fuse: fix fuseblk i_blkbits for iomap partial writes
Subsystems (2)
PATCH 1/1 fuse: fix fuseblk i_blkbits for iomap partial writes
On regular fuse filesystems, i_blkbits is set to PAGE_SHIFT which means any iomap partial writes will mark the entire folio as uptodate. However fuseblk filesystems work differently and allow the blocksize to be less than the page size. As such, this may lead to data corruption if fuseblk sets its blocksize to less than the page size, uses the writeback cache, and does a partial write, then a read and the read happens before the write has undergone writeback, since the folio will not be marked uptodate from the partial write so the read will read in the entire folio from disk, which will overwrite the partial write.
The long-term solution for this, which will also be needed for fuse to enable large folios with the writeback cache on, is to have fuse also use iomap for folio reads, but until that is done, the cleanest workaround is to use the page size for fuseblk's internal kernel blksize/blkbits values while maintaining current behavior for stat().
This was verified using ntfs-3g:
$ sudo mkfs.ntfs -f -c 512 /dev/vdd1 $ sudo ntfs-3g /dev/vdd1 ~/fuseblk $ stat ~/fuseblk/hi.txt IO Block: 512
--- fs/fuse/dir.c | 2 +- fs/fuse/fuse_i.h | 3 +++ fs/fuse/inode.c | 9 +++++++++ 3 files changed, 13 insertions(+), 1 deletion(-)
fs/fuse/dir.c +1 -1
@@ -1199,7 +1199,7 @@ static void fuse_fillattr(struct mnt_idmap *idmap, struct inode *inode,if (attr->blksize !=0 ) blkbits =ilog2 (attr->blksize);else - blkbits = inode->i_sb->s_blocksize_bits; + blkbits = fc->inode_blkbits; stat->blksize =1 << blkbits; }
fs/fuse/fuse_i.h +3 -0
@@ -969,6 +969,9 @@ struct fuse_conn {/* Request timeout (in jiffies). 0 = no timeout */ unsigned int req_timeout; } timeout; + +/** This is a workaround until fuse uses iomap for reads */ +unsigned inode_blkbits; };/*
fs/fuse/inode.c +9 -0
@@ -1805,10 +1805,19 @@ int fuse_fill_super_common(struct super_block *sb, struct fuse_fs_context *ctx) err = -EINVAL ;if (!sb_set_blocksize (sb, ctx->blksize)) goto err; +/* +* This is a workaround until fuse hooks into iomap for reads. +* Else if ctx->blksize < PAGE_SIZE and the writeback cache is +* enabled, a read may overwrite partially written data. +*/ + fc->inode_blkbits = sb->s_blocksize_bits; +if (!sb_set_blocksize (sb,PAGE_SIZE )) + goto err;#endif }else { sb->s_blocksize =PAGE_SIZE ; sb->s_blocksize_bits =PAGE_SHIFT ; + fc->inode_blkbits = sb->s_blocksize_bits; } sb->s_subtype = ctx->subtype;
Signature · 1 line
2.47.3
Quoted 17 lines · Joanne Koong
On regular fuse filesystems,
i_blkbitsis set toPAGE_SHIFTwhich means any iomap partial writes will mark the entire folio as uptodate. However fuseblk filesystems work differently and allow the blocksize to be less than the page size. As such, this may lead to data corruption if fuseblk sets its blocksize to less than the page size, uses the writeback cache, and does a partial write, then a read and the read happens before the write has undergone writeback, since the folio will not be marked uptodate from the partial write so the read will read in the entire folio from disk, which will overwrite the partial write.The long-term solution for this, which will also be needed for fuse to enable large folios with the writeback cache on, is to have fuse also use iomap for folio reads, but until that is done, the cleanest workaround is to use the page size for fuseblk's internal kernel blksize/blkbits values while maintaining current behavior for
stat().
Quoted 8 lines
This was verified using ntfs-3g:
$ sudo mkfs.ntfs -f -c 512 /dev/vdd1 $ sudo ntfs-3g /dev/vdd1 ~/fuseblk $ stat ~/fuseblk/hi.txt IO Block: 512Signed-off-by Joanne Koong <[email protected]>Fixes a4c9ab1d4975 ("fuse: use iomap for buffered writes")
Please ignore this version of the patch. This is superseded by the newer version here:
https://lore.kernel.org/linux-fsdevel/[email protected]/T/#t
Thanks,
Joanne