Re: Q. drm/i915 shrinker, synchronize_rcu_expedited() from handlers

From: Andrea Arcangeli <aarcange@redhat.com>
To: Hugh Dickins <hughd@google.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>,
	"J. R. Okajima" <hooanon05g@gmail.com>,
	chris@chris-wilson.co.uk, daniel.vetter@ffwll.ch,
	jani.nikula@intel.com, linux-kernel@vger.kernel.org
Subject: Re: Q. drm/i915 shrinker, synchronize_rcu_expedited() from handlers
Date: Wed, 10 May 2017 12:43:54 +0200	[thread overview]
Message-ID: <20170510104354.GB5011@redhat.com> (raw)
In-Reply-To: <alpine.LSU.2.11.1705091952090.3043@eggly.anvils>

Hello,

On Tue, May 09, 2017 at 08:04:24PM -0700, Hugh Dickins wrote:
> On Mon, 8 May 2017, Joonas Lahtinen wrote:
> > On pe, 2017-05-05 at 14:57 -0700, Hugh Dickins wrote:
> > > On Fri, 5 May 2017, Joonas Lahtinen wrote:
> > > > On ma, 2017-05-01 at 11:05 +0900, J. R. Okajima wrote:
> > > > > Thanx for the reply.
> > > > > 
> > > > > Andrea Arcangeli:
> > > > > > 
> > > > > > Yes I already reported this, my original fix was way more efficient
> > > > > > (and also safer considering the above) than what landed upstream. My
> > > > > > feedback was ignored though.
> > > > > > 
> > > > > > https://lists.freedesktop.org/archives/intel-gfx/2017-April/125414.html
> > > > > 
> > > > > I see.
> > > > > Actually on my test system for v4.11-rc8, kthreadd, kworker, kswapd and
> > > > > others all stopped working due to the synchronize_rcu_expedited call
> > > > > from i915_gem_shrinker_count. It is definitly a show stopper for me as
> > > > > an i915 user.
> > > > 
> > > > Filing a bug in freedesktop.org with all the details is the fastest way
> > > > of getting help. Without the bug (and with such little information as
> > > > the previous e-mail) it's hard to estimate the extent and nature of the
> > > > bug.
> > > > 
> > > > I've anyway gone and prepared a patch to drop the RCU sync completely
> > > > from shrinker phase, as discussed originally with Chris.
> > > 
> > > Is that a patch that will be suitable for 4.11-stable?  Please do post
> > > it here.  I had not experienced this i915-induced hang at all when
> > > Andrea first mentioned it, nor even on 4.11-rc8; but now with 4.11
> > > final I can get it fairly easily (I haven't tried Andrea's fix yet).
> > 
> > Please try:
> > 
> > https://patchwork.freedesktop.org/patch/154713/
> > 
> > If it works, a Tested-by: would be appreciated.
> 
> Yes, that works for me, thank you.
> 
> Tested-by: Hugh Dickins <hughd@google.com>
> 
> But the linked patch seems to be lacking a Reported-by (not me) tag,
> a Fixes tag, a Cc stable tag, and any indication in the Subject or
> commit message that this patch is something needed to fix hangs
> observed by several people - it just sounds like a minor cleanup.

It works for me too. I'm running my workstation also with
synchronize_rcu removed from i915_gem_shrink_all in addition to the
above. Isn't the oom method invoked from reclaim context too? As far
as I can tell synchronize_rcu can end up throttling on a background
synchronize_rcu_expedited(), so it might end up in the same issue
unless removed too.

Tested-by: Andrea Arcangeli <aarcange@redhat.com>

(I can't reproduce the lockups 100% of the time, but they never
happened again with this patch and I happened to run the load that
reproduces them a couple of times already with v4.11 and this patch
applied)

It's also certainly improving performance by removing the
synchronize_rcu_expedited from the _count methods where it was useless
(in addition to unsafe).

Thanks,
Andrea