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, LayerManagerLockHolder};
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 : pub(crate) 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.get_vectored_concurrent_io,
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
203 0 : .layers
204 0 : .read(LayerManagerLockHolder::DetachAncestor)
205 0 : .await;
206 0 : for layer in layers.all_persistent_layers() {
207 0 : if !layer.is_delta
208 0 : && layer.lsn_range.start == image_lsn
209 0 : && overlaps_with(&key_range, &layer.key_range)
210 : {
211 0 : tracing::warn!(
212 0 : layer=%layer, "image layer at the detach LSN already exists, skipping removing aux files"
213 : );
214 0 : return Ok(None);
215 0 : }
216 : }
217 : }
218 :
219 0 : let query = VersionedKeySpaceQuery::uniform(KeySpace::single(key_range.clone()), image_lsn);
220 0 : let data = ancestor
221 0 : .get_vectored_impl(query, &mut reconstruct_state, ctx)
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 : &detached.gate,
235 0 : detached.cancel.clone(),
236 0 : ctx,
237 0 : )
238 0 : .await
239 0 : .context("failed to create image layer writer")
240 0 : .map_err(Error::Prepare)?;
241 0 : for key in data.keys() {
242 0 : image_layer_writer
243 0 : .put_image(*key, Bytes::new(), ctx)
244 0 : .await
245 0 : .context("failed to write key")
246 0 : .map_err(|e| Error::launder(e, Error::Prepare))?;
247 : }
248 0 : let (desc, path) = image_layer_writer
249 0 : .finish(ctx)
250 0 : .await
251 0 : .context("failed to finish image layer writer for removing the metadata keys")
252 0 : .map_err(|e| Error::launder(e, Error::Prepare))?;
253 0 : let generated = Layer::finish_creating(detached.conf, detached, desc, &path)
254 0 : .map_err(|e| Error::launder(e, Error::Prepare))?;
255 0 : detached
256 0 : .remote_client
257 0 : .upload_layer_file(&generated, &detached.cancel)
258 0 : .await
259 0 : .map_err(|e| Error::launder(e, Error::Prepare))?;
260 0 : tracing::info!(layer=%generated, "wrote image layer");
261 0 : Ok(Some(generated))
262 : } else {
263 0 : tracing::info!("no aux keys found in ancestor");
264 0 : Ok(None)
265 : }
266 0 : }
267 :
268 : /// See [`Timeline::prepare_to_detach_from_ancestor`]
269 0 : pub(super) async fn prepare(
270 0 : detached: &Arc<Timeline>,
271 0 : tenant: &TenantShard,
272 0 : behavior: DetachBehavior,
273 0 : options: Options,
274 0 : ctx: &RequestContext,
275 0 : ) -> Result<Progress, Error> {
276 : use Error::*;
277 :
278 0 : let Some((mut ancestor, mut ancestor_lsn)) = detached
279 0 : .ancestor_timeline
280 0 : .as_ref()
281 0 : .map(|tl| (tl.clone(), detached.ancestor_lsn))
282 : else {
283 : let ancestor_id;
284 : let ancestor_lsn;
285 0 : let still_in_progress = {
286 0 : let accessor = detached.remote_client.initialized_upload_queue()?;
287 :
288 : // we are safe to inspect the latest uploaded, because we can only witness this after
289 : // restart is complete and ancestor is no more.
290 0 : let latest = accessor.latest_uploaded_index_part();
291 0 : let Some((id, lsn)) = latest.lineage.detached_previous_ancestor() else {
292 0 : return Err(NoAncestor);
293 : };
294 0 : ancestor_id = id;
295 0 : ancestor_lsn = lsn;
296 0 :
297 0 : latest
298 0 : .gc_blocking
299 0 : .as_ref()
300 0 : .is_some_and(|b| b.blocked_by(DetachAncestor))
301 0 : };
302 0 :
303 0 : if still_in_progress {
304 : // gc is still blocked, we can still reparent and complete.
305 : // we are safe to reparent remaining, because they were locked in in the beginning.
306 0 : let attempt =
307 0 : continue_with_blocked_gc(detached, tenant, ancestor_id, ancestor_lsn).await?;
308 :
309 : // because the ancestor of detached is already set to none, we have published all
310 : // of the layers, so we are still "prepared."
311 0 : return Ok(Progress::Prepared(
312 0 : attempt,
313 0 : PreparedTimelineDetach { layers: Vec::new() },
314 0 : ));
315 0 : }
316 :
317 0 : let reparented_timelines = reparented_direct_children(detached, tenant)?;
318 0 : return Ok(Progress::Done(AncestorDetached {
319 0 : reparented_timelines,
320 0 : }));
321 : };
322 :
323 0 : if detached.is_archived() != Some(false) {
324 0 : return Err(Archived(detached.timeline_id));
325 0 : }
326 0 :
327 0 : if !ancestor_lsn.is_valid() {
328 : // rare case, probably wouldn't even load
329 0 : tracing::error!("ancestor is set, but ancestor_lsn is invalid, this timeline needs fixing");
330 0 : return Err(NoAncestor);
331 0 : }
332 0 :
333 0 : check_no_archived_children_of_ancestor(tenant, detached, &ancestor, ancestor_lsn, behavior)?;
334 :
335 0 : if let DetachBehavior::MultiLevelAndNoReparent = behavior {
336 : // 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.
337 0 : while let Some(ancestor_of_ancestor) = ancestor.ancestor_timeline.clone() {
338 0 : if ancestor_lsn != ancestor.ancestor_lsn {
339 : // non-technical requirement; we could flatten still if ancestor LSN does not match but that needs
340 : // us to copy and cut more layers.
341 0 : return Err(AncestorNotEmpty);
342 0 : }
343 0 : // Use the ancestor of the ancestor as the new ancestor (only when the ancestor LSNs are the same)
344 0 : ancestor_lsn = ancestor.ancestor_lsn; // Get the LSN first before resetting the `ancestor` variable
345 0 : ancestor = ancestor_of_ancestor;
346 0 : // TODO: do we still need to check if we don't want to reparent?
