Line data Source code
1 : use std::{
2 : ops::{Deref, DerefMut},
3 : sync::Arc,
4 : };
5 :
6 : use anyhow::Context;
7 : use pageserver_api::{models::TimelineState, shard::TenantShardId};
8 : use tokio::sync::OwnedMutexGuard;
9 : use tracing::{error, info, instrument, Instrument};
10 : use utils::{crashsafe, fs_ext, id::TimelineId, pausable_failpoint};
11 :
12 : use crate::{
13 : config::PageServerConf,
14 : task_mgr::{self, TaskKind},
15 : tenant::{
16 : metadata::TimelineMetadata,
17 : remote_timeline_client::{PersistIndexPartWithDeletedFlagError, RemoteTimelineClient},
18 : CreateTimelineCause, DeleteTimelineError, Tenant,
19 : },
20 : };
21 :
22 : use super::{Timeline, TimelineResources};
23 :
24 : /// Mark timeline as deleted in S3 so we won't pick it up next time
25 : /// during attach or pageserver restart.
26 : /// See comment in persist_index_part_with_deleted_flag.
27 0 : async fn set_deleted_in_remote_index(timeline: &Timeline) -> Result<(), DeleteTimelineError> {
28 0 : match timeline
29 0 : .remote_client
30 0 : .persist_index_part_with_deleted_flag()
31 0 : .await
32 : {
33 : // If we (now, or already) marked it successfully as deleted, we can proceed
34 0 : Ok(()) | Err(PersistIndexPartWithDeletedFlagError::AlreadyDeleted(_)) => (),
35 : // Bail out otherwise
36 : //
37 : // AlreadyInProgress shouldn't happen, because the 'delete_lock' prevents
38 : // two tasks from performing the deletion at the same time. The first task
39 : // that starts deletion should run it to completion.
40 0 : Err(e @ PersistIndexPartWithDeletedFlagError::AlreadyInProgress(_))
41 0 : | Err(e @ PersistIndexPartWithDeletedFlagError::Other(_)) => {
42 0 : return Err(DeleteTimelineError::Other(anyhow::anyhow!(e)));
43 : }
44 : }
45 0 : Ok(())
46 0 : }
47 :
48 : /// Grab the compaction and gc locks, and actually perform the deletion.
49 : ///
50 : /// The locks prevent GC or compaction from running at the same time. The background tasks do not
51 : /// register themselves with the timeline it's operating on, so it might still be running even
52 : /// though we called `shutdown_tasks`.
53 : ///
54 : /// Note that there are still other race conditions between
55 : /// GC, compaction and timeline deletion. See
56 : /// <https://github.com/neondatabase/neon/issues/2671>
57 : ///
58 : /// No timeout here, GC & Compaction should be responsive to the
59 : /// `TimelineState::Stopping` change.
60 : // pub(super): documentation link
61 0 : pub(super) async fn delete_local_timeline_directory(
62 0 : conf: &PageServerConf,
63 0 : tenant_shard_id: TenantShardId,
64 0 : timeline: &Timeline,
65 0 : ) -> anyhow::Result<()> {
66 0 : // Always ensure the lock order is compaction -> gc.
67 0 : let compaction_lock = timeline.compaction_lock.lock();
68 0 : let compaction_lock = crate::timed(
69 0 : compaction_lock,
70 0 : "acquires compaction lock",
71 0 : std::time::Duration::from_secs(5),
72 0 : )
73 0 : .await;
74 :
75 0 : let gc_lock = timeline.gc_lock.lock();
76 0 : let gc_lock = crate::timed(
77 0 : gc_lock,
78 0 : "acquires gc lock",
79 0 : std::time::Duration::from_secs(5),
80 0 : )
81 0 : .await;
82 :
83 : // NB: storage_sync upload tasks that reference these layers have been cancelled
84 : // by the caller.
85 :
86 0 : let local_timeline_directory = conf.timeline_path(&tenant_shard_id, &timeline.timeline_id);
87 0 :
88 0 : fail::fail_point!("timeline-delete-before-rm", |_| {
89 0 : Err(anyhow::anyhow!("failpoint: timeline-delete-before-rm"))?
90 0 : });
91 :
92 : // NB: This need not be atomic because the deleted flag in the IndexPart
93 : // will be observed during tenant/timeline load. The deletion will be resumed there.
94 : //
95 : // Note that here we do not bail out on std::io::ErrorKind::NotFound.
96 : // This can happen if we're called a second time, e.g.,
97 : // because of a previous failure/cancellation at/after
98 : // failpoint timeline-delete-after-rm.
99 : //
100 : // ErrorKind::NotFound can also happen if we race with tenant detach, because,
101 : // no locks are shared.
