This is a story about 4 distinct (and very old) btrfs bugs.
Commit c8b978188c ("Btrfs: Add zlib compression support") added
three data corruption bugs for inline extents (bugs #1-3).
Commit 93c82d5750 ("Btrfs: zero page past end of inline file items")
fixed bug #1: uncompressed inline extents followed by a hole and more
extents could get non-zero data in the hole as they were read. The fix
was to add a memset in btrfs_get_extent to zero out the hole.
Commit 166ae5a418 ("btrfs: fix inline compressed read err corruption")
fixed bug #2: compressed inline extents which contained non-zero bytes
might be replaced with zero bytes in some cases. This patch removed an
unhelpful memset from uncompress_inline, but the case where memset is
required was missed.
There is also a memset in the decompression code, but this only covers
decompressed data that is shorter than the ram_bytes from the extent
ref record. This memset doesn't cover the region between the end of the
decompressed data and the end of the page. It has also moved around a
few times over the years, so there's no single patch to refer to.
This patch fixes bug #3: compressed inline extents followed by a hole
and more extents could get non-zero data in the hole as they were read
(i.e. bug #3 is the same as bug #1, but s/uncompressed/compressed/).
The fix is the same: zero out the hole in the compressed case too,
by putting a memset back in uncompress_inline, but this time with
correct parameters.
The last and oldest bug, bug #0, is the cause of the offending inline
extent/hole/extent pattern. Bug #0 is a subtle and mostly-harmless quirk
of behavior somewhere in the btrfs write code. In a few special cases,
an inline extent and hole are allowed to persist where they normally
would be combined with later extents in the file.
A fast reproducer for bug #0 is presented below. A few offending extents
are also created in the wild during large rsync transfers with the -S
flag. A Linux kernel build (git checkout; make allyesconfig; make -j8)
will produce a handful of offending files as well. Once an offending
file is created, it can present different content to userspace each
time it is read.
Bug #0 is at least 4 and possibly 8 years old. I verified every vX.Y
kernel back to v3.5 has this behavior. There are fossil records of this
bug's effects in commits all the way back to v2.6.32. I have no reason
to believe bug #0 wasn't present at the beginning of btrfs compression
support in v2.6.29, but I can't easily test kernels that old to be sure.
It is not clear whether bug #0 is worth fixing. A fix would likely
require injecting extra reads into currently write-only paths, and most
of the exceptional cases caused by bug #0 are already handled now.
Whether we like them or not, bug #0's inline extents followed by holes
are part of the btrfs de-facto disk format now, and we need to be able
to read them without data corruption or an infoleak. So enough about
bug #0, let's get back to bug #3 (this patch).
An example of on-disk structure leading to data corruption found in
the wild:
item 61 key (606890 INODE_ITEM 0) itemoff 9662 itemsize 160
inode generation 50 transid 50 size 47424 nbytes 49141
block group 0 mode 100644 links 1 uid 0 gid 0
rdev 0 flags 0x0(none)
item 62 key (606890 INODE_REF 603050) itemoff 9642 itemsize 20
inode ref index 3 namelen 10 name: DB_File.so
item 63 key (606890 EXTENT_DATA 0) itemoff 8280 itemsize 1362
inline extent data size 1341 ram 4085 compress(zlib)
item 64 key (606890 EXTENT_DATA 4096) itemoff 8227 itemsize 53
extent data disk byte 5367308288 nr 20480
extent data offset 0 nr 45056 ram 45056
extent compression(zlib)
Different data appears in userspace during each read of the 11 bytes
between 4085 and 4096. The extent in item 63 is not long enough to
fill the first page of the file, so a memset is required to fill the
space between item 63 (ending at 4085) and item 64 (beginning at 4096)
with zero.
Here is a reproducer from Liu Bo, which demonstrates another method
of creating the same inline extent and hole pattern:
Using 'page_poison=on' kernel command line (or enable
CONFIG_PAGE_POISONING) run the following:
# touch foo
# chattr +c foo
# xfs_io -f -c "pwrite -W 0 1000" foo
# xfs_io -f -c "falloc 4 8188" foo
# od -x foo
# echo 3 >/proc/sys/vm/drop_caches
# od -x foo
This produce the following on my box:
Correct output: file contains 1000 data bytes followed
by zeros:
0000000 cdcd cdcd cdcd cdcd cdcd cdcd cdcd cdcd
*
0001740 cdcd cdcd cdcd cdcd 0000 0000 0000 0000
0001760 0000 0000 0000 0000 0000 0000 0000 0000
*
0020000
Actual output: the data after the first 1000 bytes
will be different each run:
0000000 cdcd cdcd cdcd cdcd cdcd cdcd cdcd cdcd
*
0001740 cdcd cdcd cdcd cdcd 6c63 7400 635f 006d
0001760 5f74 6f43 7400 435f 0053 5f74 7363 7400
0002000 435f 0056 5f74 6164 7400 645f 0062 5f74
(...)
Signed-off-by: Zygo Blaxell <ce3g8jdj@umail.furryterror.org>
Reviewed-by: Liu Bo <bo.li.liu@oracle.com>
Reviewed-by: Chris Mason <clm@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>