usb: gadget: Kernel panic (NULL pointer dereference) when using fsl_udc2_core on i.MX31 PDK

[Magnus Lilja]

On 24 January 2017 at 19:34, Felipe Balbi <balbi at kernel.org> wrote:
>
> Hi,
>
> Magnus Lilja <lilja.magnus at gmail.com> writes:
>>> Magnus Lilja <lilja.magnus at gmail.com> writes:
>>>>> Magnus Lilja <lilja.magnus at gmail.com> writes:
>>>>>>>> I tried the fsl_udc_core gadget driver on the i.MX31 PDK board and got a
>>>>>>>> kernel panic (NULL pointer dereference) when connecting the USB cable. I
>>>>>>>> had the g_serial module loaded as well.
>>>>>>>>
>>>>>>>> The NULL pointer panic comes from gadget/udc/core.c
>>>>>>>> usb_gadget_giveback_request() which calls req->complete() and in some
>>>>>>>> cases req->complete is NULL.
>>>>>>>>
>>>>>>>> Commit 304f7e5e1d08 ("usb: gadget: Refactor request completion") changed
>>>>>>>> fsl_udc2_core.c (and several other files) and in fsl_udc2_core.c a check
>>>>>>>> that req->complete is non-NULL was removed:
>>>>>>>>
>>>>>>>> --- a/drivers/usb/gadget/udc/fsl_udc_core.c
>>>>>>>> +++ b/drivers/usb/gadget/udc/fsl_udc_core.c
>>>>>>>> @@ -197,10 +197,8 @@ __acquires(ep->udc->lock)
>>>>>>>>          ep->stopped = 1;
>>>>>>>>
>>>>>>>>          spin_unlock(&ep->udc->lock);
>>>>>>>> -       /* complete() is from gadget layer,
>>>>>>>> -        * eg fsg->bulk_in_complete() */
>>>>>>>> -       if (req->req.complete)
>>>>>>>> -               req->req.complete(&ep->ep, &req->req);
>>>>>>>> +
>>>>>>>> +       usb_gadget_giveback_request(&ep->ep, &req->req);
>>>>>>>>
>>>>>>>>          spin_lock(&ep->udc->lock);
>>>>>>>>          ep->stopped = stopped;
>>>>>>>>
>>>>>>>> If I re-introduce the check (either in fsl_udc_core.c or core.c) at
>>>>>>>> least USB gadget operation using g_serial seems to work just fine.
>>>>>>>>
>>>>>>>> I don't know the logic in detail to understand whether this is a proper
>>>>>>>> fix or if there is some other more problem with the fls_udc_core driver.
>>>>>>>> Does anyone have input in this matter?
>>>>>>>>
>>>>>>>> I can produce a proper patch that fixes this problem by re-introducing
>>>>>>>> the check (in either fsl_udc_core.c or core.c) if that is a proper
>>>>>>>> solution and I can also assist in testing other fixes to the problem.
>>>>>>>
>>>>>>> ->complete() is supposed to be mandatory. Which gadget do you have that
>>>>>>> ->doesn't set ->complete() to a valid function pointer?
>>>>>>
>>>>>> I'm modprobing g_serial so the following modules are loaded (using my patch):
>>>>>>
>>>>>> ~ # lsmod
>>>>>> usb_f_acm
>>>>>> u_serial
>>>>>> g_serial
>>>>>> libcomposite
>>>>>> configfs
>>>>>> fsl_usb2_udc
>>>>>
>>>>> okay, can you figure out which request is coming without ->complete()
>>>>> set? To which endpoint is this request being queued? It would be nice to
>>>>> know these details. Maybe this is an old bug which ought to be fixed.
>>>>
>>>> Sure, I can try figure that out. Any input to make the debug of the
>>>> faster is appreciated if you have any.
>>>
>>> well, the easiest way is to add something like:
>>>
>>>         if (!req->complete)
>>>                 dump_stack();
>>>
>>> to fsl udc driver. Then you would know who queued the request without
>>> ->complete. A slightly better approach would be to:
>>>
>>> if (WARN(!req->complete,
>>>         "%s: queueing request without ->complete\n", ep->name))
>>>         return;
>>>
>>> Or something like that.
>>
>> Well, I think I found it.
>>
>> fsl_udc_core.c:ep0_prime_status() sets req->req.complete = NULL before
>> it queues a transfer and my printk()'s indicate that this is indeed
>> the offending function.
>>
>> fsl_udc_core.c:ch9getstatus() also sets complete to NULL but in my
>> tests right now I haven't seen that one.
>>
>> So it's an internal problem in the fsl_udc_core.c file.
>
> seems like it. It's rather odd that fsl_udc doesn't wanna know about
> completion of Status stage. Oh well, I guess in this case it doesn't
> matter if you add a complete function or reinstate the previous check
> for valid complete.
>
> If you decide to reinstate the check, please add a comment above the
> check explaining that fsl_udc itself queues requests with NULL
> ->complete().
>
> I must say, however, that I would suggest adding a complete callback
> since that will help us BUG with NULL pointer deref on bad gadget
> drivers ;-)

I can do that. Such a complete() callback function would be a no-op
then I assume (with a comment in it why it is a no-op).

Regards, Magnus