[Intel-gfx] [PATCH] drm/i915: Convert hangcheck from a timer into a delayed work item

[Ville Syrjälä]

On Thu, Sep 04, 2014 at 04:38:39PM +0100, Chris Wilson wrote:
> On Thu, Sep 04, 2014 at 06:25:03PM +0300, Ville Syrjälä wrote:
> > On Thu, Sep 04, 2014 at 04:09:02PM +0100, Chris Wilson wrote:
> > > When run as a timer, i915_hangcheck_elapsed() must adhere to all the
> > > rules of running in a softirq context. This is advantageous to us as we
> > > want to minimise the risk that a driver bug will prevent us from
> > > detecting a hung GPU. However, that is irrelevant if the driver bug
> > > prevents us from resetting and recovering. Still it is prudent not to
> > > rely on mutexes inside the checker, but given the coarseness of
> > > dev->struct_mutex doing so is extremely hard.
> > > 
> > > Give in and run from a work queue, i.e. outside of softirq.
> > > 
> > > v2:
> > > 
> > > The conversion does have one significant change, from the use of
> > > mod_timer to schedule_delayed_work, means that the time that we execute
> > > the first hangcheck is fixed and not continually deferred by later work.
> > > This has the advantage of not allowing userspace to fill the ring before
> > > hangcheck can finally run. At the same time, it removes the ability for
> > > the interrupt to defer the hangcheck as well. This is sensible for that
> > > an interrupt is only for a single engine, whereas we perform hangcheck
> > > globally, so whilst one ring may have hung, the other could be running
> > > normally and preventing the hangcheck from firing.
> > 
> > But doesn't this make it so that we may not detect a hang unless more
> > work gets submitted constantly? Eg.
> > 
> > 1. execbuffer batch 1 -> queue hangcheck schedules work
> > 2. execbuffer batch 2 -> queue hangcheck does nothing
> > 3. execbuffer batch 3 -> queue hangcheck does nothing
> > 4. hangcheck expires and sees progress up to batch 2 -> everything is fine
> 4.b hangcheck rearms itself as there is outstanding wrok

Indeed. I should have actually read the code and it would have been
obvious.

> > 5. batch 3 hangs
> 6. hangcheck fires, sees progress, rearms
> 7. hangcheck fires, sees no progress, shoots the user.

Sounds like we need a disclaimer about the dangers of causing a GPU
hang :)

-- 
Ville Syrjälä
Intel OTC