Re: Chipidea USB device goes infinite loop due to interrupt while hw_ep_prime

On 21-08-09 10:45:16, Jeaho Hwang wrote:

2021년 8월 9일 (월) 오전 10:27, Peter Chen [off-list ref]님이 작성:

quoted

On 21-08-04 11:03:44, Jeaho Hwang wrote:

quoted

Hi. linux-usb and linux-rt experts.

quoted

On 21-08-02 17:35:01, Jeaho Hwang wrote:

quoted

Hi.

We found an infinite loop inside the function hw_ep_set_halt
(drivers/usb/chipidea/udc.c) if a cablle is repeatedly
connnected/disconnected while ping through RNDIS with chipidea USB device.

Using ftrace tracing, we found that hw_ep_set_halt is called due to error
return of hw_ep_prime(drivers/usb/chipidea/udc.c:202) which is called from
isr_tr_complete_handler -> isr_setup_status_phase -> _ep_queue.

The comment of function hw_ep_prime says (execute without interruption) but
timer interrupt is occurred while hw_ep_prime is executing. We believe that
the interrupt causes an error return of hw_ep_prime. We tried to protect
hw_ep_prime from irqs and then no case of the infinite loop is occurred.

I want ask if it is appropriate way that turning off irq inside (threaded)
irq handlers. And should we explicitly turn off irqs before calling
hw_ep_prime?

Jeaho, do you use RT-Linux or standard Linux? The function hw_ep_prime is
only called at udc_irq which is registered as top-half irq handlers.
Why the timer interrupt is occurred when hw_ep_prime is executing?

We use preempt_RT so timer interrupt could be occurred. Now I found
out that forced threaded irq handler disables local irq on standard
linux so It is a linux-rt issue. Then should I make patch which
disables local irqs during hw_ep_prime for RT kernel and suggest it to
linux-rt maintainers?

Thanks for a kind answer Peter.

Thanks Peter.

quoted

No, that will lead to deadlock since the normal request queue API function
ep_queue disables irq which also calls into hw_ep_prime.

ep_queue disables irq by calling spin_lock_irqsave but it actually
does not disable irq for RT kernel.
Therefore local_irq_save at hw_ep_prime would not cause deadlock if
ep_queue calls it, i think.

Is it anything wrong?

Oh, I was wrong that I thought you would use spin_lock_irqsave to
disable interrupt. I am not familiar with spin_lock_irqsave behaviours
at RT kernel.

Yes, if you use local_irq_save at below sequence, that's ok.

spin_lock_irqsave(lock, flag_1);
...
func hw_ep_prime
{
	local_irq_save(flag_2);
	...
	local_irq_restore(flag_2);
}
...
spin_unlock_irqrestore(lock, flag_1);

Peter

quoted

For RT kernel, you may try to disable local irq at isr_setup_status_phase.

--

Thanks,
Peter Chen

-- 
황재호, Jay Hwang, linux team manager of RTst
010-7242-1593

-- 

Thanks,
Peter Chen