Re: [PATCH] generic/084: check inotify limit before tail many files

[Dave Chinner <david@fromorbit.com>]

On Sun, Aug 16, 2015 at 09:05:19PM -0400, Zirong Lang wrote:
> Hi Dave,
> 
> Thanks for your reply.
> 
> ----- 原始邮件 -----
> > 发件人: "Dave Chinner" <david@fromorbit.com>
> > 收件人: "Zorro Lang" <zlang@redhat.com>
> > 抄送: fstests@vger.kernel.org, eguan@redhat.com
> > 发送时间: 星期一, 2015年 8 月 17日 上午 8:03:36
> > 主题: Re: [PATCH] generic/084: check inotify limit before tail many files
> > 
> > On Fri, Aug 14, 2015 at 12:16:32AM +0800, Zorro Lang wrote:
> > > generic/084 try to run 'tail' command, tail will use
> > > inotify, and there're some limit about inotify. I think
> > > the most important is fs.inotify.max_user_instances, then
> > > fs.inotify.max_user_watches is importand too.
> > >
> > > When I test on a machine with 154 cpu cores, this case
> > > run failed, and hit many warning likes:
> > > 
> > >     tail: inotify cannot be used, reverting to polling: Too many
> > >     open files
> > > 
> > > Because the fs.inotify.max_user_instances is 128, so if
> > > we try to tail 154 files, it will be failed.
> > 
> > We use 'tail' all over the place in xfstests, so why is only
> > generic/084 affected?
> 
> Because generic/084 use try to create $nr_cpu tail processes:
> for i in `seq 1 $nr_cpu`; do
>     ...
>     tail -f $testfile &
>     ...
> done
> 
> And nr_cpu=`$here/src/feature -o`.
> 
> Generally fs.inotify.max_user_instances is 128, when a machine
> have more than(or nearly the same) this number, this test will
> failed.

This information should have been in the patch description - it's
concurrently run tail commands that are the problem, not a single
execution of tail...

> Maybe other cases don't try to create so many tail processes, so
> they passed.

Exactly. The answer is obvious when you explain it fully :)

So, we have other tests that use hundreds of open unlinked files and
they don't have this problem. That means the issue is how
generic/084 is creating the unlinked files, not an issue with
inotify config.

Go look at src/multi_open_unlink.c and tests/xfs/1[28]1, and then
rewrite generic/084 to use multi_open_unlink to create and hold 
open unlinked files for a specified amount of time.

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com