Line data Source code
1 : //! VirtualFile is like a normal File, but it's not bound directly to
2 : //! a file descriptor.
3 : //!
4 : //! Instead, the file is opened when it's read from,
5 : //! and if too many files are open globally in the system, least-recently
6 : //! used ones are closed.
7 : //!
8 : //! To track which files have been recently used, we use the clock algorithm
9 : //! with a 'recently_used' flag on each slot.
10 : //!
11 : //! This is similar to PostgreSQL's virtual file descriptor facility in
12 : //! src/backend/storage/file/fd.c
13 : //!
14 : use crate::context::RequestContext;
15 : use crate::metrics::{StorageIoOperation, STORAGE_IO_SIZE, STORAGE_IO_TIME_METRIC};
16 :
17 : use crate::page_cache::{PageWriteGuard, PAGE_SZ};
18 : use crate::tenant::TENANTS_SEGMENT_NAME;
19 : use camino::{Utf8Path, Utf8PathBuf};
20 : use once_cell::sync::OnceCell;
21 : use owned_buffers_io::io_buf_ext::FullSlice;
22 : use pageserver_api::config::defaults::DEFAULT_IO_BUFFER_ALIGNMENT;
23 : use pageserver_api::shard::TenantShardId;
24 : use std::fs::File;
25 : use std::io::{Error, ErrorKind, Seek, SeekFrom};
26 : use tokio_epoll_uring::{BoundedBuf, IoBuf, IoBufMut, Slice};
27 :
28 : use std::os::fd::{AsRawFd, FromRawFd, IntoRawFd, OwnedFd, RawFd};
29 : use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
30 : use tokio::sync::{RwLock, RwLockReadGuard, RwLockWriteGuard};
31 : use tokio::time::Instant;
32 :
33 : pub use pageserver_api::models::virtual_file as api;
34 : pub(crate) mod io_engine;
35 : pub use io_engine::feature_test as io_engine_feature_test;
36 : pub use io_engine::io_engine_for_bench;
37 : pub use io_engine::FeatureTestResult as IoEngineFeatureTestResult;
38 : mod metadata;
39 : mod open_options;
40 : use self::owned_buffers_io::write::OwnedAsyncWriter;
41 : pub(crate) use api::DirectIoMode;
42 : pub(crate) use io_engine::IoEngineKind;
43 : pub(crate) use metadata::Metadata;
44 : pub(crate) use open_options::*;
45 :
46 : pub(crate) mod owned_buffers_io {
47 : //! Abstractions for IO with owned buffers.
48 : //!
49 : //! Not actually tied to [`crate::virtual_file`] specifically, but, it's the primary
50 : //! reason we need this abstraction.
51 : //!
52 : //! Over time, this could move into the `tokio-epoll-uring` crate, maybe `uring-common`,
53 : //! but for the time being we're proving out the primitives in the neon.git repo
54 : //! for faster iteration.
55 :
56 : pub(crate) mod io_buf_ext;
57 : pub(crate) mod slice;
58 : pub(crate) mod write;
59 : pub(crate) mod util {
60 : pub(crate) mod size_tracking_writer;
61 : }
62 : }
63 :
64 : ///
65 : /// A virtual file descriptor. You can use this just like std::fs::File, but internally
66 : /// the underlying file is closed if the system is low on file descriptors,
67 : /// and re-opened when it's accessed again.
68 : ///
69 : /// Like with std::fs::File, multiple threads can read/write the file concurrently,
70 : /// holding just a shared reference the same VirtualFile, using the read_at() / write_at()
71 : /// functions from the FileExt trait. But the functions from the Read/Write/Seek traits
72 : /// require a mutable reference, because they modify the "current position".
73 : ///
74 : /// Each VirtualFile has a physical file descriptor in the global OPEN_FILES array, at the
75 : /// slot that 'handle points to, if the underlying file is currently open. If it's not
76 : /// currently open, the 'handle' can still point to the slot where it was last kept. The
77 : /// 'tag' field is used to detect whether the handle still is valid or not.
78 : ///
79 : #[derive(Debug)]
80 : pub struct VirtualFile {
81 : /// Lazy handle to the global file descriptor cache. The slot that this points to
82 : /// might contain our File, or it may be empty, or it may contain a File that
83 : /// belongs to a different VirtualFile.
84 : handle: RwLock<SlotHandle>,
85 :
86 : /// Current file position
87 : pos: u64,
88 :
89 : /// File path and options to use to open it.
90 : ///
91 : /// Note: this only contains the options needed to re-open it. For example,
92 : /// if a new file is created, we only pass the create flag when it's initially
93 : /// opened, in the VirtualFile::create() function, and strip the flag before
94 : /// storing it here.
95 : pub path: Utf8PathBuf,
96 : open_options: OpenOptions,
97 :
98 : // These are strings becase we only use them for metrics, and those expect strings.
99 : // It makes no sense for us to constantly turn the `TimelineId` and `TenantId` into
100 : // strings.
101 : tenant_id: String,
102 : shard_id: String,
103 : timeline_id: String,
104 : }
105 :
106 : #[derive(Debug, PartialEq, Clone, Copy)]
107 : struct SlotHandle {
108 : /// Index into OPEN_FILES.slots
109 : index: usize,
110 :
111 : /// Value of 'tag' in the slot. If slot's tag doesn't match, then the slot has
112 : /// been recycled and no longer contains the FD for this virtual file.
113 : tag: u64,
114 : }
115 :
116 : /// OPEN_FILES is the global array that holds the physical file descriptors that
117 : /// are currently open. Each slot in the array is protected by a separate lock,
118 : /// so that different files can be accessed independently. The lock must be held
119 : /// in write mode to replace the slot with a different file, but a read mode
120 : /// is enough to operate on the file, whether you're reading or writing to it.
121 : ///
122 : /// OPEN_FILES starts in uninitialized state, and it's initialized by
123 : /// the virtual_file::init() function. It must be called exactly once at page
124 : /// server startup.
125 : static OPEN_FILES: OnceCell<OpenFiles> = OnceCell::new();
126 :
127 : struct OpenFiles {
128 : slots: &'static [Slot],
129 :
130 : /// clock arm for the clock algorithm
131 : next: AtomicUsize,
132 : }
133 :
134 : struct Slot {
135 : inner: RwLock<SlotInner>,
136 :
137 : /// has this file been used since last clock sweep?
138 : recently_used: AtomicBool,
139 : }
140 :
141 : struct SlotInner {
142 : /// Counter that's incremented every time a different file is stored here.
143 : /// To avoid the ABA problem.
144 : tag: u64,
145 :
146 : /// the underlying file
147 : file: Option<OwnedFd>,
148 : }
149 :
150 : /// Impl of [`tokio_epoll_uring::IoBuf`] and [`tokio_epoll_uring::IoBufMut`] for [`PageWriteGuard`].
151 : struct PageWriteGuardBuf {
152 : page: PageWriteGuard<'static>,
153 : }
154 : // Safety: the [`PageWriteGuard`] gives us exclusive ownership of the page cache slot,
155 : // and the location remains stable even if [`Self`] or the [`PageWriteGuard`] is moved.
156 : // Page cache pages are zero-initialized, so, wrt uninitialized memory we're good.
157 : // (Page cache tracks separately whether the contents are valid, see `PageWriteGuard::mark_valid`.)
158 : unsafe impl tokio_epoll_uring::IoBuf for PageWriteGuardBuf {
159 385958 : fn stable_ptr(&self) -> *const u8 {
160 385958 : self.page.as_ptr()
161 385958 : }
162 723660 : fn bytes_init(&self) -> usize {
163 723660 : self.page.len()
164 723660 : }
165 289446 : fn bytes_total(&self) -> usize {
166 289446 : self.page.len()
167 289446 : }
168 : }
169 : // Safety: see above, plus: the ownership of [`PageWriteGuard`] means exclusive access,
170 : // hence it's safe to hand out the `stable_mut_ptr()`.
171 : unsafe impl tokio_epoll_uring::IoBufMut for PageWriteGuardBuf {
172 144738 : fn stable_mut_ptr(&mut self) -> *mut u8 {
173 144738 : self.page.as_mut_ptr()
174 144738 : }
175 :
176 96482 : unsafe fn set_init(&mut self, pos: usize) {
177 96482 : // There shouldn't really be any reason to call this API since bytes_init() == bytes_total().
178 96482 : assert!(pos <= self.page.len());
179 96482 : }
180 : }
181 :
182 : impl OpenFiles {
183 : /// Find a slot to use, evicting an existing file descriptor if needed.
184 : ///
185 : /// On return, we hold a lock on the slot, and its 'tag' has been updated
186 : /// recently_used has been set. It's all ready for reuse.
187 581849 : async fn find_victim_slot(&self) -> (SlotHandle, RwLockWriteGuard<SlotInner>) {
188 581849 : //
189 581849 : // Run the clock algorithm to find a slot to replace.
190 581849 : //
191 581849 : let num_slots = self.slots.len();
192 581849 : let mut retries = 0;
193 : let mut slot;
194 : let mut slot_guard;
195 : let index;
196 : loop {
197 6999953 : let next = self.next.fetch_add(1, Ordering::AcqRel) % num_slots;
198 6999953 : slot = &self.slots[next];
199 6999953 :
200 6999953 : // If the recently_used flag on this slot is set, continue the clock
201 6999953 : // sweep. Otherwise try to use this slot. If we cannot acquire the
202 6999953 : // lock, also continue the clock sweep.
203 6999953 : //
204 6999953 : // We only continue in this manner for a while, though. If we loop
205 6999953 : // through the array twice without finding a victim, just pick the
206 6999953 : // next slot and wait until we can reuse it. This way, we avoid
207 6999953 : // spinning in the extreme case that all the slots are busy with an
208 6999953 : // I/O operation.
209 6999953 : if retries < num_slots * 2 {
210 6735680 : if !slot.recently_used.swap(false, Ordering::Release) {
211 6112698 : if let Ok(guard) = slot.inner.try_write() {
212 317576 : slot_guard = guard;
213 317576 : index = next;
214 317576 : break;
215 5795122 : }
216 622982 : }
217 6418104 : retries += 1;
218 : } else {
219 264273 : slot_guard = slot.inner.write().await;
220 264273 : index = next;
221 264273 : break;
222 : }
223 : }
224 :
225 : //
226 : // We now have the victim slot locked. If it was in use previously, close the
227 : // old file.
228 : //
229 581849 : if let Some(old_file) = slot_guard.file.take() {
230 567626 : // the normal path of dropping VirtualFile uses "close", use "close-by-replace" here to
231 567626 : // distinguish the two.
232 567626 : STORAGE_IO_TIME_METRIC
233 567626 : .get(StorageIoOperation::CloseByReplace)
234 567626 : .observe_closure_duration(|| drop(old_file));
235 567626 : }
236 :
237 : // Prepare the slot for reuse and return it
238 581849 : slot_guard.tag += 1;
239 581849 : slot.recently_used.store(true, Ordering::Relaxed);
240 581849 : (
241 581849 : SlotHandle {
242 581849 : index,
243 581849 : tag: slot_guard.tag,
244 581849 : },
245 581849 : slot_guard,
246 581849 : )
247 581849 : }
248 : }
249 :
250 : /// Identify error types that should alwways terminate the process. Other
251 : /// error types may be elegible for retry.
252 6 : pub(crate) fn is_fatal_io_error(e: &std::io::Error) -> bool {
253 : use nix::errno::Errno::*;
254 6 : match e.raw_os_error().map(nix::errno::from_i32) {
255 : Some(EIO) => {
256 : // Terminate on EIO because we no longer trust the device to store
257 : // data safely, or to uphold persistence guarantees on fsync.
258 0 : true
259 : }
260 : Some(EROFS) => {
261 : // Terminate on EROFS because a filesystem is usually remounted
262 : // readonly when it has experienced some critical issue, so the same
263 : // logic as EIO applies.
264 0 : true
265 : }
266 : Some(EACCES) => {
267 : // Terminate on EACCESS because we should always have permissions
268 : // for our own data dir: if we don't, then we can't do our job and
269 : // need administrative intervention to fix permissions. Terminating
270 : // is the best way to make sure we stop cleanly rather than going
271 : // into infinite retry loops, and will make it clear to the outside
272 : // world that we need help.
273 0 : true
274 : }
275 : _ => {
276 : // Treat all other local file I/O errors are retryable. This includes:
277 : // - ENOSPC: we stay up and wait for eviction to free some space
278 : // - EINVAL, EBADF, EBADFD: this is a code bug, not a filesystem/hardware issue
279 : // - WriteZero, Interrupted: these are used internally VirtualFile
280 6 : false
281 : }
282 : }
283 6 : }
284 :
285 : /// Call this when the local filesystem gives us an error with an external
286 : /// cause: this includes EIO, EROFS, and EACCESS: all these indicate either
287 : /// bad storage or bad configuration, and we can't fix that from inside
288 : /// a running process.
289 0 : pub(crate) fn on_fatal_io_error(e: &std::io::Error, context: &str) -> ! {
290 0 : tracing::error!("Fatal I/O error: {e}: {context})");
291 0 : std::process::abort();
292 : }
293 :
294 : pub(crate) trait MaybeFatalIo<T> {
295 : fn maybe_fatal_err(self, context: &str) -> std::io::Result<T>;
296 : fn fatal_err(self, context: &str) -> T;
297 : }
298 :
299 : impl<T> MaybeFatalIo<T> for std::io::Result<T> {
300 : /// Terminate the process if the result is an error of a fatal type, else pass it through
301 : ///
302 : /// This is appropriate for writes, where we typically want to die on EIO/ACCES etc, but
303 : /// not on ENOSPC.
304 3407564 : fn maybe_fatal_err(self, context: &str) -> std::io::Result<T> {
305 3407564 : if let Err(e) = &self {
306 6 : if is_fatal_io_error(e) {
307 0 : on_fatal_io_error(e, context);
308 6 : }
309 3407558 : }
310 3407564 : self
311 3407564 : }
312 :
313 : /// Terminate the process on any I/O error.
314 : ///
315 : /// This is appropriate for reads on files that we know exist: they should always work.
316 6114 : fn fatal_err(self, context: &str) -> T {
317 6114 : match self {
318 6114 : Ok(v) => v,
319 0 : Err(e) => {
320 0 : on_fatal_io_error(&e, context);
321 : }
322 : }
323 6114 : }
324 : }
325 :
326 : /// Observe duration for the given storage I/O operation
327 : ///
328 : /// Unlike `observe_closure_duration`, this supports async,
329 : /// where "support" means that we measure wall clock time.
330 : macro_rules! observe_duration {
331 : ($op:expr, $($body:tt)*) => {{
332 : let instant = Instant::now();
333 : let result = $($body)*;
334 : let elapsed = instant.elapsed().as_secs_f64();
335 : STORAGE_IO_TIME_METRIC
336 : .get($op)
337 : .observe(elapsed);
338 : result
339 : }}
340 : }
341 :
342 : macro_rules! with_file {
343 : ($this:expr, $op:expr, | $ident:ident | $($body:tt)*) => {{
344 : let $ident = $this.lock_file().await?;
345 : observe_duration!($op, $($body)*)
346 : }};
347 : ($this:expr, $op:expr, | mut $ident:ident | $($body:tt)*) => {{
348 : let mut $ident = $this.lock_file().await?;
349 : observe_duration!($op, $($body)*)
350 : }};
351 : }
352 :
353 : impl VirtualFile {
354 : /// Open a file in read-only mode. Like File::open.
355 6594 : pub async fn open<P: AsRef<Utf8Path>>(
356 6594 : path: P,
357 6594 : ctx: &RequestContext,
358 6594 : ) -> Result<VirtualFile, std::io::Error> {
359 6594 : Self::open_with_options(path.as_ref(), OpenOptions::new().read(true), ctx).await
360 6594 : }
361 :
362 : /// Create a new file for writing. If the file exists, it will be truncated.
363 : /// Like File::create.
364 4371 : pub async fn create<P: AsRef<Utf8Path>>(
365 4371 : path: P,
366 4371 : ctx: &RequestContext,
367 4371 : ) -> Result<VirtualFile, std::io::Error> {
368 4371 : Self::open_with_options(
369 4371 : path.as_ref(),
370 4371 : OpenOptions::new().write(true).create(true).truncate(true),
371 4371 : ctx,
372 4371 : )
373 2247 : .await
374 4371 : }
375 :
376 : /// Open a file with given options.
377 : ///
378 : /// Note: If any custom flags were set in 'open_options' through OpenOptionsExt,
379 : /// they will be applied also when the file is subsequently re-opened, not only
380 : /// on the first time. Make sure that's sane!
381 17517 : pub async fn open_with_options<P: AsRef<Utf8Path>>(
382 17517 : path: P,
383 17517 : open_options: &OpenOptions,
384 17517 : _ctx: &RequestContext, /* TODO: carry a pointer to the metrics in the RequestContext instead of the parsing https://github.com/neondatabase/neon/issues/6107 */
385 17517 : ) -> Result<VirtualFile, std::io::Error> {
386 17517 : let path_ref = path.as_ref();
387 17517 : let path_str = path_ref.to_string();
388 17517 : let parts = path_str.split('/').collect::<Vec<&str>>();
389 17517 : let (tenant_id, shard_id, timeline_id) =
390 17517 : if parts.len() > 5 && parts[parts.len() - 5] == TENANTS_SEGMENT_NAME {
391 13035 : let tenant_shard_part = parts[parts.len() - 4];
392 13035 : let (tenant_id, shard_id) = match tenant_shard_part.parse::<TenantShardId>() {
393 13035 : Ok(tenant_shard_id) => (
394 13035 : tenant_shard_id.tenant_id.to_string(),
395 13035 : format!("{}", tenant_shard_id.shard_slug()),
396 13035 : ),
397 : Err(_) => {
398 : // Malformed path: this ID is just for observability, so tolerate it
399 : // and pass through
400 0 : (tenant_shard_part.to_string(), "*".to_string())
401 : }
402 : };
403 13035 : (tenant_id, shard_id, parts[parts.len() - 2].to_string())
404 : } else {
405 4482 : ("*".to_string(), "*".to_string(), "*".to_string())
406 : };
407 17517 : let (handle, mut slot_guard) = get_open_files().find_victim_slot().await;
408 :
409 : // NB: there is also StorageIoOperation::OpenAfterReplace which is for the case
410 : // where our caller doesn't get to use the returned VirtualFile before its
411 : // slot gets re-used by someone else.
412 17517 : let file = observe_duration!(StorageIoOperation::Open, {
413 17517 : open_options.open(path_ref.as_std_path()).await?
414 : });
415 :
416 : // Strip all options other than read and write.
417 : //
418 : // It would perhaps be nicer to check just for the read and write flags
419 : // explicitly, but OpenOptions doesn't contain any functions to read flags,
420 : // only to set them.
421 17517 : let mut reopen_options = open_options.clone();
422 17517 : reopen_options.create(false);
423 17517 : reopen_options.create_new(false);
424 17517 : reopen_options.truncate(false);
425 17517 :
426 17517 : let vfile = VirtualFile {
427 17517 : handle: RwLock::new(handle),
428 17517 : pos: 0,
429 17517 : path: path_ref.to_path_buf(),
430 17517 : open_options: reopen_options,
431 17517 : tenant_id,
432 17517 : shard_id,
433 17517 : timeline_id,
434 17517 : };
435 17517 :
436 17517 : // TODO: Under pressure, it's likely the slot will get re-used and
437 17517 : // the underlying file closed before they get around to using it.
438 17517 : // => https://github.com/neondatabase/neon/issues/6065
439 17517 : slot_guard.file.replace(file);
440 17517 :
441 17517 : Ok(vfile)
442 17517 : }
443 :
444 : /// Async version of [`::utils::crashsafe::overwrite`].
445 : ///
446 : /// # NB:
447 : ///
448 : /// Doesn't actually use the [`VirtualFile`] file descriptor cache, but,
449 : /// it did at an earlier time.
450 : /// And it will use this module's [`io_engine`] in the near future, so, leaving it here.
451 84 : pub async fn crashsafe_overwrite<B: BoundedBuf<Buf = Buf> + Send, Buf: IoBuf + Send>(
452 84 : final_path: Utf8PathBuf,
453 84 : tmp_path: Utf8PathBuf,
454 84 : content: B,
455 84 : ) -> std::io::Result<()> {
456 84 : // TODO: use tokio_epoll_uring if configured as `io_engine`.
457 84 : // See https://github.com/neondatabase/neon/issues/6663
458 84 :
459 84 : tokio::task::spawn_blocking(move || {
460 84 : let slice_storage;
461 84 : let content_len = content.bytes_init();
462 84 : let content = if content.bytes_init() > 0 {
463 84 : slice_storage = Some(content.slice(0..content_len));
464 84 : slice_storage.as_deref().expect("just set it to Some()")
465 : } else {
466 0 : &[]
467 : };
468 84 : utils::crashsafe::overwrite(&final_path, &tmp_path, content)
469 84 : })
470 84 : .await
471 84 : .expect("blocking task is never aborted")
472 84 : }
473 :
474 : /// Call File::sync_all() on the underlying File.
475 8139 : pub async fn sync_all(&self) -> Result<(), Error> {
476 8139 : with_file!(self, StorageIoOperation::Fsync, |file_guard| {
477 8139 : let (_file_guard, res) = io_engine::get().sync_all(file_guard).await;
478 8139 : res
479 : })
480 8139 : }
481 :
482 : /// Call File::sync_data() on the underlying File.
483 0 : pub async fn sync_data(&self) -> Result<(), Error> {
484 0 : with_file!(self, StorageIoOperation::Fsync, |file_guard| {
485 0 : let (_file_guard, res) = io_engine::get().sync_data(file_guard).await;
486 0 : res
487 : })
488 0 : }
489 :
490 5130 : pub async fn metadata(&self) -> Result<Metadata, Error> {
491 5130 : with_file!(self, StorageIoOperation::Metadata, |file_guard| {
492 5130 : let (_file_guard, res) = io_engine::get().metadata(file_guard).await;
493 5130 : res
494 : })
495 5130 : }
496 :
497 : /// Helper function internal to `VirtualFile` that looks up the underlying File,
498 : /// opens it and evicts some other File if necessary. The passed parameter is
499 : /// assumed to be a function available for the physical `File`.
500 : ///
501 : /// We are doing it via a macro as Rust doesn't support async closures that
502 : /// take on parameters with lifetimes.
503 5706981 : async fn lock_file(&self) -> Result<FileGuard, Error> {
504 5706981 : let open_files = get_open_files();
505 :
506 564332 : let mut handle_guard = {
507 : // Read the cached slot handle, and see if the slot that it points to still
508 : // contains our File.
509 : //
510 : // We only need to hold the handle lock while we read the current handle. If
511 : // another thread closes the file and recycles the slot for a different file,
512 : // we will notice that the handle we read is no longer valid and retry.
513 5706981 : let mut handle = *self.handle.read().await;
514 : loop {
515 : // Check if the slot contains our File
516 : {
517 6001295 : let slot = &open_files.slots[handle.index];
518 6001295 : let slot_guard = slot.inner.read().await;
519 6001295 : if slot_guard.tag == handle.tag && slot_guard.file.is_some() {
520 : // Found a cached file descriptor.
521 5142649 : slot.recently_used.store(true, Ordering::Relaxed);
522 5142649 : return Ok(FileGuard { slot_guard });
523 858646 : }
524 : }
525 :
526 : // The slot didn't contain our File. We will have to open it ourselves,
527 : // but before that, grab a write lock on handle in the VirtualFile, so
528 : // that no other thread will try to concurrently open the same file.
529 858646 : let handle_guard = self.handle.write().await;
530 :
531 : // If another thread changed the handle while we were not holding the lock,
532 : // then the handle might now be valid again. Loop back to retry.
533 858646 : if *handle_guard != handle {
534 294314 : handle = *handle_guard;
535 294314 : continue;
536 564332 : }
537 564332 : break handle_guard;
538 : }
539 : };
540 :
541 : // We need to open the file ourselves. The handle in the VirtualFile is
542 : // now locked in write-mode. Find a free slot to put it in.
543 564332 : let (handle, mut slot_guard) = open_files.find_victim_slot().await;
544 :
545 : // Re-open the physical file.
546 : // NB: we use StorageIoOperation::OpenAferReplace for this to distinguish this
547 : // case from StorageIoOperation::Open. This helps with identifying thrashing
548 : // of the virtual file descriptor cache.
549 564332 : let file = observe_duration!(StorageIoOperation::OpenAfterReplace, {
550 564332 : self.open_options.open(self.path.as_std_path()).await?
551 : });
552 :
553 : // Store the File in the slot and update the handle in the VirtualFile
554 : // to point to it.
555 564332 : slot_guard.file.replace(file);
556 564332 :
557 564332 : *handle_guard = handle;
558 564332 :
559 564332 : return Ok(FileGuard {
560 564332 : slot_guard: slot_guard.downgrade(),
561 564332 : });
562 5706981 : }
563 :
564 612 : pub fn remove(self) {
565 612 : let path = self.path.clone();
566 612 : drop(self);
567 612 : std::fs::remove_file(path).expect("failed to remove the virtual file");
568 612 : }
569 :
570 16038 : pub async fn seek(&mut self, pos: SeekFrom) -> Result<u64, Error> {
571 16038 : match pos {
572 16008 : SeekFrom::Start(offset) => {
573 16008 : self.pos = offset;
574 16008 : }
575 12 : SeekFrom::End(offset) => {
576 12 : self.pos = with_file!(self, StorageIoOperation::Seek, |mut file_guard| file_guard
577 12 : .with_std_file_mut(|std_file| std_file.seek(SeekFrom::End(offset))))?
578 : }
579 18 : SeekFrom::Current(offset) => {
580 18 : let pos = self.pos as i128 + offset as i128;
581 18 : if pos < 0 {
582 6 : return Err(Error::new(
583 6 : ErrorKind::InvalidInput,
584 6 : "offset would be negative",
585 6 : ));
586 12 : }
587 12 : if pos > u64::MAX as i128 {
588 0 : return Err(Error::new(ErrorKind::InvalidInput, "offset overflow"));
589 12 : }
590 12 : self.pos = pos as u64;
591 : }
592 : }
593 16026 : Ok(self.pos)
594 16038 : }
595 :
596 : /// Read the file contents in range `offset..(offset + slice.bytes_total())` into `slice[0..slice.bytes_total()]`.
597 : ///
598 : /// The returned `Slice<Buf>` is equivalent to the input `slice`, i.e., it's the same view into the same buffer.
599 2284870 : pub async fn read_exact_at<Buf>(
600 2284870 : &self,
601 2284870 : slice: Slice<Buf>,
602 2284870 : offset: u64,
603 2284870 : ctx: &RequestContext,
604 2284870 : ) -> Result<Slice<Buf>, Error>
605 2284870 : where
606 2284870 : Buf: IoBufMut + Send,
607 2284870 : {
608 2284870 : let assert_we_return_original_bounds = if cfg!(debug_assertions) {
609 2284870 : Some((slice.stable_ptr() as usize, slice.bytes_total()))
610 : } else {
611 0 : None
612 : };
613 :
614 2284870 : let original_bounds = slice.bounds();
615 2284870 : let (buf, res) =
616 2658628 : read_exact_at_impl(slice, offset, |buf, offset| self.read_at(buf, offset, ctx)).await;
617 2284870 : let res = res.map(|_| buf.slice(original_bounds));
618 :
619 2284870 : if let Some(original_bounds) = assert_we_return_original_bounds {
620 2284870 : if let Ok(slice) = &res {
621 2284870 : let returned_bounds = (slice.stable_ptr() as usize, slice.bytes_total());
622 2284870 : assert_eq!(original_bounds, returned_bounds);
623 0 : }
624 0 : }
625 :
626 2284870 : res
627 2284870 : }
628 :
629 : /// Like [`Self::read_exact_at`] but for [`PageWriteGuard`].
630 96482 : pub async fn read_exact_at_page(
631 96482 : &self,
632 96482 : page: PageWriteGuard<'static>,
633 96482 : offset: u64,
634 96482 : ctx: &RequestContext,
635 96482 : ) -> Result<PageWriteGuard<'static>, Error> {
636 96482 : let buf = PageWriteGuardBuf { page }.slice_full();
637 96482 : debug_assert_eq!(buf.bytes_total(), PAGE_SZ);
638 96482 : self.read_exact_at(buf, offset, ctx)
639 65822 : .await
640 96482 : .map(|slice| slice.into_inner().page)
641 96482 : }
642 :
643 : // Copied from https://doc.rust-lang.org/1.72.0/src/std/os/unix/fs.rs.html#219-235
644 12 : pub async fn write_all_at<Buf: IoBuf + Send>(
645 12 : &self,
646 12 : buf: FullSlice<Buf>,
647 12 : mut offset: u64,
648 12 : ctx: &RequestContext,
649 12 : ) -> (FullSlice<Buf>, Result<(), Error>) {
650 12 : let buf = buf.into_raw_slice();
651 12 : let bounds = buf.bounds();
652 12 : let restore =
653 12 : |buf: Slice<_>| FullSlice::must_new(Slice::from_buf_bounds(buf.into_inner(), bounds));
654 12 : let mut buf = buf;
655 24 : while !buf.is_empty() {
656 12 : let (tmp, res) = self.write_at(FullSlice::must_new(buf), offset, ctx).await;
657 12 : buf = tmp.into_raw_slice();
658 0 : match res {
659 : Ok(0) => {
660 0 : return (
661 0 : restore(buf),
662 0 : Err(Error::new(
663 0 : std::io::ErrorKind::WriteZero,
664 0 : "failed to write whole buffer",
665 0 : )),
666 0 : );
667 : }
668 12 : Ok(n) => {
669 12 : buf = buf.slice(n..);
670 12 : offset += n as u64;
671 12 : }
672 0 : Err(e) if e.kind() == std::io::ErrorKind::Interrupted => {}
673 0 : Err(e) => return (restore(buf), Err(e)),
674 : }
675 : }
676 12 : (restore(buf), Ok(()))
677 12 : }
678 :
679 : /// Writes `buf` to the file at the current offset.
680 : ///
681 : /// Panics if there is an uninitialized range in `buf`, as that is most likely a bug in the caller.
682 3407600 : pub async fn write_all<Buf: IoBuf + Send>(
683 3407600 : &mut self,
684 3407600 : buf: FullSlice<Buf>,
685 3407600 : ctx: &RequestContext,
686 3407600 : ) -> (FullSlice<Buf>, Result<usize, Error>) {
687 3407600 : let buf = buf.into_raw_slice();
688 3407600 : let bounds = buf.bounds();
689 3407600 : let restore =
690 3407600 : |buf: Slice<_>| FullSlice::must_new(Slice::from_buf_bounds(buf.into_inner(), bounds));
691 3407600 : let nbytes = buf.len();
692 3407600 : let mut buf = buf;
693 6815080 : while !buf.is_empty() {
694 3407486 : let (tmp, res) = self.write(FullSlice::must_new(buf), ctx).await;
695 3407486 : buf = tmp.into_raw_slice();
696 6 : match res {
697 : Ok(0) => {
698 0 : return (
699 0 : restore(buf),
700 0 : Err(Error::new(
701 0 : std::io::ErrorKind::WriteZero,
702 0 : "failed to write whole buffer",
703 0 : )),
704 0 : );
705 : }
706 3407480 : Ok(n) => {
707 3407480 : buf = buf.slice(n..);
708 3407480 : }
709 6 : Err(ref e) if e.kind() == std::io::ErrorKind::Interrupted => {}
710 6 : Err(e) => return (restore(buf), Err(e)),
711 : }
712 : }
713 3407594 : (restore(buf), Ok(nbytes))
714 3407600 : }
715 :
716 3407486 : async fn write<B: IoBuf + Send>(
717 3407486 : &mut self,
718 3407486 : buf: FullSlice<B>,
719 3407486 : ctx: &RequestContext,
720 3407486 : ) -> (FullSlice<B>, Result<usize, std::io::Error>) {
721 3407486 : let pos = self.pos;
722 3407486 : let (buf, res) = self.write_at(buf, pos, ctx).await;
723 3407486 : let n = match res {
724 3407480 : Ok(n) => n,
725 6 : Err(e) => return (buf, Err(e)),
726 : };
727 3407480 : self.pos += n as u64;
728 3407480 : (buf, Ok(n))
729 3407486 : }
730 :
731 2286202 : pub(crate) async fn read_at<Buf>(
732 2286202 : &self,
733 2286202 : buf: tokio_epoll_uring::Slice<Buf>,
734 2286202 : offset: u64,
735 2286202 : _ctx: &RequestContext, /* TODO: use for metrics: https://github.com/neondatabase/neon/issues/6107 */
736 2286202 : ) -> (tokio_epoll_uring::Slice<Buf>, Result<usize, Error>)
737 2286202 : where
738 2286202 : Buf: tokio_epoll_uring::IoBufMut + Send,
739 2286202 : {
740 2286202 : let file_guard = match self.lock_file().await {
741 2286202 : Ok(file_guard) => file_guard,
742 0 : Err(e) => return (buf, Err(e)),
743 : };
744 :
745 2286202 : observe_duration!(StorageIoOperation::Read, {
746 2286202 : let ((_file_guard, buf), res) = io_engine::get().read_at(file_guard, offset, buf).await;
747 2286202 : if let Ok(size) = res {
748 2286196 : STORAGE_IO_SIZE
749 2286196 : .with_label_values(&[
750 2286196 : "read",
751 2286196 : &self.tenant_id,
752 2286196 : &self.shard_id,
753 2286196 : &self.timeline_id,
754 2286196 : ])
755 2286196 : .add(size as i64);
756 2286196 : }
757 2286202 : (buf, res)
758 : })
759 2286202 : }
760 :
761 : /// The function aborts the process if the error is fatal.
762 3407498 : async fn write_at<B: IoBuf + Send>(
763 3407498 : &self,
764 3407498 : buf: FullSlice<B>,
765 3407498 : offset: u64,
766 3407498 : _ctx: &RequestContext, /* TODO: use for metrics: https://github.com/neondatabase/neon/issues/6107 */
767 3407498 : ) -> (FullSlice<B>, Result<usize, Error>) {
768 3407498 : let (slice, result) = self.write_at_inner(buf, offset, _ctx).await;
769 3407498 : let result = result.maybe_fatal_err("write_at");
770 3407498 : (slice, result)
771 3407498 : }
772 :
773 3407498 : async fn write_at_inner<B: IoBuf + Send>(
774 3407498 : &self,
775 3407498 : buf: FullSlice<B>,
776 3407498 : offset: u64,
777 3407498 : _ctx: &RequestContext, /* TODO: use for metrics: https://github.com/neondatabase/neon/issues/6107 */
778 3407498 : ) -> (FullSlice<B>, Result<usize, Error>) {
779 3407498 : let file_guard = match self.lock_file().await {
780 3407498 : Ok(file_guard) => file_guard,
781 0 : Err(e) => return (buf, Err(e)),
782 : };
783 3407498 : observe_duration!(StorageIoOperation::Write, {
784 3407498 : let ((_file_guard, buf), result) =
785 3407498 : io_engine::get().write_at(file_guard, offset, buf).await;
786 3407498 : if let Ok(size) = result {
787 3407492 : STORAGE_IO_SIZE
788 3407492 : .with_label_values(&[
789 3407492 : "write",
790 3407492 : &self.tenant_id,
791 3407492 : &self.shard_id,
792 3407492 : &self.timeline_id,
793 3407492 : ])
794 3407492 : .add(size as i64);
795 3407492 : }
796 3407498 : (buf, result)
797 : })
798 3407498 : }
799 : }
800 :
801 : // Adapted from https://doc.rust-lang.org/1.72.0/src/std/os/unix/fs.rs.html#117-135
802 2284894 : pub async fn read_exact_at_impl<Buf, F, Fut>(
803 2284894 : mut buf: tokio_epoll_uring::Slice<Buf>,
804 2284894 : mut offset: u64,
805 2284894 : mut read_at: F,
806 2284894 : ) -> (Buf, std::io::Result<()>)
807 2284894 : where
808 2284894 : Buf: IoBufMut + Send,
809 2284894 : F: FnMut(tokio_epoll_uring::Slice<Buf>, u64) -> Fut,
810 2284894 : Fut: std::future::Future<Output = (tokio_epoll_uring::Slice<Buf>, std::io::Result<usize>)>,
811 2284894 : {
812 4569794 : while buf.bytes_total() != 0 {
813 : let res;
814 2658628 : (buf, res) = read_at(buf, offset).await;
815 0 : match res {
816 6 : Ok(0) => break,
817 2284900 : Ok(n) => {
818 2284900 : buf = buf.slice(n..);
819 2284900 : offset += n as u64;
820 2284900 : }
821 0 : Err(ref e) if e.kind() == std::io::ErrorKind::Interrupted => {}
822 0 : Err(e) => return (buf.into_inner(), Err(e)),
823 : }
824 : }
825 : // NB: don't use `buf.is_empty()` here; it is from the
826 : // `impl Deref for Slice { Target = [u8] }`; the &[u8]
827 : // returned by it only covers the initialized portion of `buf`.
