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::IoBufMut;
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 : use utils::fs_ext;
30 :
31 : mod io_engine;
32 : mod open_options;
33 : pub 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 298020 : #[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 6296273 : fn stable_ptr(&self) -> *const u8 {
131 6296273 : self.page.as_ptr()
132 6296273 : }
133 18888819 : fn bytes_init(&self) -> usize {
134 18888819 : self.init_up_to
135 18888819 : }
136 6296273 : fn bytes_total(&self) -> usize {
137 6296273 : self.page.len()
138 6296273 : }
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 6296273 : fn stable_mut_ptr(&mut self) -> *mut u8 {
144 6296273 : self.page.as_mut_ptr()
145 6296273 : }
146 :
147 6296273 : unsafe fn set_init(&mut self, pos: usize) {
148 6296273 : assert!(pos <= self.page.len());
149 6296273 : self.init_up_to = pos;
150 6296273 : }
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 284742 : async fn find_victim_slot(&self) -> (SlotHandle, RwLockWriteGuard<SlotInner>) {
159 284742 : //
160 284742 : // Run the clock algorithm to find a slot to replace.
161 284742 : //
162 284742 : let num_slots = self.slots.len();
163 284742 : let mut retries = 0;
164 : let mut slot;
165 : let mut slot_guard;
166 : let index;
167 2393041 : loop {
168 2393041 : let next = self.next.fetch_add(1, Ordering::AcqRel) % num_slots;
169 2393041 : slot = &self.slots[next];
170 2393041 :
171 2393041 : // If the recently_used flag on this slot is set, continue the clock
172 2393041 : // sweep. Otherwise try to use this slot. If we cannot acquire the
173 2393041 : // lock, also continue the clock sweep.
174 2393041 : //
175 2393041 : // We only continue in this manner for a while, though. If we loop
176 2393041 : // through the array twice without finding a victim, just pick the
177 2393041 : // next slot and wait until we can reuse it. This way, we avoid
178 2393041 : // spinning in the extreme case that all the slots are busy with an
179 2393041 : // I/O operation.
180 2393041 : if retries < num_slots * 2 {
181 2310796 : if !slot.recently_used.swap(false, Ordering::Release) {
182 2039372 : if let Ok(guard) = slot.inner.try_write() {
183 202497 : slot_guard = guard;
184 202497 : index = next;
185 202497 : break;
186 1836875 : }
187 271424 : }
188 2108299 : retries += 1;
189 : } else {
190 82245 : slot_guard = slot.inner.write().await;
191 82245 : index = next;
192 82245 : 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 284742 : if let Some(old_file) = slot_guard.file.take() {
201 223737 : // the normal path of dropping VirtualFile uses "close", use "close-by-replace" here to
202 223737 : // distinguish the two.
203 223737 : STORAGE_IO_TIME_METRIC
204 223737 : .get(StorageIoOperation::CloseByReplace)
205 223737 : .observe_closure_duration(|| drop(old_file));
206 223737 : }
207 :
208 : // Prepare the slot for reuse and return it
209 284742 : slot_guard.tag += 1;
210 284742 : slot.recently_used.store(true, Ordering::Relaxed);
211 284742 : (
212 284742 : SlotHandle {
213 284742 : index,
214 284742 : tag: slot_guard.tag,
215 284742 : },
216 284742 : slot_guard,
217 284742 : )
218 284742 : }
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.
275 119 : fn maybe_fatal_err(self, context: &str) -> std::io::Result<T> {
276 119 : if let Err(e) = &self {
277 0 : if is_fatal_io_error(e) {
278 0 : on_fatal_io_error(e, context);
279 0 : }
280 119 : }
281 119 : self
282 119 : }
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 1434 : fn fatal_err(self, context: &str) -> T {
288 1434 : match self {
289 1434 : Ok(v) => v,
290 0 : Err(e) => {
291 0 : on_fatal_io_error(&e, context);
292 : }
293 : }
294 1434 : }
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 34437 : pub async fn open(path: &Utf8Path) -> Result<VirtualFile, std::io::Error> {
327 34437 : Self::open_with_options(path, OpenOptions::new().read(true)).await
328 34437 : }
329 :
330 : /// Create a new file for writing. If the file exists, it will be truncated.
331 : /// Like File::create.
332 15942 : pub async fn create(path: &Utf8Path) -> Result<VirtualFile, std::io::Error> {
333 15942 : Self::open_with_options(
334 15942 : path,
335 15942 : OpenOptions::new().write(true).create(true).truncate(true),
336 15942 : )
337 233 : .await
338 15942 : }
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 80227 : pub async fn open_with_options(
346 80227 : path: &Utf8Path,
347 80227 : open_options: &OpenOptions,
348 80227 : ) -> Result<VirtualFile, std::io::Error> {
349 80227 : let path_str = path.to_string();
350 80227 : let parts = path_str.split('/').collect::<Vec<&str>>();
351 80227 : let (tenant_id, shard_id, timeline_id) =
352 80227 : if parts.len() > 5 && parts[parts.len() - 5] == TENANTS_SEGMENT_NAME {
353 68903 : let tenant_shard_part = parts[parts.len() - 4];
354 68903 : let (tenant_id, shard_id) = match tenant_shard_part.parse::<TenantShardId>() {
355 68903 : Ok(tenant_shard_id) => (
356 68903 : tenant_shard_id.tenant_id.to_string(),
357 68903 : format!("{}", tenant_shard_id.shard_slug()),
358 68903 : ),
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 68903 : (tenant_id, shard_id, parts[parts.len() - 2].to_string())
366 : } else {
367 11324 : ("*".to_string(), "*".to_string(), "*".to_string())
368 : };
369 80227 : 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 80227 : let file = observe_duration!(StorageIoOperation::Open, {
375 80227 : 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 80227 : let mut reopen_options = open_options.clone();
384 80227 : reopen_options.create(false);
385 80227 : reopen_options.create_new(false);
386 80227 : reopen_options.truncate(false);
387 80227 :
388 80227 : let vfile = VirtualFile {
389 80227 : handle: RwLock::new(handle),
390 80227 : pos: 0,
391 80227 : path: path.to_path_buf(),
392 80227 : open_options: reopen_options,
393 80227 : tenant_id,
394 80227 : shard_id,
395 80227 : timeline_id,
396 80227 : };
397 80227 :
398 80227 : // TODO: Under pressure, it's likely the slot will get re-used and
399 80227 : // the underlying file closed before they get around to using it.
400 80227 : // => https://github.com/neondatabase/neon/issues/6065
401 80227 : slot_guard.file.replace(file);
402 80227 :
403 80227 : Ok(vfile)
404 80227 : }
405 :
406 : /// Writes a file to the specified `final_path` in a crash safe fasion
407 : ///
408 : /// The file is first written to the specified tmp_path, and in a second
409 : /// step, the tmp path is renamed to the final path. As renames are
410 : /// atomic, a crash during the write operation will never leave behind a
411 : /// partially written file.
412 8849 : pub async fn crashsafe_overwrite(
413 8849 : final_path: &Utf8Path,
414 8849 : tmp_path: &Utf8Path,
415 8849 : content: &[u8],
416 8849 : ) -> std::io::Result<()> {
417 8849 : let Some(final_path_parent) = final_path.parent() else {
418 0 : return Err(std::io::Error::from_raw_os_error(
419 0 : nix::errno::Errno::EINVAL as i32,
420 0 : ));
421 : };
422 8849 : std::fs::remove_file(tmp_path).or_else(fs_ext::ignore_not_found)?;
423 8849 : let mut file = Self::open_with_options(
424 8849 : tmp_path,
425 8849 : OpenOptions::new()
426 8849 : .write(true)
427 8849 : // Use `create_new` so that, if we race with ourselves or something else,
428 8849 : // we bail out instead of causing damage.
429 8849 : .create_new(true),
430 8849 : )
431 449 : .await?;
432 8849 : file.write_all(content).await?;
433 8849 : file.sync_all().await?;
434 8849 : drop(file); // before the rename, that's important!
435 8849 : // renames are atomic
436 8849 : std::fs::rename(tmp_path, final_path)?;
437 : // Only open final path parent dirfd now, so that this operation only
438 : // ever holds one VirtualFile fd at a time. That's important because
439 : // the current `find_victim_slot` impl might pick the same slot for both
440 : // VirtualFile., and it eventually does a blocking write lock instead of
441 : // try_lock.
442 8849 : let final_parent_dirfd =
443 8849 : Self::open_with_options(final_path_parent, OpenOptions::new().read(true)).await?;
444 8849 : final_parent_dirfd.sync_all().await?;
445 8849 : Ok(())
446 8849 : }
447 :
448 : /// Call File::sync_all() on the underlying File.
449 39642 : pub async fn sync_all(&self) -> Result<(), Error> {
450 39642 : with_file!(self, StorageIoOperation::Fsync, |file_guard| file_guard
451 39642 : .with_std_file(|std_file| std_file.sync_all()))
452 39642 : }
453 :
454 21944 : pub async fn metadata(&self) -> Result<fs::Metadata, Error> {
455 21944 : with_file!(self, StorageIoOperation::Metadata, |file_guard| file_guard
456 21944 : .with_std_file(|std_file| std_file.metadata()))
457 21944 : }
458 :
459 : /// Helper function internal to `VirtualFile` that looks up the underlying File,
460 : /// opens it and evicts some other File if necessary. The passed parameter is
461 : /// assumed to be a function available for the physical `File`.
462 : ///
463 : /// We are doing it via a macro as Rust doesn't support async closures that
464 : /// take on parameters with lifetimes.
465 15504468 : async fn lock_file(&self) -> Result<FileGuard, Error> {
466 15504447 : let open_files = get_open_files();
467 :
468 204515 : let mut handle_guard = {
469 : // Read the cached slot handle, and see if the slot that it points to still
470 : // contains our File.
471 : //
472 : // We only need to hold the handle lock while we read the current handle. If
473 : // another thread closes the file and recycles the slot for a different file,
474 : // we will notice that the handle we read is no longer valid and retry.
475 15504447 : let mut handle = *self.handle.read().await;
476 15597952 : loop {
477 15597952 : // Check if the slot contains our File
478 15597952 : {
479 15597952 : let slot = &open_files.slots[handle.index];
480 15597952 : let slot_guard = slot.inner.read().await;
481 15597952 : if slot_guard.tag == handle.tag && slot_guard.file.is_some() {
482 : // Found a cached file descriptor.
483 15299932 : slot.recently_used.store(true, Ordering::Relaxed);
484 15299932 : return Ok(FileGuard { slot_guard });
485 298020 : }
486 : }
487 :
488 : // The slot didn't contain our File. We will have to open it ourselves,
489 : // but before that, grab a write lock on handle in the VirtualFile, so
490 : // that no other thread will try to concurrently open the same file.
491 298020 : let handle_guard = self.handle.write().await;
492 :
493 : // If another thread changed the handle while we were not holding the lock,
494 : // then the handle might now be valid again. Loop back to retry.
495 298020 : if *handle_guard != handle {
496 93505 : handle = *handle_guard;
497 93505 : continue;
498 204515 : }
499 204515 : break handle_guard;
500 : }
501 : };
502 :
503 : // We need to open the file ourselves. The handle in the VirtualFile is
504 : // now locked in write-mode. Find a free slot to put it in.
505 204515 : let (handle, mut slot_guard) = open_files.find_victim_slot().await;
506 :
507 : // Re-open the physical file.
508 : // NB: we use StorageIoOperation::OpenAferReplace for this to distinguish this
509 : // case from StorageIoOperation::Open. This helps with identifying thrashing
510 : // of the virtual file descriptor cache.
511 204515 : let file = observe_duration!(StorageIoOperation::OpenAfterReplace, {
512 204515 : self.open_options.open(self.path.as_std_path()).await?
513 : });
514 :
515 : // Store the File in the slot and update the handle in the VirtualFile
516 : // to point to it.
517 204515 : slot_guard.file.replace(file);
518 204515 :
519 204515 : *handle_guard = handle;
520 204515 :
521 204515 : return Ok(FileGuard {
522 204515 : slot_guard: slot_guard.downgrade(),
523 204515 : });
524 15504447 : }
525 :
526 0 : pub fn remove(self) {
527 0 : let path = self.path.clone();
528 0 : drop(self);
529 0 : std::fs::remove_file(path).expect("failed to remove the virtual file");
530 0 : }
531 :
532 65853 : pub async fn seek(&mut self, pos: SeekFrom) -> Result<u64, Error> {
533 65853 : match pos {
534 65843 : SeekFrom::Start(offset) => {
535 65843 : self.pos = offset;
536 65843 : }
537 4 : SeekFrom::End(offset) => {
538 4 : self.pos = with_file!(self, StorageIoOperation::Seek, |mut file_guard| file_guard
539 4 : .with_std_file_mut(|std_file| std_file.seek(SeekFrom::End(offset))))?
540 : }
541 6 : SeekFrom::Current(offset) => {
542 6 : let pos = self.pos as i128 + offset as i128;
543 6 : if pos < 0 {
544 2 : return Err(Error::new(
545 2 : ErrorKind::InvalidInput,
546 2 : "offset would be negative",
547 2 : ));
548 4 : }
549 4 : if pos > u64::MAX as i128 {
550 0 : return Err(Error::new(ErrorKind::InvalidInput, "offset overflow"));
551 4 : }
552 4 : self.pos = pos as u64;
553 : }
554 : }
555 65849 : Ok(self.pos)
556 65853 : }
557 :
558 6516461 : pub async fn read_exact_at<B>(&self, buf: B, offset: u64) -> Result<B, Error>
559 6516461 : where
560 6516461 : B: IoBufMut + Send,
561 6516461 : {
562 6516461 : let (buf, res) =
563 6516461 : read_exact_at_impl(buf, offset, |buf, offset| self.read_at(buf, offset)).await;
564 6516461 : res.map(|()| buf)
565 6516461 : }
566 :
567 : /// Like [`Self::read_exact_at`] but for [`PageWriteGuard`].
568 6296273 : pub async fn read_exact_at_page(
569 6296273 : &self,
570 6296273 : page: PageWriteGuard<'static>,
571 6296273 : offset: u64,
572 6296273 : ) -> Result<PageWriteGuard<'static>, Error> {
573 6296259 : let buf = PageWriteGuardBuf {
574 6296259 : page,
575 6296259 : init_up_to: 0,
576 6296259 : };
577 6296259 : let res = self.read_exact_at(buf, offset).await;
578 6296259 : res.map(|PageWriteGuardBuf { page, .. }| page)
579 6296259 : .map_err(|e| Error::new(ErrorKind::Other, e))
580 6296259 : }
581 :
582 : // Copied from https://doc.rust-lang.org/1.72.0/src/std/os/unix/fs.rs.html#219-235
583 3717709 : pub async fn write_all_at(&self, mut buf: &[u8], mut offset: u64) -> Result<(), Error> {
584 7435417 : while !buf.is_empty() {
585 3717709 : match self.write_at(buf, offset).await {
586 : Ok(0) => {
587 0 : return Err(Error::new(
588 0 : std::io::ErrorKind::WriteZero,
589 0 : "failed to write whole buffer",
590 0 : ));
591 : }
592 3717708 : Ok(n) => {
593 3717708 : buf = &buf[n..];
594 3717708 : offset += n as u64;
595 3717708 : }
596 0 : Err(ref e) if e.kind() == std::io::ErrorKind::Interrupted => {}
597 0 : Err(e) => return Err(e),
598 : }
599 : }
600 3717708 : Ok(())
601 3717708 : }
602 :
603 5208282 : pub async fn write_all(&mut self, mut buf: &[u8]) -> Result<(), Error> {
604 10416516 : while !buf.is_empty() {
605 5208243 : match self.write(buf).await {
606 : Ok(0) => {
607 0 : return Err(Error::new(
608 0 : std::io::ErrorKind::WriteZero,
609 0 : "failed to write whole buffer",
610 0 : ));
611 : }
612 5208241 : Ok(n) => {
613 5208241 : buf = &buf[n..];
614 5208241 : }
615 2 : Err(ref e) if e.kind() == std::io::ErrorKind::Interrupted => {}
616 2 : Err(e) => return Err(e),
617 : }
618 : }
619 5208273 : Ok(())
620 5208275 : }
621 :
622 5208250 : async fn write(&mut self, buf: &[u8]) -> Result<usize, std::io::Error> {
623 5208243 : let pos = self.pos;
624 5208243 : let n = self.write_at(buf, pos).await?;
625 5208241 : self.pos += n as u64;
626 5208241 : Ok(n)
627 5208243 : }
628 :
629 6516905 : pub(crate) async fn read_at<B>(&self, buf: B, offset: u64) -> (B, Result<usize, Error>)
630 6516905 : where
631 6516905 : B: tokio_epoll_uring::BoundedBufMut + Send,
632 6516905 : {
633 6516905 : let file_guard = match self.lock_file().await {
634 6516905 : Ok(file_guard) => file_guard,
635 0 : Err(e) => return (buf, Err(e)),
636 : };
637 :
638 6516905 : observe_duration!(StorageIoOperation::Read, {
639 6516905 : let ((_file_guard, buf), res) = io_engine::get().read_at(file_guard, offset, buf).await;
640 6516905 : if let Ok(size) = res {
641 6516903 : STORAGE_IO_SIZE
642 6516903 : .with_label_values(&[
643 6516903 : "read",
644 6516903 : &self.tenant_id,
645 6516903 : &self.shard_id,
646 6516903 : &self.timeline_id,
647 6516903 : ])
648 6516903 : .add(size as i64);
649 6516903 : }
650 6516905 : (buf, res)
651 : })
652 6516905 : }
653 :
654 8925959 : async fn write_at(&self, buf: &[u8], offset: u64) -> Result<usize, Error> {
655 8925952 : let result = with_file!(self, StorageIoOperation::Write, |file_guard| {
656 8925952 : file_guard.with_std_file(|std_file| std_file.write_at(buf, offset))
657 : });
658 8925952 : if let Ok(size) = result {
659 8925949 : STORAGE_IO_SIZE
660 8925949 : .with_label_values(&["write", &self.tenant_id, &self.shard_id, &self.timeline_id])
661 8925949 : .add(size as i64);
662 8925949 : }
663 8925951 : result
664 8925951 : }
665 : }
666 :
667 : // Adapted from https://doc.rust-lang.org/1.72.0/src/std/os/unix/fs.rs.html#117-135
668 6516469 : pub async fn read_exact_at_impl<B, F, Fut>(
669 6516469 : buf: B,
670 6516469 : mut offset: u64,
671 6516469 : mut read_at: F,
672 6516469 : ) -> (B, std::io::Result<()>)
673 6516469 : where
674 6516469 : B: IoBufMut + Send,
675 6516469 : F: FnMut(tokio_epoll_uring::Slice<B>, u64) -> Fut,
676 6516469 : Fut: std::future::Future<Output = (tokio_epoll_uring::Slice<B>, std::io::Result<usize>)>,
677 6516469 : {
678 6516469 : use tokio_epoll_uring::BoundedBuf;
679 6516469 : let mut buf: tokio_epoll_uring::Slice<B> = buf.slice_full(); // includes all the uninitialized memory
680 13032940 : while buf.bytes_total() != 0 {
681 : let res;
682 6516473 : (buf, res) = read_at(buf, offset).await;
683 0 : match res {
684 2 : Ok(0) => break,
685 6516471 : Ok(n) => {
686 6516471 : buf = buf.slice(n..);
687 6516471 : offset += n as u64;
688 6516471 : }
689 0 : Err(ref e) if e.kind() == std::io::ErrorKind::Interrupted => {}
690 0 : Err(e) => return (buf.into_inner(), Err(e)),
691 : }
692 : }
693 : // NB: don't use `buf.is_empty()` here; it is from the
694 : // `impl Deref for Slice { Target = [u8] }`; the the &[u8]
695 : // returned by it only covers the initialized portion of `buf`.
