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::aligned_buffer::buffer::AlignedBuffer;
22 : use owned_buffers_io::aligned_buffer::{AlignedBufferMut, AlignedSlice, ConstAlign};
23 : use owned_buffers_io::io_buf_aligned::{IoBufAligned, IoBufAlignedMut};
24 : use owned_buffers_io::io_buf_ext::FullSlice;
25 : use pageserver_api::config::defaults::DEFAULT_IO_BUFFER_ALIGNMENT;
26 : use pageserver_api::shard::TenantShardId;
27 : use std::fs::File;
28 : use std::io::{Error, ErrorKind, Seek, SeekFrom};
29 : #[cfg(target_os = "linux")]
30 : use std::os::unix::fs::OpenOptionsExt;
31 : use tokio_epoll_uring::{BoundedBuf, IoBuf, IoBufMut, Slice};
32 :
33 : use std::os::fd::{AsRawFd, FromRawFd, IntoRawFd, OwnedFd, RawFd};
34 : use std::sync::atomic::{AtomicBool, AtomicU8, AtomicUsize, Ordering};
35 : use tokio::sync::{RwLock, RwLockReadGuard, RwLockWriteGuard};
36 : use tokio::time::Instant;
37 :
38 : pub use pageserver_api::models::virtual_file as api;
39 : pub(crate) mod io_engine;
40 : pub use io_engine::feature_test as io_engine_feature_test;
41 : pub use io_engine::io_engine_for_bench;
42 : pub use io_engine::FeatureTestResult as IoEngineFeatureTestResult;
43 : mod metadata;
44 : mod open_options;
45 : use self::owned_buffers_io::write::OwnedAsyncWriter;
46 : pub(crate) use api::IoMode;
47 : pub(crate) use io_engine::IoEngineKind;
48 : pub(crate) use metadata::Metadata;
49 : pub(crate) use open_options::*;
50 :
51 : pub(crate) mod owned_buffers_io {
52 : //! Abstractions for IO with owned buffers.
53 : //!
54 : //! Not actually tied to [`crate::virtual_file`] specifically, but, it's the primary
55 : //! reason we need this abstraction.
56 : //!
57 : //! Over time, this could move into the `tokio-epoll-uring` crate, maybe `uring-common`,
58 : //! but for the time being we're proving out the primitives in the neon.git repo
59 : //! for faster iteration.
60 :
61 : pub(crate) mod aligned_buffer;
62 : pub(crate) mod io_buf_aligned;
63 : pub(crate) mod io_buf_ext;
64 : pub(crate) mod slice;
65 : pub(crate) mod write;
66 : }
67 :
68 : #[derive(Debug)]
69 : pub struct VirtualFile {
70 : inner: VirtualFileInner,
71 : _mode: IoMode,
72 : }
73 :
74 : impl VirtualFile {
75 : /// Open a file in read-only mode. Like File::open.
76 2112 : pub async fn open<P: AsRef<Utf8Path>>(
77 2112 : path: P,
78 2112 : ctx: &RequestContext,
79 2112 : ) -> Result<Self, std::io::Error> {
80 2112 : let inner = VirtualFileInner::open(path, ctx).await?;
81 2112 : Ok(VirtualFile {
82 2112 : inner,
83 2112 : _mode: IoMode::Buffered,
84 2112 : })
85 2112 : }
86 :
87 : /// Open a file in read-only mode. Like File::open.
88 : ///
89 : /// `O_DIRECT` will be enabled base on `virtual_file_io_mode`.
90 2456 : pub async fn open_v2<P: AsRef<Utf8Path>>(
91 2456 : path: P,
92 2456 : ctx: &RequestContext,
93 2456 : ) -> Result<Self, std::io::Error> {
94 2456 : Self::open_with_options_v2(path.as_ref(), OpenOptions::new().read(true), ctx).await
95 2456 : }
96 :
97 3030 : pub async fn create<P: AsRef<Utf8Path>>(
98 3030 : path: P,
99 3030 : ctx: &RequestContext,
100 3030 : ) -> Result<Self, std::io::Error> {
101 3030 : let inner = VirtualFileInner::create(path, ctx).await?;
102 3030 : Ok(VirtualFile {
103 3030 : inner,
104 3030 : _mode: IoMode::Buffered,
105 3030 : })
106 3030 : }
107 :
108 0 : pub async fn create_v2<P: AsRef<Utf8Path>>(
109 0 : path: P,
110 0 : ctx: &RequestContext,
111 0 : ) -> Result<Self, std::io::Error> {
112 0 : VirtualFile::open_with_options_v2(
113 0 : path.as_ref(),
114 0 : OpenOptions::new().write(true).create(true).truncate(true),
115 0 : ctx,
116 0 : )
117 0 : .await
118 0 : }
119 :
120 1592 : pub async fn open_with_options<P: AsRef<Utf8Path>>(
121 1592 : path: P,
122 1592 : open_options: &OpenOptions,
123 1592 : ctx: &RequestContext, /* TODO: carry a pointer to the metrics in the RequestContext instead of the parsing https://github.com/neondatabase/neon/issues/6107 */
124 1592 : ) -> Result<Self, std::io::Error> {
125 1592 : let inner = VirtualFileInner::open_with_options(path, open_options, ctx).await?;
126 1592 : Ok(VirtualFile {
127 1592 : inner,
128 1592 : _mode: IoMode::Buffered,
129 1592 : })
130 1592 : }
131 :
132 5068 : pub async fn open_with_options_v2<P: AsRef<Utf8Path>>(
133 5068 : path: P,
134 5068 : open_options: &OpenOptions,
135 5068 : ctx: &RequestContext, /* TODO: carry a pointer to the metrics in the RequestContext instead of the parsing https://github.com/neondatabase/neon/issues/6107 */
136 5068 : ) -> Result<Self, std::io::Error> {
137 5068 : let file = match get_io_mode() {
138 : IoMode::Buffered => {
139 5068 : let inner = VirtualFileInner::open_with_options(path, open_options, ctx).await?;
140 5068 : VirtualFile {
141 5068 : inner,
142 5068 : _mode: IoMode::Buffered,
143 5068 : }
144 : }
145 : #[cfg(target_os = "linux")]
146 : IoMode::Direct => {
147 0 : let inner = VirtualFileInner::open_with_options(
148 0 : path,
149 0 : open_options.clone().custom_flags(nix::libc::O_DIRECT),
150 0 : ctx,
151 0 : )
152 0 : .await?;
153 0 : VirtualFile {
154 0 : inner,
155 0 : _mode: IoMode::Direct,
156 0 : }
157 : }
158 : };
159 5068 : Ok(file)
160 5068 : }
161 :
162 2376 : pub fn path(&self) -> &Utf8Path {
163 2376 : self.inner.path.as_path()
164 2376 : }
165 :
166 44 : pub async fn crashsafe_overwrite<B: BoundedBuf<Buf = Buf> + Send, Buf: IoBuf + Send>(
167 44 : final_path: Utf8PathBuf,
168 44 : tmp_path: Utf8PathBuf,
169 44 : content: B,
170 44 : ) -> std::io::Result<()> {
171 44 : VirtualFileInner::crashsafe_overwrite(final_path, tmp_path, content).await
172 44 : }
173 :
174 5634 : pub async fn sync_all(&self) -> Result<(), Error> {
175 5634 : if SYNC_MODE.load(std::sync::atomic::Ordering::Relaxed) == SyncMode::UnsafeNoSync as u8 {
176 0 : return Ok(());
177 5634 : }
178 5634 : self.inner.sync_all().await
179 5634 : }
180 :
181 0 : pub async fn sync_data(&self) -> Result<(), Error> {
182 0 : if SYNC_MODE.load(std::sync::atomic::Ordering::Relaxed) == SyncMode::UnsafeNoSync as u8 {
183 0 : return Ok(());
184 0 : }
185 0 : self.inner.sync_data().await
186 0 : }
187 :
188 3604 : pub async fn metadata(&self) -> Result<Metadata, Error> {
189 3604 : self.inner.metadata().await
190 3604 : }
191 :
192 516 : pub fn remove(self) {
193 516 : self.inner.remove();
194 516 : }
195 :
196 11352 : pub async fn seek(&mut self, pos: SeekFrom) -> Result<u64, Error> {
197 11352 : self.inner.seek(pos).await
198 11352 : }
199 :
200 461970 : pub async fn read_exact_at<Buf>(
201 461970 : &self,
202 461970 : slice: Slice<Buf>,
203 461970 : offset: u64,
204 461970 : ctx: &RequestContext,
205 461970 : ) -> Result<Slice<Buf>, Error>
206 461970 : where
207 461970 : Buf: IoBufAlignedMut + Send,
208 461970 : {
209 461970 : self.inner.read_exact_at(slice, offset, ctx).await
210 461970 : }
211 :
212 63357 : pub async fn read_exact_at_page(
213 63357 : &self,
214 63357 : page: PageWriteGuard<'static>,
215 63357 : offset: u64,
216 63357 : ctx: &RequestContext,
217 63357 : ) -> Result<PageWriteGuard<'static>, Error> {
218 63357 : self.inner.read_exact_at_page(page, offset, ctx).await
219 63357 : }
220 :
221 13222 : pub async fn write_all_at<Buf: IoBufAligned + Send>(
222 13222 : &self,
223 13222 : buf: FullSlice<Buf>,
224 13222 : offset: u64,
225 13222 : ctx: &RequestContext,
226 13222 : ) -> (FullSlice<Buf>, Result<(), Error>) {
227 13222 : self.inner.write_all_at(buf, offset, ctx).await
228 13222 : }
229 :
230 2260790 : pub async fn write_all<Buf: IoBuf + Send>(
231 2260790 : &mut self,
232 2260790 : buf: FullSlice<Buf>,
233 2260790 : ctx: &RequestContext,
234 2260790 : ) -> (FullSlice<Buf>, Result<usize, Error>) {
235 2260790 : self.inner.write_all(buf, ctx).await
236 2260790 : }
237 :
238 448 : async fn read_to_end(&mut self, buf: &mut Vec<u8>, ctx: &RequestContext) -> Result<(), Error> {
239 448 : self.inner.read_to_end(buf, ctx).await
240 448 : }
241 :
242 0 : pub(crate) async fn read_to_string(
243 0 : &mut self,
244 0 : ctx: &RequestContext,
245 0 : ) -> Result<String, anyhow::Error> {
246 0 : let mut buf = Vec::new();
247 0 : self.read_to_end(&mut buf, ctx).await?;
248 0 : Ok(String::from_utf8(buf)?)
