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::TenantShard;
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::tenant::timeline::VersionedKeySpaceQuery;
34 : use crate::virtual_file::{MaybeFatalIo, VirtualFile};
35 :
36 : #[derive(Debug, thiserror::Error)]
37 : pub(crate) enum Error {
38 : #[error("no ancestors")]
39 : NoAncestor,
40 :
41 : #[error("too many ancestors")]
42 : TooManyAncestors,
43 :
44 : #[error("ancestor is not empty")]
45 : AncestorNotEmpty,
46 :
47 : #[error("shutting down, please retry later")]
48 : ShuttingDown,
49 :
50 : #[error("archived: {}", .0)]
51 : Archived(TimelineId),
52 :
53 : #[error(transparent)]
54 : NotFound(crate::tenant::GetTimelineError),
55 :
56 : #[error("failed to reparent all candidate timelines, please retry")]
57 : FailedToReparentAll,
58 :
59 : #[error("ancestor is already being detached by: {}", .0)]
60 : OtherTimelineDetachOngoing(TimelineId),
61 :
62 : #[error("preparing to timeline ancestor detach failed")]
63 : Prepare(#[source] anyhow::Error),
64 :
65 : #[error("detaching and reparenting failed")]
66 : DetachReparent(#[source] anyhow::Error),
67 :
68 : #[error("completing ancestor detach failed")]
69 : Complete(#[source] anyhow::Error),
70 :
71 : #[error("failpoint: {}", .0)]
72 : Failpoint(&'static str),
73 : }
74 :
75 : impl Error {
76 : /// Try to catch cancellation from within the `anyhow::Error`, or wrap the anyhow as the given
77 : /// variant or fancier `or_else`.
78 0 : fn launder<F>(e: anyhow::Error, or_else: F) -> Error
79 0 : where
80 0 : F: Fn(anyhow::Error) -> Error,
81 0 : {
82 : use remote_storage::TimeoutOrCancel;
83 :
84 : use crate::tenant::remote_timeline_client::WaitCompletionError;
85 : use crate::tenant::upload_queue::NotInitialized;
86 :
87 0 : if e.is::<NotInitialized>()
88 0 : || TimeoutOrCancel::caused_by_cancel(&e)
89 0 : || e.downcast_ref::<remote_storage::DownloadError>()
90 0 : .is_some_and(|e| e.is_cancelled())
91 0 : || e.is::<WaitCompletionError>()
92 : {
93 0 : Error::ShuttingDown
94 : } else {
95 0 : or_else(e)
96 : }
97 0 : }
98 : }
99 :
100 : impl From<Error> for ApiError {
101 0 : fn from(value: Error) -> Self {
102 0 : match value {
103 0 : Error::NoAncestor => ApiError::Conflict(value.to_string()),
104 : Error::TooManyAncestors | Error::AncestorNotEmpty => {
105 0 : ApiError::BadRequest(anyhow::anyhow!("{value}"))
106 : }
107 0 : Error::ShuttingDown => ApiError::ShuttingDown,
108 0 : Error::Archived(_) => ApiError::BadRequest(anyhow::anyhow!("{value}")),
109 : Error::OtherTimelineDetachOngoing(_) | Error::FailedToReparentAll => {
110 0 : ApiError::ResourceUnavailable(value.to_string().into())
111 : }
112 0 : Error::NotFound(e) => ApiError::from(e),
113 : // these variants should have no cancellation errors because of Error::launder
114 : Error::Prepare(_)
115 : | Error::DetachReparent(_)
116 : | Error::Complete(_)
117 0 : | Error::Failpoint(_) => ApiError::InternalServerError(value.into()),
118 : }
119 0 : }
120 : }
121 :
122 : impl From<crate::tenant::upload_queue::NotInitialized> for Error {
123 0 : fn from(_: crate::tenant::upload_queue::NotInitialized) -> Self {
124 0 : // treat all as shutting down signals, even though that is not entirely correct
125 0 : // (uninitialized state)
126 0 : Error::ShuttingDown
127 0 : }
128 : }
129 : impl From<super::layer_manager::Shutdown> for Error {
130 0 : fn from(_: super::layer_manager::Shutdown) -> Self {
131 0 : Error::ShuttingDown
132 0 : }
133 : }
134 :
135 : pub(crate) enum Progress {
136 : Prepared(Attempt, PreparedTimelineDetach),
137 : Done(AncestorDetached),
138 : }
139 :
140 : pub(crate) struct PreparedTimelineDetach {
141 : layers: Vec<Layer>,
142 : }
143 :
144 : // TODO: this should be part of PageserverConf because we cannot easily modify cplane arguments.
145 : #[derive(Debug)]
146 : pub(crate) struct Options {
147 : pub(crate) rewrite_concurrency: std::num::NonZeroUsize,
148 : pub(crate) copy_concurrency: std::num::NonZeroUsize,
149 : }
150 :
151 : impl Default for Options {
152 0 : fn default() -> Self {
153 0 : Self {
154 0 : rewrite_concurrency: std::num::NonZeroUsize::new(2).unwrap(),
155 0 : copy_concurrency: std::num::NonZeroUsize::new(100).unwrap(),
156 0 : }
157 0 : }
158 : }
159 :
160 : /// Represents an across tenant reset exclusive single attempt to detach ancestor.
161 : #[derive(Debug)]
162 : pub(crate) struct Attempt {
163 : pub(crate) timeline_id: TimelineId,
164 : pub(crate) ancestor_timeline_id: TimelineId,
165 : pub(crate) ancestor_lsn: Lsn,
166 : _guard: completion::Completion,
167 : gate_entered: Option<utils::sync::gate::GateGuard>,
168 : }
169 :
170 : impl Attempt {
171 0 : pub(crate) fn before_reset_tenant(&mut self) {
172 0 : let taken = self.gate_entered.take();
173 0 : assert!(taken.is_some());
174 0 : }
175 :
176 0 : pub(crate) fn new_barrier(&self) -> completion::Barrier {
177 0 : self._guard.barrier()
178 0 : }
179 : }
180 :
181 0 : async fn generate_tombstone_image_layer(
182 0 : detached: &Arc<Timeline>,
183 0 : ancestor: &Arc<Timeline>,
184 0 : ancestor_lsn: Lsn,
185 0 : ctx: &RequestContext,
186 0 : ) -> Result<Option<ResidentLayer>, Error> {
187 0 : tracing::info!(
188 0 : "removing non-inherited keys by writing an image layer with tombstones at the detach LSN"
189 : );
190 0 : let io_concurrency = IoConcurrency::spawn_from_conf(
191 0 : detached.conf,
192 0 : detached.gate.enter().map_err(|_| Error::ShuttingDown)?,
193 : );
194 0 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
195 0 : // Directly use `get_vectored_impl` to skip the max_vectored_read_key limit check. Note that the keyspace should
196 0 : // not contain too many keys, otherwise this takes a lot of memory. Currently we limit it to 10k keys in the compute.
197 0 : let key_range = Key::sparse_non_inherited_keyspace();
198 0 : // avoid generating a "future layer" which will then be removed
199 0 : let image_lsn = ancestor_lsn;
200 :
201 : {
202 0 : let layers = detached.layers.read().await;
203 0 : for layer in layers.all_persistent_layers() {
204 0 : if !layer.is_delta
205 0 : && layer.lsn_range.start == image_lsn
206 0 : && overlaps_with(&key_range, &layer.key_range)
207 : {
208 0 : tracing::warn!(
209 0 : layer=%layer, "image layer at the detach LSN already exists, skipping removing aux files"
210 : );
211 0 : return Ok(None);
212 0 : }
213 : }
214 : }
215 :
216 0 : let query = VersionedKeySpaceQuery::uniform(KeySpace::single(key_range.clone()), image_lsn);
217 0 : let data = ancestor
218 0 : .get_vectored_impl(query, &mut reconstruct_state, ctx)
219 0 : .await
220 0 : .context("failed to retrieve aux keys")
221 0 : .map_err(|e| Error::launder(e, Error::Prepare))?;
222 0 : if !data.is_empty() {
223 : // TODO: is it possible that we can have an image at `image_lsn`? Unlikely because image layers are only generated
224 : // upon compaction but theoretically possible.
225 0 : let mut image_layer_writer = ImageLayerWriter::new(
226 0 : detached.conf,
227 0 : detached.timeline_id,
228 0 : detached.tenant_shard_id,
229 0 : &key_range,
230 0 : image_lsn,
231 0 : &detached.gate,
232 0 : detached.cancel.clone(),
233 0 : ctx,
234 0 : )
235 0 : .await
236 0 : .context("failed to create image layer writer")
237 0 : .map_err(Error::Prepare)?;
238 0 : for key in data.keys() {
239 0 : image_layer_writer
240 0 : .put_image(*key, Bytes::new(), ctx)
241 0 : .await
242 0 : .context("failed to write key")
243 0 : .map_err(|e| Error::launder(e, Error::Prepare))?;
244 : }
245 0 : let (desc, path) = image_layer_writer
246 0 : .finish(ctx)
247 0 : .await
248 0 : .context("failed to finish image layer writer for removing the metadata keys")
249 0 : .map_err(|e| Error::launder(e, Error::Prepare))?;
250 0 : let generated = Layer::finish_creating(detached.conf, detached, desc, &path)
251 0 : .map_err(|e| Error::launder(e, Error::Prepare))?;
252 0 : detached
253 0 : .remote_client
254 0 : .upload_layer_file(&generated, &detached.cancel)
255 0 : .await
256 0 : .map_err(|e| Error::launder(e, Error::Prepare))?;
257 0 : tracing::info!(layer=%generated, "wrote image layer");
258 0 : Ok(Some(generated))
259 : } else {
260 0 : tracing::info!("no aux keys found in ancestor");
261 0 : Ok(None)
262 : }
263 0 : }
264 :
265 : /// See [`Timeline::prepare_to_detach_from_ancestor`]
266 0 : pub(super) async fn prepare(
267 0 : detached: &Arc<Timeline>,
268 0 : tenant: &TenantShard,
269 0 : behavior: DetachBehavior,
270 0 : options: Options,
271 0 : ctx: &RequestContext,
272 0 : ) -> Result<Progress, Error> {
273 : use Error::*;
274 :
275 0 : let Some((mut ancestor, mut ancestor_lsn)) = detached
276 0 : .ancestor_timeline
277 0 : .as_ref()
278 0 : .map(|tl| (tl.clone(), detached.ancestor_lsn))
279 : else {
280 : let ancestor_id;
281 : let ancestor_lsn;
282 0 : let still_in_progress = {
283 0 : let accessor = detached.remote_client.initialized_upload_queue()?;
284 :
285 : // we are safe to inspect the latest uploaded, because we can only witness this after
286 : // restart is complete and ancestor is no more.
287 0 : let latest = accessor.latest_uploaded_index_part();
288 0 : let Some((id, lsn)) = latest.lineage.detached_previous_ancestor() else {
289 0 : return Err(NoAncestor);
290 : };
291 0 : ancestor_id = id;
292 0 : ancestor_lsn = lsn;
293 0 :
294 0 : latest
295 0 : .gc_blocking
296 0 : .as_ref()
297 0 : .is_some_and(|b| b.blocked_by(DetachAncestor))
298 0 : };
299 0 :
300 0 : if still_in_progress {
301 : // gc is still blocked, we can still reparent and complete.
302 : // we are safe to reparent remaining, because they were locked in in the beginning.
303 0 : let attempt =
304 0 : continue_with_blocked_gc(detached, tenant, ancestor_id, ancestor_lsn).await?;
305 :
306 : // because the ancestor of detached is already set to none, we have published all
307 : // of the layers, so we are still "prepared."
308 0 : return Ok(Progress::Prepared(
309 0 : attempt,
310 0 : PreparedTimelineDetach { layers: Vec::new() },
311 0 : ));
312 0 : }
313 :
314 0 : let reparented_timelines = reparented_direct_children(detached, tenant)?;
315 0 : return Ok(Progress::Done(AncestorDetached {
316 0 : reparented_timelines,
317 0 : }));
318 : };
319 :
320 0 : if detached.is_archived() != Some(false) {
321 0 : return Err(Archived(detached.timeline_id));
322 0 : }
323 0 :
324 0 : if !ancestor_lsn.is_valid() {
325 : // rare case, probably wouldn't even load
326 0 : tracing::error!("ancestor is set, but ancestor_lsn is invalid, this timeline needs fixing");
327 0 : return Err(NoAncestor);
328 0 : }
329 0 :
330 0 : check_no_archived_children_of_ancestor(tenant, detached, &ancestor, ancestor_lsn, behavior)?;
331 :
332 0 : if let DetachBehavior::MultiLevelAndNoReparent = behavior {
333 : // 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.
334 0 : while let Some(ancestor_of_ancestor) = ancestor.ancestor_timeline.clone() {
335 0 : if ancestor_lsn != ancestor.ancestor_lsn {
336 : // non-technical requirement; we could flatten still if ancestor LSN does not match but that needs
337 : // us to copy and cut more layers.
338 0 : return Err(AncestorNotEmpty);
339 0 : }
340 0 : // Use the ancestor of the ancestor as the new ancestor (only when the ancestor LSNs are the same)
341 0 : ancestor_lsn = ancestor.ancestor_lsn; // Get the LSN first before resetting the `ancestor` variable
342 0 : ancestor = ancestor_of_ancestor;
343 0 : // TODO: do we still need to check if we don't want to reparent?
344 0 : check_no_archived_children_of_ancestor(
345 0 : tenant,
346 0 : detached,
347 0 : &ancestor,
348 0 : ancestor_lsn,
349 0 : behavior,
350 0 : )?;
351 : }
352 0 : } else if ancestor.ancestor_timeline.is_some() {
353 : // non-technical requirement; we could flatten N ancestors just as easily but we chose
354 : // not to, at least initially
355 0 : return Err(TooManyAncestors);
356 0 : }
357 :
358 0 : tracing::info!(
359 0 : "attempt to detach the timeline from the ancestor: {}@{}, behavior={:?}",
360 0 : ancestor.timeline_id,
361 : ancestor_lsn,
362 : behavior
363 : );
364 :
365 0 : let attempt = start_new_attempt(detached, tenant, ancestor.timeline_id, ancestor_lsn).await?;
366 :
367 0 : utils::pausable_failpoint!("timeline-detach-ancestor::before_starting_after_locking-pausable");
368 :
369 0 : fail::fail_point!(
370 0 : "timeline-detach-ancestor::before_starting_after_locking",
371 0 : |_| Err(Error::Failpoint(
372 0 : "timeline-detach-ancestor::before_starting_after_locking"
373 0 : ))
374 0 : );
375 :
376 0 : if ancestor_lsn >= ancestor.get_disk_consistent_lsn() {
377 0 : let span =
378 0 : tracing::info_span!("freeze_and_flush", ancestor_timeline_id=%ancestor.timeline_id);
379 0 : async {
380 0 : let started_at = std::time::Instant::now();
381 0 : let freeze_and_flush = ancestor.freeze_and_flush0();
382 0 : let mut freeze_and_flush = std::pin::pin!(freeze_and_flush);
383 :
384 0 : let res =
385 0 : tokio::time::timeout(std::time::Duration::from_secs(1), &mut freeze_and_flush)
386 0 : .await;
387 :
388 0 : let res = match res {
389 0 : Ok(res) => res,
390 0 : Err(_elapsed) => {
391 0 : tracing::info!("freezing and flushing ancestor is still ongoing");
392 0 : freeze_and_flush.await
393 : }
394 : };
395 :
396 0 : res.map_err(|e| {
397 : use FlushLayerError::*;
398 0 : match e {
399 : Cancelled | NotRunning(_) => {
400 : // FIXME(#6424): technically statically unreachable right now, given how we never
401 : // drop the sender
402 0 : Error::ShuttingDown
403 : }
404 0 : CreateImageLayersError(_) | Other(_) => Error::Prepare(e.into()),
405 : }
406 0 : })?;
407 :
408 : // we do not need to wait for uploads to complete but we do need `struct Layer`,
409 : // copying delta prefix is unsupported currently for `InMemoryLayer`.
410 0 : tracing::info!(
411 0 : elapsed_ms = started_at.elapsed().as_millis(),
412 0 : "froze and flushed the ancestor"
413 : );
414 0 : Ok::<_, Error>(())
415 0 : }
416 0 : .instrument(span)
417 0 : .await?;
418 0 : }
419 :
420 0 : let end_lsn = ancestor_lsn + 1;
421 :
422 0 : let (filtered_layers, straddling_branchpoint, rest_of_historic) = {
423 : // we do not need to start from our layers, because they can only be layers that come
424 : // *after* ancestor_lsn
425 0 : let layers = tokio::select! {
426 0 : guard = ancestor.layers.read() => guard,
427 0 : _ = detached.cancel.cancelled() => {
428 0 : return Err(ShuttingDown);
429 : }
430 0 : _ = ancestor.cancel.cancelled() => {
431 0 : return Err(ShuttingDown);
432 : }
433 : };
434 :
435 : // between retries, these can change if compaction or gc ran in between. this will mean
436 : // we have to redo work.
437 0 : partition_work(ancestor_lsn, &layers)?
438 : };
439 :
440 : // TODO: layers are already sorted by something: use that to determine how much of remote
441 : // copies are already done -- gc is blocked, but a compaction could had happened on ancestor,
442 : // which is something to keep in mind if copy skipping is implemented.
443 0 : tracing::info!(filtered=%filtered_layers, to_rewrite = straddling_branchpoint.len(), historic=%rest_of_historic.len(), "collected layers");
444 :
445 : // TODO: copying and lsn prefix copying could be done at the same time with a single fsync after
446 0 : let mut new_layers: Vec<Layer> =
447 0 : Vec::with_capacity(straddling_branchpoint.len() + rest_of_historic.len() + 1);
448 :
449 0 : if let Some(tombstone_layer) =
450 0 : generate_tombstone_image_layer(detached, &ancestor, ancestor_lsn, ctx).await?
451 0 : {
452 0 : new_layers.push(tombstone_layer.into());
453 0 : }
454 :
455 : {
456 0 : tracing::info!(to_rewrite = %straddling_branchpoint.len(), "copying prefix of delta layers");
457 :
458 0 : let mut tasks = tokio::task::JoinSet::new();
459 0 :
460 0 : let mut wrote_any = false;
461 0 :
462 0 : let limiter = Arc::new(Semaphore::new(options.rewrite_concurrency.get()));
463 :
464 0 : for layer in straddling_branchpoint {
465 0 : let limiter = limiter.clone();
466 0 : let timeline = detached.clone();
467 0 : let ctx = ctx.detached_child(TaskKind::DetachAncestor, DownloadBehavior::Download);
468 :
469 0 : let span = tracing::info_span!("upload_rewritten_layer", %layer);
470 0 : tasks.spawn(
471 0 : async move {
472 0 : let _permit = limiter.acquire().await;
473 0 : let copied =
474 0 : upload_rewritten_layer(end_lsn, &layer, &timeline, &timeline.cancel, &ctx)
475 0 : .await?;
476 0 : if let Some(copied) = copied.as_ref() {
477 0 : tracing::info!(%copied, "rewrote and uploaded");
478 0 : }
479 0 : Ok(copied)
480 0 : }
481 0 : .instrument(span),
482 0 : );
483 0 : }
484 :
485 0 : while let Some(res) = tasks.join_next().await {
486 0 : match res {
487 0 : Ok(Ok(Some(copied))) => {
488 0 : wrote_any = true;
489 0 : new_layers.push(copied);
490 0 : }
491 0 : Ok(Ok(None)) => {}
492 0 : Ok(Err(e)) => return Err(e),
493 0 : Err(je) => return Err(Error::Prepare(je.into())),
494 : }
495 : }
496 :
497 : // FIXME: the fsync should be mandatory, after both rewrites and copies
498 0 : if wrote_any {
499 0 : fsync_timeline_dir(detached, ctx).await;
500 0 : }
501 : }
502 :
503 0 : let mut tasks = tokio::task::JoinSet::new();
504 0 : let limiter = Arc::new(Semaphore::new(options.copy_concurrency.get()));
505 0 : let cancel_eval = CancellationToken::new();
506 :
507 0 : for adopted in rest_of_historic {
508 0 : let limiter = limiter.clone();
509 0 : let timeline = detached.clone();
510 0 : let cancel_eval = cancel_eval.clone();
511 0 :
512 0 : tasks.spawn(
513 0 : async move {
514 0 : let _permit = tokio::select! {
515 0 : permit = limiter.acquire() => {
516 0 : permit
517 : }
518 : // Wait for the cancellation here instead of letting the entire task be cancelled.
