Line data Source code
1 : //!
2 : //! VirtualFile is like a normal File, but it's not bound directly to
3 : //! a file descriptor. Instead, the file is opened when it's read from,
4 : //! and if too many files are open globally in the system, least-recently
5 : //! used ones are closed.
6 : //!
7 : //! To track which files have been recently used, we use the clock algorithm
8 : //! with a 'recently_used' flag on each slot.
9 : //!
10 : //! This is similar to PostgreSQL's virtual file descriptor facility in
11 : //! src/backend/storage/file/fd.c
12 : //!
13 : use crate::metrics::{StorageIoOperation, STORAGE_IO_SIZE, STORAGE_IO_TIME_METRIC};
14 :
15 : use crate::page_cache::PageWriteGuard;
16 : use crate::tenant::TENANTS_SEGMENT_NAME;
17 : use camino::{Utf8Path, Utf8PathBuf};
18 : use once_cell::sync::OnceCell;
19 : use pageserver_api::shard::TenantShardId;
20 : use std::fs::{self, File};
21 : use std::io::{Error, ErrorKind, Seek, SeekFrom};
22 : use tokio_epoll_uring::{BoundedBuf, IoBuf, IoBufMut, Slice};
23 :
24 : use std::os::fd::{AsRawFd, FromRawFd, IntoRawFd, OwnedFd, RawFd};
25 : use std::os::unix::fs::FileExt;
26 : use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
27 : use tokio::sync::{RwLock, RwLockReadGuard, RwLockWriteGuard};
28 : use tokio::time::Instant;
29 :
30 : pub use pageserver_api::models::virtual_file as api;
31 : pub(crate) mod io_engine;
32 : mod open_options;
33 : pub(crate) use io_engine::IoEngineKind;
34 : pub(crate) use open_options::*;
35 :
36 : ///
37 : /// A virtual file descriptor. You can use this just like std::fs::File, but internally
38 : /// the underlying file is closed if the system is low on file descriptors,
39 : /// and re-opened when it's accessed again.
40 : ///
41 : /// Like with std::fs::File, multiple threads can read/write the file concurrently,
42 : /// holding just a shared reference the same VirtualFile, using the read_at() / write_at()
43 : /// functions from the FileExt trait. But the functions from the Read/Write/Seek traits
44 : /// require a mutable reference, because they modify the "current position".
45 : ///
46 : /// Each VirtualFile has a physical file descriptor in the global OPEN_FILES array, at the
47 : /// slot that 'handle points to, if the underlying file is currently open. If it's not
48 : /// currently open, the 'handle' can still point to the slot where it was last kept. The
49 : /// 'tag' field is used to detect whether the handle still is valid or not.
50 : ///
51 0 : #[derive(Debug)]
52 : pub struct VirtualFile {
53 : /// Lazy handle to the global file descriptor cache. The slot that this points to
54 : /// might contain our File, or it may be empty, or it may contain a File that
55 : /// belongs to a different VirtualFile.
56 : handle: RwLock<SlotHandle>,
57 :
58 : /// Current file position
59 : pos: u64,
60 :
61 : /// File path and options to use to open it.
62 : ///
63 : /// Note: this only contains the options needed to re-open it. For example,
64 : /// if a new file is created, we only pass the create flag when it's initially
65 : /// opened, in the VirtualFile::create() function, and strip the flag before
66 : /// storing it here.
67 : pub path: Utf8PathBuf,
68 : open_options: OpenOptions,
69 :
70 : // These are strings becase we only use them for metrics, and those expect strings.
71 : // It makes no sense for us to constantly turn the `TimelineId` and `TenantId` into
72 : // strings.
73 : tenant_id: String,
74 : shard_id: String,
75 : timeline_id: String,
76 : }
77 :
78 304215 : #[derive(Debug, PartialEq, Clone, Copy)]
79 : struct SlotHandle {
80 : /// Index into OPEN_FILES.slots
81 : index: usize,
82 :
83 : /// Value of 'tag' in the slot. If slot's tag doesn't match, then the slot has
84 : /// been recycled and no longer contains the FD for this virtual file.
85 : tag: u64,
86 : }
87 :
88 : /// OPEN_FILES is the global array that holds the physical file descriptors that
89 : /// are currently open. Each slot in the array is protected by a separate lock,
90 : /// so that different files can be accessed independently. The lock must be held
91 : /// in write mode to replace the slot with a different file, but a read mode
92 : /// is enough to operate on the file, whether you're reading or writing to it.
93 : ///
94 : /// OPEN_FILES starts in uninitialized state, and it's initialized by
95 : /// the virtual_file::init() function. It must be called exactly once at page
96 : /// server startup.
97 : static OPEN_FILES: OnceCell<OpenFiles> = OnceCell::new();
98 :
99 : struct OpenFiles {
100 : slots: &'static [Slot],
101 :
102 : /// clock arm for the clock algorithm
103 : next: AtomicUsize,
104 : }
105 :
106 : struct Slot {
107 : inner: RwLock<SlotInner>,
108 :
109 : /// has this file been used since last clock sweep?
110 : recently_used: AtomicBool,
111 : }
112 :
113 : struct SlotInner {
114 : /// Counter that's incremented every time a different file is stored here.
115 : /// To avoid the ABA problem.
116 : tag: u64,
117 :
118 : /// the underlying file
119 : file: Option<OwnedFd>,
120 : }
121 :
122 : /// Impl of [`tokio_epoll_uring::IoBuf`] and [`tokio_epoll_uring::IoBufMut`] for [`PageWriteGuard`].
123 : struct PageWriteGuardBuf {
124 : page: PageWriteGuard<'static>,
125 : init_up_to: usize,
126 : }
127 : // Safety: the [`PageWriteGuard`] gives us exclusive ownership of the page cache slot,
128 : // and the location remains stable even if [`Self`] or the [`PageWriteGuard`] is moved.
129 : unsafe impl tokio_epoll_uring::IoBuf for PageWriteGuardBuf {
130 4925336 : fn stable_ptr(&self) -> *const u8 {
131 4925336 : self.page.as_ptr()
132 4925336 : }
133 14776009 : fn bytes_init(&self) -> usize {
134 14776009 : self.init_up_to
135 14776009 : }
136 4925338 : fn bytes_total(&self) -> usize {
137 4925338 : self.page.len()
138 4925338 : }
139 : }
140 : // Safety: see above, plus: the ownership of [`PageWriteGuard`] means exclusive access,
141 : // hence it's safe to hand out the `stable_mut_ptr()`.
142 : unsafe impl tokio_epoll_uring::IoBufMut for PageWriteGuardBuf {
143 4925337 : fn stable_mut_ptr(&mut self) -> *mut u8 {
144 4925337 : self.page.as_mut_ptr()
145 4925337 : }
146 :
147 4925337 : unsafe fn set_init(&mut self, pos: usize) {
148 4925337 : assert!(pos <= self.page.len());
149 4925337 : self.init_up_to = pos;
150 4925337 : }
151 : }
152 :
153 : impl OpenFiles {
154 : /// Find a slot to use, evicting an existing file descriptor if needed.
155 : ///
156 : /// On return, we hold a lock on the slot, and its 'tag' has been updated
157 : /// recently_used has been set. It's all ready for reuse.
158 259472 : async fn find_victim_slot(&self) -> (SlotHandle, RwLockWriteGuard<SlotInner>) {
159 259472 : //
160 259472 : // Run the clock algorithm to find a slot to replace.
161 259472 : //
162 259472 : let num_slots = self.slots.len();
163 259472 : let mut retries = 0;
164 : let mut slot;
165 : let mut slot_guard;
166 : let index;
167 2787480 : loop {
168 2787480 : let next = self.next.fetch_add(1, Ordering::AcqRel) % num_slots;
169 2787480 : slot = &self.slots[next];
170 2787480 :
171 2787480 : // If the recently_used flag on this slot is set, continue the clock
172 2787480 : // sweep. Otherwise try to use this slot. If we cannot acquire the
173 2787480 : // lock, also continue the clock sweep.
174 2787480 : //
175 2787480 : // We only continue in this manner for a while, though. If we loop
176 2787480 : // through the array twice without finding a victim, just pick the
177 2787480 : // next slot and wait until we can reuse it. This way, we avoid
178 2787480 : // spinning in the extreme case that all the slots are busy with an
179 2787480 : // I/O operation.
180 2787480 : if retries < num_slots * 2 {
181 2682253 : if !slot.recently_used.swap(false, Ordering::Release) {
182 2435337 : if let Ok(guard) = slot.inner.try_write() {
183 154245 : slot_guard = guard;
184 154245 : index = next;
185 154245 : break;
186 2281092 : }
187 246916 : }
188 2528008 : retries += 1;
189 : } else {
190 105227 : slot_guard = slot.inner.write().await;
191 105227 : index = next;
192 105227 : break;
193 : }
194 : }
195 :
196 : //
197 : // We now have the victim slot locked. If it was in use previously, close the
198 : // old file.
199 : //
200 259472 : if let Some(old_file) = slot_guard.file.take() {
201 216342 : // the normal path of dropping VirtualFile uses "close", use "close-by-replace" here to
202 216342 : // distinguish the two.
203 216342 : STORAGE_IO_TIME_METRIC
204 216342 : .get(StorageIoOperation::CloseByReplace)
205 216342 : .observe_closure_duration(|| drop(old_file));
206 216342 : }
207 :
208 : // Prepare the slot for reuse and return it
209 259472 : slot_guard.tag += 1;
210 259472 : slot.recently_used.store(true, Ordering::Relaxed);
211 259472 : (
212 259472 : SlotHandle {
213 259472 : index,
214 259472 : tag: slot_guard.tag,
215 259472 : },
216 259472 : slot_guard,
217 259472 : )
218 259472 : }
219 : }
220 :
221 : /// Identify error types that should alwways terminate the process. Other
222 : /// error types may be elegible for retry.
223 0 : pub(crate) fn is_fatal_io_error(e: &std::io::Error) -> bool {
224 0 : use nix::errno::Errno::*;
225 0 : match e.raw_os_error().map(nix::errno::from_i32) {
226 : Some(EIO) => {
227 : // Terminate on EIO because we no longer trust the device to store
228 : // data safely, or to uphold persistence guarantees on fsync.
229 0 : true
230 : }
231 : Some(EROFS) => {
232 : // Terminate on EROFS because a filesystem is usually remounted
233 : // readonly when it has experienced some critical issue, so the same
234 : // logic as EIO applies.
235 0 : true
236 : }
237 : Some(EACCES) => {
238 : // Terminate on EACCESS because we should always have permissions
239 : // for our own data dir: if we don't, then we can't do our job and
240 : // need administrative intervention to fix permissions. Terminating
241 : // is the best way to make sure we stop cleanly rather than going
242 : // into infinite retry loops, and will make it clear to the outside
243 : // world that we need help.
244 0 : true
245 : }
246 : _ => {
247 : // Treat all other local file I/O errors are retryable. This includes:
248 : // - ENOSPC: we stay up and wait for eviction to free some space
249 : // - EINVAL, EBADF, EBADFD: this is a code bug, not a filesystem/hardware issue
250 : // - WriteZero, Interrupted: these are used internally VirtualFile
251 0 : false
252 : }
253 : }
254 0 : }
255 :
256 : /// Call this when the local filesystem gives us an error with an external
257 : /// cause: this includes EIO, EROFS, and EACCESS: all these indicate either
258 : /// bad storage or bad configuration, and we can't fix that from inside
259 : /// a running process.
260 0 : pub(crate) fn on_fatal_io_error(e: &std::io::Error, context: &str) -> ! {
261 0 : tracing::error!("Fatal I/O error: {e}: {context})");
262 0 : std::process::abort();
263 : }
264 :
265 : pub(crate) trait MaybeFatalIo<T> {
266 : fn maybe_fatal_err(self, context: &str) -> std::io::Result<T>;
267 : fn fatal_err(self, context: &str) -> T;
268 : }
269 :
270 : impl<T> MaybeFatalIo<T> for std::io::Result<T> {
271 : /// Terminate the process if the result is an error of a fatal type, else pass it through
272 : ///
273 : /// This is appropriate for writes, where we typically want to die on EIO/ACCES etc, but
274 : /// not on ENOSPC.
275 115 : fn maybe_fatal_err(self, context: &str) -> std::io::Result<T> {
276 115 : if let Err(e) = &self {
277 0 : if is_fatal_io_error(e) {
278 0 : on_fatal_io_error(e, context);
279 0 : }
280 115 : }
281 115 : self
282 115 : }
283 :
284 : /// Terminate the process on any I/O error.
285 : ///
286 : /// This is appropriate for reads on files that we know exist: they should always work.
287 1401 : fn fatal_err(self, context: &str) -> T {
288 1401 : match self {
289 1401 : Ok(v) => v,
290 0 : Err(e) => {
291 0 : on_fatal_io_error(&e, context);
292 : }
293 : }
294 1401 : }
295 : }
296 :
297 : /// Observe duration for the given storage I/O operation
298 : ///
299 : /// Unlike `observe_closure_duration`, this supports async,
300 : /// where "support" means that we measure wall clock time.
301 : macro_rules! observe_duration {
302 : ($op:expr, $($body:tt)*) => {{
303 : let instant = Instant::now();
304 : let result = $($body)*;
305 : let elapsed = instant.elapsed().as_secs_f64();
306 : STORAGE_IO_TIME_METRIC
307 : .get($op)
308 : .observe(elapsed);
309 : result
310 : }}
311 : }
312 :
313 : macro_rules! with_file {
314 : ($this:expr, $op:expr, | $ident:ident | $($body:tt)*) => {{
315 : let $ident = $this.lock_file().await?;
316 : observe_duration!($op, $($body)*)
317 : }};
318 : ($this:expr, $op:expr, | mut $ident:ident | $($body:tt)*) => {{
319 : let mut $ident = $this.lock_file().await?;
320 : observe_duration!($op, $($body)*)
321 : }};
322 : }
323 :
324 : impl VirtualFile {
325 : /// Open a file in read-only mode. Like File::open.
326 33410 : pub async fn open(path: &Utf8Path) -> Result<VirtualFile, std::io::Error> {
327 33410 : Self::open_with_options(path, OpenOptions::new().read(true)).await
328 33410 : }
329 :
330 : /// Create a new file for writing. If the file exists, it will be truncated.
331 : /// Like File::create.
332 15857 : pub async fn create(path: &Utf8Path) -> Result<VirtualFile, std::io::Error> {
333 15857 : Self::open_with_options(
334 15857 : path,
335 15857 : OpenOptions::new().write(true).create(true).truncate(true),
336 15857 : )
337 232 : .await
338 15857 : }
339 :
340 : /// Open a file with given options.
341 : ///
342 : /// Note: If any custom flags were set in 'open_options' through OpenOptionsExt,
343 : /// they will be applied also when the file is subsequently re-opened, not only
344 : /// on the first time. Make sure that's sane!
