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

[Yunsheng Lin]

On 2020/9/2 17:20, Eric Dumazet wrote:
> 
> 
> On 9/2/20 1:14 AM, Yunsheng Lin wrote:
>> On 2020/9/2 15:32, Eric Dumazet wrote:
>>>
>>>
>>> On 9/1/20 11:34 PM, Yunsheng Lin wrote:
>>>
>>>>
>>>> I am not familiar with TCQ_F_CAN_BYPASS.
>>>> From my understanding, the problem is that there is no order between
>>>> qdisc enqueuing and qdisc reset.
>>>
>>> Thw qdisc_reset() should be done after rcu grace period, when there is guarantee no enqueue is in progress.
>>>
>>> qdisc_destroy() already has a qdisc_reset() call, I am not sure why qdisc_deactivate() is also calling qdisc_reset()
>>
>> That is a good point.
>> Do we allow skb left in qdisc when the qdisc is deactivated state?
>> And qdisc_destroy() is not always called after qdisc_deactivate() is called.
>> If we allow skb left in qdisc when the qdisc is deactivated state, then it is
>> huge change of semantics for qdisc_deactivate(), and I am not sure how many
>> cases will affected by this change.
> 
> All I am saying is that the qdisc_reset() in qdisc_deactivate() seems
> at the wrong place.
> 
> This certainly can be deferred _after_ the rcu grace period we already have in dev_deactivate_many()
> 
> Something like this nicer patch.

Thanks for clarifying and the nicer patch.

Reusing the existing synchronize_net() is a good idea.

Some minor comment inline.

> 
>  net/sched/sch_generic.c |   22 ++++------------------
>  1 file changed, 4 insertions(+), 18 deletions(-)
> 
> 
> diff --git a/net/sched/sch_generic.c b/net/sched/sch_generic.c
> index 265a61d011dfaa7ec0f8fb8aaede920784f690c9..0eaa99e4f8de643724c0942ee1a2e9516eed65c0 100644
> --- a/net/sched/sch_generic.c
> +++ b/net/sched/sch_generic.c
> @@ -1131,24 +1131,10 @@ EXPORT_SYMBOL(dev_activate);
>  
>  static void qdisc_deactivate(struct Qdisc *qdisc)
>  {
> -       bool nolock = qdisc->flags & TCQ_F_NOLOCK;
> -
>         if (qdisc->flags & TCQ_F_BUILTIN)
>                 return;
> -       if (test_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state))
> -               return;
> -
> -       if (nolock)
> -               spin_lock_bh(&qdisc->seqlock);
> -       spin_lock_bh(qdisc_lock(qdisc));
>  
>         set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
> -
> -       qdisc_reset(qdisc);
> -
> -       spin_unlock_bh(qdisc_lock(qdisc));
> -       if (nolock)
> -               spin_unlock_bh(&qdisc->seqlock);
>  }
>  
>  static void dev_deactivate_queue(struct net_device *dev,
> @@ -1184,6 +1170,9 @@ static bool some_qdisc_is_busy(struct net_device *dev)
>                 val = (qdisc_is_running(q) ||
>                        test_bit(__QDISC_STATE_SCHED, &q->state));
>  
> +               if (!val)
> +                       qdisc_reset(q);

It seems semantics for some_qdisc_is_busy() is changed, which does not only do
the checking, but also do the reseting?

Also, qdisc_reset() could be called multi times for the same qdisc if some_qdisc_is_busy()
return true multi times?

> +
>                 spin_unlock_bh(root_lock);
>  
>                 if (val)
> @@ -1213,10 +1202,7 @@ void dev_deactivate_many(struct list_head *head)
>                 dev_watchdog_down(dev);
>         }
>  
> -       /* Wait for outstanding qdisc-less dev_queue_xmit calls.
> -        * This is avoided if all devices are in dismantle phase :
> -        * Caller will call synchronize_net() for us
> -        */
> +       /* Wait for outstanding dev_queue_xmit calls. */
>         synchronize_net();
>  
>         /* Wait for outstanding qdisc_run calls. */
> 
> 
> .
>