Message-ID: <20241029095358.GE22600@pendragon.ideasonboard.com>
Date: Tue, 29 Oct 2024 11:53:58 +0200
From: Laurent Pinchart <laurent.pinchart@...asonboard.com>
To: Hans Verkuil <hverkuil@...all.nl>
Cc: Tomi Valkeinen <tomi.valkeinen@...asonboard.com>,
	linux-media@...r.kernel.org, linux-kernel@...r.kernel.org,
	devicetree@...r.kernel.org, linux-rpi-kernel@...ts.infradead.org,
	linux-arm-kernel@...ts.infradead.org,
	Naushir Patuck <naush@...pberrypi.com>,
	Sakari Ailus <sakari.ailus@...ux.intel.com>,
	Jacopo Mondi <jacopo.mondi@...asonboard.com>,
	Kieran Bingham <kieran.bingham@...asonboard.com>,
	Mauro Carvalho Chehab <mchehab@...nel.org>,
	Raspberry Pi Kernel Maintenance <kernel-list@...pberrypi.com>,
	Rob Herring <robh+dt@...nel.org>,
	Krzysztof Kozlowski <krzysztof.kozlowski+dt@...aro.org>,
	Conor Dooley <conor+dt@...nel.org>,
	Florian Fainelli <florian.fainelli@...adcom.com>,
	Broadcom internal kernel review list <bcm-kernel-feedback-list@...adcom.com>,
	Rob Herring <robh@...nel.org>,
	Krzysztof Kozlowski <krzk+dt@...nel.org>
Subject: Re: [PATCH v6 3/4] media: raspberrypi: Add support for RP1-CFE

