[Intel-gfx] [PATCH 7/8] drm/i915: Grab execlist spinlock to avoid post-reset concurrency issues.

[Chris Wilson]

On Tue, Oct 13, 2015 at 03:46:00PM +0200, Daniel Vetter wrote:
> On Tue, Oct 13, 2015 at 12:45:58PM +0100, Chris Wilson wrote:
> > On Tue, Oct 13, 2015 at 01:46:38PM +0200, Daniel Vetter wrote:
> > > On Fri, Oct 09, 2015 at 09:45:16AM +0100, Chris Wilson wrote:
> > > > On Fri, Oct 09, 2015 at 10:38:18AM +0200, Daniel Vetter wrote:
> > > > > On Thu, Oct 08, 2015 at 07:31:39PM +0100, Tomas Elf wrote:
> > > > > > Grab execlist lock when cleaning up execlist queues after GPU reset to avoid
> > > > > > concurrency problems between the context event interrupt handler and the reset
> > > > > > path immediately following a GPU reset.
> > > > > > 
> > > > > > Signed-off-by: Tomas Elf <tomas.elf at intel.com>
> > > > > 
> > > > > Should we instead just stop any irqs from the GT completely while we do
> > > > > the reset (plus a synchronize_irq)? It's a bit heavy-weight, but probably
> > > > > safer. Well not the entire GT, but all the rings under reset (as prep for
> > > > > per-ring reset).
> > > > 
> > > > Bah, stopping IRQs is not enough for error state capture though since
> > > > requests complete asynchronously just by polling a memory address. (If
> > > > that is the goal here, this patch just makes execlist_queue access
> > > > consistent and should only be run once the GPU has been reset and so is
> > > > categorically idle.)
> > > 
> > > This is the execlist ELSP tracking, which is execlusively driven by the
> > > CTX_SWITCH interrupt signal from each engine.
> > > 
> > > At least that's been my assumption, and under that assumption I think
> > > stalling interrupts should be good enough.
> > 
> > No, because the requests and vma are not coupled to the interrupt in
> > terms of when they can disappear.
> 
> At least today execlist keeps its own reference on requests until the
> CTX_SWITCH stuff is done to exactly make sure this is the case. And even
> when we have that fixed up I think we do need to exclude this processing
> somehow, and the irqsave spinlock seems ok for that. disabling the
> interrupt itself plus synchronize_irq was really just an idea.

What I meant was that we have this problem with vma/obj disappearing not
simply on execlists because request completion is asynchronous (we just
check for the GPU's breadcrumb). Execlists is slightly special in that
its requests can't disappear, but they can elsewhere.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre