commit 968f19c5b1b7d5595423b0ac0020cc18dfed8cb5 upstream.
[BUG] It is a long known bug that VM image on btrfs can lead to data csum mismatch, if the qemu is using direct-io for the image (this is commonly known as cache mode 'none').
[CAUSE] Inside the VM, if the fs is EXT4 or XFS, or even NTFS from Windows, the fs is allowed to dirty/modify the folio even if the folio is under writeback (as long as the address space doesn't have AS_STABLE_WRITES flag inherited from the block device).
This is a valid optimization to improve the concurrency, and since these filesystems have no extra checksum on data, the content change is not a problem at all.
But the final write into the image file is handled by btrfs, which needs the content not to be modified during writeback, or the checksum will not match the data (checksum is calculated before submitting the bio).
So EXT4/XFS/NTRFS assume they can modify the folio under writeback, but btrfs requires no modification, this leads to the false csum mismatch.
This is only a controlled example, there are even cases where multi-thread programs can submit a direct IO write, then another thread modifies the direct IO buffer for whatever reason.
For such cases, btrfs has no sane way to detect such cases and leads to false data csum mismatch.
[FIX] I have considered the following ideas to solve the problem:
- Make direct IO to always skip data checksum This not only requires a new incompatible flag, as it breaks the current per-inode NODATASUM flag. But also requires extra handling for no csum found cases.
And this also reduces our checksum protection.
- Let hardware handle all the checksum AKA, just nodatasum mount option. That requires trust for hardware (which is not that trustful in a lot of cases), and it's not generic at all.
- Always fallback to buffered write if the inode requires checksum This was suggested by Christoph, and is the solution utilized by this patch.
The cost is obvious, the extra buffer copying into page cache, thus it reduces the performance. But at least it's still user configurable, if the end user still wants the zero-copy performance, just set NODATASUM flag for the inode (which is a common practice for VM images on btrfs).
Since we cannot trust user space programs to keep the buffer consistent during direct IO, we have no choice but always falling back to buffered IO. At least by this, we avoid the more deadly false data checksum mismatch error.
CC: stable@vger.kernel.org # 6.6 Suggested-by: Christoph Hellwig hch@infradead.org Reviewed-by: Filipe Manana fdmanana@suse.com Signed-off-by: Qu Wenruo wqu@suse.com Reviewed-by: David Sterba dsterba@suse.com Signed-off-by: David Sterba dsterba@suse.com [ Fix a conflict due to the movement of the function. ] --- Changelog: v2: - Remove the incorrectly included direct-io.c "git am" automatically included the not-yet-exist file into the diff. --- fs/btrfs/file.c | 17 +++++++++++++++++ 1 file changed, 17 insertions(+)
diff --git a/fs/btrfs/file.c b/fs/btrfs/file.c index e794606e7c78..f1456c745c6d 100644 --- a/fs/btrfs/file.c +++ b/fs/btrfs/file.c @@ -1515,6 +1515,23 @@ static ssize_t btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from) goto buffered; }
+ /* + * We can't control the folios being passed in, applications can write + * to them while a direct IO write is in progress. This means the + * content might change after we calculated the data checksum. + * Therefore we can end up storing a checksum that doesn't match the + * persisted data. + * + * To be extra safe and avoid false data checksum mismatch, if the + * inode requires data checksum, just fallback to buffered IO. + * For buffered IO we have full control of page cache and can ensure + * no one is modifying the content during writeback. + */ + if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) { + btrfs_inode_unlock(BTRFS_I(inode), ilock_flags); + goto buffered; + } + /* * The iov_iter can be mapped to the same file range we are writing to. * If that's the case, then we will deadlock in the iomap code, because