[0/2] drm/atomic: Allow drivers to write their own plane check for async

  Hi,

AMD hardware can do more on the async flip path than just the primary plane, so
to lift up the current restrictions, this patchset allows drivers to write their
own check for planes for async flips.

I'm not sure if adding something new to drm_plane_funcs is the right way to do,
because if we want to expand the other object types (crtc, connector) we would
need to add their own drm_XXX_funcs, so feedbacks are welcome!

	André

André Almeida (2):
  drm/atomic: Allow drivers to write their own plane check for async
    flips
  drm/amdgpu: Implement check_async_props for planes

 .../amd/display/amdgpu_dm/amdgpu_dm_plane.c   | 30 +++++++++
 drivers/gpu/drm/drm_atomic_uapi.c             | 62 ++++++++++++++-----
 include/drm/drm_atomic_uapi.h                 | 12 ++++
 include/drm/drm_plane.h                       |  5 ++
 4 files changed, 92 insertions(+), 17 deletions(-)
  

On Tue, 16 Jan 2024 01:51:57 -0300
André Almeida <andrealmeid@igalia.com> wrote:

> Hi,
> 
> AMD hardware can do more on the async flip path than just the primary plane, so
> to lift up the current restrictions, this patchset allows drivers to write their
> own check for planes for async flips.

Hi,

what's the userspace story for this, how could userspace know it could do more?
What kind of userspace would take advantage of this and in what situations?

Or is this not meant for generic userspace?


Thanks,
pq

> I'm not sure if adding something new to drm_plane_funcs is the right way to do,
> because if we want to expand the other object types (crtc, connector) we would
> need to add their own drm_XXX_funcs, so feedbacks are welcome!
> 
> 	André
> 
> André Almeida (2):
>   drm/atomic: Allow drivers to write their own plane check for async
>     flips
>   drm/amdgpu: Implement check_async_props for planes
> 
>  .../amd/display/amdgpu_dm/amdgpu_dm_plane.c   | 30 +++++++++
>  drivers/gpu/drm/drm_atomic_uapi.c             | 62 ++++++++++++++-----
>  include/drm/drm_atomic_uapi.h                 | 12 ++++
>  include/drm/drm_plane.h                       |  5 ++
>  4 files changed, 92 insertions(+), 17 deletions(-)
>
  

Hi Pekka,

Em 16/01/2024 06:45, Pekka Paalanen escreveu:
> On Tue, 16 Jan 2024 01:51:57 -0300
> André Almeida <andrealmeid@igalia.com> wrote:
> 
>> Hi,
>>
>> AMD hardware can do more on the async flip path than just the primary plane, so
>> to lift up the current restrictions, this patchset allows drivers to write their
>> own check for planes for async flips.
> 
> Hi,
> 
> what's the userspace story for this, how could userspace know it could do more?
> What kind of userspace would take advantage of this and in what situations?
> 
> Or is this not meant for generic userspace?

Sorry, I forgot to document this. So the idea is that userspace will 
query what they can do here with DRM_MODE_ATOMIC_TEST_ONLY calls, 
instead of having capabilities for each prop.

> 
> 
> Thanks,
> pq
> 
>> I'm not sure if adding something new to drm_plane_funcs is the right way to do,
>> because if we want to expand the other object types (crtc, connector) we would
>> need to add their own drm_XXX_funcs, so feedbacks are welcome!
>>
>> 	André
>>
>> André Almeida (2):
>>    drm/atomic: Allow drivers to write their own plane check for async
>>      flips
>>    drm/amdgpu: Implement check_async_props for planes
>>
>>   .../amd/display/amdgpu_dm/amdgpu_dm_plane.c   | 30 +++++++++
>>   drivers/gpu/drm/drm_atomic_uapi.c             | 62 ++++++++++++++-----
>>   include/drm/drm_atomic_uapi.h                 | 12 ++++
>>   include/drm/drm_plane.h                       |  5 ++
>>   4 files changed, 92 insertions(+), 17 deletions(-)
>>
>
  

On Tue, 16 Jan 2024 08:50:59 -0300
André Almeida <andrealmeid@igalia.com> wrote:

> Hi Pekka,
> 
> Em 16/01/2024 06:45, Pekka Paalanen escreveu:
> > On Tue, 16 Jan 2024 01:51:57 -0300
> > André Almeida <andrealmeid@igalia.com> wrote:
> >   
> >> Hi,
> >>
> >> AMD hardware can do more on the async flip path than just the primary plane, so
> >> to lift up the current restrictions, this patchset allows drivers to write their
> >> own check for planes for async flips.  
> > 
> > Hi,
> > 
> > what's the userspace story for this, how could userspace know it could do more?
> > What kind of userspace would take advantage of this and in what situations?
> > 
> > Or is this not meant for generic userspace?  
> 
> Sorry, I forgot to document this. So the idea is that userspace will 
> query what they can do here with DRM_MODE_ATOMIC_TEST_ONLY calls, 
> instead of having capabilities for each prop.

That's the theory, but do you have a practical example?

What other planes and props would one want change in some specific use
case?

Is it just "all or nothing", or would there be room to choose and pick
which props you change and which you don't based on what the driver
supports? If the latter, then relying on TEST_ONLY might be yet another
combinatorial explosion to iterate through.


Thanks,
pq

> >> I'm not sure if adding something new to drm_plane_funcs is the right way to do,
> >> because if we want to expand the other object types (crtc, connector) we would
> >> need to add their own drm_XXX_funcs, so feedbacks are welcome!
> >>
> >> 	André
> >>
> >> André Almeida (2):
> >>    drm/atomic: Allow drivers to write their own plane check for async
> >>      flips
> >>    drm/amdgpu: Implement check_async_props for planes
> >>
> >>   .../amd/display/amdgpu_dm/amdgpu_dm_plane.c   | 30 +++++++++
> >>   drivers/gpu/drm/drm_atomic_uapi.c             | 62 ++++++++++++++-----
> >>   include/drm/drm_atomic_uapi.h                 | 12 ++++
> >>   include/drm/drm_plane.h                       |  5 ++
> >>   4 files changed, 92 insertions(+), 17 deletions(-)
> >>  
> >
  

+ Joshua

Em 16/01/2024 10:14, Pekka Paalanen escreveu:
> On Tue, 16 Jan 2024 08:50:59 -0300
> André Almeida <andrealmeid@igalia.com> wrote:
> 
>> Hi Pekka,
>>
>> Em 16/01/2024 06:45, Pekka Paalanen escreveu:
>>> On Tue, 16 Jan 2024 01:51:57 -0300
>>> André Almeida <andrealmeid@igalia.com> wrote:
>>>    
>>>> Hi,
>>>>
>>>> AMD hardware can do more on the async flip path than just the primary plane, so
>>>> to lift up the current restrictions, this patchset allows drivers to write their
>>>> own check for planes for async flips.
>>>
>>> Hi,
>>>
>>> what's the userspace story for this, how could userspace know it could do more?
>>> What kind of userspace would take advantage of this and in what situations?
>>>
>>> Or is this not meant for generic userspace?
>>
>> Sorry, I forgot to document this. So the idea is that userspace will
>> query what they can do here with DRM_MODE_ATOMIC_TEST_ONLY calls,
>> instead of having capabilities for each prop.
> 
> That's the theory, but do you have a practical example?
> 
> What other planes and props would one want change in some specific use
> case?
> 
> Is it just "all or nothing", or would there be room to choose and pick
> which props you change and which you don't based on what the driver
> supports? If the latter, then relying on TEST_ONLY might be yet another
> combinatorial explosion to iterate through.
> 

That's a good question, maybe Simon, Xaver or Joshua can share how they 
were planning to use this on Gamescope or Kwin.

> 
> Thanks,
> pq
> 
>>>> I'm not sure if adding something new to drm_plane_funcs is the right way to do,
>>>> because if we want to expand the other object types (crtc, connector) we would
>>>> need to add their own drm_XXX_funcs, so feedbacks are welcome!
>>>>
>>>> 	André
>>>>
>>>> André Almeida (2):
>>>>     drm/atomic: Allow drivers to write their own plane check for async
>>>>       flips
>>>>     drm/amdgpu: Implement check_async_props for planes
>>>>
>>>>    .../amd/display/amdgpu_dm/amdgpu_dm_plane.c   | 30 +++++++++
>>>>    drivers/gpu/drm/drm_atomic_uapi.c             | 62 ++++++++++++++-----
>>>>    include/drm/drm_atomic_uapi.h                 | 12 ++++
>>>>    include/drm/drm_plane.h                       |  5 ++
>>>>    4 files changed, 92 insertions(+), 17 deletions(-)
>>>>   
>>>    
>
  

On Tue, 16 Jan 2024 17:10:18 +0100
Xaver Hugl <xaver.hugl@gmail.com> wrote:

> My plan is to require support for IN_FENCE_FD at least. If the driver
> doesn't
> allow tearing with that, then tearing just doesn't happen.

