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