102 0 : tokio::fs::remove_dir_all(local_timeline_directory)
103 0 : .await
104 0 : .or_else(fs_ext::ignore_not_found)
105 0 : .context("remove local timeline directory")?;
106 :
107 : // Make sure previous deletions are ordered before mark removal.
108 : // Otherwise there is no guarantee that they reach the disk before mark deletion.
109 : // So its possible for mark to reach disk first and for other deletions
110 : // to be reordered later and thus missed if a crash occurs.
111 : // Note that we dont need to sync after mark file is removed
112 : // because we can tolerate the case when mark file reappears on startup.
113 0 : let timeline_path = conf.timelines_path(&tenant_shard_id);
114 0 : crashsafe::fsync_async(timeline_path)
115 0 : .await
116 0 : .context("fsync_pre_mark_remove")?;
117 :
118 0 : info!("finished deleting layer files, releasing locks");
119 0 : drop(gc_lock);
120 0 : drop(compaction_lock);
121 0 :
122 0 : fail::fail_point!("timeline-delete-after-rm", |_| {
123 0 : Err(anyhow::anyhow!("failpoint: timeline-delete-after-rm"))?
124 0 : });
125 :
126 0 : Ok(())
127 0 : }
128 :
129 : /// Removes remote layers and an index file after them.
130 0 : async fn delete_remote_layers_and_index(timeline: &Timeline) -> anyhow::Result<()> {
131 0 : timeline
132 0 : .remote_client
133 0 : .delete_all()
134 0 : .await
135 0 : .context("delete_all")
136 0 : }
137 :
138 : /// It is important that this gets called when DeletionGuard is being held.
139 : /// For more context see comments in [`DeleteTimelineFlow::prepare`]
140 0 : async fn remove_timeline_from_tenant(
141 0 : tenant: &Tenant,
142 0 : timeline: &Timeline,
143 0 : _: &DeletionGuard, // using it as a witness
144 0 : ) -> anyhow::Result<()> {
145 0 : // Remove the timeline from the map.
146 0 : let mut timelines = tenant.timelines.lock().unwrap();
147 0 : let children_exist = timelines
148 0 : .iter()
149 0 : .any(|(_, entry)| entry.get_ancestor_timeline_id() == Some(timeline.timeline_id));
150 0 : // XXX this can happen because `branch_timeline` doesn't check `TimelineState::Stopping`.
151 0 : // We already deleted the layer files, so it's probably best to panic.
152 0 : // (Ideally, above remove_dir_all is atomic so we don't see this timeline after a restart)
153 0 : if children_exist {
154 0 : panic!("Timeline grew children while we removed layer files");
155 0 : }
156 0 :
157 0 : timelines
158 0 : .remove(&timeline.timeline_id)
159 0 : .expect("timeline that we were deleting was concurrently removed from 'timelines' map");
160 0 :
161 0 : drop(timelines);
162 0 :
163 0 : Ok(())
164 0 : }
165 :
166 : /// Orchestrates timeline shut down of all timeline tasks, removes its in-memory structures,
167 : /// and deletes its data from both disk and s3.
168 : /// The sequence of steps:
169 : /// 1. Set deleted_at in remote index part.
170 : /// 2. Create local mark file.
171 : /// 3. Delete local files except metadata (it is simpler this way, to be able to reuse timeline initialization code that expects metadata)
172 : /// 4. Delete remote layers
173 : /// 5. Delete index part
174 : /// 6. Delete meta, timeline directory
175 : /// 7. Delete mark file
176 : ///
177 : /// It is resumable from any step in case a crash/restart occurs.
178 : /// There are two entrypoints to the process:
179 : /// 1. [`DeleteTimelineFlow::run`] this is the main one called by a management api handler.
180 : /// 2. [`DeleteTimelineFlow::resume_deletion`] is called during restarts when local metadata is still present
181 : /// and we possibly neeed to continue deletion of remote files.
182 : ///
183 : /// Note the only other place that messes around timeline delete mark is the logic that scans directory with timelines during tenant load.
184 : #[derive(Default)]
185 : pub enum DeleteTimelineFlow {
186 : #[default]
187 : NotStarted,
188 : InProgress,
189 : Finished,
190 : }
191 :
192 : impl DeleteTimelineFlow {
193 : // These steps are run in the context of management api request handler.
194 : // Long running steps are continued to run in the background.
195 : // NB: If this fails half-way through, and is retried, the retry will go through
196 : // all the same steps again. Make sure the code here is idempotent, and don't
197 : // error out if some of the shutdown tasks have already been completed!
198 0 : #[instrument(skip_all)]
199 : pub async fn run(
200 : tenant: &Arc<Tenant>,
201 : timeline_id: TimelineId,
202 : ) -> Result<(), DeleteTimelineError> {
203 : super::debug_assert_current_span_has_tenant_and_timeline_id();
204 :
205 : let (timeline, mut guard) = Self::prepare(tenant, timeline_id)?;
206 :
207 : guard.mark_in_progress()?;
208 :
209 : // Now that the Timeline is in Stopping state, request all the related tasks to shut down.
210 : timeline.shutdown(super::ShutdownMode::Hard).await;
211 :
212 : tenant.gc_block.before_delete(&timeline);
213 :
214 0 : fail::fail_point!("timeline-delete-before-index-deleted-at", |_| {
215 0 : Err(anyhow::anyhow!(
216 0 : "failpoint: timeline-delete-before-index-deleted-at"
217 0 : ))?
218 0 : });
219 :
220 : set_deleted_in_remote_index(&timeline).await?;
221 :
222 0 : fail::fail_point!("timeline-delete-before-schedule", |_| {
223 0 : Err(anyhow::anyhow!(
224 0 : "failpoint: timeline-delete-before-schedule"
225 0 : ))?
226 0 : });
227 :
228 : Self::schedule_background(guard, tenant.conf, Arc::clone(tenant), timeline);
229 :
230 : Ok(())
231 : }
232 :
233 0 : fn mark_in_progress(&mut self) -> anyhow::Result<()> {
234 0 : match self {
235 0 : Self::Finished => anyhow::bail!("Bug. Is in finished state"),
236 0 : Self::InProgress { .. } => { /* We're in a retry */ }
237 0 : Self::NotStarted => { /* Fresh start */ }
238 : }
239 :
240 0 : *self = Self::InProgress;
241 0 :
242 0 : Ok(())
243 0 : }
244 :
245 : /// Shortcut to create Timeline in stopping state and spawn deletion task.
246 0 : #[instrument(skip_all, fields(%timeline_id))]
247 : pub async fn resume_deletion(
248 : tenant: Arc<Tenant>,
249 : timeline_id: TimelineId,
250 : local_metadata: &TimelineMetadata,
251 : remote_client: RemoteTimelineClient,
252 : ) -> anyhow::Result<()> {
253 : // Note: here we even skip populating layer map. Timeline is essentially uninitialized.
254 : // RemoteTimelineClient is the only functioning part.
255 : let timeline = tenant
256 : .create_timeline_struct(
257 : timeline_id,
258 : local_metadata,
259 : None, // Ancestor is not needed for deletion.
260 : TimelineResources {
261 : remote_client,
262 : timeline_get_throttle: tenant.timeline_get_throttle.clone(),
263 : l0_flush_global_state: tenant.l0_flush_global_state.clone(),
264 : },
265 : // Important. We dont pass ancestor above because it can be missing.
266 : // Thus we need to skip the validation here.
267 : CreateTimelineCause::Delete,
268 : // Aux file policy is not needed for deletion, assuming deletion does not read aux keyspace
269 : None,
270 : )
271 : .context("create_timeline_struct")?;
272 :
273 : let mut guard = DeletionGuard(
274 : Arc::clone(&timeline.delete_progress)
275 : .try_lock_owned()
276 : .expect("cannot happen because we're the only owner"),
277 : );
278 :
279 : // We meed to do this because when console retries delete request we shouldnt answer with 404
280 : // because 404 means successful deletion.
281 : {
282 : let mut locked = tenant.timelines.lock().unwrap();
283 : locked.insert(timeline_id, Arc::clone(&timeline));
284 : }
285 :
286 : guard.mark_in_progress()?;
287 :
288 : Self::schedule_background(guard, tenant.conf, tenant, timeline);
289 :
290 : Ok(())
291 : }
292 :
293 0 : fn prepare(
294 0 : tenant: &Tenant,
295 0 : timeline_id: TimelineId,
296 0 : ) -> Result<(Arc<Timeline>, DeletionGuard), DeleteTimelineError> {
297 0 : // Note the interaction between this guard and deletion guard.
298 0 : // Here we attempt to lock deletion guard when we're holding a lock on timelines.
299 0 : // This is important because when you take into account `remove_timeline_from_tenant`
300 0 : // we remove timeline from memory when we still hold the deletion guard.
301 0 : // So here when timeline deletion is finished timeline wont be present in timelines map at all
302 0 : // which makes the following sequence impossible:
303 0 : // T1: get preempted right before the try_lock on `Timeline::delete_progress`
304 0 : // T2: do a full deletion, acquire and drop `Timeline::delete_progress`
305 0 : // T1: acquire deletion lock, do another `DeleteTimelineFlow::run`
306 0 : // For more context see this discussion: `https://github.com/neondatabase/neon/pull/4552#discussion_r1253437346`
307 0 : let timelines = tenant.timelines.lock().unwrap();
308 :
309 0 : let timeline = match timelines.get(&timeline_id) {
310 0 : Some(t) => t,
311 0 : None => return Err(DeleteTimelineError::NotFound),
312 : };
313 :
314 : // Ensure that there are no child timelines **attached to that pageserver**,
315 : // because detach removes files, which will break child branches
316 0 : let children: Vec<TimelineId> = timelines
317 0 : .iter()
318 0 : .filter_map(|(id, entry)| {
319 0 : if entry.get_ancestor_timeline_id() == Some(timeline_id) {
320 0 : Some(*id)
321 : } else {
322 0 : None
323 : }
324 0 : })
325 0 : .collect();
326 0 :
327 0 : if !children.is_empty() {
328 0 : return Err(DeleteTimelineError::HasChildren(children));
329 0 : }
330 0 :
331 0 : // Note that using try_lock here is important to avoid a deadlock.
332 0 : // Here we take lock on timelines and then the deletion guard.
333 0 : // At the end of the operation we're holding the guard and need to lock timelines map
334 0 : // to remove the timeline from it.
335 0 : // Always if you have two locks that are taken in different order this can result in a deadlock.
336 0 :
337 0 : let delete_progress = Arc::clone(&timeline.delete_progress);
338 0 : let delete_lock_guard = match delete_progress.try_lock_owned() {
339 0 : Ok(guard) => DeletionGuard(guard),
340 : Err(_) => {
341 : // Unfortunately if lock fails arc is consumed.
342 0 : return Err(DeleteTimelineError::AlreadyInProgress(Arc::clone(
343 0 : &timeline.delete_progress,
344 0 : )));
345 : }
346 : };
347 :
348 0 : timeline.set_state(TimelineState::Stopping);
349 0 :
350 0 : Ok((Arc::clone(timeline), delete_lock_guard))
351 0 : }
352 :
353 0 : fn schedule_background(
354 0 : guard: DeletionGuard,
355 0 : conf: &'static PageServerConf,
356 0 : tenant: Arc<Tenant>,
357 0 : timeline: Arc<Timeline>,
358 0 : ) {
359 0 : let tenant_shard_id = timeline.tenant_shard_id;
360 0 : let timeline_id = timeline.timeline_id;
361 0 :
362 0 : task_mgr::spawn(
363 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
364 0 : TaskKind::TimelineDeletionWorker,
365 0 : tenant_shard_id,
366 0 : Some(timeline_id),
367 0 : "timeline_delete",
368 0 : async move {
369 0 : if let Err(err) = Self::background(guard, conf, &tenant, &timeline).await {
370 0 : error!("Error: {err:#}");
371 0 : timeline.set_broken(format!("{err:#}"))
372 0 : };
373 0 : Ok(())
374 0 : }
375 0 : .instrument(tracing::info_span!(parent: None, "delete_timeline", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),timeline_id=%timeline_id)),
376 : );
377 0 : }
378 :
379 0 : async fn background(
380 0 : mut guard: DeletionGuard,
381 0 : conf: &PageServerConf,
382 0 : tenant: &Tenant,
383 0 : timeline: &Timeline,
384 0 : ) -> Result<(), DeleteTimelineError> {
385 0 : delete_local_timeline_directory(conf, tenant.tenant_shard_id, timeline).await?;
386 :
387 0 : delete_remote_layers_and_index(timeline).await?;
388 :
389 0 : pausable_failpoint!("in_progress_delete");
390 :
391 0 : remove_timeline_from_tenant(tenant, timeline, &guard).await?;
392 :
393 0 : *guard = Self::Finished;
394 0 :
395 0 : Ok(())
396 0 : }
397 :
398 0 : pub(crate) fn is_not_started(&self) -> bool {
399 0 : matches!(self, Self::NotStarted)
400 0 : }
401 : }
402 :
403 : struct DeletionGuard(OwnedMutexGuard<DeleteTimelineFlow>);
404 :
405 : impl Deref for DeletionGuard {
406 : type Target = DeleteTimelineFlow;
407 :
408 0 : fn deref(&self) -> &Self::Target {
409 0 : &self.0
410 0 : }
411 : }
412 :
413 : impl DerefMut for DeletionGuard {
414 0 : fn deref_mut(&mut self) -> &mut Self::Target {
415 0 : &mut self.0
416 0 : }
417 : }
|
