Hi Mike, thanks for reviewing,
On 02/03/2022 15:41, Mike Pattrick wrote:
On Tue, Mar 1, 2022 at 12:25 PM Kevin Traynor <ktray...@redhat.com> wrote:
There are currently some checks for cross-numa polling cases to
ensure that they won't effect the accuracy of the PMD ALB.
If an rxq is pinned to a pmd core by the user it will not be
reassigned by OVS, so even if it is non-local numa polled it
will not impact PMD ALB accuracy.
To establish this, a check was made on whether the pmd core was
isolated or not. However, since other_config:pmd-rxq-isolate was
introduced, rxqs may be pinned but the pmd core not isolated.
It means that by setting pmd-rxq-isolate=false and doing non-local
numa pinning, PMD ALB may not run where it should.
If the core is isolated we can skip individual rxq checks but if not,
we should check the individual rxqs for pinning before we disallow
PMD ALB.
Fixes: 6193e03267c1 ("dpif-netdev: Allow pin rxq and non-isolate PMD.")
Signed-off-by: Kevin Traynor <ktray...@redhat.com>
---
lib/dpif-netdev.c | 5 ++++-
1 file changed, 4 insertions(+), 1 deletion(-)
diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c
index 9ae488d6e..9df41e0b6 100644
--- a/lib/dpif-netdev.c
+++ b/lib/dpif-netdev.c
@@ -5697,9 +5697,12 @@ sched_numa_list_cross_numa_polling(struct
sched_numa_list *numa_list)
sched_pmd = &numa->pmds[i];
+ if (sched_pmd->isolated) {
+ continue;
+ }
/* For each rxq. */
for (unsigned k = 0; k < sched_pmd->n_rxq; k++) {
struct dp_netdev_rxq *rxq = sched_pmd->rxqs[k];
- if (!sched_pmd->isolated &&
+ if (rxq->core_id == OVS_CORE_UNSPEC &&
Should this be an ||? Is there a case where core_id is set but cross numa?
yes, that case is possible. The options are,
core_id == OVS_CORE_UNSPEC == rxq non-pinned
core_id != OVS_CORE_UNSPEC == rxq pinned
pmd core numa == rxq numa == rxq local numa polled
pmd core numa != rxq numa == rxq non-local numa polled
The function wants to find if there are any rxqs that are 'non-pinned
AND non-local numa polled'.
That is part of checks that indicate the rxq may get polled by a pmd
core from a different numa after reassignment, so (for now anyway) we
don't want proceed with the dry-run as the estimations about cycles
required to process pkts from that rxq may be inaccurate.
I can add some comments in the function to make it clearer, as it's not
the most intuitive. wdyt?
Kevin.
rxq->pmd->numa_id !=
netdev_get_numa_id(rxq->port->netdev)) {
--
2.34.1
_______________________________________________
dev mailing list
d...@openvswitch.org
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
Cheers,
M
_______________________________________________
dev mailing list
d...@openvswitch.org
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
