Re: [PATCH 08/21] drm/i915/gem: Disallow bonding of virtual engines

From: Daniel Vetter <daniel@ffwll.ch>
To: Jason Ekstrand <jason@jlekstrand.net>
Cc: Matthew Brost <matthew.brost@intel.com>,
	Intel GFX <intel-gfx@lists.freedesktop.org>,
	Maling list - DRI developers <dri-devel@lists.freedesktop.org>
Subject: Re: [PATCH 08/21] drm/i915/gem: Disallow bonding of virtual engines
Date: Thu, 29 Apr 2021 14:16:05 +0200	[thread overview]
Message-ID: <YIqjhXiIuKc1Hw8r@phenom.ffwll.local> (raw)
In-Reply-To: <CAOFGe979BEBSSw_0oqGRADHWZJAg4c=iyr3FiJB5oxno+YM+uw@mail.gmail.com>

On Wed, Apr 28, 2021 at 01:58:17PM -0500, Jason Ekstrand wrote:
> On Wed, Apr 28, 2021 at 12:18 PM Jason Ekstrand <jason@jlekstrand.net> wrote:
> >
> > On Wed, Apr 28, 2021 at 5:13 AM Daniel Vetter <daniel@ffwll.ch> wrote:
> > >
> > > On Tue, Apr 27, 2021 at 08:51:08AM -0500, Jason Ekstrand wrote:
> > > > I sent a v2 of this patch because it turns out I deleted a bit too
> > > > much code.  This function in particular, has to stay, unfortunately.
> > > > When a batch is submitted with a SUBMIT_FENCE, this is used to push
> > > > the work onto a different engine than than the one it's supposed to
> > > > run in parallel with.  This means we can't dead-code this function or
> > > > the bond_execution function pointer and related stuff.
> > >
> > > Uh that's disappointing, since if I understand your point correctly, the
> > > sibling engines should all be singletons, not load balancing virtual ones.
> > > So there really should not be any need to pick the right one at execution
> > > time.
> >
> > The media driver itself seems to work fine if I delete all the code.
> > It's just an IGT testcase that blows up.  I'll do more digging to see
> > if I can better isolate why.
> 
> I did more digging and I figured out why this test hangs.  The test
> looks at an engine class where there's more than one of that class
> (currently only vcs) and creates a context where engine[0] is all of
> the engines of that class bonded together and engine[1-N] is each of
> those engines individually.  It then tests that you can submit a batch
> to one of the individual engines and then submit with
> EXEC_FENCE_SUBMIT to the balanced engine and the kernel will sort it
> out.  This doesn't seem like a use-case we care about.
> 
> If we cared about anything, I would expect it to be submitting to two
> balanced contexts and expecting "pick any two" behavior.  But that's
> not what the test is testing for.

Yeah ditch it.

Instead make sure that the bonded setparam/ctx validation makes sure that
1) no virtual engines are used
2) no engine used twice
3) anything else stupid you can come up with that we should make sure is
blocked.

Cheers, Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel