Re: [PATCH] usb: gadget: uvc_video: unlock before submitting a request to ep

[Thinh Nguyen]

On Fri, Nov 17, 2023, Thinh Nguyen wrote:
> Hi,
> 
> On Thu, Nov 16, 2023, Dan Scally wrote:
> > CC Thinh - sorry to bother you, just want to make sure we fix this in the right place.
> > 
> > On 08/11/2023 11:48, Kuen-Han Tsai wrote:
> > > On 02/11/2023 07:11, Piyush Mehta wrote:
> > > > There could be chances where the usb_ep_queue() could fail and trigger
> > > > complete() handler with error status. In this case, if usb_ep_queue()
> > > > is called with lock held and the triggered complete() handler is waiting
> > > > for the same lock to be cleared could result in a deadlock situation and
> > > > could result in system hang. To aviod this scenerio, call usb_ep_queue()
> > > > with lock removed. This patch does the same.
> > > I would like to provide more background information on this problem.
> > > 
> > > We met a deadlock issue on Android devices and the followings are stack traces.
> > > 
> > > [35845.978435][T18021] Core - Debugging Information for Hardlockup core(8) - locked CPUs mask (0x100)
> > > [35845.978442][T18021] Call trace:
> > > [*][T18021]  queued_spin_lock_slowpath+0x84/0x388
> > > [35845.978451][T18021]  uvc_video_complete+0x180/0x24c
> > > [35845.978458][T18021]  usb_gadget_giveback_request+0x38/0x14c
> > > [35845.978464][T18021]  dwc3_gadget_giveback+0xe4/0x218
> > > [35845.978469][T18021]  dwc3_gadget_ep_cleanup_cancelled_requests+0xc8/0x108
> > > [35845.978474][T18021]  __dwc3_gadget_kick_transfer+0x34c/0x368
> > > [35845.978479][T18021]  __dwc3_gadget_start_isoc+0x13c/0x3b8
> > > [35845.978483][T18021]  dwc3_gadget_ep_queue+0x150/0x2f0
> > > [35845.978488][T18021]  usb_ep_queue+0x58/0x16c
> > > [35845.978493][T18021]  uvcg_video_pump+0x22c/0x518
> > 
> > 
> > I note in the kerneldoc comment for usb_ep_queue() that calling .complete()
> > from within itself is specifically disallowed [1]:
> > 
> >     Note that @req's ->complete() callback must never be called from
> > 
> >     within usb_ep_queue() as that can create deadlock situations.
> > 
> > 
> > And it looks like that's what's happening here - is this something that
> > needs addressing in the dwc3 driver?
> > 
> 
> Looks like it. The issue is in dwc3. It should only affect isoc request
> queuing.
> 
> Can we try with this patch:
> 
> diff --git a/drivers/usb/dwc3/gadget.c b/drivers/usb/dwc3/gadget.c
> index 858fe4c299b7..37e08eed49d9 100644
> --- a/drivers/usb/dwc3/gadget.c
> +++ b/drivers/usb/dwc3/gadget.c
> @@ -1684,12 +1684,15 @@ static int __dwc3_gadget_kick_transfer(struct dwc3_ep *dep)
>  			dwc3_gadget_move_cancelled_request(req, DWC3_REQUEST_STATUS_DEQUEUED);
>  
>  		/* If ep isn't started, then there's no end transfer pending */
> -		if (!(dep->flags & DWC3_EP_END_TRANSFER_PENDING))
> +		if (!(dep->flags & DWC3_EP_PENDING_REQUEST) &&
> +		    !(dep->flags & DWC3_EP_END_TRANSFER_PENDING))
>  			dwc3_gadget_ep_cleanup_cancelled_requests(dep);
>  
>  		return ret;
>  	}
>  
> +	dep->flags &= ~DWC3_EP_PENDING_REQUEST;
> +
>  	if (dep->stream_capable && req->request.is_last &&
>  	    !DWC3_MST_CAPABLE(&dep->dwc->hwparams))
>  		dep->flags |= DWC3_EP_WAIT_TRANSFER_COMPLETE;
> 
> ---
> 

Actually, please ignore the above, that's not correct. I'll send out a
proper patch later.

Thanks,
Thinh