xref: /dpdk/drivers/net/mlx5/mlx5_flow.c (revision a4fd5eeeadc0de3961956d44332b0e330ea33544)
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 	union {
483 		struct rte_flow_item_ipv4 ipv4;
484 		struct rte_flow_item_ipv6 ipv6;
485 	} l3;
486 	union {
487 		struct rte_flow_item_udp udp;
488 		struct rte_flow_item_tcp tcp;
489 	} l4;
490 	struct rte_flow_action_queue queue;
491 };
492 
493 /* Verbs specification header. */
494 struct ibv_spec_header {
495 	enum ibv_flow_spec_type type;
496 	uint16_t size;
497 };
498 
499 /**
500  * Check support for a given item.
501  *
502  * @param item[in]
503  *   Item specification.
504  * @param mask[in]
505  *   Bit-masks covering supported fields to compare with spec, last and mask in
506  *   \item.
507  * @param size
508  *   Bit-Mask size in bytes.
509  *
510  * @return
511  *   0 on success.
512  */
513 static int
514 mlx5_flow_item_validate(const struct rte_flow_item *item,
515 			const uint8_t *mask, unsigned int size)
516 {
517 	int ret = 0;
518 
519 	if (!item->spec && (item->mask || item->last))
520 		return -1;
521 	if (item->spec && !item->mask) {
522 		unsigned int i;
523 		const uint8_t *spec = item->spec;
524 
525 		for (i = 0; i < size; ++i)
526 			if ((spec[i] | mask[i]) != mask[i])
527 				return -1;
528 	}
529 	if (item->last && !item->mask) {
530 		unsigned int i;
531 		const uint8_t *spec = item->last;
532 
533 		for (i = 0; i < size; ++i)
534 			if ((spec[i] | mask[i]) != mask[i])
535 				return -1;
536 	}
537 	if (item->mask) {
538 		unsigned int i;
539 		const uint8_t *spec = item->mask;
540 
541 		for (i = 0; i < size; ++i)
542 			if ((spec[i] | mask[i]) != mask[i])
543 				return -1;
544 	}
545 	if (item->spec && item->last) {
546 		uint8_t spec[size];
547 		uint8_t last[size];
548 		const uint8_t *apply = mask;
549 		unsigned int i;
550 
551 		if (item->mask)
552 			apply = item->mask;
553 		for (i = 0; i < size; ++i) {
554 			spec[i] = ((const uint8_t *)item->spec)[i] & apply[i];
555 			last[i] = ((const uint8_t *)item->last)[i] & apply[i];
556 		}
557 		ret = memcmp(spec, last, size);
558 	}
559 	return ret;
560 }
561 
562 /**
563  * Copy the RSS configuration from the user ones.
564  *
565  * @param priv
566  *   Pointer to private structure.
567  * @param parser
568  *   Internal parser structure.
569  * @param rss_conf
570  *   User RSS configuration to save.
571  *
572  * @return
573  *   0 on success, errno value on failure.
574  */
575 static int
576 priv_flow_convert_rss_conf(struct priv *priv,
577 			   struct mlx5_flow_parse *parser,
578 			   const struct rte_eth_rss_conf *rss_conf)
579 {
580 	const struct rte_eth_rss_conf *rss =
581 		rss_conf ? rss_conf : &priv->rss_conf;
582 
583 	if (rss->rss_key_len > 40)
584 		return EINVAL;
585 	parser->rss_conf.rss_key_len = rss->rss_key_len;
586 	parser->rss_conf.rss_hf = rss->rss_hf;
587 	memcpy(parser->rss_key, rss->rss_key, rss->rss_key_len);
588 	parser->rss_conf.rss_key = parser->rss_key;
589 	return 0;
590 }
591 
592 /**
593  * Extract attribute to the parser.
594  *
595  * @param priv
596  *   Pointer to private structure.
597  * @param[in] attr
598  *   Flow rule attributes.
599  * @param[out] error
600  *   Perform verbose error reporting if not NULL.
601  * @param[in, out] parser
602  *   Internal parser structure.
603  *
604  * @return
605  *   0 on success, a negative errno value otherwise and rte_errno is set.
606  */
607 static int
608 priv_flow_convert_attributes(struct priv *priv,
609 			     const struct rte_flow_attr *attr,
610 			     struct rte_flow_error *error,
611 			     struct mlx5_flow_parse *parser)
612 {
613 	(void)priv;
614 	(void)parser;
615 	if (attr->group) {
616 		rte_flow_error_set(error, ENOTSUP,
617 				   RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
618 				   NULL,
619 				   "groups are not supported");
620 		return -rte_errno;
621 	}
622 	if (attr->priority && attr->priority != MLX5_CTRL_FLOW_PRIORITY) {
623 		rte_flow_error_set(error, ENOTSUP,
624 				   RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
625 				   NULL,
626 				   "priorities are not supported");
627 		return -rte_errno;
628 	}
629 	if (attr->egress) {
630 		rte_flow_error_set(error, ENOTSUP,
631 				   RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
632 				   NULL,
633 				   "egress is not supported");
634 		return -rte_errno;
635 	}
636 	if (!attr->ingress) {
637 		rte_flow_error_set(error, ENOTSUP,
638 				   RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
639 				   NULL,
640 				   "only ingress is supported");
641 		return -rte_errno;
642 	}
643 	return 0;
644 }
645 
646 /**
647  * Extract actions request to the parser.
648  *
649  * @param priv
650  *   Pointer to private structure.
651  * @param[in] actions
652  *   Associated actions (list terminated by the END action).
653  * @param[out] error
654  *   Perform verbose error reporting if not NULL.
655  * @param[in, out] parser
656  *   Internal parser structure.
657  *
658  * @return
659  *   0 on success, a negative errno value otherwise and rte_errno is set.
660  */
661 static int
662 priv_flow_convert_actions(struct priv *priv,
663 			  const struct rte_flow_action actions[],
664 			  struct rte_flow_error *error,
665 			  struct mlx5_flow_parse *parser)
666 {
667 	/*
668 	 * Add default RSS configuration necessary for Verbs to create QP even
669 	 * if no RSS is necessary.
670 	 */
671 	priv_flow_convert_rss_conf(priv, parser,
672 				   (const struct rte_eth_rss_conf *)
673 				   &priv->rss_conf);
674 	for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
675 		if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
676 			continue;
677 		} else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
678 			parser->drop = 1;
679 		} else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
680 			const struct rte_flow_action_queue *queue =
681 				(const struct rte_flow_action_queue *)
682 				actions->conf;
683 			uint16_t n;
684 			uint16_t found = 0;
685 
686 			if (!queue || (queue->index > (priv->rxqs_n - 1)))
687 				goto exit_action_not_supported;
688 			for (n = 0; n < parser->queues_n; ++n) {
689 				if (parser->queues[n] == queue->index) {
690 					found = 1;
691 					break;
692 				}
693 			}
694 			if (parser->queues_n > 1 && !found) {
695 				rte_flow_error_set(error, ENOTSUP,
696 					   RTE_FLOW_ERROR_TYPE_ACTION,
697 					   actions,
698 					   "queue action not in RSS queues");
699 				return -rte_errno;
700 			}
701 			if (!found) {
702 				parser->queues_n = 1;
703 				parser->queues[0] = queue->index;
704 			}
705 		} else if (actions->type == RTE_FLOW_ACTION_TYPE_RSS) {
706 			const struct rte_flow_action_rss *rss =
707 				(const struct rte_flow_action_rss *)
708 				actions->conf;
709 			uint16_t n;
710 
711 			if (!rss || !rss->num) {
712 				rte_flow_error_set(error, EINVAL,
713 						   RTE_FLOW_ERROR_TYPE_ACTION,
714 						   actions,
715 						   "no valid queues");
716 				return -rte_errno;
717 			}
718 			if (parser->queues_n == 1) {
719 				uint16_t found = 0;
720 
721 				assert(parser->queues_n);
722 				for (n = 0; n < rss->num; ++n) {
723 					if (parser->queues[0] ==
724 					    rss->queue[n]) {
725 						found = 1;
726 						break;
727 					}
728 				}
729 				if (!found) {
730 					rte_flow_error_set(error, ENOTSUP,
731 						   RTE_FLOW_ERROR_TYPE_ACTION,
732 						   actions,
733 						   "queue action not in RSS"
734 						   " queues");
735 					return -rte_errno;
736 				}
737 			}
738 			for (n = 0; n < rss->num; ++n) {
739 				if (rss->queue[n] >= priv->rxqs_n) {
740 					rte_flow_error_set(error, EINVAL,
741 						   RTE_FLOW_ERROR_TYPE_ACTION,
742 						   actions,
743 						   "queue id > number of"
744 						   " queues");
745 					return -rte_errno;
746 				}
747 			}
748 			for (n = 0; n < rss->num; ++n)
749 				parser->queues[n] = rss->queue[n];
750 			parser->queues_n = rss->num;
751 			if (priv_flow_convert_rss_conf(priv, parser,
752 						       rss->rss_conf)) {
753 				rte_flow_error_set(error, EINVAL,
754 						   RTE_FLOW_ERROR_TYPE_ACTION,
755 						   actions,
756 						   "wrong RSS configuration");
757 				return -rte_errno;
758 			}
759 		} else if (actions->type == RTE_FLOW_ACTION_TYPE_MARK) {
760 			const struct rte_flow_action_mark *mark =
761 				(const struct rte_flow_action_mark *)
762 				actions->conf;
763 
764 			if (!mark) {
765 				rte_flow_error_set(error, EINVAL,
766 						   RTE_FLOW_ERROR_TYPE_ACTION,
767 						   actions,
768 						   "mark must be defined");
769 				return -rte_errno;
770 			} else if (mark->id >= MLX5_FLOW_MARK_MAX) {
771 				rte_flow_error_set(error, ENOTSUP,
772 						   RTE_FLOW_ERROR_TYPE_ACTION,
773 						   actions,
774 						   "mark must be between 0"
775 						   " and 16777199");
776 				return -rte_errno;
777 			}
778 			parser->mark = 1;
779 			parser->mark_id = mark->id;
780 		} else if (actions->type == RTE_FLOW_ACTION_TYPE_FLAG) {
781 			parser->mark = 1;
782 		} else if (actions->type == RTE_FLOW_ACTION_TYPE_COUNT &&
783 			   priv->counter_set_supported) {
784 			parser->count = 1;
785 		} else {
786 			goto exit_action_not_supported;
787 		}
788 	}
789 	if (parser->drop && parser->mark)
790 		parser->mark = 0;
791 	if (!parser->queues_n && !parser->drop) {
792 		rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE,
793 				   NULL, "no valid action");
794 		return -rte_errno;
795 	}
796 	return 0;
797 exit_action_not_supported:
798 	rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
799 			   actions, "action not supported");
800 	return -rte_errno;
801 }
802 
803 /**
804  * Validate items.
805  *
806  * @param priv
807  *   Pointer to private structure.
808  * @param[in] items
809  *   Pattern specification (list terminated by the END pattern item).
810  * @param[out] error
811  *   Perform verbose error reporting if not NULL.
812  * @param[in, out] parser
813  *   Internal parser structure.
814  *
815  * @return
816  *   0 on success, a negative errno value otherwise and rte_errno is set.
817  */
818 static int
819 priv_flow_convert_items_validate(struct priv *priv,
820 				 const struct rte_flow_item items[],
821 				 struct rte_flow_error *error,
822 				 struct mlx5_flow_parse *parser)
823 {
824 	const struct mlx5_flow_items *cur_item = mlx5_flow_items;
825 	unsigned int i;
826 
827 	(void)priv;
828 	/* Initialise the offsets to start after verbs attribute. */
829 	if (parser->drop) {
830 		parser->drop_q.offset = sizeof(struct ibv_flow_attr);
831 	} else {
832 		for (i = 0; i != hash_rxq_init_n; ++i)
833 			parser->queue[i].offset = sizeof(struct ibv_flow_attr);
834 	}
835 	for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
836 		const struct mlx5_flow_items *token = NULL;
837 		unsigned int n;
838 		int err;
839 
840 		if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
841 			continue;
842 		for (i = 0;
843 		     cur_item->items &&
844 		     cur_item->items[i] != RTE_FLOW_ITEM_TYPE_END;
845 		     ++i) {
846 			if (cur_item->items[i] == items->type) {
847 				token = &mlx5_flow_items[items->type];
848 				break;
849 			}
850 		}
851 		if (!token)
852 			goto exit_item_not_supported;
853 		cur_item = token;
854 		err = mlx5_flow_item_validate(items,
855 					      (const uint8_t *)cur_item->mask,
856 					      cur_item->mask_sz);
857 		if (err)
858 			goto exit_item_not_supported;
859 		if (items->type == RTE_FLOW_ITEM_TYPE_VXLAN) {
860 			if (parser->inner) {
861 				rte_flow_error_set(error, ENOTSUP,
862 						   RTE_FLOW_ERROR_TYPE_ITEM,
863 						   items,
864 						   "cannot recognize multiple"
865 						   " VXLAN encapsulations");
866 				return -rte_errno;
867 			}
868 			parser->inner = 1;
869 		}
870 		if (parser->drop) {
871 			parser->drop_q.offset += cur_item->dst_sz;
872 		} else if (parser->queues_n == 1) {
873 			parser->queue[HASH_RXQ_ETH].offset += cur_item->dst_sz;
874 		} else {
875 			for (n = 0; n != hash_rxq_init_n; ++n)
876 				parser->queue[n].offset += cur_item->dst_sz;
877 		}
878 	}
879 	if (parser->mark) {
880 		for (i = 0; i != hash_rxq_init_n; ++i)
881 			parser->queue[i].offset +=
882 				sizeof(struct ibv_flow_spec_action_tag);
883 	}
884 	if (parser->count) {
885 		unsigned int size = sizeof(struct ibv_flow_spec_counter_action);
886 
887 		if (parser->drop) {
888 			parser->drop_q.offset += size;
889 		} else {
890 			for (i = 0; i != hash_rxq_init_n; ++i)
891 				parser->queue[i].offset += size;
892 		}
893 	}
894 	return 0;
895 exit_item_not_supported:
896 	rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
897 			   items, "item not supported");
898 	return -rte_errno;
899 }
900 
901 /**
902  * Allocate memory space to store verbs flow attributes.
903  *
904  * @param priv
905  *   Pointer to private structure.
906  * @param[in] priority
907  *   Flow priority.
908  * @param[in] size
909  *   Amount of byte to allocate.
910  * @param[out] error
911  *   Perform verbose error reporting if not NULL.
912  *
913  * @return
914  *   A verbs flow attribute on success, NULL otherwise.
915  */
916 static struct ibv_flow_attr*
917 priv_flow_convert_allocate(struct priv *priv,
918 			   unsigned int priority,
919 			   unsigned int size,
920 			   struct rte_flow_error *error)
921 {
922 	struct ibv_flow_attr *ibv_attr;
923 
924 	(void)priv;
925 	ibv_attr = rte_calloc(__func__, 1, size, 0);
926 	if (!ibv_attr) {
927 		rte_flow_error_set(error, ENOMEM,
928 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
929 				   NULL,
930 				   "cannot allocate verbs spec attributes.");
931 		return NULL;
932 	}
933 	ibv_attr->priority = priority;
934 	return ibv_attr;
935 }
936 
937 /**
938  * Finalise verbs flow attributes.
939  *
940  * @param priv
941  *   Pointer to private structure.
942  * @param[in, out] parser
943  *   Internal parser structure.
944  */
945 static void
946 priv_flow_convert_finalise(struct priv *priv, struct mlx5_flow_parse *parser)
947 {
948 	const unsigned int ipv4 =
949 		hash_rxq_init[parser->layer].ip_version == MLX5_IPV4;
950 	const enum hash_rxq_type hmin = ipv4 ? HASH_RXQ_TCPV4 : HASH_RXQ_TCPV6;
951 	const enum hash_rxq_type hmax = ipv4 ? HASH_RXQ_IPV4 : HASH_RXQ_IPV6;
952 	const enum hash_rxq_type ohmin = ipv4 ? HASH_RXQ_TCPV6 : HASH_RXQ_TCPV4;
953 	const enum hash_rxq_type ohmax = ipv4 ? HASH_RXQ_IPV6 : HASH_RXQ_IPV4;
954 	const enum hash_rxq_type ip = ipv4 ? HASH_RXQ_IPV4 : HASH_RXQ_IPV6;
955 	unsigned int i;
956 
957 	(void)priv;
958 	if (parser->layer == HASH_RXQ_ETH) {
959 		goto fill;
960 	} else {
961 		/*
962 		 * This layer becomes useless as the pattern define under
963 		 * layers.
964 		 */
965 		rte_free(parser->queue[HASH_RXQ_ETH].ibv_attr);
966 		parser->queue[HASH_RXQ_ETH].ibv_attr = NULL;
967 	}
968 	/* Remove opposite kind of layer e.g. IPv6 if the pattern is IPv4. */
969 	for (i = ohmin; i != (ohmax + 1); ++i) {
970 		if (!parser->queue[i].ibv_attr)
971 			continue;
972 		rte_free(parser->queue[i].ibv_attr);
973 		parser->queue[i].ibv_attr = NULL;
974 	}
975 	/* Remove impossible flow according to the RSS configuration. */
976 	if (hash_rxq_init[parser->layer].dpdk_rss_hf &
977 	    parser->rss_conf.rss_hf) {
978 		/* Remove any other flow. */
979 		for (i = hmin; i != (hmax + 1); ++i) {
980 			if ((i == parser->layer) ||
981 			     (!parser->queue[i].ibv_attr))
982 				continue;
983 			rte_free(parser->queue[i].ibv_attr);
984 			parser->queue[i].ibv_attr = NULL;
985 		}
986 	} else  if (!parser->queue[ip].ibv_attr) {
987 		/* no RSS possible with the current configuration. */
988 		parser->queues_n = 1;
989 		return;
990 	}
991 fill:
992 	/*
993 	 * Fill missing layers in verbs specifications, or compute the correct
994 	 * offset to allocate the memory space for the attributes and
995 	 * specifications.
996 	 */
997 	for (i = 0; i != hash_rxq_init_n - 1; ++i) {
998 		union {
999 			struct ibv_flow_spec_ipv4_ext ipv4;
1000 			struct ibv_flow_spec_ipv6 ipv6;
1001 			struct ibv_flow_spec_tcp_udp udp_tcp;
1002 		} specs;
1003 		void *dst;
1004 		uint16_t size;
1005 
1006 		if (i == parser->layer)
1007 			continue;
1008 		if (parser->layer == HASH_RXQ_ETH) {
1009 			if (hash_rxq_init[i].ip_version == MLX5_IPV4) {
1010 				size = sizeof(struct ibv_flow_spec_ipv4_ext);
1011 				specs.ipv4 = (struct ibv_flow_spec_ipv4_ext){
1012 					.type = IBV_FLOW_SPEC_IPV4_EXT |
1013 						parser->inner,
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 						parser->inner,
1021 					.size = size,
1022 				};
1023 			}
1024 			if (parser->queue[i].ibv_attr) {
1025 				dst = (void *)((uintptr_t)
1026 					       parser->queue[i].ibv_attr +
1027 					       parser->queue[i].offset);
1028 				memcpy(dst, &specs, size);
1029 				++parser->queue[i].ibv_attr->num_of_specs;
1030 			}
1031 			parser->queue[i].offset += size;
1032 		}
1033 		if ((i == HASH_RXQ_UDPV4) || (i == HASH_RXQ_TCPV4) ||
1034 		    (i == HASH_RXQ_UDPV6) || (i == HASH_RXQ_TCPV6)) {
1035 			size = sizeof(struct ibv_flow_spec_tcp_udp);
1036 			specs.udp_tcp = (struct ibv_flow_spec_tcp_udp) {
1037 				.type = ((i == HASH_RXQ_UDPV4 ||
1038 					  i == HASH_RXQ_UDPV6) ?
1039 					 IBV_FLOW_SPEC_UDP :
1040 					 IBV_FLOW_SPEC_TCP) |
1041 					parser->inner,
1042 				.size = size,
1043 			};
1044 			if (parser->queue[i].ibv_attr) {
1045 				dst = (void *)((uintptr_t)
1046 					       parser->queue[i].ibv_attr +
1047 					       parser->queue[i].offset);
1048 				memcpy(dst, &specs, size);
1049 				++parser->queue[i].ibv_attr->num_of_specs;
1050 			}
1051 			parser->queue[i].offset += size;
1052 		}
1053 	}
1054 }
1055 
1056 /**
1057  * Validate and convert a flow supported by the NIC.
1058  *
1059  * @param priv
1060  *   Pointer to private structure.
1061  * @param[in] attr
1062  *   Flow rule attributes.
1063  * @param[in] pattern
1064  *   Pattern specification (list terminated by the END pattern item).
1065  * @param[in] actions
1066  *   Associated actions (list terminated by the END action).
1067  * @param[out] error
1068  *   Perform verbose error reporting if not NULL.
1069  * @param[in, out] parser
1070  *   Internal parser structure.
1071  *
1072  * @return
1073  *   0 on success, a negative errno value otherwise and rte_errno is set.
1074  */
1075 static int
1076 priv_flow_convert(struct priv *priv,
1077 		  const struct rte_flow_attr *attr,
1078 		  const struct rte_flow_item items[],
1079 		  const struct rte_flow_action actions[],
1080 		  struct rte_flow_error *error,
1081 		  struct mlx5_flow_parse *parser)
1082 {
1083 	const struct mlx5_flow_items *cur_item = mlx5_flow_items;
1084 	unsigned int i;
1085 	int ret;
1086 
1087 	/* First step. Validate the attributes, items and actions. */
1088 	*parser = (struct mlx5_flow_parse){
1089 		.create = parser->create,
1090 		.layer = HASH_RXQ_ETH,
1091 		.mark_id = MLX5_FLOW_MARK_DEFAULT,
1092 	};
1093 	ret = priv_flow_convert_attributes(priv, attr, error, parser);
1094 	if (ret)
1095 		return ret;
1096 	ret = priv_flow_convert_actions(priv, actions, error, parser);
1097 	if (ret)
1098 		return ret;
1099 	ret = priv_flow_convert_items_validate(priv, items, error, parser);
1100 	if (ret)
1101 		return ret;
1102 	priv_flow_convert_finalise(priv, parser);
1103 	/*
1104 	 * Second step.
1105 	 * Allocate the memory space to store verbs specifications.
1106 	 */
1107 	if (parser->drop) {
1108 		parser->drop_q.ibv_attr =
1109 			priv_flow_convert_allocate(priv, attr->priority,
1110 						   parser->drop_q.offset,
1111 						   error);
1112 		if (!parser->drop_q.ibv_attr)
1113 			return ENOMEM;
1114 		parser->drop_q.offset = sizeof(struct ibv_flow_attr);
1115 	} else if (parser->queues_n == 1) {
1116 		unsigned int priority =
1117 			attr->priority +
1118 			hash_rxq_init[HASH_RXQ_ETH].flow_priority;
1119 		unsigned int offset = parser->queue[HASH_RXQ_ETH].offset;
1120 
1121 		parser->queue[HASH_RXQ_ETH].ibv_attr =
1122 			priv_flow_convert_allocate(priv, priority,
1123 						   offset, error);
1124 		if (!parser->queue[HASH_RXQ_ETH].ibv_attr)
1125 			return ENOMEM;
1126 		parser->queue[HASH_RXQ_ETH].offset =
1127 			sizeof(struct ibv_flow_attr);
1128 	} else {
1129 		for (i = 0; i != hash_rxq_init_n; ++i) {
1130 			unsigned int priority =
1131 				attr->priority +
1132 				hash_rxq_init[i].flow_priority;
1133 			unsigned int offset;
1134 
1135 			if (!(parser->rss_conf.rss_hf &
1136 			      hash_rxq_init[i].dpdk_rss_hf) &&
1137 			    (i != HASH_RXQ_ETH))
1138 				continue;
1139 			offset = parser->queue[i].offset;
1140 			parser->queue[i].ibv_attr =
1141 				priv_flow_convert_allocate(priv, priority,
1142 							   offset, error);
1143 			if (!parser->queue[i].ibv_attr)
1144 				goto exit_enomem;
1145 			parser->queue[i].offset = sizeof(struct ibv_flow_attr);
1146 		}
1147 	}
1148 	/* Third step. Conversion parse, fill the specifications. */
1149 	parser->inner = 0;
1150 	for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
1151 		if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
1152 			continue;
1153 		cur_item = &mlx5_flow_items[items->type];
1154 		ret = cur_item->convert(items,
1155 					(cur_item->default_mask ?
1156 					 cur_item->default_mask :
1157 					 cur_item->mask),
1158 					parser);
1159 		if (ret) {
1160 			rte_flow_error_set(error, ENOTSUP,
1161 					   RTE_FLOW_ERROR_TYPE_ITEM,
1162 					   items, "item not supported");
1163 			goto exit_free;
1164 		}
1165 	}
1166 	if (parser->mark)
1167 		mlx5_flow_create_flag_mark(parser, parser->mark_id);
1168 	if (parser->count && parser->create) {
1169 		mlx5_flow_create_count(priv, parser);
1170 		if (!parser->cs)
1171 			goto exit_count_error;
1172 	}
1173 	/*
1174 	 * Last step. Complete missing specification to reach the RSS
1175 	 * configuration.
1176 	 */
1177 	if (parser->queues_n > 1) {
1178 		priv_flow_convert_finalise(priv, parser);
1179 	} else if (!parser->drop) {
1180 		/*
1181 		 * Action queue have their priority overridden with
1182 		 * Ethernet priority, this priority needs to be adjusted to
1183 		 * their most specific layer priority.
1184 		 */
1185 		parser->queue[HASH_RXQ_ETH].ibv_attr->priority =
1186 			attr->priority +
1187 			hash_rxq_init[parser->layer].flow_priority;
1188 	}
1189 exit_free:
1190 	/* Only verification is expected, all resources should be released. */
1191 	if (!parser->create) {
1192 		if (parser->drop) {
1193 			rte_free(parser->drop_q.ibv_attr);
1194 			parser->drop_q.ibv_attr = NULL;
1195 		}
1196 		for (i = 0; i != hash_rxq_init_n; ++i) {
1197 			if (parser->queue[i].ibv_attr) {
1198 				rte_free(parser->queue[i].ibv_attr);
1199 				parser->queue[i].ibv_attr = NULL;
1200 			}
1201 		}
1202 	}
1203 	return ret;
1204 exit_enomem:
1205 	for (i = 0; i != hash_rxq_init_n; ++i) {
1206 		if (parser->queue[i].ibv_attr) {
1207 			rte_free(parser->queue[i].ibv_attr);
1208 			parser->queue[i].ibv_attr = NULL;
1209 		}
1210 	}
1211 	rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1212 			   NULL, "cannot allocate verbs spec attributes.");
1213 	return ret;
1214 exit_count_error:
1215 	rte_flow_error_set(error, EINVAL, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1216 			   NULL, "cannot create counter.");
1217 	return rte_errno;
1218 }
1219 
1220 /**
1221  * Copy the specification created into the flow.
1222  *
1223  * @param parser
1224  *   Internal parser structure.
1225  * @param src
1226  *   Create specification.
1227  * @param size
1228  *   Size in bytes of the specification to copy.
1229  */
1230 static void
1231 mlx5_flow_create_copy(struct mlx5_flow_parse *parser, void *src,
1232 		      unsigned int size)
1233 {
1234 	unsigned int i;
1235 	void *dst;
1236 
1237 	if (parser->drop) {
1238 		dst = (void *)((uintptr_t)parser->drop_q.ibv_attr +
1239 				parser->drop_q.offset);
1240 		memcpy(dst, src, size);
1241 		++parser->drop_q.ibv_attr->num_of_specs;
1242 		parser->drop_q.offset += size;
1243 		return;
1244 	}
1245 	for (i = 0; i != hash_rxq_init_n; ++i) {
1246 		if (!parser->queue[i].ibv_attr)
1247 			continue;
1248 		/* Specification must be the same l3 type or none. */
1249 		if (parser->layer == HASH_RXQ_ETH ||
1250 		    (hash_rxq_init[parser->layer].ip_version ==
1251 		     hash_rxq_init[i].ip_version) ||
1252 		    (hash_rxq_init[i].ip_version == 0)) {
1253 			dst = (void *)((uintptr_t)parser->queue[i].ibv_attr +
1254 					parser->queue[i].offset);
1255 			memcpy(dst, src, size);
1256 			++parser->queue[i].ibv_attr->num_of_specs;
1257 			parser->queue[i].offset += size;
1258 		}
1259 	}
1260 }
1261 
1262 /**
1263  * Convert Ethernet item to Verbs specification.
1264  *
1265  * @param item[in]
1266  *   Item specification.
1267  * @param default_mask[in]
1268  *   Default bit-masks to use when item->mask is not provided.
1269  * @param data[in, out]
1270  *   User structure.
1271  */
1272 static int
1273 mlx5_flow_create_eth(const struct rte_flow_item *item,
1274 		     const void *default_mask,
1275 		     void *data)
1276 {
1277 	const struct rte_flow_item_eth *spec = item->spec;
1278 	const struct rte_flow_item_eth *mask = item->mask;
1279 	struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1280 	const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
1281 	struct ibv_flow_spec_eth eth = {
1282 		.type = parser->inner | IBV_FLOW_SPEC_ETH,
1283 		.size = eth_size,
1284 	};
1285 
1286 	parser->layer = HASH_RXQ_ETH;
1287 	if (spec) {
1288 		unsigned int i;
1289 
1290 		if (!mask)
1291 			mask = default_mask;
1292 		memcpy(&eth.val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
1293 		memcpy(&eth.val.src_mac, spec->src.addr_bytes, ETHER_ADDR_LEN);
1294 		eth.val.ether_type = spec->type;
1295 		memcpy(&eth.mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
1296 		memcpy(&eth.mask.src_mac, mask->src.addr_bytes, ETHER_ADDR_LEN);
1297 		eth.mask.ether_type = mask->type;
1298 		/* Remove unwanted bits from values. */
1299 		for (i = 0; i < ETHER_ADDR_LEN; ++i) {
1300 			eth.val.dst_mac[i] &= eth.mask.dst_mac[i];
1301 			eth.val.src_mac[i] &= eth.mask.src_mac[i];
1302 		}
1303 		eth.val.ether_type &= eth.mask.ether_type;
1304 	}
1305 	mlx5_flow_create_copy(parser, &eth, eth_size);
1306 	return 0;
1307 }
1308 
1309 /**
1310  * Convert VLAN item to Verbs specification.
1311  *
1312  * @param item[in]
1313  *   Item specification.
1314  * @param default_mask[in]
1315  *   Default bit-masks to use when item->mask is not provided.
1316  * @param data[in, out]
1317  *   User structure.
1318  */
1319 static int
1320 mlx5_flow_create_vlan(const struct rte_flow_item *item,
1321 		      const void *default_mask,
1322 		      void *data)
1323 {
1324 	const struct rte_flow_item_vlan *spec = item->spec;
1325 	const struct rte_flow_item_vlan *mask = item->mask;
1326 	struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1327 	struct ibv_flow_spec_eth *eth;
1328 	const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
1329 
1330 	if (spec) {
1331 		unsigned int i;
1332 		if (!mask)
1333 			mask = default_mask;
1334 
1335 		if (parser->drop) {
1336 			eth = (void *)((uintptr_t)parser->drop_q.ibv_attr +
1337 				       parser->drop_q.offset - eth_size);
1338 			eth->val.vlan_tag = spec->tci;
1339 			eth->mask.vlan_tag = mask->tci;
1340 			eth->val.vlan_tag &= eth->mask.vlan_tag;
1341 			return 0;
1342 		}
1343 		for (i = 0; i != hash_rxq_init_n; ++i) {
1344 			if (!parser->queue[i].ibv_attr)
1345 				continue;
1346 
1347 			eth = (void *)((uintptr_t)parser->queue[i].ibv_attr +
1348 				       parser->queue[i].offset - eth_size);
1349 			eth->val.vlan_tag = spec->tci;
1350 			eth->mask.vlan_tag = mask->tci;
1351 			eth->val.vlan_tag &= eth->mask.vlan_tag;
1352 		}
1353 	}
1354 	return 0;
1355 }
1356 
1357 /**
1358  * Convert IPv4 item to Verbs specification.
1359  *
1360  * @param item[in]
1361  *   Item specification.
1362  * @param default_mask[in]
1363  *   Default bit-masks to use when item->mask is not provided.
1364  * @param data[in, out]
1365  *   User structure.
1366  */
1367 static int
1368 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
1369 		      const void *default_mask,
1370 		      void *data)
1371 {
1372 	const struct rte_flow_item_ipv4 *spec = item->spec;
1373 	const struct rte_flow_item_ipv4 *mask = item->mask;
1374 	struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1375 	unsigned int ipv4_size = sizeof(struct ibv_flow_spec_ipv4_ext);
1376 	struct ibv_flow_spec_ipv4_ext ipv4 = {
1377 		.type = parser->inner | IBV_FLOW_SPEC_IPV4_EXT,
1378 		.size = ipv4_size,
1379 	};
1380 
1381 	parser->layer = HASH_RXQ_IPV4;
1382 	if (spec) {
1383 		if (!mask)
1384 			mask = default_mask;
1385 		ipv4.val = (struct ibv_flow_ipv4_ext_filter){
1386 			.src_ip = spec->hdr.src_addr,
1387 			.dst_ip = spec->hdr.dst_addr,
1388 			.proto = spec->hdr.next_proto_id,
1389 			.tos = spec->hdr.type_of_service,
1390 		};
1391 		ipv4.mask = (struct ibv_flow_ipv4_ext_filter){
1392 			.src_ip = mask->hdr.src_addr,
1393 			.dst_ip = mask->hdr.dst_addr,
1394 			.proto = mask->hdr.next_proto_id,
1395 			.tos = mask->hdr.type_of_service,
1396 		};
1397 		/* Remove unwanted bits from values. */
1398 		ipv4.val.src_ip &= ipv4.mask.src_ip;
1399 		ipv4.val.dst_ip &= ipv4.mask.dst_ip;
1400 		ipv4.val.proto &= ipv4.mask.proto;
1401 		ipv4.val.tos &= ipv4.mask.tos;
1402 	}
1403 	mlx5_flow_create_copy(parser, &ipv4, ipv4_size);
1404 	return 0;
1405 }
1406 
1407 /**
1408  * Convert IPv6 item to Verbs specification.
1409  *
1410  * @param item[in]
1411  *   Item specification.
1412  * @param default_mask[in]
1413  *   Default bit-masks to use when item->mask is not provided.
1414  * @param data[in, out]
1415  *   User structure.
1416  */
1417 static int
1418 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
1419 		      const void *default_mask,
1420 		      void *data)
1421 {
1422 	const struct rte_flow_item_ipv6 *spec = item->spec;
1423 	const struct rte_flow_item_ipv6 *mask = item->mask;
1424 	struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1425 	unsigned int ipv6_size = sizeof(struct ibv_flow_spec_ipv6);
1426 	struct ibv_flow_spec_ipv6 ipv6 = {
1427 		.type = parser->inner | IBV_FLOW_SPEC_IPV6,
1428 		.size = ipv6_size,
1429 	};
1430 
1431 	parser->layer = HASH_RXQ_IPV6;
1432 	if (spec) {
1433 		unsigned int i;
1434 
1435 		if (!mask)
1436 			mask = default_mask;
1437 		memcpy(&ipv6.val.src_ip, spec->hdr.src_addr,
1438 		       RTE_DIM(ipv6.val.src_ip));
1439 		memcpy(&ipv6.val.dst_ip, spec->hdr.dst_addr,
1440 		       RTE_DIM(ipv6.val.dst_ip));
1441 		memcpy(&ipv6.mask.src_ip, mask->hdr.src_addr,
1442 		       RTE_DIM(ipv6.mask.src_ip));
1443 		memcpy(&ipv6.mask.dst_ip, mask->hdr.dst_addr,
1444 		       RTE_DIM(ipv6.mask.dst_ip));
1445 		ipv6.mask.flow_label = mask->hdr.vtc_flow;
1446 		ipv6.mask.next_hdr = mask->hdr.proto;
1447 		ipv6.mask.hop_limit = mask->hdr.hop_limits;
1448 		/* Remove unwanted bits from values. */
1449 		for (i = 0; i < RTE_DIM(ipv6.val.src_ip); ++i) {
1450 			ipv6.val.src_ip[i] &= ipv6.mask.src_ip[i];
1451 			ipv6.val.dst_ip[i] &= ipv6.mask.dst_ip[i];
1452 		}
1453 		ipv6.val.flow_label &= ipv6.mask.flow_label;
1454 		ipv6.val.next_hdr &= ipv6.mask.next_hdr;
1455 		ipv6.val.hop_limit &= ipv6.mask.hop_limit;
1456 	}
1457 	mlx5_flow_create_copy(parser, &ipv6, ipv6_size);
1458 	return 0;
1459 }
1460 
1461 /**
1462  * Convert UDP item to Verbs specification.
1463  *
1464  * @param item[in]
1465  *   Item specification.
1466  * @param default_mask[in]
1467  *   Default bit-masks to use when item->mask is not provided.
1468  * @param data[in, out]
1469  *   User structure.
1470  */
1471 static int
1472 mlx5_flow_create_udp(const struct rte_flow_item *item,
1473 		     const void *default_mask,
1474 		     void *data)
1475 {
1476 	const struct rte_flow_item_udp *spec = item->spec;
1477 	const struct rte_flow_item_udp *mask = item->mask;
1478 	struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1479 	unsigned int udp_size = sizeof(struct ibv_flow_spec_tcp_udp);
1480 	struct ibv_flow_spec_tcp_udp udp = {
1481 		.type = parser->inner | IBV_FLOW_SPEC_UDP,
1482 		.size = udp_size,
1483 	};
1484 
1485 	if (parser->layer == HASH_RXQ_IPV4)
1486 		parser->layer = HASH_RXQ_UDPV4;
1487 	else
1488 		parser->layer = HASH_RXQ_UDPV6;
1489 	if (spec) {
1490 		if (!mask)
1491 			mask = default_mask;
1492 		udp.val.dst_port = spec->hdr.dst_port;
1493 		udp.val.src_port = spec->hdr.src_port;
1494 		udp.mask.dst_port = mask->hdr.dst_port;
1495 		udp.mask.src_port = mask->hdr.src_port;
1496 		/* Remove unwanted bits from values. */
1497 		udp.val.src_port &= udp.mask.src_port;
1498 		udp.val.dst_port &= udp.mask.dst_port;
1499 	}
1500 	mlx5_flow_create_copy(parser, &udp, udp_size);
1501 	return 0;
1502 }
1503 
1504 /**
1505  * Convert TCP item to Verbs specification.
1506  *
1507  * @param item[in]
1508  *   Item specification.
1509  * @param default_mask[in]
1510  *   Default bit-masks to use when item->mask is not provided.
1511  * @param data[in, out]
1512  *   User structure.
1513  */
1514 static int
1515 mlx5_flow_create_tcp(const struct rte_flow_item *item,
1516 		     const void *default_mask,
1517 		     void *data)
1518 {
1519 	const struct rte_flow_item_tcp *spec = item->spec;
1520 	const struct rte_flow_item_tcp *mask = item->mask;
1521 	struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1522 	unsigned int tcp_size = sizeof(struct ibv_flow_spec_tcp_udp);
1523 	struct ibv_flow_spec_tcp_udp tcp = {
1524 		.type = parser->inner | IBV_FLOW_SPEC_TCP,
1525 		.size = tcp_size,
1526 	};
1527 
1528 	if (parser->layer == HASH_RXQ_IPV4)
1529 		parser->layer = HASH_RXQ_TCPV4;
1530 	else
1531 		parser->layer = HASH_RXQ_TCPV6;
1532 	if (spec) {
1533 		if (!mask)
1534 			mask = default_mask;
1535 		tcp.val.dst_port = spec->hdr.dst_port;
1536 		tcp.val.src_port = spec->hdr.src_port;
1537 		tcp.mask.dst_port = mask->hdr.dst_port;
1538 		tcp.mask.src_port = mask->hdr.src_port;
1539 		/* Remove unwanted bits from values. */
1540 		tcp.val.src_port &= tcp.mask.src_port;
1541 		tcp.val.dst_port &= tcp.mask.dst_port;
1542 	}
1543 	mlx5_flow_create_copy(parser, &tcp, tcp_size);
1544 	return 0;
1545 }
1546 
1547 /**
1548  * Convert VXLAN item to Verbs specification.
1549  *
1550  * @param item[in]
1551  *   Item specification.
1552  * @param default_mask[in]
1553  *   Default bit-masks to use when item->mask is not provided.
1554  * @param data[in, out]
1555  *   User structure.
1556  */
1557 static int
1558 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
1559 		       const void *default_mask,
1560 		       void *data)
1561 {
1562 	const struct rte_flow_item_vxlan *spec = item->spec;
1563 	const struct rte_flow_item_vxlan *mask = item->mask;
1564 	struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1565 	unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
1566 	struct ibv_flow_spec_tunnel vxlan = {
1567 		.type = parser->inner | IBV_FLOW_SPEC_VXLAN_TUNNEL,
1568 		.size = size,
1569 	};
1570 	union vni {
1571 		uint32_t vlan_id;
1572 		uint8_t vni[4];
1573 	} id;
1574 
1575 	id.vni[0] = 0;
1576 	parser->inner = IBV_FLOW_SPEC_INNER;
1577 	if (spec) {
1578 		if (!mask)
1579 			mask = default_mask;
1580 		memcpy(&id.vni[1], spec->vni, 3);
1581 		vxlan.val.tunnel_id = id.vlan_id;
1582 		memcpy(&id.vni[1], mask->vni, 3);
1583 		vxlan.mask.tunnel_id = id.vlan_id;
1584 		/* Remove unwanted bits from values. */
1585 		vxlan.val.tunnel_id &= vxlan.mask.tunnel_id;
1586 	}
1587 	mlx5_flow_create_copy(parser, &vxlan, size);
1588 	return 0;
1589 }
1590 
1591 /**
1592  * Convert mark/flag action to Verbs specification.
1593  *
1594  * @param parser
1595  *   Internal parser structure.
1596  * @param mark_id
1597  *   Mark identifier.
1598  */
1599 static int
1600 mlx5_flow_create_flag_mark(struct mlx5_flow_parse *parser, uint32_t mark_id)
1601 {
1602 	unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
1603 	struct ibv_flow_spec_action_tag tag = {
1604 		.type = IBV_FLOW_SPEC_ACTION_TAG,
1605 		.size = size,
1606 		.tag_id = mlx5_flow_mark_set(mark_id),
1607 	};
1608 
1609 	assert(parser->mark);
1610 	mlx5_flow_create_copy(parser, &tag, size);
1611 	return 0;
1612 }
1613 
1614 /**
1615  * Convert count action to Verbs specification.
1616  *
1617  * @param priv
1618  *   Pointer to private structure.
1619  * @param parser
1620  *   Pointer to MLX5 flow parser structure.
1621  *
1622  * @return
1623  *   0 on success, errno value on failure.
1624  */
1625 static int
1626 mlx5_flow_create_count(struct priv *priv __rte_unused,
1627 		       struct mlx5_flow_parse *parser __rte_unused)
1628 {
1629 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
1630 	unsigned int size = sizeof(struct ibv_flow_spec_counter_action);
1631 	struct ibv_counter_set_init_attr init_attr = {0};
1632 	struct ibv_flow_spec_counter_action counter = {
1633 		.type = IBV_FLOW_SPEC_ACTION_COUNT,
1634 		.size = size,
1635 		.counter_set_handle = 0,
1636 	};
1637 
1638 	init_attr.counter_set_id = 0;
1639 	parser->cs = ibv_create_counter_set(priv->ctx, &init_attr);
1640 	if (!parser->cs)
1641 		return EINVAL;
1642 	counter.counter_set_handle = parser->cs->handle;
1643 	mlx5_flow_create_copy(parser, &counter, size);
1644 #endif
1645 	return 0;
1646 }
1647 
1648 /**
1649  * Complete flow rule creation with a drop queue.
1650  *
1651  * @param priv
1652  *   Pointer to private structure.
1653  * @param parser
1654  *   Internal parser structure.
1655  * @param flow
1656  *   Pointer to the rte_flow.
1657  * @param[out] error
1658  *   Perform verbose error reporting if not NULL.
1659  *
1660  * @return
1661  *   0 on success, errno value on failure.
1662  */
1663 static int
1664 priv_flow_create_action_queue_drop(struct priv *priv,
1665 				   struct mlx5_flow_parse *parser,
1666 				   struct rte_flow *flow,
1667 				   struct rte_flow_error *error)
1668 {
1669 	struct ibv_flow_spec_action_drop *drop;
1670 	unsigned int size = sizeof(struct ibv_flow_spec_action_drop);
1671 	int err = 0;
1672 
1673 	assert(priv->pd);
1674 	assert(priv->ctx);
1675 	flow->drop = 1;
1676 	drop = (void *)((uintptr_t)parser->drop_q.ibv_attr +
1677 			parser->drop_q.offset);
1678 	*drop = (struct ibv_flow_spec_action_drop){
1679 			.type = IBV_FLOW_SPEC_ACTION_DROP,
1680 			.size = size,
1681 	};
1682 	++parser->drop_q.ibv_attr->num_of_specs;
1683 	parser->drop_q.offset += size;
1684 	flow->drxq.ibv_attr = parser->drop_q.ibv_attr;
1685 	if (!priv->dev->data->dev_started)
1686 		return 0;
1687 	parser->drop_q.ibv_attr = NULL;
1688 	flow->drxq.ibv_flow = ibv_create_flow(priv->flow_drop_queue->qp,
1689 					      flow->drxq.ibv_attr);
1690 	if (parser->count)
1691 		flow->cs = parser->cs;
1692 	if (!flow->drxq.ibv_flow) {
1693 		rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1694 				   NULL, "flow rule creation failure");
1695 		err = ENOMEM;
1696 		goto error;
1697 	}
1698 	return 0;
1699 error:
1700 	assert(flow);
1701 	if (flow->drxq.ibv_flow) {
1702 		claim_zero(ibv_destroy_flow(flow->drxq.ibv_flow));
1703 		flow->drxq.ibv_flow = NULL;
1704 	}
1705 	if (flow->drxq.ibv_attr) {
1706 		rte_free(flow->drxq.ibv_attr);
1707 		flow->drxq.ibv_attr = NULL;
1708 	}
1709 	if (flow->cs) {
1710 		claim_zero(ibv_destroy_counter_set(flow->cs));
1711 		flow->cs = NULL;
1712 		parser->cs = NULL;
1713 	}
1714 	return err;
1715 }
1716 
1717 /**
1718  * Create hash Rx queues when RSS is enabled.
1719  *
1720  * @param priv
1721  *   Pointer to private structure.
1722  * @param parser
1723  *   Internal parser structure.
1724  * @param flow
1725  *   Pointer to the rte_flow.
1726  * @param[out] error
1727  *   Perform verbose error reporting if not NULL.
1728  *
1729  * @return
1730  *   0 on success, a errno value otherwise and rte_errno is set.
1731  */
1732 static int
1733 priv_flow_create_action_queue_rss(struct priv *priv,
1734 				  struct mlx5_flow_parse *parser,
1735 				  struct rte_flow *flow,
1736 				  struct rte_flow_error *error)
1737 {
1738 	unsigned int i;
1739 
1740 	for (i = 0; i != hash_rxq_init_n; ++i) {
1741 		uint64_t hash_fields;
1742 
1743 		if (!parser->queue[i].ibv_attr)
1744 			continue;
1745 		flow->frxq[i].ibv_attr = parser->queue[i].ibv_attr;
1746 		parser->queue[i].ibv_attr = NULL;
1747 		hash_fields = hash_rxq_init[i].hash_fields;
1748 		if (!priv->dev->data->dev_started)
1749 			continue;
1750 		flow->frxq[i].hrxq =
1751 			mlx5_priv_hrxq_get(priv,
1752 					   parser->rss_conf.rss_key,
1753 					   parser->rss_conf.rss_key_len,
1754 					   hash_fields,
1755 					   parser->queues,
1756 					   hash_fields ? parser->queues_n : 1);
1757 		if (flow->frxq[i].hrxq)
1758 			continue;
1759 		flow->frxq[i].hrxq =
1760 			mlx5_priv_hrxq_new(priv,
1761 					   parser->rss_conf.rss_key,
1762 					   parser->rss_conf.rss_key_len,
1763 					   hash_fields,
1764 					   parser->queues,
1765 					   hash_fields ? parser->queues_n : 1);
1766 		if (!flow->frxq[i].hrxq) {
1767 			rte_flow_error_set(error, ENOMEM,
1768 					   RTE_FLOW_ERROR_TYPE_HANDLE,
1769 					   NULL, "cannot create hash rxq");
1770 			return ENOMEM;
1771 		}
1772 	}
1773 	return 0;
1774 }
1775 
1776 /**
1777  * Complete flow rule creation.
1778  *
1779  * @param priv
1780  *   Pointer to private structure.
1781  * @param parser
1782  *   Internal parser structure.
1783  * @param flow
1784  *   Pointer to the rte_flow.
1785  * @param[out] error
1786  *   Perform verbose error reporting if not NULL.
1787  *
1788  * @return
1789  *   0 on success, a errno value otherwise and rte_errno is set.
1790  */
1791 static int
1792 priv_flow_create_action_queue(struct priv *priv,
1793 			      struct mlx5_flow_parse *parser,
1794 			      struct rte_flow *flow,
1795 			      struct rte_flow_error *error)
1796 {
1797 	int err = 0;
1798 	unsigned int i;
1799 
1800 	assert(priv->pd);
1801 	assert(priv->ctx);
1802 	assert(!parser->drop);
1803 	err = priv_flow_create_action_queue_rss(priv, parser, flow, error);
1804 	if (err)
1805 		goto error;
1806 	if (parser->count)
1807 		flow->cs = parser->cs;
1808 	if (!priv->dev->data->dev_started)
1809 		return 0;
1810 	for (i = 0; i != hash_rxq_init_n; ++i) {
1811 		if (!flow->frxq[i].hrxq)
1812 			continue;
1813 		flow->frxq[i].ibv_flow =
1814 			ibv_create_flow(flow->frxq[i].hrxq->qp,
1815 					flow->frxq[i].ibv_attr);
1816 		if (!flow->frxq[i].ibv_flow) {
1817 			rte_flow_error_set(error, ENOMEM,
1818 					   RTE_FLOW_ERROR_TYPE_HANDLE,
1819 					   NULL, "flow rule creation failure");
1820 			err = ENOMEM;
1821 			goto error;
1822 		}
1823 		DEBUG("%p type %d QP %p ibv_flow %p",
1824 		      (void *)flow, i,
1825 		      (void *)flow->frxq[i].hrxq,
1826 		      (void *)flow->frxq[i].ibv_flow);
1827 	}
1828 	for (i = 0; i != parser->queues_n; ++i) {
1829 		struct mlx5_rxq_data *q =
1830 			(*priv->rxqs)[parser->queues[i]];
1831 
1832 		q->mark |= parser->mark;
1833 	}
1834 	return 0;
1835 error:
1836 	assert(flow);
1837 	for (i = 0; i != hash_rxq_init_n; ++i) {
1838 		if (flow->frxq[i].ibv_flow) {
1839 			struct ibv_flow *ibv_flow = flow->frxq[i].ibv_flow;
1840 
1841 			claim_zero(ibv_destroy_flow(ibv_flow));
1842 		}
1843 		if (flow->frxq[i].hrxq)
1844 			mlx5_priv_hrxq_release(priv, flow->frxq[i].hrxq);
1845 		if (flow->frxq[i].ibv_attr)
1846 			rte_free(flow->frxq[i].ibv_attr);
1847 	}
1848 	if (flow->cs) {
1849 		claim_zero(ibv_destroy_counter_set(flow->cs));
1850 		flow->cs = NULL;
1851 		parser->cs = NULL;
1852 	}
1853 	return err;
1854 }
1855 
1856 /**
1857  * Convert a flow.
1858  *
1859  * @param priv
1860  *   Pointer to private structure.
1861  * @param list
1862  *   Pointer to a TAILQ flow list.
1863  * @param[in] attr
1864  *   Flow rule attributes.
1865  * @param[in] pattern
1866  *   Pattern specification (list terminated by the END pattern item).
1867  * @param[in] actions
1868  *   Associated actions (list terminated by the END action).
1869  * @param[out] error
1870  *   Perform verbose error reporting if not NULL.
1871  *
1872  * @return
1873  *   A flow on success, NULL otherwise.
1874  */
1875 static struct rte_flow *
1876 priv_flow_create(struct priv *priv,
1877 		 struct mlx5_flows *list,
1878 		 const struct rte_flow_attr *attr,
1879 		 const struct rte_flow_item items[],
1880 		 const struct rte_flow_action actions[],
1881 		 struct rte_flow_error *error)
1882 {
1883 	struct mlx5_flow_parse parser = { .create = 1, };
1884 	struct rte_flow *flow = NULL;
1885 	unsigned int i;
1886 	int err;
1887 
1888 	err = priv_flow_convert(priv, attr, items, actions, error, &parser);
1889 	if (err)
1890 		goto exit;
1891 	flow = rte_calloc(__func__, 1,
1892 			  sizeof(*flow) + parser.queues_n * sizeof(uint16_t),
1893 			  0);
1894 	if (!flow) {
1895 		rte_flow_error_set(error, ENOMEM,
1896 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1897 				   NULL,
1898 				   "cannot allocate flow memory");
1899 		return NULL;
1900 	}
1901 	/* Copy queues configuration. */
1902 	flow->queues = (uint16_t (*)[])(flow + 1);
1903 	memcpy(flow->queues, parser.queues, parser.queues_n * sizeof(uint16_t));
1904 	flow->queues_n = parser.queues_n;
1905 	/* Copy RSS configuration. */
1906 	flow->rss_conf = parser.rss_conf;
1907 	flow->rss_conf.rss_key = flow->rss_key;
1908 	memcpy(flow->rss_key, parser.rss_key, parser.rss_conf.rss_key_len);
1909 	/* finalise the flow. */
1910 	if (parser.drop)
1911 		err = priv_flow_create_action_queue_drop(priv, &parser, flow,
1912 							 error);
1913 	else
1914 		err = priv_flow_create_action_queue(priv, &parser, flow, error);
1915 	if (err)
1916 		goto exit;
1917 	TAILQ_INSERT_TAIL(list, flow, next);
1918 	DEBUG("Flow created %p", (void *)flow);
1919 	return flow;
1920 exit:
1921 	if (parser.drop) {
1922 		rte_free(parser.drop_q.ibv_attr);
1923 	} else {
1924 		for (i = 0; i != hash_rxq_init_n; ++i) {
1925 			if (parser.queue[i].ibv_attr)
1926 				rte_free(parser.queue[i].ibv_attr);
1927 		}
1928 	}
1929 	rte_free(flow);
1930 	return NULL;
1931 }
1932 
1933 /**
1934  * Validate a flow supported by the NIC.
1935  *
1936  * @see rte_flow_validate()
1937  * @see rte_flow_ops
1938  */
1939 int
1940 mlx5_flow_validate(struct rte_eth_dev *dev,
1941 		   const struct rte_flow_attr *attr,
1942 		   const struct rte_flow_item items[],
1943 		   const struct rte_flow_action actions[],
1944 		   struct rte_flow_error *error)
1945 {
1946 	struct priv *priv = dev->data->dev_private;
1947 	int ret;
1948 	struct mlx5_flow_parse parser = { .create = 0, };
1949 
1950 	priv_lock(priv);
1951 	ret = priv_flow_convert(priv, attr, items, actions, error, &parser);
1952 	priv_unlock(priv);
1953 	return ret;
1954 }
1955 
1956 /**
1957  * Create a flow.
1958  *
1959  * @see rte_flow_create()
1960  * @see rte_flow_ops
1961  */
1962 struct rte_flow *
1963 mlx5_flow_create(struct rte_eth_dev *dev,
1964 		 const struct rte_flow_attr *attr,
1965 		 const struct rte_flow_item items[],
1966 		 const struct rte_flow_action actions[],
1967 		 struct rte_flow_error *error)
1968 {
1969 	struct priv *priv = dev->data->dev_private;
1970 	struct rte_flow *flow;
1971 
1972 	priv_lock(priv);
1973 	flow = priv_flow_create(priv, &priv->flows, attr, items, actions,
1974 				error);
1975 	priv_unlock(priv);
1976 	return flow;
1977 }
1978 
1979 /**
1980  * Destroy a flow.
1981  *
1982  * @param priv
1983  *   Pointer to private structure.
1984  * @param list
1985  *   Pointer to a TAILQ flow list.
1986  * @param[in] flow
1987  *   Flow to destroy.
1988  */
1989 static void
1990 priv_flow_destroy(struct priv *priv,
1991 		  struct mlx5_flows *list,
1992 		  struct rte_flow *flow)
1993 {
1994 	unsigned int i;
1995 
1996 	if (flow->cs) {
1997 		claim_zero(ibv_destroy_counter_set(flow->cs));
1998 		flow->cs = NULL;
1999 	}
2000 	if (flow->drop || !flow->mark)
2001 		goto free;
2002 	for (i = 0; i != flow->queues_n; ++i) {
2003 		struct rte_flow *tmp;
2004 		int mark = 0;
2005 
2006 		/*
2007 		 * To remove the mark from the queue, the queue must not be
2008 		 * present in any other marked flow (RSS or not).
2009 		 */
2010 		TAILQ_FOREACH(tmp, list, next) {
2011 			unsigned int j;
2012 			uint16_t *tqs = NULL;
2013 			uint16_t tq_n = 0;
2014 
2015 			if (!tmp->mark)
2016 				continue;
2017 			for (j = 0; j != hash_rxq_init_n; ++j) {
2018 				if (!tmp->frxq[j].hrxq)
2019 					continue;
2020 				tqs = tmp->frxq[j].hrxq->ind_table->queues;
2021 				tq_n = tmp->frxq[j].hrxq->ind_table->queues_n;
2022 			}
2023 			if (!tq_n)
2024 				continue;
2025 			for (j = 0; (j != tq_n) && !mark; j++)
2026 				if (tqs[j] == (*flow->queues)[i])
2027 					mark = 1;
2028 		}
2029 		(*priv->rxqs)[(*flow->queues)[i]]->mark = mark;
2030 	}
2031 free:
2032 	if (flow->drop) {
2033 		if (flow->drxq.ibv_flow)
2034 			claim_zero(ibv_destroy_flow(flow->drxq.ibv_flow));
2035 		rte_free(flow->drxq.ibv_attr);
2036 	} else {
2037 		for (i = 0; i != hash_rxq_init_n; ++i) {
2038 			struct mlx5_flow *frxq = &flow->frxq[i];
2039 
2040 			if (frxq->ibv_flow)
2041 				claim_zero(ibv_destroy_flow(frxq->ibv_flow));
2042 			if (frxq->hrxq)
2043 				mlx5_priv_hrxq_release(priv, frxq->hrxq);
2044 			if (frxq->ibv_attr)
2045 				rte_free(frxq->ibv_attr);
2046 		}
2047 	}
2048 	TAILQ_REMOVE(list, flow, next);
2049 	DEBUG("Flow destroyed %p", (void *)flow);
2050 	rte_free(flow);
2051 }
2052 
2053 /**
2054  * Destroy all flows.
2055  *
2056  * @param priv
2057  *   Pointer to private structure.
2058  * @param list
2059  *   Pointer to a TAILQ flow list.
2060  */
2061 void
2062 priv_flow_flush(struct priv *priv, struct mlx5_flows *list)
2063 {
2064 	while (!TAILQ_EMPTY(list)) {
2065 		struct rte_flow *flow;
2066 
2067 		flow = TAILQ_FIRST(list);
2068 		priv_flow_destroy(priv, list, flow);
2069 	}
2070 }
2071 
2072 /**
2073  * Create drop queue.
2074  *
2075  * @param priv
2076  *   Pointer to private structure.
2077  *
2078  * @return
2079  *   0 on success.
2080  */
2081 int
2082 priv_flow_create_drop_queue(struct priv *priv)
2083 {
2084 	struct mlx5_hrxq_drop *fdq = NULL;
2085 
2086 	assert(priv->pd);
2087 	assert(priv->ctx);
2088 	fdq = rte_calloc(__func__, 1, sizeof(*fdq), 0);
2089 	if (!fdq) {
2090 		WARN("cannot allocate memory for drop queue");
2091 		goto error;
2092 	}
2093 	fdq->cq = ibv_create_cq(priv->ctx, 1, NULL, NULL, 0);
2094 	if (!fdq->cq) {
2095 		WARN("cannot allocate CQ for drop queue");
2096 		goto error;
2097 	}
2098 	fdq->wq = ibv_create_wq(priv->ctx,
2099 			&(struct ibv_wq_init_attr){
2100 			.wq_type = IBV_WQT_RQ,
2101 			.max_wr = 1,
2102 			.max_sge = 1,
2103 			.pd = priv->pd,
2104 			.cq = fdq->cq,
2105 			});
2106 	if (!fdq->wq) {
2107 		WARN("cannot allocate WQ for drop queue");
2108 		goto error;
2109 	}
2110 	fdq->ind_table = ibv_create_rwq_ind_table(priv->ctx,
2111 			&(struct ibv_rwq_ind_table_init_attr){
2112 			.log_ind_tbl_size = 0,
2113 			.ind_tbl = &fdq->wq,
2114 			.comp_mask = 0,
2115 			});
2116 	if (!fdq->ind_table) {
2117 		WARN("cannot allocate indirection table for drop queue");
2118 		goto error;
2119 	}
2120 	fdq->qp = ibv_create_qp_ex(priv->ctx,
2121 		&(struct ibv_qp_init_attr_ex){
2122 			.qp_type = IBV_QPT_RAW_PACKET,
2123 			.comp_mask =
2124 				IBV_QP_INIT_ATTR_PD |
2125 				IBV_QP_INIT_ATTR_IND_TABLE |
2126 				IBV_QP_INIT_ATTR_RX_HASH,
2127 			.rx_hash_conf = (struct ibv_rx_hash_conf){
2128 				.rx_hash_function =
2129 					IBV_RX_HASH_FUNC_TOEPLITZ,
2130 				.rx_hash_key_len = rss_hash_default_key_len,
2131 				.rx_hash_key = rss_hash_default_key,
2132 				.rx_hash_fields_mask = 0,
2133 				},
2134 			.rwq_ind_tbl = fdq->ind_table,
2135 			.pd = priv->pd
2136 		});
2137 	if (!fdq->qp) {
2138 		WARN("cannot allocate QP for drop queue");
2139 		goto error;
2140 	}
2141 	priv->flow_drop_queue = fdq;
2142 	return 0;
2143 error:
2144 	if (fdq->qp)
2145 		claim_zero(ibv_destroy_qp(fdq->qp));
2146 	if (fdq->ind_table)
2147 		claim_zero(ibv_destroy_rwq_ind_table(fdq->ind_table));
2148 	if (fdq->wq)
2149 		claim_zero(ibv_destroy_wq(fdq->wq));
2150 	if (fdq->cq)
2151 		claim_zero(ibv_destroy_cq(fdq->cq));
2152 	if (fdq)
2153 		rte_free(fdq);
2154 	priv->flow_drop_queue = NULL;
2155 	return -1;
2156 }
2157 
2158 /**
2159  * Delete drop queue.
2160  *
2161  * @param priv
2162  *   Pointer to private structure.
2163  */
2164 void
2165 priv_flow_delete_drop_queue(struct priv *priv)
2166 {
2167 	struct mlx5_hrxq_drop *fdq = priv->flow_drop_queue;
2168 
2169 	if (!fdq)
2170 		return;
2171 	if (fdq->qp)
2172 		claim_zero(ibv_destroy_qp(fdq->qp));
2173 	if (fdq->ind_table)
2174 		claim_zero(ibv_destroy_rwq_ind_table(fdq->ind_table));
2175 	if (fdq->wq)
2176 		claim_zero(ibv_destroy_wq(fdq->wq));
2177 	if (fdq->cq)
2178 		claim_zero(ibv_destroy_cq(fdq->cq));
2179 	rte_free(fdq);
2180 	priv->flow_drop_queue = NULL;
2181 }
2182 
2183 /**
2184  * Remove all flows.
2185  *
2186  * @param priv
2187  *   Pointer to private structure.
2188  * @param list
2189  *   Pointer to a TAILQ flow list.
2190  */
2191 void
2192 priv_flow_stop(struct priv *priv, struct mlx5_flows *list)
2193 {
2194 	struct rte_flow *flow;
2195 
2196 	TAILQ_FOREACH_REVERSE(flow, list, mlx5_flows, next) {
2197 		unsigned int i;
2198 
2199 		if (flow->drop) {
2200 			if (!flow->drxq.ibv_flow)
2201 				continue;
2202 			claim_zero(ibv_destroy_flow(flow->drxq.ibv_flow));
2203 			flow->drxq.ibv_flow = NULL;
2204 			/* Next flow. */
2205 			continue;
2206 		}
2207 		if (flow->mark) {
2208 			struct mlx5_ind_table_ibv *ind_tbl = NULL;
2209 
2210 			for (i = 0; i != hash_rxq_init_n; ++i) {
2211 				if (!flow->frxq[i].hrxq)
2212 					continue;
2213 				ind_tbl = flow->frxq[i].hrxq->ind_table;
2214 			}
2215 			assert(ind_tbl);
2216 			for (i = 0; i != ind_tbl->queues_n; ++i)
2217 				(*priv->rxqs)[ind_tbl->queues[i]]->mark = 0;
2218 		}
2219 		for (i = 0; i != hash_rxq_init_n; ++i) {
2220 			if (!flow->frxq[i].ibv_flow)
2221 				continue;
2222 			claim_zero(ibv_destroy_flow(flow->frxq[i].ibv_flow));
2223 			flow->frxq[i].ibv_flow = NULL;
2224 			mlx5_priv_hrxq_release(priv, flow->frxq[i].hrxq);
2225 			flow->frxq[i].hrxq = NULL;
2226 		}
2227 		DEBUG("Flow %p removed", (void *)flow);
2228 	}
2229 }
2230 
2231 /**
2232  * Add all flows.
2233  *
2234  * @param priv
2235  *   Pointer to private structure.
2236  * @param list
2237  *   Pointer to a TAILQ flow list.
2238  *
2239  * @return
2240  *   0 on success, a errno value otherwise and rte_errno is set.
2241  */
2242 int
2243 priv_flow_start(struct priv *priv, struct mlx5_flows *list)
2244 {
2245 	struct rte_flow *flow;
2246 
2247 	TAILQ_FOREACH(flow, list, next) {
2248 		unsigned int i;
2249 
2250 		if (flow->drop) {
2251 			flow->drxq.ibv_flow =
2252 				ibv_create_flow(priv->flow_drop_queue->qp,
2253 						flow->drxq.ibv_attr);
2254 			if (!flow->drxq.ibv_flow) {
2255 				DEBUG("Flow %p cannot be applied",
2256 				      (void *)flow);
2257 				rte_errno = EINVAL;
2258 				return rte_errno;
2259 			}
2260 			DEBUG("Flow %p applied", (void *)flow);
2261 			/* Next flow. */
2262 			continue;
2263 		}
2264 		for (i = 0; i != hash_rxq_init_n; ++i) {
2265 			if (!flow->frxq[i].ibv_attr)
2266 				continue;
2267 			flow->frxq[i].hrxq =
2268 				mlx5_priv_hrxq_get(priv, flow->rss_conf.rss_key,
2269 						   flow->rss_conf.rss_key_len,
2270 						   hash_rxq_init[i].hash_fields,
2271 						   (*flow->queues),
2272 						   flow->queues_n);
2273 			if (flow->frxq[i].hrxq)
2274 				goto flow_create;
2275 			flow->frxq[i].hrxq =
2276 				mlx5_priv_hrxq_new(priv, flow->rss_conf.rss_key,
2277 						   flow->rss_conf.rss_key_len,
2278 						   hash_rxq_init[i].hash_fields,
2279 						   (*flow->queues),
2280 						   flow->queues_n);
2281 			if (!flow->frxq[i].hrxq) {
2282 				DEBUG("Flow %p cannot be applied",
2283 				      (void *)flow);
2284 				rte_errno = EINVAL;
2285 				return rte_errno;
2286 			}
2287 flow_create:
2288 			flow->frxq[i].ibv_flow =
2289 				ibv_create_flow(flow->frxq[i].hrxq->qp,
2290 						flow->frxq[i].ibv_attr);
2291 			if (!flow->frxq[i].ibv_flow) {
2292 				DEBUG("Flow %p cannot be applied",
2293 				      (void *)flow);
2294 				rte_errno = EINVAL;
2295 				return rte_errno;
2296 			}
2297 			DEBUG("Flow %p applied", (void *)flow);
2298 		}
2299 		if (!flow->mark)
2300 			continue;
2301 		for (i = 0; i != flow->queues_n; ++i)
2302 			(*priv->rxqs)[(*flow->queues)[i]]->mark = 1;
2303 	}
2304 	return 0;
2305 }
2306 
2307 /**
2308  * Verify the flow list is empty
2309  *
2310  * @param priv
2311  *  Pointer to private structure.
2312  *
2313  * @return the number of flows not released.
2314  */
2315 int
2316 priv_flow_verify(struct priv *priv)
2317 {
2318 	struct rte_flow *flow;
2319 	int ret = 0;
2320 
2321 	TAILQ_FOREACH(flow, &priv->flows, next) {
2322 		DEBUG("%p: flow %p still referenced", (void *)priv,
2323 		      (void *)flow);
2324 		++ret;
2325 	}
2326 	return ret;
2327 }
2328 
2329 /**
2330  * Enable a control flow configured from the control plane.
2331  *
2332  * @param dev
2333  *   Pointer to Ethernet device.
2334  * @param eth_spec
2335  *   An Ethernet flow spec to apply.
2336  * @param eth_mask
2337  *   An Ethernet flow mask to apply.
2338  * @param vlan_spec
2339  *   A VLAN flow spec to apply.
2340  * @param vlan_mask
2341  *   A VLAN flow mask to apply.
2342  *
2343  * @return
2344  *   0 on success.
