Line data Source code
1 : //! This module acts as a switchboard to access different repositories managed by this
2 : //! page server.
3 :
4 : use camino::{Utf8DirEntry, Utf8Path, Utf8PathBuf};
5 : use futures::StreamExt;
6 : use itertools::Itertools;
7 : use pageserver_api::key::Key;
8 : use pageserver_api::models::LocationConfigMode;
9 : use pageserver_api::shard::{
10 : ShardCount, ShardIdentity, ShardIndex, ShardNumber, ShardStripeSize, TenantShardId,
11 : };
12 : use pageserver_api::upcall_api::ReAttachResponseTenant;
13 : use rand::{distributions::Alphanumeric, Rng};
14 : use std::borrow::Cow;
15 : use std::cmp::Ordering;
16 : use std::collections::{BTreeMap, HashMap, HashSet};
17 : use std::ops::Deref;
18 : use std::sync::Arc;
19 : use std::time::Duration;
20 : use sysinfo::SystemExt;
21 : use tokio::fs;
22 :
23 : use anyhow::Context;
24 : use once_cell::sync::Lazy;
25 : use tokio::task::JoinSet;
26 : use tokio_util::sync::CancellationToken;
27 : use tracing::*;
28 :
29 : use utils::{backoff, completion, crashsafe};
30 :
31 : use crate::config::PageServerConf;
32 : use crate::context::{DownloadBehavior, RequestContext};
33 : use crate::control_plane_client::{
34 : ControlPlaneClient, ControlPlaneGenerationsApi, RetryForeverError,
35 : };
36 : use crate::deletion_queue::DeletionQueueClient;
37 : use crate::http::routes::ACTIVE_TENANT_TIMEOUT;
38 : use crate::metrics::{TENANT, TENANT_MANAGER as METRICS};
39 : use crate::task_mgr::{TaskKind, BACKGROUND_RUNTIME};
40 : use crate::tenant::config::{
41 : AttachedLocationConfig, AttachmentMode, LocationConf, LocationMode, SecondaryLocationConfig,
42 : };
43 : use crate::tenant::span::debug_assert_current_span_has_tenant_id;
44 : use crate::tenant::storage_layer::inmemory_layer;
45 : use crate::tenant::timeline::ShutdownMode;
46 : use crate::tenant::{AttachedTenantConf, GcError, LoadConfigError, SpawnMode, Tenant, TenantState};
47 : use crate::virtual_file::MaybeFatalIo;
48 : use crate::{InitializationOrder, TEMP_FILE_SUFFIX};
49 :
50 : use utils::crashsafe::path_with_suffix_extension;
51 : use utils::fs_ext::PathExt;
52 : use utils::generation::Generation;
53 : use utils::id::{TenantId, TimelineId};
54 :
55 : use super::remote_timeline_client::remote_tenant_path;
56 : use super::secondary::SecondaryTenant;
57 : use super::timeline::detach_ancestor::{self, PreparedTimelineDetach};
58 : use super::{GlobalShutDown, TenantSharedResources};
59 :
60 : /// For a tenant that appears in TenantsMap, it may either be
61 : /// - `Attached`: has a full Tenant object, is elegible to service
62 : /// reads and ingest WAL.
63 : /// - `Secondary`: is only keeping a local cache warm.
64 : ///
65 : /// Secondary is a totally distinct state rather than being a mode of a `Tenant`, because
66 : /// that way we avoid having to carefully switch a tenant's ingestion etc on and off during
67 : /// its lifetime, and we can preserve some important safety invariants like `Tenant` always
68 : /// having a properly acquired generation (Secondary doesn't need a generation)
69 : #[derive(Clone)]
70 : pub(crate) enum TenantSlot {
71 : Attached(Arc<Tenant>),
72 : Secondary(Arc<SecondaryTenant>),
73 : /// In this state, other administrative operations acting on the TenantId should
74 : /// block, or return a retry indicator equivalent to HTTP 503.
75 : InProgress(utils::completion::Barrier),
76 : }
77 :
78 : impl std::fmt::Debug for TenantSlot {
79 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
80 0 : match self {
81 0 : Self::Attached(tenant) => write!(f, "Attached({})", tenant.current_state()),
82 0 : Self::Secondary(_) => write!(f, "Secondary"),
83 0 : Self::InProgress(_) => write!(f, "InProgress"),
84 : }
85 0 : }
86 : }
87 :
88 : impl TenantSlot {
89 : /// Return the `Tenant` in this slot if attached, else None
90 0 : fn get_attached(&self) -> Option<&Arc<Tenant>> {
91 0 : match self {
92 0 : Self::Attached(t) => Some(t),
93 0 : Self::Secondary(_) => None,
94 0 : Self::InProgress(_) => None,
95 : }
96 0 : }
97 : }
98 :
99 : /// The tenants known to the pageserver.
100 : /// The enum variants are used to distinguish the different states that the pageserver can be in.
101 : pub(crate) enum TenantsMap {
102 : /// [`init_tenant_mgr`] is not done yet.
103 : Initializing,
104 : /// [`init_tenant_mgr`] is done, all on-disk tenants have been loaded.
105 : /// New tenants can be added using [`tenant_map_acquire_slot`].
106 : Open(BTreeMap<TenantShardId, TenantSlot>),
107 : /// The pageserver has entered shutdown mode via [`TenantManager::shutdown`].
108 : /// Existing tenants are still accessible, but no new tenants can be created.
109 : ShuttingDown(BTreeMap<TenantShardId, TenantSlot>),
110 : }
111 :
112 : /// When resolving a TenantId to a shard, we may be looking for the 0th
113 : /// shard, or we might be looking for whichever shard holds a particular page.
114 : #[derive(Copy, Clone)]
115 : pub(crate) enum ShardSelector {
116 : /// Only return the 0th shard, if it is present. If a non-0th shard is present,
117 : /// ignore it.
118 : Zero,
119 : /// Pick the shard that holds this key
120 : Page(Key),
121 : /// The shard ID is known: pick the given shard
122 : Known(ShardIndex),
123 : }
124 :
125 : /// A convenience for use with the re_attach ControlPlaneClient function: rather
126 : /// than the serializable struct, we build this enum that encapsulates
127 : /// the invariant that attached tenants always have generations.
128 : ///
129 : /// This represents the subset of a LocationConfig that we receive during re-attach.
130 : pub(crate) enum TenantStartupMode {
131 : Attached((AttachmentMode, Generation)),
132 : Secondary,
133 : }
134 :
135 : impl TenantStartupMode {
136 : /// Return the generation & mode that should be used when starting
137 : /// this tenant.
138 : ///
139 : /// If this returns None, the re-attach struct is in an invalid state and
140 : /// should be ignored in the response.
141 0 : fn from_reattach_tenant(rart: ReAttachResponseTenant) -> Option<Self> {
142 0 : match (rart.mode, rart.gen) {
143 0 : (LocationConfigMode::Detached, _) => None,
144 0 : (LocationConfigMode::Secondary, _) => Some(Self::Secondary),
145 0 : (LocationConfigMode::AttachedMulti, Some(g)) => {
146 0 : Some(Self::Attached((AttachmentMode::Multi, Generation::new(g))))
147 : }
148 0 : (LocationConfigMode::AttachedSingle, Some(g)) => {
149 0 : Some(Self::Attached((AttachmentMode::Single, Generation::new(g))))
150 : }
151 0 : (LocationConfigMode::AttachedStale, Some(g)) => {
152 0 : Some(Self::Attached((AttachmentMode::Stale, Generation::new(g))))
153 : }
154 : _ => {
155 0 : tracing::warn!(
156 0 : "Received invalid re-attach state for tenant {}: {rart:?}",
157 : rart.id
158 : );
159 0 : None
160 : }
161 : }
162 0 : }
163 : }
164 :
165 : /// Result type for looking up a TenantId to a specific shard
166 : pub(crate) enum ShardResolveResult {
167 : NotFound,
168 : Found(Arc<Tenant>),
169 : // Wait for this barrrier, then query again
170 : InProgress(utils::completion::Barrier),
171 : }
172 :
173 : impl TenantsMap {
174 : /// Convenience function for typical usage, where we want to get a `Tenant` object, for
175 : /// working with attached tenants. If the TenantId is in the map but in Secondary state,
176 : /// None is returned.
177 0 : pub(crate) fn get(&self, tenant_shard_id: &TenantShardId) -> Option<&Arc<Tenant>> {
178 0 : match self {
179 0 : TenantsMap::Initializing => None,
180 0 : TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => {
181 0 : m.get(tenant_shard_id).and_then(|slot| slot.get_attached())
182 : }
183 : }
184 0 : }
185 :
186 : #[cfg(all(debug_assertions, not(test)))]
187 0 : pub(crate) fn len(&self) -> usize {
188 0 : match self {
189 0 : TenantsMap::Initializing => 0,
190 0 : TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => m.len(),
191 : }
192 0 : }
193 : }
194 :
195 : /// Precursor to deletion of a tenant dir: we do a fast rename to a tmp path, and then
196 : /// the slower actual deletion in the background.
197 : ///
198 : /// This is "safe" in that that it won't leave behind a partially deleted directory
199 : /// at the original path, because we rename with TEMP_FILE_SUFFIX before starting deleting
200 : /// the contents.
201 : ///
202 : /// This is pageserver-specific, as it relies on future processes after a crash to check
203 : /// for TEMP_FILE_SUFFIX when loading things.
204 0 : async fn safe_rename_tenant_dir(path: impl AsRef<Utf8Path>) -> std::io::Result<Utf8PathBuf> {
205 0 : let parent = path
206 0 : .as_ref()
207 0 : .parent()
208 0 : // It is invalid to call this function with a relative path. Tenant directories
209 0 : // should always have a parent.
210 0 : .ok_or(std::io::Error::new(
211 0 : std::io::ErrorKind::InvalidInput,
212 0 : "Path must be absolute",
213 0 : ))?;
214 0 : let rand_suffix = rand::thread_rng()
215 0 : .sample_iter(&Alphanumeric)
216 0 : .take(8)
217 0 : .map(char::from)
218 0 : .collect::<String>()
219 0 : + TEMP_FILE_SUFFIX;
220 0 : let tmp_path = path_with_suffix_extension(&path, &rand_suffix);
221 0 : fs::rename(path.as_ref(), &tmp_path).await?;
222 0 : fs::File::open(parent).await?.sync_all().await?;
223 0 : Ok(tmp_path)
224 0 : }
225 :
226 : /// See [`Self::spawn`].
227 : #[derive(Clone, Default)]
228 : pub struct BackgroundPurges(tokio_util::task::TaskTracker);
229 :
230 : impl BackgroundPurges {
231 : /// When we have moved a tenant's content to a temporary directory, we may delete it lazily in
232 : /// the background, and thereby avoid blocking any API requests on this deletion completing.
233 : ///
234 : /// Although we are cleaning up the tenant, this task is not meant to be bound by the lifetime of the tenant in memory.
235 : /// Thus the [`BackgroundPurges`] type to keep track of these tasks.
236 0 : pub fn spawn(&self, tmp_path: Utf8PathBuf) {
237 0 : // because on shutdown we close and wait, we are misusing TaskTracker a bit.
238 0 : //
239 0 : // so first acquire a token, then check if the tracker has been closed. the tracker might get closed
240 0 : // right after, but at least the shutdown will wait for what we are spawning next.
241 0 : let token = self.0.token();
242 0 :
243 0 : if self.0.is_closed() {
244 0 : warn!(
245 : %tmp_path,
246 0 : "trying to spawn background purge during shutdown, ignoring"
247 : );
248 0 : return;
249 0 : }
250 :
251 0 : let span = info_span!(parent: None, "background_purge", %tmp_path);
252 :
253 0 : let task = move || {
254 0 : let _token = token;
255 0 : let _entered = span.entered();
256 0 : if let Err(error) = std::fs::remove_dir_all(tmp_path.as_path()) {
257 : // should we fatal_io_error here?
258 0 : warn!(%error, "failed to purge tenant directory");
259 0 : }
260 0 : };
261 :
262 0 : BACKGROUND_RUNTIME.spawn_blocking(task);
263 0 : }
264 :
265 : /// When this future completes, all background purges have completed.
266 : /// The first poll of the future will already lock out new background purges spawned via [`Self::spawn`].
267 : ///
268 : /// Concurrent calls will coalesce.
269 : ///
270 : /// # Cancellation-Safety
271 : ///
272 : /// If this future is dropped before polled to completion, concurrent and subsequent
273 : /// instances of this future will continue to be correct.
274 0 : #[instrument(skip_all)]
275 : pub async fn shutdown(&self) {
276 : // forbid new tasks (can be called many times)
277 : self.0.close();
278 : self.0.wait().await;
279 : }
280 : }
281 :
282 : static TENANTS: Lazy<std::sync::RwLock<TenantsMap>> =
283 2 : Lazy::new(|| std::sync::RwLock::new(TenantsMap::Initializing));
284 :
285 : /// The TenantManager is responsible for storing and mutating the collection of all tenants
286 : /// that this pageserver process has state for. Every Tenant and SecondaryTenant instance
287 : /// lives inside the TenantManager.
288 : ///
289 : /// The most important role of the TenantManager is to prevent conflicts: e.g. trying to attach
290 : /// the same tenant twice concurrently, or trying to configure the same tenant into secondary
291 : /// and attached modes concurrently.
292 : pub struct TenantManager {
293 : conf: &'static PageServerConf,
294 : // TODO: currently this is a &'static pointing to TENANTs. When we finish refactoring
295 : // out of that static variable, the TenantManager can own this.
296 : // See https://github.com/neondatabase/neon/issues/5796
297 : tenants: &'static std::sync::RwLock<TenantsMap>,
298 : resources: TenantSharedResources,
299 :
300 : // Long-running operations that happen outside of a [`Tenant`] lifetime should respect this token.
301 : // This is for edge cases like tenant deletion. In normal cases (within a Tenant lifetime),
302 : // tenants have their own cancellation tokens, which we fire individually in [`Self::shutdown`], or
303 : // when the tenant detaches.
304 : cancel: CancellationToken,
305 :
306 : background_purges: BackgroundPurges,
307 : }
308 :
309 0 : fn emergency_generations(
310 0 : tenant_confs: &HashMap<TenantShardId, Result<LocationConf, LoadConfigError>>,
311 0 : ) -> HashMap<TenantShardId, TenantStartupMode> {
312 0 : tenant_confs
313 0 : .iter()
314 0 : .filter_map(|(tid, lc)| {
315 0 : let lc = match lc {
316 0 : Ok(lc) => lc,
317 0 : Err(_) => return None,
318 : };
319 : Some((
320 0 : *tid,
321 0 : match &lc.mode {
322 0 : LocationMode::Attached(alc) => {
323 0 : TenantStartupMode::Attached((alc.attach_mode, alc.generation))
324 : }
325 0 : LocationMode::Secondary(_) => TenantStartupMode::Secondary,
326 : },
327 : ))
328 0 : })
329 0 : .collect()
330 0 : }
331 :
332 0 : async fn init_load_generations(
333 0 : conf: &'static PageServerConf,
334 0 : tenant_confs: &HashMap<TenantShardId, Result<LocationConf, LoadConfigError>>,
335 0 : resources: &TenantSharedResources,
336 0 : cancel: &CancellationToken,
337 0 : ) -> anyhow::Result<Option<HashMap<TenantShardId, TenantStartupMode>>> {
338 0 : let generations = if conf.control_plane_emergency_mode {
339 0 : error!(
340 0 : "Emergency mode! Tenants will be attached unsafely using their last known generation"
341 : );
342 0 : emergency_generations(tenant_confs)
343 0 : } else if let Some(client) = ControlPlaneClient::new(conf, cancel) {
344 0 : info!("Calling control plane API to re-attach tenants");
345 : // If we are configured to use the control plane API, then it is the source of truth for what tenants to load.
346 0 : match client.re_attach(conf).await {
347 0 : Ok(tenants) => tenants
348 0 : .into_iter()
349 0 : .flat_map(|(id, rart)| {
350 0 : TenantStartupMode::from_reattach_tenant(rart).map(|tsm| (id, tsm))
351 0 : })
352 0 : .collect(),
353 : Err(RetryForeverError::ShuttingDown) => {
354 0 : anyhow::bail!("Shut down while waiting for control plane re-attach response")
355 : }
356 : }
357 : } else {
358 0 : info!("Control plane API not configured, tenant generations are disabled");
359 0 : return Ok(None);
360 : };
361 :
362 : // The deletion queue needs to know about the startup attachment state to decide which (if any) stored
363 : // deletion list entries may still be valid. We provide that by pushing a recovery operation into
364 : // the queue. Sequential processing of te queue ensures that recovery is done before any new tenant deletions
365 : // are processed, even though we don't block on recovery completing here.
366 0 : let attached_tenants = generations
367 0 : .iter()
368 0 : .flat_map(|(id, start_mode)| {
369 0 : match start_mode {
370 0 : TenantStartupMode::Attached((_mode, generation)) => Some(generation),
371 0 : TenantStartupMode::Secondary => None,
372 : }
373 0 : .map(|gen| (*id, *gen))
374 0 : })
375 0 : .collect();
376 0 : resources.deletion_queue_client.recover(attached_tenants)?;
377 :
378 0 : Ok(Some(generations))
379 0 : }
380 :
381 : /// Given a directory discovered in the pageserver's tenants/ directory, attempt
382 : /// to load a tenant config from it.
383 : ///
384 : /// If we cleaned up something expected (like an empty dir or a temp dir), return None.
385 0 : fn load_tenant_config(
386 0 : conf: &'static PageServerConf,
387 0 : tenant_shard_id: TenantShardId,
388 0 : dentry: Utf8DirEntry,
389 0 : ) -> Option<Result<LocationConf, LoadConfigError>> {
390 0 : let tenant_dir_path = dentry.path().to_path_buf();
391 0 : if crate::is_temporary(&tenant_dir_path) {
392 0 : info!("Found temporary tenant directory, removing: {tenant_dir_path}");
393 : // No need to use safe_remove_tenant_dir_all because this is already
394 : // a temporary path
395 0 : std::fs::remove_dir_all(&tenant_dir_path).fatal_err("delete temporary tenant dir");
396 0 : return None;
397 0 : }
398 0 :
399 0 : // This case happens if we crash during attachment before writing a config into the dir
400 0 : let is_empty = tenant_dir_path
401 0 : .is_empty_dir()
402 0 : .fatal_err("Checking for empty tenant dir");
403 0 : if is_empty {
404 0 : info!("removing empty tenant directory {tenant_dir_path:?}");
405 0 : std::fs::remove_dir(&tenant_dir_path).fatal_err("delete empty tenant dir");
406 0 : return None;
407 0 : }
408 0 :
409 0 : Some(Tenant::load_tenant_config(conf, &tenant_shard_id))
410 0 : }
411 :
412 : /// Initial stage of load: walk the local tenants directory, clean up any temp files,
413 : /// and load configurations for the tenants we found.
414 : ///
415 : /// Do this in parallel, because we expect 10k+ tenants, so serial execution can take
416 : /// seconds even on reasonably fast drives.
417 0 : async fn init_load_tenant_configs(
418 0 : conf: &'static PageServerConf,
419 0 : ) -> HashMap<TenantShardId, Result<LocationConf, LoadConfigError>> {
420 0 : let tenants_dir = conf.tenants_path();
421 :
422 0 : let dentries = tokio::task::spawn_blocking(move || -> Vec<Utf8DirEntry> {
423 0 : let context = format!("read tenants dir {tenants_dir}");
424 0 : let dir_entries = tenants_dir.read_dir_utf8().fatal_err(&context);
425 0 :
426 0 : dir_entries
427 0 : .collect::<Result<Vec<_>, std::io::Error>>()
428 0 : .fatal_err(&context)
429 0 : })
430 0 : .await
431 0 : .expect("Config load task panicked");
432 0 :
433 0 : let mut configs = HashMap::new();
434 0 :
435 0 : let mut join_set = JoinSet::new();
436 0 : for dentry in dentries {
437 0 : let tenant_shard_id = match dentry.file_name().parse::<TenantShardId>() {
438 0 : Ok(id) => id,
439 : Err(_) => {
440 0 : warn!(
441 0 : "Invalid tenant path (garbage in our repo directory?): '{}'",
442 0 : dentry.file_name()
443 : );
444 0 : continue;
445 : }
446 : };
447 :
448 0 : join_set.spawn_blocking(move || {
449 0 : (
450 0 : tenant_shard_id,
451 0 : load_tenant_config(conf, tenant_shard_id, dentry),
452 0 : )
453 0 : });
454 0 : }
455 :
456 0 : while let Some(r) = join_set.join_next().await {
457 0 : let (tenant_shard_id, tenant_config) = r.expect("Panic in config load task");
458 0 : if let Some(tenant_config) = tenant_config {
459 0 : configs.insert(tenant_shard_id, tenant_config);
460 0 : }
461 : }
462 :
463 0 : configs
464 0 : }
465 :
466 0 : #[derive(Debug, thiserror::Error)]
467 : pub(crate) enum DeleteTenantError {
468 : #[error("Tenant map slot error {0}")]
469 : SlotError(#[from] TenantSlotError),
470 :
471 : #[error("Cancelled")]
472 : Cancelled,
473 :
474 : #[error(transparent)]
475 : Other(#[from] anyhow::Error),
476 : }
477 :
478 : /// Initialize repositories with locally available timelines.
479 : /// Timelines that are only partially available locally (remote storage has more data than this pageserver)
480 : /// are scheduled for download and added to the tenant once download is completed.
481 0 : #[instrument(skip_all)]
482 : pub async fn init_tenant_mgr(
483 : conf: &'static PageServerConf,
484 : background_purges: BackgroundPurges,
485 : resources: TenantSharedResources,
486 : init_order: InitializationOrder,
487 : cancel: CancellationToken,
488 : ) -> anyhow::Result<TenantManager> {
489 : let mut tenants = BTreeMap::new();
490 :
491 : let ctx = RequestContext::todo_child(TaskKind::Startup, DownloadBehavior::Warn);
492 :
493 : // Initialize dynamic limits that depend on system resources
494 : let system_memory =
495 : sysinfo::System::new_with_specifics(sysinfo::RefreshKind::new().with_memory())
496 : .total_memory();
497 : let max_ephemeral_layer_bytes =
498 : conf.ephemeral_bytes_per_memory_kb as u64 * (system_memory / 1024);
499 : tracing::info!("Initialized ephemeral layer size limit to {max_ephemeral_layer_bytes}, for {system_memory} bytes of memory");
500 : inmemory_layer::GLOBAL_RESOURCES.max_dirty_bytes.store(
501 : max_ephemeral_layer_bytes,
502 : std::sync::atomic::Ordering::Relaxed,
503 : );
504 :
505 : // Scan local filesystem for attached tenants
506 : let tenant_configs = init_load_tenant_configs(conf).await;
507 :
508 : // Determine which tenants are to be secondary or attached, and in which generation
509 : let tenant_modes = init_load_generations(conf, &tenant_configs, &resources, &cancel).await?;
510 :
511 : tracing::info!(
512 : "Attaching {} tenants at startup, warming up {} at a time",
513 : tenant_configs.len(),
514 : conf.concurrent_tenant_warmup.initial_permits()
515 : );
516 : TENANT.startup_scheduled.inc_by(tenant_configs.len() as u64);
517 :
518 : // Accumulate futures for writing tenant configs, so that we can execute in parallel
519 : let mut config_write_futs = Vec::new();
520 :
521 : // Update the location configs according to the re-attach response and persist them to disk
522 : tracing::info!("Updating {} location configs", tenant_configs.len());
523 : for (tenant_shard_id, location_conf) in tenant_configs {
524 : let tenant_dir_path = conf.tenant_path(&tenant_shard_id);
525 :
526 : let mut location_conf = match location_conf {
527 : Ok(l) => l,
528 : Err(e) => {
529 : // This should only happen in the case of a serialization bug or critical local I/O error: we cannot load this tenant
530 : error!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Failed to load tenant config, failed to {e:#}");
531 : continue;
532 : }
533 : };
534 :
535 : // FIXME: if we were attached, and get demoted to secondary on re-attach, we
536 : // don't have a place to get a config.
537 : // (https://github.com/neondatabase/neon/issues/5377)
538 : const DEFAULT_SECONDARY_CONF: SecondaryLocationConfig =
539 : SecondaryLocationConfig { warm: true };
540 :
541 : if let Some(tenant_modes) = &tenant_modes {
542 : // We have a generation map: treat it as the authority for whether
543 : // this tenant is really attached.
544 : match tenant_modes.get(&tenant_shard_id) {
545 : None => {
546 : 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");
547 :
548 : match safe_rename_tenant_dir(&tenant_dir_path).await {
549 : Ok(tmp_path) => {
550 : background_purges.spawn(tmp_path);
551 : }
552 : Err(e) => {
553 : error!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
554 : "Failed to move detached tenant directory '{tenant_dir_path}': {e:?}");
555 : }
556 : };
557 :
558 : // We deleted local content: move on to next tenant, don't try and spawn this one.
559 : continue;
560 : }
561 : Some(TenantStartupMode::Secondary) => {
562 : if !matches!(location_conf.mode, LocationMode::Secondary(_)) {
563 : location_conf.mode = LocationMode::Secondary(DEFAULT_SECONDARY_CONF);
564 : }
565 : }
566 : Some(TenantStartupMode::Attached((attach_mode, generation))) => {
567 : let old_gen_higher = match &location_conf.mode {
568 : LocationMode::Attached(AttachedLocationConfig {
569 : generation: old_generation,
570 : attach_mode: _attach_mode,
571 : }) => {
572 : if old_generation > generation {
573 : Some(old_generation)
574 : } else {
575 : None
576 : }
577 : }
578 : _ => None,
579 : };
580 : if let Some(old_generation) = old_gen_higher {
581 : tracing::error!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
582 : "Control plane gave decreasing generation ({generation:?}) in re-attach response for tenant that was attached in generation {:?}, demoting to secondary",
583 : old_generation
584 : );
585 :
586 : // We cannot safely attach this tenant given a bogus generation number, but let's avoid throwing away
587 : // local disk content: demote to secondary rather than detaching.
588 : location_conf.mode = LocationMode::Secondary(DEFAULT_SECONDARY_CONF);
589 : } else {
590 : location_conf.attach_in_generation(*attach_mode, *generation);
591 : }
592 : }
593 : }
594 : } else {
595 : // Legacy mode: no generation information, any tenant present
596 : // on local disk may activate
597 : info!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Starting tenant in legacy mode, no generation",);
598 : };
599 :
600 : // Presence of a generation number implies attachment: attach the tenant
601 : // if it wasn't already, and apply the generation number.
602 0 : config_write_futs.push(async move {
603 0 : let r = Tenant::persist_tenant_config(conf, &tenant_shard_id, &location_conf).await;
604 0 : (tenant_shard_id, location_conf, r)
605 0 : });
606 : }
607 :
608 : // Execute config writes with concurrency, to avoid bottlenecking on local FS write latency
609 : tracing::info!(
610 : "Writing {} location config files...",
611 : config_write_futs.len()
612 : );
613 : let config_write_results = futures::stream::iter(config_write_futs)
614 : .buffer_unordered(16)
615 : .collect::<Vec<_>>()
616 : .await;
617 :
618 : tracing::info!(
619 : "Spawning {} tenant shard locations...",
620 : config_write_results.len()
621 : );
622 : // For those shards that have live configurations, construct `Tenant` or `SecondaryTenant` objects and start them running
623 : for (tenant_shard_id, location_conf, config_write_result) in config_write_results {
624 : // Writing a config to local disk is foundational to startup up tenants: panic if we can't.
625 : config_write_result.fatal_err("write tenant shard config file");
626 :
627 : let tenant_dir_path = conf.tenant_path(&tenant_shard_id);
628 : let shard_identity = location_conf.shard;
629 : let slot = match location_conf.mode {
630 : LocationMode::Attached(attached_conf) => TenantSlot::Attached(
631 : tenant_spawn(
632 : conf,
633 : tenant_shard_id,
634 : &tenant_dir_path,
635 : resources.clone(),
636 : AttachedTenantConf::new(location_conf.tenant_conf, attached_conf),
637 : shard_identity,
638 : Some(init_order.clone()),
639 : SpawnMode::Lazy,
640 : &ctx,
641 : )
642 : .expect("global shutdown during init_tenant_mgr cannot happen"),
643 : ),
644 : LocationMode::Secondary(secondary_conf) => {
645 : info!(
646 : tenant_id = %tenant_shard_id.tenant_id,
647 : shard_id = %tenant_shard_id.shard_slug(),
648 : "Starting secondary tenant"
649 : );
650 : TenantSlot::Secondary(SecondaryTenant::new(
651 : tenant_shard_id,
652 : shard_identity,
653 : location_conf.tenant_conf,
654 : &secondary_conf,
655 : ))
656 : }
657 : };
658 :
659 : METRICS.slot_inserted(&slot);
660 : tenants.insert(tenant_shard_id, slot);
661 : }
662 :
663 : info!("Processed {} local tenants at startup", tenants.len());
664 :
665 : let mut tenants_map = TENANTS.write().unwrap();
666 : assert!(matches!(&*tenants_map, &TenantsMap::Initializing));
667 :
668 : *tenants_map = TenantsMap::Open(tenants);
669 :
670 : Ok(TenantManager {
671 : conf,
672 : tenants: &TENANTS,
673 : resources,
674 : cancel: CancellationToken::new(),
675 : background_purges,
676 : })
677 : }
678 :
679 : /// Wrapper for Tenant::spawn that checks invariants before running
680 : #[allow(clippy::too_many_arguments)]
681 0 : fn tenant_spawn(
682 0 : conf: &'static PageServerConf,
683 0 : tenant_shard_id: TenantShardId,
684 0 : tenant_path: &Utf8Path,
685 0 : resources: TenantSharedResources,
686 0 : location_conf: AttachedTenantConf,
687 0 : shard_identity: ShardIdentity,
688 0 : init_order: Option<InitializationOrder>,
689 0 : mode: SpawnMode,
690 0 : ctx: &RequestContext,
691 0 : ) -> Result<Arc<Tenant>, GlobalShutDown> {
692 0 : // All these conditions should have been satisfied by our caller: the tenant dir exists, is a well formed
693 0 : // path, and contains a configuration file. Assertions that do synchronous I/O are limited to debug mode
694 0 : // to avoid impacting prod runtime performance.
695 0 : assert!(!crate::is_temporary(tenant_path));
696 0 : debug_assert!(tenant_path.is_dir());
697 0 : debug_assert!(conf
698 0 : .tenant_location_config_path(&tenant_shard_id)
699 0 : .try_exists()
700 0 : .unwrap());
701 :
702 0 : Tenant::spawn(
703 0 : conf,
704 0 : tenant_shard_id,
705 0 : resources,
706 0 : location_conf,
707 0 : shard_identity,
708 0 : init_order,
709 0 : mode,
710 0 : ctx,
711 0 : )
712 0 : }
713 :
714 2 : async fn shutdown_all_tenants0(tenants: &std::sync::RwLock<TenantsMap>) {
715 2 : let mut join_set = JoinSet::new();
716 0 :
717 0 : #[cfg(all(debug_assertions, not(test)))]
718 0 : {
719 0 : // Check that our metrics properly tracked the size of the tenants map. This is a convenient location to check,
720 0 : // as it happens implicitly at the end of tests etc.
721 0 : let m = tenants.read().unwrap();
722 0 : debug_assert_eq!(METRICS.slots_total(), m.len() as u64);
723 : }
724 :
725 : // Atomically, 1. create the shutdown tasks and 2. prevent creation of new tenants.
726 2 : let (total_in_progress, total_attached) = {
727 2 : let mut m = tenants.write().unwrap();
728 2 : match &mut *m {
729 : TenantsMap::Initializing => {
730 0 : *m = TenantsMap::ShuttingDown(BTreeMap::default());
731 0 : info!("tenants map is empty");
732 0 : return;
733 : }
734 2 : TenantsMap::Open(tenants) => {
735 2 : let mut shutdown_state = BTreeMap::new();
736 2 : let mut total_in_progress = 0;
737 2 : let mut total_attached = 0;
738 :
739 2 : for (tenant_shard_id, v) in std::mem::take(tenants).into_iter() {
740 2 : match v {
741 0 : TenantSlot::Attached(t) => {
742 0 : shutdown_state.insert(tenant_shard_id, TenantSlot::Attached(t.clone()));
743 0 : join_set.spawn(
744 0 : async move {
745 0 : let res = {
746 0 : let (_guard, shutdown_progress) = completion::channel();
747 0 : t.shutdown(shutdown_progress, ShutdownMode::FreezeAndFlush).await
748 : };
749 :
750 0 : if let Err(other_progress) = res {
751 : // join the another shutdown in progress
752 0 : other_progress.wait().await;
753 0 : }
754 :
755 : // we cannot afford per tenant logging here, because if s3 is degraded, we are
756 : // going to log too many lines
757 0 : debug!("tenant successfully stopped");
758 0 : }
759 0 : .instrument(info_span!("shutdown", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug())),
760 : );
761 :
762 0 : total_attached += 1;
763 : }
764 0 : TenantSlot::Secondary(state) => {
765 0 : // We don't need to wait for this individually per-tenant: the
766 0 : // downloader task will be waited on eventually, this cancel
767 0 : // is just to encourage it to drop out if it is doing work
768 0 : // for this tenant right now.
769 0 : state.cancel.cancel();
770 0 :
771 0 : shutdown_state.insert(tenant_shard_id, TenantSlot::Secondary(state));
772 0 : }
773 2 : TenantSlot::InProgress(notify) => {
774 2 : // InProgress tenants are not visible in TenantsMap::ShuttingDown: we will
775 2 : // wait for their notifications to fire in this function.
776 2 : join_set.spawn(async move {
777 2 : notify.wait().await;
778 2 : });
779 2 :
780 2 : total_in_progress += 1;
781 2 : }
782 : }
783 : }
784 2 : *m = TenantsMap::ShuttingDown(shutdown_state);
785 2 : (total_in_progress, total_attached)
786 : }
787 : TenantsMap::ShuttingDown(_) => {
788 0 : error!("already shutting down, this function isn't supposed to be called more than once");
789 0 : return;
790 : }
791 : }
792 : };
793 :
794 2 : let started_at = std::time::Instant::now();
795 2 :
796 2 : info!(
797 0 : "Waiting for {} InProgress tenants and {} Attached tenants to shut down",
798 : total_in_progress, total_attached
799 : );
800 :
801 2 : let total = join_set.len();
802 2 : let mut panicked = 0;
803 2 : let mut buffering = true;
804 2 : const BUFFER_FOR: std::time::Duration = std::time::Duration::from_millis(500);
805 2 : let mut buffered = std::pin::pin!(tokio::time::sleep(BUFFER_FOR));
806 :
807 6 : while !join_set.is_empty() {
808 : tokio::select! {
809 : Some(joined) = join_set.join_next() => {
810 : match joined {
811 : Ok(()) => {},
812 : Err(join_error) if join_error.is_cancelled() => {
813 : unreachable!("we are not cancelling any of the tasks");
814 : }
815 : Err(join_error) if join_error.is_panic() => {
816 : // cannot really do anything, as this panic is likely a bug
817 : panicked += 1;
818 : }
819 : Err(join_error) => {
820 : warn!("unknown kind of JoinError: {join_error}");
821 : }
822 : }
823 : if !buffering {
824 : // buffer so that every 500ms since the first update (or starting) we'll log
825 : // how far away we are; this is because we will get SIGKILL'd at 10s, and we
826 : // are not able to log *then*.
827 : buffering = true;
828 : buffered.as_mut().reset(tokio::time::Instant::now() + BUFFER_FOR);
829 : }
830 : },
831 : _ = &mut buffered, if buffering => {
832 : buffering = false;
833 : info!(remaining = join_set.len(), total, elapsed_ms = started_at.elapsed().as_millis(), "waiting for tenants to shutdown");
834 : }
835 2 : }
836 2 : }
837 2 :
838 2 : if panicked > 0 {
839 0 : warn!(
840 : panicked,
841 0 : total, "observed panicks while shutting down tenants"
842 : );
843 2 : }
844 :
845 : // caller will log how long we took
846 2 : }
847 :
848 0 : #[derive(thiserror::Error, Debug)]
849 : pub(crate) enum UpsertLocationError {
850 : #[error("Bad config request: {0}")]
851 : BadRequest(anyhow::Error),
852 :
853 : #[error("Cannot change config in this state: {0}")]
854 : Unavailable(#[from] TenantMapError),
855 :
856 : #[error("Tenant is already being modified")]
857 : InProgress,
858 :
859 : #[error("Failed to flush: {0}")]
860 : Flush(anyhow::Error),
861 :
862 : /// This error variant is for unexpected situations (soft assertions) where the system is in an unexpected state.
863 : #[error("Internal error: {0}")]
864 : InternalError(anyhow::Error),
865 : }
866 :
867 : impl TenantManager {
868 : /// Convenience function so that anyone with a TenantManager can get at the global configuration, without
869 : /// having to pass it around everywhere as a separate object.
870 0 : pub(crate) fn get_conf(&self) -> &'static PageServerConf {
871 0 : self.conf
872 0 : }
873 :
874 : /// Gets the attached tenant from the in-memory data, erroring if it's absent, in secondary mode, or currently
875 : /// undergoing a state change (i.e. slot is InProgress).
876 : ///
877 : /// The return Tenant is not guaranteed to be active: check its status after obtaing it, or
878 : /// use [`Tenant::wait_to_become_active`] before using it if you will do I/O on it.
879 0 : pub(crate) fn get_attached_tenant_shard(
880 0 : &self,
881 0 : tenant_shard_id: TenantShardId,
882 0 : ) -> Result<Arc<Tenant>, GetTenantError> {
883 0 : let locked = self.tenants.read().unwrap();
884 :
885 0 : let peek_slot = tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Read)?;
886 :
887 0 : match peek_slot {
888 0 : Some(TenantSlot::Attached(tenant)) => Ok(Arc::clone(tenant)),
889 0 : Some(TenantSlot::InProgress(_)) => Err(GetTenantError::NotActive(tenant_shard_id)),
890 : None | Some(TenantSlot::Secondary(_)) => {
891 0 : Err(GetTenantError::NotFound(tenant_shard_id.tenant_id))
892 : }
893 : }
894 0 : }
895 :
896 0 : pub(crate) fn get_secondary_tenant_shard(
897 0 : &self,
898 0 : tenant_shard_id: TenantShardId,
899 0 : ) -> Option<Arc<SecondaryTenant>> {
900 0 : let locked = self.tenants.read().unwrap();
901 0 :
902 0 : let peek_slot = tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Read)
903 0 : .ok()
904 0 : .flatten();
905 :
906 0 : match peek_slot {
907 0 : Some(TenantSlot::Secondary(s)) => Some(s.clone()),
908 0 : _ => None,
909 : }
910 0 : }
911 :
912 : /// Whether the `TenantManager` is responsible for the tenant shard
913 0 : pub(crate) fn manages_tenant_shard(&self, tenant_shard_id: TenantShardId) -> bool {
914 0 : let locked = self.tenants.read().unwrap();
915 0 :
916 0 : let peek_slot = tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Read)
917 0 : .ok()
918 0 : .flatten();
919 0 :
920 0 : peek_slot.is_some()
921 0 : }
922 :
923 0 : #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
924 : pub(crate) async fn upsert_location(
925 : &self,
926 : tenant_shard_id: TenantShardId,
927 : new_location_config: LocationConf,
928 : flush: Option<Duration>,
929 : mut spawn_mode: SpawnMode,
930 : ctx: &RequestContext,
931 : ) -> Result<Option<Arc<Tenant>>, UpsertLocationError> {
932 : debug_assert_current_span_has_tenant_id();
933 : info!("configuring tenant location to state {new_location_config:?}");
934 :
935 : enum FastPathModified {
936 : Attached(Arc<Tenant>),
937 : Secondary(Arc<SecondaryTenant>),
938 : }
939 :
940 : // Special case fast-path for updates to existing slots: if our upsert is only updating configuration,
941 : // then we do not need to set the slot to InProgress, we can just call into the
942 : // existng tenant.
943 : let fast_path_taken = {
944 : let locked = self.tenants.read().unwrap();
945 : let peek_slot =
946 : tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Write)?;
947 : match (&new_location_config.mode, peek_slot) {
948 : (LocationMode::Attached(attach_conf), Some(TenantSlot::Attached(tenant))) => {
949 : match attach_conf.generation.cmp(&tenant.generation) {
950 : Ordering::Equal => {
951 : // A transition from Attached to Attached in the same generation, we may
952 : // take our fast path and just provide the updated configuration
953 : // to the tenant.
954 : tenant.set_new_location_config(
955 : AttachedTenantConf::try_from(new_location_config.clone())
956 : .map_err(UpsertLocationError::BadRequest)?,
957 : );
958 :
959 : Some(FastPathModified::Attached(tenant.clone()))
960 : }
961 : Ordering::Less => {
962 : return Err(UpsertLocationError::BadRequest(anyhow::anyhow!(
963 : "Generation {:?} is less than existing {:?}",
964 : attach_conf.generation,
965 : tenant.generation
966 : )));
967 : }
968 : Ordering::Greater => {
969 : // Generation advanced, fall through to general case of replacing `Tenant` object
970 : None
971 : }
972 : }
973 : }
974 : (
975 : LocationMode::Secondary(secondary_conf),
976 : Some(TenantSlot::Secondary(secondary_tenant)),
977 : ) => {
978 : secondary_tenant.set_config(secondary_conf);
979 : secondary_tenant.set_tenant_conf(&new_location_config.tenant_conf);
980 : Some(FastPathModified::Secondary(secondary_tenant.clone()))
981 : }
982 : _ => {
983 : // Not an Attached->Attached transition, fall through to general case
984 : None
985 : }
986 : }
987 : };
988 :
989 : // Fast-path continued: having dropped out of the self.tenants lock, do the async
990 : // phase of writing config and/or waiting for flush, before returning.
991 : match fast_path_taken {
992 : Some(FastPathModified::Attached(tenant)) => {
993 : Tenant::persist_tenant_config(self.conf, &tenant_shard_id, &new_location_config)
994 : .await
995 : .fatal_err("write tenant shard config");
996 :
997 : // Transition to AttachedStale means we may well hold a valid generation
998 : // still, and have been requested to go stale as part of a migration. If
999 : // the caller set `flush`, then flush to remote storage.
1000 : if let LocationMode::Attached(AttachedLocationConfig {
1001 : generation: _,
1002 : attach_mode: AttachmentMode::Stale,
1003 : }) = &new_location_config.mode
1004 : {
1005 : if let Some(flush_timeout) = flush {
1006 : match tokio::time::timeout(flush_timeout, tenant.flush_remote()).await {
1007 : Ok(Err(e)) => {
1008 : return Err(UpsertLocationError::Flush(e));
1009 : }
1010 : Ok(Ok(_)) => return Ok(Some(tenant)),
1011 : Err(_) => {
1012 : tracing::warn!(
1013 : timeout_ms = flush_timeout.as_millis(),
1014 : "Timed out waiting for flush to remote storage, proceeding anyway."
1015 : )
1016 : }
1017 : }
1018 : }
1019 : }
1020 :
1021 : return Ok(Some(tenant));
1022 : }
1023 : Some(FastPathModified::Secondary(_secondary_tenant)) => {
1024 : Tenant::persist_tenant_config(self.conf, &tenant_shard_id, &new_location_config)
1025 : .await
1026 : .fatal_err("write tenant shard config");
1027 :
1028 : return Ok(None);
1029 : }
1030 : None => {
1031 : // Proceed with the general case procedure, where we will shutdown & remove any existing
1032 : // slot contents and replace with a fresh one
1033 : }
1034 : };
1035 :
1036 : // General case for upserts to TenantsMap, excluding the case above: we will substitute an
1037 : // InProgress value to the slot while we make whatever changes are required. The state for
1038 : // the tenant is inaccessible to the outside world while we are doing this, but that is sensible:
1039 : // the state is ill-defined while we're in transition. Transitions are async, but fast: we do
1040 : // not do significant I/O, and shutdowns should be prompt via cancellation tokens.
1041 : let mut slot_guard = tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)
1042 0 : .map_err(|e| match e {
1043 : TenantSlotError::NotFound(_) => {
1044 0 : unreachable!("Called with mode Any")
1045 : }
1046 0 : TenantSlotError::InProgress => UpsertLocationError::InProgress,
1047 0 : TenantSlotError::MapState(s) => UpsertLocationError::Unavailable(s),
1048 0 : })?;
1049 :
1050 : match slot_guard.get_old_value() {
1051 : Some(TenantSlot::Attached(tenant)) => {
1052 : // The case where we keep a Tenant alive was covered above in the special case
1053 : // for Attached->Attached transitions in the same generation. By this point,
1054 : // if we see an attached tenant we know it will be discarded and should be
1055 : // shut down.
1056 : let (_guard, progress) = utils::completion::channel();
1057 :
1058 : match tenant.get_attach_mode() {
1059 : AttachmentMode::Single | AttachmentMode::Multi => {
1060 : // Before we leave our state as the presumed holder of the latest generation,
1061 : // flush any outstanding deletions to reduce the risk of leaking objects.
1062 : self.resources.deletion_queue_client.flush_advisory()
1063 : }
1064 : AttachmentMode::Stale => {
1065 : // If we're stale there's not point trying to flush deletions
1066 : }
1067 : };
1068 :
1069 : info!("Shutting down attached tenant");
1070 : match tenant.shutdown(progress, ShutdownMode::Hard).await {
1071 : Ok(()) => {}
1072 : Err(barrier) => {
1073 : info!("Shutdown already in progress, waiting for it to complete");
1074 : barrier.wait().await;
1075 : }
1076 : }
1077 : slot_guard.drop_old_value().expect("We just shut it down");
1078 :
1079 : // Edge case: if we were called with SpawnMode::Create, but a Tenant already existed, then
1080 : // the caller thinks they're creating but the tenant already existed. We must switch to
1081 : // Eager mode so that when starting this Tenant we properly probe remote storage for timelines,
1082 : // rather than assuming it to be empty.
1083 : spawn_mode = SpawnMode::Eager;
1084 : }
1085 : Some(TenantSlot::Secondary(state)) => {
1086 : info!("Shutting down secondary tenant");
1087 : state.shutdown().await;
1088 : }
1089 : Some(TenantSlot::InProgress(_)) => {
1090 : // This should never happen: acquire_slot should error out
1091 : // if the contents of a slot were InProgress.
1092 : return Err(UpsertLocationError::InternalError(anyhow::anyhow!(
1093 : "Acquired an InProgress slot, this is a bug."
1094 : )));
1095 : }
1096 : None => {
1097 : // Slot was vacant, nothing needs shutting down.
1098 : }
1099 : }
1100 :
1101 : let tenant_path = self.conf.tenant_path(&tenant_shard_id);
1102 : let timelines_path = self.conf.timelines_path(&tenant_shard_id);
1103 :
1104 : // Directory structure is the same for attached and secondary modes:
1105 : // create it if it doesn't exist. Timeline load/creation expects the
1106 : // timelines/ subdir to already exist.
1107 : //
1108 : // Does not need to be fsync'd because local storage is just a cache.
1109 : tokio::fs::create_dir_all(&timelines_path)
1110 : .await
1111 : .fatal_err("create timelines/ dir");
1112 :
1113 : // Before activating either secondary or attached mode, persist the
1114 : // configuration, so that on restart we will re-attach (or re-start
1115 : // secondary) on the tenant.
1116 : Tenant::persist_tenant_config(self.conf, &tenant_shard_id, &new_location_config)
1117 : .await
1118 : .fatal_err("write tenant shard config");
1119 :
1120 : let new_slot = match &new_location_config.mode {
1121 : LocationMode::Secondary(secondary_config) => {
1122 : let shard_identity = new_location_config.shard;
1123 : TenantSlot::Secondary(SecondaryTenant::new(
1124 : tenant_shard_id,
1125 : shard_identity,
1126 : new_location_config.tenant_conf,
1127 : secondary_config,
1128 : ))
1129 : }
1130 : LocationMode::Attached(_attach_config) => {
1131 : let shard_identity = new_location_config.shard;
1132 :
1133 : // Testing hack: if we are configured with no control plane, then drop the generation
1134 : // from upserts. This enables creating generation-less tenants even though neon_local
1135 : // always uses generations when calling the location conf API.
1136 : let attached_conf = if cfg!(feature = "testing") {
1137 : let mut conf = AttachedTenantConf::try_from(new_location_config)
1138 : .map_err(UpsertLocationError::BadRequest)?;
1139 : if self.conf.control_plane_api.is_none() {
1140 : conf.location.generation = Generation::none();
1141 : }
1142 : conf
1143 : } else {
1144 : AttachedTenantConf::try_from(new_location_config)
1145 : .map_err(UpsertLocationError::BadRequest)?
1146 : };
1147 :
1148 : let tenant = tenant_spawn(
1149 : self.conf,
1150 : tenant_shard_id,
1151 : &tenant_path,
1152 : self.resources.clone(),
1153 : attached_conf,
1154 : shard_identity,
1155 : None,
1156 : spawn_mode,
1157 : ctx,
1158 : )
1159 0 : .map_err(|_: GlobalShutDown| {
1160 0 : UpsertLocationError::Unavailable(TenantMapError::ShuttingDown)
1161 0 : })?;
1162 :
1163 : TenantSlot::Attached(tenant)
1164 : }
1165 : };
1166 :
1167 : let attached_tenant = if let TenantSlot::Attached(tenant) = &new_slot {
1168 : Some(tenant.clone())
1169 : } else {
1170 : None
1171 : };
1172 :
1173 : match slot_guard.upsert(new_slot) {
1174 : Err(TenantSlotUpsertError::InternalError(e)) => {
1175 : Err(UpsertLocationError::InternalError(anyhow::anyhow!(e)))
1176 : }
1177 : Err(TenantSlotUpsertError::MapState(e)) => Err(UpsertLocationError::Unavailable(e)),
1178 : Err(TenantSlotUpsertError::ShuttingDown((new_slot, _completion))) => {
1179 : // If we just called tenant_spawn() on a new tenant, and can't insert it into our map, then
1180 : // we must not leak it: this would violate the invariant that after shutdown_all_tenants, all tenants
1181 : // are shutdown.
1182 : //
1183 : // We must shut it down inline here.
1184 : match new_slot {
1185 : TenantSlot::InProgress(_) => {
1186 : // Unreachable because we never insert an InProgress
1187 : unreachable!()
1188 : }
1189 : TenantSlot::Attached(tenant) => {
1190 : let (_guard, progress) = utils::completion::channel();
1191 : info!("Shutting down just-spawned tenant, because tenant manager is shut down");
1192 : match tenant.shutdown(progress, ShutdownMode::Hard).await {
1193 : Ok(()) => {
1194 : info!("Finished shutting down just-spawned tenant");
1195 : }
1196 : Err(barrier) => {
1197 : info!("Shutdown already in progress, waiting for it to complete");
1198 : barrier.wait().await;
1199 : }
1200 : }
1201 : }
1202 : TenantSlot::Secondary(secondary_tenant) => {
1203 : secondary_tenant.shutdown().await;
1204 : }
1205 : }
1206 :
1207 : Err(UpsertLocationError::Unavailable(
1208 : TenantMapError::ShuttingDown,
1209 : ))
1210 : }
1211 : Ok(()) => Ok(attached_tenant),
1212 : }
1213 : }
1214 :
1215 : /// Resetting a tenant is equivalent to detaching it, then attaching it again with the same
1216 : /// LocationConf that was last used to attach it. Optionally, the local file cache may be
1217 : /// dropped before re-attaching.
1218 : ///
1219 : /// This is not part of a tenant's normal lifecycle: it is used for debug/support, in situations
1220 : /// where an issue is identified that would go away with a restart of the tenant.
1221 : ///
1222 : /// This does not have any special "force" shutdown of a tenant: it relies on the tenant's tasks
1223 : /// to respect the cancellation tokens used in normal shutdown().
1224 0 : #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), %drop_cache))]
1225 : pub(crate) async fn reset_tenant(
1226 : &self,
1227 : tenant_shard_id: TenantShardId,
1228 : drop_cache: bool,
1229 : ctx: &RequestContext,
1230 : ) -> anyhow::Result<()> {
1231 : let mut slot_guard = tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)?;
1232 : let Some(old_slot) = slot_guard.get_old_value() else {
1233 : anyhow::bail!("Tenant not found when trying to reset");
1234 : };
1235 :
1236 : let Some(tenant) = old_slot.get_attached() else {
1237 : slot_guard.revert();
1238 : anyhow::bail!("Tenant is not in attached state");
1239 : };
1240 :
1241 : let (_guard, progress) = utils::completion::channel();
1242 : match tenant.shutdown(progress, ShutdownMode::Hard).await {
1243 : Ok(()) => {
1244 : slot_guard.drop_old_value()?;
1245 : }
1246 : Err(_barrier) => {
1247 : slot_guard.revert();
1248 : anyhow::bail!("Cannot reset Tenant, already shutting down");
1249 : }
1250 : }
1251 :
1252 : let tenant_path = self.conf.tenant_path(&tenant_shard_id);
1253 : let timelines_path = self.conf.timelines_path(&tenant_shard_id);
1254 : let config = Tenant::load_tenant_config(self.conf, &tenant_shard_id)?;
1255 :
1256 : if drop_cache {
1257 : tracing::info!("Dropping local file cache");
1258 :
1259 : match tokio::fs::read_dir(&timelines_path).await {
1260 : Err(e) => {
1261 : tracing::warn!("Failed to list timelines while dropping cache: {}", e);
1262 : }
1263 : Ok(mut entries) => {
1264 : while let Some(entry) = entries.next_entry().await? {
1265 : tokio::fs::remove_dir_all(entry.path()).await?;
1266 : }
1267 : }
1268 : }
1269 : }
1270 :
1271 : let shard_identity = config.shard;
1272 : let tenant = tenant_spawn(
1273 : self.conf,
1274 : tenant_shard_id,
1275 : &tenant_path,
1276 : self.resources.clone(),
1277 : AttachedTenantConf::try_from(config)?,
1278 : shard_identity,
1279 : None,
1280 : SpawnMode::Eager,
1281 : ctx,
1282 : )?;
1283 :
1284 : slot_guard.upsert(TenantSlot::Attached(tenant))?;
1285 :
1286 : Ok(())
1287 : }
1288 :
1289 0 : pub(crate) fn get_attached_active_tenant_shards(&self) -> Vec<Arc<Tenant>> {
1290 0 : let locked = self.tenants.read().unwrap();
1291 0 : match &*locked {
1292 0 : TenantsMap::Initializing => Vec::new(),
1293 0 : TenantsMap::Open(map) | TenantsMap::ShuttingDown(map) => map
1294 0 : .values()
1295 0 : .filter_map(|slot| {
1296 0 : slot.get_attached()
1297 0 : .and_then(|t| if t.is_active() { Some(t.clone()) } else { None })
1298 0 : })
1299 0 : .collect(),
1300 : }
1301 0 : }
1302 : // Do some synchronous work for all tenant slots in Secondary state. The provided
1303 : // callback should be small and fast, as it will be called inside the global
1304 : // TenantsMap lock.
1305 0 : pub(crate) fn foreach_secondary_tenants<F>(&self, mut func: F)
1306 0 : where
1307 0 : // TODO: let the callback return a hint to drop out of the loop early
1308 0 : F: FnMut(&TenantShardId, &Arc<SecondaryTenant>),
1309 0 : {
1310 0 : let locked = self.tenants.read().unwrap();
1311 :
1312 0 : let map = match &*locked {
1313 0 : TenantsMap::Initializing | TenantsMap::ShuttingDown(_) => return,
1314 0 : TenantsMap::Open(m) => m,
1315 : };
1316 :
1317 0 : for (tenant_id, slot) in map {
1318 0 : if let TenantSlot::Secondary(state) = slot {
1319 : // Only expose secondary tenants that are not currently shutting down
1320 0 : if !state.cancel.is_cancelled() {
1321 0 : func(tenant_id, state)
1322 0 : }
1323 0 : }
1324 : }
1325 0 : }
1326 :
1327 : /// Total list of all tenant slots: this includes attached, secondary, and InProgress.
1328 0 : pub(crate) fn list(&self) -> Vec<(TenantShardId, TenantSlot)> {
1329 0 : let locked = self.tenants.read().unwrap();
1330 0 : match &*locked {
1331 0 : TenantsMap::Initializing => Vec::new(),
1332 0 : TenantsMap::Open(map) | TenantsMap::ShuttingDown(map) => {
1333 0 : map.iter().map(|(k, v)| (*k, v.clone())).collect()
1334 : }
1335 : }
1336 0 : }
1337 :
1338 0 : pub(crate) fn get(&self, tenant_shard_id: TenantShardId) -> Option<TenantSlot> {
1339 0 : let locked = self.tenants.read().unwrap();
1340 0 : match &*locked {
1341 0 : TenantsMap::Initializing => None,
1342 0 : TenantsMap::Open(map) | TenantsMap::ShuttingDown(map) => {
1343 0 : map.get(&tenant_shard_id).cloned()
1344 : }
1345 : }
1346 0 : }
1347 :
1348 0 : async fn delete_tenant_remote(
1349 0 : &self,
1350 0 : tenant_shard_id: TenantShardId,
1351 0 : ) -> Result<(), DeleteTenantError> {
1352 0 : let remote_path = remote_tenant_path(&tenant_shard_id);
1353 0 : let mut keys_stream = self.resources.remote_storage.list_streaming(
1354 0 : Some(&remote_path),
1355 0 : remote_storage::ListingMode::NoDelimiter,
1356 0 : None,
1357 0 : &self.cancel,
1358 0 : );
1359 0 : while let Some(chunk) = keys_stream.next().await {
1360 0 : let keys = match chunk {
1361 0 : Ok(listing) => listing.keys,
1362 : Err(remote_storage::DownloadError::Cancelled) => {
1363 0 : return Err(DeleteTenantError::Cancelled)
1364 : }
1365 0 : Err(remote_storage::DownloadError::NotFound) => return Ok(()),
1366 0 : Err(other) => return Err(DeleteTenantError::Other(anyhow::anyhow!(other))),
1367 : };
1368 :
1369 0 : if keys.is_empty() {
1370 0 : tracing::info!("Remote storage already deleted");
1371 : } else {
1372 0 : tracing::info!("Deleting {} keys from remote storage", keys.len());
1373 0 : let keys = keys.into_iter().map(|o| o.key).collect::<Vec<_>>();
1374 0 : self.resources
1375 0 : .remote_storage
1376 0 : .delete_objects(&keys, &self.cancel)
1377 0 : .await?;
1378 : }
1379 : }
1380 :
1381 0 : Ok(())
1382 0 : }
1383 :
1384 : /// If a tenant is attached, detach it. Then remove its data from remote storage.
1385 : ///
1386 : /// A tenant is considered deleted once it is gone from remote storage. It is the caller's
1387 : /// responsibility to avoid trying to attach the tenant again or use it any way once deletion
1388 : /// has started: this operation is not atomic, and must be retried until it succeeds.
1389 0 : pub(crate) async fn delete_tenant(
1390 0 : &self,
1391 0 : tenant_shard_id: TenantShardId,
1392 0 : ) -> Result<(), DeleteTenantError> {
1393 0 : super::span::debug_assert_current_span_has_tenant_id();
1394 :
1395 0 : async fn delete_local(
1396 0 : conf: &PageServerConf,
1397 0 : background_purges: &BackgroundPurges,
1398 0 : tenant_shard_id: &TenantShardId,
1399 0 : ) -> anyhow::Result<()> {
1400 0 : let local_tenant_directory = conf.tenant_path(tenant_shard_id);
1401 0 : let tmp_dir = safe_rename_tenant_dir(&local_tenant_directory)
1402 0 : .await
1403 0 : .with_context(|| {
1404 0 : format!("local tenant directory {local_tenant_directory:?} rename")
1405 0 : })?;
1406 0 : background_purges.spawn(tmp_dir);
1407 0 : Ok(())
1408 0 : }
1409 :
1410 0 : let slot_guard = tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)?;
1411 0 : match &slot_guard.old_value {
1412 0 : Some(TenantSlot::Attached(tenant)) => {
1413 0 : // Legacy deletion flow: the tenant remains attached, goes to Stopping state, and
1414 0 : // deletion will be resumed across restarts.
1415 0 : let tenant = tenant.clone();
1416 0 : let (_guard, progress) = utils::completion::channel();
1417 0 : match tenant.shutdown(progress, ShutdownMode::Hard).await {
1418 0 : Ok(()) => {}
1419 0 : Err(barrier) => {
1420 0 : info!("Shutdown already in progress, waiting for it to complete");
1421 0 : barrier.wait().await;
1422 : }
1423 : }
1424 0 : delete_local(self.conf, &self.background_purges, &tenant_shard_id).await?;
1425 : }
1426 0 : Some(TenantSlot::Secondary(secondary_tenant)) => {
1427 0 : secondary_tenant.shutdown().await;
1428 :
1429 0 : delete_local(self.conf, &self.background_purges, &tenant_shard_id).await?;
1430 : }
1431 0 : Some(TenantSlot::InProgress(_)) => unreachable!(),
1432 0 : None => {}
1433 : };
1434 :
1435 : // Fall through: local state for this tenant is no longer present, proceed with remote delete.
1436 : // - We use a retry wrapper here so that common transient S3 errors (e.g. 503, 429) do not result
1437 : // in 500 responses to delete requests.
1438 : // - We keep the `SlotGuard` during this I/O, so that if a concurrent delete request comes in, it will
1439 : // 503/retry, rather than kicking off a wasteful concurrent deletion.
1440 0 : match backoff::retry(
1441 0 : || async move { self.delete_tenant_remote(tenant_shard_id).await },
1442 0 : |e| match e {
1443 0 : DeleteTenantError::Cancelled => true,
1444 : DeleteTenantError::SlotError(_) => {
1445 0 : unreachable!("Remote deletion doesn't touch slots")
1446 : }
1447 0 : _ => false,
1448 0 : },
1449 0 : 1,
1450 0 : 3,
1451 0 : &format!("delete_tenant[tenant_shard_id={tenant_shard_id}]"),
1452 0 : &self.cancel,
1453 0 : )
1454 0 : .await
1455 : {
1456 0 : Some(r) => r,
1457 0 : None => Err(DeleteTenantError::Cancelled),
1458 : }
1459 0 : }
1460 :
1461 0 : #[instrument(skip_all, fields(tenant_id=%tenant.get_tenant_shard_id().tenant_id, shard_id=%tenant.get_tenant_shard_id().shard_slug(), new_shard_count=%new_shard_count.literal()))]
1462 : pub(crate) async fn shard_split(
1463 : &self,
1464 : tenant: Arc<Tenant>,
1465 : new_shard_count: ShardCount,
1466 : new_stripe_size: Option<ShardStripeSize>,
1467 : ctx: &RequestContext,
1468 : ) -> anyhow::Result<Vec<TenantShardId>> {
1469 : let tenant_shard_id = *tenant.get_tenant_shard_id();
1470 : let r = self
1471 : .do_shard_split(tenant, new_shard_count, new_stripe_size, ctx)
1472 : .await;
1473 : if r.is_err() {
1474 : // Shard splitting might have left the original shard in a partially shut down state (it
1475 : // stops the shard's remote timeline client). Reset it to ensure we leave things in
1476 : // a working state.
1477 : if self.get(tenant_shard_id).is_some() {
1478 : tracing::warn!("Resetting after shard split failure");
1479 : if let Err(e) = self.reset_tenant(tenant_shard_id, false, ctx).await {
1480 : // Log this error because our return value will still be the original error, not this one. This is
1481 : // a severe error: if this happens, we might be leaving behind a tenant that is not fully functional
1482 : // (e.g. has uploads disabled). We can't do anything else: if reset fails then shutting the tenant down or
1483 : // setting it broken probably won't help either.
1484 : tracing::error!("Failed to reset: {e}");
1485 : }
1486 : }
1487 : }
1488 :
1489 : r
1490 : }
1491 :
1492 0 : pub(crate) async fn do_shard_split(
1493 0 : &self,
1494 0 : tenant: Arc<Tenant>,
1495 0 : new_shard_count: ShardCount,
1496 0 : new_stripe_size: Option<ShardStripeSize>,
1497 0 : ctx: &RequestContext,
1498 0 : ) -> anyhow::Result<Vec<TenantShardId>> {
1499 0 : let tenant_shard_id = *tenant.get_tenant_shard_id();
1500 0 :
1501 0 : // Validate the incoming request
1502 0 : if new_shard_count.count() <= tenant_shard_id.shard_count.count() {
1503 0 : anyhow::bail!("Requested shard count is not an increase");
1504 0 : }
1505 0 : let expansion_factor = new_shard_count.count() / tenant_shard_id.shard_count.count();
1506 0 : if !expansion_factor.is_power_of_two() {
1507 0 : anyhow::bail!("Requested split is not a power of two");
1508 0 : }
1509 :
1510 0 : if let Some(new_stripe_size) = new_stripe_size {
1511 0 : if tenant.get_shard_stripe_size() != new_stripe_size
1512 0 : && tenant_shard_id.shard_count.count() > 1
1513 : {
1514 : // This tenant already has multiple shards, it is illegal to try and change its stripe size
1515 0 : anyhow::bail!(
1516 0 : "Shard stripe size may not be modified once tenant has multiple shards"
1517 0 : );
1518 0 : }
1519 0 : }
1520 :
1521 : // Plan: identify what the new child shards will be
1522 0 : let child_shards = tenant_shard_id.split(new_shard_count);
1523 0 : tracing::info!(
1524 0 : "Shard {} splits into: {}",
1525 0 : tenant_shard_id.to_index(),
1526 0 : child_shards
1527 0 : .iter()
1528 0 : .map(|id| format!("{}", id.to_index()))
1529 0 : .join(",")
1530 : );
1531 :
1532 0 : fail::fail_point!("shard-split-pre-prepare", |_| Err(anyhow::anyhow!(
1533 0 : "failpoint"
1534 0 : )));
1535 :
1536 0 : let parent_shard_identity = tenant.shard_identity;
1537 0 : let parent_tenant_conf = tenant.get_tenant_conf();
1538 0 : let parent_generation = tenant.generation;
1539 :
1540 : // Phase 1: Write out child shards' remote index files, in the parent tenant's current generation
1541 0 : if let Err(e) = tenant.split_prepare(&child_shards).await {
1542 : // If [`Tenant::split_prepare`] fails, we must reload the tenant, because it might
1543 : // have been left in a partially-shut-down state.
1544 0 : tracing::warn!("Failed to prepare for split: {e}, reloading Tenant before returning");
1545 0 : return Err(e);
1546 0 : }
1547 0 :
1548 0 : fail::fail_point!("shard-split-post-prepare", |_| Err(anyhow::anyhow!(
1549 0 : "failpoint"
1550 0 : )));
1551 :
1552 0 : self.resources.deletion_queue_client.flush_advisory();
1553 0 :
1554 0 : // Phase 2: Put the parent shard to InProgress and grab a reference to the parent Tenant
1555 0 : drop(tenant);
1556 0 : let mut parent_slot_guard =
1557 0 : tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)?;
1558 0 : let parent = match parent_slot_guard.get_old_value() {
1559 0 : Some(TenantSlot::Attached(t)) => t,
1560 0 : Some(TenantSlot::Secondary(_)) => anyhow::bail!("Tenant location in secondary mode"),
1561 : Some(TenantSlot::InProgress(_)) => {
1562 : // tenant_map_acquire_slot never returns InProgress, if a slot was InProgress
1563 : // it would return an error.
1564 0 : unreachable!()
1565 : }
1566 : None => {
1567 : // We don't actually need the parent shard to still be attached to do our work, but it's
1568 : // a weird enough situation that the caller probably didn't want us to continue working
1569 : // if they had detached the tenant they requested the split on.
1570 0 : anyhow::bail!("Detached parent shard in the middle of split!")
1571 : }
1572 : };
1573 0 : fail::fail_point!("shard-split-pre-hardlink", |_| Err(anyhow::anyhow!(
1574 0 : "failpoint"
1575 0 : )));
1576 : // Optimization: hardlink layers from the parent into the children, so that they don't have to
1577 : // re-download & duplicate the data referenced in their initial IndexPart
1578 0 : self.shard_split_hardlink(parent, child_shards.clone())
1579 0 : .await?;
1580 0 : fail::fail_point!("shard-split-post-hardlink", |_| Err(anyhow::anyhow!(
1581 0 : "failpoint"
1582 0 : )));
1583 :
1584 : // Take a snapshot of where the parent's WAL ingest had got to: we will wait for
1585 : // child shards to reach this point.
1586 0 : let mut target_lsns = HashMap::new();
1587 0 : for timeline in parent.timelines.lock().unwrap().clone().values() {
1588 0 : target_lsns.insert(timeline.timeline_id, timeline.get_last_record_lsn());
1589 0 : }
1590 :
1591 : // TODO: we should have the parent shard stop its WAL ingest here, it's a waste of resources
1592 : // and could slow down the children trying to catch up.
1593 :
1594 : // Phase 3: Spawn the child shards
1595 0 : for child_shard in &child_shards {
1596 0 : let mut child_shard_identity = parent_shard_identity;
1597 0 : if let Some(new_stripe_size) = new_stripe_size {
1598 0 : child_shard_identity.stripe_size = new_stripe_size;
1599 0 : }
1600 0 : child_shard_identity.count = child_shard.shard_count;
1601 0 : child_shard_identity.number = child_shard.shard_number;
1602 0 :
1603 0 : let child_location_conf = LocationConf {
1604 0 : mode: LocationMode::Attached(AttachedLocationConfig {
1605 0 : generation: parent_generation,
1606 0 : attach_mode: AttachmentMode::Single,
1607 0 : }),
1608 0 : shard: child_shard_identity,
1609 0 : tenant_conf: parent_tenant_conf.clone(),
1610 0 : };
1611 0 :
1612 0 : self.upsert_location(
1613 0 : *child_shard,
1614 0 : child_location_conf,
1615 0 : None,
1616 0 : SpawnMode::Eager,
1617 0 : ctx,
1618 0 : )
1619 0 : .await?;
1620 : }
1621 :
1622 0 : fail::fail_point!("shard-split-post-child-conf", |_| Err(anyhow::anyhow!(
1623 0 : "failpoint"
1624 0 : )));
1625 :
1626 : // Phase 4: wait for child chards WAL ingest to catch up to target LSN
1627 0 : for child_shard_id in &child_shards {
1628 0 : let child_shard_id = *child_shard_id;
1629 0 : let child_shard = {
1630 0 : let locked = self.tenants.read().unwrap();
1631 0 : let peek_slot =
1632 0 : tenant_map_peek_slot(&locked, &child_shard_id, TenantSlotPeekMode::Read)?;
1633 0 : peek_slot.and_then(|s| s.get_attached()).cloned()
1634 : };
1635 0 : if let Some(t) = child_shard {
1636 : // Wait for the child shard to become active: this should be very quick because it only
1637 : // has to download the index_part that we just uploaded when creating it.
1638 0 : if let Err(e) = t.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await {
1639 : // This is not fatal: we have durably created the child shard. It just makes the
1640 : // split operation less seamless for clients, as we will may detach the parent
1641 : // shard before the child shards are fully ready to serve requests.
1642 0 : tracing::warn!("Failed to wait for shard {child_shard_id} to activate: {e}");
1643 0 : continue;
1644 0 : }
1645 0 :
1646 0 : let timelines = t.timelines.lock().unwrap().clone();
1647 0 : for timeline in timelines.values() {
1648 0 : let Some(target_lsn) = target_lsns.get(&timeline.timeline_id) else {
1649 0 : continue;
1650 : };
1651 :
1652 0 : tracing::info!(
1653 0 : "Waiting for child shard {}/{} to reach target lsn {}...",
1654 0 : child_shard_id,
1655 0 : timeline.timeline_id,
1656 : target_lsn
1657 : );
1658 :
1659 0 : fail::fail_point!("shard-split-lsn-wait", |_| Err(anyhow::anyhow!(
1660 0 : "failpoint"
1661 0 : )));
1662 0 : if let Err(e) = timeline
1663 0 : .wait_lsn(
1664 0 : *target_lsn,
1665 0 : crate::tenant::timeline::WaitLsnWaiter::Tenant,
1666 0 : ctx,
1667 0 : )
1668 0 : .await
1669 : {
1670 : // Failure here might mean shutdown, in any case this part is an optimization
1671 : // and we shouldn't hold up the split operation.
1672 0 : tracing::warn!(
1673 0 : "Failed to wait for timeline {} to reach lsn {target_lsn}: {e}",
1674 0 : timeline.timeline_id
1675 : );
1676 : } else {
1677 0 : tracing::info!(
1678 0 : "Child shard {}/{} reached target lsn {}",
1679 0 : child_shard_id,
1680 0 : timeline.timeline_id,
1681 : target_lsn
1682 : );
1683 : }
1684 : }
1685 0 : }
1686 : }
1687 :
1688 : // Phase 5: Shut down the parent shard, and erase it from disk
1689 0 : let (_guard, progress) = completion::channel();
1690 0 : match parent.shutdown(progress, ShutdownMode::Hard).await {
1691 0 : Ok(()) => {}
1692 0 : Err(other) => {
1693 0 : other.wait().await;
1694 : }
1695 : }
1696 0 : let local_tenant_directory = self.conf.tenant_path(&tenant_shard_id);
1697 0 : let tmp_path = safe_rename_tenant_dir(&local_tenant_directory)
1698 0 : .await
1699 0 : .with_context(|| format!("local tenant directory {local_tenant_directory:?} rename"))?;
1700 0 : self.background_purges.spawn(tmp_path);
1701 0 :
1702 0 : fail::fail_point!("shard-split-pre-finish", |_| Err(anyhow::anyhow!(
1703 0 : "failpoint"
1704 0 : )));
1705 :
1706 0 : parent_slot_guard.drop_old_value()?;
1707 :
1708 : // Phase 6: Release the InProgress on the parent shard
1709 0 : drop(parent_slot_guard);
1710 0 :
1711 0 : Ok(child_shards)
1712 0 : }
1713 :
1714 : /// Part of [`Self::shard_split`]: hard link parent shard layers into child shards, as an optimization
1715 : /// to avoid the children downloading them again.
1716 : ///
1717 : /// For each resident layer in the parent shard, we will hard link it into all of the child shards.
1718 0 : async fn shard_split_hardlink(
1719 0 : &self,
1720 0 : parent_shard: &Tenant,
1721 0 : child_shards: Vec<TenantShardId>,
1722 0 : ) -> anyhow::Result<()> {
1723 0 : debug_assert_current_span_has_tenant_id();
1724 0 :
1725 0 : let parent_path = self.conf.tenant_path(parent_shard.get_tenant_shard_id());
1726 0 : let (parent_timelines, parent_layers) = {
1727 0 : let mut parent_layers = Vec::new();
1728 0 : let timelines = parent_shard.timelines.lock().unwrap().clone();
1729 0 : let parent_timelines = timelines.keys().cloned().collect::<Vec<_>>();
1730 0 : for timeline in timelines.values() {
1731 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Loading list of layers to hardlink");
1732 0 : let layers = timeline.layers.read().await;
1733 :
1734 0 : for layer in layers.likely_resident_layers() {
1735 0 : let relative_path = layer
1736 0 : .local_path()
1737 0 : .strip_prefix(&parent_path)
1738 0 : .context("Removing prefix from parent layer path")?;
1739 0 : parent_layers.push(relative_path.to_owned());
1740 : }
1741 : }
1742 0 : debug_assert!(
1743 0 : !parent_layers.is_empty(),
1744 0 : "shutdown cannot empty the layermap"
1745 : );
1746 0 : (parent_timelines, parent_layers)
1747 0 : };
1748 0 :
1749 0 : let mut child_prefixes = Vec::new();
1750 0 : let mut create_dirs = Vec::new();
1751 :
1752 0 : for child in child_shards {
1753 0 : let child_prefix = self.conf.tenant_path(&child);
1754 0 : create_dirs.push(child_prefix.clone());
1755 0 : create_dirs.extend(
1756 0 : parent_timelines
1757 0 : .iter()
1758 0 : .map(|t| self.conf.timeline_path(&child, t)),
1759 0 : );
1760 0 :
1761 0 : child_prefixes.push(child_prefix);
1762 0 : }
1763 :
1764 : // Since we will do a large number of small filesystem metadata operations, batch them into
1765 : // spawn_blocking calls rather than doing each one as a tokio::fs round-trip.
1766 0 : let span = tracing::Span::current();
1767 0 : let jh = tokio::task::spawn_blocking(move || -> anyhow::Result<usize> {
1768 0 : // Run this synchronous code in the same log context as the outer function that spawned it.
1769 0 : let _span = span.enter();
1770 0 :
1771 0 : tracing::info!("Creating {} directories", create_dirs.len());
1772 0 : for dir in &create_dirs {
1773 0 : if let Err(e) = std::fs::create_dir_all(dir) {
1774 : // Ignore AlreadyExists errors, drop out on all other errors
1775 0 : match e.kind() {
1776 0 : std::io::ErrorKind::AlreadyExists => {}
1777 : _ => {
1778 0 : return Err(anyhow::anyhow!(e).context(format!("Creating {dir}")));
1779 : }
1780 : }
1781 0 : }
1782 : }
1783 :
1784 0 : for child_prefix in child_prefixes {
1785 0 : tracing::info!(
1786 0 : "Hard-linking {} parent layers into child path {}",
1787 0 : parent_layers.len(),
1788 : child_prefix
1789 : );
1790 0 : for relative_layer in &parent_layers {
1791 0 : let parent_path = parent_path.join(relative_layer);
1792 0 : let child_path = child_prefix.join(relative_layer);
1793 0 : if let Err(e) = std::fs::hard_link(&parent_path, &child_path) {
1794 0 : match e.kind() {
1795 0 : std::io::ErrorKind::AlreadyExists => {}
1796 : std::io::ErrorKind::NotFound => {
1797 0 : tracing::info!(
1798 0 : "Layer {} not found during hard-linking, evicted during split?",
1799 : relative_layer
1800 : );
1801 : }
1802 : _ => {
1803 0 : return Err(anyhow::anyhow!(e).context(format!(
1804 0 : "Hard linking {relative_layer} into {child_prefix}"
1805 0 : )))
1806 : }
1807 : }
1808 0 : }
1809 : }
1810 : }
1811 :
1812 : // Durability is not required for correctness, but if we crashed during split and
1813 : // then came restarted with empty timeline dirs, it would be very inefficient to
1814 : // re-populate from remote storage.
1815 0 : tracing::info!("fsyncing {} directories", create_dirs.len());
1816 0 : for dir in create_dirs {
1817 0 : if let Err(e) = crashsafe::fsync(&dir) {
1818 : // Something removed a newly created timeline dir out from underneath us? Extremely
1819 : // unexpected, but not worth panic'ing over as this whole function is just an
1820 : // optimization.
1821 0 : tracing::warn!("Failed to fsync directory {dir}: {e}")
1822 0 : }
1823 : }
1824 :
1825 0 : Ok(parent_layers.len())
1826 0 : });
1827 0 :
1828 0 : match jh.await {
1829 0 : Ok(Ok(layer_count)) => {
1830 0 : tracing::info!(count = layer_count, "Hard linked layers into child shards");
1831 : }
1832 0 : Ok(Err(e)) => {
1833 0 : // This is an optimization, so we tolerate failure.
1834 0 : tracing::warn!("Error hard-linking layers, proceeding anyway: {e}")
1835 : }
1836 0 : Err(e) => {
1837 0 : // This is something totally unexpected like a panic, so bail out.
1838 0 : anyhow::bail!("Error joining hard linking task: {e}");
1839 : }
1840 : }
1841 :
1842 0 : Ok(())
1843 0 : }
1844 :
1845 : ///
1846 : /// Shut down all tenants. This runs as part of pageserver shutdown.
1847 : ///
1848 : /// NB: We leave the tenants in the map, so that they remain accessible through
1849 : /// the management API until we shut it down. If we removed the shut-down tenants
1850 : /// from the tenants map, the management API would return 404 for these tenants,
1851 : /// because TenantsMap::get() now returns `None`.
1852 : /// That could be easily misinterpreted by control plane, the consumer of the
1853 : /// management API. For example, it could attach the tenant on a different pageserver.
1854 : /// We would then be in split-brain once this pageserver restarts.
1855 0 : #[instrument(skip_all)]
1856 : pub(crate) async fn shutdown(&self) {
1857 : self.cancel.cancel();
1858 :
1859 : shutdown_all_tenants0(self.tenants).await
1860 : }
1861 :
1862 0 : pub(crate) async fn detach_tenant(
1863 0 : &self,
1864 0 : conf: &'static PageServerConf,
1865 0 : tenant_shard_id: TenantShardId,
1866 0 : deletion_queue_client: &DeletionQueueClient,
1867 0 : ) -> Result<(), TenantStateError> {
1868 0 : let tmp_path = self
1869 0 : .detach_tenant0(conf, tenant_shard_id, deletion_queue_client)
1870 0 : .await?;
1871 0 : self.background_purges.spawn(tmp_path);
1872 0 :
1873 0 : Ok(())
1874 0 : }
1875 :
1876 0 : async fn detach_tenant0(
1877 0 : &self,
1878 0 : conf: &'static PageServerConf,
1879 0 : tenant_shard_id: TenantShardId,
1880 0 : deletion_queue_client: &DeletionQueueClient,
1881 0 : ) -> Result<Utf8PathBuf, TenantStateError> {
1882 0 : let tenant_dir_rename_operation = |tenant_id_to_clean: TenantShardId| async move {
1883 0 : let local_tenant_directory = conf.tenant_path(&tenant_id_to_clean);
1884 0 : safe_rename_tenant_dir(&local_tenant_directory)
1885 0 : .await
1886 0 : .with_context(|| {
1887 0 : format!("local tenant directory {local_tenant_directory:?} rename")
1888 0 : })
1889 0 : };
1890 :
1891 0 : let removal_result = remove_tenant_from_memory(
1892 0 : self.tenants,
1893 0 : tenant_shard_id,
1894 0 : tenant_dir_rename_operation(tenant_shard_id),
1895 0 : )
1896 0 : .await;
1897 :
1898 : // Flush pending deletions, so that they have a good chance of passing validation
1899 : // before this tenant is potentially re-attached elsewhere.
1900 0 : deletion_queue_client.flush_advisory();
1901 0 :
1902 0 : removal_result
1903 0 : }
1904 :
1905 0 : pub(crate) fn list_tenants(
1906 0 : &self,
1907 0 : ) -> Result<Vec<(TenantShardId, TenantState, Generation)>, TenantMapListError> {
1908 0 : let tenants = self.tenants.read().unwrap();
1909 0 : let m = match &*tenants {
1910 0 : TenantsMap::Initializing => return Err(TenantMapListError::Initializing),
1911 0 : TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => m,
1912 0 : };
1913 0 : Ok(m.iter()
1914 0 : .filter_map(|(id, tenant)| match tenant {
1915 0 : TenantSlot::Attached(tenant) => {
1916 0 : Some((*id, tenant.current_state(), tenant.generation()))
1917 : }
1918 0 : TenantSlot::Secondary(_) => None,
1919 0 : TenantSlot::InProgress(_) => None,
1920 0 : })
1921 0 : .collect())
1922 0 : }
1923 :
1924 : /// Completes an earlier prepared timeline detach ancestor.
1925 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
1926 0 : &self,
1927 0 : tenant_shard_id: TenantShardId,
1928 0 : timeline_id: TimelineId,
1929 0 : prepared: PreparedTimelineDetach,
1930 0 : mut attempt: detach_ancestor::Attempt,
1931 0 : ctx: &RequestContext,
1932 0 : ) -> Result<HashSet<TimelineId>, detach_ancestor::Error> {
1933 : use detach_ancestor::Error;
1934 :
1935 0 : let slot_guard =
1936 0 : tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::MustExist).map_err(
1937 0 : |e| {
1938 : use TenantSlotError::*;
1939 :
1940 0 : match e {
1941 0 : MapState(TenantMapError::ShuttingDown) => Error::ShuttingDown,
1942 0 : NotFound(_) | InProgress | MapState(_) => Error::DetachReparent(e.into()),
1943 : }
1944 0 : },
1945 0 : )?;
1946 :
1947 0 : let tenant = {
1948 0 : let old_slot = slot_guard
1949 0 : .get_old_value()
1950 0 : .as_ref()
1951 0 : .expect("requested MustExist");
1952 :
1953 0 : let Some(tenant) = old_slot.get_attached() else {
1954 0 : return Err(Error::DetachReparent(anyhow::anyhow!(
1955 0 : "Tenant is not in attached state"
1956 0 : )));
1957 : };
1958 :
1959 0 : if !tenant.is_active() {
1960 0 : return Err(Error::DetachReparent(anyhow::anyhow!(
1961 0 : "Tenant is not active"
1962 0 : )));
1963 0 : }
1964 0 :
1965 0 : tenant.clone()
1966 : };
1967 :
1968 0 : let timeline = tenant
1969 0 : .get_timeline(timeline_id, true)
1970 0 : .map_err(Error::NotFound)?;
1971 :
1972 0 : let resp = timeline
1973 0 : .detach_from_ancestor_and_reparent(&tenant, prepared, ctx)
1974 0 : .await?;
1975 :
1976 0 : let mut slot_guard = slot_guard;
1977 :
1978 0 : let tenant = if resp.reset_tenant_required() {
1979 0 : attempt.before_reset_tenant();
1980 0 :
1981 0 : let (_guard, progress) = utils::completion::channel();
1982 0 : match tenant.shutdown(progress, ShutdownMode::Hard).await {
1983 0 : Ok(()) => {
1984 0 : slot_guard.drop_old_value().expect("it was just shutdown");
1985 0 : }
1986 0 : Err(_barrier) => {
1987 0 : slot_guard.revert();
1988 0 : // this really should not happen, at all, unless a shutdown without acquiring
1989 0 : // tenant slot was already going? regardless, on restart the attempt tracking
1990 0 : // will reset to retryable.
1991 0 : return Err(Error::ShuttingDown);
1992 : }
1993 : }
1994 :
1995 0 : let tenant_path = self.conf.tenant_path(&tenant_shard_id);
1996 0 : let config = Tenant::load_tenant_config(self.conf, &tenant_shard_id)
1997 0 : .map_err(|e| Error::DetachReparent(e.into()))?;
1998 :
1999 0 : let shard_identity = config.shard;
2000 0 : let tenant = tenant_spawn(
2001 0 : self.conf,
2002 0 : tenant_shard_id,
2003 0 : &tenant_path,
2004 0 : self.resources.clone(),
2005 0 : AttachedTenantConf::try_from(config).map_err(Error::DetachReparent)?,
2006 0 : shard_identity,
2007 0 : None,
2008 0 : SpawnMode::Eager,
2009 0 : ctx,
2010 0 : )
2011 0 : .map_err(|_| Error::ShuttingDown)?;
2012 :
2013 : {
2014 0 : let mut g = tenant.ongoing_timeline_detach.lock().unwrap();
2015 0 : assert!(
2016 0 : g.is_none(),
2017 0 : "there cannot be any new timeline detach ancestor on newly created tenant"
2018 : );
2019 0 : *g = Some((attempt.timeline_id, attempt.new_barrier()));
2020 0 : }
2021 0 :
2022 0 : // if we bail out here, we will not allow a new attempt, which should be fine.
2023 0 : // pageserver should be shutting down regardless? tenant_reset would help, unless it
2024 0 : // runs into the same problem.
2025 0 : slot_guard
2026 0 : .upsert(TenantSlot::Attached(tenant.clone()))
2027 0 : .map_err(|e| match e {
2028 0 : TenantSlotUpsertError::ShuttingDown(_) => Error::ShuttingDown,
2029 0 : other => Error::DetachReparent(other.into()),
2030 0 : })?;
2031 0 : tenant
2032 : } else {
2033 0 : tracing::info!("skipping tenant_reset as no changes made required it");
2034 0 : tenant
2035 : };
2036 :
2037 0 : if let Some(reparented) = resp.completed() {
2038 : // finally ask the restarted tenant to complete the detach
2039 : //
2040 : // rationale for 9999s: we don't really have a timetable here; if retried, the caller
2041 : // will get an 503.
2042 0 : tenant
2043 0 : .wait_to_become_active(std::time::Duration::from_secs(9999))
2044 0 : .await
2045 0 : .map_err(|e| {
2046 : use pageserver_api::models::TenantState;
2047 : use GetActiveTenantError::{Cancelled, WillNotBecomeActive};
2048 0 : match e {
2049 : Cancelled | WillNotBecomeActive(TenantState::Stopping { .. }) => {
2050 0 : Error::ShuttingDown
2051 : }
2052 0 : other => Error::Complete(other.into()),
2053 : }
2054 0 : })?;
2055 :
2056 : utils::pausable_failpoint!(
2057 : "timeline-detach-ancestor::after_activating_before_finding-pausable"
2058 : );
2059 :
2060 0 : let timeline = tenant
2061 0 : .get_timeline(attempt.timeline_id, true)
2062 0 : .map_err(Error::NotFound)?;
2063 :
2064 0 : timeline
2065 0 : .complete_detaching_timeline_ancestor(&tenant, attempt, ctx)
2066 0 : .await
2067 0 : .map(|()| reparented)
2068 : } else {
2069 : // at least the latest versions have now been downloaded and refreshed; be ready to
2070 : // retry another time.
2071 0 : Err(Error::FailedToReparentAll)
2072 : }
2073 0 : }
2074 :
2075 : /// A page service client sends a TenantId, and to look up the correct Tenant we must
2076 : /// resolve this to a fully qualified TenantShardId.
2077 : ///
2078 : /// During shard splits: we shall see parent shards in InProgress state and skip them, and
2079 : /// instead match on child shards which should appear in Attached state. Very early in a shard
2080 : /// split, or in other cases where a shard is InProgress, we will return our own InProgress result
2081 : /// to instruct the caller to wait for that to finish before querying again.
2082 0 : pub(crate) fn resolve_attached_shard(
2083 0 : &self,
2084 0 : tenant_id: &TenantId,
2085 0 : selector: ShardSelector,
2086 0 : ) -> ShardResolveResult {
2087 0 : let tenants = self.tenants.read().unwrap();
2088 0 : let mut want_shard = None;
2089 0 : let mut any_in_progress = None;
2090 0 :
2091 0 : match &*tenants {
2092 0 : TenantsMap::Initializing => ShardResolveResult::NotFound,
2093 0 : TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => {
2094 0 : for slot in m.range(TenantShardId::tenant_range(*tenant_id)) {
2095 : // Ignore all slots that don't contain an attached tenant
2096 0 : let tenant = match &slot.1 {
2097 0 : TenantSlot::Attached(t) => t,
2098 0 : TenantSlot::InProgress(barrier) => {
2099 0 : // We might still find a usable shard, but in case we don't, remember that
2100 0 : // we saw at least one InProgress slot, so that we can distinguish this case
2101 0 : // from a simple NotFound in our return value.
2102 0 : any_in_progress = Some(barrier.clone());
2103 0 : continue;
2104 : }
2105 0 : _ => continue,
2106 : };
2107 :
2108 0 : match selector {
2109 0 : ShardSelector::Zero if slot.0.shard_number == ShardNumber(0) => {
2110 0 : return ShardResolveResult::Found(tenant.clone())
2111 : }
2112 0 : ShardSelector::Page(key) => {
2113 0 : // First slot we see for this tenant, calculate the expected shard number
2114 0 : // for the key: we will use this for checking if this and subsequent
2115 0 : // slots contain the key, rather than recalculating the hash each time.
2116 0 : if want_shard.is_none() {
2117 0 : want_shard = Some(tenant.shard_identity.get_shard_number(&key));
2118 0 : }
2119 :
2120 0 : if Some(tenant.shard_identity.number) == want_shard {
2121 0 : return ShardResolveResult::Found(tenant.clone());
2122 0 : }
2123 : }
2124 0 : ShardSelector::Known(shard)
2125 0 : if tenant.shard_identity.shard_index() == shard =>
2126 0 : {
2127 0 : return ShardResolveResult::Found(tenant.clone());
2128 : }
2129 0 : _ => continue,
2130 : }
2131 : }
2132 :
2133 : // Fall through: we didn't find a slot that was in Attached state & matched our selector. If
2134 : // we found one or more InProgress slot, indicate to caller that they should retry later. Otherwise
2135 : // this requested shard simply isn't found.
2136 0 : if let Some(barrier) = any_in_progress {
2137 0 : ShardResolveResult::InProgress(barrier)
2138 : } else {
2139 0 : ShardResolveResult::NotFound
2140 : }
2141 : }
2142 : }
2143 0 : }
2144 :
2145 : /// Calculate the tenant shards' contributions to this pageserver's utilization metrics. The
2146 : /// returned values are:
2147 : /// - the number of bytes of local disk space this pageserver's shards are requesting, i.e.
2148 : /// how much space they would use if not impacted by disk usage eviction.
2149 : /// - the number of tenant shards currently on this pageserver, including attached
2150 : /// and secondary.
2151 : ///
2152 : /// This function is quite expensive: callers are expected to cache the result and
2153 : /// limit how often they call it.
2154 0 : pub(crate) fn calculate_utilization(&self) -> Result<(u64, u32), TenantMapListError> {
2155 0 : let tenants = self.tenants.read().unwrap();
2156 0 : let m = match &*tenants {
2157 0 : TenantsMap::Initializing => return Err(TenantMapListError::Initializing),
2158 0 : TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => m,
2159 0 : };
2160 0 : let shard_count = m.len();
2161 0 : let mut wanted_bytes = 0;
2162 :
2163 0 : for tenant_slot in m.values() {
2164 0 : match tenant_slot {
2165 0 : TenantSlot::InProgress(_barrier) => {
2166 0 : // While a slot is being changed, we can't know how much storage it wants. This
2167 0 : // means this function's output can fluctuate if a lot of changes are going on
2168 0 : // (such as transitions from secondary to attached).
2169 0 : //
2170 0 : // We could wait for the barrier and retry, but it's important that the utilization
2171 0 : // API is responsive, and the data quality impact is not very significant.
2172 0 : continue;
2173 : }
2174 0 : TenantSlot::Attached(tenant) => {
2175 0 : wanted_bytes += tenant.local_storage_wanted();
2176 0 : }
2177 0 : TenantSlot::Secondary(secondary) => {
2178 0 : let progress = secondary.progress.lock().unwrap();
2179 0 : wanted_bytes += if progress.heatmap_mtime.is_some() {
2180 : // If we have heatmap info, then we will 'want' the sum
2181 : // of the size of layers in the heatmap: this is how much space
2182 : // we would use if not doing any eviction.
2183 0 : progress.bytes_total
2184 : } else {
2185 : // In the absence of heatmap info, assume that the secondary location simply
2186 : // needs as much space as it is currently using.
2187 0 : secondary.resident_size_metric.get()
2188 : }
2189 : }
2190 : }
2191 : }
2192 :
2193 0 : Ok((wanted_bytes, shard_count as u32))
2194 0 : }
2195 : }
2196 :
2197 0 : #[derive(Debug, thiserror::Error)]
2198 : pub(crate) enum GetTenantError {
2199 : /// NotFound is a TenantId rather than TenantShardId, because this error type is used from
2200 : /// getters that use a TenantId and a ShardSelector, not just getters that target a specific shard.
2201 : #[error("Tenant {0} not found")]
2202 : NotFound(TenantId),
2203 :
2204 : #[error("Tenant {0} is not active")]
2205 : NotActive(TenantShardId),
2206 :
2207 : // Initializing or shutting down: cannot authoritatively say whether we have this tenant
2208 : #[error("Tenant map is not available: {0}")]
2209 : MapState(#[from] TenantMapError),
2210 : }
2211 :
2212 0 : #[derive(thiserror::Error, Debug)]
2213 : pub(crate) enum GetActiveTenantError {
2214 : /// We may time out either while TenantSlot is InProgress, or while the Tenant
2215 : /// is in a non-Active state
2216 : #[error(
2217 : "Timed out waiting {wait_time:?} for tenant active state. Latest state: {latest_state:?}"
2218 : )]
2219 : WaitForActiveTimeout {
2220 : latest_state: Option<TenantState>,
2221 : wait_time: Duration,
2222 : },
2223 :
2224 : /// The TenantSlot is absent, or in secondary mode
2225 : #[error(transparent)]
2226 : NotFound(#[from] GetTenantError),
2227 :
2228 : /// Cancellation token fired while we were waiting
2229 : #[error("cancelled")]
2230 : Cancelled,
2231 :
2232 : /// Tenant exists, but is in a state that cannot become active (e.g. Stopping, Broken)
2233 : #[error("will not become active. Current state: {0}")]
2234 : WillNotBecomeActive(TenantState),
2235 :
2236 : /// Broken is logically a subset of WillNotBecomeActive, but a distinct error is useful as
2237 : /// WillNotBecomeActive is a permitted error under some circumstances, whereas broken should
2238 : /// never happen.
2239 : #[error("Tenant is broken: {0}")]
2240 : Broken(String),
2241 :
2242 : #[error("reconnect to switch tenant id")]
2243 : SwitchedTenant,
2244 : }
2245 :
2246 0 : #[derive(Debug, thiserror::Error)]
2247 : pub(crate) enum DeleteTimelineError {
2248 : #[error("Tenant {0}")]
2249 : Tenant(#[from] GetTenantError),
2250 :
2251 : #[error("Timeline {0}")]
2252 : Timeline(#[from] crate::tenant::DeleteTimelineError),
2253 : }
2254 :
2255 0 : #[derive(Debug, thiserror::Error)]
2256 : pub(crate) enum TenantStateError {
2257 : #[error("Tenant {0} is stopping")]
2258 : IsStopping(TenantShardId),
2259 : #[error(transparent)]
2260 : SlotError(#[from] TenantSlotError),
2261 : #[error(transparent)]
2262 : SlotUpsertError(#[from] TenantSlotUpsertError),
2263 : #[error(transparent)]
2264 : Other(#[from] anyhow::Error),
2265 : }
2266 :
2267 0 : #[derive(Debug, thiserror::Error)]
2268 : pub(crate) enum TenantMapListError {
2269 : #[error("tenant map is still initiailizing")]
2270 : Initializing,
2271 : }
2272 :
2273 0 : #[derive(Debug, thiserror::Error)]
2274 : pub(crate) enum TenantMapInsertError {
2275 : #[error(transparent)]
2276 : SlotError(#[from] TenantSlotError),
2277 : #[error(transparent)]
2278 : SlotUpsertError(#[from] TenantSlotUpsertError),
2279 : #[error(transparent)]
2280 : Other(#[from] anyhow::Error),
2281 : }
2282 :
2283 : /// Superset of TenantMapError: issues that can occur when acquiring a slot
2284 : /// for a particular tenant ID.
2285 0 : #[derive(Debug, thiserror::Error)]
2286 : pub(crate) enum TenantSlotError {
2287 : /// When acquiring a slot with the expectation that the tenant already exists.
2288 : #[error("Tenant {0} not found")]
2289 : NotFound(TenantShardId),
2290 :
2291 : // Tried to read a slot that is currently being mutated by another administrative
2292 : // operation.
2293 : #[error("tenant has a state change in progress, try again later")]
2294 : InProgress,
2295 :
2296 : #[error(transparent)]
2297 : MapState(#[from] TenantMapError),
2298 : }
2299 :
2300 : /// Superset of TenantMapError: issues that can occur when using a SlotGuard
2301 : /// to insert a new value.
2302 0 : #[derive(thiserror::Error)]
2303 : pub(crate) enum TenantSlotUpsertError {
2304 : /// An error where the slot is in an unexpected state, indicating a code bug
2305 : #[error("Internal error updating Tenant")]
2306 : InternalError(Cow<'static, str>),
2307 :
2308 : #[error(transparent)]
2309 : MapState(TenantMapError),
2310 :
2311 : // If we encounter TenantManager shutdown during upsert, we must carry the Completion
2312 : // from the SlotGuard, so that the caller can hold it while they clean up: otherwise
2313 : // TenantManager shutdown might race ahead before we're done cleaning up any Tenant that
2314 : // was protected by the SlotGuard.
2315 : #[error("Shutting down")]
2316 : ShuttingDown((TenantSlot, utils::completion::Completion)),
2317 : }
2318 :
2319 : impl std::fmt::Debug for TenantSlotUpsertError {
2320 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
2321 0 : match self {
2322 0 : Self::InternalError(reason) => write!(f, "Internal Error {reason}"),
2323 0 : Self::MapState(map_error) => write!(f, "Tenant map state: {map_error:?}"),
2324 0 : Self::ShuttingDown(_completion) => write!(f, "Tenant map shutting down"),
2325 : }
2326 0 : }
2327 : }
2328 :
2329 0 : #[derive(Debug, thiserror::Error)]
2330 : enum TenantSlotDropError {
2331 : /// It is only legal to drop a TenantSlot if its contents are fully shut down
2332 : #[error("Tenant was not shut down")]
2333 : NotShutdown,
2334 : }
2335 :
2336 : /// Errors that can happen any time we are walking the tenant map to try and acquire
2337 : /// the TenantSlot for a particular tenant.
2338 0 : #[derive(Debug, thiserror::Error)]
2339 : pub enum TenantMapError {
2340 : // Tried to read while initializing
2341 : #[error("tenant map is still initializing")]
2342 : StillInitializing,
2343 :
2344 : // Tried to read while shutting down
2345 : #[error("tenant map is shutting down")]
2346 : ShuttingDown,
2347 : }
2348 :
2349 : /// Guards a particular tenant_id's content in the TenantsMap. While this
2350 : /// structure exists, the TenantsMap will contain a [`TenantSlot::InProgress`]
2351 : /// for this tenant, which acts as a marker for any operations targeting
2352 : /// this tenant to retry later, or wait for the InProgress state to end.
2353 : ///
2354 : /// This structure enforces the important invariant that we do not have overlapping
2355 : /// tasks that will try use local storage for a the same tenant ID: we enforce that
2356 : /// the previous contents of a slot have been shut down before the slot can be
2357 : /// left empty or used for something else
2358 : ///
2359 : /// Holders of a SlotGuard should explicitly dispose of it, using either `upsert`
2360 : /// to provide a new value, or `revert` to put the slot back into its initial
2361 : /// state. If the SlotGuard is dropped without calling either of these, then
2362 : /// we will leave the slot empty if our `old_value` is already shut down, else
2363 : /// we will replace the slot with `old_value` (equivalent to doing a revert).
2364 : ///
2365 : /// The `old_value` may be dropped before the SlotGuard is dropped, by calling
2366 : /// `drop_old_value`. It is an error to call this without shutting down
2367 : /// the conents of `old_value`.
2368 : pub struct SlotGuard {
2369 : tenant_shard_id: TenantShardId,
2370 : old_value: Option<TenantSlot>,
2371 : upserted: bool,
2372 :
2373 : /// [`TenantSlot::InProgress`] carries the corresponding Barrier: it will
2374 : /// release any waiters as soon as this SlotGuard is dropped.
2375 : completion: utils::completion::Completion,
2376 : }
2377 :
2378 : impl SlotGuard {
2379 2 : fn new(
2380 2 : tenant_shard_id: TenantShardId,
2381 2 : old_value: Option<TenantSlot>,
2382 2 : completion: utils::completion::Completion,
2383 2 : ) -> Self {
2384 2 : Self {
2385 2 : tenant_shard_id,
2386 2 : old_value,
2387 2 : upserted: false,
2388 2 : completion,
2389 2 : }
2390 2 : }
2391 :
2392 : /// Get any value that was present in the slot before we acquired ownership
2393 : /// of it: in state transitions, this will be the old state.
2394 : ///
2395 : // FIXME: get_ prefix
2396 : // FIXME: this should be .as_ref() -- unsure why no clippy
2397 2 : fn get_old_value(&self) -> &Option<TenantSlot> {
2398 2 : &self.old_value
2399 2 : }
2400 :
2401 : /// Emplace a new value in the slot. This consumes the guard, and after
2402 : /// returning, the slot is no longer protected from concurrent changes.
2403 0 : fn upsert(mut self, new_value: TenantSlot) -> Result<(), TenantSlotUpsertError> {
2404 0 : if !self.old_value_is_shutdown() {
2405 : // This is a bug: callers should never try to drop an old value without
2406 : // shutting it down
2407 0 : return Err(TenantSlotUpsertError::InternalError(
2408 0 : "Old TenantSlot value not shut down".into(),
2409 0 : ));
2410 0 : }
2411 :
2412 0 : let replaced = {
2413 0 : let mut locked = TENANTS.write().unwrap();
2414 0 :
2415 0 : if let TenantSlot::InProgress(_) = new_value {
2416 : // It is never expected to try and upsert InProgress via this path: it should
2417 : // only be written via the tenant_map_acquire_slot path. If we hit this it's a bug.
2418 0 : return Err(TenantSlotUpsertError::InternalError(
2419 0 : "Attempt to upsert an InProgress state".into(),
2420 0 : ));
2421 0 : }
2422 :
2423 0 : let m = match &mut *locked {
2424 : TenantsMap::Initializing => {
2425 0 : return Err(TenantSlotUpsertError::MapState(
2426 0 : TenantMapError::StillInitializing,
2427 0 : ))
2428 : }
2429 : TenantsMap::ShuttingDown(_) => {
2430 0 : return Err(TenantSlotUpsertError::ShuttingDown((
2431 0 : new_value,
2432 0 : self.completion.clone(),
2433 0 : )));
2434 : }
2435 0 : TenantsMap::Open(m) => m,
2436 0 : };
2437 0 :
2438 0 : METRICS.slot_inserted(&new_value);
2439 0 :
2440 0 : let replaced = m.insert(self.tenant_shard_id, new_value);
2441 0 : self.upserted = true;
2442 0 : if let Some(replaced) = replaced.as_ref() {
2443 0 : METRICS.slot_removed(replaced);
2444 0 : }
2445 :
2446 0 : replaced
2447 : };
2448 :
2449 : // Sanity check: on an upsert we should always be replacing an InProgress marker
2450 0 : match replaced {
2451 : Some(TenantSlot::InProgress(_)) => {
2452 : // Expected case: we find our InProgress in the map: nothing should have
2453 : // replaced it because the code that acquires slots will not grant another
2454 : // one for the same TenantId.
2455 0 : Ok(())
2456 : }
2457 : None => {
2458 0 : METRICS.unexpected_errors.inc();
2459 0 : error!(
2460 : tenant_shard_id = %self.tenant_shard_id,
2461 0 : "Missing InProgress marker during tenant upsert, this is a bug."
2462 : );
2463 0 : Err(TenantSlotUpsertError::InternalError(
2464 0 : "Missing InProgress marker during tenant upsert".into(),
2465 0 : ))
2466 : }
2467 0 : Some(slot) => {
2468 0 : METRICS.unexpected_errors.inc();
2469 0 : error!(tenant_shard_id=%self.tenant_shard_id, "Unexpected contents of TenantSlot during upsert, this is a bug. Contents: {:?}", slot);
2470 0 : Err(TenantSlotUpsertError::InternalError(
2471 0 : "Unexpected contents of TenantSlot".into(),
2472 0 : ))
2473 : }
2474 : }
2475 0 : }
2476 :
2477 : /// Replace the InProgress slot with whatever was in the guard when we started
2478 0 : fn revert(mut self) {
2479 0 : if let Some(value) = self.old_value.take() {
2480 0 : match self.upsert(value) {
2481 0 : Err(TenantSlotUpsertError::InternalError(_)) => {
2482 0 : // We already logged the error, nothing else we can do.
2483 0 : }
2484 : Err(
2485 : TenantSlotUpsertError::MapState(_) | TenantSlotUpsertError::ShuttingDown(_),
2486 0 : ) => {
2487 0 : // If the map is shutting down, we need not replace anything
2488 0 : }
2489 0 : Ok(()) => {}
2490 : }
2491 0 : }
2492 0 : }
2493 :
2494 : /// We may never drop our old value until it is cleanly shut down: otherwise we might leave
2495 : /// rogue background tasks that would write to the local tenant directory that this guard
2496 : /// is responsible for protecting
2497 2 : fn old_value_is_shutdown(&self) -> bool {
2498 2 : match self.old_value.as_ref() {
2499 2 : Some(TenantSlot::Attached(tenant)) => tenant.gate.close_complete(),
2500 0 : Some(TenantSlot::Secondary(secondary_tenant)) => secondary_tenant.gate.close_complete(),
2501 : Some(TenantSlot::InProgress(_)) => {
2502 : // A SlotGuard cannot be constructed for a slot that was already InProgress
2503 0 : unreachable!()
2504 : }
2505 0 : None => true,
2506 : }
2507 2 : }
2508 :
2509 : /// The guard holder is done with the old value of the slot: they are obliged to already
2510 : /// shut it down before we reach this point.
2511 2 : fn drop_old_value(&mut self) -> Result<(), TenantSlotDropError> {
2512 2 : if !self.old_value_is_shutdown() {
2513 0 : Err(TenantSlotDropError::NotShutdown)
2514 : } else {
2515 2 : self.old_value.take();
2516 2 : Ok(())
2517 : }
2518 2 : }
2519 : }
2520 :
2521 : impl Drop for SlotGuard {
2522 2 : fn drop(&mut self) {
2523 2 : if self.upserted {
2524 0 : return;
2525 2 : }
2526 2 : // Our old value is already shutdown, or it never existed: it is safe
2527 2 : // for us to fully release the TenantSlot back into an empty state
2528 2 :
2529 2 : let mut locked = TENANTS.write().unwrap();
2530 :
2531 2 : let m = match &mut *locked {
2532 : TenantsMap::Initializing => {
2533 : // There is no map, this should never happen.
2534 2 : return;
2535 : }
2536 : TenantsMap::ShuttingDown(_) => {
2537 : // When we transition to shutdown, InProgress elements are removed
2538 : // from the map, so we do not need to clean up our Inprogress marker.
2539 : // See [`shutdown_all_tenants0`]
2540 0 : return;
2541 : }
2542 0 : TenantsMap::Open(m) => m,
2543 0 : };
2544 0 :
2545 0 : use std::collections::btree_map::Entry;
2546 0 : match m.entry(self.tenant_shard_id) {
2547 0 : Entry::Occupied(mut entry) => {
2548 0 : if !matches!(entry.get(), TenantSlot::InProgress(_)) {
2549 0 : METRICS.unexpected_errors.inc();
2550 0 : error!(tenant_shard_id=%self.tenant_shard_id, "Unexpected contents of TenantSlot during drop, this is a bug. Contents: {:?}", entry.get());
2551 0 : }
2552 :
2553 0 : if self.old_value_is_shutdown() {
2554 0 : METRICS.slot_removed(entry.get());
2555 0 : entry.remove();
2556 0 : } else {
2557 0 : let inserting = self.old_value.take().unwrap();
2558 0 : METRICS.slot_inserted(&inserting);
2559 0 : let replaced = entry.insert(inserting);
2560 0 : METRICS.slot_removed(&replaced);
2561 0 : }
2562 : }
2563 : Entry::Vacant(_) => {
2564 0 : METRICS.unexpected_errors.inc();
2565 0 : error!(
2566 : tenant_shard_id = %self.tenant_shard_id,
2567 0 : "Missing InProgress marker during SlotGuard drop, this is a bug."
2568 : );
2569 : }
2570 : }
2571 2 : }
2572 : }
2573 :
2574 : enum TenantSlotPeekMode {
2575 : /// In Read mode, peek will be permitted to see the slots even if the pageserver is shutting down
2576 : Read,
2577 : /// In Write mode, trying to peek at a slot while the pageserver is shutting down is an error
2578 : Write,
2579 : }
2580 :
2581 0 : fn tenant_map_peek_slot<'a>(
2582 0 : tenants: &'a std::sync::RwLockReadGuard<'a, TenantsMap>,
2583 0 : tenant_shard_id: &TenantShardId,
2584 0 : mode: TenantSlotPeekMode,
2585 0 : ) -> Result<Option<&'a TenantSlot>, TenantMapError> {
2586 0 : match tenants.deref() {
2587 0 : TenantsMap::Initializing => Err(TenantMapError::StillInitializing),
2588 0 : TenantsMap::ShuttingDown(m) => match mode {
2589 : TenantSlotPeekMode::Read => Ok(Some(
2590 : // When reading in ShuttingDown state, we must translate None results
2591 : // into a ShuttingDown error, because absence of a tenant shard ID in the map
2592 : // isn't a reliable indicator of the tenant being gone: it might have been
2593 : // InProgress when shutdown started, and cleaned up from that state such
2594 : // that it's now no longer in the map. Callers will have to wait until
2595 : // we next start up to get a proper answer. This avoids incorrect 404 API responses.
2596 0 : m.get(tenant_shard_id).ok_or(TenantMapError::ShuttingDown)?,
2597 : )),
2598 0 : TenantSlotPeekMode::Write => Err(TenantMapError::ShuttingDown),
2599 : },
2600 0 : TenantsMap::Open(m) => Ok(m.get(tenant_shard_id)),
2601 : }
2602 0 : }
2603 :
2604 : enum TenantSlotAcquireMode {
2605 : /// Acquire the slot irrespective of current state, or whether it already exists
2606 : Any,
2607 : /// Return an error if trying to acquire a slot and it doesn't already exist
2608 : MustExist,
2609 : }
2610 :
2611 0 : fn tenant_map_acquire_slot(
2612 0 : tenant_shard_id: &TenantShardId,
2613 0 : mode: TenantSlotAcquireMode,
2614 0 : ) -> Result<SlotGuard, TenantSlotError> {
2615 0 : tenant_map_acquire_slot_impl(tenant_shard_id, &TENANTS, mode)
2616 0 : }
2617 :
2618 2 : fn tenant_map_acquire_slot_impl(
2619 2 : tenant_shard_id: &TenantShardId,
2620 2 : tenants: &std::sync::RwLock<TenantsMap>,
2621 2 : mode: TenantSlotAcquireMode,
2622 2 : ) -> Result<SlotGuard, TenantSlotError> {
2623 2 : use TenantSlotAcquireMode::*;
2624 2 : METRICS.tenant_slot_writes.inc();
2625 2 :
2626 2 : let mut locked = tenants.write().unwrap();
2627 2 : let span = tracing::info_span!("acquire_slot", tenant_id=%tenant_shard_id.tenant_id, shard_id = %tenant_shard_id.shard_slug());
2628 2 : let _guard = span.enter();
2629 :
2630 2 : let m = match &mut *locked {
2631 0 : TenantsMap::Initializing => return Err(TenantMapError::StillInitializing.into()),
2632 0 : TenantsMap::ShuttingDown(_) => return Err(TenantMapError::ShuttingDown.into()),
2633 2 : TenantsMap::Open(m) => m,
2634 2 : };
2635 2 :
2636 2 : use std::collections::btree_map::Entry;
2637 2 :
2638 2 : let entry = m.entry(*tenant_shard_id);
2639 2 :
2640 2 : match entry {
2641 0 : Entry::Vacant(v) => match mode {
2642 : MustExist => {
2643 0 : tracing::debug!("Vacant && MustExist: return NotFound");
2644 0 : Err(TenantSlotError::NotFound(*tenant_shard_id))
2645 : }
2646 : _ => {
2647 0 : let (completion, barrier) = utils::completion::channel();
2648 0 : let inserting = TenantSlot::InProgress(barrier);
2649 0 : METRICS.slot_inserted(&inserting);
2650 0 : v.insert(inserting);
2651 0 : tracing::debug!("Vacant, inserted InProgress");
2652 0 : Ok(SlotGuard::new(*tenant_shard_id, None, completion))
2653 : }
2654 : },
2655 2 : Entry::Occupied(mut o) => {
2656 2 : // Apply mode-driven checks
2657 2 : match (o.get(), mode) {
2658 : (TenantSlot::InProgress(_), _) => {
2659 0 : tracing::debug!("Occupied, failing for InProgress");
2660 0 : Err(TenantSlotError::InProgress)
2661 : }
2662 : _ => {
2663 : // Happy case: the slot was not in any state that violated our mode
2664 2 : let (completion, barrier) = utils::completion::channel();
2665 2 : let in_progress = TenantSlot::InProgress(barrier);
2666 2 : METRICS.slot_inserted(&in_progress);
2667 2 : let old_value = o.insert(in_progress);
2668 2 : METRICS.slot_removed(&old_value);
2669 2 : tracing::debug!("Occupied, replaced with InProgress");
2670 2 : Ok(SlotGuard::new(
2671 2 : *tenant_shard_id,
2672 2 : Some(old_value),
2673 2 : completion,
2674 2 : ))
2675 : }
2676 : }
2677 : }
2678 : }
2679 2 : }
2680 :
2681 : /// Stops and removes the tenant from memory, if it's not [`TenantState::Stopping`] already, bails otherwise.
2682 : /// Allows to remove other tenant resources manually, via `tenant_cleanup`.
2683 : /// If the cleanup fails, tenant will stay in memory in [`TenantState::Broken`] state, and another removal
2684 : /// operation would be needed to remove it.
2685 2 : async fn remove_tenant_from_memory<V, F>(
2686 2 : tenants: &std::sync::RwLock<TenantsMap>,
2687 2 : tenant_shard_id: TenantShardId,
2688 2 : tenant_cleanup: F,
2689 2 : ) -> Result<V, TenantStateError>
2690 2 : where
2691 2 : F: std::future::Future<Output = anyhow::Result<V>>,
2692 2 : {
2693 2 : let mut slot_guard =
2694 2 : tenant_map_acquire_slot_impl(&tenant_shard_id, tenants, TenantSlotAcquireMode::MustExist)?;
2695 :
2696 : // allow pageserver shutdown to await for our completion
2697 2 : let (_guard, progress) = completion::channel();
2698 :
2699 : // The SlotGuard allows us to manipulate the Tenant object without fear of some
2700 : // concurrent API request doing something else for the same tenant ID.
2701 2 : let attached_tenant = match slot_guard.get_old_value() {
2702 2 : Some(TenantSlot::Attached(tenant)) => {
2703 2 : // whenever we remove a tenant from memory, we don't want to flush and wait for upload
2704 2 : let shutdown_mode = ShutdownMode::Hard;
2705 2 :
2706 2 : // shutdown is sure to transition tenant to stopping, and wait for all tasks to complete, so
2707 2 : // that we can continue safely to cleanup.
2708 2 : match tenant.shutdown(progress, shutdown_mode).await {
2709 2 : Ok(()) => {}
2710 0 : Err(_other) => {
2711 0 : // if pageserver shutdown or other detach/ignore is already ongoing, we don't want to
2712 0 : // wait for it but return an error right away because these are distinct requests.
2713 0 : slot_guard.revert();
2714 0 : return Err(TenantStateError::IsStopping(tenant_shard_id));
2715 : }
2716 : }
2717 2 : Some(tenant)
2718 : }
2719 0 : Some(TenantSlot::Secondary(secondary_state)) => {
2720 0 : tracing::info!("Shutting down in secondary mode");
2721 0 : secondary_state.shutdown().await;
2722 0 : None
2723 : }
2724 : Some(TenantSlot::InProgress(_)) => {
2725 : // Acquiring a slot guarantees its old value was not InProgress
2726 0 : unreachable!();
2727 : }
2728 0 : None => None,
2729 : };
2730 :
2731 2 : match tenant_cleanup
2732 2 : .await
2733 2 : .with_context(|| format!("Failed to run cleanup for tenant {tenant_shard_id}"))
2734 : {
2735 2 : Ok(hook_value) => {
2736 2 : // Success: drop the old TenantSlot::Attached.
2737 2 : slot_guard
2738 2 : .drop_old_value()
2739 2 : .expect("We just called shutdown");
2740 2 :
2741 2 : Ok(hook_value)
2742 : }
2743 0 : Err(e) => {
2744 : // If we had a Tenant, set it to Broken and put it back in the TenantsMap
2745 0 : if let Some(attached_tenant) = attached_tenant {
2746 0 : attached_tenant.set_broken(e.to_string()).await;
2747 0 : }
2748 : // Leave the broken tenant in the map
2749 0 : slot_guard.revert();
2750 0 :
2751 0 : Err(TenantStateError::Other(e))
2752 : }
2753 : }
2754 2 : }
2755 :
2756 : use {
2757 : crate::repository::GcResult, pageserver_api::models::TimelineGcRequest,
2758 : utils::http::error::ApiError,
2759 : };
2760 :
2761 0 : #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), %timeline_id))]
2762 : pub(crate) async fn immediate_gc(
2763 : tenant_shard_id: TenantShardId,
2764 : timeline_id: TimelineId,
2765 : gc_req: TimelineGcRequest,
2766 : cancel: CancellationToken,
2767 : ctx: &RequestContext,
2768 : ) -> Result<GcResult, ApiError> {
2769 : let tenant = {
2770 : let guard = TENANTS.read().unwrap();
2771 : guard
2772 : .get(&tenant_shard_id)
2773 : .cloned()
2774 0 : .with_context(|| format!("tenant {tenant_shard_id}"))
2775 0 : .map_err(|e| ApiError::NotFound(e.into()))?
2776 : };
2777 :
2778 0 : let gc_horizon = gc_req.gc_horizon.unwrap_or_else(|| tenant.get_gc_horizon());
2779 : // Use tenant's pitr setting
2780 : let pitr = tenant.get_pitr_interval();
2781 :
2782 : tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
2783 :
2784 : // Run in task_mgr to avoid race with tenant_detach operation
2785 : let ctx: RequestContext =
2786 : ctx.detached_child(TaskKind::GarbageCollector, DownloadBehavior::Download);
2787 :
2788 0 : let _gate_guard = tenant.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
2789 :
2790 : fail::fail_point!("immediate_gc_task_pre");
2791 :
2792 : #[allow(unused_mut)]
2793 : let mut result = tenant
2794 : .gc_iteration(Some(timeline_id), gc_horizon, pitr, &cancel, &ctx)
2795 : .await;
2796 : // FIXME: `gc_iteration` can return an error for multiple reasons; we should handle it
2797 : // better once the types support it.
2798 :
2799 : #[cfg(feature = "testing")]
2800 : {
2801 : // we need to synchronize with drop completion for python tests without polling for
2802 : // log messages
2803 : if let Ok(result) = result.as_mut() {
2804 : let mut js = tokio::task::JoinSet::new();
2805 : for layer in std::mem::take(&mut result.doomed_layers) {
2806 : js.spawn(layer.wait_drop());
2807 : }
2808 : tracing::info!(
2809 : total = js.len(),
2810 : "starting to wait for the gc'd layers to be dropped"
2811 : );
2812 : while let Some(res) = js.join_next().await {
2813 : res.expect("wait_drop should not panic");
2814 : }
2815 : }
2816 :
2817 : let timeline = tenant.get_timeline(timeline_id, false).ok();
2818 0 : let rtc = timeline.as_ref().map(|x| &x.remote_client);
2819 :
2820 : if let Some(rtc) = rtc {
2821 : // layer drops schedule actions on remote timeline client to actually do the
2822 : // deletions; don't care about the shutdown error, just exit fast
2823 : drop(rtc.wait_completion().await);
2824 : }
2825 : }
2826 :
2827 0 : result.map_err(|e| match e {
2828 0 : GcError::TenantCancelled | GcError::TimelineCancelled => ApiError::ShuttingDown,
2829 : GcError::TimelineNotFound => {
2830 0 : ApiError::NotFound(anyhow::anyhow!("Timeline not found").into())
2831 : }
2832 0 : other => ApiError::InternalServerError(anyhow::anyhow!(other)),
2833 0 : })
2834 : }
2835 :
2836 : #[cfg(test)]
2837 : mod tests {
2838 : use std::collections::BTreeMap;
2839 : use std::sync::Arc;
2840 : use tracing::Instrument;
2841 :
2842 : use crate::tenant::mgr::TenantSlot;
2843 :
2844 : use super::{super::harness::TenantHarness, TenantsMap};
2845 :
2846 : #[tokio::test(start_paused = true)]
2847 2 : async fn shutdown_awaits_in_progress_tenant() {
2848 2 : // Test that if an InProgress tenant is in the map during shutdown, the shutdown will gracefully
2849 2 : // wait for it to complete before proceeding.
2850 2 :
2851 2 : let h = TenantHarness::create("shutdown_awaits_in_progress_tenant")
2852 2 : .await
2853 2 : .unwrap();
2854 8 : let (t, _ctx) = h.load().await;
2855 2 :
2856 2 : // harness loads it to active, which is forced and nothing is running on the tenant
2857 2 :
2858 2 : let id = t.tenant_shard_id();
2859 2 :
2860 2 : // tenant harness configures the logging and we cannot escape it
2861 2 : let span = h.span();
2862 2 : let _e = span.enter();
2863 2 :
2864 2 : let tenants = BTreeMap::from([(id, TenantSlot::Attached(t.clone()))]);
2865 2 : let tenants = Arc::new(std::sync::RwLock::new(TenantsMap::Open(tenants)));
2866 2 :
2867 2 : // Invoke remove_tenant_from_memory with a cleanup hook that blocks until we manually
2868 2 : // permit it to proceed: that will stick the tenant in InProgress
2869 2 :
2870 2 : let (until_cleanup_completed, can_complete_cleanup) = utils::completion::channel();
2871 2 : let (until_cleanup_started, cleanup_started) = utils::completion::channel();
2872 2 : let mut remove_tenant_from_memory_task = {
2873 2 : let jh = tokio::spawn({
2874 2 : let tenants = tenants.clone();
2875 2 : async move {
2876 2 : let cleanup = async move {
2877 2 : drop(until_cleanup_started);
2878 2 : can_complete_cleanup.wait().await;
2879 2 : anyhow::Ok(())
2880 2 : };
2881 2 : super::remove_tenant_from_memory(&tenants, id, cleanup).await
2882 2 : }
2883 2 : .instrument(h.span())
2884 2 : });
2885 2 :
2886 2 : // now the long cleanup should be in place, with the stopping state
2887 2 : cleanup_started.wait().await;
2888 2 : jh
2889 2 : };
2890 2 :
2891 2 : let mut shutdown_task = {
2892 2 : let (until_shutdown_started, shutdown_started) = utils::completion::channel();
2893 2 :
2894 2 : let shutdown_task = tokio::spawn(async move {
2895 2 : drop(until_shutdown_started);
2896 4 : super::shutdown_all_tenants0(&tenants).await;
2897 2 : });
2898 2 :
2899 2 : shutdown_started.wait().await;
2900 2 : shutdown_task
2901 2 : };
2902 2 :
2903 2 : let long_time = std::time::Duration::from_secs(15);
2904 2 : tokio::select! {
2905 2 : _ = &mut shutdown_task => unreachable!("shutdown should block on remove_tenant_from_memory completing"),
2906 2 : _ = &mut remove_tenant_from_memory_task => unreachable!("remove_tenant_from_memory_task should not complete until explicitly unblocked"),
2907 2 : _ = tokio::time::sleep(long_time) => {},
2908 2 : }
2909 2 :
2910 2 : drop(until_cleanup_completed);
2911 2 :
2912 2 : // Now that we allow it to proceed, shutdown should complete immediately
2913 2 : remove_tenant_from_memory_task.await.unwrap().unwrap();
2914 2 : shutdown_task.await.unwrap();
2915 2 : }
2916 : }
|
