[Intel-gfx] [PATCH] drm/i915: Make vm eviction uninterruptible

[Daniel Vetter]

On Mon, Apr 07, 2014 at 10:42:56AM +0100, Chris Wilson wrote:
> On Sun, Apr 06, 2014 at 11:35:03AM -0700, Ben Widawsky wrote:
> > On Sat, Apr 05, 2014 at 07:45:28PM -0700, Ben Widawsky wrote:
> > > The issue I was seeing appeared to seeing from sigkill. In such a case,
> > > the process may want to die before the context/work/address space is
> > > freeable. For example:
> > > 1. evict_vm called for whatever reason
> > > 2. wait_seqno because the VMA is still active
> > 
> > hmm something isn't right here. Why did I get to wait_seqno if pin_count
> > was 0? Just FYI, this wasn't hypothetical. I did trace it all the way to
> > exactly ERESTARTSYS from wait_seqno.
> > 
> > By the way, another option in evict would be:
> > while(ret = (i915_vma_unbind(vma) == -ERESTARTSYS));
> > WARN_ON(ret);
> > 
> > > 3. receive signal break out of wait_seqno
> > > 4. return to evict_vm and the above WARN
> > > 
> > > Our error handling from there just spirals.
> > > 
> > > One issue I have with our current code is I'd really like eviction to
> > > not be able to fail (obviously extreme cases are unavoidable).
> 
> This is unrealistic since we must support X which uses sigtimer.
> 
> > > Perhaps
> > > one other solution would be to make sure the context is idled before
> > > evicting its VM.
> 
> Indeed.
> 
> Anyway, I do concur that wrapping i915_driver_preclose() with
> 
> dev_priv->mm.interruptible = false;
> 
> would make us both happy.

Isn't the backtrace just fallout from the lifetime rules being a bit
funny? We didn't uninterruptibly stall for any still active bo when the
drm fd gets closed, why do we suddenly need to do that with ppgtts? Iirc
requests hold a ref on the context, contexts hold a ref on the ppgtt and
so the entire thing should only dissipate once it's really idle.

Imo just doing uninterruptible sleeps tastes way too much like duct-tape.
I can be convinced of duct-tape if the tradeoffs really strongly suggests
it's the right thing (e.g. the shrinker lock stealing, even though we've
paid a hefty price in accidental complexity with that one), but that needs
some good justification.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch