xref: /dpdk/drivers/net/mlx5/mlx5_flow.c (revision a9263e3c5ad0b71adb95d686c1c3153865b0e1d2)
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright 2016 6WIND S.A.
5  *   Copyright 2016 Mellanox.
6  *
7  *   Redistribution and use in source and binary forms, with or without
8  *   modification, are permitted provided that the following conditions
9  *   are met:
10  *
11  *     * Redistributions of source code must retain the above copyright
12  *       notice, this list of conditions and the following disclaimer.
13  *     * Redistributions in binary form must reproduce the above copyright
14  *       notice, this list of conditions and the following disclaimer in
15  *       the documentation and/or other materials provided with the
16  *       distribution.
17  *     * Neither the name of 6WIND S.A. nor the names of its
18  *       contributors may be used to endorse or promote products derived
19  *       from this software without specific prior written permission.
20  *
21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32  */
33 
34 #include <sys/queue.h>
35 #include <string.h>
36 
37 /* Verbs header. */
38 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
39 #ifdef PEDANTIC
40 #pragma GCC diagnostic ignored "-Wpedantic"
41 #endif
42 #include <infiniband/verbs.h>
43 #ifdef PEDANTIC
44 #pragma GCC diagnostic error "-Wpedantic"
45 #endif
46 
47 #include <rte_ethdev.h>
48 #include <rte_flow.h>
49 #include <rte_flow_driver.h>
50 #include <rte_malloc.h>
51 
52 #include "mlx5.h"
53 #include "mlx5_prm.h"
54 
55 /* Define minimal priority for control plane flows. */
56 #define MLX5_CTRL_FLOW_PRIORITY 4
57 
58 /* Internet Protocol versions. */
59 #define MLX5_IPV4 4
60 #define MLX5_IPV6 6
61 
62 #ifndef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
63 struct ibv_counter_set_init_attr {
64 	int dummy;
65 };
66 struct ibv_flow_spec_counter_action {
67 	int dummy;
68 };
69 struct ibv_counter_set {
70 	int dummy;
71 };
72 
73 static inline int
74 ibv_destroy_counter_set(struct ibv_counter_set *cs)
75 {
76 	(void)cs;
77 	return -ENOTSUP;
78 }
79 #endif
80 
81 /* Dev ops structure defined in mlx5.c */
82 extern const struct eth_dev_ops mlx5_dev_ops;
83 extern const struct eth_dev_ops mlx5_dev_ops_isolate;
84 
85 static int
86 mlx5_flow_create_eth(const struct rte_flow_item *item,
87 		     const void *default_mask,
88 		     void *data);
89 
90 static int
91 mlx5_flow_create_vlan(const struct rte_flow_item *item,
92 		      const void *default_mask,
93 		      void *data);
94 
95 static int
96 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
97 		      const void *default_mask,
98 		      void *data);
99 
100 static int
101 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
102 		      const void *default_mask,
103 		      void *data);
104 
105 static int
106 mlx5_flow_create_udp(const struct rte_flow_item *item,
107 		     const void *default_mask,
108 		     void *data);
109 
110 static int
111 mlx5_flow_create_tcp(const struct rte_flow_item *item,
112 		     const void *default_mask,
113 		     void *data);
114 
115 static int
116 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
117 		       const void *default_mask,
118 		       void *data);
119 
120 struct mlx5_flow_parse;
121 
122 static void
123 mlx5_flow_create_copy(struct mlx5_flow_parse *parser, void *src,
124 		      unsigned int size);
125 
126 static int
127 mlx5_flow_create_flag_mark(struct mlx5_flow_parse *parser, uint32_t mark_id);
128 
129 static int
130 mlx5_flow_create_count(struct priv *priv, struct mlx5_flow_parse *parser);
131 
132 /* Hash RX queue types. */
133 enum hash_rxq_type {
134 	HASH_RXQ_TCPV4,
135 	HASH_RXQ_UDPV4,
136 	HASH_RXQ_IPV4,
137 	HASH_RXQ_TCPV6,
138 	HASH_RXQ_UDPV6,
139 	HASH_RXQ_IPV6,
140 	HASH_RXQ_ETH,
141 };
142 
143 /* Initialization data for hash RX queue. */
144 struct hash_rxq_init {
145 	uint64_t hash_fields; /* Fields that participate in the hash. */
146 	uint64_t dpdk_rss_hf; /* Matching DPDK RSS hash fields. */
147 	unsigned int flow_priority; /* Flow priority to use. */
148 	unsigned int ip_version; /* Internet protocol. */
149 };
150 
151 /* Initialization data for hash RX queues. */
152 const struct hash_rxq_init hash_rxq_init[] = {
153 	[HASH_RXQ_TCPV4] = {
154 		.hash_fields = (IBV_RX_HASH_SRC_IPV4 |
155 				IBV_RX_HASH_DST_IPV4 |
156 				IBV_RX_HASH_SRC_PORT_TCP |
157 				IBV_RX_HASH_DST_PORT_TCP),
158 		.dpdk_rss_hf = ETH_RSS_NONFRAG_IPV4_TCP,
159 		.flow_priority = 0,
160 		.ip_version = MLX5_IPV4,
161 	},
162 	[HASH_RXQ_UDPV4] = {
163 		.hash_fields = (IBV_RX_HASH_SRC_IPV4 |
164 				IBV_RX_HASH_DST_IPV4 |
165 				IBV_RX_HASH_SRC_PORT_UDP |
166 				IBV_RX_HASH_DST_PORT_UDP),
167 		.dpdk_rss_hf = ETH_RSS_NONFRAG_IPV4_UDP,
168 		.flow_priority = 0,
169 		.ip_version = MLX5_IPV4,
170 	},
171 	[HASH_RXQ_IPV4] = {
172 		.hash_fields = (IBV_RX_HASH_SRC_IPV4 |
173 				IBV_RX_HASH_DST_IPV4),
174 		.dpdk_rss_hf = (ETH_RSS_IPV4 |
175 				ETH_RSS_FRAG_IPV4),
176 		.flow_priority = 1,
177 		.ip_version = MLX5_IPV4,
178 	},
179 	[HASH_RXQ_TCPV6] = {
180 		.hash_fields = (IBV_RX_HASH_SRC_IPV6 |
181 				IBV_RX_HASH_DST_IPV6 |
182 				IBV_RX_HASH_SRC_PORT_TCP |
183 				IBV_RX_HASH_DST_PORT_TCP),
184 		.dpdk_rss_hf = ETH_RSS_NONFRAG_IPV6_TCP,
185 		.flow_priority = 0,
186 		.ip_version = MLX5_IPV6,
187 	},
188 	[HASH_RXQ_UDPV6] = {
189 		.hash_fields = (IBV_RX_HASH_SRC_IPV6 |
190 				IBV_RX_HASH_DST_IPV6 |
191 				IBV_RX_HASH_SRC_PORT_UDP |
192 				IBV_RX_HASH_DST_PORT_UDP),
193 		.dpdk_rss_hf = ETH_RSS_NONFRAG_IPV6_UDP,
194 		.flow_priority = 0,
195 		.ip_version = MLX5_IPV6,
196 	},
197 	[HASH_RXQ_IPV6] = {
198 		.hash_fields = (IBV_RX_HASH_SRC_IPV6 |
199 				IBV_RX_HASH_DST_IPV6),
200 		.dpdk_rss_hf = (ETH_RSS_IPV6 |
201 				ETH_RSS_FRAG_IPV6),
202 		.flow_priority = 1,
203 		.ip_version = MLX5_IPV6,
204 	},
205 	[HASH_RXQ_ETH] = {
206 		.hash_fields = 0,
207 		.dpdk_rss_hf = 0,
208 		.flow_priority = 2,
209 	},
210 };
211 
212 /* Number of entries in hash_rxq_init[]. */
213 const unsigned int hash_rxq_init_n = RTE_DIM(hash_rxq_init);
214 
215 /** Structure for holding counter stats. */
216 struct mlx5_flow_counter_stats {
217 	uint64_t hits; /**< Number of packets matched by the rule. */
218 	uint64_t bytes; /**< Number of bytes matched by the rule. */
219 };
220 
221 /** Structure for Drop queue. */
222 struct mlx5_hrxq_drop {
223 	struct ibv_rwq_ind_table *ind_table; /**< Indirection table. */
224 	struct ibv_qp *qp; /**< Verbs queue pair. */
225 	struct ibv_wq *wq; /**< Verbs work queue. */
226 	struct ibv_cq *cq; /**< Verbs completion queue. */
227 };
228 
229 /* Flows structures. */
230 struct mlx5_flow {
231 	uint64_t hash_fields; /**< Fields that participate in the hash. */
232 	struct ibv_flow_attr *ibv_attr; /**< Pointer to Verbs attributes. */
233 	struct ibv_flow *ibv_flow; /**< Verbs flow. */
234 	struct mlx5_hrxq *hrxq; /**< Hash Rx queues. */
235 };
236 
237 /* Drop flows structures. */
238 struct mlx5_flow_drop {
239 	struct ibv_flow_attr *ibv_attr; /**< Pointer to Verbs attributes. */
240 	struct ibv_flow *ibv_flow; /**< Verbs flow. */
241 };
242 
243 struct rte_flow {
244 	TAILQ_ENTRY(rte_flow) next; /**< Pointer to the next flow structure. */
245 	uint32_t mark:1; /**< Set if the flow is marked. */
246 	uint32_t drop:1; /**< Drop queue. */
247 	uint16_t queues_n; /**< Number of entries in queue[]. */
248 	uint16_t (*queues)[]; /**< Queues indexes to use. */
249 	struct rte_eth_rss_conf rss_conf; /**< RSS configuration */
250 	uint8_t rss_key[40]; /**< copy of the RSS key. */
251 	struct ibv_counter_set *cs; /**< Holds the counters for the rule. */
252 	struct mlx5_flow_counter_stats counter_stats;/**<The counter stats. */
253 	union {
254 		struct mlx5_flow frxq[RTE_DIM(hash_rxq_init)];
255 		/**< Flow with Rx queue. */
256 		struct mlx5_flow_drop drxq; /**< Flow with drop Rx queue. */
257 	};
258 };
259 
260 /** Static initializer for items. */
261 #define ITEMS(...) \
262 	(const enum rte_flow_item_type []){ \
263 		__VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \
264 	}
265 
266 /** Structure to generate a simple graph of layers supported by the NIC. */
267 struct mlx5_flow_items {
268 	/** List of possible actions for these items. */
269 	const enum rte_flow_action_type *const actions;
270 	/** Bit-masks corresponding to the possibilities for the item. */
271 	const void *mask;
272 	/**
273 	 * Default bit-masks to use when item->mask is not provided. When
274 	 * \default_mask is also NULL, the full supported bit-mask (\mask) is
275 	 * used instead.
276 	 */
277 	const void *default_mask;
278 	/** Bit-masks size in bytes. */
279 	const unsigned int mask_sz;
280 	/**
281 	 * Conversion function from rte_flow to NIC specific flow.
282 	 *
283 	 * @param item
284 	 *   rte_flow item to convert.
285 	 * @param default_mask
286 	 *   Default bit-masks to use when item->mask is not provided.
287 	 * @param data
288 	 *   Internal structure to store the conversion.
289 	 *
290 	 * @return
291 	 *   0 on success, negative value otherwise.
292 	 */
293 	int (*convert)(const struct rte_flow_item *item,
294 		       const void *default_mask,
295 		       void *data);
296 	/** Size in bytes of the destination structure. */
297 	const unsigned int dst_sz;
298 	/** List of possible following items.  */
299 	const enum rte_flow_item_type *const items;
300 };
301 
302 /** Valid action for this PMD. */
303 static const enum rte_flow_action_type valid_actions[] = {
304 	RTE_FLOW_ACTION_TYPE_DROP,
305 	RTE_FLOW_ACTION_TYPE_QUEUE,
306 	RTE_FLOW_ACTION_TYPE_MARK,
307 	RTE_FLOW_ACTION_TYPE_FLAG,
308 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
309 	RTE_FLOW_ACTION_TYPE_COUNT,
310 #endif
311 	RTE_FLOW_ACTION_TYPE_END,
312 };
313 
314 /** Graph of supported items and associated actions. */
315 static const struct mlx5_flow_items mlx5_flow_items[] = {
316 	[RTE_FLOW_ITEM_TYPE_END] = {
317 		.items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH,
318 			       RTE_FLOW_ITEM_TYPE_VXLAN),
319 	},
320 	[RTE_FLOW_ITEM_TYPE_ETH] = {
321 		.items = ITEMS(RTE_FLOW_ITEM_TYPE_VLAN,
322 			       RTE_FLOW_ITEM_TYPE_IPV4,
323 			       RTE_FLOW_ITEM_TYPE_IPV6),
324 		.actions = valid_actions,
325 		.mask = &(const struct rte_flow_item_eth){
326 			.dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
327 			.src.addr_bytes = "\xff\xff\xff\xff\xff\xff",
328 			.type = -1,
329 		},
330 		.default_mask = &rte_flow_item_eth_mask,
331 		.mask_sz = sizeof(struct rte_flow_item_eth),
332 		.convert = mlx5_flow_create_eth,
333 		.dst_sz = sizeof(struct ibv_flow_spec_eth),
334 	},
335 	[RTE_FLOW_ITEM_TYPE_VLAN] = {
336 		.items = ITEMS(RTE_FLOW_ITEM_TYPE_IPV4,
337 			       RTE_FLOW_ITEM_TYPE_IPV6),
338 		.actions = valid_actions,
339 		.mask = &(const struct rte_flow_item_vlan){
340 			.tci = -1,
341 		},
342 		.default_mask = &rte_flow_item_vlan_mask,
343 		.mask_sz = sizeof(struct rte_flow_item_vlan),
344 		.convert = mlx5_flow_create_vlan,
345 		.dst_sz = 0,
346 	},
347 	[RTE_FLOW_ITEM_TYPE_IPV4] = {
348 		.items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
349 			       RTE_FLOW_ITEM_TYPE_TCP),
350 		.actions = valid_actions,
351 		.mask = &(const struct rte_flow_item_ipv4){
352 			.hdr = {
353 				.src_addr = -1,
354 				.dst_addr = -1,
355 				.type_of_service = -1,
356 				.next_proto_id = -1,
357 			},
358 		},
359 		.default_mask = &rte_flow_item_ipv4_mask,
360 		.mask_sz = sizeof(struct rte_flow_item_ipv4),
361 		.convert = mlx5_flow_create_ipv4,
362 		.dst_sz = sizeof(struct ibv_flow_spec_ipv4_ext),
363 	},
364 	[RTE_FLOW_ITEM_TYPE_IPV6] = {
365 		.items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
366 			       RTE_FLOW_ITEM_TYPE_TCP),
367 		.actions = valid_actions,
368 		.mask = &(const struct rte_flow_item_ipv6){
369 			.hdr = {
370 				.src_addr = {
371 					0xff, 0xff, 0xff, 0xff,
372 					0xff, 0xff, 0xff, 0xff,
373 					0xff, 0xff, 0xff, 0xff,
374 					0xff, 0xff, 0xff, 0xff,
375 				},
376 				.dst_addr = {
377 					0xff, 0xff, 0xff, 0xff,
378 					0xff, 0xff, 0xff, 0xff,
379 					0xff, 0xff, 0xff, 0xff,
380 					0xff, 0xff, 0xff, 0xff,
381 				},
382 				.vtc_flow = -1,
383 				.proto = -1,
384 				.hop_limits = -1,
385 			},
386 		},
387 		.default_mask = &rte_flow_item_ipv6_mask,
388 		.mask_sz = sizeof(struct rte_flow_item_ipv6),
389 		.convert = mlx5_flow_create_ipv6,
390 		.dst_sz = sizeof(struct ibv_flow_spec_ipv6),
391 	},
392 	[RTE_FLOW_ITEM_TYPE_UDP] = {
393 		.items = ITEMS(RTE_FLOW_ITEM_TYPE_VXLAN),
394 		.actions = valid_actions,
395 		.mask = &(const struct rte_flow_item_udp){
396 			.hdr = {
397 				.src_port = -1,
398 				.dst_port = -1,
399 			},
400 		},
401 		.default_mask = &rte_flow_item_udp_mask,
402 		.mask_sz = sizeof(struct rte_flow_item_udp),
403 		.convert = mlx5_flow_create_udp,
404 		.dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
405 	},
406 	[RTE_FLOW_ITEM_TYPE_TCP] = {
407 		.actions = valid_actions,
408 		.mask = &(const struct rte_flow_item_tcp){
409 			.hdr = {
410 				.src_port = -1,
411 				.dst_port = -1,
412 			},
413 		},
414 		.default_mask = &rte_flow_item_tcp_mask,
415 		.mask_sz = sizeof(struct rte_flow_item_tcp),
416 		.convert = mlx5_flow_create_tcp,
417 		.dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
418 	},
419 	[RTE_FLOW_ITEM_TYPE_VXLAN] = {
420 		.items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH),
421 		.actions = valid_actions,
422 		.mask = &(const struct rte_flow_item_vxlan){
423 			.vni = "\xff\xff\xff",
424 		},
425 		.default_mask = &rte_flow_item_vxlan_mask,
426 		.mask_sz = sizeof(struct rte_flow_item_vxlan),
427 		.convert = mlx5_flow_create_vxlan,
428 		.dst_sz = sizeof(struct ibv_flow_spec_tunnel),
429 	},
430 };
431 
432 /** Structure to pass to the conversion function. */
433 struct mlx5_flow_parse {
434 	uint32_t inner; /**< Set once VXLAN is encountered. */
435 	uint32_t create:1;
436 	/**< Whether resources should remain after a validate. */
437 	uint32_t drop:1; /**< Target is a drop queue. */
438 	uint32_t mark:1; /**< Mark is present in the flow. */
439 	uint32_t count:1; /**< Count is present in the flow. */
440 	uint32_t mark_id; /**< Mark identifier. */
441 	uint16_t queues[RTE_MAX_QUEUES_PER_PORT]; /**< Queues indexes to use. */
442 	uint16_t queues_n; /**< Number of entries in queue[]. */
443 	struct rte_eth_rss_conf rss_conf; /**< RSS configuration */
444 	uint8_t rss_key[40]; /**< copy of the RSS key. */
445 	enum hash_rxq_type layer; /**< Last pattern layer detected. */
446 	struct ibv_counter_set *cs; /**< Holds the counter set for the rule */
447 	union {
448 		struct {
449 			struct ibv_flow_attr *ibv_attr;
450 			/**< Pointer to Verbs attributes. */
451 			unsigned int offset;
452 			/**< Current position or total size of the attribute. */
453 		} queue[RTE_DIM(hash_rxq_init)];
454 		struct {
455 			struct ibv_flow_attr *ibv_attr;
456 			/**< Pointer to Verbs attributes. */
457 			unsigned int offset;
458 			/**< Current position or total size of the attribute. */
459 		} drop_q;
460 	};
461 };
462 
463 static const struct rte_flow_ops mlx5_flow_ops = {
464 	.validate = mlx5_flow_validate,
465 	.create = mlx5_flow_create,
466 	.destroy = mlx5_flow_destroy,
467 	.flush = mlx5_flow_flush,
468 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
469 	.query = mlx5_flow_query,
470 #else
471 	.query = NULL,
472 #endif
473 	.isolate = mlx5_flow_isolate,
474 };
475 
476 /* Convert FDIR request to Generic flow. */
477 struct mlx5_fdir {
478 	struct rte_flow_attr attr;
479 	struct rte_flow_action actions[2];
480 	struct rte_flow_item items[4];
481 	struct rte_flow_item_eth l2;
482 	struct rte_flow_item_eth l2_mask;
483 	union {
484 		struct rte_flow_item_ipv4 ipv4;
485 		struct rte_flow_item_ipv6 ipv6;
486 	} l3;
487 	union {
488 		struct rte_flow_item_udp udp;
489 		struct rte_flow_item_tcp tcp;
490 	} l4;
491 	struct rte_flow_action_queue queue;
492 };
493 
494 /* Verbs specification header. */
495 struct ibv_spec_header {
496 	enum ibv_flow_spec_type type;
497 	uint16_t size;
498 };
499 
500 /**
501  * Check support for a given item.
502  *
503  * @param item[in]
504  *   Item specification.
505  * @param mask[in]
506  *   Bit-masks covering supported fields to compare with spec, last and mask in
507  *   \item.
508  * @param size
509  *   Bit-Mask size in bytes.
510  *
511  * @return
512  *   0 on success.
513  */
514 static int
515 mlx5_flow_item_validate(const struct rte_flow_item *item,
516 			const uint8_t *mask, unsigned int size)
517 {
518 	int ret = 0;
519 
520 	if (!item->spec && (item->mask || item->last))
521 		return -1;
522 	if (item->spec && !item->mask) {
523 		unsigned int i;
524 		const uint8_t *spec = item->spec;
525 
526 		for (i = 0; i < size; ++i)
527 			if ((spec[i] | mask[i]) != mask[i])
528 				return -1;
529 	}
530 	if (item->last && !item->mask) {
531 		unsigned int i;
532 		const uint8_t *spec = item->last;
533 
534 		for (i = 0; i < size; ++i)
535 			if ((spec[i] | mask[i]) != mask[i])
536 				return -1;
537 	}
538 	if (item->mask) {
539 		unsigned int i;
540 		const uint8_t *spec = item->mask;
541 
542 		for (i = 0; i < size; ++i)
543 			if ((spec[i] | mask[i]) != mask[i])
544 				return -1;
545 	}
546 	if (item->spec && item->last) {
547 		uint8_t spec[size];
548 		uint8_t last[size];
549 		const uint8_t *apply = mask;
550 		unsigned int i;
551 
552 		if (item->mask)
553 			apply = item->mask;
554 		for (i = 0; i < size; ++i) {
555 			spec[i] = ((const uint8_t *)item->spec)[i] & apply[i];
556 			last[i] = ((const uint8_t *)item->last)[i] & apply[i];
557 		}
558 		ret = memcmp(spec, last, size);
559 	}
560 	return ret;
561 }
562 
563 /**
564  * Copy the RSS configuration from the user ones.
565  *
566  * @param priv
567  *   Pointer to private structure.
568  * @param parser
569  *   Internal parser structure.
570  * @param rss_conf
571  *   User RSS configuration to save.
572  *
573  * @return
574  *   0 on success, errno value on failure.
575  */
576 static int
577 priv_flow_convert_rss_conf(struct priv *priv,
578 			   struct mlx5_flow_parse *parser,
579 			   const struct rte_eth_rss_conf *rss_conf)
580 {
581 	const struct rte_eth_rss_conf *rss =
582 		rss_conf ? rss_conf : &priv->rss_conf;
583 
584 	if (rss->rss_key_len > 40)
585 		return EINVAL;
586 	parser->rss_conf.rss_key_len = rss->rss_key_len;
587 	parser->rss_conf.rss_hf = rss->rss_hf;
588 	memcpy(parser->rss_key, rss->rss_key, rss->rss_key_len);
589 	parser->rss_conf.rss_key = parser->rss_key;
590 	return 0;
591 }
592 
593 /**
594  * Extract attribute to the parser.
595  *
596  * @param priv
597  *   Pointer to private structure.
598  * @param[in] attr
599  *   Flow rule attributes.
600  * @param[out] error
601  *   Perform verbose error reporting if not NULL.
602  * @param[in, out] parser
603  *   Internal parser structure.
604  *
605  * @return
606  *   0 on success, a negative errno value otherwise and rte_errno is set.
607  */
608 static int
609 priv_flow_convert_attributes(struct priv *priv,
610 			     const struct rte_flow_attr *attr,
611 			     struct rte_flow_error *error,
612 			     struct mlx5_flow_parse *parser)
613 {
614 	(void)priv;
615 	(void)parser;
616 	if (attr->group) {
617 		rte_flow_error_set(error, ENOTSUP,
618 				   RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
619 				   NULL,
620 				   "groups are not supported");
621 		return -rte_errno;
622 	}
623 	if (attr->priority && attr->priority != MLX5_CTRL_FLOW_PRIORITY) {
624 		rte_flow_error_set(error, ENOTSUP,
625 				   RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
626 				   NULL,
627 				   "priorities are not supported");
628 		return -rte_errno;
629 	}
630 	if (attr->egress) {
631 		rte_flow_error_set(error, ENOTSUP,
632 				   RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
633 				   NULL,
634 				   "egress is not supported");
635 		return -rte_errno;
636 	}
637 	if (!attr->ingress) {
638 		rte_flow_error_set(error, ENOTSUP,
639 				   RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
640 				   NULL,
641 				   "only ingress is supported");
642 		return -rte_errno;
643 	}
644 	return 0;
645 }
646 
647 /**
648  * Extract actions request to the parser.
649  *
650  * @param priv
651  *   Pointer to private structure.
652  * @param[in] actions
653  *   Associated actions (list terminated by the END action).
654  * @param[out] error
655  *   Perform verbose error reporting if not NULL.
656  * @param[in, out] parser
657  *   Internal parser structure.
658  *
659  * @return
660  *   0 on success, a negative errno value otherwise and rte_errno is set.
661  */
662 static int
663 priv_flow_convert_actions(struct priv *priv,
664 			  const struct rte_flow_action actions[],
665 			  struct rte_flow_error *error,
666 			  struct mlx5_flow_parse *parser)
667 {
668 	/*
669 	 * Add default RSS configuration necessary for Verbs to create QP even
670 	 * if no RSS is necessary.
671 	 */
672 	priv_flow_convert_rss_conf(priv, parser,
673 				   (const struct rte_eth_rss_conf *)
674 				   &priv->rss_conf);
675 	for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
676 		if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
677 			continue;
678 		} else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
679 			parser->drop = 1;
680 		} else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
681 			const struct rte_flow_action_queue *queue =
682 				(const struct rte_flow_action_queue *)
683 				actions->conf;
684 			uint16_t n;
685 			uint16_t found = 0;
686 
687 			if (!queue || (queue->index > (priv->rxqs_n - 1)))
688 				goto exit_action_not_supported;
689 			for (n = 0; n < parser->queues_n; ++n) {
690 				if (parser->queues[n] == queue->index) {
691 					found = 1;
692 					break;
693 				}
694 			}
695 			if (parser->queues_n > 1 && !found) {
696 				rte_flow_error_set(error, ENOTSUP,
697 					   RTE_FLOW_ERROR_TYPE_ACTION,
698 					   actions,
699 					   "queue action not in RSS queues");
700 				return -rte_errno;
701 			}
702 			if (!found) {
703 				parser->queues_n = 1;
704 				parser->queues[0] = queue->index;
705 			}
706 		} else if (actions->type == RTE_FLOW_ACTION_TYPE_RSS) {
707 			const struct rte_flow_action_rss *rss =
708 				(const struct rte_flow_action_rss *)
709 				actions->conf;
710 			uint16_t n;
711 
712 			if (!rss || !rss->num) {
713 				rte_flow_error_set(error, EINVAL,
714 						   RTE_FLOW_ERROR_TYPE_ACTION,
715 						   actions,
716 						   "no valid queues");
717 				return -rte_errno;
718 			}
719 			if (parser->queues_n == 1) {
720 				uint16_t found = 0;
721 
722 				assert(parser->queues_n);
723 				for (n = 0; n < rss->num; ++n) {
724 					if (parser->queues[0] ==
725 					    rss->queue[n]) {
726 						found = 1;
727 						break;
728 					}
729 				}
730 				if (!found) {
731 					rte_flow_error_set(error, ENOTSUP,
732 						   RTE_FLOW_ERROR_TYPE_ACTION,
733 						   actions,
734 						   "queue action not in RSS"
735 						   " queues");
736 					return -rte_errno;
737 				}
738 			}
739 			for (n = 0; n < rss->num; ++n) {
740 				if (rss->queue[n] >= priv->rxqs_n) {
741 					rte_flow_error_set(error, EINVAL,
742 						   RTE_FLOW_ERROR_TYPE_ACTION,
743 						   actions,
744 						   "queue id > number of"
745 						   " queues");
746 					return -rte_errno;
747 				}
748 			}
749 			for (n = 0; n < rss->num; ++n)
750 				parser->queues[n] = rss->queue[n];
751 			parser->queues_n = rss->num;
752 			if (priv_flow_convert_rss_conf(priv, parser,
753 						       rss->rss_conf)) {
754 				rte_flow_error_set(error, EINVAL,
755 						   RTE_FLOW_ERROR_TYPE_ACTION,
756 						   actions,
757 						   "wrong RSS configuration");
758 				return -rte_errno;
759 			}
760 		} else if (actions->type == RTE_FLOW_ACTION_TYPE_MARK) {
761 			const struct rte_flow_action_mark *mark =
762 				(const struct rte_flow_action_mark *)
763 				actions->conf;
764 
765 			if (!mark) {
766 				rte_flow_error_set(error, EINVAL,
767 						   RTE_FLOW_ERROR_TYPE_ACTION,
768 						   actions,
769 						   "mark must be defined");
770 				return -rte_errno;
771 			} else if (mark->id >= MLX5_FLOW_MARK_MAX) {
772 				rte_flow_error_set(error, ENOTSUP,
773 						   RTE_FLOW_ERROR_TYPE_ACTION,
774 						   actions,
775 						   "mark must be between 0"
776 						   " and 16777199");
777 				return -rte_errno;
778 			}
779 			parser->mark = 1;
780 			parser->mark_id = mark->id;
781 		} else if (actions->type == RTE_FLOW_ACTION_TYPE_FLAG) {
782 			parser->mark = 1;
783 		} else if (actions->type == RTE_FLOW_ACTION_TYPE_COUNT &&
784 			   priv->counter_set_supported) {
785 			parser->count = 1;
786 		} else {
787 			goto exit_action_not_supported;
788 		}
789 	}
790 	if (parser->drop && parser->mark)
791 		parser->mark = 0;
792 	if (!parser->queues_n && !parser->drop) {
793 		rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE,
794 				   NULL, "no valid action");
795 		return -rte_errno;
796 	}
797 	return 0;
798 exit_action_not_supported:
799 	rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
800 			   actions, "action not supported");
801 	return -rte_errno;
802 }
803 
804 /**
805  * Validate items.
806  *
807  * @param priv
808  *   Pointer to private structure.
809  * @param[in] items
810  *   Pattern specification (list terminated by the END pattern item).
811  * @param[out] error
812  *   Perform verbose error reporting if not NULL.
813  * @param[in, out] parser
814  *   Internal parser structure.
815  *
816  * @return
817  *   0 on success, a negative errno value otherwise and rte_errno is set.
818  */
819 static int
820 priv_flow_convert_items_validate(struct priv *priv,
821 				 const struct rte_flow_item items[],
822 				 struct rte_flow_error *error,
823 				 struct mlx5_flow_parse *parser)
824 {
825 	const struct mlx5_flow_items *cur_item = mlx5_flow_items;
826 	unsigned int i;
827 
828 	(void)priv;
829 	/* Initialise the offsets to start after verbs attribute. */
830 	if (parser->drop) {
831 		parser->drop_q.offset = sizeof(struct ibv_flow_attr);
832 	} else {
833 		for (i = 0; i != hash_rxq_init_n; ++i)
834 			parser->queue[i].offset = sizeof(struct ibv_flow_attr);
835 	}
836 	for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
837 		const struct mlx5_flow_items *token = NULL;
838 		unsigned int n;
839 		int err;
840 
841 		if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
842 			continue;
843 		for (i = 0;
844 		     cur_item->items &&
845 		     cur_item->items[i] != RTE_FLOW_ITEM_TYPE_END;
846 		     ++i) {
847 			if (cur_item->items[i] == items->type) {
848 				token = &mlx5_flow_items[items->type];
849 				break;
850 			}
851 		}
852 		if (!token)
853 			goto exit_item_not_supported;
854 		cur_item = token;
855 		err = mlx5_flow_item_validate(items,
856 					      (const uint8_t *)cur_item->mask,
857 					      cur_item->mask_sz);
858 		if (err)
859 			goto exit_item_not_supported;
860 		if (items->type == RTE_FLOW_ITEM_TYPE_VXLAN) {
861 			if (parser->inner) {
862 				rte_flow_error_set(error, ENOTSUP,
863 						   RTE_FLOW_ERROR_TYPE_ITEM,
864 						   items,
865 						   "cannot recognize multiple"
866 						   " VXLAN encapsulations");
867 				return -rte_errno;
868 			}
869 			parser->inner = IBV_FLOW_SPEC_INNER;
870 		}
871 		if (parser->drop) {
872 			parser->drop_q.offset += cur_item->dst_sz;
873 		} else if (parser->queues_n == 1) {
874 			parser->queue[HASH_RXQ_ETH].offset += cur_item->dst_sz;
875 		} else {
876 			for (n = 0; n != hash_rxq_init_n; ++n)
877 				parser->queue[n].offset += cur_item->dst_sz;
878 		}
879 	}
880 	if (parser->mark) {
881 		for (i = 0; i != hash_rxq_init_n; ++i)
882 			parser->queue[i].offset +=
883 				sizeof(struct ibv_flow_spec_action_tag);
884 	}
885 	if (parser->count) {
886 		unsigned int size = sizeof(struct ibv_flow_spec_counter_action);
887 
888 		if (parser->drop) {
889 			parser->drop_q.offset += size;
890 		} else {
891 			for (i = 0; i != hash_rxq_init_n; ++i)
892 				parser->queue[i].offset += size;
893 		}
894 	}
895 	return 0;
896 exit_item_not_supported:
897 	rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
898 			   items, "item not supported");
899 	return -rte_errno;
900 }
901 
902 /**
903  * Allocate memory space to store verbs flow attributes.
904  *
905  * @param priv
906  *   Pointer to private structure.
907  * @param[in] priority
908  *   Flow priority.
909  * @param[in] size
910  *   Amount of byte to allocate.
911  * @param[out] error
912  *   Perform verbose error reporting if not NULL.
913  *
914  * @return
915  *   A verbs flow attribute on success, NULL otherwise.
916  */
917 static struct ibv_flow_attr*
918 priv_flow_convert_allocate(struct priv *priv,
919 			   unsigned int priority,
920 			   unsigned int size,
921 			   struct rte_flow_error *error)
922 {
923 	struct ibv_flow_attr *ibv_attr;
924 
925 	(void)priv;
926 	ibv_attr = rte_calloc(__func__, 1, size, 0);
927 	if (!ibv_attr) {
928 		rte_flow_error_set(error, ENOMEM,
929 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
930 				   NULL,
931 				   "cannot allocate verbs spec attributes.");
932 		return NULL;
933 	}
934 	ibv_attr->priority = priority;
935 	return ibv_attr;
936 }
937 
938 /**
939  * Finalise verbs flow attributes.
940  *
941  * @param priv
942  *   Pointer to private structure.
943  * @param[in, out] parser
944  *   Internal parser structure.
945  */
946 static void
947 priv_flow_convert_finalise(struct priv *priv, struct mlx5_flow_parse *parser)
948 {
949 	const unsigned int ipv4 =
950 		hash_rxq_init[parser->layer].ip_version == MLX5_IPV4;
951 	const enum hash_rxq_type hmin = ipv4 ? HASH_RXQ_TCPV4 : HASH_RXQ_TCPV6;
952 	const enum hash_rxq_type hmax = ipv4 ? HASH_RXQ_IPV4 : HASH_RXQ_IPV6;
953 	const enum hash_rxq_type ohmin = ipv4 ? HASH_RXQ_TCPV6 : HASH_RXQ_TCPV4;
954 	const enum hash_rxq_type ohmax = ipv4 ? HASH_RXQ_IPV6 : HASH_RXQ_IPV4;
955 	const enum hash_rxq_type ip = ipv4 ? HASH_RXQ_IPV4 : HASH_RXQ_IPV6;
956 	unsigned int i;
957 
958 	(void)priv;
959 	if (parser->layer == HASH_RXQ_ETH) {
960 		goto fill;
961 	} else {
962 		/*
963 		 * This layer becomes useless as the pattern define under
964 		 * layers.
965 		 */
966 		rte_free(parser->queue[HASH_RXQ_ETH].ibv_attr);
967 		parser->queue[HASH_RXQ_ETH].ibv_attr = NULL;
968 	}
969 	/* Remove opposite kind of layer e.g. IPv6 if the pattern is IPv4. */
970 	for (i = ohmin; i != (ohmax + 1); ++i) {
971 		if (!parser->queue[i].ibv_attr)
972 			continue;
973 		rte_free(parser->queue[i].ibv_attr);
974 		parser->queue[i].ibv_attr = NULL;
975 	}
976 	/* Remove impossible flow according to the RSS configuration. */
977 	if (hash_rxq_init[parser->layer].dpdk_rss_hf &
978 	    parser->rss_conf.rss_hf) {
979 		/* Remove any other flow. */
980 		for (i = hmin; i != (hmax + 1); ++i) {
981 			if ((i == parser->layer) ||
982 			     (!parser->queue[i].ibv_attr))
983 				continue;
984 			rte_free(parser->queue[i].ibv_attr);
985 			parser->queue[i].ibv_attr = NULL;
986 		}
987 	} else  if (!parser->queue[ip].ibv_attr) {
988 		/* no RSS possible with the current configuration. */
989 		parser->queues_n = 1;
990 		return;
991 	}
992 fill:
993 	/*
994 	 * Fill missing layers in verbs specifications, or compute the correct
995 	 * offset to allocate the memory space for the attributes and
996 	 * specifications.
997 	 */
998 	for (i = 0; i != hash_rxq_init_n - 1; ++i) {
999 		union {
1000 			struct ibv_flow_spec_ipv4_ext ipv4;
1001 			struct ibv_flow_spec_ipv6 ipv6;
1002 			struct ibv_flow_spec_tcp_udp udp_tcp;
1003 		} specs;
1004 		void *dst;
1005 		uint16_t size;
1006 
1007 		if (i == parser->layer)
1008 			continue;
1009 		if (parser->layer == HASH_RXQ_ETH) {
1010 			if (hash_rxq_init[i].ip_version == MLX5_IPV4) {
1011 				size = sizeof(struct ibv_flow_spec_ipv4_ext);
1012 				specs.ipv4 = (struct ibv_flow_spec_ipv4_ext){
1013 					.type = IBV_FLOW_SPEC_IPV4_EXT,
1014 					.size = size,
1015 				};
1016 			} else {
1017 				size = sizeof(struct ibv_flow_spec_ipv6);
1018 				specs.ipv6 = (struct ibv_flow_spec_ipv6){
1019 					.type = IBV_FLOW_SPEC_IPV6,
1020 					.size = size,
1021 				};
1022 			}
1023 			if (parser->queue[i].ibv_attr) {
1024 				dst = (void *)((uintptr_t)
1025 					       parser->queue[i].ibv_attr +
1026 					       parser->queue[i].offset);
1027 				memcpy(dst, &specs, size);
1028 				++parser->queue[i].ibv_attr->num_of_specs;
1029 			}
1030 			parser->queue[i].offset += size;
1031 		}
1032 		if ((i == HASH_RXQ_UDPV4) || (i == HASH_RXQ_TCPV4) ||
1033 		    (i == HASH_RXQ_UDPV6) || (i == HASH_RXQ_TCPV6)) {
1034 			size = sizeof(struct ibv_flow_spec_tcp_udp);
1035 			specs.udp_tcp = (struct ibv_flow_spec_tcp_udp) {
1036 				.type = ((i == HASH_RXQ_UDPV4 ||
1037 					  i == HASH_RXQ_UDPV6) ?
1038 					 IBV_FLOW_SPEC_UDP :
1039 					 IBV_FLOW_SPEC_TCP),
1040 				.size = size,
1041 			};
1042 			if (parser->queue[i].ibv_attr) {
1043 				dst = (void *)((uintptr_t)
1044 					       parser->queue[i].ibv_attr +
1045 					       parser->queue[i].offset);
1046 				memcpy(dst, &specs, size);
1047 				++parser->queue[i].ibv_attr->num_of_specs;
1048 			}
1049 			parser->queue[i].offset += size;
1050 		}
1051 	}
1052 }
1053 
1054 /**
1055  * Validate and convert a flow supported by the NIC.
1056  *
1057  * @param priv
1058  *   Pointer to private structure.
1059  * @param[in] attr
1060  *   Flow rule attributes.
1061  * @param[in] pattern
1062  *   Pattern specification (list terminated by the END pattern item).
1063  * @param[in] actions
1064  *   Associated actions (list terminated by the END action).
1065  * @param[out] error
1066  *   Perform verbose error reporting if not NULL.
1067  * @param[in, out] parser
1068  *   Internal parser structure.
1069  *
1070  * @return
1071  *   0 on success, a negative errno value otherwise and rte_errno is set.
1072  */
1073 static int
1074 priv_flow_convert(struct priv *priv,
1075 		  const struct rte_flow_attr *attr,
1076 		  const struct rte_flow_item items[],
1077 		  const struct rte_flow_action actions[],
1078 		  struct rte_flow_error *error,
1079 		  struct mlx5_flow_parse *parser)
1080 {
1081 	const struct mlx5_flow_items *cur_item = mlx5_flow_items;
1082 	unsigned int i;
1083 	int ret;
1084 
1085 	/* First step. Validate the attributes, items and actions. */
1086 	*parser = (struct mlx5_flow_parse){
1087 		.create = parser->create,
1088 		.layer = HASH_RXQ_ETH,
1089 		.mark_id = MLX5_FLOW_MARK_DEFAULT,
1090 	};
1091 	ret = priv_flow_convert_attributes(priv, attr, error, parser);
1092 	if (ret)
1093 		return ret;
1094 	ret = priv_flow_convert_actions(priv, actions, error, parser);
1095 	if (ret)
1096 		return ret;
1097 	ret = priv_flow_convert_items_validate(priv, items, error, parser);
1098 	if (ret)
1099 		return ret;
1100 	priv_flow_convert_finalise(priv, parser);
1101 	/*
1102 	 * Second step.
1103 	 * Allocate the memory space to store verbs specifications.
1104 	 */
1105 	if (parser->drop) {
1106 		parser->drop_q.ibv_attr =
1107 			priv_flow_convert_allocate(priv, attr->priority,
1108 						   parser->drop_q.offset,
1109 						   error);
1110 		if (!parser->drop_q.ibv_attr)
1111 			return ENOMEM;
1112 		parser->drop_q.offset = sizeof(struct ibv_flow_attr);
1113 	} else if (parser->queues_n == 1) {
1114 		unsigned int priority =
1115 			attr->priority +
1116 			hash_rxq_init[HASH_RXQ_ETH].flow_priority;
1117 		unsigned int offset = parser->queue[HASH_RXQ_ETH].offset;
1118 
1119 		parser->queue[HASH_RXQ_ETH].ibv_attr =
1120 			priv_flow_convert_allocate(priv, priority,
1121 						   offset, error);
1122 		if (!parser->queue[HASH_RXQ_ETH].ibv_attr)
1123 			return ENOMEM;
1124 		parser->queue[HASH_RXQ_ETH].offset =
1125 			sizeof(struct ibv_flow_attr);
1126 	} else {
1127 		for (i = 0; i != hash_rxq_init_n; ++i) {
1128 			unsigned int priority =
1129 				attr->priority +
1130 				hash_rxq_init[i].flow_priority;
1131 			unsigned int offset;
1132 
1133 			if (!(parser->rss_conf.rss_hf &
1134 			      hash_rxq_init[i].dpdk_rss_hf) &&
1135 			    (i != HASH_RXQ_ETH))
1136 				continue;
1137 			offset = parser->queue[i].offset;
1138 			parser->queue[i].ibv_attr =
1139 				priv_flow_convert_allocate(priv, priority,
1140 							   offset, error);
1141 			if (!parser->queue[i].ibv_attr)
1142 				goto exit_enomem;
1143 			parser->queue[i].offset = sizeof(struct ibv_flow_attr);
1144 		}
1145 	}
1146 	/* Third step. Conversion parse, fill the specifications. */
1147 	parser->inner = 0;
1148 	for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
1149 		if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
1150 			continue;
1151 		cur_item = &mlx5_flow_items[items->type];
1152 		ret = cur_item->convert(items,
1153 					(cur_item->default_mask ?
1154 					 cur_item->default_mask :
1155 					 cur_item->mask),
1156 					parser);
1157 		if (ret) {
1158 			rte_flow_error_set(error, ret,
1159 					   RTE_FLOW_ERROR_TYPE_ITEM,
1160 					   items, "item not supported");
1161 			goto exit_free;
1162 		}
1163 	}
1164 	if (parser->mark)
1165 		mlx5_flow_create_flag_mark(parser, parser->mark_id);
1166 	if (parser->count && parser->create) {
1167 		mlx5_flow_create_count(priv, parser);
1168 		if (!parser->cs)
1169 			goto exit_count_error;
1170 	}
1171 	/*
1172 	 * Last step. Complete missing specification to reach the RSS
1173 	 * configuration.
1174 	 */
1175 	if (parser->drop) {
1176 		/*
1177 		 * Drop queue priority needs to be adjusted to
1178 		 * their most specific layer priority.
1179 		 */
1180 		parser->drop_q.ibv_attr->priority =
1181 			attr->priority +
1182 			hash_rxq_init[parser->layer].flow_priority;
1183 	} else if (parser->queues_n > 1) {
1184 		priv_flow_convert_finalise(priv, parser);
1185 	} else {
1186 		/*
1187 		 * Action queue have their priority overridden with
1188 		 * Ethernet priority, this priority needs to be adjusted to
1189 		 * their most specific layer priority.
1190 		 */
1191 		parser->queue[HASH_RXQ_ETH].ibv_attr->priority =
1192 			attr->priority +
1193 			hash_rxq_init[parser->layer].flow_priority;
1194 	}
1195 exit_free:
1196 	/* Only verification is expected, all resources should be released. */
1197 	if (!parser->create) {
1198 		if (parser->drop) {
1199 			rte_free(parser->drop_q.ibv_attr);
1200 			parser->drop_q.ibv_attr = NULL;
1201 		}
1202 		for (i = 0; i != hash_rxq_init_n; ++i) {
1203 			if (parser->queue[i].ibv_attr) {
1204 				rte_free(parser->queue[i].ibv_attr);
1205 				parser->queue[i].ibv_attr = NULL;
1206 			}
1207 		}
1208 	}
1209 	return ret;
1210 exit_enomem:
1211 	for (i = 0; i != hash_rxq_init_n; ++i) {
1212 		if (parser->queue[i].ibv_attr) {
1213 			rte_free(parser->queue[i].ibv_attr);
1214 			parser->queue[i].ibv_attr = NULL;
1215 		}
1216 	}
1217 	rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1218 			   NULL, "cannot allocate verbs spec attributes.");
1219 	return ret;
1220 exit_count_error:
1221 	rte_flow_error_set(error, EINVAL, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1222 			   NULL, "cannot create counter.");
1223 	return rte_errno;
1224 }
1225 
1226 /**
1227  * Copy the specification created into the flow.
1228  *
1229  * @param parser
1230  *   Internal parser structure.
1231  * @param src
1232  *   Create specification.
1233  * @param size
1234  *   Size in bytes of the specification to copy.
1235  */
1236 static void
1237 mlx5_flow_create_copy(struct mlx5_flow_parse *parser, void *src,
1238 		      unsigned int size)
1239 {
1240 	unsigned int i;
1241 	void *dst;
1242 
1243 	if (parser->drop) {
1244 		dst = (void *)((uintptr_t)parser->drop_q.ibv_attr +
1245 				parser->drop_q.offset);
1246 		memcpy(dst, src, size);
1247 		++parser->drop_q.ibv_attr->num_of_specs;
1248 		parser->drop_q.offset += size;
1249 		return;
1250 	}
1251 	for (i = 0; i != hash_rxq_init_n; ++i) {
1252 		if (!parser->queue[i].ibv_attr)
1253 			continue;
1254 		/* Specification must be the same l3 type or none. */
1255 		if (parser->layer == HASH_RXQ_ETH ||
1256 		    (hash_rxq_init[parser->layer].ip_version ==
1257 		     hash_rxq_init[i].ip_version) ||
1258 		    (hash_rxq_init[i].ip_version == 0)) {
1259 			dst = (void *)((uintptr_t)parser->queue[i].ibv_attr +
1260 					parser->queue[i].offset);
1261 			memcpy(dst, src, size);
1262 			++parser->queue[i].ibv_attr->num_of_specs;
1263 			parser->queue[i].offset += size;
1264 		}
1265 	}
1266 }
1267 
1268 /**
1269  * Convert Ethernet item to Verbs specification.
1270  *
1271  * @param item[in]
1272  *   Item specification.
1273  * @param default_mask[in]
1274  *   Default bit-masks to use when item->mask is not provided.
1275  * @param data[in, out]
1276  *   User structure.
1277  */
1278 static int
1279 mlx5_flow_create_eth(const struct rte_flow_item *item,
1280 		     const void *default_mask,
1281 		     void *data)
1282 {
1283 	const struct rte_flow_item_eth *spec = item->spec;
1284 	const struct rte_flow_item_eth *mask = item->mask;
1285 	struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1286 	const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
1287 	struct ibv_flow_spec_eth eth = {
1288 		.type = parser->inner | IBV_FLOW_SPEC_ETH,
1289 		.size = eth_size,
1290 	};
1291 
1292 	/* Don't update layer for the inner pattern. */
1293 	if (!parser->inner)
1294 		parser->layer = HASH_RXQ_ETH;
1295 	if (spec) {
1296 		unsigned int i;
1297 
1298 		if (!mask)
1299 			mask = default_mask;
1300 		memcpy(&eth.val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
1301 		memcpy(&eth.val.src_mac, spec->src.addr_bytes, ETHER_ADDR_LEN);
1302 		eth.val.ether_type = spec->type;
1303 		memcpy(&eth.mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
1304 		memcpy(&eth.mask.src_mac, mask->src.addr_bytes, ETHER_ADDR_LEN);
1305 		eth.mask.ether_type = mask->type;
1306 		/* Remove unwanted bits from values. */
1307 		for (i = 0; i < ETHER_ADDR_LEN; ++i) {
1308 			eth.val.dst_mac[i] &= eth.mask.dst_mac[i];
1309 			eth.val.src_mac[i] &= eth.mask.src_mac[i];
1310 		}
1311 		eth.val.ether_type &= eth.mask.ether_type;
1312 	}
1313 	mlx5_flow_create_copy(parser, &eth, eth_size);
1314 	return 0;
1315 }
1316 
1317 /**
1318  * Convert VLAN item to Verbs specification.
1319  *
1320  * @param item[in]
1321  *   Item specification.
1322  * @param default_mask[in]
1323  *   Default bit-masks to use when item->mask is not provided.
1324  * @param data[in, out]
1325  *   User structure.
1326  */
1327 static int
1328 mlx5_flow_create_vlan(const struct rte_flow_item *item,
1329 		      const void *default_mask,
1330 		      void *data)
1331 {
1332 	const struct rte_flow_item_vlan *spec = item->spec;
1333 	const struct rte_flow_item_vlan *mask = item->mask;
1334 	struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1335 	struct ibv_flow_spec_eth *eth;
1336 	const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
1337 
1338 	if (spec) {
1339 		unsigned int i;
1340 		if (!mask)
1341 			mask = default_mask;
1342 
1343 		if (parser->drop) {
1344 			eth = (void *)((uintptr_t)parser->drop_q.ibv_attr +
1345 				       parser->drop_q.offset - eth_size);
1346 			eth->val.vlan_tag = spec->tci;
1347 			eth->mask.vlan_tag = mask->tci;
1348 			eth->val.vlan_tag &= eth->mask.vlan_tag;
1349 			return 0;
1350 		}
1351 		for (i = 0; i != hash_rxq_init_n; ++i) {
1352 			if (!parser->queue[i].ibv_attr)
1353 				continue;
1354 
1355 			eth = (void *)((uintptr_t)parser->queue[i].ibv_attr +
1356 				       parser->queue[i].offset - eth_size);
1357 			eth->val.vlan_tag = spec->tci;
1358 			eth->mask.vlan_tag = mask->tci;
1359 			eth->val.vlan_tag &= eth->mask.vlan_tag;
1360 		}
1361 	}
1362 	return 0;
1363 }
1364 
1365 /**
1366  * Convert IPv4 item to Verbs specification.
1367  *
1368  * @param item[in]
1369  *   Item specification.
1370  * @param default_mask[in]
1371  *   Default bit-masks to use when item->mask is not provided.
1372  * @param data[in, out]
1373  *   User structure.
1374  */
1375 static int
1376 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
1377 		      const void *default_mask,
1378 		      void *data)
1379 {
1380 	const struct rte_flow_item_ipv4 *spec = item->spec;
1381 	const struct rte_flow_item_ipv4 *mask = item->mask;
1382 	struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1383 	unsigned int ipv4_size = sizeof(struct ibv_flow_spec_ipv4_ext);
1384 	struct ibv_flow_spec_ipv4_ext ipv4 = {
1385 		.type = parser->inner | IBV_FLOW_SPEC_IPV4_EXT,
1386 		.size = ipv4_size,
1387 	};
1388 
1389 	/* Don't update layer for the inner pattern. */
1390 	if (!parser->inner)
1391 		parser->layer = HASH_RXQ_IPV4;
1392 	if (spec) {
1393 		if (!mask)
1394 			mask = default_mask;
1395 		ipv4.val = (struct ibv_flow_ipv4_ext_filter){
1396 			.src_ip = spec->hdr.src_addr,
1397 			.dst_ip = spec->hdr.dst_addr,
1398 			.proto = spec->hdr.next_proto_id,
1399 			.tos = spec->hdr.type_of_service,
1400 		};
1401 		ipv4.mask = (struct ibv_flow_ipv4_ext_filter){
1402 			.src_ip = mask->hdr.src_addr,
1403 			.dst_ip = mask->hdr.dst_addr,
1404 			.proto = mask->hdr.next_proto_id,
1405 			.tos = mask->hdr.type_of_service,
1406 		};
1407 		/* Remove unwanted bits from values. */
1408 		ipv4.val.src_ip &= ipv4.mask.src_ip;
1409 		ipv4.val.dst_ip &= ipv4.mask.dst_ip;
1410 		ipv4.val.proto &= ipv4.mask.proto;
1411 		ipv4.val.tos &= ipv4.mask.tos;
1412 	}
1413 	mlx5_flow_create_copy(parser, &ipv4, ipv4_size);
1414 	return 0;
1415 }
1416 
1417 /**
1418  * Convert IPv6 item to Verbs specification.
1419  *
1420  * @param item[in]
1421  *   Item specification.
1422  * @param default_mask[in]
1423  *   Default bit-masks to use when item->mask is not provided.
1424  * @param data[in, out]
1425  *   User structure.
1426  */
1427 static int
1428 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
1429 		      const void *default_mask,
1430 		      void *data)
1431 {
1432 	const struct rte_flow_item_ipv6 *spec = item->spec;
1433 	const struct rte_flow_item_ipv6 *mask = item->mask;
1434 	struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1435 	unsigned int ipv6_size = sizeof(struct ibv_flow_spec_ipv6);
1436 	struct ibv_flow_spec_ipv6 ipv6 = {
1437 		.type = parser->inner | IBV_FLOW_SPEC_IPV6,
1438 		.size = ipv6_size,
1439 	};
1440 
1441 	/* Don't update layer for the inner pattern. */
1442 	if (!parser->inner)
1443 		parser->layer = HASH_RXQ_IPV6;
1444 	if (spec) {
1445 		unsigned int i;
1446 
1447 		if (!mask)
1448 			mask = default_mask;
1449 		memcpy(&ipv6.val.src_ip, spec->hdr.src_addr,
1450 		       RTE_DIM(ipv6.val.src_ip));
1451 		memcpy(&ipv6.val.dst_ip, spec->hdr.dst_addr,
1452 		       RTE_DIM(ipv6.val.dst_ip));
1453 		memcpy(&ipv6.mask.src_ip, mask->hdr.src_addr,
1454 		       RTE_DIM(ipv6.mask.src_ip));
1455 		memcpy(&ipv6.mask.dst_ip, mask->hdr.dst_addr,
1456 		       RTE_DIM(ipv6.mask.dst_ip));
1457 		ipv6.mask.flow_label = mask->hdr.vtc_flow;
1458 		ipv6.mask.next_hdr = mask->hdr.proto;
1459 		ipv6.mask.hop_limit = mask->hdr.hop_limits;
1460 		/* Remove unwanted bits from values. */
1461 		for (i = 0; i < RTE_DIM(ipv6.val.src_ip); ++i) {
1462 			ipv6.val.src_ip[i] &= ipv6.mask.src_ip[i];
1463 			ipv6.val.dst_ip[i] &= ipv6.mask.dst_ip[i];
1464 		}
1465 		ipv6.val.flow_label &= ipv6.mask.flow_label;
1466 		ipv6.val.next_hdr &= ipv6.mask.next_hdr;
1467 		ipv6.val.hop_limit &= ipv6.mask.hop_limit;
1468 	}
1469 	mlx5_flow_create_copy(parser, &ipv6, ipv6_size);
1470 	return 0;
1471 }
1472 
1473 /**
1474  * Convert UDP item to Verbs specification.
1475  *
1476  * @param item[in]
1477  *   Item specification.
1478  * @param default_mask[in]
1479  *   Default bit-masks to use when item->mask is not provided.
1480  * @param data[in, out]
1481  *   User structure.
1482  */
1483 static int
1484 mlx5_flow_create_udp(const struct rte_flow_item *item,
1485 		     const void *default_mask,
1486 		     void *data)
1487 {
1488 	const struct rte_flow_item_udp *spec = item->spec;
1489 	const struct rte_flow_item_udp *mask = item->mask;
1490 	struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1491 	unsigned int udp_size = sizeof(struct ibv_flow_spec_tcp_udp);
1492 	struct ibv_flow_spec_tcp_udp udp = {
1493 		.type = parser->inner | IBV_FLOW_SPEC_UDP,
1494 		.size = udp_size,
1495 	};
1496 
1497 	/* Don't update layer for the inner pattern. */
1498 	if (!parser->inner) {
1499 		if (parser->layer == HASH_RXQ_IPV4)
1500 			parser->layer = HASH_RXQ_UDPV4;
1501 		else
1502 			parser->layer = HASH_RXQ_UDPV6;
1503 	}
1504 	if (spec) {
1505 		if (!mask)
1506 			mask = default_mask;
1507 		udp.val.dst_port = spec->hdr.dst_port;
1508 		udp.val.src_port = spec->hdr.src_port;
1509 		udp.mask.dst_port = mask->hdr.dst_port;
1510 		udp.mask.src_port = mask->hdr.src_port;
1511 		/* Remove unwanted bits from values. */
1512 		udp.val.src_port &= udp.mask.src_port;
1513 		udp.val.dst_port &= udp.mask.dst_port;
1514 	}
1515 	mlx5_flow_create_copy(parser, &udp, udp_size);
1516 	return 0;
1517 }
1518 
1519 /**
1520  * Convert TCP item to Verbs specification.
1521  *
1522  * @param item[in]
1523  *   Item specification.
1524  * @param default_mask[in]
1525  *   Default bit-masks to use when item->mask is not provided.
1526  * @param data[in, out]
1527  *   User structure.
