xref: /dpdk/drivers/net/mlx5/mlx5_flow_verbs.c (revision daa02b5cddbb8e11b31d41e2bf7bb1ae64dcae2f)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2018 Mellanox Technologies, Ltd
3  */
4 
5 #include <netinet/in.h>
6 #include <sys/queue.h>
7 #include <stdalign.h>
8 #include <stdint.h>
9 #include <string.h>
10 
11 #include <rte_common.h>
12 #include <rte_ether.h>
13 #include <ethdev_driver.h>
14 #include <rte_flow.h>
15 #include <rte_flow_driver.h>
16 #include <rte_malloc.h>
17 #include <rte_ip.h>
18 
19 #include <mlx5_glue.h>
20 #include <mlx5_prm.h>
21 #include <mlx5_malloc.h>
22 
23 #include "mlx5_defs.h"
24 #include "mlx5.h"
25 #include "mlx5_flow.h"
26 #include "mlx5_rx.h"
27 
28 #define VERBS_SPEC_INNER(item_flags) \
29 	(!!((item_flags) & MLX5_FLOW_LAYER_TUNNEL) ? IBV_FLOW_SPEC_INNER : 0)
30 
31 /* Map of Verbs to Flow priority with 8 Verbs priorities. */
32 static const uint32_t priority_map_3[][MLX5_PRIORITY_MAP_MAX] = {
33 	{ 0, 1, 2 }, { 2, 3, 4 }, { 5, 6, 7 },
34 };
35 
36 /* Map of Verbs to Flow priority with 16 Verbs priorities. */
37 static const uint32_t priority_map_5[][MLX5_PRIORITY_MAP_MAX] = {
38 	{ 0, 1, 2 }, { 3, 4, 5 }, { 6, 7, 8 },
39 	{ 9, 10, 11 }, { 12, 13, 14 },
40 };
41 
42 /* Verbs specification header. */
43 struct ibv_spec_header {
44 	enum ibv_flow_spec_type type;
45 	uint16_t size;
46 };
47 
48 /**
49  * Discover the maximum number of priority available.
50  *
51  * @param[in] dev
52  *   Pointer to the Ethernet device structure.
53  *
54  * @return
55  *   number of supported flow priority on success, a negative errno
56  *   value otherwise and rte_errno is set.
57  */
58 int
59 mlx5_flow_discover_priorities(struct rte_eth_dev *dev)
60 {
61 	struct mlx5_priv *priv = dev->data->dev_private;
62 	struct {
63 		struct ibv_flow_attr attr;
64 		struct ibv_flow_spec_eth eth;
65 		struct ibv_flow_spec_action_drop drop;
66 	} flow_attr = {
67 		.attr = {
68 			.num_of_specs = 2,
69 			.port = (uint8_t)priv->dev_port,
70 		},
71 		.eth = {
72 			.type = IBV_FLOW_SPEC_ETH,
73 			.size = sizeof(struct ibv_flow_spec_eth),
74 		},
75 		.drop = {
76 			.size = sizeof(struct ibv_flow_spec_action_drop),
77 			.type = IBV_FLOW_SPEC_ACTION_DROP,
78 		},
79 	};
80 	struct ibv_flow *flow;
81 	struct mlx5_hrxq *drop = priv->drop_queue.hrxq;
82 	uint16_t vprio[] = { 8, 16 };
83 	int i;
84 	int priority = 0;
85 
86 #if defined(HAVE_MLX5DV_DR_DEVX_PORT) || defined(HAVE_MLX5DV_DR_DEVX_PORT_V35)
87 	/* If DevX supported, driver must support 16 verbs flow priorities. */
88 	priority = RTE_DIM(priority_map_5);
89 	goto out;
90 #endif
91 	if (!drop->qp) {
92 		rte_errno = ENOTSUP;
93 		return -rte_errno;
94 	}
95 	for (i = 0; i != RTE_DIM(vprio); i++) {
96 		flow_attr.attr.priority = vprio[i] - 1;
97 		flow = mlx5_glue->create_flow(drop->qp, &flow_attr.attr);
98 		if (!flow)
99 			break;
100 		claim_zero(mlx5_glue->destroy_flow(flow));
101 		priority = vprio[i];
102 	}
103 	switch (priority) {
104 	case 8:
105 		priority = RTE_DIM(priority_map_3);
106 		break;
107 	case 16:
108 		priority = RTE_DIM(priority_map_5);
109 		break;
110 	default:
111 		rte_errno = ENOTSUP;
112 		DRV_LOG(ERR,
113 			"port %u verbs maximum priority: %d expected 8/16",
114 			dev->data->port_id, priority);
115 		return -rte_errno;
116 	}
117 #if defined(HAVE_MLX5DV_DR_DEVX_PORT) || defined(HAVE_MLX5DV_DR_DEVX_PORT_V35)
118 out:
119 #endif
120 	DRV_LOG(INFO, "port %u supported flow priorities:"
121 		" 0-%d for ingress or egress root table,"
122 		" 0-%d for non-root table or transfer root table.",
123 		dev->data->port_id, priority - 2,
124 		MLX5_NON_ROOT_FLOW_MAX_PRIO - 1);
125 	return priority;
126 }
127 
128 /**
129  * Adjust flow priority based on the highest layer and the request priority.
130  *
131  * @param[in] dev
132  *   Pointer to the Ethernet device structure.
133  * @param[in] priority
134  *   The rule base priority.
135  * @param[in] subpriority
136  *   The priority based on the items.
137  *
138  * @return
139  *   The new priority.
140  */
141 uint32_t
142 mlx5_flow_adjust_priority(struct rte_eth_dev *dev, int32_t priority,
143 				   uint32_t subpriority)
144 {
145 	uint32_t res = 0;
146 	struct mlx5_priv *priv = dev->data->dev_private;
147 
148 	switch (priv->config.flow_prio) {
149 	case RTE_DIM(priority_map_3):
150 		res = priority_map_3[priority][subpriority];
151 		break;
152 	case RTE_DIM(priority_map_5):
153 		res = priority_map_5[priority][subpriority];
154 		break;
155 	}
156 	return  res;
157 }
158 
159 /**
160  * Get Verbs flow counter by index.
161  *
162  * @param[in] dev
163  *   Pointer to the Ethernet device structure.
164  * @param[in] idx
165  *   mlx5 flow counter index in the container.
166  * @param[out] ppool
167  *   mlx5 flow counter pool in the container,
168  *
169  * @return
170  *   A pointer to the counter, NULL otherwise.
171  */
172 static struct mlx5_flow_counter *
173 flow_verbs_counter_get_by_idx(struct rte_eth_dev *dev,
174 			      uint32_t idx,
175 			      struct mlx5_flow_counter_pool **ppool)
176 {
177 	struct mlx5_priv *priv = dev->data->dev_private;
178 	struct mlx5_flow_counter_mng *cmng = &priv->sh->cmng;
179 	struct mlx5_flow_counter_pool *pool;
180 
181 	idx = (idx - 1) & (MLX5_CNT_SHARED_OFFSET - 1);
182 	pool = cmng->pools[idx / MLX5_COUNTERS_PER_POOL];
183 	MLX5_ASSERT(pool);
184 	if (ppool)
185 		*ppool = pool;
186 	return MLX5_POOL_GET_CNT(pool, idx % MLX5_COUNTERS_PER_POOL);
187 }
188 
189 /**
190  * Create Verbs flow counter with Verbs library.
191  *
192  * @param[in] dev
193  *   Pointer to the Ethernet device structure.
194  * @param[in, out] counter
195  *   mlx5 flow counter object, contains the counter id,
196  *   handle of created Verbs flow counter is returned
197  *   in cs field (if counters are supported).
198  *
199  * @return
200  *   0 On success else a negative errno value is returned
201  *   and rte_errno is set.
202  */
203 static int
204 flow_verbs_counter_create(struct rte_eth_dev *dev,
205 			  struct mlx5_flow_counter *counter)
206 {
207 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42)
208 	struct mlx5_priv *priv = dev->data->dev_private;
209 	struct ibv_context *ctx = priv->sh->cdev->ctx;
210 	struct ibv_counter_set_init_attr init = {
211 			 .counter_set_id = counter->shared_info.id};
212 
213 	counter->dcs_when_free = mlx5_glue->create_counter_set(ctx, &init);
214 	if (!counter->dcs_when_free) {
215 		rte_errno = ENOTSUP;
216 		return -ENOTSUP;
217 	}
218 	return 0;
219 #elif defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
220 	struct mlx5_priv *priv = dev->data->dev_private;
221 	struct ibv_context *ctx = priv->sh->cdev->ctx;
222 	struct ibv_counters_init_attr init = {0};
223 	struct ibv_counter_attach_attr attach;
224 	int ret;
225 
226 	memset(&attach, 0, sizeof(attach));
227 	counter->dcs_when_free = mlx5_glue->create_counters(ctx, &init);
228 	if (!counter->dcs_when_free) {
229 		rte_errno = ENOTSUP;
230 		return -ENOTSUP;
231 	}
232 	attach.counter_desc = IBV_COUNTER_PACKETS;
233 	attach.index = 0;
234 	ret = mlx5_glue->attach_counters(counter->dcs_when_free, &attach, NULL);
235 	if (!ret) {
236 		attach.counter_desc = IBV_COUNTER_BYTES;
237 		attach.index = 1;
238 		ret = mlx5_glue->attach_counters
239 					(counter->dcs_when_free, &attach, NULL);
240 	}
241 	if (ret) {
242 		claim_zero(mlx5_glue->destroy_counters(counter->dcs_when_free));
243 		counter->dcs_when_free = NULL;
244 		rte_errno = ret;
245 		return -ret;
246 	}
247 	return 0;
248 #else
249 	(void)dev;
250 	(void)counter;
251 	rte_errno = ENOTSUP;
252 	return -ENOTSUP;
253 #endif
254 }
255 
256 /**
257  * Get a flow counter.
258  *
259  * @param[in] dev
260  *   Pointer to the Ethernet device structure.
261  * @param[in] id
262  *   Counter identifier.
263  *
264  * @return
265  *   Index to the counter, 0 otherwise and rte_errno is set.
