On Fri, Aug 26, 2016 at 08:45:37PM -0500, Jun He wrote:
> Hi all,
>
> I find that, with '-o discard', XFS sometimes discards more data than it
> writes. This may slow down XFS more on slow-discard SSDs. Also, could
> problematic discard implementation lead to data loss?
>
> Setup to reproduce:
> - Linux 4.5.4
> - Workload: LevelDB
>
> $ mkfs.xfs -K -f -s size=4096 -b size=4096 /dev/sdc1
> $ mount -o discard /dev/sdc1 /mnt/fsonloop
> $ leveldb/db_bench --benchmarks=overwrite --num=3000000
> --db=/mnt/fsonloop/leveldb_data0 --threads=1 --use_existing_db=0
>
> I found that XFS discarded ~2.4GB while only wrote ~1.4GB,
> by blktrace. (This is not due to the mkfs discarding the
> whole disk.)
Yup. Discard covers extents being freed, not ranges being written.
So there's metadata, preallocated (unwritten) regions, speculative
prealloc beyond EOF, etc, all of which won't be counted in "writes"
but will be counted in "discards".
> Some visual
> https://github.com/junhe/xfs-bug-report/blob/master/xfs-double-discard.png
> Some spaces are discarded twice in this case.
>
> Is it a bug?
No.
Cheers,
Dave.
--
Dave Chinner
david@xxxxxxxxxxxxx
|
