xref: /dpdk/drivers/net/mlx5/mlx5_flow_verbs.c (revision b53d106d34b5c638f5a2cbdfee0da5bd42d4383f)
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 /* Verbs specification header. */
32 struct ibv_spec_header {
33 	enum ibv_flow_spec_type type;
34 	uint16_t size;
35 };
36 
37 /**
38  * Discover the maximum number of priority available.
39  *
40  * @param[in] dev
41  *   Pointer to the Ethernet device structure.
42  * @param[in] vprio
43  *   Expected result variants.
44  * @param[in] vprio_n
45  *   Number of entries in @p vprio array.
46  * @return
47  *   Number of supported flow priority on success, a negative errno
48  *   value otherwise and rte_errno is set.
49  */
50 static int
51 flow_verbs_discover_priorities(struct rte_eth_dev *dev,
52 			       const uint16_t *vprio, int vprio_n)
53 {
54 	struct mlx5_priv *priv = dev->data->dev_private;
55 	struct {
56 		struct ibv_flow_attr attr;
57 		struct ibv_flow_spec_eth eth;
58 		struct ibv_flow_spec_action_drop drop;
59 	} flow_attr = {
60 		.attr = {
61 			.num_of_specs = 2,
62 			.port = (uint8_t)priv->dev_port,
63 		},
64 		.eth = {
65 			.type = IBV_FLOW_SPEC_ETH,
66 			.size = sizeof(struct ibv_flow_spec_eth),
67 		},
68 		.drop = {
69 			.size = sizeof(struct ibv_flow_spec_action_drop),
70 			.type = IBV_FLOW_SPEC_ACTION_DROP,
71 		},
72 	};
73 	struct ibv_flow *flow;
74 	struct mlx5_hrxq *drop = priv->drop_queue.hrxq;
75 	int i;
76 	int priority = 0;
77 
78 #if defined(HAVE_MLX5DV_DR_DEVX_PORT) || defined(HAVE_MLX5DV_DR_DEVX_PORT_V35)
79 	/* If DevX supported, driver must support 16 verbs flow priorities. */
80 	priority = 16;
81 	goto out;
82 #endif
83 	if (!drop->qp) {
84 		rte_errno = ENOTSUP;
85 		return -rte_errno;
86 	}
87 	for (i = 0; i != vprio_n; i++) {
88 		flow_attr.attr.priority = vprio[i] - 1;
89 		flow = mlx5_glue->create_flow(drop->qp, &flow_attr.attr);
90 		if (!flow)
91 			break;
92 		claim_zero(mlx5_glue->destroy_flow(flow));
93 		priority = vprio[i];
94 	}
95 #if defined(HAVE_MLX5DV_DR_DEVX_PORT) || defined(HAVE_MLX5DV_DR_DEVX_PORT_V35)
96 out:
97 #endif
98 	DRV_LOG(INFO, "port %u supported flow priorities:"
99 		" 0-%d for ingress or egress root table,"
100 		" 0-%d for non-root table or transfer root table.",
101 		dev->data->port_id, priority - 2,
102 		MLX5_NON_ROOT_FLOW_MAX_PRIO - 1);
103 	return priority;
104 }
105 
106 /**
107  * Get Verbs flow counter by index.
108  *
109  * @param[in] dev
110  *   Pointer to the Ethernet device structure.
111  * @param[in] idx
112  *   mlx5 flow counter index in the container.
113  * @param[out] ppool
114  *   mlx5 flow counter pool in the container,
115  *
116  * @return
117  *   A pointer to the counter, NULL otherwise.
118  */
119 static struct mlx5_flow_counter *
120 flow_verbs_counter_get_by_idx(struct rte_eth_dev *dev,
121 			      uint32_t idx,
122 			      struct mlx5_flow_counter_pool **ppool)
123 {
124 	struct mlx5_priv *priv = dev->data->dev_private;
125 	struct mlx5_flow_counter_mng *cmng = &priv->sh->cmng;
126 	struct mlx5_flow_counter_pool *pool;
127 
128 	idx = (idx - 1) & (MLX5_CNT_SHARED_OFFSET - 1);
129 	pool = cmng->pools[idx / MLX5_COUNTERS_PER_POOL];
130 	MLX5_ASSERT(pool);
131 	if (ppool)
132 		*ppool = pool;
133 	return MLX5_POOL_GET_CNT(pool, idx % MLX5_COUNTERS_PER_POOL);
134 }
135 
136 /**
137  * Create Verbs flow counter with Verbs library.
138  *
139  * @param[in] dev
140  *   Pointer to the Ethernet device structure.
141  * @param[in, out] counter
142  *   mlx5 flow counter object, contains the counter id,
143  *   handle of created Verbs flow counter is returned
144  *   in cs field (if counters are supported).
145  *
146  * @return
147  *   0 On success else a negative errno value is returned
148  *   and rte_errno is set.
149  */
150 static int
151 flow_verbs_counter_create(struct rte_eth_dev *dev,
152 			  struct mlx5_flow_counter *counter)
153 {
154 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42)
155 	struct mlx5_priv *priv = dev->data->dev_private;
156 	struct ibv_context *ctx = priv->sh->cdev->ctx;
157 	struct ibv_counter_set_init_attr init = {
158 			 .counter_set_id = counter->shared_info.id};
159 
160 	counter->dcs_when_free = mlx5_glue->create_counter_set(ctx, &init);
161 	if (!counter->dcs_when_free) {
162 		rte_errno = ENOTSUP;
163 		return -ENOTSUP;
164 	}
165 	return 0;
166 #elif defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
167 	struct mlx5_priv *priv = dev->data->dev_private;
168 	struct ibv_context *ctx = priv->sh->cdev->ctx;
169 	struct ibv_counters_init_attr init = {0};
170 	struct ibv_counter_attach_attr attach;
171 	int ret;
172 
173 	memset(&attach, 0, sizeof(attach));
174 	counter->dcs_when_free = mlx5_glue->create_counters(ctx, &init);
175 	if (!counter->dcs_when_free) {
176 		rte_errno = ENOTSUP;
177 		return -ENOTSUP;
178 	}
179 	attach.counter_desc = IBV_COUNTER_PACKETS;
180 	attach.index = 0;
181 	ret = mlx5_glue->attach_counters(counter->dcs_when_free, &attach, NULL);
182 	if (!ret) {
183 		attach.counter_desc = IBV_COUNTER_BYTES;
184 		attach.index = 1;
185 		ret = mlx5_glue->attach_counters
186 					(counter->dcs_when_free, &attach, NULL);
187 	}
188 	if (ret) {
189 		claim_zero(mlx5_glue->destroy_counters(counter->dcs_when_free));
190 		counter->dcs_when_free = NULL;
191 		rte_errno = ret;
192 		return -ret;
193 	}
194 	return 0;
195 #else
196 	(void)dev;
197 	(void)counter;
198 	rte_errno = ENOTSUP;
199 	return -ENOTSUP;
200 #endif
201 }
202 
203 /**
204  * Get a flow counter.
205  *
206  * @param[in] dev
207  *   Pointer to the Ethernet device structure.
208  * @param[in] id
209  *   Counter identifier.
210  *
211  * @return
212  *   Index to the counter, 0 otherwise and rte_errno is set.
213  */
214 static uint32_t
215 flow_verbs_counter_new(struct rte_eth_dev *dev, uint32_t id __rte_unused)
216 {
217 	struct mlx5_priv *priv = dev->data->dev_private;
218 	struct mlx5_flow_counter_mng *cmng = &priv->sh->cmng;
219 	struct mlx5_flow_counter_pool *pool = NULL;
220 	struct mlx5_flow_counter *cnt = NULL;
221 	uint32_t n_valid = cmng->n_valid;
222 	uint32_t pool_idx, cnt_idx;
223 	uint32_t i;
224 	int ret;
225 
226 	for (pool_idx = 0; pool_idx < n_valid; ++pool_idx) {
227 		pool = cmng->pools[pool_idx];
228 		if (!pool)
229 			continue;
230 		cnt = TAILQ_FIRST(&pool->counters[0]);
231 		if (cnt)
232 			break;
233 	}
234 	if (!cnt) {
235 		struct mlx5_flow_counter_pool **pools;
236 		uint32_t size;
237 
238 		if (n_valid == cmng->n) {
239 			/* Resize the container pool array. */
240 			size = sizeof(struct mlx5_flow_counter_pool *) *
241 				     (n_valid + MLX5_CNT_CONTAINER_RESIZE);
242 			pools = mlx5_malloc(MLX5_MEM_ZERO, size, 0,
243 					    SOCKET_ID_ANY);
244 			if (!pools)
245 				return 0;
246 			if (n_valid) {
247 				memcpy(pools, cmng->pools,
248 				       sizeof(struct mlx5_flow_counter_pool *) *
249 				       n_valid);
250 				mlx5_free(cmng->pools);
251 			}
252 			cmng->pools = pools;
253 			cmng->n += MLX5_CNT_CONTAINER_RESIZE;
254 		}
255 		/* Allocate memory for new pool*/
256 		size = sizeof(*pool) + sizeof(*cnt) * MLX5_COUNTERS_PER_POOL;
257 		pool = mlx5_malloc(MLX5_MEM_ZERO, size, 0, SOCKET_ID_ANY);
258 		if (!pool)
259 			return 0;
260 		for (i = 0; i < MLX5_COUNTERS_PER_POOL; ++i) {
261 			cnt = MLX5_POOL_GET_CNT(pool, i);
262 			TAILQ_INSERT_HEAD(&pool->counters[0], cnt, next);
263 		}
264 		cnt = MLX5_POOL_GET_CNT(pool, 0);
265 		cmng->pools[n_valid] = pool;
266 		pool_idx = n_valid;
267 		cmng->n_valid++;
268 	}
269 	TAILQ_REMOVE(&pool->counters[0], cnt, next);
270 	i = MLX5_CNT_ARRAY_IDX(pool, cnt);
271 	cnt_idx = MLX5_MAKE_CNT_IDX(pool_idx, i);
272 	/* Create counter with Verbs. */
273 	ret = flow_verbs_counter_create(dev, cnt);
274 	if (!ret) {
275 		cnt->dcs_when_active = cnt->dcs_when_free;
276 		cnt->hits = 0;
277 		cnt->bytes = 0;
278 		return cnt_idx;
279 	}
280 	TAILQ_INSERT_HEAD(&pool->counters[0], cnt, next);
281 	/* Some error occurred in Verbs library. */
282 	rte_errno = -ret;
283 	return 0;
284 }
285 
286 /**
287  * Release a flow counter.
288  *
289  * @param[in] dev
290  *   Pointer to the Ethernet device structure.
291  * @param[in] counter
292  *   Index to the counter handler.
293  */
294 static void
295 flow_verbs_counter_release(struct rte_eth_dev *dev, uint32_t counter)
296 {
297 	struct mlx5_flow_counter_pool *pool;
298 	struct mlx5_flow_counter *cnt;
299 
300 	cnt = flow_verbs_counter_get_by_idx(dev, counter, &pool);
301 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42)
302 	claim_zero(mlx5_glue->destroy_counter_set
303 			((struct ibv_counter_set *)cnt->dcs_when_active));
304 #elif defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
305 	claim_zero(mlx5_glue->destroy_counters
306 				((struct ibv_counters *)cnt->dcs_when_active));
307 #endif
308 	TAILQ_INSERT_HEAD(&pool->counters[0], cnt, next);
309 }
310 
311 /**
312  * Query a flow counter via Verbs library call.
313  *
314  * @see rte_flow_query()
315  * @see rte_flow_ops
316  */
317 static int
318 flow_verbs_counter_query(struct rte_eth_dev *dev __rte_unused,
319 			 struct rte_flow *flow, void *data,
320 			 struct rte_flow_error *error)
321 {
322 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) || \
323 	defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
324 	if (flow->counter) {
325 		struct mlx5_flow_counter_pool *pool;
326 		struct mlx5_flow_counter *cnt = flow_verbs_counter_get_by_idx
327 						(dev, flow->counter, &pool);
328 		struct rte_flow_query_count *qc = data;
329 		uint64_t counters[2] = {0, 0};
330 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42)
331 		struct ibv_query_counter_set_attr query_cs_attr = {
332 			.dcs_when_free = (struct ibv_counter_set *)
333 						cnt->dcs_when_active,
334 			.query_flags = IBV_COUNTER_SET_FORCE_UPDATE,
335 		};
336 		struct ibv_counter_set_data query_out = {
337 			.out = counters,
338 			.outlen = 2 * sizeof(uint64_t),
339 		};
340 		int err = mlx5_glue->query_counter_set(&query_cs_attr,
341 						       &query_out);
342 #elif defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
343 		int err = mlx5_glue->query_counters
344 			((struct ibv_counters *)cnt->dcs_when_active, counters,
345 				RTE_DIM(counters),
346 				IBV_READ_COUNTERS_ATTR_PREFER_CACHED);
347 #endif
348 		if (err)
349 			return rte_flow_error_set
350 				(error, err,
351 				 RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
352 				 NULL,
353 				 "cannot read counter");
354 		qc->hits_set = 1;
355 		qc->bytes_set = 1;
356 		qc->hits = counters[0] - cnt->hits;
357 		qc->bytes = counters[1] - cnt->bytes;
358 		if (qc->reset) {
359 			cnt->hits = counters[0];
360 			cnt->bytes = counters[1];
361 		}
362 		return 0;
363 	}
364 	return rte_flow_error_set(error, EINVAL,
365 				  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
366 				  NULL,
367 				  "flow does not have counter");
368 #else
369 	(void)flow;
370 	(void)data;
371 	return rte_flow_error_set(error, ENOTSUP,
372 				  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
373 				  NULL,
374 				  "counters are not available");
375 #endif
376 }
377 
378 /**
379  * Add a verbs item specification into @p verbs.
380  *
381  * @param[out] verbs
382  *   Pointer to verbs structure.
383  * @param[in] src
384  *   Create specification.
385  * @param[in] size
386  *   Size in bytes of the specification to copy.
387  */
388 static void
389 flow_verbs_spec_add(struct mlx5_flow_verbs_workspace *verbs,
390 		    void *src, unsigned int size)
391 {
392 	void *dst;
393 
394 	if (!verbs)
395 		return;
396 	MLX5_ASSERT(verbs->specs);
397 	dst = (void *)(verbs->specs + verbs->size);
398 	memcpy(dst, src, size);
399 	++verbs->attr.num_of_specs;
400 	verbs->size += size;
401 }
402 
403 /**
404  * Convert the @p item into a Verbs specification. This function assumes that
405  * the input is valid and that there is space to insert the requested item
406  * into the flow.
407  *
408  * @param[in, out] dev_flow
409  *   Pointer to dev_flow structure.
410  * @param[in] item
411  *   Item specification.
412  * @param[in] item_flags
413  *   Parsed item flags.
414  */
415 static void
416 flow_verbs_translate_item_eth(struct mlx5_flow *dev_flow,
417 			      const struct rte_flow_item *item,
418 			      uint64_t item_flags)
419 {
420 	const struct rte_flow_item_eth *spec = item->spec;
421 	const struct rte_flow_item_eth *mask = item->mask;
422 	const unsigned int size = sizeof(struct ibv_flow_spec_eth);
423 	struct ibv_flow_spec_eth eth = {
424 		.type = IBV_FLOW_SPEC_ETH | VERBS_SPEC_INNER(item_flags),
425 		.size = size,
426 	};
427 
428 	if (!mask)
429 		mask = &rte_flow_item_eth_mask;
430 	if (spec) {
431 		unsigned int i;
432 
433 		memcpy(&eth.val.dst_mac, spec->dst.addr_bytes,
434 			RTE_ETHER_ADDR_LEN);
435 		memcpy(&eth.val.src_mac, spec->src.addr_bytes,
436 			RTE_ETHER_ADDR_LEN);
437 		eth.val.ether_type = spec->type;
438 		memcpy(&eth.mask.dst_mac, mask->dst.addr_bytes,
439 			RTE_ETHER_ADDR_LEN);
440 		memcpy(&eth.mask.src_mac, mask->src.addr_bytes,
441 			RTE_ETHER_ADDR_LEN);
442 		eth.mask.ether_type = mask->type;
443 		/* Remove unwanted bits from values. */
444 		for (i = 0; i < RTE_ETHER_ADDR_LEN; ++i) {
445 			eth.val.dst_mac[i] &= eth.mask.dst_mac[i];
446 			eth.val.src_mac[i] &= eth.mask.src_mac[i];
447 		}
448 		eth.val.ether_type &= eth.mask.ether_type;
449 	}
450 	flow_verbs_spec_add(&dev_flow->verbs, &eth, size);
451 }
452 
453 /**
454  * Update the VLAN tag in the Verbs Ethernet specification.
455  * This function assumes that the input is valid and there is space to add
456  * the requested item.
457  *
458  * @param[in, out] attr
459  *   Pointer to Verbs attributes structure.
460  * @param[in] eth
461  *   Verbs structure containing the VLAN information to copy.
462  */
463 static void
464 flow_verbs_item_vlan_update(struct ibv_flow_attr *attr,
465 			    struct ibv_flow_spec_eth *eth)
466 {
467 	unsigned int i;
468 	const enum ibv_flow_spec_type search = eth->type;
469 	struct ibv_spec_header *hdr = (struct ibv_spec_header *)
470 		((uint8_t *)attr + sizeof(struct ibv_flow_attr));
471 
472 	for (i = 0; i != attr->num_of_specs; ++i) {
473 		if (hdr->type == search) {
474 			struct ibv_flow_spec_eth *e =
475 				(struct ibv_flow_spec_eth *)hdr;
476 
477 			e->val.vlan_tag = eth->val.vlan_tag;
478 			e->mask.vlan_tag = eth->mask.vlan_tag;
479 			e->val.ether_type = eth->val.ether_type;
480 			e->mask.ether_type = eth->mask.ether_type;
481 			break;
482 		}
483 		hdr = (struct ibv_spec_header *)((uint8_t *)hdr + hdr->size);
484 	}
485 }
486 
487 /**
488  * Convert the @p item into a Verbs specification. This function assumes that
489  * the input is valid and that there is space to insert the requested item
490  * into the flow.
491  *
492  * @param[in, out] dev_flow
493  *   Pointer to dev_flow structure.
494  * @param[in] item
495  *   Item specification.
496  * @param[in] item_flags
497  *   Parsed item flags.
498  */
499 static void
500 flow_verbs_translate_item_vlan(struct mlx5_flow *dev_flow,
501 			       const struct rte_flow_item *item,
502 			       uint64_t item_flags)
503 {
504 	const struct rte_flow_item_vlan *spec = item->spec;
505 	const struct rte_flow_item_vlan *mask = item->mask;
506 	unsigned int size = sizeof(struct ibv_flow_spec_eth);
507 	const int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
508 	struct ibv_flow_spec_eth eth = {
509 		.type = IBV_FLOW_SPEC_ETH | VERBS_SPEC_INNER(item_flags),
510 		.size = size,
511 	};
512 	const uint32_t l2m = tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
513 				      MLX5_FLOW_LAYER_OUTER_L2;
514 
515 	if (!mask)
516 		mask = &rte_flow_item_vlan_mask;
517 	if (spec) {
518 		eth.val.vlan_tag = spec->tci;
519 		eth.mask.vlan_tag = mask->tci;
520 		eth.val.vlan_tag &= eth.mask.vlan_tag;
521 		eth.val.ether_type = spec->inner_type;
522 		eth.mask.ether_type = mask->inner_type;
523 		eth.val.ether_type &= eth.mask.ether_type;
524 	}
525 	if (!(item_flags & l2m))
526 		flow_verbs_spec_add(&dev_flow->verbs, &eth, size);
527 	else
528 		flow_verbs_item_vlan_update(&dev_flow->verbs.attr, &eth);
529 	if (!tunnel)
530 		dev_flow->handle->vf_vlan.tag =
531 			rte_be_to_cpu_16(spec->tci) & 0x0fff;
532 }
533 
534 /**
535  * Convert the @p item into a Verbs specification. This function assumes that
536  * the input is valid and that there is space to insert the requested item
537  * into the flow.
538  *
539  * @param[in, out] dev_flow
540  *   Pointer to dev_flow structure.
541  * @param[in] item
542  *   Item specification.
543  * @param[in] item_flags
544  *   Parsed item flags.
545  */
546 static void
547 flow_verbs_translate_item_ipv4(struct mlx5_flow *dev_flow,
548 			       const struct rte_flow_item *item,
549 			       uint64_t item_flags)
550 {
551 	const struct rte_flow_item_ipv4 *spec = item->spec;
552 	const struct rte_flow_item_ipv4 *mask = item->mask;
553 	unsigned int size = sizeof(struct ibv_flow_spec_ipv4_ext);
554 	struct ibv_flow_spec_ipv4_ext ipv4 = {
555 		.type = IBV_FLOW_SPEC_IPV4_EXT | VERBS_SPEC_INNER(item_flags),
556 		.size = size,
557 	};
558 
559 	if (!mask)
560 		mask = &rte_flow_item_ipv4_mask;
561 	if (spec) {
562 		ipv4.val = (struct ibv_flow_ipv4_ext_filter){
563 			.src_ip = spec->hdr.src_addr,
564 			.dst_ip = spec->hdr.dst_addr,
565 			.proto = spec->hdr.next_proto_id,
566 			.tos = spec->hdr.type_of_service,
567 		};
568 		ipv4.mask = (struct ibv_flow_ipv4_ext_filter){
569 			.src_ip = mask->hdr.src_addr,
570 			.dst_ip = mask->hdr.dst_addr,
571 			.proto = mask->hdr.next_proto_id,
572 			.tos = mask->hdr.type_of_service,
573 		};
574 		/* Remove unwanted bits from values. */
575 		ipv4.val.src_ip &= ipv4.mask.src_ip;
576 		ipv4.val.dst_ip &= ipv4.mask.dst_ip;
577 		ipv4.val.proto &= ipv4.mask.proto;
578 		ipv4.val.tos &= ipv4.mask.tos;
579 	}
580 	flow_verbs_spec_add(&dev_flow->verbs, &ipv4, size);
581 }
582 
583 /**
584  * Convert the @p item into a Verbs specification. This function assumes that
585  * the input is valid and that there is space to insert the requested item
586  * into the flow.
587  *
588  * @param[in, out] dev_flow
589  *   Pointer to dev_flow structure.
590  * @param[in] item
591  *   Item specification.
592  * @param[in] item_flags
593  *   Parsed item flags.
594  */
595 static void
596 flow_verbs_translate_item_ipv6(struct mlx5_flow *dev_flow,
597 			       const struct rte_flow_item *item,
598 			       uint64_t item_flags)
599 {
600 	const struct rte_flow_item_ipv6 *spec = item->spec;
601 	const struct rte_flow_item_ipv6 *mask = item->mask;
602 	unsigned int size = sizeof(struct ibv_flow_spec_ipv6);
603 	struct ibv_flow_spec_ipv6 ipv6 = {
604 		.type = IBV_FLOW_SPEC_IPV6 | VERBS_SPEC_INNER(item_flags),
605 		.size = size,
606 	};
607 
608 	if (!mask)
609 		mask = &rte_flow_item_ipv6_mask;
610 	if (spec) {
611 		unsigned int i;
612 		uint32_t vtc_flow_val;
613 		uint32_t vtc_flow_mask;
614 
615 		memcpy(&ipv6.val.src_ip, spec->hdr.src_addr,
616 		       RTE_DIM(ipv6.val.src_ip));
617 		memcpy(&ipv6.val.dst_ip, spec->hdr.dst_addr,
618 		       RTE_DIM(ipv6.val.dst_ip));
619 		memcpy(&ipv6.mask.src_ip, mask->hdr.src_addr,
620 		       RTE_DIM(ipv6.mask.src_ip));
621 		memcpy(&ipv6.mask.dst_ip, mask->hdr.dst_addr,
622 		       RTE_DIM(ipv6.mask.dst_ip));
623 		vtc_flow_val = rte_be_to_cpu_32(spec->hdr.vtc_flow);
624 		vtc_flow_mask = rte_be_to_cpu_32(mask->hdr.vtc_flow);
625 		ipv6.val.flow_label =
626 			rte_cpu_to_be_32((vtc_flow_val & RTE_IPV6_HDR_FL_MASK) >>
627 					 RTE_IPV6_HDR_FL_SHIFT);
628 		ipv6.val.traffic_class = (vtc_flow_val & RTE_IPV6_HDR_TC_MASK) >>
629 					 RTE_IPV6_HDR_TC_SHIFT;
630 		ipv6.val.next_hdr = spec->hdr.proto;
631 		ipv6.mask.flow_label =
632 			rte_cpu_to_be_32((vtc_flow_mask & RTE_IPV6_HDR_FL_MASK) >>
633 					 RTE_IPV6_HDR_FL_SHIFT);
634 		ipv6.mask.traffic_class = (vtc_flow_mask & RTE_IPV6_HDR_TC_MASK) >>
635 					  RTE_IPV6_HDR_TC_SHIFT;
636 		ipv6.mask.next_hdr = mask->hdr.proto;
637 		/* Remove unwanted bits from values. */
638 		for (i = 0; i < RTE_DIM(ipv6.val.src_ip); ++i) {
639 			ipv6.val.src_ip[i] &= ipv6.mask.src_ip[i];
640 			ipv6.val.dst_ip[i] &= ipv6.mask.dst_ip[i];
641 		}
642 		ipv6.val.flow_label &= ipv6.mask.flow_label;
643 		ipv6.val.traffic_class &= ipv6.mask.traffic_class;
644 		ipv6.val.next_hdr &= ipv6.mask.next_hdr;
645 	}
646 	flow_verbs_spec_add(&dev_flow->verbs, &ipv6, size);
647 }
648 
649 /**
650  * Convert the @p item into a Verbs specification. This function assumes that
651  * the input is valid and that there is space to insert the requested item
652  * into the flow.
653  *
654  * @param[in, out] dev_flow
655  *   Pointer to dev_flow structure.
656  * @param[in] item
657  *   Item specification.
658  * @param[in] item_flags
659  *   Parsed item flags.
660  */
661 static void
662 flow_verbs_translate_item_tcp(struct mlx5_flow *dev_flow,
663 			      const struct rte_flow_item *item,
664 			      uint64_t item_flags __rte_unused)
665 {
666 	const struct rte_flow_item_tcp *spec = item->spec;
667 	const struct rte_flow_item_tcp *mask = item->mask;
668 	unsigned int size = sizeof(struct ibv_flow_spec_tcp_udp);
669 	struct ibv_flow_spec_tcp_udp tcp = {
670 		.type = IBV_FLOW_SPEC_TCP | VERBS_SPEC_INNER(item_flags),
671 		.size = size,
672 	};
673 
674 	if (!mask)
675 		mask = &rte_flow_item_tcp_mask;
676 	if (spec) {
677 		tcp.val.dst_port = spec->hdr.dst_port;
678 		tcp.val.src_port = spec->hdr.src_port;
679 		tcp.mask.dst_port = mask->hdr.dst_port;
680 		tcp.mask.src_port = mask->hdr.src_port;
681 		/* Remove unwanted bits from values. */
682 		tcp.val.src_port &= tcp.mask.src_port;
683 		tcp.val.dst_port &= tcp.mask.dst_port;
684 	}
685 	flow_verbs_spec_add(&dev_flow->verbs, &tcp, size);
686 }
687 
688 /**
689  * Convert the @p item into a Verbs specification. This function assumes that
690  * the input is valid and that there is space to insert the requested item
691  * into the flow.
692  *
693  * @param[in, out] dev_flow
694  *   Pointer to dev_flow structure.
695  * @param[in] item
696  *   Item specification.
697  * @param[in] item_flags
698  *   Parsed item flags.
699  */
700 static void
701 flow_verbs_translate_item_udp(struct mlx5_flow *dev_flow,
702 			      const struct rte_flow_item *item,
703 			      uint64_t item_flags __rte_unused)
704 {
705 	const struct rte_flow_item_udp *spec = item->spec;
706 	const struct rte_flow_item_udp *mask = item->mask;
707 	unsigned int size = sizeof(struct ibv_flow_spec_tcp_udp);
708 	struct ibv_flow_spec_tcp_udp udp = {
709 		.type = IBV_FLOW_SPEC_UDP | VERBS_SPEC_INNER(item_flags),
710 		.size = size,
711 	};
712 
713 	if (!mask)
714 		mask = &rte_flow_item_udp_mask;
715 	if (spec) {
716 		udp.val.dst_port = spec->hdr.dst_port;
717 		udp.val.src_port = spec->hdr.src_port;
718 		udp.mask.dst_port = mask->hdr.dst_port;
719 		udp.mask.src_port = mask->hdr.src_port;
720 		/* Remove unwanted bits from values. */
721 		udp.val.src_port &= udp.mask.src_port;
722 		udp.val.dst_port &= udp.mask.dst_port;
723 	}
724 	item++;
725 	while (item->type == RTE_FLOW_ITEM_TYPE_VOID)
726 		item++;
727 	if (!(udp.val.dst_port & udp.mask.dst_port)) {
728 		switch ((item)->type) {
729 		case RTE_FLOW_ITEM_TYPE_VXLAN:
730 			udp.val.dst_port = htons(MLX5_UDP_PORT_VXLAN);
731 			udp.mask.dst_port = 0xffff;
732 			break;
733 		case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
734 			udp.val.dst_port = htons(MLX5_UDP_PORT_VXLAN_GPE);
735 			udp.mask.dst_port = 0xffff;
736 			break;
737 		case RTE_FLOW_ITEM_TYPE_MPLS:
738 			udp.val.dst_port = htons(MLX5_UDP_PORT_MPLS);
739 			udp.mask.dst_port = 0xffff;
740 			break;
741 		default:
742 			break;
743 		}
744 	}
745 
746 	flow_verbs_spec_add(&dev_flow->verbs, &udp, size);
747 }
748 
749 /**
750  * Convert the @p item into a Verbs specification. This function assumes that
751  * the input is valid and that there is space to insert the requested item
752  * into the flow.
753  *
754  * @param[in, out] dev_flow
755  *   Pointer to dev_flow structure.
756  * @param[in] item
757  *   Item specification.
758  * @param[in] item_flags
759  *   Parsed item flags.
760  */
761 static void
762 flow_verbs_translate_item_vxlan(struct mlx5_flow *dev_flow,
763 				const struct rte_flow_item *item,
764 				uint64_t item_flags __rte_unused)
765 {
766 	const struct rte_flow_item_vxlan *spec = item->spec;
767 	const struct rte_flow_item_vxlan *mask = item->mask;
768 	unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
769 	struct ibv_flow_spec_tunnel vxlan = {
770 		.type = IBV_FLOW_SPEC_VXLAN_TUNNEL,
771 		.size = size,
772 	};
773 	union vni {
774 		uint32_t vlan_id;
775 		uint8_t vni[4];
776 	} id = { .vlan_id = 0, };
777 
778 	if (!mask)
779 		mask = &rte_flow_item_vxlan_mask;
780 	if (spec) {
781 		memcpy(&id.vni[1], spec->vni, 3);
782 		vxlan.val.tunnel_id = id.vlan_id;
783 		memcpy(&id.vni[1], mask->vni, 3);
784 		vxlan.mask.tunnel_id = id.vlan_id;
785 		/* Remove unwanted bits from values. */
786 		vxlan.val.tunnel_id &= vxlan.mask.tunnel_id;
787 	}
788 	flow_verbs_spec_add(&dev_flow->verbs, &vxlan, size);
789 }
790 
791 /**
792  * Convert the @p item into a Verbs specification. This function assumes that
793  * the input is valid and that there is space to insert the requested item
794  * into the flow.
795  *
796  * @param[in, out] dev_flow
797  *   Pointer to dev_flow structure.
798  * @param[in] item
799  *   Item specification.
800  * @param[in] item_flags
801  *   Parsed item flags.
802  */
803 static void
804 flow_verbs_translate_item_vxlan_gpe(struct mlx5_flow *dev_flow,
805 				    const struct rte_flow_item *item,
806 				    uint64_t item_flags __rte_unused)
807 {
808 	const struct rte_flow_item_vxlan_gpe *spec = item->spec;
809 	const struct rte_flow_item_vxlan_gpe *mask = item->mask;
810 	unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
811 	struct ibv_flow_spec_tunnel vxlan_gpe = {
812 		.type = IBV_FLOW_SPEC_VXLAN_TUNNEL,
813 		.size = size,
814 	};
815 	union vni {
816 		uint32_t vlan_id;
817 		uint8_t vni[4];
818 	} id = { .vlan_id = 0, };
819 
820 	if (!mask)
821 		mask = &rte_flow_item_vxlan_gpe_mask;
822 	if (spec) {
823 		memcpy(&id.vni[1], spec->vni, 3);
824 		vxlan_gpe.val.tunnel_id = id.vlan_id;
825 		memcpy(&id.vni[1], mask->vni, 3);
826 		vxlan_gpe.mask.tunnel_id = id.vlan_id;
827 		/* Remove unwanted bits from values. */
828 		vxlan_gpe.val.tunnel_id &= vxlan_gpe.mask.tunnel_id;
829 	}
830 	flow_verbs_spec_add(&dev_flow->verbs, &vxlan_gpe, size);
831 }
832 
833 /**
834  * Update the protocol in Verbs IPv4/IPv6 spec.
835  *
836  * @param[in, out] attr
837  *   Pointer to Verbs attributes structure.
838  * @param[in] search
839  *   Specification type to search in order to update the IP protocol.
840  * @param[in] protocol
841  *   Protocol value to set if none is present in the specification.
842  */
843 static void
844 flow_verbs_item_gre_ip_protocol_update(struct ibv_flow_attr *attr,
845 				       enum ibv_flow_spec_type search,
846 				       uint8_t protocol)
847 {
848 	unsigned int i;
849 	struct ibv_spec_header *hdr = (struct ibv_spec_header *)
850 		((uint8_t *)attr + sizeof(struct ibv_flow_attr));
851 
852 	if (!attr)
853 		return;
854 	for (i = 0; i != attr->num_of_specs; ++i) {
855 		if (hdr->type == search) {
856 			union {
857 				struct ibv_flow_spec_ipv4_ext *ipv4;
858 				struct ibv_flow_spec_ipv6 *ipv6;
859 			} ip;
860 
861 			switch (search) {
862 			case IBV_FLOW_SPEC_IPV4_EXT:
863 				ip.ipv4 = (struct ibv_flow_spec_ipv4_ext *)hdr;
864 				if (!ip.ipv4->val.proto) {
865 					ip.ipv4->val.proto = protocol;
866 					ip.ipv4->mask.proto = 0xff;
867 				}
868 				break;
869 			case IBV_FLOW_SPEC_IPV6:
870 				ip.ipv6 = (struct ibv_flow_spec_ipv6 *)hdr;
871 				if (!ip.ipv6->val.next_hdr) {
872 					ip.ipv6->val.next_hdr = protocol;
873 					ip.ipv6->mask.next_hdr = 0xff;
874 				}
875 				break;
876 			default:
877 				break;
878 			}
879 			break;
880 		}
881 		hdr = (struct ibv_spec_header *)((uint8_t *)hdr + hdr->size);
882 	}
883 }
884 
885 /**
886  * Convert the @p item into a Verbs specification. This function assumes that
887  * the input is valid and that there is space to insert the requested item
888  * into the flow.
889  *
890  * @param[in, out] dev_flow
891  *   Pointer to dev_flow structure.
892  * @param[in] item
893  *   Item specification.
894  * @param[in] item_flags
895  *   Parsed item flags.
896  */
897 static void
898 flow_verbs_translate_item_gre(struct mlx5_flow *dev_flow,
899 			      const struct rte_flow_item *item __rte_unused,
900 			      uint64_t item_flags)
901 {
902 	struct mlx5_flow_verbs_workspace *verbs = &dev_flow->verbs;
903 #ifndef HAVE_IBV_DEVICE_MPLS_SUPPORT
904 	unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
905 	struct ibv_flow_spec_tunnel tunnel = {
906 		.type = IBV_FLOW_SPEC_VXLAN_TUNNEL,
907 		.size = size,
908 	};
909 #else
910 	const struct rte_flow_item_gre *spec = item->spec;
911 	const struct rte_flow_item_gre *mask = item->mask;
912 	unsigned int size = sizeof(struct ibv_flow_spec_gre);
913 	struct ibv_flow_spec_gre tunnel = {
914 		.type = IBV_FLOW_SPEC_GRE,
915 		.size = size,
916 	};
917 
918 	if (!mask)
919 		mask = &rte_flow_item_gre_mask;
920 	if (spec) {
921 		tunnel.val.c_ks_res0_ver = spec->c_rsvd0_ver;
922 		tunnel.val.protocol = spec->protocol;
923 		tunnel.mask.c_ks_res0_ver = mask->c_rsvd0_ver;
924 		tunnel.mask.protocol = mask->protocol;
925 		/* Remove unwanted bits from values. */
926 		tunnel.val.c_ks_res0_ver &= tunnel.mask.c_ks_res0_ver;
927 		tunnel.val.protocol &= tunnel.mask.protocol;
928 		tunnel.val.key &= tunnel.mask.key;
929 	}
930 #endif
931 	if (item_flags & MLX5_FLOW_LAYER_OUTER_L3_IPV4)
932 		flow_verbs_item_gre_ip_protocol_update(&verbs->attr,
933 						       IBV_FLOW_SPEC_IPV4_EXT,
934 						       IPPROTO_GRE);
935 	else
936 		flow_verbs_item_gre_ip_protocol_update(&verbs->attr,
937 						       IBV_FLOW_SPEC_IPV6,
938 						       IPPROTO_GRE);
939 	flow_verbs_spec_add(verbs, &tunnel, size);
940 }
941 
942 /**
943  * Convert the @p action into a Verbs specification. This function assumes that
944  * the input is valid and that there is space to insert the requested action
945  * into the flow. This function also return the action that was added.
946  *
947  * @param[in, out] dev_flow
948  *   Pointer to dev_flow structure.
949  * @param[in] item
950  *   Item specification.
951  * @param[in] item_flags
952  *   Parsed item flags.
953  */
954 static void
955 flow_verbs_translate_item_mpls(struct mlx5_flow *dev_flow __rte_unused,
956 			       const struct rte_flow_item *item __rte_unused,
957 			       uint64_t item_flags __rte_unused)
958 {
959 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
960 	const struct rte_flow_item_mpls *spec = item->spec;
961 	const struct rte_flow_item_mpls *mask = item->mask;
962 	unsigned int size = sizeof(struct ibv_flow_spec_mpls);
963 	struct ibv_flow_spec_mpls mpls = {
964 		.type = IBV_FLOW_SPEC_MPLS,
965 		.size = size,
966 	};
967 
968 	if (!mask)
969 		mask = &rte_flow_item_mpls_mask;
970 	if (spec) {
971 		memcpy(&mpls.val.label, spec, sizeof(mpls.val.label));
972 		memcpy(&mpls.mask.label, mask, sizeof(mpls.mask.label));
973 		/* Remove unwanted bits from values.  */
974 		mpls.val.label &= mpls.mask.label;
975 	}
976 	flow_verbs_spec_add(&dev_flow->verbs, &mpls, size);
977 #endif
978 }
979 
980 /**
981  * Convert the @p action into a Verbs specification. This function assumes that
982  * the input is valid and that there is space to insert the requested action
983  * into the flow.
984  *
985  * @param[in] dev_flow
986  *   Pointer to mlx5_flow.
987  * @param[in] action
988  *   Action configuration.
989  */
990 static void
991 flow_verbs_translate_action_drop
992 	(struct mlx5_flow *dev_flow,
993 	 const struct rte_flow_action *action __rte_unused)
994 {
995 	unsigned int size = sizeof(struct ibv_flow_spec_action_drop);
996 	struct ibv_flow_spec_action_drop drop = {
997 			.type = IBV_FLOW_SPEC_ACTION_DROP,
998 			.size = size,
999 	};
1000 
1001 	flow_verbs_spec_add(&dev_flow->verbs, &drop, size);
1002 }
1003 
1004 /**
1005  * Convert the @p action into a Verbs specification. This function assumes that
1006  * the input is valid and that there is space to insert the requested action
1007  * into the flow.
1008  *
1009  * @param[in] rss_desc
1010  *   Pointer to mlx5_flow_rss_desc.
1011  * @param[in] action
1012  *   Action configuration.
1013  */
1014 static void
1015 flow_verbs_translate_action_queue(struct mlx5_flow_rss_desc *rss_desc,
1016 				  const struct rte_flow_action *action)
1017 {
1018 	const struct rte_flow_action_queue *queue = action->conf;
1019 
1020 	rss_desc->queue[0] = queue->index;
1021 	rss_desc->queue_num = 1;
1022 }
1023 
1024 /**
1025  * Convert the @p action into a Verbs specification. This function assumes that
1026  * the input is valid and that there is space to insert the requested action
1027  * into the flow.
1028  *
1029  * @param[in] rss_desc
1030  *   Pointer to mlx5_flow_rss_desc.
1031  * @param[in] action
1032  *   Action configuration.
1033  */
1034 static void
1035 flow_verbs_translate_action_rss(struct mlx5_flow_rss_desc *rss_desc,
1036 				const struct rte_flow_action *action)
1037 {
1038 	const struct rte_flow_action_rss *rss = action->conf;
1039 	const uint8_t *rss_key;
1040 
1041 	memcpy(rss_desc->queue, rss->queue, rss->queue_num * sizeof(uint16_t));
1042 	rss_desc->queue_num = rss->queue_num;
1043 	/* NULL RSS key indicates default RSS key. */
1044 	rss_key = !rss->key ? rss_hash_default_key : rss->key;
1045 	memcpy(rss_desc->key, rss_key, MLX5_RSS_HASH_KEY_LEN);
1046 	/*
1047 	 * rss->level and rss.types should be set in advance when expanding
1048 	 * items for RSS.
1049 	 */
1050 }
1051 
1052 /**
1053  * Convert the @p action into a Verbs specification. This function assumes that
1054  * the input is valid and that there is space to insert the requested action
1055  * into the flow.
1056  *
1057  * @param[in] dev_flow
1058  *   Pointer to mlx5_flow.
1059  * @param[in] action
1060  *   Action configuration.
1061  */
1062 static void
1063 flow_verbs_translate_action_flag
1064 	(struct mlx5_flow *dev_flow,
1065 	 const struct rte_flow_action *action __rte_unused)
1066 {
1067 	unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
1068 	struct ibv_flow_spec_action_tag tag = {
1069 		.type = IBV_FLOW_SPEC_ACTION_TAG,
1070 		.size = size,
1071 		.tag_id = mlx5_flow_mark_set(MLX5_FLOW_MARK_DEFAULT),
1072 	};
1073 
1074 	flow_verbs_spec_add(&dev_flow->verbs, &tag, size);
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] dev_flow
1083  *   Pointer to mlx5_flow.
1084  * @param[in] action
1085  *   Action configuration.
1086  */
1087 static void
1088 flow_verbs_translate_action_mark(struct mlx5_flow *dev_flow,
1089 				 const struct rte_flow_action *action)
1090 {
1091 	const struct rte_flow_action_mark *mark = action->conf;
1092 	unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
1093 	struct ibv_flow_spec_action_tag tag = {
1094 		.type = IBV_FLOW_SPEC_ACTION_TAG,
1095 		.size = size,
1096 		.tag_id = mlx5_flow_mark_set(mark->id),
1097 	};
1098 
1099 	flow_verbs_spec_add(&dev_flow->verbs, &tag, size);
1100 }
1101 
1102 /**
1103  * Convert the @p action into a Verbs specification. This function assumes that
1104  * the input is valid and that there is space to insert the requested action
1105  * into the flow.
1106  *
1107  * @param[in] dev
1108  *   Pointer to the Ethernet device structure.
1109  * @param[in] action
1110  *   Action configuration.
1111  * @param[in] dev_flow
1112  *   Pointer to mlx5_flow.
1113  * @param[out] error
1114  *   Pointer to error structure.
1115  *
1116  * @return
1117  *   0 On success else a negative errno value is returned and rte_errno is set.
1118  */
1119 static int
1120 flow_verbs_translate_action_count(struct mlx5_flow *dev_flow,
1121 				  const struct rte_flow_action *action,
1122 				  struct rte_eth_dev *dev,
1123 				  struct rte_flow_error *error)
1124 {
1125 	const struct rte_flow_action_count *count = action->conf;
1126 	struct rte_flow *flow = dev_flow->flow;
1127 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) || \
1128 	defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1129 	struct mlx5_flow_counter_pool *pool;
1130 	struct mlx5_flow_counter *cnt = NULL;
1131 	unsigned int size = sizeof(struct ibv_flow_spec_counter_action);
1132 	struct ibv_flow_spec_counter_action counter = {
1133 		.type = IBV_FLOW_SPEC_ACTION_COUNT,
1134 		.size = size,
1135 	};
1136 #endif
1137 
1138 	if (!flow->counter) {
1139 		flow->counter = flow_verbs_counter_new(dev, count->id);
1140 		if (!flow->counter)
1141 			return rte_flow_error_set(error, rte_errno,
1142 						  RTE_FLOW_ERROR_TYPE_ACTION,
1143 						  action,
1144 						  "cannot get counter"
1145 						  " context.");
1146 	}
1147 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42)
1148 	cnt = flow_verbs_counter_get_by_idx(dev, flow->counter, &pool);
1149 	counter.counter_set_handle =
1150 		((struct ibv_counter_set *)cnt->dcs_when_active)->handle;
1151 	flow_verbs_spec_add(&dev_flow->verbs, &counter, size);
1152 #elif defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1153 	cnt = flow_verbs_counter_get_by_idx(dev, flow->counter, &pool);
1154 	counter.counters = (struct ibv_counters *)cnt->dcs_when_active;
1155 	flow_verbs_spec_add(&dev_flow->verbs, &counter, size);
1156 #endif
1157 	return 0;
1158 }
1159 
1160 /**
1161  * Internal validation function. For validating both actions and items.
1162  *
1163  * @param[in] dev
1164  *   Pointer to the Ethernet device structure.
1165  * @param[in] attr
1166  *   Pointer to the flow attributes.
1167  * @param[in] items
1168  *   Pointer to the list of items.
1169  * @param[in] actions
1170  *   Pointer to the list of actions.
1171  * @param[in] external
1172  *   This flow rule is created by request external to PMD.
1173  * @param[in] hairpin
1174  *   Number of hairpin TX actions, 0 means classic flow.
1175  * @param[out] error
1176  *   Pointer to the error structure.
1177  *
1178  * @return
1179  *   0 on success, a negative errno value otherwise and rte_errno is set.
1180  */
1181 static int
1182 flow_verbs_validate(struct rte_eth_dev *dev,
1183 		    const struct rte_flow_attr *attr,
1184 		    const struct rte_flow_item items[],
1185 		    const struct rte_flow_action actions[],
1186 		    bool external __rte_unused,
1187 		    int hairpin __rte_unused,
1188 		    struct rte_flow_error *error)
1189 {
1190 	int ret;
1191 	uint64_t action_flags = 0;
1192 	uint64_t item_flags = 0;
1193 	uint64_t last_item = 0;
1194 	uint8_t next_protocol = 0xff;
1195 	uint16_t ether_type = 0;
1196 	bool is_empty_vlan = false;
1197 	uint16_t udp_dport = 0;
1198 
1199 	if (items == NULL)
1200 		return -1;
1201 	ret = mlx5_flow_validate_attributes(dev, attr, error);
1202 	if (ret < 0)
1203 		return ret;
1204 	for (; items->type != RTE_FLOW_ITEM_TYPE_END; items++) {
1205 		int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
1206 		int ret = 0;
1207 
1208 		switch (items->type) {
1209 		case RTE_FLOW_ITEM_TYPE_VOID:
1210 			break;
1211 		case RTE_FLOW_ITEM_TYPE_ETH:
1212 			ret = mlx5_flow_validate_item_eth(items, item_flags,
1213 							  false, error);
1214 			if (ret < 0)
1215 				return ret;
1216 			last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
1217 					     MLX5_FLOW_LAYER_OUTER_L2;
1218 			if (items->mask != NULL && items->spec != NULL) {
1219 				ether_type =
1220 					((const struct rte_flow_item_eth *)
1221 					 items->spec)->type;
1222 				ether_type &=
1223 					((const struct rte_flow_item_eth *)
1224 					 items->mask)->type;
1225 				if (ether_type == RTE_BE16(RTE_ETHER_TYPE_VLAN))
1226 					is_empty_vlan = true;
1227 				ether_type = rte_be_to_cpu_16(ether_type);
1228 			} else {
1229 				ether_type = 0;
1230 			}
1231 			break;
1232 		case RTE_FLOW_ITEM_TYPE_VLAN:
1233 			ret = mlx5_flow_validate_item_vlan(items, item_flags,
1234 							   dev, error);
1235 			if (ret < 0)
1236 				return ret;
1237 			last_item = tunnel ? (MLX5_FLOW_LAYER_INNER_L2 |
1238 					      MLX5_FLOW_LAYER_INNER_VLAN) :
1239 					     (MLX5_FLOW_LAYER_OUTER_L2 |
1240 					      MLX5_FLOW_LAYER_OUTER_VLAN);
1241 			if (items->mask != NULL && items->spec != NULL) {
1242 				ether_type =
1243 					((const struct rte_flow_item_vlan *)
1244 					 items->spec)->inner_type;
1245 				ether_type &=
1246 					((const struct rte_flow_item_vlan *)
1247 					 items->mask)->inner_type;
1248 				ether_type = rte_be_to_cpu_16(ether_type);
1249 			} else {
1250 				ether_type = 0;
1251 			}
1252 			is_empty_vlan = false;
1253 			break;
1254 		case RTE_FLOW_ITEM_TYPE_IPV4:
1255 			ret = mlx5_flow_validate_item_ipv4
1256 						(items, item_flags,
1257 						 last_item, ether_type, NULL,
1258 						 MLX5_ITEM_RANGE_NOT_ACCEPTED,
1259 						 error);
1260 			if (ret < 0)
1261 				return ret;
1262 			last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV4 :
1263 					     MLX5_FLOW_LAYER_OUTER_L3_IPV4;
1264 			if (items->mask != NULL &&
1265 			    ((const struct rte_flow_item_ipv4 *)
1266 			     items->mask)->hdr.next_proto_id) {
1267 				next_protocol =
1268 					((const struct rte_flow_item_ipv4 *)
1269 					 (items->spec))->hdr.next_proto_id;
1270 				next_protocol &=
1271 					((const struct rte_flow_item_ipv4 *)
1272 					 (items->mask))->hdr.next_proto_id;
1273 			} else {
1274 				/* Reset for inner layer. */
1275 				next_protocol = 0xff;
1276 			}
1277 			break;
1278 		case RTE_FLOW_ITEM_TYPE_IPV6:
1279 			ret = mlx5_flow_validate_item_ipv6(items, item_flags,
1280 							   last_item,
1281 							   ether_type, NULL,
1282 							   error);
1283 			if (ret < 0)
1284 				return ret;
1285 			last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV6 :
1286 					     MLX5_FLOW_LAYER_OUTER_L3_IPV6;
1287 			if (items->mask != NULL &&
1288 			    ((const struct rte_flow_item_ipv6 *)
1289 			     items->mask)->hdr.proto) {
1290 				next_protocol =
1291 					((const struct rte_flow_item_ipv6 *)
1292 					 items->spec)->hdr.proto;
1293 				next_protocol &=
1294 					((const struct rte_flow_item_ipv6 *)
1295 					 items->mask)->hdr.proto;
1296 			} else {
1297 				/* Reset for inner layer. */
1298 				next_protocol = 0xff;
1299 			}
1300 			break;
1301 		case RTE_FLOW_ITEM_TYPE_UDP:
1302 			ret = mlx5_flow_validate_item_udp(items, item_flags,
1303 							  next_protocol,
1304 							  error);
1305 			const struct rte_flow_item_udp *spec = items->spec;
1306 			const struct rte_flow_item_udp *mask = items->mask;
1307 			if (!mask)
1308 				mask = &rte_flow_item_udp_mask;
1309 			if (spec != NULL)
1310 				udp_dport = rte_be_to_cpu_16
1311 						(spec->hdr.dst_port &
1312 						 mask->hdr.dst_port);
1313 
1314 			if (ret < 0)
1315 				return ret;
1316 			last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L4_UDP :
1317 					     MLX5_FLOW_LAYER_OUTER_L4_UDP;
1318 			break;
1319 		case RTE_FLOW_ITEM_TYPE_TCP:
1320 			ret = mlx5_flow_validate_item_tcp
1321 						(items, item_flags,
1322 						 next_protocol,
1323 						 &rte_flow_item_tcp_mask,
1324 						 error);
1325 			if (ret < 0)
1326 				return ret;
1327 			last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L4_TCP :
1328 					     MLX5_FLOW_LAYER_OUTER_L4_TCP;
1329 			break;
1330 		case RTE_FLOW_ITEM_TYPE_VXLAN:
1331 			ret = mlx5_flow_validate_item_vxlan(dev, udp_dport,
1332 							    items, item_flags,
1333 							    attr, error);
1334 			if (ret < 0)
1335 				return ret;
1336 			last_item = MLX5_FLOW_LAYER_VXLAN;
1337 			break;
1338 		case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
1339 			ret = mlx5_flow_validate_item_vxlan_gpe(items,
1340 								item_flags,
1341 								dev, error);
1342 			if (ret < 0)
1343 				return ret;
1344 			last_item = MLX5_FLOW_LAYER_VXLAN_GPE;
1345 			break;
1346 		case RTE_FLOW_ITEM_TYPE_GRE:
1347 			ret = mlx5_flow_validate_item_gre(items, item_flags,
1348 							  next_protocol, error);
1349 			if (ret < 0)
1350 				return ret;
1351 			last_item = MLX5_FLOW_LAYER_GRE;
1352 			break;
1353 		case RTE_FLOW_ITEM_TYPE_MPLS:
1354 			ret = mlx5_flow_validate_item_mpls(dev, items,
1355 							   item_flags,
1356 							   last_item, error);
1357 			if (ret < 0)
1358 				return ret;
1359 			last_item = MLX5_FLOW_LAYER_MPLS;
1360 			break;
1361 		case RTE_FLOW_ITEM_TYPE_ICMP:
1362 		case RTE_FLOW_ITEM_TYPE_ICMP6:
1363 			return rte_flow_error_set(error, ENOTSUP,
1364 						  RTE_FLOW_ERROR_TYPE_ITEM,
1365 						  NULL, "ICMP/ICMP6 "
1366 						  "item not supported");
1367 		default:
1368 			return rte_flow_error_set(error, ENOTSUP,
1369 						  RTE_FLOW_ERROR_TYPE_ITEM,
1370 						  NULL, "item not supported");
1371 		}
1372 		item_flags |= last_item;
1373 	}
1374 	if (is_empty_vlan)
1375 		return rte_flow_error_set(error, ENOTSUP,
1376 						 RTE_FLOW_ERROR_TYPE_ITEM, NULL,
1377 		    "VLAN matching without vid specification is not supported");
1378 	for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
1379 		switch (actions->type) {
1380 		case RTE_FLOW_ACTION_TYPE_VOID:
1381 			break;
1382 		case RTE_FLOW_ACTION_TYPE_FLAG:
1383 			ret = mlx5_flow_validate_action_flag(action_flags,
1384 							     attr,
1385 							     error);
1386 			if (ret < 0)
1387 				return ret;
1388 			action_flags |= MLX5_FLOW_ACTION_FLAG;
1389 			break;
1390 		case RTE_FLOW_ACTION_TYPE_MARK:
1391 			ret = mlx5_flow_validate_action_mark(actions,
1392 							     action_flags,
1393 							     attr,
1394 							     error);
1395 			if (ret < 0)
1396 				return ret;
1397 			action_flags |= MLX5_FLOW_ACTION_MARK;
1398 			break;
1399 		case RTE_FLOW_ACTION_TYPE_DROP:
1400 			ret = mlx5_flow_validate_action_drop(action_flags,
1401 							     attr,
1402 							     error);
1403 			if (ret < 0)
1404 				return ret;
1405 			action_flags |= MLX5_FLOW_ACTION_DROP;
1406 			break;
1407 		case RTE_FLOW_ACTION_TYPE_QUEUE:
1408 			ret = mlx5_flow_validate_action_queue(actions,
1409 							      action_flags, dev,
1410 							      attr,
1411 							      error);
1412 			if (ret < 0)
1413 				return ret;
1414 			action_flags |= MLX5_FLOW_ACTION_QUEUE;
1415 			break;
1416 		case RTE_FLOW_ACTION_TYPE_RSS:
1417 			ret = mlx5_flow_validate_action_rss(actions,
1418 							    action_flags, dev,
1419 							    attr, item_flags,
1420 							    error);
1421 			if (ret < 0)
1422 				return ret;
1423 			action_flags |= MLX5_FLOW_ACTION_RSS;
1424 			break;
1425 		case RTE_FLOW_ACTION_TYPE_COUNT:
1426 			ret = mlx5_flow_validate_action_count(dev, attr, error);
1427 			if (ret < 0)
1428 				return ret;
1429 			action_flags |= MLX5_FLOW_ACTION_COUNT;
1430 			break;
1431 		default:
1432 			return rte_flow_error_set(error, ENOTSUP,
1433 						  RTE_FLOW_ERROR_TYPE_ACTION,
1434 						  actions,
1435 						  "action not supported");
1436 		}
1437 	}
1438 	/*
1439 	 * Validate the drop action mutual exclusion with other actions.
1440 	 * Drop action is mutually-exclusive with any other action, except for
1441 	 * Count action.
1442 	 */
1443 	if ((action_flags & MLX5_FLOW_ACTION_DROP) &&
1444 	    (action_flags & ~(MLX5_FLOW_ACTION_DROP | MLX5_FLOW_ACTION_COUNT)))
1445 		return rte_flow_error_set(error, EINVAL,
1446 					  RTE_FLOW_ERROR_TYPE_ACTION, NULL,
1447 					  "Drop action is mutually-exclusive "
1448 					  "with any other action, except for "
1449 					  "Count action");
1450 	if (!(action_flags & MLX5_FLOW_FATE_ACTIONS))
1451 		return rte_flow_error_set(error, EINVAL,
1452 					  RTE_FLOW_ERROR_TYPE_ACTION, actions,
1453 					  "no fate action is found");
1454 	return 0;
1455 }
1456 
1457 /**
1458  * Calculate the required bytes that are needed for the action part of the verbs
1459  * flow.
1460  *
1461  * @param[in] actions
1462  *   Pointer to the list of actions.
1463  *
1464  * @return
1465  *   The size of the memory needed for all actions.
1466  */
1467 static int
1468 flow_verbs_get_actions_size(const struct rte_flow_action actions[])
1469 {
1470 	int size = 0;
1471 
1472 	for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
1473 		switch (actions->type) {
1474 		case RTE_FLOW_ACTION_TYPE_VOID:
1475 			break;
1476 		case RTE_FLOW_ACTION_TYPE_FLAG:
1477 			size += sizeof(struct ibv_flow_spec_action_tag);
1478 			break;
1479 		case RTE_FLOW_ACTION_TYPE_MARK:
1480 			size += sizeof(struct ibv_flow_spec_action_tag);
1481 			break;
1482 		case RTE_FLOW_ACTION_TYPE_DROP:
1483 			size += sizeof(struct ibv_flow_spec_action_drop);
1484 			break;
1485 		case RTE_FLOW_ACTION_TYPE_QUEUE:
1486 			break;
1487 		case RTE_FLOW_ACTION_TYPE_RSS:
1488 			break;
1489 		case RTE_FLOW_ACTION_TYPE_COUNT:
1490 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) || \
1491 	defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1492 			size += sizeof(struct ibv_flow_spec_counter_action);
1493 #endif
1494 			break;
1495 		default:
1496 			break;
1497 		}
1498 	}
1499 	return size;
1500 }
1501 
1502 /**
1503  * Calculate the required bytes that are needed for the item part of the verbs
1504  * flow.
1505  *
1506  * @param[in] items
1507  *   Pointer to the list of items.
1508  *
1509  * @return
1510  *   The size of the memory needed for all items.
1511  */
1512 static int
1513 flow_verbs_get_items_size(const struct rte_flow_item items[])
1514 {
1515 	int size = 0;
1516 
1517 	for (; items->type != RTE_FLOW_ITEM_TYPE_END; items++) {
1518 		switch (items->type) {
1519 		case RTE_FLOW_ITEM_TYPE_VOID:
1520 			break;
1521 		case RTE_FLOW_ITEM_TYPE_ETH:
1522 			size += sizeof(struct ibv_flow_spec_eth);
1523 			break;
1524 		case RTE_FLOW_ITEM_TYPE_VLAN:
1525 			size += sizeof(struct ibv_flow_spec_eth);
1526 			break;
1527 		case RTE_FLOW_ITEM_TYPE_IPV4:
1528 			size += sizeof(struct ibv_flow_spec_ipv4_ext);
1529 			break;
1530 		case RTE_FLOW_ITEM_TYPE_IPV6:
1531 			size += sizeof(struct ibv_flow_spec_ipv6);
1532 			break;
1533 		case RTE_FLOW_ITEM_TYPE_UDP:
1534 			size += sizeof(struct ibv_flow_spec_tcp_udp);
1535 			break;
1536 		case RTE_FLOW_ITEM_TYPE_TCP:
1537 			size += sizeof(struct ibv_flow_spec_tcp_udp);
1538 			break;
1539 		case RTE_FLOW_ITEM_TYPE_VXLAN:
1540 			size += sizeof(struct ibv_flow_spec_tunnel);
1541 			break;
1542 		case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
1543 			size += sizeof(struct ibv_flow_spec_tunnel);
1544 			break;
1545 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1546 		case RTE_FLOW_ITEM_TYPE_GRE:
1547 			size += sizeof(struct ibv_flow_spec_gre);
1548 			break;
1549 		case RTE_FLOW_ITEM_TYPE_MPLS:
1550 			size += sizeof(struct ibv_flow_spec_mpls);
1551 			break;
1552 #else
1553 		case RTE_FLOW_ITEM_TYPE_GRE:
1554 			size += sizeof(struct ibv_flow_spec_tunnel);
1555 			break;
1556 #endif
1557 		default:
1558 			break;
1559 		}
1560 	}
1561 	return size;
1562 }
1563 
1564 /**
1565  * Internal preparation function. Allocate mlx5_flow with the required size.
1566  * The required size is calculate based on the actions and items. This function
1567  * also returns the detected actions and items for later use.
1568  *
1569  * @param[in] dev
1570  *   Pointer to Ethernet device.
1571  * @param[in] attr
1572  *   Pointer to the flow attributes.
1573  * @param[in] items
1574  *   Pointer to the list of items.
1575  * @param[in] actions
1576  *   Pointer to the list of actions.
1577  * @param[out] error
1578  *   Pointer to the error structure.
1579  *
1580  * @return
1581  *   Pointer to mlx5_flow object on success, otherwise NULL and rte_errno
1582  *   is set.
1583  */
1584 static struct mlx5_flow *
1585 flow_verbs_prepare(struct rte_eth_dev *dev,
1586 		   const struct rte_flow_attr *attr __rte_unused,
1587 		   const struct rte_flow_item items[],
1588 		   const struct rte_flow_action actions[],
1589 		   struct rte_flow_error *error)
1590 {
1591 	size_t size = 0;
1592 	uint32_t handle_idx = 0;
1593 	struct mlx5_flow *dev_flow;
1594 	struct mlx5_flow_handle *dev_handle;
1595 	struct mlx5_priv *priv = dev->data->dev_private;
1596 	struct mlx5_flow_workspace *wks = mlx5_flow_get_thread_workspace();
1597 
1598 	MLX5_ASSERT(wks);
1599 	size += flow_verbs_get_actions_size(actions);
1600 	size += flow_verbs_get_items_size(items);
1601 	if (size > MLX5_VERBS_MAX_SPEC_ACT_SIZE) {
1602 		rte_flow_error_set(error, E2BIG,
1603 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1604 				   "Verbs spec/action size too large");
1605 		return NULL;
1606 	}
1607 	/* In case of corrupting the memory. */
1608 	if (wks->flow_idx >= MLX5_NUM_MAX_DEV_FLOWS) {
1609 		rte_flow_error_set(error, ENOSPC,
1610 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1611 				   "not free temporary device flow");
1612 		return NULL;
1613 	}
1614 	dev_handle = mlx5_ipool_zmalloc(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW],
1615 				   &handle_idx);
1616 	if (!dev_handle) {
1617 		rte_flow_error_set(error, ENOMEM,
1618 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1619 				   "not enough memory to create flow handle");
1620 		return NULL;
1621 	}
1622 	MLX5_ASSERT(wks->flow_idx + 1 < RTE_DIM(wks->flows));
1623 	dev_flow = &wks->flows[wks->flow_idx++];
1624 	dev_flow->handle = dev_handle;
1625 	dev_flow->handle_idx = handle_idx;
1626 	/* Memcpy is used, only size needs to be cleared to 0. */
1627 	dev_flow->verbs.size = 0;
1628 	dev_flow->verbs.attr.num_of_specs = 0;
1629 	dev_flow->ingress = attr->ingress;
1630 	dev_flow->hash_fields = 0;
1631 	/* Need to set transfer attribute: not supported in Verbs mode. */
1632 	return dev_flow;
1633 }
1634 
1635 /**
1636  * Fill the flow with verb spec.
1637  *
1638  * @param[in] dev
1639  *   Pointer to Ethernet device.
1640  * @param[in, out] dev_flow
1641  *   Pointer to the mlx5 flow.
1642  * @param[in] attr
1643  *   Pointer to the flow attributes.
1644  * @param[in] items
1645  *   Pointer to the list of items.
1646  * @param[in] actions
1647  *   Pointer to the list of actions.
1648  * @param[out] error
1649  *   Pointer to the error structure.
1650  *
1651  * @return
1652  *   0 on success, else a negative errno value otherwise and rte_errno is set.
1653  */
1654 static int
1655 flow_verbs_translate(struct rte_eth_dev *dev,
1656 		     struct mlx5_flow *dev_flow,
1657 		     const struct rte_flow_attr *attr,
1658 		     const struct rte_flow_item items[],
1659 		     const struct rte_flow_action actions[],
1660 		     struct rte_flow_error *error)
1661 {
1662 	uint64_t item_flags = 0;
1663 	uint64_t action_flags = 0;
1664 	uint64_t priority = attr->priority;
1665 	uint32_t subpriority = 0;
1666 	struct mlx5_priv *priv = dev->data->dev_private;
1667 	struct mlx5_flow_workspace *wks = mlx5_flow_get_thread_workspace();
1668 	struct mlx5_flow_rss_desc *rss_desc;
1669 
1670 	MLX5_ASSERT(wks);
1671 	rss_desc = &wks->rss_desc;
1672 	if (priority == MLX5_FLOW_LOWEST_PRIO_INDICATOR)
1673 		priority = priv->sh->flow_max_priority - 1;
1674 	for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
1675 		int ret;
1676 
1677 		switch (actions->type) {
1678 		case RTE_FLOW_ACTION_TYPE_VOID:
1679 			break;
1680 		case RTE_FLOW_ACTION_TYPE_FLAG:
1681 			flow_verbs_translate_action_flag(dev_flow, actions);
1682 			action_flags |= MLX5_FLOW_ACTION_FLAG;
1683 			dev_flow->handle->mark = 1;
1684 			break;
1685 		case RTE_FLOW_ACTION_TYPE_MARK:
1686 			flow_verbs_translate_action_mark(dev_flow, actions);
1687 			action_flags |= MLX5_FLOW_ACTION_MARK;
1688 			dev_flow->handle->mark = 1;
1689 			break;
1690 		case RTE_FLOW_ACTION_TYPE_DROP:
1691 			flow_verbs_translate_action_drop(dev_flow, actions);
1692 			action_flags |= MLX5_FLOW_ACTION_DROP;
1693 			dev_flow->handle->fate_action = MLX5_FLOW_FATE_DROP;
1694 			break;
1695 		case RTE_FLOW_ACTION_TYPE_QUEUE:
1696 			flow_verbs_translate_action_queue(rss_desc, actions);
1697 			action_flags |= MLX5_FLOW_ACTION_QUEUE;
1698 			dev_flow->handle->fate_action = MLX5_FLOW_FATE_QUEUE;
1699 			break;
1700 		case RTE_FLOW_ACTION_TYPE_RSS:
1701 			flow_verbs_translate_action_rss(rss_desc, actions);
1702 			action_flags |= MLX5_FLOW_ACTION_RSS;
1703 			dev_flow->handle->fate_action = MLX5_FLOW_FATE_QUEUE;
1704 			break;
1705 		case RTE_FLOW_ACTION_TYPE_COUNT:
1706 			ret = flow_verbs_translate_action_count(dev_flow,
1707 								actions,
1708 								dev, error);
1709 			if (ret < 0)
1710 				return ret;
1711 			action_flags |= MLX5_FLOW_ACTION_COUNT;
1712 			break;
1713 		default:
1714 			return rte_flow_error_set(error, ENOTSUP,
1715 						  RTE_FLOW_ERROR_TYPE_ACTION,
1716 						  actions,
1717 						  "action not supported");
1718 		}
1719 	}
1720 	dev_flow->act_flags = action_flags;
1721 	for (; items->type != RTE_FLOW_ITEM_TYPE_END; items++) {
1722 		int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
1723 
1724 		switch (items->type) {
1725 		case RTE_FLOW_ITEM_TYPE_VOID:
1726 			break;
1727 		case RTE_FLOW_ITEM_TYPE_ETH:
1728 			flow_verbs_translate_item_eth(dev_flow, items,
1729 						      item_flags);
1730 			subpriority = MLX5_PRIORITY_MAP_L2;
1731 			item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
1732 					       MLX5_FLOW_LAYER_OUTER_L2;
1733 			break;
1734 		case RTE_FLOW_ITEM_TYPE_VLAN:
1735 			flow_verbs_translate_item_vlan(dev_flow, items,
1736 						       item_flags);
1737 			subpriority = MLX5_PRIORITY_MAP_L2;
1738 			item_flags |= tunnel ? (MLX5_FLOW_LAYER_INNER_L2 |
1739 						MLX5_FLOW_LAYER_INNER_VLAN) :
1740 					       (MLX5_FLOW_LAYER_OUTER_L2 |
1741 						MLX5_FLOW_LAYER_OUTER_VLAN);
1742 			break;
1743 		case RTE_FLOW_ITEM_TYPE_IPV4:
1744 			flow_verbs_translate_item_ipv4(dev_flow, items,
1745 						       item_flags);
1746 			subpriority = MLX5_PRIORITY_MAP_L3;
1747 			dev_flow->hash_fields |=
1748 				mlx5_flow_hashfields_adjust
1749 					(rss_desc, tunnel,
1750 					 MLX5_IPV4_LAYER_TYPES,
1751 					 MLX5_IPV4_IBV_RX_HASH);
1752 			item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV4 :
1753 					       MLX5_FLOW_LAYER_OUTER_L3_IPV4;
1754 			break;
1755 		case RTE_FLOW_ITEM_TYPE_IPV6:
1756 			flow_verbs_translate_item_ipv6(dev_flow, items,
1757 						       item_flags);
1758 			subpriority = MLX5_PRIORITY_MAP_L3;
1759 			dev_flow->hash_fields |=
1760 				mlx5_flow_hashfields_adjust
1761 					(rss_desc, tunnel,
1762 					 MLX5_IPV6_LAYER_TYPES,
1763 					 MLX5_IPV6_IBV_RX_HASH);
1764 			item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV6 :
1765 					       MLX5_FLOW_LAYER_OUTER_L3_IPV6;
1766 			break;
1767 		case RTE_FLOW_ITEM_TYPE_TCP:
1768 			flow_verbs_translate_item_tcp(dev_flow, items,
1769 						      item_flags);
1770 			subpriority = MLX5_PRIORITY_MAP_L4;
1771 			if (dev_flow->hash_fields != 0)
1772 				dev_flow->hash_fields |=
1773 					mlx5_flow_hashfields_adjust
1774 					(rss_desc, tunnel, RTE_ETH_RSS_TCP,
1775 					 (IBV_RX_HASH_SRC_PORT_TCP |
1776 					  IBV_RX_HASH_DST_PORT_TCP));
1777 			item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L4_TCP :
1778 					       MLX5_FLOW_LAYER_OUTER_L4_TCP;
1779 			break;
1780 		case RTE_FLOW_ITEM_TYPE_UDP:
1781 			flow_verbs_translate_item_udp(dev_flow, items,
1782 						      item_flags);
1783 			subpriority = MLX5_PRIORITY_MAP_L4;
1784 			if (dev_flow->hash_fields != 0)
1785 				dev_flow->hash_fields |=
1786 					mlx5_flow_hashfields_adjust
1787 					(rss_desc, tunnel, RTE_ETH_RSS_UDP,
1788 					 (IBV_RX_HASH_SRC_PORT_UDP |
1789 					  IBV_RX_HASH_DST_PORT_UDP));
1790 			item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L4_UDP :
1791 					       MLX5_FLOW_LAYER_OUTER_L4_UDP;
1792 			break;
1793 		case RTE_FLOW_ITEM_TYPE_VXLAN:
1794 			flow_verbs_translate_item_vxlan(dev_flow, items,
1795 							item_flags);
1796 			subpriority = MLX5_TUNNEL_PRIO_GET(rss_desc);
1797 			item_flags |= MLX5_FLOW_LAYER_VXLAN;
1798 			break;
1799 		case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
1800 			flow_verbs_translate_item_vxlan_gpe(dev_flow, items,
1801 							    item_flags);
1802 			subpriority = MLX5_TUNNEL_PRIO_GET(rss_desc);
1803 			item_flags |= MLX5_FLOW_LAYER_VXLAN_GPE;
1804 			break;
1805 		case RTE_FLOW_ITEM_TYPE_GRE:
1806 			flow_verbs_translate_item_gre(dev_flow, items,
1807 						      item_flags);
1808 			subpriority = MLX5_TUNNEL_PRIO_GET(rss_desc);
1809 			item_flags |= MLX5_FLOW_LAYER_GRE;
1810 			break;
1811 		case RTE_FLOW_ITEM_TYPE_MPLS:
1812 			flow_verbs_translate_item_mpls(dev_flow, items,
1813 						       item_flags);
1814 			subpriority = MLX5_TUNNEL_PRIO_GET(rss_desc);
1815 			item_flags |= MLX5_FLOW_LAYER_MPLS;
1816 			break;
1817 		default:
1818 			return rte_flow_error_set(error, ENOTSUP,
1819 						  RTE_FLOW_ERROR_TYPE_ITEM,
1820 						  NULL, "item not supported");
1821 		}
1822 	}
1823 	dev_flow->handle->layers = item_flags;
1824 	/* Other members of attr will be ignored. */
1825 	dev_flow->verbs.attr.priority =
1826 		mlx5_flow_adjust_priority(dev, priority, subpriority);
1827 	dev_flow->verbs.attr.port = (uint8_t)priv->dev_port;
1828 	return 0;
1829 }
1830 
1831 /**
1832  * Remove the flow from the NIC but keeps it in memory.
