[Intel-gfx] [RFC 7/8] drm/i915/tracepoints: Add backend level request in and out tracepoints

Tvrtko Ursulin tvrtko.ursulin at linux.intel.com
Fri Jan 27 13:59:54 UTC 2017


On 27/01/2017 12:27, Chris Wilson wrote:
> On Fri, Jan 27, 2017 at 12:01:26PM +0000, Tvrtko Ursulin wrote:
>> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>
>> Two new tracepoints placed at the call sites where requests are
>> actually passed to the GPU enable userspace to track engine
>> utilisation.
>>
>> These tracepoints are only enabled when the
>> DRM_I915_LOW_LEVEL_TRACEPOINTS Kconfig option is enabled.
>>
>> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>> ---
>>  drivers/gpu/drm/i915/i915_guc_submission.c |  2 ++
>>  drivers/gpu/drm/i915/i915_trace.h          | 49 ++++++++++++++++++++++++++++++
>>  drivers/gpu/drm/i915/intel_lrc.c           |  4 +++
>>  3 files changed, 55 insertions(+)
>>
>> diff --git a/drivers/gpu/drm/i915/i915_guc_submission.c b/drivers/gpu/drm/i915/i915_guc_submission.c
>> index 8ced9e26f075..beec88a30347 100644
>> --- a/drivers/gpu/drm/i915/i915_guc_submission.c
>> +++ b/drivers/gpu/drm/i915/i915_guc_submission.c
>> @@ -518,6 +518,8 @@ static void __i915_guc_submit(struct drm_i915_gem_request *rq)
>>  	if (i915_vma_is_map_and_fenceable(rq->ring->vma))
>>  		POSTING_READ_FW(GUC_STATUS);
>>
>> +	trace_i915_gem_request_in(rq, 0);
>
> But how to get out for guc? We could do a similar in for ringbuffer.
> The original way I used it was in == dispatch, out == ring-notify, which
> is why trace_i915_gem_ring_dispatch enabled the signaling.

That was my thinking as well.

> Hmm. Following on from that we can add the out tracepoint as a
> fence-callback. For the moment, I'd drop guc/legacy
> trace_i915_gem_request_in and we will try something more magical. Though
> once we do enable the fake GuC scheduler we will have an appropriate
> place to drop an trace_i915_gem_request_out. Just leaving ringbuffer as
> the odd one out, but for who arguably the in/out logic is not as
> important.

Fence signalled is very lazy if no one is listening, no? So until the 
GUC backend I thought request_in and deriving the _out from _notify in 
userspace would be OK. Meaning enable signalling stays until then.

Regards,

Tvrtko


More information about the Intel-gfx mailing list