Re: [RFC][PATCH 1/9] usb: hcd: Introduce usb_start/stop_hcd()

[Roger Quadros <rogerq@ti.com>]

On 20/03/15 08:32, Peter Chen wrote:
> On Thu, Mar 19, 2015 at 01:38:32PM +0200, Roger Quadros wrote:
>> On 18/03/15 21:49, Alan Stern wrote:
>>> On Wed, 18 Mar 2015, Roger Quadros wrote:
>>>
>>>> To support OTG we want a mechanism to start and stop
>>>> the HCD from the OTG state machine. Add usb_start_hcd()
>>>> and usb_stop_hcd().
>>>>
>>>> Signed-off-by: Roger Quadros <rogerq@ti.com>
>>>
>>> There are a few problems in this proposed patch.
>>>
>>>> +int usb_start_hcd(struct usb_hcd *hcd)
>>>> +{
>>>> +	int retval;
>>>> +	struct usb_device *rhdev = hcd->self.root_hub;
>>>> +
>>>> +	if (hcd->state != HC_STATE_HALT) {
>>>> +		dev_err(hcd->self.controller, "not starting a running HCD\n");
>>>> +		return -EINVAL;
>>>> +	}
>>>> +
>>>> +	hcd->state = HC_STATE_RUNNING;
>>>> +	retval = hcd->driver->start(hcd);
>>>> +	if (retval < 0) {
>>>> +		dev_err(hcd->self.controller, "startup error %d\n", retval);
>>>> +		hcd->state = HC_STATE_HALT;
>>>> +		return retval;
>>>> +	}
>>>> +
>>>> +	/* starting here, usbcore will pay attention to this root hub */
>>>> +	if ((retval = register_root_hub(hcd)) != 0)
>>>> +		goto err_register_root_hub;
>>>
>>> If the host controller is started more than once, you will end up
>>> unregistering and re-registering the root hub.  The device core does
>>> not allow this.  Once a device has been unregistered, you must not try
>>> to register it again -- you have to allocate a new device and register
>>> it instead.
>>
>> Understood.
>>
>>>
>>> Also, although you call the driver's ->start method multiple times, the
>>> ->reset method is called only once, when the controller is first 
>>> probed.  It's not clear that this will work in a situation where the HC 
>>> and the UDC share hardware state; after the UDC is stopped it may be 
>>> necessary to reset the HC before it can run again.
>>
>> Yes, good point.
>>>
>>> It might be possible to make this work, but I suspect quite a few 
>>> drivers would need rewriting first.  As another example of the problems 
>>> you face, consider how stopping a host controller will interact with 
>>> the driver's PM support (both system suspend and runtime suspend).
>>
>> Right. This needs more work than I thought.
>>>
>>> It would be a lot simpler to unbind the host controller driver
>>> completely when switching to device mode and rebind it when switching
>>> back.  I guess that is the sort of heavy-duty approach you want to
>>> avoid, but it may be the only practical way forward.
>>
>> So you mean directly calling usb_add/remove_hcd() from the OTG core?
>> I don't see any issues with that other than it being a heavy-duty operation
>> like you said and hope that it doesn't violate the OTG spec timing.
>>
>> Looking at Figure 5-3: "HNP Sequence of Events (FS)" of the OTG 2.0 spec
>> we have about 150ms (X10) to switch from B-Device detected A connect (b_wait_acon state)
>> to driving bus reset (b_host state). I don't think this should be an issue in modern SoCs
>> but I'm not very sure.
>>
> 
> It is not related toadd/remove hcd, it is the time from receiving PCD
> to issue BUS_RESET, the Linux stack can't satisfy OTG spec (150ms) due
> to there are some de-bounce waitings.

OK.

> 
>> In any case I can migrate to the add/remove hcd approach to simplify things.
>>
> 
> It should be no problem, we use it more than 1 years.
> 

Good to know this.

cheers,
-roger