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 : let guards = async { tokio::join!(timeline.gc_lock.lock(), timeline.compaction_lock.lock()) };
67 0 : let guards = crate::timed(
68 0 : guards,
69 0 : "acquire gc and compaction locks",
70 0 : std::time::Duration::from_secs(5),
71 0 : )
72 0 : .await;
73 :
74 : // NB: storage_sync upload tasks that reference these layers have been cancelled
75 : // by the caller.
76 :
77 0 : let local_timeline_directory = conf.timeline_path(&tenant_shard_id, &timeline.timeline_id);
78 0 :
79 0 : fail::fail_point!("timeline-delete-before-rm", |_| {
80 0 : Err(anyhow::anyhow!("failpoint: timeline-delete-before-rm"))?
81 0 : });
82 :
83 : // NB: This need not be atomic because the deleted flag in the IndexPart
84 : // will be observed during tenant/timeline load. The deletion will be resumed there.
85 : //
86 : // Note that here we do not bail out on std::io::ErrorKind::NotFound.
87 : // This can happen if we're called a second time, e.g.,
88 : // because of a previous failure/cancellation at/after
89 : // failpoint timeline-delete-after-rm.
90 : //
91 : // ErrorKind::NotFound can also happen if we race with tenant detach, because,
92 : // no locks are shared.
93 0 : tokio::fs::remove_dir_all(local_timeline_directory)
94 0 : .await
95 0 : .or_else(fs_ext::ignore_not_found)
96 0 : .context("remove local timeline directory")?;
97 :
98 : // Make sure previous deletions are ordered before mark removal.
99 : // Otherwise there is no guarantee that they reach the disk before mark deletion.
100 : // So its possible for mark to reach disk first and for other deletions
101 : // to be reordered later and thus missed if a crash occurs.
102 : // Note that we dont need to sync after mark file is removed
103 : // because we can tolerate the case when mark file reappears on startup.
104 0 : let timeline_path = conf.timelines_path(&tenant_shard_id);
105 0 : crashsafe::fsync_async(timeline_path)
106 0 : .await
107 0 : .context("fsync_pre_mark_remove")?;
108 :
109 0 : info!("finished deleting layer files, releasing locks");
110 0 : drop(guards);
111 0 :
112 0 : fail::fail_point!("timeline-delete-after-rm", |_| {
113 0 : Err(anyhow::anyhow!("failpoint: timeline-delete-after-rm"))?
114 0 : });
115 :
116 0 : Ok(())
117 0 : }
118 :
119 : /// Removes remote layers and an index file after them.
120 0 : async fn delete_remote_layers_and_index(timeline: &Timeline) -> anyhow::Result<()> {
121 0 : timeline
122 0 : .remote_client
123 0 : .delete_all()
124 0 : .await
125 0 : .context("delete_all")
126 0 : }
127 :
128 : // This function removs remaining traces of a timeline on disk.
129 : // Namely: metadata file, timeline directory, delete mark.
130 : // Note: io::ErrorKind::NotFound are ignored for metadata and timeline dir.
131 : // delete mark should be present because it is the last step during deletion.
132 : // (nothing can fail after its deletion)
133 0 : async fn cleanup_remaining_timeline_fs_traces(
134 0 : conf: &PageServerConf,
135 0 : tenant_shard_id: TenantShardId,
136 0 : timeline_id: TimelineId,
137 0 : ) -> anyhow::Result<()> {
138 0 : // Remove delete mark
139 0 : // TODO: once we are confident that no more exist in the field, remove this
140 0 : // line. It cleans up a legacy marker file that might in rare cases be present.
141 0 : tokio::fs::remove_file(conf.timeline_delete_mark_file_path(tenant_shard_id, timeline_id))
142 0 : .await
143 0 : .or_else(fs_ext::ignore_not_found)
144 0 : .context("remove delete mark")
145 0 : }
146 :
147 : /// It is important that this gets called when DeletionGuard is being held.
148 : /// For more context see comments in [`DeleteTimelineFlow::prepare`]
149 0 : async fn remove_timeline_from_tenant(
150 0 : tenant: &Tenant,
151 0 : timeline_id: TimelineId,
152 0 : _: &DeletionGuard, // using it as a witness
153 0 : ) -> anyhow::Result<()> {
154 0 : // Remove the timeline from the map.
155 0 : let mut timelines = tenant.timelines.lock().unwrap();
156 0 : let children_exist = timelines
157 0 : .iter()
158 0 : .any(|(_, entry)| entry.get_ancestor_timeline_id() == Some(timeline_id));
159 0 : // XXX this can happen because `branch_timeline` doesn't check `TimelineState::Stopping`.
160 0 : // We already deleted the layer files, so it's probably best to panic.
161 0 : // (Ideally, above remove_dir_all is atomic so we don't see this timeline after a restart)
162 0 : if children_exist {
163 0 : panic!("Timeline grew children while we removed layer files");
164 0 : }
165 0 :
166 0 : timelines
167 0 : .remove(&timeline_id)
168 0 : .expect("timeline that we were deleting was concurrently removed from 'timelines' map");
169 0 :
170 0 : drop(timelines);
171 0 :
172 0 : Ok(())
173 0 : }
174 :
175 : /// Orchestrates timeline shut down of all timeline tasks, removes its in-memory structures,
176 : /// and deletes its data from both disk and s3.
177 : /// The sequence of steps:
178 : /// 1. Set deleted_at in remote index part.
179 : /// 2. Create local mark file.
180 : /// 3. Delete local files except metadata (it is simpler this way, to be able to reuse timeline initialization code that expects metadata)
181 : /// 4. Delete remote layers
182 : /// 5. Delete index part
183 : /// 6. Delete meta, timeline directory
184 : /// 7. Delete mark file
185 : ///
186 : /// It is resumable from any step in case a crash/restart occurs.
187 : /// There are three entrypoints to the process:
188 : /// 1. [`DeleteTimelineFlow::run`] this is the main one called by a management api handler.
189 : /// 2. [`DeleteTimelineFlow::resume_deletion`] is called during restarts when local metadata is still present
190 : /// and we possibly neeed to continue deletion of remote files.
191 : /// 3. [`DeleteTimelineFlow::cleanup_remaining_timeline_fs_traces`] is used when we deleted remote
192 : /// index but still have local metadata, timeline directory and delete mark.
193 : ///
194 : /// Note the only other place that messes around timeline delete mark is the logic that scans directory with timelines during tenant load.
195 : #[derive(Default)]
196 : pub enum DeleteTimelineFlow {
197 : #[default]
198 : NotStarted,
199 : InProgress,
200 : Finished,
201 : }
202 :
203 : impl DeleteTimelineFlow {
204 : // These steps are run in the context of management api request handler.
205 : // Long running steps are continued to run in the background.
206 : // NB: If this fails half-way through, and is retried, the retry will go through
207 : // all the same steps again. Make sure the code here is idempotent, and don't
208 : // error out if some of the shutdown tasks have already been completed!
209 0 : #[instrument(skip_all, fields(%inplace))]
210 : pub async fn run(
211 : tenant: &Arc<Tenant>,
212 : timeline_id: TimelineId,
213 : inplace: bool,
214 : ) -> Result<(), DeleteTimelineError> {
215 : super::debug_assert_current_span_has_tenant_and_timeline_id();
216 :
217 : let (timeline, mut guard) = Self::prepare(tenant, timeline_id)?;
218 :
219 : guard.mark_in_progress()?;
220 :
221 : // Now that the Timeline is in Stopping state, request all the related tasks to shut down.
222 : timeline.shutdown(super::ShutdownMode::Hard).await;
223 :
224 0 : fail::fail_point!("timeline-delete-before-index-deleted-at", |_| {
225 0 : Err(anyhow::anyhow!(
226 0 : "failpoint: timeline-delete-before-index-deleted-at"
227 0 : ))?
228 0 : });
229 :
230 : set_deleted_in_remote_index(&timeline).await?;
231 :
232 0 : fail::fail_point!("timeline-delete-before-schedule", |_| {
233 0 : Err(anyhow::anyhow!(
234 0 : "failpoint: timeline-delete-before-schedule"
235 0 : ))?
236 0 : });
237 :
238 : if inplace {
239 : Self::background(guard, tenant.conf, tenant, &timeline).await?
