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