Line data Source code
1 : use anyhow::Context;
2 : use camino::{Utf8Path, Utf8PathBuf};
3 : use pageserver_api::keyspace::KeySpace;
4 : use pageserver_api::models::{
5 : HistoricLayerInfo, LayerAccessKind, LayerResidenceEventReason, LayerResidenceStatus,
6 : };
7 : use pageserver_api::shard::{ShardIdentity, ShardIndex, TenantShardId};
8 : use std::ops::Range;
9 : use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
10 : use std::sync::{Arc, Weak};
11 : use std::time::{Duration, SystemTime};
12 : use tracing::Instrument;
13 : use utils::id::TimelineId;
14 : use utils::lsn::Lsn;
15 : use utils::sync::{gate, heavier_once_cell};
16 :
17 : use crate::config::PageServerConf;
18 : use crate::context::{DownloadBehavior, RequestContext};
19 : use crate::repository::Key;
20 : use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
21 : use crate::task_mgr::TaskKind;
22 : use crate::tenant::timeline::GetVectoredError;
23 : use crate::tenant::{remote_timeline_client::LayerFileMetadata, Timeline};
24 :
25 : use super::delta_layer::{self, DeltaEntry};
26 : use super::image_layer::{self};
27 : use super::{
28 : AsLayerDesc, ImageLayerWriter, LayerAccessStats, LayerAccessStatsReset, LayerName,
29 : PersistentLayerDesc, ValueReconstructResult, ValueReconstructState, ValuesReconstructState,
30 : };
31 :
32 : use utils::generation::Generation;
33 :
34 : #[cfg(test)]
35 : mod tests;
36 :
37 : #[cfg(test)]
38 : mod failpoints;
39 :
40 : /// A Layer contains all data in a "rectangle" consisting of a range of keys and
41 : /// range of LSNs.
42 : ///
43 : /// There are two kinds of layers, in-memory and on-disk layers. In-memory
44 : /// layers are used to ingest incoming WAL, and provide fast access to the
45 : /// recent page versions. On-disk layers are stored as files on disk, and are
46 : /// immutable. This type represents the on-disk kind while in-memory kind are represented by
47 : /// [`InMemoryLayer`].
48 : ///
49 : /// Furthermore, there are two kinds of on-disk layers: delta and image layers.
50 : /// A delta layer contains all modifications within a range of LSNs and keys.
51 : /// An image layer is a snapshot of all the data in a key-range, at a single
52 : /// LSN.
53 : ///
54 : /// This type models the on-disk layers, which can be evicted and on-demand downloaded. As a
55 : /// general goal, read accesses should always win eviction and eviction should not wait for
56 : /// download.
57 : ///
58 : /// ### State transitions
59 : ///
60 : /// The internal state of `Layer` is composed of most importantly the on-filesystem state and the
61 : /// [`ResidentOrWantedEvicted`] enum. On-filesystem state can be either present (fully downloaded,
62 : /// right size) or deleted.
63 : ///
64 : /// Reads will always win requests to evict until `wait_for_turn_and_evict` has acquired the
65 : /// `heavier_once_cell::InitPermit` and has started to `evict_blocking`. Before the
66 : /// `heavier_once_cell::InitPermit` has been acquired, any read request
67 : /// (`get_or_maybe_download`) can "re-initialize" using the existing downloaded file and thus
68 : /// cancelling the eviction.
69 : ///
70 : /// ```text
71 : /// +-----------------+ get_or_maybe_download +--------------------------------+
72 : /// | not initialized |--------------------------->| Resident(Arc<DownloadedLayer>) |
73 : /// | ENOENT | /->| |
74 : /// +-----------------+ | +--------------------------------+
75 : /// ^ | | ^
76 : /// | get_or_maybe_download | | | get_or_maybe_download, either:
77 : /// evict_blocking | /-------------------------/ | | - upgrade weak to strong
78 : /// | | | | - re-initialize without download
79 : /// | | evict_and_wait | |
80 : /// +-----------------+ v |
81 : /// | not initialized | on_downloaded_layer_drop +--------------------------------------+
82 : /// | file is present |<---------------------------| WantedEvicted(Weak<DownloadedLayer>) |
83 : /// +-----------------+ +--------------------------------------+
84 : /// ```
85 : ///
86 : /// ### Unsupported
87 : ///
88 : /// - Evicting by the operator deleting files from the filesystem
89 : ///
90 : /// [`InMemoryLayer`]: super::inmemory_layer::InMemoryLayer
91 : #[derive(Clone)]
92 : pub(crate) struct Layer(Arc<LayerInner>);
93 :
94 : impl std::fmt::Display for Layer {
95 1878 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
96 1878 : if matches!(self.0.generation, Generation::Broken) {
97 0 : write!(f, "{}-broken", self.layer_desc().short_id())
98 : } else {
99 1878 : write!(
100 1878 : f,
101 1878 : "{}{}",
102 1878 : self.layer_desc().short_id(),
103 1878 : self.0.generation.get_suffix()
104 1878 : )
105 : }
106 1878 : }
107 : }
108 :
109 : impl std::fmt::Debug for Layer {
110 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
111 0 : write!(f, "{}", self)
112 0 : }
113 : }
114 :
115 : impl AsLayerDesc for Layer {
116 720090 : fn layer_desc(&self) -> &PersistentLayerDesc {
117 720090 : self.0.layer_desc()
118 720090 : }
119 : }
120 :
121 : impl PartialEq for Layer {
122 4 : fn eq(&self, other: &Self) -> bool {
123 4 : Arc::as_ptr(&self.0) == Arc::as_ptr(&other.0)
124 4 : }
125 : }
126 :
127 1570 : pub(crate) fn local_layer_path(
128 1570 : conf: &PageServerConf,
129 1570 : tenant_shard_id: &TenantShardId,
130 1570 : timeline_id: &TimelineId,
131 1570 : layer_file_name: &LayerName,
132 1570 : generation: &Generation,
133 1570 : ) -> Utf8PathBuf {
134 1570 : let timeline_path = conf.timeline_path(tenant_shard_id, timeline_id);
135 1570 :
136 1570 : if generation.is_none() {
137 : // Without a generation, we may only use legacy path style
138 0 : timeline_path.join(layer_file_name.to_string())
139 : } else {
140 1570 : timeline_path.join(format!("{}-v1{}", layer_file_name, generation.get_suffix()))
141 : }
142 1570 : }
143 :
144 : impl Layer {
145 : /// Creates a layer value for a file we know to not be resident.
146 0 : pub(crate) fn for_evicted(
147 0 : conf: &'static PageServerConf,
148 0 : timeline: &Arc<Timeline>,
149 0 : file_name: LayerName,
150 0 : metadata: LayerFileMetadata,
151 0 : ) -> Self {
152 0 : let local_path = local_layer_path(
153 0 : conf,
154 0 : &timeline.tenant_shard_id,
155 0 : &timeline.timeline_id,
156 0 : &file_name,
157 0 : &metadata.generation,
158 0 : );
159 0 :
160 0 : let desc = PersistentLayerDesc::from_filename(
161 0 : timeline.tenant_shard_id,
162 0 : timeline.timeline_id,
163 0 : file_name,
164 0 : metadata.file_size,
165 0 : );
166 0 :
167 0 : let access_stats = LayerAccessStats::for_loading_layer(LayerResidenceStatus::Evicted);
168 0 :
169 0 : let owner = Layer(Arc::new(LayerInner::new(
170 0 : conf,
171 0 : timeline,
172 0 : local_path,
173 0 : access_stats,
174 0 : desc,
175 0 : None,
176 0 : metadata.generation,
177 0 : metadata.shard,
178 0 : )));
179 0 :
180 0 : debug_assert!(owner.0.needs_download_blocking().unwrap().is_some());
181 :
182 0 : owner
183 0 : }
184 :
185 : /// Creates a Layer value for a file we know to be resident in timeline directory.
186 24 : pub(crate) fn for_resident(
187 24 : conf: &'static PageServerConf,
188 24 : timeline: &Arc<Timeline>,
189 24 : local_path: Utf8PathBuf,
190 24 : file_name: LayerName,
191 24 : metadata: LayerFileMetadata,
192 24 : ) -> ResidentLayer {
193 24 : let desc = PersistentLayerDesc::from_filename(
194 24 : timeline.tenant_shard_id,
195 24 : timeline.timeline_id,
196 24 : file_name,
197 24 : metadata.file_size,
198 24 : );
199 24 :
200 24 : let access_stats = LayerAccessStats::for_loading_layer(LayerResidenceStatus::Resident);
201 24 :
202 24 : let mut resident = None;
203 24 :
204 24 : let owner = Layer(Arc::new_cyclic(|owner| {
205 24 : let inner = Arc::new(DownloadedLayer {
206 24 : owner: owner.clone(),
207 24 : kind: tokio::sync::OnceCell::default(),
208 24 : version: 0,
209 24 : });
210 24 : resident = Some(inner.clone());
211 24 :
212 24 : LayerInner::new(
213 24 : conf,
214 24 : timeline,
215 24 : local_path,
216 24 : access_stats,
217 24 : desc,
218 24 : Some(inner),
219 24 : metadata.generation,
220 24 : metadata.shard,
221 24 : )
222 24 : }));
223 24 :
224 24 : let downloaded = resident.expect("just initialized");
225 24 :
226 24 : debug_assert!(owner.0.needs_download_blocking().unwrap().is_none());
227 :
228 24 : timeline
229 24 : .metrics
230 24 : .resident_physical_size_add(metadata.file_size);
231 24 :
232 24 : ResidentLayer { downloaded, owner }
233 24 : }
234 :
235 : /// Creates a Layer value for freshly written out new layer file by renaming it from a
236 : /// temporary path.
237 1562 : pub(crate) fn finish_creating(
238 1562 : conf: &'static PageServerConf,
239 1562 : timeline: &Arc<Timeline>,
240 1562 : desc: PersistentLayerDesc,
241 1562 : temp_path: &Utf8Path,
242 1562 : ) -> anyhow::Result<ResidentLayer> {
243 1562 : let mut resident = None;
244 1562 :
245 1562 : let owner = Layer(Arc::new_cyclic(|owner| {
246 1562 : let inner = Arc::new(DownloadedLayer {
247 1562 : owner: owner.clone(),
248 1562 : kind: tokio::sync::OnceCell::default(),
249 1562 : version: 0,
250 1562 : });
251 1562 : resident = Some(inner.clone());
252 1562 : let access_stats = LayerAccessStats::empty_will_record_residence_event_later();
253 1562 : access_stats.record_residence_event(
254 1562 : LayerResidenceStatus::Resident,
255 1562 : LayerResidenceEventReason::LayerCreate,
256 1562 : );
257 1562 :
258 1562 : let local_path = local_layer_path(
259 1562 : conf,
260 1562 : &timeline.tenant_shard_id,
261 1562 : &timeline.timeline_id,
262 1562 : &desc.layer_name(),
263 1562 : &timeline.generation,
264 1562 : );
265 1562 :
266 1562 : LayerInner::new(
267 1562 : conf,
268 1562 : timeline,
269 1562 : local_path,
270 1562 : access_stats,
271 1562 : desc,
272 1562 : Some(inner),
273 1562 : timeline.generation,
274 1562 : timeline.get_shard_index(),
275 1562 : )
276 1562 : }));
277 1562 :
278 1562 : let downloaded = resident.expect("just initialized");
279 1562 :
280 1562 : // We never want to overwrite an existing file, so we use `RENAME_NOREPLACE`.
281 1562 : // TODO: this leaves the temp file in place if the rename fails, risking us running
282 1562 : // out of space. Should we clean it up here or does the calling context deal with this?
283 1562 : utils::fs_ext::rename_noreplace(temp_path.as_std_path(), owner.local_path().as_std_path())
284 1562 : .with_context(|| format!("rename temporary file as correct path for {owner}"))?;
285 :
286 1562 : Ok(ResidentLayer { downloaded, owner })
287 1562 : }
288 :
289 : /// Requests the layer to be evicted and waits for this to be done.
290 : ///
291 : /// If the file is not resident, an [`EvictionError::NotFound`] is returned.
292 : ///
293 : /// If for a bad luck or blocking of the executor, we miss the actual eviction and the layer is
294 : /// re-downloaded, [`EvictionError::Downloaded`] is returned.
295 : ///
296 : /// Timeout is mandatory, because waiting for eviction is only needed for our tests; eviction
297 : /// will happen regardless the future returned by this method completing unless there is a
298 : /// read access before eviction gets to complete.
299 : ///
300 : /// Technically cancellation safe, but cancelling might shift the viewpoint of what generation
301 : /// of download-evict cycle on retry.
302 36 : pub(crate) async fn evict_and_wait(&self, timeout: Duration) -> Result<(), EvictionError> {
303 48 : self.0.evict_and_wait(timeout).await
304 32 : }
305 :
306 : /// Delete the layer file when the `self` gets dropped, also try to schedule a remote index upload
307 : /// then.
308 : ///
309 : /// On drop, this will cause a call to [`crate::tenant::remote_timeline_client::RemoteTimelineClient::schedule_deletion_of_unlinked`].
310 : /// This means that the unlinking by [gc] or [compaction] must have happened strictly before
311 : /// the value this is called on gets dropped.
312 : ///
313 : /// This is ensured by both of those methods accepting references to Layer.
314 : ///
315 : /// [gc]: [`RemoteTimelineClient::schedule_gc_update`]
316 : /// [compaction]: [`RemoteTimelineClient::schedule_compaction_update`]
317 438 : pub(crate) fn delete_on_drop(&self) {
318 438 : self.0.delete_on_drop();
319 438 : }
320 :
321 : /// Return data needed to reconstruct given page at LSN.
322 : ///
323 : /// It is up to the caller to collect more data from the previous layer and
324 : /// perform WAL redo, if necessary.
325 : ///
326 : /// # Cancellation-Safety
327 : ///
328 : /// This method is cancellation-safe.
329 211782 : pub(crate) async fn get_value_reconstruct_data(
330 211782 : &self,
331 211782 : key: Key,
332 211782 : lsn_range: Range<Lsn>,
333 211782 : reconstruct_data: &mut ValueReconstructState,
334 211782 : ctx: &RequestContext,
335 211782 : ) -> anyhow::Result<ValueReconstructResult> {
336 : use anyhow::ensure;
337 :
338 211782 : let layer = self.0.get_or_maybe_download(true, Some(ctx)).await?;
339 211782 : self.0
340 211782 : .access_stats
341 211782 : .record_access(LayerAccessKind::GetValueReconstructData, ctx);
342 211782 :
343 211782 : if self.layer_desc().is_delta {
344 204764 : ensure!(lsn_range.start >= self.layer_desc().lsn_range.start);
345 204764 : ensure!(self.layer_desc().key_range.contains(&key));
346 : } else {
347 7018 : ensure!(self.layer_desc().key_range.contains(&key));
348 7018 : ensure!(lsn_range.start >= self.layer_desc().image_layer_lsn());
349 7018 : ensure!(lsn_range.end >= self.layer_desc().image_layer_lsn());
350 : }
351 :
352 211782 : layer
353 211782 : .get_value_reconstruct_data(key, lsn_range, reconstruct_data, &self.0, ctx)
354 211782 : .instrument(tracing::debug_span!("get_value_reconstruct_data", layer=%self))
355 30535 : .await
356 211782 : .with_context(|| format!("get_value_reconstruct_data for layer {self}"))
357 211782 : }
358 :
359 228 : pub(crate) async fn get_values_reconstruct_data(
360 228 : &self,
361 228 : keyspace: KeySpace,
362 228 : lsn_range: Range<Lsn>,
363 228 : reconstruct_data: &mut ValuesReconstructState,
364 228 : ctx: &RequestContext,
365 228 : ) -> Result<(), GetVectoredError> {
366 228 : let layer = self
367 228 : .0
368 228 : .get_or_maybe_download(true, Some(ctx))
369 0 : .await
370 228 : .map_err(|err| match err {
371 0 : DownloadError::DownloadCancelled => GetVectoredError::Cancelled,
372 0 : other => GetVectoredError::Other(anyhow::anyhow!(other)),
373 228 : })?;
374 :
375 228 : self.0
376 228 : .access_stats
377 228 : .record_access(LayerAccessKind::GetValueReconstructData, ctx);
378 228 :
379 228 : layer
380 228 : .get_values_reconstruct_data(keyspace, lsn_range, reconstruct_data, &self.0, ctx)
381 228 : .instrument(tracing::debug_span!("get_values_reconstruct_data", layer=%self))
382 11426 : .await
383 228 : .map_err(|err| match err {
384 0 : GetVectoredError::Other(err) => GetVectoredError::Other(
385 0 : err.context(format!("get_values_reconstruct_data for layer {self}")),
386 0 : ),
387 0 : err => err,
388 228 : })
389 228 : }
390 :
391 : /// Get all key/values in the layer. Should be replaced with an iterator-based API in the future.
392 16 : pub(crate) async fn load_key_values(
393 16 : &self,
394 16 : ctx: &RequestContext,
395 16 : ) -> anyhow::Result<Vec<(Key, Lsn, crate::repository::Value)>> {
396 16 : let layer = self
397 16 : .0
398 16 : .get_or_maybe_download(true, Some(ctx))
399 0 : .await
400 16 : .map_err(|err| match err {
401 0 : DownloadError::DownloadCancelled => GetVectoredError::Cancelled,
402 0 : other => GetVectoredError::Other(anyhow::anyhow!(other)),
403 16 : })?;
404 16 : layer.load_key_values(&self.0, ctx).await
405 16 : }
406 :
407 : /// Download the layer if evicted.
408 : ///
409 : /// Will not error when the layer is already downloaded.
410 0 : pub(crate) async fn download(&self) -> anyhow::Result<()> {
411 0 : self.0.get_or_maybe_download(true, None).await?;
412 0 : Ok(())
413 0 : }
414 :
415 : /// Assuming the layer is already downloaded, returns a guard which will prohibit eviction
416 : /// while the guard exists.
417 : ///
418 : /// Returns None if the layer is currently evicted or becoming evicted.
419 : #[cfg(test)]
420 20 : pub(crate) async fn keep_resident(&self) -> Option<ResidentLayer> {
421 20 : let downloaded = self.0.inner.get().and_then(|rowe| rowe.get())?;
422 :
423 14 : Some(ResidentLayer {
424 14 : downloaded,
425 14 : owner: self.clone(),
426 14 : })
427 20 : }
428 :
429 : /// Weak indicator of is the layer resident or not. Good enough for eviction, which can deal
430 : /// with `EvictionError::NotFound`.
431 : ///
432 : /// Returns `true` if this layer might be resident, or `false`, if it most likely evicted or
433 : /// will be unless a read happens soon.
434 44 : pub(crate) fn is_likely_resident(&self) -> bool {
435 44 : self.0
436 44 : .inner
437 44 : .get()
438 44 : .map(|rowe| rowe.is_likely_resident())
439 44 : .unwrap_or(false)
440 44 : }
441 :
442 : /// Downloads if necessary and creates a guard, which will keep this layer from being evicted.
443 414 : pub(crate) async fn download_and_keep_resident(&self) -> anyhow::Result<ResidentLayer> {
444 414 : let downloaded = self.0.get_or_maybe_download(true, None).await?;
445 :
446 414 : Ok(ResidentLayer {
447 414 : downloaded,
448 414 : owner: self.clone(),
449 414 : })
450 414 : }
451 :
452 0 : pub(crate) fn info(&self, reset: LayerAccessStatsReset) -> HistoricLayerInfo {
453 0 : self.0.info(reset)
454 0 : }
455 :
456 0 : pub(crate) fn access_stats(&self) -> &LayerAccessStats {
457 0 : &self.0.access_stats
458 0 : }
459 :
460 1564 : pub(crate) fn local_path(&self) -> &Utf8Path {
461 1564 : &self.0.path
462 1564 : }
463 :
464 211778 : pub(crate) fn debug_str(&self) -> &Arc<str> {
465 211778 : &self.0.debug_str
466 211778 : }
467 :
468 1496 : pub(crate) fn metadata(&self) -> LayerFileMetadata {
469 1496 : self.0.metadata()
470 1496 : }
471 :
472 0 : pub(crate) fn get_timeline_id(&self) -> Option<TimelineId> {
473 0 : self.0
474 0 : .timeline
475 0 : .upgrade()
476 0 : .map(|timeline| timeline.timeline_id)
477 0 : }
478 :
479 : /// Traditional debug dumping facility
480 : #[allow(unused)]
481 4 : pub(crate) async fn dump(&self, verbose: bool, ctx: &RequestContext) -> anyhow::Result<()> {
482 4 : self.0.desc.dump();
483 4 :
484 4 : if verbose {
485 : // for now, unconditionally download everything, even if that might not be wanted.
486 4 : let l = self.0.get_or_maybe_download(true, Some(ctx)).await?;
487 8 : l.dump(&self.0, ctx).await?
488 0 : }
489 :
490 4 : Ok(())
491 4 : }
492 :
493 : /// Waits until this layer has been dropped (and if needed, local file deletion and remote
494 : /// deletion scheduling has completed).
495 : ///
496 : /// Does not start local deletion, use [`Self::delete_on_drop`] for that
497 : /// separatedly.
498 : #[cfg(any(feature = "testing", test))]
499 2 : pub(crate) fn wait_drop(&self) -> impl std::future::Future<Output = ()> + 'static {
500 2 : let mut rx = self.0.status.as_ref().unwrap().subscribe();
501 :
502 2 : async move {
503 : loop {
504 6 : if rx.changed().await.is_err() {
505 2 : break;
506 0 : }
507 : }
508 2 : }
509 2 : }
510 : }
511 :
512 : /// The download-ness ([`DownloadedLayer`]) can be either resident or wanted evicted.
513 : ///
514 : /// However when we want something evicted, we cannot evict it right away as there might be current
515 : /// reads happening on it. For example: it has been searched from [`LayerMap::search`] but not yet
516 : /// read with [`Layer::get_value_reconstruct_data`].
517 : ///
518 : /// [`LayerMap::search`]: crate::tenant::layer_map::LayerMap::search
519 : #[derive(Debug)]
520 : enum ResidentOrWantedEvicted {
521 : Resident(Arc<DownloadedLayer>),
522 : WantedEvicted(Weak<DownloadedLayer>, usize),
523 : }
524 :
525 : impl ResidentOrWantedEvicted {
526 : /// Non-mutating access to the a DownloadedLayer, if possible.
527 : ///
528 : /// This is not used on the read path (anything that calls
529 : /// [`LayerInner::get_or_maybe_download`]) because it was decided that reads always win
530 : /// evictions, and part of that winning is using [`ResidentOrWantedEvicted::get_and_upgrade`].
531 : #[cfg(test)]
532 14 : fn get(&self) -> Option<Arc<DownloadedLayer>> {
533 14 : match self {
534 14 : ResidentOrWantedEvicted::Resident(strong) => Some(strong.clone()),
535 0 : ResidentOrWantedEvicted::WantedEvicted(weak, _) => weak.upgrade(),
536 : }
537 14 : }
538 :
539 : /// Best-effort query for residency right now, not as strong guarantee as receiving a strong
540 : /// reference from `ResidentOrWantedEvicted::get`.
541 38 : fn is_likely_resident(&self) -> bool {
542 38 : match self {
543 32 : ResidentOrWantedEvicted::Resident(_) => true,
544 6 : ResidentOrWantedEvicted::WantedEvicted(weak, _) => weak.strong_count() > 0,
545 : }
546 38 : }
547 :
548 : /// Upgrades any weak to strong if possible.
549 : ///
550 : /// Returns a strong reference if possible, along with a boolean telling if an upgrade
551 : /// happened.
552 212448 : fn get_and_upgrade(&mut self) -> Option<(Arc<DownloadedLayer>, bool)> {
553 212448 : match self {
554 212440 : ResidentOrWantedEvicted::Resident(strong) => Some((strong.clone(), false)),
555 8 : ResidentOrWantedEvicted::WantedEvicted(weak, _) => match weak.upgrade() {
556 0 : Some(strong) => {
557 0 : LAYER_IMPL_METRICS.inc_raced_wanted_evicted_accesses();
558 0 :
559 0 : *self = ResidentOrWantedEvicted::Resident(strong.clone());
560 0 :
561 0 : Some((strong, true))
562 : }
563 8 : None => None,
564 : },
565 : }
566 212448 : }
567 :
568 : /// When eviction is first requested, drop down to holding a [`Weak`].
