1
2
3
4
5#include "test_perf_common.h"
6
7
8
9static inline int
10perf_queue_nb_event_queues(struct evt_options *opt)
11{
12
13 uint8_t nb_prod = opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR ?
14 rte_eth_dev_count_avail() : evt_nr_active_lcores(opt->plcores);
15 return nb_prod * opt->nb_stages;
16}
17
18static __rte_always_inline void
19mark_fwd_latency(struct rte_event *const ev,
20 const uint8_t nb_stages)
21{
22 if (unlikely((ev->queue_id % nb_stages) == 0)) {
23 struct perf_elt *const m = ev->event_ptr;
24
25 m->timestamp = rte_get_timer_cycles();
26 }
27}
28
29static __rte_always_inline void
30fwd_event(struct rte_event *const ev, uint8_t *const sched_type_list,
31 const uint8_t nb_stages)
32{
33 ev->queue_id++;
34 ev->sched_type = sched_type_list[ev->queue_id % nb_stages];
35 ev->op = RTE_EVENT_OP_FORWARD;
36 ev->event_type = RTE_EVENT_TYPE_CPU;
37}
38
39static int
40perf_queue_worker(void *arg, const int enable_fwd_latency)
41{
42 PERF_WORKER_INIT;
43 struct rte_event ev;
44
45 while (t->done == false) {
46 uint16_t event = rte_event_dequeue_burst(dev, port, &ev, 1, 0);
47
48 if (!event) {
49 rte_pause();
50 continue;
51 }
52 if (enable_fwd_latency && !prod_timer_type)
53
54 mark_fwd_latency(&ev, nb_stages);
55
56
57 if (unlikely((ev.queue_id % nb_stages) == laststage)) {
58 if (enable_fwd_latency)
59 cnt = perf_process_last_stage_latency(pool,
60 &ev, w, bufs, sz, cnt);
61 else
62 cnt = perf_process_last_stage(pool,
63 &ev, w, bufs, sz, cnt);
64 } else {
65 fwd_event(&ev, sched_type_list, nb_stages);
66 while (rte_event_enqueue_burst(dev, port, &ev, 1) != 1)
67 rte_pause();
68 }
69 }
70 return 0;
71}
72
73static int
74perf_queue_worker_burst(void *arg, const int enable_fwd_latency)
75{
76 PERF_WORKER_INIT;
77 uint16_t i;
78
79 struct rte_event ev[BURST_SIZE + 1];
80
81 while (t->done == false) {
82 uint16_t const nb_rx = rte_event_dequeue_burst(dev, port, ev,
83 BURST_SIZE, 0);
84
85 if (!nb_rx) {
86 rte_pause();
87 continue;
88 }
89
90 for (i = 0; i < nb_rx; i++) {
91 if (enable_fwd_latency && !prod_timer_type) {
92 rte_prefetch0(ev[i+1].event_ptr);
93
94
95
96 mark_fwd_latency(&ev[i], nb_stages);
97 }
98
99 if (unlikely((ev[i].queue_id % nb_stages) ==
100 laststage)) {
101 if (enable_fwd_latency)
102 cnt = perf_process_last_stage_latency(
103 pool, &ev[i], w, bufs, sz, cnt);
104 else
105 cnt = perf_process_last_stage(pool,
106 &ev[i], w, bufs, sz, cnt);
107
108 ev[i].op = RTE_EVENT_OP_RELEASE;
109 } else {
110 fwd_event(&ev[i], sched_type_list, nb_stages);
111 }
112 }
113
114 uint16_t enq;
115
116 enq = rte_event_enqueue_burst(dev, port, ev, nb_rx);
117 while (enq < nb_rx) {
118 enq += rte_event_enqueue_burst(dev, port,
119 ev + enq, nb_rx - enq);
120 }
121 }
122 return 0;
123}
124
125static int
126worker_wrapper(void *arg)
127{
128 struct worker_data *w = arg;
129 struct evt_options *opt = w->t->opt;
130
131 const bool burst = evt_has_burst_mode(w->dev_id);
132 const int fwd_latency = opt->fwd_latency;
133
134
135 if (!burst && !fwd_latency)
136 return perf_queue_worker(arg, 0);
137 else if (!burst && fwd_latency)
138 return perf_queue_worker(arg, 1);
139 else if (burst && !fwd_latency)
140 return perf_queue_worker_burst(arg, 0);
141 else if (burst && fwd_latency)
142 return perf_queue_worker_burst(arg, 1);
143
144 rte_panic("invalid worker\n");
145}
146
147static int
148perf_queue_launch_lcores(struct evt_test *test, struct evt_options *opt)
149{
150 return perf_launch_lcores(test, opt, worker_wrapper);
151}
152
153static int
154perf_queue_eventdev_setup(struct evt_test *test, struct evt_options *opt)
155{
156 uint8_t queue;
157 int nb_stages = opt->nb_stages;
158 int ret;
159 int nb_ports;
160 int nb_queues;
161 uint16_t prod;
162 struct rte_event_dev_info dev_info;
163 struct test_perf *t = evt_test_priv(test);
164
165 nb_ports = evt_nr_active_lcores(opt->wlcores);
166 nb_ports += opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR ||
167 opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR ? 0 :
168 evt_nr_active_lcores(opt->plcores);
169
170 nb_queues = perf_queue_nb_event_queues(opt);
171
172 memset(&dev_info, 0, sizeof(struct rte_event_dev_info));
173 ret = rte_event_dev_info_get(opt->dev_id, &dev_info);
174 if (ret) {
175 evt_err("failed to get eventdev info %d", opt->dev_id);
176 return ret;
177 }
178
179 ret = evt_configure_eventdev(opt, nb_queues, nb_ports);
180 if (ret) {
181 evt_err("failed to configure eventdev %d", opt->dev_id);
182 return ret;
183 }
184
185 struct rte_event_queue_conf q_conf = {
186 .priority = RTE_EVENT_DEV_PRIORITY_NORMAL,
187 .nb_atomic_flows = opt->nb_flows,
188 .nb_atomic_order_sequences = opt->nb_flows,
189 };
190
191 for (queue = 0; queue < nb_queues; queue++) {
192 q_conf.schedule_type =
193 (opt->sched_type_list[queue % nb_stages]);
194
195 if (opt->q_priority) {
196 uint8_t stage_pos = queue % nb_stages;
197
198
199
200
201 uint8_t step = RTE_EVENT_DEV_PRIORITY_LOWEST /
202 (nb_stages - 1);
203
204 q_conf.priority = RTE_EVENT_DEV_PRIORITY_LOWEST -
205 (step * stage_pos);
206 }
207 ret = rte_event_queue_setup(opt->dev_id, queue, &q_conf);
208 if (ret) {
209 evt_err("failed to setup queue=%d", queue);
210 return ret;
211 }
212 }
213
214 if (opt->wkr_deq_dep > dev_info.max_event_port_dequeue_depth)
215 opt->wkr_deq_dep = dev_info.max_event_port_dequeue_depth;
216
217
218 const struct rte_event_port_conf p_conf = {
219 .dequeue_depth = opt->wkr_deq_dep,
220 .enqueue_depth = dev_info.max_event_port_dequeue_depth,
221 .new_event_threshold = dev_info.max_num_events,
222 };
223
224 ret = perf_event_dev_port_setup(test, opt, nb_stages ,
225 nb_queues, &p_conf);
226 if (ret)
227 return ret;
228
229 if (!evt_has_distributed_sched(opt->dev_id)) {
230 uint32_t service_id;
231 rte_event_dev_service_id_get(opt->dev_id, &service_id);
232 ret = evt_service_setup(service_id);
233 if (ret) {
234 evt_err("No service lcore found to run event dev.");
235 return ret;
236 }
237 }
238
239 ret = rte_event_dev_start(opt->dev_id);
240 if (ret) {
241 evt_err("failed to start eventdev %d", opt->dev_id);
242 return ret;
243 }
244
245 if (opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR) {
246 RTE_ETH_FOREACH_DEV(prod) {
247 ret = rte_eth_dev_start(prod);
248 if (ret) {
249 evt_err("Ethernet dev [%d] failed to start. Using synthetic producer",
250 prod);
251 return ret;
252 }
253
254 ret = rte_event_eth_rx_adapter_start(prod);
255 if (ret) {
256 evt_err("Rx adapter[%d] start failed", prod);
257 return ret;
258 }
259 printf("%s: Port[%d] using Rx adapter[%d] started\n",
260 __func__, prod, prod);
261 }
262 } else if (opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
263 for (prod = 0; prod < opt->nb_timer_adptrs; prod++) {
264 ret = rte_event_timer_adapter_start(
265 t->timer_adptr[prod]);
266 if (ret) {
267 evt_err("failed to Start event timer adapter %d"
268 , prod);
269 return ret;
270 }
271 }
272 }
273
274 return 0;
275}
276
277static void
278perf_queue_opt_dump(struct evt_options *opt)
279{
280 evt_dump_fwd_latency(opt);
281 perf_opt_dump(opt, perf_queue_nb_event_queues(opt));
282}
283
284static int
285perf_queue_opt_check(struct evt_options *opt)
286{
287 return perf_opt_check(opt, perf_queue_nb_event_queues(opt));
288}
289
290static bool
291perf_queue_capability_check(struct evt_options *opt)
292{
293 struct rte_event_dev_info dev_info;
294
295 rte_event_dev_info_get(opt->dev_id, &dev_info);
296 if (dev_info.max_event_queues < perf_queue_nb_event_queues(opt) ||
297 dev_info.max_event_ports < perf_nb_event_ports(opt)) {
298 evt_err("not enough eventdev queues=%d/%d or ports=%d/%d",
299 perf_queue_nb_event_queues(opt),
300 dev_info.max_event_queues,
301 perf_nb_event_ports(opt), dev_info.max_event_ports);
302 }
303
304 return true;
305}
306
307static const struct evt_test_ops perf_queue = {
308 .cap_check = perf_queue_capability_check,
309 .opt_check = perf_queue_opt_check,
310 .opt_dump = perf_queue_opt_dump,
311 .test_setup = perf_test_setup,
312 .mempool_setup = perf_mempool_setup,
313 .ethdev_setup = perf_ethdev_setup,
314 .eventdev_setup = perf_queue_eventdev_setup,
315 .launch_lcores = perf_queue_launch_lcores,
316 .eventdev_destroy = perf_eventdev_destroy,
317 .mempool_destroy = perf_mempool_destroy,
318 .ethdev_destroy = perf_ethdev_destroy,
319 .test_result = perf_test_result,
320 .test_destroy = perf_test_destroy,
321};
322
323EVT_TEST_REGISTER(perf_queue);
324
