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