Line data Source code
1 : use anyhow::Context;
2 : use camino::{Utf8Path, Utf8PathBuf};
3 : use pageserver_api::keyspace::KeySpace;
4 : use pageserver_api::models::HistoricLayerInfo;
5 : use pageserver_api::shard::{ShardIdentity, ShardIndex, TenantShardId};
6 : use std::ops::Range;
7 : use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
8 : use std::sync::{Arc, Weak};
9 : use std::time::{Duration, SystemTime};
10 : use tracing::Instrument;
11 : use utils::id::TimelineId;
12 : use utils::lsn::Lsn;
13 : use utils::sync::{gate, heavier_once_cell};
14 :
15 : use crate::config::PageServerConf;
16 : use crate::context::{DownloadBehavior, RequestContext, RequestContextBuilder};
17 : use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
18 : use crate::task_mgr::TaskKind;
19 : use crate::tenant::timeline::{CompactionError, GetVectoredError};
20 : use crate::tenant::{remote_timeline_client::LayerFileMetadata, Timeline};
21 :
22 : use super::delta_layer::{self};
23 : use super::image_layer::{self};
24 : use super::{
25 : AsLayerDesc, ImageLayerWriter, LayerAccessStats, LayerAccessStatsReset, LayerName,
26 : LayerVisibilityHint, PersistentLayerDesc, ValuesReconstructState,
27 : };
28 :
29 : use utils::generation::Generation;
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 4292 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
95 4292 : write!(
96 4292 : f,
97 4292 : "{}{}",
98 4292 : self.layer_desc().short_id(),
99 4292 : self.0.generation.get_suffix()
100 4292 : )
101 4292 : }
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 4005478 : fn layer_desc(&self) -> &PersistentLayerDesc {
112 4005478 : self.0.layer_desc()
113 4005478 : }
114 : }
115 :
116 : impl PartialEq for Layer {
117 8 : fn eq(&self, other: &Self) -> bool {
118 8 : Arc::as_ptr(&self.0) == Arc::as_ptr(&other.0)
119 8 : }
120 : }
121 :
122 3824 : pub(crate) fn local_layer_path(
123 3824 : conf: &PageServerConf,
124 3824 : tenant_shard_id: &TenantShardId,
125 3824 : timeline_id: &TimelineId,
126 3824 : layer_file_name: &LayerName,
127 3824 : generation: &Generation,
128 3824 : ) -> Utf8PathBuf {
129 3824 : let timeline_path = conf.timeline_path(tenant_shard_id, timeline_id);
130 3824 :
131 3824 : 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 3824 : timeline_path.join(format!("{}-v1{}", layer_file_name, generation.get_suffix()))
136 : }
137 3824 : }
138 :
139 : impl Layer {
140 : /// Creates a layer value for a file we know to not be resident.
141 0 : pub(crate) fn for_evicted(
142 0 : conf: &'static PageServerConf,
143 0 : timeline: &Arc<Timeline>,
144 0 : file_name: LayerName,
145 0 : metadata: LayerFileMetadata,
146 0 : ) -> Self {
147 0 : let local_path = local_layer_path(
148 0 : conf,
149 0 : &timeline.tenant_shard_id,
150 0 : &timeline.timeline_id,
151 0 : &file_name,
152 0 : &metadata.generation,
153 0 : );
154 0 :
155 0 : let desc = PersistentLayerDesc::from_filename(
156 0 : timeline.tenant_shard_id,
157 0 : timeline.timeline_id,
158 0 : file_name,
159 0 : metadata.file_size,
160 0 : );
161 0 :
162 0 : let owner = Layer(Arc::new(LayerInner::new(
163 0 : conf,
164 0 : timeline,
165 0 : local_path,
166 0 : desc,
167 0 : None,
168 0 : metadata.generation,
169 0 : metadata.shard,
170 0 : )));
171 0 :
172 0 : debug_assert!(owner.0.needs_download_blocking().unwrap().is_some());
173 :
174 0 : owner
175 0 : }
176 :
177 : /// Creates a Layer value for a file we know to be resident in timeline directory.
178 244 : pub(crate) fn for_resident(
179 244 : conf: &'static PageServerConf,
180 244 : timeline: &Arc<Timeline>,
181 244 : local_path: Utf8PathBuf,
182 244 : file_name: LayerName,
183 244 : metadata: LayerFileMetadata,
184 244 : ) -> ResidentLayer {
185 244 : let desc = PersistentLayerDesc::from_filename(
186 244 : timeline.tenant_shard_id,
187 244 : timeline.timeline_id,
188 244 : file_name,
189 244 : metadata.file_size,
190 244 : );
191 244 :
192 244 : let mut resident = None;
193 244 :
194 244 : let owner = Layer(Arc::new_cyclic(|owner| {
195 244 : let inner = Arc::new(DownloadedLayer {
196 244 : owner: owner.clone(),
197 244 : kind: tokio::sync::OnceCell::default(),
198 244 : version: 0,
199 244 : });
200 244 : resident = Some(inner.clone());
201 244 :
202 244 : LayerInner::new(
203 244 : conf,
204 244 : timeline,
205 244 : local_path,
206 244 : desc,
207 244 : Some(inner),
208 244 : metadata.generation,
209 244 : metadata.shard,
210 244 : )
211 244 : }));
212 244 :
213 244 : let downloaded = resident.expect("just initialized");
214 244 :
215 244 : debug_assert!(owner.0.needs_download_blocking().unwrap().is_none());
216 :
217 244 : timeline
218 244 : .metrics
219 244 : .resident_physical_size_add(metadata.file_size);
220 244 :
221 244 : ResidentLayer { downloaded, owner }
222 244 : }
223 :
224 : /// Creates a Layer value for freshly written out new layer file by renaming it from a
225 : /// temporary path.
226 3612 : pub(crate) fn finish_creating(
227 3612 : conf: &'static PageServerConf,
228 3612 : timeline: &Arc<Timeline>,
229 3612 : desc: PersistentLayerDesc,
230 3612 : temp_path: &Utf8Path,
231 3612 : ) -> anyhow::Result<ResidentLayer> {
232 3612 : let mut resident = None;
233 3612 :
234 3612 : let owner = Layer(Arc::new_cyclic(|owner| {
235 3612 : let inner = Arc::new(DownloadedLayer {
236 3612 : owner: owner.clone(),
237 3612 : kind: tokio::sync::OnceCell::default(),
238 3612 : version: 0,
239 3612 : });
240 3612 : resident = Some(inner.clone());
241 3612 :
242 3612 : let local_path = local_layer_path(
243 3612 : conf,
244 3612 : &timeline.tenant_shard_id,
245 3612 : &timeline.timeline_id,
246 3612 : &desc.layer_name(),
247 3612 : &timeline.generation,
248 3612 : );
249 3612 :
250 3612 : LayerInner::new(
251 3612 : conf,
252 3612 : timeline,
253 3612 : local_path,
254 3612 : desc,
255 3612 : Some(inner),
256 3612 : timeline.generation,
257 3612 : timeline.get_shard_index(),
258 3612 : )
259 3612 : }));
260 3612 :
261 3612 : let downloaded = resident.expect("just initialized");
262 3612 :
263 3612 : // We never want to overwrite an existing file, so we use `RENAME_NOREPLACE`.
264 3612 : // TODO: this leaves the temp file in place if the rename fails, risking us running
265 3612 : // out of space. Should we clean it up here or does the calling context deal with this?
266 3612 : utils::fs_ext::rename_noreplace(temp_path.as_std_path(), owner.local_path().as_std_path())
267 3612 : .with_context(|| format!("rename temporary file as correct path for {owner}"))?;
268 :
269 3612 : Ok(ResidentLayer { downloaded, owner })
270 3612 : }
271 :
272 : /// Requests the layer to be evicted and waits for this to be done.
273 : ///
274 : /// If the file is not resident, an [`EvictionError::NotFound`] is returned.
275 : ///
276 : /// If for a bad luck or blocking of the executor, we miss the actual eviction and the layer is
277 : /// re-downloaded, [`EvictionError::Downloaded`] is returned.
278 : ///
279 : /// Timeout is mandatory, because waiting for eviction is only needed for our tests; eviction
280 : /// will happen regardless the future returned by this method completing unless there is a
281 : /// read access before eviction gets to complete.
282 : ///
283 : /// Technically cancellation safe, but cancelling might shift the viewpoint of what generation
284 : /// of download-evict cycle on retry.
285 72 : pub(crate) async fn evict_and_wait(&self, timeout: Duration) -> Result<(), EvictionError> {
286 72 : self.0.evict_and_wait(timeout).await
287 64 : }
288 :
289 : /// Delete the layer file when the `self` gets dropped, also try to schedule a remote index upload
290 : /// then.
291 : ///
292 : /// On drop, this will cause a call to [`crate::tenant::remote_timeline_client::RemoteTimelineClient::schedule_deletion_of_unlinked`].
293 : /// This means that the unlinking by [gc] or [compaction] must have happened strictly before
294 : /// the value this is called on gets dropped.
295 : ///
296 : /// This is ensured by both of those methods accepting references to Layer.
297 : ///
298 : /// [gc]: [`RemoteTimelineClient::schedule_gc_update`]
299 : /// [compaction]: [`RemoteTimelineClient::schedule_compaction_update`]
300 1040 : pub(crate) fn delete_on_drop(&self) {
301 1040 : self.0.delete_on_drop();
302 1040 : }
303 :
304 554744 : pub(crate) async fn get_values_reconstruct_data(
305 554744 : &self,
306 554744 : keyspace: KeySpace,
307 554744 : lsn_range: Range<Lsn>,
308 554744 : reconstruct_data: &mut ValuesReconstructState,
309 554744 : ctx: &RequestContext,
310 554744 : ) -> Result<(), GetVectoredError> {
311 554744 : let downloaded = self
312 554744 : .0
313 554744 : .get_or_maybe_download(true, Some(ctx))
314 554744 : .await
315 554744 : .map_err(|err| match err {
316 : DownloadError::TimelineShutdown | DownloadError::DownloadCancelled => {
317 0 : GetVectoredError::Cancelled
318 : }
319 0 : other => GetVectoredError::Other(anyhow::anyhow!(other)),
320 554744 : })?;
321 554744 : let this = ResidentLayer {
322 554744 : downloaded: downloaded.clone(),
323 554744 : owner: self.clone(),
324 554744 : };
325 554744 :
326 554744 : self.record_access(ctx);
327 554744 :
328 554744 : downloaded
329 554744 : .get_values_reconstruct_data(this, keyspace, lsn_range, reconstruct_data, ctx)
330 554744 : .instrument(tracing::debug_span!("get_values_reconstruct_data", layer=%self))
331 554744 : .await
332 554744 : .map_err(|err| match err {
333 0 : GetVectoredError::Other(err) => GetVectoredError::Other(
334 0 : err.context(format!("get_values_reconstruct_data for layer {self}")),
335 0 : ),
336 0 : err => err,
337 554744 : })
338 554744 : }
339 :
340 : /// Download the layer if evicted.
341 : ///
342 : /// Will not error when the layer is already downloaded.
343 0 : pub(crate) async fn download(&self) -> Result<(), DownloadError> {
344 0 : self.0.get_or_maybe_download(true, None).await?;
345 0 : Ok(())
346 0 : }
347 :
348 608 : pub(crate) async fn needs_download(&self) -> Result<Option<NeedsDownload>, std::io::Error> {
349 608 : self.0.needs_download().await
350 608 : }
351 :
352 : /// Assuming the layer is already downloaded, returns a guard which will prohibit eviction
353 : /// while the guard exists.
