On Wed 21-11-12 12:38:21, Dave Chinner wrote:
> On Wed, Nov 21, 2012 at 01:24:59AM +0100, Jan Kara wrote:
> > 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.
>
> I think it's detter because we now don't scan every inode in the
> inode cache doing a mapping_tagged(PAGECACHE_TAG_DIRTY) check to see
> if they are dirty or not to decide whether it needs writeback. The
> overhead of doing that adds up very quickly when you have lots of
> cached inodes and you are scanning them for every page that is
> written to....
That may be for other workloads but in my case I had a test case where
I'm pretty sure only a couple of inodes were in the cache (I just mounted a
filesystem and beated one file in one directory on it).
Honza
--
Jan Kara <jack@xxxxxxx>
SUSE Labs, CR
_______________________________________________
xfs mailing list
xfs@xxxxxxxxxxx
http://oss.sgi.com/mailman/listinfo/xfs