345 61790 : pub async fn open_with_options(
346 61790 : path: &Utf8Path,
347 61790 : open_options: &OpenOptions,
348 61790 : ) -> Result<VirtualFile, std::io::Error> {
349 61790 : let path_str = path.to_string();
350 61790 : let parts = path_str.split('/').collect::<Vec<&str>>();
351 61790 : let (tenant_id, shard_id, timeline_id) =
352 61790 : if parts.len() > 5 && parts[parts.len() - 5] == TENANTS_SEGMENT_NAME {
353 61330 : let tenant_shard_part = parts[parts.len() - 4];
354 61330 : let (tenant_id, shard_id) = match tenant_shard_part.parse::<TenantShardId>() {
355 61330 : Ok(tenant_shard_id) => (
356 61330 : tenant_shard_id.tenant_id.to_string(),
357 61330 : format!("{}", tenant_shard_id.shard_slug()),
358 61330 : ),
359 : Err(_) => {
360 : // Malformed path: this ID is just for observability, so tolerate it
361 : // and pass through
362 0 : (tenant_shard_part.to_string(), "*".to_string())
363 : }
364 : };
365 61330 : (tenant_id, shard_id, parts[parts.len() - 2].to_string())
366 : } else {
367 460 : ("*".to_string(), "*".to_string(), "*".to_string())
368 : };
369 61790 : let (handle, mut slot_guard) = get_open_files().find_victim_slot().await;
370 :
371 : // NB: there is also StorageIoOperation::OpenAfterReplace which is for the case
372 : // where our caller doesn't get to use the returned VirtualFile before its
373 : // slot gets re-used by someone else.
374 61790 : let file = observe_duration!(StorageIoOperation::Open, {
375 61790 : open_options.open(path.as_std_path()).await?
376 : });
377 :
378 : // Strip all options other than read and write.
379 : //
380 : // It would perhaps be nicer to check just for the read and write flags
381 : // explicitly, but OpenOptions doesn't contain any functions to read flags,
382 : // only to set them.
383 61790 : let mut reopen_options = open_options.clone();
384 61790 : reopen_options.create(false);
385 61790 : reopen_options.create_new(false);
386 61790 : reopen_options.truncate(false);
387 61790 :
388 61790 : let vfile = VirtualFile {
389 61790 : handle: RwLock::new(handle),
390 61790 : pos: 0,
391 61790 : path: path.to_path_buf(),
392 61790 : open_options: reopen_options,
393 61790 : tenant_id,
394 61790 : shard_id,
395 61790 : timeline_id,
396 61790 : };
397 61790 :
398 61790 : // TODO: Under pressure, it's likely the slot will get re-used and
399 61790 : // the underlying file closed before they get around to using it.
400 61790 : // => https://github.com/neondatabase/neon/issues/6065
401 61790 : slot_guard.file.replace(file);
402 61790 :
403 61790 : Ok(vfile)
404 61790 : }
405 :
406 : /// Async version of [`::utils::crashsafe::overwrite`].
407 : ///
408 : /// # NB:
409 : ///
410 : /// Doesn't actually use the [`VirtualFile`] file descriptor cache, but,
411 : /// it did at an earlier time.
412 : /// And it will use this module's [`io_engine`] in the near future, so, leaving it here.
413 8859 : pub async fn crashsafe_overwrite<B: BoundedBuf<Buf = Buf> + Send, Buf: IoBuf + Send>(
414 8859 : final_path: Utf8PathBuf,
415 8859 : tmp_path: Utf8PathBuf,
416 8859 : content: B,
417 8859 : ) -> std::io::Result<()> {
418 8859 : // TODO: use tokio_epoll_uring if configured as `io_engine`.
419 8859 : // See https://github.com/neondatabase/neon/issues/6663
420 8859 :
421 8859 : tokio::task::spawn_blocking(move || {
422 8859 : let slice_storage;
423 8859 : let content_len = content.bytes_init();
424 8859 : let content = if content.bytes_init() > 0 {
425 8859 : slice_storage = Some(content.slice(0..content_len));
426 8859 : slice_storage.as_deref().expect("just set it to Some()")
427 : } else {
428 0 : &[]
429 : };
430 8859 : utils::crashsafe::overwrite(&final_path, &tmp_path, content)
431 8859 : })
432 8896 : .await
433 8859 : .expect("blocking task is never aborted")
434 8859 : }
435 :
436 : /// Call File::sync_all() on the underlying File.
437 22304 : pub async fn sync_all(&self) -> Result<(), Error> {
438 22304 : with_file!(self, StorageIoOperation::Fsync, |file_guard| file_guard
439 22304 : .with_std_file(|std_file| std_file.sync_all()))
440 22304 : }
441 :
442 22304 : pub async fn metadata(&self) -> Result<fs::Metadata, Error> {
443 22304 : with_file!(self, StorageIoOperation::Metadata, |file_guard| file_guard
444 22304 : .with_std_file(|std_file| std_file.metadata()))
445 22304 : }
446 :
447 : /// Helper function internal to `VirtualFile` that looks up the underlying File,
448 : /// opens it and evicts some other File if necessary. The passed parameter is
449 : /// assumed to be a function available for the physical `File`.
450 : ///
451 : /// We are doing it via a macro as Rust doesn't support async closures that
452 : /// take on parameters with lifetimes.
453 13854409 : async fn lock_file(&self) -> Result<FileGuard, Error> {
454 13854409 : let open_files = get_open_files();
455 :
456 197682 : let mut handle_guard = {
457 : // Read the cached slot handle, and see if the slot that it points to still
458 : // contains our File.
459 : //
460 : // We only need to hold the handle lock while we read the current handle. If
461 : // another thread closes the file and recycles the slot for a different file,
462 : // we will notice that the handle we read is no longer valid and retry.
463 13854409 : let mut handle = *self.handle.read().await;
464 13960942 : loop {
465 13960942 : // Check if the slot contains our File
466 13960942 : {
467 13960942 : let slot = &open_files.slots[handle.index];
468 13960942 : let slot_guard = slot.inner.read().await;
469 13960942 : if slot_guard.tag == handle.tag && slot_guard.file.is_some() {
470 : // Found a cached file descriptor.
471 13656727 : slot.recently_used.store(true, Ordering::Relaxed);
472 13656727 : return Ok(FileGuard { slot_guard });
473 304215 : }
474 : }
475 :
476 : // The slot didn't contain our File. We will have to open it ourselves,
477 : // but before that, grab a write lock on handle in the VirtualFile, so
478 : // that no other thread will try to concurrently open the same file.
479 304215 : let handle_guard = self.handle.write().await;
480 :
481 : // If another thread changed the handle while we were not holding the lock,
482 : // then the handle might now be valid again. Loop back to retry.
483 304215 : if *handle_guard != handle {
484 106533 : handle = *handle_guard;
485 106533 : continue;
486 197682 : }
487 197682 : break handle_guard;
488 : }
489 : };
490 :
491 : // We need to open the file ourselves. The handle in the VirtualFile is
492 : // now locked in write-mode. Find a free slot to put it in.
493 197682 : let (handle, mut slot_guard) = open_files.find_victim_slot().await;
494 :
495 : // Re-open the physical file.
496 : // NB: we use StorageIoOperation::OpenAferReplace for this to distinguish this
497 : // case from StorageIoOperation::Open. This helps with identifying thrashing
498 : // of the virtual file descriptor cache.
499 197682 : let file = observe_duration!(StorageIoOperation::OpenAfterReplace, {
500 197682 : self.open_options.open(self.path.as_std_path()).await?
501 : });
502 :
503 : // Store the File in the slot and update the handle in the VirtualFile
504 : // to point to it.
505 197682 : slot_guard.file.replace(file);
506 197682 :
507 197682 : *handle_guard = handle;
508 197682 :
509 197682 : return Ok(FileGuard {
510 197682 : slot_guard: slot_guard.downgrade(),
511 197682 : });
512 13854409 : }
513 :
514 0 : pub fn remove(self) {
515 0 : let path = self.path.clone();
516 0 : drop(self);
517 0 : std::fs::remove_file(path).expect("failed to remove the virtual file");
518 0 : }
519 :
520 66934 : pub async fn seek(&mut self, pos: SeekFrom) -> Result<u64, Error> {
521 66934 : match pos {
522 66924 : SeekFrom::Start(offset) => {
523 66924 : self.pos = offset;
524 66924 : }
525 4 : SeekFrom::End(offset) => {
526 4 : self.pos = with_file!(self, StorageIoOperation::Seek, |mut file_guard| file_guard
527 4 : .with_std_file_mut(|std_file| std_file.seek(SeekFrom::End(offset))))?
528 : }
529 6 : SeekFrom::Current(offset) => {
530 6 : let pos = self.pos as i128 + offset as i128;
531 6 : if pos < 0 {
532 2 : return Err(Error::new(
533 2 : ErrorKind::InvalidInput,
534 2 : "offset would be negative",
535 2 : ));
536 4 : }
537 4 : if pos > u64::MAX as i128 {
538 0 : return Err(Error::new(ErrorKind::InvalidInput, "offset overflow"));
539 4 : }
540 4 : self.pos = pos as u64;
541 : }
542 : }
543 66930 : Ok(self.pos)
544 66934 : }
545 :
546 5145545 : pub async fn read_exact_at<B>(&self, buf: B, offset: u64) -> Result<B, Error>
547 5145545 : where
548 5145545 : B: IoBufMut + Send,
549 5145545 : {
550 5145545 : let (buf, res) =
551 5145545 : read_exact_at_impl(buf, offset, |buf, offset| self.read_at(buf, offset)).await;
552 5145545 : res.map(|()| buf)
553 5145545 : }
554 :
555 : /// Like [`Self::read_exact_at`] but for [`PageWriteGuard`].
556 4925338 : pub async fn read_exact_at_page(
557 4925338 : &self,
558 4925338 : page: PageWriteGuard<'static>,
559 4925338 : offset: u64,
560 4925343 : ) -> Result<PageWriteGuard<'static>, Error> {
561 4925343 : let buf = PageWriteGuardBuf {
562 4925343 : page,
563 4925343 : init_up_to: 0,
564 4925343 : };
565 4925343 : let res = self.read_exact_at(buf, offset).await;
566 4925343 : res.map(|PageWriteGuardBuf { page, .. }| page)
567 4925343 : .map_err(|e| Error::new(ErrorKind::Other, e))
568 4925343 : }
569 :
570 : // Copied from https://doc.rust-lang.org/1.72.0/src/std/os/unix/fs.rs.html#219-235
571 3696832 : pub async fn write_all_at<B: BoundedBuf>(
572 3696832 : &self,
573 3696832 : buf: B,
574 3696832 : mut offset: u64,
575 3696832 : ) -> (B::Buf, Result<(), Error>) {
576 3696832 : let buf_len = buf.bytes_init();
577 3696832 : if buf_len == 0 {
578 0 : return (Slice::into_inner(buf.slice_full()), Ok(()));
579 3696832 : }
580 3696832 : let mut buf = buf.slice(0..buf_len);
581 7393664 : while !buf.is_empty() {
582 : // TODO: push `buf` further down
583 3696832 : match self.write_at(&buf, offset).await {
584 : Ok(0) => {
585 0 : return (
586 0 : Slice::into_inner(buf),
587 0 : Err(Error::new(
588 0 : std::io::ErrorKind::WriteZero,
589 0 : "failed to write whole buffer",
590 0 : )),
591 0 : );
592 : }
593 3696832 : Ok(n) => {
594 3696832 : buf = buf.slice(n..);
595 3696832 : offset += n as u64;
596 3696832 : }
597 0 : Err(ref e) if e.kind() == std::io::ErrorKind::Interrupted => {}
598 0 : Err(e) => return (Slice::into_inner(buf), Err(e)),
599 : }
600 : }
601 3696832 : (Slice::into_inner(buf), Ok(()))
602 3696832 : }
603 :
604 : /// Writes `buf.slice(0..buf.bytes_init())`.
605 : /// Returns the IoBuf that is underlying the BoundedBuf `buf`.
606 : /// I.e., the returned value's `bytes_init()` method returns something different than the `bytes_init()` that was passed in.
607 : /// It's quite brittle and easy to mis-use, so, we return the size in the Ok() variant.
608 4967006 : pub async fn write_all<B: BoundedBuf>(&mut self, buf: B) -> (B::Buf, Result<usize, Error>) {
609 4967006 : let nbytes = buf.bytes_init();
610 4967006 : if nbytes == 0 {
611 30 : return (Slice::into_inner(buf.slice_full()), Ok(0));
612 4966976 : }
613 4966976 : let mut buf = buf.slice(0..nbytes);
614 9933950 : while !buf.is_empty() {
615 : // TODO: push `Slice` further down
616 4966976 : match self.write(&buf).await {
617 : Ok(0) => {
618 0 : return (
619 0 : Slice::into_inner(buf),
620 0 : Err(Error::new(
621 0 : std::io::ErrorKind::WriteZero,
622 0 : "failed to write whole buffer",
623 0 : )),
624 0 : );
625 : }
626 4966974 : Ok(n) => {
627 4966974 : buf = buf.slice(n..);
628 4966974 : }
629 2 : Err(ref e) if e.kind() == std::io::ErrorKind::Interrupted => {}
630 2 : Err(e) => return (Slice::into_inner(buf), Err(e)),
631 : }
632 : }
633 4966974 : (Slice::into_inner(buf), Ok(nbytes))
634 4967006 : }
635 :
636 4966976 : async fn write(&mut self, buf: &[u8]) -> Result<usize, std::io::Error> {
637 4966976 : let pos = self.pos;
638 4966976 : let n = self.write_at(buf, pos).await?;
639 4966974 : self.pos += n as u64;
640 4966974 : Ok(n)
641 4966976 : }
642 :
643 5145989 : pub(crate) async fn read_at<B>(&self, buf: B, offset: u64) -> (B, Result<usize, Error>)
644 5145989 : where
645 5145989 : B: tokio_epoll_uring::BoundedBufMut + Send,
646 5145989 : {
647 5145989 : let file_guard = match self.lock_file().await {
648 5145989 : Ok(file_guard) => file_guard,
649 0 : Err(e) => return (buf, Err(e)),
650 : };
651 :
652 5145989 : observe_duration!(StorageIoOperation::Read, {
653 5145989 : let ((_file_guard, buf), res) = io_engine::get().read_at(file_guard, offset, buf).await;
654 5145989 : if let Ok(size) = res {
655 5145987 : STORAGE_IO_SIZE
656 5145987 : .with_label_values(&[
657 5145987 : "read",
658 5145987 : &self.tenant_id,
659 5145987 : &self.shard_id,
660 5145987 : &self.timeline_id,
661 5145987 : ])
662 5145987 : .add(size as i64);
663 5145987 : }
664 5145989 : (buf, res)
665 : })
666 5145989 : }
667 :
668 8663808 : async fn write_at(&self, buf: &[u8], offset: u64) -> Result<usize, Error> {
669 8663808 : let result = with_file!(self, StorageIoOperation::Write, |file_guard| {
670 8663808 : file_guard.with_std_file(|std_file| std_file.write_at(buf, offset))
671 : });
672 8663808 : if let Ok(size) = result {
673 8663806 : STORAGE_IO_SIZE
674 8663806 : .with_label_values(&["write", &self.tenant_id, &self.shard_id, &self.timeline_id])
675 8663806 : .add(size as i64);
676 8663806 : }
677 8663808 : result
678 8663808 : }
679 : }
680 :
681 : // Adapted from https://doc.rust-lang.org/1.72.0/src/std/os/unix/fs.rs.html#117-135
682 5145553 : pub async fn read_exact_at_impl<B, F, Fut>(
683 5145553 : buf: B,
684 5145553 : mut offset: u64,
685 5145553 : mut read_at: F,
686 5145553 : ) -> (B, std::io::Result<()>)
687 5145553 : where
688 5145553 : B: IoBufMut + Send,
689 5145553 : F: FnMut(tokio_epoll_uring::Slice<B>, u64) -> Fut,
690 5145553 : Fut: std::future::Future<Output = (tokio_epoll_uring::Slice<B>, std::io::Result<usize>)>,
691 5145553 : {
692 5145553 : let mut buf: tokio_epoll_uring::Slice<B> = buf.slice_full(); // includes all the uninitialized memory
693 10291108 : while buf.bytes_total() != 0 {
694 : let res;
695 5145557 : (buf, res) = read_at(buf, offset).await;
696 0 : match res {
697 2 : Ok(0) => break,
698 5145555 : Ok(n) => {
699 5145555 : buf = buf.slice(n..);
700 5145555 : offset += n as u64;
701 5145555 : }
702 0 : Err(ref e) if e.kind() == std::io::ErrorKind::Interrupted => {}
703 0 : Err(e) => return (buf.into_inner(), Err(e)),
704 : }
705 : }
706 : // NB: don't use `buf.is_empty()` here; it is from the
707 : // `impl Deref for Slice { Target = [u8] }`; the the &[u8]
708 : // returned by it only covers the initialized portion of `buf`.
709 : // Whereas we're interested in ensuring that we filled the entire
710 : // buffer that the user passed in.
711 5145553 : if buf.bytes_total() != 0 {
712 2 : (
713 2 : buf.into_inner(),
714 2 : Err(std::io::Error::new(
715 2 : std::io::ErrorKind::UnexpectedEof,
716 2 : "failed to fill whole buffer",
717 2 : )),
718 2 : )
719 : } else {
720 5145551 : assert_eq!(buf.len(), buf.bytes_total());
721 5145551 : (buf.into_inner(), Ok(()))
722 : }
723 5145553 : }
724 :
725 : #[cfg(test)]
726 : mod test_read_exact_at_impl {
727 :
728 : use std::{collections::VecDeque, sync::Arc};
729 :
730 : use tokio_epoll_uring::{BoundedBuf, BoundedBufMut};
731 :
732 : use super::read_exact_at_impl;
733 :
734 : struct Expectation {
735 : offset: u64,
736 : bytes_total: usize,
737 : result: std::io::Result<Vec<u8>>,
738 : }
739 : struct MockReadAt {
740 : expectations: VecDeque<Expectation>,
741 : }
742 :
743 : impl MockReadAt {
744 12 : async fn read_at(
745 12 : &mut self,
746 12 : mut buf: tokio_epoll_uring::Slice<Vec<u8>>,
747 12 : offset: u64,
748 12 : ) -> (tokio_epoll_uring::Slice<Vec<u8>>, std::io::Result<usize>) {
749 12 : let exp = self
750 12 : .expectations
751 12 : .pop_front()
752 12 : .expect("read_at called but we have no expectations left");
753 12 : assert_eq!(exp.offset, offset);
754 12 : assert_eq!(exp.bytes_total, buf.bytes_total());
755 12 : match exp.result {
756 12 : Ok(bytes) => {
757 12 : assert!(bytes.len() <= buf.bytes_total());
758 12 : buf.put_slice(&bytes);
759 12 : (buf, Ok(bytes.len()))
760 : }
761 0 : Err(e) => (buf, Err(e)),
762 : }
763 12 : }
764 : }
765 :
766 : impl Drop for MockReadAt {
767 8 : fn drop(&mut self) {
768 8 : assert_eq!(self.expectations.len(), 0);
769 8 : }
770 : }
771 :
772 2 : #[tokio::test]
773 2 : async fn test_basic() {
774 2 : let buf = Vec::with_capacity(5);
775 2 : let mock_read_at = Arc::new(tokio::sync::Mutex::new(MockReadAt {
776 2 : expectations: VecDeque::from(vec![Expectation {
777 2 : offset: 0,
778 2 : bytes_total: 5,
779 2 : result: Ok(vec![b'a', b'b', b'c', b'd', b'e']),
780 2 : }]),
781 2 : }));
782 2 : let (buf, res) = read_exact_at_impl(buf, 0, |buf, offset| {
783 2 : let mock_read_at = Arc::clone(&mock_read_at);
784 2 : async move { mock_read_at.lock().await.read_at(buf, offset).await }
785 2 : })
786 2 : .await;
787 2 : assert!(res.is_ok());
788 2 : assert_eq!(buf, vec![b'a', b'b', b'c', b'd', b'e']);
789 2 : }
790 :
791 2 : #[tokio::test]
792 2 : async fn test_empty_buf_issues_no_syscall() {
793 2 : let buf = Vec::new();
794 2 : let mock_read_at = Arc::new(tokio::sync::Mutex::new(MockReadAt {
795 2 : expectations: VecDeque::new(),
796 2 : }));
797 2 : let (_buf, res) = read_exact_at_impl(buf, 0, |buf, offset| {
798 0 : let mock_read_at = Arc::clone(&mock_read_at);
799 2 : async move { mock_read_at.lock().await.read_at(buf, offset).await }
800 2 : })
801 2 : .await;
802 2 : assert!(res.is_ok());
803 2 : }
804 :
805 2 : #[tokio::test]
806 2 : async fn test_two_read_at_calls_needed_until_buf_filled() {
807 2 : let buf = Vec::with_capacity(4);
808 2 : let mock_read_at = Arc::new(tokio::sync::Mutex::new(MockReadAt {
809 2 : expectations: VecDeque::from(vec![
810 2 : Expectation {
811 2 : offset: 0,
812 2 : bytes_total: 4,
813 2 : result: Ok(vec![b'a', b'b']),
814 2 : },
815 2 : Expectation {
816 2 : offset: 2,
817 2 : bytes_total: 2,
818 2 : result: Ok(vec![b'c', b'd']),
819 2 : },
820 2 : ]),
821 2 : }));
822 4 : let (buf, res) = read_exact_at_impl(buf, 0, |buf, offset| {
823 4 : let mock_read_at = Arc::clone(&mock_read_at);
824 4 : async move { mock_read_at.lock().await.read_at(buf, offset).await }
825 4 : })
826 2 : .await;
827 2 : assert!(res.is_ok());
828 2 : assert_eq!(buf, vec![b'a', b'b', b'c', b'd']);
829 2 : }
830 :
831 2 : #[tokio::test]
832 2 : async fn test_eof_before_buffer_full() {
833 2 : let buf = Vec::with_capacity(3);
834 2 : let mock_read_at = Arc::new(tokio::sync::Mutex::new(MockReadAt {
835 2 : expectations: VecDeque::from(vec![
836 2 : Expectation {
837 2 : offset: 0,
838 2 : bytes_total: 3,
839 2 : result: Ok(vec![b'a']),
840 2 : },
841 2 : Expectation {
842 2 : offset: 1,
843 2 : bytes_total: 2,
844 2 : result: Ok(vec![b'b']),
845 2 : },
846 2 : Expectation {
847 2 : offset: 2,
848 2 : bytes_total: 1,
849 2 : result: Ok(vec![]),
850 2 : },
851 2 : ]),
852 2 : }));
853 6 : let (_buf, res) = read_exact_at_impl(buf, 0, |buf, offset| {
854 6 : let mock_read_at = Arc::clone(&mock_read_at);
855 6 : async move { mock_read_at.lock().await.read_at(buf, offset).await }
856 6 : })
857 2 : .await;
858 2 : let Err(err) = res else {
859 2 : panic!("should return an error");
860 2 : };
861 2 : assert_eq!(err.kind(), std::io::ErrorKind::UnexpectedEof);
862 2 : assert_eq!(format!("{err}"), "failed to fill whole buffer");
863 2 : // buffer contents on error are unspecified
864 2 : }
865 : }
866 :
867 : struct FileGuard {
868 : slot_guard: RwLockReadGuard<'static, SlotInner>,
869 : }
870 :
871 : impl AsRef<OwnedFd> for FileGuard {
872 13854401 : fn as_ref(&self) -> &OwnedFd {
873 13854401 : // This unwrap is safe because we only create `FileGuard`s
874 13854401 : // if we know that the file is Some.
875 13854401 : self.slot_guard.file.as_ref().unwrap()
876 13854401 : }
877 : }
878 :
879 : impl FileGuard {
880 : /// Soft deprecation: we'll move VirtualFile to async APIs and remove this function eventually.
881 13709694 : fn with_std_file<F, R>(&self, with: F) -> R
882 13709694 : where
883 13709694 : F: FnOnce(&File) -> R,
884 13709694 : {
885 13709694 : // SAFETY:
886 13709694 : // - lifetime of the fd: `file` doesn't outlive the OwnedFd stored in `self`.
887 13709694 : // - `&` usage below: `self` is `&`, hence Rust typesystem guarantees there are is no `&mut`
888 13709694 : let file = unsafe { File::from_raw_fd(self.as_ref().as_raw_fd()) };
889 13709694 : let res = with(&file);
890 13709694 : let _ = file.into_raw_fd();
891 13709694 : res
892 13709694 : }
893 : /// Soft deprecation: we'll move VirtualFile to async APIs and remove this function eventually.
894 4 : fn with_std_file_mut<F, R>(&mut self, with: F) -> R
895 4 : where
896 4 : F: FnOnce(&mut File) -> R,
897 4 : {
898 4 : // SAFETY:
899 4 : // - lifetime of the fd: `file` doesn't outlive the OwnedFd stored in `self`.
900 4 : // - &mut usage below: `self` is `&mut`, hence this call is the only task/thread that has control over the underlying fd
901 4 : let mut file = unsafe { File::from_raw_fd(self.as_ref().as_raw_fd()) };
902 4 : let res = with(&mut file);
903 4 : let _ = file.into_raw_fd();
904 4 : res
905 4 : }
906 : }
907 :
908 : impl tokio_epoll_uring::IoFd for FileGuard {
909 144711 : unsafe fn as_fd(&self) -> RawFd {
910 144711 : let owned_fd: &OwnedFd = self.as_ref();
911 144711 : owned_fd.as_raw_fd()
912 144711 : }
913 : }
914 :
915 : #[cfg(test)]
916 : impl VirtualFile {
917 20200 : pub(crate) async fn read_blk(
918 20200 : &self,
919 20200 : blknum: u32,
920 20200 : ) -> Result<crate::tenant::block_io::BlockLease<'_>, std::io::Error> {
921 20200 : use crate::page_cache::PAGE_SZ;
922 20200 : let buf = vec![0; PAGE_SZ];
923 20200 : let buf = self
924 20200 : .read_exact_at(buf, blknum as u64 * (PAGE_SZ as u64))
925 10256 : .await?;
926 20200 : Ok(crate::tenant::block_io::BlockLease::Vec(buf))
927 20200 : }
928 :
929 224 : async fn read_to_end(&mut self, buf: &mut Vec<u8>) -> Result<(), Error> {
930 224 : let mut tmp = vec![0; 128];
931 : loop {
932 : let res;
933 444 : (tmp, res) = self.read_at(tmp, self.pos).await;
934 2 : match res {
935 222 : Ok(0) => return Ok(()),
936 220 : Ok(n) => {
937 220 : self.pos += n as u64;
938 220 : buf.extend_from_slice(&tmp[..n]);
939 220 : }
940 2 : Err(ref e) if e.kind() == std::io::ErrorKind::Interrupted => {}
941 2 : Err(e) => return Err(e),
942 : }
943 : }
944 224 : }
945 : }
946 :
947 : impl Drop for VirtualFile {
948 : /// If a VirtualFile is dropped, close the underlying file if it was open.
949 47312 : fn drop(&mut self) {
950 47312 : let handle = self.handle.get_mut();
951 47312 :
952 47312 : fn clean_slot(slot: &Slot, mut slot_guard: RwLockWriteGuard<'_, SlotInner>, tag: u64) {
953 47312 : if slot_guard.tag == tag {
954 47312 : slot.recently_used.store(false, Ordering::Relaxed);
955 47312 : // there is also operation "close-by-replace" for closes done on eviction for
956 47312 : // comparison.
957 47312 : if let Some(fd) = slot_guard.file.take() {
958 37605 : STORAGE_IO_TIME_METRIC
959 37605 : .get(StorageIoOperation::Close)
960 37605 : .observe_closure_duration(|| drop(fd));
961 37605 : }
962 47312 : }
963 47312 : }
964 47312 :
965 47312 : // We don't have async drop so we cannot directly await the lock here.
966 47312 : // Instead, first do a best-effort attempt at closing the underlying
967 47312 : // file descriptor by using `try_write`, and if that fails, spawn
968 47312 : // a tokio task to do it asynchronously: we just want it to be
969 47312 : // cleaned up eventually.
970 47312 : // Most of the time, the `try_lock` should succeed though,
971 47312 : // as we have `&mut self` access. In other words, if the slot
972 47312 : // is still occupied by our file, there should be no access from
973 47312 : // other I/O operations; the only other possible place to lock
974 47312 : // the slot is the lock algorithm looking for free slots.
975 47312 : let slot = &get_open_files().slots[handle.index];
976 47312 : if let Ok(slot_guard) = slot.inner.try_write() {
977 47312 : clean_slot(slot, slot_guard, handle.tag);
978 47312 : } else {
979 0 : let tag = handle.tag;
980 0 : tokio::spawn(async move {
981 0 : let slot_guard = slot.inner.write().await;
982 0 : clean_slot(slot, slot_guard, tag);
983 0 : });
984 0 : };
985 47312 : }
986 : }
987 :
988 : impl OpenFiles {
989 721 : fn new(num_slots: usize) -> OpenFiles {
990 721 : let mut slots = Box::new(Vec::with_capacity(num_slots));
991 63460 : for _ in 0..num_slots {
992 63460 : let slot = Slot {
993 63460 : recently_used: AtomicBool::new(false),
994 63460 : inner: RwLock::new(SlotInner { tag: 0, file: None }),
995 63460 : };
996 63460 : slots.push(slot);
997 63460 : }
998 :
999 721 : OpenFiles {
1000 721 : next: AtomicUsize::new(0),
1001 721 : slots: Box::leak(slots),
1002 721 : }
1003 721 : }
1004 : }
1005 :
1006 : ///
1007 : /// Initialize the virtual file module. This must be called once at page
1008 : /// server startup.
1009 : ///
1010 : #[cfg(not(test))]
1011 625 : pub fn init(num_slots: usize, engine: IoEngineKind) {
1012 625 : if OPEN_FILES.set(OpenFiles::new(num_slots)).is_err() {
1013 0 : panic!("virtual_file::init called twice");
1014 625 : }
1015 625 : io_engine::init(engine);
1016 625 : crate::metrics::virtual_file_descriptor_cache::SIZE_MAX.set(num_slots as u64);
1017 625 : }
1018 :
1019 : const TEST_MAX_FILE_DESCRIPTORS: usize = 10;
1020 :
1021 : // Get a handle to the global slots array.
1022 13963504 : fn get_open_files() -> &'static OpenFiles {
1023 13963504 : //
1024 13963504 : // In unit tests, page server startup doesn't happen and no one calls
1025 13963504 : // virtual_file::init(). Initialize it here, with a small array.
1026 13963504 : //
1027 13963504 : // This applies to the virtual file tests below, but all other unit
1028 13963504 : // tests too, so the virtual file facility is always usable in
1029 13963504 : // unit tests.
