On 12/8/21 21:25, David Marchand wrote:
On Wed, Nov 24, 2021 at 10:24 PM Maxime Coquelin <maxime.coque...@redhat.com> wrote:@@ -2385,6 +2408,7 @@ netdev_dpdk_vhost_update_rx_counters(struct netdev_dpdk *dev, } stats->rx_bytes += packet_size; + q_stats->bytes += packet_size; } if (OVS_UNLIKELY(qos_drops)) { @@ -2437,7 +2461,7 @@ netdev_dpdk_vhost_rxq_recv(struct netdev_rxq *rxq, } rte_spinlock_lock(&dev->stats_lock); - netdev_dpdk_vhost_update_rx_counters(dev, batch->packets, + netdev_dpdk_vhost_update_rx_counters(dev, qid, batch->packets,After some testing, instead of qid here, we should use rxq->queue_id :-) The tx part is ok.
Oops. Yes, it needs to be rxq->queue_id.
nb_rx, qos_drops); rte_spinlock_unlock(&dev->stats_lock);[snip] With this fix, I start a vm with a 4 rxq vhost port. This vhost port has a virtio counterpart bound to the vm virtio-net kmod. After a while, I can see some stats for rx and tx packets: # ovs-vsctl get interface vhost4 statistics | sed -e 's#[{}]##g' -e 's#, #\n#g' | grep -v '=0$' rx_1_to_64_packets=15 rx_256_to_511_packets=25 rx_65_to_127_packets=35 rx_bytes=12185 rx_packets=75 rx_q1_bytes=242 rx_q1_packets=3 rx_q2_bytes=242 rx_q2_packets=3 rx_q3_bytes=11701 rx_q3_packets=69 tx_bytes=2414 tx_packets=20 tx_q0_bytes=2414 tx_q0_packets=20 All good, and per q stats nicely sums to the same value than the "global" stats. I then stop the vm and look at stats: # ovs-vsctl get interface vhost4 statistics | sed -e 's#[{}]##g' -e 's#, #\n#g' | grep -v '=0$' rx_1_to_64_packets=15 rx_256_to_511_packets=26 rx_65_to_127_packets=38 rx_bytes=12776 rx_packets=79 tx_bytes=2414 tx_dropped=1 tx_packets=20 tx_q0_bytes=2414 tx_q0_packets=20 So we end up with misalignment of "global" stats and per q stats in this situation. Can we do something about it? I am unsure about resetting all stats to 0, but it could be the simpler solution...
I would vote for resetting all counters on number of queues changes, both per-queue and global, as discussed off-list. That's the only way to keep consistency between global and per-queue counters. Maxime
_______________________________________________ dev mailing list d...@openvswitch.org https://mail.openvswitch.org/mailman/listinfo/ovs-dev
