 : }
503 :
504 0 : #[derive(Debug, thiserror::Error)]
505 : pub(crate) enum DeleteTenantError {
506 : #[error("Tenant map slot error {0}")]
507 : SlotError(#[from] TenantSlotError),
508 :
509 : #[error("Cancelled")]
510 : Cancelled,
511 :
512 : #[error(transparent)]
513 : Other(#[from] anyhow::Error),
514 : }
515 :
516 : /// Initialize repositories with locally available timelines.
517 : /// Timelines that are only partially available locally (remote storage has more data than this pageserver)
518 : /// are scheduled for download and added to the tenant once download is completed.
519 0 : #[instrument(skip_all)]
520 : pub async fn init_tenant_mgr(
521 : conf: &'static PageServerConf,
522 : background_purges: BackgroundPurges,
523 : resources: TenantSharedResources,
524 : init_order: InitializationOrder,
525 : cancel: CancellationToken,
526 : ) -> anyhow::Result<TenantManager> {
527 : let mut tenants = BTreeMap::new();
528 :
529 : let ctx = RequestContext::todo_child(TaskKind::Startup, DownloadBehavior::Warn);
530 :
531 : // Initialize dynamic limits that depend on system resources
532 : let system_memory =
533 : sysinfo::System::new_with_specifics(sysinfo::RefreshKind::new().with_memory())
534 : .total_memory();
535 : let max_ephemeral_layer_bytes =
536 : conf.ephemeral_bytes_per_memory_kb as u64 * (system_memory / 1024);
537 : tracing::info!("Initialized ephemeral layer size limit to {max_ephemeral_layer_bytes}, for {system_memory} bytes of memory");
538 : inmemory_layer::GLOBAL_RESOURCES.max_dirty_bytes.store(
539 : max_ephemeral_layer_bytes,
540 : std::sync::atomic::Ordering::Relaxed,
541 : );
542 :
543 : // Scan local filesystem for attached tenants
544 : let tenant_configs = init_load_tenant_configs(conf).await;
545 :
546 : // Determine which tenants are to be secondary or attached, and in which generation
547 : let tenant_modes = init_load_generations(conf, &tenant_configs, &resources, &cancel).await?;
548 :
549 : tracing::info!(
550 : "Attaching {} tenants at startup, warming up {} at a time",
551 : tenant_configs.len(),
552 : conf.concurrent_tenant_warmup.initial_permits()
553 : );
554 : TENANT.startup_scheduled.inc_by(tenant_configs.len() as u64);
555 :
556 : // Accumulate futures for writing tenant configs, so that we can execute in parallel
557 : let mut config_write_futs = Vec::new();
558 :
559 : // Update the location configs according to the re-attach response and persist them to disk
560 : tracing::info!("Updating {} location configs", tenant_configs.len());
561 : for (tenant_shard_id, location_conf) in tenant_configs {
562 : let tenant_dir_path = conf.tenant_path(&tenant_shard_id);
563 :
564 : let mut location_conf = match location_conf {
565 : Ok(l) => l,
566 : Err(e) => {
567 : // This should only happen in the case of a serialization bug or critical local I/O error: we cannot load this tenant
568 : error!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Failed to load tenant config, failed to {e:#}");
569 : continue;
570 : }
571 : };
572 :
573 : // FIXME: if we were attached, and get demoted to secondary on re-attach, we
574 : // don't have a place to get a config.
575 : // (https://github.com/neondatabase/neon/issues/5377)
576 : const DEFAULT_SECONDARY_CONF: SecondaryLocationConfig =
577 : SecondaryLocationConfig { warm: true };
578 :
579 : if let Some(tenant_modes) = &tenant_modes {
580 : // We have a generation map: treat it as the authority for whether
581 : // this tenant is really attached.
582 : match tenant_modes.get(&tenant_shard_id) {
583 : None => {
584 : info!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Detaching tenant, control plane omitted it in re-attach response");
585 :
586 : match safe_rename_tenant_dir(&tenant_dir_path).await {
587 : Ok(tmp_path) => {
588 : background_purges.spawn(tmp_path);
589 : }
590 : Err(e) => {
591 : error!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
592 : "Failed to move detached tenant directory '{tenant_dir_path}': {e:?}");
593 : }
594 : };
595 :
596 : // We deleted local content: move on to next tenant, don't try and spawn this one.
597 : continue;
598 : }
599 : Some(TenantStartupMode::Secondary) => {
600 : if !matches!(location_conf.mode, LocationMode::Secondary(_)) {
601 : location_conf.mode = LocationMode::Secondary(DEFAULT_SECONDARY_CONF);
602 : }
603 : }
604 : Some(TenantStartupMode::Attached((attach_mode, generation))) => {
605 : let old_gen_higher = match &location_conf.mode {
606 : LocationMode::Attached(AttachedLocationConfig {
607 : generation: old_generation,
608 : attach_mode: _attach_mode,
609 : }) => {
610 : if old_generation > generation {
611 : Some(old_generation)
612 : } else {
613 : None
614 : }
615 : }
616 : _ => None,
617 : };
618 : if let Some(old_generation) = old_gen_higher {
619 : tracing::error!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
620 : "Control plane gave decreasing generation ({generation:?}) in re-attach response for tenant that was attached in generation {:?}, demoting to secondary",
621 : old_generation
622 : );
623 :
624 : // We cannot safely attach this tenant given a bogus generation number, but let's avoid throwing away
625 : // local disk content: demote to secondary rather than detaching.
626 : location_conf.mode = LocationMode::Secondary(DEFAULT_SECONDARY_CONF);
627 : } else {
628 : location_conf.attach_in_generation(*attach_mode, *generation);
629 : }
630 : }
631 : }
632 : } else {
633 : // Legacy mode: no generation information, any tenant present
634 : // on local disk may activate
635 : info!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Starting tenant in legacy mode, no generation",);
636 : };
637 :
638 : // Presence of a generation number implies attachment: attach the tenant
639 : // if it wasn't already, and apply the generation number.
640 0 : config_write_futs.push(async move {
641 0 : let r = Tenant::persist_tenant_config(conf, &tenant_shard_id, &location_conf).await;
642 0 : (tenant_shard_id, location_conf, r)
643 0 : });
644 : }
645 :
646 : // Execute config writes with concurrency, to avoid bottlenecking on local FS write latency
647 : tracing::info!(
648 : "Writing {} location config files...",
649 : config_write_futs.len()
650 : );
651 : let config_write_results = futures::stream::iter(config_write_futs)
652 : .buffer_unordered(16)
653 : .collect::<Vec<_>>()
654 : .await;
655 :
656 : tracing::info!(
657 : "Spawning {} tenant shard locations...",
658 : config_write_results.len()
659 : );
660 : // For those shards that have live configurations, construct `Tenant` or `SecondaryTenant` objects and start them running
661 : for (tenant_shard_id, location_conf, config_write_result) in config_write_results {
662 : // Writing a config to local disk is foundational to startup up tenants: panic if we can't.
663 : config_write_result.fatal_err("write tenant shard config file");
664 :
665 : let tenant_dir_path = conf.tenant_path(&tenant_shard_id);
666 : let shard_identity = location_conf.shard;
667 : let slot = match location_conf.mode {
668 : LocationMode::Attached(attached_conf) => TenantSlot::Attached(
669 : tenant_spawn(
670 : conf,
671 : tenant_shard_id,
672 : &tenant_dir_path,
673 : resources.clone(),
674 : AttachedTenantConf::new(location_conf.tenant_conf, attached_conf),
675 : shard_identity,
676 : Some(init_order.clone()),
677 : SpawnMode::Lazy,
678 : &ctx,
679 : )
680 : .expect("global shutdown during init_tenant_mgr cannot happen"),
681 : ),
682 : LocationMode::Secondary(secondary_conf) => {
683 : info!(
684 : tenant_id = %tenant_shard_id.tenant_id,
685 : shard_id = %tenant_shard_id.shard_slug(),
686 : "Starting secondary tenant"
687 : );
688 : TenantSlot::Secondary(SecondaryTenant::new(
689 : tenant_shard_id,
690 : shard_identity,
691 : location_conf.tenant_conf,
692 : &secondary_conf,
693 : ))
694 : }
695 : };
696 :
697 : METRICS.slot_inserted(&slot);
698 : tenants.insert(tenant_shard_id, slot);
699 : }
700 :
701 : info!("Processed {} local tenants at startup", tenants.len());
702 :
703 : let mut tenants_map = TENANTS.write().unwrap();
704 : assert!(matches!(&*tenants_map, &TenantsMap::Initializing));
705 :
706 : *tenants_map = TenantsMap::Open(tenants);
707 :
708 : Ok(TenantManager {
709 : conf,
710 : tenants: &TENANTS,
711 : resources,
712 : cancel: CancellationToken::new(),
713 : background_purges,
714 : })
715 : }
716 :
717 : /// Wrapper for Tenant::spawn that checks invariants before running
718 : #[allow(clippy::too_many_arguments)]
719 0 : fn tenant_spawn(
720 0 : conf: &'static PageServerConf,
721 0 : tenant_shard_id: TenantShardId,
722 0 : tenant_path: &Utf8Path,
723 0 : resources: TenantSharedResources,
724 0 : location_conf: AttachedTenantConf,
725 0 : shard_identity: ShardIdentity,
726 0 : init_order: Option<InitializationOrder>,
727 0 : mode: SpawnMode,
728 0 : ctx: &RequestContext,
729 0 : ) -> Result<Arc<Tenant>, GlobalShutDown> {
730 0 : // All these conditions should have been satisfied by our caller: the tenant dir exists, is a well formed
731 0 : // path, and contains a configuration file. Assertions that do synchronous I/O are limited to debug mode
732 0 : // to avoid impacting prod runtime performance.
733 0 : assert!(!crate::is_temporary(tenant_path));
734 0 : debug_assert!(tenant_path.is_dir());
735 0 : debug_assert!(conf
736 0 : .tenant_location_config_path(&tenant_shard_id)
737 0 : .try_exists()
738 0 : .unwrap());
739 :
740 0 : Tenant::spawn(
741 0 : conf,
742 0 : tenant_shard_id,
743 0 : resources,
744 0 : location_conf,
745 0 : shard_identity,
746 0 : init_order,
747 0 : mode,
748 0 : ctx,
749 0 : )
750 0 : }
751 :
752 2 : async fn shutdown_all_tenants0(tenants: &std::sync::RwLock<TenantsMap>) {
753 2 : let mut join_set = JoinSet::new();
754 0 :
755 0 : #[cfg(all(debug_assertions, not(test)))]
756 0 : {
757 0 : // Check that our metrics properly tracked the size of the tenants map. This is a convenient location to check,
758 0 : // as it happens implicitly at the end of tests etc.
759 0 : let m = tenants.read().unwrap();
760 0 : debug_assert_eq!(METRICS.slots_total(), m.len() as u64);
761 : }
762 :
763 : // Atomically, 1. create the shutdown tasks and 2. prevent creation of new tenants.
764 2 : let (total_in_progress, total_attached) = {
765 2 : let mut m = tenants.write().unwrap();
766 2 : match &mut *m {
767 : TenantsMap::Initializing => {
768 0 : *m = TenantsMap::ShuttingDown(BTreeMap::default());
769 0 : info!("tenants map is empty");
770 0 : return;
771 : }
772 2 : TenantsMap::Open(tenants) => {
773 2 : let mut shutdown_state = BTreeMap::new();
774 2 : let mut total_in_progress = 0;
775 2 : let mut total_attached = 0;
776 :
777 2 : for (tenant_shard_id, v) in std::mem::take(tenants).into_iter() {
778 2 : match v {
779 0 : TenantSlot::Attached(t) => {
780 0 : shutdown_state.insert(tenant_shard_id, TenantSlot::Attached(t.clone()));
781 0 : join_set.spawn(
782 0 : async move {
783 0 : let res = {
784 0 : let (_guard, shutdown_progress) = completion::channel();
785 0 : t.shutdown(shutdown_progress, ShutdownMode::FreezeAndFlush).await
786 : };
787 :
788 0 : if let Err(other_progress) = res {
789 : // join the another shutdown in progress
790 0 : other_progress.wait().await;
791 0 : }
792 :
793 : // we cannot afford per tenant logging here, because if s3 is degraded, we are
794 : // going to log too many lines
795 0 : debug!("tenant successfully stopped");
796 0 : }
797 0 : .instrument(info_span!("shutdown", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug())),
798 : );
799 :
800 0 : total_attached += 1;
801 : }
802 0 : TenantSlot::Secondary(state) => {
803 0 : // We don't need to wait for this individually per-tenant: the
804 0 : // downloader task will be waited on eventually, this cancel
805 0 : // is just to encourage it to drop out if it is doing work
806 0 : // for this tenant right now.