240 : } else {
241 : Self::schedule_background(guard, tenant.conf, Arc::clone(tenant), timeline);
242 : }
243 :
244 : Ok(())
245 : }
246 :
247 0 : fn mark_in_progress(&mut self) -> anyhow::Result<()> {
248 0 : match self {
249 0 : Self::Finished => anyhow::bail!("Bug. Is in finished state"),
250 0 : Self::InProgress { .. } => { /* We're in a retry */ }
251 0 : Self::NotStarted => { /* Fresh start */ }
252 : }
253 :
254 0 : *self = Self::InProgress;
255 0 :
256 0 : Ok(())
257 0 : }
258 :
259 : /// Shortcut to create Timeline in stopping state and spawn deletion task.
260 0 : #[instrument(skip_all, fields(%timeline_id))]
261 : pub async fn resume_deletion(
262 : tenant: Arc<Tenant>,
263 : timeline_id: TimelineId,
264 : local_metadata: &TimelineMetadata,
265 : remote_client: RemoteTimelineClient,
266 : ) -> anyhow::Result<()> {
267 : // Note: here we even skip populating layer map. Timeline is essentially uninitialized.
268 : // RemoteTimelineClient is the only functioning part.
269 : let timeline = tenant
270 : .create_timeline_struct(
271 : timeline_id,
272 : local_metadata,
273 : None, // Ancestor is not needed for deletion.
274 : TimelineResources {
275 : remote_client,
276 : timeline_get_throttle: tenant.timeline_get_throttle.clone(),
277 : l0_flush_global_state: tenant.l0_flush_global_state.clone(),
278 : },
279 : // Important. We dont pass ancestor above because it can be missing.
280 : // Thus we need to skip the validation here.
281 : CreateTimelineCause::Delete,
282 : // Aux file policy is not needed for deletion, assuming deletion does not read aux keyspace
283 : None,
284 : )
285 : .context("create_timeline_struct")?;
286 :
287 : let mut guard = DeletionGuard(
288 : Arc::clone(&timeline.delete_progress)
289 : .try_lock_owned()
290 : .expect("cannot happen because we're the only owner"),
291 : );
292 :
293 : // We meed to do this because when console retries delete request we shouldnt answer with 404
294 : // because 404 means successful deletion.
295 : {
296 : let mut locked = tenant.timelines.lock().unwrap();
297 : locked.insert(timeline_id, Arc::clone(&timeline));
298 : }
299 :
300 : guard.mark_in_progress()?;
301 :
302 : Self::schedule_background(guard, tenant.conf, tenant, timeline);
303 :
304 : Ok(())
305 : }
306 :
307 0 : #[instrument(skip_all, fields(%timeline_id))]
308 : pub async fn cleanup_remaining_timeline_fs_traces(
309 : tenant: &Tenant,
310 : timeline_id: TimelineId,
311 : ) -> anyhow::Result<()> {
312 : let r =
313 : cleanup_remaining_timeline_fs_traces(tenant.conf, tenant.tenant_shard_id, timeline_id)
314 : .await;
315 : info!("Done");
316 : r
317 : }
318 :
319 0 : fn prepare(
320 0 : tenant: &Tenant,
321 0 : timeline_id: TimelineId,
322 0 : ) -> Result<(Arc<Timeline>, DeletionGuard), DeleteTimelineError> {
323 0 : // Note the interaction between this guard and deletion guard.
324 0 : // Here we attempt to lock deletion guard when we're holding a lock on timelines.
325 0 : // This is important because when you take into account `remove_timeline_from_tenant`
326 0 : // we remove timeline from memory when we still hold the deletion guard.
327 0 : // So here when timeline deletion is finished timeline wont be present in timelines map at all
328 0 : // which makes the following sequence impossible:
329 0 : // T1: get preempted right before the try_lock on `Timeline::delete_progress`
330 0 : // T2: do a full deletion, acquire and drop `Timeline::delete_progress`
331 0 : // T1: acquire deletion lock, do another `DeleteTimelineFlow::run`
332 0 : // For more context see this discussion: `https://github.com/neondatabase/neon/pull/4552#discussion_r1253437346`
333 0 : let timelines = tenant.timelines.lock().unwrap();
334 :
335 0 : let timeline = match timelines.get(&timeline_id) {
336 0 : Some(t) => t,
337 0 : None => return Err(DeleteTimelineError::NotFound),
338 : };
339 :
340 : // Ensure that there are no child timelines **attached to that pageserver**,
341 : // because detach removes files, which will break child branches
342 0 : let children: Vec<TimelineId> = timelines
343 0 : .iter()
344 0 : .filter_map(|(id, entry)| {
345 0 : if entry.get_ancestor_timeline_id() == Some(timeline_id) {
346 0 : Some(*id)
347 : } else {
348 0 : None
349 : }
350 0 : })
351 0 : .collect();
352 0 :
353 0 : if !children.is_empty() {
354 0 : return Err(DeleteTimelineError::HasChildren(children));
355 0 : }
356 0 :
357 0 : // Note that using try_lock here is important to avoid a deadlock.
358 0 : // Here we take lock on timelines and then the deletion guard.
359 0 : // At the end of the operation we're holding the guard and need to lock timelines map
360 0 : // to remove the timeline from it.
361 0 : // Always if you have two locks that are taken in different order this can result in a deadlock.
362 0 :
363 0 : let delete_progress = Arc::clone(&timeline.delete_progress);
364 0 : let delete_lock_guard = match delete_progress.try_lock_owned() {
365 0 : Ok(guard) => DeletionGuard(guard),
366 : Err(_) => {
367 : // Unfortunately if lock fails arc is consumed.
368 0 : return Err(DeleteTimelineError::AlreadyInProgress(Arc::clone(
369 0 : &timeline.delete_progress,
370 0 : )));
371 : }
372 : };
373 :
374 0 : timeline.set_state(TimelineState::Stopping);
375 0 :
376 0 : Ok((Arc::clone(timeline), delete_lock_guard))
377 0 : }
378 :
379 0 : fn schedule_background(
380 0 : guard: DeletionGuard,
381 0 : conf: &'static PageServerConf,
382 0 : tenant: Arc<Tenant>,
383 0 : timeline: Arc<Timeline>,
384 0 : ) {
385 0 : let tenant_shard_id = timeline.tenant_shard_id;
386 0 : let timeline_id = timeline.timeline_id;
387 0 :
388 0 : task_mgr::spawn(
389 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
390 0 : TaskKind::TimelineDeletionWorker,
391 0 : Some(tenant_shard_id),
392 0 : Some(timeline_id),
393 0 : "timeline_delete",
394 : false,
395 0 : async move {
396 0 : if let Err(err) = Self::background(guard, conf, &tenant, &timeline).await {
397 0 : error!("Error: {err:#}");
398 0 : timeline.set_broken(format!("{err:#}"))
399 0 : };
400 0 : Ok(())
401 0 : }
402 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)),
403 : );
404 0 : }
405 :
406 0 : async fn background(
407 0 : mut guard: DeletionGuard,
408 0 : conf: &PageServerConf,
409 0 : tenant: &Tenant,
410 0 : timeline: &Timeline,
411 0 : ) -> Result<(), DeleteTimelineError> {
412 0 : delete_local_timeline_directory(conf, tenant.tenant_shard_id, timeline).await?;
413 :
414 0 : delete_remote_layers_and_index(timeline).await?;
415 :
416 : pausable_failpoint!("in_progress_delete");
417 :
418 0 : remove_timeline_from_tenant(tenant, timeline.timeline_id, &guard).await?;
419 :
420 0 : *guard = Self::Finished;
421 0 :
422 0 : Ok(())
423 0 : }
424 :
425 0 : pub(crate) fn is_not_started(&self) -> bool {
426 0 : matches!(self, Self::NotStarted)
427 0 : }
428 : }
429 :
430 : struct DeletionGuard(OwnedMutexGuard<DeleteTimelineFlow>);
431 :
432 : impl Deref for DeletionGuard {
433 : type Target = DeleteTimelineFlow;
434 :
435 0 : fn deref(&self) -> &Self::Target {
436 0 : &self.0
437 0 : }
438 : }
439 :
440 : impl DerefMut for DeletionGuard {
441 0 : fn deref_mut(&mut self) -> &mut Self::Target {
442 0 : &mut self.0
443 0 : }
444 : }
|