249 0 : }
250 : }
251 :
252 : /// Indicates whether to enable fsync, fdatasync, or O_SYNC/O_DSYNC when writing
253 : /// files. Switching this off is unsafe and only used for testing on machines
254 : /// with slow drives.
255 : #[repr(u8)]
256 : pub enum SyncMode {
257 : Sync,
258 : UnsafeNoSync,
259 : }
260 :
261 : impl TryFrom<u8> for SyncMode {
262 : type Error = u8;
263 :
264 0 : fn try_from(value: u8) -> Result<Self, Self::Error> {
265 0 : Ok(match value {
266 0 : v if v == (SyncMode::Sync as u8) => SyncMode::Sync,
267 0 : v if v == (SyncMode::UnsafeNoSync as u8) => SyncMode::UnsafeNoSync,
268 0 : x => return Err(x),
269 : })
270 0 : }
271 : }
272 :
273 : ///
274 : /// A virtual file descriptor. You can use this just like std::fs::File, but internally
275 : /// the underlying file is closed if the system is low on file descriptors,
276 : /// and re-opened when it's accessed again.
277 : ///
278 : /// Like with std::fs::File, multiple threads can read/write the file concurrently,
279 : /// holding just a shared reference the same VirtualFile, using the read_at() / write_at()
280 : /// functions from the FileExt trait. But the functions from the Read/Write/Seek traits
281 : /// require a mutable reference, because they modify the "current position".
282 : ///
283 : /// Each VirtualFile has a physical file descriptor in the global OPEN_FILES array, at the
284 : /// slot that 'handle points to, if the underlying file is currently open. If it's not
285 : /// currently open, the 'handle' can still point to the slot where it was last kept. The
286 : /// 'tag' field is used to detect whether the handle still is valid or not.
287 : ///
288 : #[derive(Debug)]
289 : pub struct VirtualFileInner {
290 : /// Lazy handle to the global file descriptor cache. The slot that this points to
291 : /// might contain our File, or it may be empty, or it may contain a File that
292 : /// belongs to a different VirtualFile.
293 : handle: RwLock<SlotHandle>,
294 :
295 : /// Current file position
296 : pos: u64,
297 :
298 : /// File path and options to use to open it.
299 : ///
300 : /// Note: this only contains the options needed to re-open it. For example,
301 : /// if a new file is created, we only pass the create flag when it's initially
302 : /// opened, in the VirtualFile::create() function, and strip the flag before
303 : /// storing it here.
304 : pub path: Utf8PathBuf,
305 : open_options: OpenOptions,
306 :
307 : // These are strings becase we only use them for metrics, and those expect strings.
308 : // It makes no sense for us to constantly turn the `TimelineId` and `TenantId` into
309 : // strings.
310 : tenant_id: String,
311 : shard_id: String,
312 : timeline_id: String,
313 : }
314 :
315 : #[derive(Debug, PartialEq, Clone, Copy)]
316 : struct SlotHandle {
317 : /// Index into OPEN_FILES.slots
318 : index: usize,
319 :
320 : /// Value of 'tag' in the slot. If slot's tag doesn't match, then the slot has
321 : /// been recycled and no longer contains the FD for this virtual file.
322 : tag: u64,
323 : }
324 :
325 : /// OPEN_FILES is the global array that holds the physical file descriptors that
326 : /// are currently open. Each slot in the array is protected by a separate lock,
327 : /// so that different files can be accessed independently. The lock must be held
328 : /// in write mode to replace the slot with a different file, but a read mode
329 : /// is enough to operate on the file, whether you're reading or writing to it.
330 : ///
331 : /// OPEN_FILES starts in uninitialized state, and it's initialized by
332 : /// the virtual_file::init() function. It must be called exactly once at page
333 : /// server startup.
334 : static OPEN_FILES: OnceCell<OpenFiles> = OnceCell::new();
335 :
336 : struct OpenFiles {
337 : slots: &'static [Slot],
338 :
339 : /// clock arm for the clock algorithm
340 : next: AtomicUsize,
341 : }
342 :
343 : struct Slot {
344 : inner: RwLock<SlotInner>,
345 :
346 : /// has this file been used since last clock sweep?
347 : recently_used: AtomicBool,
348 : }
349 :
350 : struct SlotInner {
351 : /// Counter that's incremented every time a different file is stored here.
352 : /// To avoid the ABA problem.
353 : tag: u64,
354 :
355 : /// the underlying file
356 : file: Option<OwnedFd>,
357 : }
358 :
359 : /// Impl of [`tokio_epoll_uring::IoBuf`] and [`tokio_epoll_uring::IoBufMut`] for [`PageWriteGuard`].
360 : struct PageWriteGuardBuf {
361 : page: PageWriteGuard<'static>,
362 : }
363 : // Safety: the [`PageWriteGuard`] gives us exclusive ownership of the page cache slot,
364 : // and the location remains stable even if [`Self`] or the [`PageWriteGuard`] is moved.
365 : // Page cache pages are zero-initialized, so, wrt uninitialized memory we're good.
366 : // (Page cache tracks separately whether the contents are valid, see `PageWriteGuard::mark_valid`.)
367 : unsafe impl tokio_epoll_uring::IoBuf for PageWriteGuardBuf {
368 253377 : fn stable_ptr(&self) -> *const u8 {
369 253377 : self.page.as_ptr()
370 253377 : }
371 475101 : fn bytes_init(&self) -> usize {
372 475101 : self.page.len()
373 475101 : }
374 190071 : fn bytes_total(&self) -> usize {
375 190071 : self.page.len()
376 190071 : }
377 : }
378 : // Safety: see above, plus: the ownership of [`PageWriteGuard`] means exclusive access,
379 : // hence it's safe to hand out the `stable_mut_ptr()`.
380 : unsafe impl tokio_epoll_uring::IoBufMut for PageWriteGuardBuf {
381 95010 : fn stable_mut_ptr(&mut self) -> *mut u8 {
382 95010 : self.page.as_mut_ptr()
383 95010 : }
384 :
385 63357 : unsafe fn set_init(&mut self, pos: usize) {
386 63357 : // There shouldn't really be any reason to call this API since bytes_init() == bytes_total().
387 63357 : assert!(pos <= self.page.len());
388 63357 : }
389 : }
390 :
391 : impl OpenFiles {
392 : /// Find a slot to use, evicting an existing file descriptor if needed.
393 : ///
394 : /// On return, we hold a lock on the slot, and its 'tag' has been updated
395 : /// recently_used has been set. It's all ready for reuse.
396 390284 : async fn find_victim_slot(&self) -> (SlotHandle, RwLockWriteGuard<SlotInner>) {
397 390284 : //
398 390284 : // Run the clock algorithm to find a slot to replace.
399 390284 : //
400 390284 : let num_slots = self.slots.len();
401 390284 : let mut retries = 0;
402 : let mut slot;
403 : let mut slot_guard;
404 : let index;
405 : loop {
406 5058137 : let next = self.next.fetch_add(1, Ordering::AcqRel) % num_slots;
407 5058137 : slot = &self.slots[next];
408 5058137 :
409 5058137 : // If the recently_used flag on this slot is set, continue the clock
410 5058137 : // sweep. Otherwise try to use this slot. If we cannot acquire the
411 5058137 : // lock, also continue the clock sweep.
412 5058137 : //
413 5058137 : // We only continue in this manner for a while, though. If we loop
414 5058137 : // through the array twice without finding a victim, just pick the
415 5058137 : // next slot and wait until we can reuse it. This way, we avoid
416 5058137 : // spinning in the extreme case that all the slots are busy with an
417 5058137 : // I/O operation.
418 5058137 : if retries < num_slots * 2 {
419 4860430 : if !slot.recently_used.swap(false, Ordering::Release) {
420 4444087 : if let Ok(guard) = slot.inner.try_write() {
421 192577 : slot_guard = guard;
422 192577 : index = next;
423 192577 : break;
424 4251510 : }
425 416343 : }
426 4667853 : retries += 1;
427 : } else {
428 197707 : slot_guard = slot.inner.write().await;
429 197707 : index = next;
430 197707 : break;
431 : }
432 : }
433 :
434 : //
435 : // We now have the victim slot locked. If it was in use previously, close the
436 : // old file.
437 : //
438 390284 : if let Some(old_file) = slot_guard.file.take() {
439 380306 : // the normal path of dropping VirtualFile uses "close", use "close-by-replace" here to
440 380306 : // distinguish the two.
441 380306 : STORAGE_IO_TIME_METRIC
442 380306 : .get(StorageIoOperation::CloseByReplace)
443 380306 : .observe_closure_duration(|| drop(old_file));
444 380306 : }
445 :
446 : // Prepare the slot for reuse and return it
447 390284 : slot_guard.tag += 1;
448 390284 : slot.recently_used.store(true, Ordering::Relaxed);
449 390284 : (
450 390284 : SlotHandle {
451 390284 : index,
452 390284 : tag: slot_guard.tag,
453 390284 : },
454 390284 : slot_guard,
455 390284 : )
456 390284 : }
457 : }
458 :
459 : /// Identify error types that should alwways terminate the process. Other
460 : /// error types may be elegible for retry.
461 8 : pub(crate) fn is_fatal_io_error(e: &std::io::Error) -> bool {
462 : use nix::errno::Errno::*;
463 8 : match e.raw_os_error().map(nix::errno::from_i32) {
464 : Some(EIO) => {
465 : // Terminate on EIO because we no longer trust the device to store
466 : // data safely, or to uphold persistence guarantees on fsync.
467 0 : true
468 : }
469 : Some(EROFS) => {
470 : // Terminate on EROFS because a filesystem is usually remounted
471 : // readonly when it has experienced some critical issue, so the same
472 : // logic as EIO applies.
473 0 : true
474 : }
475 : Some(EACCES) => {
476 : // Terminate on EACCESS because we should always have permissions
477 : // for our own data dir: if we don't, then we can't do our job and
478 : // need administrative intervention to fix permissions. Terminating
479 : // is the best way to make sure we stop cleanly rather than going
480 : // into infinite retry loops, and will make it clear to the outside
481 : // world that we need help.
