Re: [PATCH v6] drm/i915/selftests: Implement frequency logging for energy reading validation

["Anirban, Sk" <sk.anirban@intel.com>]

On 20-11-2024 20:20, Nilawar, Badal wrote:
>
>
> On 20-11-2024 16:00, Gupta, Anshuman wrote:
>>
>>
>>> -----Original Message-----
>>> From: Nilawar, Badal <badal.nilawar@intel.com>
>>> Sent: Wednesday, November 20, 2024 1:44 PM
>>> To: Anirban, Sk <sk.anirban@intel.com>; intel-gfx@lists.freedesktop.org
>>> Cc: Gupta, Anshuman <anshuman.gupta@intel.com>; Poosa, Karthik
>>> <karthik.poosa@intel.com>; Pottumuttu, Sai Teja
>>> <sai.teja.pottumuttu@intel.com>
>>> Subject: Re: [PATCH v6] drm/i915/selftests: Implement frequency 
>>> logging for
>>> energy reading validation
>>>
>>>
>>>
>>> On 13-11-2024 15:20, Sk Anirban wrote:
>>>> Introduce RC6 & RC0 frequency logging mechanism to ensure accurate
>>>> energy readings aimed at addressing GPU energy leaks and power
>>>> measurement failures.
>>>> This enhancement will help ensure the accuracy of energy readings.
>>>>
>>>> v2:
>>>>     - Improved commit message.
>>>> v3:
>>>>     - Used pr_err log to display frequency (Anshuman)
>>>>     - Sorted headers alphabetically (Sai Teja)
>>>> v4:
>>>>     - Improved commit message.
>>>>     - Fix pr_err log (Sai Teja)
>>>> v5:
>>>>     - Add error & debug logging for RC0 power and frequency checks
>>>> (Anshuman)
>>>> v6:
>>>>     - Modify debug logging for RC0 power and frequency checks (Sai
>>>> Teja)
>>>>
>>>> Signed-off-by: Sk Anirban <sk.anirban@intel.com>
>>>> Reviewed-by: Sai Teja Pottumuttu <sai.teja.pottumuttu@intel.com>
>>>> ---
>>>>    drivers/gpu/drm/i915/gt/selftest_rc6.c | 15 +++++++++++++--
>>>>    1 file changed, 13 insertions(+), 2 deletions(-)
>>>>
>>>> diff --git a/drivers/gpu/drm/i915/gt/selftest_rc6.c
>>>> b/drivers/gpu/drm/i915/gt/selftest_rc6.c
>>>> index 1aa1446c8fb0..a8776f88d6a1 100644
>>>> --- a/drivers/gpu/drm/i915/gt/selftest_rc6.c
>>>> +++ b/drivers/gpu/drm/i915/gt/selftest_rc6.c
>>>> @@ -8,6 +8,7 @@
>>>>    #include "intel_gpu_commands.h"
>>>>    #include "intel_gt_requests.h"
>>>>    #include "intel_ring.h"
>>>> +#include "intel_rps.h"
>>>>    #include "selftest_rc6.h"
>>>>
>>>>    #include "selftests/i915_random.h"
>>>> @@ -38,6 +39,9 @@ int live_rc6_manual(void *arg)
>>>>        ktime_t dt;
>>>>        u64 res[2];
>>>>        int err = 0;
>>>> +    u32 rc0_freq = 0;
>>>> +    u32 rc6_freq = 0;
>>>> +    struct intel_rps *rps = &gt->rps;
>>>>
>>>>        /*
>>>>         * Our claim is that we can "encourage" the GPU to enter rc6 
>>>> at will.
>>>> @@ -66,6 +70,7 @@ int live_rc6_manual(void *arg)
>>>>        rc0_power = librapl_energy_uJ() - rc0_power;
>>>>        dt = ktime_sub(ktime_get(), dt);
>>>>        res[1] = rc6_residency(rc6);
>>>> +    rc0_freq = intel_rps_read_actual_frequency(rps);
>>>>        if ((res[1] - res[0]) >> 10) {
>>>>            pr_err("RC6 residency increased by %lldus while disabled 
>>>> for
>>> 1000ms!\n",
>>>>                   (res[1] - res[0]) >> 10);
>>>> @@ -77,7 +82,11 @@ int live_rc6_manual(void *arg)
>>>>            rc0_power = div64_u64(NSEC_PER_SEC * rc0_power,
>>>>                          ktime_to_ns(dt));
>>>>            if (!rc0_power) {
>>>> -            pr_err("No power measured while in RC0\n");
>>>> +            if (rc0_freq)
>>>> +                pr_err("No power measured while in RC0!
>>> GPU Freq: %u in RC0\n",
>>>> +                       rc0_freq);
>> If rc0 frequency is there then, this has to be pr_dbg, otherwise what 
>> is the purpose of this patch.
>
> I too didn't understand purpose of this. How this going to help for 
> accurate energy readings.
in case of rc0 power is 0 , I just want to confirm whether the freq is 
available there or not ? Also it is defined as pr_err because there is 
no point of validate rc0 power below if no power is being measured.
>
>>>> +            else
>>>> +                pr_err("No power and freq measured while in
>>> RC0\n");
>>>>                err = -EINVAL;
>>>>                goto out_unlock;
>>>>            }
>>>> @@ -91,6 +100,7 @@ int live_rc6_manual(void *arg)
>>>>        dt = ktime_get();
>>>>        rc6_power = librapl_energy_uJ();
>>>>        msleep(100);
>>>> +    rc6_freq = intel_rps_read_actual_frequency(rps);
>>>
>>> I think intention of reading frequency here is to know if device was 
>>> not in RC6
>>> when there is failure. But for the platforms below gen12 reading act 
>>> frequency
>>> will cause gt wake as GEN6_RPSTAT reg requires forcewake.
>>> To avoid wake when device is in RC6 read actual frequency without 
>>> applying
>>> forcewake.
>> If reading act_freq will wake the device, How to read frequency 
>> without forcewake then ?
>
> Use this api intel_rps_read_actual_frequency_fw to read freq without 
> forcewake.
>
> Regards,
> Badal
>
>>
>> Thanks,
>> Anshuaman
>>>
>>> Additionally add delay, may be delay of 1 seconds after re-enabling RC6
>>> manually and forcewake flush.
yeah, I can use that to read the actual freq but there is already a 
check involved just to cross verify whether the system has entered RC6 
or not.
>>>
>>> Regards,
>>> Badal
Thanks,
Anirban
>>>
>>>>        rc6_power = librapl_energy_uJ() - rc6_power;
>>>>        dt = ktime_sub(ktime_get(), dt);
>>>>        res[1] = rc6_residency(rc6);
>>>> @@ -108,7 +118,8 @@ int live_rc6_manual(void *arg)
>>>>            pr_info("GPU consumed %llduW in RC0 and %llduW in
>>> RC6\n",
>>>>                rc0_power, rc6_power);
>>>>            if (2 * rc6_power > rc0_power) {
>>>> -            pr_err("GPU leaked energy while in RC6!\n");
>>>> +            pr_err("GPU leaked energy while in RC6! GPU Freq:
>>> %u in RC6 and %u in RC0\n",
>>>> +                   rc6_freq, rc0_freq);
>>>>                err = -EINVAL;
>>>>                goto out_unlock;
>>>>            }
>>
>