[PATCH 5/6] drm: Prevent vblank counter jumps with timestamp based update method.

[Daniel Vetter]

On Wed, Feb 10, 2016 at 7:36 PM, Mario Kleiner
<mario.kleiner.de at gmail.com> wrote:
> On 02/10/2016 06:17 PM, Daniel Vetter wrote:
>>
>> On Wed, Feb 10, 2016 at 5:28 PM, Mario Kleiner
>> <mario.kleiner.de at gmail.com> wrote:
>>>
>>> There's another scenario where this zero-ts case can be hit. If the
>>> driver
>>> drm_vblank_init()'s - setting all timestamps to zero - and then code
>>> starts
>>> using vblanks (drm_vblank_get()) without drm_vblank_on beforehand, which
>>> is
>>> afaics the case with nouveau. Unless that's considered an error as well,
>>> we'd need to init the timestamps to something non-zero but harmless like
>>> 1
>>> usecs at drm_vblank_init() time?
>>
>>
>> Both legacy modeset helpers and atomic ones assume by default that you
>> start out with everything disabled. Pre-atomic drivers make that
>> happen by calling disable_unused_functions() to shut down anything the
>> bios has enabled. I think this can't happen.
>>
>> For drivers that do take over bootloader display config they must call
>> vblank_on explicitly themselves, which i915 does.
>>
>>> What makes sense as output here? DRM_WARN_ONCE?
>>
>>
>> I'd go with WARN_ON and tune it down if it's offensive. But WARN_ON
>> patch for 4.6 of course.
>> -Daniel
>>
>
> Ok, so does this one have your R-b for stable as is? What about a proper
> nouveau fix if i find one? Probably also for 4.6 then, given that this patch
> fixes it up good enough for stable?

If possible I'd prefer we just fix nouveau up for stable, and do a
WARN_ON patch when this ever happens for 4.6. It sounded like you've
figured out already what nouveau needs?
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch