[Intel-gfx] [PATCH v2 1/3] drm/i915: Mark i915_request.timeline as a volatile, rcu pointer

[Chris Wilson]

Quoting Tvrtko Ursulin (2019-09-19 18:11:14)
> 
> On 19/09/2019 14:26, Chris Wilson wrote:
> > Quoting Tvrtko Ursulin (2019-09-19 14:02:19)
> >>
> >> On 19/09/2019 12:19, Chris Wilson wrote:
> >>> +static struct intel_timeline *get_timeline(struct i915_request *rq)
> >>> +{
> >>> +     struct intel_timeline *tl;
> >>> +
> >>> +     /*
> >>> +      * Even though we are holding the engine->active.lock here, there
> >>> +      * is no control over the submission queue per-se and we are
> >>> +      * inspecting the active state at a random point in time, with an
> >>> +      * unknown queue. Play safe and make sure the timeline remains valid.
> >>> +      * (Only being used for pretty printing, one extra kref shouldn't
> >>> +      * cause a camel stampede!)
> >>> +      */
> >>> +     rcu_read_lock();
> >>> +     tl = rcu_dereference(rq->timeline);
> >>> +     if (!kref_get_unless_zero(&tl->kref))
> >>> +             tl = NULL;
> >>> +     rcu_read_unlock();
> >>
> >> How can it be NULL under the active lock? Isn't that the same assertion
> >> from i915_timeline_get_active.
> > 
> > Not NULL, but retired. The difference is that during submission we know
> > that this request's context/timeline must be currently pinned until
> > a subsequent request (containing the idle-barriers) is submitted. The
> > danger I worry about here is that subsequent idle request may be already
> > submitted and since the queued requests may *already* have been retired,
> > the timeline may be unpinned and indeed dropped it's last reference.
> 
> But here it is under the engine->active.lock with interrupts disabled 
> and the requests are fetched from execlists ports. Timeline is not 
> guaranteed to be kept alive under these conditions? intel_context 
> reference will be held until process_csb schedules it out so I'd expect 
> timeline and hwsp to be there. But I could be lost in the new scheme of 
> things.

I felt it was prudent to only rely on the active pin. You are right in
that we have a context reference if it is in active, and that context
holds a reference to the timeline. But... engine->active.lock is not
the lock that guards rq->timeline, so I feel uneasy on extending
i915_request_active_timeline() too far. Outside of the submission
pathway, inside a pretty printer, it feels safer (whatever changes may
come we don't have to worry about it) to not assume anything and just
use the failsafe rcu_dereference() + kref.
-Chris