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 : // This function removs remaining traces of a timeline on disk.
139 : // Namely: metadata file, timeline directory, delete mark.
140 : // Note: io::ErrorKind::NotFound are ignored for metadata and timeline dir.
141 : // delete mark should be present because it is the last step during deletion.
142 : // (nothing can fail after its deletion)
143 0 : async fn cleanup_remaining_timeline_fs_traces(
144 0 : conf: &PageServerConf,
145 0 : tenant_shard_id: TenantShardId,
146 0 : timeline_id: TimelineId,
147 0 : ) -> anyhow::Result<()> {
148 0 : // Remove delete mark
149 0 : // TODO: once we are confident that no more exist in the field, remove this
150 0 : // line. It cleans up a legacy marker file that might in rare cases be present.
151 0 : tokio::fs::remove_file(conf.timeline_delete_mark_file_path(tenant_shard_id, timeline_id))
152 0 : .await
153 0 : .or_else(fs_ext::ignore_not_found)
154 0 : .context("remove delete mark")
155 0 : }
156 :
157 : /// It is important that this gets called when DeletionGuard is being held.
158 : /// For more context see comments in [`DeleteTimelineFlow::prepare`]
159 0 : async fn remove_timeline_from_tenant(
160 0 : tenant: &Tenant,
161 0 : timeline: &Timeline,
162 0 : _: &DeletionGuard, // using it as a witness
163 0 : ) -> anyhow::Result<()> {
164 0 : // Remove the timeline from the map.
165 0 : let mut timelines = tenant.timelines.lock().unwrap();
166 0 : let children_exist = timelines
167 0 : .iter()
168 0 : .any(|(_, entry)| entry.get_ancestor_timeline_id() == Some(timeline.timeline_id));
169 0 : // XXX this can happen because `branch_timeline` doesn't check `TimelineState::Stopping`.
170 0 : // We already deleted the layer files, so it's probably best to panic.
171 0 : // (Ideally, above remove_dir_all is atomic so we don't see this timeline after a restart)
172 0 : if children_exist {
173 0 : panic!("Timeline grew children while we removed layer files");
174 0 : }
175 0 :
176 0 : timelines
177 0 : .remove(&timeline.timeline_id)
178 0 : .expect("timeline that we were deleting was concurrently removed from 'timelines' map");
179 0 :
180 0 : drop(timelines);
181 0 :
182 0 : Ok(())
183 0 : }
184 :
185 : /// Orchestrates timeline shut down of all timeline tasks, removes its in-memory structures,
186 : /// and deletes its data from both disk and s3.
187 : /// The sequence of steps:
188 : /// 1. Set deleted_at in remote index part.
189 : /// 2. Create local mark file.
190 : /// 3. Delete local files except metadata (it is simpler this way, to be able to reuse timeline initialization code that expects metadata)
191 : /// 4. Delete remote layers
192 : /// 5. Delete index part
193 : /// 6. Delete meta, timeline directory
194 : /// 7. Delete mark file
195 : ///
196 : /// It is resumable from any step in case a crash/restart occurs.
197 : /// There are three entrypoints to the process:
198 : /// 1. [`DeleteTimelineFlow::run`] this is the main one called by a management api handler.
199 : /// 2. [`DeleteTimelineFlow::resume_deletion`] is called during restarts when local metadata is still present
200 : /// and we possibly neeed to continue deletion of remote files.
201 : /// 3. [`DeleteTimelineFlow::cleanup_remaining_timeline_fs_traces`] is used when we deleted remote
202 : /// index but still have local metadata, timeline directory and delete mark.
203 : ///
204 : /// Note the only other place that messes around timeline delete mark is the logic that scans directory with timelines during tenant load.
205 : #[derive(Default)]
206 : pub enum DeleteTimelineFlow {
207 : #[default]
208 : NotStarted,
209 : InProgress,
210 : Finished,
211 : }
212 :
213 : impl DeleteTimelineFlow {
214 : // These steps are run in the context of management api request handler.
215 : // Long running steps are continued to run in the background.
216 : // NB: If this fails half-way through, and is retried, the retry will go through
217 : // all the same steps again. Make sure the code here is idempotent, and don't
218 : // error out if some of the shutdown tasks have already been completed!
219 0 : #[instrument(skip_all)]
220 : pub async fn run(
221 : tenant: &Arc<Tenant>,
222 : timeline_id: TimelineId,
223 : ) -> Result<(), DeleteTimelineError> {
224 : super::debug_assert_current_span_has_tenant_and_timeline_id();
225 :
226 : let (timeline, mut guard) = Self::prepare(tenant, timeline_id)?;
227 :
228 : guard.mark_in_progress()?;
229 :
230 : // Now that the Timeline is in Stopping state, request all the related tasks to shut down.