2345  */
2346 int
2347 mlx5_ctrl_flow_vlan(struct rte_eth_dev *dev,
2348 		    struct rte_flow_item_eth *eth_spec,
2349 		    struct rte_flow_item_eth *eth_mask,
2350 		    struct rte_flow_item_vlan *vlan_spec,
2351 		    struct rte_flow_item_vlan *vlan_mask)
2352 {
2353 	struct priv *priv = dev->data->dev_private;
2354 	const struct rte_flow_attr attr = {
2355 		.ingress = 1,
2356 		.priority = MLX5_CTRL_FLOW_PRIORITY,
2357 	};
2358 	struct rte_flow_item items[] = {
2359 		{
2360 			.type = RTE_FLOW_ITEM_TYPE_ETH,
2361 			.spec = eth_spec,
2362 			.last = NULL,
2363 			.mask = eth_mask,
2364 		},
2365 		{
2366 			.type = (vlan_spec) ? RTE_FLOW_ITEM_TYPE_VLAN :
2367 				RTE_FLOW_ITEM_TYPE_END,
2368 			.spec = vlan_spec,
2369 			.last = NULL,
2370 			.mask = vlan_mask,
2371 		},
2372 		{
2373 			.type = RTE_FLOW_ITEM_TYPE_END,
2374 		},
2375 	};
2376 	struct rte_flow_action actions[] = {
2377 		{
2378 			.type = RTE_FLOW_ACTION_TYPE_RSS,
2379 		},
2380 		{
2381 			.type = RTE_FLOW_ACTION_TYPE_END,
2382 		},
2383 	};
2384 	struct rte_flow *flow;
2385 	struct rte_flow_error error;
2386 	unsigned int i;
2387 	union {
2388 		struct rte_flow_action_rss rss;
2389 		struct {
2390 			const struct rte_eth_rss_conf *rss_conf;
2391 			uint16_t num;
2392 			uint16_t queue[RTE_MAX_QUEUES_PER_PORT];
2393 		} local;
2394 	} action_rss;
2395 
2396 	if (!priv->reta_idx_n)
2397 		return EINVAL;
2398 	for (i = 0; i != priv->reta_idx_n; ++i)
2399 		action_rss.local.queue[i] = (*priv->reta_idx)[i];
2400 	action_rss.local.rss_conf = &priv->rss_conf;
2401 	action_rss.local.num = priv->reta_idx_n;
2402 	actions[0].conf = (const void *)&action_rss.rss;
2403 	flow = priv_flow_create(priv, &priv->ctrl_flows, &attr, items, actions,
2404 				&error);
2405 	if (!flow)
2406 		return rte_errno;
2407 	return 0;
2408 }
2409 
2410 /**
2411  * Enable a flow control configured from the control plane.
2412  *
2413  * @param dev
2414  *   Pointer to Ethernet device.
2415  * @param eth_spec
2416  *   An Ethernet flow spec to apply.
2417  * @param eth_mask
2418  *   An Ethernet flow mask to apply.
2419  *
2420  * @return
2421  *   0 on success.
2422  */
2423 int
2424 mlx5_ctrl_flow(struct rte_eth_dev *dev,
2425 	       struct rte_flow_item_eth *eth_spec,
2426 	       struct rte_flow_item_eth *eth_mask)
2427 {
2428 	return mlx5_ctrl_flow_vlan(dev, eth_spec, eth_mask, NULL, NULL);
2429 }
2430 
2431 /**
2432  * Destroy a flow.
2433  *
2434  * @see rte_flow_destroy()
2435  * @see rte_flow_ops
2436  */
2437 int
2438 mlx5_flow_destroy(struct rte_eth_dev *dev,
2439 		  struct rte_flow *flow,
2440 		  struct rte_flow_error *error)
2441 {
2442 	struct priv *priv = dev->data->dev_private;
2443 
2444 	(void)error;
2445 	priv_lock(priv);
2446 	priv_flow_destroy(priv, &priv->flows, flow);
2447 	priv_unlock(priv);
2448 	return 0;
2449 }
2450 
2451 /**
2452  * Destroy all flows.
2453  *
2454  * @see rte_flow_flush()
2455  * @see rte_flow_ops
2456  */
2457 int
2458 mlx5_flow_flush(struct rte_eth_dev *dev,
2459 		struct rte_flow_error *error)
2460 {
2461 	struct priv *priv = dev->data->dev_private;
2462 
2463 	(void)error;
2464 	priv_lock(priv);
2465 	priv_flow_flush(priv, &priv->flows);
2466 	priv_unlock(priv);
2467 	return 0;
2468 }
2469 
2470 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
2471 /**
2472  * Query flow counter.
2473  *
2474  * @param cs
2475  *   the counter set.
2476  * @param counter_value
2477  *   returned data from the counter.
2478  *
2479  * @return
2480  *   0 on success, a errno value otherwise and rte_errno is set.
2481  */
2482 static int
2483 priv_flow_query_count(struct ibv_counter_set *cs,
2484 		      struct mlx5_flow_counter_stats *counter_stats,
2485 		      struct rte_flow_query_count *query_count,
2486 		      struct rte_flow_error *error)
2487 {
2488 	uint64_t counters[2];
2489 	struct ibv_query_counter_set_attr query_cs_attr = {
2490 		.cs = cs,
2491 		.query_flags = IBV_COUNTER_SET_FORCE_UPDATE,
2492 	};
2493 	struct ibv_counter_set_data query_out = {
2494 		.out = counters,
2495 		.outlen = 2 * sizeof(uint64_t),
2496 	};
2497 	int res = ibv_query_counter_set(&query_cs_attr, &query_out);
2498 
2499 	if (res) {
2500 		rte_flow_error_set(error, -res,
2501 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2502 				   NULL,
2503 				   "cannot read counter");
2504 		return -res;
2505 	}
2506 	query_count->hits_set = 1;
2507 	query_count->bytes_set = 1;
2508 	query_count->hits = counters[0] - counter_stats->hits;
2509 	query_count->bytes = counters[1] - counter_stats->bytes;
2510 	if (query_count->reset) {
2511 		counter_stats->hits = counters[0];
2512 		counter_stats->bytes = counters[1];
2513 	}
2514 	return 0;
2515 }
2516 
2517 /**
2518  * Query a flows.
2519  *
2520  * @see rte_flow_query()
2521  * @see rte_flow_ops
2522  */
2523 int
2524 mlx5_flow_query(struct rte_eth_dev *dev,
2525 		struct rte_flow *flow,
2526 		enum rte_flow_action_type action __rte_unused,
2527 		void *data,
2528 		struct rte_flow_error *error)
2529 {
2530 	struct priv *priv = dev->data->dev_private;
2531 	int res = EINVAL;
2532 
2533 	priv_lock(priv);
2534 	if (flow->cs) {
2535 		res = priv_flow_query_count(flow->cs,
2536 					&flow->counter_stats,
2537 					(struct rte_flow_query_count *)data,
2538 					error);
2539 	} else {
2540 		rte_flow_error_set(error, res,
2541 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2542 				   NULL,
2543 				   "no counter found for flow");
2544 	}
2545 	priv_unlock(priv);
2546 	return -res;
2547 }
2548 #endif
2549 
2550 /**
2551  * Isolated mode.
2552  *
2553  * @see rte_flow_isolate()
2554  * @see rte_flow_ops
2555  */
2556 int
2557 mlx5_flow_isolate(struct rte_eth_dev *dev,
2558 		  int enable,
2559 		  struct rte_flow_error *error)
2560 {
2561 	struct priv *priv = dev->data->dev_private;
2562 
2563 	priv_lock(priv);
2564 	if (dev->data->dev_started) {
2565 		rte_flow_error_set(error, EBUSY,
2566 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2567 				   NULL,
2568 				   "port must be stopped first");
2569 		priv_unlock(priv);
2570 		return -rte_errno;
2571 	}
2572 	priv->isolated = !!enable;
2573 	if (enable)
2574 		priv->dev->dev_ops = &mlx5_dev_ops_isolate;
2575 	else
2576 		priv->dev->dev_ops = &mlx5_dev_ops;
2577 	priv_unlock(priv);
2578 	return 0;
2579 }
2580 
2581 /**
2582  * Convert a flow director filter to a generic flow.
2583  *
2584  * @param priv
2585  *   Private structure.
2586  * @param fdir_filter
2587  *   Flow director filter to add.
2588  * @param attributes
2589  *   Generic flow parameters structure.
2590  *
2591  * @return
2592  *  0 on success, errno value on error.
2593  */
2594 static int
2595 priv_fdir_filter_convert(struct priv *priv,
2596 			 const struct rte_eth_fdir_filter *fdir_filter,
2597 			 struct mlx5_fdir *attributes)
2598 {
2599 	const struct rte_eth_fdir_input *input = &fdir_filter->input;
2600 
2601 	/* Validate queue number. */
2602 	if (fdir_filter->action.rx_queue >= priv->rxqs_n) {
2603 		ERROR("invalid queue number %d", fdir_filter->action.rx_queue);
2604 		return EINVAL;
2605 	}
2606 	attributes->attr.ingress = 1;
2607 	attributes->items[0] = (struct rte_flow_item) {
2608 		.type = RTE_FLOW_ITEM_TYPE_ETH,
2609 		.spec = &attributes->l2,
2610 	};
2611 	switch (fdir_filter->action.behavior) {
2612 	case RTE_ETH_FDIR_ACCEPT:
2613 		attributes->actions[0] = (struct rte_flow_action){
2614 			.type = RTE_FLOW_ACTION_TYPE_QUEUE,
2615 			.conf = &attributes->queue,
2616 		};
2617 		break;
2618 	case RTE_ETH_FDIR_REJECT:
2619 		attributes->actions[0] = (struct rte_flow_action){
2620 			.type = RTE_FLOW_ACTION_TYPE_DROP,
2621 		};
2622 		break;
2623 	default:
2624 		ERROR("invalid behavior %d", fdir_filter->action.behavior);
2625 		return ENOTSUP;
2626 	}
2627 	attributes->queue.index = fdir_filter->action.rx_queue;
2628 	switch (fdir_filter->input.flow_type) {
2629 	case RTE_ETH_FLOW_NONFRAG_IPV4_UDP:
2630 		attributes->l3.ipv4.hdr = (struct ipv4_hdr){
2631 			.src_addr = input->flow.udp4_flow.ip.src_ip,
2632 			.dst_addr = input->flow.udp4_flow.ip.dst_ip,
2633 			.time_to_live = input->flow.udp4_flow.ip.ttl,
2634 			.type_of_service = input->flow.udp4_flow.ip.tos,
2635 			.next_proto_id = input->flow.udp4_flow.ip.proto,
2636 		};
2637 		attributes->l4.udp.hdr = (struct udp_hdr){
2638 			.src_port = input->flow.udp4_flow.src_port,
2639 			.dst_port = input->flow.udp4_flow.dst_port,
2640 		};
2641 		attributes->items[1] = (struct rte_flow_item){
2642 			.type = RTE_FLOW_ITEM_TYPE_IPV4,
2643 			.spec = &attributes->l3,
2644 		};
2645 		attributes->items[2] = (struct rte_flow_item){
2646 			.type = RTE_FLOW_ITEM_TYPE_UDP,
2647 			.spec = &attributes->l4,
2648 		};
2649 		break;
2650 	case RTE_ETH_FLOW_NONFRAG_IPV4_TCP:
2651 		attributes->l3.ipv4.hdr = (struct ipv4_hdr){
2652 			.src_addr = input->flow.tcp4_flow.ip.src_ip,
2653 			.dst_addr = input->flow.tcp4_flow.ip.dst_ip,
2654 			.time_to_live = input->flow.tcp4_flow.ip.ttl,
2655 			.type_of_service = input->flow.tcp4_flow.ip.tos,
2656 			.next_proto_id = input->flow.tcp4_flow.ip.proto,
2657 		};
2658 		attributes->l4.tcp.hdr = (struct tcp_hdr){
2659 			.src_port = input->flow.tcp4_flow.src_port,
2660 			.dst_port = input->flow.tcp4_flow.dst_port,
2661 		};
2662 		attributes->items[1] = (struct rte_flow_item){
2663 			.type = RTE_FLOW_ITEM_TYPE_IPV4,
2664 			.spec = &attributes->l3,
2665 		};
