Line data Source code
1 : //! This module acts as a switchboard to access different repositories managed by this
2 : //! page server.
3 :
4 : use std::borrow::Cow;
5 : use std::cmp::Ordering;
6 : use std::collections::{BTreeMap, HashMap, HashSet};
7 : use std::ops::Deref;
8 : use std::sync::Arc;
9 : use std::time::Duration;
10 :
11 : use anyhow::Context;
12 : use camino::{Utf8DirEntry, Utf8Path, Utf8PathBuf};
13 : use futures::StreamExt;
14 : use itertools::Itertools;
15 : use once_cell::sync::Lazy;
16 : use pageserver_api::key::Key;
17 : use pageserver_api::models::{DetachBehavior, LocationConfigMode};
18 : use pageserver_api::shard::{
19 : ShardCount, ShardIdentity, ShardIndex, ShardNumber, ShardStripeSize, TenantShardId,
20 : };
21 : use pageserver_api::upcall_api::ReAttachResponseTenant;
22 : use rand::Rng;
23 : use rand::distributions::Alphanumeric;
24 : use remote_storage::TimeoutOrCancel;
25 : use sysinfo::SystemExt;
26 : use tokio::fs;
27 : use tokio::task::JoinSet;
28 : use tokio_util::sync::CancellationToken;
29 : use tracing::*;
30 : use utils::crashsafe::path_with_suffix_extension;
31 : use utils::fs_ext::PathExt;
32 : use utils::generation::Generation;
33 : use utils::id::{TenantId, TimelineId};
34 : use utils::{backoff, completion, crashsafe};
35 :
36 : use super::remote_timeline_client::remote_tenant_path;
37 : use super::secondary::SecondaryTenant;
38 : use super::timeline::detach_ancestor::{self, PreparedTimelineDetach};
39 : use super::{GlobalShutDown, TenantSharedResources};
40 : use crate::config::PageServerConf;
41 : use crate::context::{DownloadBehavior, RequestContext};
42 : use crate::controller_upcall_client::{
43 : RetryForeverError, StorageControllerUpcallApi, StorageControllerUpcallClient,
44 : };
45 : use crate::deletion_queue::DeletionQueueClient;
46 : use crate::http::routes::ACTIVE_TENANT_TIMEOUT;
47 : use crate::metrics::{TENANT, TENANT_MANAGER as METRICS};
48 : use crate::task_mgr::{BACKGROUND_RUNTIME, TaskKind};
49 : use crate::tenant::config::{
50 : AttachedLocationConfig, AttachmentMode, LocationConf, LocationMode, SecondaryLocationConfig,
51 : };
52 : use crate::tenant::span::debug_assert_current_span_has_tenant_id;
53 : use crate::tenant::storage_layer::inmemory_layer;
54 : use crate::tenant::timeline::ShutdownMode;
55 : use crate::tenant::{
56 : AttachedTenantConf, GcError, LoadConfigError, SpawnMode, TenantShard, TenantState,
57 : };
58 : use crate::virtual_file::MaybeFatalIo;
59 : use crate::{InitializationOrder, TEMP_FILE_SUFFIX};
60 :
61 : /// For a tenant that appears in TenantsMap, it may either be
62 : /// - `Attached`: has a full Tenant object, is elegible to service
63 : /// reads and ingest WAL.
64 : /// - `Secondary`: is only keeping a local cache warm.
65 : ///
66 : /// Secondary is a totally distinct state rather than being a mode of a `Tenant`, because
67 : /// that way we avoid having to carefully switch a tenant's ingestion etc on and off during
68 : /// its lifetime, and we can preserve some important safety invariants like `Tenant` always
69 : /// having a properly acquired generation (Secondary doesn't need a generation)
70 : #[derive(Clone)]
71 : pub(crate) enum TenantSlot {
72 : Attached(Arc<TenantShard>),
73 : Secondary(Arc<SecondaryTenant>),
74 : /// In this state, other administrative operations acting on the TenantId should
75 : /// block, or return a retry indicator equivalent to HTTP 503.
76 : InProgress(utils::completion::Barrier),
77 : }
78 :
79 : impl std::fmt::Debug for TenantSlot {
80 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
81 0 : match self {
82 0 : Self::Attached(tenant) => write!(f, "Attached({})", tenant.current_state()),
83 0 : Self::Secondary(_) => write!(f, "Secondary"),
84 0 : Self::InProgress(_) => write!(f, "InProgress"),
85 : }
86 0 : }
87 : }
88 :
89 : impl TenantSlot {
90 : /// Return the `Tenant` in this slot if attached, else None
91 0 : fn get_attached(&self) -> Option<&Arc<TenantShard>> {
92 0 : match self {
93 0 : Self::Attached(t) => Some(t),
94 0 : Self::Secondary(_) => None,
95 0 : Self::InProgress(_) => None,
96 : }
97 0 : }
98 : }
99 :
100 : /// The tenants known to the pageserver.
101 : /// The enum variants are used to distinguish the different states that the pageserver can be in.
102 : pub(crate) enum TenantsMap {
103 : /// [`init_tenant_mgr`] is not done yet.
104 : Initializing,
105 : /// [`init_tenant_mgr`] is done, all on-disk tenants have been loaded.
106 : /// New tenants can be added using [`tenant_map_acquire_slot`].
107 : Open(BTreeMap<TenantShardId, TenantSlot>),
108 : /// The pageserver has entered shutdown mode via [`TenantManager::shutdown`].
109 : /// Existing tenants are still accessible, but no new tenants can be created.
110 : ShuttingDown(BTreeMap<TenantShardId, TenantSlot>),
111 : }
112 :
113 : /// When resolving a TenantId to a shard, we may be looking for the 0th
114 : /// shard, or we might be looking for whichever shard holds a particular page.
115 : #[derive(Copy, Clone)]
116 : pub(crate) enum ShardSelector {
117 : /// Only return the 0th shard, if it is present. If a non-0th shard is present,
118 : /// ignore it.
119 : Zero,
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 ControllerUpcallClient 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.r#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<TenantShard>),
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<TenantShard>> {
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)
224 0 : .await?
225 0 : .sync_all()
226 0 : .await
227 0 : .maybe_fatal_err("safe_rename_tenant_dir")?;
228 0 : Ok(tmp_path)
229 0 : }
230 :
231 : /// See [`Self::spawn`].
232 : #[derive(Clone, Default)]
233 : pub struct BackgroundPurges(tokio_util::task::TaskTracker);
234 :
235 : impl BackgroundPurges {
236 : /// When we have moved a tenant's content to a temporary directory, we may delete it lazily in
237 : /// the background, and thereby avoid blocking any API requests on this deletion completing.
238 : ///
239 : /// Although we are cleaning up the tenant, this task is not meant to be bound by the lifetime of the tenant in memory.
240 : /// Thus the [`BackgroundPurges`] type to keep track of these tasks.
241 0 : pub fn spawn(&self, tmp_path: Utf8PathBuf) {
242 0 : // because on shutdown we close and wait, we are misusing TaskTracker a bit.
243 0 : //
244 0 : // so first acquire a token, then check if the tracker has been closed. the tracker might get closed
245 0 : // right after, but at least the shutdown will wait for what we are spawning next.
246 0 : let token = self.0.token();
247 0 :
248 0 : if self.0.is_closed() {
249 0 : warn!(
250 : %tmp_path,
251 0 : "trying to spawn background purge during shutdown, ignoring"
252 : );
253 0 : return;
254 0 : }
255 :
256 0 : let span = info_span!(parent: None, "background_purge", %tmp_path);
257 :
258 0 : let task = move || {
259 0 : let _token = token;
260 0 : let _entered = span.entered();
261 0 : if let Err(error) = std::fs::remove_dir_all(tmp_path.as_path()) {
262 : // should we fatal_io_error here?
263 0 : warn!(%error, "failed to purge tenant directory");
264 0 : }
265 0 : };
266 :
267 0 : BACKGROUND_RUNTIME.spawn_blocking(task);
268 0 : }
269 :
270 : /// When this future completes, all background purges have completed.
271 : /// The first poll of the future will already lock out new background purges spawned via [`Self::spawn`].
272 : ///
273 : /// Concurrent calls will coalesce.
274 : ///
275 : /// # Cancellation-Safety
276 : ///
277 : /// If this future is dropped before polled to completion, concurrent and subsequent
278 : /// instances of this future will continue to be correct.
279 : #[instrument(skip_all)]
280 : pub async fn shutdown(&self) {
281 : // forbid new tasks (can be called many times)
282 : self.0.close();
283 : self.0.wait().await;
284 : }
285 : }
286 :
287 : static TENANTS: Lazy<std::sync::RwLock<TenantsMap>> =
288 12 : Lazy::new(|| std::sync::RwLock::new(TenantsMap::Initializing));
289 :
290 : /// Responsible for storing and mutating the collection of all tenants
291 : /// that this pageserver has state for.
292 : ///
293 : /// Every Tenant and SecondaryTenant instance lives inside the TenantManager.
294 : ///
295 : /// The most important role of the TenantManager is to prevent conflicts: e.g. trying to attach
296 : /// the same tenant twice concurrently, or trying to configure the same tenant into secondary
297 : /// and attached modes concurrently.
298 : pub struct TenantManager {
299 : conf: &'static PageServerConf,
300 : // TODO: currently this is a &'static pointing to TENANTs. When we finish refactoring
301 : // out of that static variable, the TenantManager can own this.
302 : // See https://github.com/neondatabase/neon/issues/5796
303 : tenants: &'static std::sync::RwLock<TenantsMap>,
304 : resources: TenantSharedResources,
305 :
306 : // Long-running operations that happen outside of a [`Tenant`] lifetime should respect this token.
307 : // This is for edge cases like tenant deletion. In normal cases (within a Tenant lifetime),
308 : // tenants have their own cancellation tokens, which we fire individually in [`Self::shutdown`], or
309 : // when the tenant detaches.
310 : cancel: CancellationToken,
311 :
312 : background_purges: BackgroundPurges,
313 : }
314 :
315 0 : fn emergency_generations(
316 0 : tenant_confs: &HashMap<TenantShardId, Result<LocationConf, LoadConfigError>>,
317 0 : ) -> HashMap<TenantShardId, TenantStartupMode> {
318 0 : tenant_confs
319 0 : .iter()
320 0 : .filter_map(|(tid, lc)| {
321 0 : let lc = match lc {
322 0 : Ok(lc) => lc,
323 0 : Err(_) => return None,
324 : };
325 : Some((
326 0 : *tid,
327 0 : match &lc.mode {
328 0 : LocationMode::Attached(alc) => {
329 0 : TenantStartupMode::Attached((alc.attach_mode, alc.generation))
330 : }
331 0 : LocationMode::Secondary(_) => TenantStartupMode::Secondary,
332 : },
333 : ))
334 0 : })
335 0 : .collect()
336 0 : }
337 :
338 0 : async fn init_load_generations(
339 0 : conf: &'static PageServerConf,
340 0 : tenant_confs: &HashMap<TenantShardId, Result<LocationConf, LoadConfigError>>,
341 0 : resources: &TenantSharedResources,
342 0 : cancel: &CancellationToken,
343 0 : ) -> anyhow::Result<Option<HashMap<TenantShardId, TenantStartupMode>>> {
344 0 : let generations = if conf.control_plane_emergency_mode {
345 0 : error!(
346 0 : "Emergency mode! Tenants will be attached unsafely using their last known generation"
347 : );
348 0 : emergency_generations(tenant_confs)
349 0 : } else if let Some(client) = StorageControllerUpcallClient::new(conf, cancel)? {
350 0 : info!("Calling {} API to re-attach tenants", client.base_url());
351 : // If we are configured to use the control plane API, then it is the source of truth for what tenants to load.
352 0 : match client.re_attach(conf).await {
353 0 : Ok(tenants) => tenants
354 0 : .into_iter()
355 0 : .flat_map(|(id, rart)| {
356 0 : TenantStartupMode::from_reattach_tenant(rart).map(|tsm| (id, tsm))
357 0 : })
358 0 : .collect(),
359 : Err(RetryForeverError::ShuttingDown) => {
360 0 : anyhow::bail!("Shut down while waiting for control plane re-attach response")
361 : }
362 : }
363 : } else {
364 0 : info!("Control plane API not configured, tenant generations are disabled");
365 0 : return Ok(None);
366 : };
367 :
368 : // The deletion queue needs to know about the startup attachment state to decide which (if any) stored
369 : // deletion list entries may still be valid. We provide that by pushing a recovery operation into
370 : // the queue. Sequential processing of te queue ensures that recovery is done before any new tenant deletions
371 : // are processed, even though we don't block on recovery completing here.
372 0 : let attached_tenants = generations
373 0 : .iter()
374 0 : .flat_map(|(id, start_mode)| {
375 0 : match start_mode {
376 0 : TenantStartupMode::Attached((_mode, generation)) => Some(generation),
377 0 : TenantStartupMode::Secondary => None,
378 : }
379 0 : .map(|gen_| (*id, *gen_))
380 0 : })
381 0 : .collect();
382 0 : resources.deletion_queue_client.recover(attached_tenants)?;
383 :
384 0 : Ok(Some(generations))
385 0 : }
386 :
387 : /// Given a directory discovered in the pageserver's tenants/ directory, attempt
388 : /// to load a tenant config from it.
389 : ///
390 : /// If we cleaned up something expected (like an empty dir or a temp dir), return None.
391 0 : fn load_tenant_config(
392 0 : conf: &'static PageServerConf,
393 0 : tenant_shard_id: TenantShardId,
394 0 : dentry: Utf8DirEntry,
395 0 : ) -> Option<Result<LocationConf, LoadConfigError>> {
396 0 : let tenant_dir_path = dentry.path().to_path_buf();
397 0 : if crate::is_temporary(&tenant_dir_path) {
398 0 : info!("Found temporary tenant directory, removing: {tenant_dir_path}");
399 : // No need to use safe_remove_tenant_dir_all because this is already
400 : // a temporary path
401 0 : std::fs::remove_dir_all(&tenant_dir_path).fatal_err("delete temporary tenant dir");
402 0 : return None;
403 0 : }
404 0 :
405 0 : // This case happens if we crash during attachment before writing a config into the dir
406 0 : let is_empty = tenant_dir_path
407 0 : .is_empty_dir()
408 0 : .fatal_err("Checking for empty tenant dir");
409 0 : if is_empty {
410 0 : info!("removing empty tenant directory {tenant_dir_path:?}");
411 0 : std::fs::remove_dir(&tenant_dir_path).fatal_err("delete empty tenant dir");
412 0 : return None;
413 0 : }
414 0 :
415 0 : Some(TenantShard::load_tenant_config(conf, &tenant_shard_id))
416 0 : }
417 :
418 : /// Initial stage of load: walk the local tenants directory, clean up any temp files,
419 : /// and load configurations for the tenants we found.
420 : ///
421 : /// Do this in parallel, because we expect 10k+ tenants, so serial execution can take
422 : /// seconds even on reasonably fast drives.
423 0 : async fn init_load_tenant_configs(
424 0 : conf: &'static PageServerConf,
425 0 : ) -> HashMap<TenantShardId, Result<LocationConf, LoadConfigError>> {
426 0 : let tenants_dir = conf.tenants_path();
427 :
428 0 : let dentries = tokio::task::spawn_blocking(move || -> Vec<Utf8DirEntry> {
429 0 : let context = format!("read tenants dir {tenants_dir}");
430 0 : let dir_entries = tenants_dir.read_dir_utf8().fatal_err(&context);
431 0 :
432 0 : dir_entries
433 0 : .collect::<Result<Vec<_>, std::io::Error>>()
434 0 : .fatal_err(&context)
435 0 : })
436 0 : .await
437 0 : .expect("Config load task panicked");
438 0 :
439 0 : let mut configs = HashMap::new();
440 0 :
441 0 : let mut join_set = JoinSet::new();
442 0 : for dentry in dentries {
443 0 : let tenant_shard_id = match dentry.file_name().parse::<TenantShardId>() {
444 0 : Ok(id) => id,
445 : Err(_) => {
446 0 : warn!(
447 0 : "Invalid tenant path (garbage in our repo directory?): '{}'",
448 0 : dentry.file_name()
449 : );
450 0 : continue;
451 : }
452 : };
453 :
454 0 : join_set.spawn_blocking(move || {
455 0 : (
456 0 : tenant_shard_id,
457 0 : load_tenant_config(conf, tenant_shard_id, dentry),
458 0 : )
459 0 : });
460 0 : }
461 :
462 0 : while let Some(r) = join_set.join_next().await {
463 0 : let (tenant_shard_id, tenant_config) = r.expect("Panic in config load task");
464 0 : if let Some(tenant_config) = tenant_config {
465 0 : configs.insert(tenant_shard_id, tenant_config);
466 0 : }
467 : }
468 :
469 0 : configs
470 0 : }
471 :
472 : #[derive(Debug, thiserror::Error)]
473 : pub(crate) enum DeleteTenantError {
474 : #[error("Tenant map slot error {0}")]
475 : SlotError(#[from] TenantSlotError),
476 :
477 : #[error("Cancelled")]
478 : Cancelled,
479 :
480 : #[error(transparent)]
481 : Other(#[from] anyhow::Error),
482 : }
483 :
484 : /// Initialize repositories with locally available timelines.
485 : /// Timelines that are only partially available locally (remote storage has more data than this pageserver)
486 : /// are scheduled for download and added to the tenant once download is completed.
487 : #[instrument(skip_all)]
488 : pub async fn init_tenant_mgr(
489 : conf: &'static PageServerConf,
490 : background_purges: BackgroundPurges,
491 : resources: TenantSharedResources,
492 : init_order: InitializationOrder,
493 : cancel: CancellationToken,
494 : ) -> anyhow::Result<TenantManager> {
495 : let mut tenants = BTreeMap::new();
496 :
497 : let ctx = RequestContext::todo_child(TaskKind::Startup, DownloadBehavior::Warn);
498 :
499 : // Initialize dynamic limits that depend on system resources
500 : let system_memory =
501 : sysinfo::System::new_with_specifics(sysinfo::RefreshKind::new().with_memory())
502 : .total_memory();
503 : let max_ephemeral_layer_bytes =
504 : conf.ephemeral_bytes_per_memory_kb as u64 * (system_memory / 1024);
505 : tracing::info!(
506 : "Initialized ephemeral layer size limit to {max_ephemeral_layer_bytes}, for {system_memory} bytes of memory"
507 : );
508 : inmemory_layer::GLOBAL_RESOURCES.max_dirty_bytes.store(
509 : max_ephemeral_layer_bytes,
510 : std::sync::atomic::Ordering::Relaxed,
511 : );
512 :
513 : // Scan local filesystem for attached tenants
514 : let tenant_configs = init_load_tenant_configs(conf).await;
515 :
516 : // Determine which tenants are to be secondary or attached, and in which generation
517 : let tenant_modes = init_load_generations(conf, &tenant_configs, &resources, &cancel).await?;
518 :
519 : tracing::info!(
520 : "Attaching {} tenants at startup, warming up {} at a time",
521 : tenant_configs.len(),
522 : conf.concurrent_tenant_warmup.initial_permits()
523 : );
524 : TENANT.startup_scheduled.inc_by(tenant_configs.len() as u64);
525 :
526 : // Accumulate futures for writing tenant configs, so that we can execute in parallel
527 : let mut config_write_futs = Vec::new();
528 :
529 : // Update the location configs according to the re-attach response and persist them to disk
530 : tracing::info!("Updating {} location configs", tenant_configs.len());
531 : for (tenant_shard_id, location_conf) in tenant_configs {
532 : let tenant_dir_path = conf.tenant_path(&tenant_shard_id);
533 :
534 : let mut location_conf = match location_conf {
535 : Ok(l) => l,
536 : Err(e) => {
537 : // This should only happen in the case of a serialization bug or critical local I/O error: we cannot load this tenant
538 : error!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Failed to load tenant config, failed to {e:#}");
539 : continue;
540 : }
541 : };
542 :
543 : // FIXME: if we were attached, and get demoted to secondary on re-attach, we
544 : // don't have a place to get a config.
545 : // (https://github.com/neondatabase/neon/issues/5377)
546 : const DEFAULT_SECONDARY_CONF: SecondaryLocationConfig =
547 : SecondaryLocationConfig { warm: true };
548 :
549 : if let Some(tenant_modes) = &tenant_modes {
550 : // We have a generation map: treat it as the authority for whether
551 : // this tenant is really attached.
552 : match tenant_modes.get(&tenant_shard_id) {
553 : None => {
554 : 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");
555 :
556 : match safe_rename_tenant_dir(&tenant_dir_path).await {
557 : Ok(tmp_path) => {
558 : background_purges.spawn(tmp_path);
559 : }
560 : Err(e) => {
561 : error!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
562 : "Failed to move detached tenant directory '{tenant_dir_path}': {e:?}");
563 : }
564 : };
565 :
566 : // We deleted local content: move on to next tenant, don't try and spawn this one.
567 : continue;
568 : }
569 : Some(TenantStartupMode::Secondary) => {
570 : if !matches!(location_conf.mode, LocationMode::Secondary(_)) {
571 : location_conf.mode = LocationMode::Secondary(DEFAULT_SECONDARY_CONF);
572 : }
573 : }
574 : Some(TenantStartupMode::Attached((attach_mode, generation))) => {
575 : let old_gen_higher = match &location_conf.mode {
576 : LocationMode::Attached(AttachedLocationConfig {
577 : generation: old_generation,
578 : attach_mode: _attach_mode,
579 : }) => {
580 : if old_generation > generation {
581 : Some(old_generation)
582 : } else {
583 : None
584 : }
585 : }
586 : _ => None,
587 : };
588 : if let Some(old_generation) = old_gen_higher {
589 : tracing::error!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
590 : "Control plane gave decreasing generation ({generation:?}) in re-attach response for tenant that was attached in generation {:?}, demoting to secondary",
591 : old_generation
592 : );
593 :
594 : // We cannot safely attach this tenant given a bogus generation number, but let's avoid throwing away
595 : // local disk content: demote to secondary rather than detaching.
596 : location_conf.mode = LocationMode::Secondary(DEFAULT_SECONDARY_CONF);
597 : } else {
598 : location_conf.attach_in_generation(*attach_mode, *generation);
599 : }
600 : }
601 : }
602 : } else {
603 : // Legacy mode: no generation information, any tenant present
604 : // on local disk may activate
605 : info!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Starting tenant in legacy mode, no generation",);
606 : };
607 :
608 : // Presence of a generation number implies attachment: attach the tenant
609 : // if it wasn't already, and apply the generation number.
610 0 : config_write_futs.push(async move {
611 0 : let r =
612 0 : TenantShard::persist_tenant_config(conf, &tenant_shard_id, &location_conf).await;
613 0 : (tenant_shard_id, location_conf, r)
614 0 : });
615 : }
616 :
617 : // Execute config writes with concurrency, to avoid bottlenecking on local FS write latency
618 : tracing::info!(
619 : "Writing {} location config files...",
620 : config_write_futs.len()
621 : );
622 : let config_write_results = futures::stream::iter(config_write_futs)
623 : .buffer_unordered(16)
624 : .collect::<Vec<_>>()
625 : .await;
626 :
627 : tracing::info!(
628 : "Spawning {} tenant shard locations...",
629 : config_write_results.len()
630 : );
631 : // For those shards that have live configurations, construct `Tenant` or `SecondaryTenant` objects and start them running
632 : for (tenant_shard_id, location_conf, config_write_result) in config_write_results {
633 : // Writing a config to local disk is foundational to startup up tenants: panic if we can't.
634 : config_write_result.fatal_err("write tenant shard config file");
635 :
636 : let tenant_dir_path = conf.tenant_path(&tenant_shard_id);
637 : let shard_identity = location_conf.shard;
638 : let slot = match location_conf.mode {
639 : LocationMode::Attached(attached_conf) => TenantSlot::Attached(
640 : tenant_spawn(
641 : conf,
642 : tenant_shard_id,
643 : &tenant_dir_path,
644 : resources.clone(),
645 : AttachedTenantConf::new(location_conf.tenant_conf, attached_conf),
646 : shard_identity,
647 : Some(init_order.clone()),
648 : SpawnMode::Lazy,
649 : &ctx,
650 : )
651 : .expect("global shutdown during init_tenant_mgr cannot happen"),
652 : ),
653 : LocationMode::Secondary(secondary_conf) => {
654 : info!(
655 : tenant_id = %tenant_shard_id.tenant_id,
656 : shard_id = %tenant_shard_id.shard_slug(),
657 : "Starting secondary tenant"
658 : );
659 : TenantSlot::Secondary(SecondaryTenant::new(
660 : tenant_shard_id,
661 : shard_identity,
662 : location_conf.tenant_conf,
663 : &secondary_conf,
664 : ))
665 : }
666 : };
667 :
668 : METRICS.slot_inserted(&slot);
669 : tenants.insert(tenant_shard_id, slot);
670 : }
671 :
672 : info!("Processed {} local tenants at startup", tenants.len());
673 :
674 : let mut tenants_map = TENANTS.write().unwrap();
675 : assert!(matches!(&*tenants_map, &TenantsMap::Initializing));
676 :
677 : *tenants_map = TenantsMap::Open(tenants);
678 :
679 : Ok(TenantManager {
680 : conf,
681 : tenants: &TENANTS,
682 : resources,
683 : cancel: CancellationToken::new(),
684 : background_purges,
685 : })
686 : }
687 :
688 : /// Wrapper for Tenant::spawn that checks invariants before running
689 : #[allow(clippy::too_many_arguments)]
690 0 : fn tenant_spawn(
691 0 : conf: &'static PageServerConf,
692 0 : tenant_shard_id: TenantShardId,
693 0 : tenant_path: &Utf8Path,
694 0 : resources: TenantSharedResources,
695 0 : location_conf: AttachedTenantConf,
696 0 : shard_identity: ShardIdentity,
697 0 : init_order: Option<InitializationOrder>,
698 0 : mode: SpawnMode,
699 0 : ctx: &RequestContext,
700 0 : ) -> Result<Arc<TenantShard>, GlobalShutDown> {
701 0 : // All these conditions should have been satisfied by our caller: the tenant dir exists, is a well formed
702 0 : // path, and contains a configuration file. Assertions that do synchronous I/O are limited to debug mode
703 0 : // to avoid impacting prod runtime performance.
704 0 : assert!(!crate::is_temporary(tenant_path));
705 0 : debug_assert!(tenant_path.is_dir());
706 0 : debug_assert!(
707 0 : conf.tenant_location_config_path(&tenant_shard_id)
708 0 : .try_exists()
709 0 : .unwrap()
710 : );
711 :
712 0 : TenantShard::spawn(
713 0 : conf,
714 0 : tenant_shard_id,
715 0 : resources,
716 0 : location_conf,
717 0 : shard_identity,
718 0 : init_order,
719 0 : mode,
720 0 : ctx,
721 0 : )
722 0 : }
723 :
724 12 : async fn shutdown_all_tenants0(tenants: &std::sync::RwLock<TenantsMap>) {
725 12 : let mut join_set = JoinSet::new();
726 0 :
727 0 : #[cfg(all(debug_assertions, not(test)))]
728 0 : {
729 0 : // Check that our metrics properly tracked the size of the tenants map. This is a convenient location to check,
730 0 : // as it happens implicitly at the end of tests etc.
731 0 : let m = tenants.read().unwrap();
732 0 : debug_assert_eq!(METRICS.slots_total(), m.len() as u64);
733 : }
734 :
735 : // Atomically, 1. create the shutdown tasks and 2. prevent creation of new tenants.
736 12 : let (total_in_progress, total_attached) = {
737 12 : let mut m = tenants.write().unwrap();
738 12 : match &mut *m {
739 : TenantsMap::Initializing => {
740 0 : *m = TenantsMap::ShuttingDown(BTreeMap::default());
741 0 : info!("tenants map is empty");
742 0 : return;
743 : }
744 12 : TenantsMap::Open(tenants) => {
745 12 : let mut shutdown_state = BTreeMap::new();
746 12 : let mut total_in_progress = 0;
747 12 : let mut total_attached = 0;
748 :
749 12 : for (tenant_shard_id, v) in std::mem::take(tenants).into_iter() {
750 12 : match v {
751 0 : TenantSlot::Attached(t) => {
752 0 : shutdown_state.insert(tenant_shard_id, TenantSlot::Attached(t.clone()));
753 0 : join_set.spawn(
754 0 : async move {
755 0 : let res = {
756 0 : let (_guard, shutdown_progress) = completion::channel();
757 0 : t.shutdown(shutdown_progress, ShutdownMode::FreezeAndFlush).await
758 : };
759 :
760 0 : if let Err(other_progress) = res {
761 : // join the another shutdown in progress
762 0 : other_progress.wait().await;
763 0 : }
764 :
765 : // we cannot afford per tenant logging here, because if s3 is degraded, we are
766 : // going to log too many lines
767 0 : debug!("tenant successfully stopped");
768 0 : }
769 0 : .instrument(info_span!("shutdown", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug())),
770 : );
771 :
772 0 : total_attached += 1;
773 : }
774 0 : TenantSlot::Secondary(state) => {
775 0 : // We don't need to wait for this individually per-tenant: the
776 0 : // downloader task will be waited on eventually, this cancel
777 0 : // is just to encourage it to drop out if it is doing work
778 0 : // for this tenant right now.
779 0 : state.cancel.cancel();
780 0 :
781 0 : shutdown_state.insert(tenant_shard_id, TenantSlot::Secondary(state));
782 0 : }
783 12 : TenantSlot::InProgress(notify) => {
784 12 : // InProgress tenants are not visible in TenantsMap::ShuttingDown: we will
785 12 : // wait for their notifications to fire in this function.
786 12 : join_set.spawn(async move {
787 12 : notify.wait().await;
788 12 : });
789 12 :
790 12 : total_in_progress += 1;
791 12 : }
792 : }
793 : }
794 12 : *m = TenantsMap::ShuttingDown(shutdown_state);
795 12 : (total_in_progress, total_attached)
796 : }
797 : TenantsMap::ShuttingDown(_) => {
798 0 : error!(
799 0 : "already shutting down, this function isn't supposed to be called more than once"
800 : );
801 0 : return;
802 : }
803 : }
804 : };
805 :
806 12 : let started_at = std::time::Instant::now();
807 12 :
808 12 : info!(
809 0 : "Waiting for {} InProgress tenants and {} Attached tenants to shut down",
810 : total_in_progress, total_attached
811 : );
812 :
813 12 : let total = join_set.len();
814 12 : let mut panicked = 0;
815 12 : let mut buffering = true;
816 : const BUFFER_FOR: std::time::Duration = std::time::Duration::from_millis(500);
817 12 : let mut buffered = std::pin::pin!(tokio::time::sleep(BUFFER_FOR));
818 :
819 36 : while !join_set.is_empty() {
820 24 : tokio::select! {
821 24 : Some(joined) = join_set.join_next() => {
822 0 : match joined {
823 12 : Ok(()) => {},
824 0 : Err(join_error) if join_error.is_cancelled() => {
825 0 : unreachable!("we are not cancelling any of the tasks");
826 : }
827 0 : Err(join_error) if join_error.is_panic() => {
828 0 : // cannot really do anything, as this panic is likely a bug
829 0 : panicked += 1;
830 0 : }
831 0 : Err(join_error) => {
832 0 : warn!("unknown kind of JoinError: {join_error}");
833 : }
834 : }
835 12 : if !buffering {
836 12 : // buffer so that every 500ms since the first update (or starting) we'll log
837 12 : // how far away we are; this is because we will get SIGKILL'd at 10s, and we
838 12 : // are not able to log *then*.
839 12 : buffering = true;
840 12 : buffered.as_mut().reset(tokio::time::Instant::now() + BUFFER_FOR);
841 12 : }
842 : },
843 24 : _ = &mut buffered, if buffering => {
844 12 : buffering = false;
845 12 : info!(remaining = join_set.len(), total, elapsed_ms = started_at.elapsed().as_millis(), "waiting for tenants to shutdown");
846 : }
847 : }
848 : }
849 :
850 12 : if panicked > 0 {
851 0 : warn!(
852 : panicked,
853 0 : total, "observed panicks while shutting down tenants"
854 : );
855 12 : }
856 :
857 : // caller will log how long we took
858 12 : }
859 :
860 : #[derive(thiserror::Error, Debug)]
861 : pub(crate) enum UpsertLocationError {
862 : #[error("Bad config request: {0}")]
863 : BadRequest(anyhow::Error),
864 :
865 : #[error("Cannot change config in this state: {0}")]
866 : Unavailable(#[from] TenantMapError),
867 :
868 : #[error("Tenant is already being modified")]
869 : InProgress,
870 :
871 : #[error("Failed to flush: {0}")]
872 : Flush(anyhow::Error),
873 :
874 : /// This error variant is for unexpected situations (soft assertions) where the system is in an unexpected state.
875 : #[error("Internal error: {0}")]
876 : InternalError(anyhow::Error),
877 : }
878 :
879 : impl TenantManager {
880 : /// Convenience function so that anyone with a TenantManager can get at the global configuration, without
881 : /// having to pass it around everywhere as a separate object.
882 0 : pub(crate) fn get_conf(&self) -> &'static PageServerConf {
883 0 : self.conf
884 0 : }
885 :
886 : /// Gets the attached tenant from the in-memory data, erroring if it's absent, in secondary mode, or currently
887 : /// undergoing a state change (i.e. slot is InProgress).
888 : ///
889 : /// The return TenantShard is not guaranteed to be active: check its status after obtaing it, or
890 : /// use [`TenantShard::wait_to_become_active`] before using it if you will do I/O on it.
891 0 : pub(crate) fn get_attached_tenant_shard(
892 0 : &self,
893 0 : tenant_shard_id: TenantShardId,
894 0 : ) -> Result<Arc<TenantShard>, GetTenantError> {
895 0 : let locked = self.tenants.read().unwrap();
896 :
897 0 : let peek_slot = tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Read)?;
898 :
899 0 : match peek_slot {
900 0 : Some(TenantSlot::Attached(tenant)) => Ok(Arc::clone(tenant)),
901 0 : Some(TenantSlot::InProgress(_)) => Err(GetTenantError::NotActive(tenant_shard_id)),
902 : None | Some(TenantSlot::Secondary(_)) => {
903 0 : Err(GetTenantError::ShardNotFound(tenant_shard_id))
904 : }
905 : }
906 0 : }
907 :
908 0 : pub(crate) fn get_secondary_tenant_shard(
909 0 : &self,
910 0 : tenant_shard_id: TenantShardId,
911 0 : ) -> Option<Arc<SecondaryTenant>> {
912 0 : let locked = self.tenants.read().unwrap();
913 0 :
914 0 : let peek_slot = tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Read)
915 0 : .ok()
916 0 : .flatten();
917 :
918 0 : match peek_slot {
919 0 : Some(TenantSlot::Secondary(s)) => Some(s.clone()),
920 0 : _ => None,
921 : }
922 0 : }
923 :
924 : /// Whether the `TenantManager` is responsible for the tenant shard
925 0 : pub(crate) fn manages_tenant_shard(&self, tenant_shard_id: TenantShardId) -> bool {
926 0 : let locked = self.tenants.read().unwrap();
927 0 :
928 0 : let peek_slot = tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Read)
929 0 : .ok()
930 0 : .flatten();
931 0 :
932 0 : peek_slot.is_some()
933 0 : }
934 :
935 : #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
936 : pub(crate) async fn upsert_location(
937 : &self,
938 : tenant_shard_id: TenantShardId,
939 : new_location_config: LocationConf,
940 : flush: Option<Duration>,
941 : mut spawn_mode: SpawnMode,
942 : ctx: &RequestContext,
943 : ) -> Result<Option<Arc<TenantShard>>, UpsertLocationError> {
944 : debug_assert_current_span_has_tenant_id();
945 : info!("configuring tenant location to state {new_location_config:?}");
946 :
947 : enum FastPathModified {
948 : Attached(Arc<TenantShard>),
949 : Secondary(Arc<SecondaryTenant>),
950 : }
951 :
952 : // Special case fast-path for updates to existing slots: if our upsert is only updating configuration,
953 : // then we do not need to set the slot to InProgress, we can just call into the
954 : // existng tenant.
955 : let fast_path_taken = {
956 : let locked = self.tenants.read().unwrap();
957 : let peek_slot =
958 : tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Write)?;
959 : match (&new_location_config.mode, peek_slot) {
960 : (LocationMode::Attached(attach_conf), Some(TenantSlot::Attached(tenant))) => {
961 : match attach_conf.generation.cmp(&tenant.generation) {
962 : Ordering::Equal => {
963 : // A transition from Attached to Attached in the same generation, we may
964 : // take our fast path and just provide the updated configuration
965 : // to the tenant.
966 : tenant.set_new_location_config(
967 : AttachedTenantConf::try_from(new_location_config.clone())
968 : .map_err(UpsertLocationError::BadRequest)?,
969 : );
970 :
971 : Some(FastPathModified::Attached(tenant.clone()))
972 : }
973 : Ordering::Less => {
974 : return Err(UpsertLocationError::BadRequest(anyhow::anyhow!(
975 : "Generation {:?} is less than existing {:?}",
976 : attach_conf.generation,
977 : tenant.generation
978 : )));
979 : }
980 : Ordering::Greater => {
981 : // Generation advanced, fall through to general case of replacing `Tenant` object
982 : None
983 : }
984 : }
985 : }
986 : (
987 : LocationMode::Secondary(secondary_conf),
988 : Some(TenantSlot::Secondary(secondary_tenant)),
989 : ) => {
990 : secondary_tenant.set_config(secondary_conf);
991 : secondary_tenant.set_tenant_conf(&new_location_config.tenant_conf);
992 : Some(FastPathModified::Secondary(secondary_tenant.clone()))
993 : }
994 : _ => {
995 : // Not an Attached->Attached transition, fall through to general case
996 : None
997 : }
998 : }
999 : };
1000 :
1001 : // Fast-path continued: having dropped out of the self.tenants lock, do the async
1002 : // phase of writing config and/or waiting for flush, before returning.
1003 : match fast_path_taken {
1004 : Some(FastPathModified::Attached(tenant)) => {
1005 : TenantShard::persist_tenant_config(
1006 : self.conf,
1007 : &tenant_shard_id,
1008 : &new_location_config,
1009 : )
1010 : .await
1011 : .fatal_err("write tenant shard config");
1012 :
1013 : // Transition to AttachedStale means we may well hold a valid generation
1014 : // still, and have been requested to go stale as part of a migration. If
1015 : // the caller set `flush`, then flush to remote storage.
1016 : if let LocationMode::Attached(AttachedLocationConfig {
1017 : generation: _,
1018 : attach_mode: AttachmentMode::Stale,
1019 : }) = &new_location_config.mode
1020 : {
1021 : if let Some(flush_timeout) = flush {
1022 : match tokio::time::timeout(flush_timeout, tenant.flush_remote()).await {
1023 : Ok(Err(e)) => {
1024 : return Err(UpsertLocationError::Flush(e));
1025 : }
1026 : Ok(Ok(_)) => return Ok(Some(tenant)),
1027 : Err(_) => {
1028 : tracing::warn!(
1029 : timeout_ms = flush_timeout.as_millis(),
1030 : "Timed out waiting for flush to remote storage, proceeding anyway."
1031 : )
1032 : }
1033 : }
1034 : }
1035 : }
1036 :
1037 : return Ok(Some(tenant));
1038 : }
1039 : Some(FastPathModified::Secondary(_secondary_tenant)) => {
1040 : TenantShard::persist_tenant_config(
1041 : self.conf,
1042 : &tenant_shard_id,
1043 : &new_location_config,
1044 : )
1045 : .await
1046 : .fatal_err("write tenant shard config");
1047 :
1048 : return Ok(None);
1049 : }
1050 : None => {
1051 : // Proceed with the general case procedure, where we will shutdown & remove any existing
1052 : // slot contents and replace with a fresh one
1053 : }
1054 : };
1055 :
1056 : // General case for upserts to TenantsMap, excluding the case above: we will substitute an
1057 : // InProgress value to the slot while we make whatever changes are required. The state for
1058 : // the tenant is inaccessible to the outside world while we are doing this, but that is sensible:
1059 : // the state is ill-defined while we're in transition. Transitions are async, but fast: we do
1060 : // not do significant I/O, and shutdowns should be prompt via cancellation tokens.
1061 : let mut slot_guard = tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)
1062 0 : .map_err(|e| match e {
1063 : TenantSlotError::NotFound(_) => {
1064 0 : unreachable!("Called with mode Any")
1065 : }
1066 0 : TenantSlotError::InProgress => UpsertLocationError::InProgress,
1067 0 : TenantSlotError::MapState(s) => UpsertLocationError::Unavailable(s),
1068 0 : })?;
1069 :
1070 : match slot_guard.get_old_value() {
1071 : Some(TenantSlot::Attached(tenant)) => {
1072 : // The case where we keep a Tenant alive was covered above in the special case
1073 : // for Attached->Attached transitions in the same generation. By this point,
1074 : // if we see an attached tenant we know it will be discarded and should be
1075 : // shut down.
1076 : let (_guard, progress) = utils::completion::channel();
1077 :
1078 : match tenant.get_attach_mode() {
1079 : AttachmentMode::Single | AttachmentMode::Multi => {
1080 : // Before we leave our state as the presumed holder of the latest generation,
1081 : // flush any outstanding deletions to reduce the risk of leaking objects.
1082 : self.resources.deletion_queue_client.flush_advisory()
1083 : }
1084 : AttachmentMode::Stale => {
1085 : // If we're stale there's not point trying to flush deletions
1086 : }
1087 : };
1088 :
1089 : info!("Shutting down attached tenant");
1090 : match tenant.shutdown(progress, ShutdownMode::Hard).await {
1091 : Ok(()) => {}
1092 : Err(barrier) => {
1093 : info!("Shutdown already in progress, waiting for it to complete");
1094 : barrier.wait().await;
1095 : }
1096 : }
1097 : slot_guard.drop_old_value().expect("We just shut it down");
1098 :
1099 : // Edge case: if we were called with SpawnMode::Create, but a Tenant already existed, then
1100 : // the caller thinks they're creating but the tenant already existed. We must switch to
1101 : // Eager mode so that when starting this Tenant we properly probe remote storage for timelines,
1102 : // rather than assuming it to be empty.
1103 : spawn_mode = SpawnMode::Eager;
1104 : }
1105 : Some(TenantSlot::Secondary(state)) => {
1106 : info!("Shutting down secondary tenant");
1107 : state.shutdown().await;
1108 : }
1109 : Some(TenantSlot::InProgress(_)) => {
1110 : // This should never happen: acquire_slot should error out
1111 : // if the contents of a slot were InProgress.
1112 : return Err(UpsertLocationError::InternalError(anyhow::anyhow!(
1113 : "Acquired an InProgress slot, this is a bug."
1114 : )));
1115 : }
1116 : None => {
1117 : // Slot was vacant, nothing needs shutting down.
1118 : }
1119 : }
1120 :
1121 : let tenant_path = self.conf.tenant_path(&tenant_shard_id);
1122 : let timelines_path = self.conf.timelines_path(&tenant_shard_id);
1123 :
1124 : // Directory structure is the same for attached and secondary modes:
1125 : // create it if it doesn't exist. Timeline load/creation expects the
1126 : // timelines/ subdir to already exist.
1127 : //
1128 : // Does not need to be fsync'd because local storage is just a cache.
1129 : tokio::fs::create_dir_all(&timelines_path)
1130 : .await
1131 : .fatal_err("create timelines/ dir");
1132 :
1133 : // Before activating either secondary or attached mode, persist the
1134 : // configuration, so that on restart we will re-attach (or re-start
1135 : // secondary) on the tenant.
1136 : TenantShard::persist_tenant_config(self.conf, &tenant_shard_id, &new_location_config)
1137 : .await
1138 : .fatal_err("write tenant shard config");
1139 :
1140 : let new_slot = match &new_location_config.mode {
1141 : LocationMode::Secondary(secondary_config) => {
1142 : let shard_identity = new_location_config.shard;
1143 : TenantSlot::Secondary(SecondaryTenant::new(
1144 : tenant_shard_id,
1145 : shard_identity,
1146 : new_location_config.tenant_conf,
1147 : secondary_config,
1148 : ))
1149 : }
1150 : LocationMode::Attached(_attach_config) => {
1151 : let shard_identity = new_location_config.shard;
1152 :
1153 : // Testing hack: if we are configured with no control plane, then drop the generation
1154 : // from upserts. This enables creating generation-less tenants even though neon_local
1155 : // always uses generations when calling the location conf API.
1156 : let attached_conf = if cfg!(feature = "testing") {
1157 : let mut conf = AttachedTenantConf::try_from(new_location_config)
1158 : .map_err(UpsertLocationError::BadRequest)?;
1159 : if self.conf.control_plane_api.is_none() {
1160 : conf.location.generation = Generation::none();
1161 : }
1162 : conf
1163 : } else {
1164 : AttachedTenantConf::try_from(new_location_config)
1165 : .map_err(UpsertLocationError::BadRequest)?
1166 : };
1167 :
1168 : let tenant = tenant_spawn(
1169 : self.conf,
1170 : tenant_shard_id,
1171 : &tenant_path,
1172 : self.resources.clone(),
1173 : attached_conf,
1174 : shard_identity,
1175 : None,
1176 : spawn_mode,
1177 : ctx,
1178 : )
1179 0 : .map_err(|_: GlobalShutDown| {
1180 0 : UpsertLocationError::Unavailable(TenantMapError::ShuttingDown)
1181 0 : })?;
1182 :
1183 : TenantSlot::Attached(tenant)
1184 : }
1185 : };
1186 :
1187 : let attached_tenant = if let TenantSlot::Attached(tenant) = &new_slot {
1188 : Some(tenant.clone())
1189 : } else {
1190 : None
1191 : };
1192 :
1193 : match slot_guard.upsert(new_slot) {
1194 : Err(TenantSlotUpsertError::InternalError(e)) => {
1195 : Err(UpsertLocationError::InternalError(anyhow::anyhow!(e)))
1196 : }
1197 : Err(TenantSlotUpsertError::MapState(e)) => Err(UpsertLocationError::Unavailable(e)),
1198 : Err(TenantSlotUpsertError::ShuttingDown((new_slot, _completion))) => {
1199 : // If we just called tenant_spawn() on a new tenant, and can't insert it into our map, then
1200 : // we must not leak it: this would violate the invariant that after shutdown_all_tenants, all tenants
1201 : // are shutdown.
1202 : //
1203 : // We must shut it down inline here.
1204 : match new_slot {
1205 : TenantSlot::InProgress(_) => {
1206 : // Unreachable because we never insert an InProgress
1207 : unreachable!()
1208 : }
1209 : TenantSlot::Attached(tenant) => {
1210 : let (_guard, progress) = utils::completion::channel();
1211 : info!(
1212 : "Shutting down just-spawned tenant, because tenant manager is shut down"
1213 : );
1214 : match tenant.shutdown(progress, ShutdownMode::Hard).await {
1215 : Ok(()) => {
1216 : info!("Finished shutting down just-spawned tenant");
1217 : }
1218 : Err(barrier) => {
1219 : info!("Shutdown already in progress, waiting for it to complete");
1220 : barrier.wait().await;
1221 : }
1222 : }
1223 : }
1224 : TenantSlot::Secondary(secondary_tenant) => {
1225 : secondary_tenant.shutdown().await;
1226 : }
1227 : }
1228 :
1229 : Err(UpsertLocationError::Unavailable(
1230 : TenantMapError::ShuttingDown,
1231 : ))
1232 : }
1233 : Ok(()) => Ok(attached_tenant),
1234 : }
1235 : }
1236 :
1237 : /// Resetting a tenant is equivalent to detaching it, then attaching it again with the same
1238 : /// LocationConf that was last used to attach it. Optionally, the local file cache may be
1239 : /// dropped before re-attaching.
1240 : ///
1241 : /// This is not part of a tenant's normal lifecycle: it is used for debug/support, in situations
1242 : /// where an issue is identified that would go away with a restart of the tenant.
1243 : ///
1244 : /// This does not have any special "force" shutdown of a tenant: it relies on the tenant's tasks
1245 : /// to respect the cancellation tokens used in normal shutdown().
1246 : #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), %drop_cache))]
1247 : pub(crate) async fn reset_tenant(
1248 : &self,
1249 : tenant_shard_id: TenantShardId,
1250 : drop_cache: bool,
1251 : ctx: &RequestContext,
1252 : ) -> anyhow::Result<()> {
1253 : let mut slot_guard = tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)?;
1254 : let Some(old_slot) = slot_guard.get_old_value() else {
1255 : anyhow::bail!("Tenant not found when trying to reset");
1256 : };
1257 :
1258 : let Some(tenant) = old_slot.get_attached() else {
1259 : slot_guard.revert();
1260 : anyhow::bail!("Tenant is not in attached state");
1261 : };
1262 :
1263 : let (_guard, progress) = utils::completion::channel();
1264 : match tenant.shutdown(progress, ShutdownMode::Hard).await {
1265 : Ok(()) => {
1266 : slot_guard.drop_old_value()?;
1267 : }
1268 : Err(_barrier) => {
1269 : slot_guard.revert();
1270 : anyhow::bail!("Cannot reset Tenant, already shutting down");
1271 : }
1272 : }
1273 :
1274 : let tenant_path = self.conf.tenant_path(&tenant_shard_id);
1275 : let timelines_path = self.conf.timelines_path(&tenant_shard_id);
1276 : let config = TenantShard::load_tenant_config(self.conf, &tenant_shard_id)?;
1277 :
1278 : if drop_cache {
1279 : tracing::info!("Dropping local file cache");
1280 :
1281 : match tokio::fs::read_dir(&timelines_path).await {
1282 : Err(e) => {
1283 : tracing::warn!("Failed to list timelines while dropping cache: {}", e);
1284 : }
1285 : Ok(mut entries) => {
1286 : while let Some(entry) = entries.next_entry().await? {
1287 : tokio::fs::remove_dir_all(entry.path()).await?;
1288 : }
1289 : }
1290 : }
1291 : }
1292 :
1293 : let shard_identity = config.shard;
1294 : let tenant = tenant_spawn(
1295 : self.conf,
1296 : tenant_shard_id,
1297 : &tenant_path,
1298 : self.resources.clone(),
1299 : AttachedTenantConf::try_from(config)?,
1300 : shard_identity,
1301 : None,
1302 : SpawnMode::Eager,
1303 : ctx,
1304 : )?;
1305 :
1306 : slot_guard.upsert(TenantSlot::Attached(tenant))?;
1307 :
1308 : Ok(())
1309 : }
1310 :
1311 0 : pub(crate) fn get_attached_active_tenant_shards(&self) -> Vec<Arc<TenantShard>> {
1312 0 : let locked = self.tenants.read().unwrap();
1313 0 : match &*locked {
1314 0 : TenantsMap::Initializing => Vec::new(),
1315 0 : TenantsMap::Open(map) | TenantsMap::ShuttingDown(map) => map
1316 0 : .values()
1317 0 : .filter_map(|slot| {
1318 0 : slot.get_attached()
1319 0 : .and_then(|t| if t.is_active() { Some(t.clone()) } else { None })
1320 0 : })
1321 0 : .collect(),
1322 : }
1323 0 : }
1324 : // Do some synchronous work for all tenant slots in Secondary state. The provided
1325 : // callback should be small and fast, as it will be called inside the global
1326 : // TenantsMap lock.
1327 0 : pub(crate) fn foreach_secondary_tenants<F>(&self, mut func: F)
1328 0 : where
1329 0 : // TODO: let the callback return a hint to drop out of the loop early
1330 0 : F: FnMut(&TenantShardId, &Arc<SecondaryTenant>),
1331 0 : {
1332 0 : let locked = self.tenants.read().unwrap();
1333 :
1334 0 : let map = match &*locked {
1335 0 : TenantsMap::Initializing | TenantsMap::ShuttingDown(_) => return,
1336 0 : TenantsMap::Open(m) => m,
1337 : };
1338 :
1339 0 : for (tenant_id, slot) in map {
1340 0 : if let TenantSlot::Secondary(state) = slot {
1341 : // Only expose secondary tenants that are not currently shutting down
1342 0 : if !state.cancel.is_cancelled() {
1343 0 : func(tenant_id, state)
1344 0 : }
1345 0 : }
1346 : }
1347 0 : }
1348 :
1349 : /// Total list of all tenant slots: this includes attached, secondary, and InProgress.
1350 0 : pub(crate) fn list(&self) -> Vec<(TenantShardId, TenantSlot)> {
1351 0 : let locked = self.tenants.read().unwrap();
1352 0 : match &*locked {
1353 0 : TenantsMap::Initializing => Vec::new(),
1354 0 : TenantsMap::Open(map) | TenantsMap::ShuttingDown(map) => {
1355 0 : map.iter().map(|(k, v)| (*k, v.clone())).collect()
1356 : }
1357 : }
1358 0 : }
1359 :
1360 0 : pub(crate) fn get(&self, tenant_shard_id: TenantShardId) -> Option<TenantSlot> {
1361 0 : let locked = self.tenants.read().unwrap();
1362 0 : match &*locked {
1363 0 : TenantsMap::Initializing => None,
1364 0 : TenantsMap::Open(map) | TenantsMap::ShuttingDown(map) => {
1365 0 : map.get(&tenant_shard_id).cloned()
1366 : }
1367 : }
1368 0 : }
1369 :
1370 : /// If a tenant is attached, detach it. Then remove its data from remote storage.
1371 : ///
1372 : /// A tenant is considered deleted once it is gone from remote storage. It is the caller's
1373 : /// responsibility to avoid trying to attach the tenant again or use it any way once deletion
1374 : /// has started: this operation is not atomic, and must be retried until it succeeds.
1375 : ///
1376 : /// As a special case, if an unsharded tenant ID is given for a sharded tenant, it will remove
1377 : /// all tenant shards in remote storage (removing all paths with the tenant prefix). The storage
1378 : /// controller uses this to purge all remote tenant data, including any stale parent shards that
1379 : /// may remain after splits. Ideally, this special case would be handled elsewhere. See:
1380 : /// <https://github.com/neondatabase/neon/pull/9394>.
1381 0 : pub(crate) async fn delete_tenant(
1382 0 : &self,
1383 0 : tenant_shard_id: TenantShardId,
1384 0 : ) -> Result<(), DeleteTenantError> {
1385 0 : super::span::debug_assert_current_span_has_tenant_id();
1386 :
1387 0 : async fn delete_local(
1388 0 : conf: &PageServerConf,
1389 0 : background_purges: &BackgroundPurges,
1390 0 : tenant_shard_id: &TenantShardId,
1391 0 : ) -> anyhow::Result<()> {
1392 0 : let local_tenant_directory = conf.tenant_path(tenant_shard_id);
1393 0 : let tmp_dir = safe_rename_tenant_dir(&local_tenant_directory)
1394 0 : .await
1395 0 : .with_context(|| {
1396 0 : format!("local tenant directory {local_tenant_directory:?} rename")
1397 0 : })?;
1398 0 : background_purges.spawn(tmp_dir);
1399 0 : Ok(())
1400 0 : }
1401 :
1402 0 : let slot_guard = tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)?;
1403 0 : match &slot_guard.old_value {
1404 0 : Some(TenantSlot::Attached(tenant)) => {
1405 0 : // Legacy deletion flow: the tenant remains attached, goes to Stopping state, and
1406 0 : // deletion will be resumed across restarts.
1407 0 : let tenant = tenant.clone();
1408 0 : let (_guard, progress) = utils::completion::channel();
1409 0 : match tenant.shutdown(progress, ShutdownMode::Hard).await {
1410 0 : Ok(()) => {}
1411 0 : Err(barrier) => {
1412 0 : info!("Shutdown already in progress, waiting for it to complete");
1413 0 : barrier.wait().await;
1414 : }
1415 : }
1416 0 : delete_local(self.conf, &self.background_purges, &tenant_shard_id).await?;
1417 : }
1418 0 : Some(TenantSlot::Secondary(secondary_tenant)) => {
1419 0 : secondary_tenant.shutdown().await;
1420 :
1421 0 : delete_local(self.conf, &self.background_purges, &tenant_shard_id).await?;
1422 : }
1423 0 : Some(TenantSlot::InProgress(_)) => unreachable!(),
1424 0 : None => {}
1425 : };
1426 :
1427 : // Fall through: local state for this tenant is no longer present, proceed with remote delete.
1428 : // - We use a retry wrapper here so that common transient S3 errors (e.g. 503, 429) do not result
1429 : // in 500 responses to delete requests.
1430 : // - We keep the `SlotGuard` during this I/O, so that if a concurrent delete request comes in, it will
1431 : // 503/retry, rather than kicking off a wasteful concurrent deletion.
1432 : // NB: this also deletes partial prefixes, i.e. a <tenant_id> path will delete all
1433 : // <tenant_id>_<shard_id>/* objects. See method comment for why.
1434 0 : backoff::retry(
1435 0 : || async move {
1436 0 : self.resources
1437 0 : .remote_storage
1438 0 : .delete_prefix(&remote_tenant_path(&tenant_shard_id), &self.cancel)
1439 0 : .await
1440 0 : },
1441 0 : |_| false, // backoff::retry handles cancellation
1442 0 : 1,
1443 0 : 3,
1444 0 : &format!("delete_tenant[tenant_shard_id={tenant_shard_id}]"),
1445 0 : &self.cancel,
1446 0 : )
1447 0 : .await
1448 0 : .unwrap_or(Err(TimeoutOrCancel::Cancel.into()))
1449 0 : .map_err(|err| {
1450 0 : if TimeoutOrCancel::caused_by_cancel(&err) {
1451 0 : return DeleteTenantError::Cancelled;
1452 0 : }
1453 0 : DeleteTenantError::Other(err)
1454 0 : })
1455 0 : }
1456 :
1457 : #[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()))]
1458 : pub(crate) async fn shard_split(
1459 : &self,
1460 : tenant: Arc<TenantShard>,
1461 : new_shard_count: ShardCount,
1462 : new_stripe_size: Option<ShardStripeSize>,
1463 : ctx: &RequestContext,
1464 : ) -> anyhow::Result<Vec<TenantShardId>> {
1465 : let tenant_shard_id = *tenant.get_tenant_shard_id();
1466 : let r = self
1467 : .do_shard_split(tenant, new_shard_count, new_stripe_size, ctx)
1468 : .await;
1469 : if r.is_err() {
1470 : // Shard splitting might have left the original shard in a partially shut down state (it
1471 : // stops the shard's remote timeline client). Reset it to ensure we leave things in
1472 : // a working state.
1473 : if self.get(tenant_shard_id).is_some() {
1474 : tracing::warn!("Resetting after shard split failure");
1475 : if let Err(e) = self.reset_tenant(tenant_shard_id, false, ctx).await {
1476 : // Log this error because our return value will still be the original error, not this one. This is
1477 : // a severe error: if this happens, we might be leaving behind a tenant that is not fully functional
1478 : // (e.g. has uploads disabled). We can't do anything else: if reset fails then shutting the tenant down or
1479 : // setting it broken probably won't help either.
1480 : tracing::error!("Failed to reset: {e}");
1481 : }
1482 : }
1483 : }
1484 :
1485 : r
1486 : }
1487 :
1488 0 : pub(crate) async fn do_shard_split(
1489 0 : &self,
1490 0 : tenant: Arc<TenantShard>,
1491 0 : new_shard_count: ShardCount,
1492 0 : new_stripe_size: Option<ShardStripeSize>,
1493 0 : ctx: &RequestContext,
1494 0 : ) -> anyhow::Result<Vec<TenantShardId>> {
1495 0 : let tenant_shard_id = *tenant.get_tenant_shard_id();
1496 0 :
1497 0 : // Validate the incoming request
1498 0 : if new_shard_count.count() <= tenant_shard_id.shard_count.count() {
1499 0 : anyhow::bail!("Requested shard count is not an increase");
1500 0 : }
1501 0 : let expansion_factor = new_shard_count.count() / tenant_shard_id.shard_count.count();
1502 0 : if !expansion_factor.is_power_of_two() {
1503 0 : anyhow::bail!("Requested split is not a power of two");
1504 0 : }
1505 :
1506 0 : if let Some(new_stripe_size) = new_stripe_size {
1507 0 : if tenant.get_shard_stripe_size() != new_stripe_size
1508 0 : && tenant_shard_id.shard_count.count() > 1
1509 : {
1510 : // This tenant already has multiple shards, it is illegal to try and change its stripe size
1511 0 : anyhow::bail!(
1512 0 : "Shard stripe size may not be modified once tenant has multiple shards"
1513 0 : );
1514 0 : }
1515 0 : }
1516 :
1517 : // Plan: identify what the new child shards will be
1518 0 : let child_shards = tenant_shard_id.split(new_shard_count);
1519 0 : tracing::info!(
1520 0 : "Shard {} splits into: {}",
1521 0 : tenant_shard_id.to_index(),
1522 0 : child_shards
1523 0 : .iter()
1524 0 : .map(|id| format!("{}", id.to_index()))
1525 0 : .join(",")
1526 : );
1527 :
1528 0 : fail::fail_point!("shard-split-pre-prepare", |_| Err(anyhow::anyhow!(
1529 0 : "failpoint"
1530 0 : )));
1531 :
1532 0 : let parent_shard_identity = tenant.shard_identity;
1533 0 : let parent_tenant_conf = tenant.get_tenant_conf();
1534 0 : let parent_generation = tenant.generation;
1535 :
1536 : // Phase 1: Write out child shards' remote index files, in the parent tenant's current generation
1537 0 : if let Err(e) = tenant.split_prepare(&child_shards).await {
1538 : // If [`Tenant::split_prepare`] fails, we must reload the tenant, because it might
1539 : // have been left in a partially-shut-down state.
1540 0 : tracing::warn!("Failed to prepare for split: {e}, reloading Tenant before returning");
1541 0 : return Err(e);
1542 0 : }
1543 0 :
1544 0 : fail::fail_point!("shard-split-post-prepare", |_| Err(anyhow::anyhow!(
1545 0 : "failpoint"
1546 0 : )));
1547 :
1548 0 : self.resources.deletion_queue_client.flush_advisory();
1549 0 :
1550 0 : // Phase 2: Put the parent shard to InProgress and grab a reference to the parent Tenant
1551 0 : drop(tenant);
1552 0 : let mut parent_slot_guard =
1553 0 : tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)?;
1554 0 : let parent = match parent_slot_guard.get_old_value() {
1555 0 : Some(TenantSlot::Attached(t)) => t,
1556 0 : Some(TenantSlot::Secondary(_)) => anyhow::bail!("Tenant location in secondary mode"),
1557 : Some(TenantSlot::InProgress(_)) => {
1558 : // tenant_map_acquire_slot never returns InProgress, if a slot was InProgress
1559 : // it would return an error.
1560 0 : unreachable!()
1561 : }
1562 : None => {
1563 : // We don't actually need the parent shard to still be attached to do our work, but it's
1564 : // a weird enough situation that the caller probably didn't want us to continue working
1565 : // if they had detached the tenant they requested the split on.
1566 0 : anyhow::bail!("Detached parent shard in the middle of split!")
1567 : }
1568 : };
1569 0 : fail::fail_point!("shard-split-pre-hardlink", |_| Err(anyhow::anyhow!(
1570 0 : "failpoint"
1571 0 : )));
1572 : // Optimization: hardlink layers from the parent into the children, so that they don't have to
1573 : // re-download & duplicate the data referenced in their initial IndexPart
1574 0 : self.shard_split_hardlink(parent, child_shards.clone())
1575 0 : .await?;
1576 0 : fail::fail_point!("shard-split-post-hardlink", |_| Err(anyhow::anyhow!(
1577 0 : "failpoint"
1578 0 : )));
1579 :
1580 : // Take a snapshot of where the parent's WAL ingest had got to: we will wait for
1581 : // child shards to reach this point.
1582 0 : let mut target_lsns = HashMap::new();
1583 0 : for timeline in parent.timelines.lock().unwrap().clone().values() {
1584 0 : target_lsns.insert(timeline.timeline_id, timeline.get_last_record_lsn());
1585 0 : }
1586 :
1587 : // TODO: we should have the parent shard stop its WAL ingest here, it's a waste of resources
1588 : // and could slow down the children trying to catch up.
1589 :
1590 : // Phase 3: Spawn the child shards
1591 0 : for child_shard in &child_shards {
1592 0 : let mut child_shard_identity = parent_shard_identity;
1593 0 : if let Some(new_stripe_size) = new_stripe_size {
1594 0 : child_shard_identity.stripe_size = new_stripe_size;
1595 0 : }
1596 0 : child_shard_identity.count = child_shard.shard_count;
1597 0 : child_shard_identity.number = child_shard.shard_number;
1598 0 :
1599 0 : let child_location_conf = LocationConf {
1600 0 : mode: LocationMode::Attached(AttachedLocationConfig {
1601 0 : generation: parent_generation,
1602 0 : attach_mode: AttachmentMode::Single,
1603 0 : }),
1604 0 : shard: child_shard_identity,
1605 0 : tenant_conf: parent_tenant_conf.clone(),
1606 0 : };
1607 0 :
1608 0 : self.upsert_location(
1609 0 : *child_shard,
1610 0 : child_location_conf,
1611 0 : None,
1612 0 : SpawnMode::Eager,
1613 0 : ctx,
1614 0 : )
1615 0 : .await?;
1616 : }
1617 :
1618 0 : fail::fail_point!("shard-split-post-child-conf", |_| Err(anyhow::anyhow!(
1619 0 : "failpoint"
1620 0 : )));
1621 :
1622 : // Phase 4: wait for child chards WAL ingest to catch up to target LSN
1623 0 : for child_shard_id in &child_shards {
1624 0 : let child_shard_id = *child_shard_id;
1625 0 : let child_shard = {
1626 0 : let locked = self.tenants.read().unwrap();
1627 0 : let peek_slot =
1628 0 : tenant_map_peek_slot(&locked, &child_shard_id, TenantSlotPeekMode::Read)?;
1629 0 : peek_slot.and_then(|s| s.get_attached()).cloned()
1630 : };
1631 0 : if let Some(t) = child_shard {
1632 : // Wait for the child shard to become active: this should be very quick because it only
1633 : // has to download the index_part that we just uploaded when creating it.
1634 0 : if let Err(e) = t.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await {
1635 : // This is not fatal: we have durably created the child shard. It just makes the
1636 : // split operation less seamless for clients, as we will may detach the parent
1637 : // shard before the child shards are fully ready to serve requests.
1638 0 : tracing::warn!("Failed to wait for shard {child_shard_id} to activate: {e}");
1639 0 : continue;
1640 0 : }
1641 0 :
1642 0 : let timelines = t.timelines.lock().unwrap().clone();
1643 0 : for timeline in timelines.values() {
1644 0 : let Some(target_lsn) = target_lsns.get(&timeline.timeline_id) else {
1645 0 : continue;
1646 : };
1647 :
1648 0 : tracing::info!(
1649 0 : "Waiting for child shard {}/{} to reach target lsn {}...",
1650 0 : child_shard_id,
1651 0 : timeline.timeline_id,
1652 : target_lsn
1653 : );
1654 :
1655 0 : fail::fail_point!("shard-split-lsn-wait", |_| Err(anyhow::anyhow!(
1656 0 : "failpoint"
1657 0 : )));
1658 0 : if let Err(e) = timeline
1659 0 : .wait_lsn(
1660 0 : *target_lsn,
1661 0 : crate::tenant::timeline::WaitLsnWaiter::Tenant,
1662 0 : crate::tenant::timeline::WaitLsnTimeout::Default,
1663 0 : ctx,
1664 0 : )
1665 0 : .await
1666 : {
1667 : // Failure here might mean shutdown, in any case this part is an optimization
1668 : // and we shouldn't hold up the split operation.
1669 0 : tracing::warn!(
1670 0 : "Failed to wait for timeline {} to reach lsn {target_lsn}: {e}",
1671 0 : timeline.timeline_id
1672 : );
1673 : } else {
1674 0 : tracing::info!(
1675 0 : "Child shard {}/{} reached target lsn {}",
1676 0 : child_shard_id,
1677 0 : timeline.timeline_id,
1678 : target_lsn
1679 : );
1680 : }
1681 : }
1682 0 : }
1683 : }
1684 :
1685 : // Phase 5: Shut down the parent shard, and erase it from disk
1686 0 : let (_guard, progress) = completion::channel();
1687 0 : match parent.shutdown(progress, ShutdownMode::Hard).await {
1688 0 : Ok(()) => {}
1689 0 : Err(other) => {
1690 0 : other.wait().await;
1691 : }
1692 : }
1693 0 : let local_tenant_directory = self.conf.tenant_path(&tenant_shard_id);
1694 0 : let tmp_path = safe_rename_tenant_dir(&local_tenant_directory)
1695 0 : .await
1696 0 : .with_context(|| format!("local tenant directory {local_tenant_directory:?} rename"))?;
1697 0 : self.background_purges.spawn(tmp_path);
1698 0 :
1699 0 : fail::fail_point!("shard-split-pre-finish", |_| Err(anyhow::anyhow!(
1700 0 : "failpoint"
1701 0 : )));
1702 :
1703 0 : parent_slot_guard.drop_old_value()?;
1704 :
1705 : // Phase 6: Release the InProgress on the parent shard
1706 0 : drop(parent_slot_guard);
1707 0 :
1708 0 : Ok(child_shards)
1709 0 : }
1710 :
1711 : /// Part of [`Self::shard_split`]: hard link parent shard layers into child shards, as an optimization
1712 : /// to avoid the children downloading them again.
1713 : ///
1714 : /// For each resident layer in the parent shard, we will hard link it into all of the child shards.
1715 0 : async fn shard_split_hardlink(
1716 0 : &self,
1717 0 : parent_shard: &TenantShard,
1718 0 : child_shards: Vec<TenantShardId>,
1719 0 : ) -> anyhow::Result<()> {
1720 0 : debug_assert_current_span_has_tenant_id();
1721 0 :
1722 0 : let parent_path = self.conf.tenant_path(parent_shard.get_tenant_shard_id());
1723 0 : let (parent_timelines, parent_layers) = {
1724 0 : let mut parent_layers = Vec::new();
1725 0 : let timelines = parent_shard.timelines.lock().unwrap().clone();
1726 0 : let parent_timelines = timelines.keys().cloned().collect::<Vec<_>>();
1727 0 : for timeline in timelines.values() {
1728 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Loading list of layers to hardlink");
1729 0 : let layers = timeline.layers.read().await;
1730 :
1731 0 : for layer in layers.likely_resident_layers() {
1732 0 : let relative_path = layer
1733 0 : .local_path()
1734 0 : .strip_prefix(&parent_path)
1735 0 : .context("Removing prefix from parent layer path")?;
1736 0 : parent_layers.push(relative_path.to_owned());
1737 : }
1738 : }
1739 :
1740 0 : if parent_layers.is_empty() {
1741 0 : tracing::info!("Ancestor shard has no resident layer to hard link");
1742 0 : }
1743 :
1744 0 : (parent_timelines, parent_layers)
1745 0 : };
1746 0 :
1747 0 : let mut child_prefixes = Vec::new();
1748 0 : let mut create_dirs = Vec::new();
1749 :
1750 0 : for child in child_shards {
1751 0 : let child_prefix = self.conf.tenant_path(&child);
1752 0 : create_dirs.push(child_prefix.clone());
1753 0 : create_dirs.extend(
1754 0 : parent_timelines
1755 0 : .iter()
1756 0 : .map(|t| self.conf.timeline_path(&child, t)),
1757 0 : );
1758 0 :
1759 0 : child_prefixes.push(child_prefix);
1760 0 : }
1761 :
1762 : // Since we will do a large number of small filesystem metadata operations, batch them into
1763 : // spawn_blocking calls rather than doing each one as a tokio::fs round-trip.
1764 0 : let span = tracing::Span::current();
1765 0 : let jh = tokio::task::spawn_blocking(move || -> anyhow::Result<usize> {
1766 0 : // Run this synchronous code in the same log context as the outer function that spawned it.
1767 0 : let _span = span.enter();
1768 0 :
1769 0 : tracing::info!("Creating {} directories", create_dirs.len());
1770 0 : for dir in &create_dirs {
1771 0 : if let Err(e) = std::fs::create_dir_all(dir) {
1772 : // Ignore AlreadyExists errors, drop out on all other errors
1773 0 : match e.kind() {
1774 0 : std::io::ErrorKind::AlreadyExists => {}
1775 : _ => {
1776 0 : return Err(anyhow::anyhow!(e).context(format!("Creating {dir}")));
1777 : }
1778 : }
1779 0 : }
1780 : }
1781 :
1782 0 : for child_prefix in child_prefixes {
1783 0 : tracing::info!(
1784 0 : "Hard-linking {} parent layers into child path {}",
1785 0 : parent_layers.len(),
1786 : child_prefix
1787 : );
1788 0 : for relative_layer in &parent_layers {
1789 0 : let parent_path = parent_path.join(relative_layer);
1790 0 : let child_path = child_prefix.join(relative_layer);
1791 0 : if let Err(e) = std::fs::hard_link(&parent_path, &child_path) {
1792 0 : match e.kind() {
1793 0 : std::io::ErrorKind::AlreadyExists => {}
1794 : std::io::ErrorKind::NotFound => {
1795 0 : tracing::info!(
1796 0 : "Layer {} not found during hard-linking, evicted during split?",
1797 : relative_layer
1798 : );
1799 : }
1800 : _ => {
1801 0 : return Err(anyhow::anyhow!(e).context(format!(
1802 0 : "Hard linking {relative_layer} into {child_prefix}"
1803 0 : )));
1804 : }
1805 : }
1806 0 : }
1807 : }
1808 : }
1809 :
1810 : // Durability is not required for correctness, but if we crashed during split and
1811 : // then came restarted with empty timeline dirs, it would be very inefficient to
1812 : // re-populate from remote storage.
1813 0 : tracing::info!("fsyncing {} directories", create_dirs.len());
1814 0 : for dir in create_dirs {
1815 0 : if let Err(e) = crashsafe::fsync(&dir) {
1816 : // Something removed a newly created timeline dir out from underneath us? Extremely
1817 : // unexpected, but not worth panic'ing over as this whole function is just an
1818 : // optimization.
1819 0 : tracing::warn!("Failed to fsync directory {dir}: {e}")
1820 0 : }
1821 : }
1822 :
1823 0 : Ok(parent_layers.len())
1824 0 : });
1825 0 :
1826 0 : match jh.await {
1827 0 : Ok(Ok(layer_count)) => {
1828 0 : tracing::info!(count = layer_count, "Hard linked layers into child shards");
1829 : }
1830 0 : Ok(Err(e)) => {
1831 0 : // This is an optimization, so we tolerate failure.
1832 0 : tracing::warn!("Error hard-linking layers, proceeding anyway: {e}")
1833 : }
1834 0 : Err(e) => {
1835 0 : // This is something totally unexpected like a panic, so bail out.
1836 0 : anyhow::bail!("Error joining hard linking task: {e}");
1837 : }
1838 : }
1839 :
1840 0 : Ok(())
1841 0 : }
1842 :
1843 : ///
1844 : /// Shut down all tenants. This runs as part of pageserver shutdown.
1845 : ///
1846 : /// NB: We leave the tenants in the map, so that they remain accessible through
1847 : /// the management API until we shut it down. If we removed the shut-down tenants
1848 : /// from the tenants map, the management API would return 404 for these tenants,
1849 : /// because TenantsMap::get() now returns `None`.
1850 : /// That could be easily misinterpreted by control plane, the consumer of the
1851 : /// management API. For example, it could attach the tenant on a different pageserver.
1852 : /// We would then be in split-brain once this pageserver restarts.
1853 : #[instrument(skip_all)]
1854 : pub(crate) async fn shutdown(&self) {
1855 : self.cancel.cancel();
1856 :
1857 : shutdown_all_tenants0(self.tenants).await
1858 : }
1859 :
1860 0 : pub(crate) async fn detach_tenant(
1861 0 : &self,
1862 0 : conf: &'static PageServerConf,
1863 0 : tenant_shard_id: TenantShardId,
1864 0 : deletion_queue_client: &DeletionQueueClient,
1865 0 : ) -> Result<(), TenantStateError> {
1866 0 : let tmp_path = self
1867 0 : .detach_tenant0(conf, tenant_shard_id, deletion_queue_client)
1868 0 : .await?;
1869 0 : self.background_purges.spawn(tmp_path);
1870 0 :
1871 0 : Ok(())
1872 0 : }
1873 :
1874 0 : async fn detach_tenant0(
1875 0 : &self,
1876 0 : conf: &'static PageServerConf,
1877 0 : tenant_shard_id: TenantShardId,
1878 0 : deletion_queue_client: &DeletionQueueClient,
1879 0 : ) -> Result<Utf8PathBuf, TenantStateError> {
1880 0 : let tenant_dir_rename_operation = |tenant_id_to_clean: TenantShardId| async move {
1881 0 : let local_tenant_directory = conf.tenant_path(&tenant_id_to_clean);
1882 0 : safe_rename_tenant_dir(&local_tenant_directory)
1883 0 : .await
1884 0 : .with_context(|| {
1885 0 : format!("local tenant directory {local_tenant_directory:?} rename")
1886 0 : })
1887 0 : };
1888 :
1889 0 : let removal_result = remove_tenant_from_memory(
1890 0 : self.tenants,
1891 0 : tenant_shard_id,
1892 0 : tenant_dir_rename_operation(tenant_shard_id),
1893 0 : )
1894 0 : .await;
1895 :
1896 : // Flush pending deletions, so that they have a good chance of passing validation
1897 : // before this tenant is potentially re-attached elsewhere.
