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 crate::config::PageServerConf;
14 : use crate::context::{AccessStatsBehavior, RequestContext};
15 : use bytes::Bytes;
16 : use futures::stream::FuturesUnordered;
17 : use futures::StreamExt;
18 : use pageserver_api::key::Key;
19 : use pageserver_api::keyspace::{KeySpace, KeySpaceRandomAccum};
20 : use pageserver_api::record::NeonWalRecord;
21 : use pageserver_api::value::Value;
22 : use std::cmp::Ordering;
23 : use std::collections::hash_map::Entry;
24 : use std::collections::{BinaryHeap, HashMap};
25 : use std::future::Future;
26 : use std::ops::Range;
27 : use std::pin::Pin;
28 : use std::sync::atomic::AtomicUsize;
29 : use std::sync::Arc;
30 : use std::time::{Duration, SystemTime, UNIX_EPOCH};
31 : use tracing::{trace, Instrument};
32 : use utils::sync::gate::GateGuard;
33 :
34 : use utils::lsn::Lsn;
35 :
36 : pub use batch_split_writer::{BatchLayerWriter, SplitDeltaLayerWriter, SplitImageLayerWriter};
37 : pub use delta_layer::{DeltaLayer, DeltaLayerWriter, ValueRef};
38 : pub use image_layer::{ImageLayer, ImageLayerWriter};
39 : pub use inmemory_layer::InMemoryLayer;
40 : pub use layer_desc::{PersistentLayerDesc, PersistentLayerKey};
41 : pub use layer_name::{DeltaLayerName, ImageLayerName, LayerName};
42 :
43 : pub(crate) use layer::{EvictionError, Layer, ResidentLayer};
44 :
45 : use self::inmemory_layer::InMemoryLayerFileId;
46 :
47 : use super::timeline::GetVectoredError;
48 : use super::PageReconstructError;
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 : #[derive(Clone, Copy, Debug, Default, Eq, PartialEq)]
84 : pub(crate) enum ValueReconstructSituation {
85 : Complete,
86 : #[default]
87 : Continue,
88 : }
89 :
90 : /// On disk representation of a value loaded in a buffer
91 : #[derive(Debug)]
92 : pub(crate) enum OnDiskValue {
93 : /// Unencoded [`Value::Image`]
94 : RawImage(Bytes),
95 : /// Encoded [`Value`]. Can deserialize into an image or a WAL record
96 : WalRecordOrImage(Bytes),
97 : }
98 :
99 : /// Reconstruct data accumulated for a single key during a vectored get
100 : #[derive(Debug, Default)]
101 : pub(crate) struct VectoredValueReconstructState {
102 : pub(crate) on_disk_values: Vec<(Lsn, OnDiskValueIoWaiter)>,
103 :
104 : pub(crate) situation: ValueReconstructSituation,
105 : }
106 :
107 : #[derive(Debug)]
108 : pub(crate) struct OnDiskValueIoWaiter {
109 : rx: tokio::sync::oneshot::Receiver<OnDiskValueIoResult>,
110 : }
111 :
112 : #[derive(Debug)]
113 : #[must_use]
114 : pub(crate) enum OnDiskValueIo {
115 : /// Traversal identified this IO as required to complete the vectored get.
116 : Required {
117 : num_active_ios: Arc<AtomicUsize>,
118 : tx: tokio::sync::oneshot::Sender<OnDiskValueIoResult>,
119 : },
120 : /// Sparse keyspace reads always read all the values for a given key,
121 : /// even though only the first value is needed.
122 : ///
123 : /// This variant represents the unnecessary IOs for those values at lower LSNs
124 : /// that aren't needed, but are currently still being done.
125 : ///
126 : /// The execution of unnecessary IOs was a pre-existing behavior before concurrent IO.
127 : /// We added this explicit representation here so that we can drop
128 : /// unnecessary IO results immediately, instead of buffering them in
129 : /// `oneshot` channels inside [`VectoredValueReconstructState`] until
130 : /// [`VectoredValueReconstructState::collect_pending_ios`] gets called.
131 : Unnecessary,
132 : }
133 :
134 : type OnDiskValueIoResult = Result<OnDiskValue, std::io::Error>;
135 :
136 : impl OnDiskValueIo {
137 1482755 : pub(crate) fn complete(self, res: OnDiskValueIoResult) {
138 1482755 : match self {
139 1337380 : OnDiskValueIo::Required { num_active_ios, tx } => {
140 1337380 : num_active_ios.fetch_sub(1, std::sync::atomic::Ordering::Release);
141 1337380 : let _ = tx.send(res);
142 1337380 : }
143 145375 : OnDiskValueIo::Unnecessary => {
144 145375 : // Nobody cared, see variant doc comment.
145 145375 : }
146 : }
147 1482755 : }
148 : }
149 :
150 : #[derive(Debug, thiserror::Error)]
151 : pub(crate) enum WaitCompletionError {
152 : #[error("OnDiskValueIo was dropped without completing, likely the sidecar task panicked")]
153 : IoDropped,
154 : }
155 :
156 : impl OnDiskValueIoWaiter {
157 1337378 : pub(crate) async fn wait_completion(self) -> Result<OnDiskValueIoResult, WaitCompletionError> {
158 1337378 : // NB: for Unnecessary IOs, this method never gets called because we don't add them to `on_disk_values`.
159 1337378 : self.rx.await.map_err(|_| WaitCompletionError::IoDropped)
160 1337378 : }
161 : }
162 :
163 : impl VectoredValueReconstructState {
164 : /// # Cancel-Safety
165 : ///
166 : /// Technically fine to stop polling this future, but, the IOs will still
167 : /// be executed to completion by the sidecar task and hold on to / consume resources.
168 : /// Better not do it to make reasonsing about the system easier.
169 1336066 : pub(crate) async fn collect_pending_ios(
170 1336066 : self,
171 1336066 : ) -> Result<ValueReconstructState, PageReconstructError> {
172 : use utils::bin_ser::BeSer;
173 :
174 1336066 : let mut res = Ok(ValueReconstructState::default());
175 :
176 : // We should try hard not to bail early, so that by the time we return from this
177 : // function, all IO for this value is done. It's not required -- we could totally
178 : // stop polling the IO futures in the sidecar task, they need to support that,
179 : // but just stopping to poll doesn't reduce the IO load on the disk. It's easier
180 : // to reason about the system if we just wait for all IO to complete, even if
181 : // we're no longer interested in the result.
182 : //
183 : // Revisit this when IO futures are replaced with a more sophisticated IO system
184 : // and an IO scheduler, where we know which IOs were submitted and which ones
185 : // just queued. Cf the comment on IoConcurrency::spawn_io.
186 2673444 : for (lsn, waiter) in self.on_disk_values {
187 1337378 : let value_recv_res = waiter
188 1337378 : .wait_completion()
189 1337378 : // we rely on the caller to poll us to completion, so this is not a bail point
190 1337378 : .await;
191 : // Force not bailing early by wrapping the code into a closure.
192 : #[allow(clippy::redundant_closure_call)]
193 1337378 : let _: () = (|| {
194 1337378 : match (&mut res, value_recv_res) {
195 0 : (Err(_), _) => {
196 0 : // We've already failed, no need to process more.
197 0 : }
198 0 : (Ok(_), Err(wait_err)) => {
199 0 : // This shouldn't happen - likely the sidecar task panicked.
200 0 : res = Err(PageReconstructError::Other(wait_err.into()));
201 0 : }
202 0 : (Ok(_), Ok(Err(err))) => {
203 0 : let err: std::io::Error = err;
204 0 : // TODO: returning IO error here will fail a compute query.
205 0 : // Probably not what we want, we're not doing `maybe_fatal_err`
206 0 : // in the IO futures.
207 0 : // But it's been like that for a long time, not changing it
208 0 : // as part of concurrent IO.
209 0 : // => https://github.com/neondatabase/neon/issues/10454
210 0 : res = Err(PageReconstructError::Other(err.into()));
211 0 : }
212 38198 : (Ok(ok), Ok(Ok(OnDiskValue::RawImage(img)))) => {
213 38198 : assert!(ok.img.is_none());
214 38198 : ok.img = Some((lsn, img));
215 : }
216 1299180 : (Ok(ok), Ok(Ok(OnDiskValue::WalRecordOrImage(buf)))) => {
217 1299180 : match Value::des(&buf) {
218 1400 : Ok(Value::WalRecord(rec)) => {
219 1400 : ok.records.push((lsn, rec));
220 1400 : }
221 1297780 : Ok(Value::Image(img)) => {
222 1297780 : assert!(ok.img.is_none());
223 1297780 : ok.img = Some((lsn, img));
224 : }
225 0 : Err(err) => {
226 0 : res = Err(PageReconstructError::Other(err.into()));
227 0 : }
228 : }
229 : }
230 : }
231 1337378 : })();
232 1337378 : }
233 :
234 1336066 : res
235 1336066 : }
236 : }
237 :
238 : /// Bag of data accumulated during a vectored get..
239 : pub(crate) struct ValuesReconstructState {
240 : /// The keys will be removed after `get_vectored` completes. The caller outside `Timeline`
241 : /// should not expect to get anything from this hashmap.
242 : pub(crate) keys: HashMap<Key, VectoredValueReconstructState>,
243 : /// The keys which are already retrieved
244 : keys_done: KeySpaceRandomAccum,
245 :
246 : /// The keys covered by the image layers
247 : keys_with_image_coverage: Option<Range<Key>>,
248 :
249 : // Statistics that are still accessible as a caller of `get_vectored_impl`.
250 : layers_visited: u32,
251 : delta_layers_visited: u32,
252 :
253 : pub(crate) io_concurrency: IoConcurrency,
254 : num_active_ios: Arc<AtomicUsize>,
255 : }
256 :
257 : /// The level of IO concurrency to be used on the read path
258 : ///
259 : /// The desired end state is that we always do parallel IO.
260 : /// This struct and the dispatching in the impl will be removed once
261 : /// we've built enough confidence.
262 : pub(crate) enum IoConcurrency {
263 : Sequential,
264 : SidecarTask {
265 : task_id: usize,
266 : ios_tx: tokio::sync::mpsc::UnboundedSender<IoFuture>,
267 : },
268 : }
269 :
270 : type IoFuture = Pin<Box<dyn Send + Future<Output = ()>>>;
271 :
272 : pub(crate) enum SelectedIoConcurrency {
273 : Sequential,
274 : SidecarTask(GateGuard),
275 : }
276 :
277 : impl std::fmt::Debug for IoConcurrency {
278 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
279 0 : match self {
280 0 : IoConcurrency::Sequential => write!(f, "Sequential"),
281 0 : IoConcurrency::SidecarTask { .. } => write!(f, "SidecarTask"),
282 : }
283 0 : }
284 : }
285 :
286 : impl std::fmt::Debug for SelectedIoConcurrency {
287 64 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
288 64 : match self {
289 32 : SelectedIoConcurrency::Sequential => write!(f, "Sequential"),
290 32 : SelectedIoConcurrency::SidecarTask(_) => write!(f, "SidecarTask"),
291 : }
292 64 : }
293 : }
294 :
295 : impl IoConcurrency {
296 : /// Force sequential IO. This is a temporary workaround until we have
297 : /// moved plumbing-through-the-call-stack
298 : /// of IoConcurrency into `RequestContextq.
299 : ///
300 : /// DO NOT USE for new code.
301 : ///
302 : /// Tracking issue: <https://github.com/neondatabase/neon/issues/10460>.
303 1215206 : pub(crate) fn sequential() -> Self {
304 1215206 : Self::spawn(SelectedIoConcurrency::Sequential)
305 1215206 : }
306 :
307 1012 : pub(crate) fn spawn_from_conf(
308 1012 : conf: &'static PageServerConf,
309 1012 : gate_guard: GateGuard,
310 1012 : ) -> IoConcurrency {
311 : use pageserver_api::config::GetVectoredConcurrentIo;
312 1012 : let selected = match conf.get_vectored_concurrent_io {
313 1012 : GetVectoredConcurrentIo::Sequential => SelectedIoConcurrency::Sequential,
314 0 : GetVectoredConcurrentIo::SidecarTask => SelectedIoConcurrency::SidecarTask(gate_guard),
315 : };
316 1012 : Self::spawn(selected)
317 1012 : }
318 :
319 1216282 : pub(crate) fn spawn(io_concurrency: SelectedIoConcurrency) -> Self {
320 1216282 : match io_concurrency {
321 1216250 : SelectedIoConcurrency::Sequential => IoConcurrency::Sequential,
322 32 : SelectedIoConcurrency::SidecarTask(gate_guard) => {
323 32 : let (ios_tx, ios_rx) = tokio::sync::mpsc::unbounded_channel();
324 : static TASK_ID: AtomicUsize = AtomicUsize::new(0);
325 32 : let task_id = TASK_ID.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
326 : // TODO: enrich the span with more context (tenant,shard,timeline) + (basebackup|pagestream|...)
327 32 : let span =
328 32 : tracing::info_span!(parent: None, "IoConcurrency_sidecar", task_id = task_id);
329 32 : trace!(task_id, "spawning sidecar task");
330 32 : tokio::spawn(async move {
331 32 : trace!("start");
332 32 : scopeguard::defer!{ trace!("end") };
333 : type IosRx = tokio::sync::mpsc::UnboundedReceiver<IoFuture>;
334 : enum State {
335 : Waiting {
336 : // invariant: is_empty(), but we recycle the allocation
337 : empty_futures: FuturesUnordered<IoFuture>,
338 : ios_rx: IosRx,
339 : },
340 : Executing {
341 : futures: FuturesUnordered<IoFuture>,
342 : ios_rx: IosRx,
343 : },
344 : ShuttingDown {
345 : futures: FuturesUnordered<IoFuture>,
346 : },
347 : }
348 32 : let mut state = State::Waiting {
349 32 : empty_futures: FuturesUnordered::new(),
350 32 : ios_rx,
351 32 : };
352 : loop {
353 39084 : match state {
354 : State::Waiting {
355 18531 : empty_futures,
356 18531 : mut ios_rx,
357 18531 : } => {
358 18531 : assert!(empty_futures.is_empty());
359 18531 : tokio::select! {
360 18531 : fut = ios_rx.recv() => {
361 18499 : if let Some(fut) = fut {
362 18499 : trace!("received new io future");
363 18499 : empty_futures.push(fut);
364 18499 : state = State::Executing { futures: empty_futures, ios_rx };
365 : } else {
366 0 : state = State::ShuttingDown { futures: empty_futures }
367 : }
368 : }
369 : }
370 : }
371 : State::Executing {
372 20553 : mut futures,
373 20553 : mut ios_rx,
374 20553 : } => {
375 20553 : tokio::select! {
376 20553 : res = futures.next() => {
377 19526 : trace!("io future completed");
378 19526 : assert!(res.is_some());
379 19526 : if futures.is_empty() {
380 18499 : state = State::Waiting { empty_futures: futures, ios_rx};
381 18499 : } else {
382 1027 : state = State::Executing { futures, ios_rx };
383 1027 : }
384 : }
385 20553 : fut = ios_rx.recv() => {
386 1027 : if let Some(fut) = fut {
387 1027 : trace!("received new io future");
388 1027 : futures.push(fut);
389 1027 : state = State::Executing { futures, ios_rx};
390 0 : } else {
391 0 : state = State::ShuttingDown { futures };
392 0 : }
393 : }
394 : }
395 : }
396 : State::ShuttingDown {
397 0 : mut futures,
398 0 : } => {
399 0 : trace!("shutting down");
400 0 : while let Some(()) = futures.next().await {
401 0 : trace!("io future completed (shutdown)");
402 : // drain
403 : }
404 0 : trace!("shutdown complete");
405 0 : break;
406 0 : }
407 0 : }
408 0 : }
409 0 : drop(gate_guard); // drop it right before we exit
410 32 : }.instrument(span));
411 32 : IoConcurrency::SidecarTask { task_id, ios_tx }
412 : }
413 : }
414 1216282 : }
415 :
416 76338 : pub(crate) fn clone(&self) -> Self {
417 76338 : match self {
418 39460 : IoConcurrency::Sequential => IoConcurrency::Sequential,
419 36878 : IoConcurrency::SidecarTask { task_id, ios_tx } => IoConcurrency::SidecarTask {
420 36878 : task_id: *task_id,
421 36878 : ios_tx: ios_tx.clone(),
422 36878 : },
423 : }
424 76338 : }
425 :
426 : /// Submit an IO to be executed in the background. DEADLOCK RISK, read the full doc string.
427 : ///
428 : /// The IO is represented as an opaque future.
429 : /// IO completion must be handled inside the future, e.g., through a oneshot channel.
430 : ///
431 : /// The API seems simple but there are multiple **pitfalls** involving
432 : /// DEADLOCK RISK.
433 : ///
434 : /// First, there are no guarantees about the exexecution of the IO.
435 : /// It may be `await`ed in-place before this function returns.
436 : /// It may be polled partially by this task and handed off to another task to be finished.
437 : /// It may be polled and then dropped before returning ready.
438 : ///
439 : /// This means that submitted IOs must not be interedependent.
440 : /// Interdependence may be through shared limited resources, e.g.,
441 : /// - VirtualFile file descriptor cache slot acquisition
442 : /// - tokio-epoll-uring slot
443 : ///
444 : /// # Why current usage is safe from deadlocks
445 : ///
446 : /// Textbook condition for a deadlock is that _all_ of the following be given
447 : /// - Mutual exclusion
448 : /// - Hold and wait
449 : /// - No preemption
450 : /// - Circular wait
451 : ///
452 : /// The current usage is safe because:
453 : /// - Mutual exclusion: IO futures definitely use mutexes, no way around that for now
454 : /// - Hold and wait: IO futures currently hold two kinds of locks/resources while waiting
455 : /// for acquisition of other resources:
456 : /// - VirtualFile file descriptor cache slot tokio mutex
457 : /// - tokio-epoll-uring slot (uses tokio notify => wait queue, much like mutex)
458 : /// - No preemption: there's no taking-away of acquired locks/resources => given
459 : /// - Circular wait: this is the part of the condition that isn't met: all IO futures
460 : /// first acquire VirtualFile mutex, then tokio-epoll-uring slot.
461 : /// There is no IO future that acquires slot before VirtualFile.
462 : /// Hence there can be no circular waiting.
463 : /// Hence there cannot be a deadlock.
464 : ///
465 : /// This is a very fragile situation and must be revisited whenver any code called from
466 : /// inside the IO futures is changed.
467 : ///
468 : /// We will move away from opaque IO futures towards well-defined IOs at some point in
469 : /// the future when we have shipped this first version of concurrent IO to production
470 : /// and are ready to retire the Sequential mode which runs the futures in place.
471 : /// Right now, while brittle, the opaque IO approach allows us to ship the feature
472 : /// with minimal changes to the code and minimal changes to existing behavior in Sequential mode.
473 : ///
474 : /// Also read the comment in `collect_pending_ios`.
475 1526361 : pub(crate) async fn spawn_io<F>(&mut self, fut: F)
476 1526361 : where
477 1526361 : F: std::future::Future<Output = ()> + Send + 'static,
478 1526361 : {
479 1526361 : match self {
480 1506835 : IoConcurrency::Sequential => fut.await,
481 19526 : IoConcurrency::SidecarTask { ios_tx, .. } => {
482 19526 : let fut = Box::pin(fut);
483 19526 : // NB: experiments showed that doing an opportunistic poll of `fut` here was bad for throughput
484 19526 : // while insignificant for latency.
485 19526 : // It would make sense to revisit the tokio-epoll-uring API in the future such that we can try
486 19526 : // a submission here, but never poll the future. That way, io_uring can make proccess while
487 19526 : // the future sits in the ios_tx queue.
488 19526 : match ios_tx.send(fut) {
489 19526 : Ok(()) => {}
490 : Err(_) => {
491 0 : unreachable!("the io task must have exited, likely it panicked")
492 : }
493 : }
494 : }
495 : }
496 1526361 : }
497 :
498 : #[cfg(test)]
499 64 : pub(crate) fn spawn_for_test() -> impl std::ops::DerefMut<Target = Self> {
500 : use std::ops::{Deref, DerefMut};
501 : use tracing::info;
502 : use utils::sync::gate::Gate;
503 :
504 : // Spawn needs a Gate, give it one.
505 : struct Wrapper {
506 : inner: IoConcurrency,
507 : #[allow(dead_code)]
508 : gate: Box<Gate>,
509 : }
510 : impl Deref for Wrapper {
511 : type Target = IoConcurrency;
512 :
513 36974 : fn deref(&self) -> &Self::Target {
514 36974 : &self.inner
515 36974 : }
516 : }
517 : impl DerefMut for Wrapper {
518 0 : fn deref_mut(&mut self) -> &mut Self::Target {
519 0 : &mut self.inner
520 0 : }
521 : }
522 64 : let gate = Box::new(Gate::default());
523 :
524 : // The default behavior when running Rust unit tests without any further
525 : // flags is to use the new behavior.
526 : // The CI uses the following environment variable to unit test both old
527 : // and new behavior.
528 : // NB: the Python regression & perf tests take the `else` branch
529 : // below and have their own defaults management.
530 64 : let selected = {
531 : // The pageserver_api::config type is unsuitable because it's internally tagged.
532 64 : #[derive(serde::Deserialize)]
533 : #[serde(rename_all = "kebab-case")]
534 : enum TestOverride {
535 : Sequential,
536 : SidecarTask,
537 : }
538 : use once_cell::sync::Lazy;
539 64 : static TEST_OVERRIDE: Lazy<TestOverride> = Lazy::new(|| {
540 64 : utils::env::var_serde_json_string(
541 64 : "NEON_PAGESERVER_UNIT_TEST_GET_VECTORED_CONCURRENT_IO",
542 64 : )
543 64 : .unwrap_or(TestOverride::SidecarTask)
544 64 : });
545 :
546 64 : match *TEST_OVERRIDE {
547 32 : TestOverride::Sequential => SelectedIoConcurrency::Sequential,
548 : TestOverride::SidecarTask => {
549 32 : SelectedIoConcurrency::SidecarTask(gate.enter().expect("just created it"))
550 : }
551 : }
552 : };
553 :
554 64 : info!(?selected, "get_vectored_concurrent_io test");
555 :
556 64 : Wrapper {
557 64 : inner: Self::spawn(selected),
558 64 : gate,
559 64 : }
560 64 : }
561 : }
562 :
563 : /// Make noise in case the [`ValuesReconstructState`] gets dropped while
564 : /// there are still IOs in flight.
565 : /// Refer to `collect_pending_ios` for why we prefer not to do that.
566 : //
567 : /// We log from here instead of from the sidecar task because the [`ValuesReconstructState`]
568 : /// gets dropped in a tracing span with more context.
569 : /// We repeat the sidecar tasks's `task_id` so we can correlate what we emit here with
570 : /// the logs / panic handler logs from the sidecar task, which also logs the `task_id`.
571 : impl Drop for ValuesReconstructState {
572 1255316 : fn drop(&mut self) {
573 1255316 : let num_active_ios = self
574 1255316 : .num_active_ios
575 1255316 : .load(std::sync::atomic::Ordering::Acquire);
576 1255316 : if num_active_ios == 0 {
577 1255314 : return;
578 2 : }
579 2 : let sidecar_task_id = match &self.io_concurrency {
580 0 : IoConcurrency::Sequential => None,
581 2 : IoConcurrency::SidecarTask { task_id, .. } => Some(*task_id),
582 : };
583 2 : tracing::warn!(
584 : num_active_ios,
585 : ?sidecar_task_id,
586 0 : backtrace=%std::backtrace::Backtrace::force_capture(),
587 0 : "dropping ValuesReconstructState while some IOs have not been completed",
588 : );
589 1255316 : }
590 : }
591 :
592 : impl ValuesReconstructState {
593 1255316 : pub(crate) fn new(io_concurrency: IoConcurrency) -> Self {
594 1255316 : Self {
595 1255316 : keys: HashMap::new(),
596 1255316 : keys_done: KeySpaceRandomAccum::new(),
597 1255316 : keys_with_image_coverage: None,
598 1255316 : layers_visited: 0,
599 1255316 : delta_layers_visited: 0,
600 1255316 : io_concurrency,
601 1255316 : num_active_ios: Arc::new(AtomicUsize::new(0)),
602 1255316 : }
603 1255316 : }
604 :
605 : /// Absolutely read [`IoConcurrency::spawn_io`] to learn about assumptions & pitfalls.
606 1526361 : pub(crate) async fn spawn_io<F>(&mut self, fut: F)
607 1526361 : where
608 1526361 : F: std::future::Future<Output = ()> + Send + 'static,
609 1526361 : {
610 1526361 : self.io_concurrency.spawn_io(fut).await;
611 1526361 : }
612 :
613 1696103 : pub(crate) fn on_layer_visited(&mut self, layer: &ReadableLayer) {
614 1696103 : self.layers_visited += 1;
615 1696103 : if let ReadableLayer::PersistentLayer(layer) = layer {
616 482851 : if layer.layer_desc().is_delta() {
617 437759 : self.delta_layers_visited += 1;
618 437759 : }
619 1213252 : }
620 1696103 : }
621 :
622 456 : pub(crate) fn get_delta_layers_visited(&self) -> u32 {
623 456 : self.delta_layers_visited
624 456 : }
625 :
626 1255258 : pub(crate) fn get_layers_visited(&self) -> u32 {
627 1255258 : self.layers_visited
628 1255258 : }
629 :
630 : /// On hitting image layer, we can mark all keys in this range as done, because
631 : /// if the image layer does not contain a key, it is deleted/never added.
632 45116 : pub(crate) fn on_image_layer_visited(&mut self, key_range: &Range<Key>) {
633 45116 : let prev_val = self.keys_with_image_coverage.replace(key_range.clone());
634 45116 : assert_eq!(
635 : prev_val, None,
636 0 : "should consume the keyspace before the next iteration"
637 : );
638 45116 : }
639 :
640 : /// Update the state collected for a given key.
641 : /// Returns true if this was the last value needed for the key and false otherwise.
642 : ///
643 : /// If the key is done after the update, mark it as such.
644 : ///
645 : /// If the key is in the sparse keyspace (i.e., aux files), we do not track them in
646 : /// `key_done`.
647 : // TODO: rename this method & update description.
648 1482755 : pub(crate) fn update_key(&mut self, key: &Key, lsn: Lsn, completes: bool) -> OnDiskValueIo {
649 1482755 : let state = self.keys.entry(*key).or_default();
650 1482755 :
651 1482755 : let is_sparse_key = key.is_sparse();
652 :
653 1482755 : let required_io = match state.situation {
654 : ValueReconstructSituation::Complete => {
655 145375 : if is_sparse_key {
656 : // Sparse keyspace might be visited multiple times because
657 : // we don't track unmapped keyspaces.
658 145375 : return OnDiskValueIo::Unnecessary;
659 : } else {
660 0 : unreachable!()
661 : }
662 : }
663 : ValueReconstructSituation::Continue => {
664 1337380 : self.num_active_ios
665 1337380 : .fetch_add(1, std::sync::atomic::Ordering::Release);
666 1337380 : let (tx, rx) = tokio::sync::oneshot::channel();
667 1337380 : state.on_disk_values.push((lsn, OnDiskValueIoWaiter { rx }));
668 1337380 : OnDiskValueIo::Required {
669 1337380 : tx,
670 1337380 : num_active_ios: Arc::clone(&self.num_active_ios),
671 1337380 : }
672 1337380 : }
673 1337380 : };
674 1337380 :
675 1337380 : if completes && state.situation == ValueReconstructSituation::Continue {
676 1336068 : state.situation = ValueReconstructSituation::Complete;
677 1336068 : if !is_sparse_key {
678 1208552 : self.keys_done.add_key(*key);
679 1208552 : }
680 1312 : }
681 :
682 1337380 : required_io
683 1482755 : }
684 :
685 : /// Returns the key space describing the keys that have
686 : /// been marked as completed since the last call to this function.
687 : /// Returns individual keys done, and the image layer coverage.
688 3403623 : pub(crate) fn consume_done_keys(&mut self) -> (KeySpace, Option<Range<Key>>) {
689 3403623 : (
690 3403623 : self.keys_done.consume_keyspace(),
691 3403623 : self.keys_with_image_coverage.take(),
692 3403623 : )
693 3403623 : }
694 : }
695 :
696 : /// A key that uniquely identifies a layer in a timeline
697 : #[derive(Debug, PartialEq, Eq, Clone, Hash)]
698 : pub(crate) enum LayerId {
699 : PersitentLayerId(PersistentLayerKey),
700 : InMemoryLayerId(InMemoryLayerFileId),
701 : }
702 :
703 : /// Uniquely identify a layer visit by the layer
704 : /// and LSN floor (or start LSN) of the reads.
705 : /// The layer itself is not enough since we may
706 : /// have different LSN lower bounds for delta layer reads.
707 : #[derive(Debug, PartialEq, Eq, Clone, Hash)]
708 : struct LayerToVisitId {
709 : layer_id: LayerId,
710 : lsn_floor: Lsn,
711 : }
712 :
713 : /// Layer wrapper for the read path. Note that it is valid
714 : /// to use these layers even after external operations have
715 : /// been performed on them (compaction, freeze, etc.).
716 : #[derive(Debug)]
717 : pub(crate) enum ReadableLayer {
718 : PersistentLayer(Layer),
719 : InMemoryLayer(Arc<InMemoryLayer>),
720 : }
721 :
722 : /// A partial description of a read to be done.
723 : #[derive(Debug, Clone)]
724 : struct LayerVisit {
725 : /// An id used to resolve the readable layer within the fringe
726 : layer_to_visit_id: LayerToVisitId,
727 : /// Lsn range for the read, used for selecting the next read
728 : lsn_range: Range<Lsn>,
729 : }
730 :
731 : /// Data structure which maintains a fringe of layers for the
732 : /// read path. The fringe is the set of layers which intersects
733 : /// the current keyspace that the search is descending on.
734 : /// Each layer tracks the keyspace that intersects it.
735 : ///
736 : /// The fringe must appear sorted by Lsn. Hence, it uses
737 : /// a two layer indexing scheme.
738 : #[derive(Debug)]
739 : pub(crate) struct LayerFringe {
740 : planned_visits_by_lsn: BinaryHeap<LayerVisit>,
741 : visit_reads: HashMap<LayerToVisitId, LayerVisitReads>,
742 : }
743 :
744 : #[derive(Debug)]
745 : struct LayerVisitReads {
746 : layer: ReadableLayer,
747 : target_keyspace: KeySpaceRandomAccum,
748 : }
749 :
750 : impl LayerFringe {
751 1707520 : pub(crate) fn new() -> Self {
752 1707520 : LayerFringe {
753 1707520 : planned_visits_by_lsn: BinaryHeap::new(),
754 1707520 : visit_reads: HashMap::new(),
755 1707520 : }
756 1707520 : }
757 :
758 3403623 : pub(crate) fn next_layer(&mut self) -> Option<(ReadableLayer, KeySpace, Range<Lsn>)> {
759 3403623 : let read_desc = self.planned_visits_by_lsn.pop()?;
760 :
761 1696103 : let removed = self.visit_reads.remove_entry(&read_desc.layer_to_visit_id);
762 1696103 :
763 1696103 : match removed {
764 : Some((
765 : _,
766 : LayerVisitReads {
767 1696103 : layer,
768 1696103 : mut target_keyspace,
769 1696103 : },
770 1696103 : )) => Some((
771 1696103 : layer,
772 1696103 : target_keyspace.consume_keyspace(),
773 1696103 : read_desc.lsn_range,
774 1696103 : )),
775 0 : None => unreachable!("fringe internals are always consistent"),
776 : }
777 3403623 : }
778 :
779 1696131 : pub(crate) fn update(
780 1696131 : &mut self,
781 1696131 : layer: ReadableLayer,
782 1696131 : keyspace: KeySpace,
783 1696131 : lsn_range: Range<Lsn>,
784 1696131 : ) {
785 1696131 : let layer_to_visit_id = LayerToVisitId {
786 1696131 : layer_id: layer.id(),
787 1696131 : lsn_floor: lsn_range.start,
788 1696131 : };
789 1696131 :
790 1696131 : let entry = self.visit_reads.entry(layer_to_visit_id.clone());
791 1696131 : match entry {
792 28 : Entry::Occupied(mut entry) => {
793 28 : entry.get_mut().target_keyspace.add_keyspace(keyspace);
794 28 : }
795 1696103 : Entry::Vacant(entry) => {
796 1696103 : self.planned_visits_by_lsn.push(LayerVisit {
797 1696103 : lsn_range,
798 1696103 : layer_to_visit_id: layer_to_visit_id.clone(),
799 1696103 : });
800 1696103 : let mut accum = KeySpaceRandomAccum::new();
801 1696103 : accum.add_keyspace(keyspace);
802 1696103 : entry.insert(LayerVisitReads {
803 1696103 : layer,
804 1696103 : target_keyspace: accum,
805 1696103 : });
806 1696103 : }
807 : }
808 1696131 : }
809 : }
810 :
811 : impl Default for LayerFringe {
812 0 : fn default() -> Self {
813 0 : Self::new()
814 0 : }
815 : }
816 :
817 : impl Ord for LayerVisit {
818 60 : fn cmp(&self, other: &Self) -> Ordering {
819 60 : let ord = self.lsn_range.end.cmp(&other.lsn_range.end);
820 60 : if ord == std::cmp::Ordering::Equal {
821 44 : self.lsn_range.start.cmp(&other.lsn_range.start).reverse()
822 : } else {
823 16 : ord
824 : }
825 60 : }
826 : }
827 :
828 : impl PartialOrd for LayerVisit {
829 60 : fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
830 60 : Some(self.cmp(other))
831 60 : }
832 : }
833 :
834 : impl PartialEq for LayerVisit {
835 0 : fn eq(&self, other: &Self) -> bool {
836 0 : self.lsn_range == other.lsn_range
837 0 : }
838 : }
839 :
840 : impl Eq for LayerVisit {}
841 :
842 : impl ReadableLayer {
843 1696131 : pub(crate) fn id(&self) -> LayerId {
844 1696131 : match self {
845 482879 : Self::PersistentLayer(layer) => LayerId::PersitentLayerId(layer.layer_desc().key()),
846 1213252 : Self::InMemoryLayer(layer) => LayerId::InMemoryLayerId(layer.file_id()),
847 : }
848 1696131 : }
849 :
850 1696103 : pub(crate) async fn get_values_reconstruct_data(
851 1696103 : &self,
852 1696103 : keyspace: KeySpace,
853 1696103 : lsn_range: Range<Lsn>,
854 1696103 : reconstruct_state: &mut ValuesReconstructState,
855 1696103 : ctx: &RequestContext,
856 1696103 : ) -> Result<(), GetVectoredError> {
857 1696103 : match self {
858 482851 : ReadableLayer::PersistentLayer(layer) => {
859 482851 : layer
860 482851 : .get_values_reconstruct_data(keyspace, lsn_range, reconstruct_state, ctx)
861 482851 : .await
862 : }
863 1213252 : ReadableLayer::InMemoryLayer(layer) => {
864 1213252 : layer
865 1213252 : .get_values_reconstruct_data(keyspace, lsn_range.end, reconstruct_state, ctx)
866 1213252 : .await
867 : }
868 : }
869 1696103 : }
870 : }
871 :
872 : /// Layers contain a hint indicating whether they are likely to be used for reads.
873 : ///
874 : /// This is a hint rather than an authoritative value, so that we do not have to update it synchronously
875 : /// when changing the visibility of layers (for example when creating a branch that makes some previously
876 : /// covered layers visible). It should be used for cache management but not for correctness-critical checks.
877 : #[derive(Debug, Clone, PartialEq, Eq)]
878 : pub enum LayerVisibilityHint {
879 : /// A Visible layer might be read while serving a read, because there is not an image layer between it
880 : /// and a readable LSN (the tip of the branch or a child's branch point)
881 : Visible,
882 : /// A Covered layer probably won't be read right now, but _can_ be read in future if someone creates
883 : /// a branch or ephemeral endpoint at an LSN below the layer that covers this.
884 : Covered,
885 : }
886 :
887 : pub(crate) struct LayerAccessStats(std::sync::atomic::AtomicU64);
888 :
889 0 : #[derive(Clone, Copy, strum_macros::EnumString)]
890 : pub(crate) enum LayerAccessStatsReset {
891 : NoReset,
892 : AllStats,
893 : }
894 :
895 : impl Default for LayerAccessStats {
896 3836 : fn default() -> Self {
897 3836 : // Default value is to assume resident since creation time, and visible.
898 3836 : let (_mask, mut value) = Self::to_low_res_timestamp(Self::RTIME_SHIFT, SystemTime::now());
899 3836 : value |= 0x1 << Self::VISIBILITY_SHIFT;
900 3836 :
901 3836 : Self(std::sync::atomic::AtomicU64::new(value))
902 3836 : }
903 : }
904 :
905 : // Efficient store of two very-low-resolution timestamps and some bits. Used for storing last access time and
906 : // last residence change time.
907 : impl LayerAccessStats {
908 : // How many high bits to drop from a u32 timestamp?
909 : // - Only storing up to a u32 timestamp will work fine until 2038 (if this code is still in use
910 : // after that, this software has been very successful!)
911 : // - Dropping the top bit is implicitly safe because unix timestamps are meant to be
912 : // stored in an i32, so they never used it.
913 : // - Dropping the next two bits is safe because this code is only running on systems in
914 : // years >= 2024, and these bits have been 1 since 2021
915 : //
916 : // Therefore we may store only 28 bits for a timestamp with one second resolution. We do
917 : // this truncation to make space for some flags in the high bits of our u64.
918 : const TS_DROP_HIGH_BITS: u32 = u32::count_ones(Self::TS_ONES) + 1;
919 : const TS_MASK: u32 = 0x1f_ff_ff_ff;
920 : const TS_ONES: u32 = 0x60_00_00_00;
921 :
922 : const ATIME_SHIFT: u32 = 0;
923 : const RTIME_SHIFT: u32 = 32 - Self::TS_DROP_HIGH_BITS;
924 : const VISIBILITY_SHIFT: u32 = 64 - 2 * Self::TS_DROP_HIGH_BITS;
925 :
926 483091 : fn write_bits(&self, mask: u64, value: u64) -> u64 {
927 483091 : self.0
928 483091 : .fetch_update(
929 483091 : // TODO: decide what orderings are correct
930 483091 : std::sync::atomic::Ordering::Relaxed,
931 483091 : std::sync::atomic::Ordering::Relaxed,
932 483091 : |v| Some((v & !mask) | (value & mask)),
933 483091 : )
934 483091 : .expect("Inner function is infallible")
935 483091 : }
936 :
937 486207 : fn to_low_res_timestamp(shift: u32, time: SystemTime) -> (u64, u64) {
938 486207 : // Drop the low three bits of the timestamp, for an ~8s accuracy
939 486207 : let timestamp = time.duration_since(UNIX_EPOCH).unwrap().as_secs() & (Self::TS_MASK as u64);
940 486207 :
941 486207 : ((Self::TS_MASK as u64) << shift, timestamp << shift)
942 486207 : }
943 :
944 124 : fn read_low_res_timestamp(&self, shift: u32) -> Option<SystemTime> {
945 124 : let read = self.0.load(std::sync::atomic::Ordering::Relaxed);
946 124 :
947 124 : let ts_bits = (read & ((Self::TS_MASK as u64) << shift)) >> shift;
948 124 : if ts_bits == 0 {
949 48 : None
950 : } else {
951 76 : Some(UNIX_EPOCH + Duration::from_secs(ts_bits | (Self::TS_ONES as u64)))
952 : }
953 124 : }
954 :
955 : /// Record a change in layer residency.
956 : ///
957 : /// Recording the event must happen while holding the layer map lock to
958 : /// ensure that latest-activity-threshold-based layer eviction (eviction_task.rs)
959 : /// can do an "imitate access" to this layer, before it observes `now-latest_activity() > threshold`.
960 : ///
961 : /// If we instead recorded the residence event with a timestamp from before grabbing the layer map lock,
962 : /// the following race could happen:
963 : ///
964 : /// - Compact: Write out an L1 layer from several L0 layers. This records residence event LayerCreate with the current timestamp.
965 : /// - Eviction: imitate access logical size calculation. This accesses the L0 layers because the L1 layer is not yet in the layer map.
966 : /// - Compact: Grab layer map lock, add the new L1 to layer map and remove the L0s, release layer map lock.
967 : /// - Eviction: observes the new L1 layer whose only activity timestamp is the LayerCreate event.
968 52 : pub(crate) fn record_residence_event_at(&self, now: SystemTime) {
969 52 : let (mask, value) = Self::to_low_res_timestamp(Self::RTIME_SHIFT, now);
970 52 : self.write_bits(mask, value);
971 52 : }
972 :
973 48 : pub(crate) fn record_residence_event(&self) {
974 48 : self.record_residence_event_at(SystemTime::now())
975 48 : }
976 :
977 482319 : fn record_access_at(&self, now: SystemTime) -> bool {
978 482319 : let (mut mask, mut value) = Self::to_low_res_timestamp(Self::ATIME_SHIFT, now);
979 482319 :
980 482319 : // A layer which is accessed must be visible.
981 482319 : mask |= 0x1 << Self::VISIBILITY_SHIFT;
982 482319 : value |= 0x1 << Self::VISIBILITY_SHIFT;
983 482319 :
984 482319 : let old_bits = self.write_bits(mask, value);
985 4 : !matches!(
986 482319 : self.decode_visibility(old_bits),
987 : LayerVisibilityHint::Visible
988 : )
989 482319 : }
990 :
991 : /// Returns true if we modified the layer's visibility to set it to Visible implicitly
992 : /// as a result of this access
993 482875 : pub(crate) fn record_access(&self, ctx: &RequestContext) -> bool {
994 482875 : if ctx.access_stats_behavior() == AccessStatsBehavior::Skip {
995 568 : return false;
996 482307 : }
997 482307 :
998 482307 : self.record_access_at(SystemTime::now())
999 482875 : }
1000 :
1001 0 : fn as_api_model(
1002 0 : &self,
1003 0 : reset: LayerAccessStatsReset,
1004 0 : ) -> pageserver_api::models::LayerAccessStats {
1005 0 : let ret = pageserver_api::models::LayerAccessStats {
1006 0 : access_time: self
1007 0 : .read_low_res_timestamp(Self::ATIME_SHIFT)
1008 0 : .unwrap_or(UNIX_EPOCH),
1009 0 : residence_time: self
1010 0 : .read_low_res_timestamp(Self::RTIME_SHIFT)
1011 0 : .unwrap_or(UNIX_EPOCH),
1012 0 : visible: matches!(self.visibility(), LayerVisibilityHint::Visible),
1013 : };
1014 0 : match reset {
1015 0 : LayerAccessStatsReset::NoReset => {}
1016 0 : LayerAccessStatsReset::AllStats => {
1017 0 : self.write_bits((Self::TS_MASK as u64) << Self::ATIME_SHIFT, 0x0);
1018 0 : self.write_bits((Self::TS_MASK as u64) << Self::RTIME_SHIFT, 0x0);
1019 0 : }
1020 : }
1021 0 : ret
1022 0 : }
1023 :
1024 : /// Get the latest access timestamp, falling back to latest residence event. The latest residence event
1025 : /// will be this Layer's construction time, if its residence hasn't changed since then.
1026 32 : pub(crate) fn latest_activity(&self) -> SystemTime {
1027 32 : if let Some(t) = self.read_low_res_timestamp(Self::ATIME_SHIFT) {
1028 12 : t
1029 : } else {
1030 20 : self.read_low_res_timestamp(Self::RTIME_SHIFT)
1031 20 : .expect("Residence time is set on construction")
1032 : }
1033 32 : }
1034 :
1035 : /// Whether this layer has been accessed (excluding in [`AccessStatsBehavior::Skip`]).
1036 : ///
1037 : /// This indicates whether the layer has been used for some purpose that would motivate
1038 : /// us to keep it on disk, such as for serving a getpage request.
1039 36 : fn accessed(&self) -> bool {
1040 36 : // Consider it accessed if the most recent access is more recent than
1041 36 : // the most recent change in residence status.
1042 36 : match (
1043 36 : self.read_low_res_timestamp(Self::ATIME_SHIFT),
1044 36 : self.read_low_res_timestamp(Self::RTIME_SHIFT),
1045 : ) {
1046 28 : (None, _) => false,
1047 0 : (Some(_), None) => true,
1048 8 : (Some(a), Some(r)) => a >= r,
1049 : }
1050 36 : }
1051 :
1052 : /// Helper for extracting the visibility hint from the literal value of our inner u64
1053 484464 : fn decode_visibility(&self, bits: u64) -> LayerVisibilityHint {
1054 484464 : match (bits >> Self::VISIBILITY_SHIFT) & 0x1 {
1055 484420 : 1 => LayerVisibilityHint::Visible,
1056 44 : 0 => LayerVisibilityHint::Covered,
1057 0 : _ => unreachable!(),
1058 : }
1059 484464 : }
1060 :
1061 : /// Returns the old value which has been replaced
1062 720 : pub(crate) fn set_visibility(&self, visibility: LayerVisibilityHint) -> LayerVisibilityHint {
1063 720 : let value = match visibility {
1064 616 : LayerVisibilityHint::Visible => 0x1 << Self::VISIBILITY_SHIFT,
1065 104 : LayerVisibilityHint::Covered => 0x0,
1066 : };
1067 :
1068 720 : let old_bits = self.write_bits(0x1 << Self::VISIBILITY_SHIFT, value);
1069 720 : self.decode_visibility(old_bits)
1070 720 : }
1071 :
1072 1425 : pub(crate) fn visibility(&self) -> LayerVisibilityHint {
1073 1425 : let read = self.0.load(std::sync::atomic::Ordering::Relaxed);
1074 1425 : self.decode_visibility(read)
1075 1425 : }
1076 : }
1077 :
1078 : /// Get a layer descriptor from a layer.
1079 : pub(crate) trait AsLayerDesc {
1080 : /// Get the layer descriptor.
1081 : fn layer_desc(&self) -> &PersistentLayerDesc;
1082 : }
1083 :
1084 : pub mod tests {
1085 : use pageserver_api::shard::TenantShardId;
1086 : use utils::id::TimelineId;
1087 :
1088 : use super::*;
1089 :
1090 : impl From<DeltaLayerName> for PersistentLayerDesc {
1091 0 : fn from(value: DeltaLayerName) -> Self {
1092 0 : PersistentLayerDesc::new_delta(
1093 0 : TenantShardId::from([0; 18]),
1094 0 : TimelineId::from_array([0; 16]),
1095 0 : value.key_range,
1096 0 : value.lsn_range,
1097 0 : 233,
1098 0 : )
1099 0 : }
1100 : }
1101 :
1102 : impl From<ImageLayerName> for PersistentLayerDesc {
1103 0 : fn from(value: ImageLayerName) -> Self {
1104 0 : PersistentLayerDesc::new_img(
1105 0 : TenantShardId::from([0; 18]),
1106 0 : TimelineId::from_array([0; 16]),
1107 0 : value.key_range,
1108 0 : value.lsn,
1109 0 : 233,
1110 0 : )
1111 0 : }
1112 : }
1113 :
1114 : impl From<LayerName> for PersistentLayerDesc {
1115 0 : fn from(value: LayerName) -> Self {
1116 0 : match value {
1117 0 : LayerName::Delta(d) => Self::from(d),
1118 0 : LayerName::Image(i) => Self::from(i),
1119 : }
1120 0 : }
1121 : }
1122 : }
1123 :
1124 : /// Range wrapping newtype, which uses display to render Debug.
1125 : ///
1126 : /// Useful with `Key`, which has too verbose `{:?}` for printing multiple layers.
1127 : struct RangeDisplayDebug<'a, T: std::fmt::Display>(&'a Range<T>);
1128 :
1129 : impl<T: std::fmt::Display> std::fmt::Debug for RangeDisplayDebug<'_, T> {
1130 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
1131 0 : write!(f, "{}..{}", self.0.start, self.0.end)
1132 0 : }
1133 : }
1134 :
1135 : #[cfg(test)]
1136 : mod tests2 {
1137 : use pageserver_api::key::DBDIR_KEY;
1138 : use tracing::info;
1139 :
1140 : use super::*;
1141 : use crate::tenant::storage_layer::IoConcurrency;
1142 :
1143 : /// TODO: currently this test relies on manual visual inspection of the --no-capture output.
1144 : /// Should look like so:
1145 : /// ```text
1146 : /// RUST_LOG=trace cargo nextest run --features testing --no-capture test_io_concurrency_noise
1147 : /// running 1 test
1148 : /// 2025-01-21T17:42:01.335679Z INFO get_vectored_concurrent_io test selected=SidecarTask
1149 : /// 2025-01-21T17:42:01.335680Z TRACE spawning sidecar task task_id=0
1150 : /// 2025-01-21T17:42:01.335937Z TRACE IoConcurrency_sidecar{task_id=0}: start
1151 : /// 2025-01-21T17:42:01.335972Z TRACE IoConcurrency_sidecar{task_id=0}: received new io future
1152 : /// 2025-01-21T17:42:01.335999Z INFO IoConcurrency_sidecar{task_id=0}: waiting for signal to complete IO
1153 : /// 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
1154 : /// at ./src/tenant/storage_layer.rs:553:24
1155 : /// 1: core::ptr::drop_in_place<pageserver::tenant::storage_layer::ValuesReconstructState>
1156 : /// 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
1157 : /// 2: core::mem::drop
1158 : /// 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
1159 : /// 3: pageserver::tenant::storage_layer::tests2::test_io_concurrency_noise::{{closure}}
1160 : /// at ./src/tenant/storage_layer.rs:1159:9
1161 : /// ...
1162 : /// 49: <unknown>
1163 : /// 2025-01-21T17:42:01.452293Z INFO IoConcurrency_sidecar{task_id=0}: completing IO
1164 : /// 2025-01-21T17:42:01.452357Z TRACE IoConcurrency_sidecar{task_id=0}: io future completed
1165 : /// 2025-01-21T17:42:01.452473Z TRACE IoConcurrency_sidecar{task_id=0}: end
1166 : /// test tenant::storage_layer::tests2::test_io_concurrency_noise ... ok
1167 : ///
1168 : /// ```
1169 : #[tokio::test]
1170 4 : async fn test_io_concurrency_noise() {
1171 4 : crate::tenant::harness::setup_logging();
1172 4 :
1173 4 : let io_concurrency = IoConcurrency::spawn_for_test();
1174 4 : match *io_concurrency {
1175 4 : IoConcurrency::Sequential => {
1176 4 : // This test asserts behavior in sidecar mode, doesn't make sense in sequential mode.
1177 4 : return;
1178 4 : }
1179 4 : IoConcurrency::SidecarTask { .. } => {}
1180 2 : }
1181 2 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
1182 2 :
1183 2 : let (io_fut_is_waiting_tx, io_fut_is_waiting) = tokio::sync::oneshot::channel();
1184 2 : let (do_complete_io, should_complete_io) = tokio::sync::oneshot::channel();
1185 2 : let (io_fut_exiting_tx, io_fut_exiting) = tokio::sync::oneshot::channel();
1186 2 :
1187 2 : let io = reconstruct_state.update_key(&DBDIR_KEY, Lsn(8), true);
1188 2 : reconstruct_state
1189 2 : .spawn_io(async move {
1190 2 : info!("waiting for signal to complete IO");
1191 4 : io_fut_is_waiting_tx.send(()).unwrap();
1192 2 : should_complete_io.await.unwrap();
1193 2 : info!("completing IO");
1194 4 : io.complete(Ok(OnDiskValue::RawImage(Bytes::new())));
1195 2 : io_fut_exiting_tx.send(()).unwrap();
1196 2 : })
1197 2 : .await;
1198 4 :
1199 4 : io_fut_is_waiting.await.unwrap();
1200 2 :
1201 2 : // this is what makes the noise
1202 2 : drop(reconstruct_state);
1203 2 :
1204 2 : do_complete_io.send(()).unwrap();
1205 2 :
1206 2 : io_fut_exiting.await.unwrap();
1207 4 : }
1208 : }
|