807 0 : state.cancel.cancel();
808 0 :
809 0 : shutdown_state.insert(tenant_shard_id, TenantSlot::Secondary(state));
810 0 : }
811 2 : TenantSlot::InProgress(notify) => {
812 2 : // InProgress tenants are not visible in TenantsMap::ShuttingDown: we will
813 2 : // wait for their notifications to fire in this function.
814 2 : join_set.spawn(async move {
815 2 : notify.wait().await;
816 2 : });
817 2 :
818 2 : total_in_progress += 1;
819 2 : }
820 : }
821 : }
822 2 : *m = TenantsMap::ShuttingDown(shutdown_state);
823 2 : (total_in_progress, total_attached)
824 : }
825 : TenantsMap::ShuttingDown(_) => {
826 0 : error!("already shutting down, this function isn't supposed to be called more than once");
827 0 : return;
828 : }
829 : }
830 : };
831 :
832 2 : let started_at = std::time::Instant::now();
833 2 :
834 2 : info!(
835 0 : "Waiting for {} InProgress tenants and {} Attached tenants to shut down",
836 : total_in_progress, total_attached
837 : );
838 :
839 2 : let total = join_set.len();
840 2 : let mut panicked = 0;
841 2 : let mut buffering = true;
842 2 : const BUFFER_FOR: std::time::Duration = std::time::Duration::from_millis(500);
843 2 : let mut buffered = std::pin::pin!(tokio::time::sleep(BUFFER_FOR));
844 :
845 6 : while !join_set.is_empty() {
846 : tokio::select! {
847 : Some(joined) = join_set.join_next() => {
848 : match joined {
849 : Ok(()) => {},
850 : Err(join_error) if join_error.is_cancelled() => {
851 : unreachable!("we are not cancelling any of the tasks");
852 : }
853 : Err(join_error) if join_error.is_panic() => {
854 : // cannot really do anything, as this panic is likely a bug
855 : panicked += 1;
856 : }
857 : Err(join_error) => {
858 : warn!("unknown kind of JoinError: {join_error}");
859 : }
860 : }
861 : if !buffering {
862 : // buffer so that every 500ms since the first update (or starting) we'll log
863 : // how far away we are; this is because we will get SIGKILL'd at 10s, and we
864 : // are not able to log *then*.
865 : buffering = true;
866 : buffered.as_mut().reset(tokio::time::Instant::now() + BUFFER_FOR);
867 : }
868 : },
869 : _ = &mut buffered, if buffering => {
870 : buffering = false;
871 : info!(remaining = join_set.len(), total, elapsed_ms = started_at.elapsed().as_millis(), "waiting for tenants to shutdown");
872 : }
873 2 : }
874 2 : }
875 2 :
876 2 : if panicked > 0 {
877 0 : warn!(
878 : panicked,
879 0 : total, "observed panicks while shutting down tenants"
880 : );
881 2 : }
882 :
883 : // caller will log how long we took
884 2 : }
885 :
886 0 : #[derive(thiserror::Error, Debug)]
887 : pub(crate) enum UpsertLocationError {
888 : #[error("Bad config request: {0}")]
889 : BadRequest(anyhow::Error),
890 :
891 : #[error("Cannot change config in this state: {0}")]
892 : Unavailable(#[from] TenantMapError),
893 :
894 : #[error("Tenant is already being modified")]
895 : InProgress,
896 :
897 : #[error("Failed to flush: {0}")]
898 : Flush(anyhow::Error),
899 :
900 : /// This error variant is for unexpected situations (soft assertions) where the system is in an unexpected state.
901 : #[error("Internal error: {0}")]
902 : InternalError(anyhow::Error),
903 : }
904 :
905 : impl TenantManager {
906 : /// Convenience function so that anyone with a TenantManager can get at the global configuration, without
907 : /// having to pass it around everywhere as a separate object.
908 0 : pub(crate) fn get_conf(&self) -> &'static PageServerConf {
909 0 : self.conf
910 0 : }
911 :
912 : /// Gets the attached tenant from the in-memory data, erroring if it's absent, in secondary mode, or currently
913 : /// undergoing a state change (i.e. slot is InProgress).
914 : ///
915 : /// The return Tenant is not guaranteed to be active: check its status after obtaing it, or
916 : /// use [`Tenant::wait_to_become_active`] before using it if you will do I/O on it.
917 0 : pub(crate) fn get_attached_tenant_shard(
918 0 : &self,
919 0 : tenant_shard_id: TenantShardId,
920 0 : ) -> Result<Arc<Tenant>, GetTenantError> {
921 0 : let locked = self.tenants.read().unwrap();
922 :
923 0 : let peek_slot = tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Read)?;
924 :
925 0 : match peek_slot {
926 0 : Some(TenantSlot::Attached(tenant)) => Ok(Arc::clone(tenant)),
927 0 : Some(TenantSlot::InProgress(_)) => Err(GetTenantError::NotActive(tenant_shard_id)),
928 : None | Some(TenantSlot::Secondary(_)) => {
929 0 : Err(GetTenantError::NotFound(tenant_shard_id.tenant_id))
930 : }
931 : }
932 0 : }
933 :
934 0 : pub(crate) fn get_secondary_tenant_shard(
935 0 : &self,
936 0 : tenant_shard_id: TenantShardId,
937 0 : ) -> Option<Arc<SecondaryTenant>> {
938 0 : let locked = self.tenants.read().unwrap();
939 0 :
940 0 : let peek_slot = tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Read)
941 0 : .ok()
942 0 : .flatten();
943 :
944 0 : match peek_slot {
945 0 : Some(TenantSlot::Secondary(s)) => Some(s.clone()),
946 0 : _ => None,
947 : }
948 0 : }
949 :
950 : /// Whether the `TenantManager` is responsible for the tenant shard
951 0 : pub(crate) fn manages_tenant_shard(&self, tenant_shard_id: TenantShardId) -> bool {
952 0 : let locked = self.tenants.read().unwrap();
953 0 :
954 0 : let peek_slot = tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Read)
955 0 : .ok()
956 0 : .flatten();
957 0 :
958 0 : peek_slot.is_some()
959 0 : }
960 :
961 0 : #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
962 : pub(crate) async fn upsert_location(
963 : &self,
964 : tenant_shard_id: TenantShardId,
965 : new_location_config: LocationConf,
966 : flush: Option<Duration>,
967 : mut spawn_mode: SpawnMode,
968 : ctx: &RequestContext,
969 : ) -> Result<Option<Arc<Tenant>>, UpsertLocationError> {
970 : debug_assert_current_span_has_tenant_id();
971 : info!("configuring tenant location to state {new_location_config:?}");
972 :
973 : enum FastPathModified {
974 : Attached(Arc<Tenant>),
975 : Secondary(Arc<SecondaryTenant>),
976 : }
977 :
978 : // Special case fast-path for updates to existing slots: if our upsert is only updating configuration,
979 : // then we do not need to set the slot to InProgress, we can just call into the
980 : // existng tenant.
981 : let fast_path_taken = {
982 : let locked = self.tenants.read().unwrap();
983 : let peek_slot =
984 : tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Write)?;
985 : match (&new_location_config.mode, peek_slot) {
986 : (LocationMode::Attached(attach_conf), Some(TenantSlot::Attached(tenant))) => {
987 : match attach_conf.generation.cmp(&tenant.generation) {
988 : Ordering::Equal => {
989 : // A transition from Attached to Attached in the same generation, we may
990 : // take our fast path and just provide the updated configuration
991 : // to the tenant.
992 : tenant.set_new_location_config(
993 : AttachedTenantConf::try_from(new_location_config.clone())
994 : .map_err(UpsertLocationError::BadRequest)?,
995 : );
996 :
997 : Some(FastPathModified::Attached(tenant.clone()))
998 : }
999 : Ordering::Less => {
1000 : return Err(UpsertLocationError::BadRequest(anyhow::anyhow!(
1001 : "Generation {:?} is less than existing {:?}",
1002 : attach_conf.generation,
1003 : tenant.generation
1004 : )));
1005 : }
1006 : Ordering::Greater => {
1007 : // Generation advanced, fall through to general case of replacing `Tenant` object
1008 : None
1009 : }
1010 : }
1011 : }
1012 : (
1013 : LocationMode::Secondary(secondary_conf),
1014 : Some(TenantSlot::Secondary(secondary_tenant)),
1015 : ) => {
1016 : secondary_tenant.set_config(secondary_conf);
1017 : secondary_tenant.set_tenant_conf(&new_location_config.tenant_conf);
1018 : Some(FastPathModified::Secondary(secondary_tenant.clone()))
1019 : }
1020 : _ => {
1021 : // Not an Attached->Attached transition, fall through to general case
1022 : None
1023 : }
1024 : }
1025 : };
1026 :
1027 : // Fast-path continued: having dropped out of the self.tenants lock, do the async
1028 : // phase of writing config and/or waiting for flush, before returning.
1029 : match fast_path_taken {
1030 : Some(FastPathModified::Attached(tenant)) => {
1031 : Tenant::persist_tenant_config(self.conf, &tenant_shard_id, &new_location_config)
1032 : .await
1033 : .fatal_err("write tenant shard config");
1034 :
1035 : // Transition to AttachedStale means we may well hold a valid generation
1036 : // still, and have been requested to go stale as part of a migration. If
1037 : // the caller set `flush`, then flush to remote storage.
1038 : if let LocationMode::Attached(AttachedLocationConfig {
1039 : generation: _,
1040 : attach_mode: AttachmentMode::Stale,
1041 : }) = &new_location_config.mode
1042 : {
1043 : if let Some(flush_timeout) = flush {
1044 : match tokio::time::timeout(flush_timeout, tenant.flush_remote()).await {
1045 : Ok(Err(e)) => {
1046 : return Err(UpsertLocationError::Flush(e));
1047 : }
1048 : Ok(Ok(_)) => return Ok(Some(tenant)),
1049 : Err(_) => {
1050 : tracing::warn!(
1051 : timeout_ms = flush_timeout.as_millis(),
1052 : "Timed out waiting for flush to remote storage, proceeding anyway."
1053 : )
1054 : }
1055 : }
1056 : }
1057 : }
1058 :
1059 : return Ok(Some(tenant));
1060 : }
1061 : Some(FastPathModified::Secondary(_secondary_tenant)) => {
1062 : Tenant::persist_tenant_config(self.conf, &tenant_shard_id, &new_location_config)
1063 : .await
1064 : .fatal_err("write tenant shard config");
1065 :
1066 : return Ok(None);
1067 : }
1068 : None => {
1069 : // Proceed with the general case procedure, where we will shutdown & remove any existing
1070 : // slot contents and replace with a fresh one
1071 : }
1072 : };
1073 :
1074 : // General case for upserts to TenantsMap, excluding the case above: we will substitute an
1075 : // InProgress value to the slot while we make whatever changes are required. The state for
1076 : // the tenant is inaccessible to the outside world while we are doing this, but that is sensible:
1077 : // the state is ill-defined while we're in transition. Transitions are async, but fast: we do
1078 : // not do significant I/O, and shutdowns should be prompt via cancellation tokens.
1079 : let mut slot_guard = tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)
1080 0 : .map_err(|e| match e {
1081 : TenantSlotError::NotFound(_) => {
1082 0 : unreachable!("Called with mode Any")
1083 : }
1084 0 : TenantSlotError::InProgress => UpsertLocationError::InProgress,
1085 0 : TenantSlotError::MapState(s) => UpsertLocationError::Unavailable(s),
1086 0 : })?;
1087 :
1088 : match slot_guard.get_old_value() {
1089 : Some(TenantSlot::Attached(tenant)) => {
1090 : // The case where we keep a Tenant alive was covered above in the special case
1091 : // for Attached->Attached transitions in the same generation. By this point,
1092 : // if we see an attached tenant we know it will be discarded and should be
1093 : // shut down.
1094 : let (_guard, progress) = utils::completion::channel();
1095 :
1096 : match tenant.get_attach_mode() {
1097 : AttachmentMode::Single | AttachmentMode::Multi => {
1098 : // Before we leave our state as the presumed holder of the latest generation,
1099 : // flush any outstanding deletions to reduce the risk of leaking objects.
1100 : self.resources.deletion_queue_client.flush_advisory()
1101 : }
1102 : AttachmentMode::Stale => {
1103 : // If we're stale there's not point trying to flush deletions
1104 : }
1105 : };
1106 :
1107 : info!("Shutting down attached tenant");
1108 : match tenant.shutdown(progress, ShutdownMode::Hard).await {
1109 : Ok(()) => {}
1110 : Err(barrier) => {
1111 : info!("Shutdown already in progress, waiting for it to complete");
1112 : barrier.wait().await;
1113 : }
1114 : }
1115 : slot_guard.drop_old_value().expect("We just shut it down");
1116 :
1117 : // Edge case: if we were called with SpawnMode::Create, but a Tenant already existed, then
1118 : // the caller thinks they're creating but the tenant already existed. We must switch to
1119 : // Eager mode so that when starting this Tenant we properly probe remote storage for timelines,
1120 : // rather than assuming it to be empty.
1121 : spawn_mode = SpawnMode::Eager;
1122 : }
1123 : Some(TenantSlot::Secondary(state)) => {
1124 : info!("Shutting down secondary tenant");
1125 : state.shutdown().await;
1126 : }
1127 : Some(TenantSlot::InProgress(_)) => {
1128 : // This should never happen: acquire_slot should error out
1129 : // if the contents of a slot were InProgress.
1130 : return Err(UpsertLocationError::InternalError(anyhow::anyhow!(
1131 : "Acquired an InProgress slot, this is a bug."
1132 : )));
1133 : }
1134 : None => {
1135 : // Slot was vacant, nothing needs shutting down.
1136 : }
1137 : }
1138 :
1139 : let tenant_path = self.conf.tenant_path(&tenant_shard_id);
1140 : let timelines_path = self.conf.timelines_path(&tenant_shard_id);
1141 :
1142 : // Directory structure is the same for attached and secondary modes:
1143 : // create it if it doesn't exist. Timeline load/creation expects the
1144 : // timelines/ subdir to already exist.
1145 : //
1146 : // Does not need to be fsync'd because local storage is just a cache.
1147 : tokio::fs::create_dir_all(&timelines_path)
1148 : .await
1149 : .fatal_err("create timelines/ dir");
1150 :
1151 : // Before activating either secondary or attached mode, persist the
1152 : // configuration, so that on restart we will re-attach (or re-start
1153 : // secondary) on the tenant.
1154 : Tenant::persist_tenant_config(self.conf, &tenant_shard_id, &new_location_config)
1155 : .await
1156 : .fatal_err("write tenant shard config");
1157 :
1158 : let new_slot = match &new_location_config.mode {
1159 : LocationMode::Secondary(secondary_config) => {
1160 : let shard_identity = new_location_config.shard;
1161 : TenantSlot::Secondary(SecondaryTenant::new(
1162 : tenant_shard_id,
1163 : shard_identity,
1164 : new_location_config.tenant_conf,
1165 : secondary_config,
1166 : ))
1167 : }
1168 : LocationMode::Attached(_attach_config) => {
1169 : let shard_identity = new_location_config.shard;
1170 :
1171 : // Testing hack: if we are configured with no control plane, then drop the generation
1172 : // from upserts. This enables creating generation-less tenants even though neon_local
1173 : // always uses generations when calling the location conf API.
1174 : let attached_conf = if cfg!(feature = "testing") {
1175 : let mut conf = AttachedTenantConf::try_from(new_location_config)
1176 : .map_err(UpsertLocationError::BadRequest)?;
1177 : if self.conf.control_plane_api.is_none() {
1178 : conf.location.generation = Generation::none();
1179 : }
1180 : conf
1181 : } else {
1182 : AttachedTenantConf::try_from(new_location_config)
1183 : .map_err(UpsertLocationError::BadRequest)?
1184 : };
1185 :
1186 : let tenant = tenant_spawn(
1187 : self.conf,
1188 : tenant_shard_id,
1189 : &tenant_path,
1190 : self.resources.clone(),
1191 : attached_conf,
1192 : shard_identity,
1193 : None,
1194 : spawn_mode,
1195 : ctx,
1196 : )
1197 0 : .map_err(|_: GlobalShutDown| {
1198 0 : UpsertLocationError::Unavailable(TenantMapError::ShuttingDown)
1199 0 : })?;
1200 :
1201 : TenantSlot::Attached(tenant)
1202 : }
1203 : };
1204 :
1205 : let attached_tenant = if let TenantSlot::Attached(tenant) = &new_slot {
1206 : Some(tenant.clone())
1207 : } else {
1208 : None
1209 : };
1210 :
1211 : match slot_guard.upsert(new_slot) {
1212 : Err(TenantSlotUpsertError::InternalError(e)) => {
1213 : Err(UpsertLocationError::InternalError(anyhow::anyhow!(e)))
1214 : }
1215 : Err(TenantSlotUpsertError::MapState(e)) => Err(UpsertLocationError::Unavailable(e)),
1216 : Err(TenantSlotUpsertError::ShuttingDown((new_slot, _completion))) => {
1217 : // If we just called tenant_spawn() on a new tenant, and can't insert it into our map, then
1218 : // we must not leak it: this would violate the invariant that after shutdown_all_tenants, all tenants
1219 : // are shutdown.
1220 : //
1221 : // We must shut it down inline here.
1222 : match new_slot {
1223 : TenantSlot::InProgress(_) => {
1224 : // Unreachable because we never insert an InProgress
1225 : unreachable!()
1226 : }
1227 : TenantSlot::Attached(tenant) => {
1228 : let (_guard, progress) = utils::completion::channel();
1229 : info!("Shutting down just-spawned tenant, because tenant manager is shut down");
1230 : match tenant.shutdown(progress, ShutdownMode::Hard).await {
1231 : Ok(()) => {
1232 : info!("Finished shutting down just-spawned tenant");
1233 : }
1234 : Err(barrier) => {
1235 : info!("Shutdown already in progress, waiting for it to complete");
1236 : barrier.wait().await;
1237 : }
1238 : }
1239 : }
1240 : TenantSlot::Secondary(secondary_tenant) => {
1241 : secondary_tenant.shutdown().await;
1242 : }
1243 : }
1244 :
1245 : Err(UpsertLocationError::Unavailable(
1246 : TenantMapError::ShuttingDown,
1247 : ))
1248 : }
1249 : Ok(()) => Ok(attached_tenant),
1250 : }
1251 : }
1252 :
1253 : /// Resetting a tenant is equivalent to detaching it, then attaching it again with the same
1254 : /// LocationConf that was last used to attach it. Optionally, the local file cache may be
1255 : /// dropped before re-attaching.
1256 : ///
1257 : /// This is not part of a tenant's normal lifecycle: it is used for debug/support, in situations
1258 : /// where an issue is identified that would go away with a restart of the tenant.
1259 : ///
1260 : /// This does not have any special "force" shutdown of a tenant: it relies on the tenant's tasks
1261 : /// to respect the cancellation tokens used in normal shutdown().
1262 0 : #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), %drop_cache))]
1263 : pub(crate) async fn reset_tenant(
1264 : &self,
1265 : tenant_shard_id: TenantShardId,
1266 : drop_cache: bool,
1267 : ctx: &RequestContext,
1268 : ) -> anyhow::Result<()> {
1269 : let mut slot_guard = tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)?;
1270 : let Some(old_slot) = slot_guard.get_old_value() else {
1271 : anyhow::bail!("Tenant not found when trying to reset");
1272 : };
1273 :
1274 : let Some(tenant) = old_slot.get_attached() else {
1275 : slot_guard.revert();
1276 : anyhow::bail!("Tenant is not in attached state");
1277 : };
1278 :
1279 : let (_guard, progress) = utils::completion::channel();
1280 : match tenant.shutdown(progress, ShutdownMode::Hard).await {
1281 : Ok(()) => {
1282 : slot_guard.drop_old_value()?;
1283 : }
1284 : Err(_barrier) => {
1285 : slot_guard.revert();
1286 : anyhow::bail!("Cannot reset Tenant, already shutting down");
1287 : }
1288 : }
1289 :
1290 : let tenant_path = self.conf.tenant_path(&tenant_shard_id);
1291 : let timelines_path = self.conf.timelines_path(&tenant_shard_id);
1292 : let config = Tenant::load_tenant_config(self.conf, &tenant_shard_id)?;
1293 :
1294 : if drop_cache {
1295 : tracing::info!("Dropping local file cache");
1296 :
1297 : match tokio::fs::read_dir(&timelines_path).await {
1298 : Err(e) => {
1299 : tracing::warn!("Failed to list timelines while dropping cache: {}", e);
1300 : }
1301 : Ok(mut entries) => {
1302 : while let Some(entry) = entries.next_entry().await? {
1303 : tokio::fs::remove_dir_all(entry.path()).await?;
1304 : }
1305 : }
1306 : }
1307 : }
1308 :
1309 : let shard_identity = config.shard;
1310 : let tenant = tenant_spawn(
1311 : self.conf,
1312 : tenant_shard_id,
1313 : &tenant_path,
1314 : self.resources.clone(),
1315 : AttachedTenantConf::try_from(config)?,
1316 : shard_identity,
1317 : None,
1318 : SpawnMode::Eager,
1319 : ctx,
1320 : )?;
1321 :
1322 : slot_guard.upsert(TenantSlot::Attached(tenant))?;
1323 :
1324 : Ok(())
1325 : }
1326 :
1327 0 : pub(crate) fn get_attached_active_tenant_shards(&self) -> Vec<Arc<Tenant>> {
1328 0 : let locked = self.tenants.read().unwrap();
1329 0 : match &*locked {
1330 0 : TenantsMap::Initializing => Vec::new(),
1331 0 : TenantsMap::Open(map) | TenantsMap::ShuttingDown(map) => map
1332 0 : .values()
1333 0 : .filter_map(|slot| {
1334 0 : slot.get_attached()
1335 0 : .and_then(|t| if t.is_active() { Some(t.clone()) } else { None })
1336 0 : })
1337 0 : .collect(),
1338 : }
1339 0 : }
1340 : // Do some synchronous work for all tenant slots in Secondary state. The provided
1341 : // callback should be small and fast, as it will be called inside the global
1342 : // TenantsMap lock.
1343 0 : pub(crate) fn foreach_secondary_tenants<F>(&self, mut func: F)
1344 0 : where
1345 0 : // TODO: let the callback return a hint to drop out of the loop early
1346 0 : F: FnMut(&TenantShardId, &Arc<SecondaryTenant>),
1347 0 : {
1348 0 : let locked = self.tenants.read().unwrap();
1349 :
1350 0 : let map = match &*locked {
1351 0 : TenantsMap::Initializing | TenantsMap::ShuttingDown(_) => return,
1352 0 : TenantsMap::Open(m) => m,
1353 : };
1354 :
1355 0 : for (tenant_id, slot) in map {
1356 0 : if let TenantSlot::Secondary(state) = slot {
1357 : // Only expose secondary tenants that are not currently shutting down
1358 0 : if !state.cancel.is_cancelled() {
1359 0 : func(tenant_id, state)
1360 0 : }
1361 0 : }
1362 : }
1363 0 : }
1364 :
1365 : /// Total list of all tenant slots: this includes attached, secondary, and InProgress.
1366 0 : pub(crate) fn list(&self) -> Vec<(TenantShardId, TenantSlot)> {
1367 0 : let locked = self.tenants.read().unwrap();
1368 0 : match &*locked {
1369 0 : TenantsMap::Initializing => Vec::new(),
1370 0 : TenantsMap::Open(map) | TenantsMap::ShuttingDown(map) => {
1371 0 : map.iter().map(|(k, v)| (*k, v.clone())).collect()
1372 : }
1373 : }
1374 0 : }
1375 :
1376 0 : pub(crate) fn get(&self, tenant_shard_id: TenantShardId) -> Option<TenantSlot> {
1377 0 : let locked = self.tenants.read().unwrap();
1378 0 : match &*locked {
1379 0 : TenantsMap::Initializing => None,
1380 0 : TenantsMap::Open(map) | TenantsMap::ShuttingDown(map) => {
1381 0 : map.get(&tenant_shard_id).cloned()
1382 : }
1383 : }
1384 0 : }
1385 :
1386 0 : async fn delete_tenant_remote(
1387 0 : &self,
1388 0 : tenant_shard_id: TenantShardId,
1389 0 : ) -> Result<(), DeleteTenantError> {
1390 0 : let remote_path = remote_tenant_path(&tenant_shard_id);
1391 0 : let mut keys_stream = self.resources.remote_storage.list_streaming(
1392 0 : Some(&remote_path),
1393 0 : remote_storage::ListingMode::NoDelimiter,
1394 0 : None,
1395 0 : &self.cancel,
1396 0 : );
1397 0 : while let Some(chunk) = keys_stream.next().await {
1398 0 : let keys = match chunk {
1399 0 : Ok(listing) => listing.keys,
1400 : Err(remote_storage::DownloadError::Cancelled) => {
1401 0 : return Err(DeleteTenantError::Cancelled)
1402 : }
1403 0 : Err(remote_storage::DownloadError::NotFound) => return Ok(()),
1404 0 : Err(other) => return Err(DeleteTenantError::Other(anyhow::anyhow!(other))),
1405 : };
1406 :
1407 0 : if keys.is_empty() {
1408 0 : tracing::info!("Remote storage already deleted");
1409 : } else {
1410 0 : tracing::info!("Deleting {} keys from remote storage", keys.len());
1411 0 : let keys = keys.into_iter().map(|o| o.key).collect::<Vec<_>>();
1412 0 : self.resources
1413 0 : .remote_storage
1414 0 : .delete_objects(&keys, &self.cancel)
1415 0 : .await?;
1416 : }
1417 : }
1418 :
1419 0 : Ok(())
1420 0 : }
1421 :
1422 : /// If a tenant is attached, detach it. Then remove its data from remote storage.