347 0 : check_no_archived_children_of_ancestor(
348 0 : tenant,
349 0 : detached,
350 0 : &ancestor,
351 0 : ancestor_lsn,
352 0 : behavior,
353 0 : )?;
354 : }
355 0 : } else if ancestor.ancestor_timeline.is_some() {
356 : // non-technical requirement; we could flatten N ancestors just as easily but we chose
357 : // not to, at least initially
358 0 : return Err(TooManyAncestors);
359 0 : }
360 :
361 0 : tracing::info!(
362 0 : "attempt to detach the timeline from the ancestor: {}@{}, behavior={:?}",
363 0 : ancestor.timeline_id,
364 : ancestor_lsn,
365 : behavior
366 : );
367 :
368 0 : let attempt = start_new_attempt(detached, tenant, ancestor.timeline_id, ancestor_lsn).await?;
369 :
370 0 : utils::pausable_failpoint!("timeline-detach-ancestor::before_starting_after_locking-pausable");
371 :
372 0 : fail::fail_point!(
373 0 : "timeline-detach-ancestor::before_starting_after_locking",
374 0 : |_| Err(Error::Failpoint(
375 0 : "timeline-detach-ancestor::before_starting_after_locking"
376 0 : ))
377 0 : );
378 :
379 0 : if ancestor_lsn >= ancestor.get_disk_consistent_lsn() {
380 0 : let span =
381 0 : tracing::info_span!("freeze_and_flush", ancestor_timeline_id=%ancestor.timeline_id);
382 0 : async {
383 0 : let started_at = std::time::Instant::now();
384 0 : let freeze_and_flush = ancestor.freeze_and_flush0();
385 0 : let mut freeze_and_flush = std::pin::pin!(freeze_and_flush);
386 :
387 0 : let res =
388 0 : tokio::time::timeout(std::time::Duration::from_secs(1), &mut freeze_and_flush)
389 0 : .await;
390 :
391 0 : let res = match res {
392 0 : Ok(res) => res,
393 0 : Err(_elapsed) => {
394 0 : tracing::info!("freezing and flushing ancestor is still ongoing");
395 0 : freeze_and_flush.await
396 : }
397 : };
398 :
399 0 : res.map_err(|e| {
400 : use FlushLayerError::*;
401 0 : match e {
402 : Cancelled | NotRunning(_) => {
403 : // FIXME(#6424): technically statically unreachable right now, given how we never
404 : // drop the sender
405 0 : Error::ShuttingDown
406 : }
407 0 : CreateImageLayersError(_) | Other(_) => Error::Prepare(e.into()),
408 : }
409 0 : })?;
410 :
411 : // we do not need to wait for uploads to complete but we do need `struct Layer`,
412 : // copying delta prefix is unsupported currently for `InMemoryLayer`.
413 0 : tracing::info!(
414 0 : elapsed_ms = started_at.elapsed().as_millis(),
415 0 : "froze and flushed the ancestor"
416 : );
417 0 : Ok::<_, Error>(())
418 0 : }
419 0 : .instrument(span)
420 0 : .await?;
421 0 : }
422 :
423 0 : let end_lsn = ancestor_lsn + 1;
424 :
425 0 : let (filtered_layers, straddling_branchpoint, rest_of_historic) = {
426 : // we do not need to start from our layers, because they can only be layers that come
427 : // *after* ancestor_lsn
428 0 : let layers = tokio::select! {
429 0 : guard = ancestor.layers.read(LayerManagerLockHolder::DetachAncestor) => guard,
430 0 : _ = detached.cancel.cancelled() => {
431 0 : return Err(ShuttingDown);
432 : }
433 0 : _ = ancestor.cancel.cancelled() => {
434 0 : return Err(ShuttingDown);
435 : }
436 : };
437 :
438 : // between retries, these can change if compaction or gc ran in between. this will mean
439 : // we have to redo work.
440 0 : partition_work(ancestor_lsn, &layers)?
441 : };
442 :
443 : // TODO: layers are already sorted by something: use that to determine how much of remote
444 : // copies are already done -- gc is blocked, but a compaction could had happened on ancestor,
445 : // which is something to keep in mind if copy skipping is implemented.
446 0 : tracing::info!(filtered=%filtered_layers, to_rewrite = straddling_branchpoint.len(), historic=%rest_of_historic.len(), "collected layers");
447 :
448 : // TODO: copying and lsn prefix copying could be done at the same time with a single fsync after
449 0 : let mut new_layers: Vec<Layer> =
450 0 : Vec::with_capacity(straddling_branchpoint.len() + rest_of_historic.len() + 1);
451 :
452 0 : if let Some(tombstone_layer) =
453 0 : generate_tombstone_image_layer(detached, &ancestor, ancestor_lsn, ctx).await?
454 0 : {
455 0 : new_layers.push(tombstone_layer.into());
456 0 : }
457 :
458 : {
459 0 : tracing::info!(to_rewrite = %straddling_branchpoint.len(), "copying prefix of delta layers");
460 :
461 0 : let mut tasks = tokio::task::JoinSet::new();
462 0 :
463 0 : let mut wrote_any = false;
464 0 :
465 0 : let limiter = Arc::new(Semaphore::new(options.rewrite_concurrency.get()));
466 :
467 0 : for layer in straddling_branchpoint {
468 0 : let limiter = limiter.clone();
469 0 : let timeline = detached.clone();
470 0 : let ctx = ctx.detached_child(TaskKind::DetachAncestor, DownloadBehavior::Download);
471 :
472 0 : let span = tracing::info_span!("upload_rewritten_layer", %layer);
473 0 : tasks.spawn(
474 0 : async move {
475 0 : let _permit = limiter.acquire().await;
476 0 : let copied =
477 0 : upload_rewritten_layer(end_lsn, &layer, &timeline, &timeline.cancel, &ctx)
478 0 : .await?;
479 0 : if let Some(copied) = copied.as_ref() {
480 0 : tracing::info!(%copied, "rewrote and uploaded");
481 0 : }
482 0 : Ok(copied)
483 0 : }
484 0 : .instrument(span),
485 0 : );
486 0 : }
487 :
488 0 : while let Some(res) = tasks.join_next().await {
489 0 : match res {
490 0 : Ok(Ok(Some(copied))) => {
491 0 : wrote_any = true;
492 0 : new_layers.push(copied);
493 0 : }
494 0 : Ok(Ok(None)) => {}
495 0 : Ok(Err(e)) => return Err(e),
496 0 : Err(je) => return Err(Error::Prepare(je.into())),
497 : }
498 : }
499 :
500 : // FIXME: the fsync should be mandatory, after both rewrites and copies
501 0 : if wrote_any {
502 0 : fsync_timeline_dir(detached, ctx).await;
503 0 : }
504 : }
505 :
506 0 : let mut tasks = tokio::task::JoinSet::new();
507 0 : let limiter = Arc::new(Semaphore::new(options.copy_concurrency.get()));
508 0 : let cancel_eval = CancellationToken::new();
509 :
510 0 : for adopted in rest_of_historic {
511 0 : let limiter = limiter.clone();
512 0 : let timeline = detached.clone();
513 0 : let cancel_eval = cancel_eval.clone();
514 0 :
515 0 : tasks.spawn(
516 0 : async move {
517 0 : let _permit = tokio::select! {
518 0 : permit = limiter.acquire() => {
519 0 : permit
520 : }
521 : // Wait for the cancellation here instead of letting the entire task be cancelled.
522 : // Cancellations are racy in that they might leave layers on disk.
523 0 : _ = cancel_eval.cancelled() => {
524 0 : Err(Error::ShuttingDown)?
525 : }
526 : };
527 0 : let (owned, did_hardlink) = remote_copy(
528 0 : &adopted,
529 0 : &timeline,
530 0 : timeline.generation,
531 0 : timeline.shard_identity,
532 0 : &timeline.cancel,
533 0 : )
534 0 : .await?;
535 0 : tracing::info!(layer=%owned, did_hard_link=%did_hardlink, "remote copied");
536 0 : Ok((owned, did_hardlink))
537 0 : }
538 0 : .in_current_span(),
539 0 : );
540 0 : }
541 :
542 0 : fn delete_layers(timeline: &Timeline, layers: Vec<Layer>) -> Result<(), Error> {
543 : // We are deleting layers, so we must hold the gate
544 0 : let _gate = timeline.gate.enter().map_err(|e| match e {
545 0 : GateError::GateClosed => Error::ShuttingDown,
546 0 : })?;
547 0 : {
548 0 : layers.into_iter().for_each(|l: Layer| {
549 0 : l.delete_on_drop();
550 0 : std::mem::drop(l);
551 0 : });
552 0 : }
553 0 : Ok(())
554 0 : }
555 :
556 0 : let mut should_fsync = false;
557 0 : let mut first_err = None;
558 0 : while let Some(res) = tasks.join_next().await {
559 0 : match res {
560 0 : Ok(Ok((owned, did_hardlink))) => {
561 0 : if did_hardlink {
562 0 : should_fsync = true;
563 0 : }
564 0 : new_layers.push(owned);
565 : }
566 :
567 : // Don't stop the evaluation on errors, so that we get the full set of hardlinked layers to delete.
568 0 : Ok(Err(failed)) => {
569 0 : cancel_eval.cancel();
570 0 : first_err.get_or_insert(failed);
571 0 : }
572 0 : Err(je) => {
573 0 : cancel_eval.cancel();
574 0 : first_err.get_or_insert(Error::Prepare(je.into()));
575 0 : }
576 : }
577 : }
578 :
579 0 : if let Some(failed) = first_err {
580 0 : delete_layers(detached, new_layers)?;
581 0 : return Err(failed);
582 0 : }
583 0 :
584 0 : // fsync directory again if we hardlinked something
585 0 : if should_fsync {
586 0 : fsync_timeline_dir(detached, ctx).await;
587 0 : }
588 :
589 0 : let prepared = PreparedTimelineDetach { layers: new_layers };
590 0 :
591 0 : Ok(Progress::Prepared(attempt, prepared))
592 0 : }
593 :
594 0 : async fn start_new_attempt(
595 0 : detached: &Timeline,
596 0 : tenant: &TenantShard,
597 0 : ancestor_timeline_id: TimelineId,
598 0 : ancestor_lsn: Lsn,
599 0 : ) -> Result<Attempt, Error> {
600 0 : let attempt = obtain_exclusive_attempt(detached, tenant, ancestor_timeline_id, ancestor_lsn)?;
601 :
602 : // insert the block in the index_part.json, if not already there.
603 0 : let _dont_care = tenant
604 0 : .gc_block
605 0 : .insert(
606 0 : detached,
607 0 : crate::tenant::remote_timeline_client::index::GcBlockingReason::DetachAncestor,
608 0 : )
609 0 : .await
610 0 : .map_err(|e| Error::launder(e, Error::Prepare))?;
611 :
612 0 : Ok(attempt)
613 0 : }
614 :
615 0 : async fn continue_with_blocked_gc(
616 0 : detached: &Timeline,
617 0 : tenant: &TenantShard,
618 0 : ancestor_timeline_id: TimelineId,
619 0 : ancestor_lsn: Lsn,
620 0 : ) -> Result<Attempt, Error> {
621 0 : // FIXME: it would be nice to confirm that there is an in-memory version, since we've just
622 0 : // verified there is a persistent one?
623 0 : obtain_exclusive_attempt(detached, tenant, ancestor_timeline_id, ancestor_lsn)
624 0 : }
625 :
626 0 : fn obtain_exclusive_attempt(
627 0 : detached: &Timeline,
628 0 : tenant: &TenantShard,
629 0 : ancestor_timeline_id: TimelineId,
630 0 : ancestor_lsn: Lsn,
631 0 : ) -> Result<Attempt, Error> {
632 : use Error::{OtherTimelineDetachOngoing, ShuttingDown};
633 :
634 : // ensure we are the only active attempt for this tenant
635 0 : let (guard, barrier) = completion::channel();
636 0 : {
637 0 : let mut guard = tenant.ongoing_timeline_detach.lock().unwrap();
638 0 : if let Some((tl, other)) = guard.as_ref() {
639 0 : if !other.is_ready() {
640 0 : return Err(OtherTimelineDetachOngoing(*tl));
641 0 : }
642 : // FIXME: no test enters here
643 0 : }
644 0 : *guard = Some((detached.timeline_id, barrier));
645 : }
646 :
647 : // ensure the gate is still open
648 0 : let _gate_entered = detached.gate.enter().map_err(|_| ShuttingDown)?;
649 :
650 0 : Ok(Attempt {
651 0 : timeline_id: detached.timeline_id,
652 0 : ancestor_timeline_id,
653 0 : ancestor_lsn,
654 0 : _guard: guard,
655 0 : gate_entered: Some(_gate_entered),
656 0 : })
657 0 : }
658 :
659 0 : fn reparented_direct_children(
660 0 : detached: &Arc<Timeline>,
661 0 : tenant: &TenantShard,
662 0 : ) -> Result<HashSet<TimelineId>, Error> {
663 0 : let mut all_direct_children = tenant
664 0 : .timelines
665 0 : .lock()
666 0 : .unwrap()
667 0 : .values()
668 0 : .filter_map(|tl| {
669 0 : let is_direct_child = matches!(tl.ancestor_timeline.as_ref(), Some(ancestor) if Arc::ptr_eq(ancestor, detached));
670 :
671 0 : if is_direct_child {
672 0 : Some(tl.clone())
673 : } else {
674 0 : if let Some(timeline) = tl.ancestor_timeline.as_ref() {
675 0 : assert_ne!(timeline.timeline_id, detached.timeline_id, "we cannot have two timelines with the same timeline_id live");
676 0 : }
677 0 : None
678 : }
679 0 : })
680 0 : // Collect to avoid lock taking order problem with Tenant::timelines and
681 0 : // Timeline::remote_client
682 0 : .collect::<Vec<_>>();
683 0 :
684 0 : let mut any_shutdown = false;
685 0 :
686 0 : all_direct_children.retain(|tl| match tl.remote_client.initialized_upload_queue() {
687 0 : Ok(accessor) => accessor
688 0 : .latest_uploaded_index_part()
689 0 : .lineage
690 0 : .is_reparented(),
691 0 : Err(_shutdownalike) => {
692 0 : // not 100% a shutdown, but let's bail early not to give inconsistent results in
693 0 : // sharded enviroment.
694 0 : any_shutdown = true;
695 0 : true
696 : }
697 0 : });
698 0 :
699 0 : if any_shutdown {
700 : // it could be one or many being deleted; have client retry
701 0 : return Err(Error::ShuttingDown);
702 0 : }
703 0 :
704 0 : Ok(all_direct_children
705 0 : .into_iter()
706 0 : .map(|tl| tl.timeline_id)
707 0 : .collect())
708 0 : }
709 :
710 0 : fn partition_work(
711 0 : ancestor_lsn: Lsn,
712 0 : source: &LayerManager,
713 0 : ) -> Result<(usize, Vec<Layer>, Vec<Layer>), Error> {
714 0 : let mut straddling_branchpoint = vec![];
715 0 : let mut rest_of_historic = vec![];
716 0 :
717 0 : let mut later_by_lsn = 0;
718 :
719 0 : for desc in source.layer_map()?.iter_historic_layers() {
720 : // off by one chances here:
721 : // - start is inclusive
722 : // - end is exclusive
723 0 : if desc.lsn_range.start > ancestor_lsn {
724 0 : later_by_lsn += 1;
725 0 : continue;
726 0 : }
727 :
728 0 : let target = if desc.lsn_range.start <= ancestor_lsn
729 0 : && desc.lsn_range.end > ancestor_lsn
730 0 : && desc.is_delta
731 : {
732 : // TODO: image layer at Lsn optimization
733 0 : &mut straddling_branchpoint
734 : } else {
735 0 : &mut rest_of_historic
736 : };
737 :
738 0 : target.push(source.get_from_desc(&desc));
739 : }
740 :
741 0 : Ok((later_by_lsn, straddling_branchpoint, rest_of_historic))
742 0 : }
743 :
744 0 : async fn upload_rewritten_layer(
745 0 : end_lsn: Lsn,
746 0 : layer: &Layer,
747 0 : target: &Arc<Timeline>,
748 0 : cancel: &CancellationToken,
749 0 : ctx: &RequestContext,
750 0 : ) -> Result<Option<Layer>, Error> {
751 0 : let copied = copy_lsn_prefix(end_lsn, layer, target, ctx).await?;
752 :
753 0 : let Some(copied) = copied else {
754 0 : return Ok(None);
755 : };
756 :
757 0 : target
758 0 : .remote_client
759 0 : .upload_layer_file(&copied, cancel)
760 0 : .await
761 0 : .map_err(|e| Error::launder(e, Error::Prepare))?;
762 :
763 0 : Ok(Some(copied.into()))
764 0 : }
765 :
766 0 : async fn copy_lsn_prefix(
767 0 : end_lsn: Lsn,
768 0 : layer: &Layer,
769 0 : target_timeline: &Arc<Timeline>,
770 0 : ctx: &RequestContext,
771 0 : ) -> Result<Option<ResidentLayer>, Error> {
772 0 : if target_timeline.cancel.is_cancelled() {
773 0 : return Err(Error::ShuttingDown);
774 0 : }
775 0 :
776 0 : tracing::debug!(%layer, %end_lsn, "copying lsn prefix");
777 :
778 0 : let mut writer = DeltaLayerWriter::new(
779 0 : target_timeline.conf,
780 0 : target_timeline.timeline_id,
781 0 : target_timeline.tenant_shard_id,
782 0 : layer.layer_desc().key_range.start,
783 0 : layer.layer_desc().lsn_range.start..end_lsn,
784 0 : &target_timeline.gate,
785 0 : target_timeline.cancel.clone(),
786 0 : ctx,
787 0 : )
788 0 : .await
789 0 : .with_context(|| format!("prepare to copy lsn prefix of ancestors {layer}"))
790 0 : .map_err(Error::Prepare)?;
791 :
792 0 : let resident = layer.download_and_keep_resident(ctx).await.map_err(|e| {
793 0 : if e.is_cancelled() {
794 0 : Error::ShuttingDown
795 : } else {
796 0 : Error::Prepare(e.into())
797 : }
798 0 : })?;
799 :
800 0 : let records = resident
801 0 : .copy_delta_prefix(&mut writer, end_lsn, ctx)
802 0 : .await
803 0 : .with_context(|| format!("copy lsn prefix of ancestors {layer}"))
804 0 : .map_err(Error::Prepare)?;
805 :
806 0 : drop(resident);
807 0 :
808 0 : tracing::debug!(%layer, records, "copied records");
809 :
810 0 : if records == 0 {
811 0 : drop(writer);
812 0 : // TODO: we might want to store an empty marker in remote storage for this
813 0 : // layer so that we will not needlessly walk `layer` on repeated attempts.
814 0 : Ok(None)
815 : } else {
816 : // reuse the key instead of adding more holes between layers by using the real
817 : // highest key in the layer.
818 0 : let reused_highest_key = layer.layer_desc().key_range.end;
819 0 : let (desc, path) = writer
820 0 : .finish(reused_highest_key, ctx)
821 0 : .await
822 0 : .map_err(Error::Prepare)?;
823 0 : let copied = Layer::finish_creating(target_timeline.conf, target_timeline, desc, &path)
824 0 : .map_err(Error::Prepare)?;
825 :
826 0 : tracing::debug!(%layer, %copied, "new layer produced");
827 :
828 0 : Ok(Some(copied))
829 : }
830 0 : }
831 :
832 : /// Creates a new Layer instance for the adopted layer, and ensures it is found in the remote
833 : /// storage on successful return. without the adopted layer being added to `index_part.json`.
834 : /// Returns (Layer, did hardlink)
835 0 : async fn remote_copy(
836 0 : adopted: &Layer,
837 0 : adoptee: &Arc<Timeline>,
838 0 : generation: Generation,
839 0 : shard_identity: ShardIdentity,
840 0 : cancel: &CancellationToken,
841 0 : ) -> Result<(Layer, bool), Error> {
842 0 : let mut metadata = adopted.metadata();
843 0 : debug_assert!(metadata.generation <= generation);
844 0 : metadata.generation = generation;
845 0 : metadata.shard = shard_identity.shard_index();
846 0 :
847 0 : let conf = adoptee.conf;
848 0 : let file_name = adopted.layer_desc().layer_name();
849 :
850 : // We don't want to shut the timeline down during this operation because we do `delete_on_drop` below
851 0 : let _gate = adoptee.gate.enter().map_err(|e| match e {
852 0 : GateError::GateClosed => Error::ShuttingDown,
853 0 : })?;
854 :
855 : // depending if Layer::keep_resident, do a hardlink
856 : let did_hardlink;
857 0 : let owned = if let Some(adopted_resident) = adopted.keep_resident().await {
858 0 : let adopted_path = adopted_resident.local_path();
859 0 : let adoptee_path = local_layer_path(
860 0 : conf,
861 0 : &adoptee.tenant_shard_id,
862 0 : &adoptee.timeline_id,
863 0 : &file_name,
864 0 : &metadata.generation,
865 0 : );
866 0 :
867 0 : match std::fs::hard_link(adopted_path, &adoptee_path) {
868 0 : Ok(()) => {}
869 0 : Err(e) if e.kind() == std::io::ErrorKind::AlreadyExists => {
870 0 : // In theory we should not get into this situation as we are doing cleanups of the layer file after errors.
871 0 : // However, we don't do cleanups for errors past `prepare`, so there is the slight chance to get to this branch.
872 0 :
873 0 : // Double check that the file is orphan (probably from an earlier attempt), then delete it
874 0 : let key = file_name.clone().into();
875 0 : if adoptee
876 0 : .layers
877 0 : .read(LayerManagerLockHolder::DetachAncestor)
878 0 : .await
879 0 : .contains_key(&key)
880 : {
881 : // We are supposed to filter out such cases before coming to this function
882 0 : return Err(Error::Prepare(anyhow::anyhow!(
883 0 : "layer file {file_name} already present and inside layer map"
884 0 : )));
885 0 : }
886 0 : tracing::info!("Deleting orphan layer file to make way for hard linking");
887 : // Delete orphan layer file and try again, to ensure this layer has a well understood source
888 0 : std::fs::remove_file(adopted_path)
889 0 : .map_err(|e| Error::launder(e.into(), Error::Prepare))?;
890 0 : std::fs::hard_link(adopted_path, &adoptee_path)
891 0 : .map_err(|e| Error::launder(e.into(), Error::Prepare))?;
892 : }
893 0 : Err(e) => {
894 0 : return Err(Error::launder(e.into(), Error::Prepare));
895 : }
896 : };
897 0 : did_hardlink = true;
898 0 : Layer::for_resident(conf, adoptee, adoptee_path, file_name, metadata).drop_eviction_guard()
899 : } else {
900 0 : did_hardlink = false;
901 0 : Layer::for_evicted(conf, adoptee, file_name, metadata)
902 : };
903 :
904 0 : let layer = match adoptee
905 0 : .remote_client
906 0 : .copy_timeline_layer(adopted, &owned, cancel)
907 0 : .await
908 : {
909 0 : Ok(()) => owned,
910 0 : Err(e) => {
911 0 : {
912 0 : // Clean up the layer so that on a retry we don't get errors that the file already exists
913 0 : owned.delete_on_drop();
914 0 : std::mem::drop(owned);
915 0 : }
916 0 : return Err(Error::launder(e, Error::Prepare));
917 : }
918 : };
919 :
920 0 : Ok((layer, did_hardlink))
921 0 : }
922 :
923 : pub(crate) enum DetachingAndReparenting {
924 : /// All of the following timeline ids were reparented and the timeline ancestor detach must be
925 : /// marked as completed.