266  */
267 static uint32_t
268 flow_verbs_counter_new(struct rte_eth_dev *dev, uint32_t id __rte_unused)
269 {
270 	struct mlx5_priv *priv = dev->data->dev_private;
271 	struct mlx5_flow_counter_mng *cmng = &priv->sh->cmng;
272 	struct mlx5_flow_counter_pool *pool = NULL;
273 	struct mlx5_flow_counter *cnt = NULL;
274 	uint32_t n_valid = cmng->n_valid;
275 	uint32_t pool_idx, cnt_idx;
276 	uint32_t i;
277 	int ret;
278 
279 	for (pool_idx = 0; pool_idx < n_valid; ++pool_idx) {
280 		pool = cmng->pools[pool_idx];
281 		if (!pool)
282 			continue;
283 		cnt = TAILQ_FIRST(&pool->counters[0]);
284 		if (cnt)
285 			break;
286 	}
287 	if (!cnt) {
288 		struct mlx5_flow_counter_pool **pools;
289 		uint32_t size;
290 
291 		if (n_valid == cmng->n) {
292 			/* Resize the container pool array. */
293 			size = sizeof(struct mlx5_flow_counter_pool *) *
294 				     (n_valid + MLX5_CNT_CONTAINER_RESIZE);
295 			pools = mlx5_malloc(MLX5_MEM_ZERO, size, 0,
296 					    SOCKET_ID_ANY);
297 			if (!pools)
298 				return 0;
299 			if (n_valid) {
300 				memcpy(pools, cmng->pools,
301 				       sizeof(struct mlx5_flow_counter_pool *) *
302 				       n_valid);
303 				mlx5_free(cmng->pools);
304 			}
305 			cmng->pools = pools;
306 			cmng->n += MLX5_CNT_CONTAINER_RESIZE;
307 		}
308 		/* Allocate memory for new pool*/
309 		size = sizeof(*pool) + sizeof(*cnt) * MLX5_COUNTERS_PER_POOL;
310 		pool = mlx5_malloc(MLX5_MEM_ZERO, size, 0, SOCKET_ID_ANY);
311 		if (!pool)
312 			return 0;
313 		for (i = 0; i < MLX5_COUNTERS_PER_POOL; ++i) {
314 			cnt = MLX5_POOL_GET_CNT(pool, i);
315 			TAILQ_INSERT_HEAD(&pool->counters[0], cnt, next);
316 		}
317 		cnt = MLX5_POOL_GET_CNT(pool, 0);
318 		cmng->pools[n_valid] = pool;
319 		pool_idx = n_valid;
320 		cmng->n_valid++;
321 	}
322 	TAILQ_REMOVE(&pool->counters[0], cnt, next);
323 	i = MLX5_CNT_ARRAY_IDX(pool, cnt);
324 	cnt_idx = MLX5_MAKE_CNT_IDX(pool_idx, i);
325 	/* Create counter with Verbs. */
326 	ret = flow_verbs_counter_create(dev, cnt);
327 	if (!ret) {
328 		cnt->dcs_when_active = cnt->dcs_when_free;
329 		cnt->hits = 0;
330 		cnt->bytes = 0;
331 		return cnt_idx;
332 	}
333 	TAILQ_INSERT_HEAD(&pool->counters[0], cnt, next);
334 	/* Some error occurred in Verbs library. */
335 	rte_errno = -ret;
336 	return 0;
337 }
338 
339 /**
340  * Release a flow counter.
341  *
342  * @param[in] dev
343  *   Pointer to the Ethernet device structure.
344  * @param[in] counter
345  *   Index to the counter handler.
346  */
347 static void
348 flow_verbs_counter_release(struct rte_eth_dev *dev, uint32_t counter)
349 {
350 	struct mlx5_flow_counter_pool *pool;
351 	struct mlx5_flow_counter *cnt;
352 
353 	cnt = flow_verbs_counter_get_by_idx(dev, counter, &pool);
354 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42)
355 	claim_zero(mlx5_glue->destroy_counter_set
356 			((struct ibv_counter_set *)cnt->dcs_when_active));
357 #elif defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
358 	claim_zero(mlx5_glue->destroy_counters
359 				((struct ibv_counters *)cnt->dcs_when_active));
360 #endif
361 	TAILQ_INSERT_HEAD(&pool->counters[0], cnt, next);
362 }
363 
364 /**
365  * Query a flow counter via Verbs library call.
366  *
367  * @see rte_flow_query()
368  * @see rte_flow_ops
369  */
370 static int
371 flow_verbs_counter_query(struct rte_eth_dev *dev __rte_unused,
372 			 struct rte_flow *flow, void *data,
373 			 struct rte_flow_error *error)
374 {
375 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) || \
376 	defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
377 	if (flow->counter) {
378 		struct mlx5_flow_counter_pool *pool;
379 		struct mlx5_flow_counter *cnt = flow_verbs_counter_get_by_idx
380 						(dev, flow->counter, &pool);
381 		struct rte_flow_query_count *qc = data;
382 		uint64_t counters[2] = {0, 0};
383 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42)
384 		struct ibv_query_counter_set_attr query_cs_attr = {
385 			.dcs_when_free = (struct ibv_counter_set *)
386 						cnt->dcs_when_active,
387 			.query_flags = IBV_COUNTER_SET_FORCE_UPDATE,
388 		};
389 		struct ibv_counter_set_data query_out = {
390 			.out = counters,
391 			.outlen = 2 * sizeof(uint64_t),
392 		};
393 		int err = mlx5_glue->query_counter_set(&query_cs_attr,
394 						       &query_out);
395 #elif defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
396 		int err = mlx5_glue->query_counters
397 			((struct ibv_counters *)cnt->dcs_when_active, counters,
398 				RTE_DIM(counters),
399 				IBV_READ_COUNTERS_ATTR_PREFER_CACHED);
400 #endif
401 		if (err)
402 			return rte_flow_error_set
403 				(error, err,
404 				 RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
405 				 NULL,
406 				 "cannot read counter");
407 		qc->hits_set = 1;
408 		qc->bytes_set = 1;
409 		qc->hits = counters[0] - cnt->hits;
410 		qc->bytes = counters[1] - cnt->bytes;
411 		if (qc->reset) {
412 			cnt->hits = counters[0];
413 			cnt->bytes = counters[1];
414 		}
415 		return 0;
416 	}
417 	return rte_flow_error_set(error, EINVAL,
418 				  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
419 				  NULL,
420 				  "flow does not have counter");
421 #else
422 	(void)flow;
423 	(void)data;
424 	return rte_flow_error_set(error, ENOTSUP,
425 				  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
426 				  NULL,
427 				  "counters are not available");
428 #endif
429 }
430 
431 /**
432  * Add a verbs item specification into @p verbs.
433  *
434  * @param[out] verbs
435  *   Pointer to verbs structure.
436  * @param[in] src
437  *   Create specification.
438  * @param[in] size
439  *   Size in bytes of the specification to copy.
440  */
441 static void
442 flow_verbs_spec_add(struct mlx5_flow_verbs_workspace *verbs,
443 		    void *src, unsigned int size)
444 {
445 	void *dst;
446 
447 	if (!verbs)
448 		return;
449 	MLX5_ASSERT(verbs->specs);
450 	dst = (void *)(verbs->specs + verbs->size);
451 	memcpy(dst, src, size);
452 	++verbs->attr.num_of_specs;
453 	verbs->size += size;
454 }
455 
456 /**
457  * Convert the @p item into a Verbs specification. This function assumes that
458  * the input is valid and that there is space to insert the requested item
459  * into the flow.
460  *
461  * @param[in, out] dev_flow
462  *   Pointer to dev_flow structure.
463  * @param[in] item
464  *   Item specification.
465  * @param[in] item_flags
466  *   Parsed item flags.
467  */
468 static void
469 flow_verbs_translate_item_eth(struct mlx5_flow *dev_flow,
470 			      const struct rte_flow_item *item,
471 			      uint64_t item_flags)
472 {
473 	const struct rte_flow_item_eth *spec = item->spec;
474 	const struct rte_flow_item_eth *mask = item->mask;
475 	const unsigned int size = sizeof(struct ibv_flow_spec_eth);
476 	struct ibv_flow_spec_eth eth = {
477 		.type = IBV_FLOW_SPEC_ETH | VERBS_SPEC_INNER(item_flags),
478 		.size = size,
479 	};
480 
481 	if (!mask)
482 		mask = &rte_flow_item_eth_mask;
483 	if (spec) {
484 		unsigned int i;
485 
486 		memcpy(&eth.val.dst_mac, spec->dst.addr_bytes,
487 			RTE_ETHER_ADDR_LEN);
488 		memcpy(&eth.val.src_mac, spec->src.addr_bytes,
489 			RTE_ETHER_ADDR_LEN);
490 		eth.val.ether_type = spec->type;
491 		memcpy(&eth.mask.dst_mac, mask->dst.addr_bytes,
492 			RTE_ETHER_ADDR_LEN);
493 		memcpy(&eth.mask.src_mac, mask->src.addr_bytes,
494 			RTE_ETHER_ADDR_LEN);
495 		eth.mask.ether_type = mask->type;
496 		/* Remove unwanted bits from values. */
497 		for (i = 0; i < RTE_ETHER_ADDR_LEN; ++i) {
498 			eth.val.dst_mac[i] &= eth.mask.dst_mac[i];
499 			eth.val.src_mac[i] &= eth.mask.src_mac[i];
500 		}
501 		eth.val.ether_type &= eth.mask.ether_type;
502 	}
503 	flow_verbs_spec_add(&dev_flow->verbs, &eth, size);
504 }
505 
506 /**
507  * Update the VLAN tag in the Verbs Ethernet specification.
508  * This function assumes that the input is valid and there is space to add
509  * the requested item.
510  *
511  * @param[in, out] attr
512  *   Pointer to Verbs attributes structure.
513  * @param[in] eth
514  *   Verbs structure containing the VLAN information to copy.
515  */
516 static void
517 flow_verbs_item_vlan_update(struct ibv_flow_attr *attr,
518 			    struct ibv_flow_spec_eth *eth)
519 {
520 	unsigned int i;
521 	const enum ibv_flow_spec_type search = eth->type;
522 	struct ibv_spec_header *hdr = (struct ibv_spec_header *)
523 		((uint8_t *)attr + sizeof(struct ibv_flow_attr));
524 
525 	for (i = 0; i != attr->num_of_specs; ++i) {
526 		if (hdr->type == search) {
527 			struct ibv_flow_spec_eth *e =
528 				(struct ibv_flow_spec_eth *)hdr;
529 
530 			e->val.vlan_tag = eth->val.vlan_tag;
531 			e->mask.vlan_tag = eth->mask.vlan_tag;
532 			e->val.ether_type = eth->val.ether_type;
533 			e->mask.ether_type = eth->mask.ether_type;
534 			break;
535 		}
536 		hdr = (struct ibv_spec_header *)((uint8_t *)hdr + hdr->size);
537 	}
538 }
539 
540 /**
541  * Convert the @p item into a Verbs specification. This function assumes that
542  * the input is valid and that there is space to insert the requested item
543  * into the flow.
544  *
545  * @param[in, out] dev_flow
546  *   Pointer to dev_flow structure.
547  * @param[in] item
548  *   Item specification.
549  * @param[in] item_flags
550  *   Parsed item flags.
551  */
552 static void
553 flow_verbs_translate_item_vlan(struct mlx5_flow *dev_flow,
554 			       const struct rte_flow_item *item,
555 			       uint64_t item_flags)
556 {
557 	const struct rte_flow_item_vlan *spec = item->spec;
558 	const struct rte_flow_item_vlan *mask = item->mask;
559 	unsigned int size = sizeof(struct ibv_flow_spec_eth);
560 	const int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
561 	struct ibv_flow_spec_eth eth = {
562 		.type = IBV_FLOW_SPEC_ETH | VERBS_SPEC_INNER(item_flags),
563 		.size = size,
564 	};
565 	const uint32_t l2m = tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
566 				      MLX5_FLOW_LAYER_OUTER_L2;
567 
568 	if (!mask)
569 		mask = &rte_flow_item_vlan_mask;
570 	if (spec) {
571 		eth.val.vlan_tag = spec->tci;
572 		eth.mask.vlan_tag = mask->tci;
573 		eth.val.vlan_tag &= eth.mask.vlan_tag;
574 		eth.val.ether_type = spec->inner_type;
575 		eth.mask.ether_type = mask->inner_type;
576 		eth.val.ether_type &= eth.mask.ether_type;
577 	}
578 	if (!(item_flags & l2m))
579 		flow_verbs_spec_add(&dev_flow->verbs, &eth, size);
580 	else
581 		flow_verbs_item_vlan_update(&dev_flow->verbs.attr, &eth);
582 	if (!tunnel)
583 		dev_flow->handle->vf_vlan.tag =
584 			rte_be_to_cpu_16(spec->tci) & 0x0fff;
585 }
586 
587 /**
588  * Convert the @p item into a Verbs specification. This function assumes that
589  * the input is valid and that there is space to insert the requested item
590  * into the flow.
591  *
592  * @param[in, out] dev_flow
593  *   Pointer to dev_flow structure.
594  * @param[in] item
595  *   Item specification.
596  * @param[in] item_flags
597  *   Parsed item flags.
598  */
599 static void
600 flow_verbs_translate_item_ipv4(struct mlx5_flow *dev_flow,
601 			       const struct rte_flow_item *item,
602 			       uint64_t item_flags)
603 {
604 	const struct rte_flow_item_ipv4 *spec = item->spec;
605 	const struct rte_flow_item_ipv4 *mask = item->mask;
606 	unsigned int size = sizeof(struct ibv_flow_spec_ipv4_ext);
607 	struct ibv_flow_spec_ipv4_ext ipv4 = {
608 		.type = IBV_FLOW_SPEC_IPV4_EXT | VERBS_SPEC_INNER(item_flags),
609 		.size = size,
610 	};
611 
612 	if (!mask)
613 		mask = &rte_flow_item_ipv4_mask;
614 	if (spec) {
615 		ipv4.val = (struct ibv_flow_ipv4_ext_filter){
616 			.src_ip = spec->hdr.src_addr,
617 			.dst_ip = spec->hdr.dst_addr,
618 			.proto = spec->hdr.next_proto_id,
619 			.tos = spec->hdr.type_of_service,
620 		};
621 		ipv4.mask = (struct ibv_flow_ipv4_ext_filter){
622 			.src_ip = mask->hdr.src_addr,
623 			.dst_ip = mask->hdr.dst_addr,
624 			.proto = mask->hdr.next_proto_id,
625 			.tos = mask->hdr.type_of_service,
626 		};
627 		/* Remove unwanted bits from values. */
628 		ipv4.val.src_ip &= ipv4.mask.src_ip;
629 		ipv4.val.dst_ip &= ipv4.mask.dst_ip;
630 		ipv4.val.proto &= ipv4.mask.proto;
631 		ipv4.val.tos &= ipv4.mask.tos;
632 	}
633 	flow_verbs_spec_add(&dev_flow->verbs, &ipv4, size);
634 }
635 
636 /**
637  * Convert the @p item into a Verbs specification. This function assumes that
638  * the input is valid and that there is space to insert the requested item
639  * into the flow.
640  *
641  * @param[in, out] dev_flow
642  *   Pointer to dev_flow structure.
643  * @param[in] item
644  *   Item specification.
645  * @param[in] item_flags
646  *   Parsed item flags.
647  */
648 static void
649 flow_verbs_translate_item_ipv6(struct mlx5_flow *dev_flow,
650 			       const struct rte_flow_item *item,
651 			       uint64_t item_flags)
652 {
653 	const struct rte_flow_item_ipv6 *spec = item->spec;
654 	const struct rte_flow_item_ipv6 *mask = item->mask;
655 	unsigned int size = sizeof(struct ibv_flow_spec_ipv6);
656 	struct ibv_flow_spec_ipv6 ipv6 = {
657 		.type = IBV_FLOW_SPEC_IPV6 | VERBS_SPEC_INNER(item_flags),
658 		.size = size,
659 	};
660 
661 	if (!mask)
662 		mask = &rte_flow_item_ipv6_mask;
663 	if (spec) {
664 		unsigned int i;
665 		uint32_t vtc_flow_val;
666 		uint32_t vtc_flow_mask;
667 
668 		memcpy(&ipv6.val.src_ip, spec->hdr.src_addr,
669 		       RTE_DIM(ipv6.val.src_ip));
670 		memcpy(&ipv6.val.dst_ip, spec->hdr.dst_addr,
671 		       RTE_DIM(ipv6.val.dst_ip));
672 		memcpy(&ipv6.mask.src_ip, mask->hdr.src_addr,
673 		       RTE_DIM(ipv6.mask.src_ip));
674 		memcpy(&ipv6.mask.dst_ip, mask->hdr.dst_addr,
675 		       RTE_DIM(ipv6.mask.dst_ip));
676 		vtc_flow_val = rte_be_to_cpu_32(spec->hdr.vtc_flow);
677 		vtc_flow_mask = rte_be_to_cpu_32(mask->hdr.vtc_flow);
678 		ipv6.val.flow_label =
679 			rte_cpu_to_be_32((vtc_flow_val & RTE_IPV6_HDR_FL_MASK) >>
680 					 RTE_IPV6_HDR_FL_SHIFT);
681 		ipv6.val.traffic_class = (vtc_flow_val & RTE_IPV6_HDR_TC_MASK) >>
682 					 RTE_IPV6_HDR_TC_SHIFT;
683 		ipv6.val.next_hdr = spec->hdr.proto;
684 		ipv6.mask.flow_label =
685 			rte_cpu_to_be_32((vtc_flow_mask & RTE_IPV6_HDR_FL_MASK) >>
686 					 RTE_IPV6_HDR_FL_SHIFT);
687 		ipv6.mask.traffic_class = (vtc_flow_mask & RTE_IPV6_HDR_TC_MASK) >>
688 					  RTE_IPV6_HDR_TC_SHIFT;
689 		ipv6.mask.next_hdr = mask->hdr.proto;
690 		/* Remove unwanted bits from values. */
691 		for (i = 0; i < RTE_DIM(ipv6.val.src_ip); ++i) {
692 			ipv6.val.src_ip[i] &= ipv6.mask.src_ip[i];
693 			ipv6.val.dst_ip[i] &= ipv6.mask.dst_ip[i];
694 		}
695 		ipv6.val.flow_label &= ipv6.mask.flow_label;
696 		ipv6.val.traffic_class &= ipv6.mask.traffic_class;
697 		ipv6.val.next_hdr &= ipv6.mask.next_hdr;
698 	}
699 	flow_verbs_spec_add(&dev_flow->verbs, &ipv6, size);
700 }
701 
702 /**
703  * Convert the @p item into a Verbs specification. This function assumes that
704  * the input is valid and that there is space to insert the requested item
705  * into the flow.
706  *
707  * @param[in, out] dev_flow
708  *   Pointer to dev_flow structure.
709  * @param[in] item
710  *   Item specification.
711  * @param[in] item_flags
712  *   Parsed item flags.
713  */
714 static void
715 flow_verbs_translate_item_tcp(struct mlx5_flow *dev_flow,
716 			      const struct rte_flow_item *item,
717 			      uint64_t item_flags __rte_unused)
718 {
719 	const struct rte_flow_item_tcp *spec = item->spec;
720 	const struct rte_flow_item_tcp *mask = item->mask;
721 	unsigned int size = sizeof(struct ibv_flow_spec_tcp_udp);
722 	struct ibv_flow_spec_tcp_udp tcp = {
723 		.type = IBV_FLOW_SPEC_TCP | VERBS_SPEC_INNER(item_flags),
724 		.size = size,
725 	};
726 
727 	if (!mask)
728 		mask = &rte_flow_item_tcp_mask;
729 	if (spec) {
730 		tcp.val.dst_port = spec->hdr.dst_port;
731 		tcp.val.src_port = spec->hdr.src_port;
732 		tcp.mask.dst_port = mask->hdr.dst_port;
733 		tcp.mask.src_port = mask->hdr.src_port;
734 		/* Remove unwanted bits from values. */
735 		tcp.val.src_port &= tcp.mask.src_port;
736 		tcp.val.dst_port &= tcp.mask.dst_port;
737 	}
738 	flow_verbs_spec_add(&dev_flow->verbs, &tcp, size);
739 }
740 
741 /**
742  * Convert the @p item into a Verbs specification. This function assumes that
743  * the input is valid and that there is space to insert the requested item
744  * into the flow.
745  *
746  * @param[in, out] dev_flow
747  *   Pointer to dev_flow structure.
748  * @param[in] item
749  *   Item specification.
750  * @param[in] item_flags
751  *   Parsed item flags.
752  */
753 static void
754 flow_verbs_translate_item_udp(struct mlx5_flow *dev_flow,
755 			      const struct rte_flow_item *item,
756 			      uint64_t item_flags __rte_unused)
757 {
758 	const struct rte_flow_item_udp *spec = item->spec;
759 	const struct rte_flow_item_udp *mask = item->mask;
760 	unsigned int size = sizeof(struct ibv_flow_spec_tcp_udp);
761 	struct ibv_flow_spec_tcp_udp udp = {
762 		.type = IBV_FLOW_SPEC_UDP | VERBS_SPEC_INNER(item_flags),
763 		.size = size,
764 	};
765 
766 	if (!mask)
767 		mask = &rte_flow_item_udp_mask;
768 	if (spec) {
769 		udp.val.dst_port = spec->hdr.dst_port;
770 		udp.val.src_port = spec->hdr.src_port;
771 		udp.mask.dst_port = mask->hdr.dst_port;
772 		udp.mask.src_port = mask->hdr.src_port;
773 		/* Remove unwanted bits from values. */
774 		udp.val.src_port &= udp.mask.src_port;
775 		udp.val.dst_port &= udp.mask.dst_port;
776 	}
777 	item++;
778 	while (item->type == RTE_FLOW_ITEM_TYPE_VOID)
779 		item++;
780 	if (!(udp.val.dst_port & udp.mask.dst_port)) {
781 		switch ((item)->type) {
782 		case RTE_FLOW_ITEM_TYPE_VXLAN:
783 			udp.val.dst_port = htons(MLX5_UDP_PORT_VXLAN);
784 			udp.mask.dst_port = 0xffff;
785 			break;
786 		case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
787 			udp.val.dst_port = htons(MLX5_UDP_PORT_VXLAN_GPE);
788 			udp.mask.dst_port = 0xffff;
789 			break;
790 		case RTE_FLOW_ITEM_TYPE_MPLS:
791 			udp.val.dst_port = htons(MLX5_UDP_PORT_MPLS);
792 			udp.mask.dst_port = 0xffff;
793 			break;
794 		default:
795 			break;
796 		}
797 	}
798 
799 	flow_verbs_spec_add(&dev_flow->verbs, &udp, size);
800 }
801 
802 /**
803  * Convert the @p item into a Verbs specification. This function assumes that
804  * the input is valid and that there is space to insert the requested item
805  * into the flow.
806  *
807  * @param[in, out] dev_flow
808  *   Pointer to dev_flow structure.
809  * @param[in] item
810  *   Item specification.
811  * @param[in] item_flags
812  *   Parsed item flags.
813  */
814 static void
815 flow_verbs_translate_item_vxlan(struct mlx5_flow *dev_flow,
816 				const struct rte_flow_item *item,
817 				uint64_t item_flags __rte_unused)
818 {
819 	const struct rte_flow_item_vxlan *spec = item->spec;
820 	const struct rte_flow_item_vxlan *mask = item->mask;
821 	unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
822 	struct ibv_flow_spec_tunnel vxlan = {
823 		.type = IBV_FLOW_SPEC_VXLAN_TUNNEL,
824 		.size = size,
825 	};
826 	union vni {
827 		uint32_t vlan_id;
828 		uint8_t vni[4];
829 	} id = { .vlan_id = 0, };
830 
831 	if (!mask)
832 		mask = &rte_flow_item_vxlan_mask;
833 	if (spec) {
834 		memcpy(&id.vni[1], spec->vni, 3);
835 		vxlan.val.tunnel_id = id.vlan_id;
836 		memcpy(&id.vni[1], mask->vni, 3);
837 		vxlan.mask.tunnel_id = id.vlan_id;
838 		/* Remove unwanted bits from values. */
839 		vxlan.val.tunnel_id &= vxlan.mask.tunnel_id;
840 	}
841 	flow_verbs_spec_add(&dev_flow->verbs, &vxlan, size);
842 }
843 
844 /**
845  * Convert the @p item into a Verbs specification. This function assumes that
846  * the input is valid and that there is space to insert the requested item
847  * into the flow.
848  *
849  * @param[in, out] dev_flow
850  *   Pointer to dev_flow structure.
851  * @param[in] item
852  *   Item specification.
853  * @param[in] item_flags
854  *   Parsed item flags.
855  */
856 static void
857 flow_verbs_translate_item_vxlan_gpe(struct mlx5_flow *dev_flow,
858 				    const struct rte_flow_item *item,
859 				    uint64_t item_flags __rte_unused)
860 {
861 	const struct rte_flow_item_vxlan_gpe *spec = item->spec;
862 	const struct rte_flow_item_vxlan_gpe *mask = item->mask;
863 	unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
864 	struct ibv_flow_spec_tunnel vxlan_gpe = {
865 		.type = IBV_FLOW_SPEC_VXLAN_TUNNEL,
866 		.size = size,
867 	};
868 	union vni {
869 		uint32_t vlan_id;
870 		uint8_t vni[4];
871 	} id = { .vlan_id = 0, };
872 
873 	if (!mask)
874 		mask = &rte_flow_item_vxlan_gpe_mask;
875 	if (spec) {
876 		memcpy(&id.vni[1], spec->vni, 3);
877 		vxlan_gpe.val.tunnel_id = id.vlan_id;
878 		memcpy(&id.vni[1], mask->vni, 3);
879 		vxlan_gpe.mask.tunnel_id = id.vlan_id;
880 		/* Remove unwanted bits from values. */
881 		vxlan_gpe.val.tunnel_id &= vxlan_gpe.mask.tunnel_id;
882 	}
883 	flow_verbs_spec_add(&dev_flow->verbs, &vxlan_gpe, size);
884 }
885 
886 /**
887  * Update the protocol in Verbs IPv4/IPv6 spec.
888  *
889  * @param[in, out] attr
890  *   Pointer to Verbs attributes structure.
891  * @param[in] search
892  *   Specification type to search in order to update the IP protocol.
893  * @param[in] protocol
894  *   Protocol value to set if none is present in the specification.
895  */
896 static void
897 flow_verbs_item_gre_ip_protocol_update(struct ibv_flow_attr *attr,
898 				       enum ibv_flow_spec_type search,
899 				       uint8_t protocol)
900 {
901 	unsigned int i;
902 	struct ibv_spec_header *hdr = (struct ibv_spec_header *)
903 		((uint8_t *)attr + sizeof(struct ibv_flow_attr));
904 
905 	if (!attr)
906 		return;
907 	for (i = 0; i != attr->num_of_specs; ++i) {
908 		if (hdr->type == search) {
909 			union {
910 				struct ibv_flow_spec_ipv4_ext *ipv4;
911 				struct ibv_flow_spec_ipv6 *ipv6;
912 			} ip;
913 
914 			switch (search) {
915 			case IBV_FLOW_SPEC_IPV4_EXT:
916 				ip.ipv4 = (struct ibv_flow_spec_ipv4_ext *)hdr;
917 				if (!ip.ipv4->val.proto) {
918 					ip.ipv4->val.proto = protocol;
919 					ip.ipv4->mask.proto = 0xff;
920 				}
921 				break;
922 			case IBV_FLOW_SPEC_IPV6:
923 				ip.ipv6 = (struct ibv_flow_spec_ipv6 *)hdr;
924 				if (!ip.ipv6->val.next_hdr) {
925 					ip.ipv6->val.next_hdr = protocol;
926 					ip.ipv6->mask.next_hdr = 0xff;
927 				}
928 				break;
929 			default:
930 				break;
931 			}
932 			break;
933 		}
934 		hdr = (struct ibv_spec_header *)((uint8_t *)hdr + hdr->size);
935 	}
936 }
937 
938 /**
939  * Convert the @p item into a Verbs specification. This function assumes that
940  * the input is valid and that there is space to insert the requested item
941  * into the flow.
942  *
943  * @param[in, out] dev_flow
944  *   Pointer to dev_flow structure.
945  * @param[in] item
946  *   Item specification.
947  * @param[in] item_flags
948  *   Parsed item flags.
949  */
950 static void
951 flow_verbs_translate_item_gre(struct mlx5_flow *dev_flow,
952 			      const struct rte_flow_item *item __rte_unused,
953 			      uint64_t item_flags)
954 {
955 	struct mlx5_flow_verbs_workspace *verbs = &dev_flow->verbs;
956 #ifndef HAVE_IBV_DEVICE_MPLS_SUPPORT
957 	unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
958 	struct ibv_flow_spec_tunnel tunnel = {
959 		.type = IBV_FLOW_SPEC_VXLAN_TUNNEL,
960 		.size = size,
961 	};
962 #else
963 	const struct rte_flow_item_gre *spec = item->spec;
964 	const struct rte_flow_item_gre *mask = item->mask;
965 	unsigned int size = sizeof(struct ibv_flow_spec_gre);
966 	struct ibv_flow_spec_gre tunnel = {
967 		.type = IBV_FLOW_SPEC_GRE,
968 		.size = size,
969 	};
970 
971 	if (!mask)
972 		mask = &rte_flow_item_gre_mask;
973 	if (spec) {
974 		tunnel.val.c_ks_res0_ver = spec->c_rsvd0_ver;
975 		tunnel.val.protocol = spec->protocol;
976 		tunnel.mask.c_ks_res0_ver = mask->c_rsvd0_ver;
977 		tunnel.mask.protocol = mask->protocol;
978 		/* Remove unwanted bits from values. */
979 		tunnel.val.c_ks_res0_ver &= tunnel.mask.c_ks_res0_ver;
980 		tunnel.val.protocol &= tunnel.mask.protocol;
981 		tunnel.val.key &= tunnel.mask.key;
982 	}
983 #endif
984 	if (item_flags & MLX5_FLOW_LAYER_OUTER_L3_IPV4)
985 		flow_verbs_item_gre_ip_protocol_update(&verbs->attr,
986 						       IBV_FLOW_SPEC_IPV4_EXT,
987 						       IPPROTO_GRE);
988 	else
989 		flow_verbs_item_gre_ip_protocol_update(&verbs->attr,
990 						       IBV_FLOW_SPEC_IPV6,
991 						       IPPROTO_GRE);
992 	flow_verbs_spec_add(verbs, &tunnel, size);
993 }
994 
995 /**
996  * Convert the @p action into a Verbs specification. This function assumes that
997  * the input is valid and that there is space to insert the requested action
998  * into the flow. This function also return the action that was added.
999  *
1000  * @param[in, out] dev_flow
1001  *   Pointer to dev_flow structure.
1002  * @param[in] item
1003  *   Item specification.
1004  * @param[in] item_flags
1005  *   Parsed item flags.
1006  */
1007 static void
1008 flow_verbs_translate_item_mpls(struct mlx5_flow *dev_flow __rte_unused,
1009 			       const struct rte_flow_item *item __rte_unused,
1010 			       uint64_t item_flags __rte_unused)
1011 {
1012 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1013 	const struct rte_flow_item_mpls *spec = item->spec;
1014 	const struct rte_flow_item_mpls *mask = item->mask;
1015 	unsigned int size = sizeof(struct ibv_flow_spec_mpls);
1016 	struct ibv_flow_spec_mpls mpls = {
1017 		.type = IBV_FLOW_SPEC_MPLS,
1018 		.size = size,
1019 	};
1020 
1021 	if (!mask)
1022 		mask = &rte_flow_item_mpls_mask;
1023 	if (spec) {
1024 		memcpy(&mpls.val.label, spec, sizeof(mpls.val.label));
1025 		memcpy(&mpls.mask.label, mask, sizeof(mpls.mask.label));
1026 		/* Remove unwanted bits from values.  */
1027 		mpls.val.label &= mpls.mask.label;
1028 	}
1029 	flow_verbs_spec_add(&dev_flow->verbs, &mpls, size);
1030 #endif
1031 }
1032 
1033 /**
1034  * Convert the @p action into a Verbs specification. This function assumes that
1035  * the input is valid and that there is space to insert the requested action
1036  * into the flow.
1037  *
1038  * @param[in] dev_flow
1039  *   Pointer to mlx5_flow.
1040  * @param[in] action
1041  *   Action configuration.
1042  */
1043 static void
1044 flow_verbs_translate_action_drop
1045 	(struct mlx5_flow *dev_flow,
1046 	 const struct rte_flow_action *action __rte_unused)
1047 {
1048 	unsigned int size = sizeof(struct ibv_flow_spec_action_drop);
1049 	struct ibv_flow_spec_action_drop drop = {
1050 			.type = IBV_FLOW_SPEC_ACTION_DROP,
1051 			.size = size,
1052 	};
1053 
1054 	flow_verbs_spec_add(&dev_flow->verbs, &drop, size);
1055 }
1056 
1057 /**
1058  * Convert the @p action into a Verbs specification. This function assumes that
1059  * the input is valid and that there is space to insert the requested action
1060  * into the flow.
1061  *
1062  * @param[in] rss_desc
1063  *   Pointer to mlx5_flow_rss_desc.
1064  * @param[in] action
1065  *   Action configuration.
1066  */
1067 static void
1068 flow_verbs_translate_action_queue(struct mlx5_flow_rss_desc *rss_desc,
1069 				  const struct rte_flow_action *action)
1070 {
1071 	const struct rte_flow_action_queue *queue = action->conf;
1072 
1073 	rss_desc->queue[0] = queue->index;
1074 	rss_desc->queue_num = 1;
1075 }
1076 
1077 /**
1078  * Convert the @p action into a Verbs specification. This function assumes that
1079  * the input is valid and that there is space to insert the requested action
1080  * into the flow.
1081  *
1082  * @param[in] rss_desc
1083  *   Pointer to mlx5_flow_rss_desc.
1084  * @param[in] action
1085  *   Action configuration.
1086  */
1087 static void
1088 flow_verbs_translate_action_rss(struct mlx5_flow_rss_desc *rss_desc,
1089 				const struct rte_flow_action *action)
1090 {
1091 	const struct rte_flow_action_rss *rss = action->conf;
1092 	const uint8_t *rss_key;
1093 
1094 	memcpy(rss_desc->queue, rss->queue, rss->queue_num * sizeof(uint16_t));
1095 	rss_desc->queue_num = rss->queue_num;
1096 	/* NULL RSS key indicates default RSS key. */
1097 	rss_key = !rss->key ? rss_hash_default_key : rss->key;
1098 	memcpy(rss_desc->key, rss_key, MLX5_RSS_HASH_KEY_LEN);
1099 	/*
1100 	 * rss->level and rss.types should be set in advance when expanding
1101 	 * items for RSS.
1102 	 */
1103 }
1104 
1105 /**
1106  * Convert the @p action into a Verbs specification. This function assumes that
1107  * the input is valid and that there is space to insert the requested action
1108  * into the flow.
1109  *
1110  * @param[in] dev_flow
1111  *   Pointer to mlx5_flow.
1112  * @param[in] action
1113  *   Action configuration.
1114  */
1115 static void
1116 flow_verbs_translate_action_flag
1117 	(struct mlx5_flow *dev_flow,
1118 	 const struct rte_flow_action *action __rte_unused)
1119 {
1120 	unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
1121 	struct ibv_flow_spec_action_tag tag = {
1122 		.type = IBV_FLOW_SPEC_ACTION_TAG,
1123 		.size = size,
1124 		.tag_id = mlx5_flow_mark_set(MLX5_FLOW_MARK_DEFAULT),
1125 	};
1126 
1127 	flow_verbs_spec_add(&dev_flow->verbs, &tag, size);
1128 }
1129 
1130 /**
1131  * Convert the @p action into a Verbs specification. This function assumes that
1132  * the input is valid and that there is space to insert the requested action
1133  * into the flow.
1134  *
1135  * @param[in] dev_flow
1136  *   Pointer to mlx5_flow.
1137  * @param[in] action
1138  *   Action configuration.
1139  */
1140 static void
1141 flow_verbs_translate_action_mark(struct mlx5_flow *dev_flow,
1142 				 const struct rte_flow_action *action)
1143 {
1144 	const struct rte_flow_action_mark *mark = action->conf;
1145 	unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
1146 	struct ibv_flow_spec_action_tag tag = {
1147 		.type = IBV_FLOW_SPEC_ACTION_TAG,
1148 		.size = size,
1149 		.tag_id = mlx5_flow_mark_set(mark->id),
1150 	};
1151 
1152 	flow_verbs_spec_add(&dev_flow->verbs, &tag, size);
1153 }
1154 
1155 /**
1156  * Convert the @p action into a Verbs specification. This function assumes that
1157  * the input is valid and that there is space to insert the requested action
1158  * into the flow.
1159  *
1160  * @param[in] dev
1161  *   Pointer to the Ethernet device structure.
1162  * @param[in] action
1163  *   Action configuration.
1164  * @param[in] dev_flow
1165  *   Pointer to mlx5_flow.
1166  * @param[out] error
1167  *   Pointer to error structure.
1168  *
1169  * @return
1170  *   0 On success else a negative errno value is returned and rte_errno is set.
1171  */
1172 static int
1173 flow_verbs_translate_action_count(struct mlx5_flow *dev_flow,
1174 				  const struct rte_flow_action *action,
1175 				  struct rte_eth_dev *dev,
1176 				  struct rte_flow_error *error)
1177 {
1178 	const struct rte_flow_action_count *count = action->conf;
1179 	struct rte_flow *flow = dev_flow->flow;
1180 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) || \
1181 	defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1182 	struct mlx5_flow_counter_pool *pool;
1183 	struct mlx5_flow_counter *cnt = NULL;
1184 	unsigned int size = sizeof(struct ibv_flow_spec_counter_action);
1185 	struct ibv_flow_spec_counter_action counter = {
1186 		.type = IBV_FLOW_SPEC_ACTION_COUNT,
1187 		.size = size,
1188 	};
1189 #endif
1190 
1191 	if (!flow->counter) {
1192 		flow->counter = flow_verbs_counter_new(dev, count->id);
1193 		if (!flow->counter)
1194 			return rte_flow_error_set(error, rte_errno,
1195 						  RTE_FLOW_ERROR_TYPE_ACTION,
1196 						  action,
1197 						  "cannot get counter"
1198 						  " context.");
1199 	}
1200 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42)
1201 	cnt = flow_verbs_counter_get_by_idx(dev, flow->counter, &pool);
1202 	counter.counter_set_handle =
1203 		((struct ibv_counter_set *)cnt->dcs_when_active)->handle;
1204 	flow_verbs_spec_add(&dev_flow->verbs, &counter, size);
1205 #elif defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1206 	cnt = flow_verbs_counter_get_by_idx(dev, flow->counter, &pool);
1207 	counter.counters = (struct ibv_counters *)cnt->dcs_when_active;
1208 	flow_verbs_spec_add(&dev_flow->verbs, &counter, size);
1209 #endif
1210 	return 0;
1211 }
1212 
1213 /**
1214  * Internal validation function. For validating both actions and items.
1215  *
1216  * @param[in] dev
1217  *   Pointer to the Ethernet device structure.
1218  * @param[in] attr
1219  *   Pointer to the flow attributes.
1220  * @param[in] items
1221  *   Pointer to the list of items.
1222  * @param[in] actions
1223  *   Pointer to the list of actions.
1224  * @param[in] external
1225  *   This flow rule is created by request external to PMD.
1226  * @param[in] hairpin
1227  *   Number of hairpin TX actions, 0 means classic flow.
1228  * @param[out] error
1229  *   Pointer to the error structure.
1230  *
1231  * @return
1232  *   0 on success, a negative errno value otherwise and rte_errno is set.
1233  */
1234 static int
1235 flow_verbs_validate(struct rte_eth_dev *dev,
1236 		    const struct rte_flow_attr *attr,
1237 		    const struct rte_flow_item items[],
1238 		    const struct rte_flow_action actions[],
1239 		    bool external __rte_unused,
1240 		    int hairpin __rte_unused,
1241 		    struct rte_flow_error *error)
1242 {
1243 	int ret;
1244 	uint64_t action_flags = 0;
1245 	uint64_t item_flags = 0;
1246 	uint64_t last_item = 0;
1247 	uint8_t next_protocol = 0xff;
1248 	uint16_t ether_type = 0;
1249 	bool is_empty_vlan = false;
1250 	uint16_t udp_dport = 0;
1251 
1252 	if (items == NULL)
1253 		return -1;
1254 	ret = mlx5_flow_validate_attributes(dev, attr, error);
1255 	if (ret < 0)
1256 		return ret;
1257 	for (; items->type != RTE_FLOW_ITEM_TYPE_END; items++) {
1258 		int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
1259 		int ret = 0;
1260 
1261 		switch (items->type) {
1262 		case RTE_FLOW_ITEM_TYPE_VOID:
1263 			break;
1264 		case RTE_FLOW_ITEM_TYPE_ETH:
1265 			ret = mlx5_flow_validate_item_eth(items, item_flags,
1266 							  false, error);
1267 			if (ret < 0)
1268 				return ret;
1269 			last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
1270 					     MLX5_FLOW_LAYER_OUTER_L2;
1271 			if (items->mask != NULL && items->spec != NULL) {
1272 				ether_type =
1273 					((const struct rte_flow_item_eth *)
1274 					 items->spec)->type;
1275 				ether_type &=
1276 					((const struct rte_flow_item_eth *)
1277 					 items->mask)->type;
1278 				if (ether_type == RTE_BE16(RTE_ETHER_TYPE_VLAN))
1279 					is_empty_vlan = true;
1280 				ether_type = rte_be_to_cpu_16(ether_type);
1281 			} else {
1282 				ether_type = 0;
1283 			}
1284 			break;
1285 		case RTE_FLOW_ITEM_TYPE_VLAN:
1286 			ret = mlx5_flow_validate_item_vlan(items, item_flags,
1287 							   dev, error);
1288 			if (ret < 0)
1289 				return ret;
1290 			last_item = tunnel ? (MLX5_FLOW_LAYER_INNER_L2 |
1291 					      MLX5_FLOW_LAYER_INNER_VLAN) :
1292 					     (MLX5_FLOW_LAYER_OUTER_L2 |
1293 					      MLX5_FLOW_LAYER_OUTER_VLAN);
1294 			if (items->mask != NULL && items->spec != NULL) {
1295 				ether_type =
1296 					((const struct rte_flow_item_vlan *)
1297 					 items->spec)->inner_type;
1298 				ether_type &=
1299 					((const struct rte_flow_item_vlan *)
1300 					 items->mask)->inner_type;
1301 				ether_type = rte_be_to_cpu_16(ether_type);
1302 			} else {
1303 				ether_type = 0;
1304 			}
1305 			is_empty_vlan = false;
1306 			break;
1307 		case RTE_FLOW_ITEM_TYPE_IPV4:
1308 			ret = mlx5_flow_validate_item_ipv4
1309 						(items, item_flags,
1310 						 last_item, ether_type, NULL,
1311 						 MLX5_ITEM_RANGE_NOT_ACCEPTED,
1312 						 error);
1313 			if (ret < 0)
1314 				return ret;
1315 			last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV4 :
1316 					     MLX5_FLOW_LAYER_OUTER_L3_IPV4;
1317 			if (items->mask != NULL &&
1318 			    ((const struct rte_flow_item_ipv4 *)
1319 			     items->mask)->hdr.next_proto_id) {
1320 				next_protocol =
1321 					((const struct rte_flow_item_ipv4 *)
1322 					 (items->spec))->hdr.next_proto_id;
1323 				next_protocol &=
1324 					((const struct rte_flow_item_ipv4 *)
1325 					 (items->mask))->hdr.next_proto_id;
1326 			} else {
1327 				/* Reset for inner layer. */
1328 				next_protocol = 0xff;
1329 			}
1330 			break;
1331 		case RTE_FLOW_ITEM_TYPE_IPV6:
