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, ReadPath};
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 : impl ValueReconstructState {
84 : /// Returns the number of page deltas applied to the page image.
85 1335951 : pub fn num_deltas(&self) -> usize {
86 1335951 : match self.img {
87 1335863 : Some(_) => self.records.len(),
88 88 : None => self.records.len() - 1, // omit will_init record
89 : }
90 1335951 : }
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 1482275 : pub(crate) fn complete(self, res: OnDiskValueIoResult) {
148 1482275 : match self {
149 1337405 : OnDiskValueIo::Required { num_active_ios, tx } => {
150 1337405 : num_active_ios.fetch_sub(1, std::sync::atomic::Ordering::Release);
151 1337405 : let _ = tx.send(res);
152 1337405 : }
153 144870 : OnDiskValueIo::Unnecessary => {
154 144870 : // Nobody cared, see variant doc comment.
155 144870 : }
156 : }
157 1482275 : }
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 1337403 : pub(crate) async fn wait_completion(self) -> Result<OnDiskValueIoResult, WaitCompletionError> {
168 1337403 : // NB: for Unnecessary IOs, this method never gets called because we don't add them to `on_disk_values`.
169 1337403 : self.rx.await.map_err(|_| WaitCompletionError::IoDropped)
170 1337403 : }
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 1336091 : pub(crate) async fn collect_pending_ios(
180 1336091 : self,
181 1336091 : ) -> Result<ValueReconstructState, PageReconstructError> {
182 : use utils::bin_ser::BeSer;
183 :
184 1336091 : 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 2673494 : for (lsn, waiter) in self.on_disk_values {
197 1337403 : let value_recv_res = waiter
198 1337403 : .wait_completion()
199 1337403 : // we rely on the caller to poll us to completion, so this is not a bail point
200 1337403 : .await;
201 : // Force not bailing early by wrapping the code into a closure.
202 : #[allow(clippy::redundant_closure_call)]
203 1337403 : let _: () = (|| {
204 1337403 : 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 38497 : (Ok(ok), Ok(Ok(OnDiskValue::RawImage(img)))) => {
223 38497 : assert!(ok.img.is_none());
224 38497 : ok.img = Some((lsn, img));
225 : }
226 1298906 : (Ok(ok), Ok(Ok(OnDiskValue::WalRecordOrImage(buf)))) => {
227 1298906 : match Value::des(&buf) {
228 1400 : Ok(Value::WalRecord(rec)) => {
229 1400 : ok.records.push((lsn, rec));
230 1400 : }
231 1297506 : Ok(Value::Image(img)) => {
232 1297506 : assert!(ok.img.is_none());
233 1297506 : ok.img = Some((lsn, img));
234 : }
235 0 : Err(err) => {
236 0 : res = Err(PageReconstructError::Other(err.into()));
237 0 : }
238 : }
239 : }
240 : }
241 1337403 : })();
242 1337403 : }
243 :
244 1336091 : res
245 1336091 : }
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 1215203 : pub(crate) fn sequential() -> Self {
316 1215203 : Self::spawn(SelectedIoConcurrency::Sequential)
317 1215203 : }
318 :
319 1020 : pub(crate) fn spawn_from_conf(
320 1020 : conf: &'static PageServerConf,
321 1020 : gate_guard: GateGuard,
322 1020 : ) -> IoConcurrency {
323 : use pageserver_api::config::GetVectoredConcurrentIo;
324 1020 : let selected = match conf.get_vectored_concurrent_io {
325 1020 : GetVectoredConcurrentIo::Sequential => SelectedIoConcurrency::Sequential,
326 0 : GetVectoredConcurrentIo::SidecarTask => SelectedIoConcurrency::SidecarTask(gate_guard),
327 : };
328 1020 : Self::spawn(selected)
329 1020 : }
330 :
331 1216287 : pub(crate) fn spawn(io_concurrency: SelectedIoConcurrency) -> Self {
332 1216287 : match io_concurrency {
333 1216255 : 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 39092 : match state {
366 : State::Waiting {
367 18551 : empty_futures,
368 18551 : mut ios_rx,
369 18551 : } => {
370 18551 : assert!(empty_futures.is_empty());
371 18551 : tokio::select! {
372 18551 : fut = ios_rx.recv() => {
373 18519 : if let Some(fut) = fut {
374 18519 : trace!("received new io future");
375 18519 : empty_futures.push(fut);
376 18519 : 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 20541 : mut futures,
385 20541 : mut ios_rx,
386 20541 : } => {
387 20541 : tokio::select! {
388 20541 : res = futures.next() => {
389 19530 : trace!("io future completed");
390 19530 : assert!(res.is_some());
391 19530 : if futures.is_empty() {
392 18519 : state = State::Waiting { empty_futures: futures, ios_rx};
393 18519 : } else {
394 1011 : state = State::Executing { futures, ios_rx };
395 1011 : }
396 : }
397 20541 : fut = ios_rx.recv() => {
398 1011 : if let Some(fut) = fut {
399 1011 : trace!("received new io future");
400 1011 : futures.push(fut);
401 1011 : 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 1216287 : }
427 :
428 76362 : pub(crate) fn clone(&self) -> Self {
429 76362 : match self {
430 39484 : 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 76362 : }
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 1525814 : pub(crate) async fn spawn_io<F>(&mut self, fut: F)
488 1525814 : where
489 1525814 : F: std::future::Future<Output = ()> + Send + 'static,
490 1525814 : {
491 1525814 : match self {
492 1506284 : IoConcurrency::Sequential => fut.await,
493 19530 : IoConcurrency::SidecarTask { ios_tx, .. } => {
494 19530 : let fut = Box::pin(fut);
495 19530 : // NB: experiments showed that doing an opportunistic poll of `fut` here was bad for throughput
496 19530 : // while insignificant for latency.
497 19530 : // It would make sense to revisit the tokio-epoll-uring API in the future such that we can try
498 19530 : // a submission here, but never poll the future. That way, io_uring can make proccess while
499 19530 : // the future sits in the ios_tx queue.
500 19530 : match ios_tx.send(fut) {
501 19530 : Ok(()) => {}
502 : Err(_) => {
503 0 : unreachable!("the io task must have exited, likely it panicked")
504 : }
505 : }
506 : }
507 : }
508 1525814 : }
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 : use tracing::info;
514 : use utils::sync::gate::Gate;
515 :
516 : // Spawn needs a Gate, give it one.
517 : struct Wrapper {
518 : inner: IoConcurrency,
519 : #[allow(dead_code)]
520 : gate: Box<Gate>,
521 : }
522 : impl Deref for Wrapper {
523 : type Target = IoConcurrency;
524 :
525 36974 : fn deref(&self) -> &Self::Target {
526 36974 : &self.inner
527 36974 : }
528 : }
529 : impl DerefMut for Wrapper {
530 0 : fn deref_mut(&mut self) -> &mut Self::Target {
531 0 : &mut self.inner
532 0 : }
533 : }
534 64 : let gate = Box::new(Gate::default());
535 :
536 : // The default behavior when running Rust unit tests without any further
537 : // flags is to use the new behavior.
538 : // The CI uses the following environment variable to unit test both old
539 : // and new behavior.
540 : // NB: the Python regression & perf tests take the `else` branch
541 : // below and have their own defaults management.
542 64 : let selected = {
543 : // The pageserver_api::config type is unsuitable because it's internally tagged.
544 64 : #[derive(serde::Deserialize)]
545 : #[serde(rename_all = "kebab-case")]
546 : enum TestOverride {
547 : Sequential,
548 : SidecarTask,
549 : }
550 : use once_cell::sync::Lazy;
551 64 : static TEST_OVERRIDE: Lazy<TestOverride> = Lazy::new(|| {
552 64 : utils::env::var_serde_json_string(
553 64 : "NEON_PAGESERVER_UNIT_TEST_GET_VECTORED_CONCURRENT_IO",
554 64 : )
555 64 : .unwrap_or(TestOverride::SidecarTask)
556 64 : });
557 :
558 64 : match *TEST_OVERRIDE {
559 32 : TestOverride::Sequential => SelectedIoConcurrency::Sequential,
560 : TestOverride::SidecarTask => {
561 32 : SelectedIoConcurrency::SidecarTask(gate.enter().expect("just created it"))
562 : }
563 : }
564 : };
565 :
566 64 : info!(?selected, "get_vectored_concurrent_io test");
567 :
568 64 : Wrapper {
569 64 : inner: Self::spawn(selected),
570 64 : gate,
571 64 : }
572 64 : }
573 : }
574 :
575 : /// Make noise in case the [`ValuesReconstructState`] gets dropped while
576 : /// there are still IOs in flight.
577 : /// Refer to `collect_pending_ios` for why we prefer not to do that.
578 : //
579 : /// We log from here instead of from the sidecar task because the [`ValuesReconstructState`]
580 : /// gets dropped in a tracing span with more context.
581 : /// We repeat the sidecar tasks's `task_id` so we can correlate what we emit here with
582 : /// the logs / panic handler logs from the sidecar task, which also logs the `task_id`.
583 : impl Drop for ValuesReconstructState {
584 1255341 : fn drop(&mut self) {
585 1255341 : let num_active_ios = self
586 1255341 : .num_active_ios
587 1255341 : .load(std::sync::atomic::Ordering::Acquire);
588 1255341 : if num_active_ios == 0 {
589 1255339 : return;
590 2 : }
591 2 : let sidecar_task_id = match &self.io_concurrency {
592 0 : IoConcurrency::Sequential => None,
593 2 : IoConcurrency::SidecarTask { task_id, .. } => Some(*task_id),
594 : };
595 2 : tracing::warn!(
596 : num_active_ios,
597 : ?sidecar_task_id,
598 0 : backtrace=%std::backtrace::Backtrace::force_capture(),
599 0 : "dropping ValuesReconstructState while some IOs have not been completed",
600 : );
601 1255341 : }
602 : }
603 :
604 : impl ValuesReconstructState {
605 1255341 : pub(crate) fn new(io_concurrency: IoConcurrency) -> Self {
606 1255341 : Self {
607 1255341 : keys: HashMap::new(),
608 1255341 : keys_done: KeySpaceRandomAccum::new(),
609 1255341 : keys_with_image_coverage: None,
610 1255341 : layers_visited: 0,
611 1255341 : delta_layers_visited: 0,
612 1255341 : io_concurrency,
613 1255341 : num_active_ios: Arc::new(AtomicUsize::new(0)),
614 1255341 : read_path: None,
615 1255341 : }
616 1255341 : }
617 :
618 : /// Absolutely read [`IoConcurrency::spawn_io`] to learn about assumptions & pitfalls.
619 1525814 : pub(crate) async fn spawn_io<F>(&mut self, fut: F)
620 1525814 : where
621 1525814 : F: std::future::Future<Output = ()> + Send + 'static,
622 1525814 : {
623 1525814 : self.io_concurrency.spawn_io(fut).await;
624 1525814 : }
625 :
626 1692763 : pub(crate) fn on_layer_visited(&mut self, layer: &ReadableLayer) {
627 1692763 : self.layers_visited += 1;
628 1692763 : if let ReadableLayer::PersistentLayer(layer) = layer {
629 479416 : if layer.layer_desc().is_delta() {
630 434324 : self.delta_layers_visited += 1;
631 434324 : }
632 1213347 : }
633 1692763 : }
634 :
635 460 : pub(crate) fn get_delta_layers_visited(&self) -> u32 {
636 460 : self.delta_layers_visited
637 460 : }
638 :
639 1255283 : pub(crate) fn get_layers_visited(&self) -> u32 {
640 1255283 : self.layers_visited
641 1255283 : }
642 :
643 : /// On hitting image layer, we can mark all keys in this range as done, because
644 : /// if the image layer does not contain a key, it is deleted/never added.
645 45116 : pub(crate) fn on_image_layer_visited(&mut self, key_range: &Range<Key>) {
646 45116 : let prev_val = self.keys_with_image_coverage.replace(key_range.clone());
647 45116 : assert_eq!(
648 : prev_val, None,
649 0 : "should consume the keyspace before the next iteration"
650 : );
651 45116 : }
652 :
653 : /// Update the state collected for a given key.
654 : /// Returns true if this was the last value needed for the key and false otherwise.
655 : ///
656 : /// If the key is done after the update, mark it as such.
657 : ///
658 : /// If the key is in the sparse keyspace (i.e., aux files), we do not track them in
659 : /// `key_done`.
660 : // TODO: rename this method & update description.
661 1482275 : pub(crate) fn update_key(&mut self, key: &Key, lsn: Lsn, completes: bool) -> OnDiskValueIo {
662 1482275 : let state = self.keys.entry(*key).or_default();
663 1482275 :
664 1482275 : let is_sparse_key = key.is_sparse();
665 :
666 1482275 : let required_io = match state.situation {
667 : ValueReconstructSituation::Complete => {
668 144870 : if is_sparse_key {
669 : // Sparse keyspace might be visited multiple times because
670 : // we don't track unmapped keyspaces.
671 144870 : return OnDiskValueIo::Unnecessary;
672 : } else {
673 0 : unreachable!()
674 : }
675 : }
676 : ValueReconstructSituation::Continue => {
677 1337405 : self.num_active_ios
678 1337405 : .fetch_add(1, std::sync::atomic::Ordering::Release);
679 1337405 : let (tx, rx) = tokio::sync::oneshot::channel();
680 1337405 : state.on_disk_values.push((lsn, OnDiskValueIoWaiter { rx }));
681 1337405 : OnDiskValueIo::Required {
682 1337405 : tx,
683 1337405 : num_active_ios: Arc::clone(&self.num_active_ios),
684 1337405 : }
685 1337405 : }
686 1337405 : };
687 1337405 :
688 1337405 : if completes && state.situation == ValueReconstructSituation::Continue {
689 1336093 : state.situation = ValueReconstructSituation::Complete;
690 1336093 : if !is_sparse_key {
691 1208577 : self.keys_done.add_key(*key);
692 1208577 : }
693 1312 : }
694 :
695 1337405 : required_io
696 1482275 : }
697 :
698 : /// Returns the key space describing the keys that have
699 : /// been marked as completed since the last call to this function.
700 : /// Returns individual keys done, and the image layer coverage.
701 3400578 : pub(crate) fn consume_done_keys(&mut self) -> (KeySpace, Option<Range<Key>>) {
702 3400578 : (
703 3400578 : self.keys_done.consume_keyspace(),
704 3400578 : self.keys_with_image_coverage.take(),
705 3400578 : )
706 3400578 : }
707 : }
708 :
709 : /// A key that uniquely identifies a layer in a timeline
710 : #[derive(Debug, PartialEq, Eq, Clone, Hash)]
711 : pub(crate) enum LayerId {
712 : PersitentLayerId(PersistentLayerKey),
713 : InMemoryLayerId(InMemoryLayerFileId),
714 : }
715 :
716 : /// Uniquely identify a layer visit by the layer
717 : /// and LSN floor (or start LSN) of the reads.
718 : /// The layer itself is not enough since we may
719 : /// have different LSN lower bounds for delta layer reads.
720 : #[derive(Debug, PartialEq, Eq, Clone, Hash)]
721 : struct LayerToVisitId {
722 : layer_id: LayerId,
723 : lsn_floor: Lsn,
724 : }
725 :
726 : /// Layer wrapper for the read path. Note that it is valid
727 : /// to use these layers even after external operations have
728 : /// been performed on them (compaction, freeze, etc.).
729 : #[derive(Debug)]
730 : pub(crate) enum ReadableLayer {
731 : PersistentLayer(Layer),
732 : InMemoryLayer(Arc<InMemoryLayer>),
733 : }
734 :
735 : /// A partial description of a read to be done.
736 : #[derive(Debug, Clone)]
737 : struct LayerVisit {
738 : /// An id used to resolve the readable layer within the fringe
739 : layer_to_visit_id: LayerToVisitId,
740 : /// Lsn range for the read, used for selecting the next read
741 : lsn_range: Range<Lsn>,
742 : }
743 :
744 : /// Data structure which maintains a fringe of layers for the
745 : /// read path. The fringe is the set of layers which intersects
746 : /// the current keyspace that the search is descending on.
747 : /// Each layer tracks the keyspace that intersects it.
748 : ///
749 : /// The fringe must appear sorted by Lsn. Hence, it uses
750 : /// a two layer indexing scheme.
751 : #[derive(Debug)]
752 : pub(crate) struct LayerFringe {
753 : planned_visits_by_lsn: BinaryHeap<LayerVisit>,
754 : visit_reads: HashMap<LayerToVisitId, LayerVisitReads>,
755 : }
756 :
757 : #[derive(Debug)]
758 : struct LayerVisitReads {
759 : layer: ReadableLayer,
760 : target_keyspace: KeySpaceRandomAccum,
761 : }
762 :
763 : impl LayerFringe {
764 1707815 : pub(crate) fn new() -> Self {
765 1707815 : LayerFringe {
766 1707815 : planned_visits_by_lsn: BinaryHeap::new(),
767 1707815 : visit_reads: HashMap::new(),
768 1707815 : }
769 1707815 : }
770 :
771 3400578 : pub(crate) fn next_layer(&mut self) -> Option<(ReadableLayer, KeySpace, Range<Lsn>)> {
772 3400578 : let read_desc = self.planned_visits_by_lsn.pop()?;
773 :
774 1692763 : let removed = self.visit_reads.remove_entry(&read_desc.layer_to_visit_id);
775 1692763 :
776 1692763 : match removed {
777 : Some((
778 : _,
779 : LayerVisitReads {
780 1692763 : layer,
781 1692763 : mut target_keyspace,
782 1692763 : },
783 1692763 : )) => Some((
784 1692763 : layer,
785 1692763 : target_keyspace.consume_keyspace(),
786 1692763 : read_desc.lsn_range,
787 1692763 : )),
788 0 : None => unreachable!("fringe internals are always consistent"),
789 : }
790 3400578 : }
791 :
792 1692851 : pub(crate) fn update(
793 1692851 : &mut self,
794 1692851 : layer: ReadableLayer,
795 1692851 : keyspace: KeySpace,
796 1692851 : lsn_range: Range<Lsn>,
797 1692851 : ) {
798 1692851 : let layer_to_visit_id = LayerToVisitId {
799 1692851 : layer_id: layer.id(),
800 1692851 : lsn_floor: lsn_range.start,
801 1692851 : };
802 1692851 :
803 1692851 : let entry = self.visit_reads.entry(layer_to_visit_id.clone());
804 1692851 : match entry {
805 88 : Entry::Occupied(mut entry) => {
806 88 : entry.get_mut().target_keyspace.add_keyspace(keyspace);
807 88 : }
808 1692763 : Entry::Vacant(entry) => {
809 1692763 : self.planned_visits_by_lsn.push(LayerVisit {
810 1692763 : lsn_range,
811 1692763 : layer_to_visit_id: layer_to_visit_id.clone(),
812 1692763 : });
813 1692763 : let mut accum = KeySpaceRandomAccum::new();
814 1692763 : accum.add_keyspace(keyspace);
815 1692763 : entry.insert(LayerVisitReads {
816 1692763 : layer,
817 1692763 : target_keyspace: accum,
818 1692763 : });
819 1692763 : }
820 : }
821 1692851 : }
822 : }
823 :
824 : impl Default for LayerFringe {
825 0 : fn default() -> Self {
826 0 : Self::new()
827 0 : }
828 : }
829 :
830 : impl Ord for LayerVisit {
831 60 : fn cmp(&self, other: &Self) -> Ordering {
832 60 : let ord = self.lsn_range.end.cmp(&other.lsn_range.end);
833 60 : if ord == std::cmp::Ordering::Equal {
834 44 : self.lsn_range.start.cmp(&other.lsn_range.start).reverse()
835 : } else {
836 16 : ord
837 : }
838 60 : }
839 : }
840 :
841 : impl PartialOrd for LayerVisit {
842 60 : fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
843 60 : Some(self.cmp(other))
844 60 : }
845 : }
846 :
847 : impl PartialEq for LayerVisit {
848 0 : fn eq(&self, other: &Self) -> bool {
849 0 : self.lsn_range == other.lsn_range
850 0 : }
851 : }
852 :
853 : impl Eq for LayerVisit {}
854 :
855 : impl ReadableLayer {
856 1692851 : pub(crate) fn id(&self) -> LayerId {
857 1692851 : match self {
858 479504 : Self::PersistentLayer(layer) => LayerId::PersitentLayerId(layer.layer_desc().key()),
859 1213347 : Self::InMemoryLayer(layer) => LayerId::InMemoryLayerId(layer.file_id()),
860 : }
861 1692851 : }
862 :
863 1692763 : pub(crate) async fn get_values_reconstruct_data(
864 1692763 : &self,
865 1692763 : keyspace: KeySpace,
866 1692763 : lsn_range: Range<Lsn>,
867 1692763 : reconstruct_state: &mut ValuesReconstructState,
868 1692763 : ctx: &RequestContext,
869 1692763 : ) -> Result<(), GetVectoredError> {
870 1692763 : match self {
871 479416 : ReadableLayer::PersistentLayer(layer) => {
872 479416 : layer
873 479416 : .get_values_reconstruct_data(keyspace, lsn_range, reconstruct_state, ctx)
874 479416 : .await
875 : }
876 1213347 : ReadableLayer::InMemoryLayer(layer) => {
877 1213347 : layer
878 1213347 : .get_values_reconstruct_data(keyspace, lsn_range.end, reconstruct_state, ctx)
879 1213347 : .await
880 : }
881 : }
882 1692763 : }
883 : }
884 :
885 : /// Layers contain a hint indicating whether they are likely to be used for reads.
886 : ///
887 : /// This is a hint rather than an authoritative value, so that we do not have to update it synchronously
888 : /// when changing the visibility of layers (for example when creating a branch that makes some previously
889 : /// covered layers visible). It should be used for cache management but not for correctness-critical checks.
890 : #[derive(Debug, Clone, PartialEq, Eq)]
891 : pub enum LayerVisibilityHint {
892 : /// A Visible layer might be read while serving a read, because there is not an image layer between it
893 : /// and a readable LSN (the tip of the branch or a child's branch point)
894 : Visible,
895 : /// A Covered layer probably won't be read right now, but _can_ be read in future if someone creates
896 : /// a branch or ephemeral endpoint at an LSN below the layer that covers this.
897 : Covered,
898 : }
899 :
900 : pub(crate) struct LayerAccessStats(std::sync::atomic::AtomicU64);
901 :
902 0 : #[derive(Clone, Copy, strum_macros::EnumString)]
903 : pub(crate) enum LayerAccessStatsReset {
904 : NoReset,
905 : AllStats,
906 : }
907 :
908 : impl Default for LayerAccessStats {
909 3848 : fn default() -> Self {
910 3848 : // Default value is to assume resident since creation time, and visible.
911 3848 : let (_mask, mut value) = Self::to_low_res_timestamp(Self::RTIME_SHIFT, SystemTime::now());
912 3848 : value |= 0x1 << Self::VISIBILITY_SHIFT;
913 3848 :
914 3848 : Self(std::sync::atomic::AtomicU64::new(value))
915 3848 : }
916 : }
917 :
918 : // Efficient store of two very-low-resolution timestamps and some bits. Used for storing last access time and
919 : // last residence change time.
920 : impl LayerAccessStats {
921 : // How many high bits to drop from a u32 timestamp?
922 : // - Only storing up to a u32 timestamp will work fine until 2038 (if this code is still in use
923 : // after that, this software has been very successful!)
924 : // - Dropping the top bit is implicitly safe because unix timestamps are meant to be
925 : // stored in an i32, so they never used it.
926 : // - Dropping the next two bits is safe because this code is only running on systems in
927 : // years >= 2024, and these bits have been 1 since 2021
928 : //
929 : // Therefore we may store only 28 bits for a timestamp with one second resolution. We do
930 : // this truncation to make space for some flags in the high bits of our u64.
931 : const TS_DROP_HIGH_BITS: u32 = u32::count_ones(Self::TS_ONES) + 1;
932 : const TS_MASK: u32 = 0x1f_ff_ff_ff;
933 : const TS_ONES: u32 = 0x60_00_00_00;
934 :
935 : const ATIME_SHIFT: u32 = 0;
936 : const RTIME_SHIFT: u32 = 32 - Self::TS_DROP_HIGH_BITS;
937 : const VISIBILITY_SHIFT: u32 = 64 - 2 * Self::TS_DROP_HIGH_BITS;
938 :
939 479716 : fn write_bits(&self, mask: u64, value: u64) -> u64 {
940 479716 : self.0
941 479716 : .fetch_update(
942 479716 : // TODO: decide what orderings are correct
943 479716 : std::sync::atomic::Ordering::Relaxed,
944 479716 : std::sync::atomic::Ordering::Relaxed,
945 479716 : |v| Some((v & !mask) | (value & mask)),
946 479716 : )
947 479716 : .expect("Inner function is infallible")
948 479716 : }
949 :
950 482828 : fn to_low_res_timestamp(shift: u32, time: SystemTime) -> (u64, u64) {
951 482828 : // Drop the low three bits of the timestamp, for an ~8s accuracy
952 482828 : let timestamp = time.duration_since(UNIX_EPOCH).unwrap().as_secs() & (Self::TS_MASK as u64);
953 482828 :
954 482828 : ((Self::TS_MASK as u64) << shift, timestamp << shift)
955 482828 : }
956 :
957 292 : fn read_low_res_timestamp(&self, shift: u32) -> Option<SystemTime> {
958 292 : let read = self.0.load(std::sync::atomic::Ordering::Relaxed);
959 292 :
960 292 : let ts_bits = (read & ((Self::TS_MASK as u64) << shift)) >> shift;
961 292 : if ts_bits == 0 {
962 132 : None
963 : } else {
964 160 : Some(UNIX_EPOCH + Duration::from_secs(ts_bits | (Self::TS_ONES as u64)))
965 : }
966 292 : }
967 :
968 : /// Record a change in layer residency.
969 : ///
970 : /// Recording the event must happen while holding the layer map lock to
971 : /// ensure that latest-activity-threshold-based layer eviction (eviction_task.rs)
972 : /// can do an "imitate access" to this layer, before it observes `now-latest_activity() > threshold`.
973 : ///
974 : /// If we instead recorded the residence event with a timestamp from before grabbing the layer map lock,
975 : /// the following race could happen:
976 : ///
977 : /// - Compact: Write out an L1 layer from several L0 layers. This records residence event LayerCreate with the current timestamp.
978 : /// - Eviction: imitate access logical size calculation. This accesses the L0 layers because the L1 layer is not yet in the layer map.
979 : /// - Compact: Grab layer map lock, add the new L1 to layer map and remove the L0s, release layer map lock.
980 : /// - Eviction: observes the new L1 layer whose only activity timestamp is the LayerCreate event.
981 100 : pub(crate) fn record_residence_event_at(&self, now: SystemTime) {
982 100 : let (mask, value) = Self::to_low_res_timestamp(Self::RTIME_SHIFT, now);
983 100 : self.write_bits(mask, value);
984 100 : }
985 :
986 96 : pub(crate) fn record_residence_event(&self) {
987 96 : self.record_residence_event_at(SystemTime::now())
988 96 : }
989 :
990 478880 : fn record_access_at(&self, now: SystemTime) -> bool {
991 478880 : let (mut mask, mut value) = Self::to_low_res_timestamp(Self::ATIME_SHIFT, now);
992 478880 :
993 478880 : // A layer which is accessed must be visible.
994 478880 : mask |= 0x1 << Self::VISIBILITY_SHIFT;
995 478880 : value |= 0x1 << Self::VISIBILITY_SHIFT;
996 478880 :
997 478880 : let old_bits = self.write_bits(mask, value);
998 4 : !matches!(
999 478880 : self.decode_visibility(old_bits),
1000 : LayerVisibilityHint::Visible
1001 : )
1002 478880 : }
1003 :
1004 : /// Returns true if we modified the layer's visibility to set it to Visible implicitly
1005 : /// as a result of this access
1006 479440 : pub(crate) fn record_access(&self, ctx: &RequestContext) -> bool {
1007 479440 : if ctx.access_stats_behavior() == AccessStatsBehavior::Skip {
1008 572 : return false;
1009 478868 : }
1010 478868 :
1011 478868 : self.record_access_at(SystemTime::now())
1012 479440 : }
1013 :
1014 0 : fn as_api_model(
1015 0 : &self,
1016 0 : reset: LayerAccessStatsReset,
1017 0 : ) -> pageserver_api::models::LayerAccessStats {
1018 0 : let ret = pageserver_api::models::LayerAccessStats {
1019 0 : access_time: self
1020 0 : .read_low_res_timestamp(Self::ATIME_SHIFT)
1021 0 : .unwrap_or(UNIX_EPOCH),
1022 0 : residence_time: self
1023 0 : .read_low_res_timestamp(Self::RTIME_SHIFT)
1024 0 : .unwrap_or(UNIX_EPOCH),
1025 0 : visible: matches!(self.visibility(), LayerVisibilityHint::Visible),
1026 : };
1027 0 : match reset {
1028 0 : LayerAccessStatsReset::NoReset => {}
1029 0 : LayerAccessStatsReset::AllStats => {
1030 0 : self.write_bits((Self::TS_MASK as u64) << Self::ATIME_SHIFT, 0x0);
1031 0 : self.write_bits((Self::TS_MASK as u64) << Self::RTIME_SHIFT, 0x0);
1032 0 : }
1033 : }
1034 0 : ret
1035 0 : }
1036 :
1037 : /// Get the latest access timestamp, falling back to latest residence event. The latest residence event
1038 : /// will be this Layer's construction time, if its residence hasn't changed since then.
1039 84 : pub(crate) fn latest_activity(&self) -> SystemTime {
1040 84 : if let Some(t) = self.read_low_res_timestamp(Self::ATIME_SHIFT) {
1041 12 : t
1042 : } else {
1043 72 : self.read_low_res_timestamp(Self::RTIME_SHIFT)
1044 72 : .expect("Residence time is set on construction")
1045 : }
1046 84 : }
1047 :
1048 : /// Whether this layer has been accessed (excluding in [`AccessStatsBehavior::Skip`]).
1049 : ///
1050 : /// This indicates whether the layer has been used for some purpose that would motivate
1051 : /// us to keep it on disk, such as for serving a getpage request.
1052 68 : fn accessed(&self) -> bool {
1053 68 : // Consider it accessed if the most recent access is more recent than
1054 68 : // the most recent change in residence status.
1055 68 : match (
1056 68 : self.read_low_res_timestamp(Self::ATIME_SHIFT),
1057 68 : self.read_low_res_timestamp(Self::RTIME_SHIFT),
1058 : ) {
1059 60 : (None, _) => false,
1060 0 : (Some(_), None) => true,
1061 8 : (Some(a), Some(r)) => a >= r,
1062 : }
1063 68 : }
1064 :
1065 : /// Helper for extracting the visibility hint from the literal value of our inner u64
1066 481206 : fn decode_visibility(&self, bits: u64) -> LayerVisibilityHint {
1067 481206 : match (bits >> Self::VISIBILITY_SHIFT) & 0x1 {
1068 481158 : 1 => LayerVisibilityHint::Visible,
1069 48 : 0 => LayerVisibilityHint::Covered,
1070 0 : _ => unreachable!(),
1071 : }
1072 481206 : }
1073 :
1074 : /// Returns the old value which has been replaced
1075 736 : pub(crate) fn set_visibility(&self, visibility: LayerVisibilityHint) -> LayerVisibilityHint {
1076 736 : let value = match visibility {
1077 632 : LayerVisibilityHint::Visible => 0x1 << Self::VISIBILITY_SHIFT,
1078 104 : LayerVisibilityHint::Covered => 0x0,
1079 : };
1080 :
1081 736 : let old_bits = self.write_bits(0x1 << Self::VISIBILITY_SHIFT, value);
1082 736 : self.decode_visibility(old_bits)
1083 736 : }
1084 :
1085 1590 : pub(crate) fn visibility(&self) -> LayerVisibilityHint {
1086 1590 : let read = self.0.load(std::sync::atomic::Ordering::Relaxed);
1087 1590 : self.decode_visibility(read)
1088 1590 : }
1089 : }
1090 :
1091 : /// Get a layer descriptor from a layer.
1092 : pub(crate) trait AsLayerDesc {
1093 : /// Get the layer descriptor.
1094 : fn layer_desc(&self) -> &PersistentLayerDesc;
1095 : }
1096 :
1097 : pub mod tests {
1098 : use pageserver_api::shard::TenantShardId;
1099 : use utils::id::TimelineId;
1100 :
1101 : use super::*;
1102 :
1103 : impl From<DeltaLayerName> for PersistentLayerDesc {
1104 44 : fn from(value: DeltaLayerName) -> Self {
1105 44 : PersistentLayerDesc::new_delta(
1106 44 : TenantShardId::from([0; 18]),
1107 44 : TimelineId::from_array([0; 16]),
1108 44 : value.key_range,
1109 44 : value.lsn_range,
1110 44 : 233,
1111 44 : )
1112 44 : }
1113 : }
1114 :
1115 : impl From<ImageLayerName> for PersistentLayerDesc {
1116 48 : fn from(value: ImageLayerName) -> Self {
1117 48 : PersistentLayerDesc::new_img(
1118 48 : TenantShardId::from([0; 18]),
1119 48 : TimelineId::from_array([0; 16]),
1120 48 : value.key_range,
1121 48 : value.lsn,
1122 48 : 233,
1123 48 : )
1124 48 : }
1125 : }
1126 :
1127 : impl From<LayerName> for PersistentLayerDesc {
1128 92 : fn from(value: LayerName) -> Self {
1129 92 : match value {
1130 44 : LayerName::Delta(d) => Self::from(d),
1131 48 : LayerName::Image(i) => Self::from(i),
1132 : }
1133 92 : }
1134 : }
1135 : }
1136 :
1137 : /// Range wrapping newtype, which uses display to render Debug.
1138 : ///
1139 : /// Useful with `Key`, which has too verbose `{:?}` for printing multiple layers.
1140 : struct RangeDisplayDebug<'a, T: std::fmt::Display>(&'a Range<T>);
1141 :
1142 : impl<T: std::fmt::Display> std::fmt::Debug for RangeDisplayDebug<'_, T> {
1143 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
1144 0 : write!(f, "{}..{}", self.0.start, self.0.end)
1145 0 : }
1146 : }
1147 :
1148 : #[cfg(test)]
1149 : mod tests2 {
1150 : use pageserver_api::key::DBDIR_KEY;
1151 : use tracing::info;
1152 :
1153 : use super::*;
1154 : use crate::tenant::storage_layer::IoConcurrency;
1155 :
1156 : /// TODO: currently this test relies on manual visual inspection of the --no-capture output.
1157 : /// Should look like so:
1158 : /// ```text
1159 : /// RUST_LOG=trace cargo nextest run --features testing --no-capture test_io_concurrency_noise
1160 : /// running 1 test
1161 : /// 2025-01-21T17:42:01.335679Z INFO get_vectored_concurrent_io test selected=SidecarTask
1162 : /// 2025-01-21T17:42:01.335680Z TRACE spawning sidecar task task_id=0
1163 : /// 2025-01-21T17:42:01.335937Z TRACE IoConcurrency_sidecar{task_id=0}: start
1164 : /// 2025-01-21T17:42:01.335972Z TRACE IoConcurrency_sidecar{task_id=0}: received new io future
1165 : /// 2025-01-21T17:42:01.335999Z INFO IoConcurrency_sidecar{task_id=0}: waiting for signal to complete IO
1166 : /// 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
1167 : /// at ./src/tenant/storage_layer.rs:553:24
1168 : /// 1: core::ptr::drop_in_place<pageserver::tenant::storage_layer::ValuesReconstructState>
1169 : /// 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
1170 : /// 2: core::mem::drop
1171 : /// 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
1172 : /// 3: pageserver::tenant::storage_layer::tests2::test_io_concurrency_noise::{{closure}}
1173 : /// at ./src/tenant/storage_layer.rs:1159:9
1174 : /// ...
1175 : /// 49: <unknown>
1176 : /// 2025-01-21T17:42:01.452293Z INFO IoConcurrency_sidecar{task_id=0}: completing IO
1177 : /// 2025-01-21T17:42:01.452357Z TRACE IoConcurrency_sidecar{task_id=0}: io future completed
1178 : /// 2025-01-21T17:42:01.452473Z TRACE IoConcurrency_sidecar{task_id=0}: end
1179 : /// test tenant::storage_layer::tests2::test_io_concurrency_noise ... ok
1180 : ///
1181 : /// ```
1182 : #[tokio::test]
1183 4 : async fn test_io_concurrency_noise() {
1184 4 : crate::tenant::harness::setup_logging();
1185 4 :
1186 4 : let io_concurrency = IoConcurrency::spawn_for_test();
1187 4 : match *io_concurrency {
1188 4 : IoConcurrency::Sequential => {
1189 4 : // This test asserts behavior in sidecar mode, doesn't make sense in sequential mode.
1190 4 : return;
1191 4 : }
1192 4 : IoConcurrency::SidecarTask { .. } => {}
1193 2 : }
1194 2 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
1195 2 :
1196 2 : let (io_fut_is_waiting_tx, io_fut_is_waiting) = tokio::sync::oneshot::channel();
1197 2 : let (do_complete_io, should_complete_io) = tokio::sync::oneshot::channel();
1198 2 : let (io_fut_exiting_tx, io_fut_exiting) = tokio::sync::oneshot::channel();
1199 2 :
1200 2 : let io = reconstruct_state.update_key(&DBDIR_KEY, Lsn(8), true);
1201 2 : reconstruct_state
1202 2 : .spawn_io(async move {
1203 2 : info!("waiting for signal to complete IO");
1204 4 : io_fut_is_waiting_tx.send(()).unwrap();
1205 2 : should_complete_io.await.unwrap();
1206 2 : info!("completing IO");
1207 4 : io.complete(Ok(OnDiskValue::RawImage(Bytes::new())));
1208 2 : io_fut_exiting_tx.send(()).unwrap();
1209 2 : })
1210 2 : .await;
1211 4 :
1212 4 : io_fut_is_waiting.await.unwrap();
1213 2 :
1214 2 : // this is what makes the noise
1215 2 : drop(reconstruct_state);
1216 2 :
1217 2 : do_complete_io.send(()).unwrap();
1218 2 :
1219 2 : io_fut_exiting.await.unwrap();
1220 4 : }
1221 : }
|