On 04.06.2018 11:07, Eelco Chaudron wrote:
> When ping-pong'in a live VM migration between two machines running
> OVS-DPDK every now and then the ping misses would increase
> dramatically. For example:
>
> ===========Stream Rate: 3Mpps===========
> No Stream_Rate Downtime Totaltime Ping_Loss Moongen_Loss
> 0 3Mpps 128 13974 115 7168374
> 1 3Mpps 145 13620 17 1169770
> 2 3Mpps 140 14499 116 7141175
> 3 3Mpps 142 13358 16 1150606
> 4 3Mpps 136 14004 16 1124020
> 5 3Mpps 139 15494 214 13170452
> 6 3Mpps 136 15610 217 13282413
> 7 3Mpps 146 13194 17 1167512
> 8 3Mpps 148 12871 16 1162655
> 9 3Mpps 137 15615 214 13170656
>
> I identified this issue being introduced in OVS commit,
> f3e7ec254738 ("Update relevant artifacts to add support for DPDK 17.05.1.")
> and more specific due to DPDK commit,
> af1475918124 ("vhost: introduce API to start a specific driver").
>
> The combined changes no longer have OVS start the vhost socket polling
> thread at startup, but DPDK will do it on its own when the first vhost
> client is started.
>
> Figuring out the reason why this happens kept me puzzled for quite some
> time...
> What happens is that the callbacks called from the vhost thread are
> calling ovsrcu_synchronize() as part of destroy_device(). This will
> end-up calling seq_wait__().
>
> By default, all created threads outside of OVS will get thread id 0,
> which is equal to the main ovs thread. So for example in the
> seq_wait__() function above if the main thread is waiting already we
> won't add ourselves as a waiter.
>
> The fix below assigns OVSTHREAD_ID_UNSET to none OVS created threads,
> which will get updated to a valid ID on the first call to
> ovsthread_id_self().
>
> Signed-off-by: Eelco Chaudron <[email protected]>
Fixes: f3e7ec254738 ("Update relevant artifacts to add support for DPDK
17.05.1.")
Acked-by: Ilya Maximets <[email protected]>
_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
