Re: [PATCH v2] usb: core: Solve race condition in anchor cleanup functions

[Eli Billauer <eli.billauer@gmail.com>]

On 29/07/20 16:38, Alan Stern wrote:
> With a small amount of restructuring you can eliminate three unlock-lock
> pairs and avoid the need for usb_anchor_safe_empty():
> [ ... ]
> All you have to do is move this spin_lock_irq() above the start of the
> outer loop...
> [ ... ]
> .. and move this spin_unlock_irq() below the end of the outer loop.
> Likewise for the two other routines.
>
>    
I'm afraid that might not work. The whole purpose of the outer loop is 
to kick in when urb_list is empty, but there's this unanchor-completer 
race going on. So the inner loop will be skipped, because 
list_empty(&anchor->urb_list) will evaluate true. As a result, the 
spinlock will be held as the loop spins, until the completer has finished.

But if the completer tries to take the same lock, we're deadlocked. For 
example, if it resubmits the URB, which is pretty much the point of this 
extra while loop.

This is also the reason why I didn't just modify the original 
while-loop's condition, so it would go on spinning as long the race 
condition is in effect. It mustn't spin with the lock held.
>> >  +	} while (unlikely(!usb_anchor_safe_empty(anchor)));
>>      
> likely() and unlikely() are frowned upon unless you can provide actual
> measurements showing that they make a significant difference.  In this
> case they don't matter, since the bottleneck is the usb_kill_urb() call.
>    
The irony is that I added this "unlikely" for the human reader, and not 
for the compiler: I wanted to communicate that the outer loop is 
unlikely to kick in. I'll keep that in mind for v3 of this patch.

Thanks,
    Eli