828 : // Whereas we're interested in ensuring that we filled the entire
829 : // buffer that the user passed in.
830 2284894 : if buf.bytes_total() != 0 {
831 6 : (
832 6 : buf.into_inner(),
833 6 : Err(std::io::Error::new(
834 6 : std::io::ErrorKind::UnexpectedEof,
835 6 : "failed to fill whole buffer",
836 6 : )),
837 6 : )
838 : } else {
839 2284888 : assert_eq!(buf.len(), buf.bytes_total());
840 2284888 : (buf.into_inner(), Ok(()))
841 : }
842 2284894 : }
843 :
844 : #[cfg(test)]
845 : mod test_read_exact_at_impl {
846 :
847 : use std::{collections::VecDeque, sync::Arc};
848 :
849 : use tokio_epoll_uring::{BoundedBuf, BoundedBufMut};
850 :
851 : use super::read_exact_at_impl;
852 :
853 : struct Expectation {
854 : offset: u64,
855 : bytes_total: usize,
856 : result: std::io::Result<Vec<u8>>,
857 : }
858 : struct MockReadAt {
859 : expectations: VecDeque<Expectation>,
860 : }
861 :
862 : impl MockReadAt {
863 36 : async fn read_at(
864 36 : &mut self,
865 36 : mut buf: tokio_epoll_uring::Slice<Vec<u8>>,
866 36 : offset: u64,
867 36 : ) -> (tokio_epoll_uring::Slice<Vec<u8>>, std::io::Result<usize>) {
868 36 : let exp = self
869 36 : .expectations
870 36 : .pop_front()
871 36 : .expect("read_at called but we have no expectations left");
872 36 : assert_eq!(exp.offset, offset);
873 36 : assert_eq!(exp.bytes_total, buf.bytes_total());
874 36 : match exp.result {
875 36 : Ok(bytes) => {
876 36 : assert!(bytes.len() <= buf.bytes_total());
877 36 : buf.put_slice(&bytes);
878 36 : (buf, Ok(bytes.len()))
879 : }
880 0 : Err(e) => (buf, Err(e)),
881 : }
882 36 : }
883 : }
884 :
885 : impl Drop for MockReadAt {
886 24 : fn drop(&mut self) {
887 24 : assert_eq!(self.expectations.len(), 0);
888 24 : }
889 : }
890 :
891 : #[tokio::test]
892 6 : async fn test_basic() {
893 6 : let buf = Vec::with_capacity(5).slice_full();
894 6 : let mock_read_at = Arc::new(tokio::sync::Mutex::new(MockReadAt {
895 6 : expectations: VecDeque::from(vec![Expectation {
896 6 : offset: 0,
897 6 : bytes_total: 5,
898 6 : result: Ok(vec![b'a', b'b', b'c', b'd', b'e']),
899 6 : }]),
900 6 : }));
901 6 : let (buf, res) = read_exact_at_impl(buf, 0, |buf, offset| {
902 6 : let mock_read_at = Arc::clone(&mock_read_at);
903 6 : async move { mock_read_at.lock().await.read_at(buf, offset).await }
904 6 : })
905 6 : .await;
906 6 : assert!(res.is_ok());
907 6 : assert_eq!(buf, vec![b'a', b'b', b'c', b'd', b'e']);
908 6 : }
909 :
910 : #[tokio::test]
911 6 : async fn test_empty_buf_issues_no_syscall() {
912 6 : let buf = Vec::new().slice_full();
913 6 : let mock_read_at = Arc::new(tokio::sync::Mutex::new(MockReadAt {
914 6 : expectations: VecDeque::new(),
915 6 : }));
916 6 : let (_buf, res) = read_exact_at_impl(buf, 0, |buf, offset| {
917 0 : let mock_read_at = Arc::clone(&mock_read_at);
918 6 : async move { mock_read_at.lock().await.read_at(buf, offset).await }
919 6 : })
920 6 : .await;
921 6 : assert!(res.is_ok());
922 6 : }
923 :
924 : #[tokio::test]
925 6 : async fn test_two_read_at_calls_needed_until_buf_filled() {
926 6 : let buf = Vec::with_capacity(4).slice_full();
927 6 : let mock_read_at = Arc::new(tokio::sync::Mutex::new(MockReadAt {
928 6 : expectations: VecDeque::from(vec![
929 6 : Expectation {
930 6 : offset: 0,
931 6 : bytes_total: 4,
932 6 : result: Ok(vec![b'a', b'b']),
933 6 : },
934 6 : Expectation {
935 6 : offset: 2,
936 6 : bytes_total: 2,
937 6 : result: Ok(vec![b'c', b'd']),
938 6 : },
939 6 : ]),
940 6 : }));
941 12 : let (buf, res) = read_exact_at_impl(buf, 0, |buf, offset| {
942 12 : let mock_read_at = Arc::clone(&mock_read_at);
943 12 : async move { mock_read_at.lock().await.read_at(buf, offset).await }
944 12 : })
945 6 : .await;
946 6 : assert!(res.is_ok());
947 6 : assert_eq!(buf, vec![b'a', b'b', b'c', b'd']);
948 6 : }
949 :
950 : #[tokio::test]
951 6 : async fn test_eof_before_buffer_full() {
952 6 : let buf = Vec::with_capacity(3).slice_full();
953 6 : let mock_read_at = Arc::new(tokio::sync::Mutex::new(MockReadAt {
954 6 : expectations: VecDeque::from(vec![
955 6 : Expectation {
956 6 : offset: 0,
957 6 : bytes_total: 3,
958 6 : result: Ok(vec![b'a']),
959 6 : },
960 6 : Expectation {
961 6 : offset: 1,
962 6 : bytes_total: 2,
963 6 : result: Ok(vec![b'b']),
964 6 : },
965 6 : Expectation {
966 6 : offset: 2,
967 6 : bytes_total: 1,
968 6 : result: Ok(vec![]),
969 6 : },
970 6 : ]),
971 6 : }));
972 18 : let (_buf, res) = read_exact_at_impl(buf, 0, |buf, offset| {
973 18 : let mock_read_at = Arc::clone(&mock_read_at);
974 18 : async move { mock_read_at.lock().await.read_at(buf, offset).await }
975 18 : })
976 6 : .await;
977 6 : let Err(err) = res else {
978 6 : panic!("should return an error");
979 6 : };
980 6 : assert_eq!(err.kind(), std::io::ErrorKind::UnexpectedEof);
981 6 : assert_eq!(format!("{err}"), "failed to fill whole buffer");
982 6 : // buffer contents on error are unspecified
983 6 : }
984 : }
985 :
986 : struct FileGuard {
987 : slot_guard: RwLockReadGuard<'static, SlotInner>,
988 : }
989 :
990 : impl AsRef<OwnedFd> for FileGuard {
991 5706981 : fn as_ref(&self) -> &OwnedFd {
992 5706981 : // This unwrap is safe because we only create `FileGuard`s
993 5706981 : // if we know that the file is Some.
994 5706981 : self.slot_guard.file.as_ref().unwrap()
995 5706981 : }
996 : }
997 :
998 : impl FileGuard {
999 : /// Soft deprecation: we'll move VirtualFile to async APIs and remove this function eventually.
1000 2853466 : fn with_std_file<F, R>(&self, with: F) -> R
1001 2853466 : where
1002 2853466 : F: FnOnce(&File) -> R,
1003 2853466 : {
1004 2853466 : // SAFETY:
1005 2853466 : // - lifetime of the fd: `file` doesn't outlive the OwnedFd stored in `self`.
1006 2853466 : // - `&` usage below: `self` is `&`, hence Rust typesystem guarantees there are is no `&mut`
1007 2853466 : let file = unsafe { File::from_raw_fd(self.as_ref().as_raw_fd()) };
1008 2853466 : let res = with(&file);
1009 2853466 : let _ = file.into_raw_fd();
1010 2853466 : res
1011 2853466 : }
1012 : /// Soft deprecation: we'll move VirtualFile to async APIs and remove this function eventually.
1013 12 : fn with_std_file_mut<F, R>(&mut self, with: F) -> R
1014 12 : where
1015 12 : F: FnOnce(&mut File) -> R,
1016 12 : {
1017 12 : // SAFETY:
1018 12 : // - lifetime of the fd: `file` doesn't outlive the OwnedFd stored in `self`.