231 : timeline.shutdown(super::ShutdownMode::Hard).await;
232 :
233 : tenant.gc_block.before_delete(&timeline);
234 :
235 0 : fail::fail_point!("timeline-delete-before-index-deleted-at", |_| {
236 0 : Err(anyhow::anyhow!(
237 0 : "failpoint: timeline-delete-before-index-deleted-at"
238 0 : ))?
239 0 : });
240 :
241 : set_deleted_in_remote_index(&timeline).await?;
242 :
243 0 : fail::fail_point!("timeline-delete-before-schedule", |_| {
244 0 : Err(anyhow::anyhow!(
245 0 : "failpoint: timeline-delete-before-schedule"
246 0 : ))?
247 0 : });
248 :
249 : Self::schedule_background(guard, tenant.conf, Arc::clone(tenant), timeline);
250 :
251 : Ok(())
252 : }
253 :
254 0 : fn mark_in_progress(&mut self) -> anyhow::Result<()> {
255 0 : match self {
256 0 : Self::Finished => anyhow::bail!("Bug. Is in finished state"),
257 0 : Self::InProgress { .. } => { /* We're in a retry */ }
258 0 : Self::NotStarted => { /* Fresh start */ }
259 : }
260 :
261 0 : *self = Self::InProgress;
262 0 :
263 0 : Ok(())
264 0 : }
265 :
266 : /// Shortcut to create Timeline in stopping state and spawn deletion task.
267 0 : #[instrument(skip_all, fields(%timeline_id))]
268 : pub async fn resume_deletion(
269 : tenant: Arc<Tenant>,
270 : timeline_id: TimelineId,
271 : local_metadata: &TimelineMetadata,
272 : remote_client: RemoteTimelineClient,
273 : ) -> anyhow::Result<()> {
274 : // Note: here we even skip populating layer map. Timeline is essentially uninitialized.
275 : // RemoteTimelineClient is the only functioning part.
276 : let timeline = tenant
277 : .create_timeline_struct(
278 : timeline_id,
279 : local_metadata,
280 : None, // Ancestor is not needed for deletion.
281 : TimelineResources {
282 : remote_client,
283 : timeline_get_throttle: tenant.timeline_get_throttle.clone(),
284 : l0_flush_global_state: tenant.l0_flush_global_state.clone(),
285 : },
286 : // Important. We dont pass ancestor above because it can be missing.
287 : // Thus we need to skip the validation here.
288 : CreateTimelineCause::Delete,
289 : // Aux file policy is not needed for deletion, assuming deletion does not read aux keyspace
290 : None,
291 : )
292 : .context("create_timeline_struct")?;
293 :
294 : let mut guard = DeletionGuard(
295 : Arc::clone(&timeline.delete_progress)
296 : .try_lock_owned()
297 : .expect("cannot happen because we're the only owner"),
298 : );
299 :
300 : // We meed to do this because when console retries delete request we shouldnt answer with 404
301 : // because 404 means successful deletion.
302 : {
303 : let mut locked = tenant.timelines.lock().unwrap();
304 : locked.insert(timeline_id, Arc::clone(&timeline));
305 : }
306 :
307 : guard.mark_in_progress()?;
308 :
309 : Self::schedule_background(guard, tenant.conf, tenant, timeline);
310 :
311 : Ok(())
312 : }
313 :
314 0 : #[instrument(skip_all, fields(%timeline_id))]
315 : pub async fn cleanup_remaining_timeline_fs_traces(
316 : tenant: &Tenant,
317 : timeline_id: TimelineId,
318 : ) -> anyhow::Result<()> {
319 : let r =
320 : cleanup_remaining_timeline_fs_traces(tenant.conf, tenant.tenant_shard_id, timeline_id)
321 : .await;
322 : info!("Done");
323 : r
324 : }
325 :
326 0 : fn prepare(
327 0 : tenant: &Tenant,
328 0 : timeline_id: TimelineId,
329 0 : ) -> Result<(Arc<Timeline>, DeletionGuard), DeleteTimelineError> {
330 0 : // Note the interaction between this guard and deletion guard.
331 0 : // Here we attempt to lock deletion guard when we're holding a lock on timelines.
332 0 : // This is important because when you take into account `remove_timeline_from_tenant`
333 0 : // we remove timeline from memory when we still hold the deletion guard.
