[ovs-dev] [netdev-dpdk 4/5] netdev-dpdk: Create 'number of dpdk ifaces on same cpu socket' rx queues and 'number of cpu cores' tx queues for each dpdk interface.

[Alex Wang]

>
> On 06/23/14 at 12:37pm, Alex Wang wrote:
> > This will be the default setting for dpdk iface.
> >
> > Signed-off-by: Alex Wang <alexw at nicira.com>
>
> I think you can remove the NR_QUEUE #define in lib/dpif-netdev.h
> with this patch.
>



I'll remove it.




> > ---
> >  lib/netdev-dpdk.c |   47 ++++++++++++++++++++++++++++++++++++++++-------
> >  1 file changed, 40 insertions(+), 7 deletions(-)
> >
> > @@ -725,7 +758,7 @@ netdev_dpdk_send(struct netdev *netdev, struct
> dpif_packet **pkts, int cnt,
> >          int next_tx_idx = 0;
> >          int dropped = 0;
> >
> > -        qid = rte_lcore_id() % NR_QUEUE;
> > +        qid = rte_lcore_id();
>
> This worries me a bit. In a CPU hotplug scenario and if a handler
> is manually re-pinned you might access out of bands here, right?
>


I'll add options for user to configure the cpu mask in my other series of
patches.  So, when the conflict happens, user can reset the cpu mask
and it will recreate/repin all pmd threads.

Pravin also suggests add command to repin certain pmd thread.  I'll explore
this option as well.

Thanks,
Alex Wang,