[Intel-gfx] [PATCH 1/2] drm/i915: Keep a count of requests waiting for a slot on GPU

[Tvrtko Ursulin]

On 22/11/2017 13:35, Chris Wilson wrote:
> Quoting Tvrtko Ursulin (2017-11-22 13:31:56)
>>
>> On 22/11/2017 12:59, Chris Wilson wrote:
>>> Quoting Tvrtko Ursulin (2017-11-22 12:46:21)
>>>> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>>>
>>>> Keep a per-engine number of runnable (waiting for GPU time) requests.
>>>>
>>>> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>>> ---
>>>>    drivers/gpu/drm/i915/i915_gem_request.c | 5 +++++
>>>>    drivers/gpu/drm/i915/intel_engine_cs.c  | 5 +++--
>>>>    drivers/gpu/drm/i915/intel_lrc.c        | 1 +
>>>>    drivers/gpu/drm/i915/intel_ringbuffer.h | 8 ++++++++
>>>>    4 files changed, 17 insertions(+), 2 deletions(-)
>>>>
>>>> diff --git a/drivers/gpu/drm/i915/i915_gem_request.c b/drivers/gpu/drm/i915/i915_gem_request.c
>>>> index 7325469ce754..e3c74cafa7d4 100644
>>>> --- a/drivers/gpu/drm/i915/i915_gem_request.c
>>>> +++ b/drivers/gpu/drm/i915/i915_gem_request.c
>>>> @@ -480,6 +480,9 @@ void __i915_gem_request_submit(struct drm_i915_gem_request *request)
>>>>           engine->emit_breadcrumb(request,
>>>>                                   request->ring->vaddr + request->postfix);
>>>>    
>>>> +       GEM_BUG_ON(engine->queued == 0);
>>>> +       engine->queued--;
>>>
>>> Ok, so under engine->timeline->lock.
>>>
>>>> +
>>>>           spin_lock(&request->timeline->lock);
>>>>           list_move_tail(&request->link, &timeline->requests);
>>>>           spin_unlock(&request->timeline->lock);
>>>> @@ -525,6 +528,8 @@ void __i915_gem_request_unsubmit(struct drm_i915_gem_request *request)
>>>>           timeline = request->timeline;
>>>>           GEM_BUG_ON(timeline == engine->timeline);
>>>>    
>>>> +       engine->queued++;
>>>> +
>>>>           spin_lock(&timeline->lock);
>>>>           list_move(&request->link, &timeline->requests);
>>>>           spin_unlock(&timeline->lock);
>>>> diff --git a/drivers/gpu/drm/i915/intel_engine_cs.c b/drivers/gpu/drm/i915/intel_engine_cs.c
>>>> index d53680c08cb0..cc9d60130ddd 100644
>>>> --- a/drivers/gpu/drm/i915/intel_engine_cs.c
>>>> +++ b/drivers/gpu/drm/i915/intel_engine_cs.c
>>>> @@ -1675,12 +1675,13 @@ void intel_engine_dump(struct intel_engine_cs *engine, struct drm_printer *m)
>>>>           u64 addr;
>>>>    
>>>>           drm_printf(m, "%s\n", engine->name);
>>>> -       drm_printf(m, "    current seqno %x, last %x, hangcheck %x [%d ms], inflight %d\n",
>>>> +       drm_printf(m, "    current seqno %x, last %x, hangcheck %x [%d ms], inflight %d, queued %d\n",
>>>>                      intel_engine_get_seqno(engine),
>>>>                      intel_engine_last_submit(engine),
>>>>                      engine->hangcheck.seqno,
>>>>                      jiffies_to_msecs(jiffies - engine->hangcheck.action_timestamp),
>>>> -                  engine->timeline->inflight_seqnos);
>>>> +                  engine->timeline->inflight_seqnos,
>>>> +                  INTEL_GEN(dev_priv) >= 8 ? engine->queued : -1);
>>>
>>> Not gen8 specific, just add engine->queued++ to i9xx_submit_request().
>>
>> But where to put the decrement, and more importantly, how not make it
>> lag the reality from the retire worker? :(
> 
> The decrement is in __i915_gem_request_submit as before.
> So basically it should remain 0, since we aren't keeping a queue of work
> for the HW and just submitting into the ringbuffer as soon as we are
> ready. (This may not always remain so...) Hence why the (last_seqno -
> current_seqno) was so important.

I keep getting lost in these callbacks...

Right, so if we put the increment in i9xx_submit_request, it calls 
i915_gem_request_submit which immediately decrements it, and as you say 
it remains at zero then.

So you just want to look at the last_seqno - current_seqno on <gen8. 
Which will be the totality of submitted stuff so sounds OK. Makes the 
metric usable on those platforms as well. Sorry for the confusion.

Regards,

Tvrtko