On Thu, 10 Jul 2014 15:43:40 +0800 Ian Kent <raven@xxxxxxxxxx> wrote:
> On Thu, 2014-07-10 at 09:41 +1000, NeilBrown wrote:
> > autofs4 currently doesn't support RCU-walk - it immediately
> > aborts any attempt at RCU-walk to force REF-walk for path name
> > lookup.
> >
> > This can cause a significant performance impact on multi-core
> > systems.
> > I have a client with a test case which spends >80% of its time
> > waiting for spinlocks with a "make -j 40" on a 40 core system.
>
> Right, sounds worth the effort.
>
> >
> > This patchset aims to remove most of these spinlocks. To be fully
> > effective in the particular case it needs a second patch set which
> > makes NFS RCU-walk friendly, but one thing at a time.
> >
> > This has only been lightly tested so far so I'm really after feed-back
> > rather than to have the patch set accepted, though the first two
> > patches are trivial and could be taken immediately.
>
> I've only scanned the patches so far, I'll need to spend a bit more time
> on them before I can comment.
>
> I'm going to be pressed for time for at least several days so I won't be
> able to get to this right away.
>
> I expect the submount_test I use to stress path walking and expire to
> mount transitions will likely be a good test to use. I haven't used it
> in my personal environment for quite a while now so I'll need to have a
> look around and see if I can still find a suitable set of scripts.
> Otherwise I'll need to decouple it from the RedHat automated test
> environment.
>
> >
> > The last two patches are the most interesting so review comments on
> > those are particularly welcome.
>
> Again I haven't looked closely at these but don't you mean the last
> three patches or am I just fussing over an obviously straight forward
> patch 3?
Exactly right - that thirds last patch was "obviously straight forward", so
is naturally the one that I have already found a bug in (the patch assumes
that autofs4_check_leaves returns a different dentry, which clearly isn't
true).
I'll repost it, probably on Monday.
>
> Thanks for your effort Bruce,
> Ian
>
http://www.youtube.com/watch?v=_f_p0CgPeyA
(Usually when people get my name wrong they call me "Ian", so you calling me
Bruce is both slightly ironic and quite refreshing!)
Thanks,
NeilBrown
Attachment:
signature.asc
Description: PGP signature
