Re: [PATCH v4] media: uvcvideo: Fix race condition with usb_kill_urb

[Laurent Pinchart]

Hi Ricardo,

Thank you for the patch.

On Wed, Dec 14, 2022 at 05:31:55PM +0100, Ricardo Ribalda wrote:
> usb_kill_urb warranties that all the handlers are finished when it
> returns, but does not protect against threads that might be handling
> asynchronously the urb.
> 
> For UVC, the function uvc_ctrl_status_event_async() takes care of
> control changes asynchronously.
> 
>  If the code is executed in the following order:
> 
> CPU 0					CPU 1
> ===== 					=====
> uvc_status_complete()
> 					uvc_status_stop()
> uvc_ctrl_status_event_work()
> 					uvc_status_start() -> FAIL
> 
> Then uvc_status_start will keep failing and this error will be shown:
> 
> <4>[    5.540139] URB 0000000000000000 submitted while active
> drivers/usb/core/urb.c:378 usb_submit_urb+0x4c3/0x528
> 
> Let's improve the current situation, by not re-submiting the urb if
> we are stopping the status event. Also process the queued work
> (if any) during stop.
> 
> CPU 0					CPU 1
> ===== 					=====
> uvc_status_complete()
> 					uvc_status_stop()
> 					uvc_status_start()
> uvc_ctrl_status_event_work() -> FAIL
> 
> Hopefully, with the usb layer protection this should be enough to cover
> all the cases.
> 
> Cc: stable@xxxxxxxxxxxxxxx
> Fixes: e5225c820c05 ("media: uvcvideo: Send a control event when a Control Change interrupt arrives")
> Reviewed-by: Yunke Cao <yunkec@xxxxxxxxxxxx>
> Signed-off-by: Ricardo Ribalda <ribalda@xxxxxxxxxxxx>
> ---
> uvc: Fix race condition on uvc
> 
> Make sure that all the async work is finished when we stop the status urb.
> 
> To: Yunke Cao <yunkec@xxxxxxxxxxxx>
> To: Sergey Senozhatsky <senozhatsky@xxxxxxxxxxxx>
> To: Max Staudt <mstaudt@xxxxxxxxxx>
> To: Laurent Pinchart <laurent.pinchart@xxxxxxxxxxxxxxxx>
> To: Mauro Carvalho Chehab <mchehab@xxxxxxxxxx>
> Cc: linux-media@xxxxxxxxxxxxxxx
> Cc: linux-kernel@xxxxxxxxxxxxxxx
> ---
> Changes in v4:
> - Replace bool with atomic_t to avoid compiler reordering
> - First complete the async work and then kill the urb to avoid race (Thanks Laurent!)
> - Link to v3: https://lore.kernel.org/r/20221212-uvc-race-v3-0-954efc752c9a@xxxxxxxxxxxx
> 
> Changes in v3:
> - Remove the patch for dev->status, makes more sense in another series, and makes
>   the zero day less nervous.
> - Update reviewed-by (thanks Yunke!).
> - Link to v2: https://lore.kernel.org/r/20221212-uvc-race-v2-0-54496cc3b8ab@xxxxxxxxxxxx
> 
> Changes in v2:
> - Add a patch for not kalloc dev->status
> - Redo the logic mechanism, so it also works with suspend (Thanks Yunke!)
> - Link to v1: https://lore.kernel.org/r/20221212-uvc-race-v1-0-c52e1783c31d@xxxxxxxxxxxx
> ---
>  drivers/media/usb/uvc/uvc_ctrl.c   | 3 +++
>  drivers/media/usb/uvc/uvc_status.c | 6 ++++++
>  drivers/media/usb/uvc/uvcvideo.h   | 1 +
>  3 files changed, 10 insertions(+)
> 
> diff --git a/drivers/media/usb/uvc/uvc_ctrl.c b/drivers/media/usb/uvc/uvc_ctrl.c
> index c95a2229f4fa..1be6897a7d6d 100644
> --- a/drivers/media/usb/uvc/uvc_ctrl.c
> +++ b/drivers/media/usb/uvc/uvc_ctrl.c
> @@ -1442,6 +1442,9 @@ static void uvc_ctrl_status_event_work(struct work_struct *work)
>  
>  	uvc_ctrl_status_event(w->chain, w->ctrl, w->data);
>  
> +	if (atomic_read(&dev->flush_status))
> +		return;
> +
>  	/* Resubmit the URB. */
>  	w->urb->interval = dev->int_ep->desc.bInterval;
>  	ret = usb_submit_urb(w->urb, GFP_KERNEL);
> diff --git a/drivers/media/usb/uvc/uvc_status.c b/drivers/media/usb/uvc/uvc_status.c
> index 7518ffce22ed..4a95850cdc1b 100644
> --- a/drivers/media/usb/uvc/uvc_status.c
> +++ b/drivers/media/usb/uvc/uvc_status.c
> @@ -304,10 +304,16 @@ int uvc_status_start(struct uvc_device *dev, gfp_t flags)
>  	if (dev->int_urb == NULL)
>  		return 0;
>  
> +	atomic_set(&dev->flush_status, 0);
>  	return usb_submit_urb(dev->int_urb, flags);
>  }
>  
>  void uvc_status_stop(struct uvc_device *dev)
>  {
> +	struct uvc_ctrl_work *w = &dev->async_ctrl;
> +
> +	atomic_set(&dev->flush_status, 1);

Note that atomic_read() and atomic_set() do no imply any memory barrier
on most architectures, as far as I can tell. They essentially become
READ_ONCE() and WRITE_ONCE() calls, which guarantee that the compiler
will not merge or split loads or stores, or reorder them with respect to
load and stores to the *same* memory location, but nothing else. I think
you need to add proper barriers, and you can then probably also drop
usage of atomic_t.

> +	if (cancel_work_sync(&w->work))
> +		uvc_ctrl_status_event(w->chain, w->ctrl, w->data);
>  	usb_kill_urb(dev->int_urb);

This should get rid of the main race (possibly save the barrier issue),
but it's not the most efficient option, and can still be problematic.
Consider the following case:

CPU0							CPU1
----							----

void uvc_status_stop(struct uvc_device *dev)		uvc_ctrl_status_event_async()
{							{
	...
	atomic_set(&dev->flush_status, 1);			...
	if (cancel_work_sync())
		...
								schedule_work();
	usb_kill_urb();					}
}

The usb_kill_urb() call ensures that uvc_ctrl_status_event_async()
completes before uvc_status_stop() returns, but there will still be work
scheduled in that case. uvc_ctrl_status_event_work() will be run later,
and as flush_status is set to 1, the function will not resubmit the URB.
That fixes the main race, but leaves the asynchronous control status
event handling for after uvc_status_stop() returns, which isn't great.

Now, if we consider that uvc_status_start() could be called shortly
after uvc_status_stop(), we may get in a situation where
uvc_status_start() will reset flush_status to 0 before
uvc_ctrl_status_event_async() runs. Both uvc_ctrl_status_event_async()
and uvc_status_start() will thus submit the same URB.

You can't fix this by first killing the URB and then cancelling the
work, as there would then be a risk that uvc_ctrl_status_event_work()
would be running in parallel, going past the flush_status check before
flush_status gets set to 1 in uvc_status_stop(), and submitting the URB
after usb_kill_urb() is called.

I think a good fix would be to check flush_status in
uvc_ctrl_status_event_async() and avoid the schedule_work() call in that
case. You could then also move the atomic_set(..., 0) call from
uvc_status_start() to the end of uvc_status_stop() (again with proper
barriers).

Could you please check the memory barriers and test the above proposal
(after double-checking it of course, I may have missed something) ?

>  }
> diff --git a/drivers/media/usb/uvc/uvcvideo.h b/drivers/media/usb/uvc/uvcvideo.h
> index df93db259312..1274691f157f 100644
> --- a/drivers/media/usb/uvc/uvcvideo.h
> +++ b/drivers/media/usb/uvc/uvcvideo.h
> @@ -560,6 +560,7 @@ struct uvc_device {
>  	struct usb_host_endpoint *int_ep;
>  	struct urb *int_urb;
>  	u8 *status;
> +	atomic_t flush_status;
>  	struct input_dev *input;
>  	char input_phys[64];
>  

-- 
Regards,

Laurent Pinchart