Line data Source code
1 : use std::ops::Range;
2 : use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
3 : use std::sync::{Arc, Weak};
4 : use std::time::{Duration, SystemTime};
5 :
6 : use anyhow::Context;
7 : use camino::{Utf8Path, Utf8PathBuf};
8 : use pageserver_api::keyspace::KeySpace;
9 : use pageserver_api::models::HistoricLayerInfo;
10 : use pageserver_api::shard::{ShardIdentity, ShardIndex, TenantShardId};
11 : use tracing::Instrument;
12 : use utils::generation::Generation;
13 : use utils::id::TimelineId;
14 : use utils::lsn::Lsn;
15 : use utils::sync::{gate, heavier_once_cell};
16 :
17 : use super::delta_layer::{self};
18 : use super::image_layer::{self};
19 : use super::{
20 : AsLayerDesc, ImageLayerWriter, LayerAccessStats, LayerAccessStatsReset, LayerName,
21 : LayerVisibilityHint, PersistentLayerDesc, ValuesReconstructState,
22 : };
23 : use crate::config::PageServerConf;
24 : use crate::context::{DownloadBehavior, RequestContext, RequestContextBuilder};
25 : use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
26 : use crate::task_mgr::TaskKind;
27 : use crate::tenant::Timeline;
28 : use crate::tenant::remote_timeline_client::LayerFileMetadata;
29 : use crate::tenant::timeline::{CompactionError, GetVectoredError};
30 :
31 : #[cfg(test)]
32 : mod tests;
33 :
34 : #[cfg(test)]
35 : mod failpoints;
36 :
37 : pub const S3_UPLOAD_LIMIT: u64 = 4_500_000_000;
38 :
39 : /// A Layer contains all data in a "rectangle" consisting of a range of keys and
40 : /// range of LSNs.
41 : ///
42 : /// There are two kinds of layers, in-memory and on-disk layers. In-memory
43 : /// layers are used to ingest incoming WAL, and provide fast access to the
44 : /// recent page versions. On-disk layers are stored as files on disk, and are
45 : /// immutable. This type represents the on-disk kind while in-memory kind are represented by
46 : /// [`InMemoryLayer`].
47 : ///
48 : /// Furthermore, there are two kinds of on-disk layers: delta and image layers.
49 : /// A delta layer contains all modifications within a range of LSNs and keys.
50 : /// An image layer is a snapshot of all the data in a key-range, at a single
51 : /// LSN.
52 : ///
53 : /// This type models the on-disk layers, which can be evicted and on-demand downloaded. As a
54 : /// general goal, read accesses should always win eviction and eviction should not wait for
55 : /// download.
56 : ///
57 : /// ### State transitions
58 : ///
59 : /// The internal state of `Layer` is composed of most importantly the on-filesystem state and the
60 : /// [`ResidentOrWantedEvicted`] enum. On-filesystem state can be either present (fully downloaded,
61 : /// right size) or deleted.
62 : ///
63 : /// Reads will always win requests to evict until `wait_for_turn_and_evict` has acquired the
64 : /// `heavier_once_cell::InitPermit` and has started to `evict_blocking`. Before the
65 : /// `heavier_once_cell::InitPermit` has been acquired, any read request
66 : /// (`get_or_maybe_download`) can "re-initialize" using the existing downloaded file and thus
67 : /// cancelling the eviction.
68 : ///
69 : /// ```text
70 : /// +-----------------+ get_or_maybe_download +--------------------------------+
71 : /// | not initialized |--------------------------->| Resident(Arc<DownloadedLayer>) |
72 : /// | ENOENT | /->| |
73 : /// +-----------------+ | +--------------------------------+
74 : /// ^ | | ^
75 : /// | get_or_maybe_download | | | get_or_maybe_download, either:
76 : /// evict_blocking | /-------------------------/ | | - upgrade weak to strong
77 : /// | | | | - re-initialize without download
78 : /// | | evict_and_wait | |
79 : /// +-----------------+ v |
80 : /// | not initialized | on_downloaded_layer_drop +--------------------------------------+
81 : /// | file is present |<---------------------------| WantedEvicted(Weak<DownloadedLayer>) |
82 : /// +-----------------+ +--------------------------------------+
83 : /// ```
84 : ///
85 : /// ### Unsupported
86 : ///
87 : /// - Evicting by the operator deleting files from the filesystem
88 : ///
89 : /// [`InMemoryLayer`]: super::inmemory_layer::InMemoryLayer
90 : #[derive(Clone)]
91 : pub(crate) struct Layer(Arc<LayerInner>);
92 :
93 : impl std::fmt::Display for Layer {
94 4312 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
95 4312 : write!(
96 4312 : f,
97 4312 : "{}{}",
98 4312 : self.layer_desc().short_id(),
99 4312 : self.0.generation.get_suffix()
100 4312 : )
101 4312 : }
102 : }
103 :
104 : impl std::fmt::Debug for Layer {
105 8 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
106 8 : write!(f, "{}", self)
107 8 : }
108 : }
109 :
110 : impl AsLayerDesc for Layer {
111 3422044 : fn layer_desc(&self) -> &PersistentLayerDesc {
112 3422044 : self.0.layer_desc()
113 3422044 : }
114 : }
115 :
116 : impl PartialEq for Layer {
117 5 : fn eq(&self, other: &Self) -> bool {
118 5 : Arc::as_ptr(&self.0) == Arc::as_ptr(&other.0)
119 5 : }
120 : }
121 :
122 3828 : pub(crate) fn local_layer_path(
123 3828 : conf: &PageServerConf,
124 3828 : tenant_shard_id: &TenantShardId,
125 3828 : timeline_id: &TimelineId,
126 3828 : layer_file_name: &LayerName,
127 3828 : generation: &Generation,
128 3828 : ) -> Utf8PathBuf {
129 3828 : let timeline_path = conf.timeline_path(tenant_shard_id, timeline_id);
130 3828 :
131 3828 : if generation.is_none() {
132 : // Without a generation, we may only use legacy path style
133 0 : timeline_path.join(layer_file_name.to_string())
134 : } else {
135 3828 : timeline_path.join(format!("{}-v1{}", layer_file_name, generation.get_suffix()))
136 : }
137 3828 : }
138 :
139 : pub(crate) enum LastEviction {
140 : Never,
141 : At(std::time::Instant),
142 : Evicting,
143 : }
144 :
145 : impl LastEviction {
146 52 : pub(crate) fn happened_after(&self, timepoint: std::time::Instant) -> bool {
147 52 : match self {
148 0 : LastEviction::Never => false,
149 52 : LastEviction::At(evicted_at) => evicted_at > &timepoint,
150 0 : LastEviction::Evicting => true,
151 : }
152 52 : }
153 : }
154 :
155 : impl Layer {
156 : /// Creates a layer value for a file we know to not be resident.
157 0 : pub(crate) fn for_evicted(
158 0 : conf: &'static PageServerConf,
159 0 : timeline: &Arc<Timeline>,
160 0 : file_name: LayerName,
161 0 : metadata: LayerFileMetadata,
162 0 : ) -> Self {
163 0 : let local_path = local_layer_path(
164 0 : conf,
165 0 : &timeline.tenant_shard_id,
166 0 : &timeline.timeline_id,
167 0 : &file_name,
168 0 : &metadata.generation,
169 0 : );
170 0 :
171 0 : let desc = PersistentLayerDesc::from_filename(
172 0 : timeline.tenant_shard_id,
173 0 : timeline.timeline_id,
174 0 : file_name,
175 0 : metadata.file_size,
176 0 : );
177 0 :
178 0 : let owner = Layer(Arc::new(LayerInner::new(
179 0 : conf,
180 0 : timeline,
181 0 : local_path,
182 0 : desc,
183 0 : None,
184 0 : metadata.generation,
185 0 : metadata.shard,
186 0 : )));
187 0 :
188 0 : debug_assert!(owner.0.needs_download_blocking().unwrap().is_some());
189 :
190 0 : owner
191 0 : }
192 :
193 : /// Creates a Layer value for a file we know to be resident in timeline directory.
194 244 : pub(crate) fn for_resident(
195 244 : conf: &'static PageServerConf,
196 244 : timeline: &Arc<Timeline>,
197 244 : local_path: Utf8PathBuf,
198 244 : file_name: LayerName,
199 244 : metadata: LayerFileMetadata,
200 244 : ) -> ResidentLayer {
201 244 : let desc = PersistentLayerDesc::from_filename(
202 244 : timeline.tenant_shard_id,
203 244 : timeline.timeline_id,
204 244 : file_name,
205 244 : metadata.file_size,
206 244 : );
207 244 :
208 244 : let mut resident = None;
209 244 :
210 244 : let owner = Layer(Arc::new_cyclic(|owner| {
211 244 : let inner = Arc::new(DownloadedLayer {
212 244 : owner: owner.clone(),
213 244 : kind: tokio::sync::OnceCell::default(),
214 244 : version: 0,
215 244 : });
216 244 : resident = Some(inner.clone());
217 244 :
218 244 : LayerInner::new(
219 244 : conf,
220 244 : timeline,
221 244 : local_path,
222 244 : desc,
223 244 : Some(inner),
224 244 : metadata.generation,
225 244 : metadata.shard,
226 244 : )
227 244 : }));
228 244 :
229 244 : let downloaded = resident.expect("just initialized");
230 244 :
231 244 : debug_assert!(owner.0.needs_download_blocking().unwrap().is_none());
232 :
233 244 : timeline
234 244 : .metrics
235 244 : .resident_physical_size_add(metadata.file_size);
236 244 :
237 244 : ResidentLayer { downloaded, owner }
238 244 : }
239 :
240 : /// Creates a Layer value for freshly written out new layer file by renaming it from a
241 : /// temporary path.
242 3616 : pub(crate) fn finish_creating(
243 3616 : conf: &'static PageServerConf,
244 3616 : timeline: &Arc<Timeline>,
245 3616 : desc: PersistentLayerDesc,
246 3616 : temp_path: &Utf8Path,
247 3616 : ) -> anyhow::Result<ResidentLayer> {
248 3616 : let mut resident = None;
249 3616 :
250 3616 : let owner = Layer(Arc::new_cyclic(|owner| {
251 3616 : let inner = Arc::new(DownloadedLayer {
252 3616 : owner: owner.clone(),
253 3616 : kind: tokio::sync::OnceCell::default(),
254 3616 : version: 0,
255 3616 : });
256 3616 : resident = Some(inner.clone());
257 3616 :
258 3616 : let local_path = local_layer_path(
259 3616 : conf,
260 3616 : &timeline.tenant_shard_id,
261 3616 : &timeline.timeline_id,
262 3616 : &desc.layer_name(),
263 3616 : &timeline.generation,
264 3616 : );
265 3616 :
266 3616 : LayerInner::new(
267 3616 : conf,
268 3616 : timeline,
269 3616 : local_path,
270 3616 : desc,
271 3616 : Some(inner),
272 3616 : timeline.generation,
273 3616 : timeline.get_shard_index(),
274 3616 : )
275 3616 : }));
276 3616 :
277 3616 : let downloaded = resident.expect("just initialized");
278 3616 :
279 3616 : // We never want to overwrite an existing file, so we use `RENAME_NOREPLACE`.
280 3616 : // TODO: this leaves the temp file in place if the rename fails, risking us running
281 3616 : // out of space. Should we clean it up here or does the calling context deal with this?
282 3616 : utils::fs_ext::rename_noreplace(temp_path.as_std_path(), owner.local_path().as_std_path())
283 3616 : .with_context(|| format!("rename temporary file as correct path for {owner}"))?;
284 :
285 3616 : Ok(ResidentLayer { downloaded, owner })
286 3616 : }
287 :
288 : /// Requests the layer to be evicted and waits for this to be done.
289 : ///
290 : /// If the file is not resident, an [`EvictionError::NotFound`] is returned.
291 : ///
292 : /// If for a bad luck or blocking of the executor, we miss the actual eviction and the layer is
293 : /// re-downloaded, [`EvictionError::Downloaded`] is returned.
294 : ///
295 : /// Timeout is mandatory, because waiting for eviction is only needed for our tests; eviction
296 : /// will happen regardless the future returned by this method completing unless there is a
297 : /// read access before eviction gets to complete.
298 : ///
299 : /// Technically cancellation safe, but cancelling might shift the viewpoint of what generation
300 : /// of download-evict cycle on retry.
301 104 : pub(crate) async fn evict_and_wait(&self, timeout: Duration) -> Result<(), EvictionError> {
302 104 : self.0.evict_and_wait(timeout).await
303 96 : }
304 :
305 : /// Delete the layer file when the `self` gets dropped, also try to schedule a remote index upload
306 : /// then.
307 : ///
308 : /// On drop, this will cause a call to [`crate::tenant::remote_timeline_client::RemoteTimelineClient::schedule_deletion_of_unlinked`].
309 : /// This means that the unlinking by [gc] or [compaction] must have happened strictly before
310 : /// the value this is called on gets dropped.
311 : ///
312 : /// This is ensured by both of those methods accepting references to Layer.
313 : ///
314 : /// [gc]: [`RemoteTimelineClient::schedule_gc_update`]
315 : /// [compaction]: [`RemoteTimelineClient::schedule_compaction_update`]
316 1040 : pub(crate) fn delete_on_drop(&self) {
317 1040 : self.0.delete_on_drop();
318 1040 : }
319 :
320 479472 : pub(crate) async fn get_values_reconstruct_data(
321 479472 : &self,
322 479472 : keyspace: KeySpace,
323 479472 : lsn_range: Range<Lsn>,
324 479472 : reconstruct_data: &mut ValuesReconstructState,
325 479472 : ctx: &RequestContext,
326 479472 : ) -> Result<(), GetVectoredError> {
327 479472 : let downloaded =
328 479472 : self.0
329 479472 : .get_or_maybe_download(true, ctx)
330 479472 : .await
331 479472 : .map_err(|err| match err {
332 : DownloadError::TimelineShutdown | DownloadError::DownloadCancelled => {
333 0 : GetVectoredError::Cancelled
334 : }
335 0 : other => GetVectoredError::Other(anyhow::anyhow!(other)),
336 479472 : })?;
337 479472 : let this = ResidentLayer {
338 479472 : downloaded: downloaded.clone(),
339 479472 : owner: self.clone(),
340 479472 : };
341 479472 :
342 479472 : self.record_access(ctx);
343 479472 :
344 479472 : downloaded
345 479472 : .get_values_reconstruct_data(this, keyspace, lsn_range, reconstruct_data, ctx)
346 479472 : .instrument(tracing::debug_span!("get_values_reconstruct_data", layer=%self))
347 479472 : .await
348 479472 : .map_err(|err| match err {
349 0 : GetVectoredError::Other(err) => GetVectoredError::Other(
350 0 : err.context(format!("get_values_reconstruct_data for layer {self}")),
351 0 : ),
352 0 : err => err,
353 479472 : })
354 479472 : }
355 :
356 : /// Download the layer if evicted.
357 : ///
358 : /// Will not error when the layer is already downloaded.
359 0 : pub(crate) async fn download(&self, ctx: &RequestContext) -> Result<(), DownloadError> {
360 0 : self.0.get_or_maybe_download(true, ctx).await?;
361 0 : Ok(())
362 0 : }
363 :
364 608 : pub(crate) async fn needs_download(&self) -> Result<Option<NeedsDownload>, std::io::Error> {
365 608 : self.0.needs_download().await
366 608 : }
367 :
368 : /// Assuming the layer is already downloaded, returns a guard which will prohibit eviction
369 : /// while the guard exists.
370 : ///
371 : /// Returns None if the layer is currently evicted or becoming evicted.
372 40 : pub(crate) async fn keep_resident(&self) -> Option<ResidentLayer> {
373 40 : let downloaded = self.0.inner.get().and_then(|rowe| rowe.get())?;
374 :
375 28 : Some(ResidentLayer {
376 28 : downloaded,
377 28 : owner: self.clone(),
378 28 : })
379 40 : }
380 :
381 : /// Weak indicator of is the layer resident or not. Good enough for eviction, which can deal
382 : /// with `EvictionError::NotFound`.
383 : ///
384 : /// Returns `true` if this layer might be resident, or `false`, if it most likely evicted or
385 : /// will be unless a read happens soon.
386 357 : pub(crate) fn is_likely_resident(&self) -> bool {
387 357 : self.0
388 357 : .inner
389 357 : .get()
390 357 : .map(|rowe| rowe.is_likely_resident())
391 357 : .unwrap_or(false)
392 357 : }
393 :
394 : /// Downloads if necessary and creates a guard, which will keep this layer from being evicted.
395 1148 : pub(crate) async fn download_and_keep_resident(
396 1148 : &self,
397 1148 : ctx: &RequestContext,
398 1148 : ) -> Result<ResidentLayer, DownloadError> {
399 1148 : let downloaded = self.0.get_or_maybe_download(true, ctx).await?;
400 :
401 1148 : Ok(ResidentLayer {
402 1148 : downloaded,
403 1148 : owner: self.clone(),
404 1148 : })
405 1148 : }
406 :
407 0 : pub(crate) fn info(&self, reset: LayerAccessStatsReset) -> HistoricLayerInfo {
408 0 : self.0.info(reset)
409 0 : }
410 :
411 68 : pub(crate) fn latest_activity(&self) -> SystemTime {
412 68 : self.0.access_stats.latest_activity()
413 68 : }
414 :
415 184 : pub(crate) fn visibility(&self) -> LayerVisibilityHint {
416 184 : self.0.access_stats.visibility()
417 184 : }
418 :
419 3620 : pub(crate) fn local_path(&self) -> &Utf8Path {
420 3620 : &self.0.path
421 3620 : }
422 :
423 5264 : pub(crate) fn metadata(&self) -> LayerFileMetadata {
424 5264 : self.0.metadata()
425 5264 : }
426 :
427 52 : pub(crate) fn last_evicted_at(&self) -> LastEviction {
428 52 : match self.0.last_evicted_at.try_lock() {
429 52 : Ok(lock) => match *lock {
430 0 : None => LastEviction::Never,
431 52 : Some(at) => LastEviction::At(at),
432 : },
433 0 : Err(std::sync::TryLockError::WouldBlock) => LastEviction::Evicting,
434 0 : Err(std::sync::TryLockError::Poisoned(p)) => panic!("Lock poisoned: {p}"),
435 : }
436 52 : }
437 :
438 0 : pub(crate) fn get_timeline_id(&self) -> Option<TimelineId> {
439 0 : self.0
440 0 : .timeline
441 0 : .upgrade()
442 0 : .map(|timeline| timeline.timeline_id)
443 0 : }
444 :
445 : /// Traditional debug dumping facility
446 : #[allow(unused)]
447 8 : pub(crate) async fn dump(&self, verbose: bool, ctx: &RequestContext) -> anyhow::Result<()> {
448 8 : self.0.desc.dump();
449 8 :
450 8 : if verbose {
451 : // for now, unconditionally download everything, even if that might not be wanted.
452 8 : let l = self.0.get_or_maybe_download(true, ctx).await?;
453 8 : l.dump(&self.0, ctx).await?
454 0 : }
455 :
456 8 : Ok(())
457 8 : }
458 :
459 : /// Waits until this layer has been dropped (and if needed, local file deletion and remote
460 : /// deletion scheduling has completed).
461 : ///
462 : /// Does not start local deletion, use [`Self::delete_on_drop`] for that
463 : /// separatedly.
464 : #[cfg(any(feature = "testing", test))]
465 4 : pub(crate) fn wait_drop(&self) -> impl std::future::Future<Output = ()> + 'static {
466 4 : let mut rx = self.0.status.as_ref().unwrap().subscribe();
467 :
468 4 : async move {
469 : loop {
470 4 : if rx.changed().await.is_err() {
471 4 : break;
472 0 : }
473 : }
474 4 : }
475 4 : }
476 :
477 479472 : fn record_access(&self, ctx: &RequestContext) {
478 479472 : if self.0.access_stats.record_access(ctx) {
479 : // Visibility was modified to Visible: maybe log about this
480 0 : match ctx.task_kind() {
481 : TaskKind::CalculateSyntheticSize
482 : | TaskKind::OndemandLogicalSizeCalculation
483 : | TaskKind::GarbageCollector
484 0 : | TaskKind::MgmtRequest => {
485 0 : // This situation is expected in code paths do binary searches of the LSN space to resolve
486 0 : // an LSN to a timestamp, which happens during GC, during GC cutoff calculations in synthetic size,
487 0 : // and on-demand for certain HTTP API requests. On-demand logical size calculation is also included
488 0 : // because it is run as a sub-task of synthetic size.
489 0 : }
490 : _ => {
491 : // In all other contexts, it is unusual to do I/O involving layers which are not visible at
492 : // some branch tip, so we log the fact that we are accessing something that the visibility
493 : // calculation thought should not be visible.
494 : //
495 : // This case is legal in brief time windows: for example an in-flight getpage request can hold on to a layer object
496 : // which was covered by a concurrent compaction.
497 0 : tracing::info!(
498 : layer=%self,
499 0 : "became visible as a result of access",
500 : );
501 : }
502 : }
503 :
504 : // Update the timeline's visible bytes count
505 0 : if let Some(tl) = self.0.timeline.upgrade() {
506 0 : tl.metrics
507 0 : .visible_physical_size_gauge
508 0 : .add(self.0.desc.file_size)
509 0 : }
510 479472 : }
511 479472 : }
512 :
513 716 : pub(crate) fn set_visibility(&self, visibility: LayerVisibilityHint) {
514 716 : let old_visibility = self.0.access_stats.set_visibility(visibility.clone());
515 : use LayerVisibilityHint::*;
516 716 : match (old_visibility, visibility) {
517 : (Visible, Covered) => {
518 : // Subtract this layer's contribution to the visible size metric
519 72 : if let Some(tl) = self.0.timeline.upgrade() {
520 72 : debug_assert!(
521 72 : tl.metrics.visible_physical_size_gauge.get() >= self.0.desc.file_size
522 : );
523 72 : tl.metrics
524 72 : .visible_physical_size_gauge
525 72 : .sub(self.0.desc.file_size)
526 0 : }
527 : }
528 : (Covered, Visible) => {
529 : // Add this layer's contribution to the visible size metric
530 0 : if let Some(tl) = self.0.timeline.upgrade() {
531 0 : tl.metrics
532 0 : .visible_physical_size_gauge
533 0 : .add(self.0.desc.file_size)
534 0 : }
535 : }
536 644 : (Covered, Covered) | (Visible, Visible) => {
537 644 : // no change
538 644 : }
539 : }
540 716 : }
541 : }
542 :
543 : /// The download-ness ([`DownloadedLayer`]) can be either resident or wanted evicted.
544 : ///
545 : /// However when we want something evicted, we cannot evict it right away as there might be current
546 : /// reads happening on it. For example: it has been searched from [`LayerMap::search`] but not yet
547 : /// read with [`Layer::get_values_reconstruct_data`].
548 : ///
549 : /// [`LayerMap::search`]: crate::tenant::layer_map::LayerMap::search
550 : #[derive(Debug)]
551 : enum ResidentOrWantedEvicted {
552 : Resident(Arc<DownloadedLayer>),
553 : WantedEvicted(Weak<DownloadedLayer>, usize),
554 : }
555 :
556 : impl ResidentOrWantedEvicted {
557 : /// Non-mutating access to the a DownloadedLayer, if possible.
558 : ///
559 : /// This is not used on the read path (anything that calls
560 : /// [`LayerInner::get_or_maybe_download`]) because it was decided that reads always win
561 : /// evictions, and part of that winning is using [`ResidentOrWantedEvicted::get_and_upgrade`].
562 28 : fn get(&self) -> Option<Arc<DownloadedLayer>> {
563 28 : match self {
564 28 : ResidentOrWantedEvicted::Resident(strong) => Some(strong.clone()),
565 0 : ResidentOrWantedEvicted::WantedEvicted(weak, _) => weak.upgrade(),
566 : }
567 28 : }
568 :
569 : /// Best-effort query for residency right now, not as strong guarantee as receiving a strong
570 : /// reference from `ResidentOrWantedEvicted::get`.
571 221 : fn is_likely_resident(&self) -> bool {
572 221 : match self {
573 209 : ResidentOrWantedEvicted::Resident(_) => true,
574 12 : ResidentOrWantedEvicted::WantedEvicted(weak, _) => weak.strong_count() > 0,
575 : }
576 221 : }
577 :
578 : /// Upgrades any weak to strong if possible.
579 : ///
580 : /// Returns a strong reference if possible, along with a boolean telling if an upgrade
581 : /// happened.
582 480620 : fn get_and_upgrade(&mut self) -> Option<(Arc<DownloadedLayer>, bool)> {
583 480620 : match self {
584 480604 : ResidentOrWantedEvicted::Resident(strong) => Some((strong.clone(), false)),
585 16 : ResidentOrWantedEvicted::WantedEvicted(weak, _) => match weak.upgrade() {
586 0 : Some(strong) => {
587 0 : LAYER_IMPL_METRICS.inc_raced_wanted_evicted_accesses();
588 0 :
589 0 : *self = ResidentOrWantedEvicted::Resident(strong.clone());
590 0 :
591 0 : Some((strong, true))
592 : }
593 16 : None => None,
594 : },
595 : }
596 480620 : }
597 :
598 : /// When eviction is first requested, drop down to holding a [`Weak`].
599 : ///
600 : /// Returns `Some` if this was the first time eviction was requested. Care should be taken to
601 : /// drop the possibly last strong reference outside of the mutex of
602 : /// [`heavier_once_cell::OnceCell`].
603 92 : fn downgrade(&mut self) -> Option<Arc<DownloadedLayer>> {
604 92 : match self {
605 84 : ResidentOrWantedEvicted::Resident(strong) => {
606 84 : let weak = Arc::downgrade(strong);
607 84 : let mut temp = ResidentOrWantedEvicted::WantedEvicted(weak, strong.version);
608 84 : std::mem::swap(self, &mut temp);
609 84 : match temp {
610 84 : ResidentOrWantedEvicted::Resident(strong) => Some(strong),
611 0 : ResidentOrWantedEvicted::WantedEvicted(..) => unreachable!("just swapped"),
612 : }
613 : }
614 8 : ResidentOrWantedEvicted::WantedEvicted(..) => None,
615 : }
616 92 : }
617 : }
618 :
619 : struct LayerInner {
620 : /// Only needed to check ondemand_download_behavior_treat_error_as_warn and creation of
621 : /// [`Self::path`].
622 : conf: &'static PageServerConf,
623 :
624 : /// Full path to the file; unclear if this should exist anymore.
625 : path: Utf8PathBuf,
626 :
627 : desc: PersistentLayerDesc,
628 :
629 : /// Timeline access is needed for remote timeline client and metrics.
630 : ///
631 : /// There should not be an access to timeline for any reason without entering the
632 : /// [`Timeline::gate`] at the same time.
633 : timeline: Weak<Timeline>,
634 :
635 : access_stats: LayerAccessStats,
636 :
637 : /// This custom OnceCell is backed by std mutex, but only held for short time periods.
638 : ///
639 : /// Filesystem changes (download, evict) are only done while holding a permit which the
640 : /// `heavier_once_cell` provides.
641 : ///
642 : /// A number of fields in `Layer` are meant to only be updated when holding the InitPermit, but
643 : /// possibly read while not holding it.
644 : inner: heavier_once_cell::OnceCell<ResidentOrWantedEvicted>,
645 :
646 : /// Do we want to delete locally and remotely this when `LayerInner` is dropped
647 : wanted_deleted: AtomicBool,
648 :
649 : /// Version is to make sure we will only evict a specific initialization of the downloaded file.
650 : ///
651 : /// Incremented for each initialization, stored in `DownloadedLayer::version` or
652 : /// `ResidentOrWantedEvicted::WantedEvicted`.
653 : version: AtomicUsize,
654 :
655 : /// Allow subscribing to when the layer actually gets evicted, a non-cancellable download
656 : /// starts, or completes.
657 : ///
658 : /// Updates must only be posted while holding the InitPermit or the heavier_once_cell::Guard.
659 : /// Holding the InitPermit is the only time we can do state transitions, but we also need to
660 : /// cancel a pending eviction on upgrading a [`ResidentOrWantedEvicted::WantedEvicted`] back to
661 : /// [`ResidentOrWantedEvicted::Resident`] on access.
662 : ///
663 : /// The sender is wrapped in an Option to facilitate moving it out on [`LayerInner::drop`].
664 : status: Option<tokio::sync::watch::Sender<Status>>,
665 :
666 : /// Counter for exponential backoff with the download.
667 : ///
668 : /// This is atomic only for the purposes of having additional data only accessed while holding
669 : /// the InitPermit.
670 : consecutive_failures: AtomicUsize,
671 :
672 : /// The generation of this Layer.
673 : ///
674 : /// For loaded layers (resident or evicted) this comes from [`LayerFileMetadata::generation`],
675 : /// for created layers from [`Timeline::generation`].
676 : generation: Generation,
677 :
678 : /// The shard of this Layer.
679 : ///
680 : /// For layers created in this process, this will always be the [`ShardIndex`] of the
681 : /// current `ShardIdentity`` (TODO: add link once it's introduced).
682 : ///
683 : /// For loaded layers, this may be some other value if the tenant has undergone
684 : /// a shard split since the layer was originally written.
685 : shard: ShardIndex,
686 :
687 : /// When the Layer was last evicted but has not been downloaded since.
688 : ///
689 : /// This is used for skipping evicted layers from the previous heatmap (see
690 : /// `[Timeline::generate_heatmap]`) and for updating metrics
691 : /// (see [`LayerImplMetrics::redownload_after`]).
692 : last_evicted_at: std::sync::Mutex<Option<std::time::Instant>>,
693 :
694 : #[cfg(test)]
695 : failpoints: std::sync::Mutex<Vec<failpoints::Failpoint>>,
696 : }
697 :
698 : impl std::fmt::Display for LayerInner {
699 108 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
700 108 : write!(f, "{}", self.layer_desc().short_id())
701 108 : }
702 : }
703 :
704 : impl AsLayerDesc for LayerInner {
705 3427297 : fn layer_desc(&self) -> &PersistentLayerDesc {
706 3427297 : &self.desc
707 3427297 : }
708 : }
709 :
710 : #[derive(Debug, Clone, Copy)]
711 : enum Status {
712 : Resident,
713 : Evicted,
714 : Downloading,
715 : }
716 :
717 : impl Drop for LayerInner {
718 1421 : fn drop(&mut self) {
719 1421 : // if there was a pending eviction, mark it cancelled here to balance metrics
720 1421 : if let Some((ResidentOrWantedEvicted::WantedEvicted(..), _)) = self.inner.take_and_deinit()
721 4 : {
722 4 : // eviction has already been started
723 4 : LAYER_IMPL_METRICS.inc_eviction_cancelled(EvictionCancelled::LayerGone);
724 4 :
725 4 : // eviction request is intentionally not honored as no one is present to wait for it
726 4 : // and we could be delaying shutdown for nothing.
727 1417 : }
728 :
729 1421 : let timeline = self.timeline.upgrade();
730 :
731 1421 : if let Some(timeline) = timeline.as_ref() {
732 : // Only need to decrement metrics if the timeline still exists: otherwise
733 : // it will have already de-registered these metrics via TimelineMetrics::shutdown
734 1389 : timeline.metrics.dec_layer(&self.desc);
735 :
736 1389 : if matches!(self.access_stats.visibility(), LayerVisibilityHint::Visible) {
737 1389 : debug_assert!(
738 1389 : timeline.metrics.visible_physical_size_gauge.get() >= self.desc.file_size
739 : );
740 1389 : timeline
741 1389 : .metrics
742 1389 : .visible_physical_size_gauge
743 1389 : .sub(self.desc.file_size);
744 0 : }
745 32 : }
746 :
747 1421 : if !*self.wanted_deleted.get_mut() {
748 392 : return;
749 1029 : }
750 :
751 1029 : 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);
752 :
753 1029 : let path = std::mem::take(&mut self.path);
754 1029 : let file_name = self.layer_desc().layer_name();
755 1029 : let file_size = self.layer_desc().file_size;
756 1029 : let meta = self.metadata();
757 1029 : let status = self.status.take();
758 1029 :
759 1029 : Self::spawn_blocking(move || {
760 1024 : let _g = span.entered();
761 1024 :
762 1024 : // carry this until we are finished for [`Layer::wait_drop`] support
763 1024 : let _status = status;
764 :
765 1024 : let Some(timeline) = timeline else {
766 : // no need to nag that timeline is gone: under normal situation on
767 : // task_mgr::remove_tenant_from_memory the timeline is gone before we get dropped.
768 0 : LAYER_IMPL_METRICS.inc_deletes_failed(DeleteFailed::TimelineGone);
769 0 : return;
770 : };
771 :
772 1024 : let Ok(_guard) = timeline.gate.enter() else {
773 0 : LAYER_IMPL_METRICS.inc_deletes_failed(DeleteFailed::TimelineGone);
774 0 : return;
775 : };
776 :
777 1024 : let removed = match std::fs::remove_file(path) {
778 1020 : Ok(()) => true,
779 4 : Err(e) if e.kind() == std::io::ErrorKind::NotFound => {
780 4 : // until we no longer do detaches by removing all local files before removing the
781 4 : // tenant from the global map, we will always get these errors even if we knew what
782 4 : // is the latest state.
783 4 : //
784 4 : // we currently do not track the latest state, so we'll also end up here on evicted
785 4 : // layers.
786 4 : false
787 : }
788 0 : Err(e) => {
789 0 : tracing::error!("failed to remove wanted deleted layer: {e}");
790 0 : LAYER_IMPL_METRICS.inc_delete_removes_failed();
791 0 : false
792 : }
793 : };
794 :
795 1024 : if removed {
796 1020 : timeline.metrics.resident_physical_size_sub(file_size);
797 1020 : }
798 1024 : let res = timeline
799 1024 : .remote_client
800 1024 : .schedule_deletion_of_unlinked(vec![(file_name, meta)]);
801 :
802 1024 : if let Err(e) = res {
803 : // test_timeline_deletion_with_files_stuck_in_upload_queue is good at
804 : // demonstrating this deadlock (without spawn_blocking): stop will drop
805 : // queued items, which will have ResidentLayer's, and those drops would try
806 : // to re-entrantly lock the RemoteTimelineClient inner state.
807 4 : if !timeline.is_active() {
808 4 : tracing::info!("scheduling deletion on drop failed: {e:#}");
809 : } else {
810 0 : tracing::warn!("scheduling deletion on drop failed: {e:#}");
811 : }
812 0 : LAYER_IMPL_METRICS.inc_deletes_failed(DeleteFailed::DeleteSchedulingFailed);
813 1020 : } else {
814 1020 : LAYER_IMPL_METRICS.inc_completed_deletes();
815 1020 : }
816 1029 : });
817 1421 : }
818 : }
819 :
820 : impl LayerInner {
821 : #[allow(clippy::too_many_arguments)]
822 3860 : fn new(
823 3860 : conf: &'static PageServerConf,
824 3860 : timeline: &Arc<Timeline>,
825 3860 : local_path: Utf8PathBuf,
826 3860 : desc: PersistentLayerDesc,
827 3860 : downloaded: Option<Arc<DownloadedLayer>>,
828 3860 : generation: Generation,
829 3860 : shard: ShardIndex,
830 3860 : ) -> Self {
831 3860 : let (inner, version, init_status) = if let Some(inner) = downloaded {
832 3860 : let version = inner.version;
833 3860 : let resident = ResidentOrWantedEvicted::Resident(inner);
834 3860 : (
835 3860 : heavier_once_cell::OnceCell::new(resident),
836 3860 : version,
837 3860 : Status::Resident,
838 3860 : )
839 : } else {
840 0 : (heavier_once_cell::OnceCell::default(), 0, Status::Evicted)
841 : };
842 :
843 : // This object acts as a RAII guard on these metrics: increment on construction
844 3860 : timeline.metrics.inc_layer(&desc);
845 3860 :
846 3860 : // New layers are visible by default. This metric is later updated on drop or in set_visibility
847 3860 : timeline
848 3860 : .metrics
849 3860 : .visible_physical_size_gauge
850 3860 : .add(desc.file_size);
851 3860 :
852 3860 : LayerInner {
853 3860 : conf,
854 3860 : path: local_path,
855 3860 : desc,
856 3860 : timeline: Arc::downgrade(timeline),
857 3860 : access_stats: Default::default(),
858 3860 : wanted_deleted: AtomicBool::new(false),
859 3860 : inner,
860 3860 : version: AtomicUsize::new(version),
861 3860 : status: Some(tokio::sync::watch::channel(init_status).0),
862 3860 : consecutive_failures: AtomicUsize::new(0),
863 3860 : generation,
864 3860 : shard,
865 3860 : last_evicted_at: std::sync::Mutex::default(),
866 3860 : #[cfg(test)]
867 3860 : failpoints: Default::default(),
868 3860 : }
869 3860 : }
870 :
871 1040 : fn delete_on_drop(&self) {
872 1040 : let res =
873 1040 : self.wanted_deleted
874 1040 : .compare_exchange(false, true, Ordering::Release, Ordering::Relaxed);
875 1040 :
876 1040 : if res.is_ok() {
877 1032 : LAYER_IMPL_METRICS.inc_started_deletes();
878 1032 : }
879 1040 : }
880 :
881 : /// Cancellation safe, however dropping the future and calling this method again might result
882 : /// in a new attempt to evict OR join the previously started attempt.
883 104 : #[tracing::instrument(level = tracing::Level::DEBUG, skip_all, ret, err(level = tracing::Level::DEBUG), fields(layer=%self))]
884 : pub(crate) async fn evict_and_wait(&self, timeout: Duration) -> Result<(), EvictionError> {
885 : let mut rx = self.status.as_ref().unwrap().subscribe();
886 :
887 : {
888 : let current = rx.borrow_and_update();
889 : match &*current {
890 : Status::Resident => {
891 : // we might get lucky and evict this; continue
892 : }
893 : Status::Evicted | Status::Downloading => {
894 : // it is already evicted
895 : return Err(EvictionError::NotFound);
896 : }
897 : }
898 : }
899 :
900 : let strong = {
901 : match self.inner.get() {
902 : Some(mut either) => either.downgrade(),
903 : None => {
904 : // we already have a scheduled eviction, which just has not gotten to run yet.
905 : // it might still race with a read access, but that could also get cancelled,
906 : // so let's say this is not evictable.
907 : return Err(EvictionError::NotFound);
908 : }
909 : }
910 : };
911 :
912 : if strong.is_some() {
913 : // drop the DownloadedLayer outside of the holding the guard
914 : drop(strong);
915 :
916 : // idea here is that only one evicter should ever get to witness a strong reference,
917 : // which means whenever get_or_maybe_download upgrades a weak, it must mark up a
918 : // cancelled eviction and signal us, like it currently does.
919 : //
920 : // a second concurrent evict_and_wait will not see a strong reference.
921 : LAYER_IMPL_METRICS.inc_started_evictions();
922 : }
923 :
924 : let changed = rx.changed();
925 : let changed = tokio::time::timeout(timeout, changed).await;
926 :
927 : let Ok(changed) = changed else {
928 : return Err(EvictionError::Timeout);
929 : };
930 :
931 : let _: () = changed.expect("cannot be closed, because we are holding a strong reference");
932 :
933 : let current = rx.borrow_and_update();
934 :
935 : match &*current {
936 : // the easiest case
937 : Status::Evicted => Ok(()),
938 : // it surely was evicted in between, but then there was a new access now; we can't know
939 : // if it'll succeed so lets just call it evicted
940 : Status::Downloading => Ok(()),
941 : // either the download which was started after eviction completed already, or it was
942 : // never evicted
943 : Status::Resident => Err(EvictionError::Downloaded),
944 : }
945 : }
946 :
947 : /// Cancellation safe.
948 480652 : async fn get_or_maybe_download(
949 480652 : self: &Arc<Self>,
950 480652 : allow_download: bool,
951 480652 : ctx: &RequestContext,
952 480652 : ) -> Result<Arc<DownloadedLayer>, DownloadError> {
953 48 : let (weak, permit) = {
954 : // get_or_init_detached can:
955 : // - be fast (mutex lock) OR uncontested semaphore permit acquire
956 : // - be slow (wait for semaphore permit or closing)
957 480652 : let init_cancelled = scopeguard::guard((), |_| LAYER_IMPL_METRICS.inc_init_cancelled());
958 :
959 480652 : let locked = self
960 480652 : .inner
961 480652 : .get_or_init_detached()
962 480652 : .await
963 480652 : .map(|mut guard| guard.get_and_upgrade().ok_or(guard));
964 480652 :
965 480652 : scopeguard::ScopeGuard::into_inner(init_cancelled);
966 :
967 480604 : match locked {
968 : // this path could had been a RwLock::read
969 480604 : Ok(Ok((strong, upgraded))) if !upgraded => return Ok(strong),
970 0 : Ok(Ok((strong, _))) => {
971 0 : // when upgraded back, the Arc<DownloadedLayer> is still available, but
972 0 : // previously a `evict_and_wait` was received. this is the only place when we
973 0 : // send out an update without holding the InitPermit.
974 0 : //
975 0 : // note that we also have dropped the Guard; this is fine, because we just made
976 0 : // a state change and are holding a strong reference to be returned.
977 0 : self.status.as_ref().unwrap().send_replace(Status::Resident);
978 0 : LAYER_IMPL_METRICS
979 0 : .inc_eviction_cancelled(EvictionCancelled::UpgradedBackOnAccess);
980 0 :
981 0 : return Ok(strong);
982 : }
983 16 : Ok(Err(guard)) => {
984 16 : // path to here: we won the eviction, the file should still be on the disk.
985 16 : let (weak, permit) = guard.take_and_deinit();
986 16 : (Some(weak), permit)
987 : }
988 32 : Err(permit) => (None, permit),
989 : }
990 : };
991 :
992 48 : if let Some(weak) = weak {
993 : // only drop the weak after dropping the heavier_once_cell guard
994 16 : assert!(
995 16 : matches!(weak, ResidentOrWantedEvicted::WantedEvicted(..)),
996 0 : "unexpected {weak:?}, ResidentOrWantedEvicted::get_and_upgrade has a bug"
997 : );
998 32 : }
999 :
1000 48 : let timeline = self
1001 48 : .timeline
1002 48 : .upgrade()
1003 48 : .ok_or(DownloadError::TimelineShutdown)?;
1004 :
1005 : // count cancellations, which currently remain largely unexpected
1006 48 : let init_cancelled = scopeguard::guard((), |_| LAYER_IMPL_METRICS.inc_init_cancelled());
1007 :
1008 : // check if we really need to be downloaded: this can happen if a read access won the
1009 : // semaphore before eviction.
1010 : //
1011 : // if we are cancelled while doing this `stat` the `self.inner` will be uninitialized. a
1012 : // pending eviction will try to evict even upon finding an uninitialized `self.inner`.
1013 48 : let needs_download = self
1014 48 : .needs_download()
1015 48 : .await
1016 48 : .map_err(DownloadError::PreStatFailed);
1017 48 :
1018 48 : scopeguard::ScopeGuard::into_inner(init_cancelled);
1019 :
1020 48 : let needs_download = needs_download?;
1021 :
1022 48 : let Some(reason) = needs_download else {
1023 : // the file is present locally because eviction has not had a chance to run yet
1024 :
1025 : #[cfg(test)]
1026 16 : self.failpoint(failpoints::FailpointKind::AfterDeterminingLayerNeedsNoDownload)
1027 16 : .await?;
1028 :
1029 12 : LAYER_IMPL_METRICS.inc_init_needed_no_download();
1030 12 :
1031 12 : return Ok(self.initialize_after_layer_is_on_disk(permit));
1032 : };
1033 :
1034 : // we must download; getting cancelled before spawning the download is not an issue as
1035 : // any still running eviction would not find anything to evict.
1036 :
1037 32 : if let NeedsDownload::NotFile(ft) = reason {
1038 0 : return Err(DownloadError::NotFile(ft));
1039 32 : }
1040 32 :
1041 32 : self.check_expected_download(ctx)?;
1042 :
1043 32 : if !allow_download {
1044 : // this is only used from tests, but it is hard to test without the boolean
1045 4 : return Err(DownloadError::DownloadRequired);
1046 28 : }
1047 28 :
1048 28 : let download_ctx = ctx.detached_child(TaskKind::LayerDownload, DownloadBehavior::Download);
1049 :
1050 28 : async move {
1051 28 : tracing::info!(%reason, "downloading on-demand");
1052 :
1053 28 : let init_cancelled = scopeguard::guard((), |_| LAYER_IMPL_METRICS.inc_init_cancelled());
1054 28 : let res = self
1055 28 : .download_init_and_wait(timeline, permit, download_ctx)
1056 28 : .await?;
1057 28 : scopeguard::ScopeGuard::into_inner(init_cancelled);
1058 28 : Ok(res)
1059 28 : }
1060 28 : .instrument(tracing::info_span!("get_or_maybe_download", layer=%self))
1061 28 : .await
1062 480652 : }
1063 :
1064 : /// Nag or fail per RequestContext policy
1065 32 : fn check_expected_download(&self, ctx: &RequestContext) -> Result<(), DownloadError> {
1066 : use crate::context::DownloadBehavior::*;
1067 32 : let b = ctx.download_behavior();
1068 32 : match b {
1069 32 : Download => Ok(()),
1070 : Warn | Error => {
1071 0 : tracing::info!(
1072 0 : "unexpectedly on-demand downloading for task kind {:?}",
1073 0 : ctx.task_kind()
1074 : );
1075 0 : crate::metrics::UNEXPECTED_ONDEMAND_DOWNLOADS.inc();
1076 :
1077 0 : let really_error =
1078 0 : matches!(b, Error) && !self.conf.ondemand_download_behavior_treat_error_as_warn;
1079 :
1080 0 : if really_error {
1081 : // this check is only probablistic, seems like flakyness footgun
1082 0 : Err(DownloadError::ContextAndConfigReallyDeniesDownloads)
1083 : } else {
1084 0 : Ok(())
1085 : }
1086 : }
1087 : }
1088 32 : }
1089 :
1090 : /// Actual download, at most one is executed at the time.
1091 28 : async fn download_init_and_wait(
1092 28 : self: &Arc<Self>,
1093 28 : timeline: Arc<Timeline>,
1094 28 : permit: heavier_once_cell::InitPermit,
1095 28 : ctx: RequestContext,
1096 28 : ) -> Result<Arc<DownloadedLayer>, DownloadError> {
1097 28 : debug_assert_current_span_has_tenant_and_timeline_id();
1098 28 :
1099 28 : let (tx, rx) = tokio::sync::oneshot::channel();
1100 28 :
1101 28 : let this: Arc<Self> = self.clone();
1102 :
1103 28 : let guard = timeline
1104 28 : .gate
1105 28 : .enter()
1106 28 : .map_err(|_| DownloadError::DownloadCancelled)?;
1107 :
1108 28 : Self::spawn(
1109 28 : async move {
1110 0 : let _guard = guard;
1111 0 :
1112 0 : // now that we have commited to downloading, send out an update to:
1113 0 : // - unhang any pending eviction
1114 0 : // - break out of evict_and_wait
1115 0 : this.status
1116 0 : .as_ref()
1117 0 : .unwrap()
1118 0 : .send_replace(Status::Downloading);
1119 28 :
1120 28 : #[cfg(test)]
1121 28 : this.failpoint(failpoints::FailpointKind::WaitBeforeDownloading)
1122 28 : .await
1123 28 : .unwrap();
1124 :
1125 28 : let res = this.download_and_init(timeline, permit, &ctx).await;
1126 :
1127 28 : if let Err(res) = tx.send(res) {
1128 0 : match res {
1129 0 : Ok(_res) => {
1130 0 : tracing::debug!("layer initialized, but caller has been cancelled");
1131 0 : LAYER_IMPL_METRICS.inc_init_completed_without_requester();
1132 : }
1133 0 : Err(e) => {
1134 0 : tracing::info!(
1135 0 : "layer file download failed, and caller has been cancelled: {e:?}"
1136 : );
1137 0 : LAYER_IMPL_METRICS.inc_download_failed_without_requester();
1138 : }
1139 : }
1140 28 : }
1141 28 : }
1142 28 : .in_current_span(),
1143 28 : );
1144 28 :
1145 28 : match rx.await {
1146 28 : Ok(Ok(res)) => Ok(res),
1147 : Ok(Err(remote_storage::DownloadError::Cancelled)) => {
1148 0 : Err(DownloadError::DownloadCancelled)
1149 : }
1150 0 : Ok(Err(_)) => Err(DownloadError::DownloadFailed),
1151 0 : Err(_gone) => Err(DownloadError::DownloadCancelled),
1152 : }
1153 28 : }
1154 :
1155 28 : async fn download_and_init(
1156 28 : self: &Arc<LayerInner>,
1157 28 : timeline: Arc<Timeline>,
1158 28 : permit: heavier_once_cell::InitPermit,
1159 28 : ctx: &RequestContext,
1160 28 : ) -> Result<Arc<DownloadedLayer>, remote_storage::DownloadError> {
1161 28 : let result = timeline
1162 28 : .remote_client
1163 28 : .download_layer_file(
1164 28 : &self.desc.layer_name(),
1165 28 : &self.metadata(),
1166 28 : &self.path,
1167 28 : &timeline.gate,
1168 28 : &timeline.cancel,
1169 28 : ctx,
1170 28 : )
1171 28 : .await;
1172 :
1173 28 : match result {
1174 28 : Ok(size) => {
1175 28 : assert_eq!(size, self.desc.file_size);
1176 :
1177 28 : match self.needs_download().await {
1178 0 : Ok(Some(reason)) => {
1179 0 : // this is really a bug in needs_download or remote timeline client
1180 0 : panic!("post-condition failed: needs_download returned {reason:?}");
1181 : }
1182 28 : Ok(None) => {
1183 28 : // as expected
1184 28 : }
1185 0 : Err(e) => {
1186 0 : panic!("post-condition failed: needs_download errored: {e:?}");
1187 : }
1188 : }
1189 :
1190 28 : tracing::info!(size=%self.desc.file_size, "on-demand download successful");
1191 28 : timeline
1192 28 : .metrics
1193 28 : .resident_physical_size_add(self.desc.file_size);
1194 28 : self.consecutive_failures.store(0, Ordering::Relaxed);
1195 28 :
1196 28 : let since_last_eviction = self
1197 28 : .last_evicted_at
1198 28 : .lock()
1199 28 : .unwrap()
1200 28 : .take()
1201 28 : .map(|ts| ts.elapsed());
1202 28 : if let Some(since_last_eviction) = since_last_eviction {
1203 28 : LAYER_IMPL_METRICS.record_redownloaded_after(since_last_eviction);
1204 28 : }
1205 :
1206 28 : self.access_stats.record_residence_event();
1207 28 :
1208 28 : Ok(self.initialize_after_layer_is_on_disk(permit))
1209 : }
1210 0 : Err(e) => {
1211 0 : let consecutive_failures =
1212 0 : 1 + self.consecutive_failures.fetch_add(1, Ordering::Relaxed);
1213 0 :
1214 0 : if timeline.cancel.is_cancelled() {
1215 : // If we're shutting down, drop out before logging the error
1216 0 : return Err(e);
1217 0 : }
1218 0 :
1219 0 : tracing::error!(consecutive_failures, "layer file download failed: {e:#}");
1220 :
1221 0 : let backoff = utils::backoff::exponential_backoff_duration_seconds(
1222 0 : consecutive_failures.min(u32::MAX as usize) as u32,
1223 0 : 1.5,
1224 0 : 60.0,
1225 0 : );
1226 0 :
1227 0 : let backoff = std::time::Duration::from_secs_f64(backoff);
1228 0 :
1229 0 : tokio::select! {
1230 0 : _ = tokio::time::sleep(backoff) => {},
1231 0 : _ = timeline.cancel.cancelled() => {},
1232 : };
1233 :
1234 0 : Err(e)
1235 : }
1236 : }
1237 28 : }
1238 :
1239 : /// Initializes the `Self::inner` to a "resident" state.
1240 : ///
1241 : /// Callers are assumed to ensure that the file is actually on disk with `Self::needs_download`
1242 : /// before calling this method.
1243 : ///
1244 : /// If this method is ever made async, it needs to be cancellation safe so that no state
1245 : /// changes are made before we can write to the OnceCell in non-cancellable fashion.
1246 40 : fn initialize_after_layer_is_on_disk(
1247 40 : self: &Arc<LayerInner>,
1248 40 : permit: heavier_once_cell::InitPermit,
1249 40 : ) -> Arc<DownloadedLayer> {
1250 40 : debug_assert_current_span_has_tenant_and_timeline_id();
1251 40 :
1252 40 : // disable any scheduled but not yet running eviction deletions for this initialization
1253 40 : let next_version = 1 + self.version.fetch_add(1, Ordering::Relaxed);
1254 40 : self.status.as_ref().unwrap().send_replace(Status::Resident);
1255 40 :
1256 40 : let res = Arc::new(DownloadedLayer {
1257 40 : owner: Arc::downgrade(self),
1258 40 : kind: tokio::sync::OnceCell::default(),
1259 40 : version: next_version,
1260 40 : });
1261 40 :
1262 40 : let waiters = self.inner.initializer_count();
1263 40 : if waiters > 0 {
1264 0 : tracing::info!(waiters, "completing layer init for other tasks");
1265 40 : }
1266 :
1267 40 : let value = ResidentOrWantedEvicted::Resident(res.clone());
1268 40 :
1269 40 : self.inner.set(value, permit);
1270 40 :
1271 40 : res
1272 40 : }
1273 :
1274 688 : async fn needs_download(&self) -> Result<Option<NeedsDownload>, std::io::Error> {
1275 688 : match tokio::fs::metadata(&self.path).await {
1276 656 : Ok(m) => Ok(self.is_file_present_and_good_size(&m).err()),
1277 32 : Err(e) if e.kind() == std::io::ErrorKind::NotFound => Ok(Some(NeedsDownload::NotFound)),
1278 0 : Err(e) => Err(e),
1279 : }
1280 688 : }
1281 :
1282 244 : fn needs_download_blocking(&self) -> Result<Option<NeedsDownload>, std::io::Error> {
1283 244 : match self.path.metadata() {
1284 244 : Ok(m) => Ok(self.is_file_present_and_good_size(&m).err()),
1285 0 : Err(e) if e.kind() == std::io::ErrorKind::NotFound => Ok(Some(NeedsDownload::NotFound)),
1286 0 : Err(e) => Err(e),
1287 : }
1288 244 : }
1289 :
1290 900 : fn is_file_present_and_good_size(&self, m: &std::fs::Metadata) -> Result<(), NeedsDownload> {
1291 900 : // in future, this should include sha2-256 validation of the file.
1292 900 : if !m.is_file() {
1293 0 : Err(NeedsDownload::NotFile(m.file_type()))
1294 900 : } else if m.len() != self.desc.file_size {
1295 0 : Err(NeedsDownload::WrongSize {
1296 0 : actual: m.len(),
1297 0 : expected: self.desc.file_size,
1298 0 : })
1299 : } else {
1300 900 : Ok(())
1301 : }
1302 900 : }
1303 :
1304 0 : fn info(&self, reset: LayerAccessStatsReset) -> HistoricLayerInfo {
1305 0 : let layer_name = self.desc.layer_name().to_string();
1306 0 :
1307 0 : let resident = self
1308 0 : .inner
1309 0 : .get()
1310 0 : .map(|rowe| rowe.is_likely_resident())
1311 0 : .unwrap_or(false);
1312 0 :
1313 0 : let access_stats = self.access_stats.as_api_model(reset);
1314 0 :
1315 0 : if self.desc.is_delta {
1316 0 : let lsn_range = &self.desc.lsn_range;
1317 0 :
1318 0 : HistoricLayerInfo::Delta {
1319 0 : layer_file_name: layer_name,
1320 0 : layer_file_size: self.desc.file_size,
1321 0 : lsn_start: lsn_range.start,
1322 0 : lsn_end: lsn_range.end,
1323 0 : remote: !resident,
1324 0 : access_stats,
1325 0 : l0: crate::tenant::layer_map::LayerMap::is_l0(
1326 0 : &self.layer_desc().key_range,
1327 0 : self.layer_desc().is_delta,
1328 0 : ),
1329 0 : }
1330 : } else {
1331 0 : let lsn = self.desc.image_layer_lsn();
1332 0 :
1333 0 : HistoricLayerInfo::Image {
1334 0 : layer_file_name: layer_name,
1335 0 : layer_file_size: self.desc.file_size,
1336 0 : lsn_start: lsn,
1337 0 : remote: !resident,
1338 0 : access_stats,
1339 0 : }
1340 : }
1341 0 : }
1342 :
1343 : /// `DownloadedLayer` is being dropped, so it calls this method.
1344 80 : fn on_downloaded_layer_drop(self: Arc<LayerInner>, only_version: usize) {
1345 : // we cannot know without inspecting LayerInner::inner if we should evict or not, even
1346 : // though here it is very likely
1347 80 : 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);
1348 :
1349 : // NOTE: this scope *must* never call `self.inner.get` because evict_and_wait might
1350 : // drop while the `self.inner` is being locked, leading to a deadlock.
1351 :
1352 80 : let start_evicting = async move {
1353 80 : #[cfg(test)]
1354 80 : self.failpoint(failpoints::FailpointKind::WaitBeforeStartingEvicting)
1355 80 : .await
1356 80 : .expect("failpoint should not have errored");
1357 80 :
1358 80 : tracing::debug!("eviction started");
1359 :
1360 80 : let res = self.wait_for_turn_and_evict(only_version).await;
1361 : // metrics: ignore the Ok branch, it is not done yet
1362 80 : if let Err(e) = res {
1363 12 : tracing::debug!(res=?Err::<(), _>(&e), "eviction completed");
1364 12 : LAYER_IMPL_METRICS.inc_eviction_cancelled(e);
1365 68 : }
1366 80 : };
1367 :
1368 80 : Self::spawn(start_evicting.instrument(span));
1369 80 : }
1370 :
1371 80 : async fn wait_for_turn_and_evict(
1372 80 : self: Arc<LayerInner>,
1373 80 : only_version: usize,
1374 80 : ) -> Result<(), EvictionCancelled> {
1375 156 : fn is_good_to_continue(status: &Status) -> Result<(), EvictionCancelled> {
1376 : use Status::*;
1377 156 : match status {
1378 152 : Resident => Ok(()),
1379 4 : Evicted => Err(EvictionCancelled::UnexpectedEvictedState),
1380 0 : Downloading => Err(EvictionCancelled::LostToDownload),
1381 : }
1382 156 : }
1383 :
1384 80 : let timeline = self
1385 80 : .timeline
1386 80 : .upgrade()
1387 80 : .ok_or(EvictionCancelled::TimelineGone)?;
1388 :
1389 80 : let mut rx = self
1390 80 : .status
1391 80 : .as_ref()
1392 80 : .expect("LayerInner cannot be dropped, holding strong ref")
1393 80 : .subscribe();
1394 80 :
1395 80 : is_good_to_continue(&rx.borrow_and_update())?;
1396 :
1397 76 : let Ok(gate) = timeline.gate.enter() else {
1398 0 : return Err(EvictionCancelled::TimelineGone);
1399 : };
1400 :
1401 68 : let permit = {
1402 : // we cannot just `std::fs::remove_file` because there might already be an
1403 : // get_or_maybe_download which will inspect filesystem and reinitialize. filesystem
1404 : // operations must be done while holding the heavier_once_cell::InitPermit
1405 76 : let mut wait = std::pin::pin!(self.inner.get_or_init_detached());
1406 :
1407 76 : let waited = loop {
1408 : // we must race to the Downloading starting, otherwise we would have to wait until the
1409 : // completion of the download. waiting for download could be long and hinder our
1410 : // efforts to alert on "hanging" evictions.
1411 76 : tokio::select! {
1412 76 : res = &mut wait => break res,
1413 76 : _ = rx.changed() => {
1414 0 : is_good_to_continue(&rx.borrow_and_update())?;
1415 : // two possibilities for Status::Resident:
1416 : // - the layer was found locally from disk by a read
1417 : // - we missed a bunch of updates and now the layer is
1418 : // again downloaded -- assume we'll fail later on with
1419 : // version check or AlreadyReinitialized
1420 : }
1421 : }
1422 : };
1423 :
1424 : // re-check now that we have the guard or permit; all updates should have happened
1425 : // while holding the permit.
1426 76 : is_good_to_continue(&rx.borrow_and_update())?;
1427 :
1428 : // the term deinitialize is used here, because we clearing out the Weak will eventually
1429 : // lead to deallocating the reference counted value, and the value we
1430 : // `Guard::take_and_deinit` is likely to be the last because the Weak is never cloned.
1431 76 : let (_weak, permit) = match waited {
1432 72 : Ok(guard) => {
1433 72 : match &*guard {
1434 68 : ResidentOrWantedEvicted::WantedEvicted(_weak, version)
1435 68 : if *version == only_version =>
1436 64 : {
1437 64 : tracing::debug!(version, "deinitializing matching WantedEvicted");
1438 64 : let (weak, permit) = guard.take_and_deinit();
1439 64 : (Some(weak), permit)
1440 : }
1441 4 : ResidentOrWantedEvicted::WantedEvicted(_, version) => {
1442 4 : // if we were not doing the version check, we would need to try to
1443 4 : // upgrade the weak here to see if it really is dropped. version check
1444 4 : // is done instead assuming that it is cheaper.
1445 4 : tracing::debug!(
1446 : version,
1447 : only_version,
1448 0 : "version mismatch, not deinitializing"
1449 : );
1450 4 : return Err(EvictionCancelled::VersionCheckFailed);
1451 : }
1452 : ResidentOrWantedEvicted::Resident(_) => {
1453 4 : return Err(EvictionCancelled::AlreadyReinitialized);
1454 : }
1455 : }
1456 : }
1457 4 : Err(permit) => {
1458 4 : tracing::debug!("continuing after cancelled get_or_maybe_download or eviction");
1459 4 : (None, permit)
1460 : }
1461 : };
1462 :
1463 68 : permit
1464 68 : };
1465 68 :
1466 68 : let span = tracing::Span::current();
1467 68 :
1468 68 : let spawned_at = std::time::Instant::now();
1469 68 :
1470 68 : // this is on purpose a detached spawn; we don't need to wait for it
1471 68 : //
1472 68 : // eviction completion reporting is the only thing hinging on this, and it can be just as
1473 68 : // well from a spawn_blocking thread.
1474 68 : //
1475 68 : // important to note that now that we've acquired the permit we have made sure the evicted
1476 68 : // file is either the exact `WantedEvicted` we wanted to evict, or uninitialized in case
1477 68 : // there are multiple evictions. The rest is not cancellable, and we've now commited to
1478 68 : // evicting.
1479 68 : //
1480 68 : // If spawn_blocking has a queue and maximum number of threads are in use, we could stall
1481 68 : // reads. We will need to add cancellation for that if necessary.
1482 68 : Self::spawn_blocking(move || {
1483 68 : let _span = span.entered();
1484 68 :
1485 68 : let res = self.evict_blocking(&timeline, &gate, &permit);
1486 68 :
1487 68 : let waiters = self.inner.initializer_count();
1488 68 :
1489 68 : if waiters > 0 {
1490 0 : LAYER_IMPL_METRICS.inc_evicted_with_waiters();
1491 68 : }
1492 :
1493 68 : let completed_in = spawned_at.elapsed();
1494 68 : LAYER_IMPL_METRICS.record_time_to_evict(completed_in);
1495 68 :
1496 68 : match res {
1497 68 : Ok(()) => LAYER_IMPL_METRICS.inc_completed_evictions(),
1498 0 : Err(e) => LAYER_IMPL_METRICS.inc_eviction_cancelled(e),
1499 : }
1500 :
1501 68 : tracing::debug!(?res, elapsed_ms=%completed_in.as_millis(), %waiters, "eviction completed");
1502 68 : });
1503 68 :
1504 68 : Ok(())
1505 80 : }
1506 :
1507 : /// This is blocking only to do just one spawn_blocking hop compared to multiple via tokio::fs.
1508 68 : fn evict_blocking(
1509 68 : &self,
1510 68 : timeline: &Timeline,
1511 68 : _gate: &gate::GateGuard,
1512 68 : _permit: &heavier_once_cell::InitPermit,
1513 68 : ) -> Result<(), EvictionCancelled> {
1514 68 : // now accesses to `self.inner.get_or_init*` wait on the semaphore or the `_permit`
1515 68 :
1516 68 : match capture_mtime_and_remove(&self.path) {
1517 68 : Ok(local_layer_mtime) => {
1518 68 : let duration = SystemTime::now().duration_since(local_layer_mtime);
1519 68 : match duration {
1520 68 : Ok(elapsed) => {
1521 68 : let accessed_and_visible = self.access_stats.accessed()
1522 8 : && self.access_stats.visibility() == LayerVisibilityHint::Visible;
1523 68 : if accessed_and_visible {
1524 8 : // Only layers used for reads contribute to our "low residence" metric that is used
1525 8 : // to detect thrashing. Layers promoted for other reasons (e.g. compaction) are allowed
1526 8 : // to be rapidly evicted without contributing to this metric.
1527 8 : timeline
1528 8 : .metrics
1529 8 : .evictions_with_low_residence_duration
1530 8 : .read()
1531 8 : .unwrap()
1532 8 : .observe(elapsed);
1533 60 : }
1534 :
1535 68 : tracing::info!(
1536 0 : residence_millis = elapsed.as_millis(),
1537 0 : accessed_and_visible,
1538 0 : "evicted layer after known residence period"
1539 : );
1540 : }
1541 : Err(_) => {
1542 0 : tracing::info!("evicted layer after unknown residence period");
1543 : }
1544 : }
1545 68 : timeline.metrics.evictions.inc();
1546 68 : timeline
1547 68 : .metrics
1548 68 : .resident_physical_size_sub(self.desc.file_size);
1549 : }
1550 0 : Err(e) if e.kind() == std::io::ErrorKind::NotFound => {
1551 0 : tracing::error!(
1552 : layer_size = %self.desc.file_size,
1553 0 : "failed to evict layer from disk, it was already gone"
1554 : );
1555 0 : return Err(EvictionCancelled::FileNotFound);
1556 : }
1557 0 : Err(e) => {
1558 0 : // FIXME: this should probably be an abort
1559 0 : tracing::error!("failed to evict file from disk: {e:#}");
1560 0 : return Err(EvictionCancelled::RemoveFailed);
1561 : }
1562 : }
1563 :
1564 68 : self.access_stats.record_residence_event();
1565 68 :
1566 68 : *self.last_evicted_at.lock().unwrap() = Some(std::time::Instant::now());
1567 68 :
1568 68 : self.status.as_ref().unwrap().send_replace(Status::Evicted);
1569 68 :
1570 68 : Ok(())
1571 68 : }
1572 :
1573 6321 : fn metadata(&self) -> LayerFileMetadata {
1574 6321 : LayerFileMetadata::new(self.desc.file_size, self.generation, self.shard)
1575 6321 : }
1576 :
1577 : /// Needed to use entered runtime in tests, but otherwise use BACKGROUND_RUNTIME.
1578 : ///
1579 : /// Synchronizing with spawned tasks is very complicated otherwise.
1580 108 : fn spawn<F>(fut: F)
1581 108 : where
1582 108 : F: std::future::Future<Output = ()> + Send + 'static,
1583 108 : {
1584 108 : #[cfg(test)]
1585 108 : tokio::task::spawn(fut);
1586 108 : #[cfg(not(test))]
1587 108 : crate::task_mgr::BACKGROUND_RUNTIME.spawn(fut);
1588 108 : }
1589 :
1590 : /// Needed to use entered runtime in tests, but otherwise use BACKGROUND_RUNTIME.
1591 1097 : fn spawn_blocking<F>(f: F)
1592 1097 : where
1593 1097 : F: FnOnce() + Send + 'static,
1594 1097 : {
1595 1097 : #[cfg(test)]
1596 1097 : tokio::task::spawn_blocking(f);
1597 1097 : #[cfg(not(test))]
1598 1097 : crate::task_mgr::BACKGROUND_RUNTIME.spawn_blocking(f);
1599 1097 : }
1600 : }
1601 :
1602 68 : fn capture_mtime_and_remove(path: &Utf8Path) -> Result<SystemTime, std::io::Error> {
1603 68 : let m = path.metadata()?;
1604 68 : let local_layer_mtime = m.modified()?;
1605 68 : std::fs::remove_file(path)?;
1606 68 : Ok(local_layer_mtime)
1607 68 : }
1608 :
1609 : #[derive(Debug, thiserror::Error)]
1610 : pub(crate) enum EvictionError {
1611 : #[error("layer was already evicted")]
1612 : NotFound,
1613 :
1614 : /// Evictions must always lose to downloads in races, and this time it happened.
1615 : #[error("layer was downloaded instead")]
1616 : Downloaded,
1617 :
1618 : #[error("eviction did not happen within timeout")]
1619 : Timeout,
1620 : }
1621 :
1622 : /// Error internal to the [`LayerInner::get_or_maybe_download`]
1623 : #[derive(Debug, thiserror::Error)]
1624 : pub(crate) enum DownloadError {
1625 : #[error("timeline has already shutdown")]
1626 : TimelineShutdown,
1627 : #[error("context denies downloading")]
1628 : ContextAndConfigReallyDeniesDownloads,
1629 : #[error("downloading is really required but not allowed by this method")]
1630 : DownloadRequired,
1631 : #[error("layer path exists, but it is not a file: {0:?}")]
1632 : NotFile(std::fs::FileType),
1633 : /// Why no error here? Because it will be reported by page_service. We should had also done
1634 : /// retries already.
1635 : #[error("downloading evicted layer file failed")]
1636 : DownloadFailed,
1637 : #[error("downloading failed, possibly for shutdown")]
1638 : DownloadCancelled,
1639 : #[error("pre-condition: stat before download failed")]
1640 : PreStatFailed(#[source] std::io::Error),
1641 :
1642 : #[cfg(test)]
1643 : #[error("failpoint: {0:?}")]
1644 : Failpoint(failpoints::FailpointKind),
1645 : }
1646 :
1647 : impl DownloadError {
1648 0 : pub(crate) fn is_cancelled(&self) -> bool {
1649 0 : matches!(self, DownloadError::DownloadCancelled)
1650 0 : }
1651 : }
1652 :
1653 : #[derive(Debug, PartialEq)]
1654 : pub(crate) enum NeedsDownload {
1655 : NotFound,
1656 : NotFile(std::fs::FileType),
1657 : WrongSize { actual: u64, expected: u64 },
1658 : }
1659 :
1660 : impl std::fmt::Display for NeedsDownload {
1661 28 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
1662 28 : match self {
1663 28 : NeedsDownload::NotFound => write!(f, "file was not found"),
1664 0 : NeedsDownload::NotFile(ft) => write!(f, "path is not a file; {ft:?}"),
1665 0 : NeedsDownload::WrongSize { actual, expected } => {
1666 0 : write!(f, "file size mismatch {actual} vs. {expected}")
1667 : }
1668 : }
1669 28 : }
1670 : }
1671 :
1672 : /// Existence of `DownloadedLayer` means that we have the file locally, and can later evict it.
1673 : pub(crate) struct DownloadedLayer {
1674 : owner: Weak<LayerInner>,
1675 : // Use tokio OnceCell as we do not need to deinitialize this, it'll just get dropped with the
1676 : // DownloadedLayer
1677 : kind: tokio::sync::OnceCell<anyhow::Result<LayerKind>>,
1678 : version: usize,
1679 : }
1680 :
1681 : impl std::fmt::Debug for DownloadedLayer {
1682 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
1683 0 : f.debug_struct("DownloadedLayer")
1684 0 : // owner omitted because it is always "Weak"
1685 0 : .field("kind", &self.kind)
1686 0 : .field("version", &self.version)
1687 0 : .finish()
1688 0 : }
1689 : }
1690 :
1691 : impl Drop for DownloadedLayer {
1692 1497 : fn drop(&mut self) {
1693 1497 : if let Some(owner) = self.owner.upgrade() {
1694 80 : owner.on_downloaded_layer_drop(self.version);
1695 1417 : } else {
1696 1417 : // Layer::drop will handle cancelling the eviction; because of drop order and
1697 1417 : // `DownloadedLayer` never leaking, we cannot know here if eviction was requested.
1698 1417 : }
1699 1497 : }
1700 : }
1701 :
1702 : impl DownloadedLayer {
1703 : /// Initializes the `DeltaLayerInner` or `ImageLayerInner` within [`LayerKind`].
1704 : /// Failure to load the layer is sticky, i.e., future `get()` calls will return
1705 : /// the initial load failure immediately.
1706 : ///
1707 : /// `owner` parameter is a strong reference at the same `LayerInner` as the
1708 : /// `DownloadedLayer::owner` would be when upgraded. Given how this method ends up called,
1709 : /// we will always have the LayerInner on the callstack, so we can just use it.
1710 481560 : async fn get<'a>(
1711 481560 : &'a self,
1712 481560 : owner: &Arc<LayerInner>,
1713 481560 : ctx: &RequestContext,
1714 481560 : ) -> anyhow::Result<&'a LayerKind> {
1715 481560 : let init = || async {
1716 2464 : assert_eq!(
1717 2464 : Weak::as_ptr(&self.owner),
1718 2464 : Arc::as_ptr(owner),
1719 0 : "these are the same, just avoiding the upgrade"
1720 : );
1721 :
1722 2464 : let res = if owner.desc.is_delta {
1723 2188 : let ctx = RequestContextBuilder::extend(ctx)
1724 2188 : .page_content_kind(crate::context::PageContentKind::DeltaLayerSummary)
1725 2188 : .build();
1726 2188 : let summary = Some(delta_layer::Summary::expected(
1727 2188 : owner.desc.tenant_shard_id.tenant_id,
1728 2188 : owner.desc.timeline_id,
1729 2188 : owner.desc.key_range.clone(),
1730 2188 : owner.desc.lsn_range.clone(),
1731 2188 : ));
1732 2188 : delta_layer::DeltaLayerInner::load(
1733 2188 : &owner.path,
1734 2188 : summary,
1735 2188 : Some(owner.conf.max_vectored_read_bytes),
1736 2188 : &ctx,
1737 2188 : )
1738 2188 : .await
1739 2188 : .map(LayerKind::Delta)
1740 : } else {
1741 276 : let ctx = RequestContextBuilder::extend(ctx)
1742 276 : .page_content_kind(crate::context::PageContentKind::ImageLayerSummary)
1743 276 : .build();
1744 276 : let lsn = owner.desc.image_layer_lsn();
1745 276 : let summary = Some(image_layer::Summary::expected(
1746 276 : owner.desc.tenant_shard_id.tenant_id,
1747 276 : owner.desc.timeline_id,
1748 276 : owner.desc.key_range.clone(),
1749 276 : lsn,
1750 276 : ));
1751 276 : image_layer::ImageLayerInner::load(
1752 276 : &owner.path,
1753 276 : lsn,
1754 276 : summary,
1755 276 : Some(owner.conf.max_vectored_read_bytes),
1756 276 : &ctx,
1757 276 : )
1758 276 : .await
1759 276 : .map(LayerKind::Image)
1760 : };
1761 :
1762 2464 : match res {
1763 2464 : Ok(layer) => Ok(layer),
1764 0 : Err(err) => {
1765 0 : LAYER_IMPL_METRICS.inc_permanent_loading_failures();
1766 0 : // We log this message once over the lifetime of `Self`
1767 0 : // => Ok and good to log backtrace and path here.
1768 0 : tracing::error!(
1769 0 : "layer load failed, assuming permanent failure: {}: {err:?}",
1770 0 : owner.path
1771 : );
1772 0 : Err(err)
1773 : }
1774 : }
1775 4928 : };
1776 481560 : self.kind
1777 481560 : .get_or_init(init)
1778 481560 : .await
1779 481560 : .as_ref()
1780 481560 : // We already logged the full backtrace above, once. Don't repeat that here.
1781 481560 : .map_err(|e| anyhow::anyhow!("layer load failed earlier: {e}"))
1782 481560 : }
1783 :
1784 479472 : async fn get_values_reconstruct_data(
1785 479472 : &self,
1786 479472 : this: ResidentLayer,
1787 479472 : keyspace: KeySpace,
1788 479472 : lsn_range: Range<Lsn>,
1789 479472 : reconstruct_data: &mut ValuesReconstructState,
1790 479472 : ctx: &RequestContext,
1791 479472 : ) -> Result<(), GetVectoredError> {
1792 : use LayerKind::*;
1793 :
1794 479472 : match self
1795 479472 : .get(&this.owner.0, ctx)
1796 479472 : .await
1797 479472 : .map_err(GetVectoredError::Other)?
1798 : {
1799 434348 : Delta(d) => {
1800 434348 : d.get_values_reconstruct_data(this, keyspace, lsn_range, reconstruct_data, ctx)
1801 434348 : .await
1802 : }
1803 45124 : Image(i) => {
1804 45124 : i.get_values_reconstruct_data(this, keyspace, reconstruct_data, ctx)
1805 45124 : .await
1806 : }
1807 : }
1808 479472 : }
1809 :
1810 8 : async fn dump(&self, owner: &Arc<LayerInner>, ctx: &RequestContext) -> anyhow::Result<()> {
1811 : use LayerKind::*;
1812 8 : match self.get(owner, ctx).await? {
1813 8 : Delta(d) => d.dump(ctx).await?,
1814 0 : Image(i) => i.dump(ctx).await?,
1815 : }
1816 :
1817 8 : Ok(())
1818 8 : }
1819 : }
1820 :
1821 : /// Wrapper around an actual layer implementation.
1822 : #[derive(Debug)]
1823 : enum LayerKind {
1824 : Delta(delta_layer::DeltaLayerInner),
1825 : Image(image_layer::ImageLayerInner),
1826 : }
1827 :
1828 : /// Guard for forcing a layer be resident while it exists.
1829 : #[derive(Clone)]
1830 : pub struct ResidentLayer {
1831 : owner: Layer,
1832 : downloaded: Arc<DownloadedLayer>,
1833 : }
1834 :
1835 : impl std::fmt::Display for ResidentLayer {
1836 4288 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
1837 4288 : write!(f, "{}", self.owner)
1838 4288 : }
1839 : }
1840 :
1841 : impl std::fmt::Debug for ResidentLayer {
1842 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
1843 0 : write!(f, "{}", self.owner)
1844 0 : }
1845 : }
1846 :
1847 : impl ResidentLayer {
1848 : /// Release the eviction guard, converting back into a plain [`Layer`].
1849 : ///
1850 : /// You can access the [`Layer`] also by using `as_ref`.
1851 840 : pub(crate) fn drop_eviction_guard(self) -> Layer {
1852 840 : self.into()
1853 840 : }
1854 :
1855 : /// Loads all keys stored in the layer. Returns key, lsn and value size.
1856 0 : #[tracing::instrument(level = tracing::Level::DEBUG, skip_all, fields(layer=%self))]
1857 : pub(crate) async fn load_keys<'a>(
1858 : &'a self,
1859 : ctx: &RequestContext,
1860 : ) -> anyhow::Result<Vec<pageserver_api::key::Key>> {
1861 : use LayerKind::*;
1862 :
1863 : let owner = &self.owner.0;
1864 : let inner = self.downloaded.get(owner, ctx).await?;
1865 :
1866 : // this is valid because the DownloadedLayer::kind is a OnceCell, not a
1867 : // Mutex<OnceCell>, so we cannot go and deinitialize the value with OnceCell::take
1868 : // while it's being held.
1869 : self.owner.record_access(ctx);
1870 :
1871 : let res = match inner {
1872 : Delta(d) => delta_layer::DeltaLayerInner::load_keys(d, ctx).await,
1873 : Image(i) => image_layer::ImageLayerInner::load_keys(i, ctx).await,
1874 : };
1875 0 : res.with_context(|| format!("Layer index is corrupted for {self}"))
1876 : }
1877 :
1878 : /// Read all they keys in this layer which match the ShardIdentity, and write them all to
1879 : /// the provided writer. Return the number of keys written.
1880 16 : #[tracing::instrument(level = tracing::Level::DEBUG, skip_all, fields(layer=%self))]
1881 : pub(crate) async fn filter(
1882 : &self,
1883 : shard_identity: &ShardIdentity,
1884 : writer: &mut ImageLayerWriter,
1885 : ctx: &RequestContext,
1886 : ) -> Result<usize, CompactionError> {
1887 : use LayerKind::*;
1888 :
1889 : match self
1890 : .downloaded
1891 : .get(&self.owner.0, ctx)
1892 : .await
1893 : .map_err(CompactionError::Other)?
1894 : {
1895 : Delta(_) => {
1896 : return Err(CompactionError::Other(anyhow::anyhow!(format!(
1897 : "cannot filter() on a delta layer {self}"
1898 : ))));
1899 : }
1900 : Image(i) => i
1901 : .filter(shard_identity, writer, ctx)
1902 : .await
1903 : .map_err(CompactionError::Other),
1904 : }
1905 : }
1906 :
1907 : /// Returns the amount of keys and values written to the writer.
1908 20 : pub(crate) async fn copy_delta_prefix(
1909 20 : &self,
1910 20 : writer: &mut super::delta_layer::DeltaLayerWriter,
1911 20 : until: Lsn,
1912 20 : ctx: &RequestContext,
1913 20 : ) -> anyhow::Result<usize> {
1914 : use LayerKind::*;
1915 :
1916 20 : let owner = &self.owner.0;
1917 20 :
1918 20 : match self.downloaded.get(owner, ctx).await? {
1919 20 : Delta(d) => d
1920 20 : .copy_prefix(writer, until, ctx)
1921 20 : .await
1922 20 : .with_context(|| format!("copy_delta_prefix until {until} of {self}")),
1923 0 : Image(_) => anyhow::bail!(format!("cannot copy_lsn_prefix of image layer {self}")),
1924 : }
1925 20 : }
1926 :
1927 3678 : pub(crate) fn local_path(&self) -> &Utf8Path {
1928 3678 : &self.owner.0.path
1929 3678 : }
1930 :
1931 4400 : pub(crate) fn metadata(&self) -> LayerFileMetadata {
1932 4400 : self.owner.metadata()
1933 4400 : }
1934 :
1935 : /// Cast the layer to a delta, return an error if it is an image layer.
1936 1908 : pub(crate) async fn get_as_delta(
1937 1908 : &self,
1938 1908 : ctx: &RequestContext,
1939 1908 : ) -> anyhow::Result<&delta_layer::DeltaLayerInner> {
1940 : use LayerKind::*;
1941 1908 : match self.downloaded.get(&self.owner.0, ctx).await? {
1942 1908 : Delta(d) => Ok(d),
1943 0 : Image(_) => Err(anyhow::anyhow!("image layer")),
1944 : }
1945 1908 : }
1946 :
1947 : /// Cast the layer to an image, return an error if it is a delta layer.
1948 136 : pub(crate) async fn get_as_image(
1949 136 : &self,
1950 136 : ctx: &RequestContext,
1951 136 : ) -> anyhow::Result<&image_layer::ImageLayerInner> {
1952 : use LayerKind::*;
1953 136 : match self.downloaded.get(&self.owner.0, ctx).await? {
1954 136 : Image(d) => Ok(d),
1955 0 : Delta(_) => Err(anyhow::anyhow!("delta layer")),
1956 : }
1957 136 : }
1958 : }
1959 :
1960 : impl AsLayerDesc for ResidentLayer {
1961 1961756 : fn layer_desc(&self) -> &PersistentLayerDesc {
1962 1961756 : self.owner.layer_desc()
1963 1961756 : }
1964 : }
1965 :
1966 : impl AsRef<Layer> for ResidentLayer {
1967 4084 : fn as_ref(&self) -> &Layer {
1968 4084 : &self.owner
1969 4084 : }
1970 : }
1971 :
1972 : /// Drop the eviction guard.
1973 : impl From<ResidentLayer> for Layer {
1974 840 : fn from(value: ResidentLayer) -> Self {
1975 840 : value.owner
1976 840 : }
1977 : }
1978 :
1979 : use metrics::IntCounter;
1980 :
1981 : pub(crate) struct LayerImplMetrics {
1982 : started_evictions: IntCounter,
1983 : completed_evictions: IntCounter,
1984 : cancelled_evictions: enum_map::EnumMap<EvictionCancelled, IntCounter>,
1985 :
1986 : started_deletes: IntCounter,
1987 : completed_deletes: IntCounter,
1988 : failed_deletes: enum_map::EnumMap<DeleteFailed, IntCounter>,
1989 :
1990 : rare_counters: enum_map::EnumMap<RareEvent, IntCounter>,
1991 : inits_cancelled: metrics::core::GenericCounter<metrics::core::AtomicU64>,
1992 : redownload_after: metrics::Histogram,
1993 : time_to_evict: metrics::Histogram,
1994 : }
1995 :
1996 : impl Default for LayerImplMetrics {
1997 108 : fn default() -> Self {
1998 : use enum_map::Enum;
1999 :
2000 : // reminder: these will be pageserver_layer_* with "_total" suffix
2001 :
2002 108 : let started_evictions = metrics::register_int_counter!(
2003 108 : "pageserver_layer_started_evictions",
2004 108 : "Evictions started in the Layer implementation"
2005 108 : )
2006 108 : .unwrap();
2007 108 : let completed_evictions = metrics::register_int_counter!(
2008 108 : "pageserver_layer_completed_evictions",
2009 108 : "Evictions completed in the Layer implementation"
2010 108 : )
2011 108 : .unwrap();
2012 108 :
2013 108 : let cancelled_evictions = metrics::register_int_counter_vec!(
2014 108 : "pageserver_layer_cancelled_evictions_count",
2015 108 : "Different reasons for evictions to have been cancelled or failed",
2016 108 : &["reason"]
2017 108 : )
2018 108 : .unwrap();
2019 108 :
2020 972 : let cancelled_evictions = enum_map::EnumMap::from_array(std::array::from_fn(|i| {
2021 972 : let reason = EvictionCancelled::from_usize(i);
2022 972 : let s = reason.as_str();
2023 972 : cancelled_evictions.with_label_values(&[s])
2024 972 : }));
2025 108 :
2026 108 : let started_deletes = metrics::register_int_counter!(
2027 108 : "pageserver_layer_started_deletes",
2028 108 : "Deletions on drop pending in the Layer implementation"
2029 108 : )
2030 108 : .unwrap();
2031 108 : let completed_deletes = metrics::register_int_counter!(
2032 108 : "pageserver_layer_completed_deletes",
2033 108 : "Deletions on drop completed in the Layer implementation"
2034 108 : )
2035 108 : .unwrap();
2036 108 :
2037 108 : let failed_deletes = metrics::register_int_counter_vec!(
2038 108 : "pageserver_layer_failed_deletes_count",
2039 108 : "Different reasons for deletions on drop to have failed",
2040 108 : &["reason"]
2041 108 : )
2042 108 : .unwrap();
2043 108 :
2044 216 : let failed_deletes = enum_map::EnumMap::from_array(std::array::from_fn(|i| {
2045 216 : let reason = DeleteFailed::from_usize(i);
2046 216 : let s = reason.as_str();
2047 216 : failed_deletes.with_label_values(&[s])
2048 216 : }));
2049 108 :
2050 108 : let rare_counters = metrics::register_int_counter_vec!(
2051 108 : "pageserver_layer_assumed_rare_count",
2052 108 : "Times unexpected or assumed rare event happened",
2053 108 : &["event"]
2054 108 : )
2055 108 : .unwrap();
2056 108 :
2057 756 : let rare_counters = enum_map::EnumMap::from_array(std::array::from_fn(|i| {
2058 756 : let event = RareEvent::from_usize(i);
2059 756 : let s = event.as_str();
2060 756 : rare_counters.with_label_values(&[s])
2061 756 : }));
2062 108 :
2063 108 : let inits_cancelled = metrics::register_int_counter!(
2064 108 : "pageserver_layer_inits_cancelled_count",
2065 108 : "Times Layer initialization was cancelled",
2066 108 : )
2067 108 : .unwrap();
2068 108 :
2069 108 : let redownload_after = {
2070 108 : let minute = 60.0;
2071 108 : let hour = 60.0 * minute;
2072 108 : metrics::register_histogram!(
2073 108 : "pageserver_layer_redownloaded_after",
2074 108 : "Time between evicting and re-downloading.",
2075 108 : vec![
2076 108 : 10.0,
2077 108 : 30.0,
2078 108 : minute,
2079 108 : 5.0 * minute,
2080 108 : 15.0 * minute,
2081 108 : 30.0 * minute,
2082 108 : hour,
2083 108 : 12.0 * hour,
2084 108 : ]
2085 108 : )
2086 108 : .unwrap()
2087 108 : };
2088 108 :
2089 108 : let time_to_evict = metrics::register_histogram!(
2090 108 : "pageserver_layer_eviction_held_permit_seconds",
2091 108 : "Time eviction held the permit.",
2092 108 : vec![0.001, 0.010, 0.100, 0.500, 1.000, 5.000]
2093 108 : )
2094 108 : .unwrap();
2095 108 :
2096 108 : Self {
2097 108 : started_evictions,
2098 108 : completed_evictions,
2099 108 : cancelled_evictions,
2100 108 :
2101 108 : started_deletes,
2102 108 : completed_deletes,
2103 108 : failed_deletes,
2104 108 :
2105 108 : rare_counters,
2106 108 : inits_cancelled,
2107 108 : redownload_after,
2108 108 : time_to_evict,
2109 108 : }
2110 108 : }
2111 : }
2112 :
2113 : impl LayerImplMetrics {
2114 84 : fn inc_started_evictions(&self) {
2115 84 : self.started_evictions.inc();
2116 84 : }
2117 68 : fn inc_completed_evictions(&self) {
2118 68 : self.completed_evictions.inc();
2119 68 : }
2120 16 : fn inc_eviction_cancelled(&self, reason: EvictionCancelled) {
2121 16 : self.cancelled_evictions[reason].inc()
2122 16 : }
2123 :
2124 1032 : fn inc_started_deletes(&self) {
2125 1032 : self.started_deletes.inc();
2126 1032 : }
2127 1020 : fn inc_completed_deletes(&self) {
2128 1020 : self.completed_deletes.inc();
2129 1020 : }
2130 0 : fn inc_deletes_failed(&self, reason: DeleteFailed) {
2131 0 : self.failed_deletes[reason].inc();
2132 0 : }
2133 :
2134 : /// Counted separatedly from failed layer deletes because we will complete the layer deletion
2135 : /// attempt regardless of failure to delete local file.
2136 0 : fn inc_delete_removes_failed(&self) {
2137 0 : self.rare_counters[RareEvent::RemoveOnDropFailed].inc();
2138 0 : }
2139 :
2140 : /// Expected rare just as cancellations are rare, but we could have cancellations separate from
2141 : /// the single caller which can start the download, so use this counter to separte them.
2142 0 : fn inc_init_completed_without_requester(&self) {
2143 0 : self.rare_counters[RareEvent::InitCompletedWithoutRequester].inc();
2144 0 : }
2145 :
2146 : /// Expected rare because cancellations are unexpected, and failures are unexpected
2147 0 : fn inc_download_failed_without_requester(&self) {
2148 0 : self.rare_counters[RareEvent::DownloadFailedWithoutRequester].inc();
2149 0 : }
2150 :
2151 : /// The Weak in ResidentOrWantedEvicted::WantedEvicted was successfully upgraded.
2152 : ///
2153 : /// If this counter is always zero, we should replace ResidentOrWantedEvicted type with an
2154 : /// Option.
2155 0 : fn inc_raced_wanted_evicted_accesses(&self) {
2156 0 : self.rare_counters[RareEvent::UpgradedWantedEvicted].inc();
2157 0 : }
2158 :
2159 : /// These are only expected for [`Self::inc_init_cancelled`] amount when
2160 : /// running with remote storage.
2161 12 : fn inc_init_needed_no_download(&self) {
2162 12 : self.rare_counters[RareEvent::InitWithoutDownload].inc();
2163 12 : }
2164 :
2165 : /// Expected rare because all layer files should be readable and good
2166 0 : fn inc_permanent_loading_failures(&self) {
2167 0 : self.rare_counters[RareEvent::PermanentLoadingFailure].inc();
2168 0 : }
2169 :
2170 0 : fn inc_init_cancelled(&self) {
2171 0 : self.inits_cancelled.inc()
2172 0 : }
2173 :
2174 28 : fn record_redownloaded_after(&self, duration: std::time::Duration) {
2175 28 : self.redownload_after.observe(duration.as_secs_f64())
2176 28 : }
2177 :
2178 : /// This would be bad if it ever happened, or mean extreme disk pressure. We should probably
2179 : /// instead cancel eviction if we would have read waiters. We cannot however separate reads
2180 : /// from other evictions, so this could have noise as well.
2181 0 : fn inc_evicted_with_waiters(&self) {
2182 0 : self.rare_counters[RareEvent::EvictedWithWaiters].inc();
2183 0 : }
2184 :
2185 : /// Recorded at least initially as the permit is now acquired in async context before
2186 : /// spawn_blocking action.
2187 68 : fn record_time_to_evict(&self, duration: std::time::Duration) {
2188 68 : self.time_to_evict.observe(duration.as_secs_f64())
2189 68 : }
2190 : }
2191 :
2192 : #[derive(Debug, Clone, Copy, enum_map::Enum)]
2193 : enum EvictionCancelled {
2194 : LayerGone,
2195 : TimelineGone,
2196 : VersionCheckFailed,
2197 : FileNotFound,
2198 : RemoveFailed,
2199 : AlreadyReinitialized,
2200 : /// Not evicted because of a pending reinitialization
2201 : LostToDownload,
2202 : /// After eviction, there was a new layer access which cancelled the eviction.
2203 : UpgradedBackOnAccess,
2204 : UnexpectedEvictedState,
2205 : }
2206 :
2207 : impl EvictionCancelled {
2208 972 : fn as_str(&self) -> &'static str {
2209 972 : match self {
2210 108 : EvictionCancelled::LayerGone => "layer_gone",
2211 108 : EvictionCancelled::TimelineGone => "timeline_gone",
2212 108 : EvictionCancelled::VersionCheckFailed => "version_check_fail",
2213 108 : EvictionCancelled::FileNotFound => "file_not_found",
2214 108 : EvictionCancelled::RemoveFailed => "remove_failed",
2215 108 : EvictionCancelled::AlreadyReinitialized => "already_reinitialized",
2216 108 : EvictionCancelled::LostToDownload => "lost_to_download",
2217 108 : EvictionCancelled::UpgradedBackOnAccess => "upgraded_back_on_access",
2218 108 : EvictionCancelled::UnexpectedEvictedState => "unexpected_evicted_state",
2219 : }
2220 972 : }
2221 : }
2222 :
2223 : #[derive(enum_map::Enum)]
2224 : enum DeleteFailed {
2225 : TimelineGone,
2226 : DeleteSchedulingFailed,
2227 : }
2228 :
2229 : impl DeleteFailed {
2230 216 : fn as_str(&self) -> &'static str {
2231 216 : match self {
2232 108 : DeleteFailed::TimelineGone => "timeline_gone",
2233 108 : DeleteFailed::DeleteSchedulingFailed => "delete_scheduling_failed",
2234 : }
2235 216 : }
2236 : }
2237 :
2238 : #[derive(enum_map::Enum)]
2239 : enum RareEvent {
2240 : RemoveOnDropFailed,
2241 : InitCompletedWithoutRequester,
2242 : DownloadFailedWithoutRequester,
2243 : UpgradedWantedEvicted,
2244 : InitWithoutDownload,
2245 : PermanentLoadingFailure,
2246 : EvictedWithWaiters,
2247 : }
2248 :
2249 : impl RareEvent {
2250 756 : fn as_str(&self) -> &'static str {
2251 : use RareEvent::*;
2252 :
2253 756 : match self {
2254 108 : RemoveOnDropFailed => "remove_on_drop_failed",
2255 108 : InitCompletedWithoutRequester => "init_completed_without",
2256 108 : DownloadFailedWithoutRequester => "download_failed_without",
2257 108 : UpgradedWantedEvicted => "raced_wanted_evicted",
2258 108 : InitWithoutDownload => "init_needed_no_download",
2259 108 : PermanentLoadingFailure => "permanent_loading_failure",
2260 108 : EvictedWithWaiters => "evicted_with_waiters",
2261 : }
2262 756 : }
2263 : }
2264 :
2265 : pub(crate) static LAYER_IMPL_METRICS: once_cell::sync::Lazy<LayerImplMetrics> =
2266 : once_cell::sync::Lazy::new(LayerImplMetrics::default);
|