1423 : ///
1424 : /// A tenant is considered deleted once it is gone from remote storage. It is the caller's
1425 : /// responsibility to avoid trying to attach the tenant again or use it any way once deletion
1426 : /// has started: this operation is not atomic, and must be retried until it succeeds.
1427 0 : pub(crate) async fn delete_tenant(
1428 0 : &self,
1429 0 : tenant_shard_id: TenantShardId,
1430 0 : ) -> Result<(), DeleteTenantError> {
1431 0 : super::span::debug_assert_current_span_has_tenant_id();
1432 :
1433 0 : async fn delete_local(
1434 0 : conf: &PageServerConf,
1435 0 : background_purges: &BackgroundPurges,
1436 0 : tenant_shard_id: &TenantShardId,
1437 0 : ) -> anyhow::Result<()> {
1438 0 : let local_tenant_directory = conf.tenant_path(tenant_shard_id);
1439 0 : let tmp_dir = safe_rename_tenant_dir(&local_tenant_directory)
1440 0 : .await
1441 0 : .with_context(|| {
1442 0 : format!("local tenant directory {local_tenant_directory:?} rename")
1443 0 : })?;
1444 0 : background_purges.spawn(tmp_dir);
1445 0 : Ok(())
1446 0 : }
1447 :
1448 0 : let slot_guard = tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)?;
1449 0 : match &slot_guard.old_value {
1450 0 : Some(TenantSlot::Attached(tenant)) => {
1451 0 : // Legacy deletion flow: the tenant remains attached, goes to Stopping state, and
1452 0 : // deletion will be resumed across restarts.
1453 0 : let tenant = tenant.clone();
1454 0 : let (_guard, progress) = utils::completion::channel();
1455 0 : match tenant.shutdown(progress, ShutdownMode::Hard).await {
1456 0 : Ok(()) => {}
1457 0 : Err(barrier) => {
1458 0 : info!("Shutdown already in progress, waiting for it to complete");
1459 0 : barrier.wait().await;
1460 : }
1461 : }
1462 0 : delete_local(self.conf, &self.background_purges, &tenant_shard_id).await?;
1463 : }
1464 0 : Some(TenantSlot::Secondary(secondary_tenant)) => {
1465 0 : secondary_tenant.shutdown().await;
1466 :
1467 0 : delete_local(self.conf, &self.background_purges, &tenant_shard_id).await?;
1468 : }
1469 0 : Some(TenantSlot::InProgress(_)) => unreachable!(),
1470 0 : None => {}
1471 : };
1472 :
1473 : // Fall through: local state for this tenant is no longer present, proceed with remote delete.
1474 : // - We use a retry wrapper here so that common transient S3 errors (e.g. 503, 429) do not result
1475 : // in 500 responses to delete requests.
1476 : // - We keep the `SlotGuard` during this I/O, so that if a concurrent delete request comes in, it will
1477 : // 503/retry, rather than kicking off a wasteful concurrent deletion.
1478 0 : match backoff::retry(
1479 0 : || async move { self.delete_tenant_remote(tenant_shard_id).await },
1480 0 : |e| match e {
1481 0 : DeleteTenantError::Cancelled => true,
1482 : DeleteTenantError::SlotError(_) => {
1483 0 : unreachable!("Remote deletion doesn't touch slots")
1484 : }
1485 0 : _ => false,
1486 0 : },
1487 0 : 1,
1488 0 : 3,
1489 0 : &format!("delete_tenant[tenant_shard_id={tenant_shard_id}]"),
1490 0 : &self.cancel,
1491 0 : )
1492 0 : .await
1493 : {
1494 0 : Some(r) => r,
1495 0 : None => Err(DeleteTenantError::Cancelled),
1496 : }
1497 0 : }
1498 :
1499 0 : #[instrument(skip_all, fields(tenant_id=%tenant.get_tenant_shard_id().tenant_id, shard_id=%tenant.get_tenant_shard_id().shard_slug(), new_shard_count=%new_shard_count.literal()))]
1500 : pub(crate) async fn shard_split(
1501 : &self,
1502 : tenant: Arc<Tenant>,
1503 : new_shard_count: ShardCount,
1504 : new_stripe_size: Option<ShardStripeSize>,
1505 : ctx: &RequestContext,
1506 : ) -> anyhow::Result<Vec<TenantShardId>> {
1507 : let tenant_shard_id = *tenant.get_tenant_shard_id();
1508 : let r = self
1509 : .do_shard_split(tenant, new_shard_count, new_stripe_size, ctx)
1510 : .await;
1511 : if r.is_err() {
1512 : // Shard splitting might have left the original shard in a partially shut down state (it
1513 : // stops the shard's remote timeline client). Reset it to ensure we leave things in
1514 : // a working state.
1515 : if self.get(tenant_shard_id).is_some() {
1516 : tracing::warn!("Resetting after shard split failure");
1517 : if let Err(e) = self.reset_tenant(tenant_shard_id, false, ctx).await {
1518 : // Log this error because our return value will still be the original error, not this one. This is
1519 : // a severe error: if this happens, we might be leaving behind a tenant that is not fully functional
1520 : // (e.g. has uploads disabled). We can't do anything else: if reset fails then shutting the tenant down or
1521 : // setting it broken probably won't help either.
1522 : tracing::error!("Failed to reset: {e}");
1523 : }
1524 : }
1525 : }
1526 :
1527 : r
1528 : }
1529 :
1530 0 : pub(crate) async fn do_shard_split(
1531 0 : &self,
1532 0 : tenant: Arc<Tenant>,
1533 0 : new_shard_count: ShardCount,
1534 0 : new_stripe_size: Option<ShardStripeSize>,
1535 0 : ctx: &RequestContext,
1536 0 : ) -> anyhow::Result<Vec<TenantShardId>> {
1537 0 : let tenant_shard_id = *tenant.get_tenant_shard_id();
1538 0 :
1539 0 : // Validate the incoming request
1540 0 : if new_shard_count.count() <= tenant_shard_id.shard_count.count() {
1541 0 : anyhow::bail!("Requested shard count is not an increase");
1542 0 : }
1543 0 : let expansion_factor = new_shard_count.count() / tenant_shard_id.shard_count.count();
1544 0 : if !expansion_factor.is_power_of_two() {
1545 0 : anyhow::bail!("Requested split is not a power of two");
1546 0 : }
1547 :
1548 0 : if let Some(new_stripe_size) = new_stripe_size {
1549 0 : if tenant.get_shard_stripe_size() != new_stripe_size
1550 0 : && tenant_shard_id.shard_count.count() > 1
1551 : {
1552 : // This tenant already has multiple shards, it is illegal to try and change its stripe size
1553 0 : anyhow::bail!(
1554 0 : "Shard stripe size may not be modified once tenant has multiple shards"
1555 0 : );
1556 0 : }
1557 0 : }
1558 :
1559 : // Plan: identify what the new child shards will be
1560 0 : let child_shards = tenant_shard_id.split(new_shard_count);
1561 0 : tracing::info!(
1562 0 : "Shard {} splits into: {}",
1563 0 : tenant_shard_id.to_index(),
1564 0 : child_shards
1565 0 : .iter()
1566 0 : .map(|id| format!("{}", id.to_index()))
1567 0 : .join(",")
1568 : );
1569 :
1570 0 : fail::fail_point!("shard-split-pre-prepare", |_| Err(anyhow::anyhow!(
1571 0 : "failpoint"
1572 0 : )));
1573 :
1574 0 : let parent_shard_identity = tenant.shard_identity;
1575 0 : let parent_tenant_conf = tenant.get_tenant_conf();
1576 0 : let parent_generation = tenant.generation;
1577 :
1578 : // Phase 1: Write out child shards' remote index files, in the parent tenant's current generation
1579 0 : if let Err(e) = tenant.split_prepare(&child_shards).await {
1580 : // If [`Tenant::split_prepare`] fails, we must reload the tenant, because it might
1581 : // have been left in a partially-shut-down state.
1582 0 : tracing::warn!("Failed to prepare for split: {e}, reloading Tenant before returning");
1583 0 : return Err(e);
1584 0 : }
1585 0 :
1586 0 : fail::fail_point!("shard-split-post-prepare", |_| Err(anyhow::anyhow!(
1587 0 : "failpoint"
1588 0 : )));
1589 :
1590 0 : self.resources.deletion_queue_client.flush_advisory();
1591 0 :
1592 0 : // Phase 2: Put the parent shard to InProgress and grab a reference to the parent Tenant
1593 0 : drop(tenant);
1594 0 : let mut parent_slot_guard =
1595 0 : tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)?;
1596 0 : let parent = match parent_slot_guard.get_old_value() {
1597 0 : Some(TenantSlot::Attached(t)) => t,
1598 0 : Some(TenantSlot::Secondary(_)) => anyhow::bail!("Tenant location in secondary mode"),
1599 : Some(TenantSlot::InProgress(_)) => {
1600 : // tenant_map_acquire_slot never returns InProgress, if a slot was InProgress
1601 : // it would return an error.
1602 0 : unreachable!()
1603 : }
1604 : None => {
1605 : // We don't actually need the parent shard to still be attached to do our work, but it's
1606 : // a weird enough situation that the caller probably didn't want us to continue working
1607 : // if they had detached the tenant they requested the split on.
1608 0 : anyhow::bail!("Detached parent shard in the middle of split!")
1609 : }
1610 : };
1611 0 : fail::fail_point!("shard-split-pre-hardlink", |_| Err(anyhow::anyhow!(
1612 0 : "failpoint"
1613 0 : )));
1614 : // Optimization: hardlink layers from the parent into the children, so that they don't have to
1615 : // re-download & duplicate the data referenced in their initial IndexPart
1616 0 : self.shard_split_hardlink(parent, child_shards.clone())
1617 0 : .await?;
1618 0 : fail::fail_point!("shard-split-post-hardlink", |_| Err(anyhow::anyhow!(
1619 0 : "failpoint"
1620 0 : )));
1621 :
1622 : // Take a snapshot of where the parent's WAL ingest had got to: we will wait for
1623 : // child shards to reach this point.
1624 0 : let mut target_lsns = HashMap::new();
1625 0 : for timeline in parent.timelines.lock().unwrap().clone().values() {
1626 0 : target_lsns.insert(timeline.timeline_id, timeline.get_last_record_lsn());
1627 0 : }
1628 :
1629 : // TODO: we should have the parent shard stop its WAL ingest here, it's a waste of resources
1630 : // and could slow down the children trying to catch up.
1631 :
1632 : // Phase 3: Spawn the child shards
1633 0 : for child_shard in &child_shards {
1634 0 : let mut child_shard_identity = parent_shard_identity;
1635 0 : if let Some(new_stripe_size) = new_stripe_size {
1636 0 : child_shard_identity.stripe_size = new_stripe_size;
1637 0 : }
1638 0 : child_shard_identity.count = child_shard.shard_count;
1639 0 : child_shard_identity.number = child_shard.shard_number;
1640 0 :
1641 0 : let child_location_conf = LocationConf {
1642 0 : mode: LocationMode::Attached(AttachedLocationConfig {
1643 0 : generation: parent_generation,
1644 0 : attach_mode: AttachmentMode::Single,
1645 0 : }),
1646 0 : shard: child_shard_identity,
1647 0 : tenant_conf: parent_tenant_conf.clone(),
1648 0 : };
1649 0 :
1650 0 : self.upsert_location(
1651 0 : *child_shard,
1652 0 : child_location_conf,
1653 0 : None,
1654 0 : SpawnMode::Eager,
1655 0 : ctx,
1656 0 : )
1657 0 : .await?;
1658 : }
1659 :
1660 0 : fail::fail_point!("shard-split-post-child-conf", |_| Err(anyhow::anyhow!(
1661 0 : "failpoint"
1662 0 : )));
1663 :
1664 : // Phase 4: wait for child chards WAL ingest to catch up to target LSN
1665 0 : for child_shard_id in &child_shards {
1666 0 : let child_shard_id = *child_shard_id;
1667 0 : let child_shard = {
1668 0 : let locked = self.tenants.read().unwrap();
1669 0 : let peek_slot =
1670 0 : tenant_map_peek_slot(&locked, &child_shard_id, TenantSlotPeekMode::Read)?;
1671 0 : peek_slot.and_then(|s| s.get_attached()).cloned()
1672 : };
1673 0 : if let Some(t) = child_shard {
1674 : // Wait for the child shard to become active: this should be very quick because it only
1675 : // has to download the index_part that we just uploaded when creating it.
1676 0 : if let Err(e) = t.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await {
1677 : // This is not fatal: we have durably created the child shard. It just makes the
1678 : // split operation less seamless for clients, as we will may detach the parent
1679 : // shard before the child shards are fully ready to serve requests.
