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