1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2018 Ericsson AB
7 #ifdef DSW_SORT_DEQUEUED
14 #include <rte_cycles.h>
15 #include <rte_memcpy.h>
16 #include <rte_random.h>
19 dsw_port_acquire_credits(struct dsw_evdev *dsw, struct dsw_port *port,
22 int32_t inflight_credits = port->inflight_credits;
23 int32_t missing_credits = credits - inflight_credits;
24 int32_t total_on_loan;
26 int32_t acquired_credits;
27 int32_t new_total_on_loan;
29 if (likely(missing_credits <= 0)) {
30 port->inflight_credits -= credits;
35 __atomic_load_n(&dsw->credits_on_loan, __ATOMIC_RELAXED);
36 available = dsw->max_inflight - total_on_loan;
37 acquired_credits = RTE_MAX(missing_credits, DSW_PORT_MIN_CREDITS);
39 if (available < acquired_credits)
42 /* This is a race, no locks are involved, and thus some other
43 * thread can allocate tokens in between the check and the
47 __atomic_add_fetch(&dsw->credits_on_loan, acquired_credits,
50 if (unlikely(new_total_on_loan > dsw->max_inflight)) {
51 /* Some other port took the last credits */
52 __atomic_sub_fetch(&dsw->credits_on_loan, acquired_credits,
57 DSW_LOG_DP_PORT(DEBUG, port->id, "Acquired %d tokens from pool.\n",
60 port->inflight_credits += acquired_credits;
61 port->inflight_credits -= credits;
67 dsw_port_return_credits(struct dsw_evdev *dsw, struct dsw_port *port,
70 port->inflight_credits += credits;
72 if (unlikely(port->inflight_credits > DSW_PORT_MAX_CREDITS)) {
73 int32_t leave_credits = DSW_PORT_MIN_CREDITS;
74 int32_t return_credits =
75 port->inflight_credits - leave_credits;
77 port->inflight_credits = leave_credits;
79 __atomic_sub_fetch(&dsw->credits_on_loan, return_credits,
82 DSW_LOG_DP_PORT(DEBUG, port->id,
83 "Returned %d tokens to pool.\n",
89 dsw_port_enqueue_stats(struct dsw_port *port, uint16_t num_new,
90 uint16_t num_forward, uint16_t num_release)
92 port->new_enqueued += num_new;
93 port->forward_enqueued += num_forward;
94 port->release_enqueued += num_release;
98 dsw_port_queue_enqueue_stats(struct dsw_port *source_port, uint8_t queue_id)
100 source_port->queue_enqueued[queue_id]++;
104 dsw_port_dequeue_stats(struct dsw_port *port, uint16_t num)
106 port->dequeued += num;
110 dsw_port_queue_dequeued_stats(struct dsw_port *source_port, uint8_t queue_id)
112 source_port->queue_dequeued[queue_id]++;
116 dsw_port_load_record(struct dsw_port *port, unsigned int dequeued)
118 if (dequeued > 0 && port->busy_start == 0)
119 /* work period begins */
120 port->busy_start = rte_get_timer_cycles();
121 else if (dequeued == 0 && port->busy_start > 0) {
122 /* work period ends */
123 uint64_t work_period =
124 rte_get_timer_cycles() - port->busy_start;
125 port->busy_cycles += work_period;
126 port->busy_start = 0;
131 dsw_port_load_close_period(struct dsw_port *port, uint64_t now)
133 uint64_t passed = now - port->measurement_start;
134 uint64_t busy_cycles = port->busy_cycles;
136 if (port->busy_start > 0) {
137 busy_cycles += (now - port->busy_start);
138 port->busy_start = now;
141 int16_t load = (DSW_MAX_LOAD * busy_cycles) / passed;
143 port->measurement_start = now;
144 port->busy_cycles = 0;
146 port->total_busy_cycles += busy_cycles;
152 dsw_port_load_update(struct dsw_port *port, uint64_t now)
158 old_load = __atomic_load_n(&port->load, __ATOMIC_RELAXED);
160 period_load = dsw_port_load_close_period(port, now);
162 new_load = (period_load + old_load*DSW_OLD_LOAD_WEIGHT) /
163 (DSW_OLD_LOAD_WEIGHT+1);
165 __atomic_store_n(&port->load, new_load, __ATOMIC_RELAXED);
167 /* The load of the recently immigrated flows should hopefully
168 * be reflected the load estimate by now.
170 __atomic_store_n(&port->immigration_load, 0, __ATOMIC_RELAXED);
174 dsw_port_consider_load_update(struct dsw_port *port, uint64_t now)
176 if (now < port->next_load_update)
179 port->next_load_update = now + port->load_update_interval;
181 dsw_port_load_update(port, now);
185 dsw_port_ctl_enqueue(struct dsw_port *port, struct dsw_ctl_msg *msg)
187 /* there's always room on the ring */
188 while (rte_ring_enqueue_elem(port->ctl_in_ring, msg, sizeof(*msg)) != 0)
193 dsw_port_ctl_dequeue(struct dsw_port *port, struct dsw_ctl_msg *msg)
195 return rte_ring_dequeue_elem(port->ctl_in_ring, msg, sizeof(*msg));
199 dsw_port_ctl_broadcast(struct dsw_evdev *dsw, struct dsw_port *source_port,
200 uint8_t type, struct dsw_queue_flow *qfs,
204 struct dsw_ctl_msg msg = {
206 .originating_port_id = source_port->id,
210 memcpy(msg.qfs, qfs, sizeof(struct dsw_queue_flow) * qfs_len);
212 for (port_id = 0; port_id < dsw->num_ports; port_id++)
213 if (port_id != source_port->id)
214 dsw_port_ctl_enqueue(&dsw->ports[port_id], &msg);
217 static __rte_always_inline bool
218 dsw_is_queue_flow_in_ary(const struct dsw_queue_flow *qfs, uint16_t qfs_len,
219 uint8_t queue_id, uint16_t flow_hash)
223 for (i = 0; i < qfs_len; i++)
224 if (qfs[i].queue_id == queue_id &&
225 qfs[i].flow_hash == flow_hash)
231 static __rte_always_inline bool
232 dsw_port_is_flow_paused(struct dsw_port *port, uint8_t queue_id,
235 return dsw_is_queue_flow_in_ary(port->paused_flows,
236 port->paused_flows_len,
237 queue_id, flow_hash);
241 dsw_port_add_paused_flows(struct dsw_port *port, struct dsw_queue_flow *qfs,
246 for (i = 0; i < qfs_len; i++) {
247 struct dsw_queue_flow *qf = &qfs[i];
249 DSW_LOG_DP_PORT(DEBUG, port->id,
250 "Pausing queue_id %d flow_hash %d.\n",
251 qf->queue_id, qf->flow_hash);
253 port->paused_flows[port->paused_flows_len] = *qf;
254 port->paused_flows_len++;
259 dsw_port_remove_paused_flow(struct dsw_port *port,
260 struct dsw_queue_flow *target_qf)
264 for (i = 0; i < port->paused_flows_len; i++) {
265 struct dsw_queue_flow *qf = &port->paused_flows[i];
267 if (qf->queue_id == target_qf->queue_id &&
268 qf->flow_hash == target_qf->flow_hash) {
269 uint16_t last_idx = port->paused_flows_len-1;
271 port->paused_flows[i] =
272 port->paused_flows[last_idx];
273 port->paused_flows_len--;
280 dsw_port_remove_paused_flows(struct dsw_port *port,
281 struct dsw_queue_flow *qfs, uint8_t qfs_len)
285 for (i = 0; i < qfs_len; i++)
286 dsw_port_remove_paused_flow(port, &qfs[i]);
291 dsw_port_flush_out_buffers(struct dsw_evdev *dsw, struct dsw_port *source_port);
294 dsw_port_handle_pause_flows(struct dsw_evdev *dsw, struct dsw_port *port,
295 uint8_t originating_port_id,
296 struct dsw_queue_flow *paused_qfs,
299 struct dsw_ctl_msg cfm = {
301 .originating_port_id = port->id
304 /* There might be already-scheduled events belonging to the
305 * paused flow in the output buffers.
307 dsw_port_flush_out_buffers(dsw, port);
309 dsw_port_add_paused_flows(port, paused_qfs, qfs_len);
311 /* Make sure any stores to the original port's in_ring is seen
312 * before the ctl message.
316 dsw_port_ctl_enqueue(&dsw->ports[originating_port_id], &cfm);
319 struct dsw_queue_flow_burst {
320 struct dsw_queue_flow queue_flow;
324 #define DSW_QF_TO_INT(_qf) \
325 ((int)((((_qf)->queue_id)<<16)|((_qf)->flow_hash)))
328 dsw_cmp_qf(const void *v_qf_a, const void *v_qf_b)
330 const struct dsw_queue_flow *qf_a = v_qf_a;
331 const struct dsw_queue_flow *qf_b = v_qf_b;
333 return DSW_QF_TO_INT(qf_a) - DSW_QF_TO_INT(qf_b);
337 dsw_sort_qfs_to_bursts(struct dsw_queue_flow *qfs, uint16_t qfs_len,
338 struct dsw_queue_flow_burst *bursts)
341 struct dsw_queue_flow_burst *current_burst = NULL;
342 uint16_t num_bursts = 0;
344 /* We don't need the stable property, and the list is likely
345 * large enough for qsort() to outperform dsw_stable_sort(),
346 * so we use qsort() here.
348 qsort(qfs, qfs_len, sizeof(qfs[0]), dsw_cmp_qf);
350 /* arrange the (now-consecutive) events into bursts */
351 for (i = 0; i < qfs_len; i++) {
353 dsw_cmp_qf(&qfs[i], ¤t_burst->queue_flow) != 0) {
354 current_burst = &bursts[num_bursts];
355 current_burst->queue_flow = qfs[i];
356 current_burst->count = 0;
359 current_burst->count++;
366 dsw_retrieve_port_loads(struct dsw_evdev *dsw, int16_t *port_loads,
369 bool below_limit = false;
372 for (i = 0; i < dsw->num_ports; i++) {
373 int16_t measured_load =
374 __atomic_load_n(&dsw->ports[i].load, __ATOMIC_RELAXED);
375 int32_t immigration_load =
376 __atomic_load_n(&dsw->ports[i].immigration_load,
378 int32_t load = measured_load + immigration_load;
380 load = RTE_MIN(load, DSW_MAX_LOAD);
382 if (load < load_limit)
384 port_loads[i] = load;
390 dsw_flow_load(uint16_t num_events, int16_t port_load)
392 return ((int32_t)port_load * (int32_t)num_events) /
393 DSW_MAX_EVENTS_RECORDED;
397 dsw_evaluate_migration(int16_t source_load, int16_t target_load,
400 int32_t res_target_load;
403 if (target_load > DSW_MAX_TARGET_LOAD_FOR_MIGRATION)
406 imbalance = source_load - target_load;
408 if (imbalance < DSW_REBALANCE_THRESHOLD)
411 res_target_load = target_load + flow_load;
413 /* If the estimated load of the target port will be higher
414 * than the source port's load, it doesn't make sense to move
417 if (res_target_load > source_load)
420 /* The more idle the target will be, the better. This will
421 * make migration prefer moving smaller flows, and flows to
422 * lightly loaded ports.
424 return DSW_MAX_LOAD - res_target_load;
428 dsw_is_serving_port(struct dsw_evdev *dsw, uint8_t port_id, uint8_t queue_id)
430 struct dsw_queue *queue = &dsw->queues[queue_id];
433 for (i = 0; i < queue->num_serving_ports; i++)
434 if (queue->serving_ports[i] == port_id)
441 dsw_select_emigration_target(struct dsw_evdev *dsw,
442 struct dsw_queue_flow_burst *bursts,
443 uint16_t num_bursts, uint8_t source_port_id,
444 int16_t *port_loads, uint16_t num_ports,
445 uint8_t *target_port_ids,
446 struct dsw_queue_flow *target_qfs,
447 uint8_t *targets_len)
449 int16_t source_port_load = port_loads[source_port_id];
450 struct dsw_queue_flow *candidate_qf = NULL;
451 uint8_t candidate_port_id = 0;
452 int16_t candidate_weight = -1;
453 int16_t candidate_flow_load = -1;
456 if (source_port_load < DSW_MIN_SOURCE_LOAD_FOR_MIGRATION)
459 for (i = 0; i < num_bursts; i++) {
460 struct dsw_queue_flow_burst *burst = &bursts[i];
461 struct dsw_queue_flow *qf = &burst->queue_flow;
465 if (dsw_is_queue_flow_in_ary(target_qfs, *targets_len,
466 qf->queue_id, qf->flow_hash))
469 flow_load = dsw_flow_load(burst->count, source_port_load);
471 for (port_id = 0; port_id < num_ports; port_id++) {
474 if (port_id == source_port_id)
477 if (!dsw_is_serving_port(dsw, port_id, qf->queue_id))
480 weight = dsw_evaluate_migration(source_port_load,
484 if (weight > candidate_weight) {
486 candidate_port_id = port_id;
487 candidate_weight = weight;
488 candidate_flow_load = flow_load;
493 if (candidate_weight < 0)
496 DSW_LOG_DP_PORT(DEBUG, source_port_id, "Selected queue_id %d "
497 "flow_hash %d (with flow load %d) for migration "
498 "to port %d.\n", candidate_qf->queue_id,
499 candidate_qf->flow_hash,
500 DSW_LOAD_TO_PERCENT(candidate_flow_load),
503 port_loads[candidate_port_id] += candidate_flow_load;
504 port_loads[source_port_id] -= candidate_flow_load;
506 target_port_ids[*targets_len] = candidate_port_id;
507 target_qfs[*targets_len] = *candidate_qf;
510 __atomic_add_fetch(&dsw->ports[candidate_port_id].immigration_load,
511 candidate_flow_load, __ATOMIC_RELAXED);
517 dsw_select_emigration_targets(struct dsw_evdev *dsw,
518 struct dsw_port *source_port,
519 struct dsw_queue_flow_burst *bursts,
520 uint16_t num_bursts, int16_t *port_loads)
522 struct dsw_queue_flow *target_qfs = source_port->emigration_target_qfs;
523 uint8_t *target_port_ids = source_port->emigration_target_port_ids;
524 uint8_t *targets_len = &source_port->emigration_targets_len;
527 for (i = 0; i < DSW_MAX_FLOWS_PER_MIGRATION; i++) {
530 found = dsw_select_emigration_target(dsw, bursts, num_bursts,
532 port_loads, dsw->num_ports,
540 if (*targets_len == 0)
541 DSW_LOG_DP_PORT(DEBUG, source_port->id,
542 "For the %d flows considered, no target port "
543 "was found.\n", num_bursts);
547 dsw_schedule(struct dsw_evdev *dsw, uint8_t queue_id, uint16_t flow_hash)
549 struct dsw_queue *queue = &dsw->queues[queue_id];
552 if (queue->num_serving_ports > 1)
553 port_id = queue->flow_to_port_map[flow_hash];
555 /* A single-link queue, or atomic/ordered/parallel but
556 * with just a single serving port.
558 port_id = queue->serving_ports[0];
560 DSW_LOG_DP(DEBUG, "Event with queue_id %d flow_hash %d is scheduled "
561 "to port %d.\n", queue_id, flow_hash, port_id);
567 dsw_port_transmit_buffered(struct dsw_evdev *dsw, struct dsw_port *source_port,
568 uint8_t dest_port_id)
570 struct dsw_port *dest_port = &(dsw->ports[dest_port_id]);
571 uint16_t *buffer_len = &source_port->out_buffer_len[dest_port_id];
572 struct rte_event *buffer = source_port->out_buffer[dest_port_id];
573 uint16_t enqueued = 0;
575 if (*buffer_len == 0)
578 /* The rings are dimensioned to fit all in-flight events (even
579 * on a single ring), so looping will work.
583 rte_event_ring_enqueue_burst(dest_port->in_ring,
585 *buffer_len-enqueued,
587 } while (unlikely(enqueued != *buffer_len));
593 dsw_port_get_parallel_flow_id(struct dsw_port *port)
595 uint16_t flow_id = port->next_parallel_flow_id;
597 port->next_parallel_flow_id =
598 (port->next_parallel_flow_id + 1) % DSW_PARALLEL_FLOWS;
604 dsw_port_buffer_paused(struct dsw_port *port,
605 const struct rte_event *paused_event)
607 port->paused_events[port->paused_events_len] = *paused_event;
608 port->paused_events_len++;
612 dsw_port_buffer_non_paused(struct dsw_evdev *dsw, struct dsw_port *source_port,
613 uint8_t dest_port_id, const struct rte_event *event)
615 struct rte_event *buffer = source_port->out_buffer[dest_port_id];
616 uint16_t *buffer_len = &source_port->out_buffer_len[dest_port_id];
618 if (*buffer_len == DSW_MAX_PORT_OUT_BUFFER)
619 dsw_port_transmit_buffered(dsw, source_port, dest_port_id);
621 buffer[*buffer_len] = *event;
626 #define DSW_FLOW_ID_BITS (24)
628 dsw_flow_id_hash(uint32_t flow_id)
634 hash ^= ((flow_id >> offset) & DSW_MAX_FLOWS_MASK);
635 offset += DSW_MAX_FLOWS_BITS;
636 } while (offset < DSW_FLOW_ID_BITS);
642 dsw_port_buffer_parallel(struct dsw_evdev *dsw, struct dsw_port *source_port,
643 struct rte_event event)
645 uint8_t dest_port_id;
647 event.flow_id = dsw_port_get_parallel_flow_id(source_port);
649 dest_port_id = dsw_schedule(dsw, event.queue_id,
650 dsw_flow_id_hash(event.flow_id));
652 dsw_port_buffer_non_paused(dsw, source_port, dest_port_id, &event);
656 dsw_port_buffer_event(struct dsw_evdev *dsw, struct dsw_port *source_port,
657 const struct rte_event *event)
660 uint8_t dest_port_id;
662 if (unlikely(dsw->queues[event->queue_id].schedule_type ==
663 RTE_SCHED_TYPE_PARALLEL)) {
664 dsw_port_buffer_parallel(dsw, source_port, *event);
668 flow_hash = dsw_flow_id_hash(event->flow_id);
670 if (unlikely(dsw_port_is_flow_paused(source_port, event->queue_id,
672 dsw_port_buffer_paused(source_port, event);
676 dest_port_id = dsw_schedule(dsw, event->queue_id, flow_hash);
678 dsw_port_buffer_non_paused(dsw, source_port, dest_port_id, event);
682 dsw_port_flush_paused_events(struct dsw_evdev *dsw,
683 struct dsw_port *source_port,
684 const struct dsw_queue_flow *qf)
686 uint16_t paused_events_len = source_port->paused_events_len;
687 struct rte_event paused_events[paused_events_len];
688 uint8_t dest_port_id;
691 if (paused_events_len == 0)
694 if (dsw_port_is_flow_paused(source_port, qf->queue_id, qf->flow_hash))
697 rte_memcpy(paused_events, source_port->paused_events,
698 paused_events_len * sizeof(struct rte_event));
700 source_port->paused_events_len = 0;
702 dest_port_id = dsw_schedule(dsw, qf->queue_id, qf->flow_hash);
704 for (i = 0; i < paused_events_len; i++) {
705 struct rte_event *event = &paused_events[i];
708 flow_hash = dsw_flow_id_hash(event->flow_id);
710 if (event->queue_id == qf->queue_id &&
711 flow_hash == qf->flow_hash)
712 dsw_port_buffer_non_paused(dsw, source_port,
713 dest_port_id, event);
715 dsw_port_buffer_paused(source_port, event);
720 dsw_port_emigration_stats(struct dsw_port *port, uint8_t finished)
722 uint64_t flow_migration_latency;
724 flow_migration_latency =
725 (rte_get_timer_cycles() - port->emigration_start);
726 port->emigration_latency += (flow_migration_latency * finished);
727 port->emigrations += finished;
731 dsw_port_end_emigration(struct dsw_evdev *dsw, struct dsw_port *port,
732 uint8_t schedule_type)
735 struct dsw_queue_flow left_qfs[DSW_MAX_FLOWS_PER_MIGRATION];
736 uint8_t left_port_ids[DSW_MAX_FLOWS_PER_MIGRATION];
737 uint8_t left_qfs_len = 0;
740 for (i = 0; i < port->emigration_targets_len; i++) {
741 struct dsw_queue_flow *qf = &port->emigration_target_qfs[i];
742 uint8_t queue_id = qf->queue_id;
743 uint8_t queue_schedule_type =
744 dsw->queues[queue_id].schedule_type;
745 uint16_t flow_hash = qf->flow_hash;
747 if (queue_schedule_type != schedule_type) {
748 left_port_ids[left_qfs_len] =
749 port->emigration_target_port_ids[i];
750 left_qfs[left_qfs_len] = *qf;
755 DSW_LOG_DP_PORT(DEBUG, port->id, "Migration completed for "
756 "queue_id %d flow_hash %d.\n", queue_id,
759 if (queue_schedule_type == RTE_SCHED_TYPE_ATOMIC) {
760 dsw_port_remove_paused_flow(port, qf);
761 dsw_port_flush_paused_events(dsw, port, qf);
765 finished = port->emigration_targets_len - left_qfs_len;
768 dsw_port_emigration_stats(port, finished);
770 for (i = 0; i < left_qfs_len; i++) {
771 port->emigration_target_port_ids[i] = left_port_ids[i];
772 port->emigration_target_qfs[i] = left_qfs[i];
774 port->emigration_targets_len = left_qfs_len;
776 if (port->emigration_targets_len == 0) {
777 port->migration_state = DSW_MIGRATION_STATE_IDLE;
778 port->seen_events_len = 0;
783 dsw_port_move_parallel_flows(struct dsw_evdev *dsw,
784 struct dsw_port *source_port)
788 for (i = 0; i < source_port->emigration_targets_len; i++) {
789 struct dsw_queue_flow *qf =
790 &source_port->emigration_target_qfs[i];
791 uint8_t queue_id = qf->queue_id;
793 if (dsw->queues[queue_id].schedule_type ==
794 RTE_SCHED_TYPE_PARALLEL) {
795 uint8_t dest_port_id =
796 source_port->emigration_target_port_ids[i];
797 uint16_t flow_hash = qf->flow_hash;
799 /* Single byte-sized stores are always atomic. */
800 dsw->queues[queue_id].flow_to_port_map[flow_hash] =
807 dsw_port_end_emigration(dsw, source_port, RTE_SCHED_TYPE_PARALLEL);
811 dsw_port_consider_emigration(struct dsw_evdev *dsw,
812 struct dsw_port *source_port,
815 bool any_port_below_limit;
816 struct dsw_queue_flow *seen_events = source_port->seen_events;
817 uint16_t seen_events_len = source_port->seen_events_len;
818 struct dsw_queue_flow_burst bursts[DSW_MAX_EVENTS_RECORDED];
820 int16_t source_port_load;
821 int16_t port_loads[dsw->num_ports];
823 if (now < source_port->next_emigration)
826 if (dsw->num_ports == 1)
829 if (seen_events_len < DSW_MAX_EVENTS_RECORDED)
832 DSW_LOG_DP_PORT(DEBUG, source_port->id, "Considering emigration.\n");
834 /* Randomize interval to avoid having all threads considering
835 * emigration at the same in point in time, which might lead
836 * to all choosing the same target port.
838 source_port->next_emigration = now +
839 source_port->migration_interval / 2 +
840 rte_rand() % source_port->migration_interval;
842 if (source_port->migration_state != DSW_MIGRATION_STATE_IDLE) {
843 DSW_LOG_DP_PORT(DEBUG, source_port->id,
844 "Emigration already in progress.\n");
848 /* For simplicity, avoid migration in the unlikely case there
849 * is still events to consume in the in_buffer (from the last
852 if (source_port->in_buffer_len > 0) {
853 DSW_LOG_DP_PORT(DEBUG, source_port->id, "There are still "
854 "events in the input buffer.\n");
859 __atomic_load_n(&source_port->load, __ATOMIC_RELAXED);
860 if (source_port_load < DSW_MIN_SOURCE_LOAD_FOR_MIGRATION) {
861 DSW_LOG_DP_PORT(DEBUG, source_port->id,
862 "Load %d is below threshold level %d.\n",
863 DSW_LOAD_TO_PERCENT(source_port_load),
864 DSW_LOAD_TO_PERCENT(DSW_MIN_SOURCE_LOAD_FOR_MIGRATION));
868 /* Avoid starting any expensive operations (sorting etc), in
869 * case of a scenario with all ports above the load limit.
871 any_port_below_limit =
872 dsw_retrieve_port_loads(dsw, port_loads,
873 DSW_MAX_TARGET_LOAD_FOR_MIGRATION);
874 if (!any_port_below_limit) {
875 DSW_LOG_DP_PORT(DEBUG, source_port->id,
876 "Candidate target ports are all too highly "
881 num_bursts = dsw_sort_qfs_to_bursts(seen_events, seen_events_len,
884 /* For non-big-little systems, there's no point in moving the
887 if (num_bursts < 2) {
888 DSW_LOG_DP_PORT(DEBUG, source_port->id, "Only a single flow "
889 "queue_id %d flow_hash %d has been seen.\n",
890 bursts[0].queue_flow.queue_id,
891 bursts[0].queue_flow.flow_hash);
895 dsw_select_emigration_targets(dsw, source_port, bursts, num_bursts,
898 if (source_port->emigration_targets_len == 0)
901 source_port->migration_state = DSW_MIGRATION_STATE_PAUSING;
902 source_port->emigration_start = rte_get_timer_cycles();
904 /* No need to go through the whole pause procedure for
905 * parallel queues, since atomic/ordered semantics need not to
908 dsw_port_move_parallel_flows(dsw, source_port);
910 /* All flows were on PARALLEL queues. */
911 if (source_port->migration_state == DSW_MIGRATION_STATE_IDLE)
914 /* There might be 'loopback' events already scheduled in the
917 dsw_port_flush_out_buffers(dsw, source_port);
919 dsw_port_add_paused_flows(source_port,
920 source_port->emigration_target_qfs,
921 source_port->emigration_targets_len);
923 dsw_port_ctl_broadcast(dsw, source_port, DSW_CTL_PAUS_REQ,
924 source_port->emigration_target_qfs,
925 source_port->emigration_targets_len);
926 source_port->cfm_cnt = 0;
930 dsw_port_flush_paused_events(struct dsw_evdev *dsw,
931 struct dsw_port *source_port,
932 const struct dsw_queue_flow *qf);
935 dsw_port_handle_unpause_flows(struct dsw_evdev *dsw, struct dsw_port *port,
936 uint8_t originating_port_id,
937 struct dsw_queue_flow *paused_qfs,
941 struct dsw_ctl_msg cfm = {
943 .originating_port_id = port->id
946 dsw_port_remove_paused_flows(port, paused_qfs, qfs_len);
950 dsw_port_ctl_enqueue(&dsw->ports[originating_port_id], &cfm);
952 for (i = 0; i < qfs_len; i++) {
953 struct dsw_queue_flow *qf = &paused_qfs[i];
955 if (dsw_schedule(dsw, qf->queue_id, qf->flow_hash) == port->id)
956 port->immigrations++;
958 dsw_port_flush_paused_events(dsw, port, qf);
962 #define FORWARD_BURST_SIZE (32)
965 dsw_port_forward_emigrated_flow(struct dsw_port *source_port,
966 struct rte_event_ring *dest_ring,
970 uint16_t events_left;
972 /* Control ring message should been seen before the ring count
973 * is read on the port's in_ring.
977 events_left = rte_event_ring_count(source_port->in_ring);
979 while (events_left > 0) {
980 uint16_t in_burst_size =
981 RTE_MIN(FORWARD_BURST_SIZE, events_left);
982 struct rte_event in_burst[in_burst_size];
986 in_len = rte_event_ring_dequeue_burst(source_port->in_ring,
988 in_burst_size, NULL);
989 /* No need to care about bursting forwarded events (to
990 * the destination port's in_ring), since migration
991 * doesn't happen very often, and also the majority of
992 * the dequeued events will likely *not* be forwarded.
994 for (i = 0; i < in_len; i++) {
995 struct rte_event *e = &in_burst[i];
996 if (e->queue_id == queue_id &&
997 dsw_flow_id_hash(e->flow_id) == flow_hash) {
998 while (rte_event_ring_enqueue_burst(dest_ring,
1003 uint16_t last_idx = source_port->in_buffer_len;
1004 source_port->in_buffer[last_idx] = *e;
1005 source_port->in_buffer_len++;
1009 events_left -= in_len;
1014 dsw_port_move_emigrating_flows(struct dsw_evdev *dsw,
1015 struct dsw_port *source_port)
1019 dsw_port_flush_out_buffers(dsw, source_port);
1023 for (i = 0; i < source_port->emigration_targets_len; i++) {
1024 struct dsw_queue_flow *qf =
1025 &source_port->emigration_target_qfs[i];
1026 uint8_t dest_port_id =
1027 source_port->emigration_target_port_ids[i];
1028 struct dsw_port *dest_port = &dsw->ports[dest_port_id];
1030 dsw->queues[qf->queue_id].flow_to_port_map[qf->flow_hash] =
1033 dsw_port_forward_emigrated_flow(source_port, dest_port->in_ring,
1034 qf->queue_id, qf->flow_hash);
1037 /* Flow table update and migration destination port's enqueues
1038 * must be seen before the control message.
1042 dsw_port_ctl_broadcast(dsw, source_port, DSW_CTL_UNPAUS_REQ,
1043 source_port->emigration_target_qfs,
1044 source_port->emigration_targets_len);
1045 source_port->cfm_cnt = 0;
1046 source_port->migration_state = DSW_MIGRATION_STATE_UNPAUSING;
1050 dsw_port_handle_confirm(struct dsw_evdev *dsw, struct dsw_port *port)
1054 if (port->cfm_cnt == (dsw->num_ports-1)) {
1055 switch (port->migration_state) {
1056 case DSW_MIGRATION_STATE_PAUSING:
1057 DSW_LOG_DP_PORT(DEBUG, port->id, "Going into forwarding "
1058 "migration state.\n");
1059 port->migration_state = DSW_MIGRATION_STATE_FORWARDING;
1061 case DSW_MIGRATION_STATE_UNPAUSING:
1062 dsw_port_end_emigration(dsw, port,
1063 RTE_SCHED_TYPE_ATOMIC);
1073 dsw_port_ctl_process(struct dsw_evdev *dsw, struct dsw_port *port)
1075 struct dsw_ctl_msg msg;
1077 if (dsw_port_ctl_dequeue(port, &msg) == 0) {
1079 case DSW_CTL_PAUS_REQ:
1080 dsw_port_handle_pause_flows(dsw, port,
1081 msg.originating_port_id,
1082 msg.qfs, msg.qfs_len);
1084 case DSW_CTL_UNPAUS_REQ:
1085 dsw_port_handle_unpause_flows(dsw, port,
1086 msg.originating_port_id,
1087 msg.qfs, msg.qfs_len);
1090 dsw_port_handle_confirm(dsw, port);
1097 dsw_port_note_op(struct dsw_port *port, uint16_t num_events)
1099 /* To pull the control ring reasonbly often on busy ports,
1100 * each dequeued/enqueued event is considered an 'op' too.
1102 port->ops_since_bg_task += (num_events+1);
1106 dsw_port_bg_process(struct dsw_evdev *dsw, struct dsw_port *port)
1108 if (unlikely(port->migration_state == DSW_MIGRATION_STATE_FORWARDING &&
1109 port->pending_releases == 0))
1110 dsw_port_move_emigrating_flows(dsw, port);
1112 /* Polling the control ring is relatively inexpensive, and
1113 * polling it often helps bringing down migration latency, so
1114 * do this for every iteration.
1116 dsw_port_ctl_process(dsw, port);
1118 /* To avoid considering migration and flushing output buffers
1119 * on every dequeue/enqueue call, the scheduler only performs
1120 * such 'background' tasks every nth
1121 * (i.e. DSW_MAX_PORT_OPS_PER_BG_TASK) operation.
1123 if (unlikely(port->ops_since_bg_task >= DSW_MAX_PORT_OPS_PER_BG_TASK)) {
1126 now = rte_get_timer_cycles();
1128 port->last_bg = now;
1130 /* Logic to avoid having events linger in the output
1133 dsw_port_flush_out_buffers(dsw, port);
1135 dsw_port_consider_load_update(port, now);
1137 dsw_port_consider_emigration(dsw, port, now);
1139 port->ops_since_bg_task = 0;
1144 dsw_port_flush_out_buffers(struct dsw_evdev *dsw, struct dsw_port *source_port)
1146 uint16_t dest_port_id;
1148 for (dest_port_id = 0; dest_port_id < dsw->num_ports; dest_port_id++)
1149 dsw_port_transmit_buffered(dsw, source_port, dest_port_id);
1153 dsw_event_enqueue(void *port, const struct rte_event *ev)
1155 return dsw_event_enqueue_burst(port, ev, unlikely(ev == NULL) ? 0 : 1);
1158 static __rte_always_inline uint16_t
1159 dsw_event_enqueue_burst_generic(struct dsw_port *source_port,
1160 const struct rte_event events[],
1161 uint16_t events_len, bool op_types_known,
1162 uint16_t num_new, uint16_t num_release,
1163 uint16_t num_non_release)
1165 struct dsw_evdev *dsw = source_port->dsw;
1166 bool enough_credits;
1169 DSW_LOG_DP_PORT(DEBUG, source_port->id, "Attempting to enqueue %d "
1170 "events to port %d.\n", events_len, source_port->id);
1172 dsw_port_bg_process(dsw, source_port);
1174 /* XXX: For performance (=ring efficiency) reasons, the
1175 * scheduler relies on internal non-ring buffers instead of
1176 * immediately sending the event to the destination ring. For
1177 * a producer that doesn't intend to produce or consume any
1178 * more events, the scheduler provides a way to flush the
1179 * buffer, by means of doing an enqueue of zero events. In
1180 * addition, a port cannot be left "unattended" (e.g. unused)
1181 * for long periods of time, since that would stall
1182 * migration. Eventdev API extensions to provide a cleaner way
1183 * to archieve both of these functions should be
1186 if (unlikely(events_len == 0)) {
1187 dsw_port_note_op(source_port, DSW_MAX_PORT_OPS_PER_BG_TASK);
1188 dsw_port_flush_out_buffers(dsw, source_port);
1192 dsw_port_note_op(source_port, events_len);
1194 if (!op_types_known)
1195 for (i = 0; i < events_len; i++) {
1196 switch (events[i].op) {
1197 case RTE_EVENT_OP_RELEASE:
1200 case RTE_EVENT_OP_NEW:
1202 /* Falls through. */
1209 /* Technically, we could allow the non-new events up to the
1210 * first new event in the array into the system, but for
1211 * simplicity reasons, we deny the whole burst if the port is
1212 * above the water mark.
