Op 12-02-18 om 18:01 schreef Ville Syrjälä:
> On Fri, Feb 09, 2018 at 06:21:08PM +0100, Maarten Lankhorst wrote:
>> Op 09-02-18 om 11:04 schreef Chris Wilson:
>>> Quoting Maarten Lankhorst (2018-02-09 09:53:59)
>>>> Some cleanups to move the uncore.lock around vblank evasion, so run
>>>> to completion without racing on uncore.lock. Hopefully this will reduce
>>>> the chance of underruns, and perhaps allows us to decrease
>>>> VBLANK_EVASION_TIME_US as well as a followup patch.
>>>> Tested on KBL and BSW.
>>> * shivers
>>> uncore.lock is a brutally contested lock. Ville's patches did work on
>>> splitting the uncore.lock into forcewake and display variants, which
>>> cuts down on the nasty side effects.
>>> Latency profiling, another item for the CI/QA wishlist.
>> Yeah, unfortunately this is not different from status quo. We already
>> require everything inside vblank evasion to run as fast as possible,
>> and it's down to a list of register writes and a few reads. Those
>> already need the uncore.lock, so all we do now is being more explicit
>> about when we take it and eliminate contention when we write out the
>> register values.
> Would be nice to have some results for this though. IIRC when I was
> benchmarking my update optimizations and the de_lock stuff I was
> simply logging how long the updates take, and staring at histograms
> of that after running a bunch of igts and whatnot. I'm not sure I
> have the results anymore, but IIRC I did see some improvement.
When testing with KBL and BSW, this patch series most updates complete in <40
us even with
all debug options set, with the highest amount of time being a single update of
93 us for BSW.
Because we take all locking including the vblank reference in advance, latency
locks no longer affects the time critical part of vblank evasion.
Tested with kms_rotation_crc,
Intel-gfx mailing list