1680 0 : tracing::warn!("Failed to wait for shard {child_shard_id} to activate: {e}");
1681 0 : continue;
1682 0 : }
1683 0 :
1684 0 : let timelines = t.timelines.lock().unwrap().clone();
1685 0 : for timeline in timelines.values() {
1686 0 : let Some(target_lsn) = target_lsns.get(&timeline.timeline_id) else {
1687 0 : continue;
1688 : };
1689 :
1690 0 : tracing::info!(
1691 0 : "Waiting for child shard {}/{} to reach target lsn {}...",
1692 0 : child_shard_id,
1693 0 : timeline.timeline_id,
1694 : target_lsn
1695 : );
1696 :
1697 0 : fail::fail_point!("shard-split-lsn-wait", |_| Err(anyhow::anyhow!(
1698 0 : "failpoint"
1699 0 : )));
1700 0 : if let Err(e) = timeline
1701 0 : .wait_lsn(
1702 0 : *target_lsn,
1703 0 : crate::tenant::timeline::WaitLsnWaiter::Tenant,
1704 0 : ctx,
1705 0 : )
1706 0 : .await
1707 : {
1708 : // Failure here might mean shutdown, in any case this part is an optimization
1709 : // and we shouldn't hold up the split operation.
1710 0 : tracing::warn!(
1711 0 : "Failed to wait for timeline {} to reach lsn {target_lsn}: {e}",
1712 0 : timeline.timeline_id
1713 : );
1714 : } else {
1715 0 : tracing::info!(
1716 0 : "Child shard {}/{} reached target lsn {}",
1717 0 : child_shard_id,
1718 0 : timeline.timeline_id,
1719 : target_lsn
1720 : );
1721 : }
1722 : }
1723 0 : }
1724 : }
1725 :
1726 : // Phase 5: Shut down the parent shard, and erase it from disk
1727 0 : let (_guard, progress) = completion::channel();
1728 0 : match parent.shutdown(progress, ShutdownMode::Hard).await {
1729 0 : Ok(()) => {}
1730 0 : Err(other) => {
1731 0 : other.wait().await;
1732 : }
1733 : }
1734 0 : let local_tenant_directory = self.conf.tenant_path(&tenant_shard_id);
1735 0 : let tmp_path = safe_rename_tenant_dir(&local_tenant_directory)
1736 0 : .await
1737 0 : .with_context(|| format!("local tenant directory {local_tenant_directory:?} rename"))?;
1738 0 : self.background_purges.spawn(tmp_path);
1739 0 :
1740 0 : fail::fail_point!("shard-split-pre-finish", |_| Err(anyhow::anyhow!(
1741 0 : "failpoint"
1742 0 : )));
1743 :
1744 0 : parent_slot_guard.drop_old_value()?;
1745 :
1746 : // Phase 6: Release the InProgress on the parent shard
1747 0 : drop(parent_slot_guard);
1748 0 :
1749 0 : Ok(child_shards)
1750 0 : }
1751 :
1752 : /// Part of [`Self::shard_split`]: hard link parent shard layers into child shards, as an optimization
1753 : /// to avoid the children downloading them again.
1754 : ///
1755 : /// For each resident layer in the parent shard, we will hard link it into all of the child shards.
1756 0 : async fn shard_split_hardlink(
1757 0 : &self,
1758 0 : parent_shard: &Tenant,
1759 0 : child_shards: Vec<TenantShardId>,
1760 0 : ) -> anyhow::Result<()> {
1761 0 : debug_assert_current_span_has_tenant_id();
1762 0 :
1763 0 : let parent_path = self.conf.tenant_path(parent_shard.get_tenant_shard_id());
1764 0 : let (parent_timelines, parent_layers) = {
1765 0 : let mut parent_layers = Vec::new();
1766 0 : let timelines = parent_shard.timelines.lock().unwrap().clone();
1767 0 : let parent_timelines = timelines.keys().cloned().collect::<Vec<_>>();
1768 0 : for timeline in timelines.values() {
1769 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Loading list of layers to hardlink");
1770 0 : let timeline_layers = timeline
1771 0 : .layers
1772 0 : .read()
1773 0 : .await
1774 0 : .likely_resident_layers()
1775 0 : .collect::<Vec<_>>();
1776 :
1777 0 : for layer in timeline_layers {
1778 0 : let relative_path = layer
1779 0 : .local_path()
1780 0 : .strip_prefix(&parent_path)
1781 0 : .context("Removing prefix from parent layer path")?;
1782 0 : parent_layers.push(relative_path.to_owned());
1783 : }
1784 : }
1785 0 : debug_assert!(
1786 0 : !parent_layers.is_empty(),
1787 0 : "shutdown cannot empty the layermap"
1788 : );
1789 0 : (parent_timelines, parent_layers)
1790 0 : };
1791 0 :
1792 0 : let mut child_prefixes = Vec::new();
1793 0 : let mut create_dirs = Vec::new();
1794 :
1795 0 : for child in child_shards {
1796 0 : let child_prefix = self.conf.tenant_path(&child);
1797 0 : create_dirs.push(child_prefix.clone());
1798 0 : create_dirs.extend(
1799 0 : parent_timelines
1800 0 : .iter()
1801 0 : .map(|t| self.conf.timeline_path(&child, t)),
1802 0 : );
1803 0 :
1804 0 : child_prefixes.push(child_prefix);
1805 0 : }
1806 :
1807 : // Since we will do a large number of small filesystem metadata operations, batch them into
1808 : // spawn_blocking calls rather than doing each one as a tokio::fs round-trip.
1809 0 : let span = tracing::Span::current();
1810 0 : let jh = tokio::task::spawn_blocking(move || -> anyhow::Result<usize> {
1811 0 : // Run this synchronous code in the same log context as the outer function that spawned it.
1812 0 : let _span = span.enter();
1813 0 :
1814 0 : tracing::info!("Creating {} directories", create_dirs.len());
1815 0 : for dir in &create_dirs {
1816 0 : if let Err(e) = std::fs::create_dir_all(dir) {
1817 : // Ignore AlreadyExists errors, drop out on all other errors
1818 0 : match e.kind() {
1819 0 : std::io::ErrorKind::AlreadyExists => {}
1820 : _ => {
1821 0 : return Err(anyhow::anyhow!(e).context(format!("Creating {dir}")));
1822 : }
1823 : }
1824 0 : }
1825 : }
1826 :
1827 0 : for child_prefix in child_prefixes {
1828 0 : tracing::info!(
1829 0 : "Hard-linking {} parent layers into child path {}",
1830 0 : parent_layers.len(),
1831 : child_prefix
1832 : );
1833 0 : for relative_layer in &parent_layers {
1834 0 : let parent_path = parent_path.join(relative_layer);
1835 0 : let child_path = child_prefix.join(relative_layer);
1836 0 : if let Err(e) = std::fs::hard_link(&parent_path, &child_path) {
1837 0 : match e.kind() {
1838 0 : std::io::ErrorKind::AlreadyExists => {}
1839 : std::io::ErrorKind::NotFound => {
1840 0 : tracing::info!(
1841 0 : "Layer {} not found during hard-linking, evicted during split?",
1842 : relative_layer
1843 : );
1844 : }
1845 : _ => {
1846 0 : return Err(anyhow::anyhow!(e).context(format!(
1847 0 : "Hard linking {relative_layer} into {child_prefix}"
1848 0 : )))
1849 : }
1850 : }
1851 0 : }
1852 : }
1853 : }
1854 :
1855 : // Durability is not required for correctness, but if we crashed during split and
1856 : // then came restarted with empty timeline dirs, it would be very inefficient to
1857 : // re-populate from remote storage.
1858 0 : tracing::info!("fsyncing {} directories", create_dirs.len());
1859 0 : for dir in create_dirs {
1860 0 : if let Err(e) = crashsafe::fsync(&dir) {
1861 : // Something removed a newly created timeline dir out from underneath us? Extremely
1862 : // unexpected, but not worth panic'ing over as this whole function is just an
1863 : // optimization.
1864 0 : tracing::warn!("Failed to fsync directory {dir}: {e}")
1865 0 : }
1866 : }
1867 :
1868 0 : Ok(parent_layers.len())
1869 0 : });
1870 0 :
1871 0 : match jh.await {
1872 0 : Ok(Ok(layer_count)) => {
1873 0 : tracing::info!(count = layer_count, "Hard linked layers into child shards");
1874 : }
1875 0 : Ok(Err(e)) => {
1876 0 : // This is an optimization, so we tolerate failure.
1877 0 : tracing::warn!("Error hard-linking layers, proceeding anyway: {e}")
1878 : }
1879 0 : Err(e) => {
1880 0 : // This is something totally unexpected like a panic, so bail out.
1881 0 : anyhow::bail!("Error joining hard linking task: {e}");
1882 : }
1883 : }
1884 :
1885 0 : Ok(())
1886 0 : }
1887 :
1888 : ///
1889 : /// Shut down all tenants. This runs as part of pageserver shutdown.
1890 : ///
1891 : /// NB: We leave the tenants in the map, so that they remain accessible through
1892 : /// the management API until we shut it down. If we removed the shut-down tenants
1893 : /// from the tenants map, the management API would return 404 for these tenants,
1894 : /// because TenantsMap::get() now returns `None`.
1895 : /// That could be easily misinterpreted by control plane, the consumer of the
1896 : /// management API. For example, it could attach the tenant on a different pageserver.
1897 : /// We would then be in split-brain once this pageserver restarts.
1898 0 : #[instrument(skip_all)]
1899 : pub(crate) async fn shutdown(&self) {
1900 : self.cancel.cancel();
1901 :
1902 : shutdown_all_tenants0(self.tenants).await
1903 : }
1904 :
1905 0 : pub(crate) async fn detach_tenant(
1906 0 : &self,
1907 0 : conf: &'static PageServerConf,
1908 0 : tenant_shard_id: TenantShardId,
1909 0 : deletion_queue_client: &DeletionQueueClient,
1910 0 : ) -> Result<(), TenantStateError> {
1911 0 : let tmp_path = self
1912 0 : .detach_tenant0(conf, tenant_shard_id, deletion_queue_client)
1913 0 : .await?;
1914 0 : self.background_purges.spawn(tmp_path);
1915 0 :
1916 0 : Ok(())
1917 0 : }
1918 :
1919 0 : async fn detach_tenant0(
1920 0 : &self,
1921 0 : conf: &'static PageServerConf,
1922 0 : tenant_shard_id: TenantShardId,
1923 0 : deletion_queue_client: &DeletionQueueClient,
1924 0 : ) -> Result<Utf8PathBuf, TenantStateError> {
1925 0 : let tenant_dir_rename_operation = |tenant_id_to_clean: TenantShardId| async move {
1926 0 : let local_tenant_directory = conf.tenant_path(&tenant_id_to_clean);
1927 0 : safe_rename_tenant_dir(&local_tenant_directory)
1928 0 : .await
1929 0 : .with_context(|| {
1930 0 : format!("local tenant directory {local_tenant_directory:?} rename")
1931 0 : })
1932 0 : };
1933 :
1934 0 : let removal_result = remove_tenant_from_memory(
1935 0 : self.tenants,
1936 0 : tenant_shard_id,
1937 0 : tenant_dir_rename_operation(tenant_shard_id),
1938 0 : )
1939 0 : .await;
1940 :
1941 : // Flush pending deletions, so that they have a good chance of passing validation
1942 : // before this tenant is potentially re-attached elsewhere.
1943 0 : deletion_queue_client.flush_advisory();
1944 0 :
1945 0 : removal_result
1946 0 : }
1947 :
1948 0 : pub(crate) fn list_tenants(
1949 0 : &self,
1950 0 : ) -> Result<Vec<(TenantShardId, TenantState, Generation)>, TenantMapListError> {
1951 0 : let tenants = self.tenants.read().unwrap();
1952 0 : let m = match &*tenants {
1953 0 : TenantsMap::Initializing => return Err(TenantMapListError::Initializing),
1954 0 : TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => m,
1955 0 : };
1956 0 : Ok(m.iter()
1957 0 : .filter_map(|(id, tenant)| match tenant {
1958 0 : TenantSlot::Attached(tenant) => {
1959 0 : Some((*id, tenant.current_state(), tenant.generation()))
1960 : }
1961 0 : TenantSlot::Secondary(_) => None,
1962 0 : TenantSlot::InProgress(_) => None,
1963 0 : })
1964 0 : .collect())
1965 0 : }
1966 :
1967 : /// Completes an earlier prepared timeline detach ancestor.
1968 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
1969 0 : &self,
1970 0 : tenant_shard_id: TenantShardId,
1971 0 : timeline_id: TimelineId,
1972 0 : prepared: PreparedTimelineDetach,
1973 0 : ctx: &RequestContext,
1974 0 : ) -> Result<Vec<TimelineId>, anyhow::Error> {
1975 : struct RevertOnDropSlot(Option<SlotGuard>);
1976 :
1977 : impl Drop for RevertOnDropSlot {
1978 0 : fn drop(&mut self) {
1979 0 : if let Some(taken) = self.0.take() {
1980 0 : taken.revert();
1981 0 : }
1982 0 : }
1983 : }
1984 :
1985 : impl RevertOnDropSlot {
1986 0 : fn into_inner(mut self) -> SlotGuard {
1987 0 : self.0.take().unwrap()
1988 0 : }
1989 : }
1990 :
1991 : impl std::ops::Deref for RevertOnDropSlot {
1992 : type Target = SlotGuard;
1993 :
1994 0 : fn deref(&self) -> &Self::Target {
1995 0 : self.0.as_ref().unwrap()
1996 0 : }
1997 : }
1998 :
1999 0 : let slot_guard = tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)?;
2000 0 : let slot_guard = RevertOnDropSlot(Some(slot_guard));
2001 :
2002 0 : let tenant = {
2003 0 : let Some(old_slot) = slot_guard.get_old_value() else {
2004 0 : anyhow::bail!(
2005 0 : "Tenant not found when trying to complete detaching timeline ancestor"
2006 0 : );
2007 : };
2008 :
2009 0 : let Some(tenant) = old_slot.get_attached() else {
2010 0 : anyhow::bail!("Tenant is not in attached state");
2011 : };
2012 :
2013 0 : if !tenant.is_active() {
2014 0 : anyhow::bail!("Tenant is not active");
2015 0 : }
2016 0 :
2017 0 : tenant.clone()
2018 : };
2019 :
2020 0 : let timeline = tenant.get_timeline(timeline_id, true)?;
2021 :
2022 0 : let reparented = timeline
2023 0 : .complete_detaching_timeline_ancestor(&tenant, prepared, ctx)
2024 0 : .await?;
2025 :
2026 0 : let mut slot_guard = slot_guard.into_inner();
2027 0 :
2028 0 : let (_guard, progress) = utils::completion::channel();
2029 0 : match tenant.shutdown(progress, ShutdownMode::Hard).await {
2030 : Ok(()) => {
2031 0 : slot_guard.drop_old_value()?;
2032 : }
2033 0 : Err(_barrier) => {
2034 0 : slot_guard.revert();
2035 0 : // this really should not happen, at all, unless shutdown was already going?
2036 0 : anyhow::bail!("Cannot restart Tenant, already shutting down");
2037 : }
2038 : }
2039 :
2040 0 : let tenant_path = self.conf.tenant_path(&tenant_shard_id);
2041 0 : let config = Tenant::load_tenant_config(self.conf, &tenant_shard_id)?;
2042 :
2043 0 : let shard_identity = config.shard;
2044 0 : let tenant = tenant_spawn(
2045 0 : self.conf,
2046 0 : tenant_shard_id,
2047 0 : &tenant_path,
2048 0 : self.resources.clone(),
2049 0 : AttachedTenantConf::try_from(config)?,
2050 0 : shard_identity,
2051 0 : None,
2052 0 : SpawnMode::Eager,
2053 0 : ctx,
2054 0 : )?;
2055 :
2056 0 : slot_guard.upsert(TenantSlot::Attached(tenant))?;
2057 :
2058 0 : Ok(reparented)
2059 0 : }
2060 :
2061 : /// A page service client sends a TenantId, and to look up the correct Tenant we must
2062 : /// resolve this to a fully qualified TenantShardId.
2063 : ///
2064 : /// During shard splits: we shall see parent shards in InProgress state and skip them, and
2065 : /// instead match on child shards which should appear in Attached state. Very early in a shard
2066 : /// split, or in other cases where a shard is InProgress, we will return our own InProgress result
2067 : /// to instruct the caller to wait for that to finish before querying again.
2068 0 : pub(crate) fn resolve_attached_shard(
2069 0 : &self,
2070 0 : tenant_id: &TenantId,
2071 0 : selector: ShardSelector,
2072 0 : ) -> ShardResolveResult {
2073 0 : let tenants = self.tenants.read().unwrap();
2074 0 : let mut want_shard = None;
2075 0 : let mut any_in_progress = None;
2076 0 :
2077 0 : match &*tenants {
2078 0 : TenantsMap::Initializing => ShardResolveResult::NotFound,
2079 0 : TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => {
2080 0 : for slot in m.range(TenantShardId::tenant_range(*tenant_id)) {
2081 : // Ignore all slots that don't contain an attached tenant
2082 0 : let tenant = match &slot.1 {
2083 0 : TenantSlot::Attached(t) => t,
2084 0 : TenantSlot::InProgress(barrier) => {
2085 0 : // We might still find a usable shard, but in case we don't, remember that
2086 0 : // we saw at least one InProgress slot, so that we can distinguish this case
2087 0 : // from a simple NotFound in our return value.
2088 0 : any_in_progress = Some(barrier.clone());
2089 0 : continue;
2090 : }
2091 0 : _ => continue,
2092 : };
2093 :
2094 0 : match selector {
2095 0 : ShardSelector::Zero if slot.0.shard_number == ShardNumber(0) => {
2096 0 : return ShardResolveResult::Found(tenant.clone())
2097 : }
2098 0 : ShardSelector::Page(key) => {
2099 0 : // First slot we see for this tenant, calculate the expected shard number
2100 0 : // for the key: we will use this for checking if this and subsequent
2101 0 : // slots contain the key, rather than recalculating the hash each time.
2102 0 : if want_shard.is_none() {
2103 0 : want_shard = Some(tenant.shard_identity.get_shard_number(&key));
2104 0 : }
2105 :
2106 0 : if Some(tenant.shard_identity.number) == want_shard {
2107 0 : return ShardResolveResult::Found(tenant.clone());
2108 0 : }
2109 : }
2110 0 : ShardSelector::Known(shard)
2111 0 : if tenant.shard_identity.shard_index() == shard =>
2112 0 : {
2113 0 : return ShardResolveResult::Found(tenant.clone());
2114 : }
2115 0 : _ => continue,
2116 : }
2117 : }
2118 :
2119 : // Fall through: we didn't find a slot that was in Attached state & matched our selector. If
2120 : // we found one or more InProgress slot, indicate to caller that they should retry later. Otherwise
2121 : // this requested shard simply isn't found.
2122 0 : if let Some(barrier) = any_in_progress {
2123 0 : ShardResolveResult::InProgress(barrier)
2124 : } else {
2125 0 : ShardResolveResult::NotFound
2126 : }
2127 : }
2128 : }
2129 0 : }
2130 : }
2131 :
2132 0 : #[derive(Debug, thiserror::Error)]
2133 : pub(crate) enum GetTenantError {
2134 : /// NotFound is a TenantId rather than TenantShardId, because this error type is used from
2135 : /// getters that use a TenantId and a ShardSelector, not just getters that target a specific shard.
2136 : #[error("Tenant {0} not found")]
2137 : NotFound(TenantId),
2138 :
2139 : #[error("Tenant {0} is not active")]
2140 : NotActive(TenantShardId),
2141 :
2142 : // Initializing or shutting down: cannot authoritatively say whether we have this tenant
2143 : #[error("Tenant map is not available: {0}")]
2144 : MapState(#[from] TenantMapError),
2145 : }
2146 :
2147 0 : #[derive(thiserror::Error, Debug)]
2148 : pub(crate) enum GetActiveTenantError {
2149 : /// We may time out either while TenantSlot is InProgress, or while the Tenant
2150 : /// is in a non-Active state
2151 : #[error(
2152 : "Timed out waiting {wait_time:?} for tenant active state. Latest state: {latest_state:?}"
2153 : )]
2154 : WaitForActiveTimeout {
2155 : latest_state: Option<TenantState>,
2156 : wait_time: Duration,
2157 : },
2158 :
2159 : /// The TenantSlot is absent, or in secondary mode
2160 : #[error(transparent)]
2161 : NotFound(#[from] GetTenantError),
2162 :
2163 : /// Cancellation token fired while we were waiting
2164 : #[error("cancelled")]
2165 : Cancelled,
2166 :
2167 : /// Tenant exists, but is in a state that cannot become active (e.g. Stopping, Broken)
2168 : #[error("will not become active. Current state: {0}")]
2169 : WillNotBecomeActive(TenantState),
2170 :
2171 : /// Broken is logically a subset of WillNotBecomeActive, but a distinct error is useful as
2172 : /// WillNotBecomeActive is a permitted error under some circumstances, whereas broken should
2173 : /// never happen.
2174 : #[error("Tenant is broken: {0}")]
2175 : Broken(String),
2176 :
2177 : #[error("reconnect to switch tenant id")]
2178 : SwitchedTenant,
2179 : }
2180 :
2181 0 : #[derive(Debug, thiserror::Error)]
2182 : pub(crate) enum DeleteTimelineError {
2183 : #[error("Tenant {0}")]
2184 : Tenant(#[from] GetTenantError),
2185 :
2186 : #[error("Timeline {0}")]
2187 : Timeline(#[from] crate::tenant::DeleteTimelineError),
2188 : }
2189 :
2190 0 : #[derive(Debug, thiserror::Error)]
2191 : pub(crate) enum TenantStateError {
2192 : #[error("Tenant {0} is stopping")]
2193 : IsStopping(TenantShardId),
2194 : #[error(transparent)]
2195 : SlotError(#[from] TenantSlotError),
2196 : #[error(transparent)]
2197 : SlotUpsertError(#[from] TenantSlotUpsertError),
2198 : #[error(transparent)]
2199 : Other(#[from] anyhow::Error),
2200 : }
2201 :
2202 0 : #[derive(Debug, thiserror::Error)]
2203 : pub(crate) enum TenantMapListError {
2204 : #[error("tenant map is still initiailizing")]
2205 : Initializing,
2206 : }
2207 :
2208 0 : #[derive(Debug, thiserror::Error)]
2209 : pub(crate) enum TenantMapInsertError {
2210 : #[error(transparent)]
2211 : SlotError(#[from] TenantSlotError),
2212 : #[error(transparent)]
2213 : SlotUpsertError(#[from] TenantSlotUpsertError),
2214 : #[error(transparent)]
2215 : Other(#[from] anyhow::Error),
2216 : }
2217 :
2218 : /// Superset of TenantMapError: issues that can occur when acquiring a slot
2219 : /// for a particular tenant ID.
2220 0 : #[derive(Debug, thiserror::Error)]
2221 : pub(crate) enum TenantSlotError {
2222 : /// When acquiring a slot with the expectation that the tenant already exists.
2223 : #[error("Tenant {0} not found")]
2224 : NotFound(TenantShardId),
2225 :
2226 : // Tried to read a slot that is currently being mutated by another administrative
2227 : // operation.
2228 : #[error("tenant has a state change in progress, try again later")]
2229 : InProgress,
2230 :
2231 : #[error(transparent)]
2232 : MapState(#[from] TenantMapError),
2233 : }
2234 :
2235 : /// Superset of TenantMapError: issues that can occur when using a SlotGuard
2236 : /// to insert a new value.
2237 0 : #[derive(thiserror::Error)]
2238 : pub(crate) enum TenantSlotUpsertError {
2239 : /// An error where the slot is in an unexpected state, indicating a code bug
2240 : #[error("Internal error updating Tenant")]
2241 : InternalError(Cow<'static, str>),
2242 :
2243 : #[error(transparent)]
2244 : MapState(TenantMapError),
2245 :
2246 : // If we encounter TenantManager shutdown during upsert, we must carry the Completion
2247 : // from the SlotGuard, so that the caller can hold it while they clean up: otherwise
2248 : // TenantManager shutdown might race ahead before we're done cleaning up any Tenant that
2249 : // was protected by the SlotGuard.
2250 : #[error("Shutting down")]
2251 : ShuttingDown((TenantSlot, utils::completion::Completion)),
2252 : }
2253 :
2254 : impl std::fmt::Debug for TenantSlotUpsertError {
2255 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
2256 0 : match self {
2257 0 : Self::InternalError(reason) => write!(f, "Internal Error {reason}"),
2258 0 : Self::MapState(map_error) => write!(f, "Tenant map state: {map_error:?}"),
2259 0 : Self::ShuttingDown(_completion) => write!(f, "Tenant map shutting down"),
2260 : }
2261 0 : }
2262 : }
2263 :
2264 0 : #[derive(Debug, thiserror::Error)]
2265 : enum TenantSlotDropError {
2266 : /// It is only legal to drop a TenantSlot if its contents are fully shut down
2267 : #[error("Tenant was not shut down")]
2268 : NotShutdown,
2269 : }
2270 :
2271 : /// Errors that can happen any time we are walking the tenant map to try and acquire
2272 : /// the TenantSlot for a particular tenant.
2273 0 : #[derive(Debug, thiserror::Error)]
2274 : pub enum TenantMapError {
2275 : // Tried to read while initializing
2276 : #[error("tenant map is still initializing")]
2277 : StillInitializing,
2278 :
2279 : // Tried to read while shutting down
2280 : #[error("tenant map is shutting down")]
2281 : ShuttingDown,
2282 : }
2283 :
2284 : /// Guards a particular tenant_id's content in the TenantsMap. While this
2285 : /// structure exists, the TenantsMap will contain a [`TenantSlot::InProgress`]
2286 : /// for this tenant, which acts as a marker for any operations targeting
2287 : /// this tenant to retry later, or wait for the InProgress state to end.
2288 : ///
2289 : /// This structure enforces the important invariant that we do not have overlapping
2290 : /// tasks that will try use local storage for a the same tenant ID: we enforce that
2291 : /// the previous contents of a slot have been shut down before the slot can be
2292 : /// left empty or used for something else
2293 : ///
2294 : /// Holders of a SlotGuard should explicitly dispose of it, using either `upsert`
2295 : /// to provide a new value, or `revert` to put the slot back into its initial
2296 : /// state. If the SlotGuard is dropped without calling either of these, then
2297 : /// we will leave the slot empty if our `old_value` is already shut down, else
2298 : /// we will replace the slot with `old_value` (equivalent to doing a revert).
2299 : ///
2300 : /// The `old_value` may be dropped before the SlotGuard is dropped, by calling
2301 : /// `drop_old_value`. It is an error to call this without shutting down
2302 : /// the conents of `old_value`.
2303 : pub struct SlotGuard {
2304 : tenant_shard_id: TenantShardId,
2305 : old_value: Option<TenantSlot>,
2306 : upserted: bool,
2307 :
2308 : /// [`TenantSlot::InProgress`] carries the corresponding Barrier: it will
2309 : /// release any waiters as soon as this SlotGuard is dropped.
2310 : completion: utils::completion::Completion,
2311 : }
2312 :
2313 : impl SlotGuard {
2314 2 : fn new(
2315 2 : tenant_shard_id: TenantShardId,
2316 2 : old_value: Option<TenantSlot>,
2317 2 : completion: utils::completion::Completion,
2318 2 : ) -> Self {
2319 2 : Self {
2320 2 : tenant_shard_id,
2321 2 : old_value,
2322 2 : upserted: false,
2323 2 : completion,
2324 2 : }
2325 2 : }
2326 :
2327 : /// Get any value that was present in the slot before we acquired ownership
2328 : /// of it: in state transitions, this will be the old state.
2329 2 : fn get_old_value(&self) -> &Option<TenantSlot> {
2330 2 : &self.old_value
2331 2 : }
2332 :
2333 : /// Emplace a new value in the slot. This consumes the guard, and after
2334 : /// returning, the slot is no longer protected from concurrent changes.
2335 0 : fn upsert(mut self, new_value: TenantSlot) -> Result<(), TenantSlotUpsertError> {
2336 0 : if !self.old_value_is_shutdown() {
2337 : // This is a bug: callers should never try to drop an old value without
2338 : // shutting it down
2339 0 : return Err(TenantSlotUpsertError::InternalError(
2340 0 : "Old TenantSlot value not shut down".into(),
2341 0 : ));
2342 0 : }
2343 :
2344 0 : let replaced = {
2345 0 : let mut locked = TENANTS.write().unwrap();
2346 0 :
2347 0 : if let TenantSlot::InProgress(_) = new_value {
2348 : // It is never expected to try and upsert InProgress via this path: it should
2349 : // only be written via the tenant_map_acquire_slot path. If we hit this it's a bug.
2350 0 : return Err(TenantSlotUpsertError::InternalError(
2351 0 : "Attempt to upsert an InProgress state".into(),
2352 0 : ));
2353 0 : }
2354 :
2355 0 : let m = match &mut *locked {
2356 : TenantsMap::Initializing => {
2357 0 : return Err(TenantSlotUpsertError::MapState(
2358 0 : TenantMapError::StillInitializing,
2359 0 : ))
2360 : }
2361 : TenantsMap::ShuttingDown(_) => {
2362 0 : return Err(TenantSlotUpsertError::ShuttingDown((
2363 0 : new_value,
2364 0 : self.completion.clone(),
2365 0 : )));
2366 : }
2367 0 : TenantsMap::Open(m) => m,
2368 0 : };
2369 0 :
2370 0 : METRICS.slot_inserted(&new_value);
2371 0 :
2372 0 : let replaced = m.insert(self.tenant_shard_id, new_value);
2373 0 : self.upserted = true;
2374 0 : if let Some(replaced) = replaced.as_ref() {
2375 0 : METRICS.slot_removed(replaced);
2376 0 : }
2377 :
2378 0 : replaced
2379 : };
2380 :
2381 : // Sanity check: on an upsert we should always be replacing an InProgress marker
2382 0 : match replaced {
2383 : Some(TenantSlot::InProgress(_)) => {
2384 : // Expected case: we find our InProgress in the map: nothing should have
2385 : // replaced it because the code that acquires slots will not grant another
2386 : // one for the same TenantId.
2387 0 : Ok(())
2388 : }
2389 : None => {
2390 0 : METRICS.unexpected_errors.inc();
2391 0 : error!(
2392 : tenant_shard_id = %self.tenant_shard_id,
2393 0 : "Missing InProgress marker during tenant upsert, this is a bug."
2394 : );
2395 0 : Err(TenantSlotUpsertError::InternalError(
2396 0 : "Missing InProgress marker during tenant upsert".into(),
2397 0 : ))
2398 : }
2399 0 : Some(slot) => {
2400 0 : METRICS.unexpected_errors.inc();
2401 0 : error!(tenant_shard_id=%self.tenant_shard_id, "Unexpected contents of TenantSlot during upsert, this is a bug. Contents: {:?}", slot);
2402 0 : Err(TenantSlotUpsertError::InternalError(
2403 0 : "Unexpected contents of TenantSlot".into(),
2404 0 : ))
2405 : }
2406 : }
2407 0 : }
2408 :
2409 : /// Replace the InProgress slot with whatever was in the guard when we started
2410 0 : fn revert(mut self) {
2411 0 : if let Some(value) = self.old_value.take() {
2412 0 : match self.upsert(value) {
2413 0 : Err(TenantSlotUpsertError::InternalError(_)) => {
2414 0 : // We already logged the error, nothing else we can do.
2415 0 : }
2416 : Err(
2417 : TenantSlotUpsertError::MapState(_) | TenantSlotUpsertError::ShuttingDown(_),
2418 0 : ) => {
2419 0 : // If the map is shutting down, we need not replace anything
2420 0 : }
2421 0 : Ok(()) => {}
2422 : }
2423 0 : }
2424 0 : }
2425 :
2426 : /// We may never drop our old value until it is cleanly shut down: otherwise we might leave
2427 : /// rogue background tasks that would write to the local tenant directory that this guard
2428 : /// is responsible for protecting
2429 2 : fn old_value_is_shutdown(&self) -> bool {
2430 2 : match self.old_value.as_ref() {
2431 2 : Some(TenantSlot::Attached(tenant)) => tenant.gate.close_complete(),
2432 0 : Some(TenantSlot::Secondary(secondary_tenant)) => secondary_tenant.gate.close_complete(),
2433 : Some(TenantSlot::InProgress(_)) => {
2434 : // A SlotGuard cannot be constructed for a slot that was already InProgress
2435 0 : unreachable!()
2436 : }
2437 0 : None => true,
2438 : }
2439 2 : }
2440 :
2441 : /// The guard holder is done with the old value of the slot: they are obliged to already
2442 : /// shut it down before we reach this point.
2443 2 : fn drop_old_value(&mut self) -> Result<(), TenantSlotDropError> {
2444 2 : if !self.old_value_is_shutdown() {
2445 0 : Err(TenantSlotDropError::NotShutdown)
2446 : } else {
2447 2 : self.old_value.take();
2448 2 : Ok(())
2449 : }
2450 2 : }
2451 : }
2452 :
2453 : impl Drop for SlotGuard {
2454 2 : fn drop(&mut self) {
2455 2 : if self.upserted {
2456 0 : return;
2457 2 : }
2458 2 : // Our old value is already shutdown, or it never existed: it is safe
2459 2 : // for us to fully release the TenantSlot back into an empty state
2460 2 :
2461 2 : let mut locked = TENANTS.write().unwrap();
2462 :
2463 2 : let m = match &mut *locked {
2464 : TenantsMap::Initializing => {
2465 : // There is no map, this should never happen.
2466 2 : return;
2467 : }
2468 : TenantsMap::ShuttingDown(_) => {
2469 : // When we transition to shutdown, InProgress elements are removed
2470 : // from the map, so we do not need to clean up our Inprogress marker.
2471 : // See [`shutdown_all_tenants0`]
2472 0 : return;
2473 : }
2474 0 : TenantsMap::Open(m) => m,
2475 0 : };
2476 0 :
2477 0 : use std::collections::btree_map::Entry;
2478 0 : match m.entry(self.tenant_shard_id) {
2479 0 : Entry::Occupied(mut entry) => {
2480 0 : if !matches!(entry.get(), TenantSlot::InProgress(_)) {
2481 0 : METRICS.unexpected_errors.inc();
2482 0 : error!(tenant_shard_id=%self.tenant_shard_id, "Unexpected contents of TenantSlot during drop, this is a bug. Contents: {:?}", entry.get());
2483 0 : }
2484 :
2485 0 : if self.old_value_is_shutdown() {
2486 0 : METRICS.slot_removed(entry.get());
2487 0 : entry.remove();
2488 0 : } else {
2489 0 : let inserting = self.old_value.take().unwrap();
2490 0 : METRICS.slot_inserted(&inserting);
2491 0 : let replaced = entry.insert(inserting);
2492 0 : METRICS.slot_removed(&replaced);
2493 0 : }
2494 : }
2495 : Entry::Vacant(_) => {
2496 0 : METRICS.unexpected_errors.inc();
2497 0 : error!(
2498 : tenant_shard_id = %self.tenant_shard_id,
2499 0 : "Missing InProgress marker during SlotGuard drop, this is a bug."
2500 : );
2501 : }
2502 : }
2503 2 : }
2504 : }
2505 :
2506 : enum TenantSlotPeekMode {
2507 : /// In Read mode, peek will be permitted to see the slots even if the pageserver is shutting down
2508 : Read,
2509 : /// In Write mode, trying to peek at a slot while the pageserver is shutting down is an error
2510 : Write,
2511 : }
2512 :
2513 0 : fn tenant_map_peek_slot<'a>(
2514 0 : tenants: &'a std::sync::RwLockReadGuard<'a, TenantsMap>,
2515 0 : tenant_shard_id: &TenantShardId,
2516 0 : mode: TenantSlotPeekMode,
2517 0 : ) -> Result<Option<&'a TenantSlot>, TenantMapError> {
2518 0 : match tenants.deref() {
2519 0 : TenantsMap::Initializing => Err(TenantMapError::StillInitializing),
2520 0 : TenantsMap::ShuttingDown(m) => match mode {
2521 : TenantSlotPeekMode::Read => Ok(Some(
2522 : // When reading in ShuttingDown state, we must translate None results
2523 : // into a ShuttingDown error, because absence of a tenant shard ID in the map
2524 : // isn't a reliable indicator of the tenant being gone: it might have been
2525 : // InProgress when shutdown started, and cleaned up from that state such
2526 : // that it's now no longer in the map. Callers will have to wait until
2527 : // we next start up to get a proper answer. This avoids incorrect 404 API responses.
2528 0 : m.get(tenant_shard_id).ok_or(TenantMapError::ShuttingDown)?,
2529 : )),
2530 0 : TenantSlotPeekMode::Write => Err(TenantMapError::ShuttingDown),
2531 : },
2532 0 : TenantsMap::Open(m) => Ok(m.get(tenant_shard_id)),
2533 : }
2534 0 : }
2535 :
2536 : enum TenantSlotAcquireMode {
2537 : /// Acquire the slot irrespective of current state, or whether it already exists
2538 : Any,
2539 : /// Return an error if trying to acquire a slot and it doesn't already exist
2540 : MustExist,
2541 : }
2542 :
2543 0 : fn tenant_map_acquire_slot(
2544 0 : tenant_shard_id: &TenantShardId,
2545 0 : mode: TenantSlotAcquireMode,
2546 0 : ) -> Result<SlotGuard, TenantSlotError> {
2547 0 : tenant_map_acquire_slot_impl(tenant_shard_id, &TENANTS, mode)
2548 0 : }
2549 :
2550 2 : fn tenant_map_acquire_slot_impl(
2551 2 : tenant_shard_id: &TenantShardId,
2552 2 : tenants: &std::sync::RwLock<TenantsMap>,
2553 2 : mode: TenantSlotAcquireMode,
2554 2 : ) -> Result<SlotGuard, TenantSlotError> {
2555 2 : use TenantSlotAcquireMode::*;
2556 2 : METRICS.tenant_slot_writes.inc();
2557 2 :
2558 2 : let mut locked = tenants.write().unwrap();
2559 2 : let span = tracing::info_span!("acquire_slot", tenant_id=%tenant_shard_id.tenant_id, shard_id = %tenant_shard_id.shard_slug());
2560 2 : let _guard = span.enter();
2561 :
2562 2 : let m = match &mut *locked {
2563 0 : TenantsMap::Initializing => return Err(TenantMapError::StillInitializing.into()),
2564 0 : TenantsMap::ShuttingDown(_) => return Err(TenantMapError::ShuttingDown.into()),
2565 2 : TenantsMap::Open(m) => m,
2566 2 : };
2567 2 :
2568 2 : use std::collections::btree_map::Entry;
2569 2 :
2570 2 : let entry = m.entry(*tenant_shard_id);
2571 2 :
2572 2 : match entry {
2573 0 : Entry::Vacant(v) => match mode {
2574 : MustExist => {
2575 0 : tracing::debug!("Vacant && MustExist: return NotFound");
2576 0 : Err(TenantSlotError::NotFound(*tenant_shard_id))
2577 : }
2578 : _ => {
2579 0 : let (completion, barrier) = utils::completion::channel();
2580 0 : let inserting = TenantSlot::InProgress(barrier);
2581 0 : METRICS.slot_inserted(&inserting);
2582 0 : v.insert(inserting);
2583 0 : tracing::debug!("Vacant, inserted InProgress");
2584 0 : Ok(SlotGuard::new(*tenant_shard_id, None, completion))
2585 : }
2586 : },
2587 2 : Entry::Occupied(mut o) => {
2588 2 : // Apply mode-driven checks
2589 2 : match (o.get(), mode) {
2590 : (TenantSlot::InProgress(_), _) => {
2591 0 : tracing::debug!("Occupied, failing for InProgress");
2592 0 : Err(TenantSlotError::InProgress)
2593 : }
2594 : _ => {
2595 : // Happy case: the slot was not in any state that violated our mode
2596 2 : let (completion, barrier) = utils::completion::channel();
2597 2 : let in_progress = TenantSlot::InProgress(barrier);
2598 2 : METRICS.slot_inserted(&in_progress);
2599 2 : let old_value = o.insert(in_progress);
2600 2 : METRICS.slot_removed(&old_value);
2601 2 : tracing::debug!("Occupied, replaced with InProgress");
2602 2 : Ok(SlotGuard::new(
2603 2 : *tenant_shard_id,
2604 2 : Some(old_value),
2605 2 : completion,
2606 2 : ))
2607 : }
2608 : }
2609 : }
2610 : }
2611 2 : }
2612 :
2613 : /// Stops and removes the tenant from memory, if it's not [`TenantState::Stopping`] already, bails otherwise.
2614 : /// Allows to remove other tenant resources manually, via `tenant_cleanup`.
2615 : /// If the cleanup fails, tenant will stay in memory in [`TenantState::Broken`] state, and another removal
2616 : /// operation would be needed to remove it.
2617 2 : async fn remove_tenant_from_memory<V, F>(
2618 2 : tenants: &std::sync::RwLock<TenantsMap>,
2619 2 : tenant_shard_id: TenantShardId,
2620 2 : tenant_cleanup: F,
2621 2 : ) -> Result<V, TenantStateError>
2622 2 : where
2623 2 : F: std::future::Future<Output = anyhow::Result<V>>,
2624 2 : {
2625 2 : let mut slot_guard =
2626 2 : tenant_map_acquire_slot_impl(&tenant_shard_id, tenants, TenantSlotAcquireMode::MustExist)?;
2627 :
2628 : // allow pageserver shutdown to await for our completion
2629 2 : let (_guard, progress) = completion::channel();
2630 :
2631 : // The SlotGuard allows us to manipulate the Tenant object without fear of some
2632 : // concurrent API request doing something else for the same tenant ID.
2633 2 : let attached_tenant = match slot_guard.get_old_value() {
2634 2 : Some(TenantSlot::Attached(tenant)) => {
2635 2 : // whenever we remove a tenant from memory, we don't want to flush and wait for upload
2636 2 : let shutdown_mode = ShutdownMode::Hard;
2637 2 :
2638 2 : // shutdown is sure to transition tenant to stopping, and wait for all tasks to complete, so
2639 2 : // that we can continue safely to cleanup.
2640 2 : match tenant.shutdown(progress, shutdown_mode).await {
2641 2 : Ok(()) => {}
2642 0 : Err(_other) => {
2643 0 : // if pageserver shutdown or other detach/ignore is already ongoing, we don't want to
2644 0 : // wait for it but return an error right away because these are distinct requests.
2645 0 : slot_guard.revert();
2646 0 : return Err(TenantStateError::IsStopping(tenant_shard_id));
2647 : }
2648 : }
2649 2 : Some(tenant)
2650 : }
2651 0 : Some(TenantSlot::Secondary(secondary_state)) => {
2652 0 : tracing::info!("Shutting down in secondary mode");
2653 0 : secondary_state.shutdown().await;
2654 0 : None
2655 : }
2656 : Some(TenantSlot::InProgress(_)) => {
2657 : // Acquiring a slot guarantees its old value was not InProgress
2658 0 : unreachable!();
2659 : }
2660 0 : None => None,
2661 : };
2662 :
2663 2 : match tenant_cleanup
2664 2 : .await
2665 2 : .with_context(|| format!("Failed to run cleanup for tenant {tenant_shard_id}"))
2666 : {
2667 2 : Ok(hook_value) => {
2668 2 : // Success: drop the old TenantSlot::Attached.
2669 2 : slot_guard
2670 2 : .drop_old_value()
2671 2 : .expect("We just called shutdown");
2672 2 :
2673 2 : Ok(hook_value)
2674 : }
2675 0 : Err(e) => {
2676 : // If we had a Tenant, set it to Broken and put it back in the TenantsMap
2677 0 : if let Some(attached_tenant) = attached_tenant {
2678 0 : attached_tenant.set_broken(e.to_string()).await;
2679 0 : }
2680 : // Leave the broken tenant in the map
2681 0 : slot_guard.revert();
2682 0 :
2683 0 : Err(TenantStateError::Other(e))
2684 : }
2685 : }
2686 2 : }
2687 :
2688 : use {
2689 : crate::repository::GcResult, pageserver_api::models::TimelineGcRequest,
2690 : utils::http::error::ApiError,
2691 : };
2692 :
2693 0 : #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), %timeline_id))]
2694 : pub(crate) async fn immediate_gc(
2695 : tenant_shard_id: TenantShardId,
2696 : timeline_id: TimelineId,
2697 : gc_req: TimelineGcRequest,
2698 : cancel: CancellationToken,
2699 : ctx: &RequestContext,
2700 : ) -> Result<GcResult, ApiError> {
2701 : let tenant = {
2702 : let guard = TENANTS.read().unwrap();
2703 : guard
2704 : .get(&tenant_shard_id)
2705 : .cloned()
2706 0 : .with_context(|| format!("tenant {tenant_shard_id}"))
2707 0 : .map_err(|e| ApiError::NotFound(e.into()))?
2708 : };
2709 :
2710 0 : let gc_horizon = gc_req.gc_horizon.unwrap_or_else(|| tenant.get_gc_horizon());
2711 : // Use tenant's pitr setting
2712 : let pitr = tenant.get_pitr_interval();
2713 :
2714 : tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
2715 :
2716 : // Run in task_mgr to avoid race with tenant_detach operation
2717 : let ctx: RequestContext =
2718 : ctx.detached_child(TaskKind::GarbageCollector, DownloadBehavior::Download);
2719 :
2720 0 : let _gate_guard = tenant.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
2721 :
2722 : fail::fail_point!("immediate_gc_task_pre");
2723 :
2724 : #[allow(unused_mut)]
2725 : let mut result = tenant
2726 : .gc_iteration(Some(timeline_id), gc_horizon, pitr, &cancel, &ctx)
2727 : .await;
2728 : // FIXME: `gc_iteration` can return an error for multiple reasons; we should handle it
2729 : // better once the types support it.
2730 :
2731 : #[cfg(feature = "testing")]
2732 : {
2733 : // we need to synchronize with drop completion for python tests without polling for
2734 : // log messages
2735 : if let Ok(result) = result.as_mut() {
2736 : let mut js = tokio::task::JoinSet::new();
2737 : for layer in std::mem::take(&mut result.doomed_layers) {
2738 : js.spawn(layer.wait_drop());
2739 : }
2740 : tracing::info!(
2741 : total = js.len(),
2742 : "starting to wait for the gc'd layers to be dropped"
2743 : );
2744 : while let Some(res) = js.join_next().await {
2745 : res.expect("wait_drop should not panic");
2746 : }
2747 : }
2748 :
2749 : let timeline = tenant.get_timeline(timeline_id, false).ok();
2750 0 : let rtc = timeline.as_ref().map(|x| &x.remote_client);
2751 :
2752 : if let Some(rtc) = rtc {
2753 : // layer drops schedule actions on remote timeline client to actually do the
2754 : // deletions; don't care about the shutdown error, just exit fast
2755 : drop(rtc.wait_completion().await);
2756 : }
2757 : }
2758 :
2759 0 : result.map_err(|e| match e {
2760 0 : GcError::TenantCancelled | GcError::TimelineCancelled => ApiError::ShuttingDown,
2761 : GcError::TimelineNotFound => {
2762 0 : ApiError::NotFound(anyhow::anyhow!("Timeline not found").into())
2763 : }
2764 0 : other => ApiError::InternalServerError(anyhow::anyhow!(other)),
2765 0 : })
2766 : }
2767 :
2768 : #[cfg(test)]
2769 : mod tests {
2770 : use std::collections::BTreeMap;
2771 : use std::sync::Arc;
2772 : use tracing::Instrument;
2773 :
2774 : use crate::tenant::mgr::TenantSlot;
2775 :
2776 : use super::{super::harness::TenantHarness, TenantsMap};
2777 :
2778 : #[tokio::test(start_paused = true)]
2779 2 : async fn shutdown_awaits_in_progress_tenant() {
2780 2 : // Test that if an InProgress tenant is in the map during shutdown, the shutdown will gracefully
2781 2 : // wait for it to complete before proceeding.
2782 2 :
2783 2 : let h = TenantHarness::create("shutdown_awaits_in_progress_tenant")
2784 2 : .await
2785 2 : .unwrap();
2786 8 : let (t, _ctx) = h.load().await;
2787 2 :
2788 2 : // harness loads it to active, which is forced and nothing is running on the tenant
2789 2 :
2790 2 : let id = t.tenant_shard_id();
2791 2 :
2792 2 : // tenant harness configures the logging and we cannot escape it
2793 2 : let span = h.span();
2794 2 : let _e = span.enter();
2795 2 :
2796 2 : let tenants = BTreeMap::from([(id, TenantSlot::Attached(t.clone()))]);
2797 2 : let tenants = Arc::new(std::sync::RwLock::new(TenantsMap::Open(tenants)));
2798 2 :
2799 2 : // Invoke remove_tenant_from_memory with a cleanup hook that blocks until we manually
2800 2 : // permit it to proceed: that will stick the tenant in InProgress
2801 2 :
2802 2 : let (until_cleanup_completed, can_complete_cleanup) = utils::completion::channel();
2803 2 : let (until_cleanup_started, cleanup_started) = utils::completion::channel();
2804 2 : let mut remove_tenant_from_memory_task = {
2805 2 : let jh = tokio::spawn({
2806 2 : let tenants = tenants.clone();
2807 2 : async move {
2808 2 : let cleanup = async move {
2809 2 : drop(until_cleanup_started);
2810 2 : can_complete_cleanup.wait().await;
2811 2 : anyhow::Ok(())
2812 2 : };
2813 2 : super::remove_tenant_from_memory(&tenants, id, cleanup).await
2814 2 : }
2815 2 : .instrument(h.span())
2816 2 : });
2817 2 :
2818 2 : // now the long cleanup should be in place, with the stopping state
2819 2 : cleanup_started.wait().await;
2820 2 : jh
2821 2 : };
2822 2 :
2823 2 : let mut shutdown_task = {
2824 2 : let (until_shutdown_started, shutdown_started) = utils::completion::channel();
2825 2 :
2826 2 : let shutdown_task = tokio::spawn(async move {
2827 2 : drop(until_shutdown_started);
2828 4 : super::shutdown_all_tenants0(&tenants).await;
2829 2 : });
2830 2 :
2831 2 : shutdown_started.wait().await;
2832 2 : shutdown_task
2833 2 : };
2834 2 :
2835 2 : let long_time = std::time::Duration::from_secs(15);
2836 2 : tokio::select! {
2837 2 : _ = &mut shutdown_task => unreachable!("shutdown should block on remove_tenant_from_memory completing"),
2838 2 : _ = &mut remove_tenant_from_memory_task => unreachable!("remove_tenant_from_memory_task should not complete until explicitly unblocked"),
2839 2 : _ = tokio::time::sleep(long_time) => {},
2840 2 : }
2841 2 :
2842 2 : drop(until_cleanup_completed);
2843 2 :
2844 2 : // Now that we allow it to proceed, shutdown should complete immediately
2845 2 : remove_tenant_from_memory_task.await.unwrap().unwrap();
2846 2 : shutdown_task.await.unwrap();
2847 2 : }
2848 : }
|
