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).
44 229 : pub fn new(mut storage_root: Utf8PathBuf, timeout: Duration) -> anyhow::Result<Self> {
45 229 : 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 : })?;
49 196 : }
50 229 : if !storage_root.is_absolute() {
51 186 : storage_root = storage_root.canonicalize_utf8().with_context(|| {
52 0 : format!("Failed to represent path {storage_root:?} as an absolute path")
53 186 : })?;
54 43 : }
55 :
56 229 : Ok(Self {
57 229 : storage_root,
58 229 : timeout,
59 229 : })
60 229 : }
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 27 : 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 27 : .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`
150 210 : async fn list_recursive(&self, folder: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>> {
151 210 : let full_path = match folder {
152 204 : 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.
160 210 : let mut initial_dir = full_path.clone();
161 210 :
162 210 : // If there's no trailing slash, we have to start looking from one above: even if
163 210 : // `initial_dir` is a directory, we should still list any prefixes in the parent
164 210 : // that start with the same string.
165 210 : if !full_path.to_string().ends_with('/') {
166 21 : initial_dir.pop();
167 189 : }
168 :
169 : loop {
170 : // Did we make it to the root?
171 756 : if initial_dir.parent().is_none() {
172 0 : anyhow::bail!("list_files: failed to find valid ancestor dir for {full_path}");
173 756 : }
174 756 :
175 756 : match fs::metadata(initial_dir.clone()).await {
176 210 : Ok(meta) if meta.is_dir() => {
177 210 : // We found a directory, break
178 210 : 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 : }
184 546 : Err(e) if e.kind() == ErrorKind::NotFound => {
185 546 : // It's not a file that exists: strip the prefix back to the parent directory
186 546 : initial_dir.pop();
187 546 : }
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.
197 210 : let prefix = full_path.as_str();
198 210 :
199 210 : let mut files = vec![];
200 210 : let mut directory_queue = vec![initial_dir];
201 477 : while let Some(cur_folder) = directory_queue.pop() {
202 267 : let mut entries = cur_folder.read_dir_utf8()?;
203 599 : 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 :
216 210 : Ok(files)
217 210 : }
218 :
219 2841 : async fn upload0(
220 2841 : &self,
221 2841 : data: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync,
222 2841 : data_size_bytes: usize,
223 2841 : to: &RemotePath,
224 2841 : metadata: Option<StorageMetadata>,
225 2841 : cancel: &CancellationToken,
226 2841 : ) -> anyhow::Result<()> {
227 2841 : let target_file_path = to.with_base(&self.storage_root);
228 2841 : 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 2839 : let temp_file_path =
242 2839 : path_with_suffix_extension(&target_file_path, LOCAL_FS_TEMP_FILE_SUFFIX);
243 2837 : let mut destination = io::BufWriter::new(
244 2839 : fs::OpenOptions::new()
245 2839 : .write(true)
246 2839 : .create(true)
247 2839 : .truncate(true)
248 2839 : .open(&temp_file_path)
249 2364 : .await
250 2837 : .with_context(|| {
251 0 : format!("Failed to open target fs destination at '{target_file_path}'")
252 2837 : })?,
253 : );
254 :
255 2837 : let from_size_bytes = data_size_bytes as u64;
256 2837 : let data = tokio_util::io::StreamReader::new(data);
257 2837 : let data = std::pin::pin!(data);
258 2837 : let mut buffer_to_read = data.take(from_size_bytes);
259 2837 :
260 2837 : // alternatively we could just write the bytes to a file, but local_fs is a testing utility
261 2837 : let copy = io::copy_buf(&mut buffer_to_read, &mut destination);
262 :
263 2837 : let bytes_read = tokio::select! {
264 : biased;
265 2837 : _ = 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 2837 : read = copy => read,
277 : };
278 :
279 2831 : let bytes_read =
280 2831 : 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 2831 : })?;
285 :
286 2831 : if bytes_read < from_size_bytes {
287 3 : bail!("Provided stream was shorter than expected: {bytes_read} vs {from_size_bytes} bytes");
288 2828 : }
289 2828 : // Check if there is any extra data after the given size.
290 2828 : let mut from = buffer_to_read.into_inner();
291 2828 : let extra_read = from.read(&mut [1]).await?;
292 2827 : ensure!(
293 2827 : extra_read == 0,
294 6 : "Provided stream was larger than expected: expected {from_size_bytes} bytes",
295 : );
296 :
297 2821 : 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 2821 : })?;
302 :
303 2821 : fs::rename(temp_file_path, &target_file_path)
304 2357 : .await
305 2821 : .with_context(|| {
306 0 : format!(
307 0 : "Failed to upload (rename) file to the local storage at '{target_file_path}'",
308 0 : )
309 2821 : })?;
310 :
311 2821 : 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 2818 : }
328 :
329 2821 : Ok(())
330 2833 : }
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 :
345 210 : async fn list(
346 210 : &self,
347 210 : prefix: Option<&RemotePath>,
348 210 : mode: ListingMode,
349 210 : max_keys: Option<NonZeroU32>,
350 210 : cancel: &CancellationToken,
351 210 : ) -> Result<Listing, DownloadError> {
352 210 : let op = async {
353 210 : let mut result = Listing::default();
354 :
355 : // Filter out directories: in S3 directories don't exist, only the keys within them do.
356 210 : let keys = self
357 210 : .list_recursive(prefix)
358 740 : .await
359 210 : .map_err(DownloadError::Other)?;
360 210 : let objects = keys
361 210 : .into_iter()
362 210 : .filter_map(|k| {
363 148 : let path = k.with_base(&self.storage_root);
364 148 : if path.is_dir() {
365 57 : None
366 : } else {
367 91 : Some(ListingObject {
368 91 : key: k.clone(),
369 91 : // LocalFs is just for testing, so just specify a dummy time
370 91 : last_modified: SystemTime::now(),
371 91 : size: 0,
372 91 : })
373 : }
374 210 : })
375 210 : .collect();
376 210 :
377 210 : if let ListingMode::NoDelimiter = mode {
378 9 : result.keys = objects;
379 9 : } else {
380 201 : let mut prefixes = HashSet::new();
381 265 : for object in objects {
382 64 : let key = object.key;
383 : // If the part after the prefix includes a "/", take only the first part and put it in `prefixes`.
384 64 : let relative_key = if let Some(prefix) = prefix {
385 55 : let mut prefix = prefix.clone();
386 55 : // We only strip the dirname of the prefix, so that when we strip it from the start of keys we
387 55 : // end up with full file/dir names.
388 55 : let prefix_full_local_path = prefix.with_base(&self.storage_root);
389 55 : let has_slash = prefix.0.to_string().ends_with('/');
390 55 : let strip_prefix = if prefix_full_local_path.is_dir() && has_slash {
391 31 : prefix
392 : } else {
393 24 : prefix.0.pop();
394 24 : prefix
395 : };
396 :
397 55 : RemotePath::new(key.strip_prefix(&strip_prefix).unwrap()).unwrap()
398 : } else {
399 9 : key
400 : };
401 :
402 64 : let relative_key = format!("{}", relative_key);
403 64 : if relative_key.contains(REMOTE_STORAGE_PREFIX_SEPARATOR) {
404 61 : let first_part = relative_key
405 61 : .split(REMOTE_STORAGE_PREFIX_SEPARATOR)
406 61 : .next()
407 61 : .unwrap()
408 61 : .to_owned();
409 61 : prefixes.insert(first_part);
410 61 : } else {
411 3 : result.keys.push(ListingObject {
412 3 : key: RemotePath::from_string(&relative_key).unwrap(),
413 3 : // LocalFs is just for testing
414 3 : last_modified: SystemTime::now(),
415 3 : size: 0,
416 3 : });
417 3 : }
418 : }
419 201 : result.prefixes = prefixes
420 201 : .into_iter()
421 201 : .map(|s| RemotePath::from_string(&s).unwrap())
422 201 : .collect();
423 201 : }
424 :
425 210 : if let Some(max_keys) = max_keys {
426 0 : result.keys.truncate(max_keys.get() as usize);
427 210 : }
428 210 : Ok(result)
429 210 : };
430 :
431 210 : let timeout = async {
432 593 : tokio::time::sleep(self.timeout).await;
433 0 : Err(DownloadError::Timeout)
434 0 : };
435 :
436 210 : let cancelled = async {
437 656 : cancel.cancelled().await;
438 0 : Err(DownloadError::Cancelled)
439 0 : };
440 :
441 210 : tokio::select! {
442 210 : res = op => res,
443 210 : res = timeout => res,
444 210 : res = cancelled => res,
445 : }
446 210 : }
447 :
448 0 : async fn head_object(
449 0 : &self,
450 0 : key: &RemotePath,
451 0 : _cancel: &CancellationToken,
452 0 : ) -> Result<ListingObject, DownloadError> {
453 0 : let target_file_path = key.with_base(&self.storage_root);
454 0 : let metadata = file_metadata(&target_file_path).await?;
455 : Ok(ListingObject {
456 0 : key: key.clone(),
457 0 : last_modified: metadata.modified()?,
458 0 : size: metadata.len(),
459 : })
460 0 : }
461 :
462 2841 : async fn upload(
463 2841 : &self,
464 2841 : data: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync,
465 2841 : data_size_bytes: usize,
466 2841 : to: &RemotePath,
467 2841 : metadata: Option<StorageMetadata>,
468 2841 : cancel: &CancellationToken,
469 2841 : ) -> anyhow::Result<()> {
