Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1615103 > unrolled thread

Re: [RFC 00/10] V4L2 explicit synchronization support

Back to article view | Back to linux.kernel

This discussion starts older than the indexed window; earlier articles aren't shown. The article labeled Started by below is the oldest one visible, not the original post.

Contents

  Re: [RFC 00/10] V4L2 explicit synchronization support Mauro Carvalho Chehab <mchehab@osg.samsung.com> - 2017-04-03 13:20 +0200
    Re: [RFC 00/10] V4L2 explicit synchronization support Javier Martinez Canillas <javier@osg.samsung.com> - 2017-04-03 21:50 +0200
      Re: [RFC 00/10] V4L2 explicit synchronization support Shuah Khan <shuahkhan@gmail.com> - 2017-04-03 22:50 +0200
      Re: [RFC 00/10] V4L2 explicit synchronization support Gustavo Padovan <gustavo.padovan@collabora.com> - 2017-04-05 17:20 +0200
        Re: [RFC 00/10] V4L2 explicit synchronization support Javier Martinez Canillas <javier@osg.samsung.com> - 2017-04-05 19:20 +0200
          Re: [RFC 00/10] V4L2 explicit synchronization support Gustavo Padovan <gustavo.padovan@collabora.com> - 2017-04-06 16:10 +0200
            Re: [RFC 00/10] V4L2 explicit synchronization support Javier Martinez Canillas <javier@osg.samsung.com> - 2017-04-06 16:40 +0200

#1615103 — Re: [RFC 00/10] V4L2 explicit synchronization support

Hi Gustavo,

Em Mon, 13 Mar 2017 16:20:25 -0300
Gustavo Padovan <gustavo@padovan.org> escreveu:

> From: Gustavo Padovan <gustavo.padovan@collabora.com>
> 
> Hi,
> 
> This RFC adds support for Explicit Synchronization of shared buffers in V4L2.
> It uses the Sync File Framework[1] as vector to communicate the fences
> between kernel and userspace.

Thanks for your work!

I looked on your patchset, and I didn't notice anything really weird
there. So, instead on reviewing patch per patch, I prefer to discuss
about the requirements and API, as depending on it, the code base will
change a lot.

I'd like to do some tests with it on devices with mem2mem drivers.
My plan is to use an Exynos board for such thing, but I guess that
the DRM driver for it currently doesn't. I'm seeing internally if someone
could be sure that Exynos driver upstream will become ready for such
tests.

Javier wrote some patches last year meant to implement implicit
fences support. What we noticed is that, while his mechanism worked
fine for pure capture and pure output devices, when we added a mem2mem
device, on a DMABUF+fences pipeline, e. g.:

	sensor -> [m2m] -> DRM

End everything using fences/DMABUF, the fences mechanism caused dead
locks on existing userspace apps.

A m2m device has both capture and output devnodes. Both should be
queued/dequeued. The capture queue is synchronized internally at the
driver with the output buffer[1].

[1] The names here are counter-intuitive: "capture" is a devnode
where userspace receives a video stream; "output" is a devnode where
userspace feeds a video stream.

The problem is that adding implicit fences changed the behavior of
the ioctls, causing gstreamer to wait forever for buffers to be ready.

I suspect that, even with explicit fences, the behavior of Q/DQ
will be incompatible with the current behavior (or will require some
dirty hacks to make it identical). 

So, IMHO, the best would be to use a new set of ioctls, when fences are
used (like VIDIOC_QFENCE/DQFENCE).

> 
> I'm sending this to start the discussion on the best approach to implement
> Explicit Synchronization, please check the TODO/OPEN section below.
> 
> Explicit Synchronization allows us to control the synchronization of
> shared buffers from userspace by passing fences to the kernel and/or 
> receiving them from the the kernel.
> 
> Fences passed to the kernel are named in-fences and the kernel should wait
> them to signal before using the buffer. On the other side, the kernel creates
> out-fences for every buffer it receives from userspace. This fence is sent back
> to userspace and it will signal when the capture, for example, has finished.
> 
> Signalling an out-fence in V4L2 would mean that the job on the buffer is done
> and the buffer can be used by other drivers.
> 
> Current RFC implementation
> --------------------------
> 
> The current implementation is not intended to be more than a PoC to start
> the discussion on how Explicit Synchronization should be supported in V4L2.
> 
> The first patch proposes an userspace API for fences, then on patch 2
> we prepare to the addition of in-fences in patch 3, by introducing the
> infrastructure on vb2 to wait on an in-fence signal before queueing the buffer
> in the driver.
> 
> Patch 4 fix uvc v4l2 event handling and patch 5 configure q->dev for vivid
> drivers to enable to subscribe and dequeue events on it.
> 
> Patches 6-7 enables support to notify BUF_QUEUED events, i.e., let userspace
> know that particular buffer was enqueued in the driver. This is needed,
> because we return the out-fence fd as an out argument in QBUF, but at the time
> it returns we don't know to which buffer the fence will be attached thus
> the BUF_QUEUED event tells which buffer is associated to the fence received in
> QBUF by userspace.
> 
> Patches 8 and 9 add more fence infrastructure to support out-fences and finally
> patch 10 adds support to out-fences.
> 
> TODO/OPEN:
> ----------
> 
> * For this first implementation we will keep the ordering of the buffers queued
> in videobuf2, that means we will only enqueue buffer whose fence was signalled
> if that buffer is the first one in the queue. Otherwise it has to wait until it
> is the first one. This is not implmented yet. Later we could create a flag to
> allow unordered queing in the drivers from vb2 if needed.

The V4L2 spec doesn't warrant that the buffers will be dequeued at the
queue order.

In practice, however, most drivers will not reorder. Yet, mem2mem codec 
drivers may reorder the buffers at the output, as the luminance information
(Y) usually comes first on JPEG/MPEG-like formats.

> * Should we have out-fences per-buffer or per-plane? or both? In this RFC, for
> simplicity they are per-buffer, but Mauro and Javier raised the option of
> doing per-plane fences. That could benefit mem2mem and V4L2 <-> GPU operation
> at least on cases when we have Capture hw that releases the Y frame before the
> other frames for example. When using V4L2 per-plane out-fences to communicate
> with KMS they would need to be merged together as currently the DRM Plane
> interface only supports one fence per DRM Plane.

That's another advantage of using a new set of ioctls for queues: with that,
queuing/dequeing per plane will be easier. On codec drivers, doing it per
plane could bring performance improvements.

> In-fences should be per-buffer as the DRM only has per-buffer fences, but
> in case of mem2mem operations per-plane fences might be useful?
> 
> So should we have both ways, per-plane and per-buffer, or just one of them
> for now?

The API should be flexible enough to support both usecases. We could
implement just per-buffer in the beginning, but, on such case, we
should deploy an API that will allow to later add per-plane fences
without breaking userspace.

So, I prefer that, on multiplane fences, we have one fence per plane,
even if, at the first implementation, all fences will be released
at the same time, when the buffer is fully filled. That would allow
us to later improve it, without changing userspace.

> * other open topics are how to deal with hw-fences and Request API.

Let's not mix issues here. Request API is already complex enough
without explicit fences. The same is true for explicit fences: it is 
also complex to add support for it with multi-planes and a not
ordered DQ. 

So, my suggestion here is to not delay Request API due to fences,
nor to delay fences due to Request API, letting them to be merged
on different Kernel releases. When both features got added, someone
can work on a patchset that would allow using Request API with fences.

-- 
Thanks,
Mauro

[toc] | [next] | [standalone]

#1615516

Hello Mauro and Gustavo,

On 04/03/2017 07:16 AM, Mauro Carvalho Chehab wrote:
> Hi Gustavo,
> 
> Em Mon, 13 Mar 2017 16:20:25 -0300
> Gustavo Padovan <gustavo@padovan.org> escreveu:
> 
>> From: Gustavo Padovan <gustavo.padovan@collabora.com>
>>
>> Hi,
>>
>> This RFC adds support for Explicit Synchronization of shared buffers in V4L2.
>> It uses the Sync File Framework[1] as vector to communicate the fences
>> between kernel and userspace.
> 
> Thanks for your work!
> 
> I looked on your patchset, and I didn't notice anything really weird
> there. So, instead on reviewing patch per patch, I prefer to discuss
> about the requirements and API, as depending on it, the code base will
> change a lot.
>

Agree that's better to first set on an uAPI and then implement based on that.
 
> I'd like to do some tests with it on devices with mem2mem drivers.
> My plan is to use an Exynos board for such thing, but I guess that
> the DRM driver for it currently doesn't. I'm seeing internally if someone
> could be sure that Exynos driver upstream will become ready for such
> tests.
>

Not sure if you should try to do testing before agreeing on an uAPI and
implementation.

> Javier wrote some patches last year meant to implement implicit
> fences support. What we noticed is that, while his mechanism worked
> fine for pure capture and pure output devices, when we added a mem2mem
> device, on a DMABUF+fences pipeline, e. g.:
> 
> 	sensor -> [m2m] -> DRM
> 
> End everything using fences/DMABUF, the fences mechanism caused dead
> locks on existing userspace apps.
>
> A m2m device has both capture and output devnodes. Both should be
> queued/dequeued. The capture queue is synchronized internally at the
> driver with the output buffer[1].
> 
> [1] The names here are counter-intuitive: "capture" is a devnode
> where userspace receives a video stream; "output" is a devnode where
> userspace feeds a video stream.
> 
> The problem is that adding implicit fences changed the behavior of
> the ioctls, causing gstreamer to wait forever for buffers to be ready.
>

The problem was related to trying to make user-space unaware of the implicit
fences support, and so it tried to QBUF a buffer that had already a pending
fence. A workaround was to block the second QBUF ioctl if the buffer had a
pending fence, but this caused the mentioned deadlock since GStreamer wasn't
expecting the QBUF ioctl to block.

> I suspect that, even with explicit fences, the behavior of Q/DQ
> will be incompatible with the current behavior (or will require some
> dirty hacks to make it identical). 
>
> So, IMHO, the best would be to use a new set of ioctls, when fences are
> used (like VIDIOC_QFENCE/DQFENCE).
> 

For explicit you can check if there's an input-fence so is different than
implicit, but still I agree that it would be better to have specific ioctls.

>>
>> I'm sending this to start the discussion on the best approach to implement
>> Explicit Synchronization, please check the TODO/OPEN section below.
>>
>> Explicit Synchronization allows us to control the synchronization of
>> shared buffers from userspace by passing fences to the kernel and/or 
>> receiving them from the the kernel.
>>
>> Fences passed to the kernel are named in-fences and the kernel should wait
>> them to signal before using the buffer. On the other side, the kernel creates
>> out-fences for every buffer it receives from userspace. This fence is sent back
>> to userspace and it will signal when the capture, for example, has finished.
>>
>> Signalling an out-fence in V4L2 would mean that the job on the buffer is done
>> and the buffer can be used by other drivers.
>>
>> Current RFC implementation
>> --------------------------
>>
>> The current implementation is not intended to be more than a PoC to start
>> the discussion on how Explicit Synchronization should be supported in V4L2.
>>
>> The first patch proposes an userspace API for fences, then on patch 2
>> we prepare to the addition of in-fences in patch 3, by introducing the
>> infrastructure on vb2 to wait on an in-fence signal before queueing the buffer
>> in the driver.
>>
>> Patch 4 fix uvc v4l2 event handling and patch 5 configure q->dev for vivid
>> drivers to enable to subscribe and dequeue events on it.
>>
>> Patches 6-7 enables support to notify BUF_QUEUED events, i.e., let userspace
>> know that particular buffer was enqueued in the driver. This is needed,
>> because we return the out-fence fd as an out argument in QBUF, but at the time
>> it returns we don't know to which buffer the fence will be attached thus
>> the BUF_QUEUED event tells which buffer is associated to the fence received in
>> QBUF by userspace.
>>
>> Patches 8 and 9 add more fence infrastructure to support out-fences and finally
>> patch 10 adds support to out-fences.
>>
>> TODO/OPEN:
>> ----------
>>
>> * For this first implementation we will keep the ordering of the buffers queued
>> in videobuf2, that means we will only enqueue buffer whose fence was signalled
>> if that buffer is the first one in the queue. Otherwise it has to wait until it
>> is the first one. This is not implmented yet. Later we could create a flag to
>> allow unordered queing in the drivers from vb2 if needed.
> 
> The V4L2 spec doesn't warrant that the buffers will be dequeued at the
> queue order.
> 
> In practice, however, most drivers will not reorder. Yet, mem2mem codec 
> drivers may reorder the buffers at the output, as the luminance information
> (Y) usually comes first on JPEG/MPEG-like formats.
> 
>> * Should we have out-fences per-buffer or per-plane? or both? In this RFC, for
>> simplicity they are per-buffer, but Mauro and Javier raised the option of
>> doing per-plane fences. That could benefit mem2mem and V4L2 <-> GPU operation
>> at least on cases when we have Capture hw that releases the Y frame before the
>> other frames for example. When using V4L2 per-plane out-fences to communicate
>> with KMS they would need to be merged together as currently the DRM Plane
>> interface only supports one fence per DRM Plane.
> 
> That's another advantage of using a new set of ioctls for queues: with that,
> queuing/dequeing per plane will be easier. On codec drivers, doing it per
> plane could bring performance improvements.
>

You don't really need to Q/DQ on a per plane basis AFAICT. Since on QBUF you
can get a set of out-fences that can be passed to the other driver and so it
should be able to wait per fence.

>> In-fences should be per-buffer as the DRM only has per-buffer fences, but

I'm not that familiar with DRM, but I thought DRM fences was also per plane
and not per buffer.

How this works without fences? For V4L2 there's a dma-buf fd per plane and
so I was expecting the DRM API to also import a dma-buf fd per DRM plane.

I only have access to an Exynos board whose display controller supports
single plane formats, so I don't know how this works for multi planar.

>> in case of mem2mem operations per-plane fences might be useful?
>>
>> So should we have both ways, per-plane and per-buffer, or just one of them
>> for now?
>
> The API should be flexible enough to support both usecases. We could
> implement just per-buffer in the beginning, but, on such case, we
> should deploy an API that will allow to later add per-plane fences
> without breaking userspace.
>
> So, I prefer that, on multiplane fences, we have one fence per plane,
> even if, at the first implementation, all fences will be released
> at the same time, when the buffer is fully filled. That would allow
> us to later improve it, without changing userspace.

It's true that vb2 can't currently signal fences per plane since the interface
between vb2 and drivers is per vb2_buffer. But the uAPI shouldn't be restricted
by this implementation detail (that can be changed) and should support per plane
fences IMHO.

That's for example the case with the V4L2 dma-buf API. There is a dma-buf fd per
plane, and internally for vb2 single planar buffers use the dma-buf associated
with plane 0.

Now when mentioning this, I noticed that in your implementation the fences are
not associated with a dma-buf. I thought the idea was for the fences to be
associated with a dma-buf's reservation object. If we do that, then fences will
be per fence since the dma-buf/reservation objet are also per fence in v4l2/vb2.