482 0 : true
483 : }
484 : _ => {
485 : // Treat all other local file I/O errors are retryable. This includes:
486 : // - ENOSPC: we stay up and wait for eviction to free some space
487 : // - EINVAL, EBADF, EBADFD: this is a code bug, not a filesystem/hardware issue
488 : // - WriteZero, Interrupted: these are used internally VirtualFile
489 8 : false
490 : }
491 : }
492 8 : }
493 :
494 : /// Call this when the local filesystem gives us an error with an external
495 : /// cause: this includes EIO, EROFS, and EACCESS: all these indicate either
496 : /// bad storage or bad configuration, and we can't fix that from inside
497 : /// a running process.
498 0 : pub(crate) fn on_fatal_io_error(e: &std::io::Error, context: &str) -> ! {
499 0 : let backtrace = std::backtrace::Backtrace::force_capture();
500 0 : tracing::error!("Fatal I/O error: {e}: {context})\n{backtrace}");
501 0 : std::process::abort();
502 : }
503 :
504 : pub(crate) trait MaybeFatalIo<T> {
505 : fn maybe_fatal_err(self, context: &str) -> std::io::Result<T>;
506 : fn fatal_err(self, context: &str) -> T;
507 : }
508 :
509 : impl<T> MaybeFatalIo<T> for std::io::Result<T> {
510 : /// Terminate the process if the result is an error of a fatal type, else pass it through
511 : ///
512 : /// This is appropriate for writes, where we typically want to die on EIO/ACCES etc, but
513 : /// not on ENOSPC.
514 4218606 : fn maybe_fatal_err(self, context: &str) -> std::io::Result<T> {
515 4218606 : if let Err(e) = &self {
516 8 : if is_fatal_io_error(e) {
517 0 : on_fatal_io_error(e, context);
518 8 : }
519 4218598 : }
520 4218606 : self
521 4218606 : }
522 :
523 : /// Terminate the process on any I/O error.
524 : ///
525 : /// This is appropriate for reads on files that we know exist: they should always work.
526 4116 : fn fatal_err(self, context: &str) -> T {
527 4116 : match self {
528 4116 : Ok(v) => v,
529 0 : Err(e) => {
530 0 : on_fatal_io_error(&e, context);
531 : }
532 : }
533 4116 : }
534 : }
535 :
536 : /// Observe duration for the given storage I/O operation
537 : ///
538 : /// Unlike `observe_closure_duration`, this supports async,
539 : /// where "support" means that we measure wall clock time.
540 : macro_rules! observe_duration {
541 : ($op:expr, $($body:tt)*) => {{
542 : let instant = Instant::now();
543 : let result = $($body)*;
544 : let elapsed = instant.elapsed().as_secs_f64();
545 : STORAGE_IO_TIME_METRIC
546 : .get($op)
547 : .observe(elapsed);
548 : result
549 : }}
550 : }
551 :
552 : macro_rules! with_file {
553 : ($this:expr, $op:expr, | $ident:ident | $($body:tt)*) => {{
554 : let $ident = $this.lock_file().await?;
555 : observe_duration!($op, $($body)*)
556 : }};
557 : ($this:expr, $op:expr, | mut $ident:ident | $($body:tt)*) => {{
558 : let mut $ident = $this.lock_file().await?;
559 : observe_duration!($op, $($body)*)
560 : }};
561 : }
562 :
563 : impl VirtualFileInner {
564 : /// Open a file in read-only mode. Like File::open.
565 2112 : pub async fn open<P: AsRef<Utf8Path>>(
566 2112 : path: P,
567 2112 : ctx: &RequestContext,
568 2112 : ) -> Result<VirtualFileInner, std::io::Error> {
569 2112 : Self::open_with_options(path.as_ref(), OpenOptions::new().read(true), ctx).await
570 2112 : }
571 :
572 : /// Create a new file for writing. If the file exists, it will be truncated.
573 : /// Like File::create.
574 3030 : pub async fn create<P: AsRef<Utf8Path>>(
575 3030 : path: P,
576 3030 : ctx: &RequestContext,
577 3030 : ) -> Result<VirtualFileInner, std::io::Error> {
578 3030 : Self::open_with_options(
579 3030 : path.as_ref(),
580 3030 : OpenOptions::new().write(true).create(true).truncate(true),
581 3030 : ctx,
582 3030 : )
583 3030 : .await
584 3030 : }
585 :
586 : /// Open a file with given options.
587 : ///
588 : /// Note: If any custom flags were set in 'open_options' through OpenOptionsExt,
589 : /// they will be applied also when the file is subsequently re-opened, not only
590 : /// on the first time. Make sure that's sane!
591 12202 : pub async fn open_with_options<P: AsRef<Utf8Path>>(
592 12202 : path: P,
593 12202 : open_options: &OpenOptions,
594 12202 : _ctx: &RequestContext, /* TODO: carry a pointer to the metrics in the RequestContext instead of the parsing https://github.com/neondatabase/neon/issues/6107 */
595 12202 : ) -> Result<VirtualFileInner, std::io::Error> {
596 12202 : let path_ref = path.as_ref();
597 12202 : let path_str = path_ref.to_string();
598 12202 : let parts = path_str.split('/').collect::<Vec<&str>>();
599 12202 : let (tenant_id, shard_id, timeline_id) =
600 12202 : if parts.len() > 5 && parts[parts.len() - 5] == TENANTS_SEGMENT_NAME {
601 9198 : let tenant_shard_part = parts[parts.len() - 4];
602 9198 : let (tenant_id, shard_id) = match tenant_shard_part.parse::<TenantShardId>() {
603 9198 : Ok(tenant_shard_id) => (
604 9198 : tenant_shard_id.tenant_id.to_string(),
605 9198 : format!("{}", tenant_shard_id.shard_slug()),
606 9198 : ),
607 : Err(_) => {
608 : // Malformed path: this ID is just for observability, so tolerate it
609 : // and pass through
610 0 : (tenant_shard_part.to_string(), "*".to_string())
611 : }
612 : };
613 9198 : (tenant_id, shard_id, parts[parts.len() - 2].to_string())
614 : } else {
615 3004 : ("*".to_string(), "*".to_string(), "*".to_string())
616 : };
617 12202 : let (handle, mut slot_guard) = get_open_files().find_victim_slot().await;
618 :
619 : // NB: there is also StorageIoOperation::OpenAfterReplace which is for the case
620 : // where our caller doesn't get to use the returned VirtualFile before its
621 : // slot gets re-used by someone else.
622 12202 : let file = observe_duration!(StorageIoOperation::Open, {
623 12202 : open_options.open(path_ref.as_std_path()).await?
624 : });
625 :
626 : // Strip all options other than read and write.
627 : //
628 : // It would perhaps be nicer to check just for the read and write flags
629 : // explicitly, but OpenOptions doesn't contain any functions to read flags,
630 : // only to set them.
631 12202 : let mut reopen_options = open_options.clone();
632 12202 : reopen_options.create(false);
633 12202 : reopen_options.create_new(false);
634 12202 : reopen_options.truncate(false);
635 12202 :
636 12202 : let vfile = VirtualFileInner {
637 12202 : handle: RwLock::new(handle),
638 12202 : pos: 0,
639 12202 : path: path_ref.to_path_buf(),
640 12202 : open_options: reopen_options,
641 12202 : tenant_id,
642 12202 : shard_id,
643 12202 : timeline_id,
644 12202 : };
645 12202 :
646 12202 : // TODO: Under pressure, it's likely the slot will get re-used and
647 12202 : // the underlying file closed before they get around to using it.
648 12202 : // => https://github.com/neondatabase/neon/issues/6065
649 12202 : slot_guard.file.replace(file);
650 12202 :
651 12202 : Ok(vfile)
652 12202 : }
653 :
654 : /// Async version of [`::utils::crashsafe::overwrite`].
655 : ///
656 : /// # NB:
657 : ///
658 : /// Doesn't actually use the [`VirtualFile`] file descriptor cache, but,
659 : /// it did at an earlier time.
660 : /// And it will use this module's [`io_engine`] in the near future, so, leaving it here.
661 56 : pub async fn crashsafe_overwrite<B: BoundedBuf<Buf = Buf> + Send, Buf: IoBuf + Send>(
662 56 : final_path: Utf8PathBuf,
663 56 : tmp_path: Utf8PathBuf,
664 56 : content: B,
665 56 : ) -> std::io::Result<()> {
666 56 : // TODO: use tokio_epoll_uring if configured as `io_engine`.
667 56 : // See https://github.com/neondatabase/neon/issues/6663
668 56 :
669 56 : tokio::task::spawn_blocking(move || {
670 56 : let slice_storage;
671 56 : let content_len = content.bytes_init();
672 56 : let content = if content.bytes_init() > 0 {
673 56 : slice_storage = Some(content.slice(0..content_len));
674 56 : slice_storage.as_deref().expect("just set it to Some()")
675 : } else {
676 0 : &[]
677 : };
678 56 : utils::crashsafe::overwrite(&final_path, &tmp_path, content)
679 56 : .maybe_fatal_err("crashsafe_overwrite")
680 56 : })
681 56 : .await
682 56 : .expect("blocking task is never aborted")
683 56 : }
684 :
685 : /// Call File::sync_all() on the underlying File.
686 5634 : pub async fn sync_all(&self) -> Result<(), Error> {
687 5634 : with_file!(self, StorageIoOperation::Fsync, |file_guard| {
688 5634 : let (_file_guard, res) = io_engine::get().sync_all(file_guard).await;
689 5634 : res.maybe_fatal_err("sync_all")
690 : })
691 5634 : }
692 :
693 : /// Call File::sync_data() on the underlying File.
694 0 : pub async fn sync_data(&self) -> Result<(), Error> {
695 0 : with_file!(self, StorageIoOperation::Fsync, |file_guard| {
696 0 : let (_file_guard, res) = io_engine::get().sync_data(file_guard).await;
697 0 : res.maybe_fatal_err("sync_data")
698 : })
699 0 : }
700 :
701 3604 : pub async fn metadata(&self) -> Result<Metadata, Error> {
702 3604 : with_file!(self, StorageIoOperation::Metadata, |file_guard| {
703 3604 : let (_file_guard, res) = io_engine::get().metadata(file_guard).await;
704 3604 : res
705 : })
706 3604 : }
707 :
708 : /// Helper function internal to `VirtualFile` that looks up the underlying File,
709 : /// opens it and evicts some other File if necessary. The passed parameter is
710 : /// assumed to be a function available for the physical `File`.
711 : ///
712 : /// We are doing it via a macro as Rust doesn't support async closures that
713 : /// take on parameters with lifetimes.
714 3250829 : async fn lock_file(&self) -> Result<FileGuard, Error> {
715 3250829 : let open_files = get_open_files();
716 :
717 378082 : let mut handle_guard = {
718 : // Read the cached slot handle, and see if the slot that it points to still
719 : // contains our File.
720 : //
721 : // We only need to hold the handle lock while we read the current handle. If
722 : // another thread closes the file and recycles the slot for a different file,
723 : // we will notice that the handle we read is no longer valid and retry.
724 3250829 : let mut handle = *self.handle.read().await;
725 : loop {
726 : // Check if the slot contains our File
727 : {
728 3456078 : let slot = &open_files.slots[handle.index];
729 3456078 : let slot_guard = slot.inner.read().await;
730 3456078 : if slot_guard.tag == handle.tag && slot_guard.file.is_some() {
731 : // Found a cached file descriptor.
732 2872747 : slot.recently_used.store(true, Ordering::Relaxed);
733 2872747 : return Ok(FileGuard { slot_guard });
734 583331 : }
735 : }
736 :
737 : // The slot didn't contain our File. We will have to open it ourselves,
738 : // but before that, grab a write lock on handle in the VirtualFile, so
739 : // that no other thread will try to concurrently open the same file.
740 583331 : let handle_guard = self.handle.write().await;
741 :
742 : // If another thread changed the handle while we were not holding the lock,
743 : // then the handle might now be valid again. Loop back to retry.
744 583331 : if *handle_guard != handle {
745 205249 : handle = *handle_guard;
746 205249 : continue;
747 378082 : }
748 378082 : break handle_guard;
749 : }
750 : };
751 :
752 : // We need to open the file ourselves. The handle in the VirtualFile is
753 : // now locked in write-mode. Find a free slot to put it in.
754 378082 : let (handle, mut slot_guard) = open_files.find_victim_slot().await;
755 :
756 : // Re-open the physical file.
757 : // NB: we use StorageIoOperation::OpenAferReplace for this to distinguish this
758 : // case from StorageIoOperation::Open. This helps with identifying thrashing
759 : // of the virtual file descriptor cache.
760 378082 : let file = observe_duration!(StorageIoOperation::OpenAfterReplace, {
761 378082 : self.open_options.open(self.path.as_std_path()).await?
762 : });
763 :
764 : // Store the File in the slot and update the handle in the VirtualFile
765 : // to point to it.
766 378082 : slot_guard.file.replace(file);
767 378082 :
768 378082 : *handle_guard = handle;
769 378082 :
770 378082 : Ok(FileGuard {
771 378082 : slot_guard: slot_guard.downgrade(),
772 378082 : })
773 3250829 : }
774 :
775 516 : pub fn remove(self) {
776 516 : let path = self.path.clone();
777 516 : drop(self);
778 516 : std::fs::remove_file(path).expect("failed to remove the virtual file");
779 516 : }
780 :
781 11352 : pub async fn seek(&mut self, pos: SeekFrom) -> Result<u64, Error> {
782 11352 : match pos {
783 11332 : SeekFrom::Start(offset) => {
784 11332 : self.pos = offset;
785 11332 : }
786 8 : SeekFrom::End(offset) => {
787 8 : self.pos = with_file!(self, StorageIoOperation::Seek, |mut file_guard| file_guard
788 8 : .with_std_file_mut(|std_file| std_file.seek(SeekFrom::End(offset))))?
789 : }
790 12 : SeekFrom::Current(offset) => {
791 12 : let pos = self.pos as i128 + offset as i128;
792 12 : if pos < 0 {
793 4 : return Err(Error::new(
794 4 : ErrorKind::InvalidInput,
795 4 : "offset would be negative",
796 4 : ));
797 8 : }
798 8 : if pos > u64::MAX as i128 {
799 0 : return Err(Error::new(ErrorKind::InvalidInput, "offset overflow"));
800 8 : }
801 8 : self.pos = pos as u64;
802 : }
803 : }
804 11344 : Ok(self.pos)
805 11352 : }
806 :
807 : /// Read the file contents in range `offset..(offset + slice.bytes_total())` into `slice[0..slice.bytes_total()]`.
808 : ///
809 : /// The returned `Slice<Buf>` is equivalent to the input `slice`, i.e., it's the same view into the same buffer.
810 966759 : pub async fn read_exact_at<Buf>(
811 966759 : &self,
812 966759 : slice: Slice<Buf>,
813 966759 : offset: u64,
814 966759 : ctx: &RequestContext,
815 966759 : ) -> Result<Slice<Buf>, Error>
816 966759 : where
817 966759 : Buf: IoBufAlignedMut + Send,
818 966759 : {
819 966759 : let assert_we_return_original_bounds = if cfg!(debug_assertions) {
820 966759 : Some((slice.stable_ptr() as usize, slice.bytes_total()))
821 : } else {
822 0 : None
823 : };
824 :
825 966759 : let original_bounds = slice.bounds();
826 966759 : let (buf, res) =
827 966759 : read_exact_at_impl(slice, offset, |buf, offset| self.read_at(buf, offset, ctx)).await;
828 966759 : let res = res.map(|_| buf.slice(original_bounds));
829 :
830 966759 : if let Some(original_bounds) = assert_we_return_original_bounds {
831 966759 : if let Ok(slice) = &res {
832 966759 : let returned_bounds = (slice.stable_ptr() as usize, slice.bytes_total());
833 966759 : assert_eq!(original_bounds, returned_bounds);
834 0 : }
835 0 : }
836 :
837 966759 : res
838 966759 : }
839 :
840 : /// Like [`Self::read_exact_at`] but for [`PageWriteGuard`].
841 63357 : pub async fn read_exact_at_page(
842 63357 : &self,
843 63357 : page: PageWriteGuard<'static>,
844 63357 : offset: u64,
845 63357 : ctx: &RequestContext,
846 63357 : ) -> Result<PageWriteGuard<'static>, Error> {
847 63357 : let buf = PageWriteGuardBuf { page }.slice_full();
848 63357 : debug_assert_eq!(buf.bytes_total(), PAGE_SZ);
849 63357 : self.read_exact_at(buf, offset, ctx)
850 63357 : .await
851 63357 : .map(|slice| slice.into_inner().page)
852 63357 : }
853 :
854 : // Copied from https://doc.rust-lang.org/1.72.0/src/std/os/unix/fs.rs.html#219-235
855 13222 : pub async fn write_all_at<Buf: IoBuf + Send>(
856 13222 : &self,
857 13222 : buf: FullSlice<Buf>,
858 13222 : mut offset: u64,
859 13222 : ctx: &RequestContext,
860 13222 : ) -> (FullSlice<Buf>, Result<(), Error>) {
861 13222 : let buf = buf.into_raw_slice();
862 13222 : let bounds = buf.bounds();
863 13222 : let restore =
864 13222 : |buf: Slice<_>| FullSlice::must_new(Slice::from_buf_bounds(buf.into_inner(), bounds));
865 13222 : let mut buf = buf;
866 26444 : while !buf.is_empty() {
867 13222 : let (tmp, res) = self.write_at(FullSlice::must_new(buf), offset, ctx).await;
868 13222 : buf = tmp.into_raw_slice();
869 0 : match res {
870 : Ok(0) => {
871 0 : return (
872 0 : restore(buf),
873 0 : Err(Error::new(
874 0 : std::io::ErrorKind::WriteZero,
875 0 : "failed to write whole buffer",
876 0 : )),
877 0 : );
878 : }
879 13222 : Ok(n) => {
880 13222 : buf = buf.slice(n..);
881 13222 : offset += n as u64;
882 13222 : }
883 0 : Err(e) if e.kind() == std::io::ErrorKind::Interrupted => {}
884 0 : Err(e) => return (restore(buf), Err(e)),
885 : }
886 : }
887 13222 : (restore(buf), Ok(()))
888 13222 : }
889 :
890 : /// Writes `buf` to the file at the current offset.
891 : ///
892 : /// Panics if there is an uninitialized range in `buf`, as that is most likely a bug in the caller.
893 2260790 : pub async fn write_all<Buf: IoBuf + Send>(
894 2260790 : &mut self,
895 2260790 : buf: FullSlice<Buf>,
896 2260790 : ctx: &RequestContext,
897 2260790 : ) -> (FullSlice<Buf>, Result<usize, Error>) {
898 2260790 : let buf = buf.into_raw_slice();
899 2260790 : let bounds = buf.bounds();
900 2260790 : let restore =
901 2260790 : |buf: Slice<_>| FullSlice::must_new(Slice::from_buf_bounds(buf.into_inner(), bounds));
902 2260790 : let nbytes = buf.len();
903 2260790 : let mut buf = buf;
904 4521500 : while !buf.is_empty() {
905 2260714 : let (tmp, res) = self.write(FullSlice::must_new(buf), ctx).await;
906 2260714 : buf = tmp.into_raw_slice();
907 4 : match res {
908 : Ok(0) => {
909 0 : return (
910 0 : restore(buf),
911 0 : Err(Error::new(
912 0 : std::io::ErrorKind::WriteZero,
913 0 : "failed to write whole buffer",
914 0 : )),
915 0 : );
916 : }
917 2260710 : Ok(n) => {
918 2260710 : buf = buf.slice(n..);
919 2260710 : }
920 4 : Err(ref e) if e.kind() == std::io::ErrorKind::Interrupted => {}
921 4 : Err(e) => return (restore(buf), Err(e)),
922 : }
923 : }
924 2260786 : (restore(buf), Ok(nbytes))
925 2260790 : }
926 :
927 2260714 : async fn write<B: IoBuf + Send>(
928 2260714 : &mut self,
929 2260714 : buf: FullSlice<B>,
930 2260714 : ctx: &RequestContext,
931 2260714 : ) -> (FullSlice<B>, Result<usize, std::io::Error>) {
932 2260714 : let pos = self.pos;
933 2260714 : let (buf, res) = self.write_at(buf, pos, ctx).await;
934 2260714 : let n = match res {
935 2260710 : Ok(n) => n,
936 4 : Err(e) => return (buf, Err(e)),
937 : };
938 2260710 : self.pos += n as u64;
939 2260710 : (buf, Ok(n))
940 2260714 : }
941 :
942 967647 : pub(crate) async fn read_at<Buf>(
943 967647 : &self,
944 967647 : buf: tokio_epoll_uring::Slice<Buf>,
945 967647 : offset: u64,
946 967647 : _ctx: &RequestContext, /* TODO: use for metrics: https://github.com/neondatabase/neon/issues/6107 */
947 967647 : ) -> (tokio_epoll_uring::Slice<Buf>, Result<usize, Error>)
948 967647 : where
949 967647 : Buf: tokio_epoll_uring::IoBufMut + Send,
950 967647 : {
951 967647 : let file_guard = match self
952 967647 : .lock_file()
953 967647 : .await
954 967647 : .maybe_fatal_err("lock_file inside VirtualFileInner::read_at")
955 : {
956 967647 : Ok(file_guard) => file_guard,
957 0 : Err(e) => return (buf, Err(e)),
958 : };
959 :
960 967647 : observe_duration!(StorageIoOperation::Read, {
961 967647 : let ((_file_guard, buf), res) = io_engine::get().read_at(file_guard, offset, buf).await;
962 967647 : let res = res.maybe_fatal_err("io_engine read_at inside VirtualFileInner::read_at");
963 967647 : if let Ok(size) = res {
964 967643 : STORAGE_IO_SIZE
965 967643 : .with_label_values(&[
966 967643 : "read",
967 967643 : &self.tenant_id,
968 967643 : &self.shard_id,
969 967643 : &self.timeline_id,
970 967643 : ])
971 967643 : .add(size as i64);
972 967643 : }
973 967647 : (buf, res)
974 : })
975 967647 : }
976 :
977 : /// The function aborts the process if the error is fatal.
978 2273936 : async fn write_at<B: IoBuf + Send>(
979 2273936 : &self,
980 2273936 : buf: FullSlice<B>,
981 2273936 : offset: u64,
982 2273936 : _ctx: &RequestContext, /* TODO: use for metrics: https://github.com/neondatabase/neon/issues/6107 */
983 2273936 : ) -> (FullSlice<B>, Result<usize, Error>) {
984 2273936 : let (slice, result) = self.write_at_inner(buf, offset, _ctx).await;
985 2273936 : let result = result.maybe_fatal_err("write_at");
986 2273936 : (slice, result)
987 2273936 : }
988 :
989 2273936 : async fn write_at_inner<B: IoBuf + Send>(
990 2273936 : &self,
991 2273936 : buf: FullSlice<B>,
992 2273936 : offset: u64,
993 2273936 : _ctx: &RequestContext, /* TODO: use for metrics: https://github.com/neondatabase/neon/issues/6107 */
994 2273936 : ) -> (FullSlice<B>, Result<usize, Error>) {
995 2273936 : let file_guard = match self.lock_file().await {
996 2273936 : Ok(file_guard) => file_guard,
997 0 : Err(e) => return (buf, Err(e)),
998 : };
999 2273936 : observe_duration!(StorageIoOperation::Write, {
1000 2273936 : let ((_file_guard, buf), result) =
1001 2273936 : io_engine::get().write_at(file_guard, offset, buf).await;
1002 2273936 : if let Ok(size) = result {
1003 2273932 : STORAGE_IO_SIZE
1004 2273932 : .with_label_values(&[
1005 2273932 : "write",
1006 2273932 : &self.tenant_id,
1007 2273932 : &self.shard_id,
1008 2273932 : &self.timeline_id,
1009 2273932 : ])
1010 2273932 : .add(size as i64);
1011 2273932 : }
1012 2273936 : (buf, result)
1013 : })
1014 2273936 : }
1015 :
1016 448 : async fn read_to_end(&mut self, buf: &mut Vec<u8>, ctx: &RequestContext) -> Result<(), Error> {
1017 448 : let mut tmp = vec![0; 128];
1018 : loop {
1019 888 : let slice = tmp.slice(..128);
1020 888 : let (slice, res) = self.read_at(slice, self.pos, ctx).await;
1021 4 : match res {
1022 444 : Ok(0) => return Ok(()),
1023 440 : Ok(n) => {
1024 440 : self.pos += n as u64;
1025 440 : buf.extend_from_slice(&slice[..n]);
1026 440 : }
1027 4 : Err(ref e) if e.kind() == std::io::ErrorKind::Interrupted => {}
1028 4 : Err(e) => return Err(e),
1029 : }
1030 440 : tmp = slice.into_inner();
1031 : }
1032 448 : }
1033 : }
1034 :
1035 : // Adapted from https://doc.rust-lang.org/1.72.0/src/std/os/unix/fs.rs.html#117-135
1036 966775 : pub async fn read_exact_at_impl<Buf, F, Fut>(
1037 966775 : mut buf: tokio_epoll_uring::Slice<Buf>,
1038 966775 : mut offset: u64,
1039 966775 : mut read_at: F,
1040 966775 : ) -> (Buf, std::io::Result<()>)
1041 966775 : where
1042 966775 : Buf: IoBufMut + Send,
1043 966775 : F: FnMut(tokio_epoll_uring::Slice<Buf>, u64) -> Fut,
1044 966775 : Fut: std::future::Future<Output = (tokio_epoll_uring::Slice<Buf>, std::io::Result<usize>)>,
1045 966775 : {
1046 1933554 : while buf.bytes_total() != 0 {
1047 : let res;
1048 966783 : (buf, res) = read_at(buf, offset).await;
1049 0 : match res {
1050 4 : Ok(0) => break,
1051 966779 : Ok(n) => {
1052 966779 : buf = buf.slice(n..);
1053 966779 : offset += n as u64;
1054 966779 : }
1055 0 : Err(ref e) if e.kind() == std::io::ErrorKind::Interrupted => {}
1056 0 : Err(e) => return (buf.into_inner(), Err(e)),
1057 : }
1058 : }
1059 : // NB: don't use `buf.is_empty()` here; it is from the
1060 : // `impl Deref for Slice { Target = [u8] }`; the &[u8]
1061 : // returned by it only covers the initialized portion of `buf`.
1062 : // Whereas we're interested in ensuring that we filled the entire
1063 : // buffer that the user passed in.
1064 966775 : if buf.bytes_total() != 0 {
1065 4 : (
1066 4 : buf.into_inner(),
1067 4 : Err(std::io::Error::new(
1068 4 : std::io::ErrorKind::UnexpectedEof,
1069 4 : "failed to fill whole buffer",
1070 4 : )),
1071 4 : )
1072 : } else {
1073 966771 : assert_eq!(buf.len(), buf.bytes_total());
1074 966771 : (buf.into_inner(), Ok(()))
1075 : }
1076 966775 : }
1077 :
1078 : #[cfg(test)]
1079 : mod test_read_exact_at_impl {
1080 :
1081 : use std::{collections::VecDeque, sync::Arc};
1082 :
1083 : use tokio_epoll_uring::{BoundedBuf, BoundedBufMut};
1084 :
1085 : use super::read_exact_at_impl;
1086 :
1087 : struct Expectation {
1088 : offset: u64,
1089 : bytes_total: usize,
1090 : result: std::io::Result<Vec<u8>>,
1091 : }
1092 : struct MockReadAt {
1093 : expectations: VecDeque<Expectation>,
1094 : }
1095 :
1096 : impl MockReadAt {
1097 24 : async fn read_at(
1098 24 : &mut self,
1099 24 : mut buf: tokio_epoll_uring::Slice<Vec<u8>>,
1100 24 : offset: u64,
1101 24 : ) -> (tokio_epoll_uring::Slice<Vec<u8>>, std::io::Result<usize>) {
1102 24 : let exp = self
1103 24 : .expectations
1104 24 : .pop_front()
1105 24 : .expect("read_at called but we have no expectations left");
1106 24 : assert_eq!(exp.offset, offset);
1107 24 : assert_eq!(exp.bytes_total, buf.bytes_total());
1108 24 : match exp.result {
1109 24 : Ok(bytes) => {
1110 24 : assert!(bytes.len() <= buf.bytes_total());
1111 24 : buf.put_slice(&bytes);
1112 24 : (buf, Ok(bytes.len()))
1113 : }
1114 0 : Err(e) => (buf, Err(e)),
1115 : }
1116 24 : }
1117 : }
1118 :
1119 : impl Drop for MockReadAt {
1120 16 : fn drop(&mut self) {
1121 16 : assert_eq!(self.expectations.len(), 0);
1122 16 : }
1123 : }
1124 :
1125 : #[tokio::test]
1126 4 : async fn test_basic() {
1127 4 : let buf = Vec::with_capacity(5).slice_full();
1128 4 : let mock_read_at = Arc::new(tokio::sync::Mutex::new(MockReadAt {
1129 4 : expectations: VecDeque::from(vec![Expectation {
1130 4 : offset: 0,
1131 4 : bytes_total: 5,
1132 4 : result: Ok(vec![b'a', b'b', b'c', b'd', b'e']),
1133 4 : }]),
1134 4 : }));
1135 4 : let (buf, res) = read_exact_at_impl(buf, 0, |buf, offset| {
1136 4 : let mock_read_at = Arc::clone(&mock_read_at);
1137 4 : async move { mock_read_at.lock().await.read_at(buf, offset).await }
1138 4 : })
1139 4 : .await;
1140 4 : assert!(res.is_ok());
1141 4 : assert_eq!(buf, vec![b'a', b'b', b'c', b'd', b'e']);
1142 4 : }
1143 :
1144 : #[tokio::test]
1145 4 : async fn test_empty_buf_issues_no_syscall() {
1146 4 : let buf = Vec::new().slice_full();
1147 4 : let mock_read_at = Arc::new(tokio::sync::Mutex::new(MockReadAt {
1148 4 : expectations: VecDeque::new(),
1149 4 : }));
1150 4 : let (_buf, res) = read_exact_at_impl(buf, 0, |buf, offset| {
1151 0 : let mock_read_at = Arc::clone(&mock_read_at);
1152 4 : async move { mock_read_at.lock().await.read_at(buf, offset).await }
1153 4 : })
1154 4 : .await;
1155 4 : assert!(res.is_ok());
1156 4 : }
1157 :
1158 : #[tokio::test]
1159 4 : async fn test_two_read_at_calls_needed_until_buf_filled() {
1160 4 : let buf = Vec::with_capacity(4).slice_full();
1161 4 : let mock_read_at = Arc::new(tokio::sync::Mutex::new(MockReadAt {
1162 4 : expectations: VecDeque::from(vec![
1163 4 : Expectation {
1164 4 : offset: 0,
1165 4 : bytes_total: 4,
1166 4 : result: Ok(vec![b'a', b'b']),
1167 4 : },
1168 4 : Expectation {
1169 4 : offset: 2,
1170 4 : bytes_total: 2,
1171 4 : result: Ok(vec![b'c', b'd']),
1172 4 : },
1173 4 : ]),
1174 4 : }));
1175 8 : let (buf, res) = read_exact_at_impl(buf, 0, |buf, offset| {
1176 8 : let mock_read_at = Arc::clone(&mock_read_at);
1177 8 : async move { mock_read_at.lock().await.read_at(buf, offset).await }
1178 8 : })
1179 4 : .await;
1180 4 : assert!(res.is_ok());
1181 4 : assert_eq!(buf, vec![b'a', b'b', b'c', b'd']);
1182 4 : }
1183 :
1184 : #[tokio::test]
1185 4 : async fn test_eof_before_buffer_full() {
1186 4 : let buf = Vec::with_capacity(3).slice_full();
1187 4 : let mock_read_at = Arc::new(tokio::sync::Mutex::new(MockReadAt {
1188 4 : expectations: VecDeque::from(vec![
1189 4 : Expectation {
1190 4 : offset: 0,
1191 4 : bytes_total: 3,
1192 4 : result: Ok(vec![b'a']),
1193 4 : },
1194 4 : Expectation {
1195 4 : offset: 1,
1196 4 : bytes_total: 2,
1197 4 : result: Ok(vec![b'b']),
1198 4 : },
1199 4 : Expectation {
1200 4 : offset: 2,
1201 4 : bytes_total: 1,
1202 4 : result: Ok(vec![]),
1203 4 : },
1204 4 : ]),
1205 4 : }));
1206 12 : let (_buf, res) = read_exact_at_impl(buf, 0, |buf, offset| {
1207 12 : let mock_read_at = Arc::clone(&mock_read_at);
1208 12 : async move { mock_read_at.lock().await.read_at(buf, offset).await }
1209 12 : })
1210 4 : .await;
1211 4 : let Err(err) = res else {
1212 4 : panic!("should return an error");
1213 4 : };
1214 4 : assert_eq!(err.kind(), std::io::ErrorKind::UnexpectedEof);
1215 4 : assert_eq!(format!("{err}"), "failed to fill whole buffer");
1216 4 : // buffer contents on error are unspecified
1217 4 : }
1218 : }
1219 :
1220 : struct FileGuard {
1221 : slot_guard: RwLockReadGuard<'static, SlotInner>,
1222 : }
1223 :
1224 : impl AsRef<OwnedFd> for FileGuard {
1225 3250829 : fn as_ref(&self) -> &OwnedFd {
1226 3250829 : // This unwrap is safe because we only create `FileGuard`s
1227 3250829 : // if we know that the file is Some.
1228 3250829 : self.slot_guard.file.as_ref().unwrap()
1229 3250829 : }
1230 : }
1231 :
1232 : impl FileGuard {
1233 : /// Soft deprecation: we'll move VirtualFile to async APIs and remove this function eventually.
1234 1625124 : fn with_std_file<F, R>(&self, with: F) -> R
1235 1625124 : where
1236 1625124 : F: FnOnce(&File) -> R,
1237 1625124 : {
1238 1625124 : // SAFETY:
1239 1625124 : // - lifetime of the fd: `file` doesn't outlive the OwnedFd stored in `self`.
1240 1625124 : // - `&` usage below: `self` is `&`, hence Rust typesystem guarantees there are is no `&mut`
1241 1625124 : let file = unsafe { File::from_raw_fd(self.as_ref().as_raw_fd()) };
1242 1625124 : let res = with(&file);
1243 1625124 : let _ = file.into_raw_fd();
1244 1625124 : res
1245 1625124 : }
1246 : /// Soft deprecation: we'll move VirtualFile to async APIs and remove this function eventually.
1247 8 : fn with_std_file_mut<F, R>(&mut self, with: F) -> R
1248 8 : where
1249 8 : F: FnOnce(&mut File) -> R,
1250 8 : {
1251 8 : // SAFETY:
1252 8 : // - lifetime of the fd: `file` doesn't outlive the OwnedFd stored in `self`.
1253 8 : // - &mut usage below: `self` is `&mut`, hence this call is the only task/thread that has control over the underlying fd
1254 8 : let mut file = unsafe { File::from_raw_fd(self.as_ref().as_raw_fd()) };
1255 8 : let res = with(&mut file);
1256 8 : let _ = file.into_raw_fd();
1257 8 : res
1258 8 : }
1259 : }
1260 :
1261 : impl tokio_epoll_uring::IoFd for FileGuard {
1262 1625697 : unsafe fn as_fd(&self) -> RawFd {
1263 1625697 : let owned_fd: &OwnedFd = self.as_ref();
1264 1625697 : owned_fd.as_raw_fd()
1265 1625697 : }
1266 : }
1267 :
1268 : #[cfg(test)]
1269 : impl VirtualFile {
1270 41832 : pub(crate) async fn read_blk(
1271 41832 : &self,
1272 41832 : blknum: u32,
1273 41832 : ctx: &RequestContext,
1274 41832 : ) -> Result<crate::tenant::block_io::BlockLease<'_>, std::io::Error> {
1275 41832 : self.inner.read_blk(blknum, ctx).await
1276 41832 : }
1277 : }
1278 :
1279 : #[cfg(test)]
1280 : impl VirtualFileInner {
1281 41832 : pub(crate) async fn read_blk(
1282 41832 : &self,
1283 41832 : blknum: u32,
1284 41832 : ctx: &RequestContext,
1285 41832 : ) -> Result<crate::tenant::block_io::BlockLease<'_>, std::io::Error> {
1286 : use crate::page_cache::PAGE_SZ;
1287 41832 : let slice = IoBufferMut::with_capacity(PAGE_SZ).slice_full();
1288 41832 : assert_eq!(slice.bytes_total(), PAGE_SZ);
1289 41832 : let slice = self
1290 41832 : .read_exact_at(slice, blknum as u64 * (PAGE_SZ as u64), ctx)
1291 41832 : .await?;
1292 41832 : Ok(crate::tenant::block_io::BlockLease::IoBufferMut(
1293 41832 : slice.into_inner(),
1294 41832 : ))
1295 41832 : }
1296 : }
1297 :
1298 : impl Drop for VirtualFileInner {
1299 : /// If a VirtualFile is dropped, close the underlying file if it was open.
1300 10590 : fn drop(&mut self) {
1301 10590 : let handle = self.handle.get_mut();
1302 :
1303 10590 : fn clean_slot(slot: &Slot, mut slot_guard: RwLockWriteGuard<'_, SlotInner>, tag: u64) {
1304 10590 : if slot_guard.tag == tag {
1305 9436 : slot.recently_used.store(false, Ordering::Relaxed);
1306 : // there is also operation "close-by-replace" for closes done on eviction for
1307 : // comparison.
1308 9436 : if let Some(fd) = slot_guard.file.take() {
1309 9436 : STORAGE_IO_TIME_METRIC
1310 9436 : .get(StorageIoOperation::Close)
1311 9436 : .observe_closure_duration(|| drop(fd));
1312 9436 : }
1313 1154 : }
1314 10590 : }
1315 :
1316 : // We don't have async drop so we cannot directly await the lock here.
1317 : // Instead, first do a best-effort attempt at closing the underlying
1318 : // file descriptor by using `try_write`, and if that fails, spawn
1319 : // a tokio task to do it asynchronously: we just want it to be
1320 : // cleaned up eventually.
1321 : // Most of the time, the `try_lock` should succeed though,
1322 : // as we have `&mut self` access. In other words, if the slot
1323 : // is still occupied by our file, there should be no access from
1324 : // other I/O operations; the only other possible place to lock
1325 : // the slot is the lock algorithm looking for free slots.
1326 10590 : let slot = &get_open_files().slots[handle.index];
1327 10590 : if let Ok(slot_guard) = slot.inner.try_write() {
1328 10590 : clean_slot(slot, slot_guard, handle.tag);
1329 10590 : } else {
1330 0 : let tag = handle.tag;
1331 0 : tokio::spawn(async move {
1332 0 : let slot_guard = slot.inner.write().await;
1333 0 : clean_slot(slot, slot_guard, tag);
1334 0 : });
1335 0 : };
1336 10590 : }
1337 : }
1338 :
1339 : impl OwnedAsyncWriter for VirtualFile {
1340 13214 : async fn write_all_at<Buf: IoBufAligned + Send>(
1341 13214 : &self,
1342 13214 : buf: FullSlice<Buf>,
1343 13214 : offset: u64,
1344 13214 : ctx: &RequestContext,
1345 13214 : ) -> std::io::Result<FullSlice<Buf>> {
1346 13214 : let (buf, res) = VirtualFile::write_all_at(self, buf, offset, ctx).await;
1347 13214 : res.map(|_| buf)
1348 13214 : }
1349 : }
1350 :
1351 : impl OpenFiles {
1352 464 : fn new(num_slots: usize) -> OpenFiles {
1353 464 : let mut slots = Box::new(Vec::with_capacity(num_slots));
1354 4640 : for _ in 0..num_slots {
1355 4640 : let slot = Slot {
1356 4640 : recently_used: AtomicBool::new(false),
1357 4640 : inner: RwLock::new(SlotInner { tag: 0, file: None }),
1358 4640 : };
1359 4640 : slots.push(slot);
1360 4640 : }
1361 :
1362 464 : OpenFiles {
1363 464 : next: AtomicUsize::new(0),
1364 464 : slots: Box::leak(slots),
1365 464 : }
1366 464 : }
1367 : }
1368 :
1369 : ///
1370 : /// Initialize the virtual file module. This must be called once at page
1371 : /// server startup.
1372 : ///
1373 : #[cfg(not(test))]
1374 0 : pub fn init(num_slots: usize, engine: IoEngineKind, mode: IoMode, sync_mode: SyncMode) {
1375 0 : if OPEN_FILES.set(OpenFiles::new(num_slots)).is_err() {
1376 0 : panic!("virtual_file::init called twice");
1377 0 : }
1378 0 : set_io_mode(mode);
1379 0 : io_engine::init(engine);
1380 0 : SYNC_MODE.store(sync_mode as u8, std::sync::atomic::Ordering::Relaxed);
1381 0 : crate::metrics::virtual_file_descriptor_cache::SIZE_MAX.set(num_slots as u64);
1382 0 : }
1383 :
1384 : const TEST_MAX_FILE_DESCRIPTORS: usize = 10;
1385 :
1386 : // Get a handle to the global slots array.
