Line data Source code
1 : //! This module acts as a switchboard to access different repositories managed by this
2 : //! page server.
3 :
4 : use std::borrow::Cow;
5 : use std::cmp::Ordering;
6 : use std::collections::{BTreeMap, HashMap, HashSet};
7 : use std::ops::Deref;
8 : use std::sync::Arc;
9 : use std::time::Duration;
10 :
11 : use anyhow::Context;
12 : use camino::{Utf8DirEntry, Utf8Path, Utf8PathBuf};
13 : use futures::StreamExt;
14 : use itertools::Itertools;
15 : use pageserver_api::key::Key;
16 : use pageserver_api::models::{DetachBehavior, LocationConfigMode};
17 : use pageserver_api::shard::{
18 : ShardCount, ShardIdentity, ShardIndex, ShardNumber, ShardStripeSize, TenantShardId,
19 : };
20 : use pageserver_api::upcall_api::ReAttachResponseTenant;
21 : use rand::Rng;
22 : use rand::distributions::Alphanumeric;
23 : use remote_storage::TimeoutOrCancel;
24 : use sysinfo::SystemExt;
25 : use tokio::fs;
26 : use tokio::task::JoinSet;
27 : use tokio_util::sync::CancellationToken;
28 : use tracing::*;
29 : use utils::crashsafe::path_with_suffix_extension;
30 : use utils::fs_ext::PathExt;
31 : use utils::generation::Generation;
32 : use utils::id::{TenantId, TimelineId};
33 : use utils::{backoff, completion, crashsafe};
34 :
35 : use super::remote_timeline_client::remote_tenant_path;
36 : use super::secondary::SecondaryTenant;
37 : use super::timeline::detach_ancestor::{self, PreparedTimelineDetach};
38 : use super::{GlobalShutDown, TenantSharedResources};
39 : use crate::config::PageServerConf;
40 : use crate::context::{DownloadBehavior, RequestContext};
41 : use crate::controller_upcall_client::{
42 : RetryForeverError, StorageControllerUpcallApi, StorageControllerUpcallClient,
43 : };
44 : use crate::deletion_queue::DeletionQueueClient;
45 : use crate::http::routes::ACTIVE_TENANT_TIMEOUT;
46 : use crate::metrics::{LOCAL_DATA_LOSS_SUSPECTED, TENANT, TENANT_MANAGER as METRICS};
47 : use crate::task_mgr::{BACKGROUND_RUNTIME, TaskKind};
48 : use crate::tenant::config::{
49 : AttachedLocationConfig, AttachmentMode, LocationConf, LocationMode, SecondaryLocationConfig,
50 : };
51 : use crate::tenant::span::debug_assert_current_span_has_tenant_id;
52 : use crate::tenant::storage_layer::inmemory_layer;
53 : use crate::tenant::timeline::ShutdownMode;
54 : use crate::tenant::timeline::layer_manager::LayerManagerLockHolder;
55 : use crate::tenant::{
56 : AttachedTenantConf, GcError, LoadConfigError, SpawnMode, TenantShard, TenantState,
57 : };
58 : use crate::virtual_file::MaybeFatalIo;
59 : use crate::{InitializationOrder, TEMP_FILE_SUFFIX};
60 :
61 : /// For a tenant that appears in TenantsMap, it may either be
62 : /// - `Attached`: has a full Tenant object, is elegible to service
63 : /// reads and ingest WAL.
64 : /// - `Secondary`: is only keeping a local cache warm.
65 : ///
66 : /// Secondary is a totally distinct state rather than being a mode of a `Tenant`, because
67 : /// that way we avoid having to carefully switch a tenant's ingestion etc on and off during
68 : /// its lifetime, and we can preserve some important safety invariants like `Tenant` always
69 : /// having a properly acquired generation (Secondary doesn't need a generation)
70 : #[derive(Clone)]
71 : pub(crate) enum TenantSlot {
72 : Attached(Arc<TenantShard>),
73 : Secondary(Arc<SecondaryTenant>),
74 : /// In this state, other administrative operations acting on the TenantId should
75 : /// block, or return a retry indicator equivalent to HTTP 503.
76 : InProgress(utils::completion::Barrier),
77 : }
78 :
79 : impl std::fmt::Debug for TenantSlot {
80 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
81 0 : match self {
82 0 : Self::Attached(tenant) => write!(f, "Attached({})", tenant.current_state()),
83 0 : Self::Secondary(_) => write!(f, "Secondary"),
84 0 : Self::InProgress(_) => write!(f, "InProgress"),
85 : }
86 0 : }
87 : }
88 :
89 : impl TenantSlot {
90 : /// Return the `Tenant` in this slot if attached, else None
91 0 : fn get_attached(&self) -> Option<&Arc<TenantShard>> {
92 0 : match self {
93 0 : Self::Attached(t) => Some(t),
94 0 : Self::Secondary(_) => None,
95 0 : Self::InProgress(_) => None,
96 : }
97 0 : }
98 : }
99 :
100 : /// The tenants known to the pageserver.
101 : /// The enum variants are used to distinguish the different states that the pageserver can be in.
102 : pub(crate) enum TenantsMap {
103 : /// [`init_tenant_mgr`] is not done yet.
104 : Initializing,
105 : /// [`init_tenant_mgr`] is done, all on-disk tenants have been loaded.
106 : /// New tenants can be added using [`TenantManager::tenant_map_acquire_slot`].
107 : Open(BTreeMap<TenantShardId, TenantSlot>),
108 : /// The pageserver has entered shutdown mode via [`TenantManager::shutdown`].
109 : /// Existing tenants are still accessible, but no new tenants can be created.
110 : ShuttingDown(BTreeMap<TenantShardId, TenantSlot>),
111 : }
112 :
113 : /// When resolving a TenantId to a shard, we may be looking for the 0th
114 : /// shard, or we might be looking for whichever shard holds a particular page.
115 : #[derive(Copy, Clone)]
116 : pub(crate) enum ShardSelector {
117 : /// Only return the 0th shard, if it is present. If a non-0th shard is present,
118 : /// ignore it.
119 : Zero,
120 : /// Pick the shard that holds this key
121 : Page(Key),
122 : /// The shard ID is known: pick the given shard
123 : Known(ShardIndex),
124 : }
125 :
126 : /// A convenience for use with the re_attach ControllerUpcallClient function: rather
127 : /// than the serializable struct, we build this enum that encapsulates
128 : /// the invariant that attached tenants always have generations.
129 : ///
130 : /// This represents the subset of a LocationConfig that we receive during re-attach.
131 : pub(crate) enum TenantStartupMode {
132 : Attached((AttachmentMode, Generation, ShardStripeSize)),
133 : Secondary,
134 : }
135 :
136 : impl TenantStartupMode {
137 : /// Return the generation & mode that should be used when starting
138 : /// this tenant.
139 : ///
140 : /// If this returns None, the re-attach struct is in an invalid state and
141 : /// should be ignored in the response.
142 0 : fn from_reattach_tenant(rart: ReAttachResponseTenant) -> Option<Self> {
143 0 : match (rart.mode, rart.r#gen) {
144 0 : (LocationConfigMode::Detached, _) => None,
145 0 : (LocationConfigMode::Secondary, _) => Some(Self::Secondary),
146 0 : (LocationConfigMode::AttachedMulti, Some(g)) => Some(Self::Attached((
147 0 : AttachmentMode::Multi,
148 0 : Generation::new(g),
149 0 : rart.stripe_size,
150 0 : ))),
151 0 : (LocationConfigMode::AttachedSingle, Some(g)) => Some(Self::Attached((
152 0 : AttachmentMode::Single,
153 0 : Generation::new(g),
154 0 : rart.stripe_size,
155 0 : ))),
156 0 : (LocationConfigMode::AttachedStale, Some(g)) => Some(Self::Attached((
157 0 : AttachmentMode::Stale,
158 0 : Generation::new(g),
159 0 : rart.stripe_size,
160 0 : ))),
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<TenantShard>),
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<TenantShard>> {
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 : #[cfg(all(debug_assertions, not(test)))]
194 0 : pub(crate) fn len(&self) -> usize {
195 0 : match self {
196 0 : TenantsMap::Initializing => 0,
197 0 : TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => m.len(),
198 : }
199 0 : }
200 : }
201 :
202 : /// Precursor to deletion of a tenant dir: we do a fast rename to a tmp path, and then
203 : /// the slower actual deletion in the background.
204 : ///
205 : /// This is "safe" in that that it won't leave behind a partially deleted directory
206 : /// at the original path, because we rename with TEMP_FILE_SUFFIX before starting deleting
207 : /// the contents.
208 : ///
209 : /// This is pageserver-specific, as it relies on future processes after a crash to check
210 : /// for TEMP_FILE_SUFFIX when loading things.
211 0 : async fn safe_rename_tenant_dir(path: impl AsRef<Utf8Path>) -> std::io::Result<Utf8PathBuf> {
212 0 : let parent = path
213 0 : .as_ref()
214 0 : .parent()
215 : // It is invalid to call this function with a relative path. Tenant directories
216 : // should always have a parent.
217 0 : .ok_or(std::io::Error::new(
218 0 : std::io::ErrorKind::InvalidInput,
219 : "Path must be absolute",
220 0 : ))?;
221 0 : let rand_suffix = rand::thread_rng()
222 0 : .sample_iter(&Alphanumeric)
223 0 : .take(8)
224 0 : .map(char::from)
225 0 : .collect::<String>()
226 0 : + TEMP_FILE_SUFFIX;
227 0 : let tmp_path = path_with_suffix_extension(&path, &rand_suffix);
228 0 : fs::rename(path.as_ref(), &tmp_path).await?;
229 0 : fs::File::open(parent)
230 0 : .await?
231 0 : .sync_all()
232 0 : .await
233 0 : .maybe_fatal_err("safe_rename_tenant_dir")?;
234 0 : Ok(tmp_path)
235 0 : }
236 :
237 : /// See [`Self::spawn`].
238 : #[derive(Clone, Default)]
239 : pub struct BackgroundPurges(tokio_util::task::TaskTracker);
240 :
241 : impl BackgroundPurges {
242 : /// When we have moved a tenant's content to a temporary directory, we may delete it lazily in
243 : /// the background, and thereby avoid blocking any API requests on this deletion completing.
244 : ///
245 : /// Although we are cleaning up the tenant, this task is not meant to be bound by the lifetime of the tenant in memory.
246 : /// Thus the [`BackgroundPurges`] type to keep track of these tasks.
247 0 : pub fn spawn(&self, tmp_path: Utf8PathBuf) {
248 : // because on shutdown we close and wait, we are misusing TaskTracker a bit.
249 : //
250 : // so first acquire a token, then check if the tracker has been closed. the tracker might get closed
251 : // right after, but at least the shutdown will wait for what we are spawning next.
252 0 : let token = self.0.token();
253 :
254 0 : if self.0.is_closed() {
255 0 : warn!(
256 : %tmp_path,
257 0 : "trying to spawn background purge during shutdown, ignoring"
258 : );
259 0 : return;
260 0 : }
261 :
262 0 : let span = info_span!(parent: None, "background_purge", %tmp_path);
263 :
264 0 : let task = move || {
265 0 : let _token = token;
266 0 : let _entered = span.entered();
267 0 : if let Err(error) = std::fs::remove_dir_all(tmp_path.as_path()) {
268 : // should we fatal_io_error here?
269 0 : warn!(%error, "failed to purge tenant directory");
270 0 : }
271 0 : };
272 :
273 0 : BACKGROUND_RUNTIME.spawn_blocking(task);
274 0 : }
275 :
276 : /// When this future completes, all background purges have completed.
277 : /// The first poll of the future will already lock out new background purges spawned via [`Self::spawn`].
278 : ///
279 : /// Concurrent calls will coalesce.
280 : ///
281 : /// # Cancellation-Safety
282 : ///
283 : /// If this future is dropped before polled to completion, concurrent and subsequent
284 : /// instances of this future will continue to be correct.
285 : #[instrument(skip_all)]
286 : pub async fn shutdown(&self) {
287 : // forbid new tasks (can be called many times)
288 : self.0.close();
289 : self.0.wait().await;
290 : }
291 : }
292 :
293 : /// Responsible for storing and mutating the collection of all tenants
294 : /// that this pageserver has state for.
295 : ///
296 : /// Every Tenant and SecondaryTenant instance lives inside the TenantManager.
297 : ///
298 : /// The most important role of the TenantManager is to prevent conflicts: e.g. trying to attach
299 : /// the same tenant twice concurrently, or trying to configure the same tenant into secondary
300 : /// and attached modes concurrently.
301 : pub struct TenantManager {
302 : conf: &'static PageServerConf,
303 : tenants: std::sync::RwLock<TenantsMap>,
304 : resources: TenantSharedResources,
305 :
306 : // Long-running operations that happen outside of a [`Tenant`] lifetime should respect this token.
307 : // This is for edge cases like tenant deletion. In normal cases (within a Tenant lifetime),
308 : // tenants have their own cancellation tokens, which we fire individually in [`Self::shutdown`], or
309 : // when the tenant detaches.
310 : cancel: CancellationToken,
311 :
312 : background_purges: BackgroundPurges,
313 : }
314 :
315 0 : fn emergency_generations(
316 0 : tenant_confs: &HashMap<TenantShardId, Result<LocationConf, LoadConfigError>>,
317 0 : ) -> HashMap<TenantShardId, TenantStartupMode> {
318 0 : tenant_confs
319 0 : .iter()
320 0 : .filter_map(|(tid, lc)| {
321 0 : let lc = match lc {
322 0 : Ok(lc) => lc,
323 0 : Err(_) => return None,
324 : };
325 : Some((
326 0 : *tid,
327 0 : match &lc.mode {
328 0 : LocationMode::Attached(alc) => TenantStartupMode::Attached((
329 0 : alc.attach_mode,
330 0 : alc.generation,
331 0 : ShardStripeSize::default(),
332 0 : )),
333 0 : LocationMode::Secondary(_) => TenantStartupMode::Secondary,
334 : },
335 : ))
336 0 : })
337 0 : .collect()
338 0 : }
339 :
340 0 : async fn init_load_generations(
341 0 : conf: &'static PageServerConf,
342 0 : tenant_confs: &HashMap<TenantShardId, Result<LocationConf, LoadConfigError>>,
343 0 : resources: &TenantSharedResources,
344 0 : cancel: &CancellationToken,
345 0 : ) -> anyhow::Result<Option<HashMap<TenantShardId, TenantStartupMode>>> {
346 0 : let generations = if conf.control_plane_emergency_mode {
347 0 : error!(
348 0 : "Emergency mode! Tenants will be attached unsafely using their last known generation"
349 : );
350 0 : emergency_generations(tenant_confs)
351 : } else {
352 0 : let client = StorageControllerUpcallClient::new(conf, cancel);
353 0 : info!("Calling {} API to re-attach tenants", client.base_url());
354 : // If we are configured to use the control plane API, then it is the source of truth for what tenants to load.
355 0 : match client.re_attach(conf).await {
356 0 : Ok(tenants) => tenants
357 0 : .into_iter()
358 0 : .flat_map(|(id, rart)| {
359 0 : TenantStartupMode::from_reattach_tenant(rart).map(|tsm| (id, tsm))
360 0 : })
361 0 : .collect(),
362 : Err(RetryForeverError::ShuttingDown) => {
363 0 : anyhow::bail!("Shut down while waiting for control plane re-attach response")
364 : }
365 : }
366 : };
367 :
368 : // The deletion queue needs to know about the startup attachment state to decide which (if any) stored
369 : // deletion list entries may still be valid. We provide that by pushing a recovery operation into
370 : // the queue. Sequential processing of te queue ensures that recovery is done before any new tenant deletions
371 : // are processed, even though we don't block on recovery completing here.
372 0 : let attached_tenants = generations
373 0 : .iter()
374 0 : .flat_map(|(id, start_mode)| {
375 0 : match start_mode {
376 0 : TenantStartupMode::Attached((_mode, generation, _stripe_size)) => Some(generation),
377 0 : TenantStartupMode::Secondary => None,
378 : }
379 0 : .map(|gen_| (*id, *gen_))
380 0 : })
381 0 : .collect();
382 0 : resources.deletion_queue_client.recover(attached_tenants)?;
383 :
384 0 : Ok(Some(generations))
385 0 : }
386 :
387 : /// Given a directory discovered in the pageserver's tenants/ directory, attempt
388 : /// to load a tenant config from it.
389 : ///
390 : /// If we cleaned up something expected (like an empty dir or a temp dir), return None.
391 0 : fn load_tenant_config(
392 0 : conf: &'static PageServerConf,
393 0 : tenant_shard_id: TenantShardId,
394 0 : dentry: Utf8DirEntry,
395 0 : ) -> Option<Result<LocationConf, LoadConfigError>> {
396 0 : let tenant_dir_path = dentry.path().to_path_buf();
397 0 : if crate::is_temporary(&tenant_dir_path) {
398 0 : info!("Found temporary tenant directory, removing: {tenant_dir_path}");
399 : // No need to use safe_remove_tenant_dir_all because this is already
400 : // a temporary path
401 0 : std::fs::remove_dir_all(&tenant_dir_path).fatal_err("delete temporary tenant dir");
402 0 : return None;
403 0 : }
404 :
405 : // This case happens if we crash during attachment before writing a config into the dir
406 0 : let is_empty = tenant_dir_path
407 0 : .is_empty_dir()
408 0 : .fatal_err("Checking for empty tenant dir");
409 0 : if is_empty {
410 0 : info!("removing empty tenant directory {tenant_dir_path:?}");
411 0 : std::fs::remove_dir(&tenant_dir_path).fatal_err("delete empty tenant dir");
412 0 : return None;
413 0 : }
414 :
415 0 : Some(TenantShard::load_tenant_config(conf, &tenant_shard_id))
416 0 : }
417 :
418 : /// Initial stage of load: walk the local tenants directory, clean up any temp files,
419 : /// and load configurations for the tenants we found.
420 : ///
421 : /// Do this in parallel, because we expect 10k+ tenants, so serial execution can take
422 : /// seconds even on reasonably fast drives.
423 0 : async fn init_load_tenant_configs(
424 0 : conf: &'static PageServerConf,
425 0 : ) -> HashMap<TenantShardId, Result<LocationConf, LoadConfigError>> {
426 0 : let tenants_dir = conf.tenants_path();
427 :
428 0 : let dentries = tokio::task::spawn_blocking(move || -> Vec<Utf8DirEntry> {
429 0 : let context = format!("read tenants dir {tenants_dir}");
430 0 : let dir_entries = tenants_dir.read_dir_utf8().fatal_err(&context);
431 :
432 0 : dir_entries
433 0 : .collect::<Result<Vec<_>, std::io::Error>>()
434 0 : .fatal_err(&context)
435 0 : })
436 0 : .await
437 0 : .expect("Config load task panicked");
438 :
439 0 : let mut configs = HashMap::new();
440 :
441 0 : let mut join_set = JoinSet::new();
442 0 : for dentry in dentries {
443 0 : let tenant_shard_id = match dentry.file_name().parse::<TenantShardId>() {
444 0 : Ok(id) => id,
445 : Err(_) => {
446 0 : warn!(
447 0 : "Invalid tenant path (garbage in our repo directory?): '{}'",
448 0 : dentry.file_name()
449 : );
450 0 : continue;
451 : }
452 : };
453 :
454 0 : join_set.spawn_blocking(move || {
455 0 : (
456 0 : tenant_shard_id,
457 0 : load_tenant_config(conf, tenant_shard_id, dentry),
458 0 : )
459 0 : });
460 : }
461 :
462 0 : while let Some(r) = join_set.join_next().await {
463 0 : let (tenant_shard_id, tenant_config) = r.expect("Panic in config load task");
464 0 : if let Some(tenant_config) = tenant_config {
465 0 : configs.insert(tenant_shard_id, tenant_config);
466 0 : }
467 : }
468 :
469 0 : configs
470 0 : }
471 :
472 : #[derive(Debug, thiserror::Error)]
473 : pub(crate) enum DeleteTenantError {
474 : #[error("Tenant map slot error {0}")]
475 : SlotError(#[from] TenantSlotError),
476 :
477 : #[error("Cancelled")]
478 : Cancelled,
479 :
480 : #[error(transparent)]
481 : Other(#[from] anyhow::Error),
482 : }
483 :
484 : /// Initialize repositories at `Initializing` state.
485 0 : pub fn init(
486 0 : conf: &'static PageServerConf,
487 0 : background_purges: BackgroundPurges,
488 0 : resources: TenantSharedResources,
489 0 : cancel: CancellationToken,
490 0 : ) -> TenantManager {
491 0 : TenantManager {
492 0 : conf,
493 0 : tenants: std::sync::RwLock::new(TenantsMap::Initializing),
494 0 : resources,
495 0 : cancel,
496 0 : background_purges,
497 0 : }
498 0 : }
499 :
500 : /// Transition repositories from `Initializing` state to `Open` state with locally available timelines.
501 : /// Timelines that are only partially available locally (remote storage has more data than this pageserver)
502 : /// are scheduled for download and added to the tenant once download is completed.
503 : #[instrument(skip_all)]
504 : pub async fn init_tenant_mgr(
505 : tenant_manager: Arc<TenantManager>,
506 : init_order: InitializationOrder,
507 : ) -> anyhow::Result<()> {
508 : debug_assert!(matches!(
509 : *tenant_manager.tenants.read().unwrap(),
510 : TenantsMap::Initializing
511 : ));
512 : let mut tenants = BTreeMap::new();
513 :
514 : let ctx = RequestContext::todo_child(TaskKind::Startup, DownloadBehavior::Warn);
515 :
516 : let conf = tenant_manager.conf;
517 : let resources = &tenant_manager.resources;
518 : let cancel = &tenant_manager.cancel;
519 : let background_purges = &tenant_manager.background_purges;
520 :
521 : // Initialize dynamic limits that depend on system resources
522 : let system_memory =
523 : sysinfo::System::new_with_specifics(sysinfo::RefreshKind::new().with_memory())
524 : .total_memory();
525 : let max_ephemeral_layer_bytes =
526 : conf.ephemeral_bytes_per_memory_kb as u64 * (system_memory / 1024);
527 : tracing::info!(
528 : "Initialized ephemeral layer size limit to {max_ephemeral_layer_bytes}, for {system_memory} bytes of memory"
529 : );
530 : inmemory_layer::GLOBAL_RESOURCES.max_dirty_bytes.store(
531 : max_ephemeral_layer_bytes,
532 : std::sync::atomic::Ordering::Relaxed,
533 : );
534 :
535 : // Scan local filesystem for attached tenants
536 : let tenant_configs = init_load_tenant_configs(conf).await;
537 :
538 : // Determine which tenants are to be secondary or attached, and in which generation
539 : let tenant_modes = init_load_generations(conf, &tenant_configs, resources, cancel).await?;
540 :
541 : // Hadron local SSD check: Raise an alert if our local filesystem does not contain any tenants but the re-attach request returned tenants.
542 : // This can happen if the PS suffered a Kubernetes node failure resulting in loss of all local data, but recovered quickly on another node
543 : // so the Storage Controller has not had the time to move tenants out.
544 : let data_loss_suspected = if let Some(tenant_modes) = &tenant_modes {
545 : tenant_configs.is_empty() && !tenant_modes.is_empty()
546 : } else {
547 : false
548 : };
549 : if data_loss_suspected {
550 : tracing::error!(
551 : "Local data loss suspected: no tenants found on local filesystem, but re-attach request returned tenants"
552 : );
553 : }
554 : LOCAL_DATA_LOSS_SUSPECTED.set(if data_loss_suspected { 1 } else { 0 });
555 :
556 : tracing::info!(
557 : "Attaching {} tenants at startup, warming up {} at a time",
558 : tenant_configs.len(),
559 : conf.concurrent_tenant_warmup.initial_permits()
560 : );
561 : TENANT.startup_scheduled.inc_by(tenant_configs.len() as u64);
562 :
563 : // Accumulate futures for writing tenant configs, so that we can execute in parallel
564 : let mut config_write_futs = Vec::new();
565 :
566 : // Update the location configs according to the re-attach response and persist them to disk
567 : tracing::info!("Updating {} location configs", tenant_configs.len());
568 : for (tenant_shard_id, location_conf) in tenant_configs {
569 : let tenant_dir_path = conf.tenant_path(&tenant_shard_id);
570 :
571 : let mut location_conf = match location_conf {
572 : Ok(l) => l,
573 : Err(e) => {
574 : // This should only happen in the case of a serialization bug or critical local I/O error: we cannot load this tenant
575 : error!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Failed to load tenant config, failed to {e:#}");
576 : continue;
577 : }
578 : };
579 :
580 : // FIXME: if we were attached, and get demoted to secondary on re-attach, we
581 : // don't have a place to get a config.
582 : // (https://github.com/neondatabase/neon/issues/5377)
583 : const DEFAULT_SECONDARY_CONF: SecondaryLocationConfig =
584 : SecondaryLocationConfig { warm: true };
585 :
586 : if let Some(tenant_modes) = &tenant_modes {
587 : // We have a generation map: treat it as the authority for whether
588 : // this tenant is really attached.
589 : match tenant_modes.get(&tenant_shard_id) {
590 : None => {
591 : 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");
592 :
593 : match safe_rename_tenant_dir(&tenant_dir_path).await {
594 : Ok(tmp_path) => {
595 : background_purges.spawn(tmp_path);
596 : }
597 : Err(e) => {
598 : error!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
599 : "Failed to move detached tenant directory '{tenant_dir_path}': {e:?}");
600 : }
601 : };
602 :
603 : // We deleted local content: move on to next tenant, don't try and spawn this one.
604 : continue;
605 : }
606 : Some(TenantStartupMode::Secondary) => {
607 : if !matches!(location_conf.mode, LocationMode::Secondary(_)) {
608 : location_conf.mode = LocationMode::Secondary(DEFAULT_SECONDARY_CONF);
609 : }
610 : }
611 : Some(TenantStartupMode::Attached((attach_mode, generation, stripe_size))) => {
612 : let old_gen_higher = match &location_conf.mode {
613 : LocationMode::Attached(AttachedLocationConfig {
614 : generation: old_generation,
615 : attach_mode: _attach_mode,
616 : }) => {
617 : if old_generation > generation {
618 : Some(old_generation)
619 : } else {
620 : None
621 : }
622 : }
623 : _ => None,
624 : };
625 : if let Some(old_generation) = old_gen_higher {
626 : tracing::error!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
627 : "Control plane gave decreasing generation ({generation:?}) in re-attach response for tenant that was attached in generation {:?}, demoting to secondary",
628 : old_generation
629 : );
630 :
631 : // We cannot safely attach this tenant given a bogus generation number, but let's avoid throwing away
632 : // local disk content: demote to secondary rather than detaching.
633 : location_conf.mode = LocationMode::Secondary(DEFAULT_SECONDARY_CONF);
634 : } else {
635 : location_conf.attach_in_generation(*attach_mode, *generation, *stripe_size);
636 : }
637 : }
638 : }
639 : } else {
640 : // Legacy mode: no generation information, any tenant present
641 : // on local disk may activate
642 : info!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Starting tenant in legacy mode, no generation",);
643 : };
644 :
645 : // Presence of a generation number implies attachment: attach the tenant
646 : // if it wasn't already, and apply the generation number.
647 0 : config_write_futs.push(async move {
648 0 : let r =
649 0 : TenantShard::persist_tenant_config(conf, &tenant_shard_id, &location_conf).await;
650 0 : (tenant_shard_id, location_conf, r)
651 0 : });
652 : }
653 :
654 : // Execute config writes with concurrency, to avoid bottlenecking on local FS write latency
655 : tracing::info!(
656 : "Writing {} location config files...",
657 : config_write_futs.len()
658 : );
659 : let config_write_results = futures::stream::iter(config_write_futs)
660 : .buffer_unordered(16)
661 : .collect::<Vec<_>>()
662 : .await;
663 :
664 : tracing::info!(
665 : "Spawning {} tenant shard locations...",
666 : config_write_results.len()
667 : );
668 : // For those shards that have live configurations, construct `Tenant` or `SecondaryTenant` objects and start them running
669 : for (tenant_shard_id, location_conf, config_write_result) in config_write_results {
670 : // Writing a config to local disk is foundational to startup up tenants: panic if we can't.
671 : config_write_result.fatal_err("write tenant shard config file");
672 :
673 : let tenant_dir_path = conf.tenant_path(&tenant_shard_id);
674 : let shard_identity = location_conf.shard;
675 : let slot = match location_conf.mode {
676 : LocationMode::Attached(attached_conf) => TenantSlot::Attached(
677 : tenant_spawn(
678 : conf,
679 : tenant_shard_id,
680 : &tenant_dir_path,
681 : resources.clone(),
682 : AttachedTenantConf::new(conf, location_conf.tenant_conf, attached_conf),
683 : shard_identity,
684 : Some(init_order.clone()),
685 : SpawnMode::Lazy,
686 : &ctx,
687 : )
688 : .expect("global shutdown during init_tenant_mgr cannot happen"),
689 : ),
690 : LocationMode::Secondary(secondary_conf) => {
691 : info!(
692 : tenant_id = %tenant_shard_id.tenant_id,
693 : shard_id = %tenant_shard_id.shard_slug(),
694 : "Starting secondary tenant"
695 : );
696 : TenantSlot::Secondary(SecondaryTenant::new(
697 : tenant_shard_id,
698 : shard_identity,
699 : location_conf.tenant_conf,
700 : &secondary_conf,
701 : ))
702 : }
703 : };
704 :
705 : METRICS.slot_inserted(&slot);
706 : tenants.insert(tenant_shard_id, slot);
707 : }
708 :
709 : info!("Processed {} local tenants at startup", tenants.len());
710 :
711 : let mut tenant_map = tenant_manager.tenants.write().unwrap();
712 : *tenant_map = TenantsMap::Open(tenants);
713 :
714 : Ok(())
715 : }
716 :
717 : /// Wrapper for Tenant::spawn that checks invariants before running
718 : #[allow(clippy::too_many_arguments)]
719 0 : fn tenant_spawn(
720 0 : conf: &'static PageServerConf,
721 0 : tenant_shard_id: TenantShardId,
722 0 : tenant_path: &Utf8Path,
723 0 : resources: TenantSharedResources,
724 0 : location_conf: AttachedTenantConf,
725 0 : shard_identity: ShardIdentity,
726 0 : init_order: Option<InitializationOrder>,
727 0 : mode: SpawnMode,
728 0 : ctx: &RequestContext,
729 0 : ) -> Result<Arc<TenantShard>, GlobalShutDown> {
730 : // All these conditions should have been satisfied by our caller: the tenant dir exists, is a well formed
731 : // path, and contains a configuration file. Assertions that do synchronous I/O are limited to debug mode
732 : // to avoid impacting prod runtime performance.
733 0 : assert!(!crate::is_temporary(tenant_path));
734 0 : debug_assert!(tenant_path.is_dir());
735 0 : debug_assert!(
736 0 : conf.tenant_location_config_path(&tenant_shard_id)
737 0 : .try_exists()
738 0 : .unwrap()
739 : );
740 :
741 0 : TenantShard::spawn(
742 0 : conf,
743 0 : tenant_shard_id,
744 0 : resources,
745 0 : location_conf,
746 0 : shard_identity,
747 0 : init_order,
748 0 : mode,
749 0 : ctx,
750 : )
751 0 : }
752 :
753 : #[derive(thiserror::Error, Debug)]
754 : pub(crate) enum UpsertLocationError {
755 : #[error("Bad config request: {0}")]
756 : BadRequest(anyhow::Error),
757 :
758 : #[error("Cannot change config in this state: {0}")]
759 : Unavailable(#[from] TenantMapError),
760 :
761 : #[error("Tenant is already being modified")]
762 : InProgress,
763 :
764 : #[error("Failed to flush: {0}")]
765 : Flush(anyhow::Error),
766 :
767 : /// This error variant is for unexpected situations (soft assertions) where the system is in an unexpected state.
768 : #[error("Internal error: {0}")]
769 : InternalError(anyhow::Error),
770 : }
771 :
772 : impl TenantManager {
773 : /// Convenience function so that anyone with a TenantManager can get at the global configuration, without
774 : /// having to pass it around everywhere as a separate object.
775 0 : pub(crate) fn get_conf(&self) -> &'static PageServerConf {
776 0 : self.conf
777 0 : }
778 :
779 : /// Gets the attached tenant from the in-memory data, erroring if it's absent, in secondary mode, or currently
780 : /// undergoing a state change (i.e. slot is InProgress).
781 : ///
782 : /// The return TenantShard is not guaranteed to be active: check its status after obtaing it, or
783 : /// use [`TenantShard::wait_to_become_active`] before using it if you will do I/O on it.
