[ovs-dev] [PATCH] dpdk: Refuse running on cpu >= RTE_MAX_LCORE.

David Marchand david.marchand at redhat.com
Tue Oct 20 12:25:35 UTC 2020


On Thu, Jul 2, 2020 at 6:46 PM David Marchand <david.marchand at redhat.com> wrote:
>
> On Thu, Jul 2, 2020 at 5:53 PM Ilya Maximets <i.maximets at ovn.org> wrote:
> > > pmd threads will never run on any core >= RTE_MAX_LCORE ok, it works.
> > >
> > > Though, it binds OVS core id and DPDK lcore id.
> > > Should OVS really care about this?
> >
> > I think this is better than random abort() in production environment.
> > And, in current code, OVS cores bound to lcores as 1 to 1.
>
> Yes, I did not like either the abort(), I added it wrt to a comment from Kevin.
> But rereading it, I might have misunderstood.
>
>
> >
> > I understand that there will be no direct binding with the new
> > rte_thread_register API, and we will need to figure out how to work
> > with it correctly.  With new API above changes will need to be reverted
> > and some new handling introduced to limit total number of cores, not
> > their IDs.  Maybe something like this (not tested):
>
> I just wanted to make sure you had the other patch in mind.
> I'll have a try with your proposal.

Tried this approach and found it quite ugly.
We would have to revert it almost entirely for the use of the new DPDK
api and it won't solve the problem of failing to register a thread in
DPDK later.

I'll submit a new revision of the lcore dpdk-latest patch.
Once we make progress on the rte_thread_register topic, I might
revisit what to do in older branches.

Thanks.

-- 
David Marchand



More information about the dev mailing list