1332 			ret = mlx5_flow_validate_item_ipv6(items, item_flags,
1333 							   last_item,
1334 							   ether_type, NULL,
1335 							   error);
1336 			if (ret < 0)
1337 				return ret;
1338 			last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV6 :
1339 					     MLX5_FLOW_LAYER_OUTER_L3_IPV6;
1340 			if (items->mask != NULL &&
1341 			    ((const struct rte_flow_item_ipv6 *)
1342 			     items->mask)->hdr.proto) {
1343 				next_protocol =
1344 					((const struct rte_flow_item_ipv6 *)
1345 					 items->spec)->hdr.proto;
1346 				next_protocol &=
1347 					((const struct rte_flow_item_ipv6 *)
1348 					 items->mask)->hdr.proto;
1349 			} else {
1350 				/* Reset for inner layer. */
1351 				next_protocol = 0xff;
1352 			}
1353 			break;
1354 		case RTE_FLOW_ITEM_TYPE_UDP:
1355 			ret = mlx5_flow_validate_item_udp(items, item_flags,
1356 							  next_protocol,
1357 							  error);
1358 			const struct rte_flow_item_udp *spec = items->spec;
1359 			const struct rte_flow_item_udp *mask = items->mask;
1360 			if (!mask)
1361 				mask = &rte_flow_item_udp_mask;
1362 			if (spec != NULL)
1363 				udp_dport = rte_be_to_cpu_16
1364 						(spec->hdr.dst_port &
1365 						 mask->hdr.dst_port);
1366 
1367 			if (ret < 0)
1368 				return ret;
1369 			last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L4_UDP :
1370 					     MLX5_FLOW_LAYER_OUTER_L4_UDP;
1371 			break;
1372 		case RTE_FLOW_ITEM_TYPE_TCP:
1373 			ret = mlx5_flow_validate_item_tcp
1374 						(items, item_flags,
1375 						 next_protocol,
1376 						 &rte_flow_item_tcp_mask,
1377 						 error);
1378 			if (ret < 0)
1379 				return ret;
1380 			last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L4_TCP :
1381 					     MLX5_FLOW_LAYER_OUTER_L4_TCP;
1382 			break;
1383 		case RTE_FLOW_ITEM_TYPE_VXLAN:
1384 			ret = mlx5_flow_validate_item_vxlan(dev, udp_dport,
1385 							    items, item_flags,
1386 							    attr, error);
1387 			if (ret < 0)
1388 				return ret;
1389 			last_item = MLX5_FLOW_LAYER_VXLAN;
1390 			break;
1391 		case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
1392 			ret = mlx5_flow_validate_item_vxlan_gpe(items,
1393 								item_flags,
1394 								dev, error);
1395 			if (ret < 0)
1396 				return ret;
1397 			last_item = MLX5_FLOW_LAYER_VXLAN_GPE;
1398 			break;
1399 		case RTE_FLOW_ITEM_TYPE_GRE:
1400 			ret = mlx5_flow_validate_item_gre(items, item_flags,
1401 							  next_protocol, error);
1402 			if (ret < 0)
1403 				return ret;
1404 			last_item = MLX5_FLOW_LAYER_GRE;
1405 			break;
1406 		case RTE_FLOW_ITEM_TYPE_MPLS:
1407 			ret = mlx5_flow_validate_item_mpls(dev, items,
1408 							   item_flags,
1409 							   last_item, error);
1410 			if (ret < 0)
1411 				return ret;
1412 			last_item = MLX5_FLOW_LAYER_MPLS;
1413 			break;
1414 		case RTE_FLOW_ITEM_TYPE_ICMP:
1415 		case RTE_FLOW_ITEM_TYPE_ICMP6:
1416 			return rte_flow_error_set(error, ENOTSUP,
1417 						  RTE_FLOW_ERROR_TYPE_ITEM,
1418 						  NULL, "ICMP/ICMP6 "
1419 						  "item not supported");
1420 		default:
1421 			return rte_flow_error_set(error, ENOTSUP,
1422 						  RTE_FLOW_ERROR_TYPE_ITEM,
1423 						  NULL, "item not supported");
1424 		}
1425 		item_flags |= last_item;
1426 	}
1427 	if (is_empty_vlan)
1428 		return rte_flow_error_set(error, ENOTSUP,
1429 						 RTE_FLOW_ERROR_TYPE_ITEM, NULL,
1430 		    "VLAN matching without vid specification is not supported");
1431 	for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
1432 		switch (actions->type) {
1433 		case RTE_FLOW_ACTION_TYPE_VOID:
1434 			break;
1435 		case RTE_FLOW_ACTION_TYPE_FLAG:
1436 			ret = mlx5_flow_validate_action_flag(action_flags,
1437 							     attr,
1438 							     error);
1439 			if (ret < 0)
1440 				return ret;
1441 			action_flags |= MLX5_FLOW_ACTION_FLAG;
1442 			break;
1443 		case RTE_FLOW_ACTION_TYPE_MARK:
1444 			ret = mlx5_flow_validate_action_mark(actions,
1445 							     action_flags,
1446 							     attr,
1447 							     error);
1448 			if (ret < 0)
1449 				return ret;
1450 			action_flags |= MLX5_FLOW_ACTION_MARK;
1451 			break;
1452 		case RTE_FLOW_ACTION_TYPE_DROP:
1453 			ret = mlx5_flow_validate_action_drop(action_flags,
1454 							     attr,
1455 							     error);
1456 			if (ret < 0)
1457 				return ret;
1458 			action_flags |= MLX5_FLOW_ACTION_DROP;
1459 			break;
1460 		case RTE_FLOW_ACTION_TYPE_QUEUE:
1461 			ret = mlx5_flow_validate_action_queue(actions,
1462 							      action_flags, dev,
1463 							      attr,
1464 							      error);
1465 			if (ret < 0)
1466 				return ret;
1467 			action_flags |= MLX5_FLOW_ACTION_QUEUE;
1468 			break;
1469 		case RTE_FLOW_ACTION_TYPE_RSS:
1470 			ret = mlx5_flow_validate_action_rss(actions,
1471 							    action_flags, dev,
1472 							    attr, item_flags,
1473 							    error);
1474 			if (ret < 0)
1475 				return ret;
1476 			action_flags |= MLX5_FLOW_ACTION_RSS;
1477 			break;
1478 		case RTE_FLOW_ACTION_TYPE_COUNT:
1479 			ret = mlx5_flow_validate_action_count(dev, attr, error);
1480 			if (ret < 0)
1481 				return ret;
1482 			action_flags |= MLX5_FLOW_ACTION_COUNT;
1483 			break;
1484 		default:
1485 			return rte_flow_error_set(error, ENOTSUP,
1486 						  RTE_FLOW_ERROR_TYPE_ACTION,
1487 						  actions,
1488 						  "action not supported");
1489 		}
1490 	}
1491 	/*
1492 	 * Validate the drop action mutual exclusion with other actions.
1493 	 * Drop action is mutually-exclusive with any other action, except for
1494 	 * Count action.
1495 	 */
1496 	if ((action_flags & MLX5_FLOW_ACTION_DROP) &&
1497 	    (action_flags & ~(MLX5_FLOW_ACTION_DROP | MLX5_FLOW_ACTION_COUNT)))
1498 		return rte_flow_error_set(error, EINVAL,
1499 					  RTE_FLOW_ERROR_TYPE_ACTION, NULL,
1500 					  "Drop action is mutually-exclusive "
1501 					  "with any other action, except for "
1502 					  "Count action");
1503 	if (!(action_flags & MLX5_FLOW_FATE_ACTIONS))
1504 		return rte_flow_error_set(error, EINVAL,
1505 					  RTE_FLOW_ERROR_TYPE_ACTION, actions,
1506 					  "no fate action is found");
1507 	return 0;
1508 }
1509 
1510 /**
1511  * Calculate the required bytes that are needed for the action part of the verbs
1512  * flow.
1513  *
1514  * @param[in] actions
1515  *   Pointer to the list of actions.
1516  *
1517  * @return
1518  *   The size of the memory needed for all actions.
1519  */
1520 static int
1521 flow_verbs_get_actions_size(const struct rte_flow_action actions[])
1522 {
1523 	int size = 0;
1524 
1525 	for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
1526 		switch (actions->type) {
1527 		case RTE_FLOW_ACTION_TYPE_VOID:
1528 			break;
1529 		case RTE_FLOW_ACTION_TYPE_FLAG:
1530 			size += sizeof(struct ibv_flow_spec_action_tag);
1531 			break;
1532 		case RTE_FLOW_ACTION_TYPE_MARK:
1533 			size += sizeof(struct ibv_flow_spec_action_tag);
1534 			break;
1535 		case RTE_FLOW_ACTION_TYPE_DROP:
1536 			size += sizeof(struct ibv_flow_spec_action_drop);
1537 			break;
1538 		case RTE_FLOW_ACTION_TYPE_QUEUE:
1539 			break;
1540 		case RTE_FLOW_ACTION_TYPE_RSS:
1541 			break;
1542 		case RTE_FLOW_ACTION_TYPE_COUNT:
1543 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) || \
1544 	defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1545 			size += sizeof(struct ibv_flow_spec_counter_action);
1546 #endif
1547 			break;
1548 		default:
1549 			break;
1550 		}
1551 	}
1552 	return size;
1553 }
1554 
1555 /**
1556  * Calculate the required bytes that are needed for the item part of the verbs
1557  * flow.
1558  *
1559  * @param[in] items
1560  *   Pointer to the list of items.
1561  *
1562  * @return
1563  *   The size of the memory needed for all items.
1564  */
1565 static int
1566 flow_verbs_get_items_size(const struct rte_flow_item items[])
1567 {
1568 	int size = 0;
1569 
1570 	for (; items->type != RTE_FLOW_ITEM_TYPE_END; items++) {
1571 		switch (items->type) {
1572 		case RTE_FLOW_ITEM_TYPE_VOID:
1573 			break;
1574 		case RTE_FLOW_ITEM_TYPE_ETH:
1575 			size += sizeof(struct ibv_flow_spec_eth);
1576 			break;
1577 		case RTE_FLOW_ITEM_TYPE_VLAN:
1578 			size += sizeof(struct ibv_flow_spec_eth);
1579 			break;
1580 		case RTE_FLOW_ITEM_TYPE_IPV4:
1581 			size += sizeof(struct ibv_flow_spec_ipv4_ext);
1582 			break;
1583 		case RTE_FLOW_ITEM_TYPE_IPV6:
1584 			size += sizeof(struct ibv_flow_spec_ipv6);
1585 			break;
1586 		case RTE_FLOW_ITEM_TYPE_UDP:
1587 			size += sizeof(struct ibv_flow_spec_tcp_udp);
1588 			break;
1589 		case RTE_FLOW_ITEM_TYPE_TCP:
1590 			size += sizeof(struct ibv_flow_spec_tcp_udp);
1591 			break;
1592 		case RTE_FLOW_ITEM_TYPE_VXLAN:
1593 			size += sizeof(struct ibv_flow_spec_tunnel);
1594 			break;
1595 		case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
1596 			size += sizeof(struct ibv_flow_spec_tunnel);
1597 			break;
1598 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1599 		case RTE_FLOW_ITEM_TYPE_GRE:
1600 			size += sizeof(struct ibv_flow_spec_gre);
1601 			break;
1602 		case RTE_FLOW_ITEM_TYPE_MPLS:
1603 			size += sizeof(struct ibv_flow_spec_mpls);
1604 			break;
1605 #else
1606 		case RTE_FLOW_ITEM_TYPE_GRE:
1607 			size += sizeof(struct ibv_flow_spec_tunnel);
1608 			break;
1609 #endif
1610 		default:
1611 			break;
1612 		}
1613 	}
1614 	return size;
1615 }
1616 
1617 /**
1618  * Internal preparation function. Allocate mlx5_flow with the required size.
