numam-dpdk/drivers/event/sw/sw_evdev_worker.c
Gage Eads 4deeb214ac event/sw: perform partial burst enqueues
Previously, the sw PMD would enqueue either all or no events, depending on
if enough inflight credits were available for the new events in the burst.
If a port is enqueueing a large burst (i.e. a multiple of the credit update
quanta), this can result in suboptimal performance, and requires an
understanding of the sw PMD implementation (in particular, its credit
scheme) to tune an application's burst size.

This affects software that enqueues large bursts of new events, such as the
ethernet event adapter which uses a 128-deep event buffer, when the input
packet rate is sufficiently high.

This change makes the sw PMD enqueue as many events as it has credits, if
there are any new events in the burst.

Signed-off-by: Gage Eads <gage.eads@intel.com>
Acked-by: Harry van Haaren <harry.van.haaren@intel.com>
2018-04-16 10:09:53 +02:00

187 lines
5.1 KiB
C

/* SPDX-License-Identifier: BSD-3-Clause
* Copyright(c) 2016-2017 Intel Corporation
*/
#include <rte_atomic.h>
#include <rte_cycles.h>
#include <rte_event_ring.h>
#include "sw_evdev.h"
#define PORT_ENQUEUE_MAX_BURST_SIZE 64
static inline void
sw_event_release(struct sw_port *p, uint8_t index)
{
/*
* Drops the next outstanding event in our history. Used on dequeue
* to clear any history before dequeuing more events.
*/
RTE_SET_USED(index);
/* create drop message */
struct rte_event ev;
ev.op = sw_qe_flag_map[RTE_EVENT_OP_RELEASE];
uint16_t free_count;
rte_event_ring_enqueue_burst(p->rx_worker_ring, &ev, 1, &free_count);
/* each release returns one credit */
p->outstanding_releases--;
p->inflight_credits++;
}
/*
* special-case of rte_event_ring enqueue, with overriding the ops member on
* the events that get written to the ring.
*/
static inline unsigned int
enqueue_burst_with_ops(struct rte_event_ring *r, const struct rte_event *events,
unsigned int n, uint8_t *ops)
{
struct rte_event tmp_evs[PORT_ENQUEUE_MAX_BURST_SIZE];
unsigned int i;
memcpy(tmp_evs, events, n * sizeof(events[0]));
for (i = 0; i < n; i++)
tmp_evs[i].op = ops[i];
return rte_event_ring_enqueue_burst(r, tmp_evs, n, NULL);
}
uint16_t
sw_event_enqueue_burst(void *port, const struct rte_event ev[], uint16_t num)
{
int32_t i;
uint8_t new_ops[PORT_ENQUEUE_MAX_BURST_SIZE];
struct sw_port *p = port;
struct sw_evdev *sw = (void *)p->sw;
uint32_t sw_inflights = rte_atomic32_read(&sw->inflights);
uint32_t credit_update_quanta = sw->credit_update_quanta;
int new = 0;
if (num > PORT_ENQUEUE_MAX_BURST_SIZE)
num = PORT_ENQUEUE_MAX_BURST_SIZE;
for (i = 0; i < num; i++)
new += (ev[i].op == RTE_EVENT_OP_NEW);
if (unlikely(new > 0 && p->inflight_max < sw_inflights))
return 0;
if (p->inflight_credits < new) {
/* check if event enqueue brings port over max threshold */
if (sw_inflights + credit_update_quanta > sw->nb_events_limit)
return 0;
rte_atomic32_add(&sw->inflights, credit_update_quanta);
p->inflight_credits += (credit_update_quanta);
/* If there are fewer inflight credits than new events, limit
* the number of enqueued events.
*/
num = (p->inflight_credits < new) ? p->inflight_credits : new;
}
for (i = 0; i < num; i++) {
int op = ev[i].op;
int outstanding = p->outstanding_releases > 0;
const uint8_t invalid_qid = (ev[i].queue_id >= sw->qid_count);
p->inflight_credits -= (op == RTE_EVENT_OP_NEW);
p->inflight_credits += (op == RTE_EVENT_OP_RELEASE) *
outstanding;
new_ops[i] = sw_qe_flag_map[op];
new_ops[i] &= ~(invalid_qid << QE_FLAG_VALID_SHIFT);
/* FWD and RELEASE packets will both resolve to taken (assuming
* correct usage of the API), providing very high correct
* prediction rate.
*/
if ((new_ops[i] & QE_FLAG_COMPLETE) && outstanding)
p->outstanding_releases--;
/* error case: branch to avoid touching p->stats */
if (unlikely(invalid_qid && op != RTE_EVENT_OP_RELEASE)) {
p->stats.rx_dropped++;
p->inflight_credits++;
}
}
/* returns number of events actually enqueued */
uint32_t enq = enqueue_burst_with_ops(p->rx_worker_ring, ev, i,
new_ops);
if (p->outstanding_releases == 0 && p->last_dequeue_burst_sz != 0) {
uint64_t burst_ticks = rte_get_timer_cycles() -
p->last_dequeue_ticks;
uint64_t burst_pkt_ticks =
burst_ticks / p->last_dequeue_burst_sz;
p->avg_pkt_ticks -= p->avg_pkt_ticks / NUM_SAMPLES;
p->avg_pkt_ticks += burst_pkt_ticks / NUM_SAMPLES;
p->last_dequeue_ticks = 0;
}
/* Replenish credits if enough releases are performed */
if (p->inflight_credits >= credit_update_quanta * 2) {
rte_atomic32_sub(&sw->inflights, credit_update_quanta);
p->inflight_credits -= credit_update_quanta;
}
return enq;
}
uint16_t
sw_event_enqueue(void *port, const struct rte_event *ev)
{
return sw_event_enqueue_burst(port, ev, 1);
}
uint16_t
sw_event_dequeue_burst(void *port, struct rte_event *ev, uint16_t num,
uint64_t wait)
{
RTE_SET_USED(wait);
struct sw_port *p = (void *)port;
struct rte_event_ring *ring = p->cq_worker_ring;
/* check that all previous dequeues have been released */
if (p->implicit_release) {
struct sw_evdev *sw = (void *)p->sw;
uint32_t credit_update_quanta = sw->credit_update_quanta;
uint16_t out_rels = p->outstanding_releases;
uint16_t i;
for (i = 0; i < out_rels; i++)
sw_event_release(p, i);
/* Replenish credits if enough releases are performed */
if (p->inflight_credits >= credit_update_quanta * 2) {
rte_atomic32_sub(&sw->inflights, credit_update_quanta);
p->inflight_credits -= credit_update_quanta;
}
}
/* returns number of events actually dequeued */
uint16_t ndeq = rte_event_ring_dequeue_burst(ring, ev, num, NULL);
if (unlikely(ndeq == 0)) {
p->zero_polls++;
p->total_polls++;
goto end;
}
p->outstanding_releases += ndeq;
p->last_dequeue_burst_sz = ndeq;
p->last_dequeue_ticks = rte_get_timer_cycles();
p->poll_buckets[(ndeq - 1) >> SW_DEQ_STAT_BUCKET_SHIFT]++;
p->total_polls++;
end:
return ndeq;
}
uint16_t
sw_event_dequeue(void *port, struct rte_event *ev, uint64_t wait)
{
return sw_event_dequeue_burst(port, ev, 1, wait);
}