> 
>> * other open topics are how to deal with hw-fences and Request API.
> 
> Let's not mix issues here. Request API is already complex enough
> without explicit fences. The same is true for explicit fences: it is 
> also complex to add support for it with multi-planes and a not
> ordered DQ. 
> 
> So, my suggestion here is to not delay Request API due to fences,
> nor to delay fences due to Request API, letting them to be merged
> on different Kernel releases. When both features got added, someone
> can work on a patchset that would allow using Request API with fences.
> 

Best regards,
-- 
Javier Martinez Canillas
Open Source Group
Samsung Research America

[toc] | [prev] | [next] | [standalone]

#1615549

Hi Gustavo,

On Mon, Apr 3, 2017 at 1:46 PM, Javier Martinez Canillas
<javier@osg.samsung.com> wrote:
> Hello Mauro and Gustavo,
>
> On 04/03/2017 07:16 AM, Mauro Carvalho Chehab wrote:
>> Hi Gustavo,
>>
>> Em Mon, 13 Mar 2017 16:20:25 -0300
>> Gustavo Padovan <gustavo@padovan.org> escreveu:
>>
>>> From: Gustavo Padovan <gustavo.padovan@collabora.com>
>>>
>>> Hi,
>>>
>>> This RFC adds support for Explicit Synchronization of shared buffers in V4L2.
>>> It uses the Sync File Framework[1] as vector to communicate the fences
>>> between kernel and userspace.
>>
>> Thanks for your work!
>>
>> I looked on your patchset, and I didn't notice anything really weird
>> there. So, instead on reviewing patch per patch, I prefer to discuss
>> about the requirements and API, as depending on it, the code base will
>> change a lot.
>>
>
> Agree that's better to first set on an uAPI and then implement based on that.

Yes. Agreeing on uAPI first would work well.

>
>> I'd like to do some tests with it on devices with mem2mem drivers.
>> My plan is to use an Exynos board for such thing, but I guess that
>> the DRM driver for it currently doesn't. I'm seeing internally if someone
>> could be sure that Exynos driver upstream will become ready for such
>> tests.
>>
>
> Not sure if you should try to do testing before agreeing on an uAPI and
> implementation.

Running some tests might give you a better feel for m2m - export buf - drm
cases without fences on exynos. This could help understand the performance
gains with fences.

>
>> Javier wrote some patches last year meant to implement implicit
>> fences support. What we noticed is that, while his mechanism worked
>> fine for pure capture and pure output devices, when we added a mem2mem
>> device, on a DMABUF+fences pipeline, e. g.:
>>
>>       sensor -> [m2m] -> DRM
>>
>> End everything using fences/DMABUF, the fences mechanism caused dead
>> locks on existing userspace apps.
>>
>> A m2m device has both capture and output devnodes. Both should be
>> queued/dequeued. The capture queue is synchronized internally at the
>> driver with the output buffer[1].
>>
>> [1] The names here are counter-intuitive: "capture" is a devnode
>> where userspace receives a video stream; "output" is a devnode where
>> userspace feeds a video stream.
>>
>> The problem is that adding implicit fences changed the behavior of
>> the ioctls, causing gstreamer to wait forever for buffers to be ready.
>>
>
> The problem was related to trying to make user-space unaware of the implicit
> fences support, and so it tried to QBUF a buffer that had already a pending
> fence. A workaround was to block the second QBUF ioctl if the buffer had a
> pending fence, but this caused the mentioned deadlock since GStreamer wasn't
> expecting the QBUF ioctl to block.
>
>> I suspect that, even with explicit fences, the behavior of Q/DQ
>> will be incompatible with the current behavior (or will require some
>> dirty hacks to make it identical).

One big difference between implicit and explicit is that use-space is aware
of fences in the case of explicit. Can that knowledge be helpful in avoiding
the the problems we have seen with explicit?

>>
>> So, IMHO, the best would be to use a new set of ioctls, when fences are
>> used (like VIDIOC_QFENCE/DQFENCE).
>>
>
> For explicit you can check if there's an input-fence so is different than
> implicit, but still I agree that it would be better to have specific ioctls.
>

It would be nice if we can avoid adding more ioctls. As I mentioned earlier,
user-space is aware that fences are in use. Can we key off of that and make
it a queue property to be used to change the user-space action for fence vs.
no fence case?

<snip.>

thanks,
-- Shuah

[toc] | [prev] | [next] | [standalone]

#1617036

2017-04-03 Javier Martinez Canillas <javier@osg.samsung.com>:

> Hello Mauro and Gustavo,
> 
> On 04/03/2017 07:16 AM, Mauro Carvalho Chehab wrote:
> > Hi Gustavo,
> > 
> > Em Mon, 13 Mar 2017 16:20:25 -0300
> > Gustavo Padovan <gustavo@padovan.org> escreveu:
> > 
> >> From: Gustavo Padovan <gustavo.padovan@collabora.com>
> >>
> >> Hi,
> >>
> >> This RFC adds support for Explicit Synchronization of shared buffers in V4L2.
> >> It uses the Sync File Framework[1] as vector to communicate the fences
> >> between kernel and userspace.
> > 
> > Thanks for your work!
> > 
> > I looked on your patchset, and I didn't notice anything really weird
> > there. So, instead on reviewing patch per patch, I prefer to discuss
> > about the requirements and API, as depending on it, the code base will
> > change a lot.
> >
> 
> Agree that's better to first set on an uAPI and then implement based on that.
>  
> > I'd like to do some tests with it on devices with mem2mem drivers.
> > My plan is to use an Exynos board for such thing, but I guess that
> > the DRM driver for it currently doesn't. I'm seeing internally if someone
> > could be sure that Exynos driver upstream will become ready for such
> > tests.
> >
> 
> Not sure if you should try to do testing before agreeing on an uAPI and
> implementation.
> 
> > Javier wrote some patches last year meant to implement implicit
> > fences support. What we noticed is that, while his mechanism worked
> > fine for pure capture and pure output devices, when we added a mem2mem
> > device, on a DMABUF+fences pipeline, e. g.:
> > 
> > 	sensor -> [m2m] -> DRM
> > 
> > End everything using fences/DMABUF, the fences mechanism caused dead
> > locks on existing userspace apps.
> >
> > A m2m device has both capture and output devnodes. Both should be
> > queued/dequeued. The capture queue is synchronized internally at the
> > driver with the output buffer[1].
> > 
> > [1] The names here are counter-intuitive: "capture" is a devnode
> > where userspace receives a video stream; "output" is a devnode where
> > userspace feeds a video stream.
> > 
> > The problem is that adding implicit fences changed the behavior of
> > the ioctls, causing gstreamer to wait forever for buffers to be ready.
> >
> 
> The problem was related to trying to make user-space unaware of the implicit
> fences support, and so it tried to QBUF a buffer that had already a pending
> fence. A workaround was to block the second QBUF ioctl if the buffer had a
> pending fence, but this caused the mentioned deadlock since GStreamer wasn't
> expecting the QBUF ioctl to block.
> 
> > I suspect that, even with explicit fences, the behavior of Q/DQ
> > will be incompatible with the current behavior (or will require some
> > dirty hacks to make it identical). 

For QBUF the only difference is that we set flags for fences and pass
and receives in and out fences. For DQBUF the behavior is exactly the
same. What incompatibles or hacks do you see?

I had the expectation that the flags would be for userspace to learn
about any different behavior.

> >
> > So, IMHO, the best would be to use a new set of ioctls, when fences are
> > used (like VIDIOC_QFENCE/DQFENCE).
> > 
> 
> For explicit you can check if there's an input-fence so is different than
> implicit, but still I agree that it would be better to have specific ioctls.

I'm pretty new to v4l2 so I don't know all use cases yet, but what I
thought was to just add extra flags to QBUF to mark when using fences
instead of having userspace  to setup completely new ioctls for fences.
The burden for userspace should be smaller with flags.

> 
> >>
> >> I'm sending this to start the discussion on the best approach to implement
> >> Explicit Synchronization, please check the TODO/OPEN section below.
> >>
> >> Explicit Synchronization allows us to control the synchronization of
> >> shared buffers from userspace by passing fences to the kernel and/or 
> >> receiving them from the the kernel.
> >>
> >> Fences passed to the kernel are named in-fences and the kernel should wait
> >> them to signal before using the buffer. On the other side, the kernel creates
> >> out-fences for every buffer it receives from userspace. This fence is sent back
> >> to userspace and it will signal when the capture, for example, has finished.
> >>
> >> Signalling an out-fence in V4L2 would mean that the job on the buffer is done
> >> and the buffer can be used by other drivers.
> >>
> >> Current RFC implementation
> >> --------------------------
> >>
> >> The current implementation is not intended to be more than a PoC to start
> >> the discussion on how Explicit Synchronization should be supported in V4L2.
> >>
> >> The first patch proposes an userspace API for fences, then on patch 2
> >> we prepare to the addition of in-fences in patch 3, by introducing the
> >> infrastructure on vb2 to wait on an in-fence signal before queueing the buffer
> >> in the driver.
> >>
> >> Patch 4 fix uvc v4l2 event handling and patch 5 configure q->dev for vivid
> >> drivers to enable to subscribe and dequeue events on it.
> >>
> >> Patches 6-7 enables support to notify BUF_QUEUED events, i.e., let userspace
> >> know that particular buffer was enqueued in the driver. This is needed,
> >> because we return the out-fence fd as an out argument in QBUF, but at the time
> >> it returns we don't know to which buffer the fence will be attached thus
> >> the BUF_QUEUED event tells which buffer is associated to the fence received in
> >> QBUF by userspace.
> >>
> >> Patches 8 and 9 add more fence infrastructure to support out-fences and finally
> >> patch 10 adds support to out-fences.
> >>
> >> TODO/OPEN:
> >> ----------
> >>
> >> * For this first implementation we will keep the ordering of the buffers queued
> >> in videobuf2, that means we will only enqueue buffer whose fence was signalled
> >> if that buffer is the first one in the queue. Otherwise it has to wait until it
> >> is the first one. This is not implmented yet. Later we could create a flag to
> >> allow unordered queing in the drivers from vb2 if needed.
> > 
> > The V4L2 spec doesn't warrant that the buffers will be dequeued at the
> > queue order.
> > 
> > In practice, however, most drivers will not reorder. Yet, mem2mem codec 
> > drivers may reorder the buffers at the output, as the luminance information
> > (Y) usually comes first on JPEG/MPEG-like formats.
> > 
> >> * Should we have out-fences per-buffer or per-plane? or both? In this RFC, for
> >> simplicity they are per-buffer, but Mauro and Javier raised the option of
> >> doing per-plane fences. That could benefit mem2mem and V4L2 <-> GPU operation
> >> at least on cases when we have Capture hw that releases the Y frame before the
> >> other frames for example. When using V4L2 per-plane out-fences to communicate
> >> with KMS they would need to be merged together as currently the DRM Plane
> >> interface only supports one fence per DRM Plane.
> > 
> > That's another advantage of using a new set of ioctls for queues: with that,
> > queuing/dequeing per plane will be easier. On codec drivers, doing it per
> > plane could bring performance improvements.
> >
> 
> You don't really need to Q/DQ on a per plane basis AFAICT. Since on QBUF you
> can get a set of out-fences that can be passed to the other driver and so it
> should be able to wait per fence.
> 
> >> In-fences should be per-buffer as the DRM only has per-buffer fences, but
> 
> I'm not that familiar with DRM, but I thought DRM fences was also per plane
> and not per buffer.

DRM plane is a different thing, its a representation of a region on the
screen and there is only one buffer for each DRM plane. 

One of the questions I raised was: how to match V4L2 per-plane fences to
DRM per-buffer fences?

> 
> How this works without fences? For V4L2 there's a dma-buf fd per plane and
> so I was expecting the DRM API to also import a dma-buf fd per DRM plane.

Yes. It should do something similar behind the Framebuffer abstraction.

> 
> I only have access to an Exynos board whose display controller supports
> single plane formats, so I don't know how this works for multi planar.
> 
> >> in case of mem2mem operations per-plane fences might be useful?
> >>
> >> So should we have both ways, per-plane and per-buffer, or just one of them
> >> for now?
> >
> > The API should be flexible enough to support both usecases. We could
> > implement just per-buffer in the beginning, but, on such case, we
> > should deploy an API that will allow to later add per-plane fences
> > without breaking userspace.

I believe we can just extend the per-plane parts of QBUF for their
fences. We could even use plane[0] for the per-buffer case.

> >
> > So, I prefer that, on multiplane fences, we have one fence per plane,
> > even if, at the first implementation, all fences will be released
> > at the same time, when the buffer is fully filled. That would allow
> > us to later improve it, without changing userspace.
> 
> It's true that vb2 can't currently signal fences per plane since the interface
> between vb2 and drivers is per vb2_buffer. But the uAPI shouldn't be restricted
> by this implementation detail (that can be changed) and should support per plane
> fences IMHO.
> 
> That's for example the case with the V4L2 dma-buf API. There is a dma-buf fd per
> plane, and internally for vb2 single planar buffers use the dma-buf associated
> with plane 0.
> 
> Now when mentioning this, I noticed that in your implementation the fences are
> not associated with a dma-buf. I thought the idea was for the fences to be
> associated with a dma-buf's reservation object. If we do that, then fences will
> be per fence since the dma-buf/reservation objet are also per fence in v4l2/vb2.

Can you explain what you were thinking on the relation between fences
and reservation objects? Not sure I follow.

Gustavo

[toc] | [prev] | [next] | [standalone]

#1617197

Hello Gustavo,

On 04/05/2017 11:09 AM, Gustavo Padovan wrote:
> 2017-04-03 Javier Martinez Canillas <javier@osg.samsung.com>:
> 
>> Hello Mauro and Gustavo,
>>
>> On 04/03/2017 07:16 AM, Mauro Carvalho Chehab wrote:
>>> Hi Gustavo,
>>>
>>> Em Mon, 13 Mar 2017 16:20:25 -0300
>>> Gustavo Padovan <gustavo@padovan.org> escreveu:
>>>
>>>> From: Gustavo Padovan <gustavo.padovan@collabora.com>
>>>>
>>>> Hi,
>>>>
>>>> This RFC adds support for Explicit Synchronization of shared buffers in V4L2.
>>>> It uses the Sync File Framework[1] as vector to communicate the fences
>>>> between kernel and userspace.
>>>
>>> Thanks for your work!
>>>
>>> I looked on your patchset, and I didn't notice anything really weird
>>> there. So, instead on reviewing patch per patch, I prefer to discuss
>>> about the requirements and API, as depending on it, the code base will
>>> change a lot.
>>>
>>
>> Agree that's better to first set on an uAPI and then implement based on that.
>>  
>>> I'd like to do some tests with it on devices with mem2mem drivers.
>>> My plan is to use an Exynos board for such thing, but I guess that
>>> the DRM driver for it currently doesn't. I'm seeing internally if someone
>>> could be sure that Exynos driver upstream will become ready for such
>>> tests.
>>>
>>
>> Not sure if you should try to do testing before agreeing on an uAPI and
>> implementation.
>>
>>> Javier wrote some patches last year meant to implement implicit
>>> fences support. What we noticed is that, while his mechanism worked
>>> fine for pure capture and pure output devices, when we added a mem2mem
>>> device, on a DMABUF+fences pipeline, e. g.:
>>>
>>> 	sensor -> [m2m] -> DRM
>>>
>>> End everything using fences/DMABUF, the fences mechanism caused dead
>>> locks on existing userspace apps.
>>>
>>> A m2m device has both capture and output devnodes. Both should be
>>> queued/dequeued. The capture queue is synchronized internally at the
>>> driver with the output buffer[1].
>>>
>>> [1] The names here are counter-intuitive: "capture" is a devnode
>>> where userspace receives a video stream; "output" is a devnode where
>>> userspace feeds a video stream.
>>>
>>> The problem is that adding implicit fences changed the behavior of
>>> the ioctls, causing gstreamer to wait forever for buffers to be ready.
>>>
>>
>> The problem was related to trying to make user-space unaware of the implicit
>> fences support, and so it tried to QBUF a buffer that had already a pending
>> fence. A workaround was to block the second QBUF ioctl if the buffer had a
>> pending fence, but this caused the mentioned deadlock since GStreamer wasn't
>> expecting the QBUF ioctl to block.
>>
>>> I suspect that, even with explicit fences, the behavior of Q/DQ
>>> will be incompatible with the current behavior (or will require some
>>> dirty hacks to make it identical). 
> 
> For QBUF the only difference is that we set flags for fences and pass
> and receives in and out fences. For DQBUF the behavior is exactly the
> same. What incompatibles or hacks do you see?
> 
> I had the expectation that the flags would be for userspace to learn
> about any different behavior.
> 
>>>
>>> So, IMHO, the best would be to use a new set of ioctls, when fences are
>>> used (like VIDIOC_QFENCE/DQFENCE).
>>>
>>
>> For explicit you can check if there's an input-fence so is different than
>> implicit, but still I agree that it would be better to have specific ioctls.
> 
> I'm pretty new to v4l2 so I don't know all use cases yet, but what I
> thought was to just add extra flags to QBUF to mark when using fences
> instead of having userspace  to setup completely new ioctls for fences.
> The burden for userspace should be smaller with flags.
>

Yes, you are right. So I guess its better indeed to just extend the current
ioctls as you propose and only move to new ones if really needed.

>>
>>>>
>>>> I'm sending this to start the discussion on the best approach to implement
>>>> Explicit Synchronization, please check the TODO/OPEN section below.
>>>>
>>>> Explicit Synchronization allows us to control the synchronization of
>>>> shared buffers from userspace by passing fences to the kernel and/or 
>>>> receiving them from the the kernel.
>>>>
>>>> Fences passed to the kernel are named in-fences and the kernel should wait
>>>> them to signal before using the buffer. On the other side, the kernel creates
>>>> out-fences for every buffer it receives from userspace. This fence is sent back
>>>> to userspace and it will signal when the capture, for example, has finished.
>>>>
>>>> Signalling an out-fence in V4L2 would mean that the job on the buffer is done
>>>> and the buffer can be used by other drivers.
>>>>
>>>> Current RFC implementation
>>>> --------------------------
>>>>
>>>> The current implementation is not intended to be more than a PoC to start
>>>> the discussion on how Explicit Synchronization should be supported in V4L2.
>>>>
>>>> The first patch proposes an userspace API for fences, then on patch 2
>>>> we prepare to the addition of in-fences in patch 3, by introducing the
>>>> infrastructure on vb2 to wait on an in-fence signal before queueing the buffer
>>>> in the driver.
>>>>
>>>> Patch 4 fix uvc v4l2 event handling and patch 5 configure q->dev for vivid
>>>> drivers to enable to subscribe and dequeue events on it.
>>>>
>>>> Patches 6-7 enables support to notify BUF_QUEUED events, i.e., let userspace
>>>> know that particular buffer was enqueued in the driver. This is needed,
>>>> because we return the out-fence fd as an out argument in QBUF, but at the time
>>>> it returns we don't know to which buffer the fence will be attached thus
>>>> the BUF_QUEUED event tells which buffer is associated to the fence received in
>>>> QBUF by userspace.
>>>>
>>>> Patches 8 and 9 add more fence infrastructure to support out-fences and finally
>>>> patch 10 adds support to out-fences.
>>>>
>>>> TODO/OPEN:
>>>> ----------
>>>>
>>>> * For this first implementation we will keep the ordering of the buffers queued
>>>> in videobuf2, that means we will only enqueue buffer whose fence was signalled
>>>> if that buffer is the first one in the queue. Otherwise it has to wait until it
>>>> is the first one. This is not implmented yet. Later we could create a flag to
>>>> allow unordered queing in the drivers from vb2 if needed.
>>>
>>> The V4L2 spec doesn't warrant that the buffers will be dequeued at the
>>> queue order.
>>>
>>> In practice, however, most drivers will not reorder. Yet, mem2mem codec 
>>> drivers may reorder the buffers at the output, as the luminance information
>>> (Y) usually comes first on JPEG/MPEG-like formats.
>>>
>>>> * Should we have out-fences per-buffer or per-plane? or both? In this RFC, for
>>>> simplicity they are per-buffer, but Mauro and Javier raised the option of
>>>> doing per-plane fences. That could benefit mem2mem and V4L2 <-> GPU operation
>>>> at least on cases when we have Capture hw that releases the Y frame before the
>>>> other frames for example. When using V4L2 per-plane out-fences to communicate
>>>> with KMS they would need to be merged together as currently the DRM Plane
>>>> interface only supports one fence per DRM Plane.
>>>
>>> That's another advantage of using a new set of ioctls for queues: with that,
>>> queuing/dequeing per plane will be easier. On codec drivers, doing it per
>>> plane could bring performance improvements.
>>>
>>
>> You don't really need to Q/DQ on a per plane basis AFAICT. Since on QBUF you
>> can get a set of out-fences that can be passed to the other driver and so it
>> should be able to wait per fence.
>>
>>>> In-fences should be per-buffer as the DRM only has per-buffer fences, but
>>
>> I'm not that familiar with DRM, but I thought DRM fences was also per plane
>> and not per buffer.
> 
> DRM plane is a different thing, its a representation of a region on the
> screen and there is only one buffer for each DRM plane. 
>

Yes, but you can still have different DRM planes, right? For example a RGB
primary plane and a YUV overlay plane for video display. That's why I thought
there would be some mapping between the v4l2 planes and DRM planes since there's
a way for DRM to import dma-bufs exported by v4l2 and vice-versa.

> One of the questions I raised was: how to match V4L2 per-plane fences to
> DRM per-buffer fences?
>

I don't think that DRM fences should be per-buffer but instead per dma-buf.
It's OK if there's a 1:1 mapping between DRM buffer and dma-buf, but that's
not the case for v4l2. You can have many dma-bufs in a v4l2_buffer, one per
each v4l2 plane.

IOW, it shouldn't be about DRM buffers waiting for fences associated with
v4l2 buffers, but instead about DRM waiting for fences associated with
dma-bufs that were exported by v4l2 and imported by DRM.

>>
>> How this works without fences? For V4L2 there's a dma-buf fd per plane and
>> so I was expecting the DRM API to also import a dma-buf fd per DRM plane.
> 
> Yes. It should do something similar behind the Framebuffer abstraction.
> 
>>
>> I only have access to an Exynos board whose display controller supports
>> single plane formats, so I don't know how this works for multi planar.
>>
>>>> in case of mem2mem operations per-plane fences might be useful?
>>>>
>>>> So should we have both ways, per-plane and per-buffer, or just one of them
>>>> for now?
>>>
>>> The API should be flexible enough to support both usecases. We could
>>> implement just per-buffer in the beginning, but, on such case, we
>>> should deploy an API that will allow to later add per-plane fences
>>> without breaking userspace.
> 
> I believe we can just extend the per-plane parts of QBUF for their
> fences. We could even use plane[0] for the per-buffer case.
>

Agreed.

>>>
>>> So, I prefer that, on multiplane fences, we have one fence per plane,
>>> even if, at the first implementation, all fences will be released
>>> at the same time, when the buffer is fully filled. That would allow
>>> us to later improve it, without changing userspace.
>>
>> It's true that vb2 can't currently signal fences per plane since the interface
>> between vb2 and drivers is per vb2_buffer. But the uAPI shouldn't be restricted
>> by this implementation detail (that can be changed) and should support per plane
>> fences IMHO.
>>
>> That's for example the case with the V4L2 dma-buf API. There is a dma-buf fd per
>> plane, and internally for vb2 single planar buffers use the dma-buf associated
>> with plane 0.
>>
>> Now when mentioning this, I noticed that in your implementation the fences are
>> not associated with a dma-buf. I thought the idea was for the fences to be
>> associated with a dma-buf's reservation object. If we do that, then fences will
>> be per fence since the dma-buf/reservation objet are also per fence in v4l2/vb2.
> 

Argh, I noticed that I made a lot of typos here. What I meant was:

"then fences will be per v4l2 plane since the dma-buf/reservation object are also
per v4l2 plane in v4l2/vb2."

So what I tried to say is that each v4l2 plane has an associated dma-buf, and so
if fences are per plane, it will also be per dma-buf. I guess at the DRM side is
similar, you have dma-buf associated with DRM planes that fences should protect.

> Can you explain what you were thinking on the relation between fences
> and reservation objects? Not sure I follow.
>

As mentioned above, I believe that fences should be associated with dma-bufs.
And a dma-buf already has a reservation object that can be used to associate
fences to a dma-buf. You can add shared or exclusive fences and that be looked
up by the other driver.

That's what implicit fences uses and I believe explicit fences should also use
this infrastructure and export these fences with sync_file_create() as you do.

> Gustavo
> 

Best regards,
-- 
Javier Martinez Canillas
Open Source Group
Samsung Research America

[toc] | [prev] | [next] | [standalone]

#1618057

Hi Javier,

2017-04-05 Javier Martinez Canillas <javier@osg.samsung.com>:

> Hello Gustavo,
> 
> On 04/05/2017 11:09 AM, Gustavo Padovan wrote:
> > 2017-04-03 Javier Martinez Canillas <javier@osg.samsung.com>:
> > 
> >> Hello Mauro and Gustavo,
> >>
> >> On 04/03/2017 07:16 AM, Mauro Carvalho Chehab wrote:
> >>> Hi Gustavo,
> >>>
> >>> Em Mon, 13 Mar 2017 16:20:25 -0300
> >>> Gustavo Padovan <gustavo@padovan.org> escreveu:
> >>>
> >>>> From: Gustavo Padovan <gustavo.padovan@collabora.com>
> >>>>
> >>>> Hi,
> >>>>
> >>>> This RFC adds support for Explicit Synchronization of shared buffers in V4L2.
> >>>> It uses the Sync File Framework[1] as vector to communicate the fences
> >>>> between kernel and userspace.
> >>>
> >>> Thanks for your work!
> >>>
> >>> I looked on your patchset, and I didn't notice anything really weird
> >>> there. So, instead on reviewing patch per patch, I prefer to discuss
> >>> about the requirements and API, as depending on it, the code base will
> >>> change a lot.
> >>>
> >>
> >> Agree that's better to first set on an uAPI and then implement based on that.
> >>  
> >>> I'd like to do some tests with it on devices with mem2mem drivers.
> >>> My plan is to use an Exynos board for such thing, but I guess that
> >>> the DRM driver for it currently doesn't. I'm seeing internally if someone
> >>> could be sure that Exynos driver upstream will become ready for such
> >>> tests.
> >>>
> >>
> >> Not sure if you should try to do testing before agreeing on an uAPI and
> >> implementation.
> >>
> >>> Javier wrote some patches last year meant to implement implicit
> >>> fences support. What we noticed is that, while his mechanism worked
> >>> fine for pure capture and pure output devices, when we added a mem2mem
> >>> device, on a DMABUF+fences pipeline, e. g.:
> >>>
> >>> 	sensor -> [m2m] -> DRM
> >>>
> >>> End everything using fences/DMABUF, the fences mechanism caused dead
> >>> locks on existing userspace apps.
> >>>
> >>> A m2m device has both capture and output devnodes. Both should be
> >>> queued/dequeued. The capture queue is synchronized internally at the
> >>> driver with the output buffer[1].
> >>>
> >>> [1] The names here are counter-intuitive: "capture" is a devnode
> >>> where userspace receives a video stream; "output" is a devnode where
> >>> userspace feeds a video stream.
> >>>
> >>> The problem is that adding implicit fences changed the behavior of
> >>> the ioctls, causing gstreamer to wait forever for buffers to be ready.
> >>>
> >>
> >> The problem was related to trying to make user-space unaware of the implicit
> >> fences support, and so it tried to QBUF a buffer that had already a pending
> >> fence. A workaround was to block the second QBUF ioctl if the buffer had a
> >> pending fence, but this caused the mentioned deadlock since GStreamer wasn't
> >> expecting the QBUF ioctl to block.
> >>
> >>> I suspect that, even with explicit fences, the behavior of Q/DQ
> >>> will be incompatible with the current behavior (or will require some
> >>> dirty hacks to make it identical). 
> > 
> > For QBUF the only difference is that we set flags for fences and pass
> > and receives in and out fences. For DQBUF the behavior is exactly the
> > same. What incompatibles or hacks do you see?
> > 
> > I had the expectation that the flags would be for userspace to learn
> > about any different behavior.
> > 
> >>>
> >>> So, IMHO, the best would be to use a new set of ioctls, when fences are
> >>> used (like VIDIOC_QFENCE/DQFENCE).
> >>>
> >>
> >> For explicit you can check if there's an input-fence so is different than
> >> implicit, but still I agree that it would be better to have specific ioctls.
> > 
> > I'm pretty new to v4l2 so I don't know all use cases yet, but what I
> > thought was to just add extra flags to QBUF to mark when using fences
> > instead of having userspace  to setup completely new ioctls for fences.
> > The burden for userspace should be smaller with flags.
> >
> 
> Yes, you are right. So I guess its better indeed to just extend the current
> ioctls as you propose and only move to new ones if really needed.
> 
> >>
> >>>>
> >>>> I'm sending this to start the discussion on the best approach to implement
> >>>> Explicit Synchronization, please check the TODO/OPEN section below.
> >>>>
> >>>> Explicit Synchronization allows us to control the synchronization of
> >>>> shared buffers from userspace by passing fences to the kernel and/or 
> >>>> receiving them from the the kernel.
> >>>>
> >>>> Fences passed to the kernel are named in-fences and the kernel should wait
> >>>> them to signal before using the buffer. On the other side, the kernel creates
> >>>> out-fences for every buffer it receives from userspace. This fence is sent back
> >>>> to userspace and it will signal when the capture, for example, has finished.
> >>>>
> >>>> Signalling an out-fence in V4L2 would mean that the job on the buffer is done
> >>>> and the buffer can be used by other drivers.
> >>>>
> >>>> Current RFC implementation
> >>>> --------------------------
> >>>>
> >>>> The current implementation is not intended to be more than a PoC to start
> >>>> the discussion on how Explicit Synchronization should be supported in V4L2.
> >>>>
> >>>> The first patch proposes an userspace API for fences, then on patch 2
> >>>> we prepare to the addition of in-fences in patch 3, by introducing the
> >>>> infrastructure on vb2 to wait on an in-fence signal before queueing the buffer
> >>>> in the driver.
> >>>>
> >>>> Patch 4 fix uvc v4l2 event handling and patch 5 configure q->dev for vivid
> >>>> drivers to enable to subscribe and dequeue events on it.
> >>>>
> >>>> Patches 6-7 enables support to notify BUF_QUEUED events, i.e., let userspace
> >>>> know that particular buffer was enqueued in the driver. This is needed,
> >>>> because we return the out-fence fd as an out argument in QBUF, but at the time
> >>>> it returns we don't know to which buffer the fence will be attached thus
> >>>> the BUF_QUEUED event tells which buffer is associated to the fence received in
> >>>> QBUF by userspace.
> >>>>
> >>>> Patches 8 and 9 add more fence infrastructure to support out-fences and finally
> >>>> patch 10 adds support to out-fences.
> >>>>
> >>>> TODO/OPEN:
> >>>> ----------
> >>>>
> >>>> * For this first implementation we will keep the ordering of the buffers queued
> >>>> in videobuf2, that means we will only enqueue buffer whose fence was signalled
> >>>> if that buffer is the first one in the queue. Otherwise it has to wait until it
> >>>> is the first one. This is not implmented yet. Later we could create a flag to
> >>>> allow unordered queing in the drivers from vb2 if needed.
> >>>
> >>> The V4L2 spec doesn't warrant that the buffers will be dequeued at the
> >>> queue order.
> >>>
> >>> In practice, however, most drivers will not reorder. Yet, mem2mem codec 
> >>> drivers may reorder the buffers at the output, as the luminance information
> >>> (Y) usually comes first on JPEG/MPEG-like formats.
> >>>
> >>>> * Should we have out-fences per-buffer or per-plane? or both? In this RFC, for
> >>>> simplicity they are per-buffer, but Mauro and Javier raised the option of
> >>>> doing per-plane fences. That could benefit mem2mem and V4L2 <-> GPU operation
> >>>> at least on cases when we have Capture hw that releases the Y frame before the
> >>>> other frames for example. When using V4L2 per-plane out-fences to communicate
> >>>> with KMS they would need to be merged together as currently the DRM Plane
> >>>> interface only supports one fence per DRM Plane.
> >>>
> >>> That's another advantage of using a new set of ioctls for queues: with that,
> >>> queuing/dequeing per plane will be easier. On codec drivers, doing it per
> >>> plane could bring performance improvements.
> >>>
> >>
> >> You don't really need to Q/DQ on a per plane basis AFAICT. Since on QBUF you
> >> can get a set of out-fences that can be passed to the other driver and so it
> >> should be able to wait per fence.
> >>
> >>>> In-fences should be per-buffer as the DRM only has per-buffer fences, but
> >>
> >> I'm not that familiar with DRM, but I thought DRM fences was also per plane
> >> and not per buffer.
> > 
> > DRM plane is a different thing, its a representation of a region on the
> > screen and there is only one buffer for each DRM plane. 
> >
> 
> Yes, but you can still have different DRM planes, right? For example a RGB
> primary plane and a YUV overlay plane for video display. That's why I thought
> there would be some mapping between the v4l2 planes and DRM planes since there's
> a way for DRM to import dma-bufs exported by v4l2 and vice-versa.
> 
> > One of the questions I raised was: how to match V4L2 per-plane fences to
> > DRM per-buffer fences?
> >
> 
> I don't think that DRM fences should be per-buffer but instead per dma-buf.
> It's OK if there's a 1:1 mapping between DRM buffer and dma-buf, but that's
> not the case for v4l2. You can have many dma-bufs in a v4l2_buffer, one per
> each v4l2 plane.
> 
> IOW, it shouldn't be about DRM buffers waiting for fences associated with
> v4l2 buffers, but instead about DRM waiting for fences associated with
> dma-bufs that were exported by v4l2 and imported by DRM.