569 : ///
570 : /// Returns `Some` if this was the first time eviction was requested. Care should be taken to
571 : /// drop the possibly last strong reference outside of the mutex of
572 : /// [`heavier_once_cell::OnceCell`].
573 30 : fn downgrade(&mut self) -> Option<Arc<DownloadedLayer>> {
574 30 : match self {
575 26 : ResidentOrWantedEvicted::Resident(strong) => {
576 26 : let weak = Arc::downgrade(strong);
577 26 : let mut temp = ResidentOrWantedEvicted::WantedEvicted(weak, strong.version);
578 26 : std::mem::swap(self, &mut temp);
579 26 : match temp {
580 26 : ResidentOrWantedEvicted::Resident(strong) => Some(strong),
581 0 : ResidentOrWantedEvicted::WantedEvicted(..) => unreachable!("just swapped"),
582 : }
583 : }
584 4 : ResidentOrWantedEvicted::WantedEvicted(..) => None,
585 : }
586 30 : }
587 : }
588 :
589 : struct LayerInner {
590 : /// Only needed to check ondemand_download_behavior_treat_error_as_warn and creation of
591 : /// [`Self::path`].
592 : conf: &'static PageServerConf,
593 :
594 : /// Full path to the file; unclear if this should exist anymore.
595 : path: Utf8PathBuf,
596 :
597 : /// String representation of the layer, used for traversal id.
598 : debug_str: Arc<str>,
599 :
600 : desc: PersistentLayerDesc,
601 :
602 : /// Timeline access is needed for remote timeline client and metrics.
603 : ///
604 : /// There should not be an access to timeline for any reason without entering the
605 : /// [`Timeline::gate`] at the same time.
606 : timeline: Weak<Timeline>,
607 :
608 : access_stats: LayerAccessStats,
609 :
610 : /// This custom OnceCell is backed by std mutex, but only held for short time periods.
611 : ///
612 : /// Filesystem changes (download, evict) are only done while holding a permit which the
613 : /// `heavier_once_cell` provides.
614 : ///
615 : /// A number of fields in `Layer` are meant to only be updated when holding the InitPermit, but
616 : /// possibly read while not holding it.
617 : inner: heavier_once_cell::OnceCell<ResidentOrWantedEvicted>,
618 :
619 : /// Do we want to delete locally and remotely this when `LayerInner` is dropped
620 : wanted_deleted: AtomicBool,
621 :
622 : /// Version is to make sure we will only evict a specific initialization of the downloaded file.
623 : ///
624 : /// Incremented for each initialization, stored in `DownloadedLayer::version` or
625 : /// `ResidentOrWantedEvicted::WantedEvicted`.
626 : version: AtomicUsize,
627 :
628 : /// Allow subscribing to when the layer actually gets evicted, a non-cancellable download
629 : /// starts, or completes.
630 : ///
631 : /// Updates must only be posted while holding the InitPermit or the heavier_once_cell::Guard.
632 : /// Holding the InitPermit is the only time we can do state transitions, but we also need to
633 : /// cancel a pending eviction on upgrading a [`ResidentOrWantedEvicted::WantedEvicted`] back to
634 : /// [`ResidentOrWantedEvicted::Resident`] on access.
635 : ///
636 : /// The sender is wrapped in an Option to facilitate moving it out on [`LayerInner::drop`].
637 : status: Option<tokio::sync::watch::Sender<Status>>,
638 :
639 : /// Counter for exponential backoff with the download.
640 : ///
641 : /// This is atomic only for the purposes of having additional data only accessed while holding
642 : /// the InitPermit.
643 : consecutive_failures: AtomicUsize,
644 :
645 : /// The generation of this Layer.
646 : ///
647 : /// For loaded layers (resident or evicted) this comes from [`LayerFileMetadata::generation`],
648 : /// for created layers from [`Timeline::generation`].
649 : generation: Generation,
650 :
651 : /// The shard of this Layer.
652 : ///
653 : /// For layers created in this process, this will always be the [`ShardIndex`] of the
654 : /// current `ShardIdentity`` (TODO: add link once it's introduced).
655 : ///
656 : /// For loaded layers, this may be some other value if the tenant has undergone
657 : /// a shard split since the layer was originally written.
658 : shard: ShardIndex,
659 :
660 : /// When the Layer was last evicted but has not been downloaded since.
661 : ///
662 : /// This is used solely for updating metrics. See [`LayerImplMetrics::redownload_after`].
663 : last_evicted_at: std::sync::Mutex<Option<std::time::Instant>>,
664 :
665 : #[cfg(test)]
666 : failpoints: std::sync::Mutex<Vec<failpoints::Failpoint>>,
667 : }
668 :
669 : impl std::fmt::Display for LayerInner {
670 30 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
671 30 : write!(f, "{}", self.layer_desc().short_id())
672 30 : }
673 : }
674 :
675 : impl AsLayerDesc for LayerInner {
676 722285 : fn layer_desc(&self) -> &PersistentLayerDesc {
677 722285 : &self.desc
678 722285 : }
679 : }
680 :
681 : #[derive(Debug, Clone, Copy)]
682 : enum Status {
683 : Resident,
684 : Evicted,
685 : Downloading,
686 : }
687 :
688 : impl Drop for LayerInner {
689 479 : fn drop(&mut self) {
690 : // if there was a pending eviction, mark it cancelled here to balance metrics
691 479 : if let Some((ResidentOrWantedEvicted::WantedEvicted(..), _)) = self.inner.take_and_deinit()
692 2 : {
693 2 : // eviction has already been started
694 2 : LAYER_IMPL_METRICS.inc_eviction_cancelled(EvictionCancelled::LayerGone);
695 2 :
696 2 : // eviction request is intentionally not honored as no one is present to wait for it
697 2 : // and we could be delaying shutdown for nothing.
698 477 : }
699 :
700 479 : if !*self.wanted_deleted.get_mut() {
701 46 : return;
702 433 : }
703 :
704 433 : let span = tracing::info_span!(parent: None, "layer_delete", tenant_id = %self.layer_desc().tenant_shard_id.tenant_id, shard_id=%self.layer_desc().tenant_shard_id.shard_slug(), timeline_id = %self.layer_desc().timeline_id);
705 :
706 433 : let path = std::mem::take(&mut self.path);
707 433 : let file_name = self.layer_desc().layer_name();
708 433 : let file_size = self.layer_desc().file_size;
709 433 : let timeline = self.timeline.clone();
710 433 : let meta = self.metadata();
711 433 : let status = self.status.take();
712 433 :
713 433 : Self::spawn_blocking(move || {
714 433 : let _g = span.entered();
715 433 :
716 433 : // carry this until we are finished for [`Layer::wait_drop`] support
717 433 : let _status = status;
718 :
719 433 : let Some(timeline) = timeline.upgrade() else {
720 : // no need to nag that timeline is gone: under normal situation on
721 : // task_mgr::remove_tenant_from_memory the timeline is gone before we get dropped.
722 0 : LAYER_IMPL_METRICS.inc_deletes_failed(DeleteFailed::TimelineGone);
723 0 : return;
724 : };
725 :
726 433 : let Ok(_guard) = timeline.gate.enter() else {
727 0 : LAYER_IMPL_METRICS.inc_deletes_failed(DeleteFailed::TimelineGone);
728 0 : return;
729 : };
730 :
731 433 : let removed = match std::fs::remove_file(path) {
732 431 : Ok(()) => true,
733 2 : Err(e) if e.kind() == std::io::ErrorKind::NotFound => {
734 2 : // until we no longer do detaches by removing all local files before removing the
735 2 : // tenant from the global map, we will always get these errors even if we knew what
736 2 : // is the latest state.
737 2 : //
738 2 : // we currently do not track the latest state, so we'll also end up here on evicted
739 2 : // layers.
740 2 : false
741 : }
742 0 : Err(e) => {
743 0 : tracing::error!("failed to remove wanted deleted layer: {e}");
744 0 : LAYER_IMPL_METRICS.inc_delete_removes_failed();
745 0 : false
746 : }
747 : };
748 :
749 433 : if removed {
750 431 : timeline.metrics.resident_physical_size_sub(file_size);
751 431 : }
752 433 : let res = timeline
753 433 : .remote_client
754 433 : .schedule_deletion_of_unlinked(vec![(file_name, meta)]);
755 :
756 433 : if let Err(e) = res {
757 : // test_timeline_deletion_with_files_stuck_in_upload_queue is good at
758 : // demonstrating this deadlock (without spawn_blocking): stop will drop
759 : // queued items, which will have ResidentLayer's, and those drops would try
760 : // to re-entrantly lock the RemoteTimelineClient inner state.
761 0 : if !timeline.is_active() {
762 0 : tracing::info!("scheduling deletion on drop failed: {e:#}");
763 : } else {
764 0 : tracing::warn!("scheduling deletion on drop failed: {e:#}");
765 : }
766 0 : LAYER_IMPL_METRICS.inc_deletes_failed(DeleteFailed::DeleteSchedulingFailed);
767 433 : } else {
768 433 : LAYER_IMPL_METRICS.inc_completed_deletes();
769 433 : }
770 433 : });
771 479 : }
772 : }
773 :
774 : impl LayerInner {
775 : #[allow(clippy::too_many_arguments)]
776 1586 : fn new(
777 1586 : conf: &'static PageServerConf,
778 1586 : timeline: &Arc<Timeline>,
779 1586 : local_path: Utf8PathBuf,
780 1586 : access_stats: LayerAccessStats,
781 1586 : desc: PersistentLayerDesc,
782 1586 : downloaded: Option<Arc<DownloadedLayer>>,
783 1586 : generation: Generation,
784 1586 : shard: ShardIndex,
785 1586 : ) -> Self {
786 1586 : let (inner, version, init_status) = if let Some(inner) = downloaded {
787 1586 : let version = inner.version;
788 1586 : let resident = ResidentOrWantedEvicted::Resident(inner);
789 1586 : (
790 1586 : heavier_once_cell::OnceCell::new(resident),
791 1586 : version,
792 1586 : Status::Resident,
793 1586 : )
794 : } else {
795 0 : (heavier_once_cell::OnceCell::default(), 0, Status::Evicted)
796 : };
797 :
798 1586 : LayerInner {
799 1586 : conf,
800 1586 : debug_str: {
801 1586 : format!("timelines/{}/{}", timeline.timeline_id, desc.layer_name()).into()
802 1586 : },
803 1586 : path: local_path,
804 1586 : desc,
805 1586 : timeline: Arc::downgrade(timeline),
806 1586 : access_stats,
807 1586 : wanted_deleted: AtomicBool::new(false),
808 1586 : inner,
809 1586 : version: AtomicUsize::new(version),
810 1586 : status: Some(tokio::sync::watch::channel(init_status).0),
811 1586 : consecutive_failures: AtomicUsize::new(0),
812 1586 : generation,
813 1586 : shard,
814 1586 : last_evicted_at: std::sync::Mutex::default(),
815 1586 : #[cfg(test)]
816 1586 : failpoints: Default::default(),
817 1586 : }
818 1586 : }
819 :
820 438 : fn delete_on_drop(&self) {
821 438 : let res =
822 438 : self.wanted_deleted
823 438 : .compare_exchange(false, true, Ordering::Release, Ordering::Relaxed);
824 438 :
825 438 : if res.is_ok() {
826 434 : LAYER_IMPL_METRICS.inc_started_deletes();
827 434 : }
828 438 : }
829 :
830 : /// Cancellation safe, however dropping the future and calling this method again might result
831 : /// in a new attempt to evict OR join the previously started attempt.