1528  */
1529 static int
1530 mlx5_flow_create_tcp(const struct rte_flow_item *item,
1531 		     const void *default_mask,
1532 		     void *data)
1533 {
1534 	const struct rte_flow_item_tcp *spec = item->spec;
1535 	const struct rte_flow_item_tcp *mask = item->mask;
1536 	struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1537 	unsigned int tcp_size = sizeof(struct ibv_flow_spec_tcp_udp);
1538 	struct ibv_flow_spec_tcp_udp tcp = {
1539 		.type = parser->inner | IBV_FLOW_SPEC_TCP,
1540 		.size = tcp_size,
1541 	};
1542 
1543 	/* Don't update layer for the inner pattern. */
1544 	if (!parser->inner) {
1545 		if (parser->layer == HASH_RXQ_IPV4)
1546 			parser->layer = HASH_RXQ_TCPV4;
1547 		else
1548 			parser->layer = HASH_RXQ_TCPV6;
1549 	}
1550 	if (spec) {
1551 		if (!mask)
1552 			mask = default_mask;
1553 		tcp.val.dst_port = spec->hdr.dst_port;
1554 		tcp.val.src_port = spec->hdr.src_port;
1555 		tcp.mask.dst_port = mask->hdr.dst_port;
1556 		tcp.mask.src_port = mask->hdr.src_port;
1557 		/* Remove unwanted bits from values. */
1558 		tcp.val.src_port &= tcp.mask.src_port;
1559 		tcp.val.dst_port &= tcp.mask.dst_port;
1560 	}
1561 	mlx5_flow_create_copy(parser, &tcp, tcp_size);
1562 	return 0;
1563 }
1564 
1565 /**
1566  * Convert VXLAN item to Verbs specification.
1567  *
1568  * @param item[in]
1569  *   Item specification.
1570  * @param default_mask[in]
1571  *   Default bit-masks to use when item->mask is not provided.
1572  * @param data[in, out]
1573  *   User structure.
1574  */
1575 static int
1576 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
1577 		       const void *default_mask,
1578 		       void *data)
1579 {
1580 	const struct rte_flow_item_vxlan *spec = item->spec;
1581 	const struct rte_flow_item_vxlan *mask = item->mask;
1582 	struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1583 	unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
1584 	struct ibv_flow_spec_tunnel vxlan = {
1585 		.type = parser->inner | IBV_FLOW_SPEC_VXLAN_TUNNEL,
1586 		.size = size,
1587 	};
1588 	union vni {
1589 		uint32_t vlan_id;
1590 		uint8_t vni[4];
1591 	} id;
1592 
1593 	id.vni[0] = 0;
1594 	parser->inner = IBV_FLOW_SPEC_INNER;
1595 	if (spec) {
1596 		if (!mask)
1597 			mask = default_mask;
1598 		memcpy(&id.vni[1], spec->vni, 3);
1599 		vxlan.val.tunnel_id = id.vlan_id;
1600 		memcpy(&id.vni[1], mask->vni, 3);
1601 		vxlan.mask.tunnel_id = id.vlan_id;
1602 		/* Remove unwanted bits from values. */
1603 		vxlan.val.tunnel_id &= vxlan.mask.tunnel_id;
1604 	}
1605 	/*
1606 	 * Tunnel id 0 is equivalent as not adding a VXLAN layer, if only this
1607 	 * layer is defined in the Verbs specification it is interpreted as
1608 	 * wildcard and all packets will match this rule, if it follows a full
1609 	 * stack layer (ex: eth / ipv4 / udp), all packets matching the layers
1610 	 * before will also match this rule.
1611 	 * To avoid such situation, VNI 0 is currently refused.
1612 	 */
1613 	if (!vxlan.val.tunnel_id)
1614 		return EINVAL;
1615 	mlx5_flow_create_copy(parser, &vxlan, size);
1616 	return 0;
1617 }
1618 
1619 /**
1620  * Convert mark/flag action to Verbs specification.
1621  *
1622  * @param parser
1623  *   Internal parser structure.
1624  * @param mark_id
1625  *   Mark identifier.
1626  */
1627 static int
1628 mlx5_flow_create_flag_mark(struct mlx5_flow_parse *parser, uint32_t mark_id)
1629 {
1630 	unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
1631 	struct ibv_flow_spec_action_tag tag = {
1632 		.type = IBV_FLOW_SPEC_ACTION_TAG,
1633 		.size = size,
1634 		.tag_id = mlx5_flow_mark_set(mark_id),
1635 	};
1636 
1637 	assert(parser->mark);
1638 	mlx5_flow_create_copy(parser, &tag, size);
1639 	return 0;
1640 }
1641 
1642 /**
1643  * Convert count action to Verbs specification.
1644  *
1645  * @param priv
1646  *   Pointer to private structure.
1647  * @param parser
1648  *   Pointer to MLX5 flow parser structure.
1649  *
1650  * @return
1651  *   0 on success, errno value on failure.
1652  */
1653 static int
1654 mlx5_flow_create_count(struct priv *priv __rte_unused,
1655 		       struct mlx5_flow_parse *parser __rte_unused)
1656 {
1657 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
1658 	unsigned int size = sizeof(struct ibv_flow_spec_counter_action);
1659 	struct ibv_counter_set_init_attr init_attr = {0};
1660 	struct ibv_flow_spec_counter_action counter = {
1661 		.type = IBV_FLOW_SPEC_ACTION_COUNT,
1662 		.size = size,
1663 		.counter_set_handle = 0,
1664 	};
1665 
1666 	init_attr.counter_set_id = 0;
1667 	parser->cs = ibv_create_counter_set(priv->ctx, &init_attr);
1668 	if (!parser->cs)
1669 		return EINVAL;
1670 	counter.counter_set_handle = parser->cs->handle;
1671 	mlx5_flow_create_copy(parser, &counter, size);
1672 #endif
1673 	return 0;
1674 }
1675 
1676 /**
1677  * Complete flow rule creation with a drop queue.
1678  *
1679  * @param priv
1680  *   Pointer to private structure.
1681  * @param parser
1682  *   Internal parser structure.
1683  * @param flow
1684  *   Pointer to the rte_flow.
1685  * @param[out] error
1686  *   Perform verbose error reporting if not NULL.
1687  *
1688  * @return
1689  *   0 on success, errno value on failure.
1690  */
1691 static int
1692 priv_flow_create_action_queue_drop(struct priv *priv,
1693 				   struct mlx5_flow_parse *parser,
1694 				   struct rte_flow *flow,
1695 				   struct rte_flow_error *error)
1696 {
1697 	struct ibv_flow_spec_action_drop *drop;
1698 	unsigned int size = sizeof(struct ibv_flow_spec_action_drop);
1699 	int err = 0;
1700 
1701 	assert(priv->pd);
1702 	assert(priv->ctx);
1703 	flow->drop = 1;
1704 	drop = (void *)((uintptr_t)parser->drop_q.ibv_attr +
1705 			parser->drop_q.offset);
1706 	*drop = (struct ibv_flow_spec_action_drop){
1707 			.type = IBV_FLOW_SPEC_ACTION_DROP,
1708 			.size = size,
1709 	};
1710 	++parser->drop_q.ibv_attr->num_of_specs;
1711 	parser->drop_q.offset += size;
1712 	flow->drxq.ibv_attr = parser->drop_q.ibv_attr;
1713 	if (parser->count)
1714 		flow->cs = parser->cs;
1715 	if (!priv->dev->data->dev_started)
1716 		return 0;
1717 	parser->drop_q.ibv_attr = NULL;
1718 	flow->drxq.ibv_flow = ibv_create_flow(priv->flow_drop_queue->qp,
1719 					      flow->drxq.ibv_attr);
1720 	if (!flow->drxq.ibv_flow) {
1721 		rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1722 				   NULL, "flow rule creation failure");
1723 		err = ENOMEM;
1724 		goto error;
1725 	}
1726 	return 0;
1727 error:
1728 	assert(flow);
1729 	if (flow->drxq.ibv_flow) {
1730 		claim_zero(ibv_destroy_flow(flow->drxq.ibv_flow));
1731 		flow->drxq.ibv_flow = NULL;
1732 	}
1733 	if (flow->drxq.ibv_attr) {
1734 		rte_free(flow->drxq.ibv_attr);
1735 		flow->drxq.ibv_attr = NULL;
1736 	}
1737 	if (flow->cs) {
1738 		claim_zero(ibv_destroy_counter_set(flow->cs));
1739 		flow->cs = NULL;
1740 		parser->cs = NULL;
1741 	}
1742 	return err;
1743 }
1744 
1745 /**
1746  * Create hash Rx queues when RSS is enabled.
1747  *
1748  * @param priv
1749  *   Pointer to private structure.
1750  * @param parser
1751  *   Internal parser structure.
1752  * @param flow
1753  *   Pointer to the rte_flow.
1754  * @param[out] error
1755  *   Perform verbose error reporting if not NULL.
1756  *
1757  * @return
1758  *   0 on success, a errno value otherwise and rte_errno is set.
1759  */
1760 static int
1761 priv_flow_create_action_queue_rss(struct priv *priv,
1762 				  struct mlx5_flow_parse *parser,
1763 				  struct rte_flow *flow,
1764 				  struct rte_flow_error *error)
1765 {
1766 	unsigned int i;
1767 
1768 	for (i = 0; i != hash_rxq_init_n; ++i) {
1769 		uint64_t hash_fields;
1770 
1771 		if (!parser->queue[i].ibv_attr)
1772 			continue;
1773 		flow->frxq[i].ibv_attr = parser->queue[i].ibv_attr;
1774 		parser->queue[i].ibv_attr = NULL;
1775 		hash_fields = hash_rxq_init[i].hash_fields;
1776 		if (!priv->dev->data->dev_started)
1777 			continue;
1778 		flow->frxq[i].hrxq =
1779 			mlx5_priv_hrxq_get(priv,
1780 					   parser->rss_conf.rss_key,
1781 					   parser->rss_conf.rss_key_len,
1782 					   hash_fields,
1783 					   parser->queues,
1784 					   parser->queues_n);
1785 		if (flow->frxq[i].hrxq)
1786 			continue;
1787 		flow->frxq[i].hrxq =
1788 			mlx5_priv_hrxq_new(priv,
1789 					   parser->rss_conf.rss_key,
1790 					   parser->rss_conf.rss_key_len,
1791 					   hash_fields,
1792 					   parser->queues,
1793 					   parser->queues_n);
1794 		if (!flow->frxq[i].hrxq) {
1795 			rte_flow_error_set(error, ENOMEM,
1796 					   RTE_FLOW_ERROR_TYPE_HANDLE,
1797 					   NULL, "cannot create hash rxq");
1798 			return ENOMEM;
1799 		}
1800 	}
1801 	return 0;
1802 }
1803 
1804 /**
1805  * Complete flow rule creation.
1806  *
1807  * @param priv
1808  *   Pointer to private structure.
1809  * @param parser
1810  *   Internal parser structure.
1811  * @param flow
1812  *   Pointer to the rte_flow.
1813  * @param[out] error
1814  *   Perform verbose error reporting if not NULL.
1815  *
1816  * @return
1817  *   0 on success, a errno value otherwise and rte_errno is set.
1818  */
1819 static int
1820 priv_flow_create_action_queue(struct priv *priv,
1821 			      struct mlx5_flow_parse *parser,
1822 			      struct rte_flow *flow,
1823 			      struct rte_flow_error *error)
1824 {
1825 	int err = 0;
1826 	unsigned int i;
1827 
1828 	assert(priv->pd);
1829 	assert(priv->ctx);
1830 	assert(!parser->drop);
1831 	err = priv_flow_create_action_queue_rss(priv, parser, flow, error);
1832 	if (err)
1833 		goto error;
1834 	if (parser->count)
1835 		flow->cs = parser->cs;
1836 	if (!priv->dev->data->dev_started)
1837 		return 0;
1838 	for (i = 0; i != hash_rxq_init_n; ++i) {
1839 		if (!flow->frxq[i].hrxq)
1840 			continue;
1841 		flow->frxq[i].ibv_flow =
1842 			ibv_create_flow(flow->frxq[i].hrxq->qp,
1843 					flow->frxq[i].ibv_attr);
1844 		if (!flow->frxq[i].ibv_flow) {
1845 			rte_flow_error_set(error, ENOMEM,
1846 					   RTE_FLOW_ERROR_TYPE_HANDLE,
1847 					   NULL, "flow rule creation failure");
1848 			err = ENOMEM;
1849 			goto error;
1850 		}
1851 		DEBUG("%p type %d QP %p ibv_flow %p",
1852 		      (void *)flow, i,
1853 		      (void *)flow->frxq[i].hrxq,
1854 		      (void *)flow->frxq[i].ibv_flow);
1855 	}
1856 	for (i = 0; i != parser->queues_n; ++i) {
1857 		struct mlx5_rxq_data *q =
1858 			(*priv->rxqs)[parser->queues[i]];
1859 
1860 		q->mark |= parser->mark;
1861 	}
1862 	return 0;
1863 error:
1864 	assert(flow);
1865 	for (i = 0; i != hash_rxq_init_n; ++i) {
1866 		if (flow->frxq[i].ibv_flow) {
1867 			struct ibv_flow *ibv_flow = flow->frxq[i].ibv_flow;
1868 
1869 			claim_zero(ibv_destroy_flow(ibv_flow));
1870 		}
1871 		if (flow->frxq[i].hrxq)
1872 			mlx5_priv_hrxq_release(priv, flow->frxq[i].hrxq);
1873 		if (flow->frxq[i].ibv_attr)
1874 			rte_free(flow->frxq[i].ibv_attr);
1875 	}
1876 	if (flow->cs) {
1877 		claim_zero(ibv_destroy_counter_set(flow->cs));
1878 		flow->cs = NULL;
1879 		parser->cs = NULL;
1880 	}
1881 	return err;
1882 }
1883 
1884 /**
1885  * Convert a flow.
1886  *
1887  * @param priv
1888  *   Pointer to private structure.
1889  * @param list
1890  *   Pointer to a TAILQ flow list.
1891  * @param[in] attr
1892  *   Flow rule attributes.
1893  * @param[in] pattern
1894  *   Pattern specification (list terminated by the END pattern item).
1895  * @param[in] actions
1896  *   Associated actions (list terminated by the END action).
1897  * @param[out] error
1898  *   Perform verbose error reporting if not NULL.
1899  *
1900  * @return
1901  *   A flow on success, NULL otherwise.
1902  */
1903 static struct rte_flow *
1904 priv_flow_create(struct priv *priv,
1905 		 struct mlx5_flows *list,
1906 		 const struct rte_flow_attr *attr,
1907 		 const struct rte_flow_item items[],
1908 		 const struct rte_flow_action actions[],
1909 		 struct rte_flow_error *error)
1910 {
1911 	struct mlx5_flow_parse parser = { .create = 1, };
1912 	struct rte_flow *flow = NULL;
1913 	unsigned int i;
1914 	int err;
1915 
1916 	err = priv_flow_convert(priv, attr, items, actions, error, &parser);
1917 	if (err)
1918 		goto exit;
1919 	flow = rte_calloc(__func__, 1,
1920 			  sizeof(*flow) + parser.queues_n * sizeof(uint16_t),
1921 			  0);
1922 	if (!flow) {
1923 		rte_flow_error_set(error, ENOMEM,
1924 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1925 				   NULL,
1926 				   "cannot allocate flow memory");
1927 		return NULL;
1928 	}
1929 	/* Copy queues configuration. */
1930 	flow->queues = (uint16_t (*)[])(flow + 1);
1931 	memcpy(flow->queues, parser.queues, parser.queues_n * sizeof(uint16_t));
1932 	flow->queues_n = parser.queues_n;
1933 	flow->mark = parser.mark;
1934 	/* Copy RSS configuration. */
1935 	flow->rss_conf = parser.rss_conf;
1936 	flow->rss_conf.rss_key = flow->rss_key;
1937 	memcpy(flow->rss_key, parser.rss_key, parser.rss_conf.rss_key_len);
1938 	/* finalise the flow. */
1939 	if (parser.drop)
1940 		err = priv_flow_create_action_queue_drop(priv, &parser, flow,
1941 							 error);
1942 	else
1943 		err = priv_flow_create_action_queue(priv, &parser, flow, error);
1944 	if (err)
1945 		goto exit;
1946 	TAILQ_INSERT_TAIL(list, flow, next);
1947 	DEBUG("Flow created %p", (void *)flow);
1948 	return flow;
1949 exit:
1950 	if (parser.drop) {
1951 		rte_free(parser.drop_q.ibv_attr);
1952 	} else {
1953 		for (i = 0; i != hash_rxq_init_n; ++i) {
1954 			if (parser.queue[i].ibv_attr)
1955 				rte_free(parser.queue[i].ibv_attr);
1956 		}
1957 	}
1958 	rte_free(flow);
1959 	return NULL;
1960 }
1961 
1962 /**
1963  * Validate a flow supported by the NIC.
1964  *
1965  * @see rte_flow_validate()
1966  * @see rte_flow_ops
1967  */
1968 int
1969 mlx5_flow_validate(struct rte_eth_dev *dev,
1970 		   const struct rte_flow_attr *attr,
1971 		   const struct rte_flow_item items[],
1972 		   const struct rte_flow_action actions[],
1973 		   struct rte_flow_error *error)
1974 {
1975 	struct priv *priv = dev->data->dev_private;
1976 	int ret;
1977 	struct mlx5_flow_parse parser = { .create = 0, };
1978 
1979 	priv_lock(priv);
1980 	ret = priv_flow_convert(priv, attr, items, actions, error, &parser);
1981 	priv_unlock(priv);
1982 	return ret;
1983 }
1984 
1985 /**
1986  * Create a flow.
1987  *
1988  * @see rte_flow_create()
1989  * @see rte_flow_ops
1990  */
1991 struct rte_flow *
1992 mlx5_flow_create(struct rte_eth_dev *dev,
1993 		 const struct rte_flow_attr *attr,
1994 		 const struct rte_flow_item items[],
1995 		 const struct rte_flow_action actions[],
1996 		 struct rte_flow_error *error)
1997 {
1998 	struct priv *priv = dev->data->dev_private;
1999 	struct rte_flow *flow;
2000 
2001 	priv_lock(priv);
2002 	flow = priv_flow_create(priv, &priv->flows, attr, items, actions,
2003 				error);
2004 	priv_unlock(priv);
2005 	return flow;
2006 }
2007 
2008 /**
2009  * Destroy a flow.
2010  *
2011  * @param priv
2012  *   Pointer to private structure.
2013  * @param list
2014  *   Pointer to a TAILQ flow list.
2015  * @param[in] flow
2016  *   Flow to destroy.
2017  */
2018 static void
2019 priv_flow_destroy(struct priv *priv,
2020 		  struct mlx5_flows *list,
2021 		  struct rte_flow *flow)
2022 {
2023 	unsigned int i;
2024 
2025 	if (flow->drop || !flow->mark)
2026 		goto free;
2027 	for (i = 0; i != flow->queues_n; ++i) {
2028 		struct rte_flow *tmp;
2029 		int mark = 0;
2030 
2031 		/*
2032 		 * To remove the mark from the queue, the queue must not be
2033 		 * present in any other marked flow (RSS or not).
2034 		 */
2035 		TAILQ_FOREACH(tmp, list, next) {
2036 			unsigned int j;
2037 			uint16_t *tqs = NULL;
2038 			uint16_t tq_n = 0;
2039 
2040 			if (!tmp->mark)
2041 				continue;
2042 			for (j = 0; j != hash_rxq_init_n; ++j) {
2043 				if (!tmp->frxq[j].hrxq)
2044 					continue;
2045 				tqs = tmp->frxq[j].hrxq->ind_table->queues;
2046 				tq_n = tmp->frxq[j].hrxq->ind_table->queues_n;
2047 			}
2048 			if (!tq_n)
2049 				continue;
2050 			for (j = 0; (j != tq_n) && !mark; j++)
2051 				if (tqs[j] == (*flow->queues)[i])
2052 					mark = 1;
2053 		}
2054 		(*priv->rxqs)[(*flow->queues)[i]]->mark = mark;
2055 	}
2056 free:
2057 	if (flow->drop) {
2058 		if (flow->drxq.ibv_flow)
2059 			claim_zero(ibv_destroy_flow(flow->drxq.ibv_flow));
2060 		rte_free(flow->drxq.ibv_attr);
2061 	} else {
2062 		for (i = 0; i != hash_rxq_init_n; ++i) {
2063 			struct mlx5_flow *frxq = &flow->frxq[i];
2064 
2065 			if (frxq->ibv_flow)
2066 				claim_zero(ibv_destroy_flow(frxq->ibv_flow));
2067 			if (frxq->hrxq)
2068 				mlx5_priv_hrxq_release(priv, frxq->hrxq);
2069 			if (frxq->ibv_attr)
2070 				rte_free(frxq->ibv_attr);
2071 		}
2072 	}
2073 	if (flow->cs) {
2074 		claim_zero(ibv_destroy_counter_set(flow->cs));
2075 		flow->cs = NULL;
2076 	}
2077 	TAILQ_REMOVE(list, flow, next);
2078 	DEBUG("Flow destroyed %p", (void *)flow);
2079 	rte_free(flow);
2080 }
2081 
2082 /**
2083  * Destroy all flows.
2084  *
2085  * @param priv
2086  *   Pointer to private structure.
2087  * @param list
2088  *   Pointer to a TAILQ flow list.
2089  */
2090 void
2091 priv_flow_flush(struct priv *priv, struct mlx5_flows *list)
2092 {
2093 	while (!TAILQ_EMPTY(list)) {
2094 		struct rte_flow *flow;
2095 
2096 		flow = TAILQ_FIRST(list);
2097 		priv_flow_destroy(priv, list, flow);
2098 	}
2099 }
2100 
2101 /**
2102  * Create drop queue.
2103  *
2104  * @param priv
2105  *   Pointer to private structure.
2106  *
2107  * @return
2108  *   0 on success.
2109  */
2110 int
2111 priv_flow_create_drop_queue(struct priv *priv)
2112 {
2113 	struct mlx5_hrxq_drop *fdq = NULL;
2114 
2115 	assert(priv->pd);
2116 	assert(priv->ctx);
2117 	fdq = rte_calloc(__func__, 1, sizeof(*fdq), 0);
2118 	if (!fdq) {
2119 		WARN("cannot allocate memory for drop queue");
2120 		goto error;
2121 	}
2122 	fdq->cq = ibv_create_cq(priv->ctx, 1, NULL, NULL, 0);
2123 	if (!fdq->cq) {
2124 		WARN("cannot allocate CQ for drop queue");
2125 		goto error;
2126 	}
2127 	fdq->wq = ibv_create_wq(priv->ctx,
2128 			&(struct ibv_wq_init_attr){
2129 			.wq_type = IBV_WQT_RQ,
2130 			.max_wr = 1,
2131 			.max_sge = 1,
2132 			.pd = priv->pd,
2133 			.cq = fdq->cq,
2134 			});
2135 	if (!fdq->wq) {
2136 		WARN("cannot allocate WQ for drop queue");
2137 		goto error;
2138 	}
2139 	fdq->ind_table = ibv_create_rwq_ind_table(priv->ctx,
2140 			&(struct ibv_rwq_ind_table_init_attr){
2141 			.log_ind_tbl_size = 0,
2142 			.ind_tbl = &fdq->wq,
2143 			.comp_mask = 0,
2144 			});
2145 	if (!fdq->ind_table) {
2146 		WARN("cannot allocate indirection table for drop queue");
2147 		goto error;
2148 	}
2149 	fdq->qp = ibv_create_qp_ex(priv->ctx,
2150 		&(struct ibv_qp_init_attr_ex){
2151 			.qp_type = IBV_QPT_RAW_PACKET,
2152 			.comp_mask =
2153 				IBV_QP_INIT_ATTR_PD |
2154 				IBV_QP_INIT_ATTR_IND_TABLE |
2155 				IBV_QP_INIT_ATTR_RX_HASH,
2156 			.rx_hash_conf = (struct ibv_rx_hash_conf){
2157 				.rx_hash_function =
2158 					IBV_RX_HASH_FUNC_TOEPLITZ,
2159 				.rx_hash_key_len = rss_hash_default_key_len,
2160 				.rx_hash_key = rss_hash_default_key,
2161 				.rx_hash_fields_mask = 0,
2162 				},
2163 			.rwq_ind_tbl = fdq->ind_table,
2164 			.pd = priv->pd
2165 		});
2166 	if (!fdq->qp) {
2167 		WARN("cannot allocate QP for drop queue");
2168 		goto error;
2169 	}
2170 	priv->flow_drop_queue = fdq;
2171 	return 0;
2172 error:
2173 	if (fdq->qp)
2174 		claim_zero(ibv_destroy_qp(fdq->qp));
2175 	if (fdq->ind_table)
2176 		claim_zero(ibv_destroy_rwq_ind_table(fdq->ind_table));
2177 	if (fdq->wq)
2178 		claim_zero(ibv_destroy_wq(fdq->wq));
2179 	if (fdq->cq)
2180 		claim_zero(ibv_destroy_cq(fdq->cq));
2181 	if (fdq)
2182 		rte_free(fdq);
2183 	priv->flow_drop_queue = NULL;
2184 	return -1;
2185 }
2186 
2187 /**
2188  * Delete drop queue.
2189  *
2190  * @param priv
2191  *   Pointer to private structure.
2192  */
2193 void
2194 priv_flow_delete_drop_queue(struct priv *priv)
2195 {
2196 	struct mlx5_hrxq_drop *fdq = priv->flow_drop_queue;
2197 
2198 	if (!fdq)
2199 		return;
2200 	if (fdq->qp)
2201 		claim_zero(ibv_destroy_qp(fdq->qp));
2202 	if (fdq->ind_table)
2203 		claim_zero(ibv_destroy_rwq_ind_table(fdq->ind_table));
2204 	if (fdq->wq)
2205 		claim_zero(ibv_destroy_wq(fdq->wq));
2206 	if (fdq->cq)
2207 		claim_zero(ibv_destroy_cq(fdq->cq));
2208 	rte_free(fdq);
2209 	priv->flow_drop_queue = NULL;
2210 }
2211 
2212 /**
2213  * Remove all flows.
2214  *
2215  * @param priv
2216  *   Pointer to private structure.
2217  * @param list
2218  *   Pointer to a TAILQ flow list.
2219  */
2220 void
2221 priv_flow_stop(struct priv *priv, struct mlx5_flows *list)
2222 {
2223 	struct rte_flow *flow;
2224 
2225 	TAILQ_FOREACH_REVERSE(flow, list, mlx5_flows, next) {
2226 		unsigned int i;
2227 
2228 		if (flow->drop) {
2229 			if (!flow->drxq.ibv_flow)
2230 				continue;
2231 			claim_zero(ibv_destroy_flow(flow->drxq.ibv_flow));
2232 			flow->drxq.ibv_flow = NULL;
2233 			/* Next flow. */
2234 			continue;
2235 		}
2236 		if (flow->mark) {
2237 			struct mlx5_ind_table_ibv *ind_tbl = NULL;
2238 
2239 			for (i = 0; i != hash_rxq_init_n; ++i) {
2240 				if (!flow->frxq[i].hrxq)
2241 					continue;
2242 				ind_tbl = flow->frxq[i].hrxq->ind_table;
2243 			}
2244 			assert(ind_tbl);
2245 			for (i = 0; i != ind_tbl->queues_n; ++i)
2246 				(*priv->rxqs)[ind_tbl->queues[i]]->mark = 0;
2247 		}
2248 		for (i = 0; i != hash_rxq_init_n; ++i) {
2249 			if (!flow->frxq[i].ibv_flow)
2250 				continue;
2251 			claim_zero(ibv_destroy_flow(flow->frxq[i].ibv_flow));
2252 			flow->frxq[i].ibv_flow = NULL;
2253 			mlx5_priv_hrxq_release(priv, flow->frxq[i].hrxq);
2254 			flow->frxq[i].hrxq = NULL;
2255 		}
2256 		DEBUG("Flow %p removed", (void *)flow);
2257 	}
2258 }
2259 
2260 /**
2261  * Add all flows.
2262  *
2263  * @param priv
2264  *   Pointer to private structure.
2265  * @param list
2266  *   Pointer to a TAILQ flow list.
2267  *
2268  * @return
2269  *   0 on success, a errno value otherwise and rte_errno is set.
2270  */
2271 int
2272 priv_flow_start(struct priv *priv, struct mlx5_flows *list)
2273 {
2274 	struct rte_flow *flow;
2275 
2276 	TAILQ_FOREACH(flow, list, next) {
2277 		unsigned int i;
2278 
2279 		if (flow->drop) {
2280 			flow->drxq.ibv_flow =
2281 				ibv_create_flow(priv->flow_drop_queue->qp,
2282 						flow->drxq.ibv_attr);
2283 			if (!flow->drxq.ibv_flow) {
2284 				DEBUG("Flow %p cannot be applied",
2285 				      (void *)flow);
2286 				rte_errno = EINVAL;
2287 				return rte_errno;
2288 			}
2289 			DEBUG("Flow %p applied", (void *)flow);
2290 			/* Next flow. */
2291 			continue;
2292 		}
2293 		for (i = 0; i != hash_rxq_init_n; ++i) {
2294 			if (!flow->frxq[i].ibv_attr)
2295 				continue;
2296 			flow->frxq[i].hrxq =
2297 				mlx5_priv_hrxq_get(priv, flow->rss_conf.rss_key,
2298 						   flow->rss_conf.rss_key_len,
2299 						   hash_rxq_init[i].hash_fields,
2300 						   (*flow->queues),
2301 						   flow->queues_n);
2302 			if (flow->frxq[i].hrxq)
2303 				goto flow_create;
2304 			flow->frxq[i].hrxq =
2305 				mlx5_priv_hrxq_new(priv, flow->rss_conf.rss_key,
2306 						   flow->rss_conf.rss_key_len,
2307 						   hash_rxq_init[i].hash_fields,
2308 						   (*flow->queues),
2309 						   flow->queues_n);
2310 			if (!flow->frxq[i].hrxq) {
2311 				DEBUG("Flow %p cannot be applied",
2312 				      (void *)flow);
2313 				rte_errno = EINVAL;
2314 				return rte_errno;
2315 			}
2316 flow_create:
2317 			flow->frxq[i].ibv_flow =
2318 				ibv_create_flow(flow->frxq[i].hrxq->qp,
2319 						flow->frxq[i].ibv_attr);
2320 			if (!flow->frxq[i].ibv_flow) {
2321 				DEBUG("Flow %p cannot be applied",
2322 				      (void *)flow);
2323 				rte_errno = EINVAL;
2324 				return rte_errno;
2325 			}
2326 			DEBUG("Flow %p applied", (void *)flow);
2327 		}
2328 		if (!flow->mark)
2329 			continue;
2330 		for (i = 0; i != flow->queues_n; ++i)
2331 			(*priv->rxqs)[(*flow->queues)[i]]->mark = 1;
2332 	}
2333 	return 0;
2334 }
2335 
2336 /**
2337  * Verify the flow list is empty
2338  *
2339  * @param priv
2340  *  Pointer to private structure.
2341  *
2342  * @return the number of flows not released.
2343  */
2344 int
2345 priv_flow_verify(struct priv *priv)
2346 {
2347 	struct rte_flow *flow;
2348 	int ret = 0;
2349 
2350 	TAILQ_FOREACH(flow, &priv->flows, next) {
2351 		DEBUG("%p: flow %p still referenced", (void *)priv,
2352 		      (void *)flow);
2353 		++ret;
2354 	}
2355 	return ret;
2356 }
2357 
2358 /**
2359  * Enable a control flow configured from the control plane.
2360  *
2361  * @param dev
2362  *   Pointer to Ethernet device.
2363  * @param eth_spec
2364  *   An Ethernet flow spec to apply.
2365  * @param eth_mask
2366  *   An Ethernet flow mask to apply.
2367  * @param vlan_spec
2368  *   A VLAN flow spec to apply.
2369  * @param vlan_mask
2370  *   A VLAN flow mask to apply.
2371  *
2372  * @return
2373  *   0 on success.
2374  */
2375 int
2376 mlx5_ctrl_flow_vlan(struct rte_eth_dev *dev,
2377 		    struct rte_flow_item_eth *eth_spec,
2378 		    struct rte_flow_item_eth *eth_mask,
2379 		    struct rte_flow_item_vlan *vlan_spec,
2380 		    struct rte_flow_item_vlan *vlan_mask)
2381 {
2382 	struct priv *priv = dev->data->dev_private;
2383 	const struct rte_flow_attr attr = {
2384 		.ingress = 1,
2385 		.priority = MLX5_CTRL_FLOW_PRIORITY,
2386 	};
2387 	struct rte_flow_item items[] = {
2388 		{
2389 			.type = RTE_FLOW_ITEM_TYPE_ETH,
2390 			.spec = eth_spec,
2391 			.last = NULL,
2392 			.mask = eth_mask,
2393 		},
2394 		{
2395 			.type = (vlan_spec) ? RTE_FLOW_ITEM_TYPE_VLAN :
2396 				RTE_FLOW_ITEM_TYPE_END,
2397 			.spec = vlan_spec,
2398 			.last = NULL,
2399 			.mask = vlan_mask,
2400 		},
2401 		{
2402 			.type = RTE_FLOW_ITEM_TYPE_END,
2403 		},
2404 	};
2405 	struct rte_flow_action actions[] = {
2406 		{
2407 			.type = RTE_FLOW_ACTION_TYPE_RSS,
2408 		},
2409 		{
2410 			.type = RTE_FLOW_ACTION_TYPE_END,
2411 		},
2412 	};
2413 	struct rte_flow *flow;
2414 	struct rte_flow_error error;
2415 	unsigned int i;
2416 	union {
2417 		struct rte_flow_action_rss rss;
2418 		struct {
2419 			const struct rte_eth_rss_conf *rss_conf;
2420 			uint16_t num;
2421 			uint16_t queue[RTE_MAX_QUEUES_PER_PORT];
2422 		} local;
2423 	} action_rss;
2424 
2425 	if (!priv->reta_idx_n)
2426 		return EINVAL;
2427 	for (i = 0; i != priv->reta_idx_n; ++i)
2428 		action_rss.local.queue[i] = (*priv->reta_idx)[i];
2429 	action_rss.local.rss_conf = &priv->rss_conf;
2430 	action_rss.local.num = priv->reta_idx_n;
2431 	actions[0].conf = (const void *)&action_rss.rss;
2432 	flow = priv_flow_create(priv, &priv->ctrl_flows, &attr, items, actions,
2433 				&error);
2434 	if (!flow)
2435 		return rte_errno;
2436 	return 0;
2437 }
2438 
2439 /**
2440  * Enable a flow control configured from the control plane.
2441  *
2442  * @param dev
2443  *   Pointer to Ethernet device.
2444  * @param eth_spec
2445  *   An Ethernet flow spec to apply.
2446  * @param eth_mask
2447  *   An Ethernet flow mask to apply.
2448  *
2449  * @return
2450  *   0 on success.
2451  */
2452 int
2453 mlx5_ctrl_flow(struct rte_eth_dev *dev,
2454 	       struct rte_flow_item_eth *eth_spec,
2455 	       struct rte_flow_item_eth *eth_mask)
2456 {
2457 	return mlx5_ctrl_flow_vlan(dev, eth_spec, eth_mask, NULL, NULL);
2458 }
2459 
2460 /**
2461  * Destroy a flow.
2462  *
2463  * @see rte_flow_destroy()
2464  * @see rte_flow_ops
2465  */
2466 int
2467 mlx5_flow_destroy(struct rte_eth_dev *dev,
2468 		  struct rte_flow *flow,
2469 		  struct rte_flow_error *error)
2470 {
2471 	struct priv *priv = dev->data->dev_private;
2472 
2473 	(void)error;
2474 	priv_lock(priv);
2475 	priv_flow_destroy(priv, &priv->flows, flow);
2476 	priv_unlock(priv);
2477 	return 0;
2478 }
2479 
2480 /**
2481  * Destroy all flows.
2482  *
2483  * @see rte_flow_flush()
2484  * @see rte_flow_ops
2485  */
2486 int
2487 mlx5_flow_flush(struct rte_eth_dev *dev,
2488 		struct rte_flow_error *error)
2489 {
2490 	struct priv *priv = dev->data->dev_private;
2491 
2492 	(void)error;
2493 	priv_lock(priv);
2494 	priv_flow_flush(priv, &priv->flows);
2495 	priv_unlock(priv);
2496 	return 0;
2497 }
2498 
2499 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
2500 /**
2501  * Query flow counter.
2502  *
2503  * @param cs
2504  *   the counter set.
2505  * @param counter_value
2506  *   returned data from the counter.
2507  *
2508  * @return
2509  *   0 on success, a errno value otherwise and rte_errno is set.
2510  */
2511 static int
2512 priv_flow_query_count(struct ibv_counter_set *cs,
2513 		      struct mlx5_flow_counter_stats *counter_stats,
2514 		      struct rte_flow_query_count *query_count,
2515 		      struct rte_flow_error *error)
2516 {
2517 	uint64_t counters[2];
2518 	struct ibv_query_counter_set_attr query_cs_attr = {
2519 		.cs = cs,
2520 		.query_flags = IBV_COUNTER_SET_FORCE_UPDATE,
2521 	};
2522 	struct ibv_counter_set_data query_out = {
2523 		.out = counters,
2524 		.outlen = 2 * sizeof(uint64_t),
2525 	};
2526 	int res = ibv_query_counter_set(&query_cs_attr, &query_out);
2527 
2528 	if (res) {
2529 		rte_flow_error_set(error, -res,
2530 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2531 				   NULL,
2532 				   "cannot read counter");
2533 		return -res;
2534 	}
2535 	query_count->hits_set = 1;
2536 	query_count->bytes_set = 1;
2537 	query_count->hits = counters[0] - counter_stats->hits;
2538 	query_count->bytes = counters[1] - counter_stats->bytes;
2539 	if (query_count->reset) {
2540 		counter_stats->hits = counters[0];
2541 		counter_stats->bytes = counters[1];
2542 	}
2543 	return 0;
2544 }
2545 
2546 /**
2547  * Query a flows.
2548  *
2549  * @see rte_flow_query()
2550  * @see rte_flow_ops
2551  */
2552 int
2553 mlx5_flow_query(struct rte_eth_dev *dev,
2554 		struct rte_flow *flow,
2555 		enum rte_flow_action_type action __rte_unused,
2556 		void *data,
2557 		struct rte_flow_error *error)
2558 {
2559 	struct priv *priv = dev->data->dev_private;
2560 	int res = EINVAL;
2561 
2562 	priv_lock(priv);
2563 	if (flow->cs) {
2564 		res = priv_flow_query_count(flow->cs,
2565 					&flow->counter_stats,
2566 					(struct rte_flow_query_count *)data,
2567 					error);
2568 	} else {
2569 		rte_flow_error_set(error, res,
2570 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2571 				   NULL,
2572 				   "no counter found for flow");
2573 	}
2574 	priv_unlock(priv);
2575 	return -res;
2576 }
2577 #endif
2578 
2579 /**
2580  * Isolated mode.
2581  *
2582  * @see rte_flow_isolate()
2583  * @see rte_flow_ops
2584  */
2585 int
2586 mlx5_flow_isolate(struct rte_eth_dev *dev,
2587 		  int enable,
2588 		  struct rte_flow_error *error)
2589 {
2590 	struct priv *priv = dev->data->dev_private;
2591 
2592 	priv_lock(priv);
2593 	if (dev->data->dev_started) {
2594 		rte_flow_error_set(error, EBUSY,
2595 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2596 				   NULL,
2597 				   "port must be stopped first");
2598 		priv_unlock(priv);
2599 		return -rte_errno;
2600 	}
2601 	priv->isolated = !!enable;
2602 	if (enable)
2603 		priv->dev->dev_ops = &mlx5_dev_ops_isolate;
2604 	else
2605 		priv->dev->dev_ops = &mlx5_dev_ops;
2606 	priv_unlock(priv);
2607 	return 0;
2608 }
2609 
2610 /**
2611  * Convert a flow director filter to a generic flow.
2612  *
2613  * @param priv
2614  *   Private structure.
2615  * @param fdir_filter
2616  *   Flow director filter to add.
2617  * @param attributes
2618  *   Generic flow parameters structure.
2619  *
2620  * @return
2621  *  0 on success, errno value on error.
2622  */
2623 static int
2624 priv_fdir_filter_convert(struct priv *priv,
2625 			 const struct rte_eth_fdir_filter *fdir_filter,
2626 			 struct mlx5_fdir *attributes)
2627 {
2628 	const struct rte_eth_fdir_input *input = &fdir_filter->input;
2629 
2630 	/* Validate queue number. */
2631 	if (fdir_filter->action.rx_queue >= priv->rxqs_n) {
2632 		ERROR("invalid queue number %d", fdir_filter->action.rx_queue);
2633 		return EINVAL;
2634 	}
2635 	attributes->attr.ingress = 1;
2636 	attributes->items[0] = (struct rte_flow_item) {
2637 		.type = RTE_FLOW_ITEM_TYPE_ETH,
2638 		.spec = &attributes->l2,
2639 		.mask = &attributes->l2_mask,
2640 	};
2641 	switch (fdir_filter->action.behavior) {
2642 	case RTE_ETH_FDIR_ACCEPT:
2643 		attributes->actions[0] = (struct rte_flow_action){
2644 			.type = RTE_FLOW_ACTION_TYPE_QUEUE,
2645 			.conf = &attributes->queue,
2646 		};
2647 		break;
2648 	case RTE_ETH_FDIR_REJECT:
2649 		attributes->actions[0] = (struct rte_flow_action){
2650 			.type = RTE_FLOW_ACTION_TYPE_DROP,
2651 		};
2652 		break;
2653 	default:
2654 		ERROR("invalid behavior %d", fdir_filter->action.behavior);
2655 		return ENOTSUP;
2656 	}
2657 	attributes->queue.index = fdir_filter->action.rx_queue;
2658 	switch (fdir_filter->input.flow_type) {
2659 	case RTE_ETH_FLOW_NONFRAG_IPV4_UDP:
2660 		attributes->l3.ipv4.hdr = (struct ipv4_hdr){
2661 			.src_addr = input->flow.udp4_flow.ip.src_ip,
2662 			.dst_addr = input->flow.udp4_flow.ip.dst_ip,
2663 			.time_to_live = input->flow.udp4_flow.ip.ttl,
2664 			.type_of_service = input->flow.udp4_flow.ip.tos,
2665 			.next_proto_id = input->flow.udp4_flow.ip.proto,
2666 		};
2667 		attributes->l4.udp.hdr = (struct udp_hdr){
2668 			.src_port = input->flow.udp4_flow.src_port,
2669 			.dst_port = input->flow.udp4_flow.dst_port,
2670 		};
2671 		attributes->items[1] = (struct rte_flow_item){
2672 			.type = RTE_FLOW_ITEM_TYPE_IPV4,
2673 			.spec = &attributes->l3,
2674 		};
2675 		attributes->items[2] = (struct rte_flow_item){
2676 			.type = RTE_FLOW_ITEM_TYPE_UDP,
2677 			.spec = &attributes->l4,
2678 		};
2679 		break;
2680 	case RTE_ETH_FLOW_NONFRAG_IPV4_TCP:
2681 		attributes->l3.ipv4.hdr = (struct ipv4_hdr){
2682 			.src_addr = input->flow.tcp4_flow.ip.src_ip,
2683 			.dst_addr = input->flow.tcp4_flow.ip.dst_ip,
2684 			.time_to_live = input->flow.tcp4_flow.ip.ttl,
2685 			.type_of_service = input->flow.tcp4_flow.ip.tos,
2686 			.next_proto_id = input->flow.tcp4_flow.ip.proto,
2687 		};
2688 		attributes->l4.tcp.hdr = (struct tcp_hdr){
2689 			.src_port = input->flow.tcp4_flow.src_port,
2690 			.dst_port = input->flow.tcp4_flow.dst_port,
2691 		};
2692 		attributes->items[1] = (struct rte_flow_item){
2693 			.type = RTE_FLOW_ITEM_TYPE_IPV4,
2694 			.spec = &attributes->l3,
2695 		};
2696 		attributes->items[2] = (struct rte_flow_item){
2697 			.type = RTE_FLOW_ITEM_TYPE_TCP,
2698 			.spec = &attributes->l4,
2699 		};
2700 		break;
2701 	case RTE_ETH_FLOW_NONFRAG_IPV4_OTHER:
2702 		attributes->l3.ipv4.hdr = (struct ipv4_hdr){
2703 			.src_addr = input->flow.ip4_flow.src_ip,
2704 			.dst_addr = input->flow.ip4_flow.dst_ip,
2705 			.time_to_live = input->flow.ip4_flow.ttl,
2706 			.type_of_service = input->flow.ip4_flow.tos,
2707 			.next_proto_id = input->flow.ip4_flow.proto,
2708 		};
2709 		attributes->items[1] = (struct rte_flow_item){
2710 			.type = RTE_FLOW_ITEM_TYPE_IPV4,
2711 			.spec = &attributes->l3,
2712 		};
2713 		break;
2714 	case RTE_ETH_FLOW_NONFRAG_IPV6_UDP:
2715 		attributes->l3.ipv6.hdr = (struct ipv6_hdr){
2716 			.hop_limits = input->flow.udp6_flow.ip.hop_limits,
2717 			.proto = input->flow.udp6_flow.ip.proto,
2718 		};
2719 		memcpy(attributes->l3.ipv6.hdr.src_addr,
2720 		       input->flow.udp6_flow.ip.src_ip,
2721 		       RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2722 		memcpy(attributes->l3.ipv6.hdr.dst_addr,
2723 		       input->flow.udp6_flow.ip.dst_ip,
2724 		       RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2725 		attributes->l4.udp.hdr = (struct udp_hdr){
2726 			.src_port = input->flow.udp6_flow.src_port,
2727 			.dst_port = input->flow.udp6_flow.dst_port,
2728 		};
2729 		attributes->items[1] = (struct rte_flow_item){
2730 			.type = RTE_FLOW_ITEM_TYPE_IPV6,
2731 			.spec = &attributes->l3,
2732 		};
2733 		attributes->items[2] = (struct rte_flow_item){
2734 			.type = RTE_FLOW_ITEM_TYPE_UDP,
2735 			.spec = &attributes->l4,
2736 		};
2737 		break;
2738 	case RTE_ETH_FLOW_NONFRAG_IPV6_TCP:
2739 		attributes->l3.ipv6.hdr = (struct ipv6_hdr){
2740 			.hop_limits = input->flow.tcp6_flow.ip.hop_limits,
2741 			.proto = input->flow.tcp6_flow.ip.proto,
2742 		};
2743 		memcpy(attributes->l3.ipv6.hdr.src_addr,
2744 		       input->flow.tcp6_flow.ip.src_ip,
2745 		       RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2746 		memcpy(attributes->l3.ipv6.hdr.dst_addr,
2747 		       input->flow.tcp6_flow.ip.dst_ip,
2748 		       RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2749 		attributes->l4.tcp.hdr = (struct tcp_hdr){
2750 			.src_port = input->flow.tcp6_flow.src_port,
2751 			.dst_port = input->flow.tcp6_flow.dst_port,
2752 		};
2753 		attributes->items[1] = (struct rte_flow_item){
2754 			.type = RTE_FLOW_ITEM_TYPE_IPV6,
2755 			.spec = &attributes->l3,
2756 		};
2757 		attributes->items[2] = (struct rte_flow_item){
2758 			.type = RTE_FLOW_ITEM_TYPE_TCP,
2759 			.spec = &attributes->l4,
2760 		};
2761 		break;
2762 	case RTE_ETH_FLOW_NONFRAG_IPV6_OTHER:
2763 		attributes->l3.ipv6.hdr = (struct ipv6_hdr){
2764 			.hop_limits = input->flow.ipv6_flow.hop_limits,
2765 			.proto = input->flow.ipv6_flow.proto,
2766 		};
2767 		memcpy(attributes->l3.ipv6.hdr.src_addr,
2768 		       input->flow.ipv6_flow.src_ip,
2769 		       RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2770 		memcpy(attributes->l3.ipv6.hdr.dst_addr,
2771 		       input->flow.ipv6_flow.dst_ip,
2772 		       RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2773 		attributes->items[1] = (struct rte_flow_item){
2774 			.type = RTE_FLOW_ITEM_TYPE_IPV6,
2775 			.spec = &attributes->l3,
2776 		};
2777 		break;
2778 	default:
2779 		ERROR("invalid flow type%d",
2780 		      fdir_filter->input.flow_type);
2781 		return ENOTSUP;
2782 	}
2783 	return 0;
2784 }
2785 
2786 /**
2787  * Add new flow director filter and store it in list.
2788  *
2789  * @param priv
2790  *   Private structure.
2791  * @param fdir_filter
2792  *   Flow director filter to add.
2793  *
2794  * @return
2795  *   0 on success, errno value on failure.
2796  */
2797 static int
2798 priv_fdir_filter_add(struct priv *priv,
2799 		     const struct rte_eth_fdir_filter *fdir_filter)
2800 {
2801 	struct mlx5_fdir attributes = {
2802 		.attr.group = 0,
2803 		.l2_mask = {
2804 			.dst.addr_bytes = "\x00\x00\x00\x00\x00\x00",
2805 			.src.addr_bytes = "\x00\x00\x00\x00\x00\x00",
2806 			.type = 0,
2807 		},
2808 	};
2809 	struct mlx5_flow_parse parser = {
2810 		.layer = HASH_RXQ_ETH,
2811 	};
2812 	struct rte_flow_error error;
2813 	struct rte_flow *flow;
2814 	int ret;
2815 
2816 	ret = priv_fdir_filter_convert(priv, fdir_filter, &attributes);
2817 	if (ret)
2818 		return -ret;
2819 	ret = priv_flow_convert(priv, &attributes.attr, attributes.items,
2820 				attributes.actions, &error, &parser);
2821 	if (ret)
2822 		return -ret;
2823 	flow = priv_flow_create(priv,
2824 				&priv->flows,
2825 				&attributes.attr,
2826 				attributes.items,
2827 				attributes.actions,
2828 				&error);
2829 	if (flow) {
2830 		DEBUG("FDIR created %p", (void *)flow);
2831 		return 0;
2832 	}
2833 	return ENOTSUP;
2834 }
2835 
2836 /**
2837  * Delete specific filter.
2838  *
2839  * @param priv
2840  *   Private structure.
2841  * @param fdir_filter
2842  *   Filter to be deleted.
2843  *
2844  * @return
2845  *   0 on success, errno value on failure.
2846  */
2847 static int
2848 priv_fdir_filter_delete(struct priv *priv,
2849 			const struct rte_eth_fdir_filter *fdir_filter)
2850 {
2851 	struct mlx5_fdir attributes = {
2852 		.attr.group = 0,
2853 	};
2854 	struct mlx5_flow_parse parser = {
2855 		.create = 1,
2856 		.layer = HASH_RXQ_ETH,
2857 	};
2858 	struct rte_flow_error error;
2859 	struct rte_flow *flow;
2860 	unsigned int i;
2861 	int ret;
2862 
2863 	ret = priv_fdir_filter_convert(priv, fdir_filter, &attributes);
2864 	if (ret)
2865 		return -ret;
2866 	ret = priv_flow_convert(priv, &attributes.attr, attributes.items,
2867 				attributes.actions, &error, &parser);
2868 	if (ret)
2869 		goto exit;
2870 	/*
2871 	 * Special case for drop action which is only set in the
2872 	 * specifications when the flow is created.  In this situation the
2873 	 * drop specification is missing.
2874 	 */
2875 	if (parser.drop) {
2876 		struct ibv_flow_spec_action_drop *drop;
2877 
2878 		drop = (void *)((uintptr_t)parser.drop_q.ibv_attr +
2879 				parser.drop_q.offset);
2880 		*drop = (struct ibv_flow_spec_action_drop){
2881 			.type = IBV_FLOW_SPEC_ACTION_DROP,
2882 			.size = sizeof(struct ibv_flow_spec_action_drop),
2883 		};
2884 		parser.drop_q.ibv_attr->num_of_specs++;
2885 	}
2886 	TAILQ_FOREACH(flow, &priv->flows, next) {
2887 		struct ibv_flow_attr *attr;
2888 		struct ibv_spec_header *attr_h;
2889 		void *spec;
2890 		struct ibv_flow_attr *flow_attr;
2891 		struct ibv_spec_header *flow_h;
2892 		void *flow_spec;
2893 		unsigned int specs_n;
2894 
2895 		if (parser.drop)
2896 			attr = parser.drop_q.ibv_attr;
2897 		else
2898 			attr = parser.queue[HASH_RXQ_ETH].ibv_attr;
2899 		if (flow->drop)
2900 			flow_attr = flow->drxq.ibv_attr;
2901 		else
2902 			flow_attr = flow->frxq[HASH_RXQ_ETH].ibv_attr;
2903 		/* Compare first the attributes. */
2904 		if (memcmp(attr, flow_attr, sizeof(struct ibv_flow_attr)))
2905 			continue;
2906 		if (attr->num_of_specs == 0)
2907 			continue;
2908 		spec = (void *)((uintptr_t)attr +
2909 				sizeof(struct ibv_flow_attr));
2910 		flow_spec = (void *)((uintptr_t)flow_attr +
2911 				     sizeof(struct ibv_flow_attr));
2912 		specs_n = RTE_MIN(attr->num_of_specs, flow_attr->num_of_specs);
2913 		for (i = 0; i != specs_n; ++i) {
2914 			attr_h = spec;
2915 			flow_h = flow_spec;
2916 			if (memcmp(spec, flow_spec,
2917 				   RTE_MIN(attr_h->size, flow_h->size)))
2918 				goto wrong_flow;
2919 			spec = (void *)((uintptr_t)spec + attr_h->size);
2920 			flow_spec = (void *)((uintptr_t)flow_spec +
2921 					     flow_h->size);
2922 		}
2923 		/* At this point, the flow match. */
2924 		break;
2925 wrong_flow:
2926 		/* The flow does not match. */
2927 		continue;
2928 	}
2929 	if (flow)
2930 		priv_flow_destroy(priv, &priv->flows, flow);
2931 exit:
2932 	if (parser.drop) {
2933 		rte_free(parser.drop_q.ibv_attr);
2934 	} else {
2935 		for (i = 0; i != hash_rxq_init_n; ++i) {
2936 			if (parser.queue[i].ibv_attr)
2937 				rte_free(parser.queue[i].ibv_attr);
2938 		}
2939 	}
2940 	return -ret;
2941 }
2942 
2943 /**
2944  * Update queue for specific filter.
2945  *
2946  * @param priv
2947  *   Private structure.
2948  * @param fdir_filter
2949  *   Filter to be updated.
2950  *
2951  * @return
2952  *   0 on success, errno value on failure.
2953  */
2954 static int
2955 priv_fdir_filter_update(struct priv *priv,
2956 			const struct rte_eth_fdir_filter *fdir_filter)
2957 {
2958 	int ret;
2959 
2960 	ret = priv_fdir_filter_delete(priv, fdir_filter);
2961 	if (ret)
2962 		return ret;
2963 	ret = priv_fdir_filter_add(priv, fdir_filter);
2964 	return ret;
2965 }
2966 
2967 /**
2968  * Flush all filters.
2969  *
2970  * @param priv
2971  *   Private structure.
2972  */
2973 static void
2974 priv_fdir_filter_flush(struct priv *priv)
2975 {
2976 	priv_flow_flush(priv, &priv->flows);
2977 }
2978 
2979 /**
2980  * Get flow director information.
2981  *
2982  * @param priv
2983  *   Private structure.
2984  * @param[out] fdir_info
2985  *   Resulting flow director information.
2986  */
2987 static void
2988 priv_fdir_info_get(struct priv *priv, struct rte_eth_fdir_info *fdir_info)
2989 {
2990 	struct rte_eth_fdir_masks *mask =
2991 		&priv->dev->data->dev_conf.fdir_conf.mask;
2992 
2993 	fdir_info->mode = priv->dev->data->dev_conf.fdir_conf.mode;
2994 	fdir_info->guarant_spc = 0;
2995 	rte_memcpy(&fdir_info->mask, mask, sizeof(fdir_info->mask));
2996 	fdir_info->max_flexpayload = 0;
2997 	fdir_info->flow_types_mask[0] = 0;
2998 	fdir_info->flex_payload_unit = 0;
2999 	fdir_info->max_flex_payload_segment_num = 0;
3000 	fdir_info->flex_payload_limit = 0;
3001 	memset(&fdir_info->flex_conf, 0, sizeof(fdir_info->flex_conf));
3002 }
3003 
3004 /**
3005  * Deal with flow director operations.
3006  *
3007  * @param priv
3008  *   Pointer to private structure.
3009  * @param filter_op
3010  *   Operation to perform.
3011  * @param arg
3012  *   Pointer to operation-specific structure.
3013  *
3014  * @return
3015  *   0 on success, errno value on failure.
3016  */
3017 static int
3018 priv_fdir_ctrl_func(struct priv *priv, enum rte_filter_op filter_op, void *arg)
3019 {
3020 	enum rte_fdir_mode fdir_mode =
3021 		priv->dev->data->dev_conf.fdir_conf.mode;
3022 	int ret = 0;
3023 
3024 	if (filter_op == RTE_ETH_FILTER_NOP)
3025 		return 0;
3026 	if (fdir_mode != RTE_FDIR_MODE_PERFECT &&
3027 	    fdir_mode != RTE_FDIR_MODE_PERFECT_MAC_VLAN) {
3028 		ERROR("%p: flow director mode %d not supported",
3029 		      (void *)priv, fdir_mode);
3030 		return EINVAL;
3031 	}
3032 	switch (filter_op) {
3033 	case RTE_ETH_FILTER_ADD:
3034 		ret = priv_fdir_filter_add(priv, arg);
3035 		break;
3036 	case RTE_ETH_FILTER_UPDATE:
3037 		ret = priv_fdir_filter_update(priv, arg);
3038 		break;
3039 	case RTE_ETH_FILTER_DELETE:
3040 		ret = priv_fdir_filter_delete(priv, arg);
3041 		break;
3042 	case RTE_ETH_FILTER_FLUSH:
3043 		priv_fdir_filter_flush(priv);
3044 		break;
3045 	case RTE_ETH_FILTER_INFO:
3046 		priv_fdir_info_get(priv, arg);
3047 		break;
3048 	default:
3049 		DEBUG("%p: unknown operation %u", (void *)priv,
3050 		      filter_op);
3051 		ret = EINVAL;
3052 		break;
3053 	}
3054 	return ret;
3055 }
3056 
3057 /**
3058  * Manage filter operations.
3059  *
3060  * @param dev
3061  *   Pointer to Ethernet device structure.
3062  * @param filter_type
3063  *   Filter type.
3064  * @param filter_op
3065  *   Operation to perform.
3066  * @param arg
3067  *   Pointer to operation-specific structure.
3068  *
3069  * @return
3070  *   0 on success, negative errno value on failure.
3071  */
3072 int
3073 mlx5_dev_filter_ctrl(struct rte_eth_dev *dev,
3074 		     enum rte_filter_type filter_type,
3075 		     enum rte_filter_op filter_op,
3076 		     void *arg)
3077 {
3078 	int ret = EINVAL;
3079 	struct priv *priv = dev->data->dev_private;
3080 
3081 	switch (filter_type) {
3082 	case RTE_ETH_FILTER_GENERIC:
3083 		if (filter_op != RTE_ETH_FILTER_GET)
3084 			return -EINVAL;
3085 		*(const void **)arg = &mlx5_flow_ops;
3086 		return 0;
3087 	case RTE_ETH_FILTER_FDIR:
3088 		priv_lock(priv);
3089 		ret = priv_fdir_ctrl_func(priv, filter_op, arg);
3090 		priv_unlock(priv);
3091 		break;
3092 	default:
3093 		ERROR("%p: filter type (%d) not supported",
3094 		      (void *)dev, filter_type);
3095 		break;
3096 	}
3097 	return -ret;
3098 }
3099