On Tue 20-11-12 11:04:28, Dave Chinner wrote:
> On Mon, Nov 19, 2012 at 10:39:13PM +0100, Jan Kara wrote:
> > On Tue 13-11-12 01:36:13, Jan Kara wrote:
> > > When project quota gets exceeded xfs_iomap_write_delay() ends up flushing
> > > inodes because ENOSPC gets returned from xfs_bmapi_delay() instead of
> > > EDQUOT.
> > > This makes handling of writes over project quota rather slow as a simple
> > > test
> > > program shows:
> > > fd = open(argv[1], O_WRONLY | O_CREAT | O_TRUNC, 0644);
> > > for (i = 0; i < 50000; i++)
> > > pwrite(fd, buf, 4096, i*4096);
> > >
> > > Writing 200 MB like this into a directory with 100 MB project quota takes
> > > around 6 minutes while it takes about 2 seconds with this patch applied.
> > > This
> > > actually happens in a real world load when nfs pushes data into a
> > > directory
> > > which is over project quota.
> > >
> > > Fix the problem by replacing XFS_QMOPT_ENOSPC flag with XFS_QMOPT_EPDQUOT.
> > > That makes xfs_trans_reserve_quota_bydquots() return new error EPDQUOT
> > > when
> > > project quota is exceeded. xfs_bmapi_delay() then uses this flag so that
> > > xfs_iomap_write_delay() can distinguish real ENOSPC (requiring flushing)
> > > from exceeded project quota (not requiring flushing).
> > >
> > > As a side effect this patch fixes inconsistency where e.g. xfs_create()
> > > returned EDQUOT even when project quota was exceeded.
> > Ping? Any opinions?
>
> FWIW, it doesn't look like it'll apply to a current XFs tree:
>
> > > @@ -441,8 +442,11 @@ retry:
> > > */
> > > if (nimaps == 0) {
> > > trace_xfs_delalloc_enospc(ip, offset, count);
> > > - if (flushed)
> > > - return XFS_ERROR(error ? error : ENOSPC);
> > > + if (flushed) {
> > > + if (error == 0 || error == EPDQUOT)
> > > + error = ENOSPC;
> > > + return XFS_ERROR(error);
> > > + }
> > >
> > > if (error == ENOSPC) {
> > > xfs_iunlock(ip, XFS_ILOCK_EXCL);
>
> This xfs_iomap_write_delay() looks like this now:
>
> /*
> * If bmapi returned us nothing, we got either ENOSPC or EDQUOT. Retry
> * without EOF preallocation.
> */
> if (nimaps == 0) {
> trace_xfs_delalloc_enospc(ip, offset, count);
> if (prealloc) {
> prealloc = 0;
> error = 0;
> goto retry;
> }
> return XFS_ERROR(error ? error : ENOSPC);
> }
>
> The flushing is now way up in xfs_file_buffered_aio_write(), and the
> implementation of xfs_flush_inodes() has changed as well. Hence it
> may or may not behave differently not....
OK, so I tested latest XFS tree and changes by commit 9aa05000 (changing
xfs_flush_inodes()) indeed improve the performace from those ~6 minutes to
~6 seconds which is good enough I believe. Thanks for the pointer! I was
thinking for a while why sync_inodes_sb() is so much faster than the
original XFS implementation and I believe it's because we don't force the
log on each sync now.
Honza
--
Jan Kara <jack@xxxxxxx>
SUSE Labs, CR
|