832 108 : #[tracing::instrument(level = tracing::Level::DEBUG, skip_all, ret, err(level = tracing::Level::DEBUG), fields(layer=%self))]
833 : pub(crate) async fn evict_and_wait(&self, timeout: Duration) -> Result<(), EvictionError> {
834 : let mut rx = self.status.as_ref().unwrap().subscribe();
835 :
836 : {
837 : let current = rx.borrow_and_update();
838 : match &*current {
839 : Status::Resident => {
840 : // we might get lucky and evict this; continue
841 : }
842 : Status::Evicted | Status::Downloading => {
843 : // it is already evicted
844 : return Err(EvictionError::NotFound);
845 : }
846 : }
847 : }
848 :
849 : let strong = {
850 : match self.inner.get() {
851 : Some(mut either) => either.downgrade(),
852 : None => {
853 : // we already have a scheduled eviction, which just has not gotten to run yet.
854 : // it might still race with a read access, but that could also get cancelled,
855 : // so let's say this is not evictable.
856 : return Err(EvictionError::NotFound);
857 : }
858 : }
859 : };
860 :
861 : if strong.is_some() {
862 : // drop the DownloadedLayer outside of the holding the guard
863 : drop(strong);
864 :
865 : // idea here is that only one evicter should ever get to witness a strong reference,
866 : // which means whenever get_or_maybe_download upgrades a weak, it must mark up a
867 : // cancelled eviction and signal us, like it currently does.
868 : //
869 : // a second concurrent evict_and_wait will not see a strong reference.
870 : LAYER_IMPL_METRICS.inc_started_evictions();
871 : }
872 :
873 : let changed = rx.changed();
874 : let changed = tokio::time::timeout(timeout, changed).await;
875 :
876 : let Ok(changed) = changed else {
877 : return Err(EvictionError::Timeout);
878 : };
879 :
880 : let _: () = changed.expect("cannot be closed, because we are holding a strong reference");
881 :
882 : let current = rx.borrow_and_update();
883 :
884 : match &*current {
885 : // the easiest case
886 : Status::Evicted => Ok(()),
887 : // it surely was evicted in between, but then there was a new access now; we can't know
888 : // if it'll succeed so lets just call it evicted
889 : Status::Downloading => Ok(()),
890 : // either the download which was started after eviction completed already, or it was
891 : // never evicted
892 : Status::Resident => Err(EvictionError::Downloaded),
893 : }
894 : }
895 :
896 : /// Cancellation safe.
897 212456 : async fn get_or_maybe_download(
898 212456 : self: &Arc<Self>,
899 212456 : allow_download: bool,
900 212456 : ctx: Option<&RequestContext>,
901 212456 : ) -> Result<Arc<DownloadedLayer>, DownloadError> {
902 16 : let (weak, permit) = {
903 : // get_or_init_detached can:
904 : // - be fast (mutex lock) OR uncontested semaphore permit acquire
905 : // - be slow (wait for semaphore permit or closing)
906 212456 : let init_cancelled = scopeguard::guard((), |_| LAYER_IMPL_METRICS.inc_init_cancelled());
907 :
908 212456 : let locked = self
909 212456 : .inner
910 212456 : .get_or_init_detached()
911 3 : .await
912 212456 : .map(|mut guard| guard.get_and_upgrade().ok_or(guard));
913 212456 :
914 212456 : scopeguard::ScopeGuard::into_inner(init_cancelled);
915 :
916 212440 : match locked {
917 : // this path could had been a RwLock::read
918 212440 : Ok(Ok((strong, upgraded))) if !upgraded => return Ok(strong),
919 0 : Ok(Ok((strong, _))) => {
920 0 : // when upgraded back, the Arc<DownloadedLayer> is still available, but
921 0 : // previously a `evict_and_wait` was received. this is the only place when we
922 0 : // send out an update without holding the InitPermit.
923 0 : //
924 0 : // note that we also have dropped the Guard; this is fine, because we just made
925 0 : // a state change and are holding a strong reference to be returned.
926 0 : self.status.as_ref().unwrap().send_replace(Status::Resident);
927 0 : LAYER_IMPL_METRICS
928 0 : .inc_eviction_cancelled(EvictionCancelled::UpgradedBackOnAccess);
929 0 :
930 0 : return Ok(strong);
931 : }
932 8 : Ok(Err(guard)) => {
933 8 : // path to here: we won the eviction, the file should still be on the disk.
934 8 : let (weak, permit) = guard.take_and_deinit();
935 8 : (Some(weak), permit)
936 : }
937 8 : Err(permit) => (None, permit),
938 : }
939 : };
940 :
941 16 : if let Some(weak) = weak {
942 : // only drop the weak after dropping the heavier_once_cell guard
943 8 : assert!(
944 8 : matches!(weak, ResidentOrWantedEvicted::WantedEvicted(..)),
945 0 : "unexpected {weak:?}, ResidentOrWantedEvicted::get_and_upgrade has a bug"
946 : );
947 8 : }
948 :
949 16 : let timeline = self
950 16 : .timeline
951 16 : .upgrade()
952 16 : .ok_or_else(|| DownloadError::TimelineShutdown)?;
953 :
954 : // count cancellations, which currently remain largely unexpected
955 16 : let init_cancelled = scopeguard::guard((), |_| LAYER_IMPL_METRICS.inc_init_cancelled());
956 :
957 : // check if we really need to be downloaded: this can happen if a read access won the
958 : // semaphore before eviction.
959 : //
960 : // if we are cancelled while doing this `stat` the `self.inner` will be uninitialized. a
961 : // pending eviction will try to evict even upon finding an uninitialized `self.inner`.
962 16 : let needs_download = self
963 16 : .needs_download()
964 14 : .await
965 16 : .map_err(DownloadError::PreStatFailed);
966 16 :
967 16 : scopeguard::ScopeGuard::into_inner(init_cancelled);
968 :
969 16 : let needs_download = needs_download?;
970 :
971 16 : let Some(reason) = needs_download else {
972 : // the file is present locally because eviction has not had a chance to run yet
973 :
974 : #[cfg(test)]
975 8 : self.failpoint(failpoints::FailpointKind::AfterDeterminingLayerNeedsNoDownload)
976 2 : .await?;
977 :
978 6 : LAYER_IMPL_METRICS.inc_init_needed_no_download();
979 6 :
980 6 : return Ok(self.initialize_after_layer_is_on_disk(permit));
981 : };
982 :
983 : // we must download; getting cancelled before spawning the download is not an issue as
984 : // any still running eviction would not find anything to evict.
985 :
986 8 : if let NeedsDownload::NotFile(ft) = reason {
987 0 : return Err(DownloadError::NotFile(ft));
988 8 : }
989 :
990 8 : if let Some(ctx) = ctx {
991 2 : self.check_expected_download(ctx)?;
992 6 : }
993 :
994 8 : if !allow_download {
995 : // this is only used from tests, but it is hard to test without the boolean
996 2 : return Err(DownloadError::DownloadRequired);
997 6 : }
998 6 :
999 6 : let download_ctx = ctx
1000 6 : .map(|ctx| ctx.detached_child(TaskKind::LayerDownload, DownloadBehavior::Download))
1001 6 : .unwrap_or(RequestContext::new(
1002 6 : TaskKind::LayerDownload,
1003 6 : DownloadBehavior::Download,
1004 6 : ));
1005 :
1006 6 : async move {
1007 6 : tracing::info!(%reason, "downloading on-demand");
1008 :
1009 6 : let init_cancelled = scopeguard::guard((), |_| LAYER_IMPL_METRICS.inc_init_cancelled());
1010 6 : let res = self
1011 6 : .download_init_and_wait(timeline, permit, download_ctx)
1012 10 : .await?;
1013 6 : scopeguard::ScopeGuard::into_inner(init_cancelled);
1014 6 : Ok(res)
1015 6 : }
1016 6 : .instrument(tracing::info_span!("get_or_maybe_download", layer=%self))
1017 10 : .await
1018 212456 : }
1019 :
1020 : /// Nag or fail per RequestContext policy
1021 2 : fn check_expected_download(&self, ctx: &RequestContext) -> Result<(), DownloadError> {
1022 2 : use crate::context::DownloadBehavior::*;
1023 2 : let b = ctx.download_behavior();
1024 2 : match b {
1025 2 : Download => Ok(()),
1026 : Warn | Error => {
1027 0 : tracing::info!(
1028 0 : "unexpectedly on-demand downloading for task kind {:?}",
1029 0 : ctx.task_kind()
1030 : );
1031 0 : crate::metrics::UNEXPECTED_ONDEMAND_DOWNLOADS.inc();
1032 :
1033 0 : let really_error =
1034 0 : matches!(b, Error) && !self.conf.ondemand_download_behavior_treat_error_as_warn;
1035 :
1036 0 : if really_error {
1037 : // this check is only probablistic, seems like flakyness footgun
1038 0 : Err(DownloadError::ContextAndConfigReallyDeniesDownloads)
1039 : } else {
1040 0 : Ok(())
1041 : }
1042 : }
1043 : }
1044 2 : }
1045 :
1046 : /// Actual download, at most one is executed at the time.
1047 6 : async fn download_init_and_wait(
1048 6 : self: &Arc<Self>,
1049 6 : timeline: Arc<Timeline>,
1050 6 : permit: heavier_once_cell::InitPermit,
1051 6 : ctx: RequestContext,
1052 6 : ) -> Result<Arc<DownloadedLayer>, DownloadError> {
1053 6 : debug_assert_current_span_has_tenant_and_timeline_id();
1054 6 :
1055 6 : let (tx, rx) = tokio::sync::oneshot::channel();
1056 6 :
1057 6 : let this: Arc<Self> = self.clone();
1058 :
1059 6 : let guard = timeline
1060 6 : .gate
1061 6 : .enter()
1062 6 : .map_err(|_| DownloadError::DownloadCancelled)?;
1063 :
1064 6 : Self::spawn(
1065 6 : async move {
1066 0 : let _guard = guard;
1067 0 :
1068 0 : // now that we have commited to downloading, send out an update to:
1069 0 : // - unhang any pending eviction
1070 0 : // - break out of evict_and_wait
1071 0 : this.status
1072 0 : .as_ref()
1073 0 : .unwrap()
1074 0 : .send_replace(Status::Downloading);
1075 6 :
1076 6 : #[cfg(test)]
1077 6 : this.failpoint(failpoints::FailpointKind::WaitBeforeDownloading)
1078 2 : .await
1079 6 : .unwrap();
1080 :
1081 92 : let res = this.download_and_init(timeline, permit, &ctx).await;
1082 :
1083 6 : if let Err(res) = tx.send(res) {
1084 0 : match res {
1085 0 : Ok(_res) => {
1086 0 : tracing::debug!("layer initialized, but caller has been cancelled");
1087 0 : LAYER_IMPL_METRICS.inc_init_completed_without_requester();
1088 : }
1089 0 : Err(e) => {
1090 0 : tracing::info!(
1091 0 : "layer file download failed, and caller has been cancelled: {e:?}"
1092 : );
1093 0 : LAYER_IMPL_METRICS.inc_download_failed_without_requester();
1094 : }
1095 : }
1096 6 : }
1097 6 : }
1098 6 : .in_current_span(),
1099 6 : );
1100 6 :
1101 10 : match rx.await {
1102 6 : Ok(Ok(res)) => Ok(res),
1103 0 : Ok(Err(e)) => {
1104 0 : // sleep already happened in the spawned task, if it was not cancelled
1105 0 : match e.downcast_ref::<remote_storage::DownloadError>() {
1106 : // If the download failed due to its cancellation token,
1107 : // propagate the cancellation error upstream.
1108 : Some(remote_storage::DownloadError::Cancelled) => {
1109 0 : Err(DownloadError::DownloadCancelled)
1110 : }
1111 : // FIXME: this is not embedding the error because historically it would had
1112 : // been output to compute, however that is no longer the case.
1113 0 : _ => Err(DownloadError::DownloadFailed),
1114 : }
1115 : }
1116 0 : Err(_gone) => Err(DownloadError::DownloadCancelled),
1117 : }
1118 6 : }
1119 :
1120 6 : async fn download_and_init(
1121 6 : self: &Arc<LayerInner>,
1122 6 : timeline: Arc<Timeline>,
1123 6 : permit: heavier_once_cell::InitPermit,
1124 6 : ctx: &RequestContext,
1125 6 : ) -> anyhow::Result<Arc<DownloadedLayer>> {
1126 6 : let result = timeline
1127 6 : .remote_client
1128 6 : .download_layer_file(
1129 6 : &self.desc.layer_name(),
1130 6 : &self.metadata(),
1131 6 : &self.path,
1132 6 : &timeline.cancel,
1133 6 : ctx,
1134 6 : )
1135 86 : .await;
1136 :
1137 6 : match result {
1138 6 : Ok(size) => {
1139 6 : assert_eq!(size, self.desc.file_size);
1140 :
1141 6 : match self.needs_download().await {
1142 0 : Ok(Some(reason)) => {
1143 0 : // this is really a bug in needs_download or remote timeline client
1144 0 : panic!("post-condition failed: needs_download returned {reason:?}");
1145 : }
1146 6 : Ok(None) => {
1147 6 : // as expected
1148 6 : }
1149 0 : Err(e) => {
1150 0 : panic!("post-condition failed: needs_download errored: {e:?}");
1151 : }
1152 : }
1153 :
1154 6 : tracing::info!(size=%self.desc.file_size, "on-demand download successful");
1155 6 : timeline
1156 6 : .metrics
1157 6 : .resident_physical_size_add(self.desc.file_size);
1158 6 : self.consecutive_failures.store(0, Ordering::Relaxed);
1159 6 :
1160 6 : let since_last_eviction = self
1161 6 : .last_evicted_at
1162 6 : .lock()
1163 6 : .unwrap()
1164 6 : .take()
1165 6 : .map(|ts| ts.elapsed());
1166 6 : if let Some(since_last_eviction) = since_last_eviction {
1167 6 : LAYER_IMPL_METRICS.record_redownloaded_after(since_last_eviction);
1168 6 : }
1169 :
1170 6 : self.access_stats.record_residence_event(
1171 6 : LayerResidenceStatus::Resident,
1172 6 : LayerResidenceEventReason::ResidenceChange,
1173 6 : );
1174 6 :
1175 6 : Ok(self.initialize_after_layer_is_on_disk(permit))
1176 : }
1177 0 : Err(e) => {
1178 0 : let consecutive_failures =
1179 0 : 1 + self.consecutive_failures.fetch_add(1, Ordering::Relaxed);
1180 0 :
1181 0 : if timeline.cancel.is_cancelled() {
1182 : // If we're shutting down, drop out before logging the error
1183 0 : return Err(e);
1184 0 : }
1185 0 :
1186 0 : tracing::error!(consecutive_failures, "layer file download failed: {e:#}");
1187 :
1188 0 : let backoff = utils::backoff::exponential_backoff_duration_seconds(
1189 0 : consecutive_failures.min(u32::MAX as usize) as u32,
1190 0 : 1.5,
1191 0 : 60.0,
1192 0 : );
1193 0 :
1194 0 : let backoff = std::time::Duration::from_secs_f64(backoff);
1195 :
1196 : tokio::select! {
1197 : _ = tokio::time::sleep(backoff) => {},
1198 : _ = timeline.cancel.cancelled() => {},
1199 : };
1200 :
1201 0 : Err(e)
1202 : }
1203 : }
1204 6 : }
1205 :
1206 : /// Initializes the `Self::inner` to a "resident" state.
1207 : ///
1208 : /// Callers are assumed to ensure that the file is actually on disk with `Self::needs_download`
1209 : /// before calling this method.
1210 : ///
1211 : /// If this method is ever made async, it needs to be cancellation safe so that no state
1212 : /// changes are made before we can write to the OnceCell in non-cancellable fashion.
1213 12 : fn initialize_after_layer_is_on_disk(
1214 12 : self: &Arc<LayerInner>,
1215 12 : permit: heavier_once_cell::InitPermit,
1216 12 : ) -> Arc<DownloadedLayer> {
1217 12 : debug_assert_current_span_has_tenant_and_timeline_id();
1218 12 :
1219 12 : // disable any scheduled but not yet running eviction deletions for this initialization
1220 12 : let next_version = 1 + self.version.fetch_add(1, Ordering::Relaxed);
1221 12 : self.status.as_ref().unwrap().send_replace(Status::Resident);
1222 12 :
1223 12 : let res = Arc::new(DownloadedLayer {
1224 12 : owner: Arc::downgrade(self),
1225 12 : kind: tokio::sync::OnceCell::default(),
1226 12 : version: next_version,
1227 12 : });
1228 12 :
1229 12 : let waiters = self.inner.initializer_count();
1230 12 : if waiters > 0 {
1231 0 : tracing::info!(waiters, "completing layer init for other tasks");
1232 12 : }
1233 :
1234 12 : let value = ResidentOrWantedEvicted::Resident(res.clone());
1235 12 :
1236 12 : self.inner.set(value, permit);
1237 12 :
1238 12 : res
1239 12 : }
1240 :
1241 24 : async fn needs_download(&self) -> Result<Option<NeedsDownload>, std::io::Error> {
1242 24 : match tokio::fs::metadata(&self.path).await {
1243 16 : Ok(m) => Ok(self.is_file_present_and_good_size(&m).err()),
1244 8 : Err(e) if e.kind() == std::io::ErrorKind::NotFound => Ok(Some(NeedsDownload::NotFound)),
1245 0 : Err(e) => Err(e),
1246 : }
1247 24 : }
1248 :
1249 24 : fn needs_download_blocking(&self) -> Result<Option<NeedsDownload>, std::io::Error> {
1250 24 : match self.path.metadata() {
1251 24 : Ok(m) => Ok(self.is_file_present_and_good_size(&m).err()),
1252 0 : Err(e) if e.kind() == std::io::ErrorKind::NotFound => Ok(Some(NeedsDownload::NotFound)),
1253 0 : Err(e) => Err(e),
1254 : }
1255 24 : }
1256 :
1257 40 : fn is_file_present_and_good_size(&self, m: &std::fs::Metadata) -> Result<(), NeedsDownload> {
1258 40 : // in future, this should include sha2-256 validation of the file.
1259 40 : if !m.is_file() {
1260 0 : Err(NeedsDownload::NotFile(m.file_type()))
1261 40 : } else if m.len() != self.desc.file_size {
1262 0 : Err(NeedsDownload::WrongSize {
1263 0 : actual: m.len(),
1264 0 : expected: self.desc.file_size,
1265 0 : })
1266 : } else {
1267 40 : Ok(())
1268 : }
1269 40 : }
1270 :
1271 0 : fn info(&self, reset: LayerAccessStatsReset) -> HistoricLayerInfo {
1272 0 : let layer_name = self.desc.layer_name().to_string();
1273 0 :
1274 0 : let resident = self
1275 0 : .inner
1276 0 : .get()
1277 0 : .map(|rowe| rowe.is_likely_resident())
1278 0 : .unwrap_or(false);
1279 0 :
1280 0 : let access_stats = self.access_stats.as_api_model(reset);
1281 0 :
1282 0 : if self.desc.is_delta {
1283 0 : let lsn_range = &self.desc.lsn_range;
1284 0 :
1285 0 : HistoricLayerInfo::Delta {
1286 0 : layer_file_name: layer_name,
1287 0 : layer_file_size: self.desc.file_size,
1288 0 : lsn_start: lsn_range.start,
1289 0 : lsn_end: lsn_range.end,
1290 0 : remote: !resident,
1291 0 : access_stats,
1292 0 : l0: crate::tenant::layer_map::LayerMap::is_l0(self.layer_desc()),
1293 0 : }
1294 : } else {
1295 0 : let lsn = self.desc.image_layer_lsn();
1296 0 :
1297 0 : HistoricLayerInfo::Image {
1298 0 : layer_file_name: layer_name,
1299 0 : layer_file_size: self.desc.file_size,
1300 0 : lsn_start: lsn,
1301 0 : remote: !resident,
1302 0 : access_stats,
1303 0 : }
1304 : }
1305 0 : }
1306 :
1307 : /// `DownloadedLayer` is being dropped, so it calls this method.
1308 24 : fn on_downloaded_layer_drop(self: Arc<LayerInner>, only_version: usize) {
1309 : // we cannot know without inspecting LayerInner::inner if we should evict or not, even
1310 : // though here it is very likely
1311 24 : let span = tracing::info_span!(parent: None, "layer_evict", tenant_id = %self.desc.tenant_shard_id.tenant_id, shard_id = %self.desc.tenant_shard_id.shard_slug(), timeline_id = %self.desc.timeline_id, layer=%self, version=%only_version);
1312 :
1313 : // NOTE: this scope *must* never call `self.inner.get` because evict_and_wait might
1314 : // drop while the `self.inner` is being locked, leading to a deadlock.
1315 :
1316 24 : let start_evicting = async move {
1317 24 : #[cfg(test)]
1318 24 : self.failpoint(failpoints::FailpointKind::WaitBeforeStartingEvicting)
1319 14 : .await
1320 24 : .expect("failpoint should not have errored");
1321 24 :
1322 24 : tracing::debug!("eviction started");
1323 :
1324 24 : let res = self.wait_for_turn_and_evict(only_version).await;
1325 : // metrics: ignore the Ok branch, it is not done yet
1326 24 : if let Err(e) = res {
1327 6 : tracing::debug!(res=?Err::<(), _>(&e), "eviction completed");
1328 6 : LAYER_IMPL_METRICS.inc_eviction_cancelled(e);
1329 18 : }
1330 24 : };
1331 :
1332 24 : Self::spawn(start_evicting.instrument(span));
1333 24 : }
1334 :
1335 24 : async fn wait_for_turn_and_evict(
1336 24 : self: Arc<LayerInner>,
1337 24 : only_version: usize,
1338 24 : ) -> Result<(), EvictionCancelled> {
1339 46 : fn is_good_to_continue(status: &Status) -> Result<(), EvictionCancelled> {
1340 46 : use Status::*;
1341 46 : match status {
1342 44 : Resident => Ok(()),
1343 2 : Evicted => Err(EvictionCancelled::UnexpectedEvictedState),
1344 0 : Downloading => Err(EvictionCancelled::LostToDownload),
1345 : }
1346 46 : }
1347 :
1348 24 : let timeline = self
1349 24 : .timeline
1350 24 : .upgrade()
1351 24 : .ok_or(EvictionCancelled::TimelineGone)?;
1352 :
1353 24 : let mut rx = self
1354 24 : .status
1355 24 : .as_ref()
1356 24 : .expect("LayerInner cannot be dropped, holding strong ref")
1357 24 : .subscribe();
1358 24 :
1359 24 : is_good_to_continue(&rx.borrow_and_update())?;
1360 :
1361 22 : let Ok(gate) = timeline.gate.enter() else {
1362 0 : return Err(EvictionCancelled::TimelineGone);
1363 : };
1364 :
1365 18 : let permit = {
1366 : // we cannot just `std::fs::remove_file` because there might already be an
1367 : // get_or_maybe_download which will inspect filesystem and reinitialize. filesystem
1368 : // operations must be done while holding the heavier_once_cell::InitPermit
1369 22 : let mut wait = std::pin::pin!(self.inner.get_or_init_detached());
1370 :
1371 22 : let waited = loop {
1372 22 : // we must race to the Downloading starting, otherwise we would have to wait until the
1373 22 : // completion of the download. waiting for download could be long and hinder our
1374 22 : // efforts to alert on "hanging" evictions.
1375 22 : tokio::select! {
1376 : res = &mut wait => break res,
1377 : _ = rx.changed() => {
1378 : is_good_to_continue(&rx.borrow_and_update())?;
1379 : // two possibilities for Status::Resident:
1380 : // - the layer was found locally from disk by a read
1381 : // - we missed a bunch of updates and now the layer is
1382 : // again downloaded -- assume we'll fail later on with
1383 : // version check or AlreadyReinitialized
1384 : }
1385 22 : }
1386 22 : };
1387 :
1388 : // re-check now that we have the guard or permit; all updates should have happened
1389 : // while holding the permit.
1390 22 : is_good_to_continue(&rx.borrow_and_update())?;
1391 :
1392 : // the term deinitialize is used here, because we clearing out the Weak will eventually
1393 : // lead to deallocating the reference counted value, and the value we
1394 : // `Guard::take_and_deinit` is likely to be the last because the Weak is never cloned.
1395 22 : let (_weak, permit) = match waited {
1396 20 : Ok(guard) => {
1397 20 : match &*guard {
1398 18 : ResidentOrWantedEvicted::WantedEvicted(_weak, version)
1399 18 : if *version == only_version =>
1400 16 : {
1401 16 : tracing::debug!(version, "deinitializing matching WantedEvicted");
1402 16 : let (weak, permit) = guard.take_and_deinit();
1403 16 : (Some(weak), permit)
1404 : }
1405 2 : ResidentOrWantedEvicted::WantedEvicted(_, version) => {
1406 2 : // if we were not doing the version check, we would need to try to
1407 2 : // upgrade the weak here to see if it really is dropped. version check
1408 2 : // is done instead assuming that it is cheaper.
1409 2 : tracing::debug!(
1410 : version,
1411 : only_version,
1412 0 : "version mismatch, not deinitializing"
1413 : );
1414 2 : return Err(EvictionCancelled::VersionCheckFailed);
1415 : }
1416 : ResidentOrWantedEvicted::Resident(_) => {
1417 2 : return Err(EvictionCancelled::AlreadyReinitialized);
1418 : }
1419 : }
1420 : }
1421 2 : Err(permit) => {
1422 2 : tracing::debug!("continuing after cancelled get_or_maybe_download or eviction");
1423 2 : (None, permit)
1424 : }
1425 : };
1426 :
1427 18 : permit
1428 18 : };
1429 18 :
1430 18 : let span = tracing::Span::current();
1431 18 :
1432 18 : let spawned_at = std::time::Instant::now();
1433 18 :
1434 18 : // this is on purpose a detached spawn; we don't need to wait for it
1435 18 : //
1436 18 : // eviction completion reporting is the only thing hinging on this, and it can be just as
1437 18 : // well from a spawn_blocking thread.
1438 18 : //
1439 18 : // important to note that now that we've acquired the permit we have made sure the evicted
1440 18 : // file is either the exact `WantedEvicted` we wanted to evict, or uninitialized in case
1441 18 : // there are multiple evictions. The rest is not cancellable, and we've now commited to
1442 18 : // evicting.
1443 18 : //
1444 18 : // If spawn_blocking has a queue and maximum number of threads are in use, we could stall
1445 18 : // reads. We will need to add cancellation for that if necessary.
1446 18 : Self::spawn_blocking(move || {
1447 18 : let _span = span.entered();
1448 18 :
1449 18 : let res = self.evict_blocking(&timeline, &gate, &permit);
1450 18 :
1451 18 : let waiters = self.inner.initializer_count();
1452 18 :
1453 18 : if waiters > 0 {
1454 0 : LAYER_IMPL_METRICS.inc_evicted_with_waiters();
1455 18 : }
1456 :
1457 18 : let completed_in = spawned_at.elapsed();
1458 18 : LAYER_IMPL_METRICS.record_time_to_evict(completed_in);
1459 18 :
1460 18 : match res {
1461 18 : Ok(()) => LAYER_IMPL_METRICS.inc_completed_evictions(),
1462 0 : Err(e) => LAYER_IMPL_METRICS.inc_eviction_cancelled(e),
1463 : }
1464 :
1465 18 : tracing::debug!(?res, elapsed_ms=%completed_in.as_millis(), %waiters, "eviction completed");
1466 18 : });
1467 18 :
1468 18 : Ok(())
1469 24 : }
1470 :
1471 : /// This is blocking only to do just one spawn_blocking hop compared to multiple via tokio::fs.
1472 18 : fn evict_blocking(
1473 18 : &self,
1474 18 : timeline: &Timeline,
1475 18 : _gate: &gate::GateGuard,
1476 18 : _permit: &heavier_once_cell::InitPermit,
1477 18 : ) -> Result<(), EvictionCancelled> {
1478 18 : // now accesses to `self.inner.get_or_init*` wait on the semaphore or the `_permit`
1479 18 :
1480 18 : match capture_mtime_and_remove(&self.path) {
1481 18 : Ok(local_layer_mtime) => {
1482 18 : let duration = SystemTime::now().duration_since(local_layer_mtime);
1483 18 : match duration {
1484 18 : Ok(elapsed) => {
1485 18 : timeline
1486 18 : .metrics
1487 18 : .evictions_with_low_residence_duration
1488 18 : .read()
1489 18 : .unwrap()
1490 18 : .observe(elapsed);
1491 18 : tracing::info!(
1492 0 : residence_millis = elapsed.as_millis(),
1493 0 : "evicted layer after known residence period"
1494 : );
1495 : }
1496 : Err(_) => {
1497 0 : tracing::info!("evicted layer after unknown residence period");
1498 : }
1499 : }
1500 18 : timeline.metrics.evictions.inc();
1501 18 : timeline
1502 18 : .metrics
1503 18 : .resident_physical_size_sub(self.desc.file_size);
1504 : }
1505 0 : Err(e) if e.kind() == std::io::ErrorKind::NotFound => {
1506 0 : tracing::error!(
1507 : layer_size = %self.desc.file_size,
1508 0 : "failed to evict layer from disk, it was already gone"
1509 : );
1510 0 : return Err(EvictionCancelled::FileNotFound);
1511 : }
1512 0 : Err(e) => {
1513 0 : // FIXME: this should probably be an abort
1514 0 : tracing::error!("failed to evict file from disk: {e:#}");
1515 0 : return Err(EvictionCancelled::RemoveFailed);
1516 : }
1517 : }
1518 :
1519 18 : self.access_stats.record_residence_event(
1520 18 : LayerResidenceStatus::Evicted,
1521 18 : LayerResidenceEventReason::ResidenceChange,
1522 18 : );
1523 18 :
1524 18 : self.status.as_ref().unwrap().send_replace(Status::Evicted);
1525 18 :
1526 18 : *self.last_evicted_at.lock().unwrap() = Some(std::time::Instant::now());
1527 18 :
1528 18 : Ok(())
1529 18 : }
1530 :
1531 1935 : fn metadata(&self) -> LayerFileMetadata {
1532 1935 : LayerFileMetadata::new(self.desc.file_size, self.generation, self.shard)
1533 1935 : }
1534 :
1535 : /// Needed to use entered runtime in tests, but otherwise use BACKGROUND_RUNTIME.
1536 : ///
1537 : /// Synchronizing with spawned tasks is very complicated otherwise.
1538 30 : fn spawn<F>(fut: F)
1539 30 : where
1540 30 : F: std::future::Future<Output = ()> + Send + 'static,
1541 30 : {
1542 30 : #[cfg(test)]
1543 30 : tokio::task::spawn(fut);
1544 30 : #[cfg(not(test))]
1545 30 : crate::task_mgr::BACKGROUND_RUNTIME.spawn(fut);
1546 30 : }
1547 :
1548 : /// Needed to use entered runtime in tests, but otherwise use BACKGROUND_RUNTIME.
1549 451 : fn spawn_blocking<F>(f: F)
1550 451 : where
1551 451 : F: FnOnce() + Send + 'static,
1552 451 : {
1553 451 : #[cfg(test)]
1554 451 : tokio::task::spawn_blocking(f);
1555 451 : #[cfg(not(test))]
1556 451 : crate::task_mgr::BACKGROUND_RUNTIME.spawn_blocking(f);
1557 451 : }
1558 : }
1559 :
1560 18 : fn capture_mtime_and_remove(path: &Utf8Path) -> Result<SystemTime, std::io::Error> {
1561 18 : let m = path.metadata()?;
1562 18 : let local_layer_mtime = m.modified()?;
1563 18 : std::fs::remove_file(path)?;
1564 18 : Ok(local_layer_mtime)
1565 18 : }
1566 :
1567 0 : #[derive(Debug, thiserror::Error)]
1568 : pub(crate) enum EvictionError {
1569 : #[error("layer was already evicted")]
1570 : NotFound,
1571 :
1572 : /// Evictions must always lose to downloads in races, and this time it happened.
1573 : #[error("layer was downloaded instead")]
1574 : Downloaded,
1575 :
1576 : #[error("eviction did not happen within timeout")]
1577 : Timeout,
1578 : }
1579 :
1580 : /// Error internal to the [`LayerInner::get_or_maybe_download`]
1581 0 : #[derive(Debug, thiserror::Error)]
1582 : pub(crate) enum DownloadError {
1583 : #[error("timeline has already shutdown")]
1584 : TimelineShutdown,
1585 : #[error("context denies downloading")]
1586 : ContextAndConfigReallyDeniesDownloads,
1587 : #[error("downloading is really required but not allowed by this method")]
1588 : DownloadRequired,
1589 : #[error("layer path exists, but it is not a file: {0:?}")]
1590 : NotFile(std::fs::FileType),
1591 : /// Why no error here? Because it will be reported by page_service. We should had also done
1592 : /// retries already.
1593 : #[error("downloading evicted layer file failed")]
1594 : DownloadFailed,
1595 : #[error("downloading failed, possibly for shutdown")]
1596 : DownloadCancelled,
1597 : #[error("pre-condition: stat before download failed")]
1598 : PreStatFailed(#[source] std::io::Error),
1599 :
1600 : #[cfg(test)]
1601 : #[error("failpoint: {0:?}")]
1602 : Failpoint(failpoints::FailpointKind),
1603 : }
1604 :
1605 : #[derive(Debug, PartialEq)]
1606 : pub(crate) enum NeedsDownload {
1607 : NotFound,
1608 : NotFile(std::fs::FileType),
1609 : WrongSize { actual: u64, expected: u64 },
1610 : }
1611 :
1612 : impl std::fmt::Display for NeedsDownload {
1613 6 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
1614 6 : match self {
1615 6 : NeedsDownload::NotFound => write!(f, "file was not found"),
1616 0 : NeedsDownload::NotFile(ft) => write!(f, "path is not a file; {ft:?}"),
1617 0 : NeedsDownload::WrongSize { actual, expected } => {
1618 0 : write!(f, "file size mismatch {actual} vs. {expected}")
1619 : }
1620 : }
1621 6 : }
1622 : }
1623 :
1624 : /// Existence of `DownloadedLayer` means that we have the file locally, and can later evict it.
1625 : pub(crate) struct DownloadedLayer {
1626 : owner: Weak<LayerInner>,
1627 : // Use tokio OnceCell as we do not need to deinitialize this, it'll just get dropped with the
1628 : // DownloadedLayer
1629 : kind: tokio::sync::OnceCell<anyhow::Result<LayerKind>>,
1630 : version: usize,
1631 : }
1632 :
1633 : impl std::fmt::Debug for DownloadedLayer {
1634 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
1635 0 : f.debug_struct("DownloadedLayer")
1636 0 : // owner omitted because it is always "Weak"
1637 0 : .field("kind", &self.kind)
1638 0 : .field("version", &self.version)
1639 0 : .finish()
1640 0 : }
1641 : }
1642 :
1643 : impl Drop for DownloadedLayer {
1644 501 : fn drop(&mut self) {
1645 501 : if let Some(owner) = self.owner.upgrade() {
1646 24 : owner.on_downloaded_layer_drop(self.version);
1647 477 : } else {
1648 477 : // Layer::drop will handle cancelling the eviction; because of drop order and
1649 477 : // `DownloadedLayer` never leaking, we cannot know here if eviction was requested.
1650 477 : }
1651 501 : }
1652 : }
1653 :
1654 : impl DownloadedLayer {
1655 : /// Initializes the `DeltaLayerInner` or `ImageLayerInner` within [`LayerKind`], or fails to
1656 : /// initialize it permanently.
1657 : ///
1658 : /// `owner` parameter is a strong reference at the same `LayerInner` as the
1659 : /// `DownloadedLayer::owner` would be when upgraded. Given how this method ends up called,
1660 : /// we will always have the LayerInner on the callstack, so we can just use it.
1661 212484 : async fn get<'a>(
1662 212484 : &'a self,
1663 212484 : owner: &Arc<LayerInner>,
1664 212484 : ctx: &RequestContext,
1665 212484 : ) -> anyhow::Result<&'a LayerKind> {
1666 212484 : let init = || async {
1667 1088 : assert_eq!(
1668 1088 : Weak::as_ptr(&self.owner),
1669 1088 : Arc::as_ptr(owner),
1670 1088 : "these are the same, just avoiding the upgrade"
1671 1088 : );
1672 1088 :
1673 1088 : let res = if owner.desc.is_delta {
1674 1088 : let summary = Some(delta_layer::Summary::expected(
1675 998 : owner.desc.tenant_shard_id.tenant_id,
1676 998 : owner.desc.timeline_id,
1677 998 : owner.desc.key_range.clone(),
1678 998 : owner.desc.lsn_range.clone(),
1679 998 : ));
1680 998 : delta_layer::DeltaLayerInner::load(
1681 998 : &owner.path,
1682 998 : summary,
1683 998 : Some(owner.conf.max_vectored_read_bytes),
1684 998 : ctx,
1685 998 : )
1686 1088 : .await
1687 1088 : .map(|res| res.map(LayerKind::Delta))
1688 1088 : } else {
1689 1088 : let lsn = owner.desc.image_layer_lsn();
1690 90 : let summary = Some(image_layer::Summary::expected(
1691 90 : owner.desc.tenant_shard_id.tenant_id,
1692 90 : owner.desc.timeline_id,
1693 90 : owner.desc.key_range.clone(),
1694 90 : lsn,
1695 90 : ));
1696 90 : image_layer::ImageLayerInner::load(
1697 90 : &owner.path,
1698 90 : lsn,
1699 90 : summary,
1700 90 : Some(owner.conf.max_vectored_read_bytes),
1701 90 : ctx,
1702 90 : )
1703 1088 : .await
1704 1088 : .map(|res| res.map(LayerKind::Image))
1705 1088 : };
1706 1088 :
1707 1088 : match res {
1708 1088 : Ok(Ok(layer)) => Ok(Ok(layer)),
1709 1088 : Ok(Err(transient)) => Err(transient),
1710 1088 : Err(permanent) => {
1711 0 : LAYER_IMPL_METRICS.inc_permanent_loading_failures();
1712 0 : // TODO(#5815): we are not logging all errors, so temporarily log them **once**
1713 0 : // here as well
1714 0 : let permanent = permanent.context("load layer");
1715 0 : tracing::error!("layer loading failed permanently: {permanent:#}");
1716 1088 : Ok(Err(permanent))
1717 1088 : }
1718 1088 : }
1719 1088 : };
1720 212484 : self.kind
1721 212484 : .get_or_try_init(init)
1722 : // return transient errors using `?`
1723 1099 : .await?
1724 212484 : .as_ref()
1725 212484 : .map_err(|e| {
1726 0 : // errors are not clonabled, cannot but stringify
1727 0 : // test_broken_timeline matches this string
1728 0 : anyhow::anyhow!("layer loading failed: {e:#}")
1729 212484 : })
1730 212484 : }
1731 :
1732 211782 : async fn get_value_reconstruct_data(
1733 211782 : &self,
1734 211782 : key: Key,
1735 211782 : lsn_range: Range<Lsn>,
1736 211782 : reconstruct_data: &mut ValueReconstructState,
1737 211782 : owner: &Arc<LayerInner>,
1738 211782 : ctx: &RequestContext,
1739 211782 : ) -> anyhow::Result<ValueReconstructResult> {
1740 211782 : use LayerKind::*;
1741 211782 :
1742 211782 : match self.get(owner, ctx).await? {
1743 204764 : Delta(d) => {
1744 204764 : d.get_value_reconstruct_data(key, lsn_range, reconstruct_data, ctx)
1745 29058 : .await
1746 : }
1747 7018 : Image(i) => {
1748 7018 : i.get_value_reconstruct_data(key, reconstruct_data, ctx)
1749 733 : .await
1750 : }
1751 : }
1752 211782 : }
1753 :
1754 228 : async fn get_values_reconstruct_data(
1755 228 : &self,
1756 228 : keyspace: KeySpace,
1757 228 : lsn_range: Range<Lsn>,
1758 228 : reconstruct_data: &mut ValuesReconstructState,
1759 228 : owner: &Arc<LayerInner>,
1760 228 : ctx: &RequestContext,
1761 228 : ) -> Result<(), GetVectoredError> {
1762 228 : use LayerKind::*;
1763 228 :
1764 228 : match self.get(owner, ctx).await.map_err(GetVectoredError::from)? {
1765 154 : Delta(d) => {
1766 154 : d.get_values_reconstruct_data(keyspace, lsn_range, reconstruct_data, ctx)
1767 10163 : .await
1768 : }
1769 74 : Image(i) => {
1770 74 : i.get_values_reconstruct_data(keyspace, reconstruct_data, ctx)
1771 1178 : .await
1772 : }
1773 : }
1774 228 : }
1775 :
1776 16 : async fn load_key_values(
1777 16 : &self,
1778 16 : owner: &Arc<LayerInner>,
1779 16 : ctx: &RequestContext,
1780 16 : ) -> anyhow::Result<Vec<(Key, Lsn, crate::repository::Value)>> {
1781 16 : use LayerKind::*;
1782 16 :
1783 16 : match self.get(owner, ctx).await? {
1784 8 : Delta(d) => d.load_key_values(ctx).await,
1785 8 : Image(i) => i.load_key_values(ctx).await,
1786 : }
1787 16 : }
1788 :
1789 4 : async fn dump(&self, owner: &Arc<LayerInner>, ctx: &RequestContext) -> anyhow::Result<()> {
1790 4 : use LayerKind::*;
1791 4 : match self.get(owner, ctx).await? {
1792 4 : Delta(d) => d.dump(ctx).await?,
1793 0 : Image(i) => i.dump(ctx).await?,
1794 : }
1795 :
1796 4 : Ok(())
1797 4 : }
1798 : }
1799 :
1800 : /// Wrapper around an actual layer implementation.
1801 : #[derive(Debug)]
1802 : enum LayerKind {
1803 : Delta(delta_layer::DeltaLayerInner),
1804 : Image(image_layer::ImageLayerInner),
1805 : }
1806 :
1807 : /// Guard for forcing a layer be resident while it exists.
1808 : #[derive(Clone)]
1809 : pub(crate) struct ResidentLayer {
1810 : owner: Layer,
1811 : downloaded: Arc<DownloadedLayer>,
1812 : }
1813 :
1814 : impl std::fmt::Display for ResidentLayer {
1815 1878 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
1816 1878 : write!(f, "{}", self.owner)
1817 1878 : }
1818 : }
1819 :
1820 : impl std::fmt::Debug for ResidentLayer {
1821 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
1822 0 : write!(f, "{}", self.owner)
1823 0 : }
1824 : }
1825 :
1826 : impl ResidentLayer {
1827 : /// Release the eviction guard, converting back into a plain [`Layer`].
1828 : ///
1829 : /// You can access the [`Layer`] also by using `as_ref`.
1830 420 : pub(crate) fn drop_eviction_guard(self) -> Layer {
1831 420 : self.into()
1832 420 : }
1833 :
1834 : /// Loads all keys stored in the layer. Returns key, lsn and value size.
1835 804 : #[tracing::instrument(level = tracing::Level::DEBUG, skip_all, fields(layer=%self))]
1836 : pub(crate) async fn load_keys<'a>(
1837 : &'a self,
1838 : ctx: &RequestContext,
1839 : ) -> anyhow::Result<Vec<DeltaEntry<'a>>> {
1840 : use LayerKind::*;
1841 :
1842 : let owner = &self.owner.0;
1843 : match self.downloaded.get(owner, ctx).await? {
1844 : Delta(ref d) => {
1845 : // this is valid because the DownloadedLayer::kind is a OnceCell, not a
1846 : // Mutex<OnceCell>, so we cannot go and deinitialize the value with OnceCell::take
1847 : // while it's being held.
1848 : owner
1849 : .access_stats
1850 : .record_access(LayerAccessKind::KeyIter, ctx);
1851 :
1852 : delta_layer::DeltaLayerInner::load_keys(d, ctx)
1853 : .await
1854 0 : .with_context(|| format!("Layer index is corrupted for {self}"))
1855 : }
1856 : Image(_) => anyhow::bail!(format!("cannot load_keys on a image layer {self}")),
1857 : }
1858 : }
1859 :
1860 : /// Read all they keys in this layer which match the ShardIdentity, and write them all to
1861 : /// the provided writer. Return the number of keys written.
1862 16 : #[tracing::instrument(level = tracing::Level::DEBUG, skip_all, fields(layer=%self))]
1863 : pub(crate) async fn filter<'a>(
1864 : &'a self,
1865 : shard_identity: &ShardIdentity,
1866 : writer: &mut ImageLayerWriter,
1867 : ctx: &RequestContext,
1868 : ) -> anyhow::Result<usize> {
1869 : use LayerKind::*;
1870 :
1871 : match self.downloaded.get(&self.owner.0, ctx).await? {
1872 : Delta(_) => anyhow::bail!(format!("cannot filter() on a delta layer {self}")),
1873 : Image(i) => i.filter(shard_identity, writer, ctx).await,
1874 : }
1875 : }
1876 :
1877 : /// Returns the amount of keys and values written to the writer.
1878 10 : pub(crate) async fn copy_delta_prefix(
1879 10 : &self,
1880 10 : writer: &mut super::delta_layer::DeltaLayerWriter,
1881 10 : until: Lsn,
1882 10 : ctx: &RequestContext,
1883 10 : ) -> anyhow::Result<usize> {
1884 10 : use LayerKind::*;
1885 10 :
1886 10 : let owner = &self.owner.0;
1887 10 :
1888 10 : match self.downloaded.get(owner, ctx).await? {
1889 10 : Delta(ref d) => d
1890 10 : .copy_prefix(writer, until, ctx)
1891 13 : .await
1892 10 : .with_context(|| format!("copy_delta_prefix until {until} of {self}")),
1893 0 : Image(_) => anyhow::bail!(format!("cannot copy_lsn_prefix of image layer {self}")),
1894 : }
1895 10 : }
1896 :
1897 1564 : pub(crate) fn local_path(&self) -> &Utf8Path {
1898 1564 : &self.owner.0.path
1899 1564 : }
1900 :
1901 1496 : pub(crate) fn metadata(&self) -> LayerFileMetadata {
1902 1496 : self.owner.metadata()
1903 1496 : }
1904 :
1905 : #[cfg(test)]
1906 32 : pub(crate) async fn get_as_delta(
1907 32 : &self,
1908 32 : ctx: &RequestContext,
1909 32 : ) -> anyhow::Result<&delta_layer::DeltaLayerInner> {
1910 32 : use LayerKind::*;
1911 32 : match self.downloaded.get(&self.owner.0, ctx).await? {
1912 32 : Delta(ref d) => Ok(d),
1913 0 : Image(_) => Err(anyhow::anyhow!("image layer")),
1914 : }
1915 32 : }
1916 :
1917 : #[cfg(test)]
1918 2 : pub(crate) async fn get_as_image(
1919 2 : &self,
1920 2 : ctx: &RequestContext,
1921 2 : ) -> anyhow::Result<&image_layer::ImageLayerInner> {
1922 2 : use LayerKind::*;
1923 2 : match self.downloaded.get(&self.owner.0, ctx).await? {
1924 2 : Image(ref d) => Ok(d),
1925 0 : Delta(_) => Err(anyhow::anyhow!("delta layer")),
1926 : }
1927 2 : }
1928 : }
1929 :
1930 : impl AsLayerDesc for ResidentLayer {
1931 5266 : fn layer_desc(&self) -> &PersistentLayerDesc {
1932 5266 : self.owner.layer_desc()
1933 5266 : }
1934 : }
1935 :
1936 : impl AsRef<Layer> for ResidentLayer {
1937 1848 : fn as_ref(&self) -> &Layer {
1938 1848 : &self.owner
1939 1848 : }
1940 : }
1941 :
1942 : /// Drop the eviction guard.
