Re: 答复: [PATCH] usb: gadget: u_serial: Fix the issue that gs_start_io crashed due to accessing null pointer

[Prashanth K <quic_prashk@quicinc.com>]

On 19-11-24 06:03 pm, 胡连勤 wrote:
> Hello Prashanth:
> 
>>> Considering that in some extreme cases, when u_serial driver is
>>> accessed by multiple threads, Thread A is executing the open operation
>>> and calling the gs_open, Thread B is executing the disconnect
>>> operation and calling the gserial_disconnect function,The
>>> port->port_usb pointer will be set to NULL.
>>>
>> [...]
>>> ---
>>>  drivers/usb/gadget/function/u_serial.c | 25 +++++++++++++++----------
>>>  1 file changed, 15 insertions(+), 10 deletions(-)
>>>
>>> diff --git a/drivers/usb/gadget/function/u_serial.c
>>> b/drivers/usb/gadget/function/u_serial.c
>>> index 0a8c05b2746b..9ab2dbed60a8 100644
>>> --- a/drivers/usb/gadget/function/u_serial.c
>>> +++ b/drivers/usb/gadget/function/u_serial.c
>>> @@ -124,6 +124,7 @@ struct gs_port {
>>>  	struct kfifo		port_write_buf;
>>>  	wait_queue_head_t	drain_wait;	/* wait while writes drain */
>>>  	bool                    write_busy;
>>> +	bool                    read_busy;
>>>  	wait_queue_head_t	close_wait;
>>>  	bool			suspended;	/* port suspended */
>>>  	bool			start_delayed;	/* delay start when
>> suspended */
>>> @@ -331,9 +332,11 @@ __acquires(&port->port_lock)
>>>  		/* drop lock while we call out; the controller driver
>>>  		 * may need to call us back (e.g. for disconnect)
>>>  		 */
>>> +		port->read_busy = true;
>>>  		spin_unlock(&port->port_lock);
>>>  		status = usb_ep_queue(out, req, GFP_ATOMIC);
>>>  		spin_lock(&port->port_lock);
>>> +		port->read_busy = false;
>>>
>>>  		if (status) {
>>>  			pr_debug("%s: %s %s err %d\n",
>>> @@ -1412,19 +1415,21 @@ void gserial_disconnect(struct gserial *gser)
>>>  	/* tell the TTY glue not to do I/O here any more */
>>>  	spin_lock(&port->port_lock);
>>>
>>> -	gs_console_disconnect(port);
>>> +	if (!port->read_busy) {
>> start_tx/rx rely on port->port_usb for queuing the requests, and if its not
>> null during disconnect, tx/rx would keep on queuing requests to UDC even
>> after disconnect (which is not ideal). Here in your case, after read_busy is set,
>> start_rx would queue something outside of spinlock, meanwhile disconnect
>> happens but port_usb is still valid (because read_busy is set) and start_rx
>> would break early. But start_tx could continue queuing into disconnected
>> UDC (if 'started' is non-zero, which could happen due to timing). Can't you try
>> something like this,
>>
>> --- a/drivers/usb/gadget/function/u_serial.c
>> +++ b/drivers/usb/gadget/function/u_serial.c
>> @@ -579,9 +579,12 @@ static int gs_start_io(struct gs_port *port)
>>                  * we didn't in gs_start_tx() */
>>                 tty_wakeup(port->port.tty);
>>         } else {
>> -               gs_free_requests(ep, head, &port->read_allocated);
>> -               gs_free_requests(port->port_usb->in, &port->write_pool,
>> -                       &port->write_allocated);
>> +               /* Free reqs only if we are still connected */
>> +               if (port->port_usb) {
>> +                       gs_free_requests(ep, head, &port->read_allocated);
>> +                       gs_free_requests(port->port_usb->in,
>> &port->write_pool,
>> +                               &port->write_allocated);
>> +               }
>>                 status = -EIO;
>>         }
>>
>> This will skip freeing reqs (and your crash) if port_usb is null and freeing
>> would be taken care by disconnect callback.
>>
>>
> First of all, the patch you gave can solve the problem we are currently facing.
> 
> When we first encountered this problem, we also thought about adding a null check operation to deal with it,
> but we saw that the entry of this function (gs_start_io) had a null check operation for port->port_usb, so I gave up
> the idea of ​​null check during free_req (maybe I made a simple problem complicated),
> and thought about optimizing it from the software logic, so that the port->port_usb pointer is always valid before gs_start_io is executed.
> 

If it solves the problem, i guess you can use the null pointer check as
I suggested and send a new patchset, the current one will introduce new
problems. Keep the issue analysis as it is in commit text since its
descriptive enough to understand the problem.

Regards,
Prashanth K