On Wed, Jun 29, 2011 at 10:01:19AM -0400, Christoph Hellwig wrote:
> The following script from Wu Fengguang shows very bad behaviour in XFS
> when aggressively dirtying data during a sync on XFS, with sync times
> up to almost 10 times as long as ext4.
>
> A large part of the issue is that XFS writes data out itself two times
> in the ->sync_fs method, overriding the lifelock protection in the core
> writeback code, and another issue is the lock-less xfs_ioend_wait call,
> which doesn't prevent new ioend from beeing queue up while waiting for
> the count to reach zero.
>
> This patch removes the XFS-internal sync calls and relies on the VFS
> to do it's work just like all other filesystems do. Note that the
> i_iocount wait which is rather suboptimal is simply removed here.
> We already do it in ->write_inode, which keeps the current supoptimal
> behaviour. We'll eventually need to remove that as well, but that's
> material for a separate commit.
>
> ------------------------------ snip ------------------------------
> #!/bin/sh
>
> umount /dev/sda7
> mkfs.xfs -f /dev/sda7
> # mkfs.ext4 /dev/sda7
> # mkfs.btrfs /dev/sda7
> mount /dev/sda7 /fs
>
> echo $((50<<20)) > /proc/sys/vm/dirty_bytes
>
> pid=
> for i in `seq 10`
> do
> dd if=/dev/zero of=/fs/zero-$i bs=1M count=1000 &
> pid="$pid $!"
> done
>
> sleep 1
>
> tic=$(date +'%s')
> sync
> tac=$(date +'%s')
>
> echo
> echo sync time: $((tac-tic))
> egrep '(Dirty|Writeback|NFS_Unstable)' /proc/meminfo
>
> pidof dd > /dev/null && { kill -9 $pid; echo sync NOT livelocked; }
> ------------------------------ snip ------------------------------
>
> Reported-by: Wu Fengguang <fengguang.wu@xxxxxxxxx>
> Signed-off-by: Christoph Hellwig <hch@xxxxxx>
Reviewed-by: Dave Chinner <dchinner@xxxxxxxxxx>
--
Dave Chinner
david@xxxxxxxxxxxxx
|