Line data Source code
1 : use std::collections::HashSet;
2 : use std::sync::Arc;
3 :
4 : use anyhow::Context;
5 : use bytes::Bytes;
6 : use http_utils::error::ApiError;
7 : use pageserver_api::key::Key;
8 : use pageserver_api::keyspace::KeySpace;
9 : use pageserver_api::models::DetachBehavior;
10 : use pageserver_api::models::detach_ancestor::AncestorDetached;
11 : use pageserver_api::shard::ShardIdentity;
12 : use pageserver_compaction::helpers::overlaps_with;
13 : use tokio::sync::Semaphore;
14 : use tokio_util::sync::CancellationToken;
15 : use tracing::Instrument;
16 : use utils::completion;
17 : use utils::generation::Generation;
18 : use utils::id::TimelineId;
19 : use utils::lsn::Lsn;
20 : use utils::sync::gate::GateError;
21 :
22 : use super::layer_manager::LayerManager;
23 : use super::{FlushLayerError, Timeline};
24 : use crate::context::{DownloadBehavior, RequestContext};
25 : use crate::task_mgr::TaskKind;
26 : use crate::tenant::Tenant;
27 : use crate::tenant::remote_timeline_client::index::GcBlockingReason::DetachAncestor;
28 : use crate::tenant::storage_layer::layer::local_layer_path;
29 : use crate::tenant::storage_layer::{
30 : AsLayerDesc as _, DeltaLayerWriter, ImageLayerWriter, IoConcurrency, Layer, ResidentLayer,
31 : ValuesReconstructState,
32 : };
33 : use crate::virtual_file::{MaybeFatalIo, VirtualFile};
34 :
35 : #[derive(Debug, thiserror::Error)]
36 : pub(crate) enum Error {
37 : #[error("no ancestors")]
38 : NoAncestor,
39 :
40 : #[error("too many ancestors")]
41 : TooManyAncestors,
42 :
43 : #[error("ancestor is not empty")]
44 : AncestorNotEmpty,
45 :
46 : #[error("shutting down, please retry later")]
47 : ShuttingDown,
48 :
49 : #[error("archived: {}", .0)]
50 : Archived(TimelineId),
51 :
52 : #[error(transparent)]
53 : NotFound(crate::tenant::GetTimelineError),
54 :
55 : #[error("failed to reparent all candidate timelines, please retry")]
56 : FailedToReparentAll,
57 :
58 : #[error("ancestor is already being detached by: {}", .0)]
59 : OtherTimelineDetachOngoing(TimelineId),
60 :
61 : #[error("preparing to timeline ancestor detach failed")]
62 : Prepare(#[source] anyhow::Error),
63 :
64 : #[error("detaching and reparenting failed")]
65 : DetachReparent(#[source] anyhow::Error),
66 :
67 : #[error("completing ancestor detach failed")]
68 : Complete(#[source] anyhow::Error),
69 :
70 : #[error("failpoint: {}", .0)]
71 : Failpoint(&'static str),
72 : }
73 :
74 : impl Error {
75 : /// Try to catch cancellation from within the `anyhow::Error`, or wrap the anyhow as the given
76 : /// variant or fancier `or_else`.
77 0 : fn launder<F>(e: anyhow::Error, or_else: F) -> Error
78 0 : where
79 0 : F: Fn(anyhow::Error) -> Error,
80 0 : {
81 : use remote_storage::TimeoutOrCancel;
82 :
83 : use crate::tenant::remote_timeline_client::WaitCompletionError;
84 : use crate::tenant::upload_queue::NotInitialized;
85 :
86 0 : if e.is::<NotInitialized>()
87 0 : || TimeoutOrCancel::caused_by_cancel(&e)
88 0 : || e.downcast_ref::<remote_storage::DownloadError>()
89 0 : .is_some_and(|e| e.is_cancelled())
90 0 : || e.is::<WaitCompletionError>()
91 : {
92 0 : Error::ShuttingDown
93 : } else {
94 0 : or_else(e)
95 : }
96 0 : }
97 : }
98 :
99 : impl From<Error> for ApiError {
100 0 : fn from(value: Error) -> Self {
101 0 : match value {
102 0 : Error::NoAncestor => ApiError::Conflict(value.to_string()),
103 : Error::TooManyAncestors | Error::AncestorNotEmpty => {
104 0 : ApiError::BadRequest(anyhow::anyhow!("{value}"))
105 : }
106 0 : Error::ShuttingDown => ApiError::ShuttingDown,
107 0 : Error::Archived(_) => ApiError::BadRequest(anyhow::anyhow!("{value}")),
108 : Error::OtherTimelineDetachOngoing(_) | Error::FailedToReparentAll => {
109 0 : ApiError::ResourceUnavailable(value.to_string().into())
110 : }
111 0 : Error::NotFound(e) => ApiError::from(e),
112 : // these variants should have no cancellation errors because of Error::launder
113 : Error::Prepare(_)
114 : | Error::DetachReparent(_)
115 : | Error::Complete(_)
116 0 : | Error::Failpoint(_) => ApiError::InternalServerError(value.into()),
117 : }
118 0 : }
119 : }
120 :
121 : impl From<crate::tenant::upload_queue::NotInitialized> for Error {
122 0 : fn from(_: crate::tenant::upload_queue::NotInitialized) -> Self {
123 0 : // treat all as shutting down signals, even though that is not entirely correct
124 0 : // (uninitialized state)
125 0 : Error::ShuttingDown
126 0 : }
127 : }
128 : impl From<super::layer_manager::Shutdown> for Error {
129 0 : fn from(_: super::layer_manager::Shutdown) -> Self {
130 0 : Error::ShuttingDown
131 0 : }
132 : }
133 :
134 : pub(crate) enum Progress {
135 : Prepared(Attempt, PreparedTimelineDetach),
136 : Done(AncestorDetached),
137 : }
138 :
139 : pub(crate) struct PreparedTimelineDetach {
140 : layers: Vec<Layer>,
141 : }
142 :
143 : // TODO: this should be part of PageserverConf because we cannot easily modify cplane arguments.
144 : #[derive(Debug)]
145 : pub(crate) struct Options {
146 : pub(crate) rewrite_concurrency: std::num::NonZeroUsize,
147 : pub(crate) copy_concurrency: std::num::NonZeroUsize,
148 : }
149 :
150 : impl Default for Options {
151 0 : fn default() -> Self {
152 0 : Self {
153 0 : rewrite_concurrency: std::num::NonZeroUsize::new(2).unwrap(),
154 0 : copy_concurrency: std::num::NonZeroUsize::new(100).unwrap(),
155 0 : }
156 0 : }
157 : }
158 :
159 : /// Represents an across tenant reset exclusive single attempt to detach ancestor.
160 : #[derive(Debug)]
161 : pub(crate) struct Attempt {
162 : pub(crate) timeline_id: TimelineId,
163 : pub(crate) ancestor_timeline_id: TimelineId,
164 : pub(crate) ancestor_lsn: Lsn,
165 : _guard: completion::Completion,
166 : gate_entered: Option<utils::sync::gate::GateGuard>,
167 : }
168 :
169 : impl Attempt {
170 0 : pub(crate) fn before_reset_tenant(&mut self) {
171 0 : let taken = self.gate_entered.take();
172 0 : assert!(taken.is_some());
173 0 : }
174 :
175 0 : pub(crate) fn new_barrier(&self) -> completion::Barrier {
176 0 : self._guard.barrier()
177 0 : }
178 : }
179 :
180 0 : async fn generate_tombstone_image_layer(
181 0 : detached: &Arc<Timeline>,
182 0 : ancestor: &Arc<Timeline>,
183 0 : ancestor_lsn: Lsn,
184 0 : ctx: &RequestContext,
185 0 : ) -> Result<Option<ResidentLayer>, Error> {
186 0 : tracing::info!(
187 0 : "removing non-inherited keys by writing an image layer with tombstones at the detach LSN"
188 : );
189 0 : let io_concurrency = IoConcurrency::spawn_from_conf(
190 0 : detached.conf,
191 0 : detached.gate.enter().map_err(|_| Error::ShuttingDown)?,
192 : );
193 0 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
194 0 : // Directly use `get_vectored_impl` to skip the max_vectored_read_key limit check. Note that the keyspace should
195 0 : // not contain too many keys, otherwise this takes a lot of memory. Currently we limit it to 10k keys in the compute.
196 0 : let key_range = Key::sparse_non_inherited_keyspace();
197 0 : // avoid generating a "future layer" which will then be removed
198 0 : let image_lsn = ancestor_lsn;
199 :
200 : {
201 0 : let layers = detached.layers.read().await;
202 0 : for layer in layers.all_persistent_layers() {
203 0 : if !layer.is_delta
204 0 : && layer.lsn_range.start == image_lsn
205 0 : && overlaps_with(&key_range, &layer.key_range)
206 : {
207 0 : tracing::warn!(
208 0 : layer=%layer, "image layer at the detach LSN already exists, skipping removing aux files"
209 : );
210 0 : return Ok(None);
211 0 : }
212 : }
213 : }
214 :
215 0 : let data = ancestor
216 0 : .get_vectored_impl(
217 0 : KeySpace::single(key_range.clone()),
218 0 : image_lsn,
219 0 : &mut reconstruct_state,
220 0 : ctx,
221 0 : )
222 0 : .await
223 0 : .context("failed to retrieve aux keys")
224 0 : .map_err(|e| Error::launder(e, Error::Prepare))?;
225 0 : if !data.is_empty() {
226 : // TODO: is it possible that we can have an image at `image_lsn`? Unlikely because image layers are only generated
227 : // upon compaction but theoretically possible.
228 0 : let mut image_layer_writer = ImageLayerWriter::new(
229 0 : detached.conf,
230 0 : detached.timeline_id,
231 0 : detached.tenant_shard_id,
232 0 : &key_range,
233 0 : image_lsn,
234 0 : ctx,
235 0 : )
236 0 : .await
237 0 : .context("failed to create image layer writer")
238 0 : .map_err(Error::Prepare)?;
239 0 : for key in data.keys() {
240 0 : image_layer_writer
241 0 : .put_image(*key, Bytes::new(), ctx)
242 0 : .await
243 0 : .context("failed to write key")
244 0 : .map_err(|e| Error::launder(e, Error::Prepare))?;
245 : }
246 0 : let (desc, path) = image_layer_writer
247 0 : .finish(ctx)
248 0 : .await
249 0 : .context("failed to finish image layer writer for removing the metadata keys")
250 0 : .map_err(|e| Error::launder(e, Error::Prepare))?;
251 0 : let generated = Layer::finish_creating(detached.conf, detached, desc, &path)
252 0 : .map_err(|e| Error::launder(e, Error::Prepare))?;
253 0 : detached
254 0 : .remote_client
255 0 : .upload_layer_file(&generated, &detached.cancel)
256 0 : .await
257 0 : .map_err(|e| Error::launder(e, Error::Prepare))?;
258 0 : tracing::info!(layer=%generated, "wrote image layer");
259 0 : Ok(Some(generated))
260 : } else {
261 0 : tracing::info!("no aux keys found in ancestor");
262 0 : Ok(None)
263 : }
264 0 : }
265 :
266 : /// See [`Timeline::prepare_to_detach_from_ancestor`]
267 0 : pub(super) async fn prepare(
268 0 : detached: &Arc<Timeline>,
269 0 : tenant: &Tenant,
270 0 : behavior: DetachBehavior,
271 0 : options: Options,
272 0 : ctx: &RequestContext,
273 0 : ) -> Result<Progress, Error> {
274 : use Error::*;
275 :
276 0 : let Some((mut ancestor, mut ancestor_lsn)) = detached
277 0 : .ancestor_timeline
278 0 : .as_ref()
279 0 : .map(|tl| (tl.clone(), detached.ancestor_lsn))
280 : else {
281 : let ancestor_id;
282 : let ancestor_lsn;
283 0 : let still_in_progress = {
284 0 : let accessor = detached.remote_client.initialized_upload_queue()?;
285 :
286 : // we are safe to inspect the latest uploaded, because we can only witness this after
287 : // restart is complete and ancestor is no more.
288 0 : let latest = accessor.latest_uploaded_index_part();
289 0 : let Some((id, lsn)) = latest.lineage.detached_previous_ancestor() else {
290 0 : return Err(NoAncestor);
291 : };
292 0 : ancestor_id = id;
293 0 : ancestor_lsn = lsn;
294 0 :
295 0 : latest
296 0 : .gc_blocking
297 0 : .as_ref()
298 0 : .is_some_and(|b| b.blocked_by(DetachAncestor))
299 0 : };
300 0 :
301 0 : if still_in_progress {
302 : // gc is still blocked, we can still reparent and complete.
303 : // we are safe to reparent remaining, because they were locked in in the beginning.
304 0 : let attempt =
305 0 : continue_with_blocked_gc(detached, tenant, ancestor_id, ancestor_lsn).await?;
306 :
307 : // because the ancestor of detached is already set to none, we have published all
308 : // of the layers, so we are still "prepared."
309 0 : return Ok(Progress::Prepared(
310 0 : attempt,
311 0 : PreparedTimelineDetach { layers: Vec::new() },
312 0 : ));
313 0 : }
314 :
315 0 : let reparented_timelines = reparented_direct_children(detached, tenant)?;
316 0 : return Ok(Progress::Done(AncestorDetached {
317 0 : reparented_timelines,
318 0 : }));
319 : };
320 :
321 0 : if detached.is_archived() != Some(false) {
322 0 : return Err(Archived(detached.timeline_id));
323 0 : }
324 0 :
325 0 : if !ancestor_lsn.is_valid() {
326 : // rare case, probably wouldn't even load
327 0 : tracing::error!("ancestor is set, but ancestor_lsn is invalid, this timeline needs fixing");
328 0 : return Err(NoAncestor);
329 0 : }
330 0 :
331 0 : check_no_archived_children_of_ancestor(tenant, detached, &ancestor, ancestor_lsn, behavior)?;
332 :
333 0 : if let DetachBehavior::MultiLevelAndNoReparent = behavior {
334 : // If the ancestor has an ancestor, we might be able to fast-path detach it if the current ancestor does not have any data written/used by the detaching timeline.
