xref: /dpdk/drivers/event/sw/sw_evdev_worker.c (revision 88ca872150d0b61b4e6ffcb96f5cecc9e781adb5)
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_dequeue_burst(void *port, struct rte_event *ev, uint16_t num,
136 		uint64_t wait)
137 {
138 	RTE_SET_USED(wait);
139 	struct sw_port *p = (void *)port;
140 	struct rte_event_ring *ring = p->cq_worker_ring;
141 
142 	/* check that all previous dequeues have been released */
143 	if (p->implicit_release) {
144 		struct sw_evdev *sw = (void *)p->sw;
145 		uint32_t credit_update_quanta = sw->credit_update_quanta;
146 		uint16_t out_rels = p->outstanding_releases;
147 		uint16_t i;
148 		for (i = 0; i < out_rels; i++)
149 			sw_event_release(p, i);
150 
151 		/* Replenish credits if enough releases are performed */
152 		if (p->inflight_credits >= credit_update_quanta * 2) {
153 			rte_atomic32_sub(&sw->inflights, credit_update_quanta);
154 			p->inflight_credits -= credit_update_quanta;
155 		}
156 	}
157 
158 	/* returns number of events actually dequeued */
159 	uint16_t ndeq = rte_event_ring_dequeue_burst(ring, ev, num, NULL);
160 	if (unlikely(ndeq == 0)) {
161 		p->zero_polls++;
162 		p->total_polls++;
163 		goto end;
164 	}
165 
166 	p->outstanding_releases += ndeq;
167 	p->last_dequeue_burst_sz = ndeq;
168 	p->last_dequeue_ticks = rte_get_timer_cycles();
169 	p->poll_buckets[(ndeq - 1) >> SW_DEQ_STAT_BUCKET_SHIFT]++;
170 	p->total_polls++;
171 
172 end:
173 	return ndeq;
174 }
175