Line data Source code
1 : //! Local filesystem acting as a remote storage.
2 : //! Multiple API users can use the same "storage" of this kind by using different storage roots.
3 : //!
4 : //! This storage used in tests, but can also be used in cases when a certain persistent
5 : //! volume is mounted to the local FS.
6 :
7 : use std::{
8 : collections::HashSet,
9 : io::ErrorKind,
10 : num::NonZeroU32,
11 : time::{Duration, SystemTime, UNIX_EPOCH},
12 : };
13 :
14 : use anyhow::{bail, ensure, Context};
15 : use bytes::Bytes;
16 : use camino::{Utf8Path, Utf8PathBuf};
17 : use futures::stream::Stream;
18 : use tokio::{
19 : fs,
20 : io::{self, AsyncReadExt, AsyncSeekExt, AsyncWriteExt},
21 : };
22 : use tokio_util::{io::ReaderStream, sync::CancellationToken};
23 : use utils::crashsafe::path_with_suffix_extension;
24 :
25 : use crate::{
26 : Download, DownloadError, DownloadOpts, Listing, ListingMode, ListingObject, RemotePath,
27 : TimeTravelError, TimeoutOrCancel, REMOTE_STORAGE_PREFIX_SEPARATOR,
28 : };
29 :
30 : use super::{RemoteStorage, StorageMetadata};
31 : use crate::Etag;
32 :
33 : const LOCAL_FS_TEMP_FILE_SUFFIX: &str = "___temp";
34 :
35 : #[derive(Debug, Clone)]
36 : pub struct LocalFs {
37 : storage_root: Utf8PathBuf,
38 : timeout: Duration,
39 : }
40 :
41 : impl LocalFs {
42 : /// Attempts to create local FS storage, along with its root directory.
43 : /// Storage root will be created (if does not exist) and transformed into an absolute path (if passed as relative).
44 233 : pub fn new(mut storage_root: Utf8PathBuf, timeout: Duration) -> anyhow::Result<Self> {
45 233 : if !storage_root.exists() {
46 33 : std::fs::create_dir_all(&storage_root).with_context(|| {
47 0 : format!("Failed to create all directories in the given root path {storage_root:?}")
48 33 : })?;
49 200 : }
50 233 : if !storage_root.is_absolute() {
51 190 : storage_root = storage_root.canonicalize_utf8().with_context(|| {
52 0 : format!("Failed to represent path {storage_root:?} as an absolute path")
53 190 : })?;
54 43 : }
55 :
56 233 : Ok(Self {
57 233 : storage_root,
58 233 : timeout,
59 233 : })
60 233 : }
61 :
62 : // mirrors S3Bucket::s3_object_to_relative_path
63 148 : fn local_file_to_relative_path(&self, key: Utf8PathBuf) -> RemotePath {
64 148 : let relative_path = key
65 148 : .strip_prefix(&self.storage_root)
66 148 : .expect("relative path must contain storage_root as prefix");
67 148 : RemotePath(relative_path.into())
68 148 : }
69 :
70 61 : async fn read_storage_metadata(
71 61 : &self,
72 61 : file_path: &Utf8Path,
73 61 : ) -> anyhow::Result<Option<StorageMetadata>> {
74 61 : let metadata_path = storage_metadata_path(file_path);
75 61 : if metadata_path.exists() && metadata_path.is_file() {
76 6 : let metadata_string = fs::read_to_string(&metadata_path).await.with_context(|| {
77 0 : format!("Failed to read metadata from the local storage at '{metadata_path}'")
78 6 : })?;
79 :
80 6 : serde_json::from_str(&metadata_string)
81 6 : .with_context(|| {
82 0 : format!(
83 0 : "Failed to deserialize metadata from the local storage at '{metadata_path}'",
84 0 : )
85 6 : })
86 6 : .map(|metadata| Some(StorageMetadata(metadata)))
87 : } else {
88 55 : Ok(None)
89 : }
90 61 : }
91 :
92 : #[cfg(test)]
93 9 : async fn list_all(&self) -> anyhow::Result<Vec<RemotePath>> {
94 : use std::{future::Future, pin::Pin};
95 27 : fn get_all_files<'a, P>(
96 27 : directory_path: P,
97 27 : ) -> Pin<Box<dyn Future<Output = anyhow::Result<Vec<Utf8PathBuf>>> + Send + Sync + 'a>>
98 27 : where
99 27 : P: AsRef<Utf8Path> + Send + Sync + 'a,
100 27 : {
101 27 : Box::pin(async move {
102 27 : let directory_path = directory_path.as_ref();
103 27 : if directory_path.exists() {
104 27 : if directory_path.is_dir() {
105 27 : let mut paths = Vec::new();
106 27 : let mut dir_contents = fs::read_dir(directory_path).await?;
107 54 : while let Some(dir_entry) = dir_contents.next_entry().await? {
108 27 : let file_type = dir_entry.file_type().await?;
109 27 : let entry_path =
110 27 : Utf8PathBuf::from_path_buf(dir_entry.path()).map_err(|pb| {
111 0 : anyhow::Error::msg(format!(
112 0 : "non-Unicode path: {}",
113 0 : pb.to_string_lossy()
114 0 : ))
115 27 : })?;
116 27 : if file_type.is_symlink() {
117 0 : tracing::debug!("{entry_path:?} is a symlink, skipping")
118 27 : } else if file_type.is_dir() {
119 24 : paths.extend(get_all_files(&entry_path).await?.into_iter())
120 9 : } else {
121 9 : paths.push(entry_path);
122 9 : }
123 : }
124 27 : Ok(paths)
125 : } else {
126 0 : bail!("Path {directory_path:?} is not a directory")
127 : }
128 : } else {
129 0 : Ok(Vec::new())
130 : }
131 27 : })
132 27 : }
133 :
134 9 : Ok(get_all_files(&self.storage_root)
135 24 : .await?
136 9 : .into_iter()
137 9 : .map(|path| {
138 9 : path.strip_prefix(&self.storage_root)
139 9 : .context("Failed to strip storage root prefix")
140 9 : .and_then(RemotePath::new)
141 9 : .expect(
142 9 : "We list files for storage root, hence should be able to remote the prefix",
143 9 : )
144 9 : })
145 9 : .collect())
146 9 : }
147 :
148 : // recursively lists all files in a directory,
149 : // mirroring the `list_files` for `s3_bucket`
150 404 : async fn list_recursive(&self, folder: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>> {
151 404 : let full_path = match folder {
152 398 : Some(folder) => folder.with_base(&self.storage_root),
153 6 : None => self.storage_root.clone(),
154 : };
155 :
156 : // If we were given a directory, we may use it as our starting point.
157 : // Otherwise, we must go up to the first ancestor dir that exists. This is because
158 : // S3 object list prefixes can be arbitrary strings, but when reading
159 : // the local filesystem we need a directory to start calling read_dir on.
160 404 : let mut initial_dir = full_path.clone();
161 404 :
162 404 : // If there's no trailing slash, we have to start looking from one above: even if
163 404 : // `initial_dir` is a directory, we should still list any prefixes in the parent
164 404 : // that start with the same string.
165 404 : if !full_path.to_string().ends_with('/') {
166 211 : initial_dir.pop();
167 211 : }
168 :
169 : loop {
170 : // Did we make it to the root?
171 1334 : if initial_dir.parent().is_none() {
172 0 : anyhow::bail!("list_files: failed to find valid ancestor dir for {full_path}");
173 1334 : }
174 1334 :
175 1334 : match fs::metadata(initial_dir.clone()).await {
176 404 : Ok(meta) if meta.is_dir() => {
177 404 : // We found a directory, break
178 404 : break;
179 : }
180 0 : Ok(_meta) => {
181 0 : // It's not a directory: strip back to the parent
182 0 : initial_dir.pop();
183 0 : }
184 930 : Err(e) if e.kind() == ErrorKind::NotFound => {
185 930 : // It's not a file that exists: strip the prefix back to the parent directory
186 930 : initial_dir.pop();
187 930 : }
188 0 : Err(e) => {
189 0 : // Unexpected I/O error
190 0 : anyhow::bail!(e)
191 : }
192 : }
193 : }
194 : // Note that Utf8PathBuf starts_with only considers full path segments, but
195 : // object prefixes are arbitrary strings, so we need the strings for doing
196 : // starts_with later.
197 404 : let prefix = full_path.as_str();
198 404 :
199 404 : let mut files = vec![];
200 404 : let mut directory_queue = vec![initial_dir];
201 865 : while let Some(cur_folder) = directory_queue.pop() {
202 461 : let mut entries = cur_folder.read_dir_utf8()?;
203 793 : while let Some(Ok(entry)) = entries.next() {
204 332 : let file_name = entry.file_name();
205 332 : let full_file_name = cur_folder.join(file_name);
206 332 : if full_file_name.as_str().starts_with(prefix) {
207 148 : let file_remote_path = self.local_file_to_relative_path(full_file_name.clone());
208 148 : files.push(file_remote_path);
209 148 : if full_file_name.is_dir() {
210 57 : directory_queue.push(full_file_name);
211 91 : }
212 184 : }
213 : }
214 : }
215 :
216 404 : Ok(files)
217 404 : }
218 :
219 2862 : async fn upload0(
220 2862 : &self,
221 2862 : data: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync,
222 2862 : data_size_bytes: usize,
223 2862 : to: &RemotePath,
224 2862 : metadata: Option<StorageMetadata>,
225 2862 : cancel: &CancellationToken,
226 2862 : ) -> anyhow::Result<()> {
227 2862 : let target_file_path = to.with_base(&self.storage_root);
228 2862 : create_target_directory(&target_file_path).await?;
229 : // We need this dance with sort of durable rename (without fsyncs)
230 : // to prevent partial uploads. This was really hit when pageserver shutdown
231 : // cancelled the upload and partial file was left on the fs
232 : // NOTE: Because temp file suffix always the same this operation is racy.
233 : // Two concurrent operations can lead to the following sequence:
234 : // T1: write(temp)
235 : // T2: write(temp) -> overwrites the content
236 : // T1: rename(temp, dst) -> succeeds
237 : // T2: rename(temp, dst) -> fails, temp no longet exists
238 : // This can be solved by supplying unique temp suffix every time, but this situation
239 : // is not normal in the first place, the error can help (and helped at least once)
240 : // to discover bugs in upper level synchronization.
241 2860 : let temp_file_path =
242 2860 : path_with_suffix_extension(&target_file_path, LOCAL_FS_TEMP_FILE_SUFFIX);
243 2853 : let mut destination = io::BufWriter::new(
244 2860 : fs::OpenOptions::new()
245 2860 : .write(true)
246 2860 : .create(true)
247 2860 : .truncate(true)
248 2860 : .open(&temp_file_path)
249 2547 : .await
250 2853 : .with_context(|| {
251 0 : format!("Failed to open target fs destination at '{target_file_path}'")
252 2853 : })?,
253 : );
254 :
255 2853 : let from_size_bytes = data_size_bytes as u64;
256 2853 : let data = tokio_util::io::StreamReader::new(data);
257 2853 : let data = std::pin::pin!(data);
258 2853 : let mut buffer_to_read = data.take(from_size_bytes);
259 2853 :
260 2853 : // alternatively we could just write the bytes to a file, but local_fs is a testing utility
261 2853 : let copy = io::copy_buf(&mut buffer_to_read, &mut destination);
262 :
263 2853 : let bytes_read = tokio::select! {
264 : biased;
265 2853 : _ = cancel.cancelled() => {
266 3 : let file = destination.into_inner();
267 3 : // wait for the inflight operation(s) to complete so that there could be a next
268 3 : // attempt right away and our writes are not directed to their file.
269 3 : file.into_std().await;
270 :
271 : // TODO: leave the temp or not? leaving is probably less racy. enabled truncate at
272 : // least.
273 3 : fs::remove_file(temp_file_path).await.context("remove temp_file_path after cancellation or timeout")?;
274 3 : return Err(TimeoutOrCancel::Cancel.into());
275 : }
276 2853 : read = copy => read,
277 : };
278 :
279 2845 : let bytes_read =
280 2845 : bytes_read.with_context(|| {
281 0 : format!(
282 0 : "Failed to upload file (write temp) to the local storage at '{temp_file_path}'",
283 0 : )
284 2845 : })?;
285 :
286 2845 : if bytes_read < from_size_bytes {
287 3 : bail!("Provided stream was shorter than expected: {bytes_read} vs {from_size_bytes} bytes");
288 2842 : }
289 2842 : // Check if there is any extra data after the given size.
290 2842 : let mut from = buffer_to_read.into_inner();
291 2842 : let extra_read = from.read(&mut [1]).await?;
292 2841 : ensure!(
293 2841 : extra_read == 0,
294 6 : "Provided stream was larger than expected: expected {from_size_bytes} bytes",
295 : );
296 :
297 2835 : destination.flush().await.with_context(|| {
298 0 : format!(
299 0 : "Failed to upload (flush temp) file to the local storage at '{temp_file_path}'",
300 0 : )
301 2835 : })?;
302 :
303 2835 : fs::rename(temp_file_path, &target_file_path)
304 2512 : .await
305 2833 : .with_context(|| {
306 0 : format!(
307 0 : "Failed to upload (rename) file to the local storage at '{target_file_path}'",
308 0 : )
309 2833 : })?;
310 :
311 2833 : if let Some(storage_metadata) = metadata {
312 : // FIXME: we must not be using metadata much, since this would forget the old metadata
313 : // for new writes? or perhaps metadata is sticky; could consider removing if it's never
314 : // used.
315 3 : let storage_metadata_path = storage_metadata_path(&target_file_path);
316 3 : fs::write(
317 3 : &storage_metadata_path,
318 3 : serde_json::to_string(&storage_metadata.0)
319 3 : .context("Failed to serialize storage metadata as json")?,
320 : )
321 3 : .await
322 3 : .with_context(|| {
323 0 : format!(
324 0 : "Failed to write metadata to the local storage at '{storage_metadata_path}'",
325 0 : )
326 3 : })?;
327 2830 : }
328 :
329 2833 : Ok(())
330 2845 : }
331 : }
332 :
333 : impl RemoteStorage for LocalFs {
334 0 : fn list_streaming(
335 0 : &self,
336 0 : prefix: Option<&RemotePath>,
337 0 : mode: ListingMode,
338 0 : max_keys: Option<NonZeroU32>,
339 0 : cancel: &CancellationToken,
340 0 : ) -> impl Stream<Item = Result<Listing, DownloadError>> {
341 0 : let listing = self.list(prefix, mode, max_keys, cancel);
342 0 : futures::stream::once(listing)
343 0 : }
344 :
345 404 : async fn list(
346 404 : &self,
347 404 : prefix: Option<&RemotePath>,
348 404 : mode: ListingMode,
349 404 : max_keys: Option<NonZeroU32>,
350 404 : cancel: &CancellationToken,
351 404 : ) -> Result<Listing, DownloadError> {
352 404 : let op = async {
353 404 : let mut result = Listing::default();
354 :
355 : // Filter out directories: in S3 directories don't exist, only the keys within them do.
356 404 : let keys = self
357 404 : .list_recursive(prefix)
358 1305 : .await
359 404 : .map_err(DownloadError::Other)?;
360 404 : let mut objects = Vec::with_capacity(keys.len());
361 552 : for key in keys {
362 148 : let path = key.with_base(&self.storage_root);
363 148 : let metadata = file_metadata(&path).await?;
364 148 : if metadata.is_dir() {
365 57 : continue;
366 91 : }
367 91 : objects.push(ListingObject {
368 91 : key: key.clone(),
369 91 : last_modified: metadata.modified()?,
370 91 : size: metadata.len(),
371 : });
372 : }
373 404 : let objects = objects;
374 404 :
375 404 : if let ListingMode::NoDelimiter = mode {
376 199 : result.keys = objects;
377 199 : } else {
378 205 : let mut prefixes = HashSet::new();
379 269 : for object in objects {
380 64 : let key = object.key;
381 : // If the part after the prefix includes a "/", take only the first part and put it in `prefixes`.
382 64 : let relative_key = if let Some(prefix) = prefix {
383 55 : let mut prefix = prefix.clone();
384 55 : // We only strip the dirname of the prefix, so that when we strip it from the start of keys we
385 55 : // end up with full file/dir names.
386 55 : let prefix_full_local_path = prefix.with_base(&self.storage_root);
387 55 : let has_slash = prefix.0.to_string().ends_with('/');
388 55 : let strip_prefix = if prefix_full_local_path.is_dir() && has_slash {
389 31 : prefix
390 : } else {
391 24 : prefix.0.pop();
392 24 : prefix
393 : };
394 :
395 55 : RemotePath::new(key.strip_prefix(&strip_prefix).unwrap()).unwrap()
396 : } else {
397 9 : key
398 : };
399 :
400 64 : let relative_key = format!("{}", relative_key);
401 64 : if relative_key.contains(REMOTE_STORAGE_PREFIX_SEPARATOR) {
402 61 : let first_part = relative_key
403 61 : .split(REMOTE_STORAGE_PREFIX_SEPARATOR)
404 61 : .next()
405 61 : .unwrap()
406 61 : .to_owned();
407 61 : prefixes.insert(first_part);
408 61 : } else {
409 3 : result.keys.push(ListingObject {
410 3 : key: RemotePath::from_string(&relative_key).unwrap(),
411 3 : last_modified: object.last_modified,
412 3 : size: object.size,
413 3 : });
414 3 : }
415 : }
416 205 : result.prefixes = prefixes
417 205 : .into_iter()
418 205 : .map(|s| RemotePath::from_string(&s).unwrap())
419 205 : .collect();
420 205 : }
421 :
422 404 : if let Some(max_keys) = max_keys {
423 0 : result.keys.truncate(max_keys.get() as usize);
424 404 : }
425 404 : Ok(result)
426 404 : };
427 :
428 404 : let timeout = async {
429 1151 : tokio::time::sleep(self.timeout).await;
430 0 : Err(DownloadError::Timeout)
431 0 : };
432 :
433 404 : let cancelled = async {
434 1281 : cancel.cancelled().await;
435 0 : Err(DownloadError::Cancelled)
436 0 : };
437 :
438 404 : tokio::select! {
439 404 : res = op => res,
440 404 : res = timeout => res,
441 404 : res = cancelled => res,
442 : }
443 404 : }
444 :
445 0 : async fn head_object(
446 0 : &self,
447 0 : key: &RemotePath,
448 0 : _cancel: &CancellationToken,
449 0 : ) -> Result<ListingObject, DownloadError> {
450 0 : let target_file_path = key.with_base(&self.storage_root);
451 0 : let metadata = file_metadata(&target_file_path).await?;
452 : Ok(ListingObject {
453 0 : key: key.clone(),
454 0 : last_modified: metadata.modified()?,
455 0 : size: metadata.len(),
456 : })
457 0 : }
458 :
459 2862 : async fn upload(
460 2862 : &self,
461 2862 : data: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync,
462 2862 : data_size_bytes: usize,
463 2862 : to: &RemotePath,
464 2862 : metadata: Option<StorageMetadata>,
465 2862 : cancel: &CancellationToken,
466 2862 : ) -> anyhow::Result<()> {
467 2862 : let cancel = cancel.child_token();
468 2862 :
469 2862 : let op = self.upload0(data, data_size_bytes, to, metadata, &cancel);
470 2862 : let mut op = std::pin::pin!(op);
471 :
472 : // race the upload0 to the timeout; if it goes over, do a graceful shutdown
473 2862 : let (res, timeout) = tokio::select! {
474 2862 : res = &mut op => (res, false),
475 2862 : _ = tokio::time::sleep(self.timeout) => {
476 0 : cancel.cancel();
477 0 : (op.await, true)
478 : }
479 : };
480 :
481 12 : match res {
482 12 : Err(e) if timeout && TimeoutOrCancel::caused_by_cancel(&e) => {
483 0 : // we caused this cancel (or they happened simultaneously) -- swap it out to
484 0 : // Timeout
485 0 : Err(TimeoutOrCancel::Timeout.into())
486 : }
487 2845 : res => res,
488 : }
489 2845 : }
490 :
491 651 : async fn download(
492 651 : &self,
493 651 : from: &RemotePath,
494 651 : opts: &DownloadOpts,
495 651 : cancel: &CancellationToken,
496 651 : ) -> Result<Download, DownloadError> {
497 651 : let target_path = from.with_base(&self.storage_root);
498 :
499 651 : let file_metadata = file_metadata(&target_path).await?;
500 64 : let etag = mock_etag(&file_metadata);
501 64 :
502 64 : if opts.etag.as_ref() == Some(&etag) {
503 0 : return Err(DownloadError::Unmodified);
504 64 : }
505 :
506 64 : let mut file = fs::OpenOptions::new()
507 64 : .read(true)
508 64 : .open(&target_path)
509 62 : .await
510 64 : .with_context(|| {
511 0 : format!("Failed to open source file {target_path:?} to use in the download")
512 64 : })
513 64 : .map_err(DownloadError::Other)?;
514 :
515 64 : let mut take = file_metadata.len();
516 64 : if let Some((start, end)) = opts.byte_range() {
517 15 : if start > 0 {
518 6 : file.seek(io::SeekFrom::Start(start))
519 6 : .await
520 6 : .context("Failed to seek to the range start in a local storage file")
521 6 : .map_err(DownloadError::Other)?;
522 9 : }
523 15 : if let Some(end) = end {
524 9 : take = end - start;
525 9 : }
526 49 : }
527 :
528 61 : let source = ReaderStream::new(file.take(take));
529 :
530 61 : let metadata = self
531 61 : .read_storage_metadata(&target_path)
532 6 : .await
533 61 : .map_err(DownloadError::Other)?;
534 :
535 61 : let cancel_or_timeout = crate::support::cancel_or_timeout(self.timeout, cancel.clone());
536 61 : let source = crate::support::DownloadStream::new(cancel_or_timeout, source);
537 61 :
538 61 : Ok(Download {
539 61 : metadata,
540 61 : last_modified: file_metadata
541 61 : .modified()
542 61 : .map_err(|e| DownloadError::Other(anyhow::anyhow!(e).context("Reading mtime")))?,
543 61 : etag,
544 61 : download_stream: Box::pin(source),
545 : })
546 648 : }
547 :
548 14 : async fn delete(&self, path: &RemotePath, _cancel: &CancellationToken) -> anyhow::Result<()> {
549 14 : let file_path = path.with_base(&self.storage_root);
550 14 : match fs::remove_file(&file_path).await {
551 11 : Ok(()) => Ok(()),
552 : // The file doesn't exist. This shouldn't yield an error to mirror S3's behaviour.
553 : // See https://docs.aws.amazon.com/AmazonS3/latest/API/API_DeleteObject.html
554 : // > If there isn't a null version, Amazon S3 does not remove any objects but will still respond that the command was successful.
555 3 : Err(e) if e.kind() == ErrorKind::NotFound => Ok(()),
556 0 : Err(e) => Err(anyhow::anyhow!(e)),
557 : }
558 14 : }
559 :
560 6 : async fn delete_objects<'a>(
561 6 : &self,
562 6 : paths: &'a [RemotePath],
563 6 : cancel: &CancellationToken,
564 6 : ) -> anyhow::Result<()> {
565 12 : for path in paths {
566 6 : self.delete(path, cancel).await?
567 : }
568 6 : Ok(())
569 6 : }
570 :
571 0 : async fn copy(
572 0 : &self,
573 0 : from: &RemotePath,
574 0 : to: &RemotePath,
575 0 : _cancel: &CancellationToken,
576 0 : ) -> anyhow::Result<()> {
577 0 : let from_path = from.with_base(&self.storage_root);
578 0 : let to_path = to.with_base(&self.storage_root);
579 0 : create_target_directory(&to_path).await?;
580 0 : fs::copy(&from_path, &to_path).await.with_context(|| {
581 0 : format!(
582 0 : "Failed to copy file from '{from_path}' to '{to_path}'",
583 0 : from_path = from_path,
584 0 : to_path = to_path
585 0 : )
586 0 : })?;
587 0 : Ok(())
588 0 : }
589 :
590 0 : async fn time_travel_recover(
591 0 : &self,
592 0 : _prefix: Option<&RemotePath>,
593 0 : _timestamp: SystemTime,
594 0 : _done_if_after: SystemTime,
595 0 : _cancel: &CancellationToken,
596 0 : ) -> Result<(), TimeTravelError> {
597 0 : Err(TimeTravelError::Unimplemented)
598 0 : }
599 : }
600 :
601 64 : fn storage_metadata_path(original_path: &Utf8Path) -> Utf8PathBuf {
602 64 : path_with_suffix_extension(original_path, "metadata")
603 64 : }
604 :
605 2862 : async fn create_target_directory(target_file_path: &Utf8Path) -> anyhow::Result<()> {
606 2862 : let target_dir = match target_file_path.parent() {
607 2862 : Some(parent_dir) => parent_dir,
608 0 : None => bail!("File path '{target_file_path}' has no parent directory"),
609 : };
610 2862 : if !target_dir.exists() {
611 404 : fs::create_dir_all(target_dir).await?;
612 2458 : }
613 2860 : Ok(())
614 2860 : }
615 :
616 799 : async fn file_metadata(file_path: &Utf8Path) -> Result<std::fs::Metadata, DownloadError> {
617 799 : tokio::fs::metadata(&file_path).await.map_err(|e| {
618 587 : if e.kind() == ErrorKind::NotFound {
619 587 : DownloadError::NotFound
620 : } else {
621 0 : DownloadError::BadInput(e.into())
622 : }
623 799 : })
624 799 : }
625 :
626 : // Use mtime as stand-in for ETag. We could calculate a meaningful one by md5'ing the contents of files we
627 : // read, but that's expensive and the local_fs test helper's whole reason for existence is to run small tests
628 : // quickly, with less overhead than using a mock S3 server.
629 64 : fn mock_etag(meta: &std::fs::Metadata) -> Etag {
630 64 : let mtime = meta.modified().expect("Filesystem mtime missing");
631 64 : format!("{}", mtime.duration_since(UNIX_EPOCH).unwrap().as_millis()).into()
632 64 : }
633 :
634 : #[cfg(test)]
635 : mod fs_tests {
636 : use super::*;
637 :
638 : use camino_tempfile::tempdir;
639 : use std::{collections::HashMap, io::Write, ops::Bound};
640 :
641 9 : async fn read_and_check_metadata(
642 9 : storage: &LocalFs,
643 9 : remote_storage_path: &RemotePath,
644 9 : expected_metadata: Option<&StorageMetadata>,
645 9 : ) -> anyhow::Result<String> {
646 9 : let cancel = CancellationToken::new();
647 9 : let download = storage
648 9 : .download(remote_storage_path, &DownloadOpts::default(), &cancel)
649 21 : .await
650 9 : .map_err(|e| anyhow::anyhow!("Download failed: {e}"))?;
651 9 : ensure!(
652 9 : download.metadata.as_ref() == expected_metadata,
653 0 : "Unexpected metadata returned for the downloaded file"
654 : );
655 :
656 9 : let contents = aggregate(download.download_stream).await?;
657 :
658 9 : String::from_utf8(contents).map_err(anyhow::Error::new)
659 9 : }
660 :
661 : #[tokio::test]
662 3 : async fn upload_file() -> anyhow::Result<()> {
663 3 : let (storage, cancel) = create_storage()?;
664 3 :
665 16 : let target_path_1 = upload_dummy_file(&storage, "upload_1", None, &cancel).await?;
666 3 : assert_eq!(
667 6 : storage.list_all().await?,
668 3 : vec![target_path_1.clone()],
669 3 : "Should list a single file after first upload"
670 3 : );
671 3 :
672 13 : let target_path_2 = upload_dummy_file(&storage, "upload_2", None, &cancel).await?;
673 3 : assert_eq!(
674 9 : list_files_sorted(&storage).await?,
675 3 : vec![target_path_1.clone(), target_path_2.clone()],
676 3 : "Should list a two different files after second upload"
677 3 : );
678 3 :
679 3 : Ok(())
680 3 : }
681 :
682 : #[tokio::test]
683 3 : async fn upload_file_negatives() -> anyhow::Result<()> {
684 3 : let (storage, cancel) = create_storage()?;
685 3 :
686 3 : let id = RemotePath::new(Utf8Path::new("dummy"))?;
687 3 : let content = Bytes::from_static(b"12345");
688 12 : let content = move || futures::stream::once(futures::future::ready(Ok(content.clone())));
689 3 :
690 3 : // Check that you get an error if the size parameter doesn't match the actual
691 3 : // size of the stream.
692 3 : storage
693 3 : .upload(content(), 0, &id, None, &cancel)
694 3 : .await
695 3 : .expect_err("upload with zero size succeeded");
696 3 : storage
697 3 : .upload(content(), 4, &id, None, &cancel)
698 6 : .await
699 3 : .expect_err("upload with too short size succeeded");
700 3 : storage
701 3 : .upload(content(), 6, &id, None, &cancel)
702 6 : .await
703 3 : .expect_err("upload with too large size succeeded");
704 3 :
705 3 : // Correct size is 5, this should succeed.
706 9 : storage.upload(content(), 5, &id, None, &cancel).await?;
707 3 :
708 3 : Ok(())
709 3 : }
710 :
711 33 : fn create_storage() -> anyhow::Result<(LocalFs, CancellationToken)> {
712 33 : let storage_root = tempdir()?.path().to_path_buf();
713 33 : LocalFs::new(storage_root, Duration::from_secs(120)).map(|s| (s, CancellationToken::new()))
714 33 : }
715 :
716 : #[tokio::test]
717 3 : async fn download_file() -> anyhow::Result<()> {
718 3 : let (storage, cancel) = create_storage()?;
719 3 : let upload_name = "upload_1";
720 17 : let upload_target = upload_dummy_file(&storage, upload_name, None, &cancel).await?;
721 3 :
722 9 : let contents = read_and_check_metadata(&storage, &upload_target, None).await?;
723 3 : assert_eq!(
724 3 : dummy_contents(upload_name),
725 3 : contents,
726 3 : "We should upload and download the same contents"
727 3 : );
728 3 :
729 3 : let non_existing_path = RemotePath::new(Utf8Path::new("somewhere/else"))?;
730 3 : match storage.download(&non_existing_path, &DownloadOpts::default(), &cancel).await {
731 3 : Err(DownloadError::NotFound) => {} // Should get NotFound for non existing keys
732 3 : other => panic!("Should get a NotFound error when downloading non-existing storage files, but got: {other:?}"),
733 3 : }
734 3 : Ok(())
735 3 : }
736 :
737 : #[tokio::test]
738 3 : async fn download_file_range_positive() -> anyhow::Result<()> {
739 3 : let (storage, cancel) = create_storage()?;
740 3 : let upload_name = "upload_1";
741 15 : let upload_target = upload_dummy_file(&storage, upload_name, None, &cancel).await?;
742 3 :
743 3 : let full_range_download_contents =
744 9 : read_and_check_metadata(&storage, &upload_target, None).await?;
745 3 : assert_eq!(
746 3 : dummy_contents(upload_name),
747 3 : full_range_download_contents,
748 3 : "Download full range should return the whole upload"
749 3 : );
750 3 :
751 3 : let uploaded_bytes = dummy_contents(upload_name).into_bytes();
752 3 : let (first_part_local, second_part_local) = uploaded_bytes.split_at(3);
753 3 :
754 3 : let first_part_download = storage
755 3 : .download(
756 3 : &upload_target,
757 3 : &DownloadOpts {
758 3 : byte_end: Bound::Excluded(first_part_local.len() as u64),
759 3 : ..Default::default()
760 3 : },
761 3 : &cancel,
762 3 : )
763 6 : .await?;
764 3 : assert!(
765 3 : first_part_download.metadata.is_none(),
766 3 : "No metadata should be returned for no metadata upload"
767 3 : );
768 3 :
769 3 : let first_part_remote = aggregate(first_part_download.download_stream).await?;
770 3 : assert_eq!(
771 3 : first_part_local, first_part_remote,
772 3 : "First part bytes should be returned when requested"
773 3 : );
774 3 :
775 3 : let second_part_download = storage
776 3 : .download(
777 3 : &upload_target,
778 3 : &DownloadOpts {
779 3 : byte_start: Bound::Included(first_part_local.len() as u64),
780 3 : byte_end: Bound::Excluded(
781 3 : (first_part_local.len() + second_part_local.len()) as u64,
782 3 : ),
783 3 : ..Default::default()
784 3 : },
785 3 : &cancel,
786 3 : )
787 9 : .await?;
788 3 : assert!(
789 3 : second_part_download.metadata.is_none(),
790 3 : "No metadata should be returned for no metadata upload"
791 3 : );
792 3 :
793 3 : let second_part_remote = aggregate(second_part_download.download_stream).await?;
794 3 : assert_eq!(
795 3 : second_part_local, second_part_remote,
796 3 : "Second part bytes should be returned when requested"
797 3 : );
798 3 :
799 3 : let suffix_bytes = storage
800 3 : .download(
801 3 : &upload_target,
802 3 : &DownloadOpts {
803 3 : byte_start: Bound::Included(13),
804 3 : ..Default::default()
805 3 : },
806 3 : &cancel,
807 3 : )
808 9 : .await?
809 3 : .download_stream;
810 6 : let suffix_bytes = aggregate(suffix_bytes).await?;
811 3 : let suffix = std::str::from_utf8(&suffix_bytes)?;
812 3 : assert_eq!(upload_name, suffix);
813 3 :
814 3 : let all_bytes = storage
815 3 : .download(&upload_target, &DownloadOpts::default(), &cancel)
816 6 : .await?
817 3 : .download_stream;
818 3 : let all_bytes = aggregate(all_bytes).await?;
819 3 : let all_bytes = std::str::from_utf8(&all_bytes)?;
820 3 : assert_eq!(dummy_contents("upload_1"), all_bytes);
821 3 :
822 3 : Ok(())
823 3 : }
824 :
825 : #[tokio::test]
826 : #[should_panic(expected = "at or before start")]
827 3 : async fn download_file_range_negative() {
828 3 : let (storage, cancel) = create_storage().unwrap();
829 3 : let upload_name = "upload_1";
830 3 : let upload_target = upload_dummy_file(&storage, upload_name, None, &cancel)
831 17 : .await
832 3 : .unwrap();
833 3 :
834 3 : storage
835 3 : .download(
836 3 : &upload_target,
837 3 : &DownloadOpts {
838 3 : byte_start: Bound::Included(10),
839 3 : byte_end: Bound::Excluded(10),
840 3 : ..Default::default()
841 3 : },
842 3 : &cancel,
843 3 : )
844 6 : .await
845 3 : .unwrap();
846 3 : }
847 :
848 : #[tokio::test]
849 3 : async fn delete_file() -> anyhow::Result<()> {
850 3 : let (storage, cancel) = create_storage()?;
851 3 : let upload_name = "upload_1";
852 17 : let upload_target = upload_dummy_file(&storage, upload_name, None, &cancel).await?;
853 3 :
854 3 : storage.delete(&upload_target, &cancel).await?;
855 9 : assert!(storage.list_all().await?.is_empty());
856 3 :
857 3 : storage
858 3 : .delete(&upload_target, &cancel)
859 3 : .await
860 3 : .expect("Should allow deleting non-existing storage files");
861 3 :
862 3 : Ok(())
863 3 : }
864 :
865 : #[tokio::test]
866 3 : async fn file_with_metadata() -> anyhow::Result<()> {
867 3 : let (storage, cancel) = create_storage()?;
868 3 : let upload_name = "upload_1";
869 3 : let metadata = StorageMetadata(HashMap::from([
870 3 : ("one".to_string(), "1".to_string()),
871 3 : ("two".to_string(), "2".to_string()),
872 3 : ]));
873 3 : let upload_target =
874 20 : upload_dummy_file(&storage, upload_name, Some(metadata.clone()), &cancel).await?;
875 3 :
876 3 : let full_range_download_contents =
877 12 : read_and_check_metadata(&storage, &upload_target, Some(&metadata)).await?;
