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

[Yunsheng Lin]

On 2020/9/1 14:48, Eric Dumazet wrote:
> 
> 
> On 8/31/20 5:55 PM, Yunsheng Lin wrote:
>> Currently there is concurrent reset and enqueue operation for the
>> same lockless qdisc when there is no lock to synchronize the
>> q->enqueue() in __dev_xmit_skb() with the qdisc reset operation in
>> qdisc_deactivate() called by dev_deactivate_queue(), which may cause
>> out-of-bounds access for priv->ring[] in hns3 driver if user has
>> requested a smaller queue num when __dev_xmit_skb() still enqueue a
>> skb with a larger queue_mapping after the corresponding qdisc is
>> reset, and call hns3_nic_net_xmit() with that skb later.
>>
>> Avoid the above concurrent op by calling synchronize_rcu_tasks()
>> after assigning new qdisc to dev_queue->qdisc and before calling
>> qdisc_deactivate() to make sure skb with larger queue_mapping
>> enqueued to old qdisc will always be reset when qdisc_deactivate()
>> is called.
>>
> 
> We request Fixes: tag for fixes in networking land.

ok.

Fixes: 6b3ba9146fe6 ("net: sched: allow qdiscs to handle locking")

> 
>> Signed-off-by: Yunsheng Lin <linyunsheng@xxxxxxxxxx>
>> ---
>>  net/sched/sch_generic.c | 7 ++++++-
>>  1 file changed, 6 insertions(+), 1 deletion(-)
>>
>> diff --git a/net/sched/sch_generic.c b/net/sched/sch_generic.c
>> index 265a61d..6e42237 100644
>> --- a/net/sched/sch_generic.c
>> +++ b/net/sched/sch_generic.c
>> @@ -1160,8 +1160,13 @@ static void dev_deactivate_queue(struct net_device *dev,
>>  
>>  	qdisc = rtnl_dereference(dev_queue->qdisc);
>>  	if (qdisc) {
>> -		qdisc_deactivate(qdisc);
>>  		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
>> +
>> +		/* Make sure lockless qdisc enqueuing is done with the
>> +		 * old qdisc in __dev_xmit_skb().
>> +		 */
>> +		synchronize_rcu_tasks();
> 
> This seems quite wrong, there is not a single use of synchronize_rcu_tasks() in net/,
> we probably do not want this.
> 
> I bet that synchronize_net() is appropriate, if not please explain/comment why we want this instead.

Using synchronize_net() seems more appropriate here, thanks.

> 
> Adding one synchronize_net() per TX queue is a killer for devices with 128 or 256 TX queues.
> 
> I would rather find a way of not calling qdisc_reset() from qdisc_deactivate().

Without calling qdisc_reset(), it seems there will always be skb left in the old qdisc.
Is above acceptable?

How about below steps to avoid the concurrent op:
1. assign new qdisc to all queue' qdisc(which is noop_qdisc).
2. call synchronize_net().
3. calling qdisc_reset() with all queue' qdisc_sleeping.

And the synchronize_net() in dev_deactivate_many() can be reused to
ensure old qdisc is not touched any more when calling qdisc_reset().


> 
> This lockless pfifo_fast is a mess really.
> 
> 
> .
>