Re: [PATCH 08/35] drm: Protect dev->filelist with its own mutex

From: Daniel Vetter <daniel@ffwll.ch>
To: Chris Wilson <chris@chris-wilson.co.uk>,
	Daniel Vetter <daniel@ffwll.ch>,
	Alex Deucher <alexdeucher@gmail.com>,
	Daniel Vetter <daniel.vetter@ffwll.ch>,
	DRI Development <dri-devel@lists.freedesktop.org>,
	Alex Deucher <alexander.deucher@amd.com>,
	Intel Graphics Development <intel-gfx@lists.freedesktop.org>,
	Daniel Vetter <daniel.vetter@intel.com>
Subject: Re: [PATCH 08/35] drm: Protect dev->filelist with its own mutex
Date: Wed, 27 Apr 2016 10:01:16 +0200	[thread overview]
Message-ID: <20160427080116.GW2558@phenom.ffwll.local> (raw)
In-Reply-To: <20160427071708.GV27856@nuc-i3427.alporthouse.com>

On Wed, Apr 27, 2016 at 08:17:08AM +0100, Chris Wilson wrote:
> On Wed, Apr 27, 2016 at 09:06:09AM +0200, Daniel Vetter wrote:
> > On Tue, Apr 26, 2016 at 05:45:44PM -0400, Alex Deucher wrote:
> > > On Tue, Apr 26, 2016 at 4:52 PM, Chris Wilson <chris@chris-wilson.co.uk> wrote:
> > > > On Tue, Apr 26, 2016 at 07:29:41PM +0200, Daniel Vetter wrote:
> > > >> amdgpu gained dev->struct_mutex usage, and that's because it's walking
> > > >> the dev->filelist list. Protect that list with it's own lock to take
> > > >> one more step towards getting rid of struct_mutex usage in drivers
> > > >> once and for all.
> > > >>
> > > >> While doing the conversion I noticed that 2 debugfs files in i915
> > > >> completely lacked appropriate locking. Fix that up too.
> > > >>
> > > >> v2: don't forget to switch to drm_gem_object_unreference_unlocked.
> > > >>
> > > >> Cc: Alex Deucher <alexander.deucher@amd.com>
> > > >> Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
> > > >
> > > > Just wondering if this worth converting over. Opening/closing isn't
> > > > going to be high contention, I hope, though we can certainly write
> > > > stress cases for it! The goal for drivers to stop using the struct_mutex
> > > > as their BKL, which doesn't preclude keeping the struct_mutex around for
> > > > where it makes sense to have a single mutex rather than a multitude.
> > > >
> > > > I have some misgivings over this, but only because I think its overkill.
> > > > Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
> > > 
> > > I agree with Chris' sentiments.
> > 
> > It's not to have more speed or less contention, but just to have fewer
> > things to worry about when reviewing locking. Hence orthogonal locks for
> > independent parts.
> > 
> > My goal is that in the end dev->struct_mutex is only used by some existing
> > drivers for their internals, plus all the legacy core stuff. And never
> > even used by modern drivers. New locks are pretty cheap, and not dragging
> > in the entire legacy horror show has benefits.
> > 
> > When/once I tackle the one thing left (master locking) I might move the
> > master handling under this lock too (since it's closely related to open
> > files). Not sure yet.
> 
> drm: s/struct_mutex/legacy_mutex/

Yeah that's the eventual plan. Probably need to convert over some of the
current gem drivers first to make it less of a flag day.

> drm/i915: s/struct_mutex/bfg9000/

Can't do that yet because holding struct_mutex prevents objects from
disappearing, speficially all our mm and lru lists only hold a weak
reference. We need to rework our shrinkers first and switch over to
gem_object_free_unlocked before we can disengage from struct_mutex in
i915.

But yeah, again that's the plan.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/intel-gfx