xref: /f-stack/dpdk/drivers/event/sw/sw_evdev.c (revision 2bfe3f2e)
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright(c) 2016-2017 Intel Corporation. All rights reserved.
5  *
6  *   Redistribution and use in source and binary forms, with or without
7  *   modification, are permitted provided that the following conditions
8  *   are met:
9  *
10  *     * Redistributions of source code must retain the above copyright
11  *       notice, this list of conditions and the following disclaimer.
12  *     * Redistributions in binary form must reproduce the above copyright
13  *       notice, this list of conditions and the following disclaimer in
14  *       the documentation and/or other materials provided with the
15  *       distribution.
16  *     * Neither the name of Intel Corporation nor the names of its
17  *       contributors may be used to endorse or promote products derived
18  *       from this software without specific prior written permission.
19  *
20  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
23  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
24  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
25  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
26  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
27  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
28  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
29  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
30  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
31  */
32 
33 #include <inttypes.h>
34 #include <string.h>
35 
36 #include <rte_bus_vdev.h>
37 #include <rte_kvargs.h>
38 #include <rte_ring.h>
39 #include <rte_errno.h>
40 #include <rte_event_ring.h>
41 #include <rte_service_component.h>
42 
43 #include "sw_evdev.h"
44 #include "iq_ring.h"
45 
46 #define EVENTDEV_NAME_SW_PMD event_sw
47 #define NUMA_NODE_ARG "numa_node"
48 #define SCHED_QUANTA_ARG "sched_quanta"
49 #define CREDIT_QUANTA_ARG "credit_quanta"
50 
51 static void
52 sw_info_get(struct rte_eventdev *dev, struct rte_event_dev_info *info);
53 
54 static int
55 sw_port_link(struct rte_eventdev *dev, void *port, const uint8_t queues[],
56 		const uint8_t priorities[], uint16_t num)
57 {
58 	struct sw_port *p = port;
59 	struct sw_evdev *sw = sw_pmd_priv(dev);
60 	int i;
61 
62 	RTE_SET_USED(priorities);
63 	for (i = 0; i < num; i++) {
64 		struct sw_qid *q = &sw->qids[queues[i]];
65 		unsigned int j;
66 
67 		/* check for qid map overflow */
68 		if (q->cq_num_mapped_cqs >= RTE_DIM(q->cq_map)) {
69 			rte_errno = -EDQUOT;
70 			break;
71 		}
72 
73 		if (p->is_directed && p->num_qids_mapped > 0) {
74 			rte_errno = -EDQUOT;
75 			break;
76 		}
77 
78 		for (j = 0; j < q->cq_num_mapped_cqs; j++) {
79 			if (q->cq_map[j] == p->id)
80 				break;
81 		}
82 
83 		/* check if port is already linked */
84 		if (j < q->cq_num_mapped_cqs)
85 			continue;
86 
87 		if (q->type == SW_SCHED_TYPE_DIRECT) {
88 			/* check directed qids only map to one port */
89 			if (p->num_qids_mapped > 0) {
90 				rte_errno = -EDQUOT;
91 				break;
92 			}
93 			/* check port only takes a directed flow */
94 			if (num > 1) {
95 				rte_errno = -EDQUOT;
96 				break;
97 			}
98 
99 			p->is_directed = 1;
100 			p->num_qids_mapped = 1;
101 		} else if (q->type == RTE_SCHED_TYPE_ORDERED) {
102 			p->num_ordered_qids++;
103 			p->num_qids_mapped++;
104 		} else if (q->type == RTE_SCHED_TYPE_ATOMIC ||
105 				q->type == RTE_SCHED_TYPE_PARALLEL) {
106 			p->num_qids_mapped++;
107 		}
108 
109 		q->cq_map[q->cq_num_mapped_cqs] = p->id;
110 		rte_smp_wmb();
111 		q->cq_num_mapped_cqs++;
112 	}
113 	return i;
114 }
115 
116 static int
117 sw_port_unlink(struct rte_eventdev *dev, void *port, uint8_t queues[],
118 		uint16_t nb_unlinks)
119 {
120 	struct sw_port *p = port;
121 	struct sw_evdev *sw = sw_pmd_priv(dev);
122 	unsigned int i, j;
123 
124 	int unlinked = 0;
125 	for (i = 0; i < nb_unlinks; i++) {
126 		struct sw_qid *q = &sw->qids[queues[i]];
127 		for (j = 0; j < q->cq_num_mapped_cqs; j++) {
128 			if (q->cq_map[j] == p->id) {
129 				q->cq_map[j] =
130 					q->cq_map[q->cq_num_mapped_cqs - 1];
131 				rte_smp_wmb();
132 				q->cq_num_mapped_cqs--;
133 				unlinked++;
134 
135 				p->num_qids_mapped--;
136 
137 				if (q->type == RTE_SCHED_TYPE_ORDERED)
138 					p->num_ordered_qids--;
139 
140 				continue;
141 			}
142 		}
143 	}
144 	return unlinked;
145 }
146 
147 static int
148 sw_port_setup(struct rte_eventdev *dev, uint8_t port_id,
149 		const struct rte_event_port_conf *conf)
150 {
151 	struct sw_evdev *sw = sw_pmd_priv(dev);
152 	struct sw_port *p = &sw->ports[port_id];
153 	char buf[RTE_RING_NAMESIZE];
154 	unsigned int i;
155 
156 	struct rte_event_dev_info info;
157 	sw_info_get(dev, &info);
158 
159 	/* detect re-configuring and return credits to instance if needed */
160 	if (p->initialized) {
161 		/* taking credits from pool is done one quanta at a time, and
162 		 * credits may be spend (counted in p->inflights) or still
163 		 * available in the port (p->inflight_credits). We must return
164 		 * the sum to no leak credits
165 		 */
166 		int possible_inflights = p->inflight_credits + p->inflights;
167 		rte_atomic32_sub(&sw->inflights, possible_inflights);
168 	}
169 
170 	*p = (struct sw_port){0}; /* zero entire structure */
171 	p->id = port_id;
172 	p->sw = sw;
173 
174 	/* check to see if rings exists - port_setup() can be called multiple
175 	 * times legally (assuming device is stopped). If ring exists, free it
176 	 * to so it gets re-created with the correct size
177 	 */
178 	snprintf(buf, sizeof(buf), "sw%d_p%u_%s", dev->data->dev_id,
179 			port_id, "rx_worker_ring");
180 	struct rte_event_ring *existing_ring = rte_event_ring_lookup(buf);
181 	if (existing_ring)
182 		rte_event_ring_free(existing_ring);
183 
184 	p->rx_worker_ring = rte_event_ring_create(buf, MAX_SW_PROD_Q_DEPTH,
185 			dev->data->socket_id,
186 			RING_F_SP_ENQ | RING_F_SC_DEQ | RING_F_EXACT_SZ);
187 	if (p->rx_worker_ring == NULL) {
188 		SW_LOG_ERR("Error creating RX worker ring for port %d\n",
189 				port_id);
190 		return -1;
191 	}
192 
193 	p->inflight_max = conf->new_event_threshold;
194 
195 	/* check if ring exists, same as rx_worker above */
196 	snprintf(buf, sizeof(buf), "sw%d_p%u, %s", dev->data->dev_id,
197 			port_id, "cq_worker_ring");
198 	existing_ring = rte_event_ring_lookup(buf);
199 	if (existing_ring)
200 		rte_event_ring_free(existing_ring);
201 
202 	p->cq_worker_ring = rte_event_ring_create(buf, conf->dequeue_depth,
203 			dev->data->socket_id,
204 			RING_F_SP_ENQ | RING_F_SC_DEQ | RING_F_EXACT_SZ);
205 	if (p->cq_worker_ring == NULL) {
206 		rte_event_ring_free(p->rx_worker_ring);
207 		SW_LOG_ERR("Error creating CQ worker ring for port %d\n",
208 				port_id);
209 		return -1;
210 	}
211 	sw->cq_ring_space[port_id] = conf->dequeue_depth;
212 
213 	/* set hist list contents to empty */
214 	for (i = 0; i < SW_PORT_HIST_LIST; i++) {
215 		p->hist_list[i].fid = -1;
216 		p->hist_list[i].qid = -1;
217 	}
218 	dev->data->ports[port_id] = p;
219 
220 	rte_smp_wmb();
221 	p->initialized = 1;
222 	return 0;
223 }
224 
225 static void
226 sw_port_release(void *port)
227 {
228 	struct sw_port *p = (void *)port;
229 	if (p == NULL)
230 		return;
231 
232 	rte_event_ring_free(p->rx_worker_ring);
233 	rte_event_ring_free(p->cq_worker_ring);
234 	memset(p, 0, sizeof(*p));
235 }
236 
237 static int32_t
238 qid_init(struct sw_evdev *sw, unsigned int idx, int type,
239 		const struct rte_event_queue_conf *queue_conf)
240 {
241 	unsigned int i;
242 	int dev_id = sw->data->dev_id;
243 	int socket_id = sw->data->socket_id;
244 	char buf[IQ_RING_NAMESIZE];
245 	struct sw_qid *qid = &sw->qids[idx];
246 
247 	for (i = 0; i < SW_IQS_MAX; i++) {
248 		snprintf(buf, sizeof(buf), "q_%u_iq_%d", idx, i);
249 		qid->iq[i] = iq_ring_create(buf, socket_id);
250 		if (!qid->iq[i]) {
251 			SW_LOG_DBG("ring create failed");
252 			goto cleanup;
253 		}
254 	}
255 
256 	/* Initialize the FID structures to no pinning (-1), and zero packets */
257 	const struct sw_fid_t fid = {.cq = -1, .pcount = 0};
258 	for (i = 0; i < RTE_DIM(qid->fids); i++)
259 		qid->fids[i] = fid;
260 
261 	qid->id = idx;
262 	qid->type = type;
263 	qid->priority = queue_conf->priority;
264 
265 	if (qid->type == RTE_SCHED_TYPE_ORDERED) {
266 		char ring_name[RTE_RING_NAMESIZE];
267 		uint32_t window_size;
268 
269 		/* rte_ring and window_size_mask require require window_size to
270 		 * be a power-of-2.