1943 : impl From<ResidentLayer> for Layer {
1944 420 : fn from(value: ResidentLayer) -> Self {
1945 420 : value.owner
1946 420 : }
1947 : }
1948 :
1949 : use metrics::IntCounter;
1950 :
1951 : pub(crate) struct LayerImplMetrics {
1952 : started_evictions: IntCounter,
1953 : completed_evictions: IntCounter,
1954 : cancelled_evictions: enum_map::EnumMap<EvictionCancelled, IntCounter>,
1955 :
1956 : started_deletes: IntCounter,
1957 : completed_deletes: IntCounter,
1958 : failed_deletes: enum_map::EnumMap<DeleteFailed, IntCounter>,
1959 :
1960 : rare_counters: enum_map::EnumMap<RareEvent, IntCounter>,
1961 : inits_cancelled: metrics::core::GenericCounter<metrics::core::AtomicU64>,
1962 : redownload_after: metrics::Histogram,
1963 : time_to_evict: metrics::Histogram,
1964 : }
1965 :
1966 : impl Default for LayerImplMetrics {
1967 36 : fn default() -> Self {
1968 36 : use enum_map::Enum;
1969 36 :
1970 36 : // reminder: these will be pageserver_layer_* with "_total" suffix
1971 36 :
1972 36 : let started_evictions = metrics::register_int_counter!(
1973 : "pageserver_layer_started_evictions",
1974 : "Evictions started in the Layer implementation"
1975 : )
1976 36 : .unwrap();
1977 36 : let completed_evictions = metrics::register_int_counter!(
1978 : "pageserver_layer_completed_evictions",
1979 : "Evictions completed in the Layer implementation"
1980 : )
1981 36 : .unwrap();
1982 36 :
1983 36 : let cancelled_evictions = metrics::register_int_counter_vec!(
1984 : "pageserver_layer_cancelled_evictions_count",
1985 : "Different reasons for evictions to have been cancelled or failed",
1986 : &["reason"]
1987 : )
1988 36 : .unwrap();
1989 36 :
1990 324 : let cancelled_evictions = enum_map::EnumMap::from_array(std::array::from_fn(|i| {
1991 324 : let reason = EvictionCancelled::from_usize(i);
1992 324 : let s = reason.as_str();
1993 324 : cancelled_evictions.with_label_values(&[s])
1994 324 : }));
1995 36 :
1996 36 : let started_deletes = metrics::register_int_counter!(
1997 : "pageserver_layer_started_deletes",
1998 : "Deletions on drop pending in the Layer implementation"
1999 : )
2000 36 : .unwrap();
2001 36 : let completed_deletes = metrics::register_int_counter!(
2002 : "pageserver_layer_completed_deletes",
2003 : "Deletions on drop completed in the Layer implementation"
2004 : )
2005 36 : .unwrap();
2006 36 :
2007 36 : let failed_deletes = metrics::register_int_counter_vec!(
2008 : "pageserver_layer_failed_deletes_count",
2009 : "Different reasons for deletions on drop to have failed",
2010 : &["reason"]
2011 : )
2012 36 : .unwrap();
2013 36 :
2014 72 : let failed_deletes = enum_map::EnumMap::from_array(std::array::from_fn(|i| {
2015 72 : let reason = DeleteFailed::from_usize(i);
2016 72 : let s = reason.as_str();
2017 72 : failed_deletes.with_label_values(&[s])
2018 72 : }));
2019 36 :
2020 36 : let rare_counters = metrics::register_int_counter_vec!(
2021 : "pageserver_layer_assumed_rare_count",
2022 : "Times unexpected or assumed rare event happened",
2023 : &["event"]
2024 : )
2025 36 : .unwrap();
2026 36 :
2027 252 : let rare_counters = enum_map::EnumMap::from_array(std::array::from_fn(|i| {
2028 252 : let event = RareEvent::from_usize(i);
2029 252 : let s = event.as_str();
2030 252 : rare_counters.with_label_values(&[s])
2031 252 : }));
2032 36 :
2033 36 : let inits_cancelled = metrics::register_int_counter!(
2034 : "pageserver_layer_inits_cancelled_count",
2035 : "Times Layer initialization was cancelled",
2036 : )
2037 36 : .unwrap();
2038 36 :
2039 36 : let redownload_after = {
2040 36 : let minute = 60.0;
2041 36 : let hour = 60.0 * minute;
2042 : metrics::register_histogram!(
2043 : "pageserver_layer_redownloaded_after",
2044 : "Time between evicting and re-downloading.",
2045 : vec![
2046 : 10.0,
2047 : 30.0,
2048 : minute,
2049 : 5.0 * minute,
2050 : 15.0 * minute,
2051 : 30.0 * minute,
2052 : hour,
2053 : 12.0 * hour,
2054 : ]
2055 : )
2056 36 : .unwrap()
2057 36 : };
2058 36 :
2059 36 : let time_to_evict = metrics::register_histogram!(
2060 : "pageserver_layer_eviction_held_permit_seconds",
2061 : "Time eviction held the permit.",
2062 : vec![0.001, 0.010, 0.100, 0.500, 1.000, 5.000]
2063 : )
2064 36 : .unwrap();
2065 36 :
2066 36 : Self {
2067 36 : started_evictions,
2068 36 : completed_evictions,
2069 36 : cancelled_evictions,
2070 36 :
2071 36 : started_deletes,
2072 36 : completed_deletes,
2073 36 : failed_deletes,
2074 36 :
2075 36 : rare_counters,
2076 36 : inits_cancelled,
2077 36 : redownload_after,
2078 36 : time_to_evict,
2079 36 : }
2080 36 : }
2081 : }
2082 :
2083 : impl LayerImplMetrics {
2084 26 : fn inc_started_evictions(&self) {
2085 26 : self.started_evictions.inc();
2086 26 : }
2087 18 : fn inc_completed_evictions(&self) {
2088 18 : self.completed_evictions.inc();
2089 18 : }
2090 8 : fn inc_eviction_cancelled(&self, reason: EvictionCancelled) {
2091 8 : self.cancelled_evictions[reason].inc()
2092 8 : }
2093 :
2094 434 : fn inc_started_deletes(&self) {
2095 434 : self.started_deletes.inc();
2096 434 : }
2097 433 : fn inc_completed_deletes(&self) {
2098 433 : self.completed_deletes.inc();
2099 433 : }
2100 0 : fn inc_deletes_failed(&self, reason: DeleteFailed) {
2101 0 : self.failed_deletes[reason].inc();
2102 0 : }
2103 :
2104 : /// Counted separatedly from failed layer deletes because we will complete the layer deletion
2105 : /// attempt regardless of failure to delete local file.
2106 0 : fn inc_delete_removes_failed(&self) {
2107 0 : self.rare_counters[RareEvent::RemoveOnDropFailed].inc();
2108 0 : }
2109 :
2110 : /// Expected rare just as cancellations are rare, but we could have cancellations separate from
2111 : /// the single caller which can start the download, so use this counter to separte them.
2112 0 : fn inc_init_completed_without_requester(&self) {
2113 0 : self.rare_counters[RareEvent::InitCompletedWithoutRequester].inc();
2114 0 : }
2115 :
2116 : /// Expected rare because cancellations are unexpected, and failures are unexpected
2117 0 : fn inc_download_failed_without_requester(&self) {
2118 0 : self.rare_counters[RareEvent::DownloadFailedWithoutRequester].inc();
2119 0 : }
2120 :
2121 : /// The Weak in ResidentOrWantedEvicted::WantedEvicted was successfully upgraded.
2122 : ///
2123 : /// If this counter is always zero, we should replace ResidentOrWantedEvicted type with an
2124 : /// Option.
2125 0 : fn inc_raced_wanted_evicted_accesses(&self) {
2126 0 : self.rare_counters[RareEvent::UpgradedWantedEvicted].inc();
2127 0 : }
2128 :
2129 : /// These are only expected for [`Self::inc_init_cancelled`] amount when
2130 : /// running with remote storage.
2131 6 : fn inc_init_needed_no_download(&self) {
2132 6 : self.rare_counters[RareEvent::InitWithoutDownload].inc();
2133 6 : }
2134 :
2135 : /// Expected rare because all layer files should be readable and good
2136 0 : fn inc_permanent_loading_failures(&self) {
2137 0 : self.rare_counters[RareEvent::PermanentLoadingFailure].inc();
2138 0 : }
2139 :
2140 0 : fn inc_init_cancelled(&self) {
2141 0 : self.inits_cancelled.inc()
2142 0 : }
2143 :
2144 6 : fn record_redownloaded_after(&self, duration: std::time::Duration) {
2145 6 : self.redownload_after.observe(duration.as_secs_f64())
2146 6 : }
2147 :
2148 : /// This would be bad if it ever happened, or mean extreme disk pressure. We should probably
2149 : /// instead cancel eviction if we would have read waiters. We cannot however separate reads
2150 : /// from other evictions, so this could have noise as well.
2151 0 : fn inc_evicted_with_waiters(&self) {
2152 0 : self.rare_counters[RareEvent::EvictedWithWaiters].inc();
2153 0 : }
2154 :
2155 : /// Recorded at least initially as the permit is now acquired in async context before
2156 : /// spawn_blocking action.
2157 18 : fn record_time_to_evict(&self, duration: std::time::Duration) {
2158 18 : self.time_to_evict.observe(duration.as_secs_f64())
2159 18 : }
2160 : }
2161 :
2162 : #[derive(Debug, Clone, Copy, enum_map::Enum)]
2163 : enum EvictionCancelled {
2164 : LayerGone,
2165 : TimelineGone,
2166 : VersionCheckFailed,
2167 : FileNotFound,
2168 : RemoveFailed,
2169 : AlreadyReinitialized,
2170 : /// Not evicted because of a pending reinitialization
2171 : LostToDownload,
2172 : /// After eviction, there was a new layer access which cancelled the eviction.
2173 : UpgradedBackOnAccess,
2174 : UnexpectedEvictedState,
2175 : }
2176 :
2177 : impl EvictionCancelled {
2178 324 : fn as_str(&self) -> &'static str {
2179 324 : match self {
2180 36 : EvictionCancelled::LayerGone => "layer_gone",
2181 36 : EvictionCancelled::TimelineGone => "timeline_gone",
2182 36 : EvictionCancelled::VersionCheckFailed => "version_check_fail",
2183 36 : EvictionCancelled::FileNotFound => "file_not_found",
2184 36 : EvictionCancelled::RemoveFailed => "remove_failed",
2185 36 : EvictionCancelled::AlreadyReinitialized => "already_reinitialized",
2186 36 : EvictionCancelled::LostToDownload => "lost_to_download",
2187 36 : EvictionCancelled::UpgradedBackOnAccess => "upgraded_back_on_access",
2188 36 : EvictionCancelled::UnexpectedEvictedState => "unexpected_evicted_state",
2189 : }
2190 324 : }
2191 : }
2192 :
2193 : #[derive(enum_map::Enum)]
2194 : enum DeleteFailed {
2195 : TimelineGone,
2196 : DeleteSchedulingFailed,
2197 : }
2198 :
2199 : impl DeleteFailed {
2200 72 : fn as_str(&self) -> &'static str {
2201 72 : match self {
2202 36 : DeleteFailed::TimelineGone => "timeline_gone",
2203 36 : DeleteFailed::DeleteSchedulingFailed => "delete_scheduling_failed",
2204 : }
2205 72 : }
2206 : }
2207 :
2208 : #[derive(enum_map::Enum)]
2209 : enum RareEvent {
2210 : RemoveOnDropFailed,
2211 : InitCompletedWithoutRequester,
2212 : DownloadFailedWithoutRequester,
2213 : UpgradedWantedEvicted,
2214 : InitWithoutDownload,
2215 : PermanentLoadingFailure,
2216 : EvictedWithWaiters,
2217 : }
2218 :
2219 : impl RareEvent {
2220 252 : fn as_str(&self) -> &'static str {
2221 252 : use RareEvent::*;
2222 252 :
2223 252 : match self {
2224 36 : RemoveOnDropFailed => "remove_on_drop_failed",
2225 36 : InitCompletedWithoutRequester => "init_completed_without",
2226 36 : DownloadFailedWithoutRequester => "download_failed_without",
2227 36 : UpgradedWantedEvicted => "raced_wanted_evicted",
2228 36 : InitWithoutDownload => "init_needed_no_download",
2229 36 : PermanentLoadingFailure => "permanent_loading_failure",
2230 36 : EvictedWithWaiters => "evicted_with_waiters",
2231 : }
2232 252 : }
2233 : }
2234 :
2235 : pub(crate) static LAYER_IMPL_METRICS: once_cell::sync::Lazy<LayerImplMetrics> =
2236 : once_cell::sync::Lazy::new(LayerImplMetrics::default);
|