784 0 : pub(crate) fn get_attached_tenant_shard(
785 0 : &self,
786 0 : tenant_shard_id: TenantShardId,
787 0 : ) -> Result<Arc<TenantShard>, GetTenantError> {
788 0 : let locked = self.tenants.read().unwrap();
789 :
790 0 : let peek_slot = tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Read)?;
791 :
792 0 : match peek_slot {
793 0 : Some(TenantSlot::Attached(tenant)) => Ok(Arc::clone(tenant)),
794 0 : Some(TenantSlot::InProgress(_)) => Err(GetTenantError::NotActive(tenant_shard_id)),
795 : None | Some(TenantSlot::Secondary(_)) => {
796 0 : Err(GetTenantError::ShardNotFound(tenant_shard_id))
797 : }
798 : }
799 0 : }
800 :
801 0 : pub(crate) fn get_secondary_tenant_shard(
802 0 : &self,
803 0 : tenant_shard_id: TenantShardId,
804 0 : ) -> Option<Arc<SecondaryTenant>> {
805 0 : let locked = self.tenants.read().unwrap();
806 :
807 0 : let peek_slot = tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Read)
808 0 : .ok()
809 0 : .flatten();
810 :
811 0 : match peek_slot {
812 0 : Some(TenantSlot::Secondary(s)) => Some(s.clone()),
813 0 : _ => None,
814 : }
815 0 : }
816 :
817 : /// Whether the `TenantManager` is responsible for the tenant shard
818 0 : pub(crate) fn manages_tenant_shard(&self, tenant_shard_id: TenantShardId) -> bool {
819 0 : let locked = self.tenants.read().unwrap();
820 :
821 0 : let peek_slot = tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Read)
822 0 : .ok()
823 0 : .flatten();
824 :
825 0 : peek_slot.is_some()
826 0 : }
827 :
828 : #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
829 : pub(crate) async fn upsert_location(
830 : &self,
831 : tenant_shard_id: TenantShardId,
832 : new_location_config: LocationConf,
833 : flush: Option<Duration>,
834 : mut spawn_mode: SpawnMode,
835 : ctx: &RequestContext,
836 : ) -> Result<Option<Arc<TenantShard>>, UpsertLocationError> {
837 : debug_assert_current_span_has_tenant_id();
838 : info!("configuring tenant location to state {new_location_config:?}");
839 :
840 : enum FastPathModified {
841 : Attached(Arc<TenantShard>),
842 : Secondary(Arc<SecondaryTenant>),
843 : }
844 :
845 : // Special case fast-path for updates to existing slots: if our upsert is only updating configuration,
846 : // then we do not need to set the slot to InProgress, we can just call into the
847 : // existng tenant.
848 : let fast_path_taken = {
849 : let locked = self.tenants.read().unwrap();
850 : let peek_slot =
851 : tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Write)?;
852 : match (&new_location_config.mode, peek_slot) {
853 : (LocationMode::Attached(attach_conf), Some(TenantSlot::Attached(tenant))) => {
854 : match attach_conf.generation.cmp(&tenant.generation) {
855 : Ordering::Equal => {
856 : // A transition from Attached to Attached in the same generation, we may
857 : // take our fast path and just provide the updated configuration
858 : // to the tenant.
859 : tenant.set_new_location_config(
860 : AttachedTenantConf::try_from(
861 : self.conf,
862 : new_location_config.clone(),
863 : )
864 : .map_err(UpsertLocationError::BadRequest)?,
865 : );
866 :
867 : Some(FastPathModified::Attached(tenant.clone()))
868 : }
869 : Ordering::Less => {
870 : return Err(UpsertLocationError::BadRequest(anyhow::anyhow!(
871 : "Generation {:?} is less than existing {:?}",
872 : attach_conf.generation,
873 : tenant.generation
874 : )));
875 : }
876 : Ordering::Greater => {
877 : // Generation advanced, fall through to general case of replacing `Tenant` object
878 : None
879 : }
880 : }
881 : }
882 : (
883 : LocationMode::Secondary(secondary_conf),
884 : Some(TenantSlot::Secondary(secondary_tenant)),
885 : ) => {
886 : secondary_tenant.set_config(secondary_conf);
887 : secondary_tenant.set_tenant_conf(&new_location_config.tenant_conf);
888 : Some(FastPathModified::Secondary(secondary_tenant.clone()))
889 : }
890 : _ => {
891 : // Not an Attached->Attached transition, fall through to general case
892 : None
893 : }
894 : }
895 : };
896 :
897 : // Fast-path continued: having dropped out of the self.tenants lock, do the async
898 : // phase of writing config and/or waiting for flush, before returning.
899 : match fast_path_taken {
900 : Some(FastPathModified::Attached(tenant)) => {
901 : tenant
902 : .shard_identity
903 : .assert_equal(new_location_config.shard);
904 : TenantShard::persist_tenant_config(
905 : self.conf,
906 : &tenant_shard_id,
907 : &new_location_config,
908 : )
909 : .await
910 : .fatal_err("write tenant shard config");
911 :
912 : // Transition to AttachedStale means we may well hold a valid generation
913 : // still, and have been requested to go stale as part of a migration. If
914 : // the caller set `flush`, then flush to remote storage.
915 : if let LocationMode::Attached(AttachedLocationConfig {
916 : generation: _,
917 : attach_mode: AttachmentMode::Stale,
918 : }) = &new_location_config.mode
919 : {
920 : if let Some(flush_timeout) = flush {
921 : match tokio::time::timeout(flush_timeout, tenant.flush_remote()).await {
922 : Ok(Err(e)) => {
923 : return Err(UpsertLocationError::Flush(e));
924 : }
925 : Ok(Ok(_)) => return Ok(Some(tenant)),
926 : Err(_) => {
927 : tracing::warn!(
928 : timeout_ms = flush_timeout.as_millis(),
929 : "Timed out waiting for flush to remote storage, proceeding anyway."
930 : )
931 : }
932 : }
933 : }
934 : }
935 :
936 : return Ok(Some(tenant));
937 : }
938 : Some(FastPathModified::Secondary(secondary_tenant)) => {
939 : secondary_tenant
940 : .shard_identity
941 : .assert_equal(new_location_config.shard);
942 : TenantShard::persist_tenant_config(
943 : self.conf,
944 : &tenant_shard_id,
945 : &new_location_config,
946 : )
947 : .await
948 : .fatal_err("write tenant shard config");
949 :
950 : return Ok(None);
951 : }
952 : None => {
953 : // Proceed with the general case procedure, where we will shutdown & remove any existing
954 : // slot contents and replace with a fresh one
955 : }
956 : };
957 :
958 : // General case for upserts to TenantsMap, excluding the case above: we will substitute an
959 : // InProgress value to the slot while we make whatever changes are required. The state for
960 : // the tenant is inaccessible to the outside world while we are doing this, but that is sensible:
961 : // the state is ill-defined while we're in transition. Transitions are async, but fast: we do
962 : // not do significant I/O, and shutdowns should be prompt via cancellation tokens.
963 : let mut slot_guard = self
964 : .tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)
965 0 : .map_err(|e| match e {
966 : TenantSlotError::NotFound(_) => {
967 0 : unreachable!("Called with mode Any")
968 : }
969 0 : TenantSlotError::InProgress => UpsertLocationError::InProgress,
970 0 : TenantSlotError::MapState(s) => UpsertLocationError::Unavailable(s),
971 0 : })?;
972 :
973 : match slot_guard.get_old_value() {
974 : Some(TenantSlot::Attached(tenant)) => {
975 : tenant
976 : .shard_identity
977 : .assert_equal(new_location_config.shard);
978 :
979 : // The case where we keep a Tenant alive was covered above in the special case
980 : // for Attached->Attached transitions in the same generation. By this point,
981 : // if we see an attached tenant we know it will be discarded and should be
982 : // shut down.
983 : let (_guard, progress) = utils::completion::channel();
984 :
985 : match tenant.get_attach_mode() {
986 : AttachmentMode::Single | AttachmentMode::Multi => {
987 : // Before we leave our state as the presumed holder of the latest generation,
988 : // flush any outstanding deletions to reduce the risk of leaking objects.
989 : self.resources.deletion_queue_client.flush_advisory()
990 : }
991 : AttachmentMode::Stale => {
992 : // If we're stale there's not point trying to flush deletions
993 : }
994 : };
995 :
996 : info!("Shutting down attached tenant");
997 : match tenant.shutdown(progress, ShutdownMode::Hard).await {
998 : Ok(()) => {}
999 : Err(barrier) => {
1000 : info!("Shutdown already in progress, waiting for it to complete");
1001 : barrier.wait().await;
1002 : }
1003 : }
1004 : slot_guard.drop_old_value().expect("We just shut it down");
1005 :
1006 : // Edge case: if we were called with SpawnMode::Create, but a Tenant already existed, then
1007 : // the caller thinks they're creating but the tenant already existed. We must switch to
1008 : // Eager mode so that when starting this Tenant we properly probe remote storage for timelines,
1009 : // rather than assuming it to be empty.
1010 : spawn_mode = SpawnMode::Eager;
1011 : }
1012 : Some(TenantSlot::Secondary(secondary_tenant)) => {
1013 : secondary_tenant
1014 : .shard_identity
1015 : .assert_equal(new_location_config.shard);
1016 :
1017 : info!("Shutting down secondary tenant");
1018 : secondary_tenant.shutdown().await;
1019 : }
1020 : Some(TenantSlot::InProgress(_)) => {
1021 : // This should never happen: acquire_slot should error out
1022 : // if the contents of a slot were InProgress.
1023 : return Err(UpsertLocationError::InternalError(anyhow::anyhow!(
1024 : "Acquired an InProgress slot, this is a bug."
1025 : )));
1026 : }
1027 : None => {
1028 : // Slot was vacant, nothing needs shutting down.
1029 : }
1030 : }
1031 :
1032 : let tenant_path = self.conf.tenant_path(&tenant_shard_id);
1033 : let timelines_path = self.conf.timelines_path(&tenant_shard_id);
1034 :
1035 : // Directory structure is the same for attached and secondary modes:
1036 : // create it if it doesn't exist. Timeline load/creation expects the
1037 : // timelines/ subdir to already exist.
1038 : //
1039 : // Does not need to be fsync'd because local storage is just a cache.
1040 : tokio::fs::create_dir_all(&timelines_path)
1041 : .await
1042 : .fatal_err("create timelines/ dir");
1043 :
1044 : // Before activating either secondary or attached mode, persist the
1045 : // configuration, so that on restart we will re-attach (or re-start
1046 : // secondary) on the tenant.
1047 : TenantShard::persist_tenant_config(self.conf, &tenant_shard_id, &new_location_config)
1048 : .await
1049 : .fatal_err("write tenant shard config");
1050 :
1051 : let new_slot = match &new_location_config.mode {
1052 : LocationMode::Secondary(secondary_config) => {
1053 : let shard_identity = new_location_config.shard;
1054 : TenantSlot::Secondary(SecondaryTenant::new(
1055 : tenant_shard_id,
1056 : shard_identity,
1057 : new_location_config.tenant_conf,
1058 : secondary_config,
1059 : ))
1060 : }
1061 : LocationMode::Attached(_attach_config) => {
1062 : let shard_identity = new_location_config.shard;
1063 :
1064 : // Testing hack: if we are configured with no control plane, then drop the generation
1065 : // from upserts. This enables creating generation-less tenants even though neon_local
1066 : // always uses generations when calling the location conf API.
1067 : let attached_conf = AttachedTenantConf::try_from(self.conf, new_location_config)
1068 : .map_err(UpsertLocationError::BadRequest)?;
1069 :
1070 : let tenant = tenant_spawn(
1071 : self.conf,
1072 : tenant_shard_id,
1073 : &tenant_path,
1074 : self.resources.clone(),
1075 : attached_conf,
1076 : shard_identity,
1077 : None,
1078 : spawn_mode,
1079 : ctx,
1080 : )
1081 0 : .map_err(|_: GlobalShutDown| {
1082 0 : UpsertLocationError::Unavailable(TenantMapError::ShuttingDown)
1083 0 : })?;
1084 :
1085 : TenantSlot::Attached(tenant)
1086 : }
1087 : };
1088 :
1089 : let attached_tenant = if let TenantSlot::Attached(tenant) = &new_slot {
1090 : Some(tenant.clone())
1091 : } else {
1092 : None
1093 : };
1094 :
1095 : match slot_guard.upsert(new_slot) {
1096 : Err(TenantSlotUpsertError::InternalError(e)) => {
1097 : Err(UpsertLocationError::InternalError(anyhow::anyhow!(e)))
1098 : }
1099 : Err(TenantSlotUpsertError::MapState(e)) => Err(UpsertLocationError::Unavailable(e)),
1100 : Err(TenantSlotUpsertError::ShuttingDown((new_slot, _completion))) => {
1101 : // If we just called tenant_spawn() on a new tenant, and can't insert it into our map, then
1102 : // we must not leak it: this would violate the invariant that after shutdown_all_tenants, all tenants
1103 : // are shutdown.
1104 : //
1105 : // We must shut it down inline here.
1106 : match new_slot {
1107 : TenantSlot::InProgress(_) => {
1108 : // Unreachable because we never insert an InProgress
1109 : unreachable!()
1110 : }
1111 : TenantSlot::Attached(tenant) => {
1112 : let (_guard, progress) = utils::completion::channel();
1113 : info!(
1114 : "Shutting down just-spawned tenant, because tenant manager is shut down"
1115 : );
1116 : match tenant.shutdown(progress, ShutdownMode::Hard).await {
1117 : Ok(()) => {
1118 : info!("Finished shutting down just-spawned tenant");
1119 : }
1120 : Err(barrier) => {
1121 : info!("Shutdown already in progress, waiting for it to complete");
1122 : barrier.wait().await;
1123 : }
1124 : }
1125 : }
1126 : TenantSlot::Secondary(secondary_tenant) => {
1127 : secondary_tenant.shutdown().await;
1128 : }
1129 : }
1130 :
1131 : Err(UpsertLocationError::Unavailable(
1132 : TenantMapError::ShuttingDown,
1133 : ))
1134 : }
1135 : Ok(()) => Ok(attached_tenant),
1136 : }
1137 : }
1138 :
1139 1 : fn tenant_map_acquire_slot(
1140 1 : &self,
1141 1 : tenant_shard_id: &TenantShardId,
1142 1 : mode: TenantSlotAcquireMode,
1143 1 : ) -> Result<SlotGuard, TenantSlotError> {
1144 : use TenantSlotAcquireMode::*;
1145 1 : METRICS.tenant_slot_writes.inc();
1146 :
1147 1 : let mut locked = self.tenants.write().unwrap();
1148 1 : let span = tracing::info_span!("acquire_slot", tenant_id=%tenant_shard_id.tenant_id, shard_id = %tenant_shard_id.shard_slug());
1149 1 : let _guard = span.enter();
1150 :
1151 1 : let m = match &mut *locked {
1152 0 : TenantsMap::Initializing => return Err(TenantMapError::StillInitializing.into()),
1153 0 : TenantsMap::ShuttingDown(_) => return Err(TenantMapError::ShuttingDown.into()),
1154 1 : TenantsMap::Open(m) => m,
1155 : };
1156 :
1157 : use std::collections::btree_map::Entry;
1158 :
1159 1 : let entry = m.entry(*tenant_shard_id);
1160 :
1161 1 : match entry {
1162 0 : Entry::Vacant(v) => match mode {
1163 : MustExist => {
1164 0 : tracing::debug!("Vacant && MustExist: return NotFound");
1165 0 : Err(TenantSlotError::NotFound(*tenant_shard_id))
1166 : }
1167 : _ => {
1168 0 : let (completion, barrier) = utils::completion::channel();
1169 0 : let inserting = TenantSlot::InProgress(barrier);
1170 0 : METRICS.slot_inserted(&inserting);
1171 0 : v.insert(inserting);
1172 0 : tracing::debug!("Vacant, inserted InProgress");
1173 0 : Ok(SlotGuard::new(
1174 0 : *tenant_shard_id,
1175 0 : None,
1176 0 : completion,
1177 0 : &self.tenants,
1178 0 : ))
1179 : }
1180 : },
1181 1 : Entry::Occupied(mut o) => {
1182 : // Apply mode-driven checks
1183 1 : match (o.get(), mode) {
1184 : (TenantSlot::InProgress(_), _) => {
1185 0 : tracing::debug!("Occupied, failing for InProgress");
1186 0 : Err(TenantSlotError::InProgress)
1187 : }
1188 : _ => {
1189 : // Happy case: the slot was not in any state that violated our mode
1190 1 : let (completion, barrier) = utils::completion::channel();
1191 1 : let in_progress = TenantSlot::InProgress(barrier);
1192 1 : METRICS.slot_inserted(&in_progress);
1193 1 : let old_value = o.insert(in_progress);
1194 1 : METRICS.slot_removed(&old_value);
1195 1 : tracing::debug!("Occupied, replaced with InProgress");
1196 1 : Ok(SlotGuard::new(
1197 1 : *tenant_shard_id,
1198 1 : Some(old_value),
1199 1 : completion,
1200 1 : &self.tenants,
1201 1 : ))
1202 : }
1203 : }
1204 : }
1205 : }
1206 1 : }
1207 :
1208 : /// Resetting a tenant is equivalent to detaching it, then attaching it again with the same
1209 : /// LocationConf that was last used to attach it. Optionally, the local file cache may be
1210 : /// dropped before re-attaching.