696 : // Whereas we're interested in ensuring that we filled the entire
697 : // buffer that the user passed in.
698 6516469 : if buf.bytes_total() != 0 {
699 2 : (
700 2 : buf.into_inner(),
701 2 : Err(std::io::Error::new(
702 2 : std::io::ErrorKind::UnexpectedEof,
703 2 : "failed to fill whole buffer",
704 2 : )),
705 2 : )
706 : } else {
707 6516467 : assert_eq!(buf.len(), buf.bytes_total());
708 6516467 : (buf.into_inner(), Ok(()))
709 : }
710 6516469 : }
711 :
712 : #[cfg(test)]
713 : mod test_read_exact_at_impl {
714 :
715 : use std::{collections::VecDeque, sync::Arc};
716 :
717 : use tokio_epoll_uring::{BoundedBuf, BoundedBufMut};
718 :
719 : use super::read_exact_at_impl;
720 :
721 : struct Expectation {
722 : offset: u64,
723 : bytes_total: usize,
724 : result: std::io::Result<Vec<u8>>,
725 : }
726 : struct MockReadAt {
727 : expectations: VecDeque<Expectation>,
728 : }
729 :
730 : impl MockReadAt {
731 12 : async fn read_at(
732 12 : &mut self,
733 12 : mut buf: tokio_epoll_uring::Slice<Vec<u8>>,
734 12 : offset: u64,
735 12 : ) -> (tokio_epoll_uring::Slice<Vec<u8>>, std::io::Result<usize>) {
736 12 : let exp = self
737 12 : .expectations
738 12 : .pop_front()
739 12 : .expect("read_at called but we have no expectations left");
740 12 : assert_eq!(exp.offset, offset);
741 12 : assert_eq!(exp.bytes_total, buf.bytes_total());
742 12 : match exp.result {
743 12 : Ok(bytes) => {
744 12 : assert!(bytes.len() <= buf.bytes_total());
745 12 : buf.put_slice(&bytes);
746 12 : (buf, Ok(bytes.len()))
747 : }
748 0 : Err(e) => (buf, Err(e)),
749 : }
750 12 : }
751 : }
752 :
753 : impl Drop for MockReadAt {
754 8 : fn drop(&mut self) {
755 8 : assert_eq!(self.expectations.len(), 0);
756 8 : }
757 : }
758 :
759 2 : #[tokio::test]
760 2 : async fn test_basic() {
761 2 : let buf = Vec::with_capacity(5);
762 2 : let mock_read_at = Arc::new(tokio::sync::Mutex::new(MockReadAt {
763 2 : expectations: VecDeque::from(vec![Expectation {
764 2 : offset: 0,
765 2 : bytes_total: 5,
766 2 : result: Ok(vec![b'a', b'b', b'c', b'd', b'e']),
767 2 : }]),
768 2 : }));
769 2 : let (buf, res) = read_exact_at_impl(buf, 0, |buf, offset| {
770 2 : let mock_read_at = Arc::clone(&mock_read_at);
771 2 : async move { mock_read_at.lock().await.read_at(buf, offset).await }
772 2 : })
773 0 : .await;
774 2 : assert!(res.is_ok());
775 2 : assert_eq!(buf, vec![b'a', b'b', b'c', b'd', b'e']);
776 : }
777 :
778 2 : #[tokio::test]
779 2 : async fn test_empty_buf_issues_no_syscall() {
780 2 : let buf = Vec::new();
781 2 : let mock_read_at = Arc::new(tokio::sync::Mutex::new(MockReadAt {
782 2 : expectations: VecDeque::new(),
783 2 : }));
784 2 : let (_buf, res) = read_exact_at_impl(buf, 0, |buf, offset| {
785 0 : let mock_read_at = Arc::clone(&mock_read_at);
786 0 : async move { mock_read_at.lock().await.read_at(buf, offset).await }
787 2 : })
788 0 : .await;
789 2 : assert!(res.is_ok());
790 : }
791 :
792 2 : #[tokio::test]
793 2 : async fn test_two_read_at_calls_needed_until_buf_filled() {
794 2 : let buf = Vec::with_capacity(4);
795 2 : let mock_read_at = Arc::new(tokio::sync::Mutex::new(MockReadAt {
796 2 : expectations: VecDeque::from(vec![
797 2 : Expectation {
798 2 : offset: 0,
799 2 : bytes_total: 4,
800 2 : result: Ok(vec![b'a', b'b']),
801 2 : },
802 2 : Expectation {
803 2 : offset: 2,
804 2 : bytes_total: 2,
805 2 : result: Ok(vec![b'c', b'd']),
806 2 : },
807 2 : ]),
808 2 : }));
809 4 : let (buf, res) = read_exact_at_impl(buf, 0, |buf, offset| {
810 4 : let mock_read_at = Arc::clone(&mock_read_at);
811 4 : async move { mock_read_at.lock().await.read_at(buf, offset).await }
812 4 : })
813 0 : .await;
814 2 : assert!(res.is_ok());
815 2 : assert_eq!(buf, vec![b'a', b'b', b'c', b'd']);
816 : }
817 :
818 2 : #[tokio::test]
819 2 : async fn test_eof_before_buffer_full() {
820 2 : let buf = Vec::with_capacity(3);
821 2 : let mock_read_at = Arc::new(tokio::sync::Mutex::new(MockReadAt {
822 2 : expectations: VecDeque::from(vec![
823 2 : Expectation {
824 2 : offset: 0,
825 2 : bytes_total: 3,
826 2 : result: Ok(vec![b'a']),
827 2 : },
828 2 : Expectation {
829 2 : offset: 1,
830 2 : bytes_total: 2,
831 2 : result: Ok(vec![b'b']),
832 2 : },
833 2 : Expectation {
834 2 : offset: 2,
835 2 : bytes_total: 1,
836 2 : result: Ok(vec![]),
837 2 : },
838 2 : ]),
839 2 : }));
840 6 : let (_buf, res) = read_exact_at_impl(buf, 0, |buf, offset| {
841 6 : let mock_read_at = Arc::clone(&mock_read_at);
842 6 : async move { mock_read_at.lock().await.read_at(buf, offset).await }
843 6 : })
844 0 : .await;
845 2 : let Err(err) = res else {
846 0 : panic!("should return an error");
847 : };
848 2 : assert_eq!(err.kind(), std::io::ErrorKind::UnexpectedEof);
849 2 : assert_eq!(format!("{err}"), "failed to fill whole buffer");
850 : // buffer contents on error are unspecified
851 : }
852 : }
853 :
854 : struct FileGuard {
855 : slot_guard: RwLockReadGuard<'static, SlotInner>,
856 : }
857 :
858 : impl AsRef<OwnedFd> for FileGuard {
859 15504468 : fn as_ref(&self) -> &OwnedFd {
860 15504468 : // This unwrap is safe because we only create `FileGuard`s
861 15504468 : // if we know that the file is Some.
862 15504468 : self.slot_guard.file.as_ref().unwrap()
863 15504468 : }
864 : }
865 :
866 : impl FileGuard {
867 : /// Soft deprecation: we'll move VirtualFile to async APIs and remove this function eventually.
868 15359618 : fn with_std_file<F, R>(&self, with: F) -> R
869 15359618 : where
870 15359618 : F: FnOnce(&File) -> R,
871 15359618 : {
872 15359618 : // SAFETY:
873 15359618 : // - lifetime of the fd: `file` doesn't outlive the OwnedFd stored in `self`.
874 15359618 : // - `&` usage below: `self` is `&`, hence Rust typesystem guarantees there are is no `&mut`
875 15359618 : let file = unsafe { File::from_raw_fd(self.as_ref().as_raw_fd()) };
876 15359618 : let res = with(&file);
877 15359618 : let _ = file.into_raw_fd();
878 15359618 : res
879 15359618 : }
880 : /// Soft deprecation: we'll move VirtualFile to async APIs and remove this function eventually.
881 4 : fn with_std_file_mut<F, R>(&mut self, with: F) -> R
882 4 : where
883 4 : F: FnOnce(&mut File) -> R,
884 4 : {
885 4 : // SAFETY:
886 4 : // - lifetime of the fd: `file` doesn't outlive the OwnedFd stored in `self`.
887 4 : // - &mut usage below: `self` is `&mut`, hence this call is the only task/thread that has control over the underlying fd
888 4 : let mut file = unsafe { File::from_raw_fd(self.as_ref().as_raw_fd()) };
889 4 : let res = with(&mut file);
890 4 : let _ = file.into_raw_fd();
891 4 : res
892 4 : }
893 : }
894 :
895 : impl tokio_epoll_uring::IoFd for FileGuard {
896 144825 : unsafe fn as_fd(&self) -> RawFd {
897 144825 : let owned_fd: &OwnedFd = self.as_ref();
898 144825 : owned_fd.as_raw_fd()
899 144825 : }
900 : }
901 :
902 : #[cfg(test)]
903 : impl VirtualFile {
904 20200 : pub(crate) async fn read_blk(
905 20200 : &self,
906 20200 : blknum: u32,
907 20200 : ) -> Result<crate::tenant::block_io::BlockLease<'_>, std::io::Error> {
908 20200 : use crate::page_cache::PAGE_SZ;
909 20200 : let buf = vec![0; PAGE_SZ];
910 20200 : let buf = self
911 20200 : .read_exact_at(buf, blknum as u64 * (PAGE_SZ as u64))
912 10256 : .await?;
913 20200 : Ok(crate::tenant::block_io::BlockLease::Vec(buf))
914 20200 : }
915 :
916 224 : async fn read_to_end(&mut self, buf: &mut Vec<u8>) -> Result<(), Error> {
917 224 : let mut tmp = vec![0; 128];
918 : loop {
919 : let res;
920 444 : (tmp, res) = self.read_at(tmp, self.pos).await;
921 2 : match res {
922 222 : Ok(0) => return Ok(()),
923 220 : Ok(n) => {
924 220 : self.pos += n as u64;
925 220 : buf.extend_from_slice(&tmp[..n]);
926 220 : }
927 2 : Err(ref e) if e.kind() == std::io::ErrorKind::Interrupted => {}
928 2 : Err(e) => return Err(e),
929 : }
930 : }
931 224 : }
932 : }
933 :
934 : impl Drop for VirtualFile {
935 : /// If a VirtualFile is dropped, close the underlying file if it was open.
936 66379 : fn drop(&mut self) {
937 66379 : let handle = self.handle.get_mut();
938 66379 :
939 66379 : fn clean_slot(slot: &Slot, mut slot_guard: RwLockWriteGuard<'_, SlotInner>, tag: u64) {
940 66379 : if slot_guard.tag == tag {
941 66379 : slot.recently_used.store(false, Ordering::Relaxed);
942 66379 : // there is also operation "close-by-replace" for closes done on eviction for
943 66379 : // comparison.
944 66379 : if let Some(fd) = slot_guard.file.take() {
945 55667 : STORAGE_IO_TIME_METRIC
946 55667 : .get(StorageIoOperation::Close)
947 55667 : .observe_closure_duration(|| drop(fd));
948 55667 : }
949 66379 : }
950 66379 : }
951 66379 :
952 66379 : // We don't have async drop so we cannot directly await the lock here.
953 66379 : // Instead, first do a best-effort attempt at closing the underlying
954 66379 : // file descriptor by using `try_write`, and if that fails, spawn
955 66379 : // a tokio task to do it asynchronously: we just want it to be
956 66379 : // cleaned up eventually.
957 66379 : // Most of the time, the `try_lock` should succeed though,
958 66379 : // as we have `&mut self` access. In other words, if the slot
959 66379 : // is still occupied by our file, there should be no access from
960 66379 : // other I/O operations; the only other possible place to lock
961 66379 : // the slot is the lock algorithm looking for free slots.
962 66379 : let slot = &get_open_files().slots[handle.index];
963 66379 : if let Ok(slot_guard) = slot.inner.try_write() {
964 66374 : clean_slot(slot, slot_guard, handle.tag);
965 66374 : } else {
966 5 : let tag = handle.tag;
967 5 : tokio::spawn(async move {
968 5 : let slot_guard = slot.inner.write().await;
969 5 : clean_slot(slot, slot_guard, tag);
970 5 : });
971 5 : };
972 66379 : }
973 : }
974 :
975 : impl OpenFiles {
976 706 : fn new(num_slots: usize) -> OpenFiles {
977 706 : let mut slots = Box::new(Vec::with_capacity(num_slots));
978 61420 : for _ in 0..num_slots {
979 61420 : let slot = Slot {
980 61420 : recently_used: AtomicBool::new(false),
981 61420 : inner: RwLock::new(SlotInner { tag: 0, file: None }),
982 61420 : };
983 61420 : slots.push(slot);
984 61420 : }
985 :
986 706 : OpenFiles {
987 706 : next: AtomicUsize::new(0),
988 706 : slots: Box::leak(slots),
989 706 : }
990 706 : }
991 : }
992 :
993 : ///
994 : /// Initialize the virtual file module. This must be called once at page
995 : /// server startup.
996 : ///
997 : #[cfg(not(test))]
998 604 : pub fn init(num_slots: usize, engine: IoEngineKind) {
999 604 : if OPEN_FILES.set(OpenFiles::new(num_slots)).is_err() {
1000 0 : panic!("virtual_file::init called twice");
1001 604 : }
1002 604 : io_engine::init(engine);
1003 604 : crate::metrics::virtual_file_descriptor_cache::SIZE_MAX.set(num_slots as u64);
1004 604 : }
1005 :
1006 : const TEST_MAX_FILE_DESCRIPTORS: usize = 10;
1007 :
1008 : // Get a handle to the global slots array.
1009 15651074 : fn get_open_files() -> &'static OpenFiles {
1010 15651074 : //
1011 15651074 : // In unit tests, page server startup doesn't happen and no one calls
1012 15651074 : // virtual_file::init(). Initialize it here, with a small array.
1013 15651074 : //
1014 15651074 : // This applies to the virtual file tests below, but all other unit
1015 15651074 : // tests too, so the virtual file facility is always usable in
1016 15651074 : // unit tests.
1017 15651074 : //
1018 15651074 : if cfg!(test) {
1019 370653 : OPEN_FILES.get_or_init(|| OpenFiles::new(TEST_MAX_FILE_DESCRIPTORS))
1020 : } else {
1021 15280421 : OPEN_FILES.get().expect("virtual_file::init not called yet")
1022 : }
1023 15651074 : }
1024 :
1025 : #[cfg(test)]
1026 : mod tests {
1027 : use super::*;
1028 : use rand::seq::SliceRandom;
1029 : use rand::thread_rng;
1030 : use rand::Rng;
1031 : use std::future::Future;
1032 : use std::io::Write;
1033 : use std::sync::Arc;
1034 :
1035 : enum MaybeVirtualFile {
1036 : VirtualFile(VirtualFile),
1037 : File(File),
1038 : }
1039 :
1040 : impl From<VirtualFile> for MaybeVirtualFile {
1041 6 : fn from(vf: VirtualFile) -> Self {
1042 6 : MaybeVirtualFile::VirtualFile(vf)
1043 6 : }
1044 : }
1045 :
1046 : impl MaybeVirtualFile {
1047 404 : async fn read_exact_at(&self, mut buf: Vec<u8>, offset: u64) -> Result<Vec<u8>, Error> {
1048 404 : match self {
1049 203 : MaybeVirtualFile::VirtualFile(file) => file.read_exact_at(buf, offset).await,
1050 202 : MaybeVirtualFile::File(file) => file.read_exact_at(&mut buf, offset).map(|()| buf),
1051 : }
1052 404 : }
1053 8 : async fn write_all_at(&self, buf: &[u8], offset: u64) -> Result<(), Error> {
1054 8 : match self {
1055 4 : MaybeVirtualFile::VirtualFile(file) => file.write_all_at(buf, offset).await,
1056 4 : MaybeVirtualFile::File(file) => file.write_all_at(buf, offset),
1057 : }
1058 8 : }
1059 36 : async fn seek(&mut self, pos: SeekFrom) -> Result<u64, Error> {
1060 36 : match self {
1061 18 : MaybeVirtualFile::VirtualFile(file) => file.seek(pos).await,
1062 18 : MaybeVirtualFile::File(file) => file.seek(pos),
1063 : }
1064 36 : }
1065 8 : async fn write_all(&mut self, buf: &[u8]) -> Result<(), Error> {
1066 8 : match self {
1067 4 : MaybeVirtualFile::VirtualFile(file) => file.write_all(buf).await,
1068 4 : MaybeVirtualFile::File(file) => file.write_all(buf),
1069 : }
1070 8 : }
1071 :
1072 : // Helper function to slurp contents of a file, starting at the current position,
1073 : // into a string
1074 442 : async fn read_string(&mut self) -> Result<String, Error> {
1075 442 : use std::io::Read;
1076 442 : let mut buf = String::new();
1077 442 : match self {
1078 224 : MaybeVirtualFile::VirtualFile(file) => {
1079 224 : let mut buf = Vec::new();
1080 226 : file.read_to_end(&mut buf).await?;
1081 222 : return Ok(String::from_utf8(buf).unwrap());
1082 : }
1083 218 : MaybeVirtualFile::File(file) => {
1084 218 : file.read_to_string(&mut buf)?;
1085 : }
1086 : }
1087 216 : Ok(buf)
1088 442 : }
1089 :
1090 : // Helper function to slurp a portion of a file into a string
1091 404 : async fn read_string_at(&mut self, pos: u64, len: usize) -> Result<String, Error> {
1092 404 : let buf = vec![0; len];
1093 404 : let buf = self.read_exact_at(buf, pos).await?;
1094 404 : Ok(String::from_utf8(buf).unwrap())
1095 404 : }
1096 : }
1097 :
1098 2 : #[tokio::test]
1099 2 : async fn test_virtual_files() -> anyhow::Result<()> {
1100 2 : // The real work is done in the test_files() helper function. This
1101 2 : // allows us to run the same set of tests against a native File, and
1102 2 : // VirtualFile. We trust the native Files and wouldn't need to test them,
1103 2 : // but this allows us to verify that the operations return the same
1104 2 : // results with VirtualFiles as with native Files. (Except that with
1105 2 : // native files, you will run out of file descriptors if the ulimit
1106 2 : // is low enough.)
1107 206 : test_files("virtual_files", |path, open_options| async move {
1108 206 : let vf = VirtualFile::open_with_options(&path, &open_options).await?;
1109 206 : Ok(MaybeVirtualFile::VirtualFile(vf))
1110 206 : })
1111 527 : .await
1112 : }
1113 :
1114 2 : #[tokio::test]
1115 2 : async fn test_physical_files() -> anyhow::Result<()> {
1116 206 : test_files("physical_files", |path, open_options| async move {
1117 206 : Ok(MaybeVirtualFile::File({
1118 206 : let owned_fd = open_options.open(path.as_std_path()).await?;
1119 206 : File::from(owned_fd)
1120 : }))
1121 206 : })
1122 104 : .await
1123 : }
1124 :
1125 4 : async fn test_files<OF, FT>(testname: &str, openfunc: OF) -> anyhow::Result<()>
1126 4 : where
1127 4 : OF: Fn(Utf8PathBuf, OpenOptions) -> FT,
1128 4 : FT: Future<Output = Result<MaybeVirtualFile, std::io::Error>>,
1129 4 : {
1130 4 : let testdir = crate::config::PageServerConf::test_repo_dir(testname);
1131 4 : std::fs::create_dir_all(&testdir)?;
1132 :
1133 4 : let path_a = testdir.join("file_a");
1134 4 : let mut file_a = openfunc(
1135 4 : path_a.clone(),
1136 4 : OpenOptions::new()
1137 4 : .write(true)
1138 4 : .create(true)
1139 4 : .truncate(true)
1140 4 : .to_owned(),
1141 4 : )
1142 4 : .await?;
1143 4 : file_a.write_all(b"foobar").await?;
1144 :
1145 : // cannot read from a file opened in write-only mode
1146 4 : let _ = file_a.read_string().await.unwrap_err();
1147 :
1148 : // Close the file and re-open for reading
1149 4 : let mut file_a = openfunc(path_a, OpenOptions::new().read(true).to_owned()).await?;
1150 :
1151 : // cannot write to a file opened in read-only mode
1152 4 : let _ = file_a.write_all(b"bar").await.unwrap_err();
1153 4 :
1154 4 : // Try simple read
1155 4 : assert_eq!("foobar", file_a.read_string().await?);
1156 :
1157 : // It's positioned at the EOF now.
1158 4 : assert_eq!("", file_a.read_string().await?);
1159 :
1160 : // Test seeks.
1161 4 : assert_eq!(file_a.seek(SeekFrom::Start(1)).await?, 1);
1162 4 : assert_eq!("oobar", file_a.read_string().await?);
1163 :
1164 4 : assert_eq!(file_a.seek(SeekFrom::End(-2)).await?, 4);
1165 4 : assert_eq!("ar", file_a.read_string().await?);
1166 :
1167 4 : assert_eq!(file_a.seek(SeekFrom::Start(1)).await?, 1);
1168 4 : assert_eq!(file_a.seek(SeekFrom::Current(2)).await?, 3);
1169 4 : assert_eq!("bar", file_a.read_string().await?);
1170 :
1171 4 : assert_eq!(file_a.seek(SeekFrom::Current(-5)).await?, 1);
1172 4 : assert_eq!("oobar", file_a.read_string().await?);
1173 :
1174 : // Test erroneous seeks to before byte 0
1175 4 : file_a.seek(SeekFrom::End(-7)).await.unwrap_err();
1176 4 : assert_eq!(file_a.seek(SeekFrom::Start(1)).await?, 1);
1177 4 : file_a.seek(SeekFrom::Current(-2)).await.unwrap_err();
1178 4 :
1179 4 : // the erroneous seek should have left the position unchanged
1180 4 : assert_eq!("oobar", file_a.read_string().await?);
1181 :
1182 : // Create another test file, and try FileExt functions on it.
1183 4 : let path_b = testdir.join("file_b");
1184 4 : let mut file_b = openfunc(
1185 4 : path_b.clone(),
1186 4 : OpenOptions::new()
1187 4 : .read(true)
1188 4 : .write(true)
1189 4 : .create(true)
1190 4 : .truncate(true)
1191 4 : .to_owned(),
1192 4 : )
1193 2 : .await?;
1194 4 : file_b.write_all_at(b"BAR", 3).await?;
1195 4 : file_b.write_all_at(b"FOO", 0).await?;
1196 :
1197 4 : assert_eq!(file_b.read_string_at(2, 3).await?, "OBA");
1198 :
1199 : // Open a lot of files, enough to cause some evictions. (Or to be precise,
1200 : // open the same file many times. The effect is the same.)
1201 : //
1202 : // leave file_a positioned at offset 1 before we start
1203 4 : assert_eq!(file_a.seek(SeekFrom::Start(1)).await?, 1);
1204 :
1205 4 : let mut vfiles = Vec::new();
1206 404 : for _ in 0..100 {
1207 400 : let mut vfile =
1208 400 : openfunc(path_b.clone(), OpenOptions::new().read(true).to_owned()).await?;
1209 400 : assert_eq!("FOOBAR", vfile.read_string().await?);
1210 400 : vfiles.push(vfile);
1211 : }
1212 :
1213 : // make sure we opened enough files to definitely cause evictions.
1214 4 : assert!(vfiles.len() > TEST_MAX_FILE_DESCRIPTORS * 2);
1215 :
1216 : // The underlying file descriptor for 'file_a' should be closed now. Try to read
1217 : // from it again. We left the file positioned at offset 1 above.
1218 4 : assert_eq!("oobar", file_a.read_string().await?);
1219 :
1220 : // Check that all the other FDs still work too. Use them in random order for
1221 : // good measure.
1222 4 : vfiles.as_mut_slice().shuffle(&mut thread_rng());
1223 400 : for vfile in vfiles.iter_mut() {
1224 400 : assert_eq!("OOBAR", vfile.read_string_at(1, 5).await?);
1225 : }
1226 :
1227 4 : Ok(())
1228 4 : }
1229 :
1230 : /// Test using VirtualFiles from many threads concurrently. This tests both using
1231 : /// a lot of VirtualFiles concurrently, causing evictions, and also using the same
1232 : /// VirtualFile from multiple threads concurrently.
1233 2 : #[tokio::test]
1234 2 : async fn test_vfile_concurrency() -> Result<(), Error> {
1235 2 : const SIZE: usize = 8 * 1024;
1236 2 : const VIRTUAL_FILES: usize = 100;
1237 2 : const THREADS: usize = 100;
1238 2 : const SAMPLE: [u8; SIZE] = [0xADu8; SIZE];
1239 2 :
1240 2 : let testdir = crate::config::PageServerConf::test_repo_dir("vfile_concurrency");
1241 2 : std::fs::create_dir_all(&testdir)?;
1242 :
1243 : // Create a test file.
1244 2 : let test_file_path = testdir.join("concurrency_test_file");
1245 : {
1246 2 : let file = File::create(&test_file_path)?;
1247 2 : file.write_all_at(&SAMPLE, 0)?;
1248 : }
1249 :
1250 : // Open the file many times.
1251 2 : let mut files = Vec::new();
1252 202 : for _ in 0..VIRTUAL_FILES {
1253 200 : let f = VirtualFile::open_with_options(&test_file_path, OpenOptions::new().read(true))
1254 101 : .await?;
1255 200 : files.push(f);
1256 : }
1257 2 : let files = Arc::new(files);
1258 2 :
1259 2 : // Launch many threads, and use the virtual files concurrently in random order.
1260 2 : let rt = tokio::runtime::Builder::new_multi_thread()
1261 2 : .worker_threads(THREADS)
1262 2 : .thread_name("test_vfile_concurrency thread")
1263 2 : .build()
1264 2 : .unwrap();
1265 2 : let mut hdls = Vec::new();
1266 202 : for _threadno in 0..THREADS {
1267 200 : let files = files.clone();
1268 200 : let hdl = rt.spawn(async move {
1269 200 : let mut buf = vec![0u8; SIZE];
1270 200 : let mut rng = rand::rngs::OsRng;
1271 200000 : for _ in 1..1000 {
1272 199800 : let f = &files[rng.gen_range(0..files.len())];
1273 569784 : buf = f.read_exact_at(buf, 0).await.unwrap();
1274 199800 : assert!(buf == SAMPLE);
1275 : }
1276 200 : });
1277 200 : hdls.push(hdl);
1278 200 : }
1279 202 : for hdl in hdls {
1280 200 : hdl.await?;
1281 : }
1282 2 : std::mem::forget(rt);
1283 2 :
1284 2 : Ok(())
1285 : }
1286 :
1287 2 : #[tokio::test]
1288 2 : async fn test_atomic_overwrite_basic() {
1289 2 : let testdir = crate::config::PageServerConf::test_repo_dir("test_atomic_overwrite_basic");
1290 2 : std::fs::create_dir_all(&testdir).unwrap();
1291 2 :
1292 2 : let path = testdir.join("myfile");
1293 2 : let tmp_path = testdir.join("myfile.tmp");
1294 2 :
1295 2 : VirtualFile::crashsafe_overwrite(&path, &tmp_path, b"foo")
1296 3 : .await
1297 2 : .unwrap();
1298 2 : let mut file = MaybeVirtualFile::from(VirtualFile::open(&path).await.unwrap());
1299 2 : let post = file.read_string().await.unwrap();
1300 2 : assert_eq!(post, "foo");
1301 2 : assert!(!tmp_path.exists());
1302 2 : drop(file);
1303 2 :
1304 2 : VirtualFile::crashsafe_overwrite(&path, &tmp_path, b"bar")
1305 2 : .await
1306 2 : .unwrap();
1307 2 : let mut file = MaybeVirtualFile::from(VirtualFile::open(&path).await.unwrap());
1308 2 : let post = file.read_string().await.unwrap();
1309 2 : assert_eq!(post, "bar");
1310 2 : assert!(!tmp_path.exists());
1311 2 : drop(file);
1312 : }
1313 :
1314 2 : #[tokio::test]
1315 2 : async fn test_atomic_overwrite_preexisting_tmp() {
1316 2 : let testdir =
1317 2 : crate::config::PageServerConf::test_repo_dir("test_atomic_overwrite_preexisting_tmp");
1318 2 : std::fs::create_dir_all(&testdir).unwrap();
1319 2 :
1320 2 : let path = testdir.join("myfile");
1321 2 : let tmp_path = testdir.join("myfile.tmp");
1322 2 :
1323 2 : std::fs::write(&tmp_path, "some preexisting junk that should be removed").unwrap();
1324 2 : assert!(tmp_path.exists());
1325 :
1326 2 : VirtualFile::crashsafe_overwrite(&path, &tmp_path, b"foo")
1327 3 : .await
1328 2 : .unwrap();
1329 :
1330 2 : let mut file = MaybeVirtualFile::from(VirtualFile::open(&path).await.unwrap());
1331 2 : let post = file.read_string().await.unwrap();
1332 2 : assert_eq!(post, "foo");
1333 2 : assert!(!tmp_path.exists());
1334 2 : drop(file);
1335 : }
1336 : }
|