1214 if (unlikely(num_new > 0 &&
1215 __atomic_load_n(&dsw->credits_on_loan, __ATOMIC_RELAXED) >
1216 source_port->new_event_threshold))
1219 enough_credits = dsw_port_acquire_credits(dsw, source_port,
1221 if (unlikely(!enough_credits))
1224 source_port->pending_releases -= num_release;
1226 dsw_port_enqueue_stats(source_port, num_new,
1227 num_non_release-num_new, num_release);
1229 for (i = 0; i < events_len; i++) {
1230 const struct rte_event *event = &events[i];
1232 if (likely(num_release == 0 ||
1233 event->op != RTE_EVENT_OP_RELEASE))
1234 dsw_port_buffer_event(dsw, source_port, event);
1235 dsw_port_queue_enqueue_stats(source_port, event->queue_id);
1238 DSW_LOG_DP_PORT(DEBUG, source_port->id, "%d non-release events "
1239 "accepted.\n", num_non_release);
1241 return (num_non_release + num_release);
1245 dsw_event_enqueue_burst(void *port, const struct rte_event events[],
1246 uint16_t events_len)
1248 struct dsw_port *source_port = port;
1250 if (unlikely(events_len > source_port->enqueue_depth))
1251 events_len = source_port->enqueue_depth;
1253 return dsw_event_enqueue_burst_generic(source_port, events,
1254 events_len, false, 0, 0, 0);
1258 dsw_event_enqueue_new_burst(void *port, const struct rte_event events[],
1259 uint16_t events_len)
1261 struct dsw_port *source_port = port;
1263 if (unlikely(events_len > source_port->enqueue_depth))
1264 events_len = source_port->enqueue_depth;
1266 return dsw_event_enqueue_burst_generic(source_port, events,
1267 events_len, true, events_len,
1272 dsw_event_enqueue_forward_burst(void *port, const struct rte_event events[],
1273 uint16_t events_len)
1275 struct dsw_port *source_port = port;
1277 if (unlikely(events_len > source_port->enqueue_depth))
1278 events_len = source_port->enqueue_depth;
1280 return dsw_event_enqueue_burst_generic(source_port, events,
1281 events_len, true, 0, 0,
1286 dsw_event_dequeue(void *port, struct rte_event *events, uint64_t wait)
1288 return dsw_event_dequeue_burst(port, events, 1, wait);
1292 dsw_port_record_seen_events(struct dsw_port *port, struct rte_event *events,
1297 dsw_port_dequeue_stats(port, num);
1299 for (i = 0; i < num; i++) {
1300 uint16_t l_idx = port->seen_events_idx;
1301 struct dsw_queue_flow *qf = &port->seen_events[l_idx];
1302 struct rte_event *event = &events[i];
1303 qf->queue_id = event->queue_id;
1304 qf->flow_hash = dsw_flow_id_hash(event->flow_id);
1306 port->seen_events_idx = (l_idx+1) % DSW_MAX_EVENTS_RECORDED;
1308 dsw_port_queue_dequeued_stats(port, event->queue_id);
1311 if (unlikely(port->seen_events_len != DSW_MAX_EVENTS_RECORDED))
1312 port->seen_events_len =
1313 RTE_MIN(port->seen_events_len + num,
1314 DSW_MAX_EVENTS_RECORDED);
1317 #ifdef DSW_SORT_DEQUEUED
1319 #define DSW_EVENT_TO_INT(_event) \
1320 ((int)((((_event)->queue_id)<<16)|((_event)->flow_id)))
1323 dsw_cmp_event(const void *v_event_a, const void *v_event_b)
1325 const struct rte_event *event_a = v_event_a;
1326 const struct rte_event *event_b = v_event_b;
1328 return DSW_EVENT_TO_INT(event_a) - DSW_EVENT_TO_INT(event_b);
1333 dsw_port_dequeue_burst(struct dsw_port *port, struct rte_event *events,
1336 if (unlikely(port->in_buffer_len > 0)) {
1337 uint16_t dequeued = RTE_MIN(num, port->in_buffer_len);
1339 rte_memcpy(events, &port->in_buffer[port->in_buffer_start],
1340 dequeued * sizeof(struct rte_event));
1342 port->in_buffer_start += dequeued;
1343 port->in_buffer_len -= dequeued;
1345 if (port->in_buffer_len == 0)
1346 port->in_buffer_start = 0;
1351 return rte_event_ring_dequeue_burst(port->in_ring, events, num, NULL);
1355 dsw_event_dequeue_burst(void *port, struct rte_event *events, uint16_t num,
1356 uint64_t wait __rte_unused)
1358 struct dsw_port *source_port = port;
1359 struct dsw_evdev *dsw = source_port->dsw;
1362 source_port->pending_releases = 0;
1364 dsw_port_bg_process(dsw, source_port);
1366 if (unlikely(num > source_port->dequeue_depth))
1367 num = source_port->dequeue_depth;
1369 dequeued = dsw_port_dequeue_burst(source_port, events, num);
1371 source_port->pending_releases = dequeued;
1373 dsw_port_load_record(source_port, dequeued);
1375 dsw_port_note_op(source_port, dequeued);
1378 DSW_LOG_DP_PORT(DEBUG, source_port->id, "Dequeued %d events.\n",
1381 dsw_port_return_credits(dsw, source_port, dequeued);
1383 /* One potential optimization one might think of is to
1384 * add a migration state (prior to 'pausing'), and
1385 * only record seen events when the port is in this
1386 * state (and transit to 'pausing' when enough events
1387 * have been gathered). However, that schema doesn't
1388 * seem to improve performance.
1390 dsw_port_record_seen_events(port, events, dequeued);
1391 } else /* Zero-size dequeue means a likely idle port, and thus
1392 * we can afford trading some efficiency for a slightly
1393 * reduced event wall-time latency.
1395 dsw_port_flush_out_buffers(dsw, port);
1397 #ifdef DSW_SORT_DEQUEUED
1398 dsw_stable_sort(events, dequeued, sizeof(events[0]), dsw_cmp_event);