1211 : ///
1212 : /// This is not part of a tenant's normal lifecycle: it is used for debug/support, in situations
1213 : /// where an issue is identified that would go away with a restart of the tenant.
1214 : ///
1215 : /// This does not have any special "force" shutdown of a tenant: it relies on the tenant's tasks
1216 : /// to respect the cancellation tokens used in normal shutdown().
1217 : #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), %drop_cache))]
1218 : pub(crate) async fn reset_tenant(
1219 : &self,
1220 : tenant_shard_id: TenantShardId,
1221 : drop_cache: bool,
1222 : ctx: &RequestContext,
1223 : ) -> anyhow::Result<()> {
1224 : let mut slot_guard =
1225 : self.tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)?;
1226 : let Some(old_slot) = slot_guard.get_old_value() else {
1227 : anyhow::bail!("Tenant not found when trying to reset");
1228 : };
1229 :
1230 : let Some(tenant) = old_slot.get_attached() else {
1231 : slot_guard.revert();
1232 : anyhow::bail!("Tenant is not in attached state");
1233 : };
1234 :
1235 : let (_guard, progress) = utils::completion::channel();
1236 : match tenant.shutdown(progress, ShutdownMode::Hard).await {
1237 : Ok(()) => {
1238 : slot_guard.drop_old_value()?;
1239 : }
1240 : Err(_barrier) => {
1241 : slot_guard.revert();
1242 : anyhow::bail!("Cannot reset Tenant, already shutting down");
1243 : }
1244 : }
1245 :
1246 : let tenant_path = self.conf.tenant_path(&tenant_shard_id);
1247 : let timelines_path = self.conf.timelines_path(&tenant_shard_id);
1248 : let config = TenantShard::load_tenant_config(self.conf, &tenant_shard_id)?;
1249 :
1250 : if drop_cache {
1251 : tracing::info!("Dropping local file cache");
1252 :
1253 : match tokio::fs::read_dir(&timelines_path).await {
1254 : Err(e) => {
1255 : tracing::warn!("Failed to list timelines while dropping cache: {}", e);
1256 : }
1257 : Ok(mut entries) => {
1258 : while let Some(entry) = entries.next_entry().await? {
1259 : tokio::fs::remove_dir_all(entry.path()).await?;
1260 : }
1261 : }
1262 : }
1263 : }
1264 :
1265 : let shard_identity = config.shard;
1266 : let tenant = tenant_spawn(
1267 : self.conf,
1268 : tenant_shard_id,
1269 : &tenant_path,
1270 : self.resources.clone(),
1271 : AttachedTenantConf::try_from(self.conf, config)?,
1272 : shard_identity,
1273 : None,
1274 : SpawnMode::Eager,
1275 : ctx,
1276 : )?;
1277 :
1278 : slot_guard.upsert(TenantSlot::Attached(tenant))?;
1279 :
1280 : Ok(())
1281 : }
1282 :
1283 0 : pub(crate) fn get_attached_active_tenant_shards(&self) -> Vec<Arc<TenantShard>> {
1284 0 : let locked = self.tenants.read().unwrap();
1285 0 : match &*locked {
1286 0 : TenantsMap::Initializing => Vec::new(),
1287 0 : TenantsMap::Open(map) | TenantsMap::ShuttingDown(map) => map
1288 0 : .values()
1289 0 : .filter_map(|slot| {
1290 0 : slot.get_attached()
1291 0 : .and_then(|t| if t.is_active() { Some(t.clone()) } else { None })
1292 0 : })
1293 0 : .collect(),
1294 : }
1295 0 : }
1296 : // Do some synchronous work for all tenant slots in Secondary state. The provided
1297 : // callback should be small and fast, as it will be called inside the global
1298 : // TenantsMap lock.
1299 0 : pub(crate) fn foreach_secondary_tenants<F>(&self, mut func: F)
1300 0 : where
1301 0 : // TODO: let the callback return a hint to drop out of the loop early
1302 0 : F: FnMut(&TenantShardId, &Arc<SecondaryTenant>),
1303 : {
1304 0 : let locked = self.tenants.read().unwrap();
1305 :
1306 0 : let map = match &*locked {
1307 0 : TenantsMap::Initializing | TenantsMap::ShuttingDown(_) => return,
1308 0 : TenantsMap::Open(m) => m,
1309 : };
1310 :
1311 0 : for (tenant_id, slot) in map {
1312 0 : if let TenantSlot::Secondary(state) = slot {
1313 : // Only expose secondary tenants that are not currently shutting down
1314 0 : if !state.cancel.is_cancelled() {
1315 0 : func(tenant_id, state)
1316 0 : }
1317 0 : }
1318 : }
1319 0 : }
1320 :
1321 : /// Total list of all tenant slots: this includes attached, secondary, and InProgress.
1322 0 : pub(crate) fn list(&self) -> Vec<(TenantShardId, TenantSlot)> {
1323 0 : let locked = self.tenants.read().unwrap();
1324 0 : match &*locked {
1325 0 : TenantsMap::Initializing => Vec::new(),
1326 0 : TenantsMap::Open(map) | TenantsMap::ShuttingDown(map) => {
1327 0 : map.iter().map(|(k, v)| (*k, v.clone())).collect()
1328 : }
1329 : }
1330 0 : }
1331 :
1332 0 : pub(crate) fn get(&self, tenant_shard_id: TenantShardId) -> Option<TenantSlot> {
1333 0 : let locked = self.tenants.read().unwrap();
1334 0 : match &*locked {
1335 0 : TenantsMap::Initializing => None,
1336 0 : TenantsMap::Open(map) | TenantsMap::ShuttingDown(map) => {
1337 0 : map.get(&tenant_shard_id).cloned()
1338 : }
1339 : }
1340 0 : }
1341 :
1342 : /// If a tenant is attached, detach it. Then remove its data from remote storage.
1343 : ///
1344 : /// A tenant is considered deleted once it is gone from remote storage. It is the caller's
1345 : /// responsibility to avoid trying to attach the tenant again or use it any way once deletion
1346 : /// has started: this operation is not atomic, and must be retried until it succeeds.
1347 : ///
1348 : /// As a special case, if an unsharded tenant ID is given for a sharded tenant, it will remove
1349 : /// all tenant shards in remote storage (removing all paths with the tenant prefix). The storage
1350 : /// controller uses this to purge all remote tenant data, including any stale parent shards that
1351 : /// may remain after splits. Ideally, this special case would be handled elsewhere. See:
1352 : /// <https://github.com/neondatabase/neon/pull/9394>.
1353 0 : pub(crate) async fn delete_tenant(
1354 0 : &self,
1355 0 : tenant_shard_id: TenantShardId,
1356 0 : ) -> Result<(), DeleteTenantError> {
1357 0 : super::span::debug_assert_current_span_has_tenant_id();
1358 :
1359 0 : async fn delete_local(
1360 0 : conf: &PageServerConf,
1361 0 : background_purges: &BackgroundPurges,
1362 0 : tenant_shard_id: &TenantShardId,
1363 0 : ) -> anyhow::Result<()> {
1364 0 : let local_tenant_directory = conf.tenant_path(tenant_shard_id);
1365 0 : let tmp_dir = safe_rename_tenant_dir(&local_tenant_directory)
1366 0 : .await
1367 0 : .with_context(|| {
1368 0 : format!("local tenant directory {local_tenant_directory:?} rename")
1369 0 : })?;
1370 0 : background_purges.spawn(tmp_dir);
1371 0 : Ok(())
1372 0 : }
1373 :
1374 0 : let slot_guard =
1375 0 : self.tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)?;
1376 0 : match &slot_guard.old_value {
1377 0 : Some(TenantSlot::Attached(tenant)) => {
1378 : // Legacy deletion flow: the tenant remains attached, goes to Stopping state, and
1379 : // deletion will be resumed across restarts.
1380 0 : let tenant = tenant.clone();
1381 0 : let (_guard, progress) = utils::completion::channel();
1382 0 : match tenant.shutdown(progress, ShutdownMode::Hard).await {
1383 0 : Ok(()) => {}
1384 0 : Err(barrier) => {
1385 0 : info!("Shutdown already in progress, waiting for it to complete");
1386 0 : barrier.wait().await;
1387 : }
1388 : }
1389 0 : delete_local(self.conf, &self.background_purges, &tenant_shard_id).await?;
1390 : }
1391 0 : Some(TenantSlot::Secondary(secondary_tenant)) => {
1392 0 : secondary_tenant.shutdown().await;
1393 :
1394 0 : delete_local(self.conf, &self.background_purges, &tenant_shard_id).await?;
1395 : }
1396 0 : Some(TenantSlot::InProgress(_)) => unreachable!(),
1397 0 : None => {}
1398 : };
1399 :
1400 : // Fall through: local state for this tenant is no longer present, proceed with remote delete.
1401 : // - We use a retry wrapper here so that common transient S3 errors (e.g. 503, 429) do not result
1402 : // in 500 responses to delete requests.
1403 : // - We keep the `SlotGuard` during this I/O, so that if a concurrent delete request comes in, it will
1404 : // 503/retry, rather than kicking off a wasteful concurrent deletion.
1405 : // NB: this also deletes partial prefixes, i.e. a <tenant_id> path will delete all
1406 : // <tenant_id>_<shard_id>/* objects. See method comment for why.
1407 0 : backoff::retry(
1408 0 : || async move {
1409 0 : self.resources
1410 0 : .remote_storage
1411 0 : .delete_prefix(&remote_tenant_path(&tenant_shard_id), &self.cancel)
1412 0 : .await
1413 0 : },
1414 : |_| false, // backoff::retry handles cancellation
1415 : 1,
1416 : 3,
1417 0 : &format!("delete_tenant[tenant_shard_id={tenant_shard_id}]"),
1418 0 : &self.cancel,
1419 : )
1420 0 : .await
1421 0 : .unwrap_or(Err(TimeoutOrCancel::Cancel.into()))
1422 0 : .map_err(|err| {
1423 0 : if TimeoutOrCancel::caused_by_cancel(&err) {
1424 0 : return DeleteTenantError::Cancelled;
1425 0 : }
1426 0 : DeleteTenantError::Other(err)
1427 0 : })
1428 0 : }
1429 :
1430 : #[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()))]
1431 : pub(crate) async fn shard_split(
1432 : &self,
1433 : tenant: Arc<TenantShard>,
1434 : new_shard_count: ShardCount,
1435 : new_stripe_size: Option<ShardStripeSize>,
1436 : ctx: &RequestContext,
1437 : ) -> anyhow::Result<Vec<TenantShardId>> {
1438 : let tenant_shard_id = *tenant.get_tenant_shard_id();
1439 : let r = self
1440 : .do_shard_split(tenant, new_shard_count, new_stripe_size, ctx)
1441 : .await;
1442 : if r.is_err() {
1443 : // Shard splitting might have left the original shard in a partially shut down state (it
1444 : // stops the shard's remote timeline client). Reset it to ensure we leave things in
1445 : // a working state.
1446 : if self.get(tenant_shard_id).is_some() {
1447 : tracing::warn!("Resetting after shard split failure");
1448 : if let Err(e) = self.reset_tenant(tenant_shard_id, false, ctx).await {
1449 : // Log this error because our return value will still be the original error, not this one. This is
1450 : // a severe error: if this happens, we might be leaving behind a tenant that is not fully functional
1451 : // (e.g. has uploads disabled). We can't do anything else: if reset fails then shutting the tenant down or
1452 : // setting it broken probably won't help either.
1453 : tracing::error!("Failed to reset: {e}");
1454 : }
1455 : }
1456 : }
1457 :
1458 : r
1459 : }
1460 :
1461 0 : pub(crate) async fn do_shard_split(
1462 0 : &self,
1463 0 : tenant: Arc<TenantShard>,
1464 0 : new_shard_count: ShardCount,
1465 0 : new_stripe_size: Option<ShardStripeSize>,
1466 0 : ctx: &RequestContext,
1467 0 : ) -> anyhow::Result<Vec<TenantShardId>> {
1468 0 : let tenant_shard_id = *tenant.get_tenant_shard_id();
1469 :
1470 : // Validate the incoming request
1471 0 : if new_shard_count.count() <= tenant_shard_id.shard_count.count() {
1472 0 : anyhow::bail!("Requested shard count is not an increase");
1473 0 : }
1474 0 : let expansion_factor = new_shard_count.count() / tenant_shard_id.shard_count.count();
1475 0 : if !expansion_factor.is_power_of_two() {
1476 0 : anyhow::bail!("Requested split is not a power of two");
1477 0 : }
1478 :
1479 0 : if let Some(new_stripe_size) = new_stripe_size {
1480 0 : if tenant.get_shard_stripe_size() != new_stripe_size
1481 0 : && tenant_shard_id.shard_count.count() > 1
1482 : {
1483 : // This tenant already has multiple shards, it is illegal to try and change its stripe size
1484 0 : anyhow::bail!(
1485 0 : "Shard stripe size may not be modified once tenant has multiple shards"
1486 : );
1487 0 : }
1488 0 : }
1489 :
1490 : // Plan: identify what the new child shards will be
1491 0 : let child_shards = tenant_shard_id.split(new_shard_count);
1492 0 : tracing::info!(
1493 0 : "Shard {} splits into: {}",
1494 0 : tenant_shard_id.to_index(),
1495 0 : child_shards
1496 0 : .iter()
1497 0 : .map(|id| format!("{}", id.to_index()))
1498 0 : .join(",")
1499 : );
1500 :
1501 0 : fail::fail_point!("shard-split-pre-prepare", |_| Err(anyhow::anyhow!(
1502 0 : "failpoint"
1503 : )));
1504 :
1505 0 : let parent_shard_identity = tenant.shard_identity;
1506 0 : let parent_tenant_conf = tenant.get_tenant_conf();
1507 0 : let parent_generation = tenant.generation;
1508 :
1509 : // Phase 1: Write out child shards' remote index files, in the parent tenant's current generation
1510 0 : if let Err(e) = tenant.split_prepare(&child_shards).await {
1511 : // If [`Tenant::split_prepare`] fails, we must reload the tenant, because it might
1512 : // have been left in a partially-shut-down state.
1513 0 : tracing::warn!("Failed to prepare for split: {e}, reloading Tenant before returning");
1514 0 : return Err(e);
1515 0 : }
1516 :
1517 0 : fail::fail_point!("shard-split-post-prepare", |_| Err(anyhow::anyhow!(
1518 0 : "failpoint"
1519 : )));
1520 :
1521 0 : self.resources.deletion_queue_client.flush_advisory();
1522 :
1523 : // Phase 2: Put the parent shard to InProgress and grab a reference to the parent Tenant
1524 0 : drop(tenant);
1525 0 : let mut parent_slot_guard =
1526 0 : self.tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)?;
1527 0 : let parent = match parent_slot_guard.get_old_value() {
1528 0 : Some(TenantSlot::Attached(t)) => t,
1529 0 : Some(TenantSlot::Secondary(_)) => anyhow::bail!("Tenant location in secondary mode"),
1530 : Some(TenantSlot::InProgress(_)) => {
1531 : // tenant_map_acquire_slot never returns InProgress, if a slot was InProgress
1532 : // it would return an error.
1533 0 : unreachable!()
1534 : }
1535 : None => {
1536 : // We don't actually need the parent shard to still be attached to do our work, but it's
1537 : // a weird enough situation that the caller probably didn't want us to continue working
1538 : // if they had detached the tenant they requested the split on.
1539 0 : anyhow::bail!("Detached parent shard in the middle of split!")
1540 : }
1541 : };
1542 0 : fail::fail_point!("shard-split-pre-hardlink", |_| Err(anyhow::anyhow!(
1543 0 : "failpoint"
1544 : )));
1545 : // Optimization: hardlink layers from the parent into the children, so that they don't have to
1546 : // re-download & duplicate the data referenced in their initial IndexPart
1547 0 : self.shard_split_hardlink(parent, child_shards.clone())
1548 0 : .await?;
1549 0 : fail::fail_point!("shard-split-post-hardlink", |_| Err(anyhow::anyhow!(
1550 0 : "failpoint"
1551 : )));
1552 :
1553 : // Take a snapshot of where the parent's WAL ingest had got to: we will wait for
1554 : // child shards to reach this point.
1555 0 : let mut target_lsns = HashMap::new();
1556 0 : for timeline in parent.timelines.lock().unwrap().clone().values() {
1557 0 : target_lsns.insert(timeline.timeline_id, timeline.get_last_record_lsn());
1558 0 : }
1559 :
1560 : // TODO: we should have the parent shard stop its WAL ingest here, it's a waste of resources
1561 : // and could slow down the children trying to catch up.
1562 :
1563 : // Phase 3: Spawn the child shards
1564 0 : for child_shard in &child_shards {
1565 0 : let mut child_shard_identity = parent_shard_identity;
1566 0 : if let Some(new_stripe_size) = new_stripe_size {
1567 0 : child_shard_identity.stripe_size = new_stripe_size;
1568 0 : }
1569 0 : child_shard_identity.count = child_shard.shard_count;
1570 0 : child_shard_identity.number = child_shard.shard_number;
1571 :
1572 0 : let child_location_conf = LocationConf {
1573 0 : mode: LocationMode::Attached(AttachedLocationConfig {
1574 0 : generation: parent_generation,
1575 0 : attach_mode: AttachmentMode::Single,
1576 0 : }),
1577 0 : shard: child_shard_identity,
1578 0 : tenant_conf: parent_tenant_conf.clone(),
1579 0 : };
1580 :
1581 0 : self.upsert_location(
1582 0 : *child_shard,
1583 0 : child_location_conf,
1584 0 : None,
1585 0 : SpawnMode::Eager,
1586 0 : ctx,
1587 0 : )
1588 0 : .await?;
1589 : }
1590 :
1591 0 : fail::fail_point!("shard-split-post-child-conf", |_| Err(anyhow::anyhow!(
1592 0 : "failpoint"
1593 : )));
1594 :
1595 : // Phase 4: wait for child chards WAL ingest to catch up to target LSN
1596 0 : for child_shard_id in &child_shards {
1597 0 : let child_shard_id = *child_shard_id;
1598 0 : let child_shard = {
1599 0 : let locked = self.tenants.read().unwrap();
1600 0 : let peek_slot =
1601 0 : tenant_map_peek_slot(&locked, &child_shard_id, TenantSlotPeekMode::Read)?;
1602 0 : peek_slot.and_then(|s| s.get_attached()).cloned()
1603 : };
1604 0 : if let Some(t) = child_shard {
1605 : // Wait for the child shard to become active: this should be very quick because it only
1606 : // has to download the index_part that we just uploaded when creating it.
1607 0 : if let Err(e) = t.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await {
1608 : // This is not fatal: we have durably created the child shard. It just makes the
1609 : // split operation less seamless for clients, as we will may detach the parent
1610 : // shard before the child shards are fully ready to serve requests.
1611 0 : tracing::warn!("Failed to wait for shard {child_shard_id} to activate: {e}");
1612 0 : continue;
1613 0 : }
1614 :
1615 0 : let timelines = t.timelines.lock().unwrap().clone();
1616 0 : for timeline in timelines.values() {
1617 0 : let Some(target_lsn) = target_lsns.get(&timeline.timeline_id) else {
1618 0 : continue;
1619 : };
1620 :
1621 0 : tracing::info!(
1622 0 : "Waiting for child shard {}/{} to reach target lsn {}...",
1623 : child_shard_id,
1624 0 : timeline.timeline_id,
1625 : target_lsn
1626 : );
1627 :
1628 0 : fail::fail_point!("shard-split-lsn-wait", |_| Err(anyhow::anyhow!(
1629 0 : "failpoint"
1630 : )));
1631 0 : if let Err(e) = timeline
1632 0 : .wait_lsn(
1633 0 : *target_lsn,
1634 0 : crate::tenant::timeline::WaitLsnWaiter::Tenant,
1635 0 : crate::tenant::timeline::WaitLsnTimeout::Default,
1636 0 : ctx,
1637 0 : )
1638 0 : .await
1639 : {
1640 : // Failure here might mean shutdown, in any case this part is an optimization
1641 : // and we shouldn't hold up the split operation.
1642 0 : tracing::warn!(
1643 0 : "Failed to wait for timeline {} to reach lsn {target_lsn}: {e}",
1644 0 : timeline.timeline_id
1645 : );
1646 : } else {
1647 0 : tracing::info!(
1648 0 : "Child shard {}/{} reached target lsn {}",
1649 : child_shard_id,
1650 0 : timeline.timeline_id,
1651 : target_lsn
1652 : );
1653 : }
1654 : }
1655 0 : }
1656 : }
1657 :
1658 : // Phase 5: Shut down the parent shard. We leave it on disk in case the split fails and we
1659 : // have to roll back to the parent shard, avoiding a cold start. It will be cleaned up once
1660 : // the storage controller commits the split, or if all else fails, on the next restart.
1661 : //
1662 : // TODO: We don't flush the ephemeral layer here, because the split is likely to succeed and
1663 : // catching up the parent should be reasonably quick. Consider using FreezeAndFlush instead.
1664 0 : let (_guard, progress) = completion::channel();
1665 0 : match parent.shutdown(progress, ShutdownMode::Hard).await {
1666 0 : Ok(()) => {}
1667 0 : Err(other) => {
1668 0 : other.wait().await;
1669 : }
1670 : }
1671 :
1672 0 : fail::fail_point!("shard-split-pre-finish", |_| Err(anyhow::anyhow!(
1673 0 : "failpoint"
1674 : )));
1675 :
1676 0 : parent_slot_guard.drop_old_value()?;
1677 :
1678 : // Phase 6: Release the InProgress on the parent shard
1679 0 : drop(parent_slot_guard);
1680 :
1681 0 : utils::pausable_failpoint!("shard-split-post-finish-pause");
1682 :
1683 0 : Ok(child_shards)
1684 0 : }
1685 :
1686 : /// Part of [`Self::shard_split`]: hard link parent shard layers into child shards, as an optimization
1687 : /// to avoid the children downloading them again.
1688 : ///
1689 : /// For each resident layer in the parent shard, we will hard link it into all of the child shards.
1690 0 : async fn shard_split_hardlink(
1691 0 : &self,
1692 0 : parent_shard: &TenantShard,
1693 0 : child_shards: Vec<TenantShardId>,
1694 0 : ) -> anyhow::Result<()> {
1695 0 : debug_assert_current_span_has_tenant_id();
1696 :
1697 0 : let parent_path = self.conf.tenant_path(parent_shard.get_tenant_shard_id());
1698 0 : let (parent_timelines, parent_layers) = {
1699 0 : let mut parent_layers = Vec::new();
1700 0 : let timelines = parent_shard.timelines.lock().unwrap().clone();
1701 0 : let parent_timelines = timelines.keys().cloned().collect::<Vec<_>>();
1702 0 : for timeline in timelines.values() {
1703 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Loading list of layers to hardlink");
1704 0 : let layers = timeline
1705 0 : .layers
1706 0 : .read(LayerManagerLockHolder::GetLayerMapInfo)
1707 0 : .await;
1708 :
1709 0 : for layer in layers.likely_resident_layers() {
1710 0 : let relative_path = layer
1711 0 : .local_path()
1712 0 : .strip_prefix(&parent_path)
1713 0 : .context("Removing prefix from parent layer path")?;
1714 0 : parent_layers.push(relative_path.to_owned());
1715 : }
1716 : }
1717 :
1718 0 : if parent_layers.is_empty() {
1719 0 : tracing::info!("Ancestor shard has no resident layer to hard link");
1720 0 : }
1721 :
1722 0 : (parent_timelines, parent_layers)
1723 : };
1724 :
1725 0 : let mut child_prefixes = Vec::new();
1726 0 : let mut create_dirs = Vec::new();
1727 :
1728 0 : for child in child_shards {
1729 0 : let child_prefix = self.conf.tenant_path(&child);
1730 0 : create_dirs.push(child_prefix.clone());
1731 0 : create_dirs.extend(
1732 0 : parent_timelines
1733 0 : .iter()
1734 0 : .map(|t| self.conf.timeline_path(&child, t)),
1735 : );
1736 :
1737 0 : child_prefixes.push(child_prefix);
1738 : }
1739 :
1740 : // Since we will do a large number of small filesystem metadata operations, batch them into
1741 : // spawn_blocking calls rather than doing each one as a tokio::fs round-trip.
1742 0 : let span = tracing::Span::current();
1743 0 : let jh = tokio::task::spawn_blocking(move || -> anyhow::Result<usize> {
1744 : // Run this synchronous code in the same log context as the outer function that spawned it.
1745 0 : let _span = span.enter();
1746 :
1747 0 : tracing::info!("Creating {} directories", create_dirs.len());
1748 0 : for dir in &create_dirs {
1749 0 : if let Err(e) = std::fs::create_dir_all(dir) {
1750 : // Ignore AlreadyExists errors, drop out on all other errors
1751 0 : match e.kind() {
1752 0 : std::io::ErrorKind::AlreadyExists => {}
1753 : _ => {
1754 0 : return Err(anyhow::anyhow!(e).context(format!("Creating {dir}")));
1755 : }
1756 : }
1757 0 : }
1758 : }
1759 :
1760 0 : for child_prefix in child_prefixes {
1761 0 : tracing::info!(
1762 0 : "Hard-linking {} parent layers into child path {}",
1763 0 : parent_layers.len(),
1764 : child_prefix
1765 : );
1766 0 : for relative_layer in &parent_layers {
1767 0 : let parent_path = parent_path.join(relative_layer);
1768 0 : let child_path = child_prefix.join(relative_layer);
1769 0 : if let Err(e) = std::fs::hard_link(&parent_path, &child_path) {
1770 0 : match e.kind() {
1771 0 : std::io::ErrorKind::AlreadyExists => {}
1772 : std::io::ErrorKind::NotFound => {
1773 0 : tracing::info!(
1774 0 : "Layer {} not found during hard-linking, evicted during split?",
1775 : relative_layer
1776 : );
1777 : }
1778 : _ => {
1779 0 : return Err(anyhow::anyhow!(e).context(format!(
1780 0 : "Hard linking {relative_layer} into {child_prefix}"
1781 0 : )));
1782 : }
1783 : }
1784 0 : }
1785 : }
1786 : }
1787 :
1788 : // Durability is not required for correctness, but if we crashed during split and
1789 : // then came restarted with empty timeline dirs, it would be very inefficient to
1790 : // re-populate from remote storage.
1791 0 : tracing::info!("fsyncing {} directories", create_dirs.len());
1792 0 : for dir in create_dirs {
1793 0 : if let Err(e) = crashsafe::fsync(&dir) {
1794 : // Something removed a newly created timeline dir out from underneath us? Extremely
1795 : // unexpected, but not worth panic'ing over as this whole function is just an
1796 : // optimization.
1797 0 : tracing::warn!("Failed to fsync directory {dir}: {e}")
1798 0 : }
1799 : }
1800 :
1801 0 : Ok(parent_layers.len())
1802 0 : });
1803 :
1804 0 : match jh.await {
1805 0 : Ok(Ok(layer_count)) => {
1806 0 : tracing::info!(count = layer_count, "Hard linked layers into child shards");
1807 : }
1808 0 : Ok(Err(e)) => {
1809 : // This is an optimization, so we tolerate failure.
1810 0 : tracing::warn!("Error hard-linking layers, proceeding anyway: {e}")
1811 : }
1812 0 : Err(e) => {
1813 : // This is something totally unexpected like a panic, so bail out.
1814 0 : anyhow::bail!("Error joining hard linking task: {e}");
1815 : }
1816 : }
1817 :
1818 0 : Ok(())
1819 0 : }
1820 :
1821 : ///
1822 : /// Shut down all tenants. This runs as part of pageserver shutdown.
1823 : ///
1824 : /// NB: We leave the tenants in the map, so that they remain accessible through
1825 : /// the management API until we shut it down. If we removed the shut-down tenants
1826 : /// from the tenants map, the management API would return 404 for these tenants,
1827 : /// because TenantsMap::get() now returns `None`.
1828 : /// That could be easily misinterpreted by control plane, the consumer of the
1829 : /// management API. For example, it could attach the tenant on a different pageserver.
1830 : /// We would then be in split-brain once this pageserver restarts.
1831 : #[instrument(skip_all)]
1832 : pub(crate) async fn shutdown(&self) {
1833 : self.cancel.cancel();
1834 :
1835 : self.shutdown_all_tenants0().await
1836 : }
1837 :
1838 1 : async fn shutdown_all_tenants0(&self) {
1839 1 : let mut join_set = JoinSet::new();
1840 :
1841 : #[cfg(all(debug_assertions, not(test)))]
1842 : {
1843 : // Check that our metrics properly tracked the size of the tenants map. This is a convenient location to check,
1844 : // as it happens implicitly at the end of tests etc.
1845 0 : let m = self.tenants.read().unwrap();
1846 0 : debug_assert_eq!(METRICS.slots_total(), m.len() as u64);
1847 : }
1848 :
1849 : // Atomically, 1. create the shutdown tasks and 2. prevent creation of new tenants.
1850 1 : let (total_in_progress, total_attached) = {
1851 1 : let mut m = self.tenants.write().unwrap();
1852 1 : match &mut *m {
1853 : TenantsMap::Initializing => {
1854 0 : *m = TenantsMap::ShuttingDown(BTreeMap::default());
1855 0 : info!("tenants map is empty");
1856 0 : return;
1857 : }
1858 1 : TenantsMap::Open(tenants) => {
1859 1 : let mut shutdown_state = BTreeMap::new();
1860 1 : let mut total_in_progress = 0;
1861 1 : let mut total_attached = 0;
1862 :
1863 1 : for (tenant_shard_id, v) in std::mem::take(tenants).into_iter() {
1864 1 : match v {
1865 0 : TenantSlot::Attached(t) => {
1866 0 : shutdown_state
1867 0 : .insert(tenant_shard_id, TenantSlot::Attached(t.clone()));
1868 0 : join_set.spawn(
1869 0 : async move {
1870 0 : let res = {
1871 0 : let (_guard, shutdown_progress) = completion::channel();
1872 0 : t.shutdown(shutdown_progress, ShutdownMode::FreezeAndFlush).await
1873 : };
1874 :
1875 0 : if let Err(other_progress) = res {
1876 : // join the another shutdown in progress
1877 0 : other_progress.wait().await;
1878 0 : }
1879 :
1880 : // we cannot afford per tenant logging here, because if s3 is degraded, we are
1881 : // going to log too many lines
1882 0 : debug!("tenant successfully stopped");
1883 0 : }
1884 0 : .instrument(info_span!("shutdown", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug())),
1885 : );
1886 :
1887 0 : total_attached += 1;
1888 : }
1889 0 : TenantSlot::Secondary(state) => {
1890 0 : // We don't need to wait for this individually per-tenant: the
1891 0 : // downloader task will be waited on eventually, this cancel
1892 0 : // is just to encourage it to drop out if it is doing work
1893 0 : // for this tenant right now.
1894 0 : state.cancel.cancel();
1895 0 :
1896 0 : shutdown_state
1897 0 : .insert(tenant_shard_id, TenantSlot::Secondary(state));
1898 0 : }
1899 1 : TenantSlot::InProgress(notify) => {
1900 : // InProgress tenants are not visible in TenantsMap::ShuttingDown: we will
1901 : // wait for their notifications to fire in this function.
1902 1 : join_set.spawn(async move {
1903 1 : notify.wait().await;
1904 1 : });
1905 :
1906 1 : total_in_progress += 1;
1907 : }
1908 : }
1909 : }
1910 1 : *m = TenantsMap::ShuttingDown(shutdown_state);
1911 1 : (total_in_progress, total_attached)
1912 : }
1913 : TenantsMap::ShuttingDown(_) => {
1914 0 : error!(
1915 0 : "already shutting down, this function isn't supposed to be called more than once"
1916 : );
1917 0 : return;
1918 : }
1919 : }
1920 : };
1921 :
1922 1 : let started_at = std::time::Instant::now();
1923 :
1924 1 : info!(
1925 0 : "Waiting for {} InProgress tenants and {} Attached tenants to shut down",
1926 : total_in_progress, total_attached
1927 : );
1928 :
1929 1 : let total = join_set.len();
1930 1 : let mut panicked = 0;
1931 1 : let mut buffering = true;
1932 : const BUFFER_FOR: std::time::Duration = std::time::Duration::from_millis(500);
1933 1 : let mut buffered = std::pin::pin!(tokio::time::sleep(BUFFER_FOR));
1934 :
1935 3 : while !join_set.is_empty() {
1936 2 : tokio::select! {
1937 2 : Some(joined) = join_set.join_next() => {
1938 0 : match joined {
1939 1 : Ok(()) => {},
1940 0 : Err(join_error) if join_error.is_cancelled() => {
1941 0 : unreachable!("we are not cancelling any of the tasks");
1942 : }
1943 0 : Err(join_error) if join_error.is_panic() => {
1944 0 : // cannot really do anything, as this panic is likely a bug
1945 0 : panicked += 1;
1946 0 : }
1947 0 : Err(join_error) => {
1948 0 : warn!("unknown kind of JoinError: {join_error}");
1949 : }
1950 : }
1951 1 : if !buffering {
1952 1 : // buffer so that every 500ms since the first update (or starting) we'll log
1953 1 : // how far away we are; this is because we will get SIGKILL'd at 10s, and we
1954 1 : // are not able to log *then*.
1955 1 : buffering = true;
1956 1 : buffered.as_mut().reset(tokio::time::Instant::now() + BUFFER_FOR);
1957 1 : }
1958 : },
1959 2 : _ = &mut buffered, if buffering => {
1960 1 : buffering = false;
1961 1 : info!(remaining = join_set.len(), total, elapsed_ms = started_at.elapsed().as_millis(), "waiting for tenants to shutdown");
1962 : }
1963 : }
1964 : }
1965 :
1966 1 : if panicked > 0 {
1967 0 : warn!(
1968 : panicked,
1969 0 : total, "observed panicks while shutting down tenants"
1970 : );
1971 1 : }
1972 :
1973 : // caller will log how long we took
1974 1 : }
1975 :
1976 : /// Detaches a tenant, and removes its local files asynchronously.
1977 : ///
1978 : /// File removal is idempotent: even if the tenant has already been removed, this will still
1979 : /// remove any local files. This is used during shard splits, where we leave the parent shard's
1980 : /// files around in case we have to roll back the split.
1981 0 : pub(crate) async fn detach_tenant(
1982 0 : &self,
1983 0 : conf: &'static PageServerConf,
1984 0 : tenant_shard_id: TenantShardId,
1985 0 : deletion_queue_client: &DeletionQueueClient,
1986 0 : ) -> Result<(), TenantStateError> {
1987 0 : if let Some(tmp_path) = self
1988 0 : .detach_tenant0(conf, tenant_shard_id, deletion_queue_client)
1989 0 : .await?
1990 0 : {
1991 0 : self.background_purges.spawn(tmp_path);
1992 0 : }
1993 :
1994 0 : Ok(())
1995 0 : }
1996 :
1997 : /// Detaches a tenant. This renames the tenant directory to a temporary path and returns it,
1998 : /// allowing the caller to delete it asynchronously. Returns None if the dir is already removed.
1999 0 : async fn detach_tenant0(
2000 0 : &self,
2001 0 : conf: &'static PageServerConf,
2002 0 : tenant_shard_id: TenantShardId,
2003 0 : deletion_queue_client: &DeletionQueueClient,
2004 0 : ) -> Result<Option<Utf8PathBuf>, TenantStateError> {
2005 0 : let tenant_dir_rename_operation = |tenant_id_to_clean: TenantShardId| async move {
2006 0 : let local_tenant_directory = conf.tenant_path(&tenant_id_to_clean);
2007 0 : if !tokio::fs::try_exists(&local_tenant_directory).await? {
2008 : // If the tenant directory doesn't exist, it's already cleaned up.
2009 0 : return Ok(None);
2010 0 : }
2011 0 : safe_rename_tenant_dir(&local_tenant_directory)
2012 0 : .await
2013 0 : .with_context(|| {
2014 0 : format!("local tenant directory {local_tenant_directory:?} rename")
2015 0 : })
2016 0 : .map(Some)
2017 0 : };
2018 :
2019 0 : let mut removal_result = self
2020 0 : .remove_tenant_from_memory(
2021 0 : tenant_shard_id,
2022 0 : tenant_dir_rename_operation(tenant_shard_id),
2023 0 : )
2024 0 : .await;
2025 :
2026 : // If the tenant was not found, it was likely already removed. Attempt to remove the tenant
2027 : // directory on disk anyway. For example, during shard splits, we shut down and remove the
2028 : // parent shard, but leave its directory on disk in case we have to roll back the split.
2029 : //
2030 : // TODO: it would be better to leave the parent shard attached until the split is committed.
2031 : // This will be needed by the gRPC page service too, such that a compute can continue to
2032 : // read from the parent shard until it's notified about the new child shards. See:
2033 : // <https://github.com/neondatabase/neon/issues/11728>.
2034 0 : if let Err(TenantStateError::SlotError(TenantSlotError::NotFound(_))) = removal_result {
2035 0 : removal_result = tenant_dir_rename_operation(tenant_shard_id)
2036 0 : .await
2037 0 : .map_err(TenantStateError::Other);
2038 0 : }
2039 :
2040 : // Flush pending deletions, so that they have a good chance of passing validation
2041 : // before this tenant is potentially re-attached elsewhere.
2042 0 : deletion_queue_client.flush_advisory();
2043 :
2044 0 : removal_result
2045 0 : }
2046 :
2047 0 : pub(crate) fn list_tenants(
2048 0 : &self,
2049 0 : ) -> Result<Vec<(TenantShardId, TenantState, Generation)>, TenantMapListError> {
2050 0 : let tenants = self.tenants.read().unwrap();
2051 0 : let m = match &*tenants {
2052 0 : TenantsMap::Initializing => return Err(TenantMapListError::Initializing),
2053 0 : TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => m,
2054 : };
2055 0 : Ok(m.iter()
2056 0 : .filter_map(|(id, tenant)| match tenant {
2057 0 : TenantSlot::Attached(tenant) => {
2058 0 : Some((*id, tenant.current_state(), tenant.generation()))
2059 : }
2060 0 : TenantSlot::Secondary(_) => None,
2061 0 : TenantSlot::InProgress(_) => None,
2062 0 : })
2063 0 : .collect())
2064 0 : }
2065 :
2066 : /// Completes an earlier prepared timeline detach ancestor.
2067 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
2068 0 : &self,
2069 0 : tenant_shard_id: TenantShardId,
2070 0 : timeline_id: TimelineId,
2071 0 : prepared: PreparedTimelineDetach,
2072 0 : behavior: DetachBehavior,
2073 0 : mut attempt: detach_ancestor::Attempt,
2074 0 : ctx: &RequestContext,
2075 0 : ) -> Result<HashSet<TimelineId>, detach_ancestor::Error> {
2076 : use detach_ancestor::Error;
2077 :
2078 0 : let slot_guard = self
2079 0 : .tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::MustExist)
2080 0 : .map_err(|e| {
2081 : use TenantSlotError::*;
2082 :
2083 0 : match e {
2084 0 : MapState(TenantMapError::ShuttingDown) => Error::ShuttingDown,
2085 0 : NotFound(_) | InProgress | MapState(_) => Error::DetachReparent(e.into()),
2086 : }
2087 0 : })?;
2088 :
2089 0 : let tenant = {
2090 0 : let old_slot = slot_guard
2091 0 : .get_old_value()
2092 0 : .as_ref()
2093 0 : .expect("requested MustExist");
2094 :
2095 0 : let Some(tenant) = old_slot.get_attached() else {
2096 0 : return Err(Error::DetachReparent(anyhow::anyhow!(
2097 0 : "Tenant is not in attached state"
2098 0 : )));
2099 : };
2100 :
2101 0 : if !tenant.is_active() {
2102 0 : return Err(Error::DetachReparent(anyhow::anyhow!(
2103 0 : "Tenant is not active"
2104 0 : )));
2105 0 : }
2106 :
2107 0 : tenant.clone()
2108 : };
2109 :
2110 0 : let timeline = tenant
2111 0 : .get_timeline(timeline_id, true)
2112 0 : .map_err(Error::NotFound)?;
2113 :
2114 0 : let resp = timeline
2115 0 : .detach_from_ancestor_and_reparent(
2116 0 : &tenant,
2117 0 : prepared,
2118 0 : attempt.ancestor_timeline_id,
2119 0 : attempt.ancestor_lsn,
2120 0 : behavior,
2121 0 : ctx,
2122 0 : )
2123 0 : .await?;
2124 :
2125 0 : let mut slot_guard = slot_guard;
2126 :
2127 0 : let tenant = if resp.reset_tenant_required() {
2128 0 : attempt.before_reset_tenant();
2129 :
2130 0 : let (_guard, progress) = utils::completion::channel();
2131 0 : match tenant.shutdown(progress, ShutdownMode::Reload).await {
2132 0 : Ok(()) => {
2133 0 : slot_guard.drop_old_value().expect("it was just shutdown");
2134 0 : }
2135 0 : Err(_barrier) => {
2136 0 : slot_guard.revert();
2137 : // this really should not happen, at all, unless a shutdown without acquiring
2138 : // tenant slot was already going? regardless, on restart the attempt tracking
2139 : // will reset to retryable.
2140 0 : return Err(Error::ShuttingDown);
2141 : }
2142 : }
2143 :
2144 0 : let tenant_path = self.conf.tenant_path(&tenant_shard_id);
2145 0 : let config = TenantShard::load_tenant_config(self.conf, &tenant_shard_id)
2146 0 : .map_err(|e| Error::DetachReparent(e.into()))?;
2147 :
2148 0 : let shard_identity = config.shard;
2149 0 : let tenant = tenant_spawn(
2150 0 : self.conf,
2151 0 : tenant_shard_id,
2152 0 : &tenant_path,
2153 0 : self.resources.clone(),
2154 0 : AttachedTenantConf::try_from(self.conf, config).map_err(Error::DetachReparent)?,
2155 0 : shard_identity,
2156 0 : None,
2157 0 : SpawnMode::Eager,
2158 0 : ctx,
2159 : )
2160 0 : .map_err(|_| Error::ShuttingDown)?;
2161 :
2162 : {
2163 0 : let mut g = tenant.ongoing_timeline_detach.lock().unwrap();
2164 0 : assert!(
2165 0 : g.is_none(),
2166 0 : "there cannot be any new timeline detach ancestor on newly created tenant"
2167 : );
2168 0 : *g = Some((attempt.timeline_id, attempt.new_barrier()));
2169 : }
2170 :
2171 : // if we bail out here, we will not allow a new attempt, which should be fine.
2172 : // pageserver should be shutting down regardless? tenant_reset would help, unless it
2173 : // runs into the same problem.
2174 0 : slot_guard
2175 0 : .upsert(TenantSlot::Attached(tenant.clone()))
2176 0 : .map_err(|e| match e {
2177 0 : TenantSlotUpsertError::ShuttingDown(_) => Error::ShuttingDown,
2178 0 : other => Error::DetachReparent(other.into()),
2179 0 : })?;
2180 0 : tenant
2181 : } else {
2182 0 : tracing::info!("skipping tenant_reset as no changes made required it");
2183 0 : tenant
2184 : };
2185 :
2186 0 : if let Some(reparented) = resp.completed() {
2187 : // finally ask the restarted tenant to complete the detach
2188 : //
2189 : // rationale for 9999s: we don't really have a timetable here; if retried, the caller
2190 : // will get an 503.
2191 0 : tenant
2192 0 : .wait_to_become_active(std::time::Duration::from_secs(9999))
2193 0 : .await
2194 0 : .map_err(|e| {
2195 : use GetActiveTenantError::{Cancelled, WillNotBecomeActive};
2196 : use pageserver_api::models::TenantState;
2197 0 : match e {
2198 : Cancelled | WillNotBecomeActive(TenantState::Stopping { .. }) => {
2199 0 : Error::ShuttingDown
2200 : }
2201 0 : other => Error::Complete(other.into()),
2202 : }
2203 0 : })?;
2204 :
2205 0 : utils::pausable_failpoint!(
2206 : "timeline-detach-ancestor::after_activating_before_finding-pausable"
2207 : );
2208 :
2209 0 : let timeline = tenant
2210 0 : .get_timeline(attempt.timeline_id, true)
2211 0 : .map_err(Error::NotFound)?;
2212 :
2213 0 : timeline
2214 0 : .complete_detaching_timeline_ancestor(&tenant, attempt, ctx)
2215 0 : .await
2216 0 : .map(|()| reparented)
2217 : } else {
2218 : // at least the latest versions have now been downloaded and refreshed; be ready to
2219 : // retry another time.
2220 0 : Err(Error::FailedToReparentAll)
2221 : }
2222 0 : }
2223 :
2224 : /// A page service client sends a TenantId, and to look up the correct Tenant we must
2225 : /// resolve this to a fully qualified TenantShardId.
2226 : ///
2227 : /// During shard splits: we shall see parent shards in InProgress state and skip them, and
2228 : /// instead match on child shards which should appear in Attached state. Very early in a shard
2229 : /// split, or in other cases where a shard is InProgress, we will return our own InProgress result
2230 : /// to instruct the caller to wait for that to finish before querying again.
2231 0 : pub(crate) fn resolve_attached_shard(
2232 0 : &self,
2233 0 : tenant_id: &TenantId,
2234 0 : selector: ShardSelector,
2235 0 : ) -> ShardResolveResult {
2236 0 : let tenants = self.tenants.read().unwrap();
2237 0 : let mut want_shard: Option<ShardIndex> = None;
2238 0 : let mut any_in_progress = None;
2239 :
2240 0 : match &*tenants {
2241 0 : TenantsMap::Initializing => ShardResolveResult::NotFound,
2242 0 : TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => {
2243 0 : for slot in m.range(TenantShardId::tenant_range(*tenant_id)) {
2244 : // Ignore all slots that don't contain an attached tenant
2245 0 : let tenant = match &slot.1 {
2246 0 : TenantSlot::Attached(t) => t,
2247 0 : TenantSlot::InProgress(barrier) => {
2248 : // We might still find a usable shard, but in case we don't, remember that
2249 : // we saw at least one InProgress slot, so that we can distinguish this case
2250 : // from a simple NotFound in our return value.
2251 0 : any_in_progress = Some(barrier.clone());
2252 0 : continue;
2253 : }
2254 0 : _ => continue,
2255 : };
2256 :
2257 0 : match selector {
2258 0 : ShardSelector::Zero if slot.0.shard_number == ShardNumber(0) => {
2259 0 : return ShardResolveResult::Found(tenant.clone());
2260 : }
2261 0 : ShardSelector::Page(key) => {
2262 : // Each time we find an attached slot with a different shard count,
2263 : // recompute the expected shard number: during shard splits we might
2264 : // have multiple shards with the old shard count.
2265 0 : if want_shard.is_none()
2266 0 : || want_shard.unwrap().shard_count != tenant.shard_identity.count
2267 0 : {
2268 0 : want_shard = Some(ShardIndex {
2269 0 : shard_number: tenant.shard_identity.get_shard_number(&key),
2270 0 : shard_count: tenant.shard_identity.count,
2271 0 : });
2272 0 : }
2273 :
2274 0 : if Some(ShardIndex {
2275 0 : shard_number: tenant.shard_identity.number,
2276 0 : shard_count: tenant.shard_identity.count,
2277 0 : }) == want_shard
2278 : {
2279 0 : return ShardResolveResult::Found(tenant.clone());
2280 0 : }
2281 : }
2282 0 : ShardSelector::Known(shard)
2283 0 : if tenant.shard_identity.shard_index() == shard =>
2284 : {
2285 0 : return ShardResolveResult::Found(tenant.clone());
2286 : }
2287 0 : _ => continue,
2288 : }
2289 : }
2290 :
2291 : // Fall through: we didn't find a slot that was in Attached state & matched our selector. If
2292 : // we found one or more InProgress slot, indicate to caller that they should retry later. Otherwise
2293 : // this requested shard simply isn't found.
2294 0 : if let Some(barrier) = any_in_progress {
2295 0 : ShardResolveResult::InProgress(barrier)
2296 : } else {
2297 0 : ShardResolveResult::NotFound
2298 : }
2299 : }
2300 : }
2301 0 : }
2302 :
2303 : /// Calculate the tenant shards' contributions to this pageserver's utilization metrics. The
2304 : /// returned values are:
2305 : /// - the number of bytes of local disk space this pageserver's shards are requesting, i.e.