1387 3273621 : fn get_open_files() -> &'static OpenFiles {
1388 3273621 : //
1389 3273621 : // In unit tests, page server startup doesn't happen and no one calls
1390 3273621 : // virtual_file::init(). Initialize it here, with a small array.
1391 3273621 : //
1392 3273621 : // This applies to the virtual file tests below, but all other unit
1393 3273621 : // tests too, so the virtual file facility is always usable in
1394 3273621 : // unit tests.
1395 3273621 : //
1396 3273621 : if cfg!(test) {
1397 3273621 : OPEN_FILES.get_or_init(|| OpenFiles::new(TEST_MAX_FILE_DESCRIPTORS))
1398 : } else {
1399 0 : OPEN_FILES.get().expect("virtual_file::init not called yet")
1400 : }
1401 3273621 : }
1402 :
1403 : /// Gets the io buffer alignment.
1404 0 : pub(crate) const fn get_io_buffer_alignment() -> usize {
1405 0 : DEFAULT_IO_BUFFER_ALIGNMENT
1406 0 : }
1407 :
1408 : pub(crate) type IoBufferMut = AlignedBufferMut<ConstAlign<{ get_io_buffer_alignment() }>>;
1409 : pub(crate) type IoBuffer = AlignedBuffer<ConstAlign<{ get_io_buffer_alignment() }>>;
1410 : pub(crate) type IoPageSlice<'a> =
1411 : AlignedSlice<'a, PAGE_SZ, ConstAlign<{ get_io_buffer_alignment() }>>;
1412 :
1413 : static IO_MODE: AtomicU8 = AtomicU8::new(IoMode::preferred() as u8);
1414 :
1415 0 : pub(crate) fn set_io_mode(mode: IoMode) {
1416 0 : IO_MODE.store(mode as u8, std::sync::atomic::Ordering::Relaxed);
1417 0 : }
1418 :
1419 5068 : pub(crate) fn get_io_mode() -> IoMode {
1420 5068 : IoMode::try_from(IO_MODE.load(Ordering::Relaxed)).unwrap()
1421 5068 : }
1422 :
1423 : static SYNC_MODE: AtomicU8 = AtomicU8::new(SyncMode::Sync as u8);
1424 :
1425 : #[cfg(test)]
1426 : mod tests {
1427 : use crate::context::DownloadBehavior;
1428 : use crate::task_mgr::TaskKind;
1429 :
1430 : use super::*;
1431 : use owned_buffers_io::io_buf_ext::IoBufExt;
1432 : use owned_buffers_io::slice::SliceMutExt;
1433 : use rand::seq::SliceRandom;
1434 : use rand::thread_rng;
1435 : use rand::Rng;
1436 : use std::io::Write;
1437 : use std::os::unix::fs::FileExt;
1438 : use std::sync::Arc;
1439 :
1440 : enum MaybeVirtualFile {
1441 : VirtualFile(VirtualFile),
1442 : File(File),
1443 : }
1444 :
1445 : impl From<VirtualFile> for MaybeVirtualFile {
1446 12 : fn from(vf: VirtualFile) -> Self {
1447 12 : MaybeVirtualFile::VirtualFile(vf)
1448 12 : }
1449 : }
1450 :
1451 : impl MaybeVirtualFile {
1452 808 : async fn read_exact_at(
1453 808 : &self,
1454 808 : mut slice: tokio_epoll_uring::Slice<IoBufferMut>,
1455 808 : offset: u64,
1456 808 : ctx: &RequestContext,
1457 808 : ) -> Result<tokio_epoll_uring::Slice<IoBufferMut>, Error> {
1458 808 : match self {
1459 404 : MaybeVirtualFile::VirtualFile(file) => file.read_exact_at(slice, offset, ctx).await,
1460 404 : MaybeVirtualFile::File(file) => {
1461 404 : let rust_slice: &mut [u8] = slice.as_mut_rust_slice_full_zeroed();
1462 404 : file.read_exact_at(rust_slice, offset).map(|()| slice)
1463 : }
1464 : }
1465 808 : }
1466 16 : async fn write_all_at<Buf: IoBufAligned + Send>(
1467 16 : &self,
1468 16 : buf: FullSlice<Buf>,
1469 16 : offset: u64,
1470 16 : ctx: &RequestContext,
1471 16 : ) -> Result<(), Error> {
1472 16 : match self {
1473 8 : MaybeVirtualFile::VirtualFile(file) => {
1474 8 : let (_buf, res) = file.write_all_at(buf, offset, ctx).await;
1475 8 : res
1476 : }
1477 8 : MaybeVirtualFile::File(file) => file.write_all_at(&buf[..], offset),
1478 : }
1479 16 : }
1480 72 : async fn seek(&mut self, pos: SeekFrom) -> Result<u64, Error> {
1481 72 : match self {
1482 36 : MaybeVirtualFile::VirtualFile(file) => file.seek(pos).await,
1483 36 : MaybeVirtualFile::File(file) => file.seek(pos),
1484 : }
1485 72 : }
1486 16 : async fn write_all<Buf: IoBuf + Send>(
1487 16 : &mut self,
1488 16 : buf: FullSlice<Buf>,
1489 16 : ctx: &RequestContext,
1490 16 : ) -> Result<(), Error> {
1491 16 : match self {
1492 8 : MaybeVirtualFile::VirtualFile(file) => {
1493 8 : let (_buf, res) = file.write_all(buf, ctx).await;
1494 8 : res.map(|_| ())
1495 : }
1496 8 : MaybeVirtualFile::File(file) => file.write_all(&buf[..]),
1497 : }
1498 16 : }
1499 :
1500 : // Helper function to slurp contents of a file, starting at the current position,
1501 : // into a string
1502 884 : async fn read_string(&mut self, ctx: &RequestContext) -> Result<String, Error> {
1503 : use std::io::Read;
1504 884 : let mut buf = String::new();
1505 884 : match self {
1506 448 : MaybeVirtualFile::VirtualFile(file) => {
1507 448 : let mut buf = Vec::new();
1508 448 : file.read_to_end(&mut buf, ctx).await?;
1509 444 : return Ok(String::from_utf8(buf).unwrap());
1510 : }
1511 436 : MaybeVirtualFile::File(file) => {
1512 436 : file.read_to_string(&mut buf)?;
1513 : }
1514 : }
1515 432 : Ok(buf)
1516 884 : }
1517 :
1518 : // Helper function to slurp a portion of a file into a string
1519 808 : async fn read_string_at(
1520 808 : &mut self,
1521 808 : pos: u64,
1522 808 : len: usize,
1523 808 : ctx: &RequestContext,
1524 808 : ) -> Result<String, Error> {
1525 808 : let slice = IoBufferMut::with_capacity(len).slice_full();
1526 808 : assert_eq!(slice.bytes_total(), len);
1527 808 : let slice = self.read_exact_at(slice, pos, ctx).await?;
1528 808 : let buf = slice.into_inner();
1529 808 : assert_eq!(buf.len(), len);
1530 :
1531 808 : Ok(String::from_utf8(buf.to_vec()).unwrap())
1532 808 : }
1533 : }
1534 :
1535 : #[tokio::test]
1536 4 : async fn test_virtual_files() -> anyhow::Result<()> {
1537 4 : // The real work is done in the test_files() helper function. This
1538 4 : // allows us to run the same set of tests against a native File, and
1539 4 : // VirtualFile. We trust the native Files and wouldn't need to test them,
1540 4 : // but this allows us to verify that the operations return the same
1541 4 : // results with VirtualFiles as with native Files. (Except that with
1542 4 : // native files, you will run out of file descriptors if the ulimit
1543 4 : // is low enough.)
1544 4 : struct A;
1545 4 :
1546 4 : impl Adapter for A {
1547 412 : async fn open(
1548 412 : path: Utf8PathBuf,
1549 412 : opts: OpenOptions,
1550 412 : ctx: &RequestContext,
1551 412 : ) -> Result<MaybeVirtualFile, anyhow::Error> {
1552 412 : let vf = VirtualFile::open_with_options(&path, &opts, ctx).await?;
1553 412 : Ok(MaybeVirtualFile::VirtualFile(vf))
1554 412 : }
1555 4 : }
1556 4 : test_files::<A>("virtual_files").await
1557 4 : }
1558 :
1559 : #[tokio::test]
1560 4 : async fn test_physical_files() -> anyhow::Result<()> {
1561 4 : struct B;
1562 4 :
1563 4 : impl Adapter for B {
1564 412 : async fn open(
1565 412 : path: Utf8PathBuf,
1566 412 : opts: OpenOptions,
1567 412 : _ctx: &RequestContext,
1568 412 : ) -> Result<MaybeVirtualFile, anyhow::Error> {
1569 4 : Ok(MaybeVirtualFile::File({
1570 412 : let owned_fd = opts.open(path.as_std_path()).await?;
1571 412 : File::from(owned_fd)
1572 4 : }))
1573 412 : }
1574 4 : }
1575 4 :
1576 4 : test_files::<B>("physical_files").await
1577 4 : }
1578 :
1579 : /// This is essentially a closure which returns a MaybeVirtualFile, but because rust edition
1580 : /// 2024 is not yet out with new lifetime capture or outlives rules, this is a async function
1581 : /// in trait which benefits from the new lifetime capture rules already.
1582 : trait Adapter {
1583 : async fn open(
1584 : path: Utf8PathBuf,
1585 : opts: OpenOptions,
1586 : ctx: &RequestContext,
1587 : ) -> Result<MaybeVirtualFile, anyhow::Error>;
1588 : }
1589 :
1590 8 : async fn test_files<A>(testname: &str) -> anyhow::Result<()>
1591 8 : where
1592 8 : A: Adapter,
1593 8 : {
1594 8 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
1595 8 : let testdir = crate::config::PageServerConf::test_repo_dir(testname);
1596 8 : std::fs::create_dir_all(&testdir)?;
1597 :
1598 8 : let path_a = testdir.join("file_a");
1599 8 : let mut file_a = A::open(
1600 8 : path_a.clone(),
1601 8 : OpenOptions::new()
1602 8 : .write(true)
1603 8 : .create(true)
1604 8 : .truncate(true)
1605 8 : .to_owned(),
1606 8 : &ctx,
1607 8 : )
1608 8 : .await?;
1609 :
1610 8 : file_a
1611 8 : .write_all(b"foobar".to_vec().slice_len(), &ctx)
1612 8 : .await?;
1613 :
1614 : // cannot read from a file opened in write-only mode
1615 8 : let _ = file_a.read_string(&ctx).await.unwrap_err();
1616 :
1617 : // Close the file and re-open for reading
1618 8 : let mut file_a = A::open(path_a, OpenOptions::new().read(true).to_owned(), &ctx).await?;
1619 :
1620 : // cannot write to a file opened in read-only mode
1621 8 : let _ = file_a
1622 8 : .write_all(b"bar".to_vec().slice_len(), &ctx)
1623 8 : .await
1624 8 : .unwrap_err();
1625 8 :
1626 8 : // Try simple read
1627 8 : assert_eq!("foobar", file_a.read_string(&ctx).await?);
1628 :
1629 : // It's positioned at the EOF now.
1630 8 : assert_eq!("", file_a.read_string(&ctx).await?);
1631 :
1632 : // Test seeks.
1633 8 : assert_eq!(file_a.seek(SeekFrom::Start(1)).await?, 1);
1634 8 : assert_eq!("oobar", file_a.read_string(&ctx).await?);
1635 :
1636 8 : assert_eq!(file_a.seek(SeekFrom::End(-2)).await?, 4);
1637 8 : assert_eq!("ar", file_a.read_string(&ctx).await?);
1638 :
1639 8 : assert_eq!(file_a.seek(SeekFrom::Start(1)).await?, 1);
1640 8 : assert_eq!(file_a.seek(SeekFrom::Current(2)).await?, 3);
1641 8 : assert_eq!("bar", file_a.read_string(&ctx).await?);
1642 :
1643 8 : assert_eq!(file_a.seek(SeekFrom::Current(-5)).await?, 1);
1644 8 : assert_eq!("oobar", file_a.read_string(&ctx).await?);
1645 :
1646 : // Test erroneous seeks to before byte 0
1647 8 : file_a.seek(SeekFrom::End(-7)).await.unwrap_err();
1648 8 : assert_eq!(file_a.seek(SeekFrom::Start(1)).await?, 1);
1649 8 : file_a.seek(SeekFrom::Current(-2)).await.unwrap_err();
1650 8 :
1651 8 : // the erroneous seek should have left the position unchanged
1652 8 : assert_eq!("oobar", file_a.read_string(&ctx).await?);
1653 :
1654 : // Create another test file, and try FileExt functions on it.
1655 8 : let path_b = testdir.join("file_b");
1656 8 : let mut file_b = A::open(
1657 8 : path_b.clone(),
1658 8 : OpenOptions::new()
1659 8 : .read(true)
1660 8 : .write(true)
1661 8 : .create(true)
1662 8 : .truncate(true)
1663 8 : .to_owned(),
1664 8 : &ctx,
1665 8 : )
1666 8 : .await?;
1667 8 : file_b
1668 8 : .write_all_at(IoBuffer::from(b"BAR").slice_len(), 3, &ctx)
1669 8 : .await?;
1670 8 : file_b
1671 8 : .write_all_at(IoBuffer::from(b"FOO").slice_len(), 0, &ctx)
1672 8 : .await?;
1673 :
1674 8 : assert_eq!(file_b.read_string_at(2, 3, &ctx).await?, "OBA");
1675 :
1676 : // Open a lot of files, enough to cause some evictions. (Or to be precise,
1677 : // open the same file many times. The effect is the same.)
1678 : //
1679 : // leave file_a positioned at offset 1 before we start
1680 8 : assert_eq!(file_a.seek(SeekFrom::Start(1)).await?, 1);
1681 :
1682 8 : let mut vfiles = Vec::new();
1683 808 : for _ in 0..100 {
1684 800 : let mut vfile = A::open(
1685 800 : path_b.clone(),
1686 800 : OpenOptions::new().read(true).to_owned(),
1687 800 : &ctx,
1688 800 : )
1689 800 : .await?;
1690 800 : assert_eq!("FOOBAR", vfile.read_string(&ctx).await?);
1691 800 : vfiles.push(vfile);
1692 : }
1693 :
1694 : // make sure we opened enough files to definitely cause evictions.
1695 8 : assert!(vfiles.len() > TEST_MAX_FILE_DESCRIPTORS * 2);
1696 :
1697 : // The underlying file descriptor for 'file_a' should be closed now. Try to read
1698 : // from it again. We left the file positioned at offset 1 above.
1699 8 : assert_eq!("oobar", file_a.read_string(&ctx).await?);
1700 :
1701 : // Check that all the other FDs still work too. Use them in random order for
1702 : // good measure.
1703 8 : vfiles.as_mut_slice().shuffle(&mut thread_rng());
1704 800 : for vfile in vfiles.iter_mut() {
1705 800 : assert_eq!("OOBAR", vfile.read_string_at(1, 5, &ctx).await?);
1706 : }
1707 :
1708 8 : Ok(())
1709 8 : }
1710 :
1711 : /// Test using VirtualFiles from many threads concurrently. This tests both using
1712 : /// a lot of VirtualFiles concurrently, causing evictions, and also using the same
1713 : /// VirtualFile from multiple threads concurrently.
1714 : #[tokio::test]
1715 4 : async fn test_vfile_concurrency() -> Result<(), Error> {
1716 4 : const SIZE: usize = 8 * 1024;
1717 4 : const VIRTUAL_FILES: usize = 100;
1718 4 : const THREADS: usize = 100;
1719 4 : const SAMPLE: [u8; SIZE] = [0xADu8; SIZE];
1720 4 :
1721 4 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
1722 4 : let testdir = crate::config::PageServerConf::test_repo_dir("vfile_concurrency");
1723 4 : std::fs::create_dir_all(&testdir)?;
1724 4 :
1725 4 : // Create a test file.
1726 4 : let test_file_path = testdir.join("concurrency_test_file");
1727 4 : {
1728 4 : let file = File::create(&test_file_path)?;
1729 4 : file.write_all_at(&SAMPLE, 0)?;
1730 4 : }
1731 4 :
1732 4 : // Open the file many times.
1733 4 : let mut files = Vec::new();
1734 404 : for _ in 0..VIRTUAL_FILES {
1735 400 : let f = VirtualFileInner::open_with_options(
1736 400 : &test_file_path,
1737 400 : OpenOptions::new().read(true),
1738 400 : &ctx,
1739 400 : )
1740 400 : .await?;
1741 400 : files.push(f);
1742 4 : }
1743 4 : let files = Arc::new(files);
1744 4 :
1745 4 : // Launch many threads, and use the virtual files concurrently in random order.
1746 4 : let rt = tokio::runtime::Builder::new_multi_thread()
1747 4 : .worker_threads(THREADS)
1748 4 : .thread_name("test_vfile_concurrency thread")
1749 4 : .build()
1750 4 : .unwrap();
1751 4 : let mut hdls = Vec::new();
1752 404 : for _threadno in 0..THREADS {
1753 400 : let files = files.clone();
1754 400 : let ctx = ctx.detached_child(TaskKind::UnitTest, DownloadBehavior::Error);
1755 400 : let hdl = rt.spawn(async move {
1756 400 : let mut buf = IoBufferMut::with_capacity_zeroed(SIZE);
1757 400 : let mut rng = rand::rngs::OsRng;
1758 400000 : for _ in 1..1000 {
1759 399600 : let f = &files[rng.gen_range(0..files.len())];
1760 399600 : buf = f
1761 399600 : .read_exact_at(buf.slice_full(), 0, &ctx)
1762 399600 : .await
1763 399600 : .unwrap()
1764 399600 : .into_inner();
1765 399600 : assert!(buf[..] == SAMPLE);
1766 4 : }
1767 400 : });
1768 400 : hdls.push(hdl);
1769 400 : }
1770 404 : for hdl in hdls {
1771 400 : hdl.await?;
1772 4 : }
1773 4 : std::mem::forget(rt);
1774 4 :
1775 4 : Ok(())
1776 4 : }
1777 :
1778 : #[tokio::test]
1779 4 : async fn test_atomic_overwrite_basic() {
1780 4 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
1781 4 : let testdir = crate::config::PageServerConf::test_repo_dir("test_atomic_overwrite_basic");
1782 4 : std::fs::create_dir_all(&testdir).unwrap();
1783 4 :
1784 4 : let path = testdir.join("myfile");
1785 4 : let tmp_path = testdir.join("myfile.tmp");
1786 4 :
1787 4 : VirtualFileInner::crashsafe_overwrite(path.clone(), tmp_path.clone(), b"foo".to_vec())
1788 4 : .await
1789 4 : .unwrap();
1790 4 : let mut file = MaybeVirtualFile::from(VirtualFile::open(&path, &ctx).await.unwrap());
1791 4 : let post = file.read_string(&ctx).await.unwrap();
1792 4 : assert_eq!(post, "foo");
1793 4 : assert!(!tmp_path.exists());
1794 4 : drop(file);
1795 4 :
1796 4 : VirtualFileInner::crashsafe_overwrite(path.clone(), tmp_path.clone(), b"bar".to_vec())
1797 4 : .await
1798 4 : .unwrap();
1799 4 : let mut file = MaybeVirtualFile::from(VirtualFile::open(&path, &ctx).await.unwrap());
1800 4 : let post = file.read_string(&ctx).await.unwrap();
1801 4 : assert_eq!(post, "bar");
1802 4 : assert!(!tmp_path.exists());
1803 4 : drop(file);
1804 4 : }
1805 :
1806 : #[tokio::test]
1807 4 : async fn test_atomic_overwrite_preexisting_tmp() {
1808 4 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
1809 4 : let testdir =
1810 4 : crate::config::PageServerConf::test_repo_dir("test_atomic_overwrite_preexisting_tmp");
1811 4 : std::fs::create_dir_all(&testdir).unwrap();
1812 4 :
1813 4 : let path = testdir.join("myfile");
1814 4 : let tmp_path = testdir.join("myfile.tmp");
1815 4 :
1816 4 : std::fs::write(&tmp_path, "some preexisting junk that should be removed").unwrap();
1817 4 : assert!(tmp_path.exists());
1818 4 :
1819 4 : VirtualFileInner::crashsafe_overwrite(path.clone(), tmp_path.clone(), b"foo".to_vec())
1820 4 : .await
1821 4 : .unwrap();
1822 4 :
1823 4 : let mut file = MaybeVirtualFile::from(VirtualFile::open(&path, &ctx).await.unwrap());
1824 4 : let post = file.read_string(&ctx).await.unwrap();
1825 4 : assert_eq!(post, "foo");
1826 4 : assert!(!tmp_path.exists());
1827 4 : drop(file);
1828 4 : }
1829 : }
|