That's an excellent point. I think this is important enough in its own
right, that it should be called out in the patch series.

Is it important enough to be special-cased, e.g. to be always allowed
with async commits?

Now that I think of it, if userspace needs to wait for the in-fence
itself before kicking KMS async, that would defeat much of the async's
point, right? And cases where in-fence is not necessary are so rare
they might not even exist?

So if driver/hardware cannot do IN_FENCE_FD with async, is there any
use of supporting async to begin with?

> For overlay planes though, it depends on how the compositor prioritizes
> things.
> If the compositor prioritizes overlay planes and would like to do tearing
> if possible,
> then this patch works.

Ok, I can see that.

> If the compositor prioritizes tearing and would like to do overlay planes
> if possible,
> it would have to know that switching to synchronous commits for a single
> frame,
> setting up the overlay planes and then switching back to async commits
> works, and
> that can't be figured out with TEST_ONLY commits.

I had to ponder a bit why. So I guess the synchronous commit in between
is because driver/hardware may not be able to enable/disable extra
planes in async, so you need a synchronous commit to set them up, but
afterwards updates can tear.

The comment about Intel needing one more synchronous commit when
switching from sync to async updates comes to mind as well, would that
be a problem?

> So I think having a CAP or immutable plane property to signal that async
> commits
> with overlay and/or cursor planes is supported would be useful.

Async cursor planes a good point, particularly moving them around. I'm
not too informed about the prior/on-going efforts to allow cursor
movement more often than refresh rate, I recall something about
amending atomic commits? How would these interact?

I suppose the kernel still prevents any new async commit while a
previous commit is not finished, so amending commits would still be
necessary for cursor plane motion? Or would it, if you time "big
commits" to finish quickly and then spam async "cursor commits" in the
mean time?


Thanks,
pq

> Am Di., 16. Jan. 2024 um 14:35 Uhr schrieb André Almeida <
> andrealmeid@igalia.com>:  
> 
> > + Joshua
> >
> > Em 16/01/2024 10:14, Pekka Paalanen escreveu:  
> > > On Tue, 16 Jan 2024 08:50:59 -0300
> > > André Almeida <andrealmeid@igalia.com> wrote:
> > >  
> > >> Hi Pekka,
> > >>
> > >> Em 16/01/2024 06:45, Pekka Paalanen escreveu:  
> > >>> On Tue, 16 Jan 2024 01:51:57 -0300
> > >>> André Almeida <andrealmeid@igalia.com> wrote:
> > >>>  
> > >>>> Hi,
> > >>>>
> > >>>> AMD hardware can do more on the async flip path than just the primary  
> > plane, so  
> > >>>> to lift up the current restrictions, this patchset allows drivers to  
> > write their  
> > >>>> own check for planes for async flips.  
> > >>>
> > >>> Hi,
> > >>>
> > >>> what's the userspace story for this, how could userspace know it could  
> > do more?  
> > >>> What kind of userspace would take advantage of this and in what  
> > situations?  
> > >>>
> > >>> Or is this not meant for generic userspace?  
> > >>
> > >> Sorry, I forgot to document this. So the idea is that userspace will
> > >> query what they can do here with DRM_MODE_ATOMIC_TEST_ONLY calls,
> > >> instead of having capabilities for each prop.  
> > >
> > > That's the theory, but do you have a practical example?
> > >
> > > What other planes and props would one want change in some specific use
> > > case?
> > >
> > > Is it just "all or nothing", or would there be room to choose and pick
> > > which props you change and which you don't based on what the driver
> > > supports? If the latter, then relying on TEST_ONLY might be yet another
> > > combinatorial explosion to iterate through.
> > >  
> >
> > That's a good question, maybe Simon, Xaver or Joshua can share how they
> > were planning to use this on Gamescope or Kwin.
> >  
> > >
> > > Thanks,
> > > pq
> > >  
> > >>>> I'm not sure if adding something new to drm_plane_funcs is the right  
> > way to do,  
> > >>>> because if we want to expand the other object types (crtc, connector)  
> > we would  
> > >>>> need to add their own drm_XXX_funcs, so feedbacks are welcome!
> > >>>>
> > >>>>    André
> > >>>>
> > >>>> André Almeida (2):
> > >>>>     drm/atomic: Allow drivers to write their own plane check for async
> > >>>>       flips
> > >>>>     drm/amdgpu: Implement check_async_props for planes
> > >>>>
> > >>>>    .../amd/display/amdgpu_dm/amdgpu_dm_plane.c   | 30 +++++++++
> > >>>>    drivers/gpu/drm/drm_atomic_uapi.c             | 62  
> > ++++++++++++++-----  
> > >>>>    include/drm/drm_atomic_uapi.h                 | 12 ++++
> > >>>>    include/drm/drm_plane.h                       |  5 ++
> > >>>>    4 files changed, 92 insertions(+), 17 deletions(-)
> > >>>>  
> > >>>  
> > >  
> >
  