2306 : /// how much space they would use if not impacted by disk usage eviction.
2307 : /// - the number of tenant shards currently on this pageserver, including attached
2308 : /// and secondary.
2309 : ///
2310 : /// This function is quite expensive: callers are expected to cache the result and
2311 : /// limit how often they call it.
2312 0 : pub(crate) fn calculate_utilization(&self) -> Result<(u64, u32), TenantMapListError> {
2313 0 : let tenants = self.tenants.read().unwrap();
2314 0 : let m = match &*tenants {
2315 0 : TenantsMap::Initializing => return Err(TenantMapListError::Initializing),
2316 0 : TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => m,
2317 : };
2318 0 : let shard_count = m.len();
2319 0 : let mut wanted_bytes = 0;
2320 :
2321 0 : for tenant_slot in m.values() {
2322 0 : match tenant_slot {
2323 0 : TenantSlot::InProgress(_barrier) => {
2324 : // While a slot is being changed, we can't know how much storage it wants. This
2325 : // means this function's output can fluctuate if a lot of changes are going on
2326 : // (such as transitions from secondary to attached).
2327 : //
2328 : // We could wait for the barrier and retry, but it's important that the utilization
2329 : // API is responsive, and the data quality impact is not very significant.
2330 0 : continue;
2331 : }
2332 0 : TenantSlot::Attached(tenant) => {
2333 0 : wanted_bytes += tenant.local_storage_wanted();
2334 0 : }
2335 0 : TenantSlot::Secondary(secondary) => {
2336 0 : let progress = secondary.progress.lock().unwrap();
2337 0 : wanted_bytes += if progress.heatmap_mtime.is_some() {
2338 : // If we have heatmap info, then we will 'want' the sum
2339 : // of the size of layers in the heatmap: this is how much space
2340 : // we would use if not doing any eviction.
2341 0 : progress.bytes_total
2342 : } else {
2343 : // In the absence of heatmap info, assume that the secondary location simply
2344 : // needs as much space as it is currently using.
2345 0 : secondary.resident_size_metric.get()
2346 : }
2347 : }
2348 : }
2349 : }
2350 :
2351 0 : Ok((wanted_bytes, shard_count as u32))
2352 0 : }
2353 :
2354 : #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), %timeline_id))]
2355 : pub(crate) async fn immediate_gc(
2356 : &self,
2357 : tenant_shard_id: TenantShardId,
2358 : timeline_id: TimelineId,
2359 : gc_req: TimelineGcRequest,
2360 : cancel: CancellationToken,
2361 : ctx: &RequestContext,
2362 : ) -> Result<GcResult, ApiError> {
2363 : let tenant = {
2364 : let guard = self.tenants.read().unwrap();
2365 : guard
2366 : .get(&tenant_shard_id)
2367 : .cloned()
2368 0 : .with_context(|| format!("tenant {tenant_shard_id}"))
2369 0 : .map_err(|e| ApiError::NotFound(e.into()))?
2370 : };
2371 :
2372 0 : let gc_horizon = gc_req.gc_horizon.unwrap_or_else(|| tenant.get_gc_horizon());
2373 : // Use tenant's pitr setting
2374 : let pitr = tenant.get_pitr_interval();
2375 :
2376 : tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
2377 :
2378 : // Run in task_mgr to avoid race with tenant_detach operation
2379 : let ctx: RequestContext =
2380 : ctx.detached_child(TaskKind::GarbageCollector, DownloadBehavior::Download);
2381 :
2382 : let _gate_guard = tenant.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
2383 :
2384 : fail::fail_point!("immediate_gc_task_pre");
2385 :
2386 : #[allow(unused_mut)]
2387 : let mut result = tenant
2388 : .gc_iteration(Some(timeline_id), gc_horizon, pitr, &cancel, &ctx)
2389 : .await;
2390 : // FIXME: `gc_iteration` can return an error for multiple reasons; we should handle it
2391 : // better once the types support it.
2392 :
2393 : #[cfg(feature = "testing")]
2394 : {
2395 : // we need to synchronize with drop completion for python tests without polling for
2396 : // log messages
2397 : if let Ok(result) = result.as_mut() {
2398 : let mut js = tokio::task::JoinSet::new();
2399 : for layer in std::mem::take(&mut result.doomed_layers) {
2400 : js.spawn(layer.wait_drop());
2401 : }
2402 : tracing::info!(
2403 : total = js.len(),
2404 : "starting to wait for the gc'd layers to be dropped"
2405 : );
2406 : while let Some(res) = js.join_next().await {
2407 : res.expect("wait_drop should not panic");
2408 : }
2409 : }
2410 :
2411 : let timeline = tenant.get_timeline(timeline_id, false).ok();
2412 0 : let rtc = timeline.as_ref().map(|x| &x.remote_client);
2413 :
2414 : if let Some(rtc) = rtc {
2415 : // layer drops schedule actions on remote timeline client to actually do the
2416 : // deletions; don't care about the shutdown error, just exit fast
2417 : drop(rtc.wait_completion().await);
2418 : }
2419 : }
2420 :
2421 0 : result.map_err(|e| match e {
2422 0 : GcError::TenantCancelled | GcError::TimelineCancelled => ApiError::ShuttingDown,
2423 : GcError::TimelineNotFound => {
2424 0 : ApiError::NotFound(anyhow::anyhow!("Timeline not found").into())
2425 : }
2426 0 : other => ApiError::InternalServerError(anyhow::anyhow!(other)),
2427 0 : })
2428 : }
2429 :
2430 : /// Stops and removes the tenant from memory, if it's not [`TenantState::Stopping`] already, bails otherwise.
2431 : /// Allows to remove other tenant resources manually, via `tenant_cleanup`.
2432 : /// If the cleanup fails, tenant will stay in memory in [`TenantState::Broken`] state, and another removal
2433 1 : async fn remove_tenant_from_memory<V, F>(
2434 1 : &self,
2435 1 : tenant_shard_id: TenantShardId,
2436 1 : tenant_cleanup: F,
2437 1 : ) -> Result<V, TenantStateError>
2438 1 : where
2439 1 : F: std::future::Future<Output = anyhow::Result<V>>,
2440 1 : {
2441 1 : let mut slot_guard =
2442 1 : self.tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::MustExist)?;
2443 :
2444 : // allow pageserver shutdown to await for our completion
2445 1 : let (_guard, progress) = completion::channel();
2446 :
2447 : // The SlotGuard allows us to manipulate the Tenant object without fear of some
2448 : // concurrent API request doing something else for the same tenant ID.
2449 1 : let attached_tenant = match slot_guard.get_old_value() {
2450 1 : Some(TenantSlot::Attached(tenant)) => {
2451 : // whenever we remove a tenant from memory, we don't want to flush and wait for upload
2452 1 : let shutdown_mode = ShutdownMode::Hard;
2453 :
2454 : // shutdown is sure to transition tenant to stopping, and wait for all tasks to complete, so
2455 : // that we can continue safely to cleanup.
2456 1 : match tenant.shutdown(progress, shutdown_mode).await {
2457 1 : Ok(()) => {}
2458 0 : Err(_other) => {
2459 : // if pageserver shutdown or other detach/ignore is already ongoing, we don't want to
2460 : // wait for it but return an error right away because these are distinct requests.
2461 0 : slot_guard.revert();
2462 0 : return Err(TenantStateError::IsStopping(tenant_shard_id));
2463 : }
2464 : }
2465 1 : Some(tenant)
2466 : }
2467 0 : Some(TenantSlot::Secondary(secondary_state)) => {
2468 0 : tracing::info!("Shutting down in secondary mode");
2469 0 : secondary_state.shutdown().await;
2470 0 : None
2471 : }
2472 : Some(TenantSlot::InProgress(_)) => {
2473 : // Acquiring a slot guarantees its old value was not InProgress
2474 0 : unreachable!();
2475 : }
2476 0 : None => None,
2477 : };
2478 :
2479 1 : match tenant_cleanup
2480 1 : .await
2481 1 : .with_context(|| format!("Failed to run cleanup for tenant {tenant_shard_id}"))
2482 : {
2483 1 : Ok(hook_value) => {
2484 : // Success: drop the old TenantSlot::Attached.
2485 1 : slot_guard
2486 1 : .drop_old_value()
2487 1 : .expect("We just called shutdown");
2488 :
2489 1 : Ok(hook_value)
2490 : }
2491 0 : Err(e) => {
2492 : // If we had a Tenant, set it to Broken and put it back in the TenantsMap
2493 0 : if let Some(attached_tenant) = attached_tenant {
2494 0 : attached_tenant.set_broken(e.to_string()).await;
2495 0 : }
2496 : // Leave the broken tenant in the map
2497 0 : slot_guard.revert();
2498 :
2499 0 : Err(TenantStateError::Other(e))
2500 : }
2501 : }
2502 1 : }
2503 : }
2504 :
2505 : #[derive(Debug, thiserror::Error)]
2506 : pub(crate) enum GetTenantError {
2507 : /// NotFound is a TenantId rather than TenantShardId, because this error type is used from
2508 : /// getters that use a TenantId and a ShardSelector, not just getters that target a specific shard.
2509 : #[error("Tenant {0} not found")]
2510 : NotFound(TenantId),
2511 :
2512 : #[error("Tenant {0} not found")]
2513 : ShardNotFound(TenantShardId),
2514 :
2515 : #[error("Tenant {0} is not active")]
2516 : NotActive(TenantShardId),
2517 :
2518 : // Initializing or shutting down: cannot authoritatively say whether we have this tenant
2519 : #[error("Tenant map is not available: {0}")]
2520 : MapState(#[from] TenantMapError),
2521 : }
2522 :
2523 : #[derive(thiserror::Error, Debug)]
2524 : pub(crate) enum GetActiveTenantError {
2525 : /// We may time out either while TenantSlot is InProgress, or while the Tenant
2526 : /// is in a non-Active state
2527 : #[error(
2528 : "Timed out waiting {wait_time:?} for tenant active state. Latest state: {latest_state:?}"
2529 : )]
2530 : WaitForActiveTimeout {
2531 : latest_state: Option<TenantState>,
2532 : wait_time: Duration,
2533 : },
2534 :
2535 : /// The TenantSlot is absent, or in secondary mode
2536 : #[error(transparent)]
2537 : NotFound(#[from] GetTenantError),
2538 :
2539 : /// Cancellation token fired while we were waiting
2540 : #[error("cancelled")]
2541 : Cancelled,
2542 :
2543 : /// Tenant exists, but is in a state that cannot become active (e.g. Stopping, Broken)
2544 : #[error("will not become active. Current state: {0}")]
2545 : WillNotBecomeActive(TenantState),
2546 :
2547 : /// Broken is logically a subset of WillNotBecomeActive, but a distinct error is useful as
2548 : /// WillNotBecomeActive is a permitted error under some circumstances, whereas broken should
2549 : /// never happen.
2550 : #[error("Tenant is broken: {0}")]
2551 : Broken(String),
2552 :
2553 : #[error("reconnect to switch tenant id")]
2554 : SwitchedTenant,
2555 : }
2556 :
2557 : #[derive(Debug, thiserror::Error)]
2558 : pub(crate) enum DeleteTimelineError {
2559 : #[error("Tenant {0}")]
2560 : Tenant(#[from] GetTenantError),
2561 :
2562 : #[error("Timeline {0}")]
2563 : Timeline(#[from] crate::tenant::DeleteTimelineError),
2564 : }
2565 :
2566 : #[derive(Debug, thiserror::Error)]
2567 : pub(crate) enum TenantStateError {
2568 : #[error("Tenant {0} is stopping")]
2569 : IsStopping(TenantShardId),
2570 : #[error(transparent)]
2571 : SlotError(#[from] TenantSlotError),
2572 : #[error(transparent)]
2573 : SlotUpsertError(#[from] TenantSlotUpsertError),
2574 : #[error(transparent)]
2575 : Other(#[from] anyhow::Error),
2576 : }
2577 :
2578 : #[derive(Debug, thiserror::Error)]
2579 : pub(crate) enum TenantMapListError {
2580 : #[error("tenant map is still initiailizing")]
2581 : Initializing,
2582 : }
2583 :
2584 : #[derive(Debug, thiserror::Error)]
2585 : pub(crate) enum TenantMapInsertError {
2586 : #[error(transparent)]
2587 : SlotError(#[from] TenantSlotError),
2588 : #[error(transparent)]
2589 : SlotUpsertError(#[from] TenantSlotUpsertError),
2590 : #[error(transparent)]
2591 : Other(#[from] anyhow::Error),
2592 : }
2593 :
2594 : /// Superset of TenantMapError: issues that can occur when acquiring a slot
2595 : /// for a particular tenant ID.
2596 : #[derive(Debug, thiserror::Error)]
2597 : pub(crate) enum TenantSlotError {
2598 : /// When acquiring a slot with the expectation that the tenant already exists.
2599 : #[error("Tenant {0} not found")]
2600 : NotFound(TenantShardId),
2601 :
2602 : // Tried to read a slot that is currently being mutated by another administrative
2603 : // operation.
2604 : #[error("tenant has a state change in progress, try again later")]
2605 : InProgress,
2606 :
2607 : #[error(transparent)]
2608 : MapState(#[from] TenantMapError),
2609 : }
2610 :
2611 : /// Superset of TenantMapError: issues that can occur when using a SlotGuard
2612 : /// to insert a new value.
2613 : #[derive(thiserror::Error)]
2614 : pub(crate) enum TenantSlotUpsertError {
2615 : /// An error where the slot is in an unexpected state, indicating a code bug
2616 : #[error("Internal error updating Tenant")]
2617 : InternalError(Cow<'static, str>),
2618 :
2619 : #[error(transparent)]
2620 : MapState(TenantMapError),
2621 :
2622 : // If we encounter TenantManager shutdown during upsert, we must carry the Completion
2623 : // from the SlotGuard, so that the caller can hold it while they clean up: otherwise
2624 : // TenantManager shutdown might race ahead before we're done cleaning up any Tenant that
2625 : // was protected by the SlotGuard.
2626 : #[error("Shutting down")]
2627 : ShuttingDown((TenantSlot, utils::completion::Completion)),
2628 : }
2629 :
2630 : impl std::fmt::Debug for TenantSlotUpsertError {
2631 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
2632 0 : match self {
2633 0 : Self::InternalError(reason) => write!(f, "Internal Error {reason}"),
2634 0 : Self::MapState(map_error) => write!(f, "Tenant map state: {map_error:?}"),
2635 0 : Self::ShuttingDown(_completion) => write!(f, "Tenant map shutting down"),
2636 : }
2637 0 : }
2638 : }
2639 :
2640 : #[derive(Debug, thiserror::Error)]
2641 : enum TenantSlotDropError {
2642 : /// It is only legal to drop a TenantSlot if its contents are fully shut down
2643 : #[error("Tenant was not shut down")]
2644 : NotShutdown,
2645 : }
2646 :
2647 : /// Errors that can happen any time we are walking the tenant map to try and acquire
2648 : /// the TenantSlot for a particular tenant.
2649 : #[derive(Debug, thiserror::Error)]
2650 : pub(crate) enum TenantMapError {
2651 : // Tried to read while initializing
2652 : #[error("tenant map is still initializing")]
2653 : StillInitializing,
2654 :
2655 : // Tried to read while shutting down
2656 : #[error("tenant map is shutting down")]
2657 : ShuttingDown,
2658 : }
2659 :
2660 : /// Guards a particular tenant_id's content in the TenantsMap.
2661 : ///
2662 : /// While this structure exists, the TenantsMap will contain a [`TenantSlot::InProgress`]
2663 : /// for this tenant, which acts as a marker for any operations targeting
2664 : /// this tenant to retry later, or wait for the InProgress state to end.
2665 : ///
2666 : /// This structure enforces the important invariant that we do not have overlapping
2667 : /// tasks that will try to use local storage for a the same tenant ID: we enforce that
2668 : /// the previous contents of a slot have been shut down before the slot can be
2669 : /// left empty or used for something else
2670 : ///
2671 : /// Holders of a SlotGuard should explicitly dispose of it, using either `upsert`
2672 : /// to provide a new value, or `revert` to put the slot back into its initial
2673 : /// state. If the SlotGuard is dropped without calling either of these, then
2674 : /// we will leave the slot empty if our `old_value` is already shut down, else
2675 : /// we will replace the slot with `old_value` (equivalent to doing a revert).
