Re: [net PATCH v2] net: sched, fix OOO packets with pfifo_fast

[John Fastabend <john.fastabend@gmail.com>]

On 03/26/2018 09:36 AM, David Miller wrote:
> From: John Fastabend <john.fastabend@gmail.com>
> Date: Sat, 24 Mar 2018 22:25:06 -0700
> 
>> After the qdisc lock was dropped in pfifo_fast we allow multiple
>> enqueue threads and dequeue threads to run in parallel. On the
>> enqueue side the skb bit ooo_okay is used to ensure all related
>> skbs are enqueued in-order. On the dequeue side though there is
>> no similar logic. What we observe is with fewer queues than CPUs
>> it is possible to re-order packets when two instances of
>> __qdisc_run() are running in parallel. Each thread will dequeue
>> a skb and then whichever thread calls the ndo op first will
>> be sent on the wire. This doesn't typically happen because
>> qdisc_run() is usually triggered by the same core that did the
>> enqueue. However, drivers will trigger __netif_schedule()
>> when queues are transitioning from stopped to awake using the
>> netif_tx_wake_* APIs. When this happens netif_schedule() calls
>> qdisc_run() on the same CPU that did the netif_tx_wake_* which
>> is usually done in the interrupt completion context. This CPU
>> is selected with the irq affinity which is unrelated to the
>> enqueue operations.
>>
>> To resolve this we add a RUNNING bit to the qdisc to ensure
>> only a single dequeue per qdisc is running. Enqueue and dequeue
>> operations can still run in parallel and also on multi queue
>> NICs we can still have a dequeue in-flight per qdisc, which
>> is typically per CPU.
>>
>> Fixes: c5ad119fb6c0 ("net: sched: pfifo_fast use skb_array")
>> Reported-by: Jakob Unterwurzacher <jakob.unterwurzacher@theobroma-systems.com>
>> Signed-off-by: John Fastabend <john.fastabend@gmail.com>
> 
> Applied, thanks John.
> 

Great, also off-list email from Jakob (I forgot to add him to the
CC list here, oops) told me to add, 

Tested-by: Jakob Unterwurzacher <jakob.unterwurzacher@theobroma-systems.com>

Also in net-next I'll look to see if we can avoid doing the extra
atomics especially in cases where they are not actually needed. For
example the 1:1 qdisc to txq mappings. It seems a bit evasive
though for net.

Finally just an FYI but I think I'll look at a distributed counter
soon so we can get a lockless token bucket. I need the counter for
BPF as well so coming soon.

Thanks,
John