event/dsw: add load balancing
authorMattias Rönnblom <mattias.ronnblom@ericsson.com>
Tue, 18 Sep 2018 12:45:11 +0000 (14:45 +0200)
committerJerin Jacob <jerin.jacob@caviumnetworks.com>
Mon, 1 Oct 2018 14:45:50 +0000 (16:45 +0200)
The DSW event device will now attempt to migrate (move) flows between
ports in order to balance the load.

Signed-off-by: Mattias Rönnblom <mattias.ronnblom@ericsson.com>
drivers/event/dsw/dsw_evdev.c
drivers/event/dsw/dsw_evdev.h
drivers/event/dsw/dsw_event.c

index bcfa17b..2ecb365 100644 (file)
@@ -20,6 +20,7 @@ dsw_port_setup(struct rte_eventdev *dev, uint8_t port_id,
        struct dsw_evdev *dsw = dsw_pmd_priv(dev);
        struct dsw_port *port;
        struct rte_event_ring *in_ring;
+       struct rte_ring *ctl_in_ring;
        char ring_name[RTE_RING_NAMESIZE];
 
        port = &dsw->ports[port_id];
@@ -42,13 +43,29 @@ dsw_port_setup(struct rte_eventdev *dev, uint8_t port_id,
        if (in_ring == NULL)
                return -ENOMEM;
 
+       snprintf(ring_name, sizeof(ring_name), "dswctl%d_p%u",
+                dev->data->dev_id, port_id);
+
+       ctl_in_ring = rte_ring_create(ring_name, DSW_CTL_IN_RING_SIZE,
+                                     dev->data->socket_id,
+                                     RING_F_SC_DEQ|RING_F_EXACT_SZ);
+
+       if (ctl_in_ring == NULL) {
+               rte_event_ring_free(in_ring);
+               return -ENOMEM;
+       }
+
        port->in_ring = in_ring;
+       port->ctl_in_ring = ctl_in_ring;
 
        rte_atomic16_init(&port->load);
 
        port->load_update_interval =
                (DSW_LOAD_UPDATE_INTERVAL * rte_get_timer_hz()) / US_PER_S;
 
+       port->migration_interval =
+               (DSW_MIGRATION_INTERVAL * rte_get_timer_hz()) / US_PER_S;
+
        dev->data->ports[port_id] = port;
 
        return 0;
@@ -72,6 +89,7 @@ dsw_port_release(void *p)
        struct dsw_port *port = p;
 
        rte_event_ring_free(port->in_ring);
+       rte_ring_free(port->ctl_in_ring);
 }
 
 static int
@@ -272,6 +290,14 @@ dsw_port_drain_buf(uint8_t dev_id, struct rte_event *buf, uint16_t buf_len,
                flush(dev_id, buf[i], flush_arg);
 }
 
+static void
+dsw_port_drain_paused(uint8_t dev_id, struct dsw_port *port,
+                     eventdev_stop_flush_t flush, void *flush_arg)
+{
+       dsw_port_drain_buf(dev_id, port->paused_events, port->paused_events_len,
+                          flush, flush_arg);
+}
+
 static void
 dsw_port_drain_out(uint8_t dev_id, struct dsw_evdev *dsw, struct dsw_port *port,
                   eventdev_stop_flush_t flush, void *flush_arg)
@@ -308,6 +334,7 @@ dsw_drain(uint8_t dev_id, struct dsw_evdev *dsw,
                struct dsw_port *port = &dsw->ports[port_id];
 
                dsw_port_drain_out(dev_id, dsw, port, flush, flush_arg);
+               dsw_port_drain_paused(dev_id, port, flush, flush_arg);
                dsw_port_drain_in_ring(dev_id, port, flush, flush_arg);
        }
 }
index a5399dd..783c418 100644 (file)
 #define DSW_LOAD_UPDATE_INTERVAL (DSW_MIGRATION_INTERVAL/4)
 #define DSW_OLD_LOAD_WEIGHT (1)
 
+/* The minimum time (in us) between two flow migrations. What puts an
+ * upper limit on the actual migration rate is primarily the pace in
+ * which the ports send and receive control messages, which in turn is
+ * largely a function of how much cycles are spent the processing of
+ * an event burst.
+ */
 #define DSW_MIGRATION_INTERVAL (1000)
+#define DSW_MIN_SOURCE_LOAD_FOR_MIGRATION (DSW_LOAD_FROM_PERCENT(70))
+#define DSW_MAX_TARGET_LOAD_FOR_MIGRATION (DSW_LOAD_FROM_PERCENT(95))
+
+#define DSW_MAX_EVENTS_RECORDED (128)
+
+/* Only one outstanding migration per port is allowed */
+#define DSW_MAX_PAUSED_FLOWS (DSW_MAX_PORTS)
+
+/* Enough room for paus request/confirm and unpaus request/confirm for
+ * all possible senders.
+ */
+#define DSW_CTL_IN_RING_SIZE ((DSW_MAX_PORTS-1)*4)
+
+struct dsw_queue_flow {
+       uint8_t queue_id;
+       uint16_t flow_hash;
+};
+
+enum dsw_migration_state {
+       DSW_MIGRATION_STATE_IDLE,
+       DSW_MIGRATION_STATE_PAUSING,
+       DSW_MIGRATION_STATE_FORWARDING,
+       DSW_MIGRATION_STATE_UNPAUSING
+};
 
 struct dsw_port {
        uint16_t id;
@@ -98,6 +128,7 @@ struct dsw_port {
 
        uint16_t ops_since_bg_task;
 
+       /* most recent 'background' processing */
        uint64_t last_bg;
 
        /* For port load measurement. */
@@ -108,11 +139,46 @@ struct dsw_port {
        uint64_t busy_cycles;
        uint64_t total_busy_cycles;
 
+       /* For the ctl interface and flow migration mechanism. */
+       uint64_t next_migration;
+       uint64_t migration_interval;
+       enum dsw_migration_state migration_state;
+
+       uint64_t migration_start;
+       uint64_t migrations;
+       uint64_t migration_latency;
+
+       uint8_t migration_target_port_id;
+       struct dsw_queue_flow migration_target_qf;
+       uint8_t cfm_cnt;
+
+       uint16_t paused_flows_len;
+       struct dsw_queue_flow paused_flows[DSW_MAX_PAUSED_FLOWS];
+
+       /* In a very contrived worst case all inflight events can be
+        * laying around paused here.
+        */
+       uint16_t paused_events_len;
+       struct rte_event paused_events[DSW_MAX_EVENTS];
+
+       uint16_t seen_events_len;
+       uint16_t seen_events_idx;
+       struct dsw_queue_flow seen_events[DSW_MAX_EVENTS_RECORDED];
+
        uint16_t out_buffer_len[DSW_MAX_PORTS];
        struct rte_event out_buffer[DSW_MAX_PORTS][DSW_MAX_PORT_OUT_BUFFER];
 
+       uint16_t in_buffer_len;
+       uint16_t in_buffer_start;
+       /* This buffer may contain events that were read up from the
+        * in_ring during the flow migration process.
+        */
+       struct rte_event in_buffer[DSW_MAX_EVENTS];
+
        struct rte_event_ring *in_ring __rte_cache_aligned;
 
+       struct rte_ring *ctl_in_ring __rte_cache_aligned;
+
        /* Estimate of current port load. */
        rte_atomic16_t load __rte_cache_aligned;
 } __rte_cache_aligned;
@@ -137,6 +203,20 @@ struct dsw_evdev {
        rte_atomic32_t credits_on_loan __rte_cache_aligned;
 };
 
+#define DSW_CTL_PAUS_REQ (0)
+#define DSW_CTL_UNPAUS_REQ (1)
+#define DSW_CTL_CFM (2)
+
+/* sizeof(struct dsw_ctl_msg) must be equal or less than
+ * sizeof(void *), to fit on the control ring.
+ */
+struct dsw_ctl_msg {
+       uint8_t type:2;
+       uint8_t originating_port_id:6;
+       uint8_t queue_id;
+       uint16_t flow_hash;
+} __rte_packed;
+
 uint16_t dsw_event_enqueue(void *port, const struct rte_event *event);
 uint16_t dsw_event_enqueue_burst(void *port,
                                 const struct rte_event events[],
index f326147..f034759 100644 (file)
@@ -5,9 +5,11 @@
 #include "dsw_evdev.h"
 
 #include <stdbool.h>
+#include <string.h>
 
 #include <rte_atomic.h>
 #include <rte_cycles.h>
+#include <rte_memcpy.h>
 #include <rte_random.h>
 
 static bool
@@ -140,6 +142,269 @@ dsw_port_consider_load_update(struct dsw_port *port, uint64_t now)
        dsw_port_load_update(port, now);
 }
 
+static void
+dsw_port_ctl_enqueue(struct dsw_port *port, struct dsw_ctl_msg *msg)
+{
+       void *raw_msg;
+
+       memcpy(&raw_msg, msg, sizeof(*msg));
+
+       /* there's always room on the ring */
+       while (rte_ring_enqueue(port->ctl_in_ring, raw_msg) != 0)
+               rte_pause();
+}
+
+static int
+dsw_port_ctl_dequeue(struct dsw_port *port, struct dsw_ctl_msg *msg)
+{
+       void *raw_msg;
+       int rc;
+
+       rc = rte_ring_dequeue(port->ctl_in_ring, &raw_msg);
+
+       if (rc == 0)
+               memcpy(msg, &raw_msg, sizeof(*msg));
+
+       return rc;
+}
+
+static void
+dsw_port_ctl_broadcast(struct dsw_evdev *dsw, struct dsw_port *source_port,
+                      uint8_t type, uint8_t queue_id, uint16_t flow_hash)
+{
+       uint16_t port_id;
+       struct dsw_ctl_msg msg = {
+               .type = type,
+               .originating_port_id = source_port->id,
+               .queue_id = queue_id,
+               .flow_hash = flow_hash
+       };
+
+       for (port_id = 0; port_id < dsw->num_ports; port_id++)
+               if (port_id != source_port->id)
+                       dsw_port_ctl_enqueue(&dsw->ports[port_id], &msg);
+}
+
+static bool
+dsw_port_is_flow_paused(struct dsw_port *port, uint8_t queue_id,
+                       uint16_t flow_hash)
+{
+       uint16_t i;
+
+       for (i = 0; i < port->paused_flows_len; i++) {
+               struct dsw_queue_flow *qf = &port->paused_flows[i];
+               if (qf->queue_id == queue_id &&
+                   qf->flow_hash == flow_hash)
+                       return true;
+       }
+       return false;
+}
+
+static void
+dsw_port_add_paused_flow(struct dsw_port *port, uint8_t queue_id,
+                        uint16_t paused_flow_hash)
+{
+       port->paused_flows[port->paused_flows_len] = (struct dsw_queue_flow) {
+               .queue_id = queue_id,
+               .flow_hash = paused_flow_hash
+       };
+       port->paused_flows_len++;
+}
+
+static void
+dsw_port_remove_paused_flow(struct dsw_port *port, uint8_t queue_id,
+                           uint16_t paused_flow_hash)
+{
+       uint16_t i;
+
+       for (i = 0; i < port->paused_flows_len; i++) {
+               struct dsw_queue_flow *qf = &port->paused_flows[i];
+
+               if (qf->queue_id == queue_id &&
+                   qf->flow_hash == paused_flow_hash) {
+                       uint16_t last_idx = port->paused_flows_len-1;
+                       if (i != last_idx)
+                               port->paused_flows[i] =
+                                       port->paused_flows[last_idx];
+                       port->paused_flows_len--;
+                       break;
+               }
+       }
+}
+
+static void
+dsw_port_flush_out_buffers(struct dsw_evdev *dsw, struct dsw_port *source_port);
+
+static void
+dsw_port_handle_pause_flow(struct dsw_evdev *dsw, struct dsw_port *port,
+                          uint8_t originating_port_id, uint8_t queue_id,
+                          uint16_t paused_flow_hash)
+{
+       struct dsw_ctl_msg cfm = {
+               .type = DSW_CTL_CFM,
+               .originating_port_id = port->id,
+               .queue_id = queue_id,
+               .flow_hash = paused_flow_hash
+       };
+
+       DSW_LOG_DP_PORT(DEBUG, port->id, "Pausing queue_id %d flow_hash %d.\n",
+                       queue_id, paused_flow_hash);
+
+       /* There might be already-scheduled events belonging to the
+        * paused flow in the output buffers.
+        */
+       dsw_port_flush_out_buffers(dsw, port);
+
+       dsw_port_add_paused_flow(port, queue_id, paused_flow_hash);
+
+       /* Make sure any stores to the original port's in_ring is seen
+        * before the ctl message.
+        */
+       rte_smp_wmb();
+
+       dsw_port_ctl_enqueue(&dsw->ports[originating_port_id], &cfm);
+}
+
+static void
+dsw_find_lowest_load_port(uint8_t *port_ids, uint16_t num_port_ids,
+                         uint8_t exclude_port_id, int16_t *port_loads,
+                         uint8_t *target_port_id, int16_t *target_load)
+{
+       int16_t candidate_port_id = -1;
+       int16_t candidate_load = DSW_MAX_LOAD;
+       uint16_t i;
+
+       for (i = 0; i < num_port_ids; i++) {
+               uint8_t port_id = port_ids[i];
+               if (port_id != exclude_port_id) {
+                       int16_t load = port_loads[port_id];
+                       if (candidate_port_id == -1 ||
+                           load < candidate_load) {
+                               candidate_port_id = port_id;
+                               candidate_load = load;
+                       }
+               }
+       }
+       *target_port_id = candidate_port_id;
+       *target_load = candidate_load;
+}
+
+struct dsw_queue_flow_burst {
+       struct dsw_queue_flow queue_flow;
+       uint16_t count;
+};
+
+static inline int
+dsw_cmp_burst(const void *v_burst_a, const void *v_burst_b)
+{
+       const struct dsw_queue_flow_burst *burst_a = v_burst_a;
+       const struct dsw_queue_flow_burst *burst_b = v_burst_b;
+
+       int a_count = burst_a->count;
+       int b_count = burst_b->count;
+
+       return a_count - b_count;
+}
+
+#define DSW_QF_TO_INT(_qf)                                     \
+       ((int)((((_qf)->queue_id)<<16)|((_qf)->flow_hash)))
+
+static inline int
+dsw_cmp_qf(const void *v_qf_a, const void *v_qf_b)
+{
+       const struct dsw_queue_flow *qf_a = v_qf_a;
+       const struct dsw_queue_flow *qf_b = v_qf_b;
+
+       return DSW_QF_TO_INT(qf_a) - DSW_QF_TO_INT(qf_b);
+}
+
+static uint16_t
+dsw_sort_qfs_to_bursts(struct dsw_queue_flow *qfs, uint16_t qfs_len,
+                      struct dsw_queue_flow_burst *bursts)
+{
+       uint16_t i;
+       struct dsw_queue_flow_burst *current_burst = NULL;
+       uint16_t num_bursts = 0;
+
+       /* We don't need the stable property, and the list is likely
+        * large enough for qsort() to outperform dsw_stable_sort(),
+        * so we use qsort() here.
+        */
+       qsort(qfs, qfs_len, sizeof(qfs[0]), dsw_cmp_qf);
+
+       /* arrange the (now-consecutive) events into bursts */
+       for (i = 0; i < qfs_len; i++) {
+               if (i == 0 ||
+                   dsw_cmp_qf(&qfs[i], &current_burst->queue_flow) != 0) {
+                       current_burst = &bursts[num_bursts];
+                       current_burst->queue_flow = qfs[i];
+                       current_burst->count = 0;
+                       num_bursts++;
+               }
+               current_burst->count++;
+       }
+
+       qsort(bursts, num_bursts, sizeof(bursts[0]), dsw_cmp_burst);
+
+       return num_bursts;
+}
+
+static bool
+dsw_retrieve_port_loads(struct dsw_evdev *dsw, int16_t *port_loads,
+                       int16_t load_limit)
+{
+       bool below_limit = false;
+       uint16_t i;
+
+       for (i = 0; i < dsw->num_ports; i++) {
+               int16_t load = rte_atomic16_read(&dsw->ports[i].load);
+               if (load < load_limit)
+                       below_limit = true;
+               port_loads[i] = load;
+       }
+       return below_limit;
+}
+
+static bool
+dsw_select_migration_target(struct dsw_evdev *dsw,
+                           struct dsw_port *source_port,
+                           struct dsw_queue_flow_burst *bursts,
+                           uint16_t num_bursts, int16_t *port_loads,
+                           int16_t max_load, struct dsw_queue_flow *target_qf,
+                           uint8_t *target_port_id)
+{
+       uint16_t source_load = port_loads[source_port->id];
+       uint16_t i;
+
+       for (i = 0; i < num_bursts; i++) {
+               struct dsw_queue_flow *qf = &bursts[i].queue_flow;
+
+               if (dsw_port_is_flow_paused(source_port, qf->queue_id,
+                                           qf->flow_hash))
+                       continue;
+
+               struct dsw_queue *queue = &dsw->queues[qf->queue_id];
+               int16_t target_load;
+
+               dsw_find_lowest_load_port(queue->serving_ports,
+                                         queue->num_serving_ports,
+                                         source_port->id, port_loads,
+                                         target_port_id, &target_load);
+
+               if (target_load < source_load &&
+                   target_load < max_load) {
+                       *target_qf = *qf;
+                       return true;
+               }
+       }
+
+       DSW_LOG_DP_PORT(DEBUG, source_port->id, "For the %d flows considered, "
+                       "no target port found with load less than %d.\n",
+                       num_bursts, DSW_LOAD_TO_PERCENT(max_load));
+
+       return false;
+}
+
 static uint8_t
 dsw_schedule(struct dsw_evdev *dsw, uint8_t queue_id, uint16_t flow_hash)
 {
@@ -197,6 +462,14 @@ dsw_port_get_parallel_flow_id(struct dsw_port *port)
        return flow_id;
 }
 
+static void
+dsw_port_buffer_paused(struct dsw_port *port,
+                      const struct rte_event *paused_event)
+{
+       port->paused_events[port->paused_events_len] = *paused_event;
+       port->paused_events_len++;
+}
+
 static void
 dsw_port_buffer_non_paused(struct dsw_evdev *dsw, struct dsw_port *source_port,
                           uint8_t dest_port_id, const struct rte_event *event)
@@ -256,11 +529,401 @@ dsw_port_buffer_event(struct dsw_evdev *dsw, struct dsw_port *source_port,
 
        flow_hash = dsw_flow_id_hash(event->flow_id);
 
+       if (unlikely(dsw_port_is_flow_paused(source_port, event->queue_id,
+                                            flow_hash))) {
+               dsw_port_buffer_paused(source_port, event);
+               return;
+       }
+
        dest_port_id = dsw_schedule(dsw, event->queue_id, flow_hash);
 
        dsw_port_buffer_non_paused(dsw, source_port, dest_port_id, event);
 }
 
+static void
+dsw_port_flush_paused_events(struct dsw_evdev *dsw,
+                            struct dsw_port *source_port,
+                            uint8_t queue_id, uint16_t paused_flow_hash)
+{
+       uint16_t paused_events_len = source_port->paused_events_len;
+       struct rte_event paused_events[paused_events_len];
+       uint8_t dest_port_id;
+       uint16_t i;
+
+       if (paused_events_len == 0)
+               return;
+
+       if (dsw_port_is_flow_paused(source_port, queue_id, paused_flow_hash))
+               return;
+
+       rte_memcpy(paused_events, source_port->paused_events,
+                  paused_events_len * sizeof(struct rte_event));
+
+       source_port->paused_events_len = 0;
+
+       dest_port_id = dsw_schedule(dsw, queue_id, paused_flow_hash);
+
+       for (i = 0; i < paused_events_len; i++) {
+               struct rte_event *event = &paused_events[i];
+               uint16_t flow_hash;
+
+               flow_hash = dsw_flow_id_hash(event->flow_id);
+
+               if (event->queue_id == queue_id &&
+                   flow_hash == paused_flow_hash)
+                       dsw_port_buffer_non_paused(dsw, source_port,
+                                                  dest_port_id, event);
+               else
+                       dsw_port_buffer_paused(source_port, event);
+       }
+}
+
+static void
+dsw_port_migration_stats(struct dsw_port *port)
+{
+       uint64_t migration_latency;
+
+       migration_latency = (rte_get_timer_cycles() - port->migration_start);
+       port->migration_latency += migration_latency;
+       port->migrations++;
+}
+
+static void
+dsw_port_end_migration(struct dsw_evdev *dsw, struct dsw_port *port)
+{
+       uint8_t queue_id = port->migration_target_qf.queue_id;
+       uint16_t flow_hash = port->migration_target_qf.flow_hash;
+
+       port->migration_state = DSW_MIGRATION_STATE_IDLE;
+       port->seen_events_len = 0;
+
+       dsw_port_migration_stats(port);
+
+       if (dsw->queues[queue_id].schedule_type != RTE_SCHED_TYPE_PARALLEL) {
+               dsw_port_remove_paused_flow(port, queue_id, flow_hash);
+               dsw_port_flush_paused_events(dsw, port, queue_id, flow_hash);
+       }
+
+       DSW_LOG_DP_PORT(DEBUG, port->id, "Migration completed for queue_id "
+                       "%d flow_hash %d.\n", queue_id, flow_hash);
+}
+
+static void
+dsw_port_consider_migration(struct dsw_evdev *dsw,
+                           struct dsw_port *source_port,
+                           uint64_t now)
+{
+       bool any_port_below_limit;
+       struct dsw_queue_flow *seen_events = source_port->seen_events;
+       uint16_t seen_events_len = source_port->seen_events_len;
+       struct dsw_queue_flow_burst bursts[DSW_MAX_EVENTS_RECORDED];
+       uint16_t num_bursts;
+       int16_t source_port_load;
+       int16_t port_loads[dsw->num_ports];
+
+       if (now < source_port->next_migration)
+               return;
+
+       if (dsw->num_ports == 1)
+               return;
+
+       DSW_LOG_DP_PORT(DEBUG, source_port->id, "Considering migration.\n");
+
+       /* Randomize interval to avoid having all threads considering
+        * migration at the same in point in time, which might lead to
+        * all choosing the same target port.
+        */
+       source_port->next_migration = now +
+               source_port->migration_interval / 2 +
+               rte_rand() % source_port->migration_interval;
+
+       if (source_port->migration_state != DSW_MIGRATION_STATE_IDLE) {
+               DSW_LOG_DP_PORT(DEBUG, source_port->id,
+                               "Migration already in progress.\n");
+               return;
+       }
+
+       /* For simplicity, avoid migration in the unlikely case there
+        * is still events to consume in the in_buffer (from the last
+        * migration).
+        */
+       if (source_port->in_buffer_len > 0) {
+               DSW_LOG_DP_PORT(DEBUG, source_port->id, "There are still "
+                               "events in the input buffer.\n");
+               return;
+       }
+
+       source_port_load = rte_atomic16_read(&source_port->load);
+       if (source_port_load < DSW_MIN_SOURCE_LOAD_FOR_MIGRATION) {
+               DSW_LOG_DP_PORT(DEBUG, source_port->id,
+                               "Load %d is below threshold level %d.\n",
+                               DSW_LOAD_TO_PERCENT(source_port_load),
+                      DSW_LOAD_TO_PERCENT(DSW_MIN_SOURCE_LOAD_FOR_MIGRATION));
+               return;
+       }
+
+       /* Avoid starting any expensive operations (sorting etc), in
+        * case of a scenario with all ports above the load limit.
+        */
+       any_port_below_limit =
+               dsw_retrieve_port_loads(dsw, port_loads,
+                                       DSW_MAX_TARGET_LOAD_FOR_MIGRATION);
+       if (!any_port_below_limit) {
+               DSW_LOG_DP_PORT(DEBUG, source_port->id,
+                               "Candidate target ports are all too highly "
+                               "loaded.\n");
+               return;
+       }
+
+       /* Sort flows into 'bursts' to allow attempting to migrating
+        * small (but still active) flows first - this it to avoid
+        * having large flows moving around the worker cores too much
+        * (to avoid cache misses, among other things). Of course, the
+        * number of recorded events (queue+flow ids) are limited, and
+        * provides only a snapshot, so only so many conclusions can
+        * be drawn from this data.
+        */
+       num_bursts = dsw_sort_qfs_to_bursts(seen_events, seen_events_len,
+                                           bursts);
+       /* For non-big-little systems, there's no point in moving the
+        * only (known) flow.
+        */
+       if (num_bursts < 2) {
+               DSW_LOG_DP_PORT(DEBUG, source_port->id, "Only a single flow "
+                               "queue_id %d flow_hash %d has been seen.\n",
+                               bursts[0].queue_flow.queue_id,
+                               bursts[0].queue_flow.flow_hash);
+               return;
+       }
+
+       /* The strategy is to first try to find a flow to move to a
+        * port with low load (below the migration-attempt
+        * threshold). If that fails, we try to find a port which is
+        * below the max threshold, and also less loaded than this
+        * port is.
+        */
+       if (!dsw_select_migration_target(dsw, source_port, bursts, num_bursts,
+                                        port_loads,
+                                        DSW_MIN_SOURCE_LOAD_FOR_MIGRATION,
+                                        &source_port->migration_target_qf,
+                                        &source_port->migration_target_port_id)
+           &&
+           !dsw_select_migration_target(dsw, source_port, bursts, num_bursts,
+                                        port_loads,
+                                        DSW_MAX_TARGET_LOAD_FOR_MIGRATION,
+                                        &source_port->migration_target_qf,
+                                      &source_port->migration_target_port_id))
+               return;
+
+       DSW_LOG_DP_PORT(DEBUG, source_port->id, "Migrating queue_id %d "
+                       "flow_hash %d from port %d to port %d.\n",
+                       source_port->migration_target_qf.queue_id,
+                       source_port->migration_target_qf.flow_hash,
+                       source_port->id, source_port->migration_target_port_id);
+
+       /* We have a winner. */
+
+       source_port->migration_state = DSW_MIGRATION_STATE_PAUSING;
+       source_port->migration_start = rte_get_timer_cycles();
+
+       /* No need to go through the whole pause procedure for
+        * parallel queues, since atomic/ordered semantics need not to
+        * be maintained.
+        */
+
+       if (dsw->queues[source_port->migration_target_qf.queue_id].schedule_type
+           == RTE_SCHED_TYPE_PARALLEL) {
+               uint8_t queue_id = source_port->migration_target_qf.queue_id;
+               uint16_t flow_hash = source_port->migration_target_qf.flow_hash;
+               uint8_t dest_port_id = source_port->migration_target_port_id;
+
+               /* Single byte-sized stores are always atomic. */
+               dsw->queues[queue_id].flow_to_port_map[flow_hash] =
+                       dest_port_id;
+               rte_smp_wmb();
+
+               dsw_port_end_migration(dsw, source_port);
+
+               return;
+       }
+
+       /* There might be 'loopback' events already scheduled in the
+        * output buffers.
+        */
+       dsw_port_flush_out_buffers(dsw, source_port);
+
+       dsw_port_add_paused_flow(source_port,
+                                source_port->migration_target_qf.queue_id,
+                                source_port->migration_target_qf.flow_hash);
+
+       dsw_port_ctl_broadcast(dsw, source_port, DSW_CTL_PAUS_REQ,
+                              source_port->migration_target_qf.queue_id,
+                              source_port->migration_target_qf.flow_hash);
+       source_port->cfm_cnt = 0;
+}
+
+static void
+dsw_port_flush_paused_events(struct dsw_evdev *dsw,
+                            struct dsw_port *source_port,
+                            uint8_t queue_id, uint16_t paused_flow_hash);
+
+static void
+dsw_port_handle_unpause_flow(struct dsw_evdev *dsw, struct dsw_port *port,
+                            uint8_t originating_port_id, uint8_t queue_id,
+                            uint16_t paused_flow_hash)
+{
+       struct dsw_ctl_msg cfm = {
+               .type = DSW_CTL_CFM,
+               .originating_port_id = port->id,
+               .queue_id = queue_id,
+               .flow_hash = paused_flow_hash
+       };
+
+       DSW_LOG_DP_PORT(DEBUG, port->id, "Un-pausing queue_id %d flow_hash %d.\n",
+                       queue_id, paused_flow_hash);
+
+       dsw_port_remove_paused_flow(port, queue_id, paused_flow_hash);
+
+       rte_smp_rmb();
+
+       dsw_port_ctl_enqueue(&dsw->ports[originating_port_id], &cfm);
+
+       dsw_port_flush_paused_events(dsw, port, queue_id, paused_flow_hash);
+}
+
+#define FORWARD_BURST_SIZE (32)
+
+static void
+dsw_port_forward_migrated_flow(struct dsw_port *source_port,
+                              struct rte_event_ring *dest_ring,
+                              uint8_t queue_id,
+                              uint16_t flow_hash)
+{
+       uint16_t events_left;
+
+       /* Control ring message should been seen before the ring count
+        * is read on the port's in_ring.
+        */
+       rte_smp_rmb();
+
+       events_left = rte_event_ring_count(source_port->in_ring);
+
+       while (events_left > 0) {
+               uint16_t in_burst_size =
+                       RTE_MIN(FORWARD_BURST_SIZE, events_left);
+               struct rte_event in_burst[in_burst_size];
+               uint16_t in_len;
+               uint16_t i;
+
+               in_len = rte_event_ring_dequeue_burst(source_port->in_ring,
+                                                     in_burst,
+                                                     in_burst_size, NULL);
+               /* No need to care about bursting forwarded events (to
+                * the destination port's in_ring), since migration
+                * doesn't happen very often, and also the majority of
+                * the dequeued events will likely *not* be forwarded.
+                */
+               for (i = 0; i < in_len; i++) {
+                       struct rte_event *e = &in_burst[i];
+                       if (e->queue_id == queue_id &&
+                           dsw_flow_id_hash(e->flow_id) == flow_hash) {
+                               while (rte_event_ring_enqueue_burst(dest_ring,
+                                                                   e, 1,
+                                                                   NULL) != 1)
+                                       rte_pause();
+                       } else {
+                               uint16_t last_idx = source_port->in_buffer_len;
+                               source_port->in_buffer[last_idx] = *e;
+                               source_port->in_buffer_len++;
+                       }
+               }
+
+               events_left -= in_len;
+       }
+}
+
+static void
+dsw_port_move_migrating_flow(struct dsw_evdev *dsw,
+                            struct dsw_port *source_port)
+{
+       uint8_t queue_id = source_port->migration_target_qf.queue_id;
+       uint16_t flow_hash = source_port->migration_target_qf.flow_hash;
+       uint8_t dest_port_id = source_port->migration_target_port_id;
+       struct dsw_port *dest_port = &dsw->ports[dest_port_id];
+
+       dsw_port_flush_out_buffers(dsw, source_port);
+
+       rte_smp_wmb();
+
+       dsw->queues[queue_id].flow_to_port_map[flow_hash] =
+               dest_port_id;
+
+       dsw_port_forward_migrated_flow(source_port, dest_port->in_ring,
+                                      queue_id, flow_hash);
+
+       /* Flow table update and migration destination port's enqueues
+        * must be seen before the control message.
+        */
+       rte_smp_wmb();
+
+       dsw_port_ctl_broadcast(dsw, source_port, DSW_CTL_UNPAUS_REQ, queue_id,
+                              flow_hash);
+       source_port->cfm_cnt = 0;
+       source_port->migration_state = DSW_MIGRATION_STATE_UNPAUSING;
+}
+
+static void
+dsw_port_handle_confirm(struct dsw_evdev *dsw, struct dsw_port *port)
+{
+       port->cfm_cnt++;
+
+       if (port->cfm_cnt == (dsw->num_ports-1)) {
+               switch (port->migration_state) {
+               case DSW_MIGRATION_STATE_PAUSING:
+                       DSW_LOG_DP_PORT(DEBUG, port->id, "Going into forwarding "
+                                       "migration state.\n");
+                       port->migration_state = DSW_MIGRATION_STATE_FORWARDING;
+                       break;
+               case DSW_MIGRATION_STATE_UNPAUSING:
+                       dsw_port_end_migration(dsw, port);
+                       break;
+               default:
+                       RTE_ASSERT(0);
+                       break;
+               }
+       }
+}
+
+static void
+dsw_port_ctl_process(struct dsw_evdev *dsw, struct dsw_port *port)
+{
+       struct dsw_ctl_msg msg;
+
+       /* So any table loads happens before the ring dequeue, in the
+        * case of a 'paus' message.
+        */
+       rte_smp_rmb();
+
+       if (dsw_port_ctl_dequeue(port, &msg) == 0) {
+               switch (msg.type) {
+               case DSW_CTL_PAUS_REQ:
+                       dsw_port_handle_pause_flow(dsw, port,
+                                                  msg.originating_port_id,
+                                                  msg.queue_id, msg.flow_hash);
+                       break;
+               case DSW_CTL_UNPAUS_REQ:
+                       dsw_port_handle_unpause_flow(dsw, port,
+                                                    msg.originating_port_id,
+                                                    msg.queue_id,
+                                                    msg.flow_hash);
+                       break;
+               case DSW_CTL_CFM:
+                       dsw_port_handle_confirm(dsw, port);
+                       break;
+               }
+       }
+}
+
 static void
 dsw_port_note_op(struct dsw_port *port, uint16_t num_events)
 {
@@ -270,12 +933,24 @@ dsw_port_note_op(struct dsw_port *port, uint16_t num_events)
        port->ops_since_bg_task += (num_events+1);
 }
 
-static void
-dsw_port_flush_out_buffers(struct dsw_evdev *dsw, struct dsw_port *source_port);
-
 static void
 dsw_port_bg_process(struct dsw_evdev *dsw, struct dsw_port *port)
 {
+       if (unlikely(port->migration_state == DSW_MIGRATION_STATE_FORWARDING &&
+                    port->pending_releases == 0))
+               dsw_port_move_migrating_flow(dsw, port);
+
+       /* Polling the control ring is relatively inexpensive, and
+        * polling it often helps bringing down migration latency, so
+        * do this for every iteration.
+        */
+       dsw_port_ctl_process(dsw, port);
+
+       /* To avoid considering migration and flushing output buffers
+        * on every dequeue/enqueue call, the scheduler only performs
+        * such 'background' tasks every nth
+        * (i.e. DSW_MAX_PORT_OPS_PER_BG_TASK) operation.
+        */
        if (unlikely(port->ops_since_bg_task >= DSW_MAX_PORT_OPS_PER_BG_TASK)) {
                uint64_t now;
 
@@ -290,6 +965,8 @@ dsw_port_bg_process(struct dsw_evdev *dsw, struct dsw_port *port)
 
                dsw_port_consider_load_update(port, now);
 
+               dsw_port_consider_migration(dsw, port, now);
+
                port->ops_since_bg_task = 0;
        }
 }
@@ -339,7 +1016,6 @@ dsw_event_enqueue_burst_generic(void *port, const struct rte_event events[],
         */
        if (unlikely(events_len == 0)) {
                dsw_port_note_op(source_port, DSW_MAX_PORT_OPS_PER_BG_TASK);
-               dsw_port_flush_out_buffers(dsw, source_port);
                return 0;
        }
 
@@ -423,10 +1099,52 @@ dsw_event_dequeue(void *port, struct rte_event *events, uint64_t wait)
        return dsw_event_dequeue_burst(port, events, 1, wait);
 }
 
+static void
+dsw_port_record_seen_events(struct dsw_port *port, struct rte_event *events,
+                           uint16_t num)
+{
+       uint16_t i;
+
+       for (i = 0; i < num; i++) {
+               uint16_t l_idx = port->seen_events_idx;
+               struct dsw_queue_flow *qf = &port->seen_events[l_idx];
+               struct rte_event *event = &events[i];
+               qf->queue_id = event->queue_id;
+               qf->flow_hash = dsw_flow_id_hash(event->flow_id);
+
+               port->seen_events_idx = (l_idx+1) % DSW_MAX_EVENTS_RECORDED;
+       }
+
+       if (unlikely(port->seen_events_len != DSW_MAX_EVENTS_RECORDED))
+               port->seen_events_len =
+                       RTE_MIN(port->seen_events_len + num,
+                               DSW_MAX_EVENTS_RECORDED);
+}
+
 static uint16_t
 dsw_port_dequeue_burst(struct dsw_port *port, struct rte_event *events,
                       uint16_t num)
 {
+       struct dsw_port *source_port = port;
+       struct dsw_evdev *dsw = source_port->dsw;
+
+       dsw_port_ctl_process(dsw, source_port);
+
+       if (unlikely(port->in_buffer_len > 0)) {
+               uint16_t dequeued = RTE_MIN(num, port->in_buffer_len);
+
+               rte_memcpy(events, &port->in_buffer[port->in_buffer_start],
+                          dequeued * sizeof(struct rte_event));
+
+               port->in_buffer_start += dequeued;
+               port->in_buffer_len -= dequeued;
+
+               if (port->in_buffer_len == 0)
+                       port->in_buffer_start = 0;
+
+               return dequeued;
+       }
+
        return rte_event_ring_dequeue_burst(port->in_ring, events, num, NULL);
 }
 
@@ -458,6 +1176,15 @@ dsw_event_dequeue_burst(void *port, struct rte_event *events, uint16_t num,
                                dequeued);
 
                dsw_port_return_credits(dsw, source_port, dequeued);
+
+               /* One potential optimization one might think of is to
+                * add a migration state (prior to 'pausing'), and
+                * only record seen events when the port is in this
+                * state (and transit to 'pausing' when enough events
+                * have been gathered). However, that schema doesn't
+                * seem to improve performance.
+                */
+               dsw_port_record_seen_events(port, events, dequeued);
        }
        /* XXX: Assuming the port can't produce any more work,
         *      consider flushing the output buffer, on dequeued ==