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, ShardNumber, ShardStripeSize, TenantShardId,
11 : };
12 : use pageserver_api::upcall_api::ReAttachResponseTenant;
13 : use rand::{distributions::Alphanumeric, Rng};
14 : use std::borrow::Cow;
15 : use std::cmp::Ordering;
16 : use std::collections::{BTreeMap, HashMap};
17 : use std::ops::Deref;
18 : use std::sync::Arc;
19 : use std::time::Duration;
20 : use sysinfo::SystemExt;
21 : use tokio::fs;
22 :
23 : use anyhow::Context;
24 : use once_cell::sync::Lazy;
25 : use tokio::task::JoinSet;
26 : use tokio_util::sync::CancellationToken;
27 : use tracing::*;
28 :
29 : use remote_storage::GenericRemoteStorage;
30 : use utils::{completion, crashsafe};
31 :
32 : use crate::config::PageServerConf;
33 : use crate::context::{DownloadBehavior, RequestContext};
34 : use crate::control_plane_client::{
35 : ControlPlaneClient, ControlPlaneGenerationsApi, RetryForeverError,
36 : };
37 : use crate::deletion_queue::DeletionQueueClient;
38 : use crate::http::routes::ACTIVE_TENANT_TIMEOUT;
39 : use crate::metrics::{TENANT, TENANT_MANAGER as METRICS};
40 : use crate::task_mgr::{self, TaskKind};
41 : use crate::tenant::config::{
42 : AttachedLocationConfig, AttachmentMode, LocationConf, LocationMode, SecondaryLocationConfig,
43 : };
44 : use crate::tenant::delete::DeleteTenantFlow;
45 : use crate::tenant::span::debug_assert_current_span_has_tenant_id;
46 : use crate::tenant::storage_layer::inmemory_layer;
47 : use crate::tenant::timeline::ShutdownMode;
48 : use crate::tenant::{AttachedTenantConf, SpawnMode, Tenant, TenantState};
49 : use crate::{InitializationOrder, IGNORED_TENANT_FILE_NAME, TEMP_FILE_SUFFIX};
50 :
51 : use utils::crashsafe::path_with_suffix_extension;
52 : use utils::fs_ext::PathExt;
53 : use utils::generation::Generation;
54 : use utils::id::{TenantId, TimelineId};
55 :
56 : use super::delete::DeleteTenantError;
57 : use super::secondary::SecondaryTenant;
58 : use super::timeline::detach_ancestor::PreparedTimelineDetach;
59 : use super::TenantSharedResources;
60 :
61 : /// For a tenant that appears in TenantsMap, it may either be
62 : /// - `Attached`: has a full Tenant object, is elegible to service
63 : /// reads and ingest WAL.
64 : /// - `Secondary`: is only keeping a local cache warm.
65 : ///
66 : /// Secondary is a totally distinct state rather than being a mode of a `Tenant`, because
67 : /// that way we avoid having to carefully switch a tenant's ingestion etc on and off during
68 : /// its lifetime, and we can preserve some important safety invariants like `Tenant` always
69 : /// having a properly acquired generation (Secondary doesn't need a generation)
70 : #[derive(Clone)]
71 : pub(crate) enum TenantSlot {
72 : Attached(Arc<Tenant>),
73 : Secondary(Arc<SecondaryTenant>),
74 : /// In this state, other administrative operations acting on the TenantId should
75 : /// block, or return a retry indicator equivalent to HTTP 503.
76 : InProgress(utils::completion::Barrier),
77 : }
78 :
79 : impl std::fmt::Debug for TenantSlot {
80 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
81 0 : match self {
82 0 : Self::Attached(tenant) => write!(f, "Attached({})", tenant.current_state()),
83 0 : Self::Secondary(_) => write!(f, "Secondary"),
84 0 : Self::InProgress(_) => write!(f, "InProgress"),
85 : }
86 0 : }
87 : }
88 :
89 : impl TenantSlot {
90 : /// Return the `Tenant` in this slot if attached, else None
91 0 : fn get_attached(&self) -> Option<&Arc<Tenant>> {
92 0 : match self {
93 0 : Self::Attached(t) => Some(t),
94 0 : Self::Secondary(_) => None,
95 0 : Self::InProgress(_) => None,
96 : }
97 0 : }
98 : }
99 :
100 : /// The tenants known to the pageserver.
101 : /// The enum variants are used to distinguish the different states that the pageserver can be in.
102 : pub(crate) enum TenantsMap {
103 : /// [`init_tenant_mgr`] is not done yet.
104 : Initializing,
105 : /// [`init_tenant_mgr`] is done, all on-disk tenants have been loaded.
106 : /// New tenants can be added using [`tenant_map_acquire_slot`].
107 : Open(BTreeMap<TenantShardId, TenantSlot>),
108 : /// The pageserver has entered shutdown mode via [`TenantManager::shutdown`].
109 : /// Existing tenants are still accessible, but no new tenants can be created.
110 : ShuttingDown(BTreeMap<TenantShardId, TenantSlot>),
111 : }
112 :
113 : pub(crate) enum TenantsMapRemoveResult {
114 : Occupied(TenantSlot),
115 : Vacant,
116 : InProgress(utils::completion::Barrier),
117 : }
118 :
119 : /// When resolving a TenantId to a shard, we may be looking for the 0th
120 : /// shard, or we might be looking for whichever shard holds a particular page.
121 : #[derive(Copy, Clone)]
122 : pub(crate) enum ShardSelector {
123 : /// Only return the 0th shard, if it is present. If a non-0th shard is present,
124 : /// ignore it.
125 : Zero,
126 : /// Pick the first shard we find for the TenantId
127 : First,
128 : /// Pick the shard that holds this key
129 : Page(Key),
130 : }
131 :
132 : /// A convenience for use with the re_attach ControlPlaneClient function: rather
133 : /// than the serializable struct, we build this enum that encapsulates
134 : /// the invariant that attached tenants always have generations.
135 : ///
136 : /// This represents the subset of a LocationConfig that we receive during re-attach.
137 : pub(crate) enum TenantStartupMode {
138 : Attached((AttachmentMode, Generation)),
139 : Secondary,
140 : }
141 :
142 : impl TenantStartupMode {
143 : /// Return the generation & mode that should be used when starting
144 : /// this tenant.
145 : ///
146 : /// If this returns None, the re-attach struct is in an invalid state and
147 : /// should be ignored in the response.
148 0 : fn from_reattach_tenant(rart: ReAttachResponseTenant) -> Option<Self> {
149 0 : match (rart.mode, rart.gen) {
150 0 : (LocationConfigMode::Detached, _) => None,
151 0 : (LocationConfigMode::Secondary, _) => Some(Self::Secondary),
152 0 : (LocationConfigMode::AttachedMulti, Some(g)) => {
153 0 : Some(Self::Attached((AttachmentMode::Multi, Generation::new(g))))
154 : }
155 0 : (LocationConfigMode::AttachedSingle, Some(g)) => {
156 0 : Some(Self::Attached((AttachmentMode::Single, Generation::new(g))))
157 : }
158 0 : (LocationConfigMode::AttachedStale, Some(g)) => {
159 0 : Some(Self::Attached((AttachmentMode::Stale, Generation::new(g))))
160 : }
161 : _ => {
162 0 : tracing::warn!(
163 0 : "Received invalid re-attach state for tenant {}: {rart:?}",
164 : rart.id
165 : );
166 0 : None
167 : }
168 : }
169 0 : }
170 : }
171 :
172 : /// Result type for looking up a TenantId to a specific shard
173 : pub(crate) enum ShardResolveResult {
174 : NotFound,
175 : Found(Arc<Tenant>),
176 : // Wait for this barrrier, then query again
177 : InProgress(utils::completion::Barrier),
178 : }
179 :
180 : impl TenantsMap {
181 : /// Convenience function for typical usage, where we want to get a `Tenant` object, for
182 : /// working with attached tenants. If the TenantId is in the map but in Secondary state,
183 : /// None is returned.
184 0 : pub(crate) fn get(&self, tenant_shard_id: &TenantShardId) -> Option<&Arc<Tenant>> {
185 0 : match self {
186 0 : TenantsMap::Initializing => None,
187 0 : TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => {
188 0 : m.get(tenant_shard_id).and_then(|slot| slot.get_attached())
189 : }
190 : }
191 0 : }
192 :
193 : /// Only for use from DeleteTenantFlow. This method directly removes a TenantSlot from the map.
194 : ///
195 : /// The normal way to remove a tenant is using a SlotGuard, which will gracefully remove the guarded
196 : /// slot if the enclosed tenant is shutdown.
197 0 : pub(crate) fn remove(&mut self, tenant_shard_id: TenantShardId) -> TenantsMapRemoveResult {
198 0 : use std::collections::btree_map::Entry;
199 0 : match self {
200 0 : TenantsMap::Initializing => TenantsMapRemoveResult::Vacant,
201 0 : TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => match m.entry(tenant_shard_id) {
202 0 : Entry::Occupied(entry) => match entry.get() {
203 0 : TenantSlot::InProgress(barrier) => {
204 0 : TenantsMapRemoveResult::InProgress(barrier.clone())
205 : }
206 0 : _ => TenantsMapRemoveResult::Occupied(entry.remove()),
207 : },
208 0 : Entry::Vacant(_entry) => TenantsMapRemoveResult::Vacant,
209 : },
210 : }
211 0 : }
212 :
213 : #[cfg(all(debug_assertions, not(test)))]
214 0 : pub(crate) fn len(&self) -> usize {
215 0 : match self {
216 0 : TenantsMap::Initializing => 0,
217 0 : TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => m.len(),
218 : }
219 0 : }
220 : }
221 :
222 : /// Precursor to deletion of a tenant dir: we do a fast rename to a tmp path, and then
223 : /// the slower actual deletion in the background.
224 : ///
225 : /// This is "safe" in that that it won't leave behind a partially deleted directory
226 : /// at the original path, because we rename with TEMP_FILE_SUFFIX before starting deleting
227 : /// the contents.
228 : ///
229 : /// This is pageserver-specific, as it relies on future processes after a crash to check
230 : /// for TEMP_FILE_SUFFIX when loading things.
231 0 : async fn safe_rename_tenant_dir(path: impl AsRef<Utf8Path>) -> std::io::Result<Utf8PathBuf> {
232 0 : let parent = path
233 0 : .as_ref()
234 0 : .parent()
235 0 : // It is invalid to call this function with a relative path. Tenant directories
236 0 : // should always have a parent.
237 0 : .ok_or(std::io::Error::new(
238 0 : std::io::ErrorKind::InvalidInput,
239 0 : "Path must be absolute",
240 0 : ))?;
241 0 : let rand_suffix = rand::thread_rng()
242 0 : .sample_iter(&Alphanumeric)
243 0 : .take(8)
244 0 : .map(char::from)
245 0 : .collect::<String>()
246 0 : + TEMP_FILE_SUFFIX;
247 0 : let tmp_path = path_with_suffix_extension(&path, &rand_suffix);
248 0 : fs::rename(path.as_ref(), &tmp_path).await?;
249 0 : fs::File::open(parent).await?.sync_all().await?;
250 0 : Ok(tmp_path)
251 0 : }
252 :
253 : /// When we have moved a tenant's content to a temporary directory, we may delete it lazily in
254 : /// the background, and thereby avoid blocking any API requests on this deletion completing.
255 0 : fn spawn_background_purge(tmp_path: Utf8PathBuf) {
256 0 : // Although we are cleaning up the tenant, this task is not meant to be bound by the lifetime of the tenant in memory.
257 0 : // After a tenant is detached, there are no more task_mgr tasks for that tenant_id.
258 0 : let task_tenant_id = None;
259 0 :
260 0 : task_mgr::spawn(
261 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
262 0 : TaskKind::MgmtRequest,
263 0 : task_tenant_id,
264 0 : None,
265 0 : "tenant_files_delete",
266 0 : false,
267 0 : async move {
268 0 : fs::remove_dir_all(tmp_path.as_path())
269 0 : .await
270 0 : .with_context(|| format!("tenant directory {:?} deletion", tmp_path))
271 0 : },
272 0 : );
273 0 : }
274 :
275 : static TENANTS: Lazy<std::sync::RwLock<TenantsMap>> =
276 2 : Lazy::new(|| std::sync::RwLock::new(TenantsMap::Initializing));
277 :
278 : /// The TenantManager is responsible for storing and mutating the collection of all tenants
279 : /// that this pageserver process has state for. Every Tenant and SecondaryTenant instance
280 : /// lives inside the TenantManager.
281 : ///
282 : /// The most important role of the TenantManager is to prevent conflicts: e.g. trying to attach
283 : /// the same tenant twice concurrently, or trying to configure the same tenant into secondary
284 : /// and attached modes concurrently.
285 : pub struct TenantManager {
286 : conf: &'static PageServerConf,
287 : // TODO: currently this is a &'static pointing to TENANTs. When we finish refactoring
288 : // out of that static variable, the TenantManager can own this.
289 : // See https://github.com/neondatabase/neon/issues/5796
290 : tenants: &'static std::sync::RwLock<TenantsMap>,
291 : resources: TenantSharedResources,
292 :
293 : // Long-running operations that happen outside of a [`Tenant`] lifetime should respect this token.
294 : // This is for edge cases like tenant deletion. In normal cases (within a Tenant lifetime),
295 : // tenants have their own cancellation tokens, which we fire individually in [`Self::shutdown`], or
296 : // when the tenant detaches.
297 : cancel: CancellationToken,
298 : }
299 :
300 0 : fn emergency_generations(
301 0 : tenant_confs: &HashMap<TenantShardId, anyhow::Result<LocationConf>>,
302 0 : ) -> HashMap<TenantShardId, TenantStartupMode> {
303 0 : tenant_confs
304 0 : .iter()
305 0 : .filter_map(|(tid, lc)| {
306 0 : let lc = match lc {
307 0 : Ok(lc) => lc,
308 0 : Err(_) => return None,
309 : };
310 : Some((
311 0 : *tid,
312 0 : match &lc.mode {
313 0 : LocationMode::Attached(alc) => {
314 0 : TenantStartupMode::Attached((alc.attach_mode, alc.generation))
315 : }
316 0 : LocationMode::Secondary(_) => TenantStartupMode::Secondary,
317 : },
318 : ))
319 0 : })
320 0 : .collect()
321 0 : }
322 :
323 0 : async fn init_load_generations(
324 0 : conf: &'static PageServerConf,
325 0 : tenant_confs: &HashMap<TenantShardId, anyhow::Result<LocationConf>>,
326 0 : resources: &TenantSharedResources,
327 0 : cancel: &CancellationToken,
328 0 : ) -> anyhow::Result<Option<HashMap<TenantShardId, TenantStartupMode>>> {
329 0 : let generations = if conf.control_plane_emergency_mode {
330 0 : error!(
331 0 : "Emergency mode! Tenants will be attached unsafely using their last known generation"
332 : );
333 0 : emergency_generations(tenant_confs)
334 0 : } else if let Some(client) = ControlPlaneClient::new(conf, cancel) {
335 0 : info!("Calling control plane API to re-attach tenants");
336 : // If we are configured to use the control plane API, then it is the source of truth for what tenants to load.
337 0 : match client.re_attach(conf).await {
338 0 : Ok(tenants) => tenants
339 0 : .into_iter()
340 0 : .flat_map(|(id, rart)| {
341 0 : TenantStartupMode::from_reattach_tenant(rart).map(|tsm| (id, tsm))
342 0 : })
343 0 : .collect(),
344 : Err(RetryForeverError::ShuttingDown) => {
345 0 : anyhow::bail!("Shut down while waiting for control plane re-attach response")
346 : }
347 : }
348 : } else {
349 0 : info!("Control plane API not configured, tenant generations are disabled");
350 0 : return Ok(None);
351 : };
352 :
353 : // The deletion queue needs to know about the startup attachment state to decide which (if any) stored
354 : // deletion list entries may still be valid. We provide that by pushing a recovery operation into
355 : // the queue. Sequential processing of te queue ensures that recovery is done before any new tenant deletions
356 : // are processed, even though we don't block on recovery completing here.
357 0 : let attached_tenants = generations
358 0 : .iter()
359 0 : .flat_map(|(id, start_mode)| {
360 0 : match start_mode {
361 0 : TenantStartupMode::Attached((_mode, generation)) => Some(generation),
362 0 : TenantStartupMode::Secondary => None,
363 : }
364 0 : .map(|gen| (*id, *gen))
365 0 : })
366 0 : .collect();
367 0 : resources.deletion_queue_client.recover(attached_tenants)?;
368 :
369 0 : Ok(Some(generations))
370 0 : }
371 :
372 : /// Given a directory discovered in the pageserver's tenants/ directory, attempt
373 : /// to load a tenant config from it.
374 : ///
375 : /// If file is missing, return Ok(None)
376 0 : fn load_tenant_config(
377 0 : conf: &'static PageServerConf,
378 0 : dentry: Utf8DirEntry,
379 0 : ) -> anyhow::Result<Option<(TenantShardId, anyhow::Result<LocationConf>)>> {
380 0 : let tenant_dir_path = dentry.path().to_path_buf();
381 0 : if crate::is_temporary(&tenant_dir_path) {
382 0 : info!("Found temporary tenant directory, removing: {tenant_dir_path}");
383 : // No need to use safe_remove_tenant_dir_all because this is already
384 : // a temporary path
385 0 : if let Err(e) = std::fs::remove_dir_all(&tenant_dir_path) {
386 0 : error!(
387 0 : "Failed to remove temporary directory '{}': {:?}",
388 : tenant_dir_path, e
389 : );
390 0 : }
391 0 : return Ok(None);
392 0 : }
393 :
394 : // This case happens if we crash during attachment before writing a config into the dir
395 0 : let is_empty = tenant_dir_path
396 0 : .is_empty_dir()
397 0 : .with_context(|| format!("Failed to check whether {tenant_dir_path:?} is an empty dir"))?;
398 0 : if is_empty {
399 0 : info!("removing empty tenant directory {tenant_dir_path:?}");
400 0 : if let Err(e) = std::fs::remove_dir(&tenant_dir_path) {
401 0 : error!(
402 0 : "Failed to remove empty tenant directory '{}': {e:#}",
403 : tenant_dir_path
404 : )
405 0 : }
406 0 : return Ok(None);
407 0 : }
408 :
409 0 : let tenant_shard_id = match tenant_dir_path
410 0 : .file_name()
411 0 : .unwrap_or_default()
412 0 : .parse::<TenantShardId>()
413 : {
414 0 : Ok(id) => id,
415 : Err(_) => {
416 0 : warn!("Invalid tenant path (garbage in our repo directory?): {tenant_dir_path}",);
417 0 : return Ok(None);
418 : }
419 : };
420 :
421 0 : let tenant_ignore_mark_file = tenant_dir_path.join(IGNORED_TENANT_FILE_NAME);
422 0 : if tenant_ignore_mark_file.exists() {
423 0 : info!("Found an ignore mark file {tenant_ignore_mark_file:?}, skipping the tenant");
424 0 : return Ok(None);
425 0 : }
426 0 :
427 0 : Ok(Some((
428 0 : tenant_shard_id,
429 0 : Tenant::load_tenant_config(conf, &tenant_shard_id),
430 0 : )))
431 0 : }
432 :
433 : /// Initial stage of load: walk the local tenants directory, clean up any temp files,
434 : /// and load configurations for the tenants we found.
435 : ///
436 : /// Do this in parallel, because we expect 10k+ tenants, so serial execution can take
437 : /// seconds even on reasonably fast drives.
438 0 : async fn init_load_tenant_configs(
439 0 : conf: &'static PageServerConf,
440 0 : ) -> anyhow::Result<HashMap<TenantShardId, anyhow::Result<LocationConf>>> {
441 0 : let tenants_dir = conf.tenants_path();
442 :
443 0 : let dentries = tokio::task::spawn_blocking(move || -> anyhow::Result<Vec<Utf8DirEntry>> {
444 0 : let dir_entries = tenants_dir
445 0 : .read_dir_utf8()
446 0 : .with_context(|| format!("Failed to list tenants dir {tenants_dir:?}"))?;
447 :
448 0 : Ok(dir_entries.collect::<Result<Vec<_>, std::io::Error>>()?)
449 0 : })
450 0 : .await??;
451 :
452 0 : let mut configs = HashMap::new();
453 0 :
454 0 : let mut join_set = JoinSet::new();
455 0 : for dentry in dentries {
456 0 : join_set.spawn_blocking(move || load_tenant_config(conf, dentry));
457 0 : }
458 :
459 0 : while let Some(r) = join_set.join_next().await {
460 0 : if let Some((tenant_id, tenant_config)) = r?? {
461 0 : configs.insert(tenant_id, tenant_config);
462 0 : }
463 : }
464 :
465 0 : Ok(configs)
466 0 : }
467 :
468 : /// Initialize repositories with locally available timelines.
469 : /// Timelines that are only partially available locally (remote storage has more data than this pageserver)
470 : /// are scheduled for download and added to the tenant once download is completed.
471 0 : #[instrument(skip_all)]
472 : pub async fn init_tenant_mgr(
473 : conf: &'static PageServerConf,
474 : resources: TenantSharedResources,
475 : init_order: InitializationOrder,
476 : cancel: CancellationToken,
477 : ) -> anyhow::Result<TenantManager> {
478 : let mut tenants = BTreeMap::new();
479 :
480 : let ctx = RequestContext::todo_child(TaskKind::Startup, DownloadBehavior::Warn);
481 :
482 : // Initialize dynamic limits that depend on system resources
483 : let system_memory =
484 : sysinfo::System::new_with_specifics(sysinfo::RefreshKind::new().with_memory())
485 : .total_memory();
486 : let max_ephemeral_layer_bytes =
487 : conf.ephemeral_bytes_per_memory_kb as u64 * (system_memory / 1024);
488 : tracing::info!("Initialized ephemeral layer size limit to {max_ephemeral_layer_bytes}, for {system_memory} bytes of memory");
489 : inmemory_layer::GLOBAL_RESOURCES.max_dirty_bytes.store(
490 : max_ephemeral_layer_bytes,
491 : std::sync::atomic::Ordering::Relaxed,
492 : );
493 :
494 : // Scan local filesystem for attached tenants
495 : let tenant_configs = init_load_tenant_configs(conf).await?;
496 :
497 : // Determine which tenants are to be secondary or attached, and in which generation
498 : let tenant_modes = init_load_generations(conf, &tenant_configs, &resources, &cancel).await?;
499 :
500 : tracing::info!(
501 : "Attaching {} tenants at startup, warming up {} at a time",
502 : tenant_configs.len(),
503 : conf.concurrent_tenant_warmup.initial_permits()
504 : );
505 : TENANT.startup_scheduled.inc_by(tenant_configs.len() as u64);
506 :
507 : // Accumulate futures for writing tenant configs, so that we can execute in parallel
508 : let mut config_write_futs = Vec::new();
509 :
510 : // Update the location configs according to the re-attach response and persist them to disk
511 : tracing::info!("Updating {} location configs", tenant_configs.len());
512 : for (tenant_shard_id, location_conf) in tenant_configs {
513 : let tenant_dir_path = conf.tenant_path(&tenant_shard_id);
514 :
515 : let mut location_conf = match location_conf {
516 : Ok(l) => l,
517 : Err(e) => {
518 : warn!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Marking tenant broken, failed to {e:#}");
519 :
520 : tenants.insert(
521 : tenant_shard_id,
522 : TenantSlot::Attached(Tenant::create_broken_tenant(
523 : conf,
524 : tenant_shard_id,
525 : resources.remote_storage.clone(),
526 : format!("{}", e),
527 : )),
528 : );
529 : continue;
530 : }
531 : };
532 :
533 : // FIXME: if we were attached, and get demoted to secondary on re-attach, we
534 : // don't have a place to get a config.
535 : // (https://github.com/neondatabase/neon/issues/5377)
536 : const DEFAULT_SECONDARY_CONF: SecondaryLocationConfig =
537 : SecondaryLocationConfig { warm: true };
538 :
539 : if let Some(tenant_modes) = &tenant_modes {
540 : // We have a generation map: treat it as the authority for whether
541 : // this tenant is really attached.
542 : match tenant_modes.get(&tenant_shard_id) {
543 : None => {
544 : 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");
545 :
546 : match safe_rename_tenant_dir(&tenant_dir_path).await {
547 : Ok(tmp_path) => {
548 : spawn_background_purge(tmp_path);
549 : }
550 : Err(e) => {
551 : error!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
552 : "Failed to move detached tenant directory '{tenant_dir_path}': {e:?}");
553 : }
554 : };
555 :
556 : // We deleted local content: move on to next tenant, don't try and spawn this one.
557 : continue;
558 : }
559 : Some(TenantStartupMode::Secondary) => {
560 : if !matches!(location_conf.mode, LocationMode::Secondary(_)) {
561 : location_conf.mode = LocationMode::Secondary(DEFAULT_SECONDARY_CONF);
562 : }
563 : }
564 : Some(TenantStartupMode::Attached((attach_mode, generation))) => {
565 : let old_gen_higher = match &location_conf.mode {
566 : LocationMode::Attached(AttachedLocationConfig {
567 : generation: old_generation,
568 : attach_mode: _attach_mode,
569 : }) => {
570 : if old_generation > generation {
571 : Some(old_generation)
572 : } else {
573 : None
574 : }
575 : }
576 : _ => None,
577 : };
578 : if let Some(old_generation) = old_gen_higher {
579 : tracing::error!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
580 : "Control plane gave decreasing generation ({generation:?}) in re-attach response for tenant that was attached in generation {:?}, demoting to secondary",
581 : old_generation
582 : );
583 :
584 : // We cannot safely attach this tenant given a bogus generation number, but let's avoid throwing away
585 : // local disk content: demote to secondary rather than detaching.
586 : location_conf.mode = LocationMode::Secondary(DEFAULT_SECONDARY_CONF);
587 : } else {
588 : location_conf.attach_in_generation(*attach_mode, *generation);
589 : }
590 : }
591 : }
592 : } else {
593 : // Legacy mode: no generation information, any tenant present
594 : // on local disk may activate
595 : info!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Starting tenant in legacy mode, no generation",);
596 : };
597 :
598 : // Presence of a generation number implies attachment: attach the tenant
599 : // if it wasn't already, and apply the generation number.
600 0 : config_write_futs.push(async move {
601 0 : let r = Tenant::persist_tenant_config(conf, &tenant_shard_id, &location_conf).await;
602 0 : (tenant_shard_id, location_conf, r)
603 0 : });
604 : }
605 :
606 : // Execute config writes with concurrency, to avoid bottlenecking on local FS write latency
607 : tracing::info!(
608 : "Writing {} location config files...",
609 : config_write_futs.len()
610 : );
611 : let config_write_results = futures::stream::iter(config_write_futs)
612 : .buffer_unordered(16)
613 : .collect::<Vec<_>>()
614 : .await;
615 :
616 : tracing::info!(
617 : "Spawning {} tenant shard locations...",
618 : config_write_results.len()
619 : );
620 : // For those shards that have live configurations, construct `Tenant` or `SecondaryTenant` objects and start them running
621 : for (tenant_shard_id, location_conf, config_write_result) in config_write_results {
622 : // Errors writing configs are fatal
623 : config_write_result?;
624 :
625 : let tenant_dir_path = conf.tenant_path(&tenant_shard_id);
626 : let shard_identity = location_conf.shard;
627 : let slot = match location_conf.mode {
628 : LocationMode::Attached(attached_conf) => {
629 : match tenant_spawn(
630 : conf,
631 : tenant_shard_id,
632 : &tenant_dir_path,
633 : resources.clone(),
634 : AttachedTenantConf::new(location_conf.tenant_conf, attached_conf),
635 : shard_identity,
636 : Some(init_order.clone()),
637 : &TENANTS,
638 : SpawnMode::Lazy,
639 : &ctx,
640 : ) {
641 : Ok(tenant) => TenantSlot::Attached(tenant),
642 : Err(e) => {
643 : error!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Failed to start tenant: {e:#}");
644 : continue;
645 : }
646 : }
647 : }
648 : LocationMode::Secondary(secondary_conf) => {
649 : info!(
650 : tenant_id = %tenant_shard_id.tenant_id,
651 : shard_id = %tenant_shard_id.shard_slug(),
652 : "Starting secondary tenant"
653 : );
654 : TenantSlot::Secondary(SecondaryTenant::new(
655 : tenant_shard_id,
656 : shard_identity,
657 : location_conf.tenant_conf,
658 : &secondary_conf,
659 : ))
660 : }
661 : };
662 :
663 : METRICS.slot_inserted(&slot);
664 : tenants.insert(tenant_shard_id, slot);
665 : }
666 :
667 : info!("Processed {} local tenants at startup", tenants.len());
668 :
669 : let mut tenants_map = TENANTS.write().unwrap();
670 : assert!(matches!(&*tenants_map, &TenantsMap::Initializing));
671 :
672 : *tenants_map = TenantsMap::Open(tenants);
673 :
674 : Ok(TenantManager {
675 : conf,
676 : tenants: &TENANTS,
677 : resources,
678 : cancel: CancellationToken::new(),
679 : })
680 : }
681 :
682 : /// Wrapper for Tenant::spawn that checks invariants before running, and inserts
683 : /// a broken tenant in the map if Tenant::spawn fails.
684 : #[allow(clippy::too_many_arguments)]
685 0 : fn tenant_spawn(
686 0 : conf: &'static PageServerConf,
687 0 : tenant_shard_id: TenantShardId,
688 0 : tenant_path: &Utf8Path,
689 0 : resources: TenantSharedResources,
690 0 : location_conf: AttachedTenantConf,
691 0 : shard_identity: ShardIdentity,
692 0 : init_order: Option<InitializationOrder>,
693 0 : tenants: &'static std::sync::RwLock<TenantsMap>,
694 0 : mode: SpawnMode,
695 0 : ctx: &RequestContext,
696 0 : ) -> anyhow::Result<Arc<Tenant>> {
697 0 : anyhow::ensure!(
698 0 : tenant_path.is_dir(),
699 0 : "Cannot load tenant from path {tenant_path:?}, it either does not exist or not a directory"
700 : );
701 0 : anyhow::ensure!(
702 0 : !crate::is_temporary(tenant_path),
703 0 : "Cannot load tenant from temporary path {tenant_path:?}"
704 : );
705 0 : anyhow::ensure!(
706 0 : !tenant_path.is_empty_dir().with_context(|| {
707 0 : format!("Failed to check whether {tenant_path:?} is an empty dir")
708 0 : })?,
709 0 : "Cannot load tenant from empty directory {tenant_path:?}"
710 : );
711 :
712 0 : let tenant_ignore_mark = conf.tenant_ignore_mark_file_path(&tenant_shard_id);
713 0 : anyhow::ensure!(
714 0 : !conf.tenant_ignore_mark_file_path(&tenant_shard_id).exists(),
715 0 : "Cannot load tenant, ignore mark found at {tenant_ignore_mark:?}"
716 : );
717 :
718 0 : let remote_storage = resources.remote_storage.clone();
719 0 : let tenant = match Tenant::spawn(
720 0 : conf,
721 0 : tenant_shard_id,
722 0 : resources,
723 0 : location_conf,
724 0 : shard_identity,
725 0 : init_order,
726 0 : tenants,
727 0 : mode,
728 0 : ctx,
729 0 : ) {
730 0 : Ok(tenant) => tenant,
731 0 : Err(e) => {
732 0 : error!("Failed to spawn tenant {tenant_shard_id}, reason: {e:#}");
733 0 : Tenant::create_broken_tenant(conf, tenant_shard_id, remote_storage, format!("{e:#}"))
734 : }
735 : };
736 :
737 0 : Ok(tenant)
738 0 : }
739 :
740 2 : async fn shutdown_all_tenants0(tenants: &std::sync::RwLock<TenantsMap>) {
741 2 : let mut join_set = JoinSet::new();
742 0 :
743 0 : #[cfg(all(debug_assertions, not(test)))]
744 0 : {
745 0 : // Check that our metrics properly tracked the size of the tenants map. This is a convenient location to check,
746 0 : // as it happens implicitly at the end of tests etc.
