Re: [Intel-gfx] [PATCH 2/3] drm/i915/guc/slpc: Add waitboost functionality for SLPC

["Belgaumkar, Vinay" <vinay.belgaumkar@intel.com>]

On 11/1/2021 1:28 PM, Dixit, Ashutosh wrote:
> On Sun, 31 Oct 2021 21:39:36 -0700, Belgaumkar, Vinay wrote:
>>
>> @@ -945,6 +960,17 @@ void intel_rps_boost(struct i915_request *rq)
>> 	if (!test_and_set_bit(I915_FENCE_FLAG_BOOST, &rq->fence.flags)) {
>> 		struct intel_rps *rps = &READ_ONCE(rq->engine)->gt->rps;
>>
>> +		if (rps_uses_slpc(rps)) {
>> +			slpc = rps_to_slpc(rps);
>> +
>> +			/* Return if old value is non zero */
>> +			if (atomic_fetch_inc(&slpc->num_waiters))
>> +				return;
>> +
>> +			if (intel_rps_get_requested_frequency(rps) < slpc->boost_freq)
> 
> I think this check is not needed because:
> 
> a. The waitboost code only changes min_freq. i915 code should not depend on
>     how GuC changes requested_freq in response to change in min_freq.
> 
> b. What is more worrisome is that when we "de-boost" we set min_freq to
>     min_freq_softlimit. If GuC e.g. has a delay in bringing requested_freq
>     down and intel_rps_boost() gets called meanwhile we will miss the one
>     opportunity we have to boost the freq (when num_waiters goes from 0 to
>     1. Asking GuC to boost when actual_freq is already boost_freq is
>     harmless in comparison). So to avoid this risk of missing the chance to
>     boost I think we should delete this check and replace the code above
>     with something like:
> 
>                  if (rps_uses_slpc(rps)) {
>                          struct intel_guc_slpc *slpc = rps_to_slpc(rps);
> 
>                          if (slpc->boost_freq <= slpc->min_freq_softlimit)
>                                  return;
> 
>                          if (!atomic_fetch_inc(&slpc->num_waiters))
>                                  schedule_work(&slpc->boost_work);
> 
>                          return;
>                  }
> 
> Note that this check:
> 
>                  if (slpc->boost_freq <= slpc->min_freq_softlimit)
>                                  return;
> 
> (which is basically a degenerate case in which we don't have to do
> anything), can be probably be implemented when boost_freq is set in sysfs,
> or may already be encompassed in "val < slpc->min_freq" in
> intel_guc_slpc_set_boost_freq() in which case this check can also be
> skipped from this function.

We already have that check in set_boost_freq function. So, just adding 
the atomic_fetch_inc check.

> 
>> +void intel_guc_slpc_dec_waiters(struct intel_guc_slpc *slpc)
>> +{
>> +	/* Return min back to the softlimit.
>> +	 * This is called during request retire,
>> +	 * so we don't need to fail that if the
>> +	 * set_param fails.
>> +	 */
> 
> nit: maybe follow kernel multi-line comment format.
> 
Ok.

Thanks,
Vinay.