usb: dwc3: Prevent indefinite sleep in _dwc3_set_mode during suspend/resume

[Felipe Balbi <balbi@kernel.org>]

Hi,

Baolin Wang <baolin.wang@linaro.org> writes:
>>>>>>>>>  void dwc3_gadget_exit(struct dwc3 *dwc)
>>>>>>>>>  {
>>>>>>>>> +  int epnum;
>>>>>>>>> +  unsigned long flags;
>>>>>>>>> +
>>>>>>>>> +  spin_lock_irqsave(&dwc->lock, flags);
>>>>>>>>> +  for (epnum = 2; epnum < DWC3_ENDPOINTS_NUM; epnum++) {
>>>>>>>>> +          struct dwc3_ep  *dep = dwc->eps[epnum];
>>>>>>>>> +
>>>>>>>>> +          if (!dep)
>>>>>>>>> +                  continue;
>>>>>>>>> +
>>>>>>>>> +          dep->flags &= ~DWC3_EP_END_TRANSFER_PENDING;
>>>>>>>>> +  }
>>>>>>>>> +  spin_unlock_irqrestore(&dwc->lock, flags);
>>>>>>>>> +
>>>>>>>>>    usb_del_gadget_udc(&dwc->gadget);
>>>>>>>>>    dwc3_gadget_free_endpoints(dwc);
>>>>>>>>
>>>>>>>> free endpoints is a better place for this. It's already going to free
>>>>>>>> the memory anyway. Might as well clear all flags to 0 there.
>>>>>>>>
>>>>>>>
>>>>>>> But it won't solve the deadlock issue. Since dwc3_gadget_free_endpoints()
>>>>>>> is called after usb_del_gadget_udc() and the deadlock happens when
>>>>>>>
>>>>>>> usb_del_gadget_udc()->udc_stop()->dwc3_gadget_stop()->wait_event_lock_irq()
>>>>>>>
>>>>>>> and DWC3_EP_END_TRANSFER_PENDING flag is set.
>>>>>>
>>>>>> indeed. Iterating twice over the entire endpoint list seems
>>>>>> wasteful. Perhaps we just shouldn't wait when removing the UDC since
>>>>>> that's essentially what this patch will do, right? If you clear the flag
>>>>>> before calling ->udc_stop(), this means the loop in dwc3_gadget_stop()
>>>>>> will do nothing. Might as well remove it.
>>>>>>
>>>>>
>>>>> This means that we will never wait for DWC3_EP_END_TRANSFER_PENDING to clear
>>>>> in dwc3_gadget_stop() like we used to. This is perfectly fine, right?
>>>>>
>>>>> It makes sense to me as dwc3_gadget_stop() calls __dwc3_gadget_stop() which
>>>>> masks all interrupts and nobody will ever clear that flag if it was set.
>>>>
>>>> I don't think so. It can not mask the endpoint events, please check
>>>> the events which will be masked in DEVTEN register. The reason why we
>>>> should wait for DWC3_EP_END_TRANSFER_PENDING to clear is that,
>>>> sometimes the DWC3_DEPEVT_EPCMDCMPLT event will be triggered later
>>>> than 100us, but now we may have freed the gadget irq which will cause
>>>> crash.
>>>
>>> We could mask command complete events as soon as ->udc_stop() is called,
>>> right? Hmm, actually, __dwc3_gadget_stop() already clears DEVTEN
>>> completely.
>>
>> But which bit in DEVTEN says Endpoint events are disabled?
>
> When we set up the DWC3_DEPCMD_ENDTRANSFER command in
> dwc3_stop_active_transfer(), we can do not set DWC3_DEPCMD_CMDIOC,
> then there will no endpoint command complete interrupts I think.
>
> cmd |= DWC3_DEPCMD_CMDIOC;

I remember some part of the databook mandating CMDIOC to be set. We
could test it out without and see if anything blows up. I would,
however, require a lengthy comment explaining that we're deviating from
databook revision x.yya, section foobar because $reasons. :-)