747 0 : let m = tenants.read().unwrap();
748 0 : debug_assert_eq!(METRICS.slots_total(), m.len() as u64);
749 : }
750 :
751 : // Atomically, 1. create the shutdown tasks and 2. prevent creation of new tenants.
752 2 : let (total_in_progress, total_attached) = {
753 2 : let mut m = tenants.write().unwrap();
754 2 : match &mut *m {
755 : TenantsMap::Initializing => {
756 0 : *m = TenantsMap::ShuttingDown(BTreeMap::default());
757 0 : info!("tenants map is empty");
758 0 : return;
759 : }
760 2 : TenantsMap::Open(tenants) => {
761 2 : let mut shutdown_state = BTreeMap::new();
762 2 : let mut total_in_progress = 0;
763 2 : let mut total_attached = 0;
764 :
765 2 : for (tenant_shard_id, v) in std::mem::take(tenants).into_iter() {
766 2 : match v {
767 0 : TenantSlot::Attached(t) => {
768 0 : shutdown_state.insert(tenant_shard_id, TenantSlot::Attached(t.clone()));
769 0 : join_set.spawn(
770 0 : async move {
771 0 : let res = {
772 0 : let (_guard, shutdown_progress) = completion::channel();
773 0 : t.shutdown(shutdown_progress, ShutdownMode::FreezeAndFlush).await
774 : };
775 :
776 0 : if let Err(other_progress) = res {
777 : // join the another shutdown in progress
778 0 : other_progress.wait().await;
779 0 : }
780 :
781 : // we cannot afford per tenant logging here, because if s3 is degraded, we are
782 : // going to log too many lines
783 0 : debug!("tenant successfully stopped");
784 0 : }
785 0 : .instrument(info_span!("shutdown", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug())),
786 : );
787 :
788 0 : total_attached += 1;
789 : }
790 0 : TenantSlot::Secondary(state) => {
791 0 : // We don't need to wait for this individually per-tenant: the
792 0 : // downloader task will be waited on eventually, this cancel
793 0 : // is just to encourage it to drop out if it is doing work
794 0 : // for this tenant right now.
795 0 : state.cancel.cancel();
796 0 :
797 0 : shutdown_state.insert(tenant_shard_id, TenantSlot::Secondary(state));
798 0 : }
799 2 : TenantSlot::InProgress(notify) => {
800 2 : // InProgress tenants are not visible in TenantsMap::ShuttingDown: we will
801 2 : // wait for their notifications to fire in this function.
802 2 : join_set.spawn(async move {
803 2 : notify.wait().await;
804 2 : });
805 2 :
806 2 : total_in_progress += 1;
807 2 : }
808 : }
809 : }
810 2 : *m = TenantsMap::ShuttingDown(shutdown_state);
811 2 : (total_in_progress, total_attached)
812 : }
813 : TenantsMap::ShuttingDown(_) => {
814 0 : error!("already shutting down, this function isn't supposed to be called more than once");
815 0 : return;
816 : }
817 : }
818 : };
819 :
820 2 : let started_at = std::time::Instant::now();
821 2 :
822 2 : info!(
823 0 : "Waiting for {} InProgress tenants and {} Attached tenants to shut down",
824 : total_in_progress, total_attached
825 : );
826 :
827 2 : let total = join_set.len();
828 2 : let mut panicked = 0;
829 2 : let mut buffering = true;
830 2 : const BUFFER_FOR: std::time::Duration = std::time::Duration::from_millis(500);
831 2 : let mut buffered = std::pin::pin!(tokio::time::sleep(BUFFER_FOR));
832 :
833 6 : while !join_set.is_empty() {
834 : tokio::select! {
835 : Some(joined) = join_set.join_next() => {
836 : match joined {
837 : Ok(()) => {},
838 : Err(join_error) if join_error.is_cancelled() => {
839 : unreachable!("we are not cancelling any of the tasks");
840 : }
841 : Err(join_error) if join_error.is_panic() => {
842 : // cannot really do anything, as this panic is likely a bug
843 : panicked += 1;
844 : }
845 : Err(join_error) => {
846 : warn!("unknown kind of JoinError: {join_error}");
847 : }
848 : }
849 : if !buffering {
850 : // buffer so that every 500ms since the first update (or starting) we'll log
851 : // how far away we are; this is because we will get SIGKILL'd at 10s, and we
852 : // are not able to log *then*.
853 : buffering = true;
854 : buffered.as_mut().reset(tokio::time::Instant::now() + BUFFER_FOR);
855 : }
856 : },
857 : _ = &mut buffered, if buffering => {
858 : buffering = false;
859 : info!(remaining = join_set.len(), total, elapsed_ms = started_at.elapsed().as_millis(), "waiting for tenants to shutdown");
860 : }
861 : }
862 : }
863 :
864 2 : if panicked > 0 {
865 0 : warn!(
866 : panicked,
867 0 : total, "observed panicks while shutting down tenants"
868 : );
869 2 : }
870 :
871 : // caller will log how long we took
872 2 : }
873 :
874 0 : #[derive(thiserror::Error, Debug)]
875 : pub(crate) enum UpsertLocationError {
876 : #[error("Bad config request: {0}")]
877 : BadRequest(anyhow::Error),
878 :
879 : #[error("Cannot change config in this state: {0}")]
880 : Unavailable(#[from] TenantMapError),
881 :
882 : #[error("Tenant is already being modified")]
883 : InProgress,
884 :
885 : #[error("Failed to flush: {0}")]
886 : Flush(anyhow::Error),
887 :
888 : #[error("Internal error: {0}")]
889 : Other(#[from] anyhow::Error),
890 : }
891 :
892 : impl TenantManager {
893 : /// Convenience function so that anyone with a TenantManager can get at the global configuration, without
894 : /// having to pass it around everywhere as a separate object.
895 0 : pub(crate) fn get_conf(&self) -> &'static PageServerConf {
896 0 : self.conf
897 0 : }
898 :
899 : /// Gets the attached tenant from the in-memory data, erroring if it's absent, in secondary mode, or currently
900 : /// undergoing a state change (i.e. slot is InProgress).
901 : ///
902 : /// The return Tenant is not guaranteed to be active: check its status after obtaing it, or
903 : /// use [`Tenant::wait_to_become_active`] before using it if you will do I/O on it.
904 0 : pub(crate) fn get_attached_tenant_shard(
905 0 : &self,
906 0 : tenant_shard_id: TenantShardId,
907 0 : ) -> Result<Arc<Tenant>, GetTenantError> {
908 0 : let locked = self.tenants.read().unwrap();
909 :
910 0 : let peek_slot = tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Read)?;
911 :
912 0 : match peek_slot {
913 0 : Some(TenantSlot::Attached(tenant)) => Ok(Arc::clone(tenant)),
914 0 : Some(TenantSlot::InProgress(_)) => Err(GetTenantError::NotActive(tenant_shard_id)),
915 : None | Some(TenantSlot::Secondary(_)) => {
916 0 : Err(GetTenantError::NotFound(tenant_shard_id.tenant_id))
917 : }
918 : }
919 0 : }
920 :
921 0 : pub(crate) fn get_secondary_tenant_shard(
922 0 : &self,
923 0 : tenant_shard_id: TenantShardId,
924 0 : ) -> Option<Arc<SecondaryTenant>> {
925 0 : let locked = self.tenants.read().unwrap();
926 0 :
927 0 : let peek_slot = tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Read)
928 0 : .ok()
929 0 : .flatten();
930 :
931 0 : match peek_slot {
932 0 : Some(TenantSlot::Secondary(s)) => Some(s.clone()),
933 0 : _ => None,
934 : }
935 0 : }
936 :
937 : /// Whether the `TenantManager` is responsible for the tenant shard
938 0 : pub(crate) fn manages_tenant_shard(&self, tenant_shard_id: TenantShardId) -> bool {
939 0 : let locked = self.tenants.read().unwrap();
940 0 :
941 0 : let peek_slot = tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Read)
942 0 : .ok()
943 0 : .flatten();
944 0 :
945 0 : peek_slot.is_some()
946 0 : }
947 :
948 0 : #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
949 : pub(crate) async fn upsert_location(
950 : &self,
951 : tenant_shard_id: TenantShardId,
952 : new_location_config: LocationConf,
953 : flush: Option<Duration>,
954 : mut spawn_mode: SpawnMode,
955 : ctx: &RequestContext,
956 : ) -> Result<Option<Arc<Tenant>>, UpsertLocationError> {
957 : debug_assert_current_span_has_tenant_id();
958 : info!("configuring tenant location to state {new_location_config:?}");
959 :
960 : enum FastPathModified {
961 : Attached(Arc<Tenant>),
962 : Secondary(Arc<SecondaryTenant>),
963 : }
964 :
965 : // Special case fast-path for updates to existing slots: if our upsert is only updating configuration,
966 : // then we do not need to set the slot to InProgress, we can just call into the
967 : // existng tenant.
968 : let fast_path_taken = {
969 : let locked = self.tenants.read().unwrap();
970 : let peek_slot =
971 : tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Write)?;
972 : match (&new_location_config.mode, peek_slot) {
973 : (LocationMode::Attached(attach_conf), Some(TenantSlot::Attached(tenant))) => {
974 : match attach_conf.generation.cmp(&tenant.generation) {
975 : Ordering::Equal => {
976 : // A transition from Attached to Attached in the same generation, we may
977 : // take our fast path and just provide the updated configuration
978 : // to the tenant.
979 : tenant.set_new_location_config(
980 : AttachedTenantConf::try_from(new_location_config.clone())
981 : .map_err(UpsertLocationError::BadRequest)?,
982 : );
983 :
984 : Some(FastPathModified::Attached(tenant.clone()))
985 : }
986 : Ordering::Less => {
987 : return Err(UpsertLocationError::BadRequest(anyhow::anyhow!(
988 : "Generation {:?} is less than existing {:?}",
989 : attach_conf.generation,
990 : tenant.generation
991 : )));
992 : }
993 : Ordering::Greater => {
994 : // Generation advanced, fall through to general case of replacing `Tenant` object
995 : None
996 : }
997 : }
998 : }
999 : (
1000 : LocationMode::Secondary(secondary_conf),
1001 : Some(TenantSlot::Secondary(secondary_tenant)),
1002 : ) => {
1003 : secondary_tenant.set_config(secondary_conf);
1004 : secondary_tenant.set_tenant_conf(&new_location_config.tenant_conf);
1005 : Some(FastPathModified::Secondary(secondary_tenant.clone()))
1006 : }
1007 : _ => {
1008 : // Not an Attached->Attached transition, fall through to general case
1009 : None
1010 : }
1011 : }
1012 : };
1013 :
1014 : // Fast-path continued: having dropped out of the self.tenants lock, do the async
1015 : // phase of writing config and/or waiting for flush, before returning.
1016 : match fast_path_taken {
1017 : Some(FastPathModified::Attached(tenant)) => {
1018 : Tenant::persist_tenant_config(self.conf, &tenant_shard_id, &new_location_config)
1019 : .await?;
1020 :
1021 : // Transition to AttachedStale means we may well hold a valid generation
1022 : // still, and have been requested to go stale as part of a migration. If
1023 : // the caller set `flush`, then flush to remote storage.
1024 : if let LocationMode::Attached(AttachedLocationConfig {
1025 : generation: _,
1026 : attach_mode: AttachmentMode::Stale,
1027 : }) = &new_location_config.mode
1028 : {
1029 : if let Some(flush_timeout) = flush {
1030 : match tokio::time::timeout(flush_timeout, tenant.flush_remote()).await {
1031 : Ok(Err(e)) => {
1032 : return Err(UpsertLocationError::Flush(e));
1033 : }
1034 : Ok(Ok(_)) => return Ok(Some(tenant)),
1035 : Err(_) => {
1036 : tracing::warn!(
1037 : timeout_ms = flush_timeout.as_millis(),
1038 : "Timed out waiting for flush to remote storage, proceeding anyway."
1039 : )
1040 : }
1041 : }
1042 : }
1043 : }
1044 :
1045 : return Ok(Some(tenant));
1046 : }
1047 : Some(FastPathModified::Secondary(_secondary_tenant)) => {
1048 : Tenant::persist_tenant_config(self.conf, &tenant_shard_id, &new_location_config)
1049 : .await?;
1050 :
1051 : return Ok(None);
1052 : }
1053 : None => {
1054 : // Proceed with the general case procedure, where we will shutdown & remove any existing
1055 : // slot contents and replace with a fresh one
1056 : }
1057 : };
1058 :
1059 : // General case for upserts to TenantsMap, excluding the case above: we will substitute an
1060 : // InProgress value to the slot while we make whatever changes are required. The state for
1061 : // the tenant is inaccessible to the outside world while we are doing this, but that is sensible:
1062 : // the state is ill-defined while we're in transition. Transitions are async, but fast: we do
1063 : // not do significant I/O, and shutdowns should be prompt via cancellation tokens.
1064 : let mut slot_guard = tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)
1065 0 : .map_err(|e| match e {
1066 : TenantSlotError::AlreadyExists(_, _) | TenantSlotError::NotFound(_) => {
1067 0 : unreachable!("Called with mode Any")
1068 : }
1069 0 : TenantSlotError::InProgress => UpsertLocationError::InProgress,
1070 0 : TenantSlotError::MapState(s) => UpsertLocationError::Unavailable(s),
1071 0 : })?;
1072 :
1073 : match slot_guard.get_old_value() {
1074 : Some(TenantSlot::Attached(tenant)) => {
1075 : // The case where we keep a Tenant alive was covered above in the special case
1076 : // for Attached->Attached transitions in the same generation. By this point,
1077 : // if we see an attached tenant we know it will be discarded and should be
1078 : // shut down.
1079 : let (_guard, progress) = utils::completion::channel();
1080 :
1081 : match tenant.get_attach_mode() {
1082 : AttachmentMode::Single | AttachmentMode::Multi => {
1083 : // Before we leave our state as the presumed holder of the latest generation,
1084 : // flush any outstanding deletions to reduce the risk of leaking objects.
1085 : self.resources.deletion_queue_client.flush_advisory()
1086 : }
1087 : AttachmentMode::Stale => {
1088 : // If we're stale there's not point trying to flush deletions
1089 : }
1090 : };
1091 :
1092 : info!("Shutting down attached tenant");
1093 : match tenant.shutdown(progress, ShutdownMode::Hard).await {
1094 : Ok(()) => {}
1095 : Err(barrier) => {
1096 : info!("Shutdown already in progress, waiting for it to complete");
1097 : barrier.wait().await;
1098 : }
1099 : }
1100 : slot_guard.drop_old_value().expect("We just shut it down");
1101 :
1102 : // Edge case: if we were called with SpawnMode::Create, but a Tenant already existed, then
1103 : // the caller thinks they're creating but the tenant already existed. We must switch to
1104 : // Eager mode so that when starting this Tenant we properly probe remote storage for timelines,
1105 : // rather than assuming it to be empty.
1106 : spawn_mode = SpawnMode::Eager;
1107 : }
1108 : Some(TenantSlot::Secondary(state)) => {
1109 : info!("Shutting down secondary tenant");
1110 : state.shutdown().await;
1111 : }
1112 : Some(TenantSlot::InProgress(_)) => {
1113 : // This should never happen: acquire_slot should error out
1114 : // if the contents of a slot were InProgress.
1115 : return Err(UpsertLocationError::Other(anyhow::anyhow!(
1116 : "Acquired an InProgress slot, this is a bug."
1117 : )));
1118 : }
1119 : None => {
1120 : // Slot was vacant, nothing needs shutting down.
1121 : }
1122 : }
1123 :
1124 : let tenant_path = self.conf.tenant_path(&tenant_shard_id);
1125 : let timelines_path = self.conf.timelines_path(&tenant_shard_id);
1126 :
1127 : // Directory structure is the same for attached and secondary modes:
1128 : // create it if it doesn't exist. Timeline load/creation expects the
1129 : // timelines/ subdir to already exist.
1130 : //
1131 : // Does not need to be fsync'd because local storage is just a cache.
1132 : tokio::fs::create_dir_all(&timelines_path)
1133 : .await
1134 0 : .with_context(|| format!("Creating {timelines_path}"))?;
1135 :
1136 : // Before activating either secondary or attached mode, persist the
1137 : // configuration, so that on restart we will re-attach (or re-start
1138 : // secondary) on the tenant.
1139 : Tenant::persist_tenant_config(self.conf, &tenant_shard_id, &new_location_config).await?;
1140 :
1141 : let new_slot = match &new_location_config.mode {
1142 : LocationMode::Secondary(secondary_config) => {
1143 : let shard_identity = new_location_config.shard;
1144 : TenantSlot::Secondary(SecondaryTenant::new(
1145 : tenant_shard_id,
1146 : shard_identity,
1147 : new_location_config.tenant_conf,
1148 : secondary_config,
1149 : ))
1150 : }
1151 : LocationMode::Attached(_attach_config) => {
1152 : let shard_identity = new_location_config.shard;
1153 :
1154 : // Testing hack: if we are configured with no control plane, then drop the generation
1155 : // from upserts. This enables creating generation-less tenants even though neon_local
1156 : // always uses generations when calling the location conf API.
1157 : let attached_conf = if cfg!(feature = "testing") {
1158 : let mut conf = AttachedTenantConf::try_from(new_location_config)?;
1159 : if self.conf.control_plane_api.is_none() {
1160 : conf.location.generation = Generation::none();
1161 : }
1162 : conf
1163 : } else {
1164 : AttachedTenantConf::try_from(new_location_config)?
1165 : };
1166 :
1167 : let tenant = tenant_spawn(
1168 : self.conf,
1169 : tenant_shard_id,
1170 : &tenant_path,
1171 : self.resources.clone(),
1172 : attached_conf,
1173 : shard_identity,
1174 : None,
1175 : self.tenants,
1176 : spawn_mode,
1177 : ctx,
1178 : )?;
1179 :
1180 : TenantSlot::Attached(tenant)
1181 : }
1182 : };
1183 :
1184 : let attached_tenant = if let TenantSlot::Attached(tenant) = &new_slot {
1185 : Some(tenant.clone())
1186 : } else {
1187 : None
1188 : };
1189 :
1190 : match slot_guard.upsert(new_slot) {
1191 : Err(TenantSlotUpsertError::InternalError(e)) => {
1192 : Err(UpsertLocationError::Other(anyhow::anyhow!(e)))
1193 : }
1194 : Err(TenantSlotUpsertError::MapState(e)) => Err(UpsertLocationError::Unavailable(e)),
1195 : Err(TenantSlotUpsertError::ShuttingDown((new_slot, _completion))) => {
1196 : // If we just called tenant_spawn() on a new tenant, and can't insert it into our map, then
1197 : // we must not leak it: this would violate the invariant that after shutdown_all_tenants, all tenants
1198 : // are shutdown.
1199 : //
1200 : // We must shut it down inline here.
1201 : match new_slot {
1202 : TenantSlot::InProgress(_) => {
1203 : // Unreachable because we never insert an InProgress
1204 : unreachable!()
1205 : }
1206 : TenantSlot::Attached(tenant) => {
1207 : let (_guard, progress) = utils::completion::channel();
1208 : info!("Shutting down just-spawned tenant, because tenant manager is shut down");
1209 : match tenant.shutdown(progress, ShutdownMode::Hard).await {
1210 : Ok(()) => {
1211 : info!("Finished shutting down just-spawned tenant");
1212 : }
1213 : Err(barrier) => {
1214 : info!("Shutdown already in progress, waiting for it to complete");
1215 : barrier.wait().await;
1216 : }
1217 : }
1218 : }
1219 : TenantSlot::Secondary(secondary_tenant) => {
1220 : secondary_tenant.shutdown().await;
1221 : }
1222 : }
1223 :
1224 : Err(UpsertLocationError::Unavailable(
1225 : TenantMapError::ShuttingDown,
1226 : ))
1227 : }
1228 : Ok(()) => Ok(attached_tenant),
1229 : }
1230 : }
1231 :
1232 : /// Resetting a tenant is equivalent to detaching it, then attaching it again with the same
1233 : /// LocationConf that was last used to attach it. Optionally, the local file cache may be
1234 : /// dropped before re-attaching.
1235 : ///
1236 : /// This is not part of a tenant's normal lifecycle: it is used for debug/support, in situations
1237 : /// where an issue is identified that would go away with a restart of the tenant.
1238 : ///
1239 : /// This does not have any special "force" shutdown of a tenant: it relies on the tenant's tasks
1240 : /// to respect the cancellation tokens used in normal shutdown().
1241 0 : #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), %drop_cache))]
1242 : pub(crate) async fn reset_tenant(
1243 : &self,
1244 : tenant_shard_id: TenantShardId,
1245 : drop_cache: bool,
1246 : ctx: &RequestContext,
1247 : ) -> anyhow::Result<()> {
1248 : let mut slot_guard = tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)?;
1249 : let Some(old_slot) = slot_guard.get_old_value() else {
1250 : anyhow::bail!("Tenant not found when trying to reset");
1251 : };
1252 :
1253 : let Some(tenant) = old_slot.get_attached() else {
1254 : slot_guard.revert();
1255 : anyhow::bail!("Tenant is not in attached state");
1256 : };
1257 :
1258 : let (_guard, progress) = utils::completion::channel();
1259 : match tenant.shutdown(progress, ShutdownMode::Hard).await {
1260 : Ok(()) => {
1261 : slot_guard.drop_old_value()?;
1262 : }
1263 : Err(_barrier) => {
1264 : slot_guard.revert();
1265 : anyhow::bail!("Cannot reset Tenant, already shutting down");
1266 : }
1267 : }
1268 :
1269 : let tenant_path = self.conf.tenant_path(&tenant_shard_id);
1270 : let timelines_path = self.conf.timelines_path(&tenant_shard_id);
1271 : let config = Tenant::load_tenant_config(self.conf, &tenant_shard_id)?;
1272 :
1273 : if drop_cache {
1274 : tracing::info!("Dropping local file cache");
1275 :
1276 : match tokio::fs::read_dir(&timelines_path).await {
1277 : Err(e) => {
1278 : tracing::warn!("Failed to list timelines while dropping cache: {}", e);
1279 : }
1280 : Ok(mut entries) => {
1281 : while let Some(entry) = entries.next_entry().await? {
1282 : tokio::fs::remove_dir_all(entry.path()).await?;
1283 : }
1284 : }
1285 : }
1286 : }
1287 :
1288 : let shard_identity = config.shard;
1289 : let tenant = tenant_spawn(
1290 : self.conf,
1291 : tenant_shard_id,
1292 : &tenant_path,
1293 : self.resources.clone(),
1294 : AttachedTenantConf::try_from(config)?,
1295 : shard_identity,
1296 : None,
1297 : self.tenants,
1298 : SpawnMode::Eager,
1299 : ctx,
1300 : )?;
1301 :
1302 : slot_guard.upsert(TenantSlot::Attached(tenant))?;
1303 :
1304 : Ok(())
1305 : }
1306 :
1307 0 : pub(crate) fn get_attached_active_tenant_shards(&self) -> Vec<Arc<Tenant>> {
1308 0 : let locked = self.tenants.read().unwrap();
1309 0 : match &*locked {
1310 0 : TenantsMap::Initializing => Vec::new(),
1311 0 : TenantsMap::Open(map) | TenantsMap::ShuttingDown(map) => map
1312 0 : .values()
1313 0 : .filter_map(|slot| {
1314 0 : slot.get_attached()
1315 0 : .and_then(|t| if t.is_active() { Some(t.clone()) } else { None })
1316 0 : })
1317 0 : .collect(),
1318 : }
1319 0 : }
1320 : // Do some synchronous work for all tenant slots in Secondary state. The provided
1321 : // callback should be small and fast, as it will be called inside the global
1322 : // TenantsMap lock.
1323 0 : pub(crate) fn foreach_secondary_tenants<F>(&self, mut func: F)
1324 0 : where
1325 0 : // TODO: let the callback return a hint to drop out of the loop early
1326 0 : F: FnMut(&TenantShardId, &Arc<SecondaryTenant>),
1327 0 : {
1328 0 : let locked = self.tenants.read().unwrap();
1329 :
1330 0 : let map = match &*locked {
1331 0 : TenantsMap::Initializing | TenantsMap::ShuttingDown(_) => return,
1332 0 : TenantsMap::Open(m) => m,
1333 : };
1334 :
1335 0 : for (tenant_id, slot) in map {
1336 0 : if let TenantSlot::Secondary(state) = slot {
1337 : // Only expose secondary tenants that are not currently shutting down
1338 0 : if !state.cancel.is_cancelled() {
1339 0 : func(tenant_id, state)
1340 0 : }
1341 0 : }
1342 : }
1343 0 : }
1344 :
1345 : /// Total list of all tenant slots: this includes attached, secondary, and InProgress.
1346 0 : pub(crate) fn list(&self) -> Vec<(TenantShardId, TenantSlot)> {
1347 0 : let locked = self.tenants.read().unwrap();
1348 0 : match &*locked {
1349 0 : TenantsMap::Initializing => Vec::new(),
1350 0 : TenantsMap::Open(map) | TenantsMap::ShuttingDown(map) => {
1351 0 : map.iter().map(|(k, v)| (*k, v.clone())).collect()
1352 : }
1353 : }
1354 0 : }
1355 :
1356 0 : pub(crate) fn get(&self, tenant_shard_id: TenantShardId) -> Option<TenantSlot> {
1357 0 : let locked = self.tenants.read().unwrap();
1358 0 : match &*locked {
1359 0 : TenantsMap::Initializing => None,
1360 0 : TenantsMap::Open(map) | TenantsMap::ShuttingDown(map) => {
1361 0 : map.get(&tenant_shard_id).cloned()
1362 : }
1363 : }
1364 0 : }
1365 :
1366 0 : pub(crate) async fn delete_tenant(
1367 0 : &self,
1368 0 : tenant_shard_id: TenantShardId,
1369 0 : activation_timeout: Duration,
1370 0 : ) -> Result<(), DeleteTenantError> {
1371 0 : super::span::debug_assert_current_span_has_tenant_id();
1372 : // We acquire a SlotGuard during this function to protect against concurrent
1373 : // changes while the ::prepare phase of DeleteTenantFlow executes, but then
1374 : // have to return the Tenant to the map while the background deletion runs.
1375 : //
1376 : // TODO: refactor deletion to happen outside the lifetime of a Tenant.
1377 : // Currently, deletion requires a reference to the tenants map in order to
1378 : // keep the Tenant in the map until deletion is complete, and then remove
1379 : // it at the end.
1380 : //
1381 : // See https://github.com/neondatabase/neon/issues/5080
1382 :
1383 0 : let slot_guard =
1384 0 : tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::MustExist)?;
1385 :
1386 : // unwrap is safe because we used MustExist mode when acquiring
1387 0 : let tenant = match slot_guard.get_old_value().as_ref().unwrap() {
1388 0 : TenantSlot::Attached(tenant) => tenant.clone(),
1389 : _ => {
1390 : // Express "not attached" as equivalent to "not found"
1391 0 : return Err(DeleteTenantError::NotAttached);
1392 : }
1393 : };
1394 :
1395 0 : match tenant.current_state() {
1396 : TenantState::Broken { .. } | TenantState::Stopping { .. } => {
1397 : // If deletion is already in progress, return success (the semantics of this
1398 : // function are to rerturn success afterr deletion is spawned in background).
1399 : // Otherwise fall through and let [`DeleteTenantFlow`] handle this state.
1400 0 : if DeleteTenantFlow::is_in_progress(&tenant) {
1401 : // The `delete_progress` lock is held: deletion is already happening
1402 : // in the bacckground
1403 0 : slot_guard.revert();
1404 0 : return Ok(());
1405 0 : }
1406 : }
1407 : _ => {
1408 0 : tenant
1409 0 : .wait_to_become_active(activation_timeout)
1410 0 : .await
1411 0 : .map_err(|e| match e {
1412 : GetActiveTenantError::WillNotBecomeActive(_)
1413 : | GetActiveTenantError::Broken(_) => {
1414 0 : DeleteTenantError::InvalidState(tenant.current_state())
1415 : }
1416 0 : GetActiveTenantError::Cancelled => DeleteTenantError::Cancelled,
1417 0 : GetActiveTenantError::NotFound(_) => DeleteTenantError::NotAttached,
1418 : GetActiveTenantError::WaitForActiveTimeout {
1419 0 : latest_state: _latest_state,
1420 0 : wait_time: _wait_time,
1421 0 : } => DeleteTenantError::InvalidState(tenant.current_state()),
1422 0 : })?;
1423 : }
1424 : }
1425 :
1426 0 : let result = DeleteTenantFlow::run(
1427 0 : self.conf,
1428 0 : self.resources.remote_storage.clone(),
1429 0 : &TENANTS,
1430 0 : tenant,
1431 0 : &self.cancel,
1432 0 : )
1433 0 : .await;
1434 :
1435 : // The Tenant goes back into the map in Stopping state, it will eventually be removed by DeleteTenantFLow
1436 0 : slot_guard.revert();
1437 0 : result
1438 0 : }
1439 :
1440 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()))]
1441 : pub(crate) async fn shard_split(
1442 : &self,
1443 : tenant: Arc<Tenant>,
1444 : new_shard_count: ShardCount,
1445 : new_stripe_size: Option<ShardStripeSize>,
1446 : ctx: &RequestContext,
1447 : ) -> anyhow::Result<Vec<TenantShardId>> {
1448 : let tenant_shard_id = *tenant.get_tenant_shard_id();
1449 : let r = self
1450 : .do_shard_split(tenant, new_shard_count, new_stripe_size, ctx)
1451 : .await;
1452 : if r.is_err() {
1453 : // Shard splitting might have left the original shard in a partially shut down state (it
1454 : // stops the shard's remote timeline client). Reset it to ensure we leave things in
1455 : // a working state.
1456 : if self.get(tenant_shard_id).is_some() {
1457 : tracing::warn!("Resetting after shard split failure");
1458 : if let Err(e) = self.reset_tenant(tenant_shard_id, false, ctx).await {
1459 : // Log this error because our return value will still be the original error, not this one. This is
1460 : // a severe error: if this happens, we might be leaving behind a tenant that is not fully functional
1461 : // (e.g. has uploads disabled). We can't do anything else: if reset fails then shutting the tenant down or
1462 : // setting it broken probably won't help either.
1463 : tracing::error!("Failed to reset: {e}");
1464 : }
1465 : }
1466 : }
1467 :
1468 : r
1469 : }
1470 :
1471 0 : pub(crate) async fn do_shard_split(
1472 0 : &self,
1473 0 : tenant: Arc<Tenant>,
1474 0 : new_shard_count: ShardCount,
1475 0 : new_stripe_size: Option<ShardStripeSize>,
1476 0 : ctx: &RequestContext,
1477 0 : ) -> anyhow::Result<Vec<TenantShardId>> {
1478 0 : let tenant_shard_id = *tenant.get_tenant_shard_id();
1479 0 :
1480 0 : // Validate the incoming request
1481 0 : if new_shard_count.count() <= tenant_shard_id.shard_count.count() {
1482 0 : anyhow::bail!("Requested shard count is not an increase");
1483 0 : }
1484 0 : let expansion_factor = new_shard_count.count() / tenant_shard_id.shard_count.count();
1485 0 : if !expansion_factor.is_power_of_two() {
1486 0 : anyhow::bail!("Requested split is not a power of two");
1487 0 : }
1488 :
1489 0 : if let Some(new_stripe_size) = new_stripe_size {
1490 0 : if tenant.get_shard_stripe_size() != new_stripe_size
1491 0 : && tenant_shard_id.shard_count.count() > 1
1492 : {
1493 : // This tenant already has multiple shards, it is illegal to try and change its stripe size
1494 0 : anyhow::bail!(
1495 0 : "Shard stripe size may not be modified once tenant has multiple shards"
1496 0 : );
1497 0 : }
1498 0 : }
1499 :
1500 : // Plan: identify what the new child shards will be
1501 0 : let child_shards = tenant_shard_id.split(new_shard_count);
1502 0 : tracing::info!(
1503 0 : "Shard {} splits into: {}",
1504 0 : tenant_shard_id.to_index(),
1505 0 : child_shards
1506 0 : .iter()
1507 0 : .map(|id| format!("{}", id.to_index()))
1508 0 : .join(",")
1509 : );
1510 :
1511 0 : fail::fail_point!("shard-split-pre-prepare", |_| Err(anyhow::anyhow!(
1512 0 : "failpoint"
1513 0 : )));
1514 :
1515 0 : let parent_shard_identity = tenant.shard_identity;
1516 0 : let parent_tenant_conf = tenant.get_tenant_conf();
1517 0 : let parent_generation = tenant.generation;
1518 :
1519 : // Phase 1: Write out child shards' remote index files, in the parent tenant's current generation
1520 0 : if let Err(e) = tenant.split_prepare(&child_shards).await {
1521 : // If [`Tenant::split_prepare`] fails, we must reload the tenant, because it might
1522 : // have been left in a partially-shut-down state.
1523 0 : tracing::warn!("Failed to prepare for split: {e}, reloading Tenant before returning");
1524 0 : return Err(e);
1525 0 : }
1526 0 :
1527 0 : fail::fail_point!("shard-split-post-prepare", |_| Err(anyhow::anyhow!(
1528 0 : "failpoint"
1529 0 : )));
1530 :
1531 0 : self.resources.deletion_queue_client.flush_advisory();
1532 0 :
1533 0 : // Phase 2: Put the parent shard to InProgress and grab a reference to the parent Tenant
1534 0 : drop(tenant);
1535 0 : let mut parent_slot_guard =
1536 0 : tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)?;
1537 0 : let parent = match parent_slot_guard.get_old_value() {
1538 0 : Some(TenantSlot::Attached(t)) => t,
1539 0 : Some(TenantSlot::Secondary(_)) => anyhow::bail!("Tenant location in secondary mode"),
1540 : Some(TenantSlot::InProgress(_)) => {
1541 : // tenant_map_acquire_slot never returns InProgress, if a slot was InProgress
1542 : // it would return an error.
1543 0 : unreachable!()
1544 : }
1545 : None => {
1546 : // We don't actually need the parent shard to still be attached to do our work, but it's
1547 : // a weird enough situation that the caller probably didn't want us to continue working
1548 : // if they had detached the tenant they requested the split on.
1549 0 : anyhow::bail!("Detached parent shard in the middle of split!")
1550 : }
1551 : };
1552 0 : fail::fail_point!("shard-split-pre-hardlink", |_| Err(anyhow::anyhow!(
1553 0 : "failpoint"
1554 0 : )));
1555 : // Optimization: hardlink layers from the parent into the children, so that they don't have to
1556 : // re-download & duplicate the data referenced in their initial IndexPart
1557 0 : self.shard_split_hardlink(parent, child_shards.clone())
1558 0 : .await?;
1559 0 : fail::fail_point!("shard-split-post-hardlink", |_| Err(anyhow::anyhow!(
1560 0 : "failpoint"
1561 0 : )));
1562 :
1563 : // Take a snapshot of where the parent's WAL ingest had got to: we will wait for
1564 : // child shards to reach this point.
1565 0 : let mut target_lsns = HashMap::new();
1566 0 : for timeline in parent.timelines.lock().unwrap().clone().values() {
1567 0 : target_lsns.insert(timeline.timeline_id, timeline.get_last_record_lsn());
1568 0 : }
1569 :
1570 : // TODO: we should have the parent shard stop its WAL ingest here, it's a waste of resources
1571 : // and could slow down the children trying to catch up.
1572 :
1573 : // Phase 3: Spawn the child shards
1574 0 : for child_shard in &child_shards {
1575 0 : let mut child_shard_identity = parent_shard_identity;
1576 0 : if let Some(new_stripe_size) = new_stripe_size {
1577 0 : child_shard_identity.stripe_size = new_stripe_size;
1578 0 : }
1579 0 : child_shard_identity.count = child_shard.shard_count;
1580 0 : child_shard_identity.number = child_shard.shard_number;
1581 0 :
1582 0 : let child_location_conf = LocationConf {
1583 0 : mode: LocationMode::Attached(AttachedLocationConfig {
1584 0 : generation: parent_generation,
1585 0 : attach_mode: AttachmentMode::Single,
1586 0 : }),
1587 0 : shard: child_shard_identity,
1588 0 : tenant_conf: parent_tenant_conf.clone(),
1589 0 : };
1590 0 :
1591 0 : self.upsert_location(
1592 0 : *child_shard,
1593 0 : child_location_conf,
1594 0 : None,
1595 0 : SpawnMode::Eager,
1596 0 : ctx,
1597 0 : )
1598 0 : .await?;
1599 : }
1600 :
1601 0 : fail::fail_point!("shard-split-post-child-conf", |_| Err(anyhow::anyhow!(
1602 0 : "failpoint"
1603 0 : )));
1604 :
1605 : // Phase 4: wait for child chards WAL ingest to catch up to target LSN
1606 0 : for child_shard_id in &child_shards {
1607 0 : let child_shard_id = *child_shard_id;
1608 0 : let child_shard = {
1609 0 : let locked = TENANTS.read().unwrap();
1610 0 : let peek_slot =
1611 0 : tenant_map_peek_slot(&locked, &child_shard_id, TenantSlotPeekMode::Read)?;
1612 0 : peek_slot.and_then(|s| s.get_attached()).cloned()
1613 : };
1614 0 : if let Some(t) = child_shard {
1615 : // Wait for the child shard to become active: this should be very quick because it only
1616 : // has to download the index_part that we just uploaded when creating it.
1617 0 : if let Err(e) = t.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await {
1618 : // This is not fatal: we have durably created the child shard. It just makes the
1619 : // split operation less seamless for clients, as we will may detach the parent
1620 : // shard before the child shards are fully ready to serve requests.
1621 0 : tracing::warn!("Failed to wait for shard {child_shard_id} to activate: {e}");
1622 0 : continue;
1623 0 : }
1624 0 :
1625 0 : let timelines = t.timelines.lock().unwrap().clone();
1626 0 : for timeline in timelines.values() {
1627 0 : let Some(target_lsn) = target_lsns.get(&timeline.timeline_id) else {
1628 0 : continue;
1629 : };
1630 :
1631 0 : tracing::info!(
1632 0 : "Waiting for child shard {}/{} to reach target lsn {}...",
1633 0 : child_shard_id,
1634 0 : timeline.timeline_id,
1635 : target_lsn
1636 : );
1637 :
1638 0 : fail::fail_point!("shard-split-lsn-wait", |_| Err(anyhow::anyhow!(
1639 0 : "failpoint"
1640 0 : )));
1641 0 : if let Err(e) = timeline
1642 0 : .wait_lsn(
1643 0 : *target_lsn,
1644 0 : crate::tenant::timeline::WaitLsnWaiter::Tenant,
1645 0 : ctx,
1646 0 : )
1647 0 : .await
1648 : {
1649 : // Failure here might mean shutdown, in any case this part is an optimization
1650 : // and we shouldn't hold up the split operation.
1651 0 : tracing::warn!(
1652 0 : "Failed to wait for timeline {} to reach lsn {target_lsn}: {e}",
1653 0 : timeline.timeline_id
1654 : );
1655 : } else {
1656 0 : tracing::info!(
1657 0 : "Child shard {}/{} reached target lsn {}",
1658 0 : child_shard_id,
1659 0 : timeline.timeline_id,
1660 : target_lsn
1661 : );
1662 : }
1663 : }
1664 0 : }
1665 : }
1666 :
1667 : // Phase 5: Shut down the parent shard, and erase it from disk
1668 0 : let (_guard, progress) = completion::channel();
1669 0 : match parent.shutdown(progress, ShutdownMode::Hard).await {
1670 0 : Ok(()) => {}
1671 0 : Err(other) => {
1672 0 : other.wait().await;
1673 : }
1674 : }
1675 0 : let local_tenant_directory = self.conf.tenant_path(&tenant_shard_id);
1676 0 : let tmp_path = safe_rename_tenant_dir(&local_tenant_directory)
1677 0 : .await
1678 0 : .with_context(|| format!("local tenant directory {local_tenant_directory:?} rename"))?;
1679 0 : spawn_background_purge(tmp_path);
1680 0 :
1681 0 : fail::fail_point!("shard-split-pre-finish", |_| Err(anyhow::anyhow!(
1682 0 : "failpoint"
1683 0 : )));
1684 :
1685 0 : parent_slot_guard.drop_old_value()?;
1686 :
1687 : // Phase 6: Release the InProgress on the parent shard
1688 0 : drop(parent_slot_guard);
1689 0 :
1690 0 : Ok(child_shards)
1691 0 : }
1692 :
1693 : /// Part of [`Self::shard_split`]: hard link parent shard layers into child shards, as an optimization
1694 : /// to avoid the children downloading them again.
1695 : ///
1696 : /// For each resident layer in the parent shard, we will hard link it into all of the child shards.
1697 0 : async fn shard_split_hardlink(
1698 0 : &self,
1699 0 : parent_shard: &Tenant,
1700 0 : child_shards: Vec<TenantShardId>,
1701 0 : ) -> anyhow::Result<()> {
1702 0 : debug_assert_current_span_has_tenant_id();
1703 0 :
1704 0 : let parent_path = self.conf.tenant_path(parent_shard.get_tenant_shard_id());
1705 0 : let (parent_timelines, parent_layers) = {
1706 0 : let mut parent_layers = Vec::new();
1707 0 : let timelines = parent_shard.timelines.lock().unwrap().clone();
1708 0 : let parent_timelines = timelines.keys().cloned().collect::<Vec<_>>();
1709 0 : for timeline in timelines.values() {
1710 0 : let timeline_layers = timeline
1711 0 : .layers
1712 0 : .read()
1713 0 : .await
1714 0 : .likely_resident_layers()
1715 0 : .collect::<Vec<_>>();
1716 :
1717 0 : for layer in timeline_layers {
1718 0 : let relative_path = layer
1719 0 : .local_path()
1720 0 : .strip_prefix(&parent_path)
1721 0 : .context("Removing prefix from parent layer path")?;
1722 0 : parent_layers.push(relative_path.to_owned());
1723 : }
1724 : }
1725 0 : debug_assert!(
1726 0 : !parent_layers.is_empty(),
1727 0 : "shutdown cannot empty the layermap"
1728 : );
1729 0 : (parent_timelines, parent_layers)
1730 0 : };
1731 0 :
1732 0 : let mut child_prefixes = Vec::new();
1733 0 : let mut create_dirs = Vec::new();
1734 :
1735 0 : for child in child_shards {
1736 0 : let child_prefix = self.conf.tenant_path(&child);
1737 0 : create_dirs.push(child_prefix.clone());
1738 0 : create_dirs.extend(
1739 0 : parent_timelines
1740 0 : .iter()
1741 0 : .map(|t| self.conf.timeline_path(&child, t)),
1742 0 : );
1743 0 :
1744 0 : child_prefixes.push(child_prefix);
1745 0 : }
1746 :
1747 : // Since we will do a large number of small filesystem metadata operations, batch them into
1748 : // spawn_blocking calls rather than doing each one as a tokio::fs round-trip.
1749 0 : let jh = tokio::task::spawn_blocking(move || -> anyhow::Result<usize> {
1750 0 : for dir in &create_dirs {
1751 0 : if let Err(e) = std::fs::create_dir_all(dir) {
1752 : // Ignore AlreadyExists errors, drop out on all other errors
1753 0 : match e.kind() {
1754 0 : std::io::ErrorKind::AlreadyExists => {}
1755 : _ => {
1756 0 : return Err(anyhow::anyhow!(e).context(format!("Creating {dir}")));
1757 : }
1758 : }
1759 0 : }
1760 : }
1761 :
1762 0 : for child_prefix in child_prefixes {
1763 0 : for relative_layer in &parent_layers {
1764 0 : let parent_path = parent_path.join(relative_layer);
1765 0 : let child_path = child_prefix.join(relative_layer);
1766 0 : if let Err(e) = std::fs::hard_link(&parent_path, &child_path) {
1767 0 : match e.kind() {
1768 0 : std::io::ErrorKind::AlreadyExists => {}
1769 : std::io::ErrorKind::NotFound => {
1770 0 : tracing::info!(
1771 0 : "Layer {} not found during hard-linking, evicted during split?",
1772 : relative_layer
1773 : );
1774 : }
1775 : _ => {
1776 0 : return Err(anyhow::anyhow!(e).context(format!(
1777 0 : "Hard linking {relative_layer} into {child_prefix}"
1778 0 : )))
1779 : }
1780 : }
1781 0 : }
1782 : }
1783 : }
1784 :
1785 : // Durability is not required for correctness, but if we crashed during split and
1786 : // then came restarted with empty timeline dirs, it would be very inefficient to
1787 : // re-populate from remote storage.
1788 0 : for dir in create_dirs {
1789 0 : if let Err(e) = crashsafe::fsync(&dir) {
1790 : // Something removed a newly created timeline dir out from underneath us? Extremely
1791 : // unexpected, but not worth panic'ing over as this whole function is just an
1792 : // optimization.
1793 0 : tracing::warn!("Failed to fsync directory {dir}: {e}")
1794 0 : }
1795 : }
1796 :
1797 0 : Ok(parent_layers.len())
1798 0 : });
1799 0 :
1800 0 : match jh.await {
1801 0 : Ok(Ok(layer_count)) => {
1802 0 : tracing::info!(count = layer_count, "Hard linked layers into child shards");
1803 : }
1804 0 : Ok(Err(e)) => {
1805 0 : // This is an optimization, so we tolerate failure.
1806 0 : tracing::warn!("Error hard-linking layers, proceeding anyway: {e}")
1807 : }
1808 0 : Err(e) => {
1809 0 : // This is something totally unexpected like a panic, so bail out.
1810 0 : anyhow::bail!("Error joining hard linking task: {e}");
1811 : }
1812 : }
1813 :
1814 0 : Ok(())
1815 0 : }
1816 :
1817 : ///
1818 : /// Shut down all tenants. This runs as part of pageserver shutdown.
1819 : ///
1820 : /// NB: We leave the tenants in the map, so that they remain accessible through
1821 : /// the management API until we shut it down. If we removed the shut-down tenants
1822 : /// from the tenants map, the management API would return 404 for these tenants,
1823 : /// because TenantsMap::get() now returns `None`.
1824 : /// That could be easily misinterpreted by control plane, the consumer of the
1825 : /// management API. For example, it could attach the tenant on a different pageserver.
1826 : /// We would then be in split-brain once this pageserver restarts.
1827 0 : #[instrument(skip_all)]
1828 : pub(crate) async fn shutdown(&self) {
1829 : self.cancel.cancel();
1830 :
1831 : shutdown_all_tenants0(self.tenants).await
1832 : }
1833 :
1834 0 : pub(crate) async fn detach_tenant(
1835 0 : &self,
1836 0 : conf: &'static PageServerConf,
1837 0 : tenant_shard_id: TenantShardId,
1838 0 : detach_ignored: bool,
1839 0 : deletion_queue_client: &DeletionQueueClient,
1840 0 : ) -> Result<(), TenantStateError> {
1841 0 : let tmp_path = self
1842 0 : .detach_tenant0(
1843 0 : conf,
1844 0 : &TENANTS,
1845 0 : tenant_shard_id,
1846 0 : detach_ignored,
1847 0 : deletion_queue_client,
1848 0 : )
1849 0 : .await?;
1850 0 : spawn_background_purge(tmp_path);
1851 0 :
1852 0 : Ok(())
1853 0 : }
1854 :
1855 0 : async fn detach_tenant0(
1856 0 : &self,
1857 0 : conf: &'static PageServerConf,
1858 0 : tenants: &std::sync::RwLock<TenantsMap>,
1859 0 : tenant_shard_id: TenantShardId,
1860 0 : detach_ignored: bool,
1861 0 : deletion_queue_client: &DeletionQueueClient,
1862 0 : ) -> Result<Utf8PathBuf, TenantStateError> {
1863 0 : let tenant_dir_rename_operation = |tenant_id_to_clean: TenantShardId| async move {
1864 0 : let local_tenant_directory = conf.tenant_path(&tenant_id_to_clean);
1865 0 : safe_rename_tenant_dir(&local_tenant_directory)
1866 0 : .await
1867 0 : .with_context(|| {
1868 0 : format!("local tenant directory {local_tenant_directory:?} rename")
1869 0 : })
1870 0 : };
1871 :
1872 0 : let removal_result = remove_tenant_from_memory(
1873 0 : tenants,
1874 0 : tenant_shard_id,
1875 0 : tenant_dir_rename_operation(tenant_shard_id),
1876 0 : )
1877 0 : .await;
1878 :
1879 : // Flush pending deletions, so that they have a good chance of passing validation
1880 : // before this tenant is potentially re-attached elsewhere.
1881 0 : deletion_queue_client.flush_advisory();
1882 0 :
1883 0 : // Ignored tenants are not present in memory and will bail the removal from memory operation.
1884 0 : // Before returning the error, check for ignored tenant removal case — we only need to clean its local files then.
1885 0 : if detach_ignored
1886 0 : && matches!(
1887 0 : removal_result,
1888 : Err(TenantStateError::SlotError(TenantSlotError::NotFound(_)))
1889 : )
1890 : {
1891 0 : let tenant_ignore_mark = conf.tenant_ignore_mark_file_path(&tenant_shard_id);
1892 0 : if tenant_ignore_mark.exists() {
1893 0 : info!("Detaching an ignored tenant");
1894 0 : let tmp_path = tenant_dir_rename_operation(tenant_shard_id)
1895 0 : .await
1896 0 : .with_context(|| {
1897 0 : format!("Ignored tenant {tenant_shard_id} local directory rename")
1898 0 : })?;
1899 0 : return Ok(tmp_path);
1900 0 : }
1901 0 : }
1902 :
1903 0 : removal_result
1904 0 : }
1905 :
1906 0 : pub(crate) fn list_tenants(
1907 0 : &self,
1908 0 : ) -> Result<Vec<(TenantShardId, TenantState, Generation)>, TenantMapListError> {
1909 0 : let tenants = TENANTS.read().unwrap();
1910 0 : let m = match &*tenants {
1911 0 : TenantsMap::Initializing => return Err(TenantMapListError::Initializing),
1912 0 : TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => m,
1913 0 : };
1914 0 : Ok(m.iter()
1915 0 : .filter_map(|(id, tenant)| match tenant {
1916 0 : TenantSlot::Attached(tenant) => {
1917 0 : Some((*id, tenant.current_state(), tenant.generation()))
1918 : }
1919 0 : TenantSlot::Secondary(_) => None,
1920 0 : TenantSlot::InProgress(_) => None,
1921 0 : })
1922 0 : .collect())
1923 0 : }
1924 :
1925 : /// Completes an earlier prepared timeline detach ancestor.
1926 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
1927 0 : &self,
1928 0 : tenant_shard_id: TenantShardId,
1929 0 : timeline_id: TimelineId,
1930 0 : prepared: PreparedTimelineDetach,
1931 0 : ctx: &RequestContext,
1932 0 : ) -> Result<Vec<TimelineId>, anyhow::Error> {
1933 : struct RevertOnDropSlot(Option<SlotGuard>);
1934 :
1935 : impl Drop for RevertOnDropSlot {
1936 0 : fn drop(&mut self) {
1937 0 : if let Some(taken) = self.0.take() {
1938 0 : taken.revert();
1939 0 : }
1940 0 : }
1941 : }
1942 :
1943 : impl RevertOnDropSlot {
1944 0 : fn into_inner(mut self) -> SlotGuard {
1945 0 : self.0.take().unwrap()
1946 0 : }
1947 : }
1948 :
1949 : impl std::ops::Deref for RevertOnDropSlot {
1950 : type Target = SlotGuard;
1951 :
1952 0 : fn deref(&self) -> &Self::Target {
1953 0 : self.0.as_ref().unwrap()
1954 0 : }
1955 : }
1956 :
1957 0 : let slot_guard = tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)?;
1958 0 : let slot_guard = RevertOnDropSlot(Some(slot_guard));
1959 :
1960 0 : let tenant = {
1961 0 : let Some(old_slot) = slot_guard.get_old_value() else {
1962 0 : anyhow::bail!(
1963 0 : "Tenant not found when trying to complete detaching timeline ancestor"
1964 0 : );
1965 : };
1966 :
1967 0 : let Some(tenant) = old_slot.get_attached() else {
1968 0 : anyhow::bail!("Tenant is not in attached state");
1969 : };
1970 :
1971 0 : if !tenant.is_active() {
1972 0 : anyhow::bail!("Tenant is not active");
1973 0 : }
1974 0 :
1975 0 : tenant.clone()
1976 : };
1977 :
1978 0 : let timeline = tenant.get_timeline(timeline_id, true)?;
1979 :
1980 0 : let reparented = timeline
1981 0 : .complete_detaching_timeline_ancestor(&tenant, prepared, ctx)
1982 0 : .await?;
1983 :
1984 0 : let mut slot_guard = slot_guard.into_inner();
1985 0 :
1986 0 : let (_guard, progress) = utils::completion::channel();
1987 0 : match tenant.shutdown(progress, ShutdownMode::Hard).await {
1988 : Ok(()) => {
1989 0 : slot_guard.drop_old_value()?;
1990 : }
1991 0 : Err(_barrier) => {
1992 0 : slot_guard.revert();
1993 0 : // this really should not happen, at all, unless shutdown was already going?
1994 0 : anyhow::bail!("Cannot restart Tenant, already shutting down");
1995 : }
1996 : }
1997 :
1998 0 : let tenant_path = self.conf.tenant_path(&tenant_shard_id);
1999 0 : let config = Tenant::load_tenant_config(self.conf, &tenant_shard_id)?;
2000 :
2001 0 : let shard_identity = config.shard;
2002 0 : let tenant = tenant_spawn(
2003 0 : self.conf,
2004 0 : tenant_shard_id,
2005 0 : &tenant_path,
2006 0 : self.resources.clone(),
2007 0 : AttachedTenantConf::try_from(config)?,
2008 0 : shard_identity,
2009 0 : None,
2010 0 : self.tenants,
2011 0 : SpawnMode::Eager,
2012 0 : ctx,
2013 0 : )?;
2014 :
2015 0 : slot_guard.upsert(TenantSlot::Attached(tenant))?;
2016 :
2017 0 : Ok(reparented)
2018 0 : }
2019 :
2020 : /// A page service client sends a TenantId, and to look up the correct Tenant we must
2021 : /// resolve this to a fully qualified TenantShardId.
2022 : ///
2023 : /// During shard splits: we shall see parent shards in InProgress state and skip them, and
2024 : /// instead match on child shards which should appear in Attached state. Very early in a shard
2025 : /// split, or in other cases where a shard is InProgress, we will return our own InProgress result
2026 : /// to instruct the caller to wait for that to finish before querying again.
2027 0 : pub(crate) fn resolve_attached_shard(
2028 0 : &self,
2029 0 : tenant_id: &TenantId,
2030 0 : selector: ShardSelector,
2031 0 : ) -> ShardResolveResult {
2032 0 : let tenants = self.tenants.read().unwrap();
2033 0 : let mut want_shard = None;
2034 0 : let mut any_in_progress = None;
2035 0 :
2036 0 : match &*tenants {
2037 0 : TenantsMap::Initializing => ShardResolveResult::NotFound,
2038 0 : TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => {
2039 0 : for slot in m.range(TenantShardId::tenant_range(*tenant_id)) {
2040 : // Ignore all slots that don't contain an attached tenant
2041 0 : let tenant = match &slot.1 {
2042 0 : TenantSlot::Attached(t) => t,
2043 0 : TenantSlot::InProgress(barrier) => {
2044 0 : // We might still find a usable shard, but in case we don't, remember that
2045 0 : // we saw at least one InProgress slot, so that we can distinguish this case
2046 0 : // from a simple NotFound in our return value.
2047 0 : any_in_progress = Some(barrier.clone());
2048 0 : continue;
2049 : }
2050 0 : _ => continue,
2051 : };
2052 :
2053 0 : match selector {
2054 0 : ShardSelector::First => return ShardResolveResult::Found(tenant.clone()),
2055 0 : ShardSelector::Zero if slot.0.shard_number == ShardNumber(0) => {
2056 0 : return ShardResolveResult::Found(tenant.clone())
2057 : }
2058 0 : ShardSelector::Page(key) => {
2059 0 : // First slot we see for this tenant, calculate the expected shard number
2060 0 : // for the key: we will use this for checking if this and subsequent
2061 0 : // slots contain the key, rather than recalculating the hash each time.
2062 0 : if want_shard.is_none() {
2063 0 : want_shard = Some(tenant.shard_identity.get_shard_number(&key));
2064 0 : }
2065 :
2066 0 : if Some(tenant.shard_identity.number) == want_shard {
2067 0 : return ShardResolveResult::Found(tenant.clone());
2068 0 : }
2069 : }
2070 0 : _ => continue,
2071 : }
2072 : }
2073 :
2074 : // Fall through: we didn't find a slot that was in Attached state & matched our selector. If
2075 : // we found one or more InProgress slot, indicate to caller that they should retry later. Otherwise
2076 : // this requested shard simply isn't found.
2077 0 : if let Some(barrier) = any_in_progress {
2078 0 : ShardResolveResult::InProgress(barrier)
2079 : } else {
2080 0 : ShardResolveResult::NotFound
2081 : }
2082 : }
2083 : }
2084 0 : }
2085 : }
2086 :
2087 0 : #[derive(Debug, thiserror::Error)]
2088 : pub(crate) enum GetTenantError {
2089 : /// NotFound is a TenantId rather than TenantShardId, because this error type is used from
2090 : /// getters that use a TenantId and a ShardSelector, not just getters that target a specific shard.
2091 : #[error("Tenant {0} not found")]
2092 : NotFound(TenantId),
2093 :
2094 : #[error("Tenant {0} is not active")]
2095 : NotActive(TenantShardId),
2096 :
2097 : // Initializing or shutting down: cannot authoritatively say whether we have this tenant
2098 : #[error("Tenant map is not available: {0}")]
2099 : MapState(#[from] TenantMapError),
2100 : }
2101 :
2102 0 : #[derive(thiserror::Error, Debug)]
2103 : pub(crate) enum GetActiveTenantError {
2104 : /// We may time out either while TenantSlot is InProgress, or while the Tenant
2105 : /// is in a non-Active state
2106 : #[error(
2107 : "Timed out waiting {wait_time:?} for tenant active state. Latest state: {latest_state:?}"
2108 : )]
2109 : WaitForActiveTimeout {
2110 : latest_state: Option<TenantState>,
2111 : wait_time: Duration,
2112 : },
2113 :
2114 : /// The TenantSlot is absent, or in secondary mode
2115 : #[error(transparent)]
2116 : NotFound(#[from] GetTenantError),
2117 :
2118 : /// Cancellation token fired while we were waiting
2119 : #[error("cancelled")]
2120 : Cancelled,
2121 :
2122 : /// Tenant exists, but is in a state that cannot become active (e.g. Stopping, Broken)
2123 : #[error("will not become active. Current state: {0}")]
2124 : WillNotBecomeActive(TenantState),
2125 :
2126 : /// Broken is logically a subset of WillNotBecomeActive, but a distinct error is useful as
2127 : /// WillNotBecomeActive is a permitted error under some circumstances, whereas broken should
2128 : /// never happen.
2129 : #[error("Tenant is broken: {0}")]
2130 : Broken(String),
2131 : }
2132 :
2133 0 : #[derive(Debug, thiserror::Error)]
2134 : pub(crate) enum DeleteTimelineError {
2135 : #[error("Tenant {0}")]
2136 : Tenant(#[from] GetTenantError),
2137 :
2138 : #[error("Timeline {0}")]
2139 : Timeline(#[from] crate::tenant::DeleteTimelineError),
2140 : }
2141 :
2142 0 : #[derive(Debug, thiserror::Error)]
2143 : pub(crate) enum TenantStateError {
2144 : #[error("Tenant {0} is stopping")]
2145 : IsStopping(TenantShardId),
2146 : #[error(transparent)]
2147 : SlotError(#[from] TenantSlotError),
2148 : #[error(transparent)]
2149 : SlotUpsertError(#[from] TenantSlotUpsertError),
2150 : #[error(transparent)]
2151 : Other(#[from] anyhow::Error),
2152 : }
2153 :
2154 0 : pub(crate) async fn load_tenant(
2155 0 : conf: &'static PageServerConf,
2156 0 : tenant_id: TenantId,
2157 0 : generation: Generation,
2158 0 : broker_client: storage_broker::BrokerClientChannel,
2159 0 : remote_storage: GenericRemoteStorage,
2160 0 : deletion_queue_client: DeletionQueueClient,
2161 0 : ctx: &RequestContext,
2162 0 : ) -> Result<(), TenantMapInsertError> {
2163 0 : // This is a legacy API (replaced by `/location_conf`). It does not support sharding
2164 0 : let tenant_shard_id = TenantShardId::unsharded(tenant_id);
2165 :
2166 0 : let slot_guard =
2167 0 : tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::MustNotExist)?;
2168 0 : let tenant_path = conf.tenant_path(&tenant_shard_id);
2169 0 :
2170 0 : let tenant_ignore_mark = conf.tenant_ignore_mark_file_path(&tenant_shard_id);
2171 0 : if tenant_ignore_mark.exists() {
2172 0 : std::fs::remove_file(&tenant_ignore_mark).with_context(|| {
2173 0 : format!(
2174 0 : "Failed to remove tenant ignore mark {tenant_ignore_mark:?} during tenant loading"
2175 0 : )
2176 0 : })?;
2177 0 : }
2178 :
2179 0 : let resources = TenantSharedResources {
2180 0 : broker_client,
2181 0 : remote_storage,
2182 0 : deletion_queue_client,
2183 0 : };
2184 :
2185 0 : let mut location_conf =
2186 0 : Tenant::load_tenant_config(conf, &tenant_shard_id).map_err(TenantMapInsertError::Other)?;
2187 0 : location_conf.attach_in_generation(AttachmentMode::Single, generation);
2188 0 :
2189 0 : Tenant::persist_tenant_config(conf, &tenant_shard_id, &location_conf).await?;
2190 :
2191 0 : let shard_identity = location_conf.shard;
2192 0 : let new_tenant = tenant_spawn(
2193 0 : conf,
2194 0 : tenant_shard_id,
2195 0 : &tenant_path,
2196 0 : resources,
2197 0 : AttachedTenantConf::try_from(location_conf)?,
2198 0 : shard_identity,
2199 0 : None,
2200 0 : &TENANTS,
2201 0 : SpawnMode::Eager,
2202 0 : ctx,
2203 0 : )
2204 0 : .with_context(|| format!("Failed to schedule tenant processing in path {tenant_path:?}"))?;
2205 :
2206 0 : slot_guard.upsert(TenantSlot::Attached(new_tenant))?;
2207 0 : Ok(())
2208 0 : }
2209 :
2210 0 : pub(crate) async fn ignore_tenant(
2211 0 : conf: &'static PageServerConf,
2212 0 : tenant_id: TenantId,
2213 0 : ) -> Result<(), TenantStateError> {
2214 0 : ignore_tenant0(conf, &TENANTS, tenant_id).await
2215 0 : }
2216 :
2217 0 : #[instrument(skip_all, fields(shard_id))]
2218 : async fn ignore_tenant0(
2219 : conf: &'static PageServerConf,
2220 : tenants: &std::sync::RwLock<TenantsMap>,
2221 : tenant_id: TenantId,
2222 : ) -> Result<(), TenantStateError> {
2223 : // This is a legacy API (replaced by `/location_conf`). It does not support sharding
2224 : let tenant_shard_id = TenantShardId::unsharded(tenant_id);
2225 : tracing::Span::current().record(
2226 : "shard_id",
2227 : tracing::field::display(tenant_shard_id.shard_slug()),
2228 : );
2229 :
2230 0 : remove_tenant_from_memory(tenants, tenant_shard_id, async {
2231 0 : let ignore_mark_file = conf.tenant_ignore_mark_file_path(&tenant_shard_id);
2232 0 : fs::File::create(&ignore_mark_file)
2233 0 : .await
2234 0 : .context("Failed to create ignore mark file")
2235 0 : .and_then(|_| {
2236 0 : crashsafe::fsync_file_and_parent(&ignore_mark_file)
2237 0 : .context("Failed to fsync ignore mark file")
2238 0 : })
2239 0 : .with_context(|| format!("Failed to crate ignore mark for tenant {tenant_shard_id}"))?;
2240 0 : Ok(())
2241 0 : })
2242 : .await
2243 : }
2244 :
2245 0 : #[derive(Debug, thiserror::Error)]
2246 : pub(crate) enum TenantMapListError {
2247 : #[error("tenant map is still initiailizing")]
2248 : Initializing,
2249 : }
2250 :
2251 0 : #[derive(Debug, thiserror::Error)]
2252 : pub(crate) enum TenantMapInsertError {
2253 : #[error(transparent)]
2254 : SlotError(#[from] TenantSlotError),
2255 : #[error(transparent)]
2256 : SlotUpsertError(#[from] TenantSlotUpsertError),
2257 : #[error(transparent)]
2258 : Other(#[from] anyhow::Error),
2259 : }
2260 :
2261 : /// Superset of TenantMapError: issues that can occur when acquiring a slot
2262 : /// for a particular tenant ID.
2263 0 : #[derive(Debug, thiserror::Error)]
2264 : pub(crate) enum TenantSlotError {
2265 : /// When acquiring a slot with the expectation that the tenant already exists.
2266 : #[error("Tenant {0} not found")]
2267 : NotFound(TenantShardId),
2268 :
2269 : /// When acquiring a slot with the expectation that the tenant does not already exist.
2270 : #[error("tenant {0} already exists, state: {1:?}")]
2271 : AlreadyExists(TenantShardId, TenantState),
2272 :
2273 : // Tried to read a slot that is currently being mutated by another administrative
2274 : // operation.
2275 : #[error("tenant has a state change in progress, try again later")]
2276 : InProgress,
2277 :
2278 : #[error(transparent)]
2279 : MapState(#[from] TenantMapError),
2280 : }
2281 :
2282 : /// Superset of TenantMapError: issues that can occur when using a SlotGuard
2283 : /// to insert a new value.
2284 0 : #[derive(thiserror::Error)]
2285 : pub(crate) enum TenantSlotUpsertError {
2286 : /// An error where the slot is in an unexpected state, indicating a code bug
2287 : #[error("Internal error updating Tenant")]
2288 : InternalError(Cow<'static, str>),
2289 :
2290 : #[error(transparent)]
2291 : MapState(TenantMapError),
2292 :
2293 : // If we encounter TenantManager shutdown during upsert, we must carry the Completion
2294 : // from the SlotGuard, so that the caller can hold it while they clean up: otherwise
2295 : // TenantManager shutdown might race ahead before we're done cleaning up any Tenant that
2296 : // was protected by the SlotGuard.
2297 : #[error("Shutting down")]
2298 : ShuttingDown((TenantSlot, utils::completion::Completion)),
2299 : }
2300 :
2301 : impl std::fmt::Debug for TenantSlotUpsertError {
2302 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
2303 0 : match self {
2304 0 : Self::InternalError(reason) => write!(f, "Internal Error {reason}"),
2305 0 : Self::MapState(map_error) => write!(f, "Tenant map state: {map_error:?}"),
2306 0 : Self::ShuttingDown(_completion) => write!(f, "Tenant map shutting down"),
2307 : }
2308 0 : }
2309 : }
2310 :
2311 0 : #[derive(Debug, thiserror::Error)]
2312 : enum TenantSlotDropError {
2313 : /// It is only legal to drop a TenantSlot if its contents are fully shut down
2314 : #[error("Tenant was not shut down")]
2315 : NotShutdown,
2316 : }
2317 :
2318 : /// Errors that can happen any time we are walking the tenant map to try and acquire
2319 : /// the TenantSlot for a particular tenant.
2320 0 : #[derive(Debug, thiserror::Error)]
2321 : pub enum TenantMapError {
2322 : // Tried to read while initializing
2323 : #[error("tenant map is still initializing")]
2324 : StillInitializing,
2325 :
2326 : // Tried to read while shutting down
2327 : #[error("tenant map is shutting down")]
2328 : ShuttingDown,
2329 : }
2330 :
2331 : /// Guards a particular tenant_id's content in the TenantsMap. While this
2332 : /// structure exists, the TenantsMap will contain a [`TenantSlot::InProgress`]
2333 : /// for this tenant, which acts as a marker for any operations targeting
2334 : /// this tenant to retry later, or wait for the InProgress state to end.
2335 : ///
2336 : /// This structure enforces the important invariant that we do not have overlapping
2337 : /// tasks that will try use local storage for a the same tenant ID: we enforce that
2338 : /// the previous contents of a slot have been shut down before the slot can be
2339 : /// left empty or used for something else
2340 : ///
2341 : /// Holders of a SlotGuard should explicitly dispose of it, using either `upsert`
2342 : /// to provide a new value, or `revert` to put the slot back into its initial
2343 : /// state. If the SlotGuard is dropped without calling either of these, then
2344 : /// we will leave the slot empty if our `old_value` is already shut down, else
2345 : /// we will replace the slot with `old_value` (equivalent to doing a revert).
2346 : ///
2347 : /// The `old_value` may be dropped before the SlotGuard is dropped, by calling
2348 : /// `drop_old_value`. It is an error to call this without shutting down
2349 : /// the conents of `old_value`.
2350 : pub struct SlotGuard {
2351 : tenant_shard_id: TenantShardId,
2352 : old_value: Option<TenantSlot>,
2353 : upserted: bool,
2354 :
2355 : /// [`TenantSlot::InProgress`] carries the corresponding Barrier: it will
2356 : /// release any waiters as soon as this SlotGuard is dropped.
2357 : completion: utils::completion::Completion,
2358 : }
2359 :
2360 : impl SlotGuard {
2361 2 : fn new(
2362 2 : tenant_shard_id: TenantShardId,
2363 2 : old_value: Option<TenantSlot>,
2364 2 : completion: utils::completion::Completion,
2365 2 : ) -> Self {
2366 2 : Self {
2367 2 : tenant_shard_id,
2368 2 : old_value,
2369 2 : upserted: false,
2370 2 : completion,
2371 2 : }
2372 2 : }
2373 :
2374 : /// Get any value that was present in the slot before we acquired ownership
2375 : /// of it: in state transitions, this will be the old state.
2376 2 : fn get_old_value(&self) -> &Option<TenantSlot> {
2377 2 : &self.old_value
2378 2 : }
2379 :
2380 : /// Emplace a new value in the slot. This consumes the guard, and after
2381 : /// returning, the slot is no longer protected from concurrent changes.
2382 0 : fn upsert(mut self, new_value: TenantSlot) -> Result<(), TenantSlotUpsertError> {
2383 0 : if !self.old_value_is_shutdown() {
2384 : // This is a bug: callers should never try to drop an old value without
2385 : // shutting it down
2386 0 : return Err(TenantSlotUpsertError::InternalError(
2387 0 : "Old TenantSlot value not shut down".into(),
2388 0 : ));
2389 0 : }
2390 :
2391 0 : let replaced = {
2392 0 : let mut locked = TENANTS.write().unwrap();
2393 0 :
2394 0 : if let TenantSlot::InProgress(_) = new_value {
2395 : // It is never expected to try and upsert InProgress via this path: it should
2396 : // only be written via the tenant_map_acquire_slot path. If we hit this it's a bug.
2397 0 : return Err(TenantSlotUpsertError::InternalError(
2398 0 : "Attempt to upsert an InProgress state".into(),
2399 0 : ));
2400 0 : }
2401 :
2402 0 : let m = match &mut *locked {
2403 : TenantsMap::Initializing => {
2404 0 : return Err(TenantSlotUpsertError::MapState(
2405 0 : TenantMapError::StillInitializing,
2406 0 : ))
2407 : }
2408 : TenantsMap::ShuttingDown(_) => {
2409 0 : return Err(TenantSlotUpsertError::ShuttingDown((
2410 0 : new_value,
2411 0 : self.completion.clone(),
2412 0 : )));
2413 : }
2414 0 : TenantsMap::Open(m) => m,
2415 0 : };
2416 0 :
2417 0 : METRICS.slot_inserted(&new_value);
2418 0 :
2419 0 : let replaced = m.insert(self.tenant_shard_id, new_value);
2420 0 : self.upserted = true;
2421 0 : if let Some(replaced) = replaced.as_ref() {
2422 0 : METRICS.slot_removed(replaced);
2423 0 : }
2424 :
2425 0 : replaced
2426 : };
2427 :
2428 : // Sanity check: on an upsert we should always be replacing an InProgress marker
2429 0 : match replaced {
2430 : Some(TenantSlot::InProgress(_)) => {
2431 : // Expected case: we find our InProgress in the map: nothing should have
2432 : // replaced it because the code that acquires slots will not grant another
2433 : // one for the same TenantId.
2434 0 : Ok(())
2435 : }
2436 : None => {
2437 0 : METRICS.unexpected_errors.inc();
2438 0 : error!(
2439 : tenant_shard_id = %self.tenant_shard_id,
2440 0 : "Missing InProgress marker during tenant upsert, this is a bug."
2441 : );
2442 0 : Err(TenantSlotUpsertError::InternalError(
2443 0 : "Missing InProgress marker during tenant upsert".into(),
2444 0 : ))
2445 : }
2446 0 : Some(slot) => {
2447 0 : METRICS.unexpected_errors.inc();
2448 0 : error!(tenant_shard_id=%self.tenant_shard_id, "Unexpected contents of TenantSlot during upsert, this is a bug. Contents: {:?}", slot);
2449 0 : Err(TenantSlotUpsertError::InternalError(
2450 0 : "Unexpected contents of TenantSlot".into(),
2451 0 : ))
2452 : }
2453 : }
2454 0 : }
2455 :
2456 : /// Replace the InProgress slot with whatever was in the guard when we started
2457 0 : fn revert(mut self) {
2458 0 : if let Some(value) = self.old_value.take() {
2459 0 : match self.upsert(value) {
2460 0 : Err(TenantSlotUpsertError::InternalError(_)) => {
2461 0 : // We already logged the error, nothing else we can do.
2462 0 : }
2463 : Err(
2464 : TenantSlotUpsertError::MapState(_) | TenantSlotUpsertError::ShuttingDown(_),
2465 0 : ) => {
2466 0 : // If the map is shutting down, we need not replace anything
2467 0 : }
2468 0 : Ok(()) => {}
2469 : }
2470 0 : }
2471 0 : }
2472 :
2473 : /// We may never drop our old value until it is cleanly shut down: otherwise we might leave
2474 : /// rogue background tasks that would write to the local tenant directory that this guard
2475 : /// is responsible for protecting
2476 2 : fn old_value_is_shutdown(&self) -> bool {
2477 2 : match self.old_value.as_ref() {
2478 2 : Some(TenantSlot::Attached(tenant)) => tenant.gate.close_complete(),
2479 0 : Some(TenantSlot::Secondary(secondary_tenant)) => secondary_tenant.gate.close_complete(),
2480 : Some(TenantSlot::InProgress(_)) => {
2481 : // A SlotGuard cannot be constructed for a slot that was already InProgress
2482 0 : unreachable!()
2483 : }
2484 0 : None => true,
2485 : }
2486 2 : }
2487 :
2488 : /// The guard holder is done with the old value of the slot: they are obliged to already
2489 : /// shut it down before we reach this point.
2490 2 : fn drop_old_value(&mut self) -> Result<(), TenantSlotDropError> {
2491 2 : if !self.old_value_is_shutdown() {
2492 0 : Err(TenantSlotDropError::NotShutdown)
2493 : } else {
2494 2 : self.old_value.take();
2495 2 : Ok(())
2496 : }
2497 2 : }
2498 : }
2499 :
2500 : impl Drop for SlotGuard {
2501 2 : fn drop(&mut self) {
2502 2 : if self.upserted {
2503 0 : return;
2504 2 : }
2505 2 : // Our old value is already shutdown, or it never existed: it is safe
2506 2 : // for us to fully release the TenantSlot back into an empty state
2507 2 :
2508 2 : let mut locked = TENANTS.write().unwrap();
2509 :
2510 2 : let m = match &mut *locked {
2511 : TenantsMap::Initializing => {
2512 : // There is no map, this should never happen.
2513 2 : return;
2514 : }
2515 : TenantsMap::ShuttingDown(_) => {
2516 : // When we transition to shutdown, InProgress elements are removed
2517 : // from the map, so we do not need to clean up our Inprogress marker.
2518 : // See [`shutdown_all_tenants0`]
2519 0 : return;
2520 : }
2521 0 : TenantsMap::Open(m) => m,
2522 0 : };
2523 0 :
2524 0 : use std::collections::btree_map::Entry;
2525 0 : match m.entry(self.tenant_shard_id) {
2526 0 : Entry::Occupied(mut entry) => {
2527 0 : if !matches!(entry.get(), TenantSlot::InProgress(_)) {
2528 0 : METRICS.unexpected_errors.inc();
2529 0 : error!(tenant_shard_id=%self.tenant_shard_id, "Unexpected contents of TenantSlot during drop, this is a bug. Contents: {:?}", entry.get());
2530 0 : }
2531 :
2532 0 : if self.old_value_is_shutdown() {
2533 0 : METRICS.slot_removed(entry.get());
2534 0 : entry.remove();
2535 0 : } else {
2536 0 : let inserting = self.old_value.take().unwrap();
2537 0 : METRICS.slot_inserted(&inserting);
2538 0 : let replaced = entry.insert(inserting);
2539 0 : METRICS.slot_removed(&replaced);
2540 0 : }
2541 : }
2542 : Entry::Vacant(_) => {
2543 0 : METRICS.unexpected_errors.inc();
2544 0 : error!(
2545 : tenant_shard_id = %self.tenant_shard_id,
2546 0 : "Missing InProgress marker during SlotGuard drop, this is a bug."
2547 : );
2548 : }
2549 : }
2550 2 : }
2551 : }
2552 :
2553 : enum TenantSlotPeekMode {
2554 : /// In Read mode, peek will be permitted to see the slots even if the pageserver is shutting down
2555 : Read,
2556 : /// In Write mode, trying to peek at a slot while the pageserver is shutting down is an error
2557 : Write,
2558 : }
2559 :
2560 0 : fn tenant_map_peek_slot<'a>(
2561 0 : tenants: &'a std::sync::RwLockReadGuard<'a, TenantsMap>,
2562 0 : tenant_shard_id: &TenantShardId,
2563 0 : mode: TenantSlotPeekMode,
2564 0 : ) -> Result<Option<&'a TenantSlot>, TenantMapError> {
2565 0 : match tenants.deref() {
2566 0 : TenantsMap::Initializing => Err(TenantMapError::StillInitializing),
2567 0 : TenantsMap::ShuttingDown(m) => match mode {
2568 : TenantSlotPeekMode::Read => Ok(Some(
2569 : // When reading in ShuttingDown state, we must translate None results
2570 : // into a ShuttingDown error, because absence of a tenant shard ID in the map
2571 : // isn't a reliable indicator of the tenant being gone: it might have been
2572 : // InProgress when shutdown started, and cleaned up from that state such
2573 : // that it's now no longer in the map. Callers will have to wait until
2574 : // we next start up to get a proper answer. This avoids incorrect 404 API responses.
2575 0 : m.get(tenant_shard_id).ok_or(TenantMapError::ShuttingDown)?,
2576 : )),
2577 0 : TenantSlotPeekMode::Write => Err(TenantMapError::ShuttingDown),
2578 : },
2579 0 : TenantsMap::Open(m) => Ok(m.get(tenant_shard_id)),
2580 : }
2581 0 : }
2582 :
2583 : enum TenantSlotAcquireMode {
2584 : /// Acquire the slot irrespective of current state, or whether it already exists
2585 : Any,
2586 : /// Return an error if trying to acquire a slot and it doesn't already exist
2587 : MustExist,
2588 : /// Return an error if trying to acquire a slot and it already exists
2589 : MustNotExist,
2590 : }
2591 :
2592 0 : fn tenant_map_acquire_slot(
2593 0 : tenant_shard_id: &TenantShardId,
2594 0 : mode: TenantSlotAcquireMode,
2595 0 : ) -> Result<SlotGuard, TenantSlotError> {
2596 0 : tenant_map_acquire_slot_impl(tenant_shard_id, &TENANTS, mode)
2597 0 : }
2598 :
2599 2 : fn tenant_map_acquire_slot_impl(
2600 2 : tenant_shard_id: &TenantShardId,
2601 2 : tenants: &std::sync::RwLock<TenantsMap>,
2602 2 : mode: TenantSlotAcquireMode,
2603 2 : ) -> Result<SlotGuard, TenantSlotError> {
2604 2 : use TenantSlotAcquireMode::*;
2605 2 : METRICS.tenant_slot_writes.inc();
2606 2 :
2607 2 : let mut locked = tenants.write().unwrap();
2608 2 : let span = tracing::info_span!("acquire_slot", tenant_id=%tenant_shard_id.tenant_id, shard_id = %tenant_shard_id.shard_slug());
2609 2 : let _guard = span.enter();
2610 :
2611 2 : let m = match &mut *locked {
2612 0 : TenantsMap::Initializing => return Err(TenantMapError::StillInitializing.into()),
2613 0 : TenantsMap::ShuttingDown(_) => return Err(TenantMapError::ShuttingDown.into()),
2614 2 : TenantsMap::Open(m) => m,
2615 2 : };
2616 2 :
2617 2 : use std::collections::btree_map::Entry;
2618 2 :
2619 2 : let entry = m.entry(*tenant_shard_id);
2620 2 :
2621 2 : match entry {
2622 0 : Entry::Vacant(v) => match mode {
2623 : MustExist => {
2624 0 : tracing::debug!("Vacant && MustExist: return NotFound");
2625 0 : Err(TenantSlotError::NotFound(*tenant_shard_id))
2626 : }
2627 : _ => {
2628 0 : let (completion, barrier) = utils::completion::channel();
2629 0 : let inserting = TenantSlot::InProgress(barrier);
2630 0 : METRICS.slot_inserted(&inserting);
2631 0 : v.insert(inserting);
2632 0 : tracing::debug!("Vacant, inserted InProgress");
2633 0 : Ok(SlotGuard::new(*tenant_shard_id, None, completion))
2634 : }
2635 : },
2636 2 : Entry::Occupied(mut o) => {
2637 2 : // Apply mode-driven checks
2638 2 : match (o.get(), mode) {
2639 : (TenantSlot::InProgress(_), _) => {
2640 0 : tracing::debug!("Occupied, failing for InProgress");
2641 0 : Err(TenantSlotError::InProgress)
2642 : }
2643 0 : (slot, MustNotExist) => match slot {
2644 0 : TenantSlot::Attached(tenant) => {
2645 0 : tracing::debug!("Attached && MustNotExist, return AlreadyExists");
2646 0 : Err(TenantSlotError::AlreadyExists(
2647 0 : *tenant_shard_id,
2648 0 : tenant.current_state(),
2649 0 : ))
2650 : }
2651 : _ => {
2652 : // FIXME: the AlreadyExists error assumes that we have a Tenant
2653 : // to get the state from
2654 0 : tracing::debug!("Occupied & MustNotExist, return AlreadyExists");
2655 0 : Err(TenantSlotError::AlreadyExists(
2656 0 : *tenant_shard_id,
2657 0 : TenantState::Broken {
2658 0 : reason: "Present but not attached".to_string(),
2659 0 : backtrace: "".to_string(),
2660 0 : },
2661 0 : ))
2662 : }
2663 : },
2664 : _ => {
2665 : // Happy case: the slot was not in any state that violated our mode
2666 2 : let (completion, barrier) = utils::completion::channel();
2667 2 : let in_progress = TenantSlot::InProgress(barrier);
2668 2 : METRICS.slot_inserted(&in_progress);
2669 2 : let old_value = o.insert(in_progress);
2670 2 : METRICS.slot_removed(&old_value);
2671 2 : tracing::debug!("Occupied, replaced with InProgress");
2672 2 : Ok(SlotGuard::new(
2673 2 : *tenant_shard_id,
2674 2 : Some(old_value),
2675 2 : completion,
2676 2 : ))
2677 : }
2678 : }
2679 : }
2680 : }
2681 2 : }
2682 :
2683 : /// Stops and removes the tenant from memory, if it's not [`TenantState::Stopping`] already, bails otherwise.
2684 : /// Allows to remove other tenant resources manually, via `tenant_cleanup`.
2685 : /// If the cleanup fails, tenant will stay in memory in [`TenantState::Broken`] state, and another removal
2686 : /// operation would be needed to remove it.
2687 2 : async fn remove_tenant_from_memory<V, F>(
2688 2 : tenants: &std::sync::RwLock<TenantsMap>,
2689 2 : tenant_shard_id: TenantShardId,
2690 2 : tenant_cleanup: F,
2691 2 : ) -> Result<V, TenantStateError>
2692 2 : where
2693 2 : F: std::future::Future<Output = anyhow::Result<V>>,
2694 2 : {
2695 2 : let mut slot_guard =
2696 2 : tenant_map_acquire_slot_impl(&tenant_shard_id, tenants, TenantSlotAcquireMode::MustExist)?;
2697 :
2698 : // allow pageserver shutdown to await for our completion
2699 2 : let (_guard, progress) = completion::channel();
2700 :
2701 : // The SlotGuard allows us to manipulate the Tenant object without fear of some
2702 : // concurrent API request doing something else for the same tenant ID.
2703 2 : let attached_tenant = match slot_guard.get_old_value() {
2704 2 : Some(TenantSlot::Attached(tenant)) => {
2705 2 : // whenever we remove a tenant from memory, we don't want to flush and wait for upload
2706 2 : let shutdown_mode = ShutdownMode::Hard;
2707 2 :
2708 2 : // shutdown is sure to transition tenant to stopping, and wait for all tasks to complete, so
2709 2 : // that we can continue safely to cleanup.
2710 2 : match tenant.shutdown(progress, shutdown_mode).await {
2711 2 : Ok(()) => {}
2712 0 : Err(_other) => {
2713 0 : // if pageserver shutdown or other detach/ignore is already ongoing, we don't want to
2714 0 : // wait for it but return an error right away because these are distinct requests.
2715 0 : slot_guard.revert();
2716 0 : return Err(TenantStateError::IsStopping(tenant_shard_id));
2717 : }
2718 : }
2719 2 : Some(tenant)
2720 : }
2721 0 : Some(TenantSlot::Secondary(secondary_state)) => {
2722 0 : tracing::info!("Shutting down in secondary mode");
2723 0 : secondary_state.shutdown().await;
2724 0 : None
2725 : }
2726 : Some(TenantSlot::InProgress(_)) => {
2727 : // Acquiring a slot guarantees its old value was not InProgress
2728 0 : unreachable!();
2729 : }
2730 0 : None => None,
2731 : };
2732 :
2733 2 : match tenant_cleanup
2734 2 : .await
2735 2 : .with_context(|| format!("Failed to run cleanup for tenant {tenant_shard_id}"))
2736 : {
2737 2 : Ok(hook_value) => {
2738 2 : // Success: drop the old TenantSlot::Attached.
2739 2 : slot_guard
2740 2 : .drop_old_value()
2741 2 : .expect("We just called shutdown");
2742 2 :
2743 2 : Ok(hook_value)
2744 : }
2745 0 : Err(e) => {
2746 : // If we had a Tenant, set it to Broken and put it back in the TenantsMap
2747 0 : if let Some(attached_tenant) = attached_tenant {
2748 0 : attached_tenant.set_broken(e.to_string()).await;
2749 0 : }
2750 : // Leave the broken tenant in the map
2751 0 : slot_guard.revert();
2752 0 :
2753 0 : Err(TenantStateError::Other(e))
2754 : }
2755 : }
2756 2 : }
2757 :
2758 : use {
2759 : crate::repository::GcResult, pageserver_api::models::TimelineGcRequest,
2760 : utils::http::error::ApiError,
2761 : };
2762 :
2763 0 : #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), %timeline_id))]
2764 : pub(crate) async fn immediate_gc(
2765 : tenant_shard_id: TenantShardId,
2766 : timeline_id: TimelineId,
2767 : gc_req: TimelineGcRequest,
2768 : cancel: CancellationToken,
2769 : ctx: &RequestContext,
2770 : ) -> Result<GcResult, ApiError> {
2771 : let tenant = {
2772 : let guard = TENANTS.read().unwrap();
2773 : guard
2774 : .get(&tenant_shard_id)
2775 : .cloned()
2776 0 : .with_context(|| format!("tenant {tenant_shard_id}"))
2777 0 : .map_err(|e| ApiError::NotFound(e.into()))?
2778 : };
2779 :
2780 0 : let gc_horizon = gc_req.gc_horizon.unwrap_or_else(|| tenant.get_gc_horizon());
2781 : // Use tenant's pitr setting
2782 : let pitr = tenant.get_pitr_interval();
2783 :
2784 : tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
2785 :
2786 : // Run in task_mgr to avoid race with tenant_detach operation
2787 : let ctx: RequestContext =
2788 : ctx.detached_child(TaskKind::GarbageCollector, DownloadBehavior::Download);
2789 :
2790 0 : let _gate_guard = tenant.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
2791 :
2792 : fail::fail_point!("immediate_gc_task_pre");
2793 :
2794 : #[allow(unused_mut)]
2795 : let mut result = tenant
2796 : .gc_iteration(Some(timeline_id), gc_horizon, pitr, &cancel, &ctx)
2797 : .await;
2798 : // FIXME: `gc_iteration` can return an error for multiple reasons; we should handle it
2799 : // better once the types support it.
2800 :
2801 : #[cfg(feature = "testing")]
2802 : {
2803 : // we need to synchronize with drop completion for python tests without polling for
2804 : // log messages
2805 : if let Ok(result) = result.as_mut() {
2806 : let mut js = tokio::task::JoinSet::new();
2807 : for layer in std::mem::take(&mut result.doomed_layers) {
2808 : js.spawn(layer.wait_drop());
2809 : }
2810 : tracing::info!(
2811 : total = js.len(),
2812 : "starting to wait for the gc'd layers to be dropped"
2813 : );
2814 : while let Some(res) = js.join_next().await {
2815 : res.expect("wait_drop should not panic");
2816 : }
2817 : }
2818 :
2819 : let timeline = tenant.get_timeline(timeline_id, false).ok();
2820 0 : let rtc = timeline.as_ref().map(|x| &x.remote_client);
2821 :
2822 : if let Some(rtc) = rtc {
2823 : // layer drops schedule actions on remote timeline client to actually do the
2824 : // deletions; don't care about the shutdown error, just exit fast
2825 : drop(rtc.wait_completion().await);
2826 : }
2827 : }
2828 :
2829 : result.map_err(ApiError::InternalServerError)
2830 : }
2831 :
2832 : #[cfg(test)]
2833 : mod tests {
2834 : use std::collections::BTreeMap;
2835 : use std::sync::Arc;
2836 : use tracing::Instrument;
2837 :
2838 : use crate::tenant::mgr::TenantSlot;
2839 :
2840 : use super::{super::harness::TenantHarness, TenantsMap};
2841 :
2842 : #[tokio::test(start_paused = true)]
2843 2 : async fn shutdown_awaits_in_progress_tenant() {
2844 2 : // Test that if an InProgress tenant is in the map during shutdown, the shutdown will gracefully
2845 2 : // wait for it to complete before proceeding.
2846 2 :
2847 2 : let h = TenantHarness::create("shutdown_awaits_in_progress_tenant").unwrap();
2848 8 : let (t, _ctx) = h.load().await;
2849 2 :
2850 2 : // harness loads it to active, which is forced and nothing is running on the tenant
2851 2 :
2852 2 : let id = t.tenant_shard_id();
2853 2 :
2854 2 : // tenant harness configures the logging and we cannot escape it
2855 2 : let span = h.span();
2856 2 : let _e = span.enter();
2857 2 :
2858 2 : let tenants = BTreeMap::from([(id, TenantSlot::Attached(t.clone()))]);
2859 2 : let tenants = Arc::new(std::sync::RwLock::new(TenantsMap::Open(tenants)));
2860 2 :
2861 2 : // Invoke remove_tenant_from_memory with a cleanup hook that blocks until we manually
2862 2 : // permit it to proceed: that will stick the tenant in InProgress
2863 2 :
2864 2 : let (until_cleanup_completed, can_complete_cleanup) = utils::completion::channel();
2865 2 : let (until_cleanup_started, cleanup_started) = utils::completion::channel();
2866 2 : let mut remove_tenant_from_memory_task = {
2867 2 : let jh = tokio::spawn({
2868 2 : let tenants = tenants.clone();
2869 2 : async move {
2870 2 : let cleanup = async move {
2871 2 : drop(until_cleanup_started);
2872 2 : can_complete_cleanup.wait().await;
2873 2 : anyhow::Ok(())
2874 2 : };
2875 2 : super::remove_tenant_from_memory(&tenants, id, cleanup).await
2876 2 : }
2877 2 : .instrument(h.span())
2878 2 : });
2879 2 :
2880 2 : // now the long cleanup should be in place, with the stopping state
2881 2 : cleanup_started.wait().await;
2882 2 : jh
2883 2 : };
2884 2 :
2885 2 : let mut shutdown_task = {
2886 2 : let (until_shutdown_started, shutdown_started) = utils::completion::channel();
2887 2 :
2888 2 : let shutdown_task = tokio::spawn(async move {
2889 2 : drop(until_shutdown_started);
2890 4 : super::shutdown_all_tenants0(&tenants).await;
2891 2 : });
2892 2 :
2893 2 : shutdown_started.wait().await;
2894 2 : shutdown_task
2895 2 : };
2896 2 :
2897 2 : let long_time = std::time::Duration::from_secs(15);
2898 2 : tokio::select! {
2899 2 : _ = &mut shutdown_task => unreachable!("shutdown should block on remove_tenant_from_memory completing"),
2900 2 : _ = &mut remove_tenant_from_memory_task => unreachable!("remove_tenant_from_memory_task should not complete until explicitly unblocked"),
2901 2 : _ = tokio::time::sleep(long_time) => {},
2902 2 : }
2903 2 :
2904 2 : drop(until_cleanup_completed);
2905 2 :
2906 2 : // Now that we allow it to proceed, shutdown should complete immediately
2907 2 : remove_tenant_from_memory_task.await.unwrap().unwrap();
2908 2 : shutdown_task.await.unwrap();
2909 2 : }
2910 : }
|