271 		 */
272 		window_size = rte_align32pow2(
273 				queue_conf->nb_atomic_order_sequences);
274 
275 		qid->window_size = window_size - 1;
276 
277 		if (!window_size) {
278 			SW_LOG_DBG(
279 				"invalid reorder_window_size for ordered queue\n"
280 				);
281 			goto cleanup;
282 		}
283 
284 		snprintf(buf, sizeof(buf), "sw%d_iq_%d_rob", dev_id, i);
285 		qid->reorder_buffer = rte_zmalloc_socket(buf,
286 				window_size * sizeof(qid->reorder_buffer[0]),
287 				0, socket_id);
288 		if (!qid->reorder_buffer) {
289 			SW_LOG_DBG("reorder_buffer malloc failed\n");
290 			goto cleanup;
291 		}
292 
293 		memset(&qid->reorder_buffer[0],
294 		       0,
295 		       window_size * sizeof(qid->reorder_buffer[0]));
296 
297 		snprintf(ring_name, sizeof(ring_name), "sw%d_q%d_freelist",
298 				dev_id, idx);
299 
300 		/* lookup the ring, and if it already exists, free it */
301 		struct rte_ring *cleanup = rte_ring_lookup(ring_name);
302 		if (cleanup)
303 			rte_ring_free(cleanup);
304 
305 		qid->reorder_buffer_freelist = rte_ring_create(ring_name,
306 				window_size,
307 				socket_id,
308 				RING_F_SP_ENQ | RING_F_SC_DEQ);
309 		if (!qid->reorder_buffer_freelist) {
310 			SW_LOG_DBG("freelist ring create failed");
311 			goto cleanup;
312 		}
313 
314 		/* Populate the freelist with reorder buffer entries. Enqueue
315 		 * 'window_size - 1' entries because the rte_ring holds only
316 		 * that many.
317 		 */
318 		for (i = 0; i < window_size - 1; i++) {
319 			if (rte_ring_sp_enqueue(qid->reorder_buffer_freelist,
320 						&qid->reorder_buffer[i]) < 0)
321 				goto cleanup;
322 		}
323 
324 		qid->reorder_buffer_index = 0;
325 		qid->cq_next_tx = 0;
326 	}
327 
328 	qid->initialized = 1;
329 
330 	return 0;
331 
332 cleanup:
333 	for (i = 0; i < SW_IQS_MAX; i++) {
334 		if (qid->iq[i])
335 			iq_ring_destroy(qid->iq[i]);
336 	}
337 
338 	if (qid->reorder_buffer) {
339 		rte_free(qid->reorder_buffer);
340 		qid->reorder_buffer = NULL;
341 	}
342 
343 	if (qid->reorder_buffer_freelist) {
344 		rte_ring_free(qid->reorder_buffer_freelist);
345 		qid->reorder_buffer_freelist = NULL;
346 	}
347 
348 	return -EINVAL;
349 }
350 
351 static void
352 sw_queue_release(struct rte_eventdev *dev, uint8_t id)
353 {
354 	struct sw_evdev *sw = sw_pmd_priv(dev);
355 	struct sw_qid *qid = &sw->qids[id];
356 	uint32_t i;
357 
358 	for (i = 0; i < SW_IQS_MAX; i++)
359 		iq_ring_destroy(qid->iq[i]);
360 
361 	if (qid->type == RTE_SCHED_TYPE_ORDERED) {
362 		rte_free(qid->reorder_buffer);
363 		rte_ring_free(qid->reorder_buffer_freelist);
364 	}
365 	memset(qid, 0, sizeof(*qid));
366 }
367 
368 static int
369 sw_queue_setup(struct rte_eventdev *dev, uint8_t queue_id,
370 		const struct rte_event_queue_conf *conf)
371 {
372 	int type;
373 
374 	type = conf->schedule_type;
375 
376 	if (RTE_EVENT_QUEUE_CFG_SINGLE_LINK & conf->event_queue_cfg) {
377 		type = SW_SCHED_TYPE_DIRECT;
378 	} else if (RTE_EVENT_QUEUE_CFG_ALL_TYPES
379 			& conf->event_queue_cfg) {
380 		SW_LOG_ERR("QUEUE_CFG_ALL_TYPES not supported\n");
381 		return -ENOTSUP;
382 	}
383 
384 	struct sw_evdev *sw = sw_pmd_priv(dev);
385 
386 	if (sw->qids[queue_id].initialized)
387 		sw_queue_release(dev, queue_id);
388 
389 	return qid_init(sw, queue_id, type, conf);
390 }
391 
392 static void
393 sw_queue_def_conf(struct rte_eventdev *dev, uint8_t queue_id,
394 				 struct rte_event_queue_conf *conf)
395 {
396 	RTE_SET_USED(dev);
397 	RTE_SET_USED(queue_id);
398 
399 	static const struct rte_event_queue_conf default_conf = {
400 		.nb_atomic_flows = 4096,
401 		.nb_atomic_order_sequences = 1,
402 		.schedule_type = RTE_SCHED_TYPE_ATOMIC,
403 		.priority = RTE_EVENT_DEV_PRIORITY_NORMAL,
404 	};
405 
406 	*conf = default_conf;
407 }
408 
409 static void
410 sw_port_def_conf(struct rte_eventdev *dev, uint8_t port_id,
411 		 struct rte_event_port_conf *port_conf)
412 {
413 	RTE_SET_USED(dev);
414 	RTE_SET_USED(port_id);
415 
416 	port_conf->new_event_threshold = 1024;
417 	port_conf->dequeue_depth = 16;
418 	port_conf->enqueue_depth = 16;
419 }
420 
421 static int
422 sw_dev_configure(const struct rte_eventdev *dev)
423 {
424 	struct sw_evdev *sw = sw_pmd_priv(dev);
425 	const struct rte_eventdev_data *data = dev->data;
426 	const struct rte_event_dev_config *conf = &data->dev_conf;
427 
428 	sw->qid_count = conf->nb_event_queues;
429 	sw->port_count = conf->nb_event_ports;
430 	sw->nb_events_limit = conf->nb_events_limit;
431 	rte_atomic32_set(&sw->inflights, 0);
432 
433 	if (conf->event_dev_cfg & RTE_EVENT_DEV_CFG_PER_DEQUEUE_TIMEOUT)
434 		return -ENOTSUP;
435 
436 	return 0;
437 }
438 
439 struct rte_eth_dev;
440 
441 static int
442 sw_eth_rx_adapter_caps_get(const struct rte_eventdev *dev,
443 			const struct rte_eth_dev *eth_dev,
444 			uint32_t *caps)
445 {
446 	RTE_SET_USED(dev);
447 	RTE_SET_USED(eth_dev);
448 	*caps = RTE_EVENT_ETH_RX_ADAPTER_SW_CAP;
449 	return 0;
450 }
451 
452 static void
453 sw_info_get(struct rte_eventdev *dev, struct rte_event_dev_info *info)
454 {
455 	RTE_SET_USED(dev);
456 
457 	static const struct rte_event_dev_info evdev_sw_info = {
458 			.driver_name = SW_PMD_NAME,
459 			.max_event_queues = RTE_EVENT_MAX_QUEUES_PER_DEV,
460 			.max_event_queue_flows = SW_QID_NUM_FIDS,
461 			.max_event_queue_priority_levels = SW_Q_PRIORITY_MAX,
462 			.max_event_priority_levels = SW_IQS_MAX,
463 			.max_event_ports = SW_PORTS_MAX,
464 			.max_event_port_dequeue_depth = MAX_SW_CONS_Q_DEPTH,
465 			.max_event_port_enqueue_depth = MAX_SW_PROD_Q_DEPTH,
466 			.max_num_events = SW_INFLIGHT_EVENTS_TOTAL,
467 			.event_dev_cap = (RTE_EVENT_DEV_CAP_QUEUE_QOS |
468 					RTE_EVENT_DEV_CAP_BURST_MODE |
469 					RTE_EVENT_DEV_CAP_EVENT_QOS),
470 	};
471 
472 	*info = evdev_sw_info;
473 }
474 
475 static void
476 sw_dump(struct rte_eventdev *dev, FILE *f)
477 {
478 	const struct sw_evdev *sw = sw_pmd_priv(dev);
479 
480 	static const char * const q_type_strings[] = {
481 			"Ordered", "Atomic", "Parallel", "Directed"
482 	};
483 	uint32_t i;
484 	fprintf(f, "EventDev %s: ports %d, qids %d\n", "todo-fix-name",
485 			sw->port_count, sw->qid_count);
486 
487 	fprintf(f, "\trx   %"PRIu64"\n\tdrop %"PRIu64"\n\ttx   %"PRIu64"\n",
488 		sw->stats.rx_pkts, sw->stats.rx_dropped, sw->stats.tx_pkts);
489 	fprintf(f, "\tsched calls: %"PRIu64"\n", sw->sched_called);
490 	fprintf(f, "\tsched cq/qid call: %"PRIu64"\n", sw->sched_cq_qid_called);
491 	fprintf(f, "\tsched no IQ enq: %"PRIu64"\n", sw->sched_no_iq_enqueues);
492 	fprintf(f, "\tsched no CQ enq: %"PRIu64"\n", sw->sched_no_cq_enqueues);
493 	uint32_t inflights = rte_atomic32_read(&sw->inflights);
494 	uint32_t credits = sw->nb_events_limit - inflights;
495 	fprintf(f, "\tinflight %d, credits: %d\n", inflights, credits);
496 
497 #define COL_RED "\x1b[31m"
498 #define COL_RESET "\x1b[0m"
499 
500 	for (i = 0; i < sw->port_count; i++) {
501 		int max, j;
502 		const struct sw_port *p = &sw->ports[i];
503 		if (!p->initialized) {
504 			fprintf(f, "  %sPort %d not initialized.%s\n",
505 				COL_RED, i, COL_RESET);
506 			continue;
507 		}
508 		fprintf(f, "  Port %d %s\n", i,
509 			p->is_directed ? " (SingleCons)" : "");
510 		fprintf(f, "\trx   %"PRIu64"\tdrop %"PRIu64"\ttx   %"PRIu64
511 			"\t%sinflight %d%s\n", sw->ports[i].stats.rx_pkts,
512 			sw->ports[i].stats.rx_dropped,
513 			sw->ports[i].stats.tx_pkts,
514 			(p->inflights == p->inflight_max) ?
515 				COL_RED : COL_RESET,
516 			sw->ports[i].inflights, COL_RESET);
517 
518 		fprintf(f, "\tMax New: %u"
519 			"\tAvg cycles PP: %"PRIu64"\tCredits: %u\n",
520 			sw->ports[i].inflight_max,
521 			sw->ports[i].avg_pkt_ticks,
522 			sw->ports[i].inflight_credits);
523 		fprintf(f, "\tReceive burst distribution:\n");
524 		float zp_percent = p->zero_polls * 100.0 / p->total_polls;
525 		fprintf(f, zp_percent < 10 ? "\t\t0:%.02f%% " : "\t\t0:%.0f%% ",
526 				zp_percent);
527 		for (max = (int)RTE_DIM(p->poll_buckets); max-- > 0;)
528 			if (p->poll_buckets[max] != 0)
529 				break;
530 		for (j = 0; j <= max; j++) {
531 			if (p->poll_buckets[j] != 0) {
532 				float poll_pc = p->poll_buckets[j] * 100.0 /
533 					p->total_polls;
534 				fprintf(f, "%u-%u:%.02f%% ",
535 					((j << SW_DEQ_STAT_BUCKET_SHIFT) + 1),
536 					((j+1) << SW_DEQ_STAT_BUCKET_SHIFT),
537 					poll_pc);
538 			}
539 		}
540 		fprintf(f, "\n");
541 
542 		if (p->rx_worker_ring) {
543 			uint64_t used = rte_event_ring_count(p->rx_worker_ring);
544 			uint64_t space = rte_event_ring_free_count(
545 					p->rx_worker_ring);
546 			const char *col = (space == 0) ? COL_RED : COL_RESET;
547 			fprintf(f, "\t%srx ring used: %4"PRIu64"\tfree: %4"
548 					PRIu64 COL_RESET"\n", col, used, space);
549 		} else
550 			fprintf(f, "\trx ring not initialized.\n");
551 
552 		if (p->cq_worker_ring) {
553 			uint64_t used = rte_event_ring_count(p->cq_worker_ring);
554 			uint64_t space = rte_event_ring_free_count(
555 					p->cq_worker_ring);
556 			const char *col = (space == 0) ? COL_RED : COL_RESET;
557 			fprintf(f, "\t%scq ring used: %4"PRIu64"\tfree: %4"
558 					PRIu64 COL_RESET"\n", col, used, space);
559 		} else
560 			fprintf(f, "\tcq ring not initialized.\n");
561 	}
562 
563 	for (i = 0; i < sw->qid_count; i++) {
564 		const struct sw_qid *qid = &sw->qids[i];
565 		if (!qid->initialized) {
566 			fprintf(f, "  %sQueue %d not initialized.%s\n",
567 				COL_RED, i, COL_RESET);
568 			continue;
569 		}
570 		int affinities_per_port[SW_PORTS_MAX] = {0};
571 		uint32_t inflights = 0;
572 
573 		fprintf(f, "  Queue %d (%s)\n", i, q_type_strings[qid->type]);
574 		fprintf(f, "\trx   %"PRIu64"\tdrop %"PRIu64"\ttx   %"PRIu64"\n",
575 			qid->stats.rx_pkts, qid->stats.rx_dropped,
576 			qid->stats.tx_pkts);
577 		if (qid->type == RTE_SCHED_TYPE_ORDERED) {
578 			struct rte_ring *rob_buf_free =
579 				qid->reorder_buffer_freelist;
580 			if (rob_buf_free)
581 				fprintf(f, "\tReorder entries in use: %u\n",
582 					rte_ring_free_count(rob_buf_free));
583 			else
584 				fprintf(f,
585 					"\tReorder buffer not initialized\n");
586 		}
587 
588 		uint32_t flow;
589 		for (flow = 0; flow < RTE_DIM(qid->fids); flow++)
590 			if (qid->fids[flow].cq != -1) {
591 				affinities_per_port[qid->fids[flow].cq]++;
592 				inflights += qid->fids[flow].pcount;
593 			}
594 
595 		uint32_t port;
596 		fprintf(f, "\tPer Port Stats:\n");
597 		for (port = 0; port < sw->port_count; port++) {
598 			fprintf(f, "\t  Port %d: Pkts: %"PRIu64, port,
599 					qid->to_port[port]);
600 			fprintf(f, "\tFlows: %d\n", affinities_per_port[port]);
601 		}
602 
603 		uint32_t iq;
604 		uint32_t iq_printed = 0;
605 		for (iq = 0; iq < SW_IQS_MAX; iq++) {
606 			if (!qid->iq[iq]) {
607 				fprintf(f, "\tiq %d is not initialized.\n", iq);
608 				iq_printed = 1;
609 				continue;
610 			}
611 			uint32_t used = iq_ring_count(qid->iq[iq]);
612 			uint32_t free = iq_ring_free_count(qid->iq[iq]);
613 			const char *col = (free == 0) ? COL_RED : COL_RESET;
614 			if (used > 0) {
615 				fprintf(f, "\t%siq %d: Used %d\tFree %d"
616 					COL_RESET"\n", col, iq, used, free);
617 				iq_printed = 1;
618 			}
619 		}
620 		if (iq_printed == 0)
621 			fprintf(f, "\t-- iqs empty --\n");
622 	}
623 }
624 
625 static int
626 sw_start(struct rte_eventdev *dev)
627 {
628 	unsigned int i, j;
629 	struct sw_evdev *sw = sw_pmd_priv(dev);
630 
631 	rte_service_component_runstate_set(sw->service_id, 1);
632 
633 	/* check a service core is mapped to this service */
634 	if (!rte_service_runstate_get(sw->service_id)) {
635 		SW_LOG_ERR("Warning: No Service core enabled on service %s\n",
636 				sw->service_name);
637 		return -ENOENT;
638 	}
639 
640 	/* check all ports are set up */
641 	for (i = 0; i < sw->port_count; i++)
642 		if (sw->ports[i].rx_worker_ring == NULL) {
643 			SW_LOG_ERR("Port %d not configured\n", i);
644 			return -ESTALE;
645 		}
646 
647 	/* check all queues are configured and mapped to ports*/
648 	for (i = 0; i < sw->qid_count; i++)
649 		if (sw->qids[i].iq[0] == NULL ||
650 				sw->qids[i].cq_num_mapped_cqs == 0) {
651 			SW_LOG_ERR("Queue %d not configured\n", i);
652 			return -ENOLINK;
653 		}
654 
655 	/* build up our prioritized array of qids */
656 	/* We don't use qsort here, as if all/multiple entries have the same
657 	 * priority, the result is non-deterministic. From "man 3 qsort":
658 	 * "If two members compare as equal, their order in the sorted
659 	 * array is undefined."
660 	 */
661 	uint32_t qidx = 0;
662 	for (j = 0; j <= RTE_EVENT_DEV_PRIORITY_LOWEST; j++) {
663 		for (i = 0; i < sw->qid_count; i++) {
664 			if (sw->qids[i].priority == j) {
665 				sw->qids_prioritized[qidx] = &sw->qids[i];
666 				qidx++;
667 			}
668 		}
669 	}
670 
671 	if (sw_xstats_init(sw) < 0)
672 		return -EINVAL;
673 
674 	rte_smp_wmb();
675 	sw->started = 1;
676 
677 	return 0;
678 }
679 
680 static void
681 sw_stop(struct rte_eventdev *dev)
682 {
683 	struct sw_evdev *sw = sw_pmd_priv(dev);
684 	sw_xstats_uninit(sw);
685 	sw->started = 0;
686 	rte_smp_wmb();
687 }
688 
689 static int
690 sw_close(struct rte_eventdev *dev)
691 {
692 	struct sw_evdev *sw = sw_pmd_priv(dev);
693 	uint32_t i;
694 
695 	for (i = 0; i < sw->qid_count; i++)
696 		sw_queue_release(dev, i);
697 	sw->qid_count = 0;
698 
699 	for (i = 0; i < sw->port_count; i++)
700 		sw_port_release(&sw->ports[i]);
701 	sw->port_count = 0;
702 
703 	memset(&sw->stats, 0, sizeof(sw->stats));
704 	sw->sched_called = 0;
705 	sw->sched_no_iq_enqueues = 0;
706 	sw->sched_no_cq_enqueues = 0;
707 	sw->sched_cq_qid_called = 0;
708 
709 	return 0;
710 }
711 
712 static int
713 assign_numa_node(const char *key __rte_unused, const char *value, void *opaque)
714 {
715 	int *socket_id = opaque;
716 	*socket_id = atoi(value);
717 	if (*socket_id >= RTE_MAX_NUMA_NODES)
718 		return -1;
719 	return 0;
720 }
721 
722 static int
723 set_sched_quanta(const char *key __rte_unused, const char *value, void *opaque)
724 {
725 	int *quanta = opaque;
726 	*quanta = atoi(value);
727 	if (*quanta < 0 || *quanta >= 4096)
728 		return -1;
729 	return 0;
730 }
731 
732 static int
733 set_credit_quanta(const char *key __rte_unused, const char *value, void *opaque)
734 {
735 	int *credit = opaque;
736 	*credit = atoi(value);
737 	if (*credit < 0 || *credit >= 128)
738 		return -1;
739 	return 0;
740 }
741 
742 
743 static int32_t sw_sched_service_func(void *args)
744 {
745 	struct rte_eventdev *dev = args;
746 	sw_event_schedule(dev);
747 	return 0;
748 }
749 
750 static int
751 sw_probe(struct rte_vdev_device *vdev)
752 {
753 	static const struct rte_eventdev_ops evdev_sw_ops = {
754 			.dev_configure = sw_dev_configure,
755 			.dev_infos_get = sw_info_get,
756 			.dev_close = sw_close,
757 			.dev_start = sw_start,
758 			.dev_stop = sw_stop,
759 			.dump = sw_dump,
760 
761 			.queue_def_conf = sw_queue_def_conf,
762 			.queue_setup = sw_queue_setup,
763 			.queue_release = sw_queue_release,
764 			.port_def_conf = sw_port_def_conf,
765 			.port_setup = sw_port_setup,
766 			.port_release = sw_port_release,
767 			.port_link = sw_port_link,
768 			.port_unlink = sw_port_unlink,
769 
770 			.eth_rx_adapter_caps_get = sw_eth_rx_adapter_caps_get,
771 
772 			.xstats_get = sw_xstats_get,
773 			.xstats_get_names = sw_xstats_get_names,
774 			.xstats_get_by_name = sw_xstats_get_by_name,
775 			.xstats_reset = sw_xstats_reset,
776 	};
777 
778 	static const char *const args[] = {
779 		NUMA_NODE_ARG,
780 		SCHED_QUANTA_ARG,
781 		CREDIT_QUANTA_ARG,
782 		NULL
783 	};
784 	const char *name;
785 	const char *params;
786 	struct rte_eventdev *dev;
787 	struct sw_evdev *sw;
788 	int socket_id = rte_socket_id();
789 	int sched_quanta  = SW_DEFAULT_SCHED_QUANTA;
790 	int credit_quanta = SW_DEFAULT_CREDIT_QUANTA;
791 
792 	name = rte_vdev_device_name(vdev);
793 	params = rte_vdev_device_args(vdev);
794 	if (params != NULL && params[0] != '\0') {
795 		struct rte_kvargs *kvlist = rte_kvargs_parse(params, args);
796 
797 		if (!kvlist) {
798 			SW_LOG_INFO(
799 				"Ignoring unsupported parameters when creating device '%s'\n",
800 				name);
801 		} else {
802 			int ret = rte_kvargs_process(kvlist, NUMA_NODE_ARG,
803 					assign_numa_node, &socket_id);
804 			if (ret != 0) {
805 				SW_LOG_ERR(
806 					"%s: Error parsing numa node parameter",
807 					name);
808 				rte_kvargs_free(kvlist);
809 				return ret;
810 			}
811 
812 			ret = rte_kvargs_process(kvlist, SCHED_QUANTA_ARG,
813 					set_sched_quanta, &sched_quanta);
814 			if (ret != 0) {
815 				SW_LOG_ERR(
816 					"%s: Error parsing sched quanta parameter",
817 					name);
818 				rte_kvargs_free(kvlist);
819 				return ret;
820 			}
821 
822 			ret = rte_kvargs_process(kvlist, CREDIT_QUANTA_ARG,
823 					set_credit_quanta, &credit_quanta);
824 			if (ret != 0) {
825 				SW_LOG_ERR(
826 					"%s: Error parsing credit quanta parameter",
827 					name);
828 				rte_kvargs_free(kvlist);
829 				return ret;
830 			}
831 
832 			rte_kvargs_free(kvlist);
833 		}
834 	}
835 
836 	SW_LOG_INFO(
837 			"Creating eventdev sw device %s, numa_node=%d, sched_quanta=%d, credit_quanta=%d\n",
838 			name, socket_id, sched_quanta, credit_quanta);
839 
840 	dev = rte_event_pmd_vdev_init(name,
841 			sizeof(struct sw_evdev), socket_id);
842 	if (dev == NULL) {
843 		SW_LOG_ERR("eventdev vdev init() failed");
844 		return -EFAULT;
845 	}
846 	dev->dev_ops = &evdev_sw_ops;
847 	dev->enqueue = sw_event_enqueue;
848 	dev->enqueue_burst = sw_event_enqueue_burst;
849 	dev->enqueue_new_burst = sw_event_enqueue_burst;
850 	dev->enqueue_forward_burst = sw_event_enqueue_burst;
851 	dev->dequeue = sw_event_dequeue;
852 	dev->dequeue_burst = sw_event_dequeue_burst;
853 
854 	if (rte_eal_process_type() != RTE_PROC_PRIMARY)
855 		return 0;
856 
857 	sw = dev->data->dev_private;
858 	sw->data = dev->data;
859 
860 	/* copy values passed from vdev command line to instance */
861 	sw->credit_update_quanta = credit_quanta;
862 	sw->sched_quanta = sched_quanta;
863 
864 	/* register service with EAL */
865 	struct rte_service_spec service;
866 	memset(&service, 0, sizeof(struct rte_service_spec));
867 	snprintf(service.name, sizeof(service.name), "%s_service", name);
868 	snprintf(sw->service_name, sizeof(sw->service_name), "%s_service",
869 			name);
870 	service.socket_id = socket_id;
871 	service.callback = sw_sched_service_func;
872 	service.callback_userdata = (void *)dev;
873 
874 	int32_t ret = rte_service_component_register(&service, &sw->service_id);
875 	if (ret) {
876 		SW_LOG_ERR("service register() failed");
877 		return -ENOEXEC;
878 	}
879 
880 	dev->data->service_inited = 1;
881 	dev->data->service_id = sw->service_id;
882 
883 	return 0;
884 }
885 
886 static int
887 sw_remove(struct rte_vdev_device *vdev)
888 {
889 	const char *name;
890 
891 	name = rte_vdev_device_name(vdev);
892 	if (name == NULL)
893 		return -EINVAL;
894 
895 	SW_LOG_INFO("Closing eventdev sw device %s\n", name);
896 
897 	return rte_event_pmd_vdev_uninit(name);
898 }
899 
900 static struct rte_vdev_driver evdev_sw_pmd_drv = {
901 	.probe = sw_probe,
902 	.remove = sw_remove
903 };
904 
905 RTE_PMD_REGISTER_VDEV(EVENTDEV_NAME_SW_PMD, evdev_sw_pmd_drv);
906 RTE_PMD_REGISTER_PARAM_STRING(event_sw, NUMA_NODE_ARG "=<int> "
907 		SCHED_QUANTA_ARG "=<int>" CREDIT_QUANTA_ARG "=<int>");
908