354 : ///
355 : /// Returns None if the layer is currently evicted or becoming evicted.
356 40 : pub(crate) async fn keep_resident(&self) -> Option<ResidentLayer> {
357 40 : let downloaded = self.0.inner.get().and_then(|rowe| rowe.get())?;
358 :
359 28 : Some(ResidentLayer {
360 28 : downloaded,
361 28 : owner: self.clone(),
362 28 : })
363 40 : }
364 :
365 : /// Weak indicator of is the layer resident or not. Good enough for eviction, which can deal
366 : /// with `EvictionError::NotFound`.
367 : ///
368 : /// Returns `true` if this layer might be resident, or `false`, if it most likely evicted or
369 : /// will be unless a read happens soon.
370 112 : pub(crate) fn is_likely_resident(&self) -> bool {
371 112 : self.0
372 112 : .inner
373 112 : .get()
374 112 : .map(|rowe| rowe.is_likely_resident())
375 112 : .unwrap_or(false)
376 112 : }
377 :
378 : /// Downloads if necessary and creates a guard, which will keep this layer from being evicted.
379 1132 : pub(crate) async fn download_and_keep_resident(&self) -> Result<ResidentLayer, DownloadError> {
380 1132 : let downloaded = self.0.get_or_maybe_download(true, None).await?;
381 :
382 1132 : Ok(ResidentLayer {
383 1132 : downloaded,
384 1132 : owner: self.clone(),
385 1132 : })
386 1132 : }
387 :
388 0 : pub(crate) fn info(&self, reset: LayerAccessStatsReset) -> HistoricLayerInfo {
389 0 : self.0.info(reset)
390 0 : }
391 :
392 16 : pub(crate) fn latest_activity(&self) -> SystemTime {
393 16 : self.0.access_stats.latest_activity()
394 16 : }
395 :
396 20 : pub(crate) fn visibility(&self) -> LayerVisibilityHint {
397 20 : self.0.access_stats.visibility()
398 20 : }
399 :
400 3616 : pub(crate) fn local_path(&self) -> &Utf8Path {
401 3616 : &self.0.path
402 3616 : }
403 :
404 5148 : pub(crate) fn metadata(&self) -> LayerFileMetadata {
405 5148 : self.0.metadata()
406 5148 : }
407 :
408 0 : pub(crate) fn get_timeline_id(&self) -> Option<TimelineId> {
409 0 : self.0
410 0 : .timeline
411 0 : .upgrade()
412 0 : .map(|timeline| timeline.timeline_id)
413 0 : }
414 :
415 : /// Traditional debug dumping facility
416 : #[allow(unused)]
417 8 : pub(crate) async fn dump(&self, verbose: bool, ctx: &RequestContext) -> anyhow::Result<()> {
418 8 : self.0.desc.dump();
419 8 :
420 8 : if verbose {
421 : // for now, unconditionally download everything, even if that might not be wanted.
422 8 : let l = self.0.get_or_maybe_download(true, Some(ctx)).await?;
423 8 : l.dump(&self.0, ctx).await?
424 0 : }
425 :
426 8 : Ok(())
427 8 : }
428 :
429 : /// Waits until this layer has been dropped (and if needed, local file deletion and remote
430 : /// deletion scheduling has completed).
431 : ///
432 : /// Does not start local deletion, use [`Self::delete_on_drop`] for that
433 : /// separatedly.
434 : #[cfg(any(feature = "testing", test))]
435 4 : pub(crate) fn wait_drop(&self) -> impl std::future::Future<Output = ()> + 'static {
436 4 : let mut rx = self.0.status.as_ref().unwrap().subscribe();
437 :
438 4 : async move {
439 : loop {
440 4 : if rx.changed().await.is_err() {
441 4 : break;
442 0 : }
443 : }
444 4 : }
445 4 : }
446 :
447 554744 : fn record_access(&self, ctx: &RequestContext) {
448 554744 : if self.0.access_stats.record_access(ctx) {
449 : // Visibility was modified to Visible: maybe log about this
450 0 : match ctx.task_kind() {
451 : TaskKind::CalculateSyntheticSize
452 : | TaskKind::OndemandLogicalSizeCalculation
453 : | TaskKind::GarbageCollector
454 0 : | TaskKind::MgmtRequest => {
455 0 : // This situation is expected in code paths do binary searches of the LSN space to resolve
456 0 : // an LSN to a timestamp, which happens during GC, during GC cutoff calculations in synthetic size,
457 0 : // and on-demand for certain HTTP API requests. On-demand logical size calculation is also included
458 0 : // because it is run as a sub-task of synthetic size.
459 0 : }
460 : _ => {
461 : // In all other contexts, it is unusual to do I/O involving layers which are not visible at
462 : // some branch tip, so we log the fact that we are accessing something that the visibility
463 : // calculation thought should not be visible.
464 : //
465 : // This case is legal in brief time windows: for example an in-flight getpage request can hold on to a layer object
466 : // which was covered by a concurrent compaction.
467 0 : tracing::info!(
468 : layer=%self,
469 0 : "became visible as a result of access",
470 : );
471 : }
472 : }
473 :
474 : // Update the timeline's visible bytes count
475 0 : if let Some(tl) = self.0.timeline.upgrade() {
476 0 : tl.metrics
477 0 : .visible_physical_size_gauge
478 0 : .add(self.0.desc.file_size)
479 0 : }
480 554744 : }
481 554744 : }
482 :
483 668 : pub(crate) fn set_visibility(&self, visibility: LayerVisibilityHint) {
484 668 : let old_visibility = self.0.access_stats.set_visibility(visibility.clone());
485 : use LayerVisibilityHint::*;
486 668 : match (old_visibility, visibility) {
487 : (Visible, Covered) => {
488 : // Subtract this layer's contribution to the visible size metric
489 96 : if let Some(tl) = self.0.timeline.upgrade() {
490 96 : debug_assert!(
491 96 : tl.metrics.visible_physical_size_gauge.get() >= self.0.desc.file_size
492 : );
493 96 : tl.metrics
494 96 : .visible_physical_size_gauge
495 96 : .sub(self.0.desc.file_size)
496 0 : }
497 : }
498 : (Covered, Visible) => {
499 : // Add this layer's contribution to the visible size metric
500 0 : if let Some(tl) = self.0.timeline.upgrade() {
501 0 : tl.metrics
502 0 : .visible_physical_size_gauge
503 0 : .add(self.0.desc.file_size)
504 0 : }
505 : }
506 572 : (Covered, Covered) | (Visible, Visible) => {
507 572 : // no change
508 572 : }
509 : }
510 668 : }
511 : }
512 :
513 : /// The download-ness ([`DownloadedLayer`]) can be either resident or wanted evicted.
514 : ///
515 : /// However when we want something evicted, we cannot evict it right away as there might be current
516 : /// reads happening on it. For example: it has been searched from [`LayerMap::search`] but not yet
517 : /// read with [`Layer::get_values_reconstruct_data`].
518 : ///
519 : /// [`LayerMap::search`]: crate::tenant::layer_map::LayerMap::search
520 : #[derive(Debug)]
521 : enum ResidentOrWantedEvicted {
522 : Resident(Arc<DownloadedLayer>),
523 : WantedEvicted(Weak<DownloadedLayer>, usize),
524 : }
525 :
526 : impl ResidentOrWantedEvicted {
527 : /// Non-mutating access to the a DownloadedLayer, if possible.
528 : ///
529 : /// This is not used on the read path (anything that calls
530 : /// [`LayerInner::get_or_maybe_download`]) because it was decided that reads always win
531 : /// evictions, and part of that winning is using [`ResidentOrWantedEvicted::get_and_upgrade`].
532 28 : fn get(&self) -> Option<Arc<DownloadedLayer>> {
533 28 : match self {
534 28 : ResidentOrWantedEvicted::Resident(strong) => Some(strong.clone()),
535 0 : ResidentOrWantedEvicted::WantedEvicted(weak, _) => weak.upgrade(),
536 : }
537 28 : }
538 :
539 : /// Best-effort query for residency right now, not as strong guarantee as receiving a strong
540 : /// reference from `ResidentOrWantedEvicted::get`.
541 100 : fn is_likely_resident(&self) -> bool {
542 100 : match self {
543 88 : ResidentOrWantedEvicted::Resident(_) => true,
544 12 : ResidentOrWantedEvicted::WantedEvicted(weak, _) => weak.strong_count() > 0,
545 : }
546 100 : }
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 555892 : fn get_and_upgrade(&mut self) -> Option<(Arc<DownloadedLayer>, bool)> {
553 555892 : match self {
554 555876 : ResidentOrWantedEvicted::Resident(strong) => Some((strong.clone(), false)),
555 16 : 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 16 : None => None,
564 : },
565 : }
566 555892 : }
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 60 : fn downgrade(&mut self) -> Option<Arc<DownloadedLayer>> {
574 60 : match self {
575 52 : ResidentOrWantedEvicted::Resident(strong) => {
576 52 : let weak = Arc::downgrade(strong);
577 52 : let mut temp = ResidentOrWantedEvicted::WantedEvicted(weak, strong.version);
578 52 : std::mem::swap(self, &mut temp);
579 52 : match temp {
580 52 : ResidentOrWantedEvicted::Resident(strong) => Some(strong),
581 0 : ResidentOrWantedEvicted::WantedEvicted(..) => unreachable!("just swapped"),
582 : }
583 : }
584 8 : ResidentOrWantedEvicted::WantedEvicted(..) => None,
585 : }
586 60 : }
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 : desc: PersistentLayerDesc,
598 :
599 : /// Timeline access is needed for remote timeline client and metrics.
600 : ///
601 : /// There should not be an access to timeline for any reason without entering the
602 : /// [`Timeline::gate`] at the same time.
603 : timeline: Weak<Timeline>,
604 :
605 : access_stats: LayerAccessStats,
606 :
607 : /// This custom OnceCell is backed by std mutex, but only held for short time periods.
608 : ///
609 : /// Filesystem changes (download, evict) are only done while holding a permit which the
610 : /// `heavier_once_cell` provides.
611 : ///
612 : /// A number of fields in `Layer` are meant to only be updated when holding the InitPermit, but
613 : /// possibly read while not holding it.
614 : inner: heavier_once_cell::OnceCell<ResidentOrWantedEvicted>,
615 :
616 : /// Do we want to delete locally and remotely this when `LayerInner` is dropped
617 : wanted_deleted: AtomicBool,
618 :
619 : /// Version is to make sure we will only evict a specific initialization of the downloaded file.
620 : ///
621 : /// Incremented for each initialization, stored in `DownloadedLayer::version` or
622 : /// `ResidentOrWantedEvicted::WantedEvicted`.
623 : version: AtomicUsize,
624 :
625 : /// Allow subscribing to when the layer actually gets evicted, a non-cancellable download
626 : /// starts, or completes.
627 : ///
628 : /// Updates must only be posted while holding the InitPermit or the heavier_once_cell::Guard.
629 : /// Holding the InitPermit is the only time we can do state transitions, but we also need to
630 : /// cancel a pending eviction on upgrading a [`ResidentOrWantedEvicted::WantedEvicted`] back to
631 : /// [`ResidentOrWantedEvicted::Resident`] on access.
632 : ///
633 : /// The sender is wrapped in an Option to facilitate moving it out on [`LayerInner::drop`].
634 : status: Option<tokio::sync::watch::Sender<Status>>,
635 :
636 : /// Counter for exponential backoff with the download.
637 : ///
638 : /// This is atomic only for the purposes of having additional data only accessed while holding
639 : /// the InitPermit.
640 : consecutive_failures: AtomicUsize,
641 :
642 : /// The generation of this Layer.
643 : ///
644 : /// For loaded layers (resident or evicted) this comes from [`LayerFileMetadata::generation`],
645 : /// for created layers from [`Timeline::generation`].
646 : generation: Generation,
647 :
648 : /// The shard of this Layer.
649 : ///
650 : /// For layers created in this process, this will always be the [`ShardIndex`] of the
651 : /// current `ShardIdentity`` (TODO: add link once it's introduced).
652 : ///
653 : /// For loaded layers, this may be some other value if the tenant has undergone
654 : /// a shard split since the layer was originally written.
655 : shard: ShardIndex,
656 :
657 : /// When the Layer was last evicted but has not been downloaded since.
658 : ///
659 : /// This is used solely for updating metrics. See [`LayerImplMetrics::redownload_after`].
660 : last_evicted_at: std::sync::Mutex<Option<std::time::Instant>>,
661 :
662 : #[cfg(test)]
663 : failpoints: std::sync::Mutex<Vec<failpoints::Failpoint>>,
664 : }
665 :
666 : impl std::fmt::Display for LayerInner {
667 60 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
668 60 : write!(f, "{}", self.layer_desc().short_id())
669 60 : }
670 : }
671 :
672 : impl AsLayerDesc for LayerInner {
673 4010628 : fn layer_desc(&self) -> &PersistentLayerDesc {
674 4010628 : &self.desc
675 4010628 : }
676 : }
677 :
678 : #[derive(Debug, Clone, Copy)]
679 : enum Status {
680 : Resident,
681 : Evicted,
682 : Downloading,
683 : }
684 :
685 : impl Drop for LayerInner {
686 1410 : fn drop(&mut self) {
687 1410 : // if there was a pending eviction, mark it cancelled here to balance metrics
688 1410 : if let Some((ResidentOrWantedEvicted::WantedEvicted(..), _)) = self.inner.take_and_deinit()
689 4 : {
690 4 : // eviction has already been started
691 4 : LAYER_IMPL_METRICS.inc_eviction_cancelled(EvictionCancelled::LayerGone);
692 4 :
693 4 : // eviction request is intentionally not honored as no one is present to wait for it
694 4 : // and we could be delaying shutdown for nothing.
695 1406 : }
696 :
697 1410 : let timeline = self.timeline.upgrade();
698 :
699 1410 : if let Some(timeline) = timeline.as_ref() {
700 : // Only need to decrement metrics if the timeline still exists: otherwise
701 : // it will have already de-registered these metrics via TimelineMetrics::shutdown
702 1378 : timeline.metrics.dec_layer(&self.desc);
703 :
704 1378 : if matches!(self.access_stats.visibility(), LayerVisibilityHint::Visible) {
705 1378 : debug_assert!(
706 1378 : timeline.metrics.visible_physical_size_gauge.get() >= self.desc.file_size
707 : );
708 1378 : timeline
709 1378 : .metrics
710 1378 : .visible_physical_size_gauge
711 1378 : .sub(self.desc.file_size);
712 0 : }
713 32 : }
714 :
715 1410 : if !*self.wanted_deleted.get_mut() {
716 392 : return;
717 1018 : }
718 :
719 1018 : 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);
720 :
721 1018 : let path = std::mem::take(&mut self.path);
722 1018 : let file_name = self.layer_desc().layer_name();
723 1018 : let file_size = self.layer_desc().file_size;
724 1018 : let meta = self.metadata();
725 1018 : let status = self.status.take();
726 1018 :
727 1018 : Self::spawn_blocking(move || {
728 1017 : let _g = span.entered();
729 1017 :
730 1017 : // carry this until we are finished for [`Layer::wait_drop`] support
731 1017 : let _status = status;
732 :
733 1017 : let Some(timeline) = timeline else {
734 : // no need to nag that timeline is gone: under normal situation on
735 : // task_mgr::remove_tenant_from_memory the timeline is gone before we get dropped.
736 0 : LAYER_IMPL_METRICS.inc_deletes_failed(DeleteFailed::TimelineGone);
737 0 : return;
738 : };
739 :
740 1017 : let Ok(_guard) = timeline.gate.enter() else {
741 0 : LAYER_IMPL_METRICS.inc_deletes_failed(DeleteFailed::TimelineGone);
742 0 : return;
743 : };
744 :
745 1017 : let removed = match std::fs::remove_file(path) {
746 1013 : Ok(()) => true,
747 4 : Err(e) if e.kind() == std::io::ErrorKind::NotFound => {
748 4 : // until we no longer do detaches by removing all local files before removing the
749 4 : // tenant from the global map, we will always get these errors even if we knew what
750 4 : // is the latest state.
751 4 : //
752 4 : // we currently do not track the latest state, so we'll also end up here on evicted
753 4 : // layers.
754 4 : false
755 : }
756 0 : Err(e) => {
757 0 : tracing::error!("failed to remove wanted deleted layer: {e}");
758 0 : LAYER_IMPL_METRICS.inc_delete_removes_failed();
759 0 : false
760 : }
761 : };
762 :
763 1017 : if removed {
764 1013 : timeline.metrics.resident_physical_size_sub(file_size);
765 1013 : }
766 1017 : let res = timeline
767 1017 : .remote_client
768 1017 : .schedule_deletion_of_unlinked(vec![(file_name, meta)]);
769 :
770 1017 : if let Err(e) = res {
771 : // test_timeline_deletion_with_files_stuck_in_upload_queue is good at
772 : // demonstrating this deadlock (without spawn_blocking): stop will drop
773 : // queued items, which will have ResidentLayer's, and those drops would try
774 : // to re-entrantly lock the RemoteTimelineClient inner state.
775 4 : if !timeline.is_active() {
776 4 : tracing::info!("scheduling deletion on drop failed: {e:#}");
777 : } else {
778 0 : tracing::warn!("scheduling deletion on drop failed: {e:#}");
779 : }
780 0 : LAYER_IMPL_METRICS.inc_deletes_failed(DeleteFailed::DeleteSchedulingFailed);
781 1013 : } else {
782 1013 : LAYER_IMPL_METRICS.inc_completed_deletes();
783 1013 : }
784 1018 : });
785 1410 : }
786 : }
787 :
788 : impl LayerInner {
789 : #[allow(clippy::too_many_arguments)]
790 3856 : fn new(
791 3856 : conf: &'static PageServerConf,
792 3856 : timeline: &Arc<Timeline>,
793 3856 : local_path: Utf8PathBuf,
794 3856 : desc: PersistentLayerDesc,
795 3856 : downloaded: Option<Arc<DownloadedLayer>>,
796 3856 : generation: Generation,
797 3856 : shard: ShardIndex,
798 3856 : ) -> Self {
799 3856 : let (inner, version, init_status) = if let Some(inner) = downloaded {
800 3856 : let version = inner.version;
801 3856 : let resident = ResidentOrWantedEvicted::Resident(inner);
802 3856 : (
803 3856 : heavier_once_cell::OnceCell::new(resident),
804 3856 : version,
805 3856 : Status::Resident,
806 3856 : )
807 : } else {
808 0 : (heavier_once_cell::OnceCell::default(), 0, Status::Evicted)
809 : };
810 :
811 : // This object acts as a RAII guard on these metrics: increment on construction
812 3856 : timeline.metrics.inc_layer(&desc);
813 3856 :
814 3856 : // New layers are visible by default. This metric is later updated on drop or in set_visibility
815 3856 : timeline
816 3856 : .metrics
817 3856 : .visible_physical_size_gauge
818 3856 : .add(desc.file_size);
819 3856 :
820 3856 : LayerInner {
821 3856 : conf,
822 3856 : path: local_path,
823 3856 : desc,
824 3856 : timeline: Arc::downgrade(timeline),
825 3856 : access_stats: Default::default(),
826 3856 : wanted_deleted: AtomicBool::new(false),
827 3856 : inner,
828 3856 : version: AtomicUsize::new(version),
829 3856 : status: Some(tokio::sync::watch::channel(init_status).0),
830 3856 : consecutive_failures: AtomicUsize::new(0),
831 3856 : generation,
832 3856 : shard,
833 3856 : last_evicted_at: std::sync::Mutex::default(),
834 3856 : #[cfg(test)]
835 3856 : failpoints: Default::default(),
836 3856 : }
837 3856 : }
838 :
839 1040 : fn delete_on_drop(&self) {
840 1040 : let res =
841 1040 : self.wanted_deleted
842 1040 : .compare_exchange(false, true, Ordering::Release, Ordering::Relaxed);
843 1040 :
844 1040 : if res.is_ok() {
845 1032 : LAYER_IMPL_METRICS.inc_started_deletes();
846 1032 : }
847 1040 : }
848 :
849 : /// Cancellation safe, however dropping the future and calling this method again might result
850 : /// in a new attempt to evict OR join the previously started attempt.
851 72 : #[tracing::instrument(level = tracing::Level::DEBUG, skip_all, ret, err(level = tracing::Level::DEBUG), fields(layer=%self))]
852 : pub(crate) async fn evict_and_wait(&self, timeout: Duration) -> Result<(), EvictionError> {
853 : let mut rx = self.status.as_ref().unwrap().subscribe();
854 :
855 : {
856 : let current = rx.borrow_and_update();
857 : match &*current {
858 : Status::Resident => {
859 : // we might get lucky and evict this; continue
860 : }
861 : Status::Evicted | Status::Downloading => {
862 : // it is already evicted
863 : return Err(EvictionError::NotFound);
864 : }
865 : }
866 : }
867 :
868 : let strong = {
869 : match self.inner.get() {
870 : Some(mut either) => either.downgrade(),
871 : None => {
872 : // we already have a scheduled eviction, which just has not gotten to run yet.
873 : // it might still race with a read access, but that could also get cancelled,
874 : // so let's say this is not evictable.
875 : return Err(EvictionError::NotFound);
876 : }
877 : }
878 : };
879 :
880 : if strong.is_some() {
881 : // drop the DownloadedLayer outside of the holding the guard
882 : drop(strong);
883 :
884 : // idea here is that only one evicter should ever get to witness a strong reference,
885 : // which means whenever get_or_maybe_download upgrades a weak, it must mark up a
886 : // cancelled eviction and signal us, like it currently does.
887 : //
888 : // a second concurrent evict_and_wait will not see a strong reference.
889 : LAYER_IMPL_METRICS.inc_started_evictions();
890 : }
891 :
892 : let changed = rx.changed();
893 : let changed = tokio::time::timeout(timeout, changed).await;
894 :
895 : let Ok(changed) = changed else {
896 : return Err(EvictionError::Timeout);
897 : };
898 :
899 : let _: () = changed.expect("cannot be closed, because we are holding a strong reference");
900 :
901 : let current = rx.borrow_and_update();
902 :
903 : match &*current {
904 : // the easiest case
905 : Status::Evicted => Ok(()),
906 : // it surely was evicted in between, but then there was a new access now; we can't know
907 : // if it'll succeed so lets just call it evicted
908 : Status::Downloading => Ok(()),
909 : // either the download which was started after eviction completed already, or it was
910 : // never evicted
911 : Status::Resident => Err(EvictionError::Downloaded),
912 : }
913 : }
914 :
915 : /// Cancellation safe.
916 555908 : async fn get_or_maybe_download(
917 555908 : self: &Arc<Self>,
918 555908 : allow_download: bool,
919 555908 : ctx: Option<&RequestContext>,
920 555908 : ) -> Result<Arc<DownloadedLayer>, DownloadError> {
921 32 : let (weak, permit) = {
922 : // get_or_init_detached can:
923 : // - be fast (mutex lock) OR uncontested semaphore permit acquire
924 : // - be slow (wait for semaphore permit or closing)
925 555908 : let init_cancelled = scopeguard::guard((), |_| LAYER_IMPL_METRICS.inc_init_cancelled());
926 :
927 555908 : let locked = self
928 555908 : .inner
929 555908 : .get_or_init_detached()
930 555908 : .await
931 555908 : .map(|mut guard| guard.get_and_upgrade().ok_or(guard));
932 555908 :
933 555908 : scopeguard::ScopeGuard::into_inner(init_cancelled);
934 :
935 555876 : match locked {
936 : // this path could had been a RwLock::read
937 555876 : Ok(Ok((strong, upgraded))) if !upgraded => return Ok(strong),
938 0 : Ok(Ok((strong, _))) => {
939 0 : // when upgraded back, the Arc<DownloadedLayer> is still available, but
940 0 : // previously a `evict_and_wait` was received. this is the only place when we
941 0 : // send out an update without holding the InitPermit.
942 0 : //
943 0 : // note that we also have dropped the Guard; this is fine, because we just made
944 0 : // a state change and are holding a strong reference to be returned.
945 0 : self.status.as_ref().unwrap().send_replace(Status::Resident);
946 0 : LAYER_IMPL_METRICS
947 0 : .inc_eviction_cancelled(EvictionCancelled::UpgradedBackOnAccess);
948 0 :
949 0 : return Ok(strong);
950 : }
951 16 : Ok(Err(guard)) => {
952 16 : // path to here: we won the eviction, the file should still be on the disk.
953 16 : let (weak, permit) = guard.take_and_deinit();
954 16 : (Some(weak), permit)
955 : }
956 16 : Err(permit) => (None, permit),
957 : }
958 : };
959 :
960 32 : if let Some(weak) = weak {
961 : // only drop the weak after dropping the heavier_once_cell guard
962 16 : assert!(
963 16 : matches!(weak, ResidentOrWantedEvicted::WantedEvicted(..)),
964 0 : "unexpected {weak:?}, ResidentOrWantedEvicted::get_and_upgrade has a bug"
965 : );
966 16 : }
967 :
968 32 : let timeline = self
969 32 : .timeline
970 32 : .upgrade()
971 32 : .ok_or(DownloadError::TimelineShutdown)?;
972 :
973 : // count cancellations, which currently remain largely unexpected
974 32 : let init_cancelled = scopeguard::guard((), |_| LAYER_IMPL_METRICS.inc_init_cancelled());
975 :
976 : // check if we really need to be downloaded: this can happen if a read access won the
977 : // semaphore before eviction.
978 : //
979 : // if we are cancelled while doing this `stat` the `self.inner` will be uninitialized. a
980 : // pending eviction will try to evict even upon finding an uninitialized `self.inner`.
981 32 : let needs_download = self
982 32 : .needs_download()
983 32 : .await
984 32 : .map_err(DownloadError::PreStatFailed);
985 32 :
986 32 : scopeguard::ScopeGuard::into_inner(init_cancelled);
987 :
988 32 : let needs_download = needs_download?;
989 :
990 32 : let Some(reason) = needs_download else {
991 : // the file is present locally because eviction has not had a chance to run yet
992 :
993 : #[cfg(test)]
994 16 : self.failpoint(failpoints::FailpointKind::AfterDeterminingLayerNeedsNoDownload)
995 16 : .await?;
996 :
997 12 : LAYER_IMPL_METRICS.inc_init_needed_no_download();
998 12 :
999 12 : return Ok(self.initialize_after_layer_is_on_disk(permit));
1000 : };
1001 :
1002 : // we must download; getting cancelled before spawning the download is not an issue as
1003 : // any still running eviction would not find anything to evict.
1004 :
1005 16 : if let NeedsDownload::NotFile(ft) = reason {
1006 0 : return Err(DownloadError::NotFile(ft));
1007 16 : }
1008 :
1009 16 : if let Some(ctx) = ctx {
1010 4 : self.check_expected_download(ctx)?;
1011 12 : }
1012 :
1013 16 : if !allow_download {
1014 : // this is only used from tests, but it is hard to test without the boolean
1015 4 : return Err(DownloadError::DownloadRequired);
1016 12 : }
1017 12 :
1018 12 : let download_ctx = ctx
1019 12 : .map(|ctx| ctx.detached_child(TaskKind::LayerDownload, DownloadBehavior::Download))
1020 12 : .unwrap_or(RequestContext::new(
1021 12 : TaskKind::LayerDownload,
1022 12 : DownloadBehavior::Download,
1023 12 : ));
1024 :
1025 12 : async move {
1026 12 : tracing::info!(%reason, "downloading on-demand");
1027 :
1028 12 : let init_cancelled = scopeguard::guard((), |_| LAYER_IMPL_METRICS.inc_init_cancelled());
1029 12 : let res = self
1030 12 : .download_init_and_wait(timeline, permit, download_ctx)
1031 12 : .await?;
1032 12 : scopeguard::ScopeGuard::into_inner(init_cancelled);
1033 12 : Ok(res)
1034 12 : }
1035 12 : .instrument(tracing::info_span!("get_or_maybe_download", layer=%self))
1036 12 : .await
1037 555908 : }
1038 :
1039 : /// Nag or fail per RequestContext policy
1040 4 : fn check_expected_download(&self, ctx: &RequestContext) -> Result<(), DownloadError> {
1041 : use crate::context::DownloadBehavior::*;
1042 4 : let b = ctx.download_behavior();
1043 4 : match b {
1044 4 : Download => Ok(()),
1045 : Warn | Error => {
1046 0 : tracing::info!(
1047 0 : "unexpectedly on-demand downloading for task kind {:?}",
1048 0 : ctx.task_kind()
1049 : );
1050 0 : crate::metrics::UNEXPECTED_ONDEMAND_DOWNLOADS.inc();
1051 :
1052 0 : let really_error =
1053 0 : matches!(b, Error) && !self.conf.ondemand_download_behavior_treat_error_as_warn;
1054 :
1055 0 : if really_error {
1056 : // this check is only probablistic, seems like flakyness footgun
1057 0 : Err(DownloadError::ContextAndConfigReallyDeniesDownloads)
1058 : } else {
1059 0 : Ok(())
1060 : }
1061 : }
1062 : }
1063 4 : }
1064 :
1065 : /// Actual download, at most one is executed at the time.
1066 12 : async fn download_init_and_wait(
1067 12 : self: &Arc<Self>,
1068 12 : timeline: Arc<Timeline>,
1069 12 : permit: heavier_once_cell::InitPermit,
1070 12 : ctx: RequestContext,
1071 12 : ) -> Result<Arc<DownloadedLayer>, DownloadError> {
1072 12 : debug_assert_current_span_has_tenant_and_timeline_id();
1073 12 :
1074 12 : let (tx, rx) = tokio::sync::oneshot::channel();
1075 12 :
1076 12 : let this: Arc<Self> = self.clone();
1077 :
1078 12 : let guard = timeline
1079 12 : .gate
1080 12 : .enter()
1081 12 : .map_err(|_| DownloadError::DownloadCancelled)?;
1082 :
1083 12 : Self::spawn(
1084 12 : async move {
1085 0 : let _guard = guard;
1086 0 :
1087 0 : // now that we have commited to downloading, send out an update to:
1088 0 : // - unhang any pending eviction
1089 0 : // - break out of evict_and_wait
1090 0 : this.status
1091 0 : .as_ref()
1092 0 : .unwrap()
1093 0 : .send_replace(Status::Downloading);
1094 12 :
1095 12 : #[cfg(test)]
1096 12 : this.failpoint(failpoints::FailpointKind::WaitBeforeDownloading)
1097 12 : .await
1098 12 : .unwrap();
1099 :
1100 12 : let res = this.download_and_init(timeline, permit, &ctx).await;
1101 :
1102 12 : if let Err(res) = tx.send(res) {
1103 0 : match res {
1104 0 : Ok(_res) => {
1105 0 : tracing::debug!("layer initialized, but caller has been cancelled");
1106 0 : LAYER_IMPL_METRICS.inc_init_completed_without_requester();
1107 : }
1108 0 : Err(e) => {
1109 0 : tracing::info!(
1110 0 : "layer file download failed, and caller has been cancelled: {e:?}"
1111 : );
1112 0 : LAYER_IMPL_METRICS.inc_download_failed_without_requester();
1113 : }
1114 : }
1115 12 : }
1116 12 : }
1117 12 : .in_current_span(),
1118 12 : );
1119 12 :
1120 12 : match rx.await {
1121 12 : Ok(Ok(res)) => Ok(res),
1122 : Ok(Err(remote_storage::DownloadError::Cancelled)) => {
1123 0 : Err(DownloadError::DownloadCancelled)
1124 : }
1125 0 : Ok(Err(_)) => Err(DownloadError::DownloadFailed),
1126 0 : Err(_gone) => Err(DownloadError::DownloadCancelled),
1127 : }
1128 12 : }
1129 :
1130 12 : async fn download_and_init(
1131 12 : self: &Arc<LayerInner>,
1132 12 : timeline: Arc<Timeline>,
1133 12 : permit: heavier_once_cell::InitPermit,
1134 12 : ctx: &RequestContext,
1135 12 : ) -> Result<Arc<DownloadedLayer>, remote_storage::DownloadError> {
1136 12 : let result = timeline
1137 12 : .remote_client
1138 12 : .download_layer_file(
1139 12 : &self.desc.layer_name(),
1140 12 : &self.metadata(),
1141 12 : &self.path,
1142 12 : &timeline.gate,
1143 12 : &timeline.cancel,
1144 12 : ctx,
1145 12 : )
1146 12 : .await;
1147 :
1148 12 : match result {
1149 12 : Ok(size) => {
1150 12 : assert_eq!(size, self.desc.file_size);
1151 :
1152 12 : match self.needs_download().await {
1153 0 : Ok(Some(reason)) => {
1154 0 : // this is really a bug in needs_download or remote timeline client
1155 0 : panic!("post-condition failed: needs_download returned {reason:?}");
1156 : }
1157 12 : Ok(None) => {
1158 12 : // as expected
1159 12 : }
1160 0 : Err(e) => {
1161 0 : panic!("post-condition failed: needs_download errored: {e:?}");
1162 : }
1163 : }
1164 :
1165 12 : tracing::info!(size=%self.desc.file_size, "on-demand download successful");
1166 12 : timeline
1167 12 : .metrics
1168 12 : .resident_physical_size_add(self.desc.file_size);
1169 12 : self.consecutive_failures.store(0, Ordering::Relaxed);
1170 12 :
1171 12 : let since_last_eviction = self
1172 12 : .last_evicted_at
1173 12 : .lock()
1174 12 : .unwrap()
1175 12 : .take()
1176 12 : .map(|ts| ts.elapsed());
1177 12 : if let Some(since_last_eviction) = since_last_eviction {
1178 12 : LAYER_IMPL_METRICS.record_redownloaded_after(since_last_eviction);
1179 12 : }
1180 :
1181 12 : self.access_stats.record_residence_event();
1182 12 :
1183 12 : Ok(self.initialize_after_layer_is_on_disk(permit))
1184 : }
1185 0 : Err(e) => {
1186 0 : let consecutive_failures =
1187 0 : 1 + self.consecutive_failures.fetch_add(1, Ordering::Relaxed);
1188 0 :
1189 0 : if timeline.cancel.is_cancelled() {
1190 : // If we're shutting down, drop out before logging the error
1191 0 : return Err(e);
1192 0 : }
1193 0 :
1194 0 : tracing::error!(consecutive_failures, "layer file download failed: {e:#}");
1195 :
1196 0 : let backoff = utils::backoff::exponential_backoff_duration_seconds(
1197 0 : consecutive_failures.min(u32::MAX as usize) as u32,
1198 0 : 1.5,
1199 0 : 60.0,
1200 0 : );
1201 0 :
1202 0 : let backoff = std::time::Duration::from_secs_f64(backoff);
1203 0 :
1204 0 : tokio::select! {
1205 0 : _ = tokio::time::sleep(backoff) => {},
1206 0 : _ = timeline.cancel.cancelled() => {},
1207 : };
1208 :
1209 0 : Err(e)
1210 : }
1211 : }
1212 12 : }
1213 :
1214 : /// Initializes the `Self::inner` to a "resident" state.
1215 : ///
1216 : /// Callers are assumed to ensure that the file is actually on disk with `Self::needs_download`
1217 : /// before calling this method.
1218 : ///
1219 : /// If this method is ever made async, it needs to be cancellation safe so that no state
1220 : /// changes are made before we can write to the OnceCell in non-cancellable fashion.
1221 24 : fn initialize_after_layer_is_on_disk(
1222 24 : self: &Arc<LayerInner>,
1223 24 : permit: heavier_once_cell::InitPermit,
1224 24 : ) -> Arc<DownloadedLayer> {
1225 24 : debug_assert_current_span_has_tenant_and_timeline_id();
1226 24 :
1227 24 : // disable any scheduled but not yet running eviction deletions for this initialization
1228 24 : let next_version = 1 + self.version.fetch_add(1, Ordering::Relaxed);
1229 24 : self.status.as_ref().unwrap().send_replace(Status::Resident);
1230 24 :
1231 24 : let res = Arc::new(DownloadedLayer {
1232 24 : owner: Arc::downgrade(self),
1233 24 : kind: tokio::sync::OnceCell::default(),
1234 24 : version: next_version,
1235 24 : });
1236 24 :
1237 24 : let waiters = self.inner.initializer_count();
1238 24 : if waiters > 0 {
1239 0 : tracing::info!(waiters, "completing layer init for other tasks");
1240 24 : }
1241 :
1242 24 : let value = ResidentOrWantedEvicted::Resident(res.clone());
1243 24 :
1244 24 : self.inner.set(value, permit);
1245 24 :
1246 24 : res
1247 24 : }
1248 :
1249 656 : async fn needs_download(&self) -> Result<Option<NeedsDownload>, std::io::Error> {
1250 656 : match tokio::fs::metadata(&self.path).await {
1251 640 : Ok(m) => Ok(self.is_file_present_and_good_size(&m).err()),
1252 16 : Err(e) if e.kind() == std::io::ErrorKind::NotFound => Ok(Some(NeedsDownload::NotFound)),
1253 0 : Err(e) => Err(e),
1254 : }
1255 656 : }
1256 :
1257 244 : fn needs_download_blocking(&self) -> Result<Option<NeedsDownload>, std::io::Error> {
1258 244 : match self.path.metadata() {
1259 244 : Ok(m) => Ok(self.is_file_present_and_good_size(&m).err()),
1260 0 : Err(e) if e.kind() == std::io::ErrorKind::NotFound => Ok(Some(NeedsDownload::NotFound)),
1261 0 : Err(e) => Err(e),
1262 : }
1263 244 : }
1264 :
1265 884 : fn is_file_present_and_good_size(&self, m: &std::fs::Metadata) -> Result<(), NeedsDownload> {
1266 884 : // in future, this should include sha2-256 validation of the file.
1267 884 : if !m.is_file() {
1268 0 : Err(NeedsDownload::NotFile(m.file_type()))
1269 884 : } else if m.len() != self.desc.file_size {
1270 0 : Err(NeedsDownload::WrongSize {
1271 0 : actual: m.len(),
1272 0 : expected: self.desc.file_size,
1273 0 : })
1274 : } else {
1275 884 : Ok(())
1276 : }
1277 884 : }
1278 :
1279 0 : fn info(&self, reset: LayerAccessStatsReset) -> HistoricLayerInfo {
1280 0 : let layer_name = self.desc.layer_name().to_string();
1281 0 :
1282 0 : let resident = self
1283 0 : .inner
1284 0 : .get()
1285 0 : .map(|rowe| rowe.is_likely_resident())
1286 0 : .unwrap_or(false);
1287 0 :
1288 0 : let access_stats = self.access_stats.as_api_model(reset);
1289 0 :
1290 0 : if self.desc.is_delta {
1291 0 : let lsn_range = &self.desc.lsn_range;
1292 0 :
1293 0 : HistoricLayerInfo::Delta {
1294 0 : layer_file_name: layer_name,
1295 0 : layer_file_size: self.desc.file_size,
1296 0 : lsn_start: lsn_range.start,
1297 0 : lsn_end: lsn_range.end,
1298 0 : remote: !resident,
1299 0 : access_stats,
1300 0 : l0: crate::tenant::layer_map::LayerMap::is_l0(
1301 0 : &self.layer_desc().key_range,
1302 0 : self.layer_desc().is_delta,
1303 0 : ),
1304 0 : }
1305 : } else {
1306 0 : let lsn = self.desc.image_layer_lsn();
1307 0 :
1308 0 : HistoricLayerInfo::Image {
1309 0 : layer_file_name: layer_name,
1310 0 : layer_file_size: self.desc.file_size,
1311 0 : lsn_start: lsn,
1312 0 : remote: !resident,
1313 0 : access_stats,
1314 0 : }
1315 : }
1316 0 : }
1317 :
1318 : /// `DownloadedLayer` is being dropped, so it calls this method.
1319 48 : fn on_downloaded_layer_drop(self: Arc<LayerInner>, only_version: usize) {
1320 : // we cannot know without inspecting LayerInner::inner if we should evict or not, even
1321 : // though here it is very likely
1322 48 : 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);
1323 :
1324 : // NOTE: this scope *must* never call `self.inner.get` because evict_and_wait might
1325 : // drop while the `self.inner` is being locked, leading to a deadlock.
1326 :
1327 48 : let start_evicting = async move {
1328 48 : #[cfg(test)]
1329 48 : self.failpoint(failpoints::FailpointKind::WaitBeforeStartingEvicting)
1330 48 : .await
1331 48 : .expect("failpoint should not have errored");
1332 48 :
1333 48 : tracing::debug!("eviction started");
1334 :
1335 48 : let res = self.wait_for_turn_and_evict(only_version).await;
1336 : // metrics: ignore the Ok branch, it is not done yet
1337 48 : if let Err(e) = res {
1338 12 : tracing::debug!(res=?Err::<(), _>(&e), "eviction completed");
1339 12 : LAYER_IMPL_METRICS.inc_eviction_cancelled(e);
1340 36 : }
1341 48 : };
1342 :
1343 48 : Self::spawn(start_evicting.instrument(span));
1344 48 : }
1345 :
1346 48 : async fn wait_for_turn_and_evict(
1347 48 : self: Arc<LayerInner>,
1348 48 : only_version: usize,
1349 48 : ) -> Result<(), EvictionCancelled> {
1350 92 : fn is_good_to_continue(status: &Status) -> Result<(), EvictionCancelled> {
1351 : use Status::*;
1352 92 : match status {
1353 88 : Resident => Ok(()),
1354 4 : Evicted => Err(EvictionCancelled::UnexpectedEvictedState),
1355 0 : Downloading => Err(EvictionCancelled::LostToDownload),
1356 : }
1357 92 : }
1358 :
1359 48 : let timeline = self
1360 48 : .timeline
1361 48 : .upgrade()
1362 48 : .ok_or(EvictionCancelled::TimelineGone)?;
1363 :
1364 48 : let mut rx = self
1365 48 : .status
1366 48 : .as_ref()
1367 48 : .expect("LayerInner cannot be dropped, holding strong ref")
1368 48 : .subscribe();
1369 48 :
1370 48 : is_good_to_continue(&rx.borrow_and_update())?;
1371 :
1372 44 : let Ok(gate) = timeline.gate.enter() else {
1373 0 : return Err(EvictionCancelled::TimelineGone);
1374 : };
1375 :
1376 36 : let permit = {
1377 : // we cannot just `std::fs::remove_file` because there might already be an
1378 : // get_or_maybe_download which will inspect filesystem and reinitialize. filesystem
1379 : // operations must be done while holding the heavier_once_cell::InitPermit
1380 44 : let mut wait = std::pin::pin!(self.inner.get_or_init_detached());
1381 :
1382 44 : let waited = loop {
1383 : // we must race to the Downloading starting, otherwise we would have to wait until the
1384 : // completion of the download. waiting for download could be long and hinder our
1385 : // efforts to alert on "hanging" evictions.
1386 44 : tokio::select! {
1387 44 : res = &mut wait => break res,
1388 44 : _ = rx.changed() => {
1389 0 : is_good_to_continue(&rx.borrow_and_update())?;
1390 : // two possibilities for Status::Resident:
1391 : // - the layer was found locally from disk by a read
1392 : // - we missed a bunch of updates and now the layer is
1393 : // again downloaded -- assume we'll fail later on with
1394 : // version check or AlreadyReinitialized
1395 : }
1396 : }
1397 : };
1398 :
1399 : // re-check now that we have the guard or permit; all updates should have happened
1400 : // while holding the permit.
1401 44 : is_good_to_continue(&rx.borrow_and_update())?;
1402 :
1403 : // the term deinitialize is used here, because we clearing out the Weak will eventually
1404 : // lead to deallocating the reference counted value, and the value we
1405 : // `Guard::take_and_deinit` is likely to be the last because the Weak is never cloned.
1406 44 : let (_weak, permit) = match waited {
1407 40 : Ok(guard) => {
1408 40 : match &*guard {
1409 36 : ResidentOrWantedEvicted::WantedEvicted(_weak, version)
1410 36 : if *version == only_version =>
1411 32 : {
1412 32 : tracing::debug!(version, "deinitializing matching WantedEvicted");
1413 32 : let (weak, permit) = guard.take_and_deinit();
1414 32 : (Some(weak), permit)
1415 : }
1416 4 : ResidentOrWantedEvicted::WantedEvicted(_, version) => {
1417 4 : // if we were not doing the version check, we would need to try to
1418 4 : // upgrade the weak here to see if it really is dropped. version check
1419 4 : // is done instead assuming that it is cheaper.
1420 4 : tracing::debug!(
1421 : version,
1422 : only_version,
1423 0 : "version mismatch, not deinitializing"
1424 : );
1425 4 : return Err(EvictionCancelled::VersionCheckFailed);
1426 : }
1427 : ResidentOrWantedEvicted::Resident(_) => {
1428 4 : return Err(EvictionCancelled::AlreadyReinitialized);
1429 : }
1430 : }
1431 : }
1432 4 : Err(permit) => {
1433 4 : tracing::debug!("continuing after cancelled get_or_maybe_download or eviction");
1434 4 : (None, permit)
1435 : }
1436 : };
1437 :
1438 36 : permit
1439 36 : };
1440 36 :
1441 36 : let span = tracing::Span::current();
1442 36 :
1443 36 : let spawned_at = std::time::Instant::now();
1444 36 :
1445 36 : // this is on purpose a detached spawn; we don't need to wait for it
1446 36 : //
1447 36 : // eviction completion reporting is the only thing hinging on this, and it can be just as
1448 36 : // well from a spawn_blocking thread.
1449 36 : //
1450 36 : // important to note that now that we've acquired the permit we have made sure the evicted
1451 36 : // file is either the exact `WantedEvicted` we wanted to evict, or uninitialized in case
1452 36 : // there are multiple evictions. The rest is not cancellable, and we've now commited to
1453 36 : // evicting.
1454 36 : //
1455 36 : // If spawn_blocking has a queue and maximum number of threads are in use, we could stall
1456 36 : // reads. We will need to add cancellation for that if necessary.
1457 36 : Self::spawn_blocking(move || {
1458 36 : let _span = span.entered();
1459 36 :
1460 36 : let res = self.evict_blocking(&timeline, &gate, &permit);
1461 36 :
1462 36 : let waiters = self.inner.initializer_count();
1463 36 :
1464 36 : if waiters > 0 {
1465 0 : LAYER_IMPL_METRICS.inc_evicted_with_waiters();
1466 36 : }
1467 :
1468 36 : let completed_in = spawned_at.elapsed();
1469 36 : LAYER_IMPL_METRICS.record_time_to_evict(completed_in);
1470 36 :
1471 36 : match res {
1472 36 : Ok(()) => LAYER_IMPL_METRICS.inc_completed_evictions(),
1473 0 : Err(e) => LAYER_IMPL_METRICS.inc_eviction_cancelled(e),
1474 : }
1475 :
1476 36 : tracing::debug!(?res, elapsed_ms=%completed_in.as_millis(), %waiters, "eviction completed");
1477 36 : });
1478 36 :
1479 36 : Ok(())
1480 48 : }
1481 :
1482 : /// This is blocking only to do just one spawn_blocking hop compared to multiple via tokio::fs.
1483 36 : fn evict_blocking(
1484 36 : &self,
1485 36 : timeline: &Timeline,
1486 36 : _gate: &gate::GateGuard,
1487 36 : _permit: &heavier_once_cell::InitPermit,
1488 36 : ) -> Result<(), EvictionCancelled> {
1489 36 : // now accesses to `self.inner.get_or_init*` wait on the semaphore or the `_permit`
1490 36 :
1491 36 : match capture_mtime_and_remove(&self.path) {
1492 36 : Ok(local_layer_mtime) => {
1493 36 : let duration = SystemTime::now().duration_since(local_layer_mtime);
1494 36 : match duration {
1495 36 : Ok(elapsed) => {
1496 36 : let accessed_and_visible = self.access_stats.accessed()
1497 8 : && self.access_stats.visibility() == LayerVisibilityHint::Visible;
1498 36 : if accessed_and_visible {
1499 8 : // Only layers used for reads contribute to our "low residence" metric that is used
1500 8 : // to detect thrashing. Layers promoted for other reasons (e.g. compaction) are allowed
1501 8 : // to be rapidly evicted without contributing to this metric.
1502 8 : timeline
1503 8 : .metrics
1504 8 : .evictions_with_low_residence_duration
1505 8 : .read()
1506 8 : .unwrap()
1507 8 : .observe(elapsed);
1508 28 : }
1509 :
1510 36 : tracing::info!(
1511 0 : residence_millis = elapsed.as_millis(),
1512 0 : accessed_and_visible,
1513 0 : "evicted layer after known residence period"
1514 : );
1515 : }
1516 : Err(_) => {
1517 0 : tracing::info!("evicted layer after unknown residence period");
1518 : }
1519 : }
1520 36 : timeline.metrics.evictions.inc();
1521 36 : timeline
1522 36 : .metrics
1523 36 : .resident_physical_size_sub(self.desc.file_size);
1524 : }
1525 0 : Err(e) if e.kind() == std::io::ErrorKind::NotFound => {
1526 0 : tracing::error!(
1527 : layer_size = %self.desc.file_size,
1528 0 : "failed to evict layer from disk, it was already gone"
1529 : );
1530 0 : return Err(EvictionCancelled::FileNotFound);
1531 : }
1532 0 : Err(e) => {
1533 0 : // FIXME: this should probably be an abort
1534 0 : tracing::error!("failed to evict file from disk: {e:#}");
1535 0 : return Err(EvictionCancelled::RemoveFailed);
1536 : }
1537 : }
1538 :
1539 36 : self.access_stats.record_residence_event();
1540 36 :
1541 36 : self.status.as_ref().unwrap().send_replace(Status::Evicted);
1542 36 :
1543 36 : *self.last_evicted_at.lock().unwrap() = Some(std::time::Instant::now());
1544 36 :
1545 36 : Ok(())
1546 36 : }
1547 :
1548 6178 : fn metadata(&self) -> LayerFileMetadata {
1549 6178 : LayerFileMetadata::new(self.desc.file_size, self.generation, self.shard)
1550 6178 : }
1551 :
1552 : /// Needed to use entered runtime in tests, but otherwise use BACKGROUND_RUNTIME.
1553 : ///
1554 : /// Synchronizing with spawned tasks is very complicated otherwise.
1555 60 : fn spawn<F>(fut: F)
1556 60 : where
1557 60 : F: std::future::Future<Output = ()> + Send + 'static,
1558 60 : {
1559 60 : #[cfg(test)]
1560 60 : tokio::task::spawn(fut);
1561 60 : #[cfg(not(test))]
1562 60 : crate::task_mgr::BACKGROUND_RUNTIME.spawn(fut);
1563 60 : }
1564 :
1565 : /// Needed to use entered runtime in tests, but otherwise use BACKGROUND_RUNTIME.
1566 1054 : fn spawn_blocking<F>(f: F)
1567 1054 : where
1568 1054 : F: FnOnce() + Send + 'static,
1569 1054 : {
1570 1054 : #[cfg(test)]
1571 1054 : tokio::task::spawn_blocking(f);
1572 1054 : #[cfg(not(test))]
1573 1054 : crate::task_mgr::BACKGROUND_RUNTIME.spawn_blocking(f);
1574 1054 : }
1575 : }
1576 :
1577 36 : fn capture_mtime_and_remove(path: &Utf8Path) -> Result<SystemTime, std::io::Error> {
1578 36 : let m = path.metadata()?;
1579 36 : let local_layer_mtime = m.modified()?;
1580 36 : std::fs::remove_file(path)?;
1581 36 : Ok(local_layer_mtime)
1582 36 : }
1583 :
1584 : #[derive(Debug, thiserror::Error)]
1585 : pub(crate) enum EvictionError {
1586 : #[error("layer was already evicted")]
1587 : NotFound,
1588 :
1589 : /// Evictions must always lose to downloads in races, and this time it happened.
1590 : #[error("layer was downloaded instead")]
1591 : Downloaded,
1592 :
1593 : #[error("eviction did not happen within timeout")]
1594 : Timeout,
1595 : }
1596 :
1597 : /// Error internal to the [`LayerInner::get_or_maybe_download`]
1598 : #[derive(Debug, thiserror::Error)]
1599 : pub(crate) enum DownloadError {
1600 : #[error("timeline has already shutdown")]
1601 : TimelineShutdown,
1602 : #[error("context denies downloading")]
1603 : ContextAndConfigReallyDeniesDownloads,
1604 : #[error("downloading is really required but not allowed by this method")]
1605 : DownloadRequired,
1606 : #[error("layer path exists, but it is not a file: {0:?}")]
1607 : NotFile(std::fs::FileType),
1608 : /// Why no error here? Because it will be reported by page_service. We should had also done
1609 : /// retries already.
1610 : #[error("downloading evicted layer file failed")]
1611 : DownloadFailed,
1612 : #[error("downloading failed, possibly for shutdown")]
1613 : DownloadCancelled,
1614 : #[error("pre-condition: stat before download failed")]
1615 : PreStatFailed(#[source] std::io::Error),
1616 :
1617 : #[cfg(test)]
1618 : #[error("failpoint: {0:?}")]
1619 : Failpoint(failpoints::FailpointKind),
1620 : }
1621 :
1622 : impl DownloadError {
1623 0 : pub(crate) fn is_cancelled(&self) -> bool {
1624 0 : matches!(self, DownloadError::DownloadCancelled)
1625 0 : }
1626 : }
1627 :
1628 : #[derive(Debug, PartialEq)]
1629 : pub(crate) enum NeedsDownload {
1630 : NotFound,
1631 : NotFile(std::fs::FileType),
1632 : WrongSize { actual: u64, expected: u64 },
1633 : }
1634 :
1635 : impl std::fmt::Display for NeedsDownload {
1636 12 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
1637 12 : match self {
1638 12 : NeedsDownload::NotFound => write!(f, "file was not found"),
1639 0 : NeedsDownload::NotFile(ft) => write!(f, "path is not a file; {ft:?}"),
1640 0 : NeedsDownload::WrongSize { actual, expected } => {
1641 0 : write!(f, "file size mismatch {actual} vs. {expected}")
1642 : }
1643 : }
1644 12 : }
1645 : }
1646 :
1647 : /// Existence of `DownloadedLayer` means that we have the file locally, and can later evict it.
1648 : pub(crate) struct DownloadedLayer {
1649 : owner: Weak<LayerInner>,
1650 : // Use tokio OnceCell as we do not need to deinitialize this, it'll just get dropped with the
1651 : // DownloadedLayer
1652 : kind: tokio::sync::OnceCell<anyhow::Result<LayerKind>>,
1653 : version: usize,
1654 : }
1655 :
1656 : impl std::fmt::Debug for DownloadedLayer {
1657 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
1658 0 : f.debug_struct("DownloadedLayer")
1659 0 : // owner omitted because it is always "Weak"
1660 0 : .field("kind", &self.kind)
1661 0 : .field("version", &self.version)
1662 0 : .finish()
1663 0 : }
1664 : }
1665 :
1666 : impl Drop for DownloadedLayer {
1667 1454 : fn drop(&mut self) {
1668 1454 : if let Some(owner) = self.owner.upgrade() {
1669 48 : owner.on_downloaded_layer_drop(self.version);
1670 1406 : } else {
1671 1406 : // Layer::drop will handle cancelling the eviction; because of drop order and
1672 1406 : // `DownloadedLayer` never leaking, we cannot know here if eviction was requested.
1673 1406 : }
1674 1454 : }
1675 : }
1676 :
1677 : impl DownloadedLayer {
1678 : /// Initializes the `DeltaLayerInner` or `ImageLayerInner` within [`LayerKind`].
1679 : /// Failure to load the layer is sticky, i.e., future `get()` calls will return
1680 : /// the initial load failure immediately.
1681 : ///
1682 : /// `owner` parameter is a strong reference at the same `LayerInner` as the
1683 : /// `DownloadedLayer::owner` would be when upgraded. Given how this method ends up called,
1684 : /// we will always have the LayerInner on the callstack, so we can just use it.
1685 556832 : async fn get<'a>(
1686 556832 : &'a self,
1687 556832 : owner: &Arc<LayerInner>,
1688 556832 : ctx: &RequestContext,
1689 556832 : ) -> anyhow::Result<&'a LayerKind> {
1690 556832 : let init = || async {
1691 2468 : assert_eq!(
1692 2468 : Weak::as_ptr(&self.owner),
1693 2468 : Arc::as_ptr(owner),
1694 0 : "these are the same, just avoiding the upgrade"
1695 : );
1696 :
1697 2468 : let res = if owner.desc.is_delta {
1698 2216 : let ctx = RequestContextBuilder::extend(ctx)
1699 2216 : .page_content_kind(crate::context::PageContentKind::DeltaLayerSummary)
1700 2216 : .build();
1701 2216 : let summary = Some(delta_layer::Summary::expected(
1702 2216 : owner.desc.tenant_shard_id.tenant_id,
1703 2216 : owner.desc.timeline_id,
1704 2216 : owner.desc.key_range.clone(),
1705 2216 : owner.desc.lsn_range.clone(),
1706 2216 : ));
1707 2216 : delta_layer::DeltaLayerInner::load(
1708 2216 : &owner.path,
1709 2216 : summary,
1710 2216 : Some(owner.conf.max_vectored_read_bytes),
1711 2216 : &ctx,
1712 2216 : )
1713 2216 : .await
1714 2216 : .map(LayerKind::Delta)
1715 : } else {
1716 252 : let ctx = RequestContextBuilder::extend(ctx)
1717 252 : .page_content_kind(crate::context::PageContentKind::ImageLayerSummary)
1718 252 : .build();
1719 252 : let lsn = owner.desc.image_layer_lsn();
1720 252 : let summary = Some(image_layer::Summary::expected(
1721 252 : owner.desc.tenant_shard_id.tenant_id,
1722 252 : owner.desc.timeline_id,
1723 252 : owner.desc.key_range.clone(),
1724 252 : lsn,
1725 252 : ));
1726 252 : image_layer::ImageLayerInner::load(
1727 252 : &owner.path,
1728 252 : lsn,
1729 252 : summary,
1730 252 : Some(owner.conf.max_vectored_read_bytes),
1731 252 : &ctx,
1732 252 : )
1733 252 : .await
1734 252 : .map(LayerKind::Image)
1735 : };
1736 :
1737 2468 : match res {
1738 2468 : Ok(layer) => Ok(layer),
1739 0 : Err(err) => {
1740 0 : LAYER_IMPL_METRICS.inc_permanent_loading_failures();
1741 0 : // We log this message once over the lifetime of `Self`
1742 0 : // => Ok and good to log backtrace and path here.
1743 0 : tracing::error!(
1744 0 : "layer load failed, assuming permanent failure: {}: {err:?}",
1745 0 : owner.path
1746 : );
1747 0 : Err(err)
1748 : }
1749 : }
1750 4936 : };
1751 556832 : self.kind
1752 556832 : .get_or_init(init)
1753 556832 : .await
1754 556832 : .as_ref()
1755 556832 : // We already logged the full backtrace above, once. Don't repeat that here.
1756 556832 : .map_err(|e| anyhow::anyhow!("layer load failed earlier: {e}"))
1757 556832 : }
1758 :
1759 554744 : async fn get_values_reconstruct_data(
1760 554744 : &self,
1761 554744 : this: ResidentLayer,
1762 554744 : keyspace: KeySpace,
1763 554744 : lsn_range: Range<Lsn>,
1764 554744 : reconstruct_data: &mut ValuesReconstructState,
1765 554744 : ctx: &RequestContext,
1766 554744 : ) -> Result<(), GetVectoredError> {
1767 : use LayerKind::*;
1768 :
1769 554744 : match self
1770 554744 : .get(&this.owner.0, ctx)
1771 554744 : .await
1772 554744 : .map_err(GetVectoredError::Other)?
1773 : {
1774 549728 : Delta(d) => {
1775 549728 : d.get_values_reconstruct_data(this, keyspace, lsn_range, reconstruct_data, ctx)
1776 549728 : .await
1777 : }
1778 5016 : Image(i) => {
1779 5016 : i.get_values_reconstruct_data(this, keyspace, reconstruct_data, ctx)
1780 5016 : .await
1781 : }
1782 : }
1783 554744 : }
1784 :
1785 8 : async fn dump(&self, owner: &Arc<LayerInner>, ctx: &RequestContext) -> anyhow::Result<()> {
1786 : use LayerKind::*;
1787 8 : match self.get(owner, ctx).await? {
1788 8 : Delta(d) => d.dump(ctx).await?,
1789 0 : Image(i) => i.dump(ctx).await?,
1790 : }
1791 :
1792 8 : Ok(())
1793 8 : }
1794 : }
1795 :
1796 : /// Wrapper around an actual layer implementation.
1797 : #[derive(Debug)]
1798 : enum LayerKind {
1799 : Delta(delta_layer::DeltaLayerInner),
1800 : Image(image_layer::ImageLayerInner),
1801 : }
1802 :
1803 : /// Guard for forcing a layer be resident while it exists.
1804 : #[derive(Clone)]
1805 : pub struct ResidentLayer {
1806 : owner: Layer,
1807 : downloaded: Arc<DownloadedLayer>,
1808 : }
1809 :
1810 : impl std::fmt::Display for ResidentLayer {
1811 4284 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
1812 4284 : write!(f, "{}", self.owner)
1813 4284 : }
1814 : }
1815 :
1816 : impl std::fmt::Debug for ResidentLayer {
1817 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
1818 0 : write!(f, "{}", self.owner)
1819 0 : }
1820 : }
1821 :
1822 : impl ResidentLayer {
1823 : /// Release the eviction guard, converting back into a plain [`Layer`].
1824 : ///
1825 : /// You can access the [`Layer`] also by using `as_ref`.
1826 840 : pub(crate) fn drop_eviction_guard(self) -> Layer {
1827 840 : self.into()
1828 840 : }
1829 :
1830 : /// Loads all keys stored in the layer. Returns key, lsn and value size.
1831 0 : #[tracing::instrument(level = tracing::Level::DEBUG, skip_all, fields(layer=%self))]
1832 : pub(crate) async fn load_keys<'a>(
1833 : &'a self,
1834 : ctx: &RequestContext,
1835 : ) -> anyhow::Result<Vec<pageserver_api::key::Key>> {
1836 : use LayerKind::*;
1837 :
1838 : let owner = &self.owner.0;
1839 : let inner = self.downloaded.get(owner, ctx).await?;
1840 :
1841 : // this is valid because the DownloadedLayer::kind is a OnceCell, not a
1842 : // Mutex<OnceCell>, so we cannot go and deinitialize the value with OnceCell::take
1843 : // while it's being held.
1844 : self.owner.record_access(ctx);
1845 :
1846 : let res = match inner {
1847 : Delta(ref d) => delta_layer::DeltaLayerInner::load_keys(d, ctx).await,
1848 : Image(ref i) => image_layer::ImageLayerInner::load_keys(i, ctx).await,
1849 : };
1850 0 : res.with_context(|| format!("Layer index is corrupted for {self}"))
1851 : }
1852 :
1853 : /// Read all they keys in this layer which match the ShardIdentity, and write them all to
1854 : /// the provided writer. Return the number of keys written.
1855 16 : #[tracing::instrument(level = tracing::Level::DEBUG, skip_all, fields(layer=%self))]
1856 : pub(crate) async fn filter(
1857 : &self,
1858 : shard_identity: &ShardIdentity,
1859 : writer: &mut ImageLayerWriter,
1860 : ctx: &RequestContext,
1861 : ) -> Result<usize, CompactionError> {
1862 : use LayerKind::*;
1863 :
1864 : match self
1865 : .downloaded
1866 : .get(&self.owner.0, ctx)
1867 : .await
1868 : .map_err(CompactionError::Other)?
1869 : {
1870 : Delta(_) => {
1871 : return Err(CompactionError::Other(anyhow::anyhow!(format!(
1872 : "cannot filter() on a delta layer {self}"
1873 : ))));
1874 : }
1875 : Image(i) => i
1876 : .filter(shard_identity, writer, ctx)
1877 : .await
1878 : .map_err(CompactionError::Other),
1879 : }
1880 : }
1881 :
1882 : /// Returns the amount of keys and values written to the writer.
1883 20 : pub(crate) async fn copy_delta_prefix(
1884 20 : &self,
1885 20 : writer: &mut super::delta_layer::DeltaLayerWriter,
1886 20 : until: Lsn,
1887 20 : ctx: &RequestContext,
1888 20 : ) -> anyhow::Result<usize> {
1889 : use LayerKind::*;
1890 :
1891 20 : let owner = &self.owner.0;
1892 20 :
1893 20 : match self.downloaded.get(owner, ctx).await? {
1894 20 : Delta(ref d) => d
1895 20 : .copy_prefix(writer, until, ctx)
1896 20 : .await
1897 20 : .with_context(|| format!("copy_delta_prefix until {until} of {self}")),
1898 0 : Image(_) => anyhow::bail!(format!("cannot copy_lsn_prefix of image layer {self}")),
1899 : }
1900 20 : }
1901 :
1902 3640 : pub(crate) fn local_path(&self) -> &Utf8Path {
1903 3640 : &self.owner.0.path
1904 3640 : }
1905 :
1906 4372 : pub(crate) fn metadata(&self) -> LayerFileMetadata {
1907 4372 : self.owner.metadata()
1908 4372 : }
1909 :
1910 : /// Cast the layer to a delta, return an error if it is an image layer.
1911 1908 : pub(crate) async fn get_as_delta(
1912 1908 : &self,
1913 1908 : ctx: &RequestContext,
1914 1908 : ) -> anyhow::Result<&delta_layer::DeltaLayerInner> {
1915 : use LayerKind::*;
1916 1908 : match self.downloaded.get(&self.owner.0, ctx).await? {
1917 1908 : Delta(ref d) => Ok(d),
1918 0 : Image(_) => Err(anyhow::anyhow!("image layer")),
1919 : }
1920 1908 : }
1921 :
1922 : /// Cast the layer to an image, return an error if it is a delta layer.
1923 136 : pub(crate) async fn get_as_image(
1924 136 : &self,
1925 136 : ctx: &RequestContext,
1926 136 : ) -> anyhow::Result<&image_layer::ImageLayerInner> {
1927 : use LayerKind::*;
1928 136 : match self.downloaded.get(&self.owner.0, ctx).await? {
1929 136 : Image(ref d) => Ok(d),
1930 0 : Delta(_) => Err(anyhow::anyhow!("delta layer")),
1931 : }
1932 136 : }
1933 : }
1934 :
1935 : impl AsLayerDesc for ResidentLayer {
1936 2320124 : fn layer_desc(&self) -> &PersistentLayerDesc {
1937 2320124 : self.owner.layer_desc()
1938 2320124 : }
1939 : }
1940 :
1941 : impl AsRef<Layer> for ResidentLayer {
1942 4116 : fn as_ref(&self) -> &Layer {
1943 4116 : &self.owner
1944 4116 : }
1945 : }
1946 :
1947 : /// Drop the eviction guard.
1948 : impl From<ResidentLayer> for Layer {
1949 840 : fn from(value: ResidentLayer) -> Self {
1950 840 : value.owner
1951 840 : }
1952 : }
1953 :
1954 : use metrics::IntCounter;
1955 :
1956 : pub(crate) struct LayerImplMetrics {
1957 : started_evictions: IntCounter,
1958 : completed_evictions: IntCounter,
1959 : cancelled_evictions: enum_map::EnumMap<EvictionCancelled, IntCounter>,
1960 :
1961 : started_deletes: IntCounter,
1962 : completed_deletes: IntCounter,
1963 : failed_deletes: enum_map::EnumMap<DeleteFailed, IntCounter>,
1964 :
1965 : rare_counters: enum_map::EnumMap<RareEvent, IntCounter>,
1966 : inits_cancelled: metrics::core::GenericCounter<metrics::core::AtomicU64>,
1967 : redownload_after: metrics::Histogram,
1968 : time_to_evict: metrics::Histogram,
1969 : }
1970 :
1971 : impl Default for LayerImplMetrics {
1972 100 : fn default() -> Self {
1973 : use enum_map::Enum;
1974 :
1975 : // reminder: these will be pageserver_layer_* with "_total" suffix
1976 :
1977 100 : let started_evictions = metrics::register_int_counter!(
1978 100 : "pageserver_layer_started_evictions",
1979 100 : "Evictions started in the Layer implementation"
1980 100 : )
1981 100 : .unwrap();
1982 100 : let completed_evictions = metrics::register_int_counter!(
1983 100 : "pageserver_layer_completed_evictions",
1984 100 : "Evictions completed in the Layer implementation"
1985 100 : )
1986 100 : .unwrap();
1987 100 :
1988 100 : let cancelled_evictions = metrics::register_int_counter_vec!(
1989 100 : "pageserver_layer_cancelled_evictions_count",
1990 100 : "Different reasons for evictions to have been cancelled or failed",
1991 100 : &["reason"]
1992 100 : )
1993 100 : .unwrap();
1994 100 :
1995 900 : let cancelled_evictions = enum_map::EnumMap::from_array(std::array::from_fn(|i| {
1996 900 : let reason = EvictionCancelled::from_usize(i);
1997 900 : let s = reason.as_str();
1998 900 : cancelled_evictions.with_label_values(&[s])
1999 900 : }));
2000 100 :
2001 100 : let started_deletes = metrics::register_int_counter!(
2002 100 : "pageserver_layer_started_deletes",
2003 100 : "Deletions on drop pending in the Layer implementation"
2004 100 : )
2005 100 : .unwrap();
2006 100 : let completed_deletes = metrics::register_int_counter!(
2007 100 : "pageserver_layer_completed_deletes",
2008 100 : "Deletions on drop completed in the Layer implementation"
2009 100 : )
2010 100 : .unwrap();
2011 100 :
2012 100 : let failed_deletes = metrics::register_int_counter_vec!(
2013 100 : "pageserver_layer_failed_deletes_count",
2014 100 : "Different reasons for deletions on drop to have failed",
2015 100 : &["reason"]
2016 100 : )
2017 100 : .unwrap();
2018 100 :
2019 200 : let failed_deletes = enum_map::EnumMap::from_array(std::array::from_fn(|i| {
2020 200 : let reason = DeleteFailed::from_usize(i);
2021 200 : let s = reason.as_str();
2022 200 : failed_deletes.with_label_values(&[s])
2023 200 : }));
2024 100 :
2025 100 : let rare_counters = metrics::register_int_counter_vec!(
2026 100 : "pageserver_layer_assumed_rare_count",
2027 100 : "Times unexpected or assumed rare event happened",
2028 100 : &["event"]
2029 100 : )
2030 100 : .unwrap();
2031 100 :
2032 700 : let rare_counters = enum_map::EnumMap::from_array(std::array::from_fn(|i| {
2033 700 : let event = RareEvent::from_usize(i);
2034 700 : let s = event.as_str();
2035 700 : rare_counters.with_label_values(&[s])
2036 700 : }));
2037 100 :
2038 100 : let inits_cancelled = metrics::register_int_counter!(
2039 100 : "pageserver_layer_inits_cancelled_count",
2040 100 : "Times Layer initialization was cancelled",
2041 100 : )
2042 100 : .unwrap();
2043 100 :
2044 100 : let redownload_after = {
2045 100 : let minute = 60.0;
2046 100 : let hour = 60.0 * minute;
2047 100 : metrics::register_histogram!(
2048 100 : "pageserver_layer_redownloaded_after",
2049 100 : "Time between evicting and re-downloading.",
2050 100 : vec![
2051 100 : 10.0,
2052 100 : 30.0,
2053 100 : minute,
2054 100 : 5.0 * minute,
2055 100 : 15.0 * minute,
2056 100 : 30.0 * minute,
2057 100 : hour,
2058 100 : 12.0 * hour,
2059 100 : ]
2060 100 : )
2061 100 : .unwrap()
2062 100 : };
2063 100 :
2064 100 : let time_to_evict = metrics::register_histogram!(
2065 100 : "pageserver_layer_eviction_held_permit_seconds",
2066 100 : "Time eviction held the permit.",
2067 100 : vec![0.001, 0.010, 0.100, 0.500, 1.000, 5.000]
2068 100 : )
2069 100 : .unwrap();
2070 100 :
2071 100 : Self {
2072 100 : started_evictions,
2073 100 : completed_evictions,
2074 100 : cancelled_evictions,
2075 100 :
2076 100 : started_deletes,
2077 100 : completed_deletes,
2078 100 : failed_deletes,
2079 100 :
2080 100 : rare_counters,
2081 100 : inits_cancelled,
2082 100 : redownload_after,
2083 100 : time_to_evict,
2084 100 : }
2085 100 : }
2086 : }
2087 :
2088 : impl LayerImplMetrics {
2089 52 : fn inc_started_evictions(&self) {
2090 52 : self.started_evictions.inc();
2091 52 : }
2092 36 : fn inc_completed_evictions(&self) {
2093 36 : self.completed_evictions.inc();
2094 36 : }
2095 16 : fn inc_eviction_cancelled(&self, reason: EvictionCancelled) {
2096 16 : self.cancelled_evictions[reason].inc()
2097 16 : }
2098 :
2099 1032 : fn inc_started_deletes(&self) {
2100 1032 : self.started_deletes.inc();
2101 1032 : }
2102 1013 : fn inc_completed_deletes(&self) {
2103 1013 : self.completed_deletes.inc();
2104 1013 : }
2105 0 : fn inc_deletes_failed(&self, reason: DeleteFailed) {
2106 0 : self.failed_deletes[reason].inc();
2107 0 : }
2108 :
2109 : /// Counted separatedly from failed layer deletes because we will complete the layer deletion
2110 : /// attempt regardless of failure to delete local file.
2111 0 : fn inc_delete_removes_failed(&self) {
2112 0 : self.rare_counters[RareEvent::RemoveOnDropFailed].inc();
2113 0 : }
2114 :
2115 : /// Expected rare just as cancellations are rare, but we could have cancellations separate from
2116 : /// the single caller which can start the download, so use this counter to separte them.
2117 0 : fn inc_init_completed_without_requester(&self) {
2118 0 : self.rare_counters[RareEvent::InitCompletedWithoutRequester].inc();
2119 0 : }
2120 :
2121 : /// Expected rare because cancellations are unexpected, and failures are unexpected
2122 0 : fn inc_download_failed_without_requester(&self) {
2123 0 : self.rare_counters[RareEvent::DownloadFailedWithoutRequester].inc();
2124 0 : }
2125 :
2126 : /// The Weak in ResidentOrWantedEvicted::WantedEvicted was successfully upgraded.
2127 : ///
2128 : /// If this counter is always zero, we should replace ResidentOrWantedEvicted type with an
2129 : /// Option.
2130 0 : fn inc_raced_wanted_evicted_accesses(&self) {
2131 0 : self.rare_counters[RareEvent::UpgradedWantedEvicted].inc();
2132 0 : }
2133 :
2134 : /// These are only expected for [`Self::inc_init_cancelled`] amount when
2135 : /// running with remote storage.
2136 12 : fn inc_init_needed_no_download(&self) {
2137 12 : self.rare_counters[RareEvent::InitWithoutDownload].inc();
2138 12 : }
2139 :
2140 : /// Expected rare because all layer files should be readable and good
2141 0 : fn inc_permanent_loading_failures(&self) {
2142 0 : self.rare_counters[RareEvent::PermanentLoadingFailure].inc();
2143 0 : }
2144 :
2145 0 : fn inc_init_cancelled(&self) {
2146 0 : self.inits_cancelled.inc()
2147 0 : }
2148 :
2149 12 : fn record_redownloaded_after(&self, duration: std::time::Duration) {
2150 12 : self.redownload_after.observe(duration.as_secs_f64())
2151 12 : }
2152 :
2153 : /// This would be bad if it ever happened, or mean extreme disk pressure. We should probably
2154 : /// instead cancel eviction if we would have read waiters. We cannot however separate reads
2155 : /// from other evictions, so this could have noise as well.
2156 0 : fn inc_evicted_with_waiters(&self) {
2157 0 : self.rare_counters[RareEvent::EvictedWithWaiters].inc();
2158 0 : }
2159 :
2160 : /// Recorded at least initially as the permit is now acquired in async context before
2161 : /// spawn_blocking action.
2162 36 : fn record_time_to_evict(&self, duration: std::time::Duration) {
2163 36 : self.time_to_evict.observe(duration.as_secs_f64())
2164 36 : }
2165 : }
2166 :
2167 : #[derive(Debug, Clone, Copy, enum_map::Enum)]
2168 : enum EvictionCancelled {
2169 : LayerGone,
2170 : TimelineGone,
2171 : VersionCheckFailed,
2172 : FileNotFound,
2173 : RemoveFailed,
2174 : AlreadyReinitialized,
2175 : /// Not evicted because of a pending reinitialization
2176 : LostToDownload,
2177 : /// After eviction, there was a new layer access which cancelled the eviction.
2178 : UpgradedBackOnAccess,
2179 : UnexpectedEvictedState,
2180 : }
2181 :
2182 : impl EvictionCancelled {
2183 900 : fn as_str(&self) -> &'static str {
2184 900 : match self {
2185 100 : EvictionCancelled::LayerGone => "layer_gone",
2186 100 : EvictionCancelled::TimelineGone => "timeline_gone",
2187 100 : EvictionCancelled::VersionCheckFailed => "version_check_fail",
2188 100 : EvictionCancelled::FileNotFound => "file_not_found",
2189 100 : EvictionCancelled::RemoveFailed => "remove_failed",
2190 100 : EvictionCancelled::AlreadyReinitialized => "already_reinitialized",
2191 100 : EvictionCancelled::LostToDownload => "lost_to_download",
2192 100 : EvictionCancelled::UpgradedBackOnAccess => "upgraded_back_on_access",
2193 100 : EvictionCancelled::UnexpectedEvictedState => "unexpected_evicted_state",
2194 : }
2195 900 : }
2196 : }
2197 :
2198 : #[derive(enum_map::Enum)]
2199 : enum DeleteFailed {
2200 : TimelineGone,
2201 : DeleteSchedulingFailed,
2202 : }
2203 :
2204 : impl DeleteFailed {
2205 200 : fn as_str(&self) -> &'static str {
2206 200 : match self {
2207 100 : DeleteFailed::TimelineGone => "timeline_gone",
2208 100 : DeleteFailed::DeleteSchedulingFailed => "delete_scheduling_failed",
2209 : }
2210 200 : }
2211 : }
2212 :
2213 : #[derive(enum_map::Enum)]
2214 : enum RareEvent {
2215 : RemoveOnDropFailed,
2216 : InitCompletedWithoutRequester,
2217 : DownloadFailedWithoutRequester,
2218 : UpgradedWantedEvicted,
2219 : InitWithoutDownload,
2220 : PermanentLoadingFailure,
2221 : EvictedWithWaiters,
2222 : }
2223 :
2224 : impl RareEvent {
2225 700 : fn as_str(&self) -> &'static str {
2226 : use RareEvent::*;
2227 :
2228 700 : match self {
2229 100 : RemoveOnDropFailed => "remove_on_drop_failed",
2230 100 : InitCompletedWithoutRequester => "init_completed_without",
2231 100 : DownloadFailedWithoutRequester => "download_failed_without",
2232 100 : UpgradedWantedEvicted => "raced_wanted_evicted",
2233 100 : InitWithoutDownload => "init_needed_no_download",
2234 100 : PermanentLoadingFailure => "permanent_loading_failure",
2235 100 : EvictedWithWaiters => "evicted_with_waiters",
2236 : }
2237 700 : }
2238 : }
2239 :
2240 : pub(crate) static LAYER_IMPL_METRICS: once_cell::sync::Lazy<LayerImplMetrics> =
2241 : once_cell::sync::Lazy::new(LayerImplMetrics::default);
|
