Re: Handling incoming ZLP in cdc-wdm

[Oliver Neukum <oneukum@suse.com>]

On 02.04.25 13:57, Robert Hodaszi wrote:
> 
> 2025. 03. 31. 11:59 keltezéssel, Oliver Neukum írta:
>>
>>>      3. we should not set res_length to 0, otherwise  we get stuck again (actually, should we clear the "if (!desc->reslength)" in wdm_read(), since that's is no longer possible anymore?
>>
>> AFAICT it can happen if two threads are racing on wdm_read()
>>
> I'm not seeing how could we get stuck. I think, with the patch I sent previously, we can just simply remove reslength as is, as it is no longer used. Nothing sets it to 0 anymore, as in that case, the setter gets skipped, and the only consumer is wdm_read(), where, it is only checking for 0.

Hi,

I see what you mean, but I am afraid this introduces an error case.

Suppose we have the following scenario with your patch applied:

wdm_in_callback() gets a ZLP -> we schedule service_interrupt_work()
service_interrupt_work() calls service_outstanding_interrupt()
service_outstanding_interrupt() decrements resp_count to zero
service_interrupt_work() sets WDM_READ and wakes the waiting task

wdm_read() is woken up and finds an empty buffer
If reslength == 0 all is well and it returns -EAGAIN to user space
If reslength != 0 it will return EOF

We must not return EOF.
Hence the longer I think about it it seems to me that

1. reslength is necessary
2. You must set reslength to 0 whenever you get a ZLP

	Regards
		Oliver