878 3 : assert_eq!(
879 3 : dummy_contents(upload_name),
880 3 : full_range_download_contents,
881 3 : "We should upload and download the same contents"
882 3 : );
883 3 :
884 3 : let uploaded_bytes = dummy_contents(upload_name).into_bytes();
885 3 : let (first_part_local, _) = uploaded_bytes.split_at(3);
886 3 :
887 3 : let partial_download_with_metadata = storage
888 3 : .download(
889 3 : &upload_target,
890 3 : &DownloadOpts {
891 3 : byte_end: Bound::Excluded(first_part_local.len() as u64),
892 3 : ..Default::default()
893 3 : },
894 3 : &cancel,
895 3 : )
896 9 : .await?;
897 3 : let first_part_remote = aggregate(partial_download_with_metadata.download_stream).await?;
898 3 : assert_eq!(
899 3 : first_part_local,
900 3 : first_part_remote.as_slice(),
901 3 : "First part bytes should be returned when requested"
902 3 : );
903 3 :
904 3 : assert_eq!(
905 3 : partial_download_with_metadata.metadata,
906 3 : Some(metadata),
907 3 : "We should get the same metadata back for partial download"
908 3 : );
909 3 :
910 3 : Ok(())
911 3 : }
912 :
913 : #[tokio::test]
914 3 : async fn list() -> anyhow::Result<()> {
915 3 : // No delimiter: should recursively list everything
916 3 : let (storage, cancel) = create_storage()?;
917 16 : let child = upload_dummy_file(&storage, "grandparent/parent/child", None, &cancel).await?;
918 3 : let child_sibling =
919 13 : upload_dummy_file(&storage, "grandparent/parent/child_sibling", None, &cancel).await?;
920 12 : let uncle = upload_dummy_file(&storage, "grandparent/uncle", None, &cancel).await?;
921 3 :
922 3 : let listing = storage
923 3 : .list(None, ListingMode::NoDelimiter, None, &cancel)
924 18 : .await?;
925 3 : assert!(listing.prefixes.is_empty());
926 3 : assert_eq!(
927 3 : listing
928 3 : .keys
929 3 : .into_iter()
930 9 : .map(|o| o.key)
931 3 : .collect::<HashSet<_>>(),
932 3 : HashSet::from([uncle.clone(), child.clone(), child_sibling.clone()])
933 3 : );
934 3 :
935 3 : // Delimiter: should only go one deep
936 3 : let listing = storage
937 3 : .list(None, ListingMode::WithDelimiter, None, &cancel)
938 19 : .await?;
939 3 :
940 3 : assert_eq!(
941 3 : listing.prefixes,
942 3 : [RemotePath::from_string("timelines").unwrap()].to_vec()
943 3 : );
944 3 : assert!(listing.keys.is_empty());
945 3 :
946 3 : // Delimiter & prefix with a trailing slash
947 3 : let listing = storage
948 3 : .list(
949 3 : Some(&RemotePath::from_string("timelines/some_timeline/grandparent/").unwrap()),
950 3 : ListingMode::WithDelimiter,
951 3 : None,
952 3 : &cancel,
953 3 : )
954 10 : .await?;
955 3 : assert_eq!(
956 3 : listing.keys.into_iter().map(|o| o.key).collect::<Vec<_>>(),
957 3 : [RemotePath::from_string("uncle").unwrap()].to_vec()
958 3 : );
959 3 : assert_eq!(
960 3 : listing.prefixes,
961 3 : [RemotePath::from_string("parent").unwrap()].to_vec()
962 3 : );
963 3 :
964 3 : // Delimiter and prefix without a trailing slash
965 3 : let listing = storage
966 3 : .list(
967 3 : Some(&RemotePath::from_string("timelines/some_timeline/grandparent").unwrap()),
968 3 : ListingMode::WithDelimiter,
969 3 : None,
970 3 : &cancel,
971 3 : )
972 11 : .await?;
973 3 : assert_eq!(listing.keys, vec![]);
974 3 : assert_eq!(
975 3 : listing.prefixes,
976 3 : [RemotePath::from_string("grandparent").unwrap()].to_vec()
977 3 : );
978 3 :
979 3 : // Delimiter and prefix that's partway through a path component
980 3 : let listing = storage
981 3 : .list(
982 3 : Some(&RemotePath::from_string("timelines/some_timeline/grandp").unwrap()),
983 3 : ListingMode::WithDelimiter,
984 3 : None,
985 3 : &cancel,
986 3 : )
987 13 : .await?;
988 3 : assert_eq!(listing.keys, vec![]);
989 3 : assert_eq!(
990 3 : listing.prefixes,
991 3 : [RemotePath::from_string("grandparent").unwrap()].to_vec()
992 3 : );
993 3 :
994 3 : Ok(())
995 3 : }
996 :
997 : #[tokio::test]
998 3 : async fn list_part_component() -> anyhow::Result<()> {
999 3 : // No delimiter: should recursively list everything
1000 3 : let (storage, cancel) = create_storage()?;
1001 3 :
1002 3 : // Imitates what happens in a tenant path when we have an unsharded path and a sharded path, and do a listing
1003 3 : // of the unsharded path: although there is a "directory" at the unsharded path, it should be handled as
1004 3 : // a freeform prefix.
1005 3 : let _child_a =
1006 18 : upload_dummy_file(&storage, "grandparent/tenant-01/child", None, &cancel).await?;
1007 3 : let _child_b =
1008 18 : upload_dummy_file(&storage, "grandparent/tenant/child", None, &cancel).await?;
1009 3 :
1010 3 : // Delimiter and prefix that's partway through a path component
1011 3 : let listing = storage
1012 3 : .list(
1013 3 : Some(
1014 3 : &RemotePath::from_string("timelines/some_timeline/grandparent/tenant").unwrap(),
1015 3 : ),
1016 3 : ListingMode::WithDelimiter,
1017 3 : None,
1018 3 : &cancel,
1019 3 : )
1020 15 : .await?;
1021 3 : assert_eq!(listing.keys, vec![]);
1022 3 :
1023 3 : let mut found_prefixes = listing.prefixes.clone();
1024 3 : found_prefixes.sort();
1025 3 : assert_eq!(
1026 3 : found_prefixes,
1027 3 : [
1028 3 : RemotePath::from_string("tenant").unwrap(),
1029 3 : RemotePath::from_string("tenant-01").unwrap(),
1030 3 : ]
1031 3 : .to_vec()
1032 3 : );
1033 3 :
1034 3 : Ok(())
1035 3 : }
1036 :
1037 : #[tokio::test]
1038 3 : async fn overwrite_shorter_file() -> anyhow::Result<()> {
1039 3 : let (storage, cancel) = create_storage()?;
1040 3 :
1041 3 : let path = RemotePath::new("does/not/matter/file".into())?;
1042 3 :
1043 3 : let body = Bytes::from_static(b"long file contents is long");
1044 3 : {
1045 3 : let len = body.len();
1046 3 : let body =
1047 3 : futures::stream::once(futures::future::ready(std::io::Result::Ok(body.clone())));
1048 11 : storage.upload(body, len, &path, None, &cancel).await?;
1049 3 : }
1050 3 :
1051 3 : let read = aggregate(
1052 3 : storage
1053 3 : .download(&path, &DownloadOpts::default(), &cancel)
1054 6 : .await?
1055 3 : .download_stream,
1056 3 : )
1057 3 : .await?;
1058 3 : assert_eq!(body, read);
1059 3 :
1060 3 : let shorter = Bytes::from_static(b"shorter body");
1061 3 : {
1062 3 : let len = shorter.len();
1063 3 : let body =
1064 3 : futures::stream::once(futures::future::ready(std::io::Result::Ok(shorter.clone())));
1065 9 : storage.upload(body, len, &path, None, &cancel).await?;
1066 3 : }
1067 3 :
1068 3 : let read = aggregate(
1069 3 : storage
1070 3 : .download(&path, &DownloadOpts::default(), &cancel)
1071 6 : .await?
1072 3 : .download_stream,
1073 3 : )
1074 3 : .await?;
1075 3 : assert_eq!(shorter, read);
1076 3 : Ok(())
1077 3 : }
1078 :
1079 : #[tokio::test]
1080 3 : async fn cancelled_upload_can_later_be_retried() -> anyhow::Result<()> {
1081 3 : let (storage, cancel) = create_storage()?;
1082 3 :
1083 3 : let path = RemotePath::new("does/not/matter/file".into())?;
1084 3 :
1085 3 : let body = Bytes::from_static(b"long file contents is long");
1086 3 : {
1087 3 : let len = body.len();
1088 3 : let body =
1089 3 : futures::stream::once(futures::future::ready(std::io::Result::Ok(body.clone())));
1090 3 : let cancel = cancel.child_token();
1091 3 : cancel.cancel();
1092 3 : let e = storage
1093 3 : .upload(body, len, &path, None, &cancel)
1094 9 : .await
1095 3 : .unwrap_err();
1096 3 :
1097 3 : assert!(TimeoutOrCancel::caused_by_cancel(&e));
1098 3 : }
1099 3 :
1100 3 : {
1101 3 : let len = body.len();
1102 3 : let body =
1103 3 : futures::stream::once(futures::future::ready(std::io::Result::Ok(body.clone())));
1104 6 : storage.upload(body, len, &path, None, &cancel).await?;
1105 3 : }
1106 3 :
1107 3 : let read = aggregate(
1108 3 : storage
1109 3 : .download(&path, &DownloadOpts::default(), &cancel)
1110 4 : .await?
1111 3 : .download_stream,
1112 3 : )
1113 3 : .await?;
1114 3 : assert_eq!(body, read);
1115 3 :
1116 3 : Ok(())
1117 3 : }
1118 :
1119 36 : async fn upload_dummy_file(
1120 36 : storage: &LocalFs,
1121 36 : name: &str,
1122 36 : metadata: Option<StorageMetadata>,
1123 36 : cancel: &CancellationToken,
1124 36 : ) -> anyhow::Result<RemotePath> {
1125 36 : let from_path = storage
1126 36 : .storage_root
1127 36 : .join("timelines")
1128 36 : .join("some_timeline")
1129 36 : .join(name);
1130 36 : let (file, size) = create_file_for_upload(&from_path, &dummy_contents(name)).await?;
1131 :
1132 36 : let relative_path = from_path
1133 36 : .strip_prefix(&storage.storage_root)
1134 36 : .context("Failed to strip storage root prefix")
1135 36 : .and_then(RemotePath::new)
1136 36 : .with_context(|| {
1137 0 : format!(
1138 0 : "Failed to resolve remote part of path {:?} for base {:?}",
1139 0 : from_path, storage.storage_root
1140 0 : )
1141 36 : })?;
1142 :
1143 36 : let file = tokio_util::io::ReaderStream::new(file);
1144 36 :
1145 36 : storage
1146 36 : .upload(file, size, &relative_path, metadata, cancel)
1147 160 : .await?;
1148 36 : Ok(relative_path)
1149 36 : }
1150 :
1151 36 : async fn create_file_for_upload(
1152 36 : path: &Utf8Path,
1153 36 : contents: &str,
1154 36 : ) -> anyhow::Result<(fs::File, usize)> {
1155 36 : std::fs::create_dir_all(path.parent().unwrap())?;
1156 36 : let mut file_for_writing = std::fs::OpenOptions::new()
1157 36 : .write(true)
1158 36 : .create_new(true)
1159 36 : .open(path)?;
1160 36 : write!(file_for_writing, "{}", contents)?;
1161 36 : drop(file_for_writing);
1162 36 : let file_size = path.metadata()?.len() as usize;
1163 36 : Ok((
1164 36 : fs::OpenOptions::new().read(true).open(&path).await?,
1165 36 : file_size,
1166 : ))
1167 36 : }
1168 :
1169 54 : fn dummy_contents(name: &str) -> String {
1170 54 : format!("contents for {name}")
1171 54 : }
1172 :
1173 3 : async fn list_files_sorted(storage: &LocalFs) -> anyhow::Result<Vec<RemotePath>> {
1174 9 : let mut files = storage.list_all().await?;
1175 3 : files.sort_by(|a, b| a.0.cmp(&b.0));
1176 3 : Ok(files)
1177 3 : }
1178 :
1179 33 : async fn aggregate(
1180 33 : stream: impl Stream<Item = std::io::Result<Bytes>>,
1181 33 : ) -> anyhow::Result<Vec<u8>> {
1182 : use futures::stream::StreamExt;
1183 33 : let mut out = Vec::new();
1184 33 : let mut stream = std::pin::pin!(stream);
1185 66 : while let Some(res) = stream.next().await {
1186 33 : out.extend_from_slice(&res?[..]);
1187 : }
1188 33 : Ok(out)
1189 33 : }
1190 : }
|
