Line data Source code
1 : use std::ops::ControlFlow;
2 :
3 : use tokio_util::sync::CancellationToken;
4 : use tracing::{Instrument, info_span, warn};
5 : use utils::sync::duplex;
6 :
7 : use super::{Buffer, CheapCloneForRead, OwnedAsyncWriter};
8 : use crate::context::RequestContext;
9 : use crate::virtual_file::MaybeFatalIo;
10 : use crate::virtual_file::owned_buffers_io::io_buf_aligned::IoBufAligned;
11 : use crate::virtual_file::owned_buffers_io::io_buf_ext::FullSlice;
12 :
13 : /// A handle to the flush task.
14 : pub struct FlushHandle<Buf, W> {
15 : inner: Option<FlushHandleInner<Buf, W>>,
16 : }
17 :
18 : pub struct FlushHandleInner<Buf, W> {
19 : /// A bi-directional channel that sends (buffer, offset) for writes,
20 : /// and receives recyled buffer.
21 : channel: duplex::mpsc::Duplex<Request<Buf>, FullSlice<Buf>>,
22 : /// Join handle for the background flush task.
23 : join_handle: tokio::task::JoinHandle<Result<W, FlushTaskError>>,
24 : }
25 :
26 : struct FlushRequest<Buf> {
27 : slice: FullSlice<Buf>,
28 : offset: u64,
29 : #[cfg(test)]
30 : ready_to_flush_rx: Option<tokio::sync::oneshot::Receiver<()>>,
31 : #[cfg(test)]
32 : done_flush_tx: Option<tokio::sync::oneshot::Sender<()>>,
33 : }
34 :
35 : pub struct ShutdownRequest {
36 : pub set_len: Option<u64>,
37 : }
38 :
39 : enum Request<Buf> {
40 : Flush(FlushRequest<Buf>),
41 : Shutdown(ShutdownRequest),
42 : }
43 :
44 : impl<Buf> Request<Buf> {
45 152904 : fn op_str(&self) -> &'static str {
46 152904 : match self {
47 141648 : Request::Flush(_) => "flush",
48 11256 : Request::Shutdown(_) => "shutdown",
49 : }
50 152904 : }
51 : }
52 :
53 : /// Constructs a request and a control object for a new flush operation.
54 : #[cfg(not(test))]
55 0 : fn new_flush_op<Buf>(slice: FullSlice<Buf>, offset: u64) -> (FlushRequest<Buf>, FlushControl) {
56 0 : let request = FlushRequest { slice, offset };
57 0 : let control = FlushControl::untracked();
58 0 :
59 0 : (request, control)
60 0 : }
61 :
62 : /// Constructs a request and a control object for a new flush operation.
63 : #[cfg(test)]
64 141648 : fn new_flush_op<Buf>(slice: FullSlice<Buf>, offset: u64) -> (FlushRequest<Buf>, FlushControl) {
65 141648 : let (ready_to_flush_tx, ready_to_flush_rx) = tokio::sync::oneshot::channel();
66 141648 : let (done_flush_tx, done_flush_rx) = tokio::sync::oneshot::channel();
67 141648 : let control = FlushControl::not_started(ready_to_flush_tx, done_flush_rx);
68 141648 :
69 141648 : let request = FlushRequest {
70 141648 : slice,
71 141648 : offset,
72 141648 : ready_to_flush_rx: Some(ready_to_flush_rx),
73 141648 : done_flush_tx: Some(done_flush_tx),
74 141648 : };
75 141648 : (request, control)
76 141648 : }
77 :
78 : /// A handle to a `FlushRequest` that allows unit tests precise control over flush behavior.
79 : #[cfg(test)]
80 : pub(crate) struct FlushControl {
81 : not_started: FlushNotStarted,
82 : }
83 :
84 : #[cfg(not(test))]
85 : pub(crate) struct FlushControl;
86 :
87 : impl FlushControl {
88 : #[cfg(test)]
89 141648 : fn not_started(
90 141648 : ready_to_flush_tx: tokio::sync::oneshot::Sender<()>,
91 141648 : done_flush_rx: tokio::sync::oneshot::Receiver<()>,
92 141648 : ) -> Self {
93 141648 : FlushControl {
94 141648 : not_started: FlushNotStarted {
95 141648 : ready_to_flush_tx,
96 141648 : done_flush_rx,
97 141648 : },
98 141648 : }
99 141648 : }
100 :
101 : #[cfg(not(test))]
102 0 : fn untracked() -> Self {
103 0 : FlushControl
104 0 : }
105 :
106 : /// In tests, turn flush control into a not started state.
107 : #[cfg(test)]
108 12 : pub(crate) fn into_not_started(self) -> FlushNotStarted {
109 12 : self.not_started
110 12 : }
111 :
112 : /// Release control to the submitted buffer.
113 : ///
114 : /// In `cfg(test)` environment, the buffer is guranteed to be flushed to disk after [`FlushControl::release`] is finishes execution.
115 130404 : pub async fn release(self) {
116 130404 : #[cfg(test)]
117 130404 : {
118 130404 : self.not_started
119 130404 : .ready_to_flush()
120 130404 : .wait_until_flush_is_done()
121 130404 : .await;
122 0 : }
123 130404 : }
124 : }
125 :
126 : impl<Buf, W> FlushHandle<Buf, W>
127 : where
128 : Buf: IoBufAligned + Send + Sync + CheapCloneForRead,
129 : W: OwnedAsyncWriter + Send + Sync + 'static + std::fmt::Debug,
130 : {
131 : /// Spawns a new background flush task and obtains a handle.
132 : ///
133 : /// Handle and background task are connected through a duplex channel.
134 : /// Dirty buffers are sent to the background task for flushing.
135 : /// Clean buffers are sent back to the handle for reuse.
136 : ///
137 : /// The queue depth is 1, and the passed-in `buf` seeds the queue depth.
138 : /// I.e., the passed-in buf is immediately available to the handle as a recycled buffer.
139 20832 : pub fn spawn_new<B>(
140 20832 : file: W,
141 20832 : buf: B,
142 20832 : gate_guard: utils::sync::gate::GateGuard,
143 20832 : cancel: CancellationToken,
144 20832 : ctx: RequestContext,
145 20832 : span: tracing::Span,
146 20832 : ) -> Self
147 20832 : where
148 20832 : B: Buffer<IoBuf = Buf> + Send + 'static,
149 20832 : {
150 20832 : let (front, back) = duplex::mpsc::channel(1);
151 20832 : back.try_send(buf.flush())
152 20832 : .expect("we just created it with capacity 1");
153 20832 :
154 20832 : let join_handle = tokio::spawn(
155 20832 : FlushBackgroundTask::new(back, file, gate_guard, cancel, ctx)
156 20832 : .run()
157 20832 : .instrument(span),
158 20832 : );
159 20832 :
160 20832 : FlushHandle {
161 20832 : inner: Some(FlushHandleInner {
162 20832 : channel: front,
163 20832 : join_handle,
164 20832 : }),
165 20832 : }
166 20832 : }
167 :
168 : /// Submits a buffer to be flushed in the background task.
169 : /// Returns a buffer that completed flushing for re-use, length reset to 0, capacity unchanged.
170 : /// If `save_buf_for_read` is true, then we save the buffer in `Self::maybe_flushed`, otherwise
171 : /// clear `maybe_flushed`.
172 141648 : pub async fn flush(
173 141648 : &mut self,
174 141648 : slice: FullSlice<Buf>,
175 141648 : offset: u64,
176 141648 : ) -> Result<(FullSlice<Buf>, FlushControl), FlushTaskError> {
177 141648 : let (request, flush_control) = new_flush_op(slice, offset);
178 141648 :
179 141648 : // Submits the buffer to the background task.
180 141648 : self.send(Request::Flush(request)).await?;
181 :
182 : // Wait for an available buffer from the background flush task.
183 : // This is the BACKPRESSURE mechanism: if the flush task can't keep up,
184 : // then the write path will eventually wait for it here.
185 141648 : let Some(recycled) = self.inner_mut().channel.recv().await else {
186 0 : return self.handle_error().await;
187 : };
188 :
189 141648 : Ok((recycled, flush_control))
190 141648 : }
191 :
192 : /// Sends poison pill to flush task and waits for it to exit.
193 11256 : pub async fn shutdown(&mut self, req: ShutdownRequest) -> Result<W, FlushTaskError> {
194 11256 : self.send(Request::Shutdown(req)).await?;
195 11256 : self.wait().await
196 11256 : }
197 :
198 152904 : async fn send(&mut self, request: Request<Buf>) -> Result<(), FlushTaskError> {
199 152904 : let submit = self.inner_mut().channel.send(request).await;
200 152904 : if submit.is_err() {
201 0 : return self.handle_error().await;
202 152904 : }
203 152904 : Ok(())
204 152904 : }
205 :
206 0 : async fn handle_error<T>(&mut self) -> Result<T, FlushTaskError> {
207 0 : Err(self
208 0 : .wait()
209 0 : .await
210 0 : .expect_err("flush task only disconnects duplex if it exits with an error"))
211 0 : }
212 :
213 11256 : async fn wait(&mut self) -> Result<W, FlushTaskError> {
214 11256 : let handle = self
215 11256 : .inner
216 11256 : .take()
217 11256 : .expect("must not use after we returned an error");
218 11256 : drop(handle.channel.tx);
219 11256 : handle.join_handle.await.unwrap()
220 11256 : }
221 :
222 : /// Gets a mutable reference to the inner handle. Panics if [`Self::inner`] is `None`.