470 2841 : let cancel = cancel.child_token();
471 2841 :
472 2841 : let op = self.upload0(data, data_size_bytes, to, metadata, &cancel);
473 2841 : let mut op = std::pin::pin!(op);
474 :
475 : // race the upload0 to the timeout; if it goes over, do a graceful shutdown
476 2841 : let (res, timeout) = tokio::select! {
477 2841 : res = &mut op => (res, false),
478 2841 : _ = tokio::time::sleep(self.timeout) => {
479 0 : cancel.cancel();
480 0 : (op.await, true)
481 : }
482 : };
483 :
484 12 : match res {
485 12 : Err(e) if timeout && TimeoutOrCancel::caused_by_cancel(&e) => {
486 0 : // we caused this cancel (or they happened simultaneously) -- swap it out to
487 0 : // Timeout
488 0 : Err(TimeoutOrCancel::Timeout.into())
489 : }
490 2833 : res => res,
491 : }
492 2833 : }
493 :
494 81 : async fn download(
495 81 : &self,
496 81 : from: &RemotePath,
497 81 : opts: &DownloadOpts,
498 81 : cancel: &CancellationToken,
499 81 : ) -> Result<Download, DownloadError> {
500 81 : let target_path = from.with_base(&self.storage_root);
501 :
502 81 : let file_metadata = file_metadata(&target_path).await?;
503 64 : let etag = mock_etag(&file_metadata);
504 64 :
505 64 : if opts.etag.as_ref() == Some(&etag) {
506 0 : return Err(DownloadError::Unmodified);
507 64 : }
508 :
509 64 : let mut file = fs::OpenOptions::new()
510 64 : .read(true)
511 64 : .open(&target_path)
512 60 : .await
513 64 : .with_context(|| {
514 0 : format!("Failed to open source file {target_path:?} to use in the download")
515 64 : })
516 64 : .map_err(DownloadError::Other)?;
517 :
518 64 : let mut take = file_metadata.len();
519 64 : if let Some((start, end)) = opts.byte_range() {
520 15 : if start > 0 {
521 6 : file.seek(io::SeekFrom::Start(start))
522 6 : .await
523 6 : .context("Failed to seek to the range start in a local storage file")
524 6 : .map_err(DownloadError::Other)?;
525 9 : }
526 15 : if let Some(end) = end {
527 9 : take = end - start;
528 9 : }
529 49 : }
530 :
531 61 : let source = ReaderStream::new(file.take(take));
532 :
533 61 : let metadata = self
534 61 : .read_storage_metadata(&target_path)
535 6 : .await
536 61 : .map_err(DownloadError::Other)?;
537 :
538 61 : let cancel_or_timeout = crate::support::cancel_or_timeout(self.timeout, cancel.clone());
539 61 : let source = crate::support::DownloadStream::new(cancel_or_timeout, source);
540 61 :
541 61 : Ok(Download {
542 61 : metadata,
543 61 : last_modified: file_metadata
544 61 : .modified()
545 61 : .map_err(|e| DownloadError::Other(anyhow::anyhow!(e).context("Reading mtime")))?,
546 61 : etag,
547 61 : download_stream: Box::pin(source),
548 : })
549 78 : }
550 :
551 14 : async fn delete(&self, path: &RemotePath, _cancel: &CancellationToken) -> anyhow::Result<()> {
552 14 : let file_path = path.with_base(&self.storage_root);
553 14 : match fs::remove_file(&file_path).await {
554 11 : Ok(()) => Ok(()),
555 : // The file doesn't exist. This shouldn't yield an error to mirror S3's behaviour.
556 : // See https://docs.aws.amazon.com/AmazonS3/latest/API/API_DeleteObject.html
557 : // > If there isn't a null version, Amazon S3 does not remove any objects but will still respond that the command was successful.
558 3 : Err(e) if e.kind() == ErrorKind::NotFound => Ok(()),
559 0 : Err(e) => Err(anyhow::anyhow!(e)),
560 : }
561 14 : }
562 :
563 6 : async fn delete_objects<'a>(
564 6 : &self,
565 6 : paths: &'a [RemotePath],
566 6 : cancel: &CancellationToken,
567 6 : ) -> anyhow::Result<()> {
568 12 : for path in paths {
569 6 : self.delete(path, cancel).await?
570 : }
571 6 : Ok(())
572 6 : }
573 :
574 0 : async fn copy(
575 0 : &self,
576 0 : from: &RemotePath,
577 0 : to: &RemotePath,
578 0 : _cancel: &CancellationToken,
579 0 : ) -> anyhow::Result<()> {
580 0 : let from_path = from.with_base(&self.storage_root);
581 0 : let to_path = to.with_base(&self.storage_root);
582 0 : create_target_directory(&to_path).await?;
583 0 : fs::copy(&from_path, &to_path).await.with_context(|| {
584 0 : format!(
585 0 : "Failed to copy file from '{from_path}' to '{to_path}'",
586 0 : from_path = from_path,
587 0 : to_path = to_path
588 0 : )
589 0 : })?;
590 0 : Ok(())
591 0 : }
592 :
593 0 : async fn time_travel_recover(
594 0 : &self,
595 0 : _prefix: Option<&RemotePath>,
596 0 : _timestamp: SystemTime,
597 0 : _done_if_after: SystemTime,
598 0 : _cancel: &CancellationToken,
599 0 : ) -> Result<(), TimeTravelError> {
600 0 : Err(TimeTravelError::Unimplemented)
601 0 : }
602 : }
603 :
604 64 : fn storage_metadata_path(original_path: &Utf8Path) -> Utf8PathBuf {
605 64 : path_with_suffix_extension(original_path, "metadata")
606 64 : }
607 :
608 2841 : async fn create_target_directory(target_file_path: &Utf8Path) -> anyhow::Result<()> {
609 2841 : let target_dir = match target_file_path.parent() {
610 2841 : Some(parent_dir) => parent_dir,
611 0 : None => bail!("File path '{target_file_path}' has no parent directory"),
612 : };
613 2841 : if !target_dir.exists() {
614 400 : fs::create_dir_all(target_dir).await?;
615 2441 : }
616 2839 : Ok(())
617 2839 : }
618 :
619 81 : async fn file_metadata(file_path: &Utf8Path) -> Result<std::fs::Metadata, DownloadError> {
620 81 : tokio::fs::metadata(&file_path).await.map_err(|e| {
621 17 : if e.kind() == ErrorKind::NotFound {
622 17 : DownloadError::NotFound
623 : } else {
624 0 : DownloadError::BadInput(e.into())
625 : }
626 81 : })
627 81 : }
628 :
629 : // Use mtime as stand-in for ETag. We could calculate a meaningful one by md5'ing the contents of files we
630 : // read, but that's expensive and the local_fs test helper's whole reason for existence is to run small tests
631 : // quickly, with less overhead than using a mock S3 server.
632 64 : fn mock_etag(meta: &std::fs::Metadata) -> Etag {
633 64 : let mtime = meta.modified().expect("Filesystem mtime missing");
634 64 : format!("{}", mtime.duration_since(UNIX_EPOCH).unwrap().as_millis()).into()
635 64 : }
636 :
637 : #[cfg(test)]
638 : mod fs_tests {
639 : use super::*;
640 :
641 : use camino_tempfile::tempdir;
642 : use std::{collections::HashMap, io::Write, ops::Bound};
643 :
644 9 : async fn read_and_check_metadata(
645 9 : storage: &LocalFs,
646 9 : remote_storage_path: &RemotePath,
647 9 : expected_metadata: Option<&StorageMetadata>,
648 9 : ) -> anyhow::Result<String> {
649 9 : let cancel = CancellationToken::new();
650 9 : let download = storage
651 9 : .download(remote_storage_path, &DownloadOpts::default(), &cancel)
652 21 : .await
653 9 : .map_err(|e| anyhow::anyhow!("Download failed: {e}"))?;
654 9 : ensure!(
655 9 : download.metadata.as_ref() == expected_metadata,
656 0 : "Unexpected metadata returned for the downloaded file"
657 : );
658 :
659 9 : let contents = aggregate(download.download_stream).await?;
660 :
661 9 : String::from_utf8(contents).map_err(anyhow::Error::new)
662 9 : }
663 :
664 : #[tokio::test]
665 3 : async fn upload_file() -> anyhow::Result<()> {
666 3 : let (storage, cancel) = create_storage()?;
667 3 :
668 17 : let target_path_1 = upload_dummy_file(&storage, "upload_1", None, &cancel).await?;
669 3 : assert_eq!(
670 9 : storage.list_all().await?,
671 3 : vec![target_path_1.clone()],
672 3 : "Should list a single file after first upload"
673 3 : );
674 3 :
675 18 : let target_path_2 = upload_dummy_file(&storage, "upload_2", None, &cancel).await?;
676 3 : assert_eq!(
677 9 : list_files_sorted(&storage).await?,
678 3 : vec![target_path_1.clone(), target_path_2.clone()],
679 3 : "Should list a two different files after second upload"
680 3 : );
681 3 :
682 3 : Ok(())
683 3 : }
684 :
685 : #[tokio::test]
686 3 : async fn upload_file_negatives() -> anyhow::Result<()> {
687 3 : let (storage, cancel) = create_storage()?;
688 3 :
689 3 : let id = RemotePath::new(Utf8Path::new("dummy"))?;
690 3 : let content = Bytes::from_static(b"12345");
691 12 : let content = move || futures::stream::once(futures::future::ready(Ok(content.clone())));
692 3 :
693 3 : // Check that you get an error if the size parameter doesn't match the actual
694 3 : // size of the stream.
695 3 : storage
696 3 : .upload(content(), 0, &id, None, &cancel)
697 3 : .await
698 3 : .expect_err("upload with zero size succeeded");
699 3 : storage
700 3 : .upload(content(), 4, &id, None, &cancel)
701 6 : .await
702 3 : .expect_err("upload with too short size succeeded");
703 3 : storage
704 3 : .upload(content(), 6, &id, None, &cancel)
705 6 : .await
706 3 : .expect_err("upload with too large size succeeded");
707 3 :
708 3 : // Correct size is 5, this should succeed.
709 9 : storage.upload(content(), 5, &id, None, &cancel).await?;
710 3 :
711 3 : Ok(())
712 3 : }
713 :
714 33 : fn create_storage() -> anyhow::Result<(LocalFs, CancellationToken)> {
715 33 : let storage_root = tempdir()?.path().to_path_buf();
716 33 : LocalFs::new(storage_root, Duration::from_secs(120)).map(|s| (s, CancellationToken::new()))
717 33 : }
718 :
719 : #[tokio::test]
720 3 : async fn download_file() -> anyhow::Result<()> {
721 3 : let (storage, cancel) = create_storage()?;
722 3 : let upload_name = "upload_1";
723 18 : let upload_target = upload_dummy_file(&storage, upload_name, None, &cancel).await?;
724 3 :
725 9 : let contents = read_and_check_metadata(&storage, &upload_target, None).await?;
726 3 : assert_eq!(
727 3 : dummy_contents(upload_name),
728 3 : contents,
729 3 : "We should upload and download the same contents"
730 3 : );
731 3 :
732 3 : let non_existing_path = RemotePath::new(Utf8Path::new("somewhere/else"))?;
733 3 : match storage.download(&non_existing_path, &DownloadOpts::default(), &cancel).await {
734 3 : Err(DownloadError::NotFound) => {} // Should get NotFound for non existing keys
735 3 : other => panic!("Should get a NotFound error when downloading non-existing storage files, but got: {other:?}"),
736 3 : }
737 3 : Ok(())
738 3 : }
739 :
740 : #[tokio::test]
741 3 : async fn download_file_range_positive() -> anyhow::Result<()> {
742 3 : let (storage, cancel) = create_storage()?;
743 3 : let upload_name = "upload_1";
744 18 : let upload_target = upload_dummy_file(&storage, upload_name, None, &cancel).await?;
745 3 :
746 3 : let full_range_download_contents =
747 9 : read_and_check_metadata(&storage, &upload_target, None).await?;
748 3 : assert_eq!(
749 3 : dummy_contents(upload_name),
750 3 : full_range_download_contents,
751 3 : "Download full range should return the whole upload"
752 3 : );
753 3 :
754 3 : let uploaded_bytes = dummy_contents(upload_name).into_bytes();
755 3 : let (first_part_local, second_part_local) = uploaded_bytes.split_at(3);
756 3 :
757 3 : let first_part_download = storage
758 3 : .download(
759 3 : &upload_target,
760 3 : &DownloadOpts {
761 3 : byte_end: Bound::Excluded(first_part_local.len() as u64),
762 3 : ..Default::default()
763 3 : },
764 3 : &cancel,
765 3 : )
766 6 : .await?;
767 3 : assert!(
768 3 : first_part_download.metadata.is_none(),
769 3 : "No metadata should be returned for no metadata upload"
770 3 : );
771 3 :
772 3 : let first_part_remote = aggregate(first_part_download.download_stream).await?;
773 3 : assert_eq!(
774 3 : first_part_local, first_part_remote,
775 3 : "First part bytes should be returned when requested"
776 3 : );
777 3 :
778 3 : let second_part_download = storage
779 3 : .download(
780 3 : &upload_target,
781 3 : &DownloadOpts {
782 3 : byte_start: Bound::Included(first_part_local.len() as u64),
783 3 : byte_end: Bound::Excluded(
784 3 : (first_part_local.len() + second_part_local.len()) as u64,
785 3 : ),
786 3 : ..Default::default()
787 3 : },
788 3 : &cancel,
789 3 : )
790 9 : .await?;
791 3 : assert!(
792 3 : second_part_download.metadata.is_none(),
793 3 : "No metadata should be returned for no metadata upload"
794 3 : );
795 3 :
796 3 : let second_part_remote = aggregate(second_part_download.download_stream).await?;
797 3 : assert_eq!(
798 3 : second_part_local, second_part_remote,
799 3 : "Second part bytes should be returned when requested"
800 3 : );
801 3 :
802 3 : let suffix_bytes = storage
803 3 : .download(
804 3 : &upload_target,
805 3 : &DownloadOpts {
806 3 : byte_start: Bound::Included(13),
807 3 : ..Default::default()
808 3 : },
809 3 : &cancel,
810 3 : )
811 9 : .await?
812 3 : .download_stream;
813 6 : let suffix_bytes = aggregate(suffix_bytes).await?;
814 3 : let suffix = std::str::from_utf8(&suffix_bytes)?;
815 3 : assert_eq!(upload_name, suffix);
816 3 :
817 3 : let all_bytes = storage
818 3 : .download(&upload_target, &DownloadOpts::default(), &cancel)
819 6 : .await?
820 3 : .download_stream;
821 3 : let all_bytes = aggregate(all_bytes).await?;
822 3 : let all_bytes = std::str::from_utf8(&all_bytes)?;
823 3 : assert_eq!(dummy_contents("upload_1"), all_bytes);
824 3 :
825 3 : Ok(())
826 3 : }
827 :
828 : #[tokio::test]
829 : #[should_panic(expected = "at or before start")]
830 3 : async fn download_file_range_negative() {
831 3 : let (storage, cancel) = create_storage().unwrap();
832 3 : let upload_name = "upload_1";
833 3 : let upload_target = upload_dummy_file(&storage, upload_name, None, &cancel)
834 16 : .await
835 3 : .unwrap();
836 3 :
837 3 : storage
838 3 : .download(
839 3 : &upload_target,
840 3 : &DownloadOpts {
841 3 : byte_start: Bound::Included(10),
842 3 : byte_end: Bound::Excluded(10),
843 3 : ..Default::default()
844 3 : },
845 3 : &cancel,
846 3 : )
847 6 : .await
848 3 : .unwrap();
849 3 : }
850 :
851 : #[tokio::test]
852 3 : async fn delete_file() -> anyhow::Result<()> {
853 3 : let (storage, cancel) = create_storage()?;
854 3 : let upload_name = "upload_1";
855 16 : let upload_target = upload_dummy_file(&storage, upload_name, None, &cancel).await?;
856 3 :
857 3 : storage.delete(&upload_target, &cancel).await?;
858 9 : assert!(storage.list_all().await?.is_empty());
859 3 :
860 3 : storage
861 3 : .delete(&upload_target, &cancel)
862 3 : .await
863 3 : .expect("Should allow deleting non-existing storage files");
864 3 :
865 3 : Ok(())
866 3 : }
867 :
868 : #[tokio::test]
869 3 : async fn file_with_metadata() -> anyhow::Result<()> {
870 3 : let (storage, cancel) = create_storage()?;
871 3 : let upload_name = "upload_1";
872 3 : let metadata = StorageMetadata(HashMap::from([
873 3 : ("one".to_string(), "1".to_string()),
874 3 : ("two".to_string(), "2".to_string()),
875 3 : ]));
876 3 : let upload_target =
877 21 : upload_dummy_file(&storage, upload_name, Some(metadata.clone()), &cancel).await?;
878 3 :
879 3 : let full_range_download_contents =
880 12 : read_and_check_metadata(&storage, &upload_target, Some(&metadata)).await?;
881 3 : assert_eq!(
882 3 : dummy_contents(upload_name),
883 3 : full_range_download_contents,
884 3 : "We should upload and download the same contents"
885 3 : );
886 3 :
887 3 : let uploaded_bytes = dummy_contents(upload_name).into_bytes();
888 3 : let (first_part_local, _) = uploaded_bytes.split_at(3);
889 3 :
890 3 : let partial_download_with_metadata = storage
891 3 : .download(
892 3 : &upload_target,
893 3 : &DownloadOpts {
894 3 : byte_end: Bound::Excluded(first_part_local.len() as u64),
895 3 : ..Default::default()
896 3 : },
897 3 : &cancel,
898 3 : )
899 9 : .await?;
900 3 : let first_part_remote = aggregate(partial_download_with_metadata.download_stream).await?;
901 3 : assert_eq!(
902 3 : first_part_local,
903 3 : first_part_remote.as_slice(),
904 3 : "First part bytes should be returned when requested"
905 3 : );
906 3 :
907 3 : assert_eq!(
908 3 : partial_download_with_metadata.metadata,
909 3 : Some(metadata),
910 3 : "We should get the same metadata back for partial download"
911 3 : );
912 3 :
913 3 : Ok(())
914 3 : }
915 :
916 : #[tokio::test]
917 3 : async fn list() -> anyhow::Result<()> {
918 3 : // No delimiter: should recursively list everything
919 3 : let (storage, cancel) = create_storage()?;
920 18 : let child = upload_dummy_file(&storage, "grandparent/parent/child", None, &cancel).await?;
921 3 : let child_sibling =
922 18 : upload_dummy_file(&storage, "grandparent/parent/child_sibling", None, &cancel).await?;
923 18 : let uncle = upload_dummy_file(&storage, "grandparent/uncle", None, &cancel).await?;
924 3 :
925 3 : let listing = storage
926 3 : .list(None, ListingMode::NoDelimiter, None, &cancel)
927 3 : .await?;
928 3 : assert!(listing.prefixes.is_empty());
929 3 : assert_eq!(
930 3 : listing
931 3 : .keys
932 3 : .into_iter()
933 9 : .map(|o| o.key)
934 3 : .collect::<HashSet<_>>(),
935 3 : HashSet::from([uncle.clone(), child.clone(), child_sibling.clone()])
936 3 : );
937 3 :
938 3 : // Delimiter: should only go one deep
939 3 : let listing = storage
940 3 : .list(None, ListingMode::WithDelimiter, None, &cancel)
941 3 : .await?;
942 3 :
943 3 : assert_eq!(
944 3 : listing.prefixes,
945 3 : [RemotePath::from_string("timelines").unwrap()].to_vec()