1833  *
1834  * @param[in] dev
1835  *   Pointer to the Ethernet device structure.
1836  * @param[in, out] flow
1837  *   Pointer to flow structure.
1838  */
1839 static void
1840 flow_verbs_remove(struct rte_eth_dev *dev, struct rte_flow *flow)
1841 {
1842 	struct mlx5_priv *priv = dev->data->dev_private;
1843 	struct mlx5_flow_handle *handle;
1844 	uint32_t handle_idx;
1845 
1846 	if (!flow)
1847 		return;
1848 	SILIST_FOREACH(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW], flow->dev_handles,
1849 		       handle_idx, handle, next) {
1850 		if (handle->drv_flow) {
1851 			claim_zero(mlx5_glue->destroy_flow(handle->drv_flow));
1852 			handle->drv_flow = NULL;
1853 		}
1854 		/* hrxq is union, don't touch it only the flag is set. */
1855 		if (handle->rix_hrxq &&
1856 		    handle->fate_action == MLX5_FLOW_FATE_QUEUE) {
1857 			mlx5_hrxq_release(dev, handle->rix_hrxq);
1858 			handle->rix_hrxq = 0;
1859 		}
1860 		if (handle->vf_vlan.tag && handle->vf_vlan.created)
1861 			mlx5_vlan_vmwa_release(dev, &handle->vf_vlan);
1862 	}
1863 }
1864 
1865 /**
1866  * Remove the flow from the NIC and the memory.
1867  *
1868  * @param[in] dev
1869  *   Pointer to the Ethernet device structure.
1870  * @param[in, out] flow
1871  *   Pointer to flow structure.
1872  */
1873 static void
1874 flow_verbs_destroy(struct rte_eth_dev *dev, struct rte_flow *flow)
1875 {
1876 	struct mlx5_priv *priv = dev->data->dev_private;
1877 	struct mlx5_flow_handle *handle;
1878 
1879 	if (!flow)
1880 		return;
1881 	flow_verbs_remove(dev, flow);
1882 	while (flow->dev_handles) {
1883 		uint32_t tmp_idx = flow->dev_handles;
1884 
1885 		handle = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW],
1886 				   tmp_idx);
1887 		if (!handle)
1888 			return;
1889 		flow->dev_handles = handle->next.next;
1890 		mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW],
1891 			   tmp_idx);
1892 	}
1893 	if (flow->counter) {
1894 		flow_verbs_counter_release(dev, flow->counter);
1895 		flow->counter = 0;
1896 	}
1897 }
1898 
1899 /**
1900  * Apply the flow to the NIC.
1901  *
1902  * @param[in] dev
1903  *   Pointer to the Ethernet device structure.
1904  * @param[in, out] flow
1905  *   Pointer to flow structure.
1906  * @param[out] error
1907  *   Pointer to error structure.
1908  *
1909  * @return
1910  *   0 on success, a negative errno value otherwise and rte_errno is set.
1911  */
1912 static int
1913 flow_verbs_apply(struct rte_eth_dev *dev, struct rte_flow *flow,
1914 		 struct rte_flow_error *error)
1915 {
1916 	struct mlx5_priv *priv = dev->data->dev_private;
1917 	struct mlx5_flow_handle *handle;
1918 	struct mlx5_flow *dev_flow;
1919 	struct mlx5_hrxq *hrxq;
1920 	uint32_t dev_handles;
1921 	int err;
1922 	int idx;
1923 	struct mlx5_flow_workspace *wks = mlx5_flow_get_thread_workspace();
1924 
1925 	MLX5_ASSERT(wks);
1926 	for (idx = wks->flow_idx - 1; idx >= 0; idx--) {
1927 		dev_flow = &wks->flows[idx];
1928 		handle = dev_flow->handle;
1929 		if (handle->fate_action == MLX5_FLOW_FATE_DROP) {
1930 			MLX5_ASSERT(priv->drop_queue.hrxq);
1931 			hrxq = priv->drop_queue.hrxq;
1932 		} else {
1933 			uint32_t hrxq_idx;
1934 			struct mlx5_flow_rss_desc *rss_desc = &wks->rss_desc;
1935 
1936 			MLX5_ASSERT(rss_desc->queue_num);
1937 			rss_desc->key_len = MLX5_RSS_HASH_KEY_LEN;
1938 			rss_desc->hash_fields = dev_flow->hash_fields;
1939 			rss_desc->tunnel = !!(handle->layers &
1940 					      MLX5_FLOW_LAYER_TUNNEL);
1941 			rss_desc->shared_rss = 0;
1942 			hrxq_idx = mlx5_hrxq_get(dev, rss_desc);
1943 			hrxq = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_HRXQ],
1944 					      hrxq_idx);
1945 			if (!hrxq) {
1946 				rte_flow_error_set
1947 					(error, rte_errno,
1948 					 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1949 					 "cannot get hash queue");
1950 				goto error;
1951 			}
1952 			handle->rix_hrxq = hrxq_idx;
1953 		}
1954 		MLX5_ASSERT(hrxq);
1955 		handle->drv_flow = mlx5_glue->create_flow
1956 					(hrxq->qp, &dev_flow->verbs.attr);
1957 		if (!handle->drv_flow) {
1958 			rte_flow_error_set(error, errno,
1959 					   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1960 					   NULL,
1961 					   "hardware refuses to create flow");
1962 			goto error;
1963 		}
1964 		if (priv->vmwa_context &&
1965 		    handle->vf_vlan.tag && !handle->vf_vlan.created) {
1966 			/*
1967 			 * The rule contains the VLAN pattern.
1968 			 * For VF we are going to create VLAN
1969 			 * interface to make hypervisor set correct
1970 			 * e-Switch vport context.
1971 			 */
1972 			mlx5_vlan_vmwa_acquire(dev, &handle->vf_vlan);
1973 		}
1974 	}
1975 	return 0;
1976 error:
1977 	err = rte_errno; /* Save rte_errno before cleanup. */
1978 	SILIST_FOREACH(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW], flow->dev_handles,
1979 		       dev_handles, handle, next) {
1980 		/* hrxq is union, don't touch it only the flag is set. */
1981 		if (handle->rix_hrxq &&
1982 		    handle->fate_action == MLX5_FLOW_FATE_QUEUE) {
1983 			mlx5_hrxq_release(dev, handle->rix_hrxq);
1984 			handle->rix_hrxq = 0;
1985 		}
1986 		if (handle->vf_vlan.tag && handle->vf_vlan.created)
1987 			mlx5_vlan_vmwa_release(dev, &handle->vf_vlan);
1988 	}
1989 	rte_errno = err; /* Restore rte_errno. */
1990 	return -rte_errno;
1991 }
1992 
1993 /**
1994  * Query a flow.
1995  *
1996  * @see rte_flow_query()
1997  * @see rte_flow_ops
1998  */
1999 static int
2000 flow_verbs_query(struct rte_eth_dev *dev,
2001 		 struct rte_flow *flow,
2002 		 const struct rte_flow_action *actions,
2003 		 void *data,
2004 		 struct rte_flow_error *error)
2005 {
2006 	int ret = -EINVAL;
2007 
2008 	for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
2009 		switch (actions->type) {
2010 		case RTE_FLOW_ACTION_TYPE_VOID:
2011 			break;
2012 		case RTE_FLOW_ACTION_TYPE_COUNT:
2013 			ret = flow_verbs_counter_query(dev, flow, data, error);
2014 			break;
2015 		default:
2016 			return rte_flow_error_set(error, ENOTSUP,
2017 						  RTE_FLOW_ERROR_TYPE_ACTION,
2018 						  actions,
2019 						  "action not supported");
2020 		}
2021 	}
2022 	return ret;
2023 }
2024 
2025 static int
2026 flow_verbs_sync_domain(struct rte_eth_dev *dev, uint32_t domains,
2027 		       uint32_t flags)
2028 {
2029 	RTE_SET_USED(dev);
2030 	RTE_SET_USED(domains);
2031 	RTE_SET_USED(flags);
2032 
2033 	return 0;
2034 }
2035 
2036 const struct mlx5_flow_driver_ops mlx5_flow_verbs_drv_ops = {
2037 	.validate = flow_verbs_validate,
2038 	.prepare = flow_verbs_prepare,
2039 	.translate = flow_verbs_translate,
2040 	.apply = flow_verbs_apply,
2041 	.remove = flow_verbs_remove,
2042 	.destroy = flow_verbs_destroy,
2043 	.query = flow_verbs_query,
2044 	.sync_domain = flow_verbs_sync_domain,
2045 	.discover_priorities = flow_verbs_discover_priorities,
2046 };
2047