@@ -1498,18 +1498,30 @@ get_n_pmd_threads_on_numa(struct dp_netdev *dp, int numa_id)
return n_pmds;
}
-/* Returns 'true' if there is a port with pmd netdev and the netdev
- * is on numa node 'numa_id'. */
+/* Returns 'true' if there is a port with pmd netdev and the netdev is on
+ * numa node 'numa_id' or its rx queue assigned to core on that numa node . */
static bool
-has_pmd_port_for_numa(struct dp_netdev *dp, int numa_id)
+has_pmd_rxq_for_numa(struct dp_netdev *dp, int numa_id)
OVS_REQUIRES(dp->port_mutex)
{
struct dp_netdev_port *port;
HMAP_FOR_EACH (port, node, &dp->ports) {
- if (netdev_is_pmd(port->netdev)
- && netdev_get_numa_id(port->netdev) == numa_id) {
- return true;
+ if (netdev_is_pmd(port->netdev)) {
+ int i;
+
+ if (netdev_get_numa_id(port->netdev) == numa_id) {
+ return true;
+ }
+
+ for (i = 0; i < port->n_rxq; i++) {
+ unsigned core_id = port->rxqs[i].core_id;
+
+ if (core_id != -1
+ && ovs_numa_get_numa_id(core_id) == numa_id) {
+ return true;
+ }
+ }
}
}
@@ -1533,7 +1545,7 @@ do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
ovs_assert(ovs_numa_numa_id_is_valid(numa_id));
/* If there is no netdev on the numa node, deletes the pmd threads
* for that numa. */
- if (!has_pmd_port_for_numa(dp, numa_id)) {
+ if (!has_pmd_rxq_for_numa(dp, numa_id)) {
dp_netdev_del_pmds_on_numa(dp, numa_id);
}
}
@@ -3738,9 +3750,33 @@ dp_netdev_reset_pmd_threads(struct dp_netdev *dp)
HMAP_FOR_EACH (port, node, &dp->ports) {
if (netdev_is_pmd(port->netdev)) {
- int numa_id = netdev_get_numa_id(port->netdev);
+ struct hmapx numas = HMAPX_INITIALIZER(&numas);
+ struct hmapx_node *numa_node;
+ uintptr_t numa_id;
+ int i;
+
+ numa_id = netdev_get_numa_id(port->netdev);
+ hmapx_add(&numas, (void *) numa_id);
+ for (i = 0; i < port->n_rxq; i++) {
+ unsigned core_id = port->rxqs[i].core_id;
- dp_netdev_set_pmds_on_numa(dp, numa_id);
+ if (core_id != -1) {
+ numa_id = ovs_numa_get_numa_id(core_id);
+ hmapx_add(&numas, (void *) numa_id);
+ }
+ }
+
+ HMAPX_FOR_EACH (numa_node, &numas) {
+ dp_netdev_set_pmds_on_numa(dp, (uintptr_t) numa_node->data);
+ }
+
+ if (hmapx_count(&numas) > 1) {
+ VLOG_WARN("Some Rx queues of port '%s"
+ "' assigned to cores on different numa node. "
+ "This may cause performance issues.",
+ netdev_get_name(port->netdev));
+ }
+ hmapx_destroy(&numas);
}
/* Distribute only pinned rx queues first to mark threads as isolated */
dp_netdev_add_port_rx_to_pmds(dp, port, &to_reload, true);
@@ -517,6 +517,62 @@ p1 3 0 2
OVS_VSWITCHD_STOP(["/dpif_netdev|WARN|There is no PMD thread on core/d"])
AT_CLEANUP
+AT_SETUP([PMD - rxq affinity - NUMA])
+OVS_VSWITCHD_START(
+ [], [], [], [--dummy-numa 0,0,0,1,1])
+AT_CHECK([ovs-appctl vlog/set dpif:dbg dpif_netdev:dbg])
+
+AT_CHECK([ovs-ofctl add-flow br0 actions=controller])
+
+AT_CHECK([ovs-vsctl set Open_vSwitch . other_config:pmd-cpu-mask=7e])
+
+AT_CHECK([ovs-vsctl add-port br0 p1 -- set Interface p1 type=dummy-pmd ofport_request=1 options:n_rxq=2 options:numa_id=0 other_config:pmd-rxq-affinity="0:1,1:2"])
+
+dnl The rxqs should be on the requested cores.
+AT_CHECK([ovs-appctl dpif-netdev/pmd-rxq-show | parse_pmd_rxq_show], [0], [dnl
+p1 0 0 1
+p1 1 0 2
+])
+
+AT_CHECK([ovs-vsctl set Interface p1 other_config:pmd-rxq-affinity="0:3,1:4"])
+
+dnl We moved the queues to different numa node. Expecting threads on
+dnl NUMA node 1 to be created.
+AT_CHECK([ovs-appctl dpif-netdev/pmd-rxq-show | parse_pmd_rxq_show], [0], [dnl
+p1 0 1 3
+p1 1 1 4
+])
+
+AT_CHECK([ovs-vsctl set Interface p1 other_config:pmd-rxq-affinity="0:3,1:1"])
+
+dnl Queues splitted between NUMA nodes.
+AT_CHECK([ovs-appctl dpif-netdev/pmd-rxq-show | parse_pmd_rxq_show], [0], [dnl
+p1 0 1 3
+p1 1 0 1
+])
+
+AT_CHECK([ovs-vsctl remove Interface p1 other_config pmd-rxq-affinity])
+
+dnl We removed the rxq-affinity request. dpif-netdev should assign queues
+dnl in a round robin fashion. We just make sure that every rxq is being
+dnl polled again.
+AT_CHECK([ovs-appctl dpif-netdev/pmd-rxq-show | parse_pmd_rxq_show | cut -f 1,2 -d ' ' | sort], [0], [dnl
+p1 0
+p1 1
+])
+
+AT_CHECK([ovs-vsctl set Interface p1 other_config:pmd-rxq-affinity='0:3'])
+
+dnl We explicitly requesting NUMA node 1 for queue 0.
+dnl Queue 1 should be polled by thread from NUMA node 0.
+AT_CHECK([ovs-appctl dpif-netdev/pmd-rxq-show | parse_pmd_rxq_show | cut -f 1,2,3 -d ' '], [0], [dnl
+p1 0 1
+p1 1 0
+])
+
+OVS_VSWITCHD_STOP(["/dpif_netdev|WARN|Some Rx queues of port 'p1' assigned to cores on different numa node. This may cause performance issues./d"])
+AT_CLEANUP
+
AT_SETUP([PMD - monitor threads])
OVS_VSWITCHD_START(
[], [], [], [--dummy-numa 0,0])
Currently, If user will set up 'pmd-rxq-affinity' to cores on different numa node, they may not be polled, because pmd threads will not be created there even if this cores are in 'pmd-cpu-mask'. Fix that by creating threads on all numa nodes rxqs assigned to. Anyway, OVS will warn about so inefficient mapping. Fixes: 3eb67853c481 ("dpif-netdev: Introduce pmd-rxq-affinity.") Signed-off-by: Ilya Maximets <i.maximets@samsung.com> --- lib/dpif-netdev.c | 54 ++++++++++++++++++++++++++++++++++++++++++++--------- tests/pmd.at | 56 +++++++++++++++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 101 insertions(+), 9 deletions(-)