1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2018 Ericsson AB
7 #ifdef DSW_SORT_DEQUEUED
14 #include <rte_atomic.h>
15 #include <rte_cycles.h>
16 #include <rte_memcpy.h>
17 #include <rte_random.h>
20 dsw_port_acquire_credits(struct dsw_evdev *dsw, struct dsw_port *port,
23 int32_t inflight_credits = port->inflight_credits;
24 int32_t missing_credits = credits - inflight_credits;
25 int32_t total_on_loan;
27 int32_t acquired_credits;
28 int32_t new_total_on_loan;
30 if (likely(missing_credits <= 0)) {
31 port->inflight_credits -= credits;
35 total_on_loan = rte_atomic32_read(&dsw->credits_on_loan);
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
46 new_total_on_loan = rte_atomic32_add_return(&dsw->credits_on_loan,
49 if (unlikely(new_total_on_loan > dsw->max_inflight)) {
50 /* Some other port took the last credits */
51 rte_atomic32_sub(&dsw->credits_on_loan, acquired_credits);
55 DSW_LOG_DP_PORT(DEBUG, port->id, "Acquired %d tokens from pool.\n",
58 port->inflight_credits += acquired_credits;
59 port->inflight_credits -= credits;
65 dsw_port_return_credits(struct dsw_evdev *dsw, struct dsw_port *port,
68 port->inflight_credits += credits;
70 if (unlikely(port->inflight_credits > DSW_PORT_MAX_CREDITS)) {
71 int32_t leave_credits = DSW_PORT_MIN_CREDITS;
72 int32_t return_credits =
73 port->inflight_credits - leave_credits;
75 port->inflight_credits = leave_credits;
77 rte_atomic32_sub(&dsw->credits_on_loan, return_credits);
79 DSW_LOG_DP_PORT(DEBUG, port->id,
80 "Returned %d tokens to pool.\n",
86 dsw_port_enqueue_stats(struct dsw_port *port, uint16_t num_new,
87 uint16_t num_forward, uint16_t num_release)
89 port->new_enqueued += num_new;
90 port->forward_enqueued += num_forward;
91 port->release_enqueued += num_release;
95 dsw_port_queue_enqueue_stats(struct dsw_port *source_port, uint8_t queue_id)
97 source_port->queue_enqueued[queue_id]++;
101 dsw_port_dequeue_stats(struct dsw_port *port, uint16_t num)
103 port->dequeued += num;
107 dsw_port_queue_dequeued_stats(struct dsw_port *source_port, uint8_t queue_id)
109 source_port->queue_dequeued[queue_id]++;
113 dsw_port_load_record(struct dsw_port *port, unsigned int dequeued)
115 if (dequeued > 0 && port->busy_start == 0)
116 /* work period begins */
117 port->busy_start = rte_get_timer_cycles();
118 else if (dequeued == 0 && port->busy_start > 0) {
119 /* work period ends */
120 uint64_t work_period =
121 rte_get_timer_cycles() - port->busy_start;
122 port->busy_cycles += work_period;
123 port->busy_start = 0;
128 dsw_port_load_close_period(struct dsw_port *port, uint64_t now)
130 uint64_t passed = now - port->measurement_start;
131 uint64_t busy_cycles = port->busy_cycles;
133 if (port->busy_start > 0) {
134 busy_cycles += (now - port->busy_start);
135 port->busy_start = now;
138 int16_t load = (DSW_MAX_LOAD * busy_cycles) / passed;
140 port->measurement_start = now;
141 port->busy_cycles = 0;
143 port->total_busy_cycles += busy_cycles;
149 dsw_port_load_update(struct dsw_port *port, uint64_t now)
155 old_load = rte_atomic16_read(&port->load);
157 period_load = dsw_port_load_close_period(port, now);
159 new_load = (period_load + old_load*DSW_OLD_LOAD_WEIGHT) /
160 (DSW_OLD_LOAD_WEIGHT+1);
162 rte_atomic16_set(&port->load, new_load);
164 /* The load of the recently immigrated flows should hopefully
165 * be reflected the load estimate by now.
