Virtualization - May 2016 - [PATCH v3 7/7] [wip] virtio-gpu: add page flip support

On Mon, May 30, 2016 at 02:06:50PM +0200, Gerd Hoffmann
wrote:
>   Hi,
> 
> > > But I'll take you up on the implied offer to help out and
test ;-)
> > 
> > git://people.freedesktop.org/~danvet/drm stuff
> 
> Tried that branch.
> 
> > Would be really awesome if you could test this on virtio. Note that
the
> > new nonblocking helpers require that your atomic backend gets the drm
> > event handling right. So if there's a bug in that logic then
you'll see
> > lots of dmesg noise about waits timing out (after 10s of waiting).
From a
> > quick inspection it should work though.
> 
> No timeouts.  Yay!
> 
> But it seems crtcs can be (temporarely) disabled now, so we might have
> to pick up the crtc from old_state in virtio_gpu_plane_atomic_update to
> figure which virtual output needs to be turned off.  Ran into this last
> week already.  Happened with multihead setups only, but the same patch
> fixes this one too ;)
> 
> https://lists.freedesktop.org/archives/dri-devel/2016-May/108772.html
Hm, smells more like virtio isn't too happy with the default ordering of
the commit operation. The default is:

- Disable any crtc/encoders that need to be disabled/change.
- Bash new plane setup into hw.
- Enable all crtcs/encoders that need to be enabled/have changed.

There's two problems:
- some hw gets real grumpy if you bash in plane state without the crtc
  state yet matching.
- if you do runtime pm nothing is enabled and the writes get lost at best,
  or hang your interconnect at worst.

That's why you can overwrite atomic_commit_tail, and use something more
sensible. See for example what it looks like for rockchip. I have a gut
feeling that should also take care of your troubles. Using the old crtc is
definitely not what you want.

Another option is that virtio isn't happy about bashing in plane state for
disabled crtc. Again helpers have you covered, look at the active_only
parameter for drm_atomic_helper_commit_planes().

Aside, if you wonder why these defaults: They match what the crtc helpers
are doing, but they are a bit surprising indeed.

Cheers, Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

Hi,
> > https://lists.freedesktop.org/archives/dri-devel/2016-May/108772.html
> 
> Hm, smells more like virtio isn't too happy with the default ordering
of
> the commit operation. The default is:
> 
> - Disable any crtc/encoders that need to be disabled/change.
> - Bash new plane setup into hw.
> - Enable all crtcs/encoders that need to be enabled/have changed.
> 
> There's two problems:
> - some hw gets real grumpy if you bash in plane state without the crtc
>   state yet matching.
> - if you do runtime pm nothing is enabled and the writes get lost at best,
>   or hang your interconnect at worst.
> 
> That's why you can overwrite atomic_commit_tail, and use something more
> sensible. See for example what it looks like for rockchip. I have a gut
> feeling that should also take care of your troubles. Using the old crtc is
> definitely not what you want.
> Another option is that virtio isn't happy about bashing in plane state
for
> disabled crtc. Again helpers have you covered, look at the active_only
> parameter for drm_atomic_helper_commit_planes().
virtio-gpu is a bit simplified compared to real hardware, so there isn't
really separate plane/crtc state.

Right now the virtual outputs are linked to drm_crtc.  To apply any
changes I need to lookup the crtc to figure which virtual output should
be updated.

So, setting active_only should make sure I have a valid crtc pointer on
plane updates, right?  It probably also skips the disable + enable crtc
steps on commit?  What happens when outputs are disabled?

Maybe it makes sense to link our virtual outputs to (primary) planes not
crtcs?

cheers,
  Gerd

On Tue, May 31, 2016 at 8:18 AM, Gerd Hoffmann <kraxel at redhat.com>
wrote:
>> >
https://lists.freedesktop.org/archives/dri-devel/2016-May/108772.html
>>
>> Hm, smells more like virtio isn't too happy with the default
ordering of
>> the commit operation. The default is:
>>
>> - Disable any crtc/encoders that need to be disabled/change.
>> - Bash new plane setup into hw.
>> - Enable all crtcs/encoders that need to be enabled/have changed.
>>
>> There's two problems:
>> - some hw gets real grumpy if you bash in plane state without the crtc
>>   state yet matching.
>> - if you do runtime pm nothing is enabled and the writes get lost at
best,
>>   or hang your interconnect at worst.
>>
>> That's why you can overwrite atomic_commit_tail, and use something
more
>> sensible. See for example what it looks like for rockchip. I have a gut
>> feeling that should also take care of your troubles. Using the old crtc
is
>> definitely not what you want.
>
>> Another option is that virtio isn't happy about bashing in plane
state for
>> disabled crtc. Again helpers have you covered, look at the active_only
>> parameter for drm_atomic_helper_commit_planes().
>
> virtio-gpu is a bit simplified compared to real hardware, so there
isn't
> really separate plane/crtc state.
>
> Right now the virtual outputs are linked to drm_crtc.  To apply any
> changes I need to lookup the crtc to figure which virtual output should
> be updated.
>
> So, setting active_only should make sure I have a valid crtc pointer on
> plane updates, right?  It probably also skips the disable + enable crtc
> steps on commit?  What happens when outputs are disabled?
>
> Maybe it makes sense to link our virtual outputs to (primary) planes not
> crtcs?
Nah, I just misunderstood your patch. If it's all about finding the
corresponding crtc, then you're all good. I thought there was some
other reason (like the virtual hw getting upset about certain things).
I'll pick up your patch into my nonblocking atomic branch to make sure
virtio isn't accidentally broken. btw can you pls drop an ack or r-b
onto my virtio conversion? I already added your tested-by.

Thanks, Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch