xref: /dpdk/drivers/net/mlx5/mlx5_flow_verbs.c (revision 30a1de105a5f40d77b344a891c4a68f79e815c43)
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 	static const struct rte_flow_item_gre empty_gre = {0,};
911 	const struct rte_flow_item_gre *spec = item->spec;
912 	const struct rte_flow_item_gre *mask = item->mask;
913 	unsigned int size = sizeof(struct ibv_flow_spec_gre);
914 	struct ibv_flow_spec_gre tunnel = {
915 		.type = IBV_FLOW_SPEC_GRE,
916 		.size = size,
917 	};
918 
919 	if (!spec) {
920 		spec = &empty_gre;
921 		mask = &empty_gre;
922 	} else {
923 		if (!mask)
924 			mask = &rte_flow_item_gre_mask;
925 	}
926 	tunnel.val.c_ks_res0_ver = spec->c_rsvd0_ver;
927 	tunnel.val.protocol = spec->protocol;
928 	tunnel.mask.c_ks_res0_ver = mask->c_rsvd0_ver;
929 	tunnel.mask.protocol = mask->protocol;
930 	/* Remove unwanted bits from values. */
931 	tunnel.val.c_ks_res0_ver &= tunnel.mask.c_ks_res0_ver;
932 	tunnel.val.key &= tunnel.mask.key;
933 	if (tunnel.mask.protocol) {
934 		tunnel.val.protocol &= tunnel.mask.protocol;
935 	} else {
936 		tunnel.val.protocol = mlx5_translate_tunnel_etypes(item_flags);
937 		if (tunnel.val.protocol) {
938 			tunnel.mask.protocol = 0xFFFF;
939 			tunnel.val.protocol =
940 				rte_cpu_to_be_16(tunnel.val.protocol);
941 		}
942 	}
943 #endif
944 	if (item_flags & MLX5_FLOW_LAYER_OUTER_L3_IPV4)
945 		flow_verbs_item_gre_ip_protocol_update(&verbs->attr,
946 						       IBV_FLOW_SPEC_IPV4_EXT,
947 						       IPPROTO_GRE);
948 	else
949 		flow_verbs_item_gre_ip_protocol_update(&verbs->attr,
950 						       IBV_FLOW_SPEC_IPV6,
951 						       IPPROTO_GRE);
952 	flow_verbs_spec_add(verbs, &tunnel, size);
953 }
954 
955 /**
956  * Convert the @p action into a Verbs specification. This function assumes that
957  * the input is valid and that there is space to insert the requested action
958  * into the flow. This function also return the action that was added.
959  *
960  * @param[in, out] dev_flow
961  *   Pointer to dev_flow structure.
962  * @param[in] item
963  *   Item specification.
964  * @param[in] item_flags
965  *   Parsed item flags.
966  */
967 static void
968 flow_verbs_translate_item_mpls(struct mlx5_flow *dev_flow __rte_unused,
969 			       const struct rte_flow_item *item __rte_unused,
970 			       uint64_t item_flags __rte_unused)
971 {
972 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
973 	const struct rte_flow_item_mpls *spec = item->spec;
974 	const struct rte_flow_item_mpls *mask = item->mask;
975 	unsigned int size = sizeof(struct ibv_flow_spec_mpls);
976 	struct ibv_flow_spec_mpls mpls = {
977 		.type = IBV_FLOW_SPEC_MPLS,
978 		.size = size,
979 	};
980 
981 	if (!mask)
982 		mask = &rte_flow_item_mpls_mask;
983 	if (spec) {
984 		memcpy(&mpls.val.label, spec, sizeof(mpls.val.label));
985 		memcpy(&mpls.mask.label, mask, sizeof(mpls.mask.label));
986 		/* Remove unwanted bits from values.  */
987 		mpls.val.label &= mpls.mask.label;
988 	}
989 	flow_verbs_spec_add(&dev_flow->verbs, &mpls, size);
990 #endif
991 }
992 
993 /**
994  * Convert the @p action into a Verbs specification. This function assumes that
995  * the input is valid and that there is space to insert the requested action
996  * into the flow.
997  *
998  * @param[in] dev_flow
999  *   Pointer to mlx5_flow.
1000  * @param[in] action
1001  *   Action configuration.
1002  */
1003 static void
1004 flow_verbs_translate_action_drop
1005 	(struct mlx5_flow *dev_flow,
1006 	 const struct rte_flow_action *action __rte_unused)
1007 {
1008 	unsigned int size = sizeof(struct ibv_flow_spec_action_drop);
1009 	struct ibv_flow_spec_action_drop drop = {
1010 			.type = IBV_FLOW_SPEC_ACTION_DROP,
1011 			.size = size,
1012 	};
1013 
1014 	flow_verbs_spec_add(&dev_flow->verbs, &drop, size);
1015 }
1016 
1017 /**
1018  * Convert the @p action into a Verbs specification. This function assumes that
1019  * the input is valid and that there is space to insert the requested action
1020  * into the flow.
1021  *
1022  * @param[in] rss_desc
1023  *   Pointer to mlx5_flow_rss_desc.
1024  * @param[in] action
1025  *   Action configuration.
1026  */
1027 static void
1028 flow_verbs_translate_action_queue(struct mlx5_flow_rss_desc *rss_desc,
1029 				  const struct rte_flow_action *action)
1030 {
1031 	const struct rte_flow_action_queue *queue = action->conf;
1032 
1033 	rss_desc->queue[0] = queue->index;
1034 	rss_desc->queue_num = 1;
1035 }
1036 
1037 /**
1038  * Convert the @p action into a Verbs specification. This function assumes that
1039  * the input is valid and that there is space to insert the requested action
1040  * into the flow.
1041  *
1042  * @param[in] rss_desc
1043  *   Pointer to mlx5_flow_rss_desc.
1044  * @param[in] action
1045  *   Action configuration.
1046  */
1047 static void
1048 flow_verbs_translate_action_rss(struct mlx5_flow_rss_desc *rss_desc,
1049 				const struct rte_flow_action *action)
1050 {
1051 	const struct rte_flow_action_rss *rss = action->conf;
1052 	const uint8_t *rss_key;
1053 
1054 	memcpy(rss_desc->queue, rss->queue, rss->queue_num * sizeof(uint16_t));
1055 	rss_desc->queue_num = rss->queue_num;
1056 	/* NULL RSS key indicates default RSS key. */
1057 	rss_key = !rss->key ? rss_hash_default_key : rss->key;
1058 	memcpy(rss_desc->key, rss_key, MLX5_RSS_HASH_KEY_LEN);
1059 	/*
1060 	 * rss->level and rss.types should be set in advance when expanding
1061 	 * items for RSS.
1062 	 */
1063 }
1064 
1065 /**
1066  * Convert the @p action into a Verbs specification. This function assumes that
1067  * the input is valid and that there is space to insert the requested action
1068  * into the flow.
1069  *
1070  * @param[in] dev_flow
1071  *   Pointer to mlx5_flow.
1072  * @param[in] action
1073  *   Action configuration.
1074  */
1075 static void
1076 flow_verbs_translate_action_flag
1077 	(struct mlx5_flow *dev_flow,
1078 	 const struct rte_flow_action *action __rte_unused)
1079 {
1080 	unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
1081 	struct ibv_flow_spec_action_tag tag = {
1082 		.type = IBV_FLOW_SPEC_ACTION_TAG,
1083 		.size = size,
1084 		.tag_id = mlx5_flow_mark_set(MLX5_FLOW_MARK_DEFAULT),
1085 	};
1086 
1087 	flow_verbs_spec_add(&dev_flow->verbs, &tag, size);
1088 }
1089 
1090 /**
1091  * Convert the @p action into a Verbs specification. This function assumes that
1092  * the input is valid and that there is space to insert the requested action
1093  * into the flow.
1094  *
1095  * @param[in] dev_flow
1096  *   Pointer to mlx5_flow.
1097  * @param[in] action
1098  *   Action configuration.
1099  */
1100 static void
1101 flow_verbs_translate_action_mark(struct mlx5_flow *dev_flow,
1102 				 const struct rte_flow_action *action)
1103 {
1104 	const struct rte_flow_action_mark *mark = action->conf;
1105 	unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
1106 	struct ibv_flow_spec_action_tag tag = {
1107 		.type = IBV_FLOW_SPEC_ACTION_TAG,
1108 		.size = size,
1109 		.tag_id = mlx5_flow_mark_set(mark->id),
1110 	};
1111 
1112 	flow_verbs_spec_add(&dev_flow->verbs, &tag, size);
1113 }
1114 
1115 /**
1116  * Convert the @p action into a Verbs specification. This function assumes that
1117  * the input is valid and that there is space to insert the requested action
1118  * into the flow.
1119  *
1120  * @param[in] dev
1121  *   Pointer to the Ethernet device structure.
1122  * @param[in] action
1123  *   Action configuration.
1124  * @param[in] dev_flow
1125  *   Pointer to mlx5_flow.
1126  * @param[out] error
1127  *   Pointer to error structure.
1128  *
1129  * @return
1130  *   0 On success else a negative errno value is returned and rte_errno is set.
1131  */
1132 static int
1133 flow_verbs_translate_action_count(struct mlx5_flow *dev_flow,
1134 				  const struct rte_flow_action *action,
1135 				  struct rte_eth_dev *dev,
1136 				  struct rte_flow_error *error)
1137 {
1138 	const struct rte_flow_action_count *count = action->conf;
1139 	struct rte_flow *flow = dev_flow->flow;
1140 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) || \
1141 	defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1142 	struct mlx5_flow_counter_pool *pool;
1143 	struct mlx5_flow_counter *cnt = NULL;
1144 	unsigned int size = sizeof(struct ibv_flow_spec_counter_action);
1145 	struct ibv_flow_spec_counter_action counter = {
1146 		.type = IBV_FLOW_SPEC_ACTION_COUNT,
1147 		.size = size,
1148 	};
1149 #endif
1150 
1151 	if (!flow->counter) {
1152 		flow->counter = flow_verbs_counter_new(dev, count->id);
1153 		if (!flow->counter)
1154 			return rte_flow_error_set(error, rte_errno,
1155 						  RTE_FLOW_ERROR_TYPE_ACTION,
1156 						  action,
1157 						  "cannot get counter"
1158 						  " context.");
1159 	}
1160 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42)
1161 	cnt = flow_verbs_counter_get_by_idx(dev, flow->counter, &pool);
1162 	counter.counter_set_handle =
1163 		((struct ibv_counter_set *)cnt->dcs_when_active)->handle;
1164 	flow_verbs_spec_add(&dev_flow->verbs, &counter, size);
1165 #elif defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1166 	cnt = flow_verbs_counter_get_by_idx(dev, flow->counter, &pool);
1167 	counter.counters = (struct ibv_counters *)cnt->dcs_when_active;
1168 	flow_verbs_spec_add(&dev_flow->verbs, &counter, size);
1169 #endif
1170 	return 0;
1171 }
1172 
1173 /**
1174  * Internal validation function. For validating both actions and items.
1175  *
1176  * @param[in] dev
1177  *   Pointer to the Ethernet device structure.
1178  * @param[in] attr
1179  *   Pointer to the flow attributes.
1180  * @param[in] items
1181  *   Pointer to the list of items.
1182  * @param[in] actions
1183  *   Pointer to the list of actions.
1184  * @param[in] external
1185  *   This flow rule is created by request external to PMD.
1186  * @param[in] hairpin
1187  *   Number of hairpin TX actions, 0 means classic flow.
1188  * @param[out] error
1189  *   Pointer to the error structure.
1190  *
1191  * @return
1192  *   0 on success, a negative errno value otherwise and rte_errno is set.
1193  */
1194 static int
1195 flow_verbs_validate(struct rte_eth_dev *dev,
1196 		    const struct rte_flow_attr *attr,
1197 		    const struct rte_flow_item items[],
1198 		    const struct rte_flow_action actions[],
1199 		    bool external __rte_unused,
1200 		    int hairpin __rte_unused,
1201 		    struct rte_flow_error *error)
1202 {
1203 	int ret;
1204 	uint64_t action_flags = 0;
1205 	uint64_t item_flags = 0;
1206 	uint64_t last_item = 0;
1207 	uint8_t next_protocol = 0xff;
1208 	uint16_t ether_type = 0;
1209 	bool is_empty_vlan = false;
1210 	uint16_t udp_dport = 0;
1211 
1212 	if (items == NULL)
1213 		return -1;
1214 	ret = mlx5_flow_validate_attributes(dev, attr, error);
1215 	if (ret < 0)
1216 		return ret;
1217 	for (; items->type != RTE_FLOW_ITEM_TYPE_END; items++) {
1218 		int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
1219 		int ret = 0;
1220 
1221 		switch (items->type) {
1222 		case RTE_FLOW_ITEM_TYPE_VOID:
1223 			break;
1224 		case RTE_FLOW_ITEM_TYPE_ETH:
1225 			ret = mlx5_flow_validate_item_eth(items, item_flags,
1226 							  false, error);
1227 			if (ret < 0)
1228 				return ret;
1229 			last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
1230 					     MLX5_FLOW_LAYER_OUTER_L2;
1231 			if (items->mask != NULL && items->spec != NULL) {
1232 				ether_type =
1233 					((const struct rte_flow_item_eth *)
1234 					 items->spec)->type;
1235 				ether_type &=
1236 					((const struct rte_flow_item_eth *)
1237 					 items->mask)->type;
1238 				if (ether_type == RTE_BE16(RTE_ETHER_TYPE_VLAN))
1239 					is_empty_vlan = true;
1240 				ether_type = rte_be_to_cpu_16(ether_type);
1241 			} else {
1242 				ether_type = 0;
1243 			}
1244 			break;
1245 		case RTE_FLOW_ITEM_TYPE_VLAN:
1246 			ret = mlx5_flow_validate_item_vlan(items, item_flags,
1247 							   dev, error);
1248 			if (ret < 0)
1249 				return ret;
1250 			last_item = tunnel ? (MLX5_FLOW_LAYER_INNER_L2 |
1251 					      MLX5_FLOW_LAYER_INNER_VLAN) :
1252 					     (MLX5_FLOW_LAYER_OUTER_L2 |
1253 					      MLX5_FLOW_LAYER_OUTER_VLAN);
1254 			if (items->mask != NULL && items->spec != NULL) {
1255 				ether_type =
1256 					((const struct rte_flow_item_vlan *)
1257 					 items->spec)->inner_type;
1258 				ether_type &=
1259 					((const struct rte_flow_item_vlan *)
1260 					 items->mask)->inner_type;
1261 				ether_type = rte_be_to_cpu_16(ether_type);
1262 			} else {
1263 				ether_type = 0;
1264 			}
1265 			is_empty_vlan = false;
1266 			break;
1267 		case RTE_FLOW_ITEM_TYPE_IPV4:
1268 			ret = mlx5_flow_validate_item_ipv4
1269 						(items, item_flags,
1270 						 last_item, ether_type, NULL,
1271 						 MLX5_ITEM_RANGE_NOT_ACCEPTED,
1272 						 error);
1273 			if (ret < 0)
1274 				return ret;
1275 			last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV4 :
1276 					     MLX5_FLOW_LAYER_OUTER_L3_IPV4;
1277 			if (items->mask != NULL &&
1278 			    ((const struct rte_flow_item_ipv4 *)
1279 			     items->mask)->hdr.next_proto_id) {
1280 				next_protocol =
1281 					((const struct rte_flow_item_ipv4 *)
1282 					 (items->spec))->hdr.next_proto_id;
1283 				next_protocol &=
1284 					((const struct rte_flow_item_ipv4 *)
1285 					 (items->mask))->hdr.next_proto_id;
1286 			} else {
1287 				/* Reset for inner layer. */
1288 				next_protocol = 0xff;
1289 			}
1290 			break;
1291 		case RTE_FLOW_ITEM_TYPE_IPV6:
1292 			ret = mlx5_flow_validate_item_ipv6(items, item_flags,
1293 							   last_item,
1294 							   ether_type, NULL,
1295 							   error);
1296 			if (ret < 0)
1297 				return ret;
1298 			last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV6 :
1299 					     MLX5_FLOW_LAYER_OUTER_L3_IPV6;
1300 			if (items->mask != NULL &&
1301 			    ((const struct rte_flow_item_ipv6 *)
1302 			     items->mask)->hdr.proto) {
1303 				next_protocol =
1304 					((const struct rte_flow_item_ipv6 *)
1305 					 items->spec)->hdr.proto;
1306 				next_protocol &=
1307 					((const struct rte_flow_item_ipv6 *)
1308 					 items->mask)->hdr.proto;
1309 			} else {
1310 				/* Reset for inner layer. */
1311 				next_protocol = 0xff;
1312 			}
1313 			break;
1314 		case RTE_FLOW_ITEM_TYPE_UDP:
1315 			ret = mlx5_flow_validate_item_udp(items, item_flags,
1316 							  next_protocol,
1317 							  error);
1318 			const struct rte_flow_item_udp *spec = items->spec;
1319 			const struct rte_flow_item_udp *mask = items->mask;
1320 			if (!mask)
1321 				mask = &rte_flow_item_udp_mask;
1322 			if (spec != NULL)
1323 				udp_dport = rte_be_to_cpu_16
1324 						(spec->hdr.dst_port &
1325 						 mask->hdr.dst_port);
1326 
1327 			if (ret < 0)
1328 				return ret;
1329 			last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L4_UDP :
1330 					     MLX5_FLOW_LAYER_OUTER_L4_UDP;
1331 			break;
1332 		case RTE_FLOW_ITEM_TYPE_TCP:
1333 			ret = mlx5_flow_validate_item_tcp
1334 						(items, item_flags,
1335 						 next_protocol,
1336 						 &rte_flow_item_tcp_mask,
1337 						 error);
1338 			if (ret < 0)
1339 				return ret;
1340 			last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L4_TCP :
1341 					     MLX5_FLOW_LAYER_OUTER_L4_TCP;
1342 			break;
1343 		case RTE_FLOW_ITEM_TYPE_VXLAN:
1344 			ret = mlx5_flow_validate_item_vxlan(dev, udp_dport,
1345 							    items, item_flags,
1346 							    attr, error);
1347 			if (ret < 0)
1348 				return ret;
1349 			last_item = MLX5_FLOW_LAYER_VXLAN;
1350 			break;
1351 		case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
1352 			ret = mlx5_flow_validate_item_vxlan_gpe(items,
1353 								item_flags,
1354 								dev, error);
1355 			if (ret < 0)
1356 				return ret;
1357 			last_item = MLX5_FLOW_LAYER_VXLAN_GPE;
1358 			break;
1359 		case RTE_FLOW_ITEM_TYPE_GRE:
1360 			ret = mlx5_flow_validate_item_gre(items, item_flags,
1361 							  next_protocol, error);
1362 			if (ret < 0)
1363 				return ret;
1364 			last_item = MLX5_FLOW_LAYER_GRE;
1365 			break;
1366 		case RTE_FLOW_ITEM_TYPE_MPLS:
1367 			ret = mlx5_flow_validate_item_mpls(dev, items,
1368 							   item_flags,
1369 							   last_item, error);
1370 			if (ret < 0)
1371 				return ret;
1372 			last_item = MLX5_FLOW_LAYER_MPLS;
1373 			break;
1374 		case RTE_FLOW_ITEM_TYPE_ICMP:
1375 		case RTE_FLOW_ITEM_TYPE_ICMP6:
1376 			return rte_flow_error_set(error, ENOTSUP,
1377 						  RTE_FLOW_ERROR_TYPE_ITEM,
1378 						  NULL, "ICMP/ICMP6 "
1379 						  "item not supported");
1380 		default:
1381 			return rte_flow_error_set(error, ENOTSUP,
1382 						  RTE_FLOW_ERROR_TYPE_ITEM,
1383 						  NULL, "item not supported");
1384 		}
1385 		item_flags |= last_item;
1386 	}
1387 	if (is_empty_vlan)
1388 		return rte_flow_error_set(error, ENOTSUP,
1389 						 RTE_FLOW_ERROR_TYPE_ITEM, NULL,
1390 		    "VLAN matching without vid specification is not supported");
1391 	for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
1392 		switch (actions->type) {
1393 		case RTE_FLOW_ACTION_TYPE_VOID:
1394 			break;
1395 		case RTE_FLOW_ACTION_TYPE_FLAG:
1396 			ret = mlx5_flow_validate_action_flag(action_flags,
1397 							     attr,
1398 							     error);
1399 			if (ret < 0)
1400 				return ret;
1401 			action_flags |= MLX5_FLOW_ACTION_FLAG;
1402 			break;
1403 		case RTE_FLOW_ACTION_TYPE_MARK:
1404 			ret = mlx5_flow_validate_action_mark(actions,
1405 							     action_flags,
1406 							     attr,
1407 							     error);
1408 			if (ret < 0)
1409 				return ret;
1410 			action_flags |= MLX5_FLOW_ACTION_MARK;
1411 			break;
1412 		case RTE_FLOW_ACTION_TYPE_DROP:
1413 			ret = mlx5_flow_validate_action_drop(action_flags,
1414 							     attr,
1415 							     error);
1416 			if (ret < 0)
1417 				return ret;
1418 			action_flags |= MLX5_FLOW_ACTION_DROP;
1419 			break;
1420 		case RTE_FLOW_ACTION_TYPE_QUEUE:
1421 			ret = mlx5_flow_validate_action_queue(actions,
1422 							      action_flags, dev,
1423 							      attr,
1424 							      error);
1425 			if (ret < 0)
1426 				return ret;
1427 			action_flags |= MLX5_FLOW_ACTION_QUEUE;
1428 			break;
1429 		case RTE_FLOW_ACTION_TYPE_RSS:
1430 			ret = mlx5_flow_validate_action_rss(actions,
1431 							    action_flags, dev,
1432 							    attr, item_flags,
1433 							    error);
1434 			if (ret < 0)
1435 				return ret;
1436 			action_flags |= MLX5_FLOW_ACTION_RSS;
1437 			break;
1438 		case RTE_FLOW_ACTION_TYPE_COUNT:
1439 			ret = mlx5_flow_validate_action_count(dev, attr, error);
1440 			if (ret < 0)
1441 				return ret;
1442 			action_flags |= MLX5_FLOW_ACTION_COUNT;
1443 			break;
1444 		default:
1445 			return rte_flow_error_set(error, ENOTSUP,
1446 						  RTE_FLOW_ERROR_TYPE_ACTION,
1447 						  actions,
1448 						  "action not supported");
1449 		}
1450 	}
1451 	/*
1452 	 * Validate the drop action mutual exclusion with other actions.
1453 	 * Drop action is mutually-exclusive with any other action, except for
1454 	 * Count action.
1455 	 */
1456 	if ((action_flags & MLX5_FLOW_ACTION_DROP) &&
1457 	    (action_flags & ~(MLX5_FLOW_ACTION_DROP | MLX5_FLOW_ACTION_COUNT)))
1458 		return rte_flow_error_set(error, EINVAL,
1459 					  RTE_FLOW_ERROR_TYPE_ACTION, NULL,
1460 					  "Drop action is mutually-exclusive "
1461 					  "with any other action, except for "
1462 					  "Count action");
1463 	if (!(action_flags & MLX5_FLOW_FATE_ACTIONS))
1464 		return rte_flow_error_set(error, EINVAL,
1465 					  RTE_FLOW_ERROR_TYPE_ACTION, actions,
1466 					  "no fate action is found");
1467 	return 0;
1468 }
1469 
1470 /**
1471  * Calculate the required bytes that are needed for the action part of the verbs
1472  * flow.
1473  *
1474  * @param[in] actions
1475  *   Pointer to the list of actions.
1476  *
1477  * @return
1478  *   The size of the memory needed for all actions.
1479  */
1480 static int
1481 flow_verbs_get_actions_size(const struct rte_flow_action actions[])
1482 {
1483 	int size = 0;
1484 
1485 	for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
1486 		switch (actions->type) {
1487 		case RTE_FLOW_ACTION_TYPE_VOID:
1488 			break;
1489 		case RTE_FLOW_ACTION_TYPE_FLAG:
1490 			size += sizeof(struct ibv_flow_spec_action_tag);
1491 			break;
1492 		case RTE_FLOW_ACTION_TYPE_MARK:
1493 			size += sizeof(struct ibv_flow_spec_action_tag);
1494 			break;
1495 		case RTE_FLOW_ACTION_TYPE_DROP:
1496 			size += sizeof(struct ibv_flow_spec_action_drop);
1497 			break;
1498 		case RTE_FLOW_ACTION_TYPE_QUEUE:
1499 			break;
1500 		case RTE_FLOW_ACTION_TYPE_RSS:
1501 			break;
1502 		case RTE_FLOW_ACTION_TYPE_COUNT:
1503 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) || \
1504 	defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1505 			size += sizeof(struct ibv_flow_spec_counter_action);
1506 #endif
1507 			break;
1508 		default:
1509 			break;
1510 		}
1511 	}
1512 	return size;
1513 }
1514 
1515 /**
1516  * Calculate the required bytes that are needed for the item part of the verbs
1517  * flow.
1518  *
1519  * @param[in] items
1520  *   Pointer to the list of items.
1521  *
1522  * @return
1523  *   The size of the memory needed for all items.
1524  */
1525 static int
1526 flow_verbs_get_items_size(const struct rte_flow_item items[])
1527 {
1528 	int size = 0;
1529 
1530 	for (; items->type != RTE_FLOW_ITEM_TYPE_END; items++) {
1531 		switch (items->type) {
1532 		case RTE_FLOW_ITEM_TYPE_VOID:
1533 			break;
1534 		case RTE_FLOW_ITEM_TYPE_ETH:
1535 			size += sizeof(struct ibv_flow_spec_eth);
1536 			break;
1537 		case RTE_FLOW_ITEM_TYPE_VLAN:
1538 			size += sizeof(struct ibv_flow_spec_eth);
1539 			break;
1540 		case RTE_FLOW_ITEM_TYPE_IPV4:
1541 			size += sizeof(struct ibv_flow_spec_ipv4_ext);
1542 			break;
1543 		case RTE_FLOW_ITEM_TYPE_IPV6:
1544 			size += sizeof(struct ibv_flow_spec_ipv6);
1545 			break;
1546 		case RTE_FLOW_ITEM_TYPE_UDP:
1547 			size += sizeof(struct ibv_flow_spec_tcp_udp);
1548 			break;
1549 		case RTE_FLOW_ITEM_TYPE_TCP:
1550 			size += sizeof(struct ibv_flow_spec_tcp_udp);
1551 			break;
1552 		case RTE_FLOW_ITEM_TYPE_VXLAN:
1553 			size += sizeof(struct ibv_flow_spec_tunnel);
1554 			break;
1555 		case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
1556 			size += sizeof(struct ibv_flow_spec_tunnel);
1557 			break;
1558 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1559 		case RTE_FLOW_ITEM_TYPE_GRE:
1560 			size += sizeof(struct ibv_flow_spec_gre);
1561 			break;
1562 		case RTE_FLOW_ITEM_TYPE_MPLS:
1563 			size += sizeof(struct ibv_flow_spec_mpls);
1564 			break;
1565 #else
1566 		case RTE_FLOW_ITEM_TYPE_GRE:
1567 			size += sizeof(struct ibv_flow_spec_tunnel);
1568 			break;
1569 #endif
1570 		default:
1571 			break;
1572 		}
1573 	}
1574 	return size;
1575 }
1576 
1577 /**
1578  * Internal preparation function. Allocate mlx5_flow with the required size.
1579  * The required size is calculate based on the actions and items. This function
1580  * also returns the detected actions and items for later use.
1581  *
1582  * @param[in] dev
1583  *   Pointer to Ethernet device.
1584  * @param[in] attr
1585  *   Pointer to the flow attributes.
1586  * @param[in] items
1587  *   Pointer to the list of items.
1588  * @param[in] actions
1589  *   Pointer to the list of actions.
1590  * @param[out] error
1591  *   Pointer to the error structure.
1592  *
1593  * @return
1594  *   Pointer to mlx5_flow object on success, otherwise NULL and rte_errno
1595  *   is set.
1596  */
1597 static struct mlx5_flow *
1598 flow_verbs_prepare(struct rte_eth_dev *dev,
1599 		   const struct rte_flow_attr *attr __rte_unused,
1600 		   const struct rte_flow_item items[],
1601 		   const struct rte_flow_action actions[],
1602 		   struct rte_flow_error *error)
1603 {
1604 	size_t size = 0;
1605 	uint32_t handle_idx = 0;
1606 	struct mlx5_flow *dev_flow;
1607 	struct mlx5_flow_handle *dev_handle;
1608 	struct mlx5_priv *priv = dev->data->dev_private;
1609 	struct mlx5_flow_workspace *wks = mlx5_flow_get_thread_workspace();
1610 
1611 	MLX5_ASSERT(wks);
1612 	size += flow_verbs_get_actions_size(actions);
1613 	size += flow_verbs_get_items_size(items);
1614 	if (size > MLX5_VERBS_MAX_SPEC_ACT_SIZE) {
1615 		rte_flow_error_set(error, E2BIG,
1616 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1617 				   "Verbs spec/action size too large");
1618 		return NULL;
1619 	}
1620 	/* In case of corrupting the memory. */
1621 	if (wks->flow_idx >= MLX5_NUM_MAX_DEV_FLOWS) {
1622 		rte_flow_error_set(error, ENOSPC,
1623 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1624 				   "not free temporary device flow");
1625 		return NULL;
1626 	}
1627 	dev_handle = mlx5_ipool_zmalloc(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW],
1628 				   &handle_idx);
1629 	if (!dev_handle) {
1630 		rte_flow_error_set(error, ENOMEM,
1631 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1632 				   "not enough memory to create flow handle");
1633 		return NULL;
1634 	}
1635 	MLX5_ASSERT(wks->flow_idx + 1 < RTE_DIM(wks->flows));
1636 	dev_flow = &wks->flows[wks->flow_idx++];
1637 	dev_flow->handle = dev_handle;
1638 	dev_flow->handle_idx = handle_idx;
1639 	/* Memcpy is used, only size needs to be cleared to 0. */
1640 	dev_flow->verbs.size = 0;
1641 	dev_flow->verbs.attr.num_of_specs = 0;
1642 	dev_flow->ingress = attr->ingress;
1643 	dev_flow->hash_fields = 0;
1644 	/* Need to set transfer attribute: not supported in Verbs mode. */
1645 	return dev_flow;
1646 }
1647 
1648 /**
1649  * Fill the flow with verb spec.
1650  *
1651  * @param[in] dev
1652  *   Pointer to Ethernet device.
1653  * @param[in, out] dev_flow
1654  *   Pointer to the mlx5 flow.
1655  * @param[in] attr
1656  *   Pointer to the flow attributes.
1657  * @param[in] items
1658  *   Pointer to the list of items.
1659  * @param[in] actions
1660  *   Pointer to the list of actions.
1661  * @param[out] error
1662  *   Pointer to the error structure.
1663  *
1664  * @return
1665  *   0 on success, else a negative errno value otherwise and rte_errno is set.
1666  */
1667 static int
1668 flow_verbs_translate(struct rte_eth_dev *dev,
1669 		     struct mlx5_flow *dev_flow,
1670 		     const struct rte_flow_attr *attr,
1671 		     const struct rte_flow_item items[],
1672 		     const struct rte_flow_action actions[],
1673 		     struct rte_flow_error *error)
1674 {
1675 	uint64_t item_flags = 0;
1676 	uint64_t action_flags = 0;
1677 	uint64_t priority = attr->priority;
1678 	uint32_t subpriority = 0;
1679 	struct mlx5_priv *priv = dev->data->dev_private;
1680 	struct mlx5_flow_workspace *wks = mlx5_flow_get_thread_workspace();
1681 	struct mlx5_flow_rss_desc *rss_desc;
1682 
1683 	MLX5_ASSERT(wks);
1684 	rss_desc = &wks->rss_desc;
1685 	if (priority == MLX5_FLOW_LOWEST_PRIO_INDICATOR)
1686 		priority = priv->sh->flow_max_priority - 1;
1687 	for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
1688 		int ret;
1689 
1690 		switch (actions->type) {
1691 		case RTE_FLOW_ACTION_TYPE_VOID:
1692 			break;
1693 		case RTE_FLOW_ACTION_TYPE_FLAG:
1694 			flow_verbs_translate_action_flag(dev_flow, actions);
1695 			action_flags |= MLX5_FLOW_ACTION_FLAG;
1696 			wks->mark = 1;
1697 			break;
1698 		case RTE_FLOW_ACTION_TYPE_MARK:
1699 			flow_verbs_translate_action_mark(dev_flow, actions);
1700 			action_flags |= MLX5_FLOW_ACTION_MARK;
1701 			wks->mark = 1;
1702 			break;
1703 		case RTE_FLOW_ACTION_TYPE_DROP:
1704 			flow_verbs_translate_action_drop(dev_flow, actions);
1705 			action_flags |= MLX5_FLOW_ACTION_DROP;
1706 			dev_flow->handle->fate_action = MLX5_FLOW_FATE_DROP;
1707 			break;
1708 		case RTE_FLOW_ACTION_TYPE_QUEUE:
1709 			flow_verbs_translate_action_queue(rss_desc, actions);
1710 			action_flags |= MLX5_FLOW_ACTION_QUEUE;
1711 			dev_flow->handle->fate_action = MLX5_FLOW_FATE_QUEUE;
1712 			break;
1713 		case RTE_FLOW_ACTION_TYPE_RSS:
1714 			flow_verbs_translate_action_rss(rss_desc, actions);
1715 			action_flags |= MLX5_FLOW_ACTION_RSS;
1716 			dev_flow->handle->fate_action = MLX5_FLOW_FATE_QUEUE;
1717 			break;
1718 		case RTE_FLOW_ACTION_TYPE_COUNT:
1719 			ret = flow_verbs_translate_action_count(dev_flow,
1720 								actions,
1721 								dev, error);
1722 			if (ret < 0)
1723 				return ret;
1724 			action_flags |= MLX5_FLOW_ACTION_COUNT;
1725 			break;
1726 		default:
1727 			return rte_flow_error_set(error, ENOTSUP,
1728 						  RTE_FLOW_ERROR_TYPE_ACTION,
1729 						  actions,
1730 						  "action not supported");
1731 		}
1732 	}
1733 	dev_flow->act_flags = action_flags;
1734 	for (; items->type != RTE_FLOW_ITEM_TYPE_END; items++) {
1735 		int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
1736 
1737 		switch (items->type) {
1738 		case RTE_FLOW_ITEM_TYPE_VOID:
1739 			break;
1740 		case RTE_FLOW_ITEM_TYPE_ETH:
1741 			flow_verbs_translate_item_eth(dev_flow, items,
1742 						      item_flags);
1743 			subpriority = MLX5_PRIORITY_MAP_L2;
1744 			item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
1745 					       MLX5_FLOW_LAYER_OUTER_L2;
1746 			break;
1747 		case RTE_FLOW_ITEM_TYPE_VLAN:
1748 			flow_verbs_translate_item_vlan(dev_flow, items,
1749 						       item_flags);
1750 			subpriority = MLX5_PRIORITY_MAP_L2;
1751 			item_flags |= tunnel ? (MLX5_FLOW_LAYER_INNER_L2 |
1752 						MLX5_FLOW_LAYER_INNER_VLAN) :
1753 					       (MLX5_FLOW_LAYER_OUTER_L2 |
1754 						MLX5_FLOW_LAYER_OUTER_VLAN);
1755 			break;
1756 		case RTE_FLOW_ITEM_TYPE_IPV4:
1757 			flow_verbs_translate_item_ipv4(dev_flow, items,
1758 						       item_flags);
1759 			subpriority = MLX5_PRIORITY_MAP_L3;
1760 			dev_flow->hash_fields |=
1761 				mlx5_flow_hashfields_adjust
1762 					(rss_desc, tunnel,
1763 					 MLX5_IPV4_LAYER_TYPES,
1764 					 MLX5_IPV4_IBV_RX_HASH);
1765 			item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV4 :
1766 					       MLX5_FLOW_LAYER_OUTER_L3_IPV4;
1767 			break;
1768 		case RTE_FLOW_ITEM_TYPE_IPV6:
1769 			flow_verbs_translate_item_ipv6(dev_flow, items,
1770 						       item_flags);
1771 			subpriority = MLX5_PRIORITY_MAP_L3;
1772 			dev_flow->hash_fields |=
1773 				mlx5_flow_hashfields_adjust
1774 					(rss_desc, tunnel,
1775 					 MLX5_IPV6_LAYER_TYPES,
1776 					 MLX5_IPV6_IBV_RX_HASH);
1777 			item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV6 :
1778 					       MLX5_FLOW_LAYER_OUTER_L3_IPV6;
1779 			break;
1780 		case RTE_FLOW_ITEM_TYPE_TCP:
1781 			flow_verbs_translate_item_tcp(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_TCP,
1788 					 (IBV_RX_HASH_SRC_PORT_TCP |
1789 					  IBV_RX_HASH_DST_PORT_TCP));
1790 			item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L4_TCP :
1791 					       MLX5_FLOW_LAYER_OUTER_L4_TCP;
1792 			break;
1793 		case RTE_FLOW_ITEM_TYPE_UDP:
1794 			flow_verbs_translate_item_udp(dev_flow, items,
1795 						      item_flags);
1796 			subpriority = MLX5_PRIORITY_MAP_L4;
1797 			if (dev_flow->hash_fields != 0)
1798 				dev_flow->hash_fields |=
1799 					mlx5_flow_hashfields_adjust
1800 					(rss_desc, tunnel, RTE_ETH_RSS_UDP,
1801 					 (IBV_RX_HASH_SRC_PORT_UDP |
1802 					  IBV_RX_HASH_DST_PORT_UDP));
1803 			item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L4_UDP :
1804 					       MLX5_FLOW_LAYER_OUTER_L4_UDP;
1805 			break;
1806 		case RTE_FLOW_ITEM_TYPE_VXLAN:
1807 			flow_verbs_translate_item_vxlan(dev_flow, items,
1808 							item_flags);
1809 			subpriority = MLX5_TUNNEL_PRIO_GET(rss_desc);
1810 			item_flags |= MLX5_FLOW_LAYER_VXLAN;
1811 			break;
1812 		case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
1813 			flow_verbs_translate_item_vxlan_gpe(dev_flow, items,
1814 							    item_flags);
1815 			subpriority = MLX5_TUNNEL_PRIO_GET(rss_desc);
1816 			item_flags |= MLX5_FLOW_LAYER_VXLAN_GPE;
1817 			break;
1818 		case RTE_FLOW_ITEM_TYPE_GRE:
1819 			subpriority = MLX5_TUNNEL_PRIO_GET(rss_desc);
1820 			item_flags |= MLX5_FLOW_LAYER_GRE;
1821 			break;
1822 		case RTE_FLOW_ITEM_TYPE_MPLS:
1823 			flow_verbs_translate_item_mpls(dev_flow, items,
1824 						       item_flags);
1825 			subpriority = MLX5_TUNNEL_PRIO_GET(rss_desc);
1826 			item_flags |= MLX5_FLOW_LAYER_MPLS;
1827 			break;
1828 		default:
1829 			return rte_flow_error_set(error, ENOTSUP,
1830 						  RTE_FLOW_ERROR_TYPE_ITEM,
1831 						  NULL, "item not supported");
1832 		}
1833 	}
1834 	if (item_flags & MLX5_FLOW_LAYER_GRE)
1835 		flow_verbs_translate_item_gre(dev_flow, items, item_flags);
1836 	dev_flow->handle->layers = item_flags;
1837 	/* Other members of attr will be ignored. */
1838 	dev_flow->verbs.attr.priority =
1839 		mlx5_flow_adjust_priority(dev, priority, subpriority);
1840 	dev_flow->verbs.attr.port = (uint8_t)priv->dev_port;
1841 	return 0;
1842 }
1843 
1844 /**
1845  * Remove the flow from the NIC but keeps it in memory.
1846  *
1847  * @param[in] dev
1848  *   Pointer to the Ethernet device structure.
1849  * @param[in, out] flow
1850  *   Pointer to flow structure.
1851  */
1852 static void
1853 flow_verbs_remove(struct rte_eth_dev *dev, struct rte_flow *flow)
1854 {
1855 	struct mlx5_priv *priv = dev->data->dev_private;
1856 	struct mlx5_flow_handle *handle;
1857 	uint32_t handle_idx;
1858 
1859 	if (!flow)
1860 		return;
1861 	SILIST_FOREACH(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW], flow->dev_handles,
1862 		       handle_idx, handle, next) {
1863 		if (handle->drv_flow) {
1864 			claim_zero(mlx5_glue->destroy_flow(handle->drv_flow));
1865 			handle->drv_flow = NULL;
1866 		}
1867 		/* hrxq is union, don't touch it only the flag is set. */
1868 		if (handle->rix_hrxq &&
1869 		    handle->fate_action == MLX5_FLOW_FATE_QUEUE) {
1870 			mlx5_hrxq_release(dev, handle->rix_hrxq);
1871 			handle->rix_hrxq = 0;
1872 		}
1873 		if (handle->vf_vlan.tag && handle->vf_vlan.created)
1874 			mlx5_vlan_vmwa_release(dev, &handle->vf_vlan);
1875 	}
1876 }
1877 
1878 /**
1879  * Remove the flow from the NIC and the memory.
1880  *
1881  * @param[in] dev
1882  *   Pointer to the Ethernet device structure.
1883  * @param[in, out] flow
1884  *   Pointer to flow structure.
1885  */
1886 static void
1887 flow_verbs_destroy(struct rte_eth_dev *dev, struct rte_flow *flow)
1888 {
1889 	struct mlx5_priv *priv = dev->data->dev_private;
1890 	struct mlx5_flow_handle *handle;
1891 
1892 	if (!flow)
1893 		return;
1894 	flow_verbs_remove(dev, flow);
1895 	while (flow->dev_handles) {
1896 		uint32_t tmp_idx = flow->dev_handles;
1897 
1898 		handle = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW],
1899 				   tmp_idx);
1900 		if (!handle)
1901 			return;
1902 		flow->dev_handles = handle->next.next;
1903 		mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW],
1904 			   tmp_idx);
1905 	}
1906 	if (flow->counter) {
1907 		flow_verbs_counter_release(dev, flow->counter);
1908 		flow->counter = 0;
1909 	}
1910 }
1911 
1912 /**
1913  * Apply the flow to the NIC.
1914  *
1915  * @param[in] dev
1916  *   Pointer to the Ethernet device structure.
1917  * @param[in, out] flow
1918  *   Pointer to flow structure.
1919  * @param[out] error
1920  *   Pointer to error structure.
1921  *
1922  * @return
1923  *   0 on success, a negative errno value otherwise and rte_errno is set.
1924  */
1925 static int
1926 flow_verbs_apply(struct rte_eth_dev *dev, struct rte_flow *flow,
1927 		 struct rte_flow_error *error)
1928 {
1929 	struct mlx5_priv *priv = dev->data->dev_private;
1930 	struct mlx5_flow_handle *handle;
1931 	struct mlx5_flow *dev_flow;
1932 	struct mlx5_hrxq *hrxq;
1933 	uint32_t dev_handles;
1934 	int err;
1935 	int idx;
1936 	struct mlx5_flow_workspace *wks = mlx5_flow_get_thread_workspace();
1937 
1938 	MLX5_ASSERT(wks);
1939 	for (idx = wks->flow_idx - 1; idx >= 0; idx--) {
1940 		dev_flow = &wks->flows[idx];
1941 		handle = dev_flow->handle;
1942 		if (handle->fate_action == MLX5_FLOW_FATE_DROP) {
1943 			MLX5_ASSERT(priv->drop_queue.hrxq);
1944 			hrxq = priv->drop_queue.hrxq;
1945 		} else {
1946 			uint32_t hrxq_idx;
1947 			struct mlx5_flow_rss_desc *rss_desc = &wks->rss_desc;
1948 
1949 			MLX5_ASSERT(rss_desc->queue_num);
1950 			rss_desc->key_len = MLX5_RSS_HASH_KEY_LEN;
1951 			rss_desc->hash_fields = dev_flow->hash_fields;
1952 			rss_desc->tunnel = !!(handle->layers &
1953 					      MLX5_FLOW_LAYER_TUNNEL);
1954 			rss_desc->shared_rss = 0;
1955 			hrxq_idx = mlx5_hrxq_get(dev, rss_desc);
1956 			hrxq = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_HRXQ],
1957 					      hrxq_idx);
1958 			if (!hrxq) {
1959 				rte_flow_error_set
1960 					(error, rte_errno,
1961 					 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1962 					 "cannot get hash queue");
1963 				goto error;
1964 			}
1965 			handle->rix_hrxq = hrxq_idx;
1966 		}
1967 		MLX5_ASSERT(hrxq);
1968 		handle->drv_flow = mlx5_glue->create_flow
1969 					(hrxq->qp, &dev_flow->verbs.attr);
1970 		if (!handle->drv_flow) {
1971 			rte_flow_error_set(error, errno,
1972 					   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1973 					   NULL,
1974 					   "hardware refuses to create flow");
1975 			goto error;
1976 		}
1977 		if (priv->vmwa_context &&
1978 		    handle->vf_vlan.tag && !handle->vf_vlan.created) {
1979 			/*
1980 			 * The rule contains the VLAN pattern.
1981 			 * For VF we are going to create VLAN
1982 			 * interface to make hypervisor set correct
1983 			 * e-Switch vport context.
1984 			 */
1985 			mlx5_vlan_vmwa_acquire(dev, &handle->vf_vlan);
1986 		}
1987 	}
1988 	return 0;
1989 error:
1990 	err = rte_errno; /* Save rte_errno before cleanup. */
1991 	SILIST_FOREACH(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW], flow->dev_handles,
1992 		       dev_handles, handle, next) {
1993 		/* hrxq is union, don't touch it only the flag is set. */
1994 		if (handle->rix_hrxq &&
1995 		    handle->fate_action == MLX5_FLOW_FATE_QUEUE) {
1996 			mlx5_hrxq_release(dev, handle->rix_hrxq);
1997 			handle->rix_hrxq = 0;
1998 		}
1999 		if (handle->vf_vlan.tag && handle->vf_vlan.created)
2000 			mlx5_vlan_vmwa_release(dev, &handle->vf_vlan);
2001 	}
2002 	rte_errno = err; /* Restore rte_errno. */
2003 	return -rte_errno;
2004 }
2005 
2006 /**
2007  * Query a flow.
2008  *
2009  * @see rte_flow_query()
2010  * @see rte_flow_ops
2011  */
2012 static int
2013 flow_verbs_query(struct rte_eth_dev *dev,
2014 		 struct rte_flow *flow,
2015 		 const struct rte_flow_action *actions,
2016 		 void *data,
2017 		 struct rte_flow_error *error)
2018 {
2019 	int ret = -EINVAL;
2020 
2021 	for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
2022 		switch (actions->type) {
2023 		case RTE_FLOW_ACTION_TYPE_VOID:
2024 			break;
2025 		case RTE_FLOW_ACTION_TYPE_COUNT:
2026 			ret = flow_verbs_counter_query(dev, flow, data, error);
2027 			break;
2028 		default:
2029 			return rte_flow_error_set(error, ENOTSUP,
2030 						  RTE_FLOW_ERROR_TYPE_ACTION,
2031 						  actions,
2032 						  "action not supported");
2033 		}
2034 	}
2035 	return ret;
2036 }
2037 
2038 static int
2039 flow_verbs_sync_domain(struct rte_eth_dev *dev, uint32_t domains,
2040 		       uint32_t flags)
2041 {
2042 	RTE_SET_USED(dev);
2043 	RTE_SET_USED(domains);
2044 	RTE_SET_USED(flags);
2045 
2046 	return 0;
2047 }
2048 
2049 const struct mlx5_flow_driver_ops mlx5_flow_verbs_drv_ops = {
2050 	.validate = flow_verbs_validate,
2051 	.prepare = flow_verbs_prepare,
2052 	.translate = flow_verbs_translate,
2053 	.apply = flow_verbs_apply,
2054 	.remove = flow_verbs_remove,
2055 	.destroy = flow_verbs_destroy,
2056 	.query = flow_verbs_query,
2057 	.sync_domain = flow_verbs_sync_domain,
2058 	.discover_priorities = flow_verbs_discover_priorities,
2059 };
2060