2676 : ///
2677 : /// The `old_value` may be dropped before the SlotGuard is dropped, by calling
2678 : /// `drop_old_value`. It is an error to call this without shutting down
2679 : /// the conents of `old_value`.
2680 : pub(crate) struct SlotGuard<'a> {
2681 : tenant_shard_id: TenantShardId,
2682 : old_value: Option<TenantSlot>,
2683 : upserted: bool,
2684 :
2685 : /// [`TenantSlot::InProgress`] carries the corresponding Barrier: it will
2686 : /// release any waiters as soon as this SlotGuard is dropped.
2687 : completion: utils::completion::Completion,
2688 :
2689 : tenants: &'a std::sync::RwLock<TenantsMap>,
2690 : }
2691 :
2692 : impl<'a> SlotGuard<'a> {
2693 1 : fn new(
2694 1 : tenant_shard_id: TenantShardId,
2695 1 : old_value: Option<TenantSlot>,
2696 1 : completion: utils::completion::Completion,
2697 1 : tenants: &'a std::sync::RwLock<TenantsMap>,
2698 1 : ) -> Self {
2699 1 : Self {
2700 1 : tenant_shard_id,
2701 1 : old_value,
2702 1 : upserted: false,
2703 1 : completion,
2704 1 : tenants,
2705 1 : }
2706 1 : }
2707 :
2708 : /// Get any value that was present in the slot before we acquired ownership
2709 : /// of it: in state transitions, this will be the old state.
2710 : ///
2711 : // FIXME: get_ prefix
2712 : // FIXME: this should be .as_ref() -- unsure why no clippy
2713 1 : fn get_old_value(&self) -> &Option<TenantSlot> {
2714 1 : &self.old_value
2715 1 : }
2716 :
2717 : /// Emplace a new value in the slot. This consumes the guard, and after
2718 : /// returning, the slot is no longer protected from concurrent changes.
2719 0 : fn upsert(mut self, new_value: TenantSlot) -> Result<(), TenantSlotUpsertError> {
2720 0 : if !self.old_value_is_shutdown() {
2721 : // This is a bug: callers should never try to drop an old value without
2722 : // shutting it down
2723 0 : return Err(TenantSlotUpsertError::InternalError(
2724 0 : "Old TenantSlot value not shut down".into(),
2725 0 : ));
2726 0 : }
2727 :
2728 0 : let replaced: Option<TenantSlot> = {
2729 0 : let mut locked = self.tenants.write().unwrap();
2730 :
2731 0 : if let TenantSlot::InProgress(_) = new_value {
2732 : // It is never expected to try and upsert InProgress via this path: it should
2733 : // only be written via the tenant_map_acquire_slot path. If we hit this it's a bug.
2734 0 : return Err(TenantSlotUpsertError::InternalError(
2735 0 : "Attempt to upsert an InProgress state".into(),
2736 0 : ));
2737 0 : }
2738 :
2739 0 : let m = match &mut *locked {
2740 : TenantsMap::Initializing => {
2741 0 : return Err(TenantSlotUpsertError::MapState(
2742 0 : TenantMapError::StillInitializing,
2743 0 : ));
2744 : }
2745 : TenantsMap::ShuttingDown(_) => {
2746 0 : return Err(TenantSlotUpsertError::ShuttingDown((
2747 0 : new_value,
2748 0 : self.completion.clone(),
2749 0 : )));
2750 : }
2751 0 : TenantsMap::Open(m) => m,
2752 : };
2753 :
2754 0 : METRICS.slot_inserted(&new_value);
2755 :
2756 0 : let replaced = m.insert(self.tenant_shard_id, new_value);
2757 0 : self.upserted = true;
2758 0 : if let Some(replaced) = replaced.as_ref() {
2759 0 : METRICS.slot_removed(replaced);
2760 0 : }
2761 :
2762 0 : replaced
2763 : };
2764 :
2765 : // Sanity check: on an upsert we should always be replacing an InProgress marker
2766 0 : match replaced {
2767 : Some(TenantSlot::InProgress(_)) => {
2768 : // Expected case: we find our InProgress in the map: nothing should have
2769 : // replaced it because the code that acquires slots will not grant another
2770 : // one for the same TenantId.
2771 0 : Ok(())
2772 : }
2773 : None => {
2774 0 : METRICS.unexpected_errors.inc();
2775 0 : error!(
2776 : tenant_shard_id = %self.tenant_shard_id,
2777 0 : "Missing InProgress marker during tenant upsert, this is a bug."
2778 : );
2779 0 : Err(TenantSlotUpsertError::InternalError(
2780 0 : "Missing InProgress marker during tenant upsert".into(),
2781 0 : ))
2782 : }
2783 0 : Some(slot) => {
2784 0 : METRICS.unexpected_errors.inc();
2785 0 : error!(tenant_shard_id=%self.tenant_shard_id, "Unexpected contents of TenantSlot during upsert, this is a bug. Contents: {:?}", slot);
2786 0 : Err(TenantSlotUpsertError::InternalError(
2787 0 : "Unexpected contents of TenantSlot".into(),
2788 0 : ))
2789 : }
2790 : }
2791 0 : }
2792 :
2793 : /// Replace the InProgress slot with whatever was in the guard when we started
2794 0 : fn revert(mut self) {
2795 0 : if let Some(value) = self.old_value.take() {
2796 0 : match self.upsert(value) {
2797 0 : Err(TenantSlotUpsertError::InternalError(_)) => {
2798 0 : // We already logged the error, nothing else we can do.
2799 0 : }
2800 : Err(
2801 : TenantSlotUpsertError::MapState(_) | TenantSlotUpsertError::ShuttingDown(_),
2802 0 : ) => {
2803 0 : // If the map is shutting down, we need not replace anything
2804 0 : }
2805 0 : Ok(()) => {}
2806 : }
2807 0 : }
2808 0 : }
2809 :
2810 : /// We may never drop our old value until it is cleanly shut down: otherwise we might leave
2811 : /// rogue background tasks that would write to the local tenant directory that this guard
2812 : /// is responsible for protecting
2813 1 : fn old_value_is_shutdown(&self) -> bool {
2814 1 : match self.old_value.as_ref() {
2815 1 : Some(TenantSlot::Attached(tenant)) => tenant.gate.close_complete(),
2816 0 : Some(TenantSlot::Secondary(secondary_tenant)) => secondary_tenant.gate.close_complete(),
2817 : Some(TenantSlot::InProgress(_)) => {
2818 : // A SlotGuard cannot be constructed for a slot that was already InProgress
2819 0 : unreachable!()
2820 : }
2821 0 : None => true,
2822 : }
2823 1 : }
2824 :
2825 : /// The guard holder is done with the old value of the slot: they are obliged to already
2826 : /// shut it down before we reach this point.
2827 1 : fn drop_old_value(&mut self) -> Result<(), TenantSlotDropError> {
2828 1 : if !self.old_value_is_shutdown() {
2829 0 : Err(TenantSlotDropError::NotShutdown)
2830 : } else {
2831 1 : self.old_value.take();
2832 1 : Ok(())
2833 : }
2834 1 : }
2835 : }
2836 :
2837 : impl<'a> Drop for SlotGuard<'a> {
2838 1 : fn drop(&mut self) {
2839 1 : if self.upserted {
2840 0 : return;
2841 1 : }
2842 : // Our old value is already shutdown, or it never existed: it is safe
2843 : // for us to fully release the TenantSlot back into an empty state
2844 :
2845 1 : let mut locked = self.tenants.write().unwrap();
2846 :
2847 1 : let m = match &mut *locked {
2848 : TenantsMap::Initializing => {
2849 : // There is no map, this should never happen.
2850 0 : return;
2851 : }
2852 : TenantsMap::ShuttingDown(_) => {
2853 : // When we transition to shutdown, InProgress elements are removed
2854 : // from the map, so we do not need to clean up our Inprogress marker.
2855 : // See [`shutdown_all_tenants0`]
2856 1 : return;
2857 : }
2858 0 : TenantsMap::Open(m) => m,
2859 : };
2860 :
2861 : use std::collections::btree_map::Entry;
2862 0 : match m.entry(self.tenant_shard_id) {
2863 0 : Entry::Occupied(mut entry) => {
2864 0 : if !matches!(entry.get(), TenantSlot::InProgress(_)) {
2865 0 : METRICS.unexpected_errors.inc();
2866 0 : error!(tenant_shard_id=%self.tenant_shard_id, "Unexpected contents of TenantSlot during drop, this is a bug. Contents: {:?}", entry.get());
2867 0 : }
2868 :
2869 0 : if self.old_value_is_shutdown() {
2870 0 : METRICS.slot_removed(entry.get());
2871 0 : entry.remove();
2872 0 : } else {
2873 0 : let inserting = self.old_value.take().unwrap();
2874 0 : METRICS.slot_inserted(&inserting);
2875 0 : let replaced = entry.insert(inserting);
2876 0 : METRICS.slot_removed(&replaced);
2877 0 : }
2878 : }
2879 : Entry::Vacant(_) => {
2880 0 : METRICS.unexpected_errors.inc();
2881 0 : error!(
2882 : tenant_shard_id = %self.tenant_shard_id,
2883 0 : "Missing InProgress marker during SlotGuard drop, this is a bug."
2884 : );
2885 : }
2886 : }
2887 1 : }
2888 : }
2889 :
2890 : enum TenantSlotPeekMode {
2891 : /// In Read mode, peek will be permitted to see the slots even if the pageserver is shutting down
2892 : Read,
2893 : /// In Write mode, trying to peek at a slot while the pageserver is shutting down is an error
2894 : Write,
2895 : }
2896 :
2897 0 : fn tenant_map_peek_slot<'a>(
2898 0 : tenants: &'a std::sync::RwLockReadGuard<'a, TenantsMap>,
2899 0 : tenant_shard_id: &TenantShardId,
2900 0 : mode: TenantSlotPeekMode,
2901 0 : ) -> Result<Option<&'a TenantSlot>, TenantMapError> {
2902 0 : match tenants.deref() {
2903 0 : TenantsMap::Initializing => Err(TenantMapError::StillInitializing),
2904 0 : TenantsMap::ShuttingDown(m) => match mode {
2905 : TenantSlotPeekMode::Read => Ok(Some(
2906 : // When reading in ShuttingDown state, we must translate None results
2907 : // into a ShuttingDown error, because absence of a tenant shard ID in the map
2908 : // isn't a reliable indicator of the tenant being gone: it might have been
2909 : // InProgress when shutdown started, and cleaned up from that state such
2910 : // that it's now no longer in the map. Callers will have to wait until
2911 : // we next start up to get a proper answer. This avoids incorrect 404 API responses.
2912 0 : m.get(tenant_shard_id).ok_or(TenantMapError::ShuttingDown)?,
2913 : )),
2914 0 : TenantSlotPeekMode::Write => Err(TenantMapError::ShuttingDown),
2915 : },
2916 0 : TenantsMap::Open(m) => Ok(m.get(tenant_shard_id)),
2917 : }
2918 0 : }
2919 :
2920 : enum TenantSlotAcquireMode {
2921 : /// Acquire the slot irrespective of current state, or whether it already exists
2922 : Any,
2923 : /// Return an error if trying to acquire a slot and it doesn't already exist
2924 : MustExist,
2925 : }
2926 :
2927 : use http_utils::error::ApiError;
2928 : use pageserver_api::models::TimelineGcRequest;
2929 :
2930 : use crate::tenant::gc_result::GcResult;
2931 :
2932 : #[cfg(test)]
2933 : mod tests {
2934 : use std::collections::BTreeMap;
2935 : use std::sync::Arc;
2936 :
2937 : use camino::Utf8PathBuf;
2938 : use storage_broker::BrokerClientChannel;
2939 : use tracing::Instrument;
2940 :
2941 : use super::super::harness::TenantHarness;
2942 : use super::TenantsMap;
2943 : use crate::{
2944 : basebackup_cache::BasebackupCache,
2945 : tenant::{
2946 : TenantSharedResources,
2947 : mgr::{BackgroundPurges, TenantManager, TenantSlot},
2948 : },
2949 : };
2950 :
2951 : #[tokio::test(start_paused = true)]
2952 1 : async fn shutdown_awaits_in_progress_tenant() {
2953 : // Test that if an InProgress tenant is in the map during shutdown, the shutdown will gracefully
2954 : // wait for it to complete before proceeding.
2955 :
2956 1 : let h = TenantHarness::create("shutdown_awaits_in_progress_tenant")
2957 1 : .await
2958 1 : .unwrap();
2959 1 : let (t, _ctx) = h.load().await;
2960 :
2961 : // harness loads it to active, which is forced and nothing is running on the tenant
2962 :
2963 1 : let id = t.tenant_shard_id();
2964 :
2965 : // tenant harness configures the logging and we cannot escape it
2966 1 : let span = h.span();
2967 1 : let _e = span.enter();
2968 :
2969 1 : let tenants = BTreeMap::from([(id, TenantSlot::Attached(t.clone()))]);
2970 :
2971 : // Invoke remove_tenant_from_memory with a cleanup hook that blocks until we manually
2972 : // permit it to proceed: that will stick the tenant in InProgress
2973 :
2974 1 : let (basebackup_cache, _) = BasebackupCache::new(Utf8PathBuf::new(), None);
2975 :
2976 1 : let tenant_manager = TenantManager {
2977 1 : tenants: std::sync::RwLock::new(TenantsMap::Open(tenants)),
2978 1 : conf: h.conf,
2979 : resources: TenantSharedResources {
2980 1 : broker_client: BrokerClientChannel::connect_lazy("foobar.com")
2981 1 : .await
2982 1 : .unwrap(),
2983 1 : remote_storage: h.remote_storage.clone(),
2984 1 : deletion_queue_client: h.deletion_queue.new_client(),
2985 1 : l0_flush_global_state: crate::l0_flush::L0FlushGlobalState::new(
2986 1 : h.conf.l0_flush.clone(),
2987 : ),
2988 1 : basebackup_cache,
2989 1 : feature_resolver: crate::feature_resolver::FeatureResolver::new_disabled(),
2990 : },
2991 1 : cancel: tokio_util::sync::CancellationToken::new(),
2992 1 : background_purges: BackgroundPurges::default(),
2993 : };
2994 :
2995 1 : let tenant_manager = Arc::new(tenant_manager);
2996 :
2997 1 : let (until_cleanup_completed, can_complete_cleanup) = utils::completion::channel();
2998 1 : let (until_cleanup_started, cleanup_started) = utils::completion::channel();
2999 1 : let mut remove_tenant_from_memory_task = {
3000 1 : let tenant_manager = tenant_manager.clone();
3001 1 : let jh = tokio::spawn({
3002 1 : async move {
3003 1 : let cleanup = async move {
3004 1 : drop(until_cleanup_started);
3005 1 : can_complete_cleanup.wait().await;
3006 1 : anyhow::Ok(())
3007 1 : };
3008 1 : tenant_manager.remove_tenant_from_memory(id, cleanup).await
3009 1 : }
3010 1 : .instrument(h.span())
3011 : });
3012 :
3013 : // now the long cleanup should be in place, with the stopping state
3014 1 : cleanup_started.wait().await;
3015 1 : jh
3016 : };
3017 :
3018 1 : let mut shutdown_task = {
3019 1 : let (until_shutdown_started, shutdown_started) = utils::completion::channel();
3020 :
3021 1 : let tenant_manager = tenant_manager.clone();
3022 :
3023 1 : let shutdown_task = tokio::spawn(async move {
3024 1 : drop(until_shutdown_started);
3025 1 : tenant_manager.shutdown_all_tenants0().await;
3026 1 : });
3027 :
3028 1 : shutdown_started.wait().await;
3029 1 : shutdown_task
3030 : };
3031 :
3032 1 : let long_time = std::time::Duration::from_secs(15);
3033 1 : tokio::select! {
3034 1 : _ = &mut shutdown_task => unreachable!("shutdown should block on remove_tenant_from_memory completing"),
3035 1 : _ = &mut remove_tenant_from_memory_task => unreachable!("remove_tenant_from_memory_task should not complete until explicitly unblocked"),
3036 1 : _ = tokio::time::sleep(long_time) => {},
3037 : }
3038 :
3039 1 : drop(until_cleanup_completed);
3040 :
3041 : // Now that we allow it to proceed, shutdown should complete immediately
3042 1 : remove_tenant_from_memory_task.await.unwrap().unwrap();
3043 1 : shutdown_task.await.unwrap();
3044 1 : }
3045 : }
|
