Re: [PATCH] staging: vc04_services: vchiq_arm: Create platform_device per device

From: Laurent Pinchart
Date: Mon Dec 26 2022 - 05:05:04 EST


Hi Stefan,

On Fri, Dec 23, 2022 at 12:24:22PM +0100, Stefan Wahren wrote:
> Am 22.12.22 um 18:35 schrieb Laurent Pinchart:
> > On Thu, Dec 22, 2022 at 01:59:28PM +0530, Umang Jain wrote:
> >> On 12/21/22 6:40 PM, Greg Kroah-Hartman wrote:
> >>> On Wed, Dec 21, 2022 at 01:14:59PM +0200, Laurent Pinchart wrote:
> >>>> On Tue, Dec 20, 2022 at 02:14:04PM +0530, Umang Jain wrote:
> >>>>> Create a proper per device platorm_device structure for all the child
> >>>>> devices that needs to be registered by vchiq platform driver. Replace
> >>>>> the vchiq_register_child() with platform_add_devices() to register the
> >>>>> child devices.
> >>>>
> >>>> This explains what the patch does, but not why.
> >>>>
> >>>>> This is part of an effort to address TODO item "Get rid of all non
> >>>>> essential global structures and create a proper per device structure"
> >>>>
> >>>> And this explains part of the reason only. Could you please expand the
> >>>> commit message with the reasoning behind this change ? It's not clear
> >>>> from the change below why this is needed and good.
> >>
> >> Ok, I thought the TODO reference was sufficient but I'll expand on it.
> >>
> >>>>> Signed-off-by: Umang Jain <umang.jain@xxxxxxxxxxxxxxxx>
> >>>>> ---
> >>>>> .../interface/vchiq_arm/vchiq_arm.c | 59 ++++++++++---------
> >>>>> 1 file changed, 31 insertions(+), 28 deletions(-)
> >>>>>
> >>>>> diff --git a/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c b/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c
> >>>>> index 22de23f3af02..fa42ea3791a7 100644
> >>>>> --- a/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c
> >>>>> +++ b/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c
> >>>>> @@ -65,8 +65,29 @@ int vchiq_susp_log_level = VCHIQ_LOG_ERROR;
> >>>>> DEFINE_SPINLOCK(msg_queue_spinlock);
> >>>>> struct vchiq_state g_state;
> >>>>>
> >>>>> -static struct platform_device *bcm2835_camera;
> >>>>> -static struct platform_device *bcm2835_audio;
> >>>>> +static u64 vchiq_device_dmamask = DMA_BIT_MASK(32);
> >>>>
> >>>> The fact that this isn't const and is used by two different
> >>>> platform_device instances is worrying. Either it can be made const, or
> >>>> it's wrong.
> >>
> >> ack.
> >>
> >>>>> +
> >>>>> +static struct platform_device bcm2835_camera = {
> >>>>> + .name = "bcm2835-camera",
> >>>>> + .id = PLATFORM_DEVID_NONE,
> >>>>> + .dev = {
> >>>>> + .dma_mask = &vchiq_device_dmamask,
> >>>>> + }
> >>>>> +};
> >>>>> +
> >>>>> +static struct platform_device bcm2835_audio = {
> >>>>> + .name = "bcm2835_audio",
> >>>>> + .id = PLATFORM_DEVID_NONE,
> >>>>> + .dev = {
> >>>>> + .dma_mask = &vchiq_device_dmamask,
> >>>>> + }
> >>>>> +
> >>>>
> >>>> Extra blank line.
> >>
> >> oops, checkpatch.pl didn't catch this :-/
> >>
> >>>>> +};
> >>>>> +
> >>>>> +static struct platform_device *vchiq_devices[] __initdata = {
> >>>>
> >>>> Make it const.
> >>>>
> >>>>> + &bcm2835_camera,
> >>>>> + &bcm2835_audio,
> >>>>> +};
> >>>>>
> >>>>> struct vchiq_drvdata {
> >>>>> const unsigned int cache_line_size;
> >>>>> @@ -1763,28 +1784,6 @@ static const struct of_device_id vchiq_of_match[] = {
> >>>>> };
> >>>>> MODULE_DEVICE_TABLE(of, vchiq_of_match);
> >>>>>
> >>>>> -static struct platform_device *
> >>>>> -vchiq_register_child(struct platform_device *pdev, const char *name)
> >>>>> -{
> >>>>> - struct platform_device_info pdevinfo;
> >>>>> - struct platform_device *child;
> >>>>> -
> >>>>> - memset(&pdevinfo, 0, sizeof(pdevinfo));
> >>>>> -
> >>>>> - pdevinfo.parent = &pdev->dev;
> >>>>> - pdevinfo.name = name;
> >>>>> - pdevinfo.id = PLATFORM_DEVID_NONE;
> >>>>> - pdevinfo.dma_mask = DMA_BIT_MASK(32);
> >>>>> -
> >>>>> - child = platform_device_register_full(&pdevinfo);
> >>>>> - if (IS_ERR(child)) {
> >>>>> - dev_warn(&pdev->dev, "%s not registered\n", name);
> >>>>> - child = NULL;
> >>>>> - }
> >>>>> -
> >>>>> - return child;
> >>>>> -}
> >>>>> -
> >>>>> static int vchiq_probe(struct platform_device *pdev)
> >>>>> {
> >>>>> struct device_node *fw_node;
> >>>>> @@ -1832,8 +1831,11 @@ static int vchiq_probe(struct platform_device *pdev)
> >>>>> goto error_exit;
> >>>>> }
> >>>>>
> >>>>> - bcm2835_camera = vchiq_register_child(pdev, "bcm2835-camera");
> >>>>> - bcm2835_audio = vchiq_register_child(pdev, "bcm2835_audio");
> >>>>> + err = platform_add_devices(vchiq_devices, ARRAY_SIZE(vchiq_devices));
> >>>>> + if (err) {
> >>>>> + dev_warn(&pdev->dev, "Failed to add vchiq child devices");
> >>>>> + goto error_exit;
> >>>>> + }
> >>>>
> >>>> If you unbind and rebind this driver, the platform_device instances
> >>>> defined as global variables will be reused, and I'm pretty sure that
> >>>> will cause issues, for instance with the kobj->state_initialized check
> >>>> in kobject_init() (called from device_initialize(), itself called from
> >>>> platform_device_register(), from platform_add_devices()). I'm not sure
> >>>> static instances of platform_device are a very good idea in general.
> >>>
> >>> static instances of any device are a horrible idea, but it seems that
> >>> many drivers do this and abuse platform devices this way :(
> >>
> >> It seems  I have been a victim of the abuse usage while looking for
> >> platform_device references in the codebase. I'm working on a new
> >> approach for this.
> >>
> >> Currently (as per the linux-next branch), the vchiq driver will happily
> >> carry on if any of the child  platform device registration fails. That
> >> means if bcm2835-audio fails to register, bcm2835-camera will  still
> >> kept registered I suppose.
> >>
> >> However with usage of platform_add_devices() in this patch, I introduced
> >> a functionality change (I'm realizing this now) - any failure of child
> >> platform device registeration will -unregister- all the other platform
> >> devices i.e. if bcm2835-audio fails, bcm2835-camera will also get
> >> unregistered.
> >>
> >> Should I be working towards the status-quo behavior ? Or it's sane to
> >> unregistered other platform devices if one of the fails like
> >> platform_add_devices() does ? (This affects my new approach as well,
> >> hence the question)
> >
> > If it doesn't cause too much extra complexity, it would be nice to skip
> > devices that can't be registered successfully, and still support the
> > other ones. I don't expect registration failures to be a occuring
> > normally, so if this causes too much completely, I think it would still
> > be fine to fail more harshly.
> >
> >>> Ideally this should be done properly, with the correct devices created
> >>> automatically based on the device tree structure, NOT hard-coded into a
> >>> .c file like this.
> >>>
> >>> So I too really do not like this change, why are these not being created
> >>> by the firware layer automatically?
> >>
> >> Not sure if this is a helpful comment, but as far I know, there can be
> >> vchiq child platform devices which probably don't have a Device tree
> >> entry. like the bcm2835-isp [1] I posted earlier.
> >>
> >> [1] https://lore.kernel.org/lkml/20221121214722.22563-1-umang.jain@xxxxxxxxxxxxxxxx/
> >
> > Those devices are implemented and exposed by the firmware running on the
> > VC4. The device tree describes the VC4 itself with the resources
> > required to communicate with it through a mailbox interface. I was going
> > to say that the platform devices are then created based on what the
> > firmware exposes, but that's not right, they're indeed hardcoded in the
> > vchiq driver. Adding corresponding DT nodes (as children of the vchiq DT
> > node) could make sense. Dave, do you have any opinion on this ?
>
> i vaguely remember the discussion how to represent audio and camera
> interface in the device tree. Representing as child nodes of the VC4 has
> been rejected on the device tree mailing some years ago, because this
> doesn't represent the physical (hardware) wiring. It's still possible to
> access e.g. the camera interface from the ARM.

For the camera, things have changed a lot since the mail thread you've
linked. The CSI-2 receiver (and camera sensors) are now described in DT
and controlled from the ARM side. I believe the firmware still supports
controlling that hardware from the VC4 side (limited to a very small set
of camera sensors), but I think we can ignore that from a mainline point
of view.

The devices that are still controlled from the VC4 side are the camera
ISP, the video codec and the audio interface. As far as I can tell,
there's no plan to change this in neither the short term or long term
future. Based on my limited understanding, this architecture makes sense
for the ISP and codec as they share resources in a way that is best
handled by the VC4 firmware. I have no idea about the audio side. Dave,
please correct me if this is incorrect.

> The whole approach with using a separate binding for all the firmware
> stuff lead to a lot of trouble on the Raspberry Pi platform (ugly
> dependencies between firmware, DT and kernel). So i would like to avoid
> this here. In case the current implementation is a no go, how about
> letting the ARM core discover the available interfaces e.g. via mailbox
> interface?

I don't know if this is possible with existing firmware, and, if not, if
it could be implemented (the firmware isn't open-source). If not, we
will need to handle the current situation in the best possible way,
which would require creating devices either in the VCHIQ driver, or
through DT. I agree the former is probably best, there would still be a
dependency between the kernel and firmware, but DT would at least be out
of the picture. A custom bus seems fine to me.

> For more inspiration take a look at this old thread [1]
>
> But i agree DT binding for vchiq itself is also a TODO
>
> and any help is appreciated.
>
> [1] - http://lists.infradead.org/pipermail/linux-rpi-kernel/2017-February/005541.html

--
Regards,

Laurent Pinchart