Re: [BUG] pfifo_fast may cause out-of-order CAN frame transmission

[Ahmad Fatoum <a.fatoum@pengutronix.de>]

Hello Paolo,

On 2/6/20 2:21 PM, Paolo Abeni wrote:
> On Tue, 2020-02-04 at 17:25 +0100, Ahmad Fatoum wrote:
>> Hello Paolo,
>>
>> On 1/20/20 5:06 PM, Ahmad Fatoum wrote:
>>> Hello Paolo,
>>>
>>> On 1/16/20 1:40 PM, Paolo Abeni wrote:
>>>> I'm sorry for this trial & error experience. I tried to reproduce the
>>>> issue on top of the vcan virtual device, but it looks like it requires
>>>> the timing imposed by a real device, and it's missing here (TL;DR: I
>>>> can't reproduce the issue locally).
>>>
>>> No worries. I don't mind testing.
>>>
>>>> Code wise, the 2nd patch closed a possible race, but it dumbly re-
>>>> opened the one addressed by the first attempt - the 'empty' field must
>>>> be cleared prior to the trylock operation, or we may end-up with such
>>>> field set and the queue not empty.
>>>>
>>>> So, could you please try the following code?
>>>
>>> Unfortunately, I still see observe reodering.
>>
>> Any news?
> 
> I'm unable to find any better solution than a revert. That will cost
> some small performace regression, so I'm a bit reluctant to go ahead.
> If there is agreement I can post the revert.

Could you draft the revert?
d518d2ed864 ("net/sched: fix race between deactivation and dequeue for NOLOCK qdisc")
looks applicable even with the revert, so I don't want to inadvertently cause
a regression by wrongly reverting.

Cheers
Ahmad


> 
> Cheers,
> 
> Paolo
> 
> 

-- 
Pengutronix e.K.                           |                             |
Steuerwalder Str. 21                       | http://www.pengutronix.de/  |
31137 Hildesheim, Germany                  | Phone: +49-5121-206917-0    |
Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-5555 |