[PATCH 00/23] drm/etnaviv: support performance counters

[Christian Gmeiner]

Hi Lucas

2017-09-13 13:01 GMT+02:00 Lucas Stach <l.stach at pengutronix.de>:
> Am Dienstag, den 12.09.2017, 17:11 +0200 schrieb Christian Gmeiner:
>> In a perfect world we would be able to read GPU registers of interest
>> via the command stream with a 'read-register' command/package. For perf
>> counters it is a must to read them synchronized with the GPU to put the
>> values in relation to a draw command. As Vivante GPUs do not provide this
>> functionality we need to emulate it in software.
>>
>>
>> We need to support three different kind of perf register types:
>>
>> 1) normal register
>>   This is the easierst case where we can simply read the register and we
>>   are done.
>>
>> 2) debug register
>>   We need to configure the mux register and then read the debug register value.
>>
>> 3) pipeline register
>>   We need to 'iterate' over all pixel pipes and sum up the values. The 'iteration'
>>   is done by select the pipe of interest via HI_CLOCK_CONTROL_DEBUG_PIXEL_PIPE.
>>   There is also need to configure the mux register.
>>
>>
>> Allowing the userspace to do it all by its own feels quite error prone and not
>> future-proof. Thats why the kernel exports all performance domains and their
>> signals to the userspace via two new ioctls. So the kernel knows all performance
>> counters and how to sample them.
>>
>> At the moment all performacne domains and signals get exported to all gpu pipe types,
>> but that can be changed in follow-up patches.
>
> Okay, after testing the V4 patches I'm convinced that we need this
> filter _before_ any of this can go in.
>

Will integrate this in the next version of this series.

> At least on GC3000 profiling the PE counters wastes the GPU bad enough
> that not even our regular GPU recovery mechanism is able to get it back
> to life. This is a trivial DOS vector, which we simply can't allow in
> the kernel driver. For this to go in, we need some way to only export
> the counters, which are _tested_ to be safe on the hardware.
>

I have tested all exported counters on a GC2000 with glxgears - one
perf counter for each glxgears
run. I have not seen any GPU hangs or similar. Will redo this 'testing
procedure on a GC3000.

Thanks for review :)

greets
--
Christian Gmeiner, MSc

https://christian-gmeiner.info