1019 12 : // - &mut usage below: `self` is `&mut`, hence this call is the only task/thread that has control over the underlying fd
1020 12 : let mut file = unsafe { File::from_raw_fd(self.as_ref().as_raw_fd()) };
1021 12 : let res = with(&mut file);
1022 12 : let _ = file.into_raw_fd();
1023 12 : res
1024 12 : }
1025 : }
1026 :
1027 : impl tokio_epoll_uring::IoFd for FileGuard {
1028 2853503 : unsafe fn as_fd(&self) -> RawFd {
1029 2853503 : let owned_fd: &OwnedFd = self.as_ref();
1030 2853503 : owned_fd.as_raw_fd()
1031 2853503 : }
1032 : }
1033 :
1034 : #[cfg(test)]
1035 : impl VirtualFile {
1036 62748 : pub(crate) async fn read_blk(
1037 62748 : &self,
1038 62748 : blknum: u32,
1039 62748 : ctx: &RequestContext,
1040 62748 : ) -> Result<crate::tenant::block_io::BlockLease<'_>, std::io::Error> {
1041 : use crate::page_cache::PAGE_SZ;
1042 62748 : let slice = Vec::with_capacity(PAGE_SZ).slice_full();
1043 62748 : assert_eq!(slice.bytes_total(), PAGE_SZ);
1044 62748 : let slice = self
1045 62748 : .read_exact_at(slice, blknum as u64 * (PAGE_SZ as u64), ctx)
1046 31857 : .await?;
1047 62748 : Ok(crate::tenant::block_io::BlockLease::Vec(slice.into_inner()))
1048 62748 : }
1049 :
1050 672 : async fn read_to_end(&mut self, buf: &mut Vec<u8>, ctx: &RequestContext) -> Result<(), Error> {
1051 672 : let mut tmp = vec![0; 128];
1052 : loop {
1053 1332 : let slice = tmp.slice(..128);
1054 1332 : let (slice, res) = self.read_at(slice, self.pos, ctx).await;
1055 6 : match res {
1056 666 : Ok(0) => return Ok(()),
1057 660 : Ok(n) => {
1058 660 : self.pos += n as u64;
1059 660 : buf.extend_from_slice(&slice[..n]);
1060 660 : }
1061 6 : Err(ref e) if e.kind() == std::io::ErrorKind::Interrupted => {}
1062 6 : Err(e) => return Err(e),
1063 : }
1064 660 : tmp = slice.into_inner();
1065 : }
1066 672 : }
1067 : }
1068 :
1069 : impl Drop for VirtualFile {
1070 : /// If a VirtualFile is dropped, close the underlying file if it was open.
1071 15159 : fn drop(&mut self) {
1072 15159 : let handle = self.handle.get_mut();
1073 :
1074 15159 : fn clean_slot(slot: &Slot, mut slot_guard: RwLockWriteGuard<'_, SlotInner>, tag: u64) {
1075 15159 : if slot_guard.tag == tag {
1076 13479 : slot.recently_used.store(false, Ordering::Relaxed);
1077 : // there is also operation "close-by-replace" for closes done on eviction for
1078 : // comparison.
1079 13479 : if let Some(fd) = slot_guard.file.take() {
1080 13479 : STORAGE_IO_TIME_METRIC
1081 13479 : .get(StorageIoOperation::Close)
1082 13479 : .observe_closure_duration(|| drop(fd));
1083 13479 : }
1084 1680 : }
1085 15159 : }
1086 :
1087 : // We don't have async drop so we cannot directly await the lock here.
1088 : // Instead, first do a best-effort attempt at closing the underlying
1089 : // file descriptor by using `try_write`, and if that fails, spawn
1090 : // a tokio task to do it asynchronously: we just want it to be
1091 : // cleaned up eventually.
1092 : // Most of the time, the `try_lock` should succeed though,
1093 : // as we have `&mut self` access. In other words, if the slot
1094 : // is still occupied by our file, there should be no access from
1095 : // other I/O operations; the only other possible place to lock
1096 : // the slot is the lock algorithm looking for free slots.
1097 15159 : let slot = &get_open_files().slots[handle.index];
1098 15159 : if let Ok(slot_guard) = slot.inner.try_write() {
1099 15159 : clean_slot(slot, slot_guard, handle.tag);
1100 15159 : } else {
1101 0 : let tag = handle.tag;
1102 0 : tokio::spawn(async move {
1103 0 : let slot_guard = slot.inner.write().await;
1104 0 : clean_slot(slot, slot_guard, tag);
1105 0 : });
1106 0 : };
1107 15159 : }
1108 : }
1109 :
1110 : impl OwnedAsyncWriter for VirtualFile {
1111 : #[inline(always)]
1112 19791 : async fn write_all<Buf: IoBuf + Send>(
1113 19791 : &mut self,
1114 19791 : buf: FullSlice<Buf>,
1115 19791 : ctx: &RequestContext,
1116 19791 : ) -> std::io::Result<(usize, FullSlice<Buf>)> {
1117 19791 : let (buf, res) = VirtualFile::write_all(self, buf, ctx).await;
1118 19791 : res.map(move |v| (v, buf))
1119 19791 : }
1120 : }
1121 :
1122 : impl OpenFiles {
1123 600 : fn new(num_slots: usize) -> OpenFiles {
1124 600 : let mut slots = Box::new(Vec::with_capacity(num_slots));
1125 6000 : for _ in 0..num_slots {
1126 6000 : let slot = Slot {
1127 6000 : recently_used: AtomicBool::new(false),
1128 6000 : inner: RwLock::new(SlotInner { tag: 0, file: None }),
1129 6000 : };
1130 6000 : slots.push(slot);
1131 6000 : }
1132 :
1133 600 : OpenFiles {
1134 600 : next: AtomicUsize::new(0),
1135 600 : slots: Box::leak(slots),
1136 600 : }
1137 600 : }
1138 : }
1139 :
1140 : ///
1141 : /// Initialize the virtual file module. This must be called once at page
1142 : /// server startup.
1143 : ///
1144 : #[cfg(not(test))]
1145 0 : pub fn init(num_slots: usize, engine: IoEngineKind, io_buffer_alignment: usize) {
1146 0 : if OPEN_FILES.set(OpenFiles::new(num_slots)).is_err() {
1147 0 : panic!("virtual_file::init called twice");
1148 0 : }
1149 0 : if set_io_buffer_alignment(io_buffer_alignment).is_err() {
1150 0 : panic!("IO buffer alignment ({io_buffer_alignment}) is not a power of two");
1151 0 : }
1152 0 : io_engine::init(engine);
1153 0 : crate::metrics::virtual_file_descriptor_cache::SIZE_MAX.set(num_slots as u64);
1154 0 : }
1155 :
1156 : const TEST_MAX_FILE_DESCRIPTORS: usize = 10;
1157 :
1158 : // Get a handle to the global slots array.
1159 5739657 : fn get_open_files() -> &'static OpenFiles {
1160 5739657 : //
1161 5739657 : // In unit tests, page server startup doesn't happen and no one calls
1162 5739657 : // virtual_file::init(). Initialize it here, with a small array.
1163 5739657 : //
1164 5739657 : // This applies to the virtual file tests below, but all other unit
1165 5739657 : // tests too, so the virtual file facility is always usable in
1166 5739657 : // unit tests.
1167 5739657 : //
1168 5739657 : if cfg!(test) {
1169 5739657 : OPEN_FILES.get_or_init(|| OpenFiles::new(TEST_MAX_FILE_DESCRIPTORS))
1170 : } else {
1171 0 : OPEN_FILES.get().expect("virtual_file::init not called yet")
1172 : }
1173 5739657 : }
1174 :
1175 : static IO_BUFFER_ALIGNMENT: AtomicUsize = AtomicUsize::new(DEFAULT_IO_BUFFER_ALIGNMENT);
1176 :
1177 : /// Returns true if `x` is zero or a power of two.
1178 1263827 : fn is_zero_or_power_of_two(x: usize) -> bool {
1179 1263827 : (x == 0) || ((x & (x - 1)) == 0)
1180 1263827 : }
1181 :
1182 : #[allow(unused)]
1183 0 : pub(crate) fn set_io_buffer_alignment(align: usize) -> Result<(), usize> {
1184 0 : if is_zero_or_power_of_two(align) {
1185 0 : IO_BUFFER_ALIGNMENT.store(align, std::sync::atomic::Ordering::Relaxed);
1186 0 : Ok(())
1187 : } else {
1188 0 : Err(align)
1189 : }
1190 0 : }
1191 :
1192 : /// Gets the io buffer alignment requirement. Returns 0 if there is no requirement specified.
1193 : ///
1194 : /// This function should be used to check the raw config value.
1195 1263827 : pub(crate) fn get_io_buffer_alignment_raw() -> usize {
1196 1263827 : let align = IO_BUFFER_ALIGNMENT.load(std::sync::atomic::Ordering::Relaxed);
1197 1263827 :
1198 1263827 : if cfg!(test) {
1199 1263827 : let env_var_name = "NEON_PAGESERVER_UNIT_TEST_IO_BUFFER_ALIGNMENT";
1200 1263827 : if let Some(test_align) = utils::env::var(env_var_name) {
1201 1263827 : if is_zero_or_power_of_two(test_align) {
1202 1263827 : test_align
1203 : } else {
1204 0 : panic!("IO buffer alignment ({test_align}) is not a power of two");
1205 : }
1206 : } else {
1207 0 : align
1208 : }
1209 : } else {
1210 0 : align
1211 : }
1212 1263827 : }
1213 :
1214 : /// Gets the io buffer alignment requirement. Returns 1 if the alignment config is set to zero.
1215 : ///
1216 : /// This function should be used for getting the actual alignment value to use.