1619  * The required size is calculate based on the actions and items. This function
1620  * also returns the detected actions and items for later use.
1621  *
1622  * @param[in] dev
1623  *   Pointer to Ethernet device.
1624  * @param[in] attr
1625  *   Pointer to the flow attributes.
1626  * @param[in] items
1627  *   Pointer to the list of items.
1628  * @param[in] actions
1629  *   Pointer to the list of actions.
1630  * @param[out] error
1631  *   Pointer to the error structure.
1632  *
1633  * @return
1634  *   Pointer to mlx5_flow object on success, otherwise NULL and rte_errno
1635  *   is set.
1636  */
1637 static struct mlx5_flow *
1638 flow_verbs_prepare(struct rte_eth_dev *dev,
1639 		   const struct rte_flow_attr *attr __rte_unused,
1640 		   const struct rte_flow_item items[],
1641 		   const struct rte_flow_action actions[],
1642 		   struct rte_flow_error *error)
1643 {
1644 	size_t size = 0;
1645 	uint32_t handle_idx = 0;
1646 	struct mlx5_flow *dev_flow;
1647 	struct mlx5_flow_handle *dev_handle;
1648 	struct mlx5_priv *priv = dev->data->dev_private;
1649 	struct mlx5_flow_workspace *wks = mlx5_flow_get_thread_workspace();
1650 
1651 	MLX5_ASSERT(wks);
1652 	size += flow_verbs_get_actions_size(actions);
1653 	size += flow_verbs_get_items_size(items);
1654 	if (size > MLX5_VERBS_MAX_SPEC_ACT_SIZE) {
1655 		rte_flow_error_set(error, E2BIG,
1656 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1657 				   "Verbs spec/action size too large");
1658 		return NULL;
1659 	}
1660 	/* In case of corrupting the memory. */
1661 	if (wks->flow_idx >= MLX5_NUM_MAX_DEV_FLOWS) {
1662 		rte_flow_error_set(error, ENOSPC,
1663 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1664 				   "not free temporary device flow");
1665 		return NULL;
1666 	}
1667 	dev_handle = mlx5_ipool_zmalloc(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW],
1668 				   &handle_idx);
1669 	if (!dev_handle) {
1670 		rte_flow_error_set(error, ENOMEM,
1671 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1672 				   "not enough memory to create flow handle");
1673 		return NULL;
1674 	}
1675 	MLX5_ASSERT(wks->flow_idx + 1 < RTE_DIM(wks->flows));
1676 	dev_flow = &wks->flows[wks->flow_idx++];
1677 	dev_flow->handle = dev_handle;
1678 	dev_flow->handle_idx = handle_idx;
1679 	/* Memcpy is used, only size needs to be cleared to 0. */
1680 	dev_flow->verbs.size = 0;
1681 	dev_flow->verbs.attr.num_of_specs = 0;
1682 	dev_flow->ingress = attr->ingress;
1683 	dev_flow->hash_fields = 0;
1684 	/* Need to set transfer attribute: not supported in Verbs mode. */
1685 	return dev_flow;
1686 }
1687 
1688 /**
1689  * Fill the flow with verb spec.
1690  *
1691  * @param[in] dev
1692  *   Pointer to Ethernet device.
1693  * @param[in, out] dev_flow
1694  *   Pointer to the mlx5 flow.
1695  * @param[in] attr
1696  *   Pointer to the flow attributes.
1697  * @param[in] items
1698  *   Pointer to the list of items.
1699  * @param[in] actions
1700  *   Pointer to the list of actions.
1701  * @param[out] error
1702  *   Pointer to the error structure.
1703  *
1704  * @return
1705  *   0 on success, else a negative errno value otherwise and rte_errno is set.
1706  */
1707 static int
1708 flow_verbs_translate(struct rte_eth_dev *dev,
1709 		     struct mlx5_flow *dev_flow,
1710 		     const struct rte_flow_attr *attr,
1711 		     const struct rte_flow_item items[],
1712 		     const struct rte_flow_action actions[],
1713 		     struct rte_flow_error *error)
1714 {
1715 	uint64_t item_flags = 0;
1716 	uint64_t action_flags = 0;
1717 	uint64_t priority = attr->priority;
1718 	uint32_t subpriority = 0;
1719 	struct mlx5_priv *priv = dev->data->dev_private;
1720 	struct mlx5_flow_workspace *wks = mlx5_flow_get_thread_workspace();
1721 	struct mlx5_flow_rss_desc *rss_desc;
1722 
1723 	MLX5_ASSERT(wks);
1724 	rss_desc = &wks->rss_desc;
1725 	if (priority == MLX5_FLOW_LOWEST_PRIO_INDICATOR)
1726 		priority = priv->config.flow_prio - 1;
1727 	for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
1728 		int ret;
1729 
1730 		switch (actions->type) {
1731 		case RTE_FLOW_ACTION_TYPE_VOID:
1732 			break;
1733 		case RTE_FLOW_ACTION_TYPE_FLAG:
1734 			flow_verbs_translate_action_flag(dev_flow, actions);
1735 			action_flags |= MLX5_FLOW_ACTION_FLAG;
1736 			dev_flow->handle->mark = 1;
1737 			break;
1738 		case RTE_FLOW_ACTION_TYPE_MARK:
1739 			flow_verbs_translate_action_mark(dev_flow, actions);
1740 			action_flags |= MLX5_FLOW_ACTION_MARK;
1741 			dev_flow->handle->mark = 1;
1742 			break;
1743 		case RTE_FLOW_ACTION_TYPE_DROP:
1744 			flow_verbs_translate_action_drop(dev_flow, actions);
1745 			action_flags |= MLX5_FLOW_ACTION_DROP;
1746 			dev_flow->handle->fate_action = MLX5_FLOW_FATE_DROP;
1747 			break;
1748 		case RTE_FLOW_ACTION_TYPE_QUEUE:
1749 			flow_verbs_translate_action_queue(rss_desc, actions);
1750 			action_flags |= MLX5_FLOW_ACTION_QUEUE;
1751 			dev_flow->handle->fate_action = MLX5_FLOW_FATE_QUEUE;
1752 			break;
1753 		case RTE_FLOW_ACTION_TYPE_RSS:
1754 			flow_verbs_translate_action_rss(rss_desc, actions);
1755 			action_flags |= MLX5_FLOW_ACTION_RSS;
1756 			dev_flow->handle->fate_action = MLX5_FLOW_FATE_QUEUE;
1757 			break;
1758 		case RTE_FLOW_ACTION_TYPE_COUNT:
1759 			ret = flow_verbs_translate_action_count(dev_flow,
1760 								actions,
1761 								dev, error);
1762 			if (ret < 0)
1763 				return ret;
1764 			action_flags |= MLX5_FLOW_ACTION_COUNT;
1765 			break;
1766 		default:
1767 			return rte_flow_error_set(error, ENOTSUP,
1768 						  RTE_FLOW_ERROR_TYPE_ACTION,
1769 						  actions,
1770 						  "action not supported");
1771 		}
1772 	}
1773 	dev_flow->act_flags = action_flags;
1774 	for (; items->type != RTE_FLOW_ITEM_TYPE_END; items++) {
1775 		int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
1776 
1777 		switch (items->type) {
1778 		case RTE_FLOW_ITEM_TYPE_VOID:
1779 			break;
1780 		case RTE_FLOW_ITEM_TYPE_ETH:
1781 			flow_verbs_translate_item_eth(dev_flow, items,
1782 						      item_flags);
1783 			subpriority = MLX5_PRIORITY_MAP_L2;
1784 			item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
1785 					       MLX5_FLOW_LAYER_OUTER_L2;
1786 			break;
1787 		case RTE_FLOW_ITEM_TYPE_VLAN:
1788 			flow_verbs_translate_item_vlan(dev_flow, items,
1789 						       item_flags);
1790 			subpriority = MLX5_PRIORITY_MAP_L2;
1791 			item_flags |= tunnel ? (MLX5_FLOW_LAYER_INNER_L2 |
1792 						MLX5_FLOW_LAYER_INNER_VLAN) :
1793 					       (MLX5_FLOW_LAYER_OUTER_L2 |
1794 						MLX5_FLOW_LAYER_OUTER_VLAN);
1795 			break;
1796 		case RTE_FLOW_ITEM_TYPE_IPV4:
1797 			flow_verbs_translate_item_ipv4(dev_flow, items,
1798 						       item_flags);
1799 			subpriority = MLX5_PRIORITY_MAP_L3;
1800 			dev_flow->hash_fields |=
1801 				mlx5_flow_hashfields_adjust
1802 					(rss_desc, tunnel,
1803 					 MLX5_IPV4_LAYER_TYPES,
1804 					 MLX5_IPV4_IBV_RX_HASH);
1805 			item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV4 :
1806 					       MLX5_FLOW_LAYER_OUTER_L3_IPV4;
1807 			break;
1808 		case RTE_FLOW_ITEM_TYPE_IPV6:
1809 			flow_verbs_translate_item_ipv6(dev_flow, items,
1810 						       item_flags);
1811 			subpriority = MLX5_PRIORITY_MAP_L3;
1812 			dev_flow->hash_fields |=
1813 				mlx5_flow_hashfields_adjust
1814 					(rss_desc, tunnel,
1815 					 MLX5_IPV6_LAYER_TYPES,
1816 					 MLX5_IPV6_IBV_RX_HASH);
1817 			item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV6 :
1818 					       MLX5_FLOW_LAYER_OUTER_L3_IPV6;
1819 			break;
1820 		case RTE_FLOW_ITEM_TYPE_TCP:
1821 			flow_verbs_translate_item_tcp(dev_flow, items,
1822 						      item_flags);
1823 			subpriority = MLX5_PRIORITY_MAP_L4;
1824 			if (dev_flow->hash_fields != 0)
1825 				dev_flow->hash_fields |=
1826 					mlx5_flow_hashfields_adjust
1827 					(rss_desc, tunnel, RTE_ETH_RSS_TCP,
1828 					 (IBV_RX_HASH_SRC_PORT_TCP |
1829 					  IBV_RX_HASH_DST_PORT_TCP));
1830 			item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L4_TCP :
1831 					       MLX5_FLOW_LAYER_OUTER_L4_TCP;
1832 			break;
1833 		case RTE_FLOW_ITEM_TYPE_UDP:
1834 			flow_verbs_translate_item_udp(dev_flow, items,
1835 						      item_flags);
1836 			subpriority = MLX5_PRIORITY_MAP_L4;
1837 			if (dev_flow->hash_fields != 0)
1838 				dev_flow->hash_fields |=
1839 					mlx5_flow_hashfields_adjust
1840 					(rss_desc, tunnel, RTE_ETH_RSS_UDP,
1841 					 (IBV_RX_HASH_SRC_PORT_UDP |
1842 					  IBV_RX_HASH_DST_PORT_UDP));
1843 			item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L4_UDP :
1844 					       MLX5_FLOW_LAYER_OUTER_L4_UDP;
1845 			break;
1846 		case RTE_FLOW_ITEM_TYPE_VXLAN:
1847 			flow_verbs_translate_item_vxlan(dev_flow, items,
1848 							item_flags);
1849 			subpriority = MLX5_TUNNEL_PRIO_GET(rss_desc);
1850 			item_flags |= MLX5_FLOW_LAYER_VXLAN;
1851 			break;
1852 		case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
1853 			flow_verbs_translate_item_vxlan_gpe(dev_flow, items,
1854 							    item_flags);
1855 			subpriority = MLX5_TUNNEL_PRIO_GET(rss_desc);
1856 			item_flags |= MLX5_FLOW_LAYER_VXLAN_GPE;
1857 			break;
1858 		case RTE_FLOW_ITEM_TYPE_GRE:
1859 			flow_verbs_translate_item_gre(dev_flow, items,
1860 						      item_flags);
1861 			subpriority = MLX5_TUNNEL_PRIO_GET(rss_desc);
1862 			item_flags |= MLX5_FLOW_LAYER_GRE;
1863 			break;
1864 		case RTE_FLOW_ITEM_TYPE_MPLS:
1865 			flow_verbs_translate_item_mpls(dev_flow, items,
1866 						       item_flags);
1867 			subpriority = MLX5_TUNNEL_PRIO_GET(rss_desc);
1868 			item_flags |= MLX5_FLOW_LAYER_MPLS;
1869 			break;
1870 		default:
1871 			return rte_flow_error_set(error, ENOTSUP,
1872 						  RTE_FLOW_ERROR_TYPE_ITEM,
1873 						  NULL, "item not supported");
1874 		}
1875 	}
1876 	dev_flow->handle->layers = item_flags;
1877 	/* Other members of attr will be ignored. */
1878 	dev_flow->verbs.attr.priority =
1879 		mlx5_flow_adjust_priority(dev, priority, subpriority);
1880 	dev_flow->verbs.attr.port = (uint8_t)priv->dev_port;
1881 	return 0;
1882 }
1883 
1884 /**
1885  * Remove the flow from the NIC but keeps it in memory.
1886  *
1887  * @param[in] dev
1888  *   Pointer to the Ethernet device structure.
1889  * @param[in, out] flow
1890  *   Pointer to flow structure.
1891  */
1892 static void
1893 flow_verbs_remove(struct rte_eth_dev *dev, struct rte_flow *flow)
1894 {
1895 	struct mlx5_priv *priv = dev->data->dev_private;
1896 	struct mlx5_flow_handle *handle;
1897 	uint32_t handle_idx;
1898 
1899 	if (!flow)
1900 		return;
1901 	SILIST_FOREACH(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW], flow->dev_handles,
1902 		       handle_idx, handle, next) {
1903 		if (handle->drv_flow) {
1904 			claim_zero(mlx5_glue->destroy_flow(handle->drv_flow));
1905 			handle->drv_flow = NULL;
1906 		}
1907 		/* hrxq is union, don't touch it only the flag is set. */
1908 		if (handle->rix_hrxq &&
1909 		    handle->fate_action == MLX5_FLOW_FATE_QUEUE) {
1910 			mlx5_hrxq_release(dev, handle->rix_hrxq);
1911 			handle->rix_hrxq = 0;
1912 		}
1913 		if (handle->vf_vlan.tag && handle->vf_vlan.created)
1914 			mlx5_vlan_vmwa_release(dev, &handle->vf_vlan);
1915 	}
1916 }
1917 
1918 /**
1919  * Remove the flow from the NIC and the memory.
1920  *
1921  * @param[in] dev
1922  *   Pointer to the Ethernet device structure.
1923  * @param[in, out] flow
1924  *   Pointer to flow structure.
1925  */
1926 static void
1927 flow_verbs_destroy(struct rte_eth_dev *dev, struct rte_flow *flow)
1928 {
1929 	struct mlx5_priv *priv = dev->data->dev_private;
1930 	struct mlx5_flow_handle *handle;
1931 
1932 	if (!flow)
1933 		return;
1934 	flow_verbs_remove(dev, flow);
1935 	while (flow->dev_handles) {
1936 		uint32_t tmp_idx = flow->dev_handles;
1937 
1938 		handle = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW],
1939 				   tmp_idx);
1940 		if (!handle)
1941 			return;
1942 		flow->dev_handles = handle->next.next;
1943 		mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW],
1944 			   tmp_idx);
1945 	}
1946 	if (flow->counter) {
1947 		flow_verbs_counter_release(dev, flow->counter);
1948 		flow->counter = 0;
1949 	}
1950 }
1951 
1952 /**
1953  * Apply the flow to the NIC.
1954  *
1955  * @param[in] dev
1956  *   Pointer to the Ethernet device structure.
1957  * @param[in, out] flow
1958  *   Pointer to flow structure.
1959  * @param[out] error
1960  *   Pointer to error structure.
1961  *
1962  * @return
1963  *   0 on success, a negative errno value otherwise and rte_errno is set.
1964  */
1965 static int
1966 flow_verbs_apply(struct rte_eth_dev *dev, struct rte_flow *flow,
1967 		 struct rte_flow_error *error)
1968 {
1969 	struct mlx5_priv *priv = dev->data->dev_private;
1970 	struct mlx5_flow_handle *handle;
1971 	struct mlx5_flow *dev_flow;
1972 	struct mlx5_hrxq *hrxq;
1973 	uint32_t dev_handles;
1974 	int err;
1975 	int idx;
1976 	struct mlx5_flow_workspace *wks = mlx5_flow_get_thread_workspace();
1977 
1978 	MLX5_ASSERT(wks);
1979 	for (idx = wks->flow_idx - 1; idx >= 0; idx--) {
1980 		dev_flow = &wks->flows[idx];
1981 		handle = dev_flow->handle;
1982 		if (handle->fate_action == MLX5_FLOW_FATE_DROP) {
1983 			MLX5_ASSERT(priv->drop_queue.hrxq);
1984 			hrxq = priv->drop_queue.hrxq;
1985 		} else {
1986 			uint32_t hrxq_idx;
1987 			struct mlx5_flow_rss_desc *rss_desc = &wks->rss_desc;
1988 
1989 			MLX5_ASSERT(rss_desc->queue_num);
1990 			rss_desc->key_len = MLX5_RSS_HASH_KEY_LEN;
1991 			rss_desc->hash_fields = dev_flow->hash_fields;
1992 			rss_desc->tunnel = !!(handle->layers &
1993 					      MLX5_FLOW_LAYER_TUNNEL);
1994 			rss_desc->shared_rss = 0;
1995 			hrxq_idx = mlx5_hrxq_get(dev, rss_desc);
1996 			hrxq = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_HRXQ],
1997 					      hrxq_idx);
1998 			if (!hrxq) {
1999 				rte_flow_error_set
2000 					(error, rte_errno,
2001 					 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
2002 					 "cannot get hash queue");
2003 				goto error;
2004 			}
2005 			handle->rix_hrxq = hrxq_idx;
2006 		}
2007 		MLX5_ASSERT(hrxq);
2008 		handle->drv_flow = mlx5_glue->create_flow
2009 					(hrxq->qp, &dev_flow->verbs.attr);
2010 		if (!handle->drv_flow) {
2011 			rte_flow_error_set(error, errno,
2012 					   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2013 					   NULL,
2014 					   "hardware refuses to create flow");
2015 			goto error;
2016 		}
2017 		if (priv->vmwa_context &&
2018 		    handle->vf_vlan.tag && !handle->vf_vlan.created) {
2019 			/*
2020 			 * The rule contains the VLAN pattern.
2021 			 * For VF we are going to create VLAN
2022 			 * interface to make hypervisor set correct
2023 			 * e-Switch vport context.
2024 			 */
2025 			mlx5_vlan_vmwa_acquire(dev, &handle->vf_vlan);
2026 		}
2027 	}
2028 	return 0;
2029 error:
2030 	err = rte_errno; /* Save rte_errno before cleanup. */
2031 	SILIST_FOREACH(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW], flow->dev_handles,
2032 		       dev_handles, handle, next) {
2033 		/* hrxq is union, don't touch it only the flag is set. */
2034 		if (handle->rix_hrxq &&
2035 		    handle->fate_action == MLX5_FLOW_FATE_QUEUE) {
2036 			mlx5_hrxq_release(dev, handle->rix_hrxq);
2037 			handle->rix_hrxq = 0;
2038 		}
2039 		if (handle->vf_vlan.tag && handle->vf_vlan.created)
2040 			mlx5_vlan_vmwa_release(dev, &handle->vf_vlan);
2041 	}
2042 	rte_errno = err; /* Restore rte_errno. */
2043 	return -rte_errno;
2044 }
2045 
2046 /**
2047  * Query a flow.
2048  *
2049  * @see rte_flow_query()
2050  * @see rte_flow_ops
2051  */
2052 static int
2053 flow_verbs_query(struct rte_eth_dev *dev,
2054 		 struct rte_flow *flow,
2055 		 const struct rte_flow_action *actions,
2056 		 void *data,
2057 		 struct rte_flow_error *error)
2058 {
2059 	int ret = -EINVAL;
2060 
2061 	for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
2062 		switch (actions->type) {
2063 		case RTE_FLOW_ACTION_TYPE_VOID:
2064 			break;
2065 		case RTE_FLOW_ACTION_TYPE_COUNT:
2066 			ret = flow_verbs_counter_query(dev, flow, data, error);
2067 			break;
2068 		default:
2069 			return rte_flow_error_set(error, ENOTSUP,
2070 						  RTE_FLOW_ERROR_TYPE_ACTION,
2071 						  actions,
2072 						  "action not supported");
2073 		}
2074 	}
2075 	return ret;
2076 }
2077 
2078 static int
2079 flow_verbs_sync_domain(struct rte_eth_dev *dev, uint32_t domains,
2080 		       uint32_t flags)
2081 {
2082 	RTE_SET_USED(dev);
2083 	RTE_SET_USED(domains);
2084 	RTE_SET_USED(flags);
2085 
2086 	return 0;
2087 }
2088 
2089 const struct mlx5_flow_driver_ops mlx5_flow_verbs_drv_ops = {
2090 	.validate = flow_verbs_validate,
2091 	.prepare = flow_verbs_prepare,
2092 	.translate = flow_verbs_translate,
2093 	.apply = flow_verbs_apply,
2094 	.remove = flow_verbs_remove,
2095 	.destroy = flow_verbs_destroy,
2096 	.query = flow_verbs_query,
2097 	.sync_domain = flow_verbs_sync_domain,
2098 };
2099