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