946 3 : );
947 3 : assert!(listing.keys.is_empty());
948 3 :
949 3 : // Delimiter & prefix with a trailing slash
950 3 : let listing = storage
951 3 : .list(
952 3 : Some(&RemotePath::from_string("timelines/some_timeline/grandparent/").unwrap()),
953 3 : ListingMode::WithDelimiter,
954 3 : None,
955 3 : &cancel,
956 3 : )
957 3 : .await?;
958 3 : assert_eq!(
959 3 : listing.keys.into_iter().map(|o| o.key).collect::<Vec<_>>(),
960 3 : [RemotePath::from_string("uncle").unwrap()].to_vec()
961 3 : );
962 3 : assert_eq!(
963 3 : listing.prefixes,
964 3 : [RemotePath::from_string("parent").unwrap()].to_vec()
965 3 : );
966 3 :
967 3 : // Delimiter and prefix without a trailing slash
968 3 : let listing = storage
969 3 : .list(
970 3 : Some(&RemotePath::from_string("timelines/some_timeline/grandparent").unwrap()),
971 3 : ListingMode::WithDelimiter,
972 3 : None,
973 3 : &cancel,
974 3 : )
975 3 : .await?;
976 3 : assert_eq!(listing.keys, vec![]);
977 3 : assert_eq!(
978 3 : listing.prefixes,
979 3 : [RemotePath::from_string("grandparent").unwrap()].to_vec()
980 3 : );
981 3 :
982 3 : // Delimiter and prefix that's partway through a path component
983 3 : let listing = storage
984 3 : .list(
985 3 : Some(&RemotePath::from_string("timelines/some_timeline/grandp").unwrap()),
986 3 : ListingMode::WithDelimiter,
987 3 : None,
988 3 : &cancel,
989 3 : )
990 3 : .await?;
991 3 : assert_eq!(listing.keys, vec![]);
992 3 : assert_eq!(
993 3 : listing.prefixes,
994 3 : [RemotePath::from_string("grandparent").unwrap()].to_vec()
995 3 : );
996 3 :
997 3 : Ok(())
998 3 : }
999 :
1000 : #[tokio::test]
1001 3 : async fn list_part_component() -> anyhow::Result<()> {
1002 3 : // No delimiter: should recursively list everything
1003 3 : let (storage, cancel) = create_storage()?;
1004 3 :
1005 3 : // Imitates what happens in a tenant path when we have an unsharded path and a sharded path, and do a listing
1006 3 : // of the unsharded path: although there is a "directory" at the unsharded path, it should be handled as
1007 3 : // a freeform prefix.
1008 3 : let _child_a =
1009 18 : upload_dummy_file(&storage, "grandparent/tenant-01/child", None, &cancel).await?;
1010 3 : let _child_b =
1011 18 : upload_dummy_file(&storage, "grandparent/tenant/child", None, &cancel).await?;
1012 3 :
1013 3 : // Delimiter and prefix that's partway through a path component
1014 3 : let listing = storage
1015 3 : .list(
1016 3 : Some(
1017 3 : &RemotePath::from_string("timelines/some_timeline/grandparent/tenant").unwrap(),
1018 3 : ),
1019 3 : ListingMode::WithDelimiter,
1020 3 : None,
1021 3 : &cancel,
1022 3 : )
1023 3 : .await?;
1024 3 : assert_eq!(listing.keys, vec![]);
1025 3 :
1026 3 : let mut found_prefixes = listing.prefixes.clone();
1027 3 : found_prefixes.sort();
1028 3 : assert_eq!(
1029 3 : found_prefixes,
1030 3 : [
1031 3 : RemotePath::from_string("tenant").unwrap(),
1032 3 : RemotePath::from_string("tenant-01").unwrap(),
1033 3 : ]
1034 3 : .to_vec()
1035 3 : );
1036 3 :
1037 3 : Ok(())
1038 3 : }
1039 :
1040 : #[tokio::test]
1041 3 : async fn overwrite_shorter_file() -> anyhow::Result<()> {
1042 3 : let (storage, cancel) = create_storage()?;
1043 3 :
1044 3 : let path = RemotePath::new("does/not/matter/file".into())?;
1045 3 :
1046 3 : let body = Bytes::from_static(b"long file contents is long");
1047 3 : {
1048 3 : let len = body.len();
1049 3 : let body =
1050 3 : futures::stream::once(futures::future::ready(std::io::Result::Ok(body.clone())));
1051 12 : storage.upload(body, len, &path, None, &cancel).await?;
1052 3 : }
1053 3 :
1054 3 : let read = aggregate(
1055 3 : storage
1056 3 : .download(&path, &DownloadOpts::default(), &cancel)
1057 6 : .await?
1058 3 : .download_stream,
1059 3 : )
1060 3 : .await?;
1061 3 : assert_eq!(body, read);
1062 3 :
1063 3 : let shorter = Bytes::from_static(b"shorter body");
1064 3 : {
1065 3 : let len = shorter.len();
1066 3 : let body =
1067 3 : futures::stream::once(futures::future::ready(std::io::Result::Ok(shorter.clone())));
1068 9 : storage.upload(body, len, &path, None, &cancel).await?;
1069 3 : }
1070 3 :
1071 3 : let read = aggregate(
1072 3 : storage
1073 3 : .download(&path, &DownloadOpts::default(), &cancel)
1074 6 : .await?
