Re: [PATCH v8 49/55] [media] media-device: add support for MEDIA_IOC_G_TOPOLOGY ioctl

From: Sakari Ailus <sakari.ailus@iki.fi>
To: Mauro Carvalho Chehab <mchehab@osg.samsung.com>
Cc: Linux Media Mailing List <linux-media@vger.kernel.org>
Subject: Re: [PATCH v8 49/55] [media] media-device: add support for MEDIA_IOC_G_TOPOLOGY ioctl
Date: Tue, 8 Sep 2015 16:34:59 +0300	[thread overview]
Message-ID: <20150908133459.GG3175@valkosipuli.retiisi.org.uk> (raw)
In-Reply-To: <20150908074945.4358c9d7@recife.lan>

Hi Mauro,

On Tue, Sep 08, 2015 at 07:49:45AM -0300, Mauro Carvalho Chehab wrote:
> Em Tue, 8 Sep 2015 10:26:29 +0300
> Sakari Ailus <sakari.ailus@iki.fi> escreveu:
> 
> > Hi Mauro,
> > 
> > On Mon, Sep 07, 2015 at 10:23:57PM -0300, Mauro Carvalho Chehab wrote:
> > > Em Tue, 8 Sep 2015 01:18:30 +0300
> > > Sakari Ailus <sakari.ailus@iki.fi> escreveu:
> > > 
> > > > Hi Mauro,
> > > > 
> > > > A few comments below.
> > > 
> > > Thanks for review!
> > 
> > You're welcome!
> > 
> > > > 
> > > > On Sun, Sep 06, 2015 at 09:03:09AM -0300, Mauro Carvalho Chehab wrote:
> > > > > Add support for the new MEDIA_IOC_G_TOPOLOGY ioctl, according
> > > > > with the RFC for the MC next generation.
> > > > > 
> > > > > Signed-off-by: Mauro Carvalho Chehab <mchehab@osg.samsung.com>
> > > > > 
> > > > > diff --git a/drivers/media/media-device.c b/drivers/media/media-device.c
> > > > > index 5b2c9f7fcd45..96a476eeb16e 100644
> > > > > --- a/drivers/media/media-device.c
> > > > > +++ b/drivers/media/media-device.c
> > > > > @@ -232,6 +232,156 @@ static long media_device_setup_link(struct media_device *mdev,
> > > > >  	return ret;
> > > > >  }
> > > > >  
> > > > > +static long __media_device_get_topology(struct media_device *mdev,
> > > > > +				      struct media_v2_topology *topo)
> > > > > +{
> > > > > +	struct media_entity *entity;
> > > > > +	struct media_interface *intf;
> > > > > +	struct media_pad *pad;
> > > > > +	struct media_link *link;
> > > > > +	struct media_v2_entity uentity;
> > > > > +	struct media_v2_interface uintf;
> > > > > +	struct media_v2_pad upad;
> > > > > +	struct media_v2_link ulink;
> > > > > +	int ret = 0, i;
> > > > 
> > > > I think i wants to be unsigned.
> > > 
> > > Yes, "i" can be unsigned. I'll change that.
> > > 
> > > > 
> > > > > +
> > > > > +	topo->topology_version = mdev->topology_version;
> > > > > +
> > > > > +	/* Get entities and number of entities */
> > > > > +	i = 0;
> > > > > +	media_device_for_each_entity(entity, mdev) {
> > > > > +		i++;
> > > > > +
> > > > > +		if (ret || !topo->entities)
> > > > > +			continue;
> > > > > +
> > > > > +		if (i > topo->num_entities) {
> > > > > +			ret = -ENOSPC;
> > > > > +			continue;
> > > > > +		}
> > > > > +
> > > > > +		/* Copy fields to userspace struct if not error */
> > > > > +		memset(&uentity, 0, sizeof(uentity));
> > > > > +		uentity.id = entity->graph_obj.id;
> > > > > +		strncpy(uentity.name, entity->name,
> > > > > +			sizeof(uentity.name));
> > > > > +
> > > > > +		if (copy_to_user(&topo->entities[i - 1], &uentity, sizeof(uentity)))
> > > > > +			ret = -EFAULT;
> > > > > +	}
> > > > > +	topo->num_entities = i;
> > > > > +
> > > > > +	/* Get interfaces and number of interfaces */
> > > > > +	i = 0;
> > > > > +	media_device_for_each_intf(intf, mdev) {
> > > > > +		i++;
> > > > > +
> > > > > +		if (ret || !topo->interfaces)
> > > > > +			continue;
> > > > > +
> > > > > +		if (i > topo->num_interfaces) {
> > > > > +			ret = -ENOSPC;
> > > > > +			continue;
> > > > > +		}
> > > > > +
> > > > > +		memset(&uintf, 0, sizeof(uintf));
> > > > > +
> > > > > +		/* Copy intf fields to userspace struct */
> > > > > +		uintf.id = intf->graph_obj.id;
> > > > > +		uintf.intf_type = intf->type;
> > > > > +		uintf.flags = intf->flags;
> > > > > +
> > > > > +		if (media_type(&intf->graph_obj) == MEDIA_GRAPH_INTF_DEVNODE) {
> > > > > +			struct media_intf_devnode *devnode;
> > > > > +
> > > > > +			devnode = intf_to_devnode(intf);
> > > > > +
> > > > > +			uintf.devnode.major = devnode->major;
> > > > > +			uintf.devnode.minor = devnode->minor;
> > > > > +		}
> > > > > +
> > > > > +		if (copy_to_user(&topo->interfaces[i - 1], &uintf, sizeof(uintf)))
> > > > > +			ret = -EFAULT;
> > > > > +	}
> > > > > +	topo->num_interfaces = i;
> > > > > +
> > > > > +	/* Get pads and number of pads */
> > > > > +	i = 0;
> > > > > +	media_device_for_each_pad(pad, mdev) {
> > > > > +		i++;
> > > > > +
> > > > > +		if (ret || !topo->pads)
> > > > > +			continue;
> > > > > +
> > > > > +		if (i > topo->num_pads) {
> > > > > +			ret = -ENOSPC;
> > > > > +			continue;
> > > > > +		}
> > > > > +
> > > > > +		memset(&upad, 0, sizeof(upad));
> > > > > +
> > > > > +		/* Copy pad fields to userspace struct */
> > > > > +		upad.id = pad->graph_obj.id;
> > > > 
> > > > How about the pad index? Shouldn't that be also passed to the user space for
> > > > every pad?
> > > 
> > > We've agreed to not pass the pad index to userspace at the MC workshop.
> > > 
> > > There are two aspects here to consider:
> > > 
> > > a) to properly represent the topology (e. g. TOPOLOGY)
> > > 
> > > Writing the userspace code to support it, I didn't find any need to
> > > pass it to userspace, as the data links are connected via the PAD object 
> > > ID.
> > > 
> > > The PAD index for userspace is just a number that it uses when
> > > generating the dot graph. See:
> > > 	https://mchehab.fedorapeople.org/mc-next-gen/au0828.png
> > > 
> > > This was generated by this tool:
> > > 	http://git.linuxtv.org/cgit.cgi/mchehab/experimental-v4l-utils.git/tree/contrib/test/mc_nextgen_test.c
> > > 
> > > And no pad index was required at all. What the tool does is that it
> > > generates the pad numbers just when the --dot option is used,
> > > at the media_show_graphviz() function.
> > > 
> > > Also, please notice that having a pad index makes harder to support
> > > dynamic PAD addition/removal, as the pad index numberspace will
> > > be discontinued.
> > > 
> > > b) using the PAD index to setup a link.
> > > 
> > > As I argued with Hans on one of his reviews, I don't think that
> > > an index is always the best way to refer to a PAD. 
> > > 
> > > See for example, the entity_1 on the au0828 (ATV decoder). It 
> > > has 3 PADs:
> > > 	- pad 0 - sink - receives an ATV signal [1]
> > > 	- pad 1 - source - it outputs a Video stream
> > > 	- pad 2 - source - it outputs a VBI stream
> > > 
> > > Pads 1 and 2 are not interchangeable, as they carry different
> > > types of data.
> > > 
> > > [1] Actually, to be honest, pad 0 there is also wrong. In a matter
> > > of fact, Composite, S-Video and Tuner interfaces are actually
> > > supported by 3 different PADs on the hardware. On some hardware,
> > > it is just a software configuration, but on others, different
> > > pins are used for each different input type. We just don't need
> > > to have all those details ATM on the Media Controller data flow,
> > > as the V4L2 input selection API at the /dev/video0 devnode hides
> > > those dirty details. Note however that, if we end by adding support
> > > for ATV decoder subdevice nodes, such level of detail may be
> > > required.
> > > 
> > > However some other ATV decoder could have mapped it on some
> > > different order, like:
> > > 	- pad 0 - sink - receives an ATV signal
> > > 	- pad 1 - source - it outputs a VBI stream
> > > 	- pad 2 - source - it outputs a Video stream
> > > 
> > > Any (generic) application would need to check the properties of
> > > pads 1 and 2 in order to identify what of those pads has video
> > > and what pad has VBI. The pad index is meaningless.
> > > 
> > > Ok, there are cases where the pad index are meaningful. I guess
> > > that the best is to use properties to properly identify the
> > > pad, or add some strings to them.
> > > 
> > > Anyway, I guess this should be covered via the properties API.
> > 
> > That's a good summary, however I believe we have to pay attention to the
> > existing API in order not to break user space. The current expectation is
> > that pad numbers start from zero, their range is contiguous and that they
> > are also stable.
> > 
> > The documentation does not say all that, but that's what's currently
> > offered. That's what I believe all hardware specific applications expect.
> > Generic applications probably have no such limitations.
> 
> That's true, but there are two situations here to consider:
> 
> 1) existing drivers/applications:
> - No dynamic support, so PAD numbers are stable. Index is provided
>   via v1;

Agreed, in such cases it's fine as long as the pad indices will be
essentially defined by the driver --- as they used to.

> 
> 2) new drivers/applications:
> - We need a way to provide stable equivalents to PAD indexes via
>   v2.

I propose to only add them once we have the use case. This could be through
the property API. My hands are pretty full of other work at the moment but
this does not seem urgent either, which is good.

I think that applications that are hardware specific would still often
prefer hard-coding the pad number there, but instead use the new interface.
Do you think we could guarantee the same pad number stability there, as long
as no dynamic hardware changes affecting the pads are made? Such
applications couldn't cope with those changes anyway.

> 
> > The pad index is indeed just a number but it's an important one. If we
> > provide the pad ID (i.e. the graph object ID) to the user space, it has none
> > of the three properties, and missing even one would be enough to break the
> > user space.
> 
> No userspace will be broken on v1 ioctls, as it will still be there.
> 
> > Then, it indeed becomes a naming issue. Names can be stable even if the
> > index wouldn't be. The user would convert the names to integers that may not
> > be stable.
> 
> Yes, I guess a name is the best alternative here, as it fits both the
> cases where a simple number index is enough and cases where userspace
> need to use some other criteria to detect the PAD he needs.

I think the names would have to be entity specific. Otherwise we'll have the
same entity naming problem there.

> 
> > Do you have a use case for dynamically adding or removing pads --- wouldn't
> > it take a change in hardware topology to do that? If it's software only, I'd
> > definitely favour solutions that didn't involve changes in topology ---
> > software changes are generally very fast, and changing the media device
> > topology for that could also affect the device nodes, forcing the user space
> > to wait for udev to create them. That's very inconvenient when you want to
> > take a photo, for instance.
> 
> I'm pretty sure we'll need dynamic pad addition/removal in some future,
> but they're not of the first use cases I want to address.
> 
> There are some hardware that allows to dynamically reconfigure it to
> provide more PADs. I've seen two different cases of that, at the DVB
> side:
> 
> 1) One hardware with a fixed number of pads (like 4 sink pads and
>    40 source pads), where it is capable of being dynamically
>    configured to become 4 independent hardware entities, each one with
>    one sink pad and a number of source pads. So, it could start like
>    4 entities, with one sink and 10 sources. However, if user needs
>    more than 10 sources, it can send a command to the hardware to get
>    some unused PADs from the other entities;

This is not completely unlike many CSI-2 receivers which can share the lanes
found in the receiver block between different receivers, say dividing five
receiver lanes in 4:1:0 or 2:2:1 configuration among three receivers
depending on the connected sensors. Those are not shown to the user space at
all.

Althrough that's configuration made in board design time whereas yours is
runtime if I understand you correctly.

What are these 40 pads in your example? Is it a DVB demux or something else?

I'm not that much worried about the number alone, but if the pads are
indistinguishable from each other, I'd look for other ways to expose (or
avoid exposing) them to the user.

> 
> 2) Firmware or FPGA-based devices, where you can dynamically
>    reconfigure the hardware any time. We have one such driver
>    merged for 4.3 whose vendor is promising to release the
>    FPGA code as an Open Source, in order to allow users to
>    change the hardware in real time.

If you replace the FPGA configuration with another, it's essentially a new
piece of hardware which will require its own driver. Of course you could do
that with some limitations, but you have to take it into account in the
driver in that case.

> 
> The first case could actually be handled by always creating 4
> entities, each with 40 source pads, and adding a new flag to tell
> that a pad was not actually created yet, but this sounds hacky
> and would allocate 4 times more memory for the PADs than actually
> needed. Btw, the memory allocation is already a problem on one of
> the hardwares I know that uses this device, as the DVB core doesn't
> support dynamic filter reconfiguration neither. The amount of memory
> required there to handle each source PAD on a demux is not small,
> as one ringbuffer to handle the I/O transfer is needed for each
> PAD source. We're working to fix it at the DVB side, but having
> the DVB to use a different number of PADs than what's there at
> the MC would be really messy.
> 
> For the second case, I don't know any alternative but to support
> dynamic PAD changes.
> 
> Please notice that, on DVB, I don't see any usage for subdevs, so 
> there's no need to wait for udev to create new devices to address 
> those new entities/pads.
> 
> The second case, e. g. a Firmware/FPGA-based device can also work at
> the V4L2 side. I remember someone (from Intel - I guess) commenting
> about one such hardware on one of our media workshops, but it was
> for some hardware he was not allowed to upstream.
> 
> Yeah, assuming that someone would upstream such driver and we would
> need to find a solution, yeah, reconfiguration while taking a photo
> can be very inconvenient.

Yes. As the "hardware topology" of such a device is essentially defined by
software, I wouldn't try to represent it as hardware. There's a subtle but
significant difference to FPGA devices: FPGAs are essentially hardware,
albeit you can change the layout to whatever you like, whereas the devices
running software require software frameworks to support the device on both
sides (the CPU and the device firmware). They more resemble programmable
GPUs in that sense: even if what the device does (computes) changes, the
drivers do not (and should not).

> Perhaps the solution for such case would be to create the new
> pads/entities when the camera software would be loaded, and not
> when the user takes a photo.
> 
> I don't expect us to solve all issues related to dynamic PAD
> addition this year, but we should be sure that no API changes
> will be needed at G_TOPOLOGY when we add support for it, and
> I guess that the right thing here is to provide properties that
> would allow userspace to retrieve what's needed to uniquely
> identify a PAD on a stable manner. Eventually, some devices
> could just use an u32 pad_index properties, while others may
> use strings.

New properties (such as the name) could be added later on without affecting
the existing users that depend on the index.

-- 
Kind regards,

Sakari Ailus
e-mail: sakari.ailus@iki.fi	XMPP: sailus@retiisi.org.uk