On Tue, Oct 29, 2024 at 09:23:35AM +0100, Hans Verkuil wrote:
> On 28/10/2024 17:32, Laurent Pinchart wrote:
> > On Mon, Oct 28, 2024 at 05:32:27PM +0200, Tomi Valkeinen wrote:
> >> On 28/10/2024 17:17, Laurent Pinchart wrote:
> >>> On Mon, Oct 28, 2024 at 12:30:45PM +0100, Hans Verkuil wrote:
> >>>> On 28/10/2024 12:25, Tomi Valkeinen wrote:
> >>>>> On 28/10/2024 13:13, Hans Verkuil wrote:
> >>>>>> On 28/10/2024 12:05, Tomi Valkeinen wrote:
> >>>>>>> On 28/10/2024 12:11, Hans Verkuil wrote:
> >>>>>>>> On 28/10/2024 10:21, Tomi Valkeinen wrote:
> >>>>>>>>> On 24/10/2024 11:20, Hans Verkuil wrote:
> >>>>>>>>>> Hi Tomi,
> >>>>>>>>>>
> >>>>>>>>>> I know this driver is already merged, but while checking for drivers that use
> >>>>>>>>>> q->max_num_buffers I stumbled on this cfe code:
> >>>>>>>>>>
> >>>>>>>>>> <snip>
> >>>>>>>>>>
> >>>>>>>>>>> +/*
> >>>>>>>>>>> + * vb2 ops
> >>>>>>>>>>> + */
> >>>>>>>>>>> +
> >>>>>>>>>>> +static int cfe_queue_setup(struct vb2_queue *vq, unsigned int *nbuffers,
> >>>>>>>>>>> +               unsigned int *nplanes, unsigned int sizes[],
> >>>>>>>>>>> +               struct device *alloc_devs[])
> >>>>>>>>>>> +{
> >>>>>>>>>>> +    struct cfe_node *node = vb2_get_drv_priv(vq);
> >>>>>>>>>>> +    struct cfe_device *cfe = node->cfe;
> >>>>>>>>>>> +    unsigned int size = is_image_node(node) ?
> >>>>>>>>>>> +                    node->vid_fmt.fmt.pix.sizeimage :
> >>>>>>>>>>> +                    node->meta_fmt.fmt.meta.buffersize;
> >>>>>>>>>>> +
> >>>>>>>>>>> +    cfe_dbg(cfe, "%s: [%s] type:%u\n", __func__, node_desc[node->id].name,
> >>>>>>>>>>> +        node->buffer_queue.type);
> >>>>>>>>>>> +
> >>>>>>>>>>> +    if (vq->max_num_buffers + *nbuffers < 3)
> >>>>>>>>>>> +        *nbuffers = 3 - vq->max_num_buffers;
> >>>>>>>>>>
> >>>>>>>>>> This makes no sense: max_num_buffers is 32, unless explicitly set when vb2_queue_init
> >>>>>>>>>> is called. So 32 + *nbuffers is never < 3.
> >>>>>>>>>>
> >>>>>>>>>> If the idea is that at least 3 buffers should be allocated by REQBUFS, then set
> >>>>>>>>>> q->min_reqbufs_allocation = 3; before calling vb2_queue_init and vb2 will handle this
> >>>>>>>>>> for you.
> >>>>>>>>>>
> >>>>>>>>>> Drivers shouldn't modify *nbuffers, except in very rare circumstances, especially
> >>>>>>>>>> since the code is almost always wrong.
> >>>>>>>>>
> >>>>>>>>> Looking at this, the original code in the old BSP tree was, which somehow, along the long way, got turned into the above:
> >>>>>>>>>
> >>>>>>>>> if (vq->num_buffers + *nbuffers < 3)
> >>>>>>>>>            *nbuffers = 3 - vq->num_buffers;
> >>>>>>>>>
> >>>>>>>>> So... I think that is the same as "q->min_reqbufs_allocation = 3"?
> >>>>>>>>>
> >>>>>>>>> The distinction between min_queued_buffers and
> >>>>>>>>> min_reqbufs_allocation, or rather the need for the latter, still
> >>>>>>>>> escapes me. If the HW/SW requires N buffers to be queued, why
> >>>>>>>>> would we require allocating more than N buffers?
> >>>>>>>>
> >>>>>>>> min_queued_buffers is easiest to explain: that represents the requirements of the DMA
> >>>>>>>> engine, i.e. how many buffers much be queued before the DMA engine can be started.
> >>>>>>>> Typically it is 0, 1 or 2.
> >>>
> >>> That's partly true only. Even if the hardware requires 2 buffers, a
> >>> driver can allocate scratch buffers to lower the requirement for
> >>> userspace. Setting min_queued_buffers to 1 is usually fine, as there are
> >>> few use cases for userspace to start the hardware before a buffer is
> >>> available to capture a frame to. A value of 2 is much more problematic,
> >>> as it prevents operating with a single buffer. I know using a single
> >>> buffer results in frame drops, but there are resource-constrained
> >>> systems where application don't always need all the frames (such as the
> >>> Raspberry Pi Zero for instance). I very strongly encourage drivers to
> >>> never set a min_queued_buffers value higher than 1.
> >>>
> >>>>>>>>
> >>>>>>>> min_reqbufs_allocation is the minimum number of buffers that will be allocated when
> >>>>>>>> calling VIDIOC_REQBUFS in order for userspace to be able to stream without blocking
> >>>>>>>> or dropping frames.
> >>>>>>>>
> >>>>>>>> Typically this is 3 for video capture: one buffer is being DMAed, another is queued up
> >>>>>>>> and the third is being processed by userspace. But sometimes drivers have other
> >>>>>>>> requirements.
> >>>
> >>> This is exactly why I dislike min_reqbufs_allocation when set based on
> >>> this logic, it encodes assumption on userspace use cases that a capture
> >>> driver really shouldn't make.
> >>>
> >>>>>>>>
> >>>>>>>> The reason is that some applications will just call VIDIOC_REQBUFS with count=1 and
> >>>>>>>> expect it to be rounded up to whatever makes sense. See the VIDIOC_REQBUFS doc in
> >>>>>>>> https://hverkuil.home.xs4all.nl/spec/userspace-api/v4l/vidioc-reqbufs.html
> >>>>>>>>
> >>>>>>>> "It can be smaller than the number requested, even zero, when the driver runs out of
> >>>>>>>>     free memory. A larger number is also possible when the driver requires more buffers
> >>>>>>>>     to function correctly."
> >>>>>>>>
> >>>>>>>> How drivers implement this is a mess, and usually the code in the driver is wrong as
> >>>>>>>> well. In particular they often did not take VIDIOC_CREATE_BUFS into account, i.e.
> >>>>>>>> instead of 'if (vq->num_buffers + *nbuffers < 3)' they would do 'if (*nbuffers < 3)'.
> >>>>>>>
> >>>>>>> Thanks, this was educational!
> >>>>>>>
> >>>>>>> So. If I have a driver that has min_queued_buffers = 1, I can use
> >>>>>>> VIDIOC_CREATE_BUFS to allocate a single buffer, and then capture
> >>>>>>> just one buffer, right? Whereas VIDIOC_REQBUFS would give me
> >>>>>>> (probably) three (or two, if the driver does not set
> >>>>>>> min_reqbufs_allocation). Three buffers makes sense for full
> >>>>>>> streaming, of course.
> >>>>>>>
> >>>>>>>> When we worked on the support for more than 32 buffers we added min_reqbufs_allocation
> >>>>>>>> to let the core take care of this. In addition, this only applies to VIDIOC_REQBUFS,
> >>>
> >>> I agree it's better to handle it in the core than in drivers, even if I
> >>> dislike the feature in the first place.
> >>>
> >>>>>>>> if you want full control over the number of allocated buffers, then use VIDIOC_CREATE_BUFS,
> >>>>>>>> with this ioctl the number of buffers will never be more than requested, although it
> >>>>>>>> may be less if you run out of memory.
> >>>
> >>> On a side note, we should transition libcamera to use VIDIOC_CREATE_BUFS
> >>> unconditionally.
> >>>
> >>>>>>>>
> >>>>>>>> I really should go through all existing drivers and fix them up if they try to
> >>>>>>>> handle this in the queue_setup function, I suspect a lot of them are quite messy.
> >>>>>>>>
> >>>>>>>> One thing that is missing in the V4L2 uAPI is a way to report the minimum number of
> >>>>>>>> buffers that need to be allocated, i.e. min_queued_buffers + 1. Since if you want
> >>>>>>>
> >>>>>>> Hmm, so what I wrote above is not correct? One needs min_queued_buffers + 1? Why is that?
> >>>>>>
> >>>>>> The DMA engine always uses min_queued_buffers, so if there are only that many buffers,
> >>>>>> then it can never return a buffer to userspace! So you need one more. That's the absolute
> >>>>>> minimum. For smooth capture you need two more to allow time for userspace to process the
> >>>>>> buffer.
> >>>>>
> >>>>> Hmm, ok, I see. Well, I guess my "I want to capture just a single frame" is not a very common case.
> >>>
> >>> It's not that uncommon, see above.
> >>>
> >>>>>
> >>>>> Can I queue one buffer, start streaming, stop streaming, and get the
> >>>>> filled buffer? But then I guess I don't when the buffer has been
> >>>>> filled, i.e. when to call stop streaming.
> >>>>
> >>>> Exactly. If you really want that, then the driver has to be adapted in the way that Laurent
> >>>> suggested, i.e. with one or more scratch buffers. But that is not always possible, esp. with
> >>>> older hardware without an IOMMU.
> >>>
> >>> Drivers can always allocate a full-frame scratch buffer in the worst
> >>> case. That can waste memory though, which is less than ideal.
> >>>
> >>>>> So, never mind, I don't actually have any use case for this, just wondering.
> >>>>>
> >>>>>>>
> >>>>>>>> to use CREATE_BUFS you need that information so you know that you have to create
> >>>>>>>> at least that number of buffers. We have the V4L2_CID_MIN_BUFFERS_FOR_CAPTURE control,
> >>>>>>>> but it is effectively codec specific. This probably should be clarified.
> >>>>>>>>
> >>>>>>>> I wonder if it wouldn't be better to add a min_num_buffers field to
> >>>>>>>> struct v4l2_create_buffers and set it to min_queued_buffers + 1.
> >>>
> >>> Don't add the +1. We should give userspace the information it needs to
> >>> make informed decisions, not make decisions on its behalf.
> >>>
> >>>>>>>
> >>>>>>> I think this makes sense (although I still don't get the +1).
> >>>>>>>
> >>>>>>> However, based on the experiences from adding the streams features
> >>>>>>> to various ioctls, let's be very careful =). The new
> >>>>>>> 'min_num_buffers' can be filled with garbage by the userspace. If
> >>>>>>> we define the 'min_num_buffers' field to be always filled by the
> >>>>>>> kernel, and any value provided from the userspace to be ignored, I
> >>>>>>> think it should work.
> >>>>>>
> >>>>>> I've posted an RFC for this.
> >>>>>
> >>>>> Thanks, I'll check it out.
> >>>>>
> >>>>> For the original issue in this thread, I think the correct fix is to
> >>>>> remove the lines from cfe_queue_setup(), and add
> >>>>> "q->min_reqbufs_allocation = 3".
> >>>
> >>> Or just don't set min_reqbufs_allocation ? This is a new driver, and it
> >>> requires a device-specific userspace to operate the ISP. I don't think
> >>> we need to care about applications blindly calling VIDIOC_REQBUFS(1) and
> >>> expecting to get more buffers.
> >>
> >> It doesn't require a device-specific userspace for plain CSI-2 capture.
> >>
> >> If I understood right, the expected behavior for VIDIOC_REQBUFS is to 
> >> return enough buffers for "smooth streaming". So even if device-specific 
> >> userspace would be required, doesn't it still make sense to have 
> >> min_reqbufs_allocation = 3?
> > 
> > "Smooth streaming" is use case-dependent, you will need different number
> > of buffers for different use cases. That's why I don't like hardcoding
> > this in a video capture driver. I'd rather expose information about the
> > driver behaviour (in particular, how many buffers it will hold on
> > without returning anything to userspace until a new buffer gets queued)
> > and let applications make a decision. I don't expect applications
> > relying on VIDIOC_REQBUFS(1) to work out-of-the-box on Pi 5 anyway, as
> > the media graph needs to be configured.
> > 
> >> Or is your point that even a device-specific userspace, which knows 
> >> exactly what it's doing, would use VIDIOC_REQBUFS, instead of 
> >> VIDIOC_CREATE_BUFS?
> > 
> > I expect a device-specific userspace not to require drivers to make
> > policy decisions on its behalf.
> 
> Remember that libcamera is a specialized library that indeed wants to
> make policy decisions itself. But many other drivers for much simpler
> pipelines (typically for video receivers) don't need this and can use
> the standard V4L2 API.
> 
> My goal is to have the standard V4L2 API behave in a well-defined manner,
> while giving enough information to specialized userspace code like libcamera
> to do their own thing.

I think we all agree on well-defined :-) Where we may not agree is that,
regardless of whether an application goes through libcamera, interfaces
with V4L2 through another framework, or directly, I don't think the
kernel should make policy decisions on behalf of userspace. I'm fine
keeping VIDIOC_REQBUFS(1) operating as it was meant to because we have
to ensure backward compatibility, but I don't think it's the kind of API
we should design today.

Let's start by making sure we expose the information userspace needs,
and then we can discuss the next step.

> >> Also, if I don't set min_reqbufs_allocation, VIDIOC_REQBUFS(1) would 
> >> still allocate two buffers, not one.

-- 
Regards,

Laurent Pinchart

Powered by blists - more mailing lists