On 24 Mar 2021 at 02:27, Darrick J. Wong wrote:
> On Tue, Mar 23, 2021 at 09:21:27PM +0530, Chandan Babu R wrote:
>> On 22 Mar 2021 at 23:26, Darrick J. Wong wrote:
>> > On Tue, Mar 09, 2021 at 10:31:16AM +0530, Chandan Babu R wrote:
>> >> Verify that XFS does not cause realtime bitmap/summary inode fork's
>> >> extent count to overflow when growing the realtime volume associated
>> >> with a filesystem.
>> >>
>> >> Reviewed-by: Darrick J. Wong <djwong@xxxxxxxxxx>
>> >> Signed-off-by: Chandan Babu R <chandanrlinux@xxxxxxxxx>
>> >
>> > Soo... I discovered that this test doesn't pass with multiblock
>> > directories:
>>
>> Thanks for the bug report and the description of the corresponding solution. I
>> am fixing the tests and will soon post corresponding patches to the mailing
>> list.
>
> Also, I found a problem with xfs/534 when it does the direct write tests
> to a pmem volume with DAX enabled:
>
> --- /tmp/fstests/tests/xfs/534.out 2021-03-21 11:44:09.384407426 -0700
> +++ /var/tmp/fstests/xfs/534.out.bad 2021-03-23 13:32:15.898301839 -0700
> @@ -5,7 +5,4 @@
> Fallocate 15 blocks
> Buffered write to every other block of fallocated space
> Verify $testfile's extent count
> -* Direct write to unwritten extent
> -Fallocate 15 blocks
> -Direct write to every other block of fallocated space
> -Verify $testfile's extent count
> +Extent count overflow check failed: nextents = 11
The inode extent overflow reported above was actually due to the buffered
write operation. But it does occur with direct write operation as well.
I was able to recreate the bug with an emulated pmem device on my qemu guest.
>
> looking at the xfs_bmap output for $testfile shows:
>
> /opt/testfile:
> EXT: FILE-OFFSET BLOCK-RANGE AG AG-OFFSET TOTAL FLAGS
> 0: [0..7]: 208..215 0 (208..215) 8 010000
> 1: [8..15]: 216..223 0 (216..223) 8 000000
> 2: [16..23]: 224..231 0 (224..231) 8 010000
> 3: [24..31]: 232..239 0 (232..239) 8 000000
> 4: [32..39]: 240..247 0 (240..247) 8 010000
> 5: [40..47]: 248..255 0 (248..255) 8 000000
> 6: [48..55]: 256..263 0 (256..263) 8 010000
> 7: [56..63]: 264..271 0 (264..271) 8 000000
> 8: [64..71]: 272..279 0 (272..279) 8 010000
> 9: [72..79]: 280..287 0 (280..287) 8 000000
> 10: [80..119]: 288..327 0 (288..327) 40 010000
>
> Which is ... odd since the same direct write gets cut off after writing
> to block 7 (like you'd expect since it's the same function) when DAX
> isn't enabled...
>
> ...OH, I see the problem. For a non-DAX direct write,
> xfs_iomap_write_direct will allocate an unwritten block into a hole, but
> if the block was already mapped (written or unwritten) it won't do
> anything at all. For that case, XFS_IEXT_ADD_NOSPLIT_CNT is sufficient,
> because in the worst case we add one extent to the data fork.
>
> For DAX writes, however, the behavior is different:
>
> if (IS_DAX(VFS_I(ip))) {
> bmapi_flags = XFS_BMAPI_CONVERT | XFS_BMAPI_ZERO;
> if (imap->br_state == XFS_EXT_UNWRITTEN) {
> force = true;
> dblocks = XFS_DIOSTRAT_SPACE_RES(mp, 0) << 1;
> }
> }
>
> This tells xfs_bmapi_write that we want to /convert/ an unwritten extent
> to written, and we want to zero the blocks. If we're dax-writing into
> the middle of an unwritten range, this will cause a split. The correct
> parameter there would be XFS_IEXT_WRITE_UNWRITTEN_CNT. Would you mind
> sending a kernel patch to fix that?
Sure, I will work on fixing both the buffered and direct IO extent overflow
issues.
Thanks for reporting the bug.
--
chandan