The explicit synchronization implementation on DRM never deals directly
with reservation objects. That is what implicit synchronization does.
To be strict in explicit we associated the to a job, and the job can be
anything. For DRM in-fences they are per-plane and the signaling of the
fence tell us the framebuffer is ready to be scanned out. For out-fences
they are per-crtc (which includes many dma-bufs). So setting them to be
per dma-buf only for V4L2 limits our possibilities to solve problems.

So instead, let's what problems we want to solve and how sane the
userspace api can be.

In which situations will we have an advantage for having per-plane
fences on V4L2? Not all userspace applications will need that and
per-buffer is much simpler as an API.

As discussed earlier we should do it in incremental steps and support
per v4l2-buffer first, then once that is done we can add per-plane to
it.

Gustavo

[toc] | [prev] | [next] | [standalone]

#1618075

Hello Gustavo,

On 04/06/2017 10:08 AM, Gustavo Padovan wrote:
> Hi Javier,
> 
> 2017-04-05 Javier Martinez Canillas <javier@osg.samsung.com>:
> 
>> Hello Gustavo,
>>
>> On 04/05/2017 11:09 AM, Gustavo Padovan wrote:
>>> 2017-04-03 Javier Martinez Canillas <javier@osg.samsung.com>:
>>>
>>>> Hello Mauro and Gustavo,
>>>>
>>>> On 04/03/2017 07:16 AM, Mauro Carvalho Chehab wrote:
>>>>> Hi Gustavo,
>>>>>
>>>>> Em Mon, 13 Mar 2017 16:20:25 -0300
>>>>> Gustavo Padovan <gustavo@padovan.org> escreveu:
>>>>>
>>>>>> From: Gustavo Padovan <gustavo.padovan@collabora.com>
>>>>>>
>>>>>> Hi,
>>>>>>
>>>>>> This RFC adds support for Explicit Synchronization of shared buffers in V4L2.
>>>>>> It uses the Sync File Framework[1] as vector to communicate the fences
>>>>>> between kernel and userspace.
>>>>>
>>>>> Thanks for your work!
>>>>>
>>>>> I looked on your patchset, and I didn't notice anything really weird
>>>>> there. So, instead on reviewing patch per patch, I prefer to discuss
>>>>> about the requirements and API, as depending on it, the code base will
>>>>> change a lot.
>>>>>
>>>>
>>>> Agree that's better to first set on an uAPI and then implement based on that.
>>>>  
>>>>> I'd like to do some tests with it on devices with mem2mem drivers.
>>>>> My plan is to use an Exynos board for such thing, but I guess that
>>>>> the DRM driver for it currently doesn't. I'm seeing internally if someone
>>>>> could be sure that Exynos driver upstream will become ready for such
>>>>> tests.
>>>>>
>>>>
>>>> Not sure if you should try to do testing before agreeing on an uAPI and
>>>> implementation.
>>>>
>>>>> Javier wrote some patches last year meant to implement implicit
>>>>> fences support. What we noticed is that, while his mechanism worked
>>>>> fine for pure capture and pure output devices, when we added a mem2mem
>>>>> device, on a DMABUF+fences pipeline, e. g.:
>>>>>
>>>>> 	sensor -> [m2m] -> DRM
>>>>>
>>>>> End everything using fences/DMABUF, the fences mechanism caused dead
>>>>> locks on existing userspace apps.
>>>>>
>>>>> A m2m device has both capture and output devnodes. Both should be
>>>>> queued/dequeued. The capture queue is synchronized internally at the
>>>>> driver with the output buffer[1].
>>>>>
>>>>> [1] The names here are counter-intuitive: "capture" is a devnode
>>>>> where userspace receives a video stream; "output" is a devnode where
>>>>> userspace feeds a video stream.
>>>>>
>>>>> The problem is that adding implicit fences changed the behavior of
>>>>> the ioctls, causing gstreamer to wait forever for buffers to be ready.
>>>>>
>>>>
>>>> The problem was related to trying to make user-space unaware of the implicit
>>>> fences support, and so it tried to QBUF a buffer that had already a pending
>>>> fence. A workaround was to block the second QBUF ioctl if the buffer had a
>>>> pending fence, but this caused the mentioned deadlock since GStreamer wasn't
>>>> expecting the QBUF ioctl to block.
>>>>
>>>>> I suspect that, even with explicit fences, the behavior of Q/DQ
>>>>> will be incompatible with the current behavior (or will require some
>>>>> dirty hacks to make it identical). 
>>>
>>> For QBUF the only difference is that we set flags for fences and pass
>>> and receives in and out fences. For DQBUF the behavior is exactly the
>>> same. What incompatibles or hacks do you see?
>>>
>>> I had the expectation that the flags would be for userspace to learn
>>> about any different behavior.
>>>
>>>>>
>>>>> So, IMHO, the best would be to use a new set of ioctls, when fences are
>>>>> used (like VIDIOC_QFENCE/DQFENCE).
>>>>>
>>>>
>>>> For explicit you can check if there's an input-fence so is different than
>>>> implicit, but still I agree that it would be better to have specific ioctls.
>>>
>>> I'm pretty new to v4l2 so I don't know all use cases yet, but what I
>>> thought was to just add extra flags to QBUF to mark when using fences
>>> instead of having userspace  to setup completely new ioctls for fences.
>>> The burden for userspace should be smaller with flags.
>>>
>>
>> Yes, you are right. So I guess its better indeed to just extend the current
>> ioctls as you propose and only move to new ones if really needed.
>>
>>>>
>>>>>>
>>>>>> I'm sending this to start the discussion on the best approach to implement
>>>>>> Explicit Synchronization, please check the TODO/OPEN section below.
>>>>>>
>>>>>> Explicit Synchronization allows us to control the synchronization of
>>>>>> shared buffers from userspace by passing fences to the kernel and/or 
>>>>>> receiving them from the the kernel.
>>>>>>
>>>>>> Fences passed to the kernel are named in-fences and the kernel should wait
>>>>>> them to signal before using the buffer. On the other side, the kernel creates
>>>>>> out-fences for every buffer it receives from userspace. This fence is sent back
>>>>>> to userspace and it will signal when the capture, for example, has finished.
>>>>>>
>>>>>> Signalling an out-fence in V4L2 would mean that the job on the buffer is done
>>>>>> and the buffer can be used by other drivers.
>>>>>>
>>>>>> Current RFC implementation
>>>>>> --------------------------
>>>>>>
>>>>>> The current implementation is not intended to be more than a PoC to start
>>>>>> the discussion on how Explicit Synchronization should be supported in V4L2.
>>>>>>
>>>>>> The first patch proposes an userspace API for fences, then on patch 2
>>>>>> we prepare to the addition of in-fences in patch 3, by introducing the
>>>>>> infrastructure on vb2 to wait on an in-fence signal before queueing the buffer
>>>>>> in the driver.
>>>>>>
>>>>>> Patch 4 fix uvc v4l2 event handling and patch 5 configure q->dev for vivid
>>>>>> drivers to enable to subscribe and dequeue events on it.
>>>>>>
>>>>>> Patches 6-7 enables support to notify BUF_QUEUED events, i.e., let userspace
>>>>>> know that particular buffer was enqueued in the driver. This is needed,
>>>>>> because we return the out-fence fd as an out argument in QBUF, but at the time
>>>>>> it returns we don't know to which buffer the fence will be attached thus
>>>>>> the BUF_QUEUED event tells which buffer is associated to the fence received in
>>>>>> QBUF by userspace.
>>>>>>
>>>>>> Patches 8 and 9 add more fence infrastructure to support out-fences and finally
>>>>>> patch 10 adds support to out-fences.
>>>>>>
>>>>>> TODO/OPEN:
>>>>>> ----------
>>>>>>
>>>>>> * For this first implementation we will keep the ordering of the buffers queued
>>>>>> in videobuf2, that means we will only enqueue buffer whose fence was signalled
>>>>>> if that buffer is the first one in the queue. Otherwise it has to wait until it
>>>>>> is the first one. This is not implmented yet. Later we could create a flag to
>>>>>> allow unordered queing in the drivers from vb2 if needed.
>>>>>
>>>>> The V4L2 spec doesn't warrant that the buffers will be dequeued at the
>>>>> queue order.
>>>>>
>>>>> In practice, however, most drivers will not reorder. Yet, mem2mem codec 
>>>>> drivers may reorder the buffers at the output, as the luminance information
>>>>> (Y) usually comes first on JPEG/MPEG-like formats.
>>>>>
>>>>>> * Should we have out-fences per-buffer or per-plane? or both? In this RFC, for
>>>>>> simplicity they are per-buffer, but Mauro and Javier raised the option of
>>>>>> doing per-plane fences. That could benefit mem2mem and V4L2 <-> GPU operation
>>>>>> at least on cases when we have Capture hw that releases the Y frame before the
>>>>>> other frames for example. When using V4L2 per-plane out-fences to communicate
>>>>>> with KMS they would need to be merged together as currently the DRM Plane
>>>>>> interface only supports one fence per DRM Plane.
>>>>>
>>>>> That's another advantage of using a new set of ioctls for queues: with that,
>>>>> queuing/dequeing per plane will be easier. On codec drivers, doing it per
>>>>> plane could bring performance improvements.
>>>>>
>>>>
>>>> You don't really need to Q/DQ on a per plane basis AFAICT. Since on QBUF you
>>>> can get a set of out-fences that can be passed to the other driver and so it
>>>> should be able to wait per fence.
>>>>
>>>>>> In-fences should be per-buffer as the DRM only has per-buffer fences, but
>>>>
>>>> I'm not that familiar with DRM, but I thought DRM fences was also per plane
>>>> and not per buffer.
>>>
>>> DRM plane is a different thing, its a representation of a region on the
>>> screen and there is only one buffer for each DRM plane. 
>>>
>>
>> Yes, but you can still have different DRM planes, right? For example a RGB
>> primary plane and a YUV overlay plane for video display. That's why I thought
>> there would be some mapping between the v4l2 planes and DRM planes since there's
>> a way for DRM to import dma-bufs exported by v4l2 and vice-versa.
>>
>>> One of the questions I raised was: how to match V4L2 per-plane fences to
>>> DRM per-buffer fences?
>>>
>>
>> I don't think that DRM fences should be per-buffer but instead per dma-buf.
>> It's OK if there's a 1:1 mapping between DRM buffer and dma-buf, but that's
>> not the case for v4l2. You can have many dma-bufs in a v4l2_buffer, one per
>> each v4l2 plane.
>>
>> IOW, it shouldn't be about DRM buffers waiting for fences associated with
>> v4l2 buffers, but instead about DRM waiting for fences associated with
>> dma-bufs that were exported by v4l2 and imported by DRM.
> 
> The explicit synchronization implementation on DRM never deals directly
> with reservation objects. That is what implicit synchronization does.
> To be strict in explicit we associated the to a job, and the job can be
> anything. For DRM in-fences they are per-plane and the signaling of the
> fence tell us the framebuffer is ready to be scanned out. For out-fences
> they are per-crtc (which includes many dma-bufs). So setting them to be

I see, thanks a lot for the clarification and sorry for my ignorance about
explicit fences support in DRM.

> per dma-buf only for V4L2 limits our possibilities to solve problems.
>

Yes, having it per dma-buf will not help when using fences with DRM indeed.

> So instead, let's what problems we want to solve and how sane the
> userspace api can be.
> 
> In which situations will we have an advantage for having per-plane
> fences on V4L2? Not all userspace applications will need that and
> per-buffer is much simpler as an API.
>

I was thinking in a case where you are using fences between two v4l2
devices that can process individual planes in a multi-planar format.

But as mentioned in a previous email, currently it won't have any
advantage since the interface between videobuf2 and v4l2 drivers uses
vb2_buffer and not vb2_planes. That is, individual planes can't be
passed between vb2 and drivers and only complete buffers (so fences
will be signaled at the same time anyways even if are per plane).

But I thought the API shouldn't be restricted by this vb2 implementation
detail since vb2 can be changed in the future to pass vb2_planes instead. 

> As discussed earlier we should do it in incremental steps and support
> per v4l2-buffer first, then once that is done we can add per-plane to
> it.
>

Yes, I misunderstood how explicit fences worked for DRM and thought that
was used to protect individual dma-bufs. But as you said, out-fences are
per CRTC so then I guess that makes sense to have per v4l2_buffer fences
in v4l2/vb2, since explicit fences are used to signal that an operation
occurred rather than a specific dma-buf is ready to be accessed.

> Gustavo
> 

Best regards,
-- 
Javier Martinez Canillas
Open Source Group
Samsung Research America

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web