335 0 : while let Some(ancestor_of_ancestor) = ancestor.ancestor_timeline.clone() {
336 0 : if ancestor_lsn != ancestor.ancestor_lsn {
337 : // non-technical requirement; we could flatten still if ancestor LSN does not match but that needs
338 : // us to copy and cut more layers.
339 0 : return Err(AncestorNotEmpty);
340 0 : }
341 0 : // Use the ancestor of the ancestor as the new ancestor (only when the ancestor LSNs are the same)
342 0 : ancestor_lsn = ancestor.ancestor_lsn; // Get the LSN first before resetting the `ancestor` variable
343 0 : ancestor = ancestor_of_ancestor;
344 0 : // TODO: do we still need to check if we don't want to reparent?
345 0 : check_no_archived_children_of_ancestor(
346 0 : tenant,
347 0 : detached,
348 0 : &ancestor,
349 0 : ancestor_lsn,
350 0 : behavior,
351 0 : )?;
352 : }
353 0 : } else if ancestor.ancestor_timeline.is_some() {
354 : // non-technical requirement; we could flatten N ancestors just as easily but we chose
355 : // not to, at least initially
356 0 : return Err(TooManyAncestors);
357 0 : }
358 :
359 0 : tracing::info!(
360 0 : "attempt to detach the timeline from the ancestor: {}@{}, behavior={:?}",
361 0 : ancestor.timeline_id,
362 : ancestor_lsn,
363 : behavior
364 : );
365 :
366 0 : let attempt = start_new_attempt(detached, tenant, ancestor.timeline_id, ancestor_lsn).await?;
367 :
368 0 : utils::pausable_failpoint!("timeline-detach-ancestor::before_starting_after_locking-pausable");
369 :
370 0 : fail::fail_point!(
371 0 : "timeline-detach-ancestor::before_starting_after_locking",
372 0 : |_| Err(Error::Failpoint(
373 0 : "timeline-detach-ancestor::before_starting_after_locking"
374 0 : ))
375 0 : );
376 :
377 0 : if ancestor_lsn >= ancestor.get_disk_consistent_lsn() {
378 0 : let span =
379 0 : tracing::info_span!("freeze_and_flush", ancestor_timeline_id=%ancestor.timeline_id);
380 0 : async {
381 0 : let started_at = std::time::Instant::now();
382 0 : let freeze_and_flush = ancestor.freeze_and_flush0();
383 0 : let mut freeze_and_flush = std::pin::pin!(freeze_and_flush);
384 :
385 0 : let res =
386 0 : tokio::time::timeout(std::time::Duration::from_secs(1), &mut freeze_and_flush)
387 0 : .await;
388 :
389 0 : let res = match res {
390 0 : Ok(res) => res,
391 0 : Err(_elapsed) => {
392 0 : tracing::info!("freezing and flushing ancestor is still ongoing");
393 0 : freeze_and_flush.await
394 : }
395 : };
396 :
397 0 : res.map_err(|e| {
398 : use FlushLayerError::*;
399 0 : match e {
400 : Cancelled | NotRunning(_) => {
401 : // FIXME(#6424): technically statically unreachable right now, given how we never
402 : // drop the sender
403 0 : Error::ShuttingDown
404 : }
405 0 : CreateImageLayersError(_) | Other(_) => Error::Prepare(e.into()),
406 : }
407 0 : })?;
408 :
409 : // we do not need to wait for uploads to complete but we do need `struct Layer`,
410 : // copying delta prefix is unsupported currently for `InMemoryLayer`.
411 0 : tracing::info!(
412 0 : elapsed_ms = started_at.elapsed().as_millis(),
413 0 : "froze and flushed the ancestor"
414 : );
415 0 : Ok::<_, Error>(())
416 0 : }
417 0 : .instrument(span)
418 0 : .await?;
419 0 : }
420 :
421 0 : let end_lsn = ancestor_lsn + 1;
422 :
423 0 : let (filtered_layers, straddling_branchpoint, rest_of_historic) = {
424 : // we do not need to start from our layers, because they can only be layers that come
425 : // *after* ancestor_lsn
426 0 : let layers = tokio::select! {
427 0 : guard = ancestor.layers.read() => guard,
428 0 : _ = detached.cancel.cancelled() => {
429 0 : return Err(ShuttingDown);
430 : }
431 0 : _ = ancestor.cancel.cancelled() => {
432 0 : return Err(ShuttingDown);
433 : }
434 : };
435 :
436 : // between retries, these can change if compaction or gc ran in between. this will mean
437 : // we have to redo work.
438 0 : partition_work(ancestor_lsn, &layers)?
439 : };
440 :
441 : // TODO: layers are already sorted by something: use that to determine how much of remote
442 : // copies are already done -- gc is blocked, but a compaction could had happened on ancestor,
443 : // which is something to keep in mind if copy skipping is implemented.
444 0 : tracing::info!(filtered=%filtered_layers, to_rewrite = straddling_branchpoint.len(), historic=%rest_of_historic.len(), "collected layers");
445 :
446 : // TODO: copying and lsn prefix copying could be done at the same time with a single fsync after
447 0 : let mut new_layers: Vec<Layer> =
448 0 : Vec::with_capacity(straddling_branchpoint.len() + rest_of_historic.len() + 1);
449 :
450 0 : if let Some(tombstone_layer) =
451 0 : generate_tombstone_image_layer(detached, &ancestor, ancestor_lsn, ctx).await?
452 0 : {
453 0 : new_layers.push(tombstone_layer.into());
454 0 : }
455 :
456 : {
457 0 : tracing::info!(to_rewrite = %straddling_branchpoint.len(), "copying prefix of delta layers");
458 :
459 0 : let mut tasks = tokio::task::JoinSet::new();
460 0 :
461 0 : let mut wrote_any = false;
462 0 :
463 0 : let limiter = Arc::new(Semaphore::new(options.rewrite_concurrency.get()));
464 :
465 0 : for layer in straddling_branchpoint {
466 0 : let limiter = limiter.clone();
467 0 : let timeline = detached.clone();
468 0 : let ctx = ctx.detached_child(TaskKind::DetachAncestor, DownloadBehavior::Download);
469 :
470 0 : let span = tracing::info_span!("upload_rewritten_layer", %layer);
471 0 : tasks.spawn(
472 0 : async move {
473 0 : let _permit = limiter.acquire().await;
474 0 : let copied =
475 0 : upload_rewritten_layer(end_lsn, &layer, &timeline, &timeline.cancel, &ctx)
476 0 : .await?;
477 0 : if let Some(copied) = copied.as_ref() {
478 0 : tracing::info!(%copied, "rewrote and uploaded");
479 0 : }
480 0 : Ok(copied)
481 0 : }
482 0 : .instrument(span),
483 0 : );
484 0 : }
485 :
486 0 : while let Some(res) = tasks.join_next().await {
487 0 : match res {
488 0 : Ok(Ok(Some(copied))) => {
489 0 : wrote_any = true;
490 0 : new_layers.push(copied);
491 0 : }
492 0 : Ok(Ok(None)) => {}
493 0 : Ok(Err(e)) => return Err(e),
494 0 : Err(je) => return Err(Error::Prepare(je.into())),
495 : }
496 : }
497 :
498 : // FIXME: the fsync should be mandatory, after both rewrites and copies
499 0 : if wrote_any {
500 0 : fsync_timeline_dir(detached, ctx).await;
501 0 : }
502 : }
503 :
504 0 : let mut tasks = tokio::task::JoinSet::new();
505 0 : let limiter = Arc::new(Semaphore::new(options.copy_concurrency.get()));
506 0 : let cancel_eval = CancellationToken::new();
507 :
508 0 : for adopted in rest_of_historic {
509 0 : let limiter = limiter.clone();
510 0 : let timeline = detached.clone();
511 0 : let cancel_eval = cancel_eval.clone();
512 0 :
513 0 : tasks.spawn(
514 0 : async move {
515 0 : let _permit = tokio::select! {
516 0 : permit = limiter.acquire() => {
517 0 : permit
518 : }
519 : // Wait for the cancellation here instead of letting the entire task be cancelled.
520 : // Cancellations are racy in that they might leave layers on disk.
521 0 : _ = cancel_eval.cancelled() => {
522 0 : Err(Error::ShuttingDown)?
523 : }
524 : };
525 0 : let (owned, did_hardlink) = remote_copy(
526 0 : &adopted,
527 0 : &timeline,
528 0 : timeline.generation,
529 0 : timeline.shard_identity,
530 0 : &timeline.cancel,
531 0 : )
532 0 : .await?;
533 0 : tracing::info!(layer=%owned, did_hard_link=%did_hardlink, "remote copied");
534 0 : Ok((owned, did_hardlink))
535 0 : }
536 0 : .in_current_span(),
537 0 : );
538 0 : }
539 :
540 0 : fn delete_layers(timeline: &Timeline, layers: Vec<Layer>) -> Result<(), Error> {
541 : // We are deleting layers, so we must hold the gate
542 0 : let _gate = timeline.gate.enter().map_err(|e| match e {
543 0 : GateError::GateClosed => Error::ShuttingDown,
544 0 : })?;
545 0 : {
546 0 : layers.into_iter().for_each(|l: Layer| {
547 0 : l.delete_on_drop();
548 0 : std::mem::drop(l);
549 0 : });
550 0 : }
551 0 : Ok(())
552 0 : }
553 :
554 0 : let mut should_fsync = false;
555 0 : let mut first_err = None;
556 0 : while let Some(res) = tasks.join_next().await {
557 0 : match res {
558 0 : Ok(Ok((owned, did_hardlink))) => {
559 0 : if did_hardlink {
560 0 : should_fsync = true;
561 0 : }
562 0 : new_layers.push(owned);
563 : }
564 :
565 : // Don't stop the evaluation on errors, so that we get the full set of hardlinked layers to delete.
566 0 : Ok(Err(failed)) => {
567 0 : cancel_eval.cancel();
568 0 : first_err.get_or_insert(failed);
569 0 : }
570 0 : Err(je) => {
571 0 : cancel_eval.cancel();
572 0 : first_err.get_or_insert(Error::Prepare(je.into()));
573 0 : }
574 : }
575 : }
576 :
577 0 : if let Some(failed) = first_err {
578 0 : delete_layers(detached, new_layers)?;
579 0 : return Err(failed);
580 0 : }
581 0 :
582 0 : // fsync directory again if we hardlinked something
583 0 : if should_fsync {
584 0 : fsync_timeline_dir(detached, ctx).await;
585 0 : }
586 :
587 0 : let prepared = PreparedTimelineDetach { layers: new_layers };
588 0 :
589 0 : Ok(Progress::Prepared(attempt, prepared))
590 0 : }
591 :
592 0 : async fn start_new_attempt(
593 0 : detached: &Timeline,
594 0 : tenant: &Tenant,
595 0 : ancestor_timeline_id: TimelineId,
596 0 : ancestor_lsn: Lsn,
597 0 : ) -> Result<Attempt, Error> {
598 0 : let attempt = obtain_exclusive_attempt(detached, tenant, ancestor_timeline_id, ancestor_lsn)?;
599 :
600 : // insert the block in the index_part.json, if not already there.
601 0 : let _dont_care = tenant
602 0 : .gc_block
603 0 : .insert(
604 0 : detached,
605 0 : crate::tenant::remote_timeline_client::index::GcBlockingReason::DetachAncestor,
606 0 : )
607 0 : .await
608 0 : .map_err(|e| Error::launder(e, Error::Prepare))?;
609 :
610 0 : Ok(attempt)
611 0 : }
612 :
613 0 : async fn continue_with_blocked_gc(
614 0 : detached: &Timeline,
615 0 : tenant: &Tenant,
616 0 : ancestor_timeline_id: TimelineId,
617 0 : ancestor_lsn: Lsn,
618 0 : ) -> Result<Attempt, Error> {
619 0 : // FIXME: it would be nice to confirm that there is an in-memory version, since we've just
620 0 : // verified there is a persistent one?
621 0 : obtain_exclusive_attempt(detached, tenant, ancestor_timeline_id, ancestor_lsn)
622 0 : }
623 :
624 0 : fn obtain_exclusive_attempt(
625 0 : detached: &Timeline,
626 0 : tenant: &Tenant,
627 0 : ancestor_timeline_id: TimelineId,
628 0 : ancestor_lsn: Lsn,
629 0 : ) -> Result<Attempt, Error> {
630 : use Error::{OtherTimelineDetachOngoing, ShuttingDown};
631 :
632 : // ensure we are the only active attempt for this tenant
633 0 : let (guard, barrier) = completion::channel();
634 0 : {
635 0 : let mut guard = tenant.ongoing_timeline_detach.lock().unwrap();
636 0 : if let Some((tl, other)) = guard.as_ref() {
637 0 : if !other.is_ready() {
638 0 : return Err(OtherTimelineDetachOngoing(*tl));
639 0 : }
640 : // FIXME: no test enters here
641 0 : }
642 0 : *guard = Some((detached.timeline_id, barrier));
643 : }
644 :
645 : // ensure the gate is still open
646 0 : let _gate_entered = detached.gate.enter().map_err(|_| ShuttingDown)?;
647 :
648 0 : Ok(Attempt {
649 0 : timeline_id: detached.timeline_id,
650 0 : ancestor_timeline_id,
651 0 : ancestor_lsn,
652 0 : _guard: guard,
653 0 : gate_entered: Some(_gate_entered),
654 0 : })
655 0 : }
656 :
657 0 : fn reparented_direct_children(
658 0 : detached: &Arc<Timeline>,
659 0 : tenant: &Tenant,
660 0 : ) -> Result<HashSet<TimelineId>, Error> {
661 0 : let mut all_direct_children = tenant
662 0 : .timelines
663 0 : .lock()
664 0 : .unwrap()
665 0 : .values()
666 0 : .filter_map(|tl| {
667 0 : let is_direct_child = matches!(tl.ancestor_timeline.as_ref(), Some(ancestor) if Arc::ptr_eq(ancestor, detached));
668 :
669 0 : if is_direct_child {
670 0 : Some(tl.clone())
671 : } else {
672 0 : if let Some(timeline) = tl.ancestor_timeline.as_ref() {
673 0 : assert_ne!(timeline.timeline_id, detached.timeline_id, "we cannot have two timelines with the same timeline_id live");
674 0 : }
675 0 : None
676 : }
677 0 : })
678 0 : // Collect to avoid lock taking order problem with Tenant::timelines and
679 0 : // Timeline::remote_client
680 0 : .collect::<Vec<_>>();
681 0 :
682 0 : let mut any_shutdown = false;
683 0 :
684 0 : all_direct_children.retain(|tl| match tl.remote_client.initialized_upload_queue() {
685 0 : Ok(accessor) => accessor
686 0 : .latest_uploaded_index_part()
687 0 : .lineage
688 0 : .is_reparented(),
689 0 : Err(_shutdownalike) => {
690 0 : // not 100% a shutdown, but let's bail early not to give inconsistent results in
691 0 : // sharded enviroment.
692 0 : any_shutdown = true;
693 0 : true
694 : }
695 0 : });
696 0 :
697 0 : if any_shutdown {
698 : // it could be one or many being deleted; have client retry
699 0 : return Err(Error::ShuttingDown);
700 0 : }
701 0 :
702 0 : Ok(all_direct_children
703 0 : .into_iter()
704 0 : .map(|tl| tl.timeline_id)
705 0 : .collect())
706 0 : }
707 :
708 0 : fn partition_work(
709 0 : ancestor_lsn: Lsn,
710 0 : source: &LayerManager,
711 0 : ) -> Result<(usize, Vec<Layer>, Vec<Layer>), Error> {
712 0 : let mut straddling_branchpoint = vec![];
713 0 : let mut rest_of_historic = vec![];
714 0 :
715 0 : let mut later_by_lsn = 0;
716 :
717 0 : for desc in source.layer_map()?.iter_historic_layers() {
718 : // off by one chances here:
719 : // - start is inclusive
720 : // - end is exclusive
721 0 : if desc.lsn_range.start > ancestor_lsn {
722 0 : later_by_lsn += 1;
723 0 : continue;
724 0 : }
725 :
726 0 : let target = if desc.lsn_range.start <= ancestor_lsn
727 0 : && desc.lsn_range.end > ancestor_lsn
728 0 : && desc.is_delta
729 : {
730 : // TODO: image layer at Lsn optimization
731 0 : &mut straddling_branchpoint
732 : } else {
733 0 : &mut rest_of_historic
734 : };
735 :
736 0 : target.push(source.get_from_desc(&desc));
737 : }
738 :
739 0 : Ok((later_by_lsn, straddling_branchpoint, rest_of_historic))
740 0 : }
741 :
742 0 : async fn upload_rewritten_layer(
743 0 : end_lsn: Lsn,
744 0 : layer: &Layer,
745 0 : target: &Arc<Timeline>,
746 0 : cancel: &CancellationToken,
747 0 : ctx: &RequestContext,
748 0 : ) -> Result<Option<Layer>, Error> {
749 0 : let copied = copy_lsn_prefix(end_lsn, layer, target, ctx).await?;
750 :
751 0 : let Some(copied) = copied else {
752 0 : return Ok(None);
753 : };
754 :
755 0 : target
756 0 : .remote_client
757 0 : .upload_layer_file(&copied, cancel)
758 0 : .await
759 0 : .map_err(|e| Error::launder(e, Error::Prepare))?;
760 :
761 0 : Ok(Some(copied.into()))
762 0 : }
763 :
764 0 : async fn copy_lsn_prefix(
765 0 : end_lsn: Lsn,
766 0 : layer: &Layer,
767 0 : target_timeline: &Arc<Timeline>,
768 0 : ctx: &RequestContext,
769 0 : ) -> Result<Option<ResidentLayer>, Error> {
770 0 : if target_timeline.cancel.is_cancelled() {
771 0 : return Err(Error::ShuttingDown);
772 0 : }
773 0 :
774 0 : tracing::debug!(%layer, %end_lsn, "copying lsn prefix");
775 :
776 0 : let mut writer = DeltaLayerWriter::new(
777 0 : target_timeline.conf,
778 0 : target_timeline.timeline_id,
779 0 : target_timeline.tenant_shard_id,
780 0 : layer.layer_desc().key_range.start,
781 0 : layer.layer_desc().lsn_range.start..end_lsn,
782 0 : ctx,
783 0 : )
784 0 : .await
785 0 : .with_context(|| format!("prepare to copy lsn prefix of ancestors {layer}"))
786 0 : .map_err(Error::Prepare)?;
787 :
788 0 : let resident = layer.download_and_keep_resident(ctx).await.map_err(|e| {
789 0 : if e.is_cancelled() {
790 0 : Error::ShuttingDown
791 : } else {
792 0 : Error::Prepare(e.into())
793 : }
794 0 : })?;
795 :
796 0 : let records = resident
797 0 : .copy_delta_prefix(&mut writer, end_lsn, ctx)
798 0 : .await
799 0 : .with_context(|| format!("copy lsn prefix of ancestors {layer}"))
800 0 : .map_err(Error::Prepare)?;
801 :
802 0 : drop(resident);
803 0 :
804 0 : tracing::debug!(%layer, records, "copied records");
805 :
806 0 : if records == 0 {
807 0 : drop(writer);
808 0 : // TODO: we might want to store an empty marker in remote storage for this
809 0 : // layer so that we will not needlessly walk `layer` on repeated attempts.
810 0 : Ok(None)
811 : } else {
812 : // reuse the key instead of adding more holes between layers by using the real
813 : // highest key in the layer.
814 0 : let reused_highest_key = layer.layer_desc().key_range.end;
815 0 : let (desc, path) = writer
816 0 : .finish(reused_highest_key, ctx)
817 0 : .await
818 0 : .map_err(Error::Prepare)?;
819 0 : let copied = Layer::finish_creating(target_timeline.conf, target_timeline, desc, &path)
820 0 : .map_err(Error::Prepare)?;
821 :
822 0 : tracing::debug!(%layer, %copied, "new layer produced");
823 :
824 0 : Ok(Some(copied))
825 : }
826 0 : }
827 :
828 : /// Creates a new Layer instance for the adopted layer, and ensures it is found in the remote
829 : /// storage on successful return. without the adopted layer being added to `index_part.json`.
830 : /// Returns (Layer, did hardlink)
831 0 : async fn remote_copy(
832 0 : adopted: &Layer,
833 0 : adoptee: &Arc<Timeline>,
834 0 : generation: Generation,
835 0 : shard_identity: ShardIdentity,
836 0 : cancel: &CancellationToken,
837 0 : ) -> Result<(Layer, bool), Error> {
838 0 : let mut metadata = adopted.metadata();
839 0 : debug_assert!(metadata.generation <= generation);
840 0 : metadata.generation = generation;
841 0 : metadata.shard = shard_identity.shard_index();
842 0 :
843 0 : let conf = adoptee.conf;
844 0 : let file_name = adopted.layer_desc().layer_name();
845 :
846 : // We don't want to shut the timeline down during this operation because we do `delete_on_drop` below
847 0 : let _gate = adoptee.gate.enter().map_err(|e| match e {
848 0 : GateError::GateClosed => Error::ShuttingDown,
849 0 : })?;
850 :
851 : // depending if Layer::keep_resident, do a hardlink
852 : let did_hardlink;
853 0 : let owned = if let Some(adopted_resident) = adopted.keep_resident().await {
854 0 : let adopted_path = adopted_resident.local_path();
855 0 : let adoptee_path = local_layer_path(
856 0 : conf,
857 0 : &adoptee.tenant_shard_id,
858 0 : &adoptee.timeline_id,
859 0 : &file_name,
860 0 : &metadata.generation,
861 0 : );
862 0 :
863 0 : match std::fs::hard_link(adopted_path, &adoptee_path) {
864 0 : Ok(()) => {}
865 0 : Err(e) if e.kind() == std::io::ErrorKind::AlreadyExists => {
866 0 : // In theory we should not get into this situation as we are doing cleanups of the layer file after errors.
867 0 : // However, we don't do cleanups for errors past `prepare`, so there is the slight chance to get to this branch.
868 0 :
869 0 : // Double check that the file is orphan (probably from an earlier attempt), then delete it
870 0 : let key = file_name.clone().into();
871 0 : if adoptee.layers.read().await.contains_key(&key) {
872 : // We are supposed to filter out such cases before coming to this function
873 0 : return Err(Error::Prepare(anyhow::anyhow!(
874 0 : "layer file {file_name} already present and inside layer map"
875 0 : )));
876 0 : }
877 0 : tracing::info!("Deleting orphan layer file to make way for hard linking");
878 : // Delete orphan layer file and try again, to ensure this layer has a well understood source
879 0 : std::fs::remove_file(adopted_path)
880 0 : .map_err(|e| Error::launder(e.into(), Error::Prepare))?;
881 0 : std::fs::hard_link(adopted_path, &adoptee_path)
882 0 : .map_err(|e| Error::launder(e.into(), Error::Prepare))?;
883 : }
884 0 : Err(e) => {
885 0 : return Err(Error::launder(e.into(), Error::Prepare));
886 : }
887 : };
888 0 : did_hardlink = true;
889 0 : Layer::for_resident(conf, adoptee, adoptee_path, file_name, metadata).drop_eviction_guard()
890 : } else {
891 0 : did_hardlink = false;
892 0 : Layer::for_evicted(conf, adoptee, file_name, metadata)
893 : };
894 :
895 0 : let layer = match adoptee
896 0 : .remote_client
897 0 : .copy_timeline_layer(adopted, &owned, cancel)
898 0 : .await
899 : {
900 0 : Ok(()) => owned,
901 0 : Err(e) => {
902 0 : {
903 0 : // Clean up the layer so that on a retry we don't get errors that the file already exists
904 0 : owned.delete_on_drop();
905 0 : std::mem::drop(owned);
906 0 : }
907 0 : return Err(Error::launder(e, Error::Prepare));
908 : }
909 : };
910 :
911 0 : Ok((layer, did_hardlink))
912 0 : }
913 :
914 : pub(crate) enum DetachingAndReparenting {
915 : /// All of the following timeline ids were reparented and the timeline ancestor detach must be
916 : /// marked as completed.
917 : Reparented(HashSet<TimelineId>),
918 :
919 : /// Some of the reparentings failed. The timeline ancestor detach must **not** be marked as
920 : /// completed.
921 : ///
922 : /// Nested `must_reset_tenant` is set to true when any restart requiring changes were made.
923 : SomeReparentingFailed { must_reset_tenant: bool },
924 :
925 : /// Detaching and reparentings were completed in a previous attempt. Timeline ancestor detach
926 : /// must be marked as completed.
927 : AlreadyDone(HashSet<TimelineId>),
928 : }
929 :
930 : impl DetachingAndReparenting {
931 0 : pub(crate) fn reset_tenant_required(&self) -> bool {
932 : use DetachingAndReparenting::*;
933 0 : match self {
934 0 : Reparented(_) => true,
935 0 : SomeReparentingFailed { must_reset_tenant } => *must_reset_tenant,
936 0 : AlreadyDone(_) => false,
937 : }
938 0 : }
939 :
940 0 : pub(crate) fn completed(self) -> Option<HashSet<TimelineId>> {
941 : use DetachingAndReparenting::*;
942 0 : match self {
943 0 : Reparented(x) | AlreadyDone(x) => Some(x),
944 0 : SomeReparentingFailed { .. } => None,
945 : }
946 0 : }
947 : }
948 :
949 : /// See [`Timeline::detach_from_ancestor_and_reparent`].
950 0 : pub(super) async fn detach_and_reparent(
951 0 : detached: &Arc<Timeline>,
952 0 : tenant: &Tenant,
953 0 : prepared: PreparedTimelineDetach,
954 0 : ancestor_timeline_id: TimelineId,
955 0 : ancestor_lsn: Lsn,
956 0 : behavior: DetachBehavior,
957 0 : _ctx: &RequestContext,
958 0 : ) -> Result<DetachingAndReparenting, Error> {
959 0 : let PreparedTimelineDetach { layers } = prepared;
960 :
961 : #[derive(Debug)]
962 : enum Ancestor {
963 : NotDetached(Arc<Timeline>, Lsn),
964 : Detached(Arc<Timeline>, Lsn),
965 : }
966 :
967 0 : let (recorded_branchpoint, still_ongoing) = {
968 0 : let access = detached.remote_client.initialized_upload_queue()?;
969 0 : let latest = access.latest_uploaded_index_part();
970 0 :
971 0 : (
972 0 : latest.lineage.detached_previous_ancestor(),
973 0 : latest
974 0 : .gc_blocking
975 0 : .as_ref()
976 0 : .is_some_and(|b| b.blocked_by(DetachAncestor)),
977 0 : )
978 0 : };
979 0 : assert!(
980 0 : still_ongoing,
981 0 : "cannot (detach? reparent)? complete if the operation is not still ongoing"
982 : );
983 :
984 0 : let ancestor_to_detach = match detached.ancestor_timeline.as_ref() {
985 0 : Some(mut ancestor) => {
986 0 : while ancestor.timeline_id != ancestor_timeline_id {
987 0 : match ancestor.ancestor_timeline.as_ref() {
988 0 : Some(found) => {
989 0 : if ancestor_lsn != ancestor.ancestor_lsn {
990 0 : return Err(Error::DetachReparent(anyhow::anyhow!(
991 0 : "cannot find the ancestor timeline to detach from: wrong ancestor lsn"
992 0 : )));
993 0 : }
994 0 : ancestor = found;
995 : }
996 : None => {
997 0 : return Err(Error::DetachReparent(anyhow::anyhow!(
998 0 : "cannot find the ancestor timeline to detach from"
999 0 : )));
1000 : }
1001 : }
1002 : }
1003 0 : Some(ancestor)
1004 : }
1005 0 : None => None,
1006 : };
1007 0 : let ancestor = match (ancestor_to_detach, recorded_branchpoint) {
1008 0 : (Some(ancestor), None) => {
1009 0 : assert!(
1010 0 : !layers.is_empty(),
1011 0 : "there should always be at least one layer to inherit"
1012 : );
1013 0 : Ancestor::NotDetached(ancestor.clone(), detached.ancestor_lsn)
1014 : }
1015 : (Some(_), Some(_)) => {
1016 0 : panic!(
1017 0 : "it should be impossible to get to here without having gone through the tenant reset; if the tenant was reset, then the ancestor_timeline would be None"
1018 0 : );
1019 : }
1020 0 : (None, Some((ancestor_id, ancestor_lsn))) => {
1021 0 : // it has been either:
1022 0 : // - detached but still exists => we can try reparenting
1023 0 : // - detached and deleted
1024 0 : //
1025 0 : // either way, we must complete
1026 0 : assert!(
1027 0 : layers.is_empty(),
1028 0 : "no layers should had been copied as detach is done"
1029 : );
1030 :
1031 0 : let existing = tenant.timelines.lock().unwrap().get(&ancestor_id).cloned();
1032 :
1033 0 : if let Some(ancestor) = existing {
1034 0 : Ancestor::Detached(ancestor, ancestor_lsn)
1035 : } else {
1036 0 : let direct_children = reparented_direct_children(detached, tenant)?;
1037 0 : return Ok(DetachingAndReparenting::AlreadyDone(direct_children));
1038 : }
1039 : }
1040 : (None, None) => {
1041 : // TODO: make sure there are no `?` before tenant_reset from after a questionmark from
1042 : // here.
1043 0 : panic!(
1044 0 : "bug: detach_and_reparent called on a timeline which has not been detached or which has no live ancestor"
1045 0 : );
1046 : }
1047 : };
1048 :
1049 : // publish the prepared layers before we reparent any of the timelines, so that on restart
1050 : // reparented timelines find layers. also do the actual detaching.
1051 : //
1052 : // if we crash after this operation, a retry will allow reparenting the remaining timelines as
1053 : // gc is blocked.
1054 :
1055 0 : let (ancestor, ancestor_lsn, was_detached) = match ancestor {
1056 0 : Ancestor::NotDetached(ancestor, ancestor_lsn) => {
1057 0 : // this has to complete before any reparentings because otherwise they would not have
1058 0 : // layers on the new parent.
1059 0 : detached
1060 0 : .remote_client
1061 0 : .schedule_adding_existing_layers_to_index_detach_and_wait(
1062 0 : &layers,
1063 0 : (ancestor.timeline_id, ancestor_lsn),
1064 0 : )
1065 0 : .await
1066 0 : .context("publish layers and detach ancestor")
1067 0 : .map_err(|e| Error::launder(e, Error::DetachReparent))?;
1068 :
1069 0 : tracing::info!(
1070 0 : ancestor=%ancestor.timeline_id,
1071 0 : %ancestor_lsn,
1072 0 : inherited_layers=%layers.len(),
1073 0 : "detached from ancestor"
1074 : );
1075 0 : (ancestor, ancestor_lsn, true)
1076 : }
1077 0 : Ancestor::Detached(ancestor, ancestor_lsn) => (ancestor, ancestor_lsn, false),
1078 : };
1079 :
1080 0 : if let DetachBehavior::MultiLevelAndNoReparent = behavior {
1081 : // Do not reparent if the user requests to behave so.
1082 0 : return Ok(DetachingAndReparenting::Reparented(HashSet::new()));
1083 0 : }
1084 0 :
1085 0 : let mut tasks = tokio::task::JoinSet::new();
1086 0 :
1087 0 : // Returns a single permit semaphore which will be used to make one reparenting succeed,
1088 0 : // others will fail as if those timelines had been stopped for whatever reason.
1089 0 : #[cfg(feature = "testing")]
1090 0 : let failpoint_sem = || -> Option<Arc<Semaphore>> {
1091 0 : fail::fail_point!("timeline-detach-ancestor::allow_one_reparented", |_| Some(
1092 0 : Arc::new(Semaphore::new(1))
1093 0 : ));
1094 0 : None
1095 0 : }();
1096 0 :
1097 0 : // because we are now keeping the slot in progress, it is unlikely that there will be any
1098 0 : // timeline deletions during this time. if we raced one, then we'll just ignore it.
1099 0 : {
1100 0 : let g = tenant.timelines.lock().unwrap();
1101 0 : reparentable_timelines(g.values(), detached, &ancestor, ancestor_lsn)
1102 0 : .cloned()
1103 0 : .for_each(|timeline| {
1104 : // important in this scope: we are holding the Tenant::timelines lock
1105 0 : let span = tracing::info_span!("reparent", reparented=%timeline.timeline_id);
1106 0 : let new_parent = detached.timeline_id;
1107 0 : #[cfg(feature = "testing")]
1108 0 : let failpoint_sem = failpoint_sem.clone();
1109 0 :
1110 0 : tasks.spawn(
1111 0 : async move {
1112 0 : let res = async {
1113 : #[cfg(feature = "testing")]
1114 0 : if let Some(failpoint_sem) = failpoint_sem {
1115 0 : let _permit = failpoint_sem.acquire().await.map_err(|_| {
1116 0 : anyhow::anyhow!(
1117 0 : "failpoint: timeline-detach-ancestor::allow_one_reparented",
1118 0 : )
1119 0 : })?;
1120 0 : failpoint_sem.close();
1121 0 : }
1122 :
1123 0 : timeline
1124 0 : .remote_client
1125 0 : .schedule_reparenting_and_wait(&new_parent)
1126 0 : .await
1127 0 : }
1128 0 : .await;
1129 :
1130 0 : match res {
1131 : Ok(()) => {
1132 0 : tracing::info!("reparented");
1133 0 : Some(timeline)
1134 : }
1135 0 : Err(e) => {
1136 0 : // with the use of tenant slot, raced timeline deletion is the most
1137 0 : // likely reason.
