examples/eventdev: probe max events
[dpdk.git] / examples / eventdev_pipeline / pipeline_worker_generic.c
1 /*
2  * SPDX-License-Identifier: BSD-3-Clause
3  * Copyright 2016 Intel Corporation.
4  * Copyright 2017 Cavium, Inc.
5  */
6
7 #include "pipeline_common.h"
8
9 static __rte_always_inline int
10 worker_generic(void *arg)
11 {
12         struct rte_event ev;
13
14         struct worker_data *data = (struct worker_data *)arg;
15         uint8_t dev_id = data->dev_id;
16         uint8_t port_id = data->port_id;
17         size_t sent = 0, received = 0;
18         unsigned int lcore_id = rte_lcore_id();
19
20         while (!fdata->done) {
21
22                 if (fdata->cap.scheduler)
23                         fdata->cap.scheduler(lcore_id);
24
25                 if (!fdata->worker_core[lcore_id]) {
26                         rte_pause();
27                         continue;
28                 }
29
30                 const uint16_t nb_rx = rte_event_dequeue_burst(dev_id, port_id,
31                                 &ev, 1, 0);
32
33                 if (nb_rx == 0) {
34                         rte_pause();
35                         continue;
36                 }
37                 received++;
38
39                 /* The first worker stage does classification */
40                 if (ev.queue_id == cdata.qid[0])
41                         ev.flow_id = ev.mbuf->hash.rss
42                                                 % cdata.num_fids;
43
44                 ev.queue_id = cdata.next_qid[ev.queue_id];
45                 ev.op = RTE_EVENT_OP_FORWARD;
46                 ev.sched_type = cdata.queue_type;
47
48                 work();
49
50                 while (rte_event_enqueue_burst(dev_id, port_id, &ev, 1) != 1)
51                         rte_pause();
52                 sent++;
53         }
54
55         if (!cdata.quiet)
56                 printf("  worker %u thread done. RX=%zu TX=%zu\n",
57                                 rte_lcore_id(), received, sent);
58
59         return 0;
60 }
61
62 static int
63 worker_generic_burst(void *arg)
64 {
65         struct rte_event events[BATCH_SIZE];
66
67         struct worker_data *data = (struct worker_data *)arg;
68         uint8_t dev_id = data->dev_id;
69         uint8_t port_id = data->port_id;
70         size_t sent = 0, received = 0;
71         unsigned int lcore_id = rte_lcore_id();
72
73         while (!fdata->done) {
74                 uint16_t i;
75
76                 if (fdata->cap.scheduler)
77                         fdata->cap.scheduler(lcore_id);
78
79                 if (!fdata->worker_core[lcore_id]) {
80                         rte_pause();
81                         continue;
82                 }
83
84                 const uint16_t nb_rx = rte_event_dequeue_burst(dev_id, port_id,
85                                 events, RTE_DIM(events), 0);
86
87                 if (nb_rx == 0) {
88                         rte_pause();
89                         continue;
90                 }
91                 received += nb_rx;
92
93                 for (i = 0; i < nb_rx; i++) {
94
95                         /* The first worker stage does classification */
96                         if (events[i].queue_id == cdata.qid[0])
97                                 events[i].flow_id = events[i].mbuf->hash.rss
98                                                         % cdata.num_fids;
99
100                         events[i].queue_id = cdata.next_qid[events[i].queue_id];
101                         events[i].op = RTE_EVENT_OP_FORWARD;
102                         events[i].sched_type = cdata.queue_type;
103
104                         work();
105                 }
106                 uint16_t nb_tx = rte_event_enqueue_burst(dev_id, port_id,
107                                 events, nb_rx);
108                 while (nb_tx < nb_rx && !fdata->done)
109                         nb_tx += rte_event_enqueue_burst(dev_id, port_id,
110                                                         events + nb_tx,
111                                                         nb_rx - nb_tx);
112                 sent += nb_tx;
113         }
114
115         if (!cdata.quiet)
116                 printf("  worker %u thread done. RX=%zu TX=%zu\n",
117                                 rte_lcore_id(), received, sent);
118
119         return 0;
120 }
121
122 static int
123 setup_eventdev_generic(struct worker_data *worker_data)
124 {
125         const uint8_t dev_id = 0;
126         /* +1 stages is for a SINGLE_LINK TX stage */
127         const uint8_t nb_queues = cdata.num_stages + 1;
128         const uint8_t nb_ports = cdata.num_workers;
129         struct rte_event_dev_config config = {
130                         .nb_event_queues = nb_queues,
131                         .nb_event_ports = nb_ports,
132                         .nb_events_limit  = 4096,
133                         .nb_event_queue_flows = 1024,
134                         .nb_event_port_dequeue_depth = 128,
135                         .nb_event_port_enqueue_depth = 128,
136         };
137         struct rte_event_port_conf wkr_p_conf = {
138                         .dequeue_depth = cdata.worker_cq_depth,
139                         .enqueue_depth = 64,
140                         .new_event_threshold = 4096,
141         };
142         struct rte_event_queue_conf wkr_q_conf = {
143                         .schedule_type = cdata.queue_type,
144                         .priority = RTE_EVENT_DEV_PRIORITY_NORMAL,
145                         .nb_atomic_flows = 1024,
146                 .nb_atomic_order_sequences = 1024,
147         };
148         struct rte_event_queue_conf tx_q_conf = {
149                         .priority = RTE_EVENT_DEV_PRIORITY_HIGHEST,
150                         .event_queue_cfg = RTE_EVENT_QUEUE_CFG_SINGLE_LINK,
151         };
152
153         struct port_link worker_queues[MAX_NUM_STAGES];
154         uint8_t disable_implicit_release;
155         unsigned int i;
156
157         int ret, ndev = rte_event_dev_count();
158         if (ndev < 1) {
159                 printf("%d: No Eventdev Devices Found\n", __LINE__);
160                 return -1;
161         }
162
163         struct rte_event_dev_info dev_info;
164         ret = rte_event_dev_info_get(dev_id, &dev_info);
165         printf("\tEventdev %d: %s\n", dev_id, dev_info.driver_name);
166
167         disable_implicit_release = (dev_info.event_dev_cap &
168                         RTE_EVENT_DEV_CAP_IMPLICIT_RELEASE_DISABLE);
169
170         wkr_p_conf.disable_implicit_release = disable_implicit_release;
171
172         if (dev_info.max_num_events < config.nb_events_limit)
173                 config.nb_events_limit = dev_info.max_num_events;
174         if (dev_info.max_event_port_dequeue_depth <
175                         config.nb_event_port_dequeue_depth)
176                 config.nb_event_port_dequeue_depth =
177                                 dev_info.max_event_port_dequeue_depth;
178         if (dev_info.max_event_port_enqueue_depth <
179                         config.nb_event_port_enqueue_depth)
180                 config.nb_event_port_enqueue_depth =
181                                 dev_info.max_event_port_enqueue_depth;
182
183         ret = rte_event_dev_configure(dev_id, &config);
184         if (ret < 0) {
185                 printf("%d: Error configuring device\n", __LINE__);
186                 return -1;
187         }
188
189         /* Q creation - one load balanced per pipeline stage*/
190         printf("  Stages:\n");
191         for (i = 0; i < cdata.num_stages; i++) {
192                 if (rte_event_queue_setup(dev_id, i, &wkr_q_conf) < 0) {
193                         printf("%d: error creating qid %d\n", __LINE__, i);
194                         return -1;
195                 }
196                 cdata.qid[i] = i;
197                 cdata.next_qid[i] = i+1;
198                 worker_queues[i].queue_id = i;
199                 if (cdata.enable_queue_priorities) {
200                         /* calculate priority stepping for each stage, leaving
201                          * headroom of 1 for the SINGLE_LINK TX below
202                          */
203                         const uint32_t prio_delta =
204                                 (RTE_EVENT_DEV_PRIORITY_LOWEST-1) /  nb_queues;
205
206                         /* higher priority for queues closer to tx */
207                         wkr_q_conf.priority =
208                                 RTE_EVENT_DEV_PRIORITY_LOWEST - prio_delta * i;
209                 }
210
211                 const char *type_str = "Atomic";
212                 switch (wkr_q_conf.schedule_type) {
213                 case RTE_SCHED_TYPE_ORDERED:
214                         type_str = "Ordered";
215                         break;
216                 case RTE_SCHED_TYPE_PARALLEL:
217                         type_str = "Parallel";
218                         break;
219                 }
220                 printf("\tStage %d, Type %s\tPriority = %d\n", i, type_str,
221                                 wkr_q_conf.priority);
222         }
223         printf("\n");
224
225         /* final queue for sending to TX core */
226         if (rte_event_queue_setup(dev_id, i, &tx_q_conf) < 0) {
227                 printf("%d: error creating qid %d\n", __LINE__, i);
228                 return -1;
229         }
230         cdata.tx_queue_id = i;
231
232         if (wkr_p_conf.new_event_threshold > config.nb_events_limit)
233                 wkr_p_conf.new_event_threshold = config.nb_events_limit;
234         if (wkr_p_conf.dequeue_depth > config.nb_event_port_dequeue_depth)
235                 wkr_p_conf.dequeue_depth = config.nb_event_port_dequeue_depth;
236         if (wkr_p_conf.enqueue_depth > config.nb_event_port_enqueue_depth)
237                 wkr_p_conf.enqueue_depth = config.nb_event_port_enqueue_depth;
238
239         /* set up one port per worker, linking to all stage queues */
240         for (i = 0; i < cdata.num_workers; i++) {
241                 struct worker_data *w = &worker_data[i];
242                 w->dev_id = dev_id;
243                 if (rte_event_port_setup(dev_id, i, &wkr_p_conf) < 0) {
244                         printf("Error setting up port %d\n", i);
245                         return -1;
246                 }
247
248                 uint32_t s;
249                 for (s = 0; s < cdata.num_stages; s++) {
250                         if (rte_event_port_link(dev_id, i,
251                                                 &worker_queues[s].queue_id,
252                                                 &worker_queues[s].priority,
253                                                 1) != 1) {
254                                 printf("%d: error creating link for port %d\n",
255                                                 __LINE__, i);
256                                 return -1;
257                         }
258                 }
259                 w->port_id = i;
260         }
261
262         ret = rte_event_dev_service_id_get(dev_id,
263                                 &fdata->evdev_service_id);
264         if (ret != -ESRCH && ret != 0) {
265                 printf("Error getting the service ID for sw eventdev\n");
266                 return -1;
267         }
268         rte_service_runstate_set(fdata->evdev_service_id, 1);
269         rte_service_set_runstate_mapped_check(fdata->evdev_service_id, 0);
270
271         return dev_id;
272 }
273
274 static void
275 init_adapters(uint16_t nb_ports)
276 {
277         int i;
278         int ret;
279         uint8_t tx_port_id = 0;
280         uint8_t evdev_id = 0;
281         struct rte_event_dev_info dev_info;
282
283         ret = rte_event_dev_info_get(evdev_id, &dev_info);
284
285         struct rte_event_port_conf adptr_p_conf = {
286                 .dequeue_depth = cdata.worker_cq_depth,
287                 .enqueue_depth = 64,
288                 .new_event_threshold = 4096,
289         };
290
291         if (adptr_p_conf.new_event_threshold > dev_info.max_num_events)
292                 adptr_p_conf.new_event_threshold = dev_info.max_num_events;
293         if (adptr_p_conf.dequeue_depth > dev_info.max_event_port_dequeue_depth)
294                 adptr_p_conf.dequeue_depth =
295                         dev_info.max_event_port_dequeue_depth;
296         if (adptr_p_conf.enqueue_depth > dev_info.max_event_port_enqueue_depth)
297                 adptr_p_conf.enqueue_depth =
298                         dev_info.max_event_port_enqueue_depth;
299
300         /* Create one adapter for all the ethernet ports. */
301         ret = rte_event_eth_rx_adapter_create(cdata.rx_adapter_id, evdev_id,
302                         &adptr_p_conf);
303         if (ret)
304                 rte_exit(EXIT_FAILURE, "failed to create rx adapter[%d]",
305                                 cdata.rx_adapter_id);
306
307         ret = rte_event_eth_tx_adapter_create(cdata.tx_adapter_id, evdev_id,
308                         &adptr_p_conf);
309         if (ret)
310                 rte_exit(EXIT_FAILURE, "failed to create tx adapter[%d]",
311                                 cdata.tx_adapter_id);
312
313         struct rte_event_eth_rx_adapter_queue_conf queue_conf;
314         memset(&queue_conf, 0, sizeof(queue_conf));
315         queue_conf.ev.sched_type = cdata.queue_type;
316         queue_conf.ev.queue_id = cdata.qid[0];
317
318         for (i = 0; i < nb_ports; i++) {
319                 ret = rte_event_eth_rx_adapter_queue_add(cdata.rx_adapter_id, i,
320                                 -1, &queue_conf);
321                 if (ret)
322                         rte_exit(EXIT_FAILURE,
323                                         "Failed to add queues to Rx adapter");
324
325                 ret = rte_event_eth_tx_adapter_queue_add(cdata.tx_adapter_id, i,
326                                 -1);
327                 if (ret)
328                         rte_exit(EXIT_FAILURE,
329                                         "Failed to add queues to Tx adapter");
330         }
331
332         ret = rte_event_eth_tx_adapter_event_port_get(cdata.tx_adapter_id,
333                         &tx_port_id);
334         if (ret)
335                 rte_exit(EXIT_FAILURE,
336                                 "Failed to get Tx adapter port id");
337         ret = rte_event_port_link(evdev_id, tx_port_id, &cdata.tx_queue_id,
338                         NULL, 1);
339         if (ret != 1)
340                 rte_exit(EXIT_FAILURE,
341                                 "Unable to link Tx adapter port to Tx queue");
342
343         ret = rte_event_eth_rx_adapter_service_id_get(cdata.rx_adapter_id,
344                                 &fdata->rxadptr_service_id);
345         if (ret != -ESRCH && ret != 0) {
346                 rte_exit(EXIT_FAILURE,
347                         "Error getting the service ID for Rx adapter\n");
348         }
349         rte_service_runstate_set(fdata->rxadptr_service_id, 1);
350         rte_service_set_runstate_mapped_check(fdata->rxadptr_service_id, 0);
351
352         ret = rte_event_eth_tx_adapter_service_id_get(cdata.tx_adapter_id,
353                                 &fdata->txadptr_service_id);
354         if (ret != -ESRCH && ret != 0) {
355                 rte_exit(EXIT_FAILURE,
356                         "Error getting the service ID for Tx adapter\n");
357         }
358         rte_service_runstate_set(fdata->txadptr_service_id, 1);
359         rte_service_set_runstate_mapped_check(fdata->txadptr_service_id, 0);
360
361         ret = rte_event_eth_rx_adapter_start(cdata.rx_adapter_id);
362         if (ret)
363                 rte_exit(EXIT_FAILURE, "Rx adapter[%d] start failed",
364                                 cdata.rx_adapter_id);
365
366         ret = rte_event_eth_tx_adapter_start(cdata.tx_adapter_id);
367         if (ret)
368                 rte_exit(EXIT_FAILURE, "Tx adapter[%d] start failed",
369                                 cdata.tx_adapter_id);
370
371         if (rte_event_dev_start(evdev_id) < 0)
372                 rte_exit(EXIT_FAILURE, "Error starting eventdev");
373 }
374
375 static void
376 generic_opt_check(void)
377 {
378         int i;
379         int ret;
380         uint32_t cap = 0;
381         uint8_t rx_needed = 0;
382         uint8_t sched_needed = 0;
383         struct rte_event_dev_info eventdev_info;
384
385         memset(&eventdev_info, 0, sizeof(struct rte_event_dev_info));
386         rte_event_dev_info_get(0, &eventdev_info);
387
388         if (cdata.all_type_queues && !(eventdev_info.event_dev_cap &
389                                 RTE_EVENT_DEV_CAP_QUEUE_ALL_TYPES))
390                 rte_exit(EXIT_FAILURE,
391                                 "Event dev doesn't support all type queues\n");
392         sched_needed = !(eventdev_info.event_dev_cap &
393                 RTE_EVENT_DEV_CAP_DISTRIBUTED_SCHED);
394
395         RTE_ETH_FOREACH_DEV(i) {
396                 ret = rte_event_eth_rx_adapter_caps_get(0, i, &cap);
397                 if (ret)
398                         rte_exit(EXIT_FAILURE,
399                                 "failed to get event rx adapter capabilities");
400                 rx_needed |=
401                         !(cap & RTE_EVENT_ETH_RX_ADAPTER_CAP_INTERNAL_PORT);
402         }
403
404         if (cdata.worker_lcore_mask == 0 ||
405                         (rx_needed && cdata.rx_lcore_mask == 0) ||
406                         (cdata.tx_lcore_mask == 0) ||
407                         (sched_needed && cdata.sched_lcore_mask == 0)) {
408                 printf("Core part of pipeline was not assigned any cores. "
409                         "This will stall the pipeline, please check core masks "
410                         "(use -h for details on setting core masks):\n"
411                         "\trx: %"PRIu64"\n\ttx: %"PRIu64"\n\tsched: %"PRIu64
412                         "\n\tworkers: %"PRIu64"\n",
413                         cdata.rx_lcore_mask, cdata.tx_lcore_mask,
414                         cdata.sched_lcore_mask,
415                         cdata.worker_lcore_mask);
416                 rte_exit(-1, "Fix core masks\n");
417         }
418
419         if (!sched_needed)
420                 memset(fdata->sched_core, 0,
421                                 sizeof(unsigned int) * MAX_NUM_CORE);
422         if (!rx_needed)
423                 memset(fdata->rx_core, 0,
424                                 sizeof(unsigned int) * MAX_NUM_CORE);
425 }
426
427 void
428 set_worker_generic_setup_data(struct setup_data *caps, bool burst)
429 {
430         if (burst) {
431                 caps->worker = worker_generic_burst;
432         } else {
433                 caps->worker = worker_generic;
434         }
435
436         caps->adptr_setup = init_adapters;
437         caps->scheduler = schedule_devices;
438         caps->evdev_setup = setup_eventdev_generic;
439         caps->check_opt = generic_opt_check;
440 }