filecache: use os.CreateTemp to create files #2088
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Signed-off-by: Achille Roussel <[email protected]>
Signed-off-by: Achille Roussel <[email protected]>
|
I submitted a second commit after realizing that we could get rid of the in-memory synchronization in the file cache and rely exclusively on the atomicity guarantees offered by the file system. Besides the code simplification, relying on the file system also offers synchronization that works when multiple processes share the same cache directory. |
|
so, the assumption essentially is A) reads the file, validates the header in other words, if this is correct we should be able to reproduce the failure on CI by crafting a cache entry with correct metadata but random or zeroed data in the body In a quick experiment I commented out lines 206-210 in: wazero/internal/engine/compiler/engine_cache.go Lines 201 to 217 in 99c057b i.e. the cache now always returns a right-sized buffer with whatever garbage the trick is similar for wazevo in wazero/internal/engine/wazevo/engine_cache.go Lines 225 to 229 in 330be2b and the resulting failures look compatible to me trace for `compiler`trace for wazevoI think it might be worth hardening the file structure on disk adding a "trailer" after the body; i.e. the file structure would become (more or less): notice that wazevo data comes with a trailing "sourceMap" but this is skipped unless a specific flag in Fn_HEAD the TRAILER might contain either a checksum of BODY or just a repetition of the LEN (found already in EXEC_HEADER). The second is cheaper because it's enough to check LEN against the LEN found in EXEC_HEADER, but it's a bit less safe. (I am not implying this should be part of the PR, just reasoning out loud) |
|
Hardening of the file format sounds great to me, I would add the checksum even if it's more compute, the in-memory cache should amortize that cost quite effectively, something like CRC would probably be good enough? That being said, this PR would still fix the problem as-is since it makes the insertion of files in the cache an atomic operation, so there is never a conditions where opening a file would cause the truncation of an existing one. |
ncruces
approved these changes
Feb 23, 2024
There was a problem hiding this comment.
If we want to add a checksum, Castagnoli has hardware support in hash/crc32.
But this PR should also fix the race, as long as no one corrupts the cache on purpose.
Signed-off-by: Achille Roussel <[email protected]>
evacchi
pushed a commit
to evacchi/wazero
that referenced
this pull request
Feb 28, 2024
Signed-off-by: Achille Roussel <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This might be a fix for #2039
Since #1816, the file cache appended
.tmpto the file path name to avoid leaving incomplete files in the cache when an error occurs. However, this is the definition ofos.Create:File creation does not include the
O_EXCLflag, which means that when the file already exists, it is opened and truncated.This PR changes the file cache to use
os.CreateTempinstead, which guarantees that the file name will be unique by injecting a random string in the file name and opening the file withO_EXCLto prevent racing calls that would accidentally use the same name from both suceeding:In #2039, we observe errors where the memory mapping containing the code was getting corrupted somehow. We first assumed that it was caused by a lifecycle issue on the compiled modules that hold these memory segments. The hypothesis I'm making with this change is that concurrent calls to
os.Createfor the same compiled module race each other; when the second call occurs, it truncates the content that was being written by the first call. After that, the program cannot rely on the content of the file anymore.I don't have a test to verify the hypothesis, but regardless, I believe this change is useful to make because it addresses a potential race condition that will manifest sooner or later.