[patch] Re: stalled 'sync' on ext3+quota over drbd

[Jan Kara]

  Hi,

> On Mon, 2004-04-19 at 19:27, Eugene Crosser wrote:
> 
> > > From the first impression, it did not help.
> > 
> > Luckily, I was wrong.  At least on the test environment I get the
> > results below.  This is 2.6.5 kernel with euivalent of Stephen's patch. 
> > 'setquota' sets user and group quota for 40,000 userids and 40,000
> > groupids.  'mktree' writes one byte into 40,000 files and sets their
> > owner and group to 40,000 unique values.
> > 
> > 			unpatched		patched
> > sync			7m50			0m0.8
> 
> OK, that's a win. :-)
> 
> > On my 'big' system sync still runs unexpectedly long (40sec - 4min) but
> > it is by far better than it was before the patch...
> 
> That's quote possibly just the raw cost of writing out a million
> dquots.  The overhead of the list traversal should be under control now,
> though.  A profile would help determine what the remaining cost is: is
> that time spent in the CPU, or in IO wait?
> 
> Jan, mind if I push the patch to 2.4?  Your locking concern seems to be
> 2.6-only; on 2.4, BKL should be sufficient protection even on SMP.
  For 2.4 I think your patch is the best solution. Please push it. For
2.6 I've written a patch which implements per-sb lists of dirty dquots
which should also fix the problem and I think it's nicer. My patch needs
a bit more testing so I'll submit it later this week. Thanks for help.

								Honza