1898 0 : deletion_queue_client.flush_advisory();
1899 0 :
1900 0 : removal_result
1901 0 : }
1902 :
1903 0 : pub(crate) fn list_tenants(
1904 0 : &self,
1905 0 : ) -> Result<Vec<(TenantShardId, TenantState, Generation)>, TenantMapListError> {
1906 0 : let tenants = self.tenants.read().unwrap();
1907 0 : let m = match &*tenants {
1908 0 : TenantsMap::Initializing => return Err(TenantMapListError::Initializing),
1909 0 : TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => m,
1910 0 : };
1911 0 : Ok(m.iter()
1912 0 : .filter_map(|(id, tenant)| match tenant {
1913 0 : TenantSlot::Attached(tenant) => {
1914 0 : Some((*id, tenant.current_state(), tenant.generation()))
1915 : }
1916 0 : TenantSlot::Secondary(_) => None,
1917 0 : TenantSlot::InProgress(_) => None,
1918 0 : })
1919 0 : .collect())
1920 0 : }
1921 :
1922 : /// Completes an earlier prepared timeline detach ancestor.
1923 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
1924 0 : &self,
1925 0 : tenant_shard_id: TenantShardId,
1926 0 : timeline_id: TimelineId,
1927 0 : prepared: PreparedTimelineDetach,
1928 0 : behavior: DetachBehavior,
1929 0 : mut attempt: detach_ancestor::Attempt,
1930 0 : ctx: &RequestContext,
1931 0 : ) -> Result<HashSet<TimelineId>, detach_ancestor::Error> {
1932 : use detach_ancestor::Error;
1933 :
1934 0 : let slot_guard =
1935 0 : tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::MustExist).map_err(
1936 0 : |e| {
1937 : use TenantSlotError::*;
1938 :
1939 0 : match e {
1940 0 : MapState(TenantMapError::ShuttingDown) => Error::ShuttingDown,
1941 0 : NotFound(_) | InProgress | MapState(_) => Error::DetachReparent(e.into()),
1942 : }
1943 0 : },
1944 0 : )?;
1945 :
1946 0 : let tenant = {
1947 0 : let old_slot = slot_guard
1948 0 : .get_old_value()
1949 0 : .as_ref()
1950 0 : .expect("requested MustExist");
1951 :
1952 0 : let Some(tenant) = old_slot.get_attached() else {
1953 0 : return Err(Error::DetachReparent(anyhow::anyhow!(
1954 0 : "Tenant is not in attached state"
1955 0 : )));
1956 : };
1957 :
1958 0 : if !tenant.is_active() {
1959 0 : return Err(Error::DetachReparent(anyhow::anyhow!(
1960 0 : "Tenant is not active"
1961 0 : )));
1962 0 : }
1963 0 :
1964 0 : tenant.clone()
1965 : };
1966 :
1967 0 : let timeline = tenant
1968 0 : .get_timeline(timeline_id, true)
1969 0 : .map_err(Error::NotFound)?;
1970 :
1971 0 : let resp = timeline
1972 0 : .detach_from_ancestor_and_reparent(
1973 0 : &tenant,
1974 0 : prepared,
1975 0 : attempt.ancestor_timeline_id,
1976 0 : attempt.ancestor_lsn,
1977 0 : behavior,
1978 0 : ctx,
1979 0 : )
1980 0 : .await?;
1981 :
1982 0 : let mut slot_guard = slot_guard;
1983 :
1984 0 : let tenant = if resp.reset_tenant_required() {
1985 0 : attempt.before_reset_tenant();
1986 0 :
1987 0 : let (_guard, progress) = utils::completion::channel();
1988 0 : match tenant.shutdown(progress, ShutdownMode::Reload).await {
1989 0 : Ok(()) => {
1990 0 : slot_guard.drop_old_value().expect("it was just shutdown");
1991 0 : }
1992 0 : Err(_barrier) => {
1993 0 : slot_guard.revert();
1994 0 : // this really should not happen, at all, unless a shutdown without acquiring
1995 0 : // tenant slot was already going? regardless, on restart the attempt tracking
1996 0 : // will reset to retryable.
1997 0 : return Err(Error::ShuttingDown);
1998 : }
1999 : }
2000 :
2001 0 : let tenant_path = self.conf.tenant_path(&tenant_shard_id);
2002 0 : let config = TenantShard::load_tenant_config(self.conf, &tenant_shard_id)
2003 0 : .map_err(|e| Error::DetachReparent(e.into()))?;
2004 :
2005 0 : let shard_identity = config.shard;
2006 0 : let tenant = tenant_spawn(
2007 0 : self.conf,
2008 0 : tenant_shard_id,
2009 0 : &tenant_path,
2010 0 : self.resources.clone(),
2011 0 : AttachedTenantConf::try_from(config).map_err(Error::DetachReparent)?,
2012 0 : shard_identity,
2013 0 : None,
2014 0 : SpawnMode::Eager,
2015 0 : ctx,
2016 0 : )
2017 0 : .map_err(|_| Error::ShuttingDown)?;
2018 :
2019 : {
2020 0 : let mut g = tenant.ongoing_timeline_detach.lock().unwrap();
2021 0 : assert!(
2022 0 : g.is_none(),
2023 0 : "there cannot be any new timeline detach ancestor on newly created tenant"
2024 : );
2025 0 : *g = Some((attempt.timeline_id, attempt.new_barrier()));
2026 0 : }
2027 0 :
2028 0 : // if we bail out here, we will not allow a new attempt, which should be fine.
2029 0 : // pageserver should be shutting down regardless? tenant_reset would help, unless it
2030 0 : // runs into the same problem.
2031 0 : slot_guard
2032 0 : .upsert(TenantSlot::Attached(tenant.clone()))
2033 0 : .map_err(|e| match e {
2034 0 : TenantSlotUpsertError::ShuttingDown(_) => Error::ShuttingDown,
2035 0 : other => Error::DetachReparent(other.into()),
2036 0 : })?;
2037 0 : tenant
2038 : } else {
2039 0 : tracing::info!("skipping tenant_reset as no changes made required it");
2040 0 : tenant
2041 : };
2042 :
2043 0 : if let Some(reparented) = resp.completed() {
2044 : // finally ask the restarted tenant to complete the detach
2045 : //
2046 : // rationale for 9999s: we don't really have a timetable here; if retried, the caller
2047 : // will get an 503.
2048 0 : tenant
2049 0 : .wait_to_become_active(std::time::Duration::from_secs(9999))
2050 0 : .await
2051 0 : .map_err(|e| {
2052 : use GetActiveTenantError::{Cancelled, WillNotBecomeActive};
2053 : use pageserver_api::models::TenantState;
2054 0 : match e {
2055 : Cancelled | WillNotBecomeActive(TenantState::Stopping { .. }) => {
2056 0 : Error::ShuttingDown
2057 : }
2058 0 : other => Error::Complete(other.into()),
2059 : }
2060 0 : })?;
2061 :
2062 0 : utils::pausable_failpoint!(
2063 0 : "timeline-detach-ancestor::after_activating_before_finding-pausable"
2064 0 : );
2065 :
2066 0 : let timeline = tenant
2067 0 : .get_timeline(attempt.timeline_id, true)
2068 0 : .map_err(Error::NotFound)?;
2069 :
2070 0 : timeline
2071 0 : .complete_detaching_timeline_ancestor(&tenant, attempt, ctx)
2072 0 : .await
2073 0 : .map(|()| reparented)
2074 : } else {
2075 : // at least the latest versions have now been downloaded and refreshed; be ready to
2076 : // retry another time.
2077 0 : Err(Error::FailedToReparentAll)
2078 : }
2079 0 : }
2080 :
2081 : /// A page service client sends a TenantId, and to look up the correct Tenant we must
2082 : /// resolve this to a fully qualified TenantShardId.
2083 : ///
2084 : /// During shard splits: we shall see parent shards in InProgress state and skip them, and
2085 : /// instead match on child shards which should appear in Attached state. Very early in a shard
2086 : /// split, or in other cases where a shard is InProgress, we will return our own InProgress result
2087 : /// to instruct the caller to wait for that to finish before querying again.
2088 0 : pub(crate) fn resolve_attached_shard(
2089 0 : &self,
2090 0 : tenant_id: &TenantId,
2091 0 : selector: ShardSelector,
2092 0 : ) -> ShardResolveResult {
2093 0 : let tenants = self.tenants.read().unwrap();
2094 0 : let mut want_shard = None;
2095 0 : let mut any_in_progress = None;
2096 0 :
2097 0 : match &*tenants {
2098 0 : TenantsMap::Initializing => ShardResolveResult::NotFound,
2099 0 : TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => {
2100 0 : for slot in m.range(TenantShardId::tenant_range(*tenant_id)) {
2101 : // Ignore all slots that don't contain an attached tenant
2102 0 : let tenant = match &slot.1 {
2103 0 : TenantSlot::Attached(t) => t,
2104 0 : TenantSlot::InProgress(barrier) => {
2105 0 : // We might still find a usable shard, but in case we don't, remember that
2106 0 : // we saw at least one InProgress slot, so that we can distinguish this case
2107 0 : // from a simple NotFound in our return value.
2108 0 : any_in_progress = Some(barrier.clone());
2109 0 : continue;
2110 : }
2111 0 : _ => continue,
2112 : };
2113 :
2114 0 : match selector {
2115 0 : ShardSelector::Zero if slot.0.shard_number == ShardNumber(0) => {
2116 0 : return ShardResolveResult::Found(tenant.clone());
2117 : }
2118 0 : ShardSelector::Page(key) => {
2119 0 : // First slot we see for this tenant, calculate the expected shard number
2120 0 : // for the key: we will use this for checking if this and subsequent
2121 0 : // slots contain the key, rather than recalculating the hash each time.
2122 0 : if want_shard.is_none() {
2123 0 : want_shard = Some(tenant.shard_identity.get_shard_number(&key));
2124 0 : }
2125 :
2126 0 : if Some(tenant.shard_identity.number) == want_shard {
2127 0 : return ShardResolveResult::Found(tenant.clone());
2128 0 : }
2129 : }
2130 0 : ShardSelector::Known(shard)
2131 0 : if tenant.shard_identity.shard_index() == shard =>
2132 0 : {
2133 0 : return ShardResolveResult::Found(tenant.clone());
2134 : }
2135 0 : _ => continue,
2136 : }
2137 : }
2138 :
2139 : // Fall through: we didn't find a slot that was in Attached state & matched our selector. If
2140 : // we found one or more InProgress slot, indicate to caller that they should retry later. Otherwise
2141 : // this requested shard simply isn't found.
2142 0 : if let Some(barrier) = any_in_progress {
2143 0 : ShardResolveResult::InProgress(barrier)
2144 : } else {
2145 0 : ShardResolveResult::NotFound
2146 : }
2147 : }
2148 : }
2149 0 : }
2150 :
2151 : /// Calculate the tenant shards' contributions to this pageserver's utilization metrics. The
2152 : /// returned values are:
2153 : /// - the number of bytes of local disk space this pageserver's shards are requesting, i.e.
2154 : /// how much space they would use if not impacted by disk usage eviction.
2155 : /// - the number of tenant shards currently on this pageserver, including attached
2156 : /// and secondary.
2157 : ///
2158 : /// This function is quite expensive: callers are expected to cache the result and
2159 : /// limit how often they call it.
2160 0 : pub(crate) fn calculate_utilization(&self) -> Result<(u64, u32), TenantMapListError> {
2161 0 : let tenants = self.tenants.read().unwrap();
2162 0 : let m = match &*tenants {
2163 0 : TenantsMap::Initializing => return Err(TenantMapListError::Initializing),
2164 0 : TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => m,
2165 0 : };
2166 0 : let shard_count = m.len();
2167 0 : let mut wanted_bytes = 0;
2168 :
2169 0 : for tenant_slot in m.values() {
2170 0 : match tenant_slot {
2171 0 : TenantSlot::InProgress(_barrier) => {
2172 0 : // While a slot is being changed, we can't know how much storage it wants. This
2173 0 : // means this function's output can fluctuate if a lot of changes are going on
2174 0 : // (such as transitions from secondary to attached).
2175 0 : //
2176 0 : // We could wait for the barrier and retry, but it's important that the utilization
2177 0 : // API is responsive, and the data quality impact is not very significant.
2178 0 : continue;
2179 : }
2180 0 : TenantSlot::Attached(tenant) => {
2181 0 : wanted_bytes += tenant.local_storage_wanted();
2182 0 : }
2183 0 : TenantSlot::Secondary(secondary) => {
2184 0 : let progress = secondary.progress.lock().unwrap();
2185 0 : wanted_bytes += if progress.heatmap_mtime.is_some() {
2186 : // If we have heatmap info, then we will 'want' the sum
2187 : // of the size of layers in the heatmap: this is how much space
2188 : // we would use if not doing any eviction.
2189 0 : progress.bytes_total
2190 : } else {
2191 : // In the absence of heatmap info, assume that the secondary location simply
2192 : // needs as much space as it is currently using.
2193 0 : secondary.resident_size_metric.get()
2194 : }
2195 : }
2196 : }
2197 : }
2198 :
2199 0 : Ok((wanted_bytes, shard_count as u32))
2200 0 : }
2201 :
2202 : #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), %timeline_id))]
2203 : pub(crate) async fn immediate_gc(
2204 : &self,
2205 : tenant_shard_id: TenantShardId,
2206 : timeline_id: TimelineId,
2207 : gc_req: TimelineGcRequest,
2208 : cancel: CancellationToken,
2209 : ctx: &RequestContext,
2210 : ) -> Result<GcResult, ApiError> {
2211 : let tenant = {
2212 : let guard = self.tenants.read().unwrap();
2213 : guard
2214 : .get(&tenant_shard_id)
2215 : .cloned()
2216 0 : .with_context(|| format!("tenant {tenant_shard_id}"))
2217 0 : .map_err(|e| ApiError::NotFound(e.into()))?
2218 : };
2219 :
2220 0 : let gc_horizon = gc_req.gc_horizon.unwrap_or_else(|| tenant.get_gc_horizon());
2221 : // Use tenant's pitr setting
2222 : let pitr = tenant.get_pitr_interval();
2223 :
2224 : tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
2225 :
2226 : // Run in task_mgr to avoid race with tenant_detach operation
2227 : let ctx: RequestContext =
2228 : ctx.detached_child(TaskKind::GarbageCollector, DownloadBehavior::Download);
2229 :
2230 0 : let _gate_guard = tenant.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
2231 :
2232 : fail::fail_point!("immediate_gc_task_pre");
2233 :
2234 : #[allow(unused_mut)]
2235 : let mut result = tenant
2236 : .gc_iteration(Some(timeline_id), gc_horizon, pitr, &cancel, &ctx)
2237 : .await;
2238 : // FIXME: `gc_iteration` can return an error for multiple reasons; we should handle it
2239 : // better once the types support it.
2240 :
2241 : #[cfg(feature = "testing")]
2242 : {
2243 : // we need to synchronize with drop completion for python tests without polling for
2244 : // log messages
2245 : if let Ok(result) = result.as_mut() {
2246 : let mut js = tokio::task::JoinSet::new();
2247 : for layer in std::mem::take(&mut result.doomed_layers) {
2248 : js.spawn(layer.wait_drop());
2249 : }
2250 : tracing::info!(
2251 : total = js.len(),
2252 : "starting to wait for the gc'd layers to be dropped"
2253 : );
2254 : while let Some(res) = js.join_next().await {
2255 : res.expect("wait_drop should not panic");
2256 : }
2257 : }
2258 :
2259 : let timeline = tenant.get_timeline(timeline_id, false).ok();
2260 0 : let rtc = timeline.as_ref().map(|x| &x.remote_client);
2261 :
2262 : if let Some(rtc) = rtc {
2263 : // layer drops schedule actions on remote timeline client to actually do the
2264 : // deletions; don't care about the shutdown error, just exit fast
2265 : drop(rtc.wait_completion().await);
2266 : }
2267 : }
2268 :
2269 0 : result.map_err(|e| match e {
2270 0 : GcError::TenantCancelled | GcError::TimelineCancelled => ApiError::ShuttingDown,
2271 : GcError::TimelineNotFound => {
2272 0 : ApiError::NotFound(anyhow::anyhow!("Timeline not found").into())
2273 : }
2274 0 : other => ApiError::InternalServerError(anyhow::anyhow!(other)),
2275 0 : })
2276 : }
2277 : }
2278 :
2279 : #[derive(Debug, thiserror::Error)]
2280 : pub(crate) enum GetTenantError {
2281 : /// NotFound is a TenantId rather than TenantShardId, because this error type is used from
2282 : /// getters that use a TenantId and a ShardSelector, not just getters that target a specific shard.
2283 : #[error("Tenant {0} not found")]
2284 : NotFound(TenantId),
2285 :
2286 : #[error("Tenant {0} not found")]
2287 : ShardNotFound(TenantShardId),
2288 :
2289 : #[error("Tenant {0} is not active")]
2290 : NotActive(TenantShardId),
2291 :
2292 : // Initializing or shutting down: cannot authoritatively say whether we have this tenant
2293 : #[error("Tenant map is not available: {0}")]
2294 : MapState(#[from] TenantMapError),
2295 : }
2296 :
2297 : #[derive(thiserror::Error, Debug)]
2298 : pub(crate) enum GetActiveTenantError {
2299 : /// We may time out either while TenantSlot is InProgress, or while the Tenant
2300 : /// is in a non-Active state
2301 : #[error(
2302 : "Timed out waiting {wait_time:?} for tenant active state. Latest state: {latest_state:?}"
2303 : )]
2304 : WaitForActiveTimeout {
2305 : latest_state: Option<TenantState>,
2306 : wait_time: Duration,
2307 : },
2308 :
2309 : /// The TenantSlot is absent, or in secondary mode
2310 : #[error(transparent)]
2311 : NotFound(#[from] GetTenantError),
2312 :
2313 : /// Cancellation token fired while we were waiting
2314 : #[error("cancelled")]
2315 : Cancelled,
2316 :
2317 : /// Tenant exists, but is in a state that cannot become active (e.g. Stopping, Broken)
2318 : #[error("will not become active. Current state: {0}")]
2319 : WillNotBecomeActive(TenantState),
2320 :
2321 : /// Broken is logically a subset of WillNotBecomeActive, but a distinct error is useful as
2322 : /// WillNotBecomeActive is a permitted error under some circumstances, whereas broken should
2323 : /// never happen.
2324 : #[error("Tenant is broken: {0}")]
2325 : Broken(String),
2326 :
2327 : #[error("reconnect to switch tenant id")]
2328 : SwitchedTenant,
2329 : }
2330 :
2331 : #[derive(Debug, thiserror::Error)]
2332 : pub(crate) enum DeleteTimelineError {
2333 : #[error("Tenant {0}")]
2334 : Tenant(#[from] GetTenantError),
2335 :
2336 : #[error("Timeline {0}")]
2337 : Timeline(#[from] crate::tenant::DeleteTimelineError),
2338 : }
2339 :
2340 : #[derive(Debug, thiserror::Error)]
2341 : pub(crate) enum TenantStateError {
2342 : #[error("Tenant {0} is stopping")]
2343 : IsStopping(TenantShardId),
2344 : #[error(transparent)]
2345 : SlotError(#[from] TenantSlotError),
2346 : #[error(transparent)]
2347 : SlotUpsertError(#[from] TenantSlotUpsertError),
2348 : #[error(transparent)]
2349 : Other(#[from] anyhow::Error),
2350 : }
2351 :
2352 : #[derive(Debug, thiserror::Error)]
2353 : pub(crate) enum TenantMapListError {
2354 : #[error("tenant map is still initiailizing")]
2355 : Initializing,
2356 : }
2357 :
2358 : #[derive(Debug, thiserror::Error)]
2359 : pub(crate) enum TenantMapInsertError {
2360 : #[error(transparent)]
2361 : SlotError(#[from] TenantSlotError),
2362 : #[error(transparent)]
2363 : SlotUpsertError(#[from] TenantSlotUpsertError),
2364 : #[error(transparent)]
2365 : Other(#[from] anyhow::Error),
2366 : }
2367 :
2368 : /// Superset of TenantMapError: issues that can occur when acquiring a slot
2369 : /// for a particular tenant ID.
2370 : #[derive(Debug, thiserror::Error)]
2371 : pub(crate) enum TenantSlotError {
2372 : /// When acquiring a slot with the expectation that the tenant already exists.
2373 : #[error("Tenant {0} not found")]
2374 : NotFound(TenantShardId),
2375 :
2376 : // Tried to read a slot that is currently being mutated by another administrative
2377 : // operation.
2378 : #[error("tenant has a state change in progress, try again later")]
2379 : InProgress,
2380 :
2381 : #[error(transparent)]
2382 : MapState(#[from] TenantMapError),
2383 : }
2384 :
2385 : /// Superset of TenantMapError: issues that can occur when using a SlotGuard
2386 : /// to insert a new value.
2387 : #[derive(thiserror::Error)]
2388 : pub(crate) enum TenantSlotUpsertError {
2389 : /// An error where the slot is in an unexpected state, indicating a code bug
2390 : #[error("Internal error updating Tenant")]
2391 : InternalError(Cow<'static, str>),
2392 :
2393 : #[error(transparent)]
2394 : MapState(TenantMapError),
2395 :
2396 : // If we encounter TenantManager shutdown during upsert, we must carry the Completion
2397 : // from the SlotGuard, so that the caller can hold it while they clean up: otherwise
2398 : // TenantManager shutdown might race ahead before we're done cleaning up any Tenant that
2399 : // was protected by the SlotGuard.
2400 : #[error("Shutting down")]
2401 : ShuttingDown((TenantSlot, utils::completion::Completion)),
2402 : }
2403 :
2404 : impl std::fmt::Debug for TenantSlotUpsertError {
2405 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
2406 0 : match self {
2407 0 : Self::InternalError(reason) => write!(f, "Internal Error {reason}"),
2408 0 : Self::MapState(map_error) => write!(f, "Tenant map state: {map_error:?}"),
2409 0 : Self::ShuttingDown(_completion) => write!(f, "Tenant map shutting down"),
2410 : }
2411 0 : }
2412 : }
2413 :
2414 : #[derive(Debug, thiserror::Error)]
2415 : enum TenantSlotDropError {
2416 : /// It is only legal to drop a TenantSlot if its contents are fully shut down
2417 : #[error("Tenant was not shut down")]
2418 : NotShutdown,
2419 : }
2420 :
2421 : /// Errors that can happen any time we are walking the tenant map to try and acquire
2422 : /// the TenantSlot for a particular tenant.
2423 : #[derive(Debug, thiserror::Error)]
2424 : pub(crate) enum TenantMapError {
2425 : // Tried to read while initializing
2426 : #[error("tenant map is still initializing")]
2427 : StillInitializing,
2428 :
2429 : // Tried to read while shutting down
2430 : #[error("tenant map is shutting down")]
2431 : ShuttingDown,
2432 : }
2433 :
2434 : /// Guards a particular tenant_id's content in the TenantsMap.
2435 : ///
2436 : /// While this structure exists, the TenantsMap will contain a [`TenantSlot::InProgress`]
2437 : /// for this tenant, which acts as a marker for any operations targeting
2438 : /// this tenant to retry later, or wait for the InProgress state to end.
2439 : ///
2440 : /// This structure enforces the important invariant that we do not have overlapping
2441 : /// tasks that will try use local storage for a the same tenant ID: we enforce that
2442 : /// the previous contents of a slot have been shut down before the slot can be
2443 : /// left empty or used for something else
2444 : ///
2445 : /// Holders of a SlotGuard should explicitly dispose of it, using either `upsert`
2446 : /// to provide a new value, or `revert` to put the slot back into its initial
2447 : /// state. If the SlotGuard is dropped without calling either of these, then
2448 : /// we will leave the slot empty if our `old_value` is already shut down, else
2449 : /// we will replace the slot with `old_value` (equivalent to doing a revert).
2450 : ///
2451 : /// The `old_value` may be dropped before the SlotGuard is dropped, by calling
2452 : /// `drop_old_value`. It is an error to call this without shutting down
2453 : /// the conents of `old_value`.
2454 : pub(crate) struct SlotGuard {
2455 : tenant_shard_id: TenantShardId,
2456 : old_value: Option<TenantSlot>,
2457 : upserted: bool,
2458 :
2459 : /// [`TenantSlot::InProgress`] carries the corresponding Barrier: it will
2460 : /// release any waiters as soon as this SlotGuard is dropped.
2461 : completion: utils::completion::Completion,
2462 : }
2463 :
2464 : impl SlotGuard {
2465 12 : fn new(
2466 12 : tenant_shard_id: TenantShardId,
2467 12 : old_value: Option<TenantSlot>,
2468 12 : completion: utils::completion::Completion,
2469 12 : ) -> Self {
2470 12 : Self {
2471 12 : tenant_shard_id,
2472 12 : old_value,
2473 12 : upserted: false,
2474 12 : completion,
2475 12 : }
2476 12 : }
2477 :
2478 : /// Get any value that was present in the slot before we acquired ownership
2479 : /// of it: in state transitions, this will be the old state.
2480 : ///
2481 : // FIXME: get_ prefix
2482 : // FIXME: this should be .as_ref() -- unsure why no clippy
2483 12 : fn get_old_value(&self) -> &Option<TenantSlot> {
2484 12 : &self.old_value
2485 12 : }
2486 :
2487 : /// Emplace a new value in the slot. This consumes the guard, and after
2488 : /// returning, the slot is no longer protected from concurrent changes.
2489 0 : fn upsert(mut self, new_value: TenantSlot) -> Result<(), TenantSlotUpsertError> {
2490 0 : if !self.old_value_is_shutdown() {
2491 : // This is a bug: callers should never try to drop an old value without
2492 : // shutting it down
2493 0 : return Err(TenantSlotUpsertError::InternalError(
2494 0 : "Old TenantSlot value not shut down".into(),
2495 0 : ));
2496 0 : }
2497 :
2498 0 : let replaced = {
2499 0 : let mut locked = TENANTS.write().unwrap();
2500 0 :
2501 0 : if let TenantSlot::InProgress(_) = new_value {
2502 : // It is never expected to try and upsert InProgress via this path: it should
2503 : // only be written via the tenant_map_acquire_slot path. If we hit this it's a bug.
2504 0 : return Err(TenantSlotUpsertError::InternalError(
2505 0 : "Attempt to upsert an InProgress state".into(),
2506 0 : ));
2507 0 : }
2508 :
2509 0 : let m = match &mut *locked {
2510 : TenantsMap::Initializing => {
2511 0 : return Err(TenantSlotUpsertError::MapState(
2512 0 : TenantMapError::StillInitializing,
2513 0 : ));
2514 : }
2515 : TenantsMap::ShuttingDown(_) => {
2516 0 : return Err(TenantSlotUpsertError::ShuttingDown((
2517 0 : new_value,
2518 0 : self.completion.clone(),
2519 0 : )));
2520 : }
2521 0 : TenantsMap::Open(m) => m,
2522 0 : };
2523 0 :
2524 0 : METRICS.slot_inserted(&new_value);
2525 0 :
2526 0 : let replaced = m.insert(self.tenant_shard_id, new_value);
2527 0 : self.upserted = true;
2528 0 : if let Some(replaced) = replaced.as_ref() {
2529 0 : METRICS.slot_removed(replaced);
2530 0 : }
2531 :
2532 0 : replaced
2533 : };
2534 :
2535 : // Sanity check: on an upsert we should always be replacing an InProgress marker
2536 0 : match replaced {
2537 : Some(TenantSlot::InProgress(_)) => {
2538 : // Expected case: we find our InProgress in the map: nothing should have
2539 : // replaced it because the code that acquires slots will not grant another
2540 : // one for the same TenantId.
2541 0 : Ok(())
2542 : }
2543 : None => {
2544 0 : METRICS.unexpected_errors.inc();
2545 0 : error!(
2546 : tenant_shard_id = %self.tenant_shard_id,
2547 0 : "Missing InProgress marker during tenant upsert, this is a bug."
2548 : );
2549 0 : Err(TenantSlotUpsertError::InternalError(
2550 0 : "Missing InProgress marker during tenant upsert".into(),
2551 0 : ))
2552 : }
2553 0 : Some(slot) => {
2554 0 : METRICS.unexpected_errors.inc();
2555 0 : error!(tenant_shard_id=%self.tenant_shard_id, "Unexpected contents of TenantSlot during upsert, this is a bug. Contents: {:?}", slot);
2556 0 : Err(TenantSlotUpsertError::InternalError(
2557 0 : "Unexpected contents of TenantSlot".into(),
2558 0 : ))
2559 : }
2560 : }
2561 0 : }
2562 :
2563 : /// Replace the InProgress slot with whatever was in the guard when we started
2564 0 : fn revert(mut self) {
2565 0 : if let Some(value) = self.old_value.take() {
2566 0 : match self.upsert(value) {
2567 0 : Err(TenantSlotUpsertError::InternalError(_)) => {
2568 0 : // We already logged the error, nothing else we can do.
2569 0 : }
2570 : Err(
2571 : TenantSlotUpsertError::MapState(_) | TenantSlotUpsertError::ShuttingDown(_),
2572 0 : ) => {
2573 0 : // If the map is shutting down, we need not replace anything
2574 0 : }
2575 0 : Ok(()) => {}
2576 : }
2577 0 : }
2578 0 : }
2579 :
2580 : /// We may never drop our old value until it is cleanly shut down: otherwise we might leave
2581 : /// rogue background tasks that would write to the local tenant directory that this guard
2582 : /// is responsible for protecting
2583 12 : fn old_value_is_shutdown(&self) -> bool {
2584 12 : match self.old_value.as_ref() {
2585 12 : Some(TenantSlot::Attached(tenant)) => tenant.gate.close_complete(),
2586 0 : Some(TenantSlot::Secondary(secondary_tenant)) => secondary_tenant.gate.close_complete(),
2587 : Some(TenantSlot::InProgress(_)) => {
2588 : // A SlotGuard cannot be constructed for a slot that was already InProgress
2589 0 : unreachable!()
2590 : }
2591 0 : None => true,
2592 : }
2593 12 : }
2594 :
2595 : /// The guard holder is done with the old value of the slot: they are obliged to already
2596 : /// shut it down before we reach this point.
2597 12 : fn drop_old_value(&mut self) -> Result<(), TenantSlotDropError> {
2598 12 : if !self.old_value_is_shutdown() {
2599 0 : Err(TenantSlotDropError::NotShutdown)
2600 : } else {
2601 12 : self.old_value.take();
2602 12 : Ok(())
2603 : }
2604 12 : }
2605 : }
2606 :
2607 : impl Drop for SlotGuard {
2608 12 : fn drop(&mut self) {
2609 12 : if self.upserted {
2610 0 : return;
2611 12 : }
2612 12 : // Our old value is already shutdown, or it never existed: it is safe
2613 12 : // for us to fully release the TenantSlot back into an empty state
2614 12 :
2615 12 : let mut locked = TENANTS.write().unwrap();
2616 :
2617 12 : let m = match &mut *locked {
2618 : TenantsMap::Initializing => {
2619 : // There is no map, this should never happen.
2620 12 : return;
2621 : }
2622 : TenantsMap::ShuttingDown(_) => {
2623 : // When we transition to shutdown, InProgress elements are removed
2624 : // from the map, so we do not need to clean up our Inprogress marker.
2625 : // See [`shutdown_all_tenants0`]
2626 0 : return;
2627 : }
2628 0 : TenantsMap::Open(m) => m,
2629 : };
2630 :
2631 : use std::collections::btree_map::Entry;
2632 0 : match m.entry(self.tenant_shard_id) {
2633 0 : Entry::Occupied(mut entry) => {
2634 0 : if !matches!(entry.get(), TenantSlot::InProgress(_)) {
2635 0 : METRICS.unexpected_errors.inc();
2636 0 : error!(tenant_shard_id=%self.tenant_shard_id, "Unexpected contents of TenantSlot during drop, this is a bug. Contents: {:?}", entry.get());
2637 0 : }
2638 :
2639 0 : if self.old_value_is_shutdown() {
2640 0 : METRICS.slot_removed(entry.get());
2641 0 : entry.remove();
2642 0 : } else {
2643 0 : let inserting = self.old_value.take().unwrap();
2644 0 : METRICS.slot_inserted(&inserting);
2645 0 : let replaced = entry.insert(inserting);
2646 0 : METRICS.slot_removed(&replaced);
2647 0 : }
2648 : }
2649 : Entry::Vacant(_) => {
2650 0 : METRICS.unexpected_errors.inc();
2651 0 : error!(
2652 : tenant_shard_id = %self.tenant_shard_id,
2653 0 : "Missing InProgress marker during SlotGuard drop, this is a bug."
2654 : );
2655 : }
2656 : }
2657 12 : }
2658 : }
2659 :
2660 : enum TenantSlotPeekMode {
2661 : /// In Read mode, peek will be permitted to see the slots even if the pageserver is shutting down
2662 : Read,
2663 : /// In Write mode, trying to peek at a slot while the pageserver is shutting down is an error
2664 : Write,
2665 : }
2666 :
2667 0 : fn tenant_map_peek_slot<'a>(
2668 0 : tenants: &'a std::sync::RwLockReadGuard<'a, TenantsMap>,
2669 0 : tenant_shard_id: &TenantShardId,
2670 0 : mode: TenantSlotPeekMode,
2671 0 : ) -> Result<Option<&'a TenantSlot>, TenantMapError> {
2672 0 : match tenants.deref() {
2673 0 : TenantsMap::Initializing => Err(TenantMapError::StillInitializing),
2674 0 : TenantsMap::ShuttingDown(m) => match mode {
2675 : TenantSlotPeekMode::Read => Ok(Some(
2676 : // When reading in ShuttingDown state, we must translate None results
2677 : // into a ShuttingDown error, because absence of a tenant shard ID in the map
2678 : // isn't a reliable indicator of the tenant being gone: it might have been
2679 : // InProgress when shutdown started, and cleaned up from that state such
2680 : // that it's now no longer in the map. Callers will have to wait until
2681 : // we next start up to get a proper answer. This avoids incorrect 404 API responses.
2682 0 : m.get(tenant_shard_id).ok_or(TenantMapError::ShuttingDown)?,
2683 : )),
2684 0 : TenantSlotPeekMode::Write => Err(TenantMapError::ShuttingDown),
2685 : },
2686 0 : TenantsMap::Open(m) => Ok(m.get(tenant_shard_id)),
2687 : }
2688 0 : }
2689 :
2690 : enum TenantSlotAcquireMode {
2691 : /// Acquire the slot irrespective of current state, or whether it already exists
2692 : Any,
2693 : /// Return an error if trying to acquire a slot and it doesn't already exist
2694 : MustExist,
2695 : }
2696 :
2697 0 : fn tenant_map_acquire_slot(
2698 0 : tenant_shard_id: &TenantShardId,
2699 0 : mode: TenantSlotAcquireMode,
2700 0 : ) -> Result<SlotGuard, TenantSlotError> {
2701 0 : tenant_map_acquire_slot_impl(tenant_shard_id, &TENANTS, mode)
2702 0 : }
2703 :
2704 12 : fn tenant_map_acquire_slot_impl(
2705 12 : tenant_shard_id: &TenantShardId,
2706 12 : tenants: &std::sync::RwLock<TenantsMap>,
2707 12 : mode: TenantSlotAcquireMode,
2708 12 : ) -> Result<SlotGuard, TenantSlotError> {
2709 : use TenantSlotAcquireMode::*;
2710 12 : METRICS.tenant_slot_writes.inc();
2711 12 :
2712 12 : let mut locked = tenants.write().unwrap();
2713 12 : let span = tracing::info_span!("acquire_slot", tenant_id=%tenant_shard_id.tenant_id, shard_id = %tenant_shard_id.shard_slug());
2714 12 : let _guard = span.enter();
2715 :
2716 12 : let m = match &mut *locked {
2717 0 : TenantsMap::Initializing => return Err(TenantMapError::StillInitializing.into()),
2718 0 : TenantsMap::ShuttingDown(_) => return Err(TenantMapError::ShuttingDown.into()),
2719 12 : TenantsMap::Open(m) => m,
2720 : };
2721 :
2722 : use std::collections::btree_map::Entry;
2723 :
2724 12 : let entry = m.entry(*tenant_shard_id);
2725 12 :
2726 12 : match entry {
2727 0 : Entry::Vacant(v) => match mode {
2728 : MustExist => {
2729 0 : tracing::debug!("Vacant && MustExist: return NotFound");
2730 0 : Err(TenantSlotError::NotFound(*tenant_shard_id))
2731 : }
2732 : _ => {
2733 0 : let (completion, barrier) = utils::completion::channel();
2734 0 : let inserting = TenantSlot::InProgress(barrier);
2735 0 : METRICS.slot_inserted(&inserting);
2736 0 : v.insert(inserting);
2737 0 : tracing::debug!("Vacant, inserted InProgress");
2738 0 : Ok(SlotGuard::new(*tenant_shard_id, None, completion))
2739 : }
2740 : },
2741 12 : Entry::Occupied(mut o) => {
2742 12 : // Apply mode-driven checks
2743 12 : match (o.get(), mode) {
2744 : (TenantSlot::InProgress(_), _) => {
2745 0 : tracing::debug!("Occupied, failing for InProgress");
2746 0 : Err(TenantSlotError::InProgress)
2747 : }
2748 : _ => {
2749 : // Happy case: the slot was not in any state that violated our mode
2750 12 : let (completion, barrier) = utils::completion::channel();
2751 12 : let in_progress = TenantSlot::InProgress(barrier);
2752 12 : METRICS.slot_inserted(&in_progress);
2753 12 : let old_value = o.insert(in_progress);
2754 12 : METRICS.slot_removed(&old_value);
2755 12 : tracing::debug!("Occupied, replaced with InProgress");
2756 12 : Ok(SlotGuard::new(
2757 12 : *tenant_shard_id,
2758 12 : Some(old_value),
2759 12 : completion,
2760 12 : ))
2761 : }
2762 : }
2763 : }
2764 : }
2765 12 : }
2766 :
2767 : /// Stops and removes the tenant from memory, if it's not [`TenantState::Stopping`] already, bails otherwise.
2768 : /// Allows to remove other tenant resources manually, via `tenant_cleanup`.
2769 : /// If the cleanup fails, tenant will stay in memory in [`TenantState::Broken`] state, and another removal
2770 : /// operation would be needed to remove it.
2771 12 : async fn remove_tenant_from_memory<V, F>(
2772 12 : tenants: &std::sync::RwLock<TenantsMap>,
2773 12 : tenant_shard_id: TenantShardId,
2774 12 : tenant_cleanup: F,
2775 12 : ) -> Result<V, TenantStateError>
2776 12 : where
2777 12 : F: std::future::Future<Output = anyhow::Result<V>>,
2778 12 : {
2779 12 : let mut slot_guard =
2780 12 : tenant_map_acquire_slot_impl(&tenant_shard_id, tenants, TenantSlotAcquireMode::MustExist)?;
2781 :
2782 : // allow pageserver shutdown to await for our completion
2783 12 : let (_guard, progress) = completion::channel();
2784 :
2785 : // The SlotGuard allows us to manipulate the Tenant object without fear of some
2786 : // concurrent API request doing something else for the same tenant ID.
2787 12 : let attached_tenant = match slot_guard.get_old_value() {
2788 12 : Some(TenantSlot::Attached(tenant)) => {
2789 12 : // whenever we remove a tenant from memory, we don't want to flush and wait for upload
2790 12 : let shutdown_mode = ShutdownMode::Hard;
2791 12 :
2792 12 : // shutdown is sure to transition tenant to stopping, and wait for all tasks to complete, so
2793 12 : // that we can continue safely to cleanup.
2794 12 : match tenant.shutdown(progress, shutdown_mode).await {
2795 12 : Ok(()) => {}
2796 0 : Err(_other) => {
2797 0 : // if pageserver shutdown or other detach/ignore is already ongoing, we don't want to
2798 0 : // wait for it but return an error right away because these are distinct requests.
2799 0 : slot_guard.revert();
2800 0 : return Err(TenantStateError::IsStopping(tenant_shard_id));
2801 : }
2802 : }
2803 12 : Some(tenant)
2804 : }
2805 0 : Some(TenantSlot::Secondary(secondary_state)) => {
2806 0 : tracing::info!("Shutting down in secondary mode");
2807 0 : secondary_state.shutdown().await;
2808 0 : None
2809 : }
2810 : Some(TenantSlot::InProgress(_)) => {
2811 : // Acquiring a slot guarantees its old value was not InProgress
2812 0 : unreachable!();
2813 : }
2814 0 : None => None,
2815 : };
2816 :
2817 12 : match tenant_cleanup
2818 12 : .await
2819 12 : .with_context(|| format!("Failed to run cleanup for tenant {tenant_shard_id}"))
2820 : {
2821 12 : Ok(hook_value) => {
2822 12 : // Success: drop the old TenantSlot::Attached.
2823 12 : slot_guard
2824 12 : .drop_old_value()
2825 12 : .expect("We just called shutdown");
2826 12 :
2827 12 : Ok(hook_value)
2828 : }
2829 0 : Err(e) => {
2830 : // If we had a Tenant, set it to Broken and put it back in the TenantsMap
2831 0 : if let Some(attached_tenant) = attached_tenant {
2832 0 : attached_tenant.set_broken(e.to_string()).await;
2833 0 : }
2834 : // Leave the broken tenant in the map
2835 0 : slot_guard.revert();
2836 0 :
2837 0 : Err(TenantStateError::Other(e))
2838 : }
2839 : }
2840 12 : }
2841 :
2842 : use http_utils::error::ApiError;
2843 : use pageserver_api::models::TimelineGcRequest;
2844 :
2845 : use crate::tenant::gc_result::GcResult;
2846 :
2847 : #[cfg(test)]
2848 : mod tests {
2849 : use std::collections::BTreeMap;
2850 : use std::sync::Arc;
2851 :
2852 : use tracing::Instrument;
2853 :
2854 : use super::super::harness::TenantHarness;
2855 : use super::TenantsMap;
2856 : use crate::tenant::mgr::TenantSlot;
2857 :
2858 : #[tokio::test(start_paused = true)]
2859 12 : async fn shutdown_awaits_in_progress_tenant() {
2860 12 : // Test that if an InProgress tenant is in the map during shutdown, the shutdown will gracefully
2861 12 : // wait for it to complete before proceeding.
2862 12 :
2863 12 : let h = TenantHarness::create("shutdown_awaits_in_progress_tenant")
2864 12 : .await
2865 12 : .unwrap();
2866 12 : let (t, _ctx) = h.load().await;
2867 12 :
2868 12 : // harness loads it to active, which is forced and nothing is running on the tenant
2869 12 :
2870 12 : let id = t.tenant_shard_id();
2871 12 :
2872 12 : // tenant harness configures the logging and we cannot escape it
2873 12 : let span = h.span();
2874 12 : let _e = span.enter();
2875 12 :
2876 12 : let tenants = BTreeMap::from([(id, TenantSlot::Attached(t.clone()))]);
2877 12 : let tenants = Arc::new(std::sync::RwLock::new(TenantsMap::Open(tenants)));
2878 12 :
2879 12 : // Invoke remove_tenant_from_memory with a cleanup hook that blocks until we manually
2880 12 : // permit it to proceed: that will stick the tenant in InProgress
2881 12 :
2882 12 : let (until_cleanup_completed, can_complete_cleanup) = utils::completion::channel();
2883 12 : let (until_cleanup_started, cleanup_started) = utils::completion::channel();
2884 12 : let mut remove_tenant_from_memory_task = {
2885 12 : let jh = tokio::spawn({
2886 12 : let tenants = tenants.clone();
2887 12 : async move {
2888 12 : let cleanup = async move {
2889 12 : drop(until_cleanup_started);
2890 12 : can_complete_cleanup.wait().await;
2891 12 : anyhow::Ok(())
2892 12 : };
2893 12 : super::remove_tenant_from_memory(&tenants, id, cleanup).await
2894 12 : }
2895 12 : .instrument(h.span())
2896 12 : });
2897 12 :
2898 12 : // now the long cleanup should be in place, with the stopping state
2899 12 : cleanup_started.wait().await;
2900 12 : jh
2901 12 : };
2902 12 :
2903 12 : let mut shutdown_task = {
2904 12 : let (until_shutdown_started, shutdown_started) = utils::completion::channel();
2905 12 :
2906 12 : let shutdown_task = tokio::spawn(async move {
2907 12 : drop(until_shutdown_started);
2908 12 : super::shutdown_all_tenants0(&tenants).await;
2909 12 : });
2910 12 :
2911 12 : shutdown_started.wait().await;
2912 12 : shutdown_task
2913 12 : };
2914 12 :
2915 12 : let long_time = std::time::Duration::from_secs(15);
2916 12 : tokio::select! {
2917 12 : _ = &mut shutdown_task => unreachable!("shutdown should block on remove_tenant_from_memory completing"),
2918 12 : _ = &mut remove_tenant_from_memory_task => unreachable!("remove_tenant_from_memory_task should not complete until explicitly unblocked"),
2919 12 : _ = tokio::time::sleep(long_time) => {},
2920 12 : }
2921 12 :
2922 12 : drop(until_cleanup_completed);
2923 12 :
2924 12 : // Now that we allow it to proceed, shutdown should complete immediately
2925 12 : remove_tenant_from_memory_task.await.unwrap().unwrap();
2926 12 : shutdown_task.await.unwrap();
2927 12 : }
2928 : }
|