519 : // Cancellations are racy in that they might leave layers on disk.
520 0 : _ = cancel_eval.cancelled() => {
521 0 : Err(Error::ShuttingDown)?
522 : }
523 : };
524 0 : let (owned, did_hardlink) = remote_copy(
525 0 : &adopted,
526 0 : &timeline,
527 0 : timeline.generation,
528 0 : timeline.shard_identity,
529 0 : &timeline.cancel,
530 0 : )
531 0 : .await?;
532 0 : tracing::info!(layer=%owned, did_hard_link=%did_hardlink, "remote copied");
533 0 : Ok((owned, did_hardlink))
534 0 : }
535 0 : .in_current_span(),
536 0 : );
537 0 : }
538 :
539 0 : fn delete_layers(timeline: &Timeline, layers: Vec<Layer>) -> Result<(), Error> {
540 : // We are deleting layers, so we must hold the gate
541 0 : let _gate = timeline.gate.enter().map_err(|e| match e {
542 0 : GateError::GateClosed => Error::ShuttingDown,
543 0 : })?;
544 0 : {
545 0 : layers.into_iter().for_each(|l: Layer| {
546 0 : l.delete_on_drop();
547 0 : std::mem::drop(l);
548 0 : });
549 0 : }
550 0 : Ok(())
551 0 : }
552 :
553 0 : let mut should_fsync = false;
554 0 : let mut first_err = None;
555 0 : while let Some(res) = tasks.join_next().await {
556 0 : match res {
557 0 : Ok(Ok((owned, did_hardlink))) => {
558 0 : if did_hardlink {
559 0 : should_fsync = true;
560 0 : }
561 0 : new_layers.push(owned);
562 : }
563 :
564 : // Don't stop the evaluation on errors, so that we get the full set of hardlinked layers to delete.
565 0 : Ok(Err(failed)) => {
566 0 : cancel_eval.cancel();
567 0 : first_err.get_or_insert(failed);
568 0 : }
569 0 : Err(je) => {
570 0 : cancel_eval.cancel();
571 0 : first_err.get_or_insert(Error::Prepare(je.into()));
572 0 : }
573 : }
574 : }
575 :
576 0 : if let Some(failed) = first_err {
577 0 : delete_layers(detached, new_layers)?;
578 0 : return Err(failed);
579 0 : }
580 0 :
581 0 : // fsync directory again if we hardlinked something
582 0 : if should_fsync {
583 0 : fsync_timeline_dir(detached, ctx).await;
584 0 : }
585 :
586 0 : let prepared = PreparedTimelineDetach { layers: new_layers };
587 0 :
588 0 : Ok(Progress::Prepared(attempt, prepared))
589 0 : }
590 :
591 0 : async fn start_new_attempt(
592 0 : detached: &Timeline,
593 0 : tenant: &TenantShard,
594 0 : ancestor_timeline_id: TimelineId,
595 0 : ancestor_lsn: Lsn,
596 0 : ) -> Result<Attempt, Error> {
597 0 : let attempt = obtain_exclusive_attempt(detached, tenant, ancestor_timeline_id, ancestor_lsn)?;
598 :
599 : // insert the block in the index_part.json, if not already there.
600 0 : let _dont_care = tenant
601 0 : .gc_block
602 0 : .insert(
603 0 : detached,
604 0 : crate::tenant::remote_timeline_client::index::GcBlockingReason::DetachAncestor,
605 0 : )
606 0 : .await
607 0 : .map_err(|e| Error::launder(e, Error::Prepare))?;
608 :
609 0 : Ok(attempt)
610 0 : }
611 :
612 0 : async fn continue_with_blocked_gc(
613 0 : detached: &Timeline,
614 0 : tenant: &TenantShard,
615 0 : ancestor_timeline_id: TimelineId,
616 0 : ancestor_lsn: Lsn,
617 0 : ) -> Result<Attempt, Error> {
618 0 : // FIXME: it would be nice to confirm that there is an in-memory version, since we've just
619 0 : // verified there is a persistent one?
620 0 : obtain_exclusive_attempt(detached, tenant, ancestor_timeline_id, ancestor_lsn)
621 0 : }
622 :
623 0 : fn obtain_exclusive_attempt(
624 0 : detached: &Timeline,
625 0 : tenant: &TenantShard,
626 0 : ancestor_timeline_id: TimelineId,
627 0 : ancestor_lsn: Lsn,
628 0 : ) -> Result<Attempt, Error> {
629 : use Error::{OtherTimelineDetachOngoing, ShuttingDown};
630 :
631 : // ensure we are the only active attempt for this tenant
632 0 : let (guard, barrier) = completion::channel();
633 0 : {
634 0 : let mut guard = tenant.ongoing_timeline_detach.lock().unwrap();
635 0 : if let Some((tl, other)) = guard.as_ref() {
636 0 : if !other.is_ready() {
637 0 : return Err(OtherTimelineDetachOngoing(*tl));
638 0 : }
639 : // FIXME: no test enters here
640 0 : }
641 0 : *guard = Some((detached.timeline_id, barrier));
642 : }
643 :
644 : // ensure the gate is still open
645 0 : let _gate_entered = detached.gate.enter().map_err(|_| ShuttingDown)?;
646 :
647 0 : Ok(Attempt {
648 0 : timeline_id: detached.timeline_id,
649 0 : ancestor_timeline_id,
650 0 : ancestor_lsn,
651 0 : _guard: guard,
652 0 : gate_entered: Some(_gate_entered),
653 0 : })
654 0 : }
655 :
656 0 : fn reparented_direct_children(
657 0 : detached: &Arc<Timeline>,
658 0 : tenant: &TenantShard,
659 0 : ) -> Result<HashSet<TimelineId>, Error> {
660 0 : let mut all_direct_children = tenant
661 0 : .timelines
662 0 : .lock()
663 0 : .unwrap()
664 0 : .values()
665 0 : .filter_map(|tl| {
666 0 : let is_direct_child = matches!(tl.ancestor_timeline.as_ref(), Some(ancestor) if Arc::ptr_eq(ancestor, detached));
667 :
668 0 : if is_direct_child {
669 0 : Some(tl.clone())
670 : } else {
671 0 : if let Some(timeline) = tl.ancestor_timeline.as_ref() {
672 0 : assert_ne!(timeline.timeline_id, detached.timeline_id, "we cannot have two timelines with the same timeline_id live");
673 0 : }
674 0 : None
675 : }
676 0 : })
677 0 : // Collect to avoid lock taking order problem with Tenant::timelines and
678 0 : // Timeline::remote_client
679 0 : .collect::<Vec<_>>();
680 0 :
681 0 : let mut any_shutdown = false;
682 0 :
683 0 : all_direct_children.retain(|tl| match tl.remote_client.initialized_upload_queue() {
684 0 : Ok(accessor) => accessor
685 0 : .latest_uploaded_index_part()
686 0 : .lineage
687 0 : .is_reparented(),
688 0 : Err(_shutdownalike) => {
689 0 : // not 100% a shutdown, but let's bail early not to give inconsistent results in
690 0 : // sharded enviroment.
691 0 : any_shutdown = true;
692 0 : true
693 : }
694 0 : });
695 0 :
696 0 : if any_shutdown {
697 : // it could be one or many being deleted; have client retry
698 0 : return Err(Error::ShuttingDown);
699 0 : }
700 0 :
701 0 : Ok(all_direct_children
702 0 : .into_iter()
703 0 : .map(|tl| tl.timeline_id)
704 0 : .collect())
705 0 : }
706 :
707 0 : fn partition_work(
708 0 : ancestor_lsn: Lsn,
709 0 : source: &LayerManager,
710 0 : ) -> Result<(usize, Vec<Layer>, Vec<Layer>), Error> {
711 0 : let mut straddling_branchpoint = vec![];
712 0 : let mut rest_of_historic = vec![];
713 0 :
714 0 : let mut later_by_lsn = 0;
715 :
716 0 : for desc in source.layer_map()?.iter_historic_layers() {
717 : // off by one chances here:
718 : // - start is inclusive
719 : // - end is exclusive
720 0 : if desc.lsn_range.start > ancestor_lsn {
721 0 : later_by_lsn += 1;
722 0 : continue;
723 0 : }
724 :
725 0 : let target = if desc.lsn_range.start <= ancestor_lsn
726 0 : && desc.lsn_range.end > ancestor_lsn
727 0 : && desc.is_delta
728 : {
729 : // TODO: image layer at Lsn optimization
730 0 : &mut straddling_branchpoint
731 : } else {
732 0 : &mut rest_of_historic
733 : };
734 :
735 0 : target.push(source.get_from_desc(&desc));
736 : }
737 :
738 0 : Ok((later_by_lsn, straddling_branchpoint, rest_of_historic))
739 0 : }
740 :
741 0 : async fn upload_rewritten_layer(
742 0 : end_lsn: Lsn,
743 0 : layer: &Layer,
744 0 : target: &Arc<Timeline>,
745 0 : cancel: &CancellationToken,
746 0 : ctx: &RequestContext,
747 0 : ) -> Result<Option<Layer>, Error> {
748 0 : let copied = copy_lsn_prefix(end_lsn, layer, target, ctx).await?;
749 :
750 0 : let Some(copied) = copied else {
751 0 : return Ok(None);
752 : };
753 :
754 0 : target
755 0 : .remote_client
756 0 : .upload_layer_file(&copied, cancel)
757 0 : .await
758 0 : .map_err(|e| Error::launder(e, Error::Prepare))?;
759 :
760 0 : Ok(Some(copied.into()))
761 0 : }
762 :
763 0 : async fn copy_lsn_prefix(
764 0 : end_lsn: Lsn,
765 0 : layer: &Layer,
766 0 : target_timeline: &Arc<Timeline>,
767 0 : ctx: &RequestContext,
768 0 : ) -> Result<Option<ResidentLayer>, Error> {
769 0 : if target_timeline.cancel.is_cancelled() {
770 0 : return Err(Error::ShuttingDown);
771 0 : }
772 0 :
773 0 : tracing::debug!(%layer, %end_lsn, "copying lsn prefix");
774 :
775 0 : let mut writer = DeltaLayerWriter::new(
776 0 : target_timeline.conf,
777 0 : target_timeline.timeline_id,
778 0 : target_timeline.tenant_shard_id,
779 0 : layer.layer_desc().key_range.start,
780 0 : layer.layer_desc().lsn_range.start..end_lsn,
781 0 : &target_timeline.gate,
782 0 : target_timeline.cancel.clone(),
783 0 : ctx,
784 0 : )
785 0 : .await
786 0 : .with_context(|| format!("prepare to copy lsn prefix of ancestors {layer}"))
787 0 : .map_err(Error::Prepare)?;
788 :
789 0 : let resident = layer.download_and_keep_resident(ctx).await.map_err(|e| {
790 0 : if e.is_cancelled() {
791 0 : Error::ShuttingDown
792 : } else {
793 0 : Error::Prepare(e.into())
794 : }
795 0 : })?;
796 :
797 0 : let records = resident
798 0 : .copy_delta_prefix(&mut writer, end_lsn, ctx)
799 0 : .await
800 0 : .with_context(|| format!("copy lsn prefix of ancestors {layer}"))
801 0 : .map_err(Error::Prepare)?;
802 :
803 0 : drop(resident);
804 0 :
805 0 : tracing::debug!(%layer, records, "copied records");
806 :
807 0 : if records == 0 {
808 0 : drop(writer);
809 0 : // TODO: we might want to store an empty marker in remote storage for this
810 0 : // layer so that we will not needlessly walk `layer` on repeated attempts.
811 0 : Ok(None)
812 : } else {
813 : // reuse the key instead of adding more holes between layers by using the real
814 : // highest key in the layer.
815 0 : let reused_highest_key = layer.layer_desc().key_range.end;
816 0 : let (desc, path) = writer
817 0 : .finish(reused_highest_key, ctx)
818 0 : .await
819 0 : .map_err(Error::Prepare)?;
820 0 : let copied = Layer::finish_creating(target_timeline.conf, target_timeline, desc, &path)
821 0 : .map_err(Error::Prepare)?;
822 :
823 0 : tracing::debug!(%layer, %copied, "new layer produced");
824 :
825 0 : Ok(Some(copied))
826 : }
827 0 : }
828 :
829 : /// Creates a new Layer instance for the adopted layer, and ensures it is found in the remote
830 : /// storage on successful return. without the adopted layer being added to `index_part.json`.
831 : /// Returns (Layer, did hardlink)
832 0 : async fn remote_copy(
833 0 : adopted: &Layer,
834 0 : adoptee: &Arc<Timeline>,
835 0 : generation: Generation,
836 0 : shard_identity: ShardIdentity,
837 0 : cancel: &CancellationToken,
838 0 : ) -> Result<(Layer, bool), Error> {
839 0 : let mut metadata = adopted.metadata();
840 0 : debug_assert!(metadata.generation <= generation);
841 0 : metadata.generation = generation;
842 0 : metadata.shard = shard_identity.shard_index();
843 0 :
844 0 : let conf = adoptee.conf;
845 0 : let file_name = adopted.layer_desc().layer_name();
846 :
847 : // We don't want to shut the timeline down during this operation because we do `delete_on_drop` below
848 0 : let _gate = adoptee.gate.enter().map_err(|e| match e {
849 0 : GateError::GateClosed => Error::ShuttingDown,
850 0 : })?;
851 :
852 : // depending if Layer::keep_resident, do a hardlink
853 : let did_hardlink;
854 0 : let owned = if let Some(adopted_resident) = adopted.keep_resident().await {
855 0 : let adopted_path = adopted_resident.local_path();
856 0 : let adoptee_path = local_layer_path(
857 0 : conf,
858 0 : &adoptee.tenant_shard_id,
859 0 : &adoptee.timeline_id,
860 0 : &file_name,
861 0 : &metadata.generation,
862 0 : );
863 0 :
864 0 : match std::fs::hard_link(adopted_path, &adoptee_path) {
865 0 : Ok(()) => {}
866 0 : Err(e) if e.kind() == std::io::ErrorKind::AlreadyExists => {
867 0 : // In theory we should not get into this situation as we are doing cleanups of the layer file after errors.
868 0 : // However, we don't do cleanups for errors past `prepare`, so there is the slight chance to get to this branch.
869 0 :
870 0 : // Double check that the file is orphan (probably from an earlier attempt), then delete it
871 0 : let key = file_name.clone().into();
872 0 : if adoptee.layers.read().await.contains_key(&key) {
873 : // We are supposed to filter out such cases before coming to this function
874 0 : return Err(Error::Prepare(anyhow::anyhow!(
875 0 : "layer file {file_name} already present and inside layer map"
876 0 : )));
877 0 : }
878 0 : tracing::info!("Deleting orphan layer file to make way for hard linking");
879 : // Delete orphan layer file and try again, to ensure this layer has a well understood source
880 0 : std::fs::remove_file(adopted_path)
881 0 : .map_err(|e| Error::launder(e.into(), Error::Prepare))?;
882 0 : std::fs::hard_link(adopted_path, &adoptee_path)
883 0 : .map_err(|e| Error::launder(e.into(), Error::Prepare))?;
884 : }
885 0 : Err(e) => {
886 0 : return Err(Error::launder(e.into(), Error::Prepare));
887 : }
888 : };
889 0 : did_hardlink = true;
890 0 : Layer::for_resident(conf, adoptee, adoptee_path, file_name, metadata).drop_eviction_guard()
891 : } else {
892 0 : did_hardlink = false;
893 0 : Layer::for_evicted(conf, adoptee, file_name, metadata)
894 : };
895 :
896 0 : let layer = match adoptee
897 0 : .remote_client
898 0 : .copy_timeline_layer(adopted, &owned, cancel)
899 0 : .await
900 : {
901 0 : Ok(()) => owned,
902 0 : Err(e) => {
903 0 : {
904 0 : // Clean up the layer so that on a retry we don't get errors that the file already exists
905 0 : owned.delete_on_drop();
906 0 : std::mem::drop(owned);
907 0 : }
908 0 : return Err(Error::launder(e, Error::Prepare));
909 : }
910 : };
911 :
912 0 : Ok((layer, did_hardlink))
913 0 : }
914 :
915 : pub(crate) enum DetachingAndReparenting {
916 : /// All of the following timeline ids were reparented and the timeline ancestor detach must be
917 : /// marked as completed.
918 : Reparented(HashSet<TimelineId>),
919 :
920 : /// Some of the reparentings failed. The timeline ancestor detach must **not** be marked as
921 : /// completed.
922 : ///
923 : /// Nested `must_reset_tenant` is set to true when any restart requiring changes were made.
924 : SomeReparentingFailed { must_reset_tenant: bool },
925 :
926 : /// Detaching and reparentings were completed in a previous attempt. Timeline ancestor detach
927 : /// must be marked as completed.
928 : AlreadyDone(HashSet<TimelineId>),
929 : }
930 :
931 : impl DetachingAndReparenting {
932 0 : pub(crate) fn reset_tenant_required(&self) -> bool {
933 : use DetachingAndReparenting::*;
934 0 : match self {
935 0 : Reparented(_) => true,
936 0 : SomeReparentingFailed { must_reset_tenant } => *must_reset_tenant,
937 0 : AlreadyDone(_) => false,
938 : }
939 0 : }
940 :
941 0 : pub(crate) fn completed(self) -> Option<HashSet<TimelineId>> {
942 : use DetachingAndReparenting::*;
943 0 : match self {
944 0 : Reparented(x) | AlreadyDone(x) => Some(x),
945 0 : SomeReparentingFailed { .. } => None,
946 : }
947 0 : }
948 : }
949 :
950 : /// See [`Timeline::detach_from_ancestor_and_reparent`].
951 0 : pub(super) async fn detach_and_reparent(
952 0 : detached: &Arc<Timeline>,
953 0 : tenant: &TenantShard,
954 0 : prepared: PreparedTimelineDetach,
955 0 : ancestor_timeline_id: TimelineId,
956 0 : ancestor_lsn: Lsn,
957 0 : behavior: DetachBehavior,
958 0 : _ctx: &RequestContext,
959 0 : ) -> Result<DetachingAndReparenting, Error> {
960 0 : let PreparedTimelineDetach { layers } = prepared;
961 :
962 : #[derive(Debug)]
963 : enum Ancestor {
964 : NotDetached(Arc<Timeline>, Lsn),
965 : Detached(Arc<Timeline>, Lsn),
966 : }
967 :
968 0 : let (recorded_branchpoint, still_ongoing) = {
969 0 : let access = detached.remote_client.initialized_upload_queue()?;
970 0 : let latest = access.latest_uploaded_index_part();
971 0 :
972 0 : (
973 0 : latest.lineage.detached_previous_ancestor(),
974 0 : latest
975 0 : .gc_blocking
976 0 : .as_ref()
977 0 : .is_some_and(|b| b.blocked_by(DetachAncestor)),
978 0 : )
979 0 : };
980 0 : assert!(
981 0 : still_ongoing,
982 0 : "cannot (detach? reparent)? complete if the operation is not still ongoing"
983 : );
984 :
985 0 : let ancestor_to_detach = match detached.ancestor_timeline.as_ref() {
986 0 : Some(mut ancestor) => {
987 0 : while ancestor.timeline_id != ancestor_timeline_id {
988 0 : match ancestor.ancestor_timeline.as_ref() {
989 0 : Some(found) => {
990 0 : if ancestor_lsn != ancestor.ancestor_lsn {
991 0 : return Err(Error::DetachReparent(anyhow::anyhow!(
992 0 : "cannot find the ancestor timeline to detach from: wrong ancestor lsn"
993 0 : )));
994 0 : }
995 0 : ancestor = found;
996 : }
997 : None => {
998 0 : return Err(Error::DetachReparent(anyhow::anyhow!(
999 0 : "cannot find the ancestor timeline to detach from"
1000 0 : )));
1001 : }
1002 : }
1003 : }
1004 0 : Some(ancestor)
1005 : }
1006 0 : None => None,
1007 : };
1008 0 : let ancestor = match (ancestor_to_detach, recorded_branchpoint) {
1009 0 : (Some(ancestor), None) => {
1010 0 : assert!(
1011 0 : !layers.is_empty(),
1012 0 : "there should always be at least one layer to inherit"
1013 : );
1014 0 : Ancestor::NotDetached(ancestor.clone(), detached.ancestor_lsn)
1015 : }
1016 : (Some(_), Some(_)) => {
1017 0 : panic!(
1018 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"
1019 0 : );
1020 : }
1021 0 : (None, Some((ancestor_id, ancestor_lsn))) => {
1022 0 : // it has been either:
1023 0 : // - detached but still exists => we can try reparenting
1024 0 : // - detached and deleted
1025 0 : //
1026 0 : // either way, we must complete
1027 0 : assert!(
1028 0 : layers.is_empty(),
1029 0 : "no layers should had been copied as detach is done"
1030 : );
1031 :
1032 0 : let existing = tenant.timelines.lock().unwrap().get(&ancestor_id).cloned();
1033 :
1034 0 : if let Some(ancestor) = existing {
1035 0 : Ancestor::Detached(ancestor, ancestor_lsn)
1036 : } else {
1037 0 : let direct_children = reparented_direct_children(detached, tenant)?;
1038 0 : return Ok(DetachingAndReparenting::AlreadyDone(direct_children));
1039 : }
1040 : }
1041 : (None, None) => {
1042 : // TODO: make sure there are no `?` before tenant_reset from after a questionmark from
1043 : // here.
1044 0 : panic!(
1045 0 : "bug: detach_and_reparent called on a timeline which has not been detached or which has no live ancestor"
1046 0 : );
1047 : }
1048 : };
1049 :
1050 : // publish the prepared layers before we reparent any of the timelines, so that on restart
1051 : // reparented timelines find layers. also do the actual detaching.
1052 : //
1053 : // if we crash after this operation, a retry will allow reparenting the remaining timelines as
1054 : // gc is blocked.
1055 :
1056 0 : let (ancestor, ancestor_lsn, was_detached) = match ancestor {
1057 0 : Ancestor::NotDetached(ancestor, ancestor_lsn) => {
1058 0 : // this has to complete before any reparentings because otherwise they would not have
1059 0 : // layers on the new parent.
1060 0 : detached
1061 0 : .remote_client
1062 0 : .schedule_adding_existing_layers_to_index_detach_and_wait(
1063 0 : &layers,
1064 0 : (ancestor.timeline_id, ancestor_lsn),
1065 0 : )
1066 0 : .await
1067 0 : .context("publish layers and detach ancestor")
1068 0 : .map_err(|e| Error::launder(e, Error::DetachReparent))?;
1069 :
1070 0 : tracing::info!(
1071 0 : ancestor=%ancestor.timeline_id,
1072 0 : %ancestor_lsn,
1073 0 : inherited_layers=%layers.len(),
1074 0 : "detached from ancestor"
1075 : );
1076 0 : (ancestor, ancestor_lsn, true)
1077 : }
1078 0 : Ancestor::Detached(ancestor, ancestor_lsn) => (ancestor, ancestor_lsn, false),
1079 : };
1080 :
1081 0 : if let DetachBehavior::MultiLevelAndNoReparent = behavior {
1082 : // Do not reparent if the user requests to behave so.
1083 0 : return Ok(DetachingAndReparenting::Reparented(HashSet::new()));
1084 0 : }
1085 0 :
1086 0 : let mut tasks = tokio::task::JoinSet::new();
1087 0 :
1088 0 : // Returns a single permit semaphore which will be used to make one reparenting succeed,
1089 0 : // others will fail as if those timelines had been stopped for whatever reason.
1090 0 : #[cfg(feature = "testing")]
1091 0 : let failpoint_sem = || -> Option<Arc<Semaphore>> {
1092 0 : fail::fail_point!("timeline-detach-ancestor::allow_one_reparented", |_| Some(
1093 0 : Arc::new(Semaphore::new(1))
1094 0 : ));
1095 0 : None
1096 0 : }();
1097 0 :
1098 0 : // because we are now keeping the slot in progress, it is unlikely that there will be any
1099 0 : // timeline deletions during this time. if we raced one, then we'll just ignore it.
1100 0 : {
1101 0 : let g = tenant.timelines.lock().unwrap();
1102 0 : reparentable_timelines(g.values(), detached, &ancestor, ancestor_lsn)
1103 0 : .cloned()
1104 0 : .for_each(|timeline| {
1105 : // important in this scope: we are holding the Tenant::timelines lock
1106 0 : let span = tracing::info_span!("reparent", reparented=%timeline.timeline_id);
1107 0 : let new_parent = detached.timeline_id;
1108 0 : #[cfg(feature = "testing")]
1109 0 : let failpoint_sem = failpoint_sem.clone();
1110 0 :
1111 0 : tasks.spawn(
1112 0 : async move {
1113 0 : let res = async {
1114 : #[cfg(feature = "testing")]
1115 0 : if let Some(failpoint_sem) = failpoint_sem {
1116 0 : let _permit = failpoint_sem.acquire().await.map_err(|_| {
1117 0 : anyhow::anyhow!(
1118 0 : "failpoint: timeline-detach-ancestor::allow_one_reparented",
1119 0 : )
1120 0 : })?;
1121 0 : failpoint_sem.close();
1122 0 : }
1123 :
1124 0 : timeline
1125 0 : .remote_client
1126 0 : .schedule_reparenting_and_wait(&new_parent)
1127 0 : .await
1128 0 : }
1129 0 : .await;
1130 :
1131 0 : match res {
1132 : Ok(()) => {
1133 0 : tracing::info!("reparented");
1134 0 : Some(timeline)
1135 : }
1136 0 : Err(e) => {
1137 0 : // with the use of tenant slot, raced timeline deletion is the most
1138 0 : // likely reason.
1139 0 : tracing::warn!("reparenting failed: {e:#}");
1140 0 : None
1141 : }
1142 : }
1143 0 : }
1144 0 : .instrument(span),
1145 0 : );
1146 0 : });
1147 0 : }
1148 0 :
1149 0 : let reparenting_candidates = tasks.len();
1150 0 : let mut reparented = HashSet::with_capacity(tasks.len());
1151 :
1152 0 : while let Some(res) = tasks.join_next().await {
1153 0 : match res {
1154 0 : Ok(Some(timeline)) => {
1155 0 : assert!(
1156 0 : reparented.insert(timeline.timeline_id),
1157 0 : "duplicate reparenting? timeline_id={}",
1158 0 : timeline.timeline_id
1159 : );
1160 : }
1161 0 : Err(je) if je.is_cancelled() => unreachable!("not used"),
1162 : // just ignore failures now, we can retry
1163 0 : Ok(None) => {}
1164 0 : Err(je) if je.is_panic() => {}
1165 0 : Err(je) => tracing::error!("unexpected join error: {je:?}"),
1166 : }
1167 : }
1168 :
1169 0 : let reparented_all = reparenting_candidates == reparented.len();
1170 0 :
1171 0 : if reparented_all {
1172 0 : Ok(DetachingAndReparenting::Reparented(reparented))
1173 : } else {
1174 0 : tracing::info!(
1175 0 : reparented = reparented.len(),
1176 0 : candidates = reparenting_candidates,
1177 0 : "failed to reparent all candidates; they can be retried after the tenant_reset",
1178 : );
1179 :
1180 0 : let must_reset_tenant = !reparented.is_empty() || was_detached;
1181 0 : Ok(DetachingAndReparenting::SomeReparentingFailed { must_reset_tenant })
1182 : }
1183 0 : }
1184 :
1185 0 : pub(super) async fn complete(
1186 0 : detached: &Arc<Timeline>,
1187 0 : tenant: &TenantShard,
1188 0 : mut attempt: Attempt,
1189 0 : _ctx: &RequestContext,
1190 0 : ) -> Result<(), Error> {
1191 0 : assert_eq!(detached.timeline_id, attempt.timeline_id);
1192 :
1193 0 : if attempt.gate_entered.is_none() {
1194 0 : let entered = detached.gate.enter().map_err(|_| Error::ShuttingDown)?;
1195 0 : attempt.gate_entered = Some(entered);
1196 0 : } else {
1197 0 : // Some(gate_entered) means the tenant was not restarted, as is not required
1198 0 : }
1199 :
1200 0 : assert!(detached.ancestor_timeline.is_none());
1201 :
1202 : // this should be an 503 at least...?
1203 0 : fail::fail_point!(
1204 0 : "timeline-detach-ancestor::complete_before_uploading",
1205 0 : |_| Err(Error::Failpoint(
1206 0 : "timeline-detach-ancestor::complete_before_uploading"
1207 0 : ))
1208 0 : );
1209 :
1210 0 : tenant
1211 0 : .gc_block
1212 0 : .remove(
1213 0 : detached,
1214 0 : crate::tenant::remote_timeline_client::index::GcBlockingReason::DetachAncestor,
1215 0 : )
1216 0 : .await
1217 0 : .map_err(|e| Error::launder(e, Error::Complete))?;
1218 :
1219 0 : Ok(())
1220 0 : }
1221 :
1222 : /// Query against a locked `Tenant::timelines`.
1223 : ///
1224 : /// A timeline is reparentable if:
1225 : ///
1226 : /// - It is not the timeline being detached.
1227 : /// - It has the same ancestor as the timeline being detached. Note that the ancestor might not be the direct ancestor.
1228 0 : fn reparentable_timelines<'a, I>(
1229 0 : timelines: I,
1230 0 : detached: &'a Arc<Timeline>,
1231 0 : ancestor: &'a Arc<Timeline>,
1232 0 : ancestor_lsn: Lsn,
1233 0 : ) -> impl Iterator<Item = &'a Arc<Timeline>> + 'a
1234 0 : where
1235 0 : I: Iterator<Item = &'a Arc<Timeline>> + 'a,
1236 0 : {
1237 0 : timelines.filter_map(move |tl| {
1238 0 : if Arc::ptr_eq(tl, detached) {
1239 0 : return None;
1240 0 : }
1241 :
1242 0 : let tl_ancestor = tl.ancestor_timeline.as_ref()?;
1243 0 : let is_same = Arc::ptr_eq(ancestor, tl_ancestor);
1244 0 : let is_earlier = tl.get_ancestor_lsn() <= ancestor_lsn;
1245 0 :
1246 0 : let is_deleting = tl
1247 0 : .delete_progress
1248 0 : .try_lock()
1249 0 : .map(|flow| !flow.is_not_started())
1250 0 : .unwrap_or(true);
1251 0 :
1252 0 : if is_same && is_earlier && !is_deleting {
1253 0 : Some(tl)
1254 : } else {
1255 0 : None
1256 : }
1257 0 : })
1258 0 : }
1259 :
1260 0 : fn check_no_archived_children_of_ancestor(
1261 0 : tenant: &TenantShard,
1262 0 : detached: &Arc<Timeline>,
1263 0 : ancestor: &Arc<Timeline>,
1264 0 : ancestor_lsn: Lsn,
1265 0 : detach_behavior: DetachBehavior,
1266 0 : ) -> Result<(), Error> {
1267 0 : match detach_behavior {
1268 : DetachBehavior::NoAncestorAndReparent => {
1269 0 : let timelines = tenant.timelines.lock().unwrap();
1270 0 : let timelines_offloaded = tenant.timelines_offloaded.lock().unwrap();
1271 :
1272 0 : for timeline in
1273 0 : reparentable_timelines(timelines.values(), detached, ancestor, ancestor_lsn)
1274 : {
1275 0 : if timeline.is_archived() == Some(true) {
1276 0 : return Err(Error::Archived(timeline.timeline_id));
1277 0 : }
1278 : }
1279 :
1280 0 : for timeline_offloaded in timelines_offloaded.values() {
1281 0 : if timeline_offloaded.ancestor_timeline_id != Some(ancestor.timeline_id) {
1282 0 : continue;
1283 0 : }
1284 : // This forbids the detach ancestor feature if flattened timelines are present,
1285 : // even if the ancestor_lsn is from after the branchpoint of the detached timeline.
1286 : // But as per current design, we don't record the ancestor_lsn of flattened timelines.
1287 : // This is a bit unfortunate, but as of writing this we don't support flattening
1288 : // anyway. Maybe we can evolve the data model in the future.
1289 0 : if let Some(retain_lsn) = timeline_offloaded.ancestor_retain_lsn {
1290 0 : let is_earlier = retain_lsn <= ancestor_lsn;
1291 0 : if !is_earlier {
1292 0 : continue;
1293 0 : }
1294 0 : }
1295 0 : return Err(Error::Archived(timeline_offloaded.timeline_id));
1296 : }
1297 : }
1298 0 : DetachBehavior::MultiLevelAndNoReparent => {
1299 0 : // We don't need to check anything if the user requested to not reparent.
1300 0 : }
1301 : }
1302 :
1303 0 : Ok(())
1304 0 : }
1305 :
1306 0 : async fn fsync_timeline_dir(timeline: &Timeline, ctx: &RequestContext) {
1307 0 : let path = &timeline
1308 0 : .conf
1309 0 : .timeline_path(&timeline.tenant_shard_id, &timeline.timeline_id);
1310 0 : let timeline_dir = VirtualFile::open(&path, ctx)
1311 0 : .await
1312 0 : .fatal_err("VirtualFile::open for timeline dir fsync");
1313 0 : timeline_dir
1314 0 : .sync_all()
1315 0 : .await
1316 0 : .fatal_err("VirtualFile::sync_all timeline dir");
1317 0 : }
|
