[PATCH 5/7] compositor-drm: Gracefully handle vblank and flip invalid timestamps

[Daniel Stone]

Hi,

On 22 March 2016 at 08:50, Daniel Vetter <daniel at ffwll.ch> wrote:
> On Tue, Mar 22, 2016 at 08:47:12AM +0000, Daniel Stone wrote:
>> This is odd. Instant queries failing is fine, but in which
>> circumstances can you fail to return a valid timestamp in a pageflip
>> event ... ? I don't know of any other kernel driver which does this -
>> at the very least, if you don't have a hardware timestamp, then you
>> log the time in your kernel driver which handles the hardware event
>> (e.g. softirq handler), and use that instead.
>
> nouveau bugs iirc. Yes, should never happen, it's a bug in the kernel
> driver that iirc Mario recently tracked down. nouveau somehow started the
> vblank machinery before the pipe was up, or something like that.
>
> So yeah happens, I'm not sure we want to let kernel drivers get away with
> plain broken behavior forever. I'm leaning towards nack, pls fix the
> kernel driver.

Yes, particularly as no-one has reported it on any other drivers. I
don't think it's behaviour we really want to support.

> And maybe make sure the corresponding igt pass ;-)

For reference, igt here is intel-gpu-tools, which despite its name is
actually a generic KMS test/conformance suite. Most of the tests are
still Intel-specific, but we are actively working on making them
generic - I was actually in the middle of reviewing a patchset to make
a rather larger chunk of it generic when I got distracted by this
thread. ;) It's a great way to find out ways in which your driver
breaks userspace expectations of a 'well-behaved' KMS driver early.

Cheers,
Daniel