David,
I believe you are right, it is possible to get into a dead lock between worker
A and B in the case where A is waiting for B and B is waiting for A.
We solved that by tail drop. Another way of doing that is by using a separate
handover worker from NAT worker.
We are exploring some new ideas for NAT session handling now, can’t say if that
will pan out or not yet. Stay tuned. ;-)
Cheers,
Ole
> On 30 Dec 2018, at 15:51, david.leitch....@gmail.com wrote:
>
> Hi Damjan
> Thnaks for your answer,
>
> I think it is possible (I encountered this problem) to have deadlock even
> when I use separate CPU for handoff and NAT processing, Or same CPU with
> congestion drop mechanism.
>
> I studied the code and separate the CPUs of handoff and NAT processing, in
> this situation I think just in one case deadlock for handoff may be happen
> suppose if worker A wants to dequeue but elt->valid value is zero so
> vlib_frame_queue_dequeue function return and do not dequeue, as result
> fq->head will never increase:
>
> int vlib_frame_queue_dequeue (vlib_main_t *vm, vlib_frame_queue_main_t *fqm)
> {
> ....
> while (1)
> {
> if (fq->head == fq->tail)
> return processed;
>
> elt = fq->elts + ((fq->head + 1) & (fq->nelts - 1));
>
> if (!elt->valid)
> {
> fq->head_hint = fq->head;
> return processed;
> }
> ....
> }
>
> and on the other side , worker B ( on handoff node) wants to enqueue but ring
> is full so wait :
>
> static inline vlib_frame_queue_elt_t *
> vlib_get_frame_queue_elt (u32 frame_queue_index, u32 index)
> {
> ...
> new_tail = __sync_add_and_fetch (&fq->tail, 1);
>
> /* Wait until a ring slot is available */
> while (new_tail >= fq->head_hint + fq->nelts)
> vlib_worker_thread_barrier_check ();
> ....
> }
>
> therefore worker B never exit from while loop , do you think it is possible ?
> -=-=-=-=-=-=-=-=-=-=-=-
> Links: You receive all messages sent to this group.
>
> View/Reply Online (#11803): https://lists.fd.io/g/vpp-dev/message/11803
> Mute This Topic: https://lists.fd.io/mt/28878687/675193
> Group Owner: vpp-dev+ow...@lists.fd.io
> Unsubscribe: https://lists.fd.io/g/vpp-dev/unsub [otr...@employees.org]
> -=-=-=-=-=-=-=-=-=-=-=-
-=-=-=-=-=-=-=-=-=-=-=-
Links: You receive all messages sent to this group.
View/Reply Online (#11804): https://lists.fd.io/g/vpp-dev/message/11804
Mute This Topic: https://lists.fd.io/mt/28878687/21656
Group Owner: vpp-dev+ow...@lists.fd.io
Unsubscribe: https://lists.fd.io/g/vpp-dev/unsub [arch...@mail-archive.com]
-=-=-=-=-=-=-=-=-=-=-=-