Am Mi., 17. Jan. 2024 um 09:55 Uhr schrieb Pekka Paalanen <ppaalanen@gmail.com>:
> Is it important enough to be special-cased, e.g. to be always allowed
> with async commits?

I thought so, and sent a patch to dri-devel to make it happen, but
there are some
concerns about untested driver paths.
https://lists.freedesktop.org/archives/dri-devel/2024-January/437511.html

> Now that I think of it, if userspace needs to wait for the in-fence
> itself before kicking KMS async, that would defeat much of the async's
> point, right? And cases where in-fence is not necessary are so rare
> they might not even exist?
>
> So if driver/hardware cannot do IN_FENCE_FD with async, is there any
> use of supporting async to begin with?

KWin never commits a buffer where IN_FENCE_FD would actually delay the
pageflip; it's really only used to disable implicit sync, as there's some edge
cases where it can wrongly delay the pageflip. The waiting for buffers to become
readable on the compositor side isn't really significant in terms of latency.

If hardware doesn't support IN_FENCE_FD with async commits, checking if the
fence is already signaled at commit time would thus still make things work, at
least for KWin.

> > If the compositor prioritizes tearing and would like to do overlay planes
> > if possible,
> > it would have to know that switching to synchronous commits for a single
> > frame,
> > setting up the overlay planes and then switching back to async commits
> > works, and
> > that can't be figured out with TEST_ONLY commits.
>
> I had to ponder a bit why. So I guess the synchronous commit in between
> is because driver/hardware may not be able to enable/disable extra
> planes in async, so you need a synchronous commit to set them up, but
> afterwards updates can tear.

The hardware could be a factor, yes, but I've been thinking more about the API.
With this patchset, the compositor is still only allowed to change a
limited set of
plane properties - but it needs to set at least SRC_X, SRC_Y, CRTC_X etc on
the overlay plane(s) to the correct values before it can only change the allowed
properties again.

> The comment about Intel needing one more synchronous commit when
> switching from sync to async updates comes to mind as well, would that
> be a problem?

With only one synchronous update, the compositor could theoretically mask the
issue by committing it right before vblank; with that one
implicitly-sync "async"
commit though, you'd definitely get one frame without async commits. Having a
flag for a sync-but-then-async-again commit could solve that issue.

In practice I don't think anyone will notice one or two frames without
async commits.
It should be a pretty rare occurance, usually when the game or match
starts or an
overlay gets opened, so I doubt it's worth putting effort in to fix that.

> > So I think having a CAP or immutable plane property to signal that async
> > commits
> > with overlay and/or cursor planes is supported would be useful.
>
> Async cursor planes a good point, particularly moving them around. I'm
> not too informed about the prior/on-going efforts to allow cursor
> movement more often than refresh rate, I recall something about
> amending atomic commits? How would these interact?
>
> I suppose the kernel still prevents any new async commit while a
> previous commit is not finished, so amending commits would still be
> necessary for cursor plane motion? Or would it, if you time "big
> commits" to finish quickly and then spam async "cursor commits" in the
> mean time?

With async commits for cursor planes I'm really only talking about
getting to use
the cursor plane while doing async commits on the primary plane.

FWIW I personally consider the amend stuff mostly solved - KWin does that
internally since a few months ago now, with a separate thread to amend and
even reorder commits in a queue, and only actually commit immediately
before vblank.

>
> Thanks,
> pq
>
> > Am Di., 16. Jan. 2024 um 14:35 Uhr schrieb André Almeida <
> > andrealmeid@igalia.com>:
> >
> > > + Joshua
> > >
> > > Em 16/01/2024 10:14, Pekka Paalanen escreveu:
> > > > On Tue, 16 Jan 2024 08:50:59 -0300
> > > > André Almeida <andrealmeid@igalia.com> wrote:
> > > >
> > > >> Hi Pekka,
> > > >>
> > > >> Em 16/01/2024 06:45, Pekka Paalanen escreveu:
> > > >>> On Tue, 16 Jan 2024 01:51:57 -0300
> > > >>> André Almeida <andrealmeid@igalia.com> wrote:
> > > >>>
> > > >>>> Hi,
> > > >>>>
> > > >>>> AMD hardware can do more on the async flip path than just the primary
> > > plane, so
> > > >>>> to lift up the current restrictions, this patchset allows drivers to
> > > write their
> > > >>>> own check for planes for async flips.
> > > >>>
> > > >>> Hi,
> > > >>>
> > > >>> what's the userspace story for this, how could userspace know it could
> > > do more?
> > > >>> What kind of userspace would take advantage of this and in what
> > > situations?
> > > >>>
> > > >>> Or is this not meant for generic userspace?
> > > >>
> > > >> Sorry, I forgot to document this. So the idea is that userspace will
> > > >> query what they can do here with DRM_MODE_ATOMIC_TEST_ONLY calls,
> > > >> instead of having capabilities for each prop.
> > > >
> > > > That's the theory, but do you have a practical example?
> > > >
> > > > What other planes and props would one want change in some specific use
> > > > case?
> > > >
> > > > Is it just "all or nothing", or would there be room to choose and pick
> > > > which props you change and which you don't based on what the driver
> > > > supports? If the latter, then relying on TEST_ONLY might be yet another
> > > > combinatorial explosion to iterate through.
> > > >
> > >
> > > That's a good question, maybe Simon, Xaver or Joshua can share how they
> > > were planning to use this on Gamescope or Kwin.
> > >
> > > >
> > > > Thanks,
> > > > pq
> > > >
> > > >>>> I'm not sure if adding something new to drm_plane_funcs is the right
> > > way to do,
> > > >>>> because if we want to expand the other object types (crtc, connector)
> > > we would
> > > >>>> need to add their own drm_XXX_funcs, so feedbacks are welcome!
> > > >>>>
> > > >>>>    André
> > > >>>>
> > > >>>> André Almeida (2):
> > > >>>>     drm/atomic: Allow drivers to write their own plane check for async
> > > >>>>       flips
> > > >>>>     drm/amdgpu: Implement check_async_props for planes
> > > >>>>
> > > >>>>    .../amd/display/amdgpu_dm/amdgpu_dm_plane.c   | 30 +++++++++
> > > >>>>    drivers/gpu/drm/drm_atomic_uapi.c             | 62
> > > ++++++++++++++-----
> > > >>>>    include/drm/drm_atomic_uapi.h                 | 12 ++++
> > > >>>>    include/drm/drm_plane.h                       |  5 ++
> > > >>>>    4 files changed, 92 insertions(+), 17 deletions(-)
> > > >>>>
> > > >>>
> > > >
> > >
>
  

On 2024-01-17 13:57, Xaver Hugl wrote:
> Am Mi., 17. Jan. 2024 um 09:55 Uhr schrieb Pekka Paalanen <ppaalanen@gmail.com>:
>> Is it important enough to be special-cased, e.g. to be always allowed
>> with async commits?
> 
> I thought so, and sent a patch to dri-devel to make it happen, but
> there are some
> concerns about untested driver paths.
> https://lists.freedesktop.org/archives/dri-devel/2024-January/437511.html
> 
>> Now that I think of it, if userspace needs to wait for the in-fence
>> itself before kicking KMS async, that would defeat much of the async's
>> point, right? And cases where in-fence is not necessary are so rare
>> they might not even exist?
>>
>> So if driver/hardware cannot do IN_FENCE_FD with async, is there any
>> use of supporting async to begin with?
> 
> KWin never commits a buffer where IN_FENCE_FD would actually delay the
> pageflip; it's really only used to disable implicit sync, as there's some edge
> cases where it can wrongly delay the pageflip. The waiting for buffers to become
> readable on the compositor side isn't really significant in terms of latency.
> 
> If hardware doesn't support IN_FENCE_FD with async commits, checking if the
> fence is already signaled at commit time would thus still make things work, at
> least for KWin.

That's how IN_FENCE_FD (and implicit sync) is handled anyway, in common code: It waits for all fences to signal before calling into the driver to commit the atomic commit.

I can't see why this wouldn't work with async commits, the same as with synchronous ones, with any driver.