223 : /// This only happens if the handle is used after an error.
224 294552 : fn inner_mut(&mut self) -> &mut FlushHandleInner<Buf, W> {
225 294552 : self.inner
226 294552 : .as_mut()
227 294552 : .expect("must not use after we returned an error")
228 294552 : }
229 : }
230 :
231 : /// A background task for flushing data to disk.
232 : pub struct FlushBackgroundTask<Buf, W> {
233 : /// A bi-directional channel that receives (buffer, offset) for writes,
234 : /// and send back recycled buffer.
235 : channel: duplex::mpsc::Duplex<FullSlice<Buf>, Request<Buf>>,
236 : /// A writter for persisting data to disk.
237 : writer: W,
238 : ctx: RequestContext,
239 : cancel: CancellationToken,
240 : /// Prevent timeline from shuting down until the flush background task finishes flushing all remaining buffers to disk.
241 : _gate_guard: utils::sync::gate::GateGuard,
242 : }
243 :
244 : #[derive(Debug, thiserror::Error)]
245 : pub enum FlushTaskError {
246 : #[error("flush task cancelled")]
247 : Cancelled,
248 : }
249 :
250 : impl<Buf, W> FlushBackgroundTask<Buf, W>
251 : where
252 : Buf: IoBufAligned + Send + Sync,
253 : W: OwnedAsyncWriter + Sync + 'static,
254 : {
255 : /// Creates a new background flush task.
256 20832 : fn new(
257 20832 : channel: duplex::mpsc::Duplex<FullSlice<Buf>, Request<Buf>>,
258 20832 : file: W,
259 20832 : gate_guard: utils::sync::gate::GateGuard,
260 20832 : cancel: CancellationToken,
261 20832 : ctx: RequestContext,
262 20832 : ) -> Self {
263 20832 : FlushBackgroundTask {
264 20832 : channel,
265 20832 : writer: file,
266 20832 : _gate_guard: gate_guard,
267 20832 : cancel,
268 20832 : ctx,
269 20832 : }
270 20832 : }
271 :
272 : /// Runs the background flush task.
273 20832 : async fn run(mut self) -> Result<W, FlushTaskError> {
274 : // Exit condition: channel is closed and there is no remaining buffer to be flushed
275 173658 : while let Some(request) = self.channel.recv().await {
276 152904 : let op_kind = request.op_str();
277 152904 :
278 152904 : // Perform the requested operation.
279 152904 : //
280 152904 : // Error handling happens according to the current policy of crashing
281 152904 : // on fatal IO errors and retrying in place otherwise (deeming all other errors retryable).
282 152904 : // (The upper layers of the Pageserver write path are not equipped to retry write errors
283 152904 : // becasuse they often deallocate the buffers that were already written).
284 152904 : //
285 152904 : // TODO: use utils::backoff::retry once async closures are actually usable
286 152904 : //
287 152904 : let mut request_storage = Some(request);
288 152904 : for attempt in 1.. {
289 152904 : if self.cancel.is_cancelled() {
290 0 : return Err(FlushTaskError::Cancelled);
291 152904 : }
292 152904 : let result = async {
293 152904 : let request: Request<Buf> = request_storage .take().expect(
294 152904 : "likely previous invocation of this future didn't get polled to completion",
295 152904 : );
296 11256 : match &request {
297 : Request::Shutdown(ShutdownRequest { set_len: None }) => {
298 11160 : request_storage = Some(request);
299 11160 : return ControlFlow::Break(());
300 : },
301 141744 : Request::Flush(_) | Request::Shutdown(ShutdownRequest { set_len: Some(_) }) => {
302 141744 : },
303 141744 : }
304 141744 : if attempt > 1 {
305 0 : warn!(op=%request.op_str(), "retrying");
306 141744 : }
307 : // borrows so we can async move the requests into async block while not moving these borrows here
308 141744 : let writer = &self.writer;
309 141744 : let request_storage = &mut request_storage;
310 141744 : let ctx = &self.ctx;
311 141744 : let io_fut = match request {
312 141648 : Request::Flush(FlushRequest { slice, offset, #[cfg(test)] ready_to_flush_rx, #[cfg(test)] done_flush_tx }) => futures::future::Either::Left(async move {
313 : #[cfg(test)]
314 141648 : if let Some(ready_to_flush_rx) = ready_to_flush_rx {
315 : {
316 : // In test, wait for control to signal that we are ready to flush.
317 141648 : if ready_to_flush_rx.await.is_err() {
318 11232 : tracing::debug!("control dropped");
319 130416 : }
320 : }
321 0 : }
322 141648 : let (slice, res) = writer.write_all_at(slice, offset, ctx).await;
323 141648 : *request_storage = Some(Request::Flush(FlushRequest {
324 141648 : slice,
325 141648 : offset,
326 141648 : #[cfg(test)]
327 141648 : ready_to_flush_rx: None, // the contract is that we notify before first attempt
328 141648 : #[cfg(test)]
329 141648 : done_flush_tx
330 141648 : }));
331 141648 : res
332 141648 : }),
333 96 : Request::Shutdown(ShutdownRequest { set_len }) => futures::future::Either::Right(async move {
334 96 : let set_len = set_len.expect("we filter out the None case above");
335 96 : let res = writer.set_len(set_len, ctx).await;
336 96 : *request_storage = Some(Request::Shutdown(ShutdownRequest {
337 96 : set_len: Some(set_len),
338 96 : }));
339 96 : res
340 96 : }),
341 : };
342 : // Don't cancel the io_fut by doing tokio::select with self.cancel.cancelled().
343 : // The underlying tokio-epoll-uring slot / kernel operation is still ongoing and occupies resources.
344 : // If we retry indefinitely, we'll deplete those resources.
345 : // Future: teach tokio-epoll-uring io_uring operation cancellation, but still,
346 : // wait for cancelled ops to complete and discard their error.
347 141744 : let res = io_fut.await;
348 141744 : let res = res.maybe_fatal_err("owned_buffers_io flush");
349 141744 : let Err(err) = res else {
350 141744 : if attempt > 1 {
351 0 : warn!(op=%op_kind, "retry succeeded");
352 141744 : }
353 141744 : return ControlFlow::Break(());
354 : };
355 0 : warn!(%err, "error flushing buffered writer buffer to disk, retrying after backoff");
356 0 : utils::backoff::exponential_backoff(attempt, 1.0, 10.0, &self.cancel).await;
357 0 : ControlFlow::Continue(())
358 152904 : }
359 152904 : .instrument(info_span!("attempt", %attempt, %op_kind))
360 152904 : .await;
361 152904 : match result {
362 152904 : ControlFlow::Break(()) => break,
363 0 : ControlFlow::Continue(()) => continue,
364 : }
365 : }
366 152904 : let request = request_storage.expect("loop must have run at least once");
367 :
368 152904 : let slice = match request {
369 : Request::Flush(FlushRequest {
370 141648 : slice,
371 141648 : #[cfg(test)]
372 141648 : mut done_flush_tx,
373 141648 : ..
374 141648 : }) => {
375 141648 : #[cfg(test)]
376 141648 : {
377 141648 : // In test, tell control we are done flushing buffer.
378 141648 : if done_flush_tx.take().expect("always Some").send(()).is_err() {
379 11162 : tracing::debug!("control dropped");
380 130486 : }
381 0 : }
382 141648 : slice
383 : }
384 : Request::Shutdown(_) => {
385 : // next iteration will observe recv() returning None
386 11256 : continue;
387 : }
388 : };
389 :
390 : // Sends the buffer back to the handle for reuse. The handle is in charged of cleaning the buffer.
391 141648 : let send_res = self.channel.send(slice).await;
392 141648 : if send_res.is_err() {
393 : // Although channel is closed. Still need to finish flushing the remaining buffers.
394 0 : continue;
395 141648 : }
396 : }
397 :
398 19859 : Ok(self.writer)
399 19859 : }
400 : }
401 :
402 : #[cfg(test)]
403 : pub(crate) struct FlushNotStarted {
404 : ready_to_flush_tx: tokio::sync::oneshot::Sender<()>,
405 : done_flush_rx: tokio::sync::oneshot::Receiver<()>,
406 : }
407 :
408 : #[cfg(test)]
409 : pub(crate) struct FlushInProgress {
410 : done_flush_rx: tokio::sync::oneshot::Receiver<()>,
411 : }
412 :
413 : #[cfg(test)]
414 : pub(crate) struct FlushDone;
415 :
416 : #[cfg(test)]
417 : impl FlushNotStarted {
418 : /// Signals the background task the buffer is ready to flush to disk.
419 130416 : pub fn ready_to_flush(self) -> FlushInProgress {
420 130416 : self.ready_to_flush_tx
421 130416 : .send(())
422 130416 : .map(|_| FlushInProgress {
423 130416 : done_flush_rx: self.done_flush_rx,
424 130416 : })
425 130416 : .unwrap()
426 130416 : }
427 : }
428 :
429 : #[cfg(test)]
430 : impl FlushInProgress {
431 : /// Waits until background flush is done.
432 130416 : pub async fn wait_until_flush_is_done(self) -> FlushDone {
433 130416 : self.done_flush_rx.await.unwrap();
434 130416 : FlushDone
435 130416 : }
436 : }
|
