Re: [PATCH v2 net] net: sch_generic: aviod concurrent reset and enqueue op for lockless qdisc

[Yunsheng Lin]

On 2020/11/3 0:55, Cong Wang wrote:
> On Fri, Oct 30, 2020 at 12:38 AM Yunsheng Lin <linyunsheng@xxxxxxxxxx> wrote:
>>
>> On 2020/10/30 3:05, Cong Wang wrote:
>>>
>>> I do not see how and why it should. synchronize_net() is merely an optimized
>>> version of synchronize_rcu(), it should wait for RCU readers, softirqs are not
>>> necessarily RCU readers, net_tx_action() does not take RCU read lock either.
>>
>> Ok, make sense.
>>
>> Taking RCU read lock in net_tx_action() does not seems to solve the problem,
>> what about the time window between __netif_reschedule() and net_tx_action()?
>>
>> It seems we need to re-dereference the qdisc whenever RCU read lock is released
>> and qdisc is still in sd->output_queue or wait for the sd->output_queue to drain?
> 
> Not suggesting you to take RCU read lock. We already wait for TX action with
> a loop of sleep. To me, the only thing missing is just moving the
> reset after that
> wait.

__QDISC_STATE_SCHED is cleared before calling qdisc_run() in net_tx_action(),
some_qdisc_is_busy does not seem to wait fully for TX action, at least
qdisc is still being accessed even if __QDISC_STATE_DEACTIVATED is set.

> 
> 
>>>>>> If we do any additional reset that is not related to qdisc in dev_reset_queue(), we
>>>>>> can move it after some_qdisc_is_busy() checking.
>>>>>
>>>>> I am not suggesting to do an additional reset, I am suggesting to move
>>>>> your reset after the busy waiting.
>>>>
>>>> There maybe a deadlock here if we reset the qdisc after the some_qdisc_is_busy() checking,
>>>> because some_qdisc_is_busy() may require the qdisc reset to clear the skb, so that
>>>
>>> some_qdisc_is_busy() checks the status of qdisc, not the skb queue.
>>
>> Is there any reason why we do not check the skb queue in the dqisc?
>> It seems there may be skb left when netdev is deactivated, maybe at least warn
>> about that when there is still skb left when netdev is deactivated?
>> Is that why we call qdisc_reset() to clear the leftover skb in qdisc_destroy()?
>>
>>>
>>>
>>>> some_qdisc_is_busy() can return false. I am not sure this is really a problem, but
>>>> sch_direct_xmit() may requeue the skb when dev_hard_start_xmit return TX_BUSY.
>>>
>>> Sounds like another reason we should move the reset as late as possible?
>>
>> Why?
> 
> You said "sch_direct_xmit() may requeue the skb", I agree. I assume you mean
> net_tx_action() calls sch_direct_xmit() which does the requeue then races with
> reset. No?
> 

Look at current code again, I think there is no race between sch_direct_xmit()
in net_tx_action() and dev_reset_queue() in dev_deactivate_many(), because
qdisc_lock(qdisc) or qdisc->seqlock has been taken when calling sch_direct_xmit()
or dev_reset_queue().


> 
>>
>> There current netdev down order is mainly below:
>>
>> netif_tx_stop_all_queues()
>>
>> dev_deactivate_queue()
>>
>> synchronize_net()
>>
>> dev_reset_queue()
>>
>> some_qdisc_is_busy()
>>
>>
>> You suggest to change it to below order, right?
>>
>> netif_tx_stop_all_queues()
>>
>> dev_deactivate_queue()
>>
>> synchronize_net()
>>
>> some_qdisc_is_busy()
>>
>> dev_reset_queue()
> 
> Yes.
> 
>>
>>
>> What is the semantics of some_qdisc_is_busy()?
> 
> Waiting for flying TX action.

It wait for __QDISC_STATE_SCHED to clear and qdisc running to finish, but
there is still time window between __QDISC_STATE_SCHED clearing and qdisc
running, right?

> 
>> From my understanding, we can do anything about the old qdisc (including
>> destorying the old qdisc) after some_qdisc_is_busy() return false.
> 
> But the current code does the reset _before_ some_qdisc_is_busy(). ;)

If lock is taken when doing reset, it does not matter if the reset is
before some_qdisc_is_busy(), right?

> 
> Thanks.
> .
>