[Intel-gfx] [PATCH] drm/i915/byt: Avoid tweaking evaluation thresholds

[Mika Kuoppala]

Chris Wilson <chris at chris-wilson.co.uk> writes:

> On Wed, Jan 25, 2017 at 03:09:04PM +0200, Mika Kuoppala wrote:
>> Chris Wilson <chris at chris-wilson.co.uk> writes:
>> 
>> > On Wed, Jan 25, 2017 at 02:31:08PM +0200, Mika Kuoppala wrote:
>> >> Certain Baytrails, namely the 4 cpu core variants, have been
>> >> plaqued by spurious system hangs, mostly occurring with light loads.
>> >> 
>> >> Multiple bisects by various people point to a commit which changes the
>> >> reclocking strategy for Baytrail to follow its bigger brethen:
>> >> commit 8fb55197e64d ("drm/i915: Agressive downclocking on Baytrail")
>> >> 
>> >> There is also a review comment attached to this commit from Deepak S
>> >> on avoiding punit access on Cherryview and thus it is excluded on
>> >> common reclocking path. By taking the same approach and omitting
>> >> the punit access by not tweaking the thresholds when the hardware
>> >> has been asked to move into different frequency, considerable gains
>> >> in stability have been observed.
>> >> 
>> >> With J1900 box, light render/video load would end up in system hang
>> >> in usually less than 12 hours. With this patch applied, the cumulative
>> >> uptime has now been 34 days without issues. To provoke system hang,
>> >> light loads on both render and bsd engines in parallel have been used:
>> >> glxgears >/dev/null 2>/dev/null &
>> >> mpv --vo=vaapi --hwdec=vaapi --loop=inf vid.mp4
>> >> 
>> >> So far, author has not witnessed system hang with above load
>> >> and this patch applied. Reports from the tenacious people at
>> >> kernel bugzilla are also promising.
>> >> 
>> >> Considering that the punit access frequency with this patch is
>> >> considerably less, there is a possibility that this will push
>> >> the, still unknown, root cause past the triggering point on most loads.
>> >> Further work on investigating the punit accesses on byt is welcomed.
>> >
>> > Please find the underlying problem and not disabling rps for all vlv
>> > for a GT specific problem.
>> 
>> This is not disabling rps.
>
> Your are disabling the key ingredients of the algorithm, making it less
> generic in order to workaround a problem elsewhere. You are tackling the
> symptoms and not the cause.

Yes, definitely we are tackling the symptoms.

We have been trying to find the root cause for 2 years.
Admittely hindered by the multiple other causes for
system hangs on baytrail platform.

One could argue that why was the deviation for Cherryview accepted,
as this just mimics the same way, omitting the sw adjustments.

It allows baytrail users to run their rigs without
intel_idle.max_cstate=1 which kind of ruins their power budget by far
bigger margin than this patch does.

-Mika

> -Chris
>
> -- 
> Chris Wilson, Intel Open Source Technology Centre