1030 13963504 : //
1031 13963504 : if cfg!(test) {
1032 365255 : OPEN_FILES.get_or_init(|| OpenFiles::new(TEST_MAX_FILE_DESCRIPTORS))
1033 : } else {
1034 13598249 : OPEN_FILES.get().expect("virtual_file::init not called yet")
1035 : }
1036 13963504 : }
1037 :
1038 : #[cfg(test)]
1039 : mod tests {
1040 : use super::*;
1041 : use rand::seq::SliceRandom;
1042 : use rand::thread_rng;
1043 : use rand::Rng;
1044 : use std::future::Future;
1045 : use std::io::Write;
1046 : use std::sync::Arc;
1047 :
1048 : enum MaybeVirtualFile {
1049 : VirtualFile(VirtualFile),
1050 : File(File),
1051 : }
1052 :
1053 : impl From<VirtualFile> for MaybeVirtualFile {
1054 6 : fn from(vf: VirtualFile) -> Self {
1055 6 : MaybeVirtualFile::VirtualFile(vf)
1056 6 : }
1057 : }
1058 :
1059 : impl MaybeVirtualFile {
1060 404 : async fn read_exact_at(&self, mut buf: Vec<u8>, offset: u64) -> Result<Vec<u8>, Error> {
1061 404 : match self {
1062 203 : MaybeVirtualFile::VirtualFile(file) => file.read_exact_at(buf, offset).await,
1063 202 : MaybeVirtualFile::File(file) => file.read_exact_at(&mut buf, offset).map(|()| buf),
1064 : }
1065 404 : }
1066 8 : async fn write_all_at<B: BoundedBuf>(&self, buf: B, offset: u64) -> Result<(), Error> {
1067 8 : match self {
1068 4 : MaybeVirtualFile::VirtualFile(file) => {
1069 4 : let (_buf, res) = file.write_all_at(buf, offset).await;
1070 4 : res
1071 : }
1072 4 : MaybeVirtualFile::File(file) => {
1073 4 : let buf_len = buf.bytes_init();
1074 4 : if buf_len == 0 {
1075 0 : return Ok(());
1076 4 : }
1077 4 : file.write_all_at(&buf.slice(0..buf_len), offset)
1078 : }
1079 : }
1080 8 : }
1081 36 : async fn seek(&mut self, pos: SeekFrom) -> Result<u64, Error> {
1082 36 : match self {
1083 18 : MaybeVirtualFile::VirtualFile(file) => file.seek(pos).await,
1084 18 : MaybeVirtualFile::File(file) => file.seek(pos),
1085 : }
1086 36 : }
1087 8 : async fn write_all<B: BoundedBuf>(&mut self, buf: B) -> Result<(), Error> {
1088 8 : match self {
1089 4 : MaybeVirtualFile::VirtualFile(file) => {
1090 4 : let (_buf, res) = file.write_all(buf).await;
1091 4 : res.map(|_| ())
1092 : }
1093 4 : MaybeVirtualFile::File(file) => {
1094 4 : let buf_len = buf.bytes_init();
1095 4 : if buf_len == 0 {
1096 0 : return Ok(());
1097 4 : }
1098 4 : file.write_all(&buf.slice(0..buf_len))
1099 : }
1100 : }
1101 8 : }
1102 :
1103 : // Helper function to slurp contents of a file, starting at the current position,
1104 : // into a string
1105 442 : async fn read_string(&mut self) -> Result<String, Error> {
1106 442 : use std::io::Read;
1107 442 : let mut buf = String::new();
1108 442 : match self {
1109 224 : MaybeVirtualFile::VirtualFile(file) => {
1110 224 : let mut buf = Vec::new();
1111 226 : file.read_to_end(&mut buf).await?;
1112 222 : return Ok(String::from_utf8(buf).unwrap());
1113 : }
1114 218 : MaybeVirtualFile::File(file) => {
1115 218 : file.read_to_string(&mut buf)?;
1116 : }
1117 : }
1118 216 : Ok(buf)
1119 442 : }
1120 :
1121 : // Helper function to slurp a portion of a file into a string
1122 404 : async fn read_string_at(&mut self, pos: u64, len: usize) -> Result<String, Error> {
1123 404 : let buf = vec![0; len];
1124 404 : let buf = self.read_exact_at(buf, pos).await?;
1125 404 : Ok(String::from_utf8(buf).unwrap())
1126 404 : }
1127 : }
1128 :
1129 2 : #[tokio::test]
1130 2 : async fn test_virtual_files() -> anyhow::Result<()> {
1131 2 : // The real work is done in the test_files() helper function. This
1132 2 : // allows us to run the same set of tests against a native File, and
1133 2 : // VirtualFile. We trust the native Files and wouldn't need to test them,
1134 2 : // but this allows us to verify that the operations return the same
1135 2 : // results with VirtualFiles as with native Files. (Except that with
1136 2 : // native files, you will run out of file descriptors if the ulimit
1137 2 : // is low enough.)
1138 206 : test_files("virtual_files", |path, open_options| async move {
1139 206 : let vf = VirtualFile::open_with_options(&path, &open_options).await?;
1140 206 : Ok(MaybeVirtualFile::VirtualFile(vf))
1141 412 : })
1142 527 : .await
1143 2 : }
1144 :
1145 2 : #[tokio::test]
1146 2 : async fn test_physical_files() -> anyhow::Result<()> {
1147 206 : test_files("physical_files", |path, open_options| async move {
1148 206 : Ok(MaybeVirtualFile::File({
1149 206 : let owned_fd = open_options.open(path.as_std_path()).await?;
1150 206 : File::from(owned_fd)
1151 2 : }))
1152 412 : })
1153 104 : .await
1154 2 : }
1155 :
1156 4 : async fn test_files<OF, FT>(testname: &str, openfunc: OF) -> anyhow::Result<()>
1157 4 : where
1158 4 : OF: Fn(Utf8PathBuf, OpenOptions) -> FT,
1159 4 : FT: Future<Output = Result<MaybeVirtualFile, std::io::Error>>,
1160 4 : {
1161 4 : let testdir = crate::config::PageServerConf::test_repo_dir(testname);
1162 4 : std::fs::create_dir_all(&testdir)?;
1163 :
1164 4 : let path_a = testdir.join("file_a");
1165 4 : let mut file_a = openfunc(
1166 4 : path_a.clone(),
1167 4 : OpenOptions::new()
1168 4 : .write(true)
1169 4 : .create(true)
1170 4 : .truncate(true)
1171 4 : .to_owned(),
1172 4 : )
1173 4 : .await?;
1174 4 : file_a.write_all(b"foobar".to_vec()).await?;
1175 :
1176 : // cannot read from a file opened in write-only mode
1177 4 : let _ = file_a.read_string().await.unwrap_err();
1178 :
1179 : // Close the file and re-open for reading
1180 4 : let mut file_a = openfunc(path_a, OpenOptions::new().read(true).to_owned()).await?;
1181 :
1182 : // cannot write to a file opened in read-only mode
1183 4 : let _ = file_a.write_all(b"bar".to_vec()).await.unwrap_err();
1184 4 :
1185 4 : // Try simple read
1186 4 : assert_eq!("foobar", file_a.read_string().await?);
1187 :
1188 : // It's positioned at the EOF now.
1189 4 : assert_eq!("", file_a.read_string().await?);
1190 :
1191 : // Test seeks.
1192 4 : assert_eq!(file_a.seek(SeekFrom::Start(1)).await?, 1);
1193 4 : assert_eq!("oobar", file_a.read_string().await?);
1194 :
1195 4 : assert_eq!(file_a.seek(SeekFrom::End(-2)).await?, 4);
1196 4 : assert_eq!("ar", file_a.read_string().await?);
1197 :
1198 4 : assert_eq!(file_a.seek(SeekFrom::Start(1)).await?, 1);
1199 4 : assert_eq!(file_a.seek(SeekFrom::Current(2)).await?, 3);
1200 4 : assert_eq!("bar", file_a.read_string().await?);
1201 :
1202 4 : assert_eq!(file_a.seek(SeekFrom::Current(-5)).await?, 1);
1203 4 : assert_eq!("oobar", file_a.read_string().await?);
1204 :
1205 : // Test erroneous seeks to before byte 0
1206 4 : file_a.seek(SeekFrom::End(-7)).await.unwrap_err();
1207 4 : assert_eq!(file_a.seek(SeekFrom::Start(1)).await?, 1);
1208 4 : file_a.seek(SeekFrom::Current(-2)).await.unwrap_err();
1209 4 :
1210 4 : // the erroneous seek should have left the position unchanged
1211 4 : assert_eq!("oobar", file_a.read_string().await?);
1212 :
1213 : // Create another test file, and try FileExt functions on it.
1214 4 : let path_b = testdir.join("file_b");
1215 4 : let mut file_b = openfunc(
1216 4 : path_b.clone(),
1217 4 : OpenOptions::new()
1218 4 : .read(true)
1219 4 : .write(true)
1220 4 : .create(true)
1221 4 : .truncate(true)
1222 4 : .to_owned(),
1223 4 : )
1224 2 : .await?;
1225 4 : file_b.write_all_at(b"BAR".to_vec(), 3).await?;
1226 4 : file_b.write_all_at(b"FOO".to_vec(), 0).await?;
1227 :
1228 4 : assert_eq!(file_b.read_string_at(2, 3).await?, "OBA");
1229 :
1230 : // Open a lot of files, enough to cause some evictions. (Or to be precise,
1231 : // open the same file many times. The effect is the same.)
1232 : //
1233 : // leave file_a positioned at offset 1 before we start
1234 4 : assert_eq!(file_a.seek(SeekFrom::Start(1)).await?, 1);
1235 :
1236 4 : let mut vfiles = Vec::new();
1237 404 : for _ in 0..100 {
1238 400 : let mut vfile =
1239 400 : openfunc(path_b.clone(), OpenOptions::new().read(true).to_owned()).await?;
1240 400 : assert_eq!("FOOBAR", vfile.read_string().await?);
1241 400 : vfiles.push(vfile);
1242 : }
1243 :
1244 : // make sure we opened enough files to definitely cause evictions.
1245 4 : assert!(vfiles.len() > TEST_MAX_FILE_DESCRIPTORS * 2);
1246 :
1247 : // The underlying file descriptor for 'file_a' should be closed now. Try to read
1248 : // from it again. We left the file positioned at offset 1 above.
1249 4 : assert_eq!("oobar", file_a.read_string().await?);
1250 :
1251 : // Check that all the other FDs still work too. Use them in random order for
1252 : // good measure.
1253 4 : vfiles.as_mut_slice().shuffle(&mut thread_rng());
1254 400 : for vfile in vfiles.iter_mut() {
1255 400 : assert_eq!("OOBAR", vfile.read_string_at(1, 5).await?);
1256 : }
1257 :
1258 4 : Ok(())
1259 4 : }
1260 :
1261 : /// Test using VirtualFiles from many threads concurrently. This tests both using
1262 : /// a lot of VirtualFiles concurrently, causing evictions, and also using the same
1263 : /// VirtualFile from multiple threads concurrently.
1264 2 : #[tokio::test]
1265 2 : async fn test_vfile_concurrency() -> Result<(), Error> {
1266 2 : const SIZE: usize = 8 * 1024;
1267 2 : const VIRTUAL_FILES: usize = 100;
1268 2 : const THREADS: usize = 100;
1269 2 : const SAMPLE: [u8; SIZE] = [0xADu8; SIZE];
1270 2 :
1271 2 : let testdir = crate::config::PageServerConf::test_repo_dir("vfile_concurrency");
1272 2 : std::fs::create_dir_all(&testdir)?;
1273 2 :
1274 2 : // Create a test file.
1275 2 : let test_file_path = testdir.join("concurrency_test_file");
1276 2 : {
1277 2 : let file = File::create(&test_file_path)?;
1278 2 : file.write_all_at(&SAMPLE, 0)?;
1279 2 : }
1280 2 :
1281 2 : // Open the file many times.
1282 2 : let mut files = Vec::new();
1283 202 : for _ in 0..VIRTUAL_FILES {
1284 200 : let f = VirtualFile::open_with_options(&test_file_path, OpenOptions::new().read(true))
1285 101 : .await?;
1286 200 : files.push(f);
1287 2 : }
1288 2 : let files = Arc::new(files);
1289 2 :
1290 2 : // Launch many threads, and use the virtual files concurrently in random order.
1291 2 : let rt = tokio::runtime::Builder::new_multi_thread()
1292 2 : .worker_threads(THREADS)
1293 2 : .thread_name("test_vfile_concurrency thread")
1294 2 : .build()
1295 2 : .unwrap();
1296 2 : let mut hdls = Vec::new();
1297 202 : for _threadno in 0..THREADS {
1298 200 : let files = files.clone();
1299 200 : let hdl = rt.spawn(async move {
1300 200 : let mut buf = vec![0u8; SIZE];
1301 200 : let mut rng = rand::rngs::OsRng;
1302 200000 : for _ in 1..1000 {
1303 199800 : let f = &files[rng.gen_range(0..files.len())];
1304 652455 : buf = f.read_exact_at(buf, 0).await.unwrap();
1305 199800 : assert!(buf == SAMPLE);
1306 2 : }
1307 200 : });
1308 200 : hdls.push(hdl);
1309 200 : }
1310 202 : for hdl in hdls {
1311 200 : hdl.await?;
1312 2 : }
1313 2 : std::mem::forget(rt);
1314 2 :
1315 2 : Ok(())
1316 2 : }
1317 :
1318 2 : #[tokio::test]
1319 2 : async fn test_atomic_overwrite_basic() {
1320 2 : let testdir = crate::config::PageServerConf::test_repo_dir("test_atomic_overwrite_basic");
1321 2 : std::fs::create_dir_all(&testdir).unwrap();
1322 2 :
1323 2 : let path = testdir.join("myfile");
1324 2 : let tmp_path = testdir.join("myfile.tmp");
1325 2 :
1326 2 : VirtualFile::crashsafe_overwrite(path.clone(), tmp_path.clone(), b"foo".to_vec())
1327 2 : .await
1328 2 : .unwrap();
1329 2 : let mut file = MaybeVirtualFile::from(VirtualFile::open(&path).await.unwrap());
1330 2 : let post = file.read_string().await.unwrap();
1331 2 : assert_eq!(post, "foo");
1332 2 : assert!(!tmp_path.exists());
1333 2 : drop(file);
1334 2 :
1335 2 : VirtualFile::crashsafe_overwrite(path.clone(), tmp_path.clone(), b"bar".to_vec())
1336 2 : .await
1337 2 : .unwrap();
1338 2 : let mut file = MaybeVirtualFile::from(VirtualFile::open(&path).await.unwrap());
1339 2 : let post = file.read_string().await.unwrap();
1340 2 : assert_eq!(post, "bar");
1341 2 : assert!(!tmp_path.exists());
1342 2 : drop(file);
1343 2 : }
1344 :
1345 2 : #[tokio::test]
1346 2 : async fn test_atomic_overwrite_preexisting_tmp() {
1347 2 : let testdir =
1348 2 : crate::config::PageServerConf::test_repo_dir("test_atomic_overwrite_preexisting_tmp");
1349 2 : std::fs::create_dir_all(&testdir).unwrap();
1350 2 :
1351 2 : let path = testdir.join("myfile");
1352 2 : let tmp_path = testdir.join("myfile.tmp");
1353 2 :
1354 2 : std::fs::write(&tmp_path, "some preexisting junk that should be removed").unwrap();
1355 2 : assert!(tmp_path.exists());
1356 2 :
1357 2 : VirtualFile::crashsafe_overwrite(path.clone(), tmp_path.clone(), b"foo".to_vec())
1358 2 : .await
1359 2 : .unwrap();
1360 2 :
1361 2 : let mut file = MaybeVirtualFile::from(VirtualFile::open(&path).await.unwrap());
1362 2 : let post = file.read_string().await.unwrap();
1363 2 : assert_eq!(post, "foo");
1364 2 : assert!(!tmp_path.exists());
1365 2 : drop(file);
1366 2 : }
1367 : }
|
