On Wed, Aug 03, 2011 at 10:49:03PM +0200, Jan Kara wrote:
> at one of customer's machines, I've spotted an issue that sync(1) called
> after writing a single huge file has been achieving rather low throughput.
> debugging this with blktrace, I've found that the culprit was in flusher
> racing with page writeout happening from XFS sync code. The patches below
> that case. Although they are not a complete solution, I belive they are useful
> anyway so please consider merging them...
We currently have three calls to xfs_flush_pages with XBF_ASYNC set:
The first one actually is a synchronous writeout, just implemented in
a rather odd way by doing the xfs_ioend_wait right after it, so your
change is actively harmful for it. The second is only called from
xfs_flush_worker, which is the workqueue offload when we hit ENOSPC.
I can see how this might race with the writeback code, but the correct
fix is to replace it with a call to writeback_inodes_sb(_if_idle)
on that one is fixed to do a trylock on s_umount and thus won't
deadlock. The third one is opportunistic writeout if a file got
truncated down on final release. filemap_flush probably is fine
here, but there's no need for a range version. If you replace it
with filemap_flush please also kill the useless wrapper while you're