drivers/event/sw/sw_evdev_worker.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2016-2017 Intel Corporation
   3  */
   4
   5 #include <rte_atomic.h>
   6 #include <rte_cycles.h>
   7 #include <rte_event_ring.h>
   8
   9 #include "sw_evdev.h"
  10
  11 #define PORT_ENQUEUE_MAX_BURST_SIZE 64
  12
  13 static inline void
  14 sw_event_release(struct sw_port *p, uint8_t index)
  15 {
  16         /*
  17          * Drops the next outstanding event in our history. Used on dequeue
  18          * to clear any history before dequeuing more events.
  19          */
  20         RTE_SET_USED(index);
  21
  22         /* create drop message */
  23         struct rte_event ev;
  24         ev.op = sw_qe_flag_map[RTE_EVENT_OP_RELEASE];
  25
  26         uint16_t free_count;
  27         rte_event_ring_enqueue_burst(p->rx_worker_ring, &ev, 1, &free_count);
  28
  29         /* each release returns one credit */
  30         p->outstanding_releases--;
  31         p->inflight_credits++;
  32 }
  33
  34 /*
  35  * special-case of rte_event_ring enqueue, with overriding the ops member on
  36  * the events that get written to the ring.
  37  */
  38 static inline unsigned int
  39 enqueue_burst_with_ops(struct rte_event_ring *r, const struct rte_event *events,
  40                 unsigned int n, uint8_t *ops)
  41 {
  42         struct rte_event tmp_evs[PORT_ENQUEUE_MAX_BURST_SIZE];
  43         unsigned int i;
  44
  45         memcpy(tmp_evs, events, n * sizeof(events[0]));
  46         for (i = 0; i < n; i++)
  47                 tmp_evs[i].op = ops[i];
  48
  49         return rte_event_ring_enqueue_burst(r, tmp_evs, n, NULL);
  50 }
  51
  52 uint16_t
  53 sw_event_enqueue_burst(void *port, const struct rte_event ev[], uint16_t num)
  54 {
  55         int32_t i;
  56         uint8_t new_ops[PORT_ENQUEUE_MAX_BURST_SIZE];
  57         struct sw_port *p = port;
  58         struct sw_evdev *sw = (void *)p->sw;
  59         uint32_t sw_inflights = rte_atomic32_read(&sw->inflights);
  60         uint32_t credit_update_quanta = sw->credit_update_quanta;
  61         int new = 0;
  62
  63         if (num > PORT_ENQUEUE_MAX_BURST_SIZE)
  64                 num = PORT_ENQUEUE_MAX_BURST_SIZE;
  65
  66         for (i = 0; i < num; i++)
  67                 new += (ev[i].op == RTE_EVENT_OP_NEW);
  68
  69         if (unlikely(new > 0 && p->inflight_max < sw_inflights))
  70                 return 0;
  71
  72         if (p->inflight_credits < new) {
  73                 /* check if event enqueue brings port over max threshold */
  74                 if (sw_inflights + credit_update_quanta > sw->nb_events_limit)
  75                         return 0;
  76
  77                 rte_atomic32_add(&sw->inflights, credit_update_quanta);
  78                 p->inflight_credits += (credit_update_quanta);
  79
  80                 /* If there are fewer inflight credits than new events, limit
  81                  * the number of enqueued events.
  82                  */
  83                 num = (p->inflight_credits < new) ? p->inflight_credits : new;
  84         }
  85
  86         for (i = 0; i < num; i++) {
  87                 int op = ev[i].op;
  88                 int outstanding = p->outstanding_releases > 0;
  89                 const uint8_t invalid_qid = (ev[i].queue_id >= sw->qid_count);
  90
  91                 p->inflight_credits -= (op == RTE_EVENT_OP_NEW);
  92                 p->inflight_credits += (op == RTE_EVENT_OP_RELEASE) *
  93                                         outstanding;
  94
  95                 new_ops[i] = sw_qe_flag_map[op];
  96                 new_ops[i] &= ~(invalid_qid << QE_FLAG_VALID_SHIFT);
  97
  98                 /* FWD and RELEASE packets will both resolve to taken (assuming
  99                  * correct usage of the API), providing very high correct
 100                  * prediction rate.
 101                  */
 102                 if ((new_ops[i] & QE_FLAG_COMPLETE) && outstanding)
 103                         p->outstanding_releases--;
 104
 105                 /* error case: branch to avoid touching p->stats */
 106                 if (unlikely(invalid_qid && op != RTE_EVENT_OP_RELEASE)) {
 107                         p->stats.rx_dropped++;
 108                         p->inflight_credits++;
 109                 }
 110         }
 111
 112         /* returns number of events actually enqueued */
 113         uint32_t enq = enqueue_burst_with_ops(p->rx_worker_ring, ev, i,
 114                                              new_ops);
 115         if (p->outstanding_releases == 0 && p->last_dequeue_burst_sz != 0) {
 116                 uint64_t burst_ticks = rte_get_timer_cycles() -
 117                                 p->last_dequeue_ticks;
 118                 uint64_t burst_pkt_ticks =
 119                         burst_ticks / p->last_dequeue_burst_sz;
 120                 p->avg_pkt_ticks -= p->avg_pkt_ticks / NUM_SAMPLES;
 121                 p->avg_pkt_ticks += burst_pkt_ticks / NUM_SAMPLES;
 122                 p->last_dequeue_ticks = 0;
 123         }
 124
 125         /* Replenish credits if enough releases are performed */
 126         if (p->inflight_credits >= credit_update_quanta * 2) {
 127                 rte_atomic32_sub(&sw->inflights, credit_update_quanta);
 128                 p->inflight_credits -= credit_update_quanta;
 129         }
 130
 131         return enq;
 132 }
 133
 134 uint16_t
 135 sw_event_enqueue(void *port, const struct rte_event *ev)
 136 {
 137         return sw_event_enqueue_burst(port, ev, 1);
 138 }
 139
 140 uint16_t
 141 sw_event_dequeue_burst(void *port, struct rte_event *ev, uint16_t num,
 142                 uint64_t wait)
 143 {
 144         RTE_SET_USED(wait);
 145         struct sw_port *p = (void *)port;
 146         struct rte_event_ring *ring = p->cq_worker_ring;
 147
 148         /* check that all previous dequeues have been released */
 149         if (p->implicit_release) {
 150                 struct sw_evdev *sw = (void *)p->sw;
 151                 uint32_t credit_update_quanta = sw->credit_update_quanta;
 152                 uint16_t out_rels = p->outstanding_releases;
 153                 uint16_t i;
 154                 for (i = 0; i < out_rels; i++)
 155                         sw_event_release(p, i);
 156
 157                 /* Replenish credits if enough releases are performed */
 158                 if (p->inflight_credits >= credit_update_quanta * 2) {
 159                         rte_atomic32_sub(&sw->inflights, credit_update_quanta);
 160                         p->inflight_credits -= credit_update_quanta;
 161                 }
 162         }
 163
 164         /* returns number of events actually dequeued */
 165         uint16_t ndeq = rte_event_ring_dequeue_burst(ring, ev, num, NULL);
 166         if (unlikely(ndeq == 0)) {
 167                 p->zero_polls++;
 168                 p->total_polls++;
 169                 goto end;
 170         }
 171
 172         p->outstanding_releases += ndeq;
 173         p->last_dequeue_burst_sz = ndeq;
 174         p->last_dequeue_ticks = rte_get_timer_cycles();
 175         p->poll_buckets[(ndeq - 1) >> SW_DEQ_STAT_BUCKET_SHIFT]++;
 176         p->total_polls++;
 177
 178 end:
 179         return ndeq;
 180 }
 181
 182 uint16_t
 183 sw_event_dequeue(void *port, struct rte_event *ev, uint64_t wait)
 184 {
 185         return sw_event_dequeue_burst(port, ev, 1, wait);
 186 }