Re: [PATCH v8 10/12] drm/i915: add oa_event_min_timer_exponent sysctl

[Robert Bragg <robert@xxxxxxxxxxxxx>]

On Wed, Nov 2, 2016 at 6:29 AM, sourab gupta <sourab.gupta@xxxxxxxxx> wrote:

On Thu, 2016-10-27 at 19:14 -0700, Robert Bragg wrote:
> The minimal sampling period is now configurable via a
> dev.i915.oa_min_timer_exponent sysctl parameter.
>
> Following the precedent set by perf, the default is the minimum that
> won't (on its own) exceed the default kernel.perf_event_max_sample_rate
> default of 100000 samples/s.
>
> Signed-off-by: Robert Bragg <robert@xxxxxxxxxxxxx>
> Reviewed-by: Matthew Auld <matthew.auld@xxxxxxxxx>
> ---
>  drivers/gpu/drm/i915/i915_perf.c | 42 ++++++++++++++++++++++++++++------------
>  1 file changed, 30 insertions(+), 12 deletions(-)
>
> diff --git a/drivers/gpu/drm/i915/i915_perf.c b/drivers/gpu/drm/i915/i915_perf.c
> index 4e42073..e3c6f51 100644
> --- a/drivers/gpu/drm/i915/i915_perf.c
> +++ b/drivers/gpu/drm/i915/i915_perf.c
> @@ -82,6 +82,22 @@ static u32 i915_perf_stream_paranoid = true;
>  #define INVALID_CTX_ID 0xffffffff
>
>
> +/* for sysctl proc_dointvec_minmax of i915_oa_min_timer_exponent */
> +static int oa_exponent_max = OA_EXPONENT_MAX;
> +
> +/* Theoretically we can program the OA unit to sample every 160ns but don't
> + * allow that by default unless root...
> + *
> + * The period is derived from the exponent as:
> + *
> + *   period = 80ns * 2^(exponent + 1)
> + *
> + * Referring to perf's kernel.perf_event_max_sample_rate for a precedent
> + * (100000 by default); with an OA exponent of 6 we get a period of 10.240
> + * microseconds - just under 100000Hz
> + */
> +static u32 i915_oa_min_timer_exponent = 6;

For HSW, the timestamp period is 80ns, so the exponent of 6 translates
to sampling rate of ~100000Hz. But the timestamp period may change for
other platforms, leading to different values of oa_min_timer_exponent
corresponding to sampling rate of ~100000Hz. Do we plan to have this
value platform specific subsequently, or the guidance value of ~100000Hz
min sampling rate needn't be strictly followed?

actually it's bothered me a bit that I've been lazy with not having this adapt for gen9+ in later patches

I think it would probably be better to make this a Hz based threshold for userspace, otherwise any userspace policy here needs to be adapted for each system with a different timestamp frequency which isn't great.

I've updated the patch locally to make this an oa_max_sample_rate parameter in Hz, which I'll aim to test on haswell tomorrow and send out.

Thanks,

- Robert

_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx