Re: Packet gets stuck in NOLOCK pfifo_fast qdisc

[Jonas Bonn <jonas.bonn@netrounds.com>]

Hi Paolo,

On 09/10/2019 21:14, Paolo Abeni wrote:
> On Wed, 2019-10-09 at 08:46 +0200, Jonas Bonn wrote:
>> Hi,
>>
>> The lockless pfifo_fast qdisc has an issue with packets getting stuck in
>> the queue.  What appears to happen is:
>>
>> i)  Thread 1 holds the 'seqlock' on the qdisc and dequeues packets.
>> ii)  Thread 1 dequeues the last packet in the queue.
>> iii)  Thread 1 iterates through the qdisc->dequeue function again and
>> determines that the queue is empty.
>>
>> iv)  Thread 2 queues up a packet.  Since 'seqlock' is busy, it just
>> assumes the packet will be dequeued by whoever is holding the lock.
>>
>> v)  Thread 1 releases 'seqlock'.
>>
>> After v), nobody will check if there are packets in the queue until a
>> new packet is enqueued.  Thereby, the packet enqueued by Thread 2 may be
>> delayed indefinitely.
> 
> I think you are right.
> 
> It looks like this possible race is present since the initial lockless
> implementation - commit 6b3ba9146fe6 ("net: sched: allow qdiscs to
> handle locking")
> 
> Anyhow the racing windows looks quite tiny - I never observed that
> issue in my tests. Do you have a working reproducer?

Yes, it's reliably reproducible.  We do network latency measurements and 
latency spikes for these packets that get stuck in the queue.

> 
> Something alike the following code - completely untested - can possibly
> address the issue, but it's a bit rough and I would prefer not adding
> additonal complexity to the lockless qdiscs, can you please have a spin
> a it?

Your change looks reasonable.  I'll give it a try.


> 
> Thanks,
> 
> Paolo
> ---
> diff --git a/include/net/pkt_sched.h b/include/net/pkt_sched.h
> index 6a70845bd9ab..65a1c03330d6 100644
> --- a/include/net/pkt_sched.h
> +++ b/include/net/pkt_sched.h
> @@ -113,18 +113,23 @@ bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
>   		     struct net_device *dev, struct netdev_queue *txq,
>   		     spinlock_t *root_lock, bool validate);
>   
> -void __qdisc_run(struct Qdisc *q);
> +int __qdisc_run(struct Qdisc *q);
>   
>   static inline void qdisc_run(struct Qdisc *q)
>   {
> +	int quota = 0;
> +
>   	if (qdisc_run_begin(q)) {
>   		/* NOLOCK qdisc must check 'state' under the qdisc seqlock
>   		 * to avoid racing with dev_qdisc_reset()
>   		 */
>   		if (!(q->flags & TCQ_F_NOLOCK) ||
>   		    likely(!test_bit(__QDISC_STATE_DEACTIVATED, &q->state)))
> -			__qdisc_run(q);
> +			quota = __qdisc_run(q);
>   		qdisc_run_end(q);
> +
> +		if (quota > 0 && q->flags & TCQ_F_NOLOCK && q->ops->peek(q))
> +			__netif_schedule(q);

Not sure this is relevant, but there's a subtle difference in the way 
that the underlying ptr_ring peeks at the queue head and checks whether 
the queue is empty.

For peek it's:

READ_ONCE(r->queue[r->consumer_head]);

For is_empty it's:

!r->queue[READ_ONCE(r->consumer_head)];

The placement of the READ_ONCE changes here.  I can't get my head around 
whether this difference is significant or not.  If it is, then perhaps 
an is_empty() method is needed on the qdisc_ops...???

/Jonas