334 0 : // So here when timeline deletion is finished timeline wont be present in timelines map at all
335 0 : // which makes the following sequence impossible:
336 0 : // T1: get preempted right before the try_lock on `Timeline::delete_progress`
337 0 : // T2: do a full deletion, acquire and drop `Timeline::delete_progress`
338 0 : // T1: acquire deletion lock, do another `DeleteTimelineFlow::run`
339 0 : // For more context see this discussion: `https://github.com/neondatabase/neon/pull/4552#discussion_r1253437346`
340 0 : let timelines = tenant.timelines.lock().unwrap();
341 :
342 0 : let timeline = match timelines.get(&timeline_id) {
343 0 : Some(t) => t,
344 0 : None => return Err(DeleteTimelineError::NotFound),
345 : };
346 :
347 : // Ensure that there are no child timelines **attached to that pageserver**,
348 : // because detach removes files, which will break child branches
349 0 : let children: Vec<TimelineId> = timelines
350 0 : .iter()
351 0 : .filter_map(|(id, entry)| {
352 0 : if entry.get_ancestor_timeline_id() == Some(timeline_id) {
353 0 : Some(*id)
354 : } else {
355 0 : None
356 : }
357 0 : })
358 0 : .collect();
359 0 :
360 0 : if !children.is_empty() {
361 0 : return Err(DeleteTimelineError::HasChildren(children));
362 0 : }
363 0 :
364 0 : // Note that using try_lock here is important to avoid a deadlock.
365 0 : // Here we take lock on timelines and then the deletion guard.
366 0 : // At the end of the operation we're holding the guard and need to lock timelines map
367 0 : // to remove the timeline from it.
368 0 : // Always if you have two locks that are taken in different order this can result in a deadlock.
369 0 :
370 0 : let delete_progress = Arc::clone(&timeline.delete_progress);
371 0 : let delete_lock_guard = match delete_progress.try_lock_owned() {
372 0 : Ok(guard) => DeletionGuard(guard),
373 : Err(_) => {
374 : // Unfortunately if lock fails arc is consumed.
375 0 : return Err(DeleteTimelineError::AlreadyInProgress(Arc::clone(
376 0 : &timeline.delete_progress,
377 0 : )));
378 : }
379 : };
380 :
381 0 : timeline.set_state(TimelineState::Stopping);
382 0 :
383 0 : Ok((Arc::clone(timeline), delete_lock_guard))
384 0 : }
385 :
386 0 : fn schedule_background(
387 0 : guard: DeletionGuard,
388 0 : conf: &'static PageServerConf,
389 0 : tenant: Arc<Tenant>,
390 0 : timeline: Arc<Timeline>,
391 0 : ) {
392 0 : let tenant_shard_id = timeline.tenant_shard_id;
393 0 : let timeline_id = timeline.timeline_id;
394 0 :
395 0 : task_mgr::spawn(
396 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
397 0 : TaskKind::TimelineDeletionWorker,
398 0 : tenant_shard_id,
399 0 : Some(timeline_id),
400 0 : "timeline_delete",
401 0 : async move {
402 0 : if let Err(err) = Self::background(guard, conf, &tenant, &timeline).await {
403 0 : error!("Error: {err:#}");
404 0 : timeline.set_broken(format!("{err:#}"))
405 0 : };
406 0 : Ok(())
407 0 : }
408 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)),
409 : );
410 0 : }
411 :
412 0 : async fn background(
413 0 : mut guard: DeletionGuard,
414 0 : conf: &PageServerConf,
415 0 : tenant: &Tenant,
416 0 : timeline: &Timeline,
417 0 : ) -> Result<(), DeleteTimelineError> {
418 0 : delete_local_timeline_directory(conf, tenant.tenant_shard_id, timeline).await?;
419 :
420 0 : delete_remote_layers_and_index(timeline).await?;
421 :
422 0 : pausable_failpoint!("in_progress_delete");
423 :
424 0 : remove_timeline_from_tenant(tenant, timeline, &guard).await?;
425 :
426 0 : *guard = Self::Finished;
427 0 :
428 0 : Ok(())
429 0 : }
430 :
431 0 : pub(crate) fn is_not_started(&self) -> bool {
432 0 : matches!(self, Self::NotStarted)
433 0 : }
434 : }
435 :
436 : struct DeletionGuard(OwnedMutexGuard<DeleteTimelineFlow>);
437 :
438 : impl Deref for DeletionGuard {
439 : type Target = DeleteTimelineFlow;
440 :
441 0 : fn deref(&self) -> &Self::Target {
442 0 : &self.0
443 0 : }
444 : }
445 :
446 : impl DerefMut for DeletionGuard {
447 0 : fn deref_mut(&mut self) -> &mut Self::Target {
448 0 : &mut self.0
449 0 : }
450 : }
|
