Re: [ovs-dev] [PATCH v4] netdev-dpdk:add set_queue datapath dpdk support

[Ilya Maximets]

On 7/6/23 10:27, wangze wrote:
> When doing set_queue action, skb_priority is marked in
> userspace, but there is no further processing in datapath
> dpdk.

Hi.  There actually is further processing for dpdk ports.

The naming is confusing, I understand that, but set_queue OpenFlow
action actually has nothing in common with hardware Tx queues of
the device.  These are QoS queues configured in the database in
QoS table.  And netdev-dpdk makes use of them via Two Rate Three
Color Marker policing (trtcm-policer).  You can find documentation
for it here:
  https://docs.openvswitch.org/en/latest/topics/dpdk/qos/#multi-queue-policer
Or in the ovs-vswitchd.conf.db (5) man page.

The general QoS with set_queue examples can be found here:
  https://docs.openvswitch.org/en/latest/faq/qos/#quality-of-service-qos

In general, we shouldn't overload the meaning of the action.
And there are no high-level OpenFlow fileds or actions that
work with hardware Rx/Tx queues, AFAIK.  And there shouldn't be
as OpenFlow is a much higher abstraction.

> This patch add set_queue support in datapath dpdk,
> sending skbs from specific queues according to the
> skb_priority.
> 
> The enqueue action can be applied to the scenario of
> double-sending arp packets in bond, usage:
> ovs-ofctl add-flow br-ext "priority=1,in_port=LOCAL,arp actions=push_vlan:
> 0x8100,set_field:4397->vlan_vid,set_queue:1,output:bond1,pop_queue,set_queue:2,
> output:bond1,pop_queue" -Oopenflow13

Why do you need to double-send ARP packets?
And why they need to go to different hardware queues?

Best regards, Ilya Maximets.
___
dev mailing list
d...@openvswitch.org
https://mail.openvswitch.org/mailman/listinfo/ovs-dev

[ovs-dev] [PATCH v4] netdev-dpdk:add set_queue datapath dpdk support

[wangze]

When doing set_queue action, skb_priority is marked in
userspace, but there is no further processing in datapath
dpdk. This patch add set_queue support in datapath dpdk,
sending skbs from specific queues according to the
skb_priority.

The enqueue action can be applied to the scenario of
double-sending arp packets in bond, usage:
ovs-ofctl add-flow br-ext "priority=1,in_port=LOCAL,arp actions=push_vlan:
0x8100,set_field:4397->vlan_vid,set_queue:1,output:bond1,pop_queue,set_queue:2,
output:bond1,pop_queue" -Oopenflow13

Signed-off-by: wangze 
Reviewed-by: wenxu 
---
 lib/dpif-netdev.c | 43 +--
 1 file changed, 37 insertions(+), 6 deletions(-)

diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c
index ab493f9..4bf36fd 100644
--- a/lib/dpif-netdev.c
+++ b/lib/dpif-netdev.c
@@ -5259,6 +5259,31 @@ pmd_perf_metrics_enabled(const struct 
dp_netdev_pmd_thread *pmd OVS_UNUSED)
 }
 #endif
 
+
+#ifdef DPDK_NETDEV
+static int
+dp_netdev_pmd_flush_txqs_on_port(struct dp_netdev_pmd_thread *pmd,
+   struct tx_port *p)
+{
+int i;
+int output_cnt = 0;
+
+int n_txq = netdev_n_txq(p->port->netdev);
+
+for (i = 0; i < n_txq; i++) {
+if (dp_packet_batch_is_empty(>txq_pkts[i])) {
+continue;
+}
+output_cnt += dp_packet_batch_size(>txq_pkts[i]);
+netdev_send(p->port->netdev, i, >txq_pkts[i], true);
+dp_packet_batch_init(>txq_pkts[i]);
+}
+
+pmd_perf_update_counter(>perf_stats, PMD_STAT_SENT_PKTS, output_cnt);
+return output_cnt;
+}
+#endif
+
 static int
 dp_netdev_pmd_flush_output_on_port(struct dp_netdev_pmd_thread *pmd,
struct tx_port *p)
@@ -7799,13 +7824,11 @@ dp_netdev_add_port_tx_to_pmd(struct 
dp_netdev_pmd_thread *pmd,
 tx->flush_time = 0LL;
 dp_packet_batch_init(>output_pkts);
 
-if (tx->port->txq_mode == TXQ_MODE_XPS_HASH) {
-int i, n_txq = netdev_n_txq(tx->port->netdev);
+int i, n_txq = netdev_n_txq(tx->port->netdev);
 
-tx->txq_pkts = xzalloc(n_txq * sizeof *tx->txq_pkts);
-for (i = 0; i < n_txq; i++) {
-dp_packet_batch_init(>txq_pkts[i]);
-}
+tx->txq_pkts = xzalloc(n_txq * sizeof *tx->txq_pkts);
+for (i = 0; i < n_txq; i++) {
+dp_packet_batch_init(>txq_pkts[i]);
 }
 
 hmap_insert(>tx_ports, >node, hash_port_no(tx->port->port_no));
@@ -8832,6 +8855,7 @@ dp_execute_output_action(struct dp_netdev_pmd_thread *pmd,
 }
 dp_packet_batch_apply_cutlen(packets_);
 #ifdef DPDK_NETDEV
+dp_netdev_pmd_flush_txqs_on_port(pmd, p);
 if (OVS_UNLIKELY(!dp_packet_batch_is_empty(>output_pkts)
  && packets_->packets[0]->source
 != p->output_pkts.packets[0]->source)) {
@@ -8851,9 +8875,16 @@ dp_execute_output_action(struct dp_netdev_pmd_thread 
*pmd,
 }
 
 struct dp_packet *packet;
+int n_txq = netdev_n_txq(p->port->netdev);
 DP_PACKET_BATCH_FOR_EACH (i, packet, packets_) {
 p->output_pkts_rxqs[dp_packet_batch_size(>output_pkts)] =
 pmd->ctx.last_rxq;
+if (packet->md.skb_priority) {
+dp_packet_batch_add(
+>txq_pkts[packet->md.skb_priority % n_txq - 1],
+packet);
+continue;
+}
 dp_packet_batch_add(>output_pkts, packet);
 }
 return true;
-- 
1.8.3.1

___
dev mailing list
d...@openvswitch.org
https://mail.openvswitch.org/mailman/listinfo/ovs-dev