Line data Source code
1 : //! Common traits and structs for layers
2 :
3 : pub mod batch_split_writer;
4 : pub mod delta_layer;
5 : pub mod filter_iterator;
6 : pub mod image_layer;
7 : pub mod inmemory_layer;
8 : pub(crate) mod layer;
9 : mod layer_desc;
10 : mod layer_name;
11 : pub mod merge_iterator;
12 :
13 : use std::cmp::Ordering;
14 : use std::collections::hash_map::Entry;
15 : use std::collections::{BinaryHeap, HashMap};
16 : use std::ops::Range;
17 : use std::pin::Pin;
18 : use std::sync::Arc;
19 : use std::sync::atomic::AtomicUsize;
20 : use std::time::{Duration, SystemTime, UNIX_EPOCH};
21 :
22 : use crate::PERF_TRACE_TARGET;
23 : pub use batch_split_writer::{BatchLayerWriter, SplitDeltaLayerWriter, SplitImageLayerWriter};
24 : use bytes::Bytes;
25 : pub use delta_layer::{DeltaLayer, DeltaLayerWriter, ValueRef};
26 : use futures::StreamExt;
27 : use futures::stream::FuturesUnordered;
28 : pub use image_layer::{ImageLayer, ImageLayerWriter};
29 : pub use inmemory_layer::InMemoryLayer;
30 : pub(crate) use layer::{EvictionError, Layer, ResidentLayer};
31 : pub use layer_desc::{PersistentLayerDesc, PersistentLayerKey};
32 : pub use layer_name::{DeltaLayerName, ImageLayerName, LayerName};
33 : use pageserver_api::key::Key;
34 : use pageserver_api::keyspace::{KeySpace, KeySpaceRandomAccum};
35 : use pageserver_api::record::NeonWalRecord;
36 : use pageserver_api::value::Value;
37 : use tracing::{Instrument, info_span, trace};
38 : use utils::lsn::Lsn;
39 : use utils::sync::gate::GateGuard;
40 :
41 : use self::inmemory_layer::InMemoryLayerFileId;
42 : use super::PageReconstructError;
43 : use super::layer_map::InMemoryLayerDesc;
44 : use super::timeline::{GetVectoredError, ReadPath};
45 : use crate::config::PageServerConf;
46 : use crate::context::{
47 : AccessStatsBehavior, PerfInstrumentFutureExt, RequestContext, RequestContextBuilder,
48 : };
49 :
50 0 : pub fn range_overlaps<T>(a: &Range<T>, b: &Range<T>) -> bool
51 0 : where
52 0 : T: PartialOrd<T>,
53 0 : {
54 0 : if a.start < b.start {
55 0 : a.end > b.start
56 : } else {
57 0 : b.end > a.start
58 : }
59 0 : }
60 :
61 : /// Struct used to communicate across calls to 'get_value_reconstruct_data'.
62 : ///
63 : /// Before first call, you can fill in 'page_img' if you have an older cached
64 : /// version of the page available. That can save work in
65 : /// 'get_value_reconstruct_data', as it can stop searching for page versions
66 : /// when all the WAL records going back to the cached image have been collected.
67 : ///
68 : /// When get_value_reconstruct_data returns Complete, 'img' is set to an image
69 : /// of the page, or the oldest WAL record in 'records' is a will_init-type
70 : /// record that initializes the page without requiring a previous image.
71 : ///
72 : /// If 'get_page_reconstruct_data' returns Continue, some 'records' may have
73 : /// been collected, but there are more records outside the current layer. Pass
74 : /// the same ValueReconstructState struct in the next 'get_value_reconstruct_data'
75 : /// call, to collect more records.
76 : ///
77 : #[derive(Debug, Default)]
78 : pub(crate) struct ValueReconstructState {
79 : pub(crate) records: Vec<(Lsn, NeonWalRecord)>,
80 : pub(crate) img: Option<(Lsn, Bytes)>,
81 : }
82 :
83 : impl ValueReconstructState {
84 : /// Returns the number of page deltas applied to the page image.
85 2672614 : pub fn num_deltas(&self) -> usize {
86 2672614 : match self.img {
87 2672414 : Some(_) => self.records.len(),
88 200 : None => self.records.len() - 1, // omit will_init record
89 : }
90 2672614 : }
91 : }
92 :
93 : #[derive(Clone, Copy, Debug, Default, Eq, PartialEq)]
94 : pub(crate) enum ValueReconstructSituation {
95 : Complete,
96 : #[default]
97 : Continue,
98 : }
99 :
100 : /// On disk representation of a value loaded in a buffer
101 : #[derive(Debug)]
102 : pub(crate) enum OnDiskValue {
103 : /// Unencoded [`Value::Image`]
104 : RawImage(Bytes),
105 : /// Encoded [`Value`]. Can deserialize into an image or a WAL record
106 : WalRecordOrImage(Bytes),
107 : }
108 :
109 : /// Reconstruct data accumulated for a single key during a vectored get
110 : #[derive(Debug, Default)]
111 : pub(crate) struct VectoredValueReconstructState {
112 : pub(crate) on_disk_values: Vec<(Lsn, OnDiskValueIoWaiter)>,
113 :
114 : pub(crate) situation: ValueReconstructSituation,
115 : }
116 :
117 : #[derive(Debug)]
118 : pub(crate) struct OnDiskValueIoWaiter {
119 : rx: tokio::sync::oneshot::Receiver<OnDiskValueIoResult>,
120 : }
121 :
122 : #[derive(Debug)]
123 : #[must_use]
124 : pub(crate) enum OnDiskValueIo {
125 : /// Traversal identified this IO as required to complete the vectored get.
126 : Required {
127 : num_active_ios: Arc<AtomicUsize>,
128 : tx: tokio::sync::oneshot::Sender<OnDiskValueIoResult>,
129 : },
130 : /// Sparse keyspace reads always read all the values for a given key,
131 : /// even though only the first value is needed.
132 : ///
133 : /// This variant represents the unnecessary IOs for those values at lower LSNs
134 : /// that aren't needed, but are currently still being done.
135 : ///
136 : /// The execution of unnecessary IOs was a pre-existing behavior before concurrent IO.
137 : /// We added this explicit representation here so that we can drop
138 : /// unnecessary IO results immediately, instead of buffering them in
139 : /// `oneshot` channels inside [`VectoredValueReconstructState`] until
140 : /// [`VectoredValueReconstructState::collect_pending_ios`] gets called.
141 : Unnecessary,
142 : }
143 :
144 : type OnDiskValueIoResult = Result<OnDiskValue, std::io::Error>;
145 :
146 : impl OnDiskValueIo {
147 1482151 : pub(crate) fn complete(self, res: OnDiskValueIoResult) {
148 1482151 : match self {
149 1337841 : OnDiskValueIo::Required { num_active_ios, tx } => {
150 1337841 : num_active_ios.fetch_sub(1, std::sync::atomic::Ordering::Release);
151 1337841 : let _ = tx.send(res);
152 1337841 : }
153 144310 : OnDiskValueIo::Unnecessary => {
154 144310 : // Nobody cared, see variant doc comment.
155 144310 : }
156 : }
157 1482151 : }
158 : }
159 :
160 : #[derive(Debug, thiserror::Error)]
161 : pub(crate) enum WaitCompletionError {
162 : #[error("OnDiskValueIo was dropped without completing, likely the sidecar task panicked")]
163 : IoDropped,
164 : }
165 :
166 : impl OnDiskValueIoWaiter {
167 1337839 : pub(crate) async fn wait_completion(self) -> Result<OnDiskValueIoResult, WaitCompletionError> {
168 1337839 : // NB: for Unnecessary IOs, this method never gets called because we don't add them to `on_disk_values`.
169 1337839 : self.rx.await.map_err(|_| WaitCompletionError::IoDropped)
170 1337839 : }
171 : }
172 :
173 : impl VectoredValueReconstructState {
174 : /// # Cancel-Safety
175 : ///
176 : /// Technically fine to stop polling this future, but, the IOs will still
177 : /// be executed to completion by the sidecar task and hold on to / consume resources.
178 : /// Better not do it to make reasonsing about the system easier.
179 1336447 : pub(crate) async fn collect_pending_ios(
180 1336447 : self,
181 1336447 : ) -> Result<ValueReconstructState, PageReconstructError> {
182 : use utils::bin_ser::BeSer;
183 :
184 1336447 : let mut res = Ok(ValueReconstructState::default());
185 :
186 : // We should try hard not to bail early, so that by the time we return from this
187 : // function, all IO for this value is done. It's not required -- we could totally
188 : // stop polling the IO futures in the sidecar task, they need to support that,
189 : // but just stopping to poll doesn't reduce the IO load on the disk. It's easier
190 : // to reason about the system if we just wait for all IO to complete, even if
191 : // we're no longer interested in the result.
192 : //
193 : // Revisit this when IO futures are replaced with a more sophisticated IO system
194 : // and an IO scheduler, where we know which IOs were submitted and which ones
195 : // just queued. Cf the comment on IoConcurrency::spawn_io.
196 2674286 : for (lsn, waiter) in self.on_disk_values {
197 1337839 : let value_recv_res = waiter
198 1337839 : .wait_completion()
199 1337839 : // we rely on the caller to poll us to completion, so this is not a bail point
200 1337839 : .await;
201 : // Force not bailing early by wrapping the code into a closure.
202 : #[allow(clippy::redundant_closure_call)]
203 1337839 : let _: () = (|| {
204 1337839 : match (&mut res, value_recv_res) {
205 0 : (Err(_), _) => {
206 0 : // We've already failed, no need to process more.
207 0 : }
208 0 : (Ok(_), Err(wait_err)) => {
209 0 : // This shouldn't happen - likely the sidecar task panicked.
210 0 : res = Err(PageReconstructError::Other(wait_err.into()));
211 0 : }
212 0 : (Ok(_), Ok(Err(err))) => {
213 0 : let err: std::io::Error = err;
214 0 : // TODO: returning IO error here will fail a compute query.
215 0 : // Probably not what we want, we're not doing `maybe_fatal_err`
216 0 : // in the IO futures.
217 0 : // But it's been like that for a long time, not changing it
218 0 : // as part of concurrent IO.
219 0 : // => https://github.com/neondatabase/neon/issues/10454
220 0 : res = Err(PageReconstructError::Other(err.into()));
221 0 : }
222 38401 : (Ok(ok), Ok(Ok(OnDiskValue::RawImage(img)))) => {
223 38401 : assert!(ok.img.is_none());
224 38401 : ok.img = Some((lsn, img));
225 : }
226 1299438 : (Ok(ok), Ok(Ok(OnDiskValue::WalRecordOrImage(buf)))) => {
227 1299438 : match Value::des(&buf) {
228 1492 : Ok(Value::WalRecord(rec)) => {
229 1492 : ok.records.push((lsn, rec));
230 1492 : }
231 1297946 : Ok(Value::Image(img)) => {
232 1297946 : assert!(ok.img.is_none());
233 1297946 : ok.img = Some((lsn, img));
234 : }
235 0 : Err(err) => {
236 0 : res = Err(PageReconstructError::Other(err.into()));
237 0 : }
238 : }
239 : }
240 : }
241 1337839 : })();
242 1337839 : }
243 :
244 1336447 : res
245 1336447 : }
246 : }
247 :
248 : /// Bag of data accumulated during a vectored get..
249 : pub(crate) struct ValuesReconstructState {
250 : /// The keys will be removed after `get_vectored` completes. The caller outside `Timeline`
251 : /// should not expect to get anything from this hashmap.
252 : pub(crate) keys: HashMap<Key, VectoredValueReconstructState>,
253 : /// The keys which are already retrieved
254 : keys_done: KeySpaceRandomAccum,
255 :
256 : /// The keys covered by the image layers
257 : keys_with_image_coverage: Option<Range<Key>>,
258 :
259 : // Statistics that are still accessible as a caller of `get_vectored_impl`.
260 : layers_visited: u32,
261 : delta_layers_visited: u32,
262 :
263 : pub(crate) io_concurrency: IoConcurrency,
264 : num_active_ios: Arc<AtomicUsize>,
265 :
266 : pub(crate) read_path: Option<ReadPath>,
267 : }
268 :
269 : /// The level of IO concurrency to be used on the read path
270 : ///
271 : /// The desired end state is that we always do parallel IO.
272 : /// This struct and the dispatching in the impl will be removed once
273 : /// we've built enough confidence.
274 : pub(crate) enum IoConcurrency {
275 : Sequential,
276 : SidecarTask {
277 : task_id: usize,
278 : ios_tx: tokio::sync::mpsc::UnboundedSender<IoFuture>,
279 : },
280 : }
281 :
282 : type IoFuture = Pin<Box<dyn Send + Future<Output = ()>>>;
283 :
284 : pub(crate) enum SelectedIoConcurrency {
285 : Sequential,
286 : SidecarTask(GateGuard),
287 : }
288 :
289 : impl std::fmt::Debug for IoConcurrency {
290 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
291 0 : match self {
292 0 : IoConcurrency::Sequential => write!(f, "Sequential"),
293 0 : IoConcurrency::SidecarTask { .. } => write!(f, "SidecarTask"),
294 : }
295 0 : }
296 : }
297 :
298 : impl std::fmt::Debug for SelectedIoConcurrency {
299 64 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
300 64 : match self {
301 32 : SelectedIoConcurrency::Sequential => write!(f, "Sequential"),
302 32 : SelectedIoConcurrency::SidecarTask(_) => write!(f, "SidecarTask"),
303 : }
304 64 : }
305 : }
306 :
307 : impl IoConcurrency {
308 : /// Force sequential IO. This is a temporary workaround until we have
309 : /// moved plumbing-through-the-call-stack
310 : /// of IoConcurrency into `RequestContextq.
311 : ///
312 : /// DO NOT USE for new code.
313 : ///
314 : /// Tracking issue: <https://github.com/neondatabase/neon/issues/10460>.
315 1215411 : pub(crate) fn sequential() -> Self {
316 1215411 : Self::spawn(SelectedIoConcurrency::Sequential)
317 1215411 : }
318 :
319 1052 : pub(crate) fn spawn_from_conf(
320 1052 : conf: &'static PageServerConf,
321 1052 : gate_guard: GateGuard,
322 1052 : ) -> IoConcurrency {
323 : use pageserver_api::config::GetVectoredConcurrentIo;
324 1052 : let selected = match conf.get_vectored_concurrent_io {
325 1052 : GetVectoredConcurrentIo::Sequential => SelectedIoConcurrency::Sequential,
326 0 : GetVectoredConcurrentIo::SidecarTask => SelectedIoConcurrency::SidecarTask(gate_guard),
327 : };
328 1052 : Self::spawn(selected)
329 1052 : }
330 :
331 1216527 : pub(crate) fn spawn(io_concurrency: SelectedIoConcurrency) -> Self {
332 1216527 : match io_concurrency {
333 1216495 : SelectedIoConcurrency::Sequential => IoConcurrency::Sequential,
334 32 : SelectedIoConcurrency::SidecarTask(gate_guard) => {
335 32 : let (ios_tx, ios_rx) = tokio::sync::mpsc::unbounded_channel();
336 : static TASK_ID: AtomicUsize = AtomicUsize::new(0);
337 32 : let task_id = TASK_ID.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
338 : // TODO: enrich the span with more context (tenant,shard,timeline) + (basebackup|pagestream|...)
339 32 : let span =
340 32 : tracing::info_span!(parent: None, "IoConcurrency_sidecar", task_id = task_id);
341 32 : trace!(task_id, "spawning sidecar task");
342 32 : tokio::spawn(async move {
343 32 : trace!("start");
344 32 : scopeguard::defer!{ trace!("end") };
345 : type IosRx = tokio::sync::mpsc::UnboundedReceiver<IoFuture>;
346 : enum State {
347 : Waiting {
348 : // invariant: is_empty(), but we recycle the allocation
349 : empty_futures: FuturesUnordered<IoFuture>,
350 : ios_rx: IosRx,
351 : },
352 : Executing {
353 : futures: FuturesUnordered<IoFuture>,
354 : ios_rx: IosRx,
355 : },
356 : ShuttingDown {
357 : futures: FuturesUnordered<IoFuture>,
358 : },
359 : }
360 32 : let mut state = State::Waiting {
361 32 : empty_futures: FuturesUnordered::new(),
362 32 : ios_rx,
363 32 : };
364 : loop {
365 39088 : match state {
366 : State::Waiting {
367 18554 : empty_futures,
368 18554 : mut ios_rx,
369 18554 : } => {
370 18554 : assert!(empty_futures.is_empty());
371 18554 : tokio::select! {
372 18554 : fut = ios_rx.recv() => {
373 18522 : if let Some(fut) = fut {
374 18522 : trace!("received new io future");
375 18522 : empty_futures.push(fut);
376 18522 : state = State::Executing { futures: empty_futures, ios_rx };
377 : } else {
378 0 : state = State::ShuttingDown { futures: empty_futures }
379 : }
380 : }
381 : }
382 : }
383 : State::Executing {
384 20534 : mut futures,
385 20534 : mut ios_rx,
386 20534 : } => {
387 20534 : tokio::select! {
388 20534 : res = futures.next() => {
389 19528 : trace!("io future completed");
390 19528 : assert!(res.is_some());
391 19528 : if futures.is_empty() {
392 18522 : state = State::Waiting { empty_futures: futures, ios_rx};
393 18522 : } else {
394 1006 : state = State::Executing { futures, ios_rx };
395 1006 : }
396 : }
397 20534 : fut = ios_rx.recv() => {
398 1006 : if let Some(fut) = fut {
399 1006 : trace!("received new io future");
400 1006 : futures.push(fut);
401 1006 : state = State::Executing { futures, ios_rx};
402 0 : } else {
403 0 : state = State::ShuttingDown { futures };
404 0 : }
405 : }
406 : }
407 : }
408 : State::ShuttingDown {
409 0 : mut futures,
410 0 : } => {
411 0 : trace!("shutting down");
412 0 : while let Some(()) = futures.next().await {
413 0 : trace!("io future completed (shutdown)");
414 : // drain
415 : }
416 0 : trace!("shutdown complete");
417 0 : break;
418 0 : }
419 0 : }
420 0 : }
421 0 : drop(gate_guard); // drop it right before we exit
422 32 : }.instrument(span));
423 32 : IoConcurrency::SidecarTask { task_id, ios_tx }
424 : }
425 : }
426 1216527 : }
427 :
428 76458 : pub(crate) fn clone(&self) -> Self {
429 76458 : match self {
430 39580 : IoConcurrency::Sequential => IoConcurrency::Sequential,
431 36878 : IoConcurrency::SidecarTask { task_id, ios_tx } => IoConcurrency::SidecarTask {
432 36878 : task_id: *task_id,
433 36878 : ios_tx: ios_tx.clone(),
434 36878 : },
435 : }
436 76458 : }
437 :
438 : /// Submit an IO to be executed in the background. DEADLOCK RISK, read the full doc string.
439 : ///
440 : /// The IO is represented as an opaque future.
441 : /// IO completion must be handled inside the future, e.g., through a oneshot channel.
442 : ///
443 : /// The API seems simple but there are multiple **pitfalls** involving
444 : /// DEADLOCK RISK.
445 : ///
446 : /// First, there are no guarantees about the exexecution of the IO.
447 : /// It may be `await`ed in-place before this function returns.
448 : /// It may be polled partially by this task and handed off to another task to be finished.
449 : /// It may be polled and then dropped before returning ready.
450 : ///
451 : /// This means that submitted IOs must not be interedependent.
452 : /// Interdependence may be through shared limited resources, e.g.,
453 : /// - VirtualFile file descriptor cache slot acquisition
454 : /// - tokio-epoll-uring slot
455 : ///
456 : /// # Why current usage is safe from deadlocks
457 : ///
458 : /// Textbook condition for a deadlock is that _all_ of the following be given
459 : /// - Mutual exclusion
460 : /// - Hold and wait
461 : /// - No preemption
462 : /// - Circular wait
463 : ///
464 : /// The current usage is safe because:
465 : /// - Mutual exclusion: IO futures definitely use mutexes, no way around that for now
466 : /// - Hold and wait: IO futures currently hold two kinds of locks/resources while waiting
467 : /// for acquisition of other resources:
468 : /// - VirtualFile file descriptor cache slot tokio mutex
469 : /// - tokio-epoll-uring slot (uses tokio notify => wait queue, much like mutex)
470 : /// - No preemption: there's no taking-away of acquired locks/resources => given
471 : /// - Circular wait: this is the part of the condition that isn't met: all IO futures
472 : /// first acquire VirtualFile mutex, then tokio-epoll-uring slot.
473 : /// There is no IO future that acquires slot before VirtualFile.
474 : /// Hence there can be no circular waiting.
475 : /// Hence there cannot be a deadlock.
476 : ///
477 : /// This is a very fragile situation and must be revisited whenver any code called from
478 : /// inside the IO futures is changed.
479 : ///
480 : /// We will move away from opaque IO futures towards well-defined IOs at some point in
481 : /// the future when we have shipped this first version of concurrent IO to production
482 : /// and are ready to retire the Sequential mode which runs the futures in place.
483 : /// Right now, while brittle, the opaque IO approach allows us to ship the feature
484 : /// with minimal changes to the code and minimal changes to existing behavior in Sequential mode.
485 : ///
486 : /// Also read the comment in `collect_pending_ios`.
487 1525842 : pub(crate) async fn spawn_io<F>(&mut self, fut: F)
488 1525842 : where
489 1525842 : F: std::future::Future<Output = ()> + Send + 'static,
490 1525842 : {
491 1525842 : match self {
492 1506314 : IoConcurrency::Sequential => fut.await,
493 19528 : IoConcurrency::SidecarTask { ios_tx, .. } => {
494 19528 : let fut = Box::pin(fut);
495 19528 : // NB: experiments showed that doing an opportunistic poll of `fut` here was bad for throughput
496 19528 : // while insignificant for latency.
497 19528 : // It would make sense to revisit the tokio-epoll-uring API in the future such that we can try
498 19528 : // a submission here, but never poll the future. That way, io_uring can make proccess while
499 19528 : // the future sits in the ios_tx queue.
500 19528 : match ios_tx.send(fut) {
501 19528 : Ok(()) => {}
502 : Err(_) => {
503 0 : unreachable!("the io task must have exited, likely it panicked")
504 : }
505 : }
506 : }
507 : }
508 1525842 : }
509 :
510 : #[cfg(test)]
511 64 : pub(crate) fn spawn_for_test() -> impl std::ops::DerefMut<Target = Self> {
512 : use std::ops::{Deref, DerefMut};
513 :
514 : use tracing::info;
515 : use utils::sync::gate::Gate;
516 :
517 : // Spawn needs a Gate, give it one.
518 : struct Wrapper {
519 : inner: IoConcurrency,
520 : #[allow(dead_code)]
521 : gate: Box<Gate>,
522 : }
523 : impl Deref for Wrapper {
524 : type Target = IoConcurrency;
525 :
526 36974 : fn deref(&self) -> &Self::Target {
527 36974 : &self.inner
528 36974 : }
529 : }
530 : impl DerefMut for Wrapper {
531 0 : fn deref_mut(&mut self) -> &mut Self::Target {
532 0 : &mut self.inner
533 0 : }
534 : }
535 64 : let gate = Box::new(Gate::default());
536 :
537 : // The default behavior when running Rust unit tests without any further
538 : // flags is to use the new behavior.
539 : // The CI uses the following environment variable to unit test both old
540 : // and new behavior.
541 : // NB: the Python regression & perf tests take the `else` branch
542 : // below and have their own defaults management.
543 64 : let selected = {
544 : // The pageserver_api::config type is unsuitable because it's internally tagged.
545 64 : #[derive(serde::Deserialize)]
546 : #[serde(rename_all = "kebab-case")]
547 : enum TestOverride {
548 : Sequential,
549 : SidecarTask,
550 : }
551 : use once_cell::sync::Lazy;
552 64 : static TEST_OVERRIDE: Lazy<TestOverride> = Lazy::new(|| {
553 64 : utils::env::var_serde_json_string(
554 64 : "NEON_PAGESERVER_UNIT_TEST_GET_VECTORED_CONCURRENT_IO",
555 64 : )
556 64 : .unwrap_or(TestOverride::SidecarTask)
557 64 : });
558 :
559 64 : match *TEST_OVERRIDE {
560 32 : TestOverride::Sequential => SelectedIoConcurrency::Sequential,
561 : TestOverride::SidecarTask => {
562 32 : SelectedIoConcurrency::SidecarTask(gate.enter().expect("just created it"))
563 : }
564 : }
565 : };
566 :
567 64 : info!(?selected, "get_vectored_concurrent_io test");
568 :
569 64 : Wrapper {
570 64 : inner: Self::spawn(selected),
571 64 : gate,
572 64 : }
573 64 : }
574 : }
575 :
576 : /// Make noise in case the [`ValuesReconstructState`] gets dropped while
577 : /// there are still IOs in flight.
578 : /// Refer to `collect_pending_ios` for why we prefer not to do that.
579 : //
580 : /// We log from here instead of from the sidecar task because the [`ValuesReconstructState`]
581 : /// gets dropped in a tracing span with more context.
582 : /// We repeat the sidecar tasks's `task_id` so we can correlate what we emit here with
583 : /// the logs / panic handler logs from the sidecar task, which also logs the `task_id`.
584 : impl Drop for ValuesReconstructState {
585 1255661 : fn drop(&mut self) {
586 1255661 : let num_active_ios = self
587 1255661 : .num_active_ios
588 1255661 : .load(std::sync::atomic::Ordering::Acquire);
589 1255661 : if num_active_ios == 0 {
590 1255659 : return;
591 2 : }
592 2 : let sidecar_task_id = match &self.io_concurrency {
593 0 : IoConcurrency::Sequential => None,
594 2 : IoConcurrency::SidecarTask { task_id, .. } => Some(*task_id),
595 : };
596 2 : tracing::warn!(
597 : num_active_ios,
598 : ?sidecar_task_id,
599 0 : backtrace=%std::backtrace::Backtrace::force_capture(),
600 0 : "dropping ValuesReconstructState while some IOs have not been completed",
601 : );
602 1255661 : }
603 : }
604 :
605 : impl ValuesReconstructState {
606 1255661 : pub(crate) fn new(io_concurrency: IoConcurrency) -> Self {
607 1255661 : Self {
608 1255661 : keys: HashMap::new(),
609 1255661 : keys_done: KeySpaceRandomAccum::new(),
610 1255661 : keys_with_image_coverage: None,
611 1255661 : layers_visited: 0,
612 1255661 : delta_layers_visited: 0,
613 1255661 : io_concurrency,
614 1255661 : num_active_ios: Arc::new(AtomicUsize::new(0)),
615 1255661 : read_path: None,
616 1255661 : }
617 1255661 : }
618 :
619 : /// Absolutely read [`IoConcurrency::spawn_io`] to learn about assumptions & pitfalls.
620 1525842 : pub(crate) async fn spawn_io<F>(&mut self, fut: F)
621 1525842 : where
622 1525842 : F: std::future::Future<Output = ()> + Send + 'static,
623 1525842 : {
624 1525842 : self.io_concurrency.spawn_io(fut).await;
625 1525842 : }
626 :
627 1693846 : pub(crate) fn on_layer_visited(&mut self, layer: &ReadableLayer) {
628 1693846 : self.layers_visited += 1;
629 1693846 : if let ReadableLayer::PersistentLayer(layer) = layer {
630 480220 : if layer.layer_desc().is_delta() {
631 435092 : self.delta_layers_visited += 1;
632 435092 : }
633 1213626 : }
634 1693846 : }
635 :
636 476 : pub(crate) fn get_delta_layers_visited(&self) -> u32 {
637 476 : self.delta_layers_visited
638 476 : }
639 :
640 1255603 : pub(crate) fn get_layers_visited(&self) -> u32 {
641 1255603 : self.layers_visited
642 1255603 : }
643 :
644 : /// On hitting image layer, we can mark all keys in this range as done, because
645 : /// if the image layer does not contain a key, it is deleted/never added.
646 45152 : pub(crate) fn on_image_layer_visited(&mut self, key_range: &Range<Key>) {
647 45152 : let prev_val = self.keys_with_image_coverage.replace(key_range.clone());
648 45152 : assert_eq!(
649 : prev_val, None,
650 0 : "should consume the keyspace before the next iteration"
651 : );
652 45152 : }
653 :
654 : /// Update the state collected for a given key.
655 : /// Returns true if this was the last value needed for the key and false otherwise.
656 : ///
657 : /// If the key is done after the update, mark it as such.
658 : ///
659 : /// If the key is in the sparse keyspace (i.e., aux files), we do not track them in
660 : /// `key_done`.
661 : // TODO: rename this method & update description.
662 1482151 : pub(crate) fn update_key(&mut self, key: &Key, lsn: Lsn, completes: bool) -> OnDiskValueIo {
663 1482151 : let state = self.keys.entry(*key).or_default();
664 1482151 :
665 1482151 : let is_sparse_key = key.is_sparse();
666 :
667 1482151 : let required_io = match state.situation {
668 : ValueReconstructSituation::Complete => {
669 144310 : if is_sparse_key {
670 : // Sparse keyspace might be visited multiple times because
671 : // we don't track unmapped keyspaces.
672 144310 : return OnDiskValueIo::Unnecessary;
673 : } else {
674 0 : unreachable!()
675 : }
676 : }
677 : ValueReconstructSituation::Continue => {
678 1337841 : self.num_active_ios
679 1337841 : .fetch_add(1, std::sync::atomic::Ordering::Release);
680 1337841 : let (tx, rx) = tokio::sync::oneshot::channel();
681 1337841 : state.on_disk_values.push((lsn, OnDiskValueIoWaiter { rx }));
682 1337841 : OnDiskValueIo::Required {
683 1337841 : tx,
684 1337841 : num_active_ios: Arc::clone(&self.num_active_ios),
685 1337841 : }
686 1337841 : }
687 1337841 : };
688 1337841 :
689 1337841 : if completes && state.situation == ValueReconstructSituation::Continue {
690 1336449 : state.situation = ValueReconstructSituation::Complete;
691 1336449 : if !is_sparse_key {
692 1208925 : self.keys_done.add_key(*key);
693 1208925 : }
694 1392 : }
695 :
696 1337841 : required_io
697 1482151 : }
698 :
699 : /// Returns the key space describing the keys that have
700 : /// been marked as completed since the last call to this function.
701 : /// Returns individual keys done, and the image layer coverage.
702 1693846 : pub(crate) fn consume_done_keys(&mut self) -> (KeySpace, Option<Range<Key>>) {
703 1693846 : (
704 1693846 : self.keys_done.consume_keyspace(),
705 1693846 : self.keys_with_image_coverage.take(),
706 1693846 : )
707 1693846 : }
708 : }
709 :
710 : /// A key that uniquely identifies a layer in a timeline
711 : #[derive(Debug, PartialEq, Eq, Clone, Hash)]
712 : pub(crate) enum LayerId {
713 : PersitentLayerId(PersistentLayerKey),
714 : InMemoryLayerId(InMemoryLayerFileId),
715 : }
716 :
717 : /// Uniquely identify a layer visit by the layer
718 : /// and LSN floor (or start LSN) of the reads.
719 : /// The layer itself is not enough since we may
720 : /// have different LSN lower bounds for delta layer reads.
721 : #[derive(Debug, PartialEq, Eq, Clone, Hash)]
722 : struct LayerToVisitId {
723 : layer_id: LayerId,
724 : lsn_floor: Lsn,
725 : }
726 :
727 : #[derive(Debug, PartialEq, Eq, Hash)]
728 : pub enum ReadableLayerWeak {
729 : PersistentLayer(Arc<PersistentLayerDesc>),
730 : InMemoryLayer(InMemoryLayerDesc),
731 : }
732 :
733 : /// Layer wrapper for the read path. Note that it is valid
734 : /// to use these layers even after external operations have
735 : /// been performed on them (compaction, freeze, etc.).
736 : #[derive(Debug)]
737 : pub(crate) enum ReadableLayer {
738 : PersistentLayer(Layer),
739 : InMemoryLayer(Arc<InMemoryLayer>),
740 : }
741 :
742 : /// A partial description of a read to be done.
743 : #[derive(Debug, Clone)]
744 : struct LayerVisit {
745 : /// An id used to resolve the readable layer within the fringe
746 : layer_to_visit_id: LayerToVisitId,
747 : /// Lsn range for the read, used for selecting the next read
748 : lsn_range: Range<Lsn>,
749 : }
750 :
751 : /// Data structure which maintains a fringe of layers for the
752 : /// read path. The fringe is the set of layers which intersects
753 : /// the current keyspace that the search is descending on.
754 : /// Each layer tracks the keyspace that intersects it.
755 : ///
756 : /// The fringe must appear sorted by Lsn. Hence, it uses
757 : /// a two layer indexing scheme.
758 : #[derive(Debug)]
759 : pub(crate) struct LayerFringe {
760 : planned_visits_by_lsn: BinaryHeap<LayerVisit>,
761 : visit_reads: HashMap<LayerToVisitId, LayerVisitReads>,
762 : }
763 :
764 : #[derive(Debug)]
765 : struct LayerVisitReads {
766 : layer: ReadableLayer,
767 : target_keyspace: KeySpaceRandomAccum,
768 : }
769 :
770 : impl LayerFringe {
771 1709813 : pub(crate) fn new() -> Self {
772 1709813 : LayerFringe {
773 1709813 : planned_visits_by_lsn: BinaryHeap::new(),
774 1709813 : visit_reads: HashMap::new(),
775 1709813 : }
776 1709813 : }
777 :
778 3403659 : pub(crate) fn next_layer(&mut self) -> Option<(ReadableLayer, KeySpace, Range<Lsn>)> {
779 3403659 : let read_desc = self.planned_visits_by_lsn.pop()?;
780 :
781 1693846 : let removed = self.visit_reads.remove_entry(&read_desc.layer_to_visit_id);
782 1693846 :
783 1693846 : match removed {
784 : Some((
785 : _,
786 : LayerVisitReads {
787 1693846 : layer,
788 1693846 : mut target_keyspace,
789 1693846 : },
790 1693846 : )) => Some((
791 1693846 : layer,
792 1693846 : target_keyspace.consume_keyspace(),
793 1693846 : read_desc.lsn_range,
794 1693846 : )),
795 0 : None => unreachable!("fringe internals are always consistent"),
796 : }
797 3403659 : }
798 :
799 1694798 : pub(crate) fn update(
800 1694798 : &mut self,
801 1694798 : layer: ReadableLayer,
802 1694798 : keyspace: KeySpace,
803 1694798 : lsn_range: Range<Lsn>,
804 1694798 : ) {
805 1694798 : let layer_to_visit_id = LayerToVisitId {
806 1694798 : layer_id: layer.id(),
807 1694798 : lsn_floor: lsn_range.start,
808 1694798 : };
809 1694798 :
810 1694798 : let entry = self.visit_reads.entry(layer_to_visit_id.clone());
811 1694798 : match entry {
812 952 : Entry::Occupied(mut entry) => {
813 952 : entry.get_mut().target_keyspace.add_keyspace(keyspace);
814 952 : }
815 1693846 : Entry::Vacant(entry) => {
816 1693846 : self.planned_visits_by_lsn.push(LayerVisit {
817 1693846 : lsn_range,
818 1693846 : layer_to_visit_id: layer_to_visit_id.clone(),
819 1693846 : });
820 1693846 : let mut accum = KeySpaceRandomAccum::new();
821 1693846 : accum.add_keyspace(keyspace);
822 1693846 : entry.insert(LayerVisitReads {
823 1693846 : layer,
824 1693846 : target_keyspace: accum,
825 1693846 : });
826 1693846 : }
827 : }
828 1694798 : }
829 : }
830 :
831 : impl Default for LayerFringe {
832 0 : fn default() -> Self {
833 0 : Self::new()
834 0 : }
835 : }
836 :
837 : impl Ord for LayerVisit {
838 72 : fn cmp(&self, other: &Self) -> Ordering {
839 72 : let ord = self.lsn_range.end.cmp(&other.lsn_range.end);
840 72 : if ord == std::cmp::Ordering::Equal {
841 48 : self.lsn_range.start.cmp(&other.lsn_range.start).reverse()
842 : } else {
843 24 : ord
844 : }
845 72 : }
846 : }
847 :
848 : impl PartialOrd for LayerVisit {
849 72 : fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
850 72 : Some(self.cmp(other))
851 72 : }
852 : }
853 :
854 : impl PartialEq for LayerVisit {
855 0 : fn eq(&self, other: &Self) -> bool {
856 0 : self.lsn_range == other.lsn_range
857 0 : }
858 : }
859 :
860 : impl Eq for LayerVisit {}
861 :
862 : impl ReadableLayer {
863 1694798 : pub(crate) fn id(&self) -> LayerId {
864 1694798 : match self {
865 480316 : Self::PersistentLayer(layer) => LayerId::PersitentLayerId(layer.layer_desc().key()),
866 1214482 : Self::InMemoryLayer(layer) => LayerId::InMemoryLayerId(layer.file_id()),
867 : }
868 1694798 : }
869 :
870 1693846 : pub(crate) async fn get_values_reconstruct_data(
871 1693846 : &self,
872 1693846 : keyspace: KeySpace,
873 1693846 : lsn_range: Range<Lsn>,
874 1693846 : reconstruct_state: &mut ValuesReconstructState,
875 1693846 : ctx: &RequestContext,
876 1693846 : ) -> Result<(), GetVectoredError> {
877 1693846 : match self {
878 480220 : ReadableLayer::PersistentLayer(layer) => {
879 480220 : let ctx = RequestContextBuilder::from(ctx)
880 480220 : .perf_span(|crnt_perf_span| {
881 0 : info_span!(
882 : target: PERF_TRACE_TARGET,
883 0 : parent: crnt_perf_span,
884 : "PLAN_LAYER",
885 : layer = %layer
886 : )
887 480220 : })
888 480220 : .attached_child();
889 480220 :
890 480220 : layer
891 480220 : .get_values_reconstruct_data(keyspace, lsn_range, reconstruct_state, &ctx)
892 480220 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
893 480220 : .await
894 : }
895 1213626 : ReadableLayer::InMemoryLayer(layer) => {
896 1213626 : let ctx = RequestContextBuilder::from(ctx)
897 1213626 : .perf_span(|crnt_perf_span| {
898 0 : info_span!(
899 : target: PERF_TRACE_TARGET,
900 0 : parent: crnt_perf_span,
901 : "PLAN_LAYER",
902 : layer = %layer
903 : )
904 1213626 : })
905 1213626 : .attached_child();
906 1213626 :
907 1213626 : layer
908 1213626 : .get_values_reconstruct_data(keyspace, lsn_range, reconstruct_state, &ctx)
909 1213626 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
910 1213626 : .await
911 : }
912 : }
913 1693846 : }
914 : }
915 :
916 : /// Layers contain a hint indicating whether they are likely to be used for reads.
917 : ///
918 : /// This is a hint rather than an authoritative value, so that we do not have to update it synchronously
919 : /// when changing the visibility of layers (for example when creating a branch that makes some previously
920 : /// covered layers visible). It should be used for cache management but not for correctness-critical checks.
921 : #[derive(Debug, Clone, PartialEq, Eq)]
922 : pub enum LayerVisibilityHint {
923 : /// A Visible layer might be read while serving a read, because there is not an image layer between it
924 : /// and a readable LSN (the tip of the branch or a child's branch point)
925 : Visible,
926 : /// A Covered layer probably won't be read right now, but _can_ be read in future if someone creates
927 : /// a branch or ephemeral endpoint at an LSN below the layer that covers this.
928 : Covered,
929 : }
930 :
931 : pub(crate) struct LayerAccessStats(std::sync::atomic::AtomicU64);
932 :
933 0 : #[derive(Clone, Copy, strum_macros::EnumString)]
934 : pub(crate) enum LayerAccessStatsReset {
935 : NoReset,
936 : AllStats,
937 : }
938 :
939 : impl Default for LayerAccessStats {
940 3880 : fn default() -> Self {
941 3880 : // Default value is to assume resident since creation time, and visible.
942 3880 : let (_mask, mut value) = Self::to_low_res_timestamp(Self::RTIME_SHIFT, SystemTime::now());
943 3880 : value |= 0x1 << Self::VISIBILITY_SHIFT;
944 3880 :
945 3880 : Self(std::sync::atomic::AtomicU64::new(value))
946 3880 : }
947 : }
948 :
949 : // Efficient store of two very-low-resolution timestamps and some bits. Used for storing last access time and
950 : // last residence change time.
951 : impl LayerAccessStats {
952 : // How many high bits to drop from a u32 timestamp?
953 : // - Only storing up to a u32 timestamp will work fine until 2038 (if this code is still in use
954 : // after that, this software has been very successful!)
955 : // - Dropping the top bit is implicitly safe because unix timestamps are meant to be
956 : // stored in an i32, so they never used it.
957 : // - Dropping the next two bits is safe because this code is only running on systems in
958 : // years >= 2024, and these bits have been 1 since 2021
959 : //
960 : // Therefore we may store only 28 bits for a timestamp with one second resolution. We do
961 : // this truncation to make space for some flags in the high bits of our u64.
962 : const TS_DROP_HIGH_BITS: u32 = u32::count_ones(Self::TS_ONES) + 1;
963 : const TS_MASK: u32 = 0x1f_ff_ff_ff;
964 : const TS_ONES: u32 = 0x60_00_00_00;
965 :
966 : const ATIME_SHIFT: u32 = 0;
967 : const RTIME_SHIFT: u32 = 32 - Self::TS_DROP_HIGH_BITS;
968 : const VISIBILITY_SHIFT: u32 = 64 - 2 * Self::TS_DROP_HIGH_BITS;
969 :
970 480536 : fn write_bits(&self, mask: u64, value: u64) -> u64 {
971 480536 : self.0
972 480536 : .fetch_update(
973 480536 : // TODO: decide what orderings are correct
974 480536 : std::sync::atomic::Ordering::Relaxed,
975 480536 : std::sync::atomic::Ordering::Relaxed,
976 480536 : |v| Some((v & !mask) | (value & mask)),
977 480536 : )
978 480536 : .expect("Inner function is infallible")
979 480536 : }
980 :
981 483664 : fn to_low_res_timestamp(shift: u32, time: SystemTime) -> (u64, u64) {
982 483664 : // Drop the low three bits of the timestamp, for an ~8s accuracy
983 483664 : let timestamp = time.duration_since(UNIX_EPOCH).unwrap().as_secs() & (Self::TS_MASK as u64);
984 483664 :
985 483664 : ((Self::TS_MASK as u64) << shift, timestamp << shift)
986 483664 : }
987 :
988 292 : fn read_low_res_timestamp(&self, shift: u32) -> Option<SystemTime> {
989 292 : let read = self.0.load(std::sync::atomic::Ordering::Relaxed);
990 292 :
991 292 : let ts_bits = (read & ((Self::TS_MASK as u64) << shift)) >> shift;
992 292 : if ts_bits == 0 {
993 132 : None
994 : } else {
995 160 : Some(UNIX_EPOCH + Duration::from_secs(ts_bits | (Self::TS_ONES as u64)))
996 : }
997 292 : }
998 :
999 : /// Record a change in layer residency.
1000 : ///
1001 : /// Recording the event must happen while holding the layer map lock to
1002 : /// ensure that latest-activity-threshold-based layer eviction (eviction_task.rs)
1003 : /// can do an "imitate access" to this layer, before it observes `now-latest_activity() > threshold`.
1004 : ///
1005 : /// If we instead recorded the residence event with a timestamp from before grabbing the layer map lock,
1006 : /// the following race could happen:
1007 : ///
1008 : /// - Compact: Write out an L1 layer from several L0 layers. This records residence event LayerCreate with the current timestamp.
1009 : /// - Eviction: imitate access logical size calculation. This accesses the L0 layers because the L1 layer is not yet in the layer map.
1010 : /// - Compact: Grab layer map lock, add the new L1 to layer map and remove the L0s, release layer map lock.
1011 : /// - Eviction: observes the new L1 layer whose only activity timestamp is the LayerCreate event.
1012 100 : pub(crate) fn record_residence_event_at(&self, now: SystemTime) {
1013 100 : let (mask, value) = Self::to_low_res_timestamp(Self::RTIME_SHIFT, now);
1014 100 : self.write_bits(mask, value);
1015 100 : }
1016 :
1017 96 : pub(crate) fn record_residence_event(&self) {
1018 96 : self.record_residence_event_at(SystemTime::now())
1019 96 : }
1020 :
1021 479684 : fn record_access_at(&self, now: SystemTime) -> bool {
1022 479684 : let (mut mask, mut value) = Self::to_low_res_timestamp(Self::ATIME_SHIFT, now);
1023 479684 :
1024 479684 : // A layer which is accessed must be visible.
1025 479684 : mask |= 0x1 << Self::VISIBILITY_SHIFT;
1026 479684 : value |= 0x1 << Self::VISIBILITY_SHIFT;
1027 479684 :
1028 479684 : let old_bits = self.write_bits(mask, value);
1029 4 : !matches!(
1030 479684 : self.decode_visibility(old_bits),
1031 : LayerVisibilityHint::Visible
1032 : )
1033 479684 : }
1034 :
1035 : /// Returns true if we modified the layer's visibility to set it to Visible implicitly
1036 : /// as a result of this access
1037 480244 : pub(crate) fn record_access(&self, ctx: &RequestContext) -> bool {
1038 480244 : if ctx.access_stats_behavior() == AccessStatsBehavior::Skip {
1039 572 : return false;
1040 479672 : }
1041 479672 :
1042 479672 : self.record_access_at(SystemTime::now())
1043 480244 : }
1044 :
1045 0 : fn as_api_model(
1046 0 : &self,
1047 0 : reset: LayerAccessStatsReset,
1048 0 : ) -> pageserver_api::models::LayerAccessStats {
1049 0 : let ret = pageserver_api::models::LayerAccessStats {
1050 0 : access_time: self
1051 0 : .read_low_res_timestamp(Self::ATIME_SHIFT)
1052 0 : .unwrap_or(UNIX_EPOCH),
1053 0 : residence_time: self
1054 0 : .read_low_res_timestamp(Self::RTIME_SHIFT)
1055 0 : .unwrap_or(UNIX_EPOCH),
1056 0 : visible: matches!(self.visibility(), LayerVisibilityHint::Visible),
1057 : };
1058 0 : match reset {
1059 0 : LayerAccessStatsReset::NoReset => {}
1060 0 : LayerAccessStatsReset::AllStats => {
1061 0 : self.write_bits((Self::TS_MASK as u64) << Self::ATIME_SHIFT, 0x0);
1062 0 : self.write_bits((Self::TS_MASK as u64) << Self::RTIME_SHIFT, 0x0);
1063 0 : }
1064 : }
1065 0 : ret
1066 0 : }
1067 :
1068 : /// Get the latest access timestamp, falling back to latest residence event. The latest residence event
1069 : /// will be this Layer's construction time, if its residence hasn't changed since then.
1070 84 : pub(crate) fn latest_activity(&self) -> SystemTime {
1071 84 : if let Some(t) = self.read_low_res_timestamp(Self::ATIME_SHIFT) {
1072 12 : t
1073 : } else {
1074 72 : self.read_low_res_timestamp(Self::RTIME_SHIFT)
1075 72 : .expect("Residence time is set on construction")
1076 : }
1077 84 : }
1078 :
1079 : /// Whether this layer has been accessed (excluding in [`AccessStatsBehavior::Skip`]).
1080 : ///
1081 : /// This indicates whether the layer has been used for some purpose that would motivate
1082 : /// us to keep it on disk, such as for serving a getpage request.
1083 68 : fn accessed(&self) -> bool {
1084 68 : // Consider it accessed if the most recent access is more recent than
1085 68 : // the most recent change in residence status.
1086 68 : match (
1087 68 : self.read_low_res_timestamp(Self::ATIME_SHIFT),
1088 68 : self.read_low_res_timestamp(Self::RTIME_SHIFT),
1089 : ) {
1090 60 : (None, _) => false,
1091 0 : (Some(_), None) => true,
1092 8 : (Some(a), Some(r)) => a >= r,
1093 : }
1094 68 : }
1095 :
1096 : /// Helper for extracting the visibility hint from the literal value of our inner u64
1097 482025 : fn decode_visibility(&self, bits: u64) -> LayerVisibilityHint {
1098 482025 : match (bits >> Self::VISIBILITY_SHIFT) & 0x1 {
1099 481977 : 1 => LayerVisibilityHint::Visible,
1100 48 : 0 => LayerVisibilityHint::Covered,
1101 0 : _ => unreachable!(),
1102 : }
1103 482025 : }
1104 :
1105 : /// Returns the old value which has been replaced
1106 752 : pub(crate) fn set_visibility(&self, visibility: LayerVisibilityHint) -> LayerVisibilityHint {
1107 752 : let value = match visibility {
1108 648 : LayerVisibilityHint::Visible => 0x1 << Self::VISIBILITY_SHIFT,
1109 104 : LayerVisibilityHint::Covered => 0x0,
1110 : };
1111 :
1112 752 : let old_bits = self.write_bits(0x1 << Self::VISIBILITY_SHIFT, value);
1113 752 : self.decode_visibility(old_bits)
1114 752 : }
1115 :
1116 1589 : pub(crate) fn visibility(&self) -> LayerVisibilityHint {
1117 1589 : let read = self.0.load(std::sync::atomic::Ordering::Relaxed);
1118 1589 : self.decode_visibility(read)
1119 1589 : }
1120 : }
1121 :
1122 : /// Get a layer descriptor from a layer.
1123 : pub(crate) trait AsLayerDesc {
1124 : /// Get the layer descriptor.
1125 : fn layer_desc(&self) -> &PersistentLayerDesc;
1126 : }
1127 :
1128 : pub mod tests {
1129 : use pageserver_api::shard::TenantShardId;
1130 : use utils::id::TimelineId;
1131 :
1132 : use super::*;
1133 :
1134 : impl From<DeltaLayerName> for PersistentLayerDesc {
1135 44 : fn from(value: DeltaLayerName) -> Self {
1136 44 : PersistentLayerDesc::new_delta(
1137 44 : TenantShardId::from([0; 18]),
1138 44 : TimelineId::from_array([0; 16]),
1139 44 : value.key_range,
1140 44 : value.lsn_range,
1141 44 : 233,
1142 44 : )
1143 44 : }
1144 : }
1145 :
1146 : impl From<ImageLayerName> for PersistentLayerDesc {
1147 48 : fn from(value: ImageLayerName) -> Self {
1148 48 : PersistentLayerDesc::new_img(
1149 48 : TenantShardId::from([0; 18]),
1150 48 : TimelineId::from_array([0; 16]),
1151 48 : value.key_range,
1152 48 : value.lsn,
1153 48 : 233,
1154 48 : )
1155 48 : }
1156 : }
1157 :
1158 : impl From<LayerName> for PersistentLayerDesc {
1159 92 : fn from(value: LayerName) -> Self {
1160 92 : match value {
1161 44 : LayerName::Delta(d) => Self::from(d),
1162 48 : LayerName::Image(i) => Self::from(i),
1163 : }
1164 92 : }
1165 : }
1166 : }
1167 :
1168 : /// Range wrapping newtype, which uses display to render Debug.
1169 : ///
1170 : /// Useful with `Key`, which has too verbose `{:?}` for printing multiple layers.
1171 : struct RangeDisplayDebug<'a, T: std::fmt::Display>(&'a Range<T>);
1172 :
1173 : impl<T: std::fmt::Display> std::fmt::Debug for RangeDisplayDebug<'_, T> {
1174 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
1175 0 : write!(f, "{}..{}", self.0.start, self.0.end)
1176 0 : }
1177 : }
1178 :
1179 : #[cfg(test)]
1180 : mod tests2 {
1181 : use pageserver_api::key::DBDIR_KEY;
1182 : use tracing::info;
1183 :
1184 : use super::*;
1185 : use crate::tenant::storage_layer::IoConcurrency;
1186 :
1187 : /// TODO: currently this test relies on manual visual inspection of the --no-capture output.
1188 : /// Should look like so:
1189 : /// ```text
1190 : /// RUST_LOG=trace cargo nextest run --features testing --no-capture test_io_concurrency_noise
1191 : /// running 1 test
1192 : /// 2025-01-21T17:42:01.335679Z INFO get_vectored_concurrent_io test selected=SidecarTask
1193 : /// 2025-01-21T17:42:01.335680Z TRACE spawning sidecar task task_id=0
1194 : /// 2025-01-21T17:42:01.335937Z TRACE IoConcurrency_sidecar{task_id=0}: start
1195 : /// 2025-01-21T17:42:01.335972Z TRACE IoConcurrency_sidecar{task_id=0}: received new io future
1196 : /// 2025-01-21T17:42:01.335999Z INFO IoConcurrency_sidecar{task_id=0}: waiting for signal to complete IO
1197 : /// 2025-01-21T17:42:01.336229Z WARN dropping ValuesReconstructState while some IOs have not been completed num_active_ios=1 sidecar_task_id=Some(0) backtrace= 0: <pageserver::tenant::storage_layer::ValuesReconstructState as core::ops::drop::Drop>::drop
1198 : /// at ./src/tenant/storage_layer.rs:553:24
1199 : /// 1: core::ptr::drop_in_place<pageserver::tenant::storage_layer::ValuesReconstructState>
1200 : /// at /home/christian/.rustup/toolchains/1.84.0-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/library/core/src/ptr/mod.rs:521:1
1201 : /// 2: core::mem::drop
1202 : /// at /home/christian/.rustup/toolchains/1.84.0-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/library/core/src/mem/mod.rs:942:24
1203 : /// 3: pageserver::tenant::storage_layer::tests2::test_io_concurrency_noise::{{closure}}
1204 : /// at ./src/tenant/storage_layer.rs:1159:9
1205 : /// ...
1206 : /// 49: <unknown>
1207 : /// 2025-01-21T17:42:01.452293Z INFO IoConcurrency_sidecar{task_id=0}: completing IO
1208 : /// 2025-01-21T17:42:01.452357Z TRACE IoConcurrency_sidecar{task_id=0}: io future completed
1209 : /// 2025-01-21T17:42:01.452473Z TRACE IoConcurrency_sidecar{task_id=0}: end
1210 : /// test tenant::storage_layer::tests2::test_io_concurrency_noise ... ok
1211 : ///
1212 : /// ```
1213 : #[tokio::test]
1214 4 : async fn test_io_concurrency_noise() {
1215 4 : crate::tenant::harness::setup_logging();
1216 4 :
1217 4 : let io_concurrency = IoConcurrency::spawn_for_test();
1218 4 : match *io_concurrency {
1219 4 : IoConcurrency::Sequential => {
1220 4 : // This test asserts behavior in sidecar mode, doesn't make sense in sequential mode.
1221 4 : return;
1222 4 : }
1223 4 : IoConcurrency::SidecarTask { .. } => {}
1224 2 : }
1225 2 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
1226 2 :
1227 2 : let (io_fut_is_waiting_tx, io_fut_is_waiting) = tokio::sync::oneshot::channel();
1228 2 : let (do_complete_io, should_complete_io) = tokio::sync::oneshot::channel();
1229 2 : let (io_fut_exiting_tx, io_fut_exiting) = tokio::sync::oneshot::channel();
1230 2 :
1231 2 : let io = reconstruct_state.update_key(&DBDIR_KEY, Lsn(8), true);
1232 2 : reconstruct_state
1233 2 : .spawn_io(async move {
1234 2 : info!("waiting for signal to complete IO");
1235 4 : io_fut_is_waiting_tx.send(()).unwrap();
1236 2 : should_complete_io.await.unwrap();
1237 2 : info!("completing IO");
1238 4 : io.complete(Ok(OnDiskValue::RawImage(Bytes::new())));
1239 2 : io_fut_exiting_tx.send(()).unwrap();
1240 2 : })
1241 2 : .await;
1242 4 :
1243 4 : io_fut_is_waiting.await.unwrap();
1244 2 :
1245 2 : // this is what makes the noise
1246 2 : drop(reconstruct_state);
1247 2 :
1248 2 : do_complete_io.send(()).unwrap();
1249 2 :
1250 2 : io_fut_exiting.await.unwrap();
1251 4 : }
1252 : }
|