926 : Reparented(HashSet<TimelineId>),
927 :
928 : /// Some of the reparentings failed. The timeline ancestor detach must **not** be marked as
929 : /// completed.
930 : ///
931 : /// Nested `must_reset_tenant` is set to true when any restart requiring changes were made.
932 : SomeReparentingFailed { must_reset_tenant: bool },
933 :
934 : /// Detaching and reparentings were completed in a previous attempt. Timeline ancestor detach
935 : /// must be marked as completed.
936 : AlreadyDone(HashSet<TimelineId>),
937 : }
938 :
939 : impl DetachingAndReparenting {
940 0 : pub(crate) fn reset_tenant_required(&self) -> bool {
941 : use DetachingAndReparenting::*;
942 0 : match self {
943 0 : Reparented(_) => true,
944 0 : SomeReparentingFailed { must_reset_tenant } => *must_reset_tenant,
945 0 : AlreadyDone(_) => false,
946 : }
947 0 : }
948 :
949 0 : pub(crate) fn completed(self) -> Option<HashSet<TimelineId>> {
950 : use DetachingAndReparenting::*;
951 0 : match self {
952 0 : Reparented(x) | AlreadyDone(x) => Some(x),
953 0 : SomeReparentingFailed { .. } => None,
954 : }
955 0 : }
956 : }
957 :
958 : /// See [`Timeline::detach_from_ancestor_and_reparent`].
959 0 : pub(super) async fn detach_and_reparent(
960 0 : detached: &Arc<Timeline>,
961 0 : tenant: &TenantShard,
962 0 : prepared: PreparedTimelineDetach,
963 0 : ancestor_timeline_id: TimelineId,
964 0 : ancestor_lsn: Lsn,
965 0 : behavior: DetachBehavior,
966 0 : _ctx: &RequestContext,
967 0 : ) -> Result<DetachingAndReparenting, Error> {
968 0 : let PreparedTimelineDetach { layers } = prepared;
969 :
970 : #[derive(Debug)]
971 : enum Ancestor {
972 : NotDetached(Arc<Timeline>, Lsn),
973 : Detached(Arc<Timeline>, Lsn),
974 : }
975 :
976 0 : let (recorded_branchpoint, still_ongoing) = {
977 0 : let access = detached.remote_client.initialized_upload_queue()?;
978 0 : let latest = access.latest_uploaded_index_part();
979 0 :
980 0 : (
981 0 : latest.lineage.detached_previous_ancestor(),
982 0 : latest
983 0 : .gc_blocking
984 0 : .as_ref()
985 0 : .is_some_and(|b| b.blocked_by(DetachAncestor)),
986 0 : )
987 0 : };
988 0 : assert!(
989 0 : still_ongoing,
990 0 : "cannot (detach? reparent)? complete if the operation is not still ongoing"
991 : );
992 :
993 0 : let ancestor_to_detach = match detached.ancestor_timeline.as_ref() {
994 0 : Some(mut ancestor) => {
995 0 : while ancestor.timeline_id != ancestor_timeline_id {
996 0 : match ancestor.ancestor_timeline.as_ref() {
997 0 : Some(found) => {
998 0 : if ancestor_lsn != ancestor.ancestor_lsn {
999 0 : return Err(Error::DetachReparent(anyhow::anyhow!(
1000 0 : "cannot find the ancestor timeline to detach from: wrong ancestor lsn"
1001 0 : )));
1002 0 : }
1003 0 : ancestor = found;
1004 : }
1005 : None => {
1006 0 : return Err(Error::DetachReparent(anyhow::anyhow!(
1007 0 : "cannot find the ancestor timeline to detach from"
1008 0 : )));
1009 : }
1010 : }
1011 : }
1012 0 : Some(ancestor)
1013 : }
1014 0 : None => None,
1015 : };
1016 0 : let ancestor = match (ancestor_to_detach, recorded_branchpoint) {
1017 0 : (Some(ancestor), None) => {
1018 0 : assert!(
1019 0 : !layers.is_empty(),
1020 0 : "there should always be at least one layer to inherit"
1021 : );
1022 0 : Ancestor::NotDetached(ancestor.clone(), detached.ancestor_lsn)
1023 : }
1024 : (Some(_), Some(_)) => {
1025 0 : panic!(
1026 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"
1027 0 : );
1028 : }
1029 0 : (None, Some((ancestor_id, ancestor_lsn))) => {
1030 0 : // it has been either:
1031 0 : // - detached but still exists => we can try reparenting
1032 0 : // - detached and deleted
1033 0 : //
1034 0 : // either way, we must complete
1035 0 : assert!(
1036 0 : layers.is_empty(),
1037 0 : "no layers should had been copied as detach is done"
1038 : );
1039 :
1040 0 : let existing = tenant.timelines.lock().unwrap().get(&ancestor_id).cloned();
1041 :
1042 0 : if let Some(ancestor) = existing {
1043 0 : Ancestor::Detached(ancestor, ancestor_lsn)
1044 : } else {
1045 0 : let direct_children = reparented_direct_children(detached, tenant)?;
1046 0 : return Ok(DetachingAndReparenting::AlreadyDone(direct_children));
1047 : }
1048 : }
1049 : (None, None) => {
1050 : // TODO: make sure there are no `?` before tenant_reset from after a questionmark from
1051 : // here.
1052 0 : panic!(
1053 0 : "bug: detach_and_reparent called on a timeline which has not been detached or which has no live ancestor"
1054 0 : );
1055 : }
1056 : };
1057 :
1058 : // publish the prepared layers before we reparent any of the timelines, so that on restart
1059 : // reparented timelines find layers. also do the actual detaching.
1060 : //
1061 : // if we crash after this operation, a retry will allow reparenting the remaining timelines as
1062 : // gc is blocked.
1063 :
1064 0 : let (ancestor, ancestor_lsn, was_detached) = match ancestor {
1065 0 : Ancestor::NotDetached(ancestor, ancestor_lsn) => {
1066 0 : // this has to complete before any reparentings because otherwise they would not have
1067 0 : // layers on the new parent.
1068 0 : detached
1069 0 : .remote_client
1070 0 : .schedule_adding_existing_layers_to_index_detach_and_wait(
1071 0 : &layers,
1072 0 : (ancestor.timeline_id, ancestor_lsn),
1073 0 : )
1074 0 : .await
1075 0 : .context("publish layers and detach ancestor")
1076 0 : .map_err(|e| Error::launder(e, Error::DetachReparent))?;
1077 :
1078 0 : tracing::info!(
1079 0 : ancestor=%ancestor.timeline_id,
1080 0 : %ancestor_lsn,
1081 0 : inherited_layers=%layers.len(),
1082 0 : "detached from ancestor"
1083 : );
1084 0 : (ancestor, ancestor_lsn, true)
1085 : }
1086 0 : Ancestor::Detached(ancestor, ancestor_lsn) => (ancestor, ancestor_lsn, false),
1087 : };
1088 :
1089 0 : if let DetachBehavior::MultiLevelAndNoReparent = behavior {
1090 : // Do not reparent if the user requests to behave so.
1091 0 : return Ok(DetachingAndReparenting::Reparented(HashSet::new()));
1092 0 : }
1093 0 :
1094 0 : let mut tasks = tokio::task::JoinSet::new();
1095 0 :
1096 0 : // Returns a single permit semaphore which will be used to make one reparenting succeed,
1097 0 : // others will fail as if those timelines had been stopped for whatever reason.
1098 0 : #[cfg(feature = "testing")]
1099 0 : let failpoint_sem = || -> Option<Arc<Semaphore>> {
1100 0 : fail::fail_point!("timeline-detach-ancestor::allow_one_reparented", |_| Some(
1101 0 : Arc::new(Semaphore::new(1))
1102 0 : ));
1103 0 : None
1104 0 : }();
1105 0 :
1106 0 : // because we are now keeping the slot in progress, it is unlikely that there will be any
1107 0 : // timeline deletions during this time. if we raced one, then we'll just ignore it.
1108 0 : {
1109 0 : let g = tenant.timelines.lock().unwrap();
1110 0 : reparentable_timelines(g.values(), detached, &ancestor, ancestor_lsn)
1111 0 : .cloned()
1112 0 : .for_each(|timeline| {
1113 : // important in this scope: we are holding the Tenant::timelines lock
1114 0 : let span = tracing::info_span!("reparent", reparented=%timeline.timeline_id);
1115 0 : let new_parent = detached.timeline_id;
1116 0 : #[cfg(feature = "testing")]
1117 0 : let failpoint_sem = failpoint_sem.clone();
1118 0 :
1119 0 : tasks.spawn(
1120 0 : async move {
1121 0 : let res = async {
1122 : #[cfg(feature = "testing")]
1123 0 : if let Some(failpoint_sem) = failpoint_sem {
1124 0 : let _permit = failpoint_sem.acquire().await.map_err(|_| {
1125 0 : anyhow::anyhow!(
1126 0 : "failpoint: timeline-detach-ancestor::allow_one_reparented",
1127 0 : )
1128 0 : })?;
1129 0 : failpoint_sem.close();
1130 0 : }
1131 :
1132 0 : timeline
1133 0 : .remote_client
1134 0 : .schedule_reparenting_and_wait(&new_parent)
1135 0 : .await
1136 0 : }
1137 0 : .await;
1138 :
1139 0 : match res {
1140 : Ok(()) => {
1141 0 : tracing::info!("reparented");
1142 0 : Some(timeline)
1143 : }
1144 0 : Err(e) => {
1145 0 : // with the use of tenant slot, raced timeline deletion is the most
1146 0 : // likely reason.
1147 0 : tracing::warn!("reparenting failed: {e:#}");
1148 0 : None
1149 : }
1150 : }
1151 0 : }
1152 0 : .instrument(span),
1153 0 : );
1154 0 : });
1155 0 : }
1156 0 :
1157 0 : let reparenting_candidates = tasks.len();
1158 0 : let mut reparented = HashSet::with_capacity(tasks.len());
1159 :
1160 0 : while let Some(res) = tasks.join_next().await {
1161 0 : match res {
1162 0 : Ok(Some(timeline)) => {
1163 0 : assert!(
1164 0 : reparented.insert(timeline.timeline_id),
1165 0 : "duplicate reparenting? timeline_id={}",
1166 0 : timeline.timeline_id
1167 : );
1168 : }
1169 0 : Err(je) if je.is_cancelled() => unreachable!("not used"),
1170 : // just ignore failures now, we can retry
1171 0 : Ok(None) => {}
1172 0 : Err(je) if je.is_panic() => {}
1173 0 : Err(je) => tracing::error!("unexpected join error: {je:?}"),
1174 : }
1175 : }
1176 :
1177 0 : let reparented_all = reparenting_candidates == reparented.len();
1178 0 :
1179 0 : if reparented_all {
1180 0 : Ok(DetachingAndReparenting::Reparented(reparented))
1181 : } else {
1182 0 : tracing::info!(
1183 0 : reparented = reparented.len(),
1184 0 : candidates = reparenting_candidates,
1185 0 : "failed to reparent all candidates; they can be retried after the tenant_reset",
1186 : );
1187 :
1188 0 : let must_reset_tenant = !reparented.is_empty() || was_detached;
1189 0 : Ok(DetachingAndReparenting::SomeReparentingFailed { must_reset_tenant })
1190 : }
1191 0 : }
1192 :
1193 0 : pub(super) async fn complete(
1194 0 : detached: &Arc<Timeline>,
1195 0 : tenant: &TenantShard,
1196 0 : mut attempt: Attempt,
1197 0 : _ctx: &RequestContext,
1198 0 : ) -> Result<(), Error> {
1199 0 : assert_eq!(detached.timeline_id, attempt.timeline_id);
1200 :
1201 0 : if attempt.gate_entered.is_none() {
1202 0 : let entered = detached.gate.enter().map_err(|_| Error::ShuttingDown)?;
1203 0 : attempt.gate_entered = Some(entered);
1204 0 : } else {
1205 0 : // Some(gate_entered) means the tenant was not restarted, as is not required
1206 0 : }
1207 :
1208 0 : assert!(detached.ancestor_timeline.is_none());
1209 :
1210 : // this should be an 503 at least...?
1211 0 : fail::fail_point!(
1212 0 : "timeline-detach-ancestor::complete_before_uploading",
1213 0 : |_| Err(Error::Failpoint(
1214 0 : "timeline-detach-ancestor::complete_before_uploading"
1215 0 : ))
1216 0 : );
1217 :
1218 0 : tenant
1219 0 : .gc_block
1220 0 : .remove(
1221 0 : detached,
1222 0 : crate::tenant::remote_timeline_client::index::GcBlockingReason::DetachAncestor,
1223 0 : )
1224 0 : .await
1225 0 : .map_err(|e| Error::launder(e, Error::Complete))?;
1226 :
1227 0 : Ok(())
1228 0 : }
1229 :
1230 : /// Query against a locked `Tenant::timelines`.
1231 : ///
1232 : /// A timeline is reparentable if:
1233 : ///
1234 : /// - It is not the timeline being detached.
1235 : /// - It has the same ancestor as the timeline being detached. Note that the ancestor might not be the direct ancestor.
1236 0 : fn reparentable_timelines<'a, I>(
1237 0 : timelines: I,
1238 0 : detached: &'a Arc<Timeline>,
1239 0 : ancestor: &'a Arc<Timeline>,
1240 0 : ancestor_lsn: Lsn,
1241 0 : ) -> impl Iterator<Item = &'a Arc<Timeline>> + 'a
1242 0 : where
1243 0 : I: Iterator<Item = &'a Arc<Timeline>> + 'a,
1244 0 : {
1245 0 : timelines.filter_map(move |tl| {
1246 0 : if Arc::ptr_eq(tl, detached) {
1247 0 : return None;
1248 0 : }
1249 :
1250 0 : let tl_ancestor = tl.ancestor_timeline.as_ref()?;
1251 0 : let is_same = Arc::ptr_eq(ancestor, tl_ancestor);
1252 0 : let is_earlier = tl.get_ancestor_lsn() <= ancestor_lsn;
1253 0 :
1254 0 : let is_deleting = tl
1255 0 : .delete_progress
1256 0 : .try_lock()
1257 0 : .map(|flow| !flow.is_not_started())
1258 0 : .unwrap_or(true);
1259 0 :
1260 0 : if is_same && is_earlier && !is_deleting {
1261 0 : Some(tl)
1262 : } else {
1263 0 : None
1264 : }
1265 0 : })
1266 0 : }
1267 :
1268 0 : fn check_no_archived_children_of_ancestor(
1269 0 : tenant: &TenantShard,
1270 0 : detached: &Arc<Timeline>,
1271 0 : ancestor: &Arc<Timeline>,
1272 0 : ancestor_lsn: Lsn,
1273 0 : detach_behavior: DetachBehavior,
1274 0 : ) -> Result<(), Error> {
1275 0 : match detach_behavior {
1276 : DetachBehavior::NoAncestorAndReparent => {
1277 0 : let timelines = tenant.timelines.lock().unwrap();
1278 0 : let timelines_offloaded = tenant.timelines_offloaded.lock().unwrap();
1279 :
1280 0 : for timeline in
1281 0 : reparentable_timelines(timelines.values(), detached, ancestor, ancestor_lsn)
1282 : {
1283 0 : if timeline.is_archived() == Some(true) {
1284 0 : return Err(Error::Archived(timeline.timeline_id));
1285 0 : }
1286 : }
1287 :
1288 0 : for timeline_offloaded in timelines_offloaded.values() {
1289 0 : if timeline_offloaded.ancestor_timeline_id != Some(ancestor.timeline_id) {
1290 0 : continue;
1291 0 : }
1292 : // This forbids the detach ancestor feature if flattened timelines are present,
1293 : // even if the ancestor_lsn is from after the branchpoint of the detached timeline.
1294 : // But as per current design, we don't record the ancestor_lsn of flattened timelines.
1295 : // This is a bit unfortunate, but as of writing this we don't support flattening
1296 : // anyway. Maybe we can evolve the data model in the future.
1297 0 : if let Some(retain_lsn) = timeline_offloaded.ancestor_retain_lsn {
1298 0 : let is_earlier = retain_lsn <= ancestor_lsn;
1299 0 : if !is_earlier {
1300 0 : continue;
1301 0 : }
1302 0 : }
1303 0 : return Err(Error::Archived(timeline_offloaded.timeline_id));
1304 : }
1305 : }
1306 0 : DetachBehavior::MultiLevelAndNoReparent => {
1307 0 : // We don't need to check anything if the user requested to not reparent.
1308 0 : }
1309 : }
1310 :
1311 0 : Ok(())
1312 0 : }
1313 :
1314 0 : async fn fsync_timeline_dir(timeline: &Timeline, ctx: &RequestContext) {
1315 0 : let path = &timeline
1316 0 : .conf
1317 0 : .timeline_path(&timeline.tenant_shard_id, &timeline.timeline_id);
1318 0 : let timeline_dir = VirtualFile::open(&path, ctx)
1319 0 : .await
1320 0 : .fatal_err("VirtualFile::open for timeline dir fsync");
1321 0 : timeline_dir
1322 0 : .sync_all()
1323 0 : .await
1324 0 : .fatal_err("VirtualFile::sync_all timeline dir");
1325 0 : }
|