1217 623942 : pub(crate) fn get_io_buffer_alignment() -> usize {
1218 623942 : let align = get_io_buffer_alignment_raw();
1219 623942 : align.max(1)
1220 623942 : }
1221 :
1222 : #[cfg(test)]
1223 : mod tests {
1224 : use crate::context::DownloadBehavior;
1225 : use crate::task_mgr::TaskKind;
1226 :
1227 : use super::*;
1228 : use owned_buffers_io::io_buf_ext::IoBufExt;
1229 : use owned_buffers_io::slice::SliceMutExt;
1230 : use rand::seq::SliceRandom;
1231 : use rand::thread_rng;
1232 : use rand::Rng;
1233 : use std::io::Write;
1234 : use std::os::unix::fs::FileExt;
1235 : use std::sync::Arc;
1236 :
1237 : enum MaybeVirtualFile {
1238 : VirtualFile(VirtualFile),
1239 : File(File),
1240 : }
1241 :
1242 : impl From<VirtualFile> for MaybeVirtualFile {
1243 18 : fn from(vf: VirtualFile) -> Self {
1244 18 : MaybeVirtualFile::VirtualFile(vf)
1245 18 : }
1246 : }
1247 :
1248 : impl MaybeVirtualFile {
1249 1212 : async fn read_exact_at(
1250 1212 : &self,
1251 1212 : mut slice: tokio_epoll_uring::Slice<Vec<u8>>,
1252 1212 : offset: u64,
1253 1212 : ctx: &RequestContext,
1254 1212 : ) -> Result<tokio_epoll_uring::Slice<Vec<u8>>, Error> {
1255 1212 : match self {
1256 609 : MaybeVirtualFile::VirtualFile(file) => file.read_exact_at(slice, offset, ctx).await,
1257 606 : MaybeVirtualFile::File(file) => {
1258 606 : let rust_slice: &mut [u8] = slice.as_mut_rust_slice_full_zeroed();
1259 606 : file.read_exact_at(rust_slice, offset).map(|()| slice)
1260 : }
1261 : }
1262 1212 : }
1263 24 : async fn write_all_at<Buf: IoBuf + Send>(
1264 24 : &self,
1265 24 : buf: FullSlice<Buf>,
1266 24 : offset: u64,
1267 24 : ctx: &RequestContext,
1268 24 : ) -> Result<(), Error> {
1269 24 : match self {
1270 12 : MaybeVirtualFile::VirtualFile(file) => {
1271 12 : let (_buf, res) = file.write_all_at(buf, offset, ctx).await;
1272 12 : res
1273 : }
1274 12 : MaybeVirtualFile::File(file) => file.write_all_at(&buf[..], offset),
1275 : }
1276 24 : }
1277 108 : async fn seek(&mut self, pos: SeekFrom) -> Result<u64, Error> {
1278 108 : match self {
1279 54 : MaybeVirtualFile::VirtualFile(file) => file.seek(pos).await,
1280 54 : MaybeVirtualFile::File(file) => file.seek(pos),
1281 : }
1282 108 : }
1283 24 : async fn write_all<Buf: IoBuf + Send>(
1284 24 : &mut self,
1285 24 : buf: FullSlice<Buf>,
1286 24 : ctx: &RequestContext,
1287 24 : ) -> Result<(), Error> {
1288 24 : match self {
1289 12 : MaybeVirtualFile::VirtualFile(file) => {
1290 12 : let (_buf, res) = file.write_all(buf, ctx).await;
1291 12 : res.map(|_| ())
1292 : }
1293 12 : MaybeVirtualFile::File(file) => file.write_all(&buf[..]),
1294 : }
1295 24 : }
1296 :
1297 : // Helper function to slurp contents of a file, starting at the current position,
1298 : // into a string
1299 1326 : async fn read_string(&mut self, ctx: &RequestContext) -> Result<String, Error> {
1300 : use std::io::Read;
1301 1326 : let mut buf = String::new();
1302 1326 : match self {
1303 672 : MaybeVirtualFile::VirtualFile(file) => {
1304 672 : let mut buf = Vec::new();
1305 678 : file.read_to_end(&mut buf, ctx).await?;
1306 666 : return Ok(String::from_utf8(buf).unwrap());
1307 : }
1308 654 : MaybeVirtualFile::File(file) => {
1309 654 : file.read_to_string(&mut buf)?;
1310 : }
1311 : }
1312 648 : Ok(buf)
1313 1326 : }
1314 :
1315 : // Helper function to slurp a portion of a file into a string
1316 1212 : async fn read_string_at(
1317 1212 : &mut self,
1318 1212 : pos: u64,
1319 1212 : len: usize,
1320 1212 : ctx: &RequestContext,
1321 1212 : ) -> Result<String, Error> {
1322 1212 : let slice = Vec::with_capacity(len).slice_full();
1323 1212 : assert_eq!(slice.bytes_total(), len);
1324 1212 : let slice = self.read_exact_at(slice, pos, ctx).await?;
1325 1212 : let vec = slice.into_inner();
1326 1212 : assert_eq!(vec.len(), len);
1327 1212 : Ok(String::from_utf8(vec).unwrap())
1328 1212 : }
1329 : }
1330 :
1331 : #[tokio::test]
1332 6 : async fn test_virtual_files() -> anyhow::Result<()> {
1333 6 : // The real work is done in the test_files() helper function. This
1334 6 : // allows us to run the same set of tests against a native File, and
1335 6 : // VirtualFile. We trust the native Files and wouldn't need to test them,
1336 6 : // but this allows us to verify that the operations return the same
1337 6 : // results with VirtualFiles as with native Files. (Except that with
1338 6 : // native files, you will run out of file descriptors if the ulimit
1339 6 : // is low enough.)
1340 6 : struct A;
1341 6 :
1342 6 : impl Adapter for A {
1343 618 : async fn open(
1344 618 : path: Utf8PathBuf,
1345 618 : opts: OpenOptions,
1346 618 : ctx: &RequestContext,
1347 618 : ) -> Result<MaybeVirtualFile, anyhow::Error> {
1348 618 : let vf = VirtualFile::open_with_options(&path, &opts, ctx).await?;
1349 618 : Ok(MaybeVirtualFile::VirtualFile(vf))
1350 618 : }
1351 6 : }
1352 1593 : test_files::<A>("virtual_files").await
1353 6 : }
1354 :
1355 : #[tokio::test]
1356 6 : async fn test_physical_files() -> anyhow::Result<()> {
1357 6 : struct B;
1358 6 :
1359 6 : impl Adapter for B {
1360 618 : async fn open(
1361 618 : path: Utf8PathBuf,
1362 618 : opts: OpenOptions,
1363 618 : _ctx: &RequestContext,
1364 618 : ) -> Result<MaybeVirtualFile, anyhow::Error> {
1365 6 : Ok(MaybeVirtualFile::File({
1366 618 : let owned_fd = opts.open(path.as_std_path()).await?;
1367 618 : File::from(owned_fd)
1368 6 : }))
1369 618 : }
1370 6 : }
1371 6 :
1372 312 : test_files::<B>("physical_files").await
1373 6 : }
1374 :
1375 : /// This is essentially a closure which returns a MaybeVirtualFile, but because rust edition
1376 : /// 2024 is not yet out with new lifetime capture or outlives rules, this is a async function
1377 : /// in trait which benefits from the new lifetime capture rules already.
1378 : trait Adapter {
1379 : async fn open(
1380 : path: Utf8PathBuf,
1381 : opts: OpenOptions,
1382 : ctx: &RequestContext,
1383 : ) -> Result<MaybeVirtualFile, anyhow::Error>;
1384 : }
1385 :
1386 12 : async fn test_files<A>(testname: &str) -> anyhow::Result<()>
1387 12 : where
1388 12 : A: Adapter,
1389 12 : {
1390 12 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
1391 12 : let testdir = crate::config::PageServerConf::test_repo_dir(testname);
1392 12 : std::fs::create_dir_all(&testdir)?;
1393 :
1394 12 : let path_a = testdir.join("file_a");
1395 12 : let mut file_a = A::open(
1396 12 : path_a.clone(),
1397 12 : OpenOptions::new()
1398 12 : .write(true)
1399 12 : .create(true)
1400 12 : .truncate(true)
1401 12 : .to_owned(),
1402 12 : &ctx,
1403 12 : )
1404 12 : .await?;
1405 12 : file_a
1406 12 : .write_all(b"foobar".to_vec().slice_len(), &ctx)
1407 3 : .await?;
1408 :
1409 : // cannot read from a file opened in write-only mode
1410 12 : let _ = file_a.read_string(&ctx).await.unwrap_err();
1411 :
1412 : // Close the file and re-open for reading
1413 12 : let mut file_a = A::open(path_a, OpenOptions::new().read(true).to_owned(), &ctx).await?;
1414 :
1415 : // cannot write to a file opened in read-only mode
1416 12 : let _ = file_a
1417 12 : .write_all(b"bar".to_vec().slice_len(), &ctx)
1418 3 : .await
1419 12 : .unwrap_err();
1420 12 :
1421 12 : // Try simple read
1422 12 : assert_eq!("foobar", file_a.read_string(&ctx).await?);
1423 :
1424 : // It's positioned at the EOF now.
1425 12 : assert_eq!("", file_a.read_string(&ctx).await?);
1426 :
1427 : // Test seeks.
1428 12 : assert_eq!(file_a.seek(SeekFrom::Start(1)).await?, 1);
1429 12 : assert_eq!("oobar", file_a.read_string(&ctx).await?);
1430 :
1431 12 : assert_eq!(file_a.seek(SeekFrom::End(-2)).await?, 4);
1432 12 : assert_eq!("ar", file_a.read_string(&ctx).await?);
1433 :
1434 12 : assert_eq!(file_a.seek(SeekFrom::Start(1)).await?, 1);
1435 12 : assert_eq!(file_a.seek(SeekFrom::Current(2)).await?, 3);
1436 12 : assert_eq!("bar", file_a.read_string(&ctx).await?);
1437 :
1438 12 : assert_eq!(file_a.seek(SeekFrom::Current(-5)).await?, 1);
1439 12 : assert_eq!("oobar", file_a.read_string(&ctx).await?);
1440 :
1441 : // Test erroneous seeks to before byte 0
1442 12 : file_a.seek(SeekFrom::End(-7)).await.unwrap_err();
1443 12 : assert_eq!(file_a.seek(SeekFrom::Start(1)).await?, 1);
1444 12 : file_a.seek(SeekFrom::Current(-2)).await.unwrap_err();
1445 12 :
1446 12 : // the erroneous seek should have left the position unchanged
1447 12 : assert_eq!("oobar", file_a.read_string(&ctx).await?);
1448 :
1449 : // Create another test file, and try FileExt functions on it.
1450 12 : let path_b = testdir.join("file_b");
1451 12 : let mut file_b = A::open(
1452 12 : path_b.clone(),
1453 12 : OpenOptions::new()
1454 12 : .read(true)
1455 12 : .write(true)
1456 12 : .create(true)
1457 12 : .truncate(true)
1458 12 : .to_owned(),
1459 12 : &ctx,
1460 12 : )
1461 6 : .await?;
1462 12 : file_b
1463 12 : .write_all_at(b"BAR".to_vec().slice_len(), 3, &ctx)
1464 3 : .await?;
1465 12 : file_b
1466 12 : .write_all_at(b"FOO".to_vec().slice_len(), 0, &ctx)
1467 3 : .await?;
1468 :
1469 12 : assert_eq!(file_b.read_string_at(2, 3, &ctx).await?, "OBA");
1470 :
1471 : // Open a lot of files, enough to cause some evictions. (Or to be precise,
1472 : // open the same file many times. The effect is the same.)
1473 : //
1474 : // leave file_a positioned at offset 1 before we start
1475 12 : assert_eq!(file_a.seek(SeekFrom::Start(1)).await?, 1);
1476 :
1477 12 : let mut vfiles = Vec::new();
1478 1212 : for _ in 0..100 {
1479 1200 : let mut vfile = A::open(
1480 1200 : path_b.clone(),
1481 1200 : OpenOptions::new().read(true).to_owned(),
1482 1200 : &ctx,
1483 1200 : )
1484 600 : .await?;
1485 1200 : assert_eq!("FOOBAR", vfile.read_string(&ctx).await?);
1486 1200 : vfiles.push(vfile);
1487 : }
1488 :
1489 : // make sure we opened enough files to definitely cause evictions.
1490 12 : assert!(vfiles.len() > TEST_MAX_FILE_DESCRIPTORS * 2);
1491 :
1492 : // The underlying file descriptor for 'file_a' should be closed now. Try to read
1493 : // from it again. We left the file positioned at offset 1 above.
1494 12 : assert_eq!("oobar", file_a.read_string(&ctx).await?);
1495 :
1496 : // Check that all the other FDs still work too. Use them in random order for
1497 : // good measure.
1498 12 : vfiles.as_mut_slice().shuffle(&mut thread_rng());
1499 1200 : for vfile in vfiles.iter_mut() {
1500 1200 : assert_eq!("OOBAR", vfile.read_string_at(1, 5, &ctx).await?);
1501 : }
1502 :
1503 12 : Ok(())
1504 12 : }
1505 :
1506 : /// Test using VirtualFiles from many threads concurrently. This tests both using
1507 : /// a lot of VirtualFiles concurrently, causing evictions, and also using the same
1508 : /// VirtualFile from multiple threads concurrently.
1509 : #[tokio::test]
1510 6 : async fn test_vfile_concurrency() -> Result<(), Error> {
1511 6 : const SIZE: usize = 8 * 1024;
1512 6 : const VIRTUAL_FILES: usize = 100;
1513 6 : const THREADS: usize = 100;
1514 6 : const SAMPLE: [u8; SIZE] = [0xADu8; SIZE];
1515 6 :
1516 6 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
1517 6 : let testdir = crate::config::PageServerConf::test_repo_dir("vfile_concurrency");
1518 6 : std::fs::create_dir_all(&testdir)?;
1519 6 :
1520 6 : // Create a test file.
1521 6 : let test_file_path = testdir.join("concurrency_test_file");
1522 6 : {
1523 6 : let file = File::create(&test_file_path)?;
1524 6 : file.write_all_at(&SAMPLE, 0)?;
1525 6 : }
1526 6 :
1527 6 : // Open the file many times.
1528 6 : let mut files = Vec::new();
1529 606 : for _ in 0..VIRTUAL_FILES {
1530 600 : let f = VirtualFile::open_with_options(
1531 600 : &test_file_path,
1532 600 : OpenOptions::new().read(true),
1533 600 : &ctx,
1534 600 : )
1535 303 : .await?;
1536 600 : files.push(f);
1537 6 : }
1538 6 : let files = Arc::new(files);
1539 6 :
1540 6 : // Launch many threads, and use the virtual files concurrently in random order.
1541 6 : let rt = tokio::runtime::Builder::new_multi_thread()
1542 6 : .worker_threads(THREADS)
1543 6 : .thread_name("test_vfile_concurrency thread")
1544 6 : .build()
1545 6 : .unwrap();
1546 6 : let mut hdls = Vec::new();
1547 606 : for _threadno in 0..THREADS {
1548 600 : let files = files.clone();
1549 600 : let ctx = ctx.detached_child(TaskKind::UnitTest, DownloadBehavior::Error);
1550 600 : let hdl = rt.spawn(async move {
1551 600 : let mut buf = vec![0u8; SIZE];
1552 600 : let mut rng = rand::rngs::OsRng;
1553 600000 : for _ in 1..1000 {
1554 599400 : let f = &files[rng.gen_range(0..files.len())];
1555 599400 : buf = f
1556 599400 : .read_exact_at(buf.slice_full(), 0, &ctx)
1557 1785251 : .await
1558 599400 : .unwrap()
1559 599400 : .into_inner();
1560 599400 : assert!(buf == SAMPLE);
1561 6 : }
1562 600 : });
1563 600 : hdls.push(hdl);
1564 600 : }
1565 606 : for hdl in hdls {
1566 600 : hdl.await?;
1567 6 : }
1568 6 : std::mem::forget(rt);
1569 6 :
1570 6 : Ok(())
1571 6 : }
1572 :
1573 : #[tokio::test]
1574 6 : async fn test_atomic_overwrite_basic() {
1575 6 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
1576 6 : let testdir = crate::config::PageServerConf::test_repo_dir("test_atomic_overwrite_basic");
1577 6 : std::fs::create_dir_all(&testdir).unwrap();
1578 6 :
1579 6 : let path = testdir.join("myfile");
1580 6 : let tmp_path = testdir.join("myfile.tmp");
1581 6 :
1582 6 : VirtualFile::crashsafe_overwrite(path.clone(), tmp_path.clone(), b"foo".to_vec())
1583 6 : .await
1584 6 : .unwrap();
1585 6 : let mut file = MaybeVirtualFile::from(VirtualFile::open(&path, &ctx).await.unwrap());
1586 6 : let post = file.read_string(&ctx).await.unwrap();
1587 6 : assert_eq!(post, "foo");
1588 6 : assert!(!tmp_path.exists());
1589 6 : drop(file);
1590 6 :
1591 6 : VirtualFile::crashsafe_overwrite(path.clone(), tmp_path.clone(), b"bar".to_vec())
1592 6 : .await
1593 6 : .unwrap();
1594 6 : let mut file = MaybeVirtualFile::from(VirtualFile::open(&path, &ctx).await.unwrap());
1595 6 : let post = file.read_string(&ctx).await.unwrap();
1596 6 : assert_eq!(post, "bar");
1597 6 : assert!(!tmp_path.exists());
1598 6 : drop(file);
1599 6 : }
1600 :
1601 : #[tokio::test]
1602 6 : async fn test_atomic_overwrite_preexisting_tmp() {
1603 6 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
1604 6 : let testdir =
1605 6 : crate::config::PageServerConf::test_repo_dir("test_atomic_overwrite_preexisting_tmp");
1606 6 : std::fs::create_dir_all(&testdir).unwrap();
1607 6 :
1608 6 : let path = testdir.join("myfile");
1609 6 : let tmp_path = testdir.join("myfile.tmp");
1610 6 :
1611 6 : std::fs::write(&tmp_path, "some preexisting junk that should be removed").unwrap();
1612 6 : assert!(tmp_path.exists());
1613 6 :
1614 6 : VirtualFile::crashsafe_overwrite(path.clone(), tmp_path.clone(), b"foo".to_vec())
1615 6 : .await
1616 6 : .unwrap();
1617 6 :
1618 6 : let mut file = MaybeVirtualFile::from(VirtualFile::open(&path, &ctx).await.unwrap());
1619 6 : let post = file.read_string(&ctx).await.unwrap();
1620 6 : assert_eq!(post, "foo");
1621 6 : assert!(!tmp_path.exists());
1622 6 : drop(file);
1623 6 : }
1624 : }
|