167 rte_atomic32_set(&port->immigration_load, 0);
171 dsw_port_consider_load_update(struct dsw_port *port, uint64_t now)
173 if (now < port->next_load_update)
176 port->next_load_update = now + port->load_update_interval;
178 dsw_port_load_update(port, now);
182 dsw_port_ctl_enqueue(struct dsw_port *port, struct dsw_ctl_msg *msg)
184 /* there's always room on the ring */
185 while (rte_ring_enqueue_elem(port->ctl_in_ring, msg, sizeof(*msg)) != 0)
190 dsw_port_ctl_dequeue(struct dsw_port *port, struct dsw_ctl_msg *msg)
192 return rte_ring_dequeue_elem(port->ctl_in_ring, msg, sizeof(*msg));
196 dsw_port_ctl_broadcast(struct dsw_evdev *dsw, struct dsw_port *source_port,
197 uint8_t type, struct dsw_queue_flow *qfs,
201 struct dsw_ctl_msg msg = {
203 .originating_port_id = source_port->id,
207 memcpy(msg.qfs, qfs, sizeof(struct dsw_queue_flow) * qfs_len);
209 for (port_id = 0; port_id < dsw->num_ports; port_id++)
210 if (port_id != source_port->id)
211 dsw_port_ctl_enqueue(&dsw->ports[port_id], &msg);
214 static __rte_always_inline bool
215 dsw_is_queue_flow_in_ary(const struct dsw_queue_flow *qfs, uint16_t qfs_len,
216 uint8_t queue_id, uint16_t flow_hash)
220 for (i = 0; i < qfs_len; i++)
221 if (qfs[i].queue_id == queue_id &&
222 qfs[i].flow_hash == flow_hash)
228 static __rte_always_inline bool
229 dsw_port_is_flow_paused(struct dsw_port *port, uint8_t queue_id,
232 return dsw_is_queue_flow_in_ary(port->paused_flows,
233 port->paused_flows_len,
234 queue_id, flow_hash);
238 dsw_port_add_paused_flows(struct dsw_port *port, struct dsw_queue_flow *qfs,
243 for (i = 0; i < qfs_len; i++) {
244 struct dsw_queue_flow *qf = &qfs[i];
246 DSW_LOG_DP_PORT(DEBUG, port->id,
247 "Pausing queue_id %d flow_hash %d.\n",
248 qf->queue_id, qf->flow_hash);
250 port->paused_flows[port->paused_flows_len] = *qf;
251 port->paused_flows_len++;
256 dsw_port_remove_paused_flow(struct dsw_port *port,
257 struct dsw_queue_flow *target_qf)
261 for (i = 0; i < port->paused_flows_len; i++) {
262 struct dsw_queue_flow *qf = &port->paused_flows[i];
264 if (qf->queue_id == target_qf->queue_id &&
265 qf->flow_hash == target_qf->flow_hash) {
266 uint16_t last_idx = port->paused_flows_len-1;
268 port->paused_flows[i] =
269 port->paused_flows[last_idx];
270 port->paused_flows_len--;
277 dsw_port_remove_paused_flows(struct dsw_port *port,
278 struct dsw_queue_flow *qfs, uint8_t qfs_len)
282 for (i = 0; i < qfs_len; i++)
283 dsw_port_remove_paused_flow(port, &qfs[i]);
288 dsw_port_flush_out_buffers(struct dsw_evdev *dsw, struct dsw_port *source_port);
291 dsw_port_handle_pause_flows(struct dsw_evdev *dsw, struct dsw_port *port,
292 uint8_t originating_port_id,
293 struct dsw_queue_flow *paused_qfs,
296 struct dsw_ctl_msg cfm = {
298 .originating_port_id = port->id
301 /* There might be already-scheduled events belonging to the
302 * paused flow in the output buffers.
304 dsw_port_flush_out_buffers(dsw, port);
306 dsw_port_add_paused_flows(port, paused_qfs, qfs_len);
308 /* Make sure any stores to the original port's in_ring is seen
309 * before the ctl message.
313 dsw_port_ctl_enqueue(&dsw->ports[originating_port_id], &cfm);
316 struct dsw_queue_flow_burst {
317 struct dsw_queue_flow queue_flow;
321 #define DSW_QF_TO_INT(_qf) \
322 ((int)((((_qf)->queue_id)<<16)|((_qf)->flow_hash)))
325 dsw_cmp_qf(const void *v_qf_a, const void *v_qf_b)
327 const struct dsw_queue_flow *qf_a = v_qf_a;
328 const struct dsw_queue_flow *qf_b = v_qf_b;
330 return DSW_QF_TO_INT(qf_a) - DSW_QF_TO_INT(qf_b);
334 dsw_sort_qfs_to_bursts(struct dsw_queue_flow *qfs, uint16_t qfs_len,
335 struct dsw_queue_flow_burst *bursts)
338 struct dsw_queue_flow_burst *current_burst = NULL;
339 uint16_t num_bursts = 0;
341 /* We don't need the stable property, and the list is likely
342 * large enough for qsort() to outperform dsw_stable_sort(),
343 * so we use qsort() here.
345 qsort(qfs, qfs_len, sizeof(qfs[0]), dsw_cmp_qf);
347 /* arrange the (now-consecutive) events into bursts */
348 for (i = 0; i < qfs_len; i++) {
350 dsw_cmp_qf(&qfs[i], ¤t_burst->queue_flow) != 0) {
351 current_burst = &bursts[num_bursts];
352 current_burst->queue_flow = qfs[i];
353 current_burst->count = 0;
356 current_burst->count++;
363 dsw_retrieve_port_loads(struct dsw_evdev *dsw, int16_t *port_loads,
366 bool below_limit = false;
369 for (i = 0; i < dsw->num_ports; i++) {
370 int16_t measured_load = rte_atomic16_read(&dsw->ports[i].load);
371 int32_t immigration_load =
372 rte_atomic32_read(&dsw->ports[i].immigration_load);
373 int32_t load = measured_load + immigration_load;
375 load = RTE_MIN(load, DSW_MAX_LOAD);
377 if (load < load_limit)
379 port_loads[i] = load;
385 dsw_flow_load(uint16_t num_events, int16_t port_load)
387 return ((int32_t)port_load * (int32_t)num_events) /
388 DSW_MAX_EVENTS_RECORDED;
392 dsw_evaluate_migration(int16_t source_load, int16_t target_load,
395 int32_t res_target_load;
398 if (target_load > DSW_MAX_TARGET_LOAD_FOR_MIGRATION)
401 imbalance = source_load - target_load;
403 if (imbalance < DSW_REBALANCE_THRESHOLD)
406 res_target_load = target_load + flow_load;
408 /* If the estimated load of the target port will be higher
409 * than the source port's load, it doesn't make sense to move
412 if (res_target_load > source_load)
415 /* The more idle the target will be, the better. This will
416 * make migration prefer moving smaller flows, and flows to
417 * lightly loaded ports.
419 return DSW_MAX_LOAD - res_target_load;
423 dsw_is_serving_port(struct dsw_evdev *dsw, uint8_t port_id, uint8_t queue_id)
425 struct dsw_queue *queue = &dsw->queues[queue_id];
428 for (i = 0; i < queue->num_serving_ports; i++)
429 if (queue->serving_ports[i] == port_id)
436 dsw_select_emigration_target(struct dsw_evdev *dsw,
437 struct dsw_queue_flow_burst *bursts,
438 uint16_t num_bursts, uint8_t source_port_id,
439 int16_t *port_loads, uint16_t num_ports,
440 uint8_t *target_port_ids,
441 struct dsw_queue_flow *target_qfs,
442 uint8_t *targets_len)
444 int16_t source_port_load = port_loads[source_port_id];
445 struct dsw_queue_flow *candidate_qf;
446 uint8_t candidate_port_id;
447 int16_t candidate_weight = -1;
448 int16_t candidate_flow_load;
451 if (source_port_load < DSW_MIN_SOURCE_LOAD_FOR_MIGRATION)
454 for (i = 0; i < num_bursts; i++) {
455 struct dsw_queue_flow_burst *burst = &bursts[i];
456 struct dsw_queue_flow *qf = &burst->queue_flow;
460 if (dsw_is_queue_flow_in_ary(target_qfs, *targets_len,
461 qf->queue_id, qf->flow_hash))
464 flow_load = dsw_flow_load(burst->count, source_port_load);
466 for (port_id = 0; port_id < num_ports; port_id++) {
469 if (port_id == source_port_id)
472 if (!dsw_is_serving_port(dsw, port_id, qf->queue_id))
475 weight = dsw_evaluate_migration(source_port_load,
479 if (weight > candidate_weight) {
481 candidate_port_id = port_id;
482 candidate_weight = weight;
483 candidate_flow_load = flow_load;
488 if (candidate_weight < 0)
491 DSW_LOG_DP_PORT(DEBUG, source_port_id, "Selected queue_id %d "
492 "flow_hash %d (with flow load %d) for migration "
493 "to port %d.\n", candidate_qf->queue_id,
494 candidate_qf->flow_hash,
495 DSW_LOAD_TO_PERCENT(candidate_flow_load),
498 port_loads[candidate_port_id] += candidate_flow_load;
499 port_loads[source_port_id] -= candidate_flow_load;
501 target_port_ids[*targets_len] = candidate_port_id;
502 target_qfs[*targets_len] = *candidate_qf;
505 rte_atomic32_add(&dsw->ports[candidate_port_id].immigration_load,
506 candidate_flow_load);
512 dsw_select_emigration_targets(struct dsw_evdev *dsw,
513 struct dsw_port *source_port,
514 struct dsw_queue_flow_burst *bursts,
515 uint16_t num_bursts, int16_t *port_loads)
517 struct dsw_queue_flow *target_qfs = source_port->emigration_target_qfs;
518 uint8_t *target_port_ids = source_port->emigration_target_port_ids;
519 uint8_t *targets_len = &source_port->emigration_targets_len;
522 for (i = 0; i < DSW_MAX_FLOWS_PER_MIGRATION; i++) {
525 found = dsw_select_emigration_target(dsw, bursts, num_bursts,
527 port_loads, dsw->num_ports,
535 if (*targets_len == 0)
536 DSW_LOG_DP_PORT(DEBUG, source_port->id,
537 "For the %d flows considered, no target port "
538 "was found.\n", num_bursts);
542 dsw_schedule(struct dsw_evdev *dsw, uint8_t queue_id, uint16_t flow_hash)
544 struct dsw_queue *queue = &dsw->queues[queue_id];
547 if (queue->num_serving_ports > 1)
548 port_id = queue->flow_to_port_map[flow_hash];
550 /* A single-link queue, or atomic/ordered/parallel but
551 * with just a single serving port.
553 port_id = queue->serving_ports[0];
555 DSW_LOG_DP(DEBUG, "Event with queue_id %d flow_hash %d is scheduled "
556 "to port %d.\n", queue_id, flow_hash, port_id);
562 dsw_port_transmit_buffered(struct dsw_evdev *dsw, struct dsw_port *source_port,
563 uint8_t dest_port_id)
565 struct dsw_port *dest_port = &(dsw->ports[dest_port_id]);
566 uint16_t *buffer_len = &source_port->out_buffer_len[dest_port_id];
567 struct rte_event *buffer = source_port->out_buffer[dest_port_id];
568 uint16_t enqueued = 0;
570 if (*buffer_len == 0)
573 /* The rings are dimensioned to fit all in-flight events (even
574 * on a single ring), so looping will work.
578 rte_event_ring_enqueue_burst(dest_port->in_ring,
580 *buffer_len-enqueued,
582 } while (unlikely(enqueued != *buffer_len));
588 dsw_port_get_parallel_flow_id(struct dsw_port *port)
590 uint16_t flow_id = port->next_parallel_flow_id;
592 port->next_parallel_flow_id =
593 (port->next_parallel_flow_id + 1) % DSW_PARALLEL_FLOWS;
599 dsw_port_buffer_paused(struct dsw_port *port,
600 const struct rte_event *paused_event)
602 port->paused_events[port->paused_events_len] = *paused_event;
603 port->paused_events_len++;
607 dsw_port_buffer_non_paused(struct dsw_evdev *dsw, struct dsw_port *source_port,
608 uint8_t dest_port_id, const struct rte_event *event)
610 struct rte_event *buffer = source_port->out_buffer[dest_port_id];
611 uint16_t *buffer_len = &source_port->out_buffer_len[dest_port_id];
613 if (*buffer_len == DSW_MAX_PORT_OUT_BUFFER)
614 dsw_port_transmit_buffered(dsw, source_port, dest_port_id);
616 buffer[*buffer_len] = *event;
621 #define DSW_FLOW_ID_BITS (24)
623 dsw_flow_id_hash(uint32_t flow_id)
629 hash ^= ((flow_id >> offset) & DSW_MAX_FLOWS_MASK);
630 offset += DSW_MAX_FLOWS_BITS;
631 } while (offset < DSW_FLOW_ID_BITS);
637 dsw_port_buffer_parallel(struct dsw_evdev *dsw, struct dsw_port *source_port,
638 struct rte_event event)
640 uint8_t dest_port_id;
642 event.flow_id = dsw_port_get_parallel_flow_id(source_port);
644 dest_port_id = dsw_schedule(dsw, event.queue_id,
645 dsw_flow_id_hash(event.flow_id));
647 dsw_port_buffer_non_paused(dsw, source_port, dest_port_id, &event);
651 dsw_port_buffer_event(struct dsw_evdev *dsw, struct dsw_port *source_port,
652 const struct rte_event *event)
655 uint8_t dest_port_id;
657 if (unlikely(dsw->queues[event->queue_id].schedule_type ==
658 RTE_SCHED_TYPE_PARALLEL)) {
659 dsw_port_buffer_parallel(dsw, source_port, *event);
663 flow_hash = dsw_flow_id_hash(event->flow_id);
665 if (unlikely(dsw_port_is_flow_paused(source_port, event->queue_id,
667 dsw_port_buffer_paused(source_port, event);
671 dest_port_id = dsw_schedule(dsw, event->queue_id, flow_hash);
673 dsw_port_buffer_non_paused(dsw, source_port, dest_port_id, event);
677 dsw_port_flush_paused_events(struct dsw_evdev *dsw,
678 struct dsw_port *source_port,
679 const struct dsw_queue_flow *qf)
681 uint16_t paused_events_len = source_port->paused_events_len;
682 struct rte_event paused_events[paused_events_len];
683 uint8_t dest_port_id;
686 if (paused_events_len == 0)
689 if (dsw_port_is_flow_paused(source_port, qf->queue_id, qf->flow_hash))
692 rte_memcpy(paused_events, source_port->paused_events,
693 paused_events_len * sizeof(struct rte_event));
695 source_port->paused_events_len = 0;
697 dest_port_id = dsw_schedule(dsw, qf->queue_id, qf->flow_hash);
699 for (i = 0; i < paused_events_len; i++) {
700 struct rte_event *event = &paused_events[i];
703 flow_hash = dsw_flow_id_hash(event->flow_id);
705 if (event->queue_id == qf->queue_id &&
706 flow_hash == qf->flow_hash)
707 dsw_port_buffer_non_paused(dsw, source_port,
708 dest_port_id, event);
710 dsw_port_buffer_paused(source_port, event);
715 dsw_port_emigration_stats(struct dsw_port *port, uint8_t finished)
717 uint64_t flow_migration_latency;
719 flow_migration_latency =
720 (rte_get_timer_cycles() - port->emigration_start);
721 port->emigration_latency += (flow_migration_latency * finished);
722 port->emigrations += finished;
726 dsw_port_end_emigration(struct dsw_evdev *dsw, struct dsw_port *port,
727 uint8_t schedule_type)
730 struct dsw_queue_flow left_qfs[DSW_MAX_FLOWS_PER_MIGRATION];
731 uint8_t left_port_ids[DSW_MAX_FLOWS_PER_MIGRATION];
732 uint8_t left_qfs_len = 0;
735 for (i = 0; i < port->emigration_targets_len; i++) {
736 struct dsw_queue_flow *qf = &port->emigration_target_qfs[i];
737 uint8_t queue_id = qf->queue_id;
738 uint8_t queue_schedule_type =
739 dsw->queues[queue_id].schedule_type;
740 uint16_t flow_hash = qf->flow_hash;
742 if (queue_schedule_type != schedule_type) {
743 left_port_ids[left_qfs_len] =
744 port->emigration_target_port_ids[i];
745 left_qfs[left_qfs_len] = *qf;
750 DSW_LOG_DP_PORT(DEBUG, port->id, "Migration completed for "
751 "queue_id %d flow_hash %d.\n", queue_id,
754 if (queue_schedule_type == RTE_SCHED_TYPE_ATOMIC) {
755 dsw_port_remove_paused_flow(port, qf);
756 dsw_port_flush_paused_events(dsw, port, qf);
760 finished = port->emigration_targets_len - left_qfs_len;
763 dsw_port_emigration_stats(port, finished);
765 for (i = 0; i < left_qfs_len; i++) {
766 port->emigration_target_port_ids[i] = left_port_ids[i];
767 port->emigration_target_qfs[i] = left_qfs[i];
769 port->emigration_targets_len = left_qfs_len;
771 if (port->emigration_targets_len == 0) {
772 port->migration_state = DSW_MIGRATION_STATE_IDLE;
773 port->seen_events_len = 0;
778 dsw_port_move_parallel_flows(struct dsw_evdev *dsw,
779 struct dsw_port *source_port)
783 for (i = 0; i < source_port->emigration_targets_len; i++) {
784 struct dsw_queue_flow *qf =
785 &source_port->emigration_target_qfs[i];
786 uint8_t queue_id = qf->queue_id;
788 if (dsw->queues[queue_id].schedule_type ==
789 RTE_SCHED_TYPE_PARALLEL) {
790 uint8_t dest_port_id =
791 source_port->emigration_target_port_ids[i];
792 uint16_t flow_hash = qf->flow_hash;
794 /* Single byte-sized stores are always atomic. */
795 dsw->queues[queue_id].flow_to_port_map[flow_hash] =
802 dsw_port_end_emigration(dsw, source_port, RTE_SCHED_TYPE_PARALLEL);
806 dsw_port_consider_emigration(struct dsw_evdev *dsw,
807 struct dsw_port *source_port,
810 bool any_port_below_limit;
811 struct dsw_queue_flow *seen_events = source_port->seen_events;
812 uint16_t seen_events_len = source_port->seen_events_len;
813 struct dsw_queue_flow_burst bursts[DSW_MAX_EVENTS_RECORDED];
815 int16_t source_port_load;
816 int16_t port_loads[dsw->num_ports];
818 if (now < source_port->next_emigration)
821 if (dsw->num_ports == 1)
824 DSW_LOG_DP_PORT(DEBUG, source_port->id, "Considering emigration.\n");
826 /* Randomize interval to avoid having all threads considering
827 * emigration at the same in point in time, which might lead
828 * to all choosing the same target port.
830 source_port->next_emigration = now +
831 source_port->migration_interval / 2 +
832 rte_rand() % source_port->migration_interval;
834 if (source_port->migration_state != DSW_MIGRATION_STATE_IDLE) {
835 DSW_LOG_DP_PORT(DEBUG, source_port->id,
836 "Emigration already in progress.\n");
840 /* For simplicity, avoid migration in the unlikely case there
841 * is still events to consume in the in_buffer (from the last
844 if (source_port->in_buffer_len > 0) {
845 DSW_LOG_DP_PORT(DEBUG, source_port->id, "There are still "
846 "events in the input buffer.\n");
850 source_port_load = rte_atomic16_read(&source_port->load);
851 if (source_port_load < DSW_MIN_SOURCE_LOAD_FOR_MIGRATION) {
852 DSW_LOG_DP_PORT(DEBUG, source_port->id,
853 "Load %d is below threshold level %d.\n",
854 DSW_LOAD_TO_PERCENT(source_port_load),
855 DSW_LOAD_TO_PERCENT(DSW_MIN_SOURCE_LOAD_FOR_MIGRATION));
859 /* Avoid starting any expensive operations (sorting etc), in
860 * case of a scenario with all ports above the load limit.
862 any_port_below_limit =
863 dsw_retrieve_port_loads(dsw, port_loads,
864 DSW_MAX_TARGET_LOAD_FOR_MIGRATION);
865 if (!any_port_below_limit) {
866 DSW_LOG_DP_PORT(DEBUG, source_port->id,
867 "Candidate target ports are all too highly "
872 num_bursts = dsw_sort_qfs_to_bursts(seen_events, seen_events_len,
875 /* For non-big-little systems, there's no point in moving the
878 if (num_bursts < 2) {
879 DSW_LOG_DP_PORT(DEBUG, source_port->id, "Only a single flow "
880 "queue_id %d flow_hash %d has been seen.\n",
881 bursts[0].queue_flow.queue_id,
882 bursts[0].queue_flow.flow_hash);
886 dsw_select_emigration_targets(dsw, source_port, bursts, num_bursts,
889 if (source_port->emigration_targets_len == 0)
892 source_port->migration_state = DSW_MIGRATION_STATE_PAUSING;
893 source_port->emigration_start = rte_get_timer_cycles();
895 /* No need to go through the whole pause procedure for
896 * parallel queues, since atomic/ordered semantics need not to
899 dsw_port_move_parallel_flows(dsw, source_port);
901 /* All flows were on PARALLEL queues. */
902 if (source_port->migration_state == DSW_MIGRATION_STATE_IDLE)
905 /* There might be 'loopback' events already scheduled in the
908 dsw_port_flush_out_buffers(dsw, source_port);
910 dsw_port_add_paused_flows(source_port,
911 source_port->emigration_target_qfs,
912 source_port->emigration_targets_len);
914 dsw_port_ctl_broadcast(dsw, source_port, DSW_CTL_PAUS_REQ,
915 source_port->emigration_target_qfs,
916 source_port->emigration_targets_len);
917 source_port->cfm_cnt = 0;
921 dsw_port_flush_paused_events(struct dsw_evdev *dsw,
922 struct dsw_port *source_port,
923 const struct dsw_queue_flow *qf);
926 dsw_port_handle_unpause_flows(struct dsw_evdev *dsw, struct dsw_port *port,
927 uint8_t originating_port_id,
928 struct dsw_queue_flow *paused_qfs,
932 struct dsw_ctl_msg cfm = {
934 .originating_port_id = port->id
937 dsw_port_remove_paused_flows(port, paused_qfs, qfs_len);
941 dsw_port_ctl_enqueue(&dsw->ports[originating_port_id], &cfm);
943 for (i = 0; i < qfs_len; i++) {
944 struct dsw_queue_flow *qf = &paused_qfs[i];
946 if (dsw_schedule(dsw, qf->queue_id, qf->flow_hash) == port->id)
947 port->immigrations++;
949 dsw_port_flush_paused_events(dsw, port, qf);
953 #define FORWARD_BURST_SIZE (32)
956 dsw_port_forward_emigrated_flow(struct dsw_port *source_port,
957 struct rte_event_ring *dest_ring,
961 uint16_t events_left;
963 /* Control ring message should been seen before the ring count
964 * is read on the port's in_ring.
968 events_left = rte_event_ring_count(source_port->in_ring);
970 while (events_left > 0) {
971 uint16_t in_burst_size =
972 RTE_MIN(FORWARD_BURST_SIZE, events_left);
973 struct rte_event in_burst[in_burst_size];
977 in_len = rte_event_ring_dequeue_burst(source_port->in_ring,
979 in_burst_size, NULL);
980 /* No need to care about bursting forwarded events (to
981 * the destination port's in_ring), since migration
982 * doesn't happen very often, and also the majority of
983 * the dequeued events will likely *not* be forwarded.
985 for (i = 0; i < in_len; i++) {
986 struct rte_event *e = &in_burst[i];
987 if (e->queue_id == queue_id &&
988 dsw_flow_id_hash(e->flow_id) == flow_hash) {
989 while (rte_event_ring_enqueue_burst(dest_ring,
994 uint16_t last_idx = source_port->in_buffer_len;
995 source_port->in_buffer[last_idx] = *e;
996 source_port->in_buffer_len++;
1000 events_left -= in_len;
1005 dsw_port_move_emigrating_flows(struct dsw_evdev *dsw,
1006 struct dsw_port *source_port)
1010 dsw_port_flush_out_buffers(dsw, source_port);
1014 for (i = 0; i < source_port->emigration_targets_len; i++) {
1015 struct dsw_queue_flow *qf =
1016 &source_port->emigration_target_qfs[i];
1017 uint8_t dest_port_id =
1018 source_port->emigration_target_port_ids[i];
1019 struct dsw_port *dest_port = &dsw->ports[dest_port_id];
1021 dsw->queues[qf->queue_id].flow_to_port_map[qf->flow_hash] =
1024 dsw_port_forward_emigrated_flow(source_port, dest_port->in_ring,
1025 qf->queue_id, qf->flow_hash);
1028 /* Flow table update and migration destination port's enqueues
1029 * must be seen before the control message.
1033 dsw_port_ctl_broadcast(dsw, source_port, DSW_CTL_UNPAUS_REQ,
1034 source_port->emigration_target_qfs,
1035 source_port->emigration_targets_len);
1036 source_port->cfm_cnt = 0;
1037 source_port->migration_state = DSW_MIGRATION_STATE_UNPAUSING;
1041 dsw_port_handle_confirm(struct dsw_evdev *dsw, struct dsw_port *port)
1045 if (port->cfm_cnt == (dsw->num_ports-1)) {
1046 switch (port->migration_state) {
1047 case DSW_MIGRATION_STATE_PAUSING:
1048 DSW_LOG_DP_PORT(DEBUG, port->id, "Going into forwarding "
1049 "migration state.\n");
1050 port->migration_state = DSW_MIGRATION_STATE_FORWARDING;
1052 case DSW_MIGRATION_STATE_UNPAUSING:
1053 dsw_port_end_emigration(dsw, port,
1054 RTE_SCHED_TYPE_ATOMIC);
1064 dsw_port_ctl_process(struct dsw_evdev *dsw, struct dsw_port *port)
1066 struct dsw_ctl_msg msg;
1068 if (dsw_port_ctl_dequeue(port, &msg) == 0) {
1070 case DSW_CTL_PAUS_REQ:
1071 dsw_port_handle_pause_flows(dsw, port,
1072 msg.originating_port_id,
1073 msg.qfs, msg.qfs_len);
1075 case DSW_CTL_UNPAUS_REQ:
1076 dsw_port_handle_unpause_flows(dsw, port,
1077 msg.originating_port_id,
1078 msg.qfs, msg.qfs_len);
1081 dsw_port_handle_confirm(dsw, port);
1088 dsw_port_note_op(struct dsw_port *port, uint16_t num_events)
1090 /* To pull the control ring reasonbly often on busy ports,
1091 * each dequeued/enqueued event is considered an 'op' too.
1093 port->ops_since_bg_task += (num_events+1);
1097 dsw_port_bg_process(struct dsw_evdev *dsw, struct dsw_port *port)
1099 if (unlikely(port->migration_state == DSW_MIGRATION_STATE_FORWARDING &&
1100 port->pending_releases == 0))
1101 dsw_port_move_emigrating_flows(dsw, port);
1103 /* Polling the control ring is relatively inexpensive, and
1104 * polling it often helps bringing down migration latency, so
1105 * do this for every iteration.
1107 dsw_port_ctl_process(dsw, port);
1109 /* To avoid considering migration and flushing output buffers
1110 * on every dequeue/enqueue call, the scheduler only performs
1111 * such 'background' tasks every nth
1112 * (i.e. DSW_MAX_PORT_OPS_PER_BG_TASK) operation.
1114 if (unlikely(port->ops_since_bg_task >= DSW_MAX_PORT_OPS_PER_BG_TASK)) {
1117 now = rte_get_timer_cycles();
1119 port->last_bg = now;
1121 /* Logic to avoid having events linger in the output
1124 dsw_port_flush_out_buffers(dsw, port);
1126 dsw_port_consider_load_update(port, now);
1128 dsw_port_consider_emigration(dsw, port, now);
1130 port->ops_since_bg_task = 0;
1135 dsw_port_flush_out_buffers(struct dsw_evdev *dsw, struct dsw_port *source_port)
1137 uint16_t dest_port_id;
1139 for (dest_port_id = 0; dest_port_id < dsw->num_ports; dest_port_id++)
1140 dsw_port_transmit_buffered(dsw, source_port, dest_port_id);
1144 dsw_event_enqueue(void *port, const struct rte_event *ev)
1146 return dsw_event_enqueue_burst(port, ev, unlikely(ev == NULL) ? 0 : 1);
1149 static __rte_always_inline uint16_t
1150 dsw_event_enqueue_burst_generic(struct dsw_port *source_port,
1151 const struct rte_event events[],
1152 uint16_t events_len, bool op_types_known,
1153 uint16_t num_new, uint16_t num_release,
1154 uint16_t num_non_release)
1156 struct dsw_evdev *dsw = source_port->dsw;
1157 bool enough_credits;
1160 DSW_LOG_DP_PORT(DEBUG, source_port->id, "Attempting to enqueue %d "
1161 "events to port %d.\n", events_len, source_port->id);
1163 dsw_port_bg_process(dsw, source_port);
1165 /* XXX: For performance (=ring efficiency) reasons, the
1166 * scheduler relies on internal non-ring buffers instead of
1167 * immediately sending the event to the destination ring. For
1168 * a producer that doesn't intend to produce or consume any
1169 * more events, the scheduler provides a way to flush the
1170 * buffer, by means of doing an enqueue of zero events. In
1171 * addition, a port cannot be left "unattended" (e.g. unused)
1172 * for long periods of time, since that would stall
1173 * migration. Eventdev API extensions to provide a cleaner way
1174 * to archieve both of these functions should be
1177 if (unlikely(events_len == 0)) {
1178 dsw_port_note_op(source_port, DSW_MAX_PORT_OPS_PER_BG_TASK);
1179 dsw_port_flush_out_buffers(dsw, source_port);
1183 dsw_port_note_op(source_port, events_len);
1185 if (!op_types_known)
1186 for (i = 0; i < events_len; i++) {
1187 switch (events[i].op) {
1188 case RTE_EVENT_OP_RELEASE:
1191 case RTE_EVENT_OP_NEW:
1193 /* Falls through. */
1200 /* Technically, we could allow the non-new events up to the
1201 * first new event in the array into the system, but for
1202 * simplicity reasons, we deny the whole burst if the port is
1203 * above the water mark.
1205 if (unlikely(num_new > 0 && rte_atomic32_read(&dsw->credits_on_loan) >
1206 source_port->new_event_threshold))
1209 enough_credits = dsw_port_acquire_credits(dsw, source_port,
1211 if (unlikely(!enough_credits))
1214 source_port->pending_releases -= num_release;
1216 dsw_port_enqueue_stats(source_port, num_new,
1217 num_non_release-num_new, num_release);
1219 for (i = 0; i < events_len; i++) {
1220 const struct rte_event *event = &events[i];
1222 if (likely(num_release == 0 ||
1223 event->op != RTE_EVENT_OP_RELEASE))
1224 dsw_port_buffer_event(dsw, source_port, event);
1225 dsw_port_queue_enqueue_stats(source_port, event->queue_id);
1228 DSW_LOG_DP_PORT(DEBUG, source_port->id, "%d non-release events "
1229 "accepted.\n", num_non_release);
1231 return num_non_release;
1235 dsw_event_enqueue_burst(void *port, const struct rte_event events[],
1236 uint16_t events_len)
1238 struct dsw_port *source_port = port;
1240 if (unlikely(events_len > source_port->enqueue_depth))
1241 events_len = source_port->enqueue_depth;
1243 return dsw_event_enqueue_burst_generic(source_port, events,
1244 events_len, false, 0, 0, 0);
1248 dsw_event_enqueue_new_burst(void *port, const struct rte_event events[],
1249 uint16_t events_len)
1251 struct dsw_port *source_port = port;
1253 if (unlikely(events_len > source_port->enqueue_depth))
1254 events_len = source_port->enqueue_depth;
1256 return dsw_event_enqueue_burst_generic(source_port, events,
1257 events_len, true, events_len,
1262 dsw_event_enqueue_forward_burst(void *port, const struct rte_event events[],
1263 uint16_t events_len)
1265 struct dsw_port *source_port = port;
1267 if (unlikely(events_len > source_port->enqueue_depth))
1268 events_len = source_port->enqueue_depth;
1270 return dsw_event_enqueue_burst_generic(source_port, events,
1271 events_len, true, 0, 0,
1276 dsw_event_dequeue(void *port, struct rte_event *events, uint64_t wait)
1278 return dsw_event_dequeue_burst(port, events, 1, wait);
1282 dsw_port_record_seen_events(struct dsw_port *port, struct rte_event *events,
1287 dsw_port_dequeue_stats(port, num);
1289 for (i = 0; i < num; i++) {
1290 uint16_t l_idx = port->seen_events_idx;
1291 struct dsw_queue_flow *qf = &port->seen_events[l_idx];
1292 struct rte_event *event = &events[i];
1293 qf->queue_id = event->queue_id;
1294 qf->flow_hash = dsw_flow_id_hash(event->flow_id);
1296 port->seen_events_idx = (l_idx+1) % DSW_MAX_EVENTS_RECORDED;
1298 dsw_port_queue_dequeued_stats(port, event->queue_id);
1301 if (unlikely(port->seen_events_len != DSW_MAX_EVENTS_RECORDED))
1302 port->seen_events_len =
1303 RTE_MIN(port->seen_events_len + num,
1304 DSW_MAX_EVENTS_RECORDED);
1307 #ifdef DSW_SORT_DEQUEUED
1309 #define DSW_EVENT_TO_INT(_event) \
1310 ((int)((((_event)->queue_id)<<16)|((_event)->flow_id)))
1313 dsw_cmp_event(const void *v_event_a, const void *v_event_b)
1315 const struct rte_event *event_a = v_event_a;
1316 const struct rte_event *event_b = v_event_b;
1318 return DSW_EVENT_TO_INT(event_a) - DSW_EVENT_TO_INT(event_b);
1323 dsw_port_dequeue_burst(struct dsw_port *port, struct rte_event *events,
1326 if (unlikely(port->in_buffer_len > 0)) {
1327 uint16_t dequeued = RTE_MIN(num, port->in_buffer_len);
1329 rte_memcpy(events, &port->in_buffer[port->in_buffer_start],
1330 dequeued * sizeof(struct rte_event));
1332 port->in_buffer_start += dequeued;
1333 port->in_buffer_len -= dequeued;
1335 if (port->in_buffer_len == 0)
1336 port->in_buffer_start = 0;
1341 return rte_event_ring_dequeue_burst(port->in_ring, events, num, NULL);
1345 dsw_event_dequeue_burst(void *port, struct rte_event *events, uint16_t num,
1346 uint64_t wait __rte_unused)
1348 struct dsw_port *source_port = port;
1349 struct dsw_evdev *dsw = source_port->dsw;
1352 source_port->pending_releases = 0;
1354 dsw_port_bg_process(dsw, source_port);
1356 if (unlikely(num > source_port->dequeue_depth))
1357 num = source_port->dequeue_depth;
1359 dequeued = dsw_port_dequeue_burst(source_port, events, num);
1361 source_port->pending_releases = dequeued;
1363 dsw_port_load_record(source_port, dequeued);
1365 dsw_port_note_op(source_port, dequeued);
1368 DSW_LOG_DP_PORT(DEBUG, source_port->id, "Dequeued %d events.\n",
1371 dsw_port_return_credits(dsw, source_port, dequeued);
1373 /* One potential optimization one might think of is to
1374 * add a migration state (prior to 'pausing'), and
1375 * only record seen events when the port is in this
1376 * state (and transit to 'pausing' when enough events
1377 * have been gathered). However, that schema doesn't
1378 * seem to improve performance.
1380 dsw_port_record_seen_events(port, events, dequeued);
1381 } else /* Zero-size dequeue means a likely idle port, and thus
1382 * we can afford trading some efficiency for a slightly
1383 * reduced event wall-time latency.
1385 dsw_port_flush_out_buffers(dsw, port);
1387 #ifdef DSW_SORT_DEQUEUED
1388 dsw_stable_sort(events, dequeued, sizeof(events[0]), dsw_cmp_event);