1075 3 : .download_stream,
1076 3 : )
1077 3 : .await?;
1078 3 : assert_eq!(shorter, read);
1079 3 : Ok(())
1080 3 : }
1081 :
1082 : #[tokio::test]
1083 3 : async fn cancelled_upload_can_later_be_retried() -> anyhow::Result<()> {
1084 3 : let (storage, cancel) = create_storage()?;
1085 3 :
1086 3 : let path = RemotePath::new("does/not/matter/file".into())?;
1087 3 :
1088 3 : let body = Bytes::from_static(b"long file contents is long");
1089 3 : {
1090 3 : let len = body.len();
1091 3 : let body =
1092 3 : futures::stream::once(futures::future::ready(std::io::Result::Ok(body.clone())));
1093 3 : let cancel = cancel.child_token();
1094 3 : cancel.cancel();
1095 3 : let e = storage
1096 3 : .upload(body, len, &path, None, &cancel)
1097 9 : .await
1098 3 : .unwrap_err();
1099 3 :
1100 3 : assert!(TimeoutOrCancel::caused_by_cancel(&e));
1101 3 : }
1102 3 :
1103 3 : {
1104 3 : let len = body.len();
1105 3 : let body =
1106 3 : futures::stream::once(futures::future::ready(std::io::Result::Ok(body.clone())));
1107 9 : storage.upload(body, len, &path, None, &cancel).await?;
1108 3 : }
1109 3 :
1110 3 : let read = aggregate(
1111 3 : storage
1112 3 : .download(&path, &DownloadOpts::default(), &cancel)
1113 6 : .await?
1114 3 : .download_stream,
1115 3 : )
1116 3 : .await?;
1117 3 : assert_eq!(body, read);
1118 3 :
1119 3 : Ok(())
1120 3 : }
1121 :
1122 36 : async fn upload_dummy_file(
1123 36 : storage: &LocalFs,
1124 36 : name: &str,
1125 36 : metadata: Option<StorageMetadata>,
1126 36 : cancel: &CancellationToken,
1127 36 : ) -> anyhow::Result<RemotePath> {
1128 36 : let from_path = storage
1129 36 : .storage_root
1130 36 : .join("timelines")
1131 36 : .join("some_timeline")
1132 36 : .join(name);
1133 36 : let (file, size) = create_file_for_upload(&from_path, &dummy_contents(name)).await?;
1134 :
1135 36 : let relative_path = from_path
1136 36 : .strip_prefix(&storage.storage_root)
1137 36 : .context("Failed to strip storage root prefix")
1138 36 : .and_then(RemotePath::new)
1139 36 : .with_context(|| {
1140 0 : format!(
1141 0 : "Failed to resolve remote part of path {:?} for base {:?}",
1142 0 : from_path, storage.storage_root
1143 0 : )
1144 36 : })?;
1145 :
1146 36 : let file = tokio_util::io::ReaderStream::new(file);
1147 36 :
1148 36 : storage
1149 36 : .upload(file, size, &relative_path, metadata, cancel)
1150 178 : .await?;
1151 36 : Ok(relative_path)
1152 36 : }
1153 :
1154 36 : async fn create_file_for_upload(
1155 36 : path: &Utf8Path,
1156 36 : contents: &str,
1157 36 : ) -> anyhow::Result<(fs::File, usize)> {
1158 36 : std::fs::create_dir_all(path.parent().unwrap())?;
1159 36 : let mut file_for_writing = std::fs::OpenOptions::new()
1160 36 : .write(true)
1161 36 : .create_new(true)
1162 36 : .open(path)?;
1163 36 : write!(file_for_writing, "{}", contents)?;
1164 36 : drop(file_for_writing);
1165 36 : let file_size = path.metadata()?.len() as usize;
1166 36 : Ok((
1167 36 : fs::OpenOptions::new().read(true).open(&path).await?,
1168 36 : file_size,
1169 : ))
1170 36 : }
1171 :
1172 54 : fn dummy_contents(name: &str) -> String {
1173 54 : format!("contents for {name}")
1174 54 : }
1175 :
1176 3 : async fn list_files_sorted(storage: &LocalFs) -> anyhow::Result<Vec<RemotePath>> {
1177 9 : let mut files = storage.list_all().await?;
1178 3 : files.sort_by(|a, b| a.0.cmp(&b.0));
1179 3 : Ok(files)
1180 3 : }
1181 :
1182 33 : async fn aggregate(
1183 33 : stream: impl Stream<Item = std::io::Result<Bytes>>,
1184 33 : ) -> anyhow::Result<Vec<u8>> {
1185 : use futures::stream::StreamExt;
1186 33 : let mut out = Vec::new();
1187 33 : let mut stream = std::pin::pin!(stream);
1188 66 : while let Some(res) = stream.next().await {
1189 33 : out.extend_from_slice(&res?[..]);
1190 : }
1191 33 : Ok(out)
1192 33 : }
1193 : }
|