1138 0 : tracing::warn!("reparenting failed: {e:#}");
1139 0 : None
1140 : }
1141 : }
1142 0 : }
1143 0 : .instrument(span),
1144 0 : );
1145 0 : });
1146 0 : }
1147 0 :
1148 0 : let reparenting_candidates = tasks.len();
1149 0 : let mut reparented = HashSet::with_capacity(tasks.len());
1150 :
1151 0 : while let Some(res) = tasks.join_next().await {
1152 0 : match res {
1153 0 : Ok(Some(timeline)) => {
1154 0 : assert!(
1155 0 : reparented.insert(timeline.timeline_id),
1156 0 : "duplicate reparenting? timeline_id={}",
1157 0 : timeline.timeline_id
1158 : );
1159 : }
1160 0 : Err(je) if je.is_cancelled() => unreachable!("not used"),
1161 : // just ignore failures now, we can retry
1162 0 : Ok(None) => {}
1163 0 : Err(je) if je.is_panic() => {}
1164 0 : Err(je) => tracing::error!("unexpected join error: {je:?}"),
1165 : }
1166 : }
1167 :
1168 0 : let reparented_all = reparenting_candidates == reparented.len();
1169 0 :
1170 0 : if reparented_all {
1171 0 : Ok(DetachingAndReparenting::Reparented(reparented))
1172 : } else {
1173 0 : tracing::info!(
1174 0 : reparented = reparented.len(),
1175 0 : candidates = reparenting_candidates,
1176 0 : "failed to reparent all candidates; they can be retried after the tenant_reset",
1177 : );
1178 :
1179 0 : let must_reset_tenant = !reparented.is_empty() || was_detached;
1180 0 : Ok(DetachingAndReparenting::SomeReparentingFailed { must_reset_tenant })
1181 : }
1182 0 : }
1183 :
1184 0 : pub(super) async fn complete(
1185 0 : detached: &Arc<Timeline>,
1186 0 : tenant: &Tenant,
1187 0 : mut attempt: Attempt,
1188 0 : _ctx: &RequestContext,
1189 0 : ) -> Result<(), Error> {
1190 0 : assert_eq!(detached.timeline_id, attempt.timeline_id);
1191 :
1192 0 : if attempt.gate_entered.is_none() {
1193 0 : let entered = detached.gate.enter().map_err(|_| Error::ShuttingDown)?;
1194 0 : attempt.gate_entered = Some(entered);
1195 0 : } else {
1196 0 : // Some(gate_entered) means the tenant was not restarted, as is not required
1197 0 : }
1198 :
1199 0 : assert!(detached.ancestor_timeline.is_none());
1200 :
1201 : // this should be an 503 at least...?
1202 0 : fail::fail_point!(
1203 0 : "timeline-detach-ancestor::complete_before_uploading",
1204 0 : |_| Err(Error::Failpoint(
1205 0 : "timeline-detach-ancestor::complete_before_uploading"
1206 0 : ))
1207 0 : );
1208 :
1209 0 : tenant
1210 0 : .gc_block
1211 0 : .remove(
1212 0 : detached,
1213 0 : crate::tenant::remote_timeline_client::index::GcBlockingReason::DetachAncestor,
1214 0 : )
1215 0 : .await
1216 0 : .map_err(|e| Error::launder(e, Error::Complete))?;
1217 :
1218 0 : Ok(())
1219 0 : }
1220 :
1221 : /// Query against a locked `Tenant::timelines`.
1222 : ///
1223 : /// A timeline is reparentable if:
1224 : ///
1225 : /// - It is not the timeline being detached.
1226 : /// - It has the same ancestor as the timeline being detached. Note that the ancestor might not be the direct ancestor.
1227 0 : fn reparentable_timelines<'a, I>(
1228 0 : timelines: I,
1229 0 : detached: &'a Arc<Timeline>,
1230 0 : ancestor: &'a Arc<Timeline>,
1231 0 : ancestor_lsn: Lsn,
1232 0 : ) -> impl Iterator<Item = &'a Arc<Timeline>> + 'a
1233 0 : where
1234 0 : I: Iterator<Item = &'a Arc<Timeline>> + 'a,
1235 0 : {
1236 0 : timelines.filter_map(move |tl| {
1237 0 : if Arc::ptr_eq(tl, detached) {
1238 0 : return None;
1239 0 : }
1240 :
1241 0 : let tl_ancestor = tl.ancestor_timeline.as_ref()?;
1242 0 : let is_same = Arc::ptr_eq(ancestor, tl_ancestor);
1243 0 : let is_earlier = tl.get_ancestor_lsn() <= ancestor_lsn;
1244 0 :
1245 0 : let is_deleting = tl
1246 0 : .delete_progress
1247 0 : .try_lock()
1248 0 : .map(|flow| !flow.is_not_started())
1249 0 : .unwrap_or(true);
1250 0 :
1251 0 : if is_same && is_earlier && !is_deleting {
1252 0 : Some(tl)
1253 : } else {
1254 0 : None
1255 : }
1256 0 : })
1257 0 : }
1258 :
1259 0 : fn check_no_archived_children_of_ancestor(
1260 0 : tenant: &Tenant,
1261 0 : detached: &Arc<Timeline>,
1262 0 : ancestor: &Arc<Timeline>,
1263 0 : ancestor_lsn: Lsn,
1264 0 : detach_behavior: DetachBehavior,
1265 0 : ) -> Result<(), Error> {
1266 0 : match detach_behavior {
1267 : DetachBehavior::NoAncestorAndReparent => {
1268 0 : let timelines = tenant.timelines.lock().unwrap();
1269 0 : let timelines_offloaded = tenant.timelines_offloaded.lock().unwrap();
1270 :
1271 0 : for timeline in
1272 0 : reparentable_timelines(timelines.values(), detached, ancestor, ancestor_lsn)
1273 : {
1274 0 : if timeline.is_archived() == Some(true) {
1275 0 : return Err(Error::Archived(timeline.timeline_id));
1276 0 : }
1277 : }
1278 :
1279 0 : for timeline_offloaded in timelines_offloaded.values() {
1280 0 : if timeline_offloaded.ancestor_timeline_id != Some(ancestor.timeline_id) {
1281 0 : continue;
1282 0 : }
1283 : // This forbids the detach ancestor feature if flattened timelines are present,
1284 : // even if the ancestor_lsn is from after the branchpoint of the detached timeline.
1285 : // But as per current design, we don't record the ancestor_lsn of flattened timelines.
1286 : // This is a bit unfortunate, but as of writing this we don't support flattening
1287 : // anyway. Maybe we can evolve the data model in the future.
1288 0 : if let Some(retain_lsn) = timeline_offloaded.ancestor_retain_lsn {
1289 0 : let is_earlier = retain_lsn <= ancestor_lsn;
1290 0 : if !is_earlier {
1291 0 : continue;
1292 0 : }
1293 0 : }
1294 0 : return Err(Error::Archived(timeline_offloaded.timeline_id));
1295 : }
1296 : }
1297 0 : DetachBehavior::MultiLevelAndNoReparent => {
1298 0 : // We don't need to check anything if the user requested to not reparent.
1299 0 : }
1300 : }
1301 :
1302 0 : Ok(())
1303 0 : }
1304 :
1305 0 : async fn fsync_timeline_dir(timeline: &Timeline, ctx: &RequestContext) {
1306 0 : let path = &timeline
1307 0 : .conf
1308 0 : .timeline_path(&timeline.tenant_shard_id, &timeline.timeline_id);
1309 0 : let timeline_dir = VirtualFile::open(&path, ctx)
1310 0 : .await
1311 0 : .fatal_err("VirtualFile::open for timeline dir fsync");
1312 0 : timeline_dir
1313 0 : .sync_all()
1314 0 : .await
1315 0 : .fatal_err("VirtualFile::sync_all timeline dir");
1316 0 : }
|