2666 		attributes->items[2] = (struct rte_flow_item){
2667 			.type = RTE_FLOW_ITEM_TYPE_TCP,
2668 			.spec = &attributes->l4,
2669 		};
2670 		break;
2671 	case RTE_ETH_FLOW_NONFRAG_IPV4_OTHER:
2672 		attributes->l3.ipv4.hdr = (struct ipv4_hdr){
2673 			.src_addr = input->flow.ip4_flow.src_ip,
2674 			.dst_addr = input->flow.ip4_flow.dst_ip,
2675 			.time_to_live = input->flow.ip4_flow.ttl,
2676 			.type_of_service = input->flow.ip4_flow.tos,
2677 			.next_proto_id = input->flow.ip4_flow.proto,
2678 		};
2679 		attributes->items[1] = (struct rte_flow_item){
2680 			.type = RTE_FLOW_ITEM_TYPE_IPV4,
2681 			.spec = &attributes->l3,
2682 		};
2683 		break;
2684 	case RTE_ETH_FLOW_NONFRAG_IPV6_UDP:
2685 		attributes->l3.ipv6.hdr = (struct ipv6_hdr){
2686 			.hop_limits = input->flow.udp6_flow.ip.hop_limits,
2687 			.proto = input->flow.udp6_flow.ip.proto,
2688 		};
2689 		memcpy(attributes->l3.ipv6.hdr.src_addr,
2690 		       input->flow.udp6_flow.ip.src_ip,
2691 		       RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2692 		memcpy(attributes->l3.ipv6.hdr.dst_addr,
2693 		       input->flow.udp6_flow.ip.dst_ip,
2694 		       RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2695 		attributes->l4.udp.hdr = (struct udp_hdr){
2696 			.src_port = input->flow.udp6_flow.src_port,
2697 			.dst_port = input->flow.udp6_flow.dst_port,
2698 		};
2699 		attributes->items[1] = (struct rte_flow_item){
2700 			.type = RTE_FLOW_ITEM_TYPE_IPV6,
2701 			.spec = &attributes->l3,
2702 		};
2703 		attributes->items[2] = (struct rte_flow_item){
2704 			.type = RTE_FLOW_ITEM_TYPE_UDP,
2705 			.spec = &attributes->l4,
2706 		};
2707 		break;
2708 	case RTE_ETH_FLOW_NONFRAG_IPV6_TCP:
2709 		attributes->l3.ipv6.hdr = (struct ipv6_hdr){
2710 			.hop_limits = input->flow.tcp6_flow.ip.hop_limits,
2711 			.proto = input->flow.tcp6_flow.ip.proto,
2712 		};
2713 		memcpy(attributes->l3.ipv6.hdr.src_addr,
2714 		       input->flow.tcp6_flow.ip.src_ip,
2715 		       RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2716 		memcpy(attributes->l3.ipv6.hdr.dst_addr,
2717 		       input->flow.tcp6_flow.ip.dst_ip,
2718 		       RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2719 		attributes->l4.tcp.hdr = (struct tcp_hdr){
2720 			.src_port = input->flow.tcp6_flow.src_port,
2721 			.dst_port = input->flow.tcp6_flow.dst_port,
2722 		};
2723 		attributes->items[1] = (struct rte_flow_item){
2724 			.type = RTE_FLOW_ITEM_TYPE_IPV6,
2725 			.spec = &attributes->l3,
2726 		};
2727 		attributes->items[2] = (struct rte_flow_item){
2728 			.type = RTE_FLOW_ITEM_TYPE_UDP,
2729 			.spec = &attributes->l4,
2730 		};
2731 		break;
2732 	case RTE_ETH_FLOW_NONFRAG_IPV6_OTHER:
2733 		attributes->l3.ipv6.hdr = (struct ipv6_hdr){
2734 			.hop_limits = input->flow.ipv6_flow.hop_limits,
2735 			.proto = input->flow.ipv6_flow.proto,
2736 		};
2737 		memcpy(attributes->l3.ipv6.hdr.src_addr,
2738 		       input->flow.ipv6_flow.src_ip,
2739 		       RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2740 		memcpy(attributes->l3.ipv6.hdr.dst_addr,
2741 		       input->flow.ipv6_flow.dst_ip,
2742 		       RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2743 		attributes->items[1] = (struct rte_flow_item){
2744 			.type = RTE_FLOW_ITEM_TYPE_IPV6,
2745 			.spec = &attributes->l3,
2746 		};
2747 		break;
2748 	default:
2749 		ERROR("invalid flow type%d",
2750 		      fdir_filter->input.flow_type);
2751 		return ENOTSUP;
2752 	}
2753 	return 0;
2754 }
2755 
2756 /**
2757  * Add new flow director filter and store it in list.
2758  *
2759  * @param priv
2760  *   Private structure.
2761  * @param fdir_filter
2762  *   Flow director filter to add.
2763  *
2764  * @return
2765  *   0 on success, errno value on failure.
2766  */
2767 static int
2768 priv_fdir_filter_add(struct priv *priv,
2769 		     const struct rte_eth_fdir_filter *fdir_filter)
2770 {
2771 	struct mlx5_fdir attributes = {
2772 		.attr.group = 0,
2773 	};
2774 	struct mlx5_flow_parse parser = {
2775 		.layer = HASH_RXQ_ETH,
2776 	};
2777 	struct rte_flow_error error;
2778 	struct rte_flow *flow;
2779 	int ret;
2780 
2781 	ret = priv_fdir_filter_convert(priv, fdir_filter, &attributes);
2782 	if (ret)
2783 		return -ret;
2784 	ret = priv_flow_convert(priv, &attributes.attr, attributes.items,
2785 				attributes.actions, &error, &parser);
2786 	if (ret)
2787 		return -ret;
2788 	flow = priv_flow_create(priv,
2789 				&priv->flows,
2790 				&attributes.attr,
2791 				attributes.items,
2792 				attributes.actions,
2793 				&error);
2794 	if (flow) {
2795 		DEBUG("FDIR created %p", (void *)flow);
2796 		return 0;
2797 	}
2798 	return ENOTSUP;
2799 }
2800 
2801 /**
2802  * Delete specific filter.
2803  *
2804  * @param priv
2805  *   Private structure.
2806  * @param fdir_filter
2807  *   Filter to be deleted.
2808  *
2809  * @return
2810  *   0 on success, errno value on failure.
2811  */
2812 static int
2813 priv_fdir_filter_delete(struct priv *priv,
2814 			const struct rte_eth_fdir_filter *fdir_filter)
2815 {
2816 	struct mlx5_fdir attributes;
2817 	struct mlx5_flow_parse parser = {
2818 		.create = 1,
2819 		.layer = HASH_RXQ_ETH,
2820 	};
2821 	struct rte_flow_error error;
2822 	struct rte_flow *flow;
2823 	unsigned int i;
2824 	int ret;
2825 
2826 	ret = priv_fdir_filter_convert(priv, fdir_filter, &attributes);
2827 	if (ret)
2828 		return -ret;
2829 	ret = priv_flow_convert(priv, &attributes.attr, attributes.items,
2830 				attributes.actions, &error, &parser);
2831 	if (ret)
2832 		goto exit;
2833 	TAILQ_FOREACH(flow, &priv->flows, next) {
2834 		struct ibv_flow_attr *attr;
2835 		struct ibv_spec_header *attr_h;
2836 		void *spec;
2837 		struct ibv_flow_attr *flow_attr;
2838 		struct ibv_spec_header *flow_h;
2839 		void *flow_spec;
2840 		unsigned int specs_n;
2841 
2842 		if (parser.drop)
2843 			attr = parser.drop_q.ibv_attr;
2844 		else
2845 			attr = parser.queue[HASH_RXQ_ETH].ibv_attr;
2846 		if (flow->drop)
2847 			flow_attr = flow->drxq.ibv_attr;
2848 		else
2849 			flow_attr = flow->frxq[HASH_RXQ_ETH].ibv_attr;
2850 		/* Compare first the attributes. */
2851 		if (memcmp(attr, flow_attr, sizeof(struct ibv_flow_attr)))
2852 			continue;
2853 		if (attr->num_of_specs == 0)
2854 			continue;
2855 		spec = (void *)((uintptr_t)attr +
2856 				sizeof(struct ibv_flow_attr));
2857 		flow_spec = (void *)((uintptr_t)flow_attr +
2858 				     sizeof(struct ibv_flow_attr));
2859 		specs_n = RTE_MIN(attr->num_of_specs, flow_attr->num_of_specs);
2860 		for (i = 0; i != specs_n; ++i) {
2861 			attr_h = spec;
2862 			flow_h = flow_spec;
2863 			if (memcmp(spec, flow_spec,
2864 				   RTE_MIN(attr_h->size, flow_h->size)))
2865 				continue;
2866 			spec = (void *)((uintptr_t)attr + attr_h->size);
2867 			flow_spec = (void *)((uintptr_t)flow_attr +
2868 					     flow_h->size);
2869 		}
2870 		/* At this point, the flow match. */
2871 		break;
2872 	}
2873 	if (flow)
2874 		priv_flow_destroy(priv, &priv->flows, flow);
2875 exit:
2876 	if (parser.drop) {
2877 		rte_free(parser.drop_q.ibv_attr);
2878 	} else {
2879 		for (i = 0; i != hash_rxq_init_n; ++i) {
2880 			if (parser.queue[i].ibv_attr)
2881 				rte_free(parser.queue[i].ibv_attr);
2882 		}
2883 	}
2884 	return -ret;
2885 }
2886 
2887 /**
2888  * Update queue for specific filter.
2889  *
2890  * @param priv
2891  *   Private structure.
2892  * @param fdir_filter
2893  *   Filter to be updated.
2894  *
2895  * @return
2896  *   0 on success, errno value on failure.
2897  */
2898 static int
2899 priv_fdir_filter_update(struct priv *priv,
2900 			const struct rte_eth_fdir_filter *fdir_filter)
2901 {
2902 	int ret;
2903 
2904 	ret = priv_fdir_filter_delete(priv, fdir_filter);
2905 	if (ret)
2906 		return ret;
2907 	ret = priv_fdir_filter_add(priv, fdir_filter);
2908 	return ret;
2909 }
2910 
2911 /**
2912  * Flush all filters.
2913  *
2914  * @param priv
2915  *   Private structure.
2916  */
2917 static void
2918 priv_fdir_filter_flush(struct priv *priv)
2919 {
2920 	priv_flow_flush(priv, &priv->flows);
2921 }
2922 
2923 /**
2924  * Get flow director information.
2925  *
2926  * @param priv
2927  *   Private structure.
2928  * @param[out] fdir_info
2929  *   Resulting flow director information.
2930  */
2931 static void
2932 priv_fdir_info_get(struct priv *priv, struct rte_eth_fdir_info *fdir_info)
2933 {
2934 	struct rte_eth_fdir_masks *mask =
2935 		&priv->dev->data->dev_conf.fdir_conf.mask;
2936 
2937 	fdir_info->mode = priv->dev->data->dev_conf.fdir_conf.mode;
2938 	fdir_info->guarant_spc = 0;
2939 	rte_memcpy(&fdir_info->mask, mask, sizeof(fdir_info->mask));
2940 	fdir_info->max_flexpayload = 0;
2941 	fdir_info->flow_types_mask[0] = 0;
2942 	fdir_info->flex_payload_unit = 0;
2943 	fdir_info->max_flex_payload_segment_num = 0;
2944 	fdir_info->flex_payload_limit = 0;
2945 	memset(&fdir_info->flex_conf, 0, sizeof(fdir_info->flex_conf));
2946 }
2947 
2948 /**
2949  * Deal with flow director operations.
2950  *
2951  * @param priv
2952  *   Pointer to private structure.
2953  * @param filter_op
2954  *   Operation to perform.
2955  * @param arg
2956  *   Pointer to operation-specific structure.
2957  *
2958  * @return
2959  *   0 on success, errno value on failure.
2960  */
2961 static int
2962 priv_fdir_ctrl_func(struct priv *priv, enum rte_filter_op filter_op, void *arg)
2963 {
2964 	enum rte_fdir_mode fdir_mode =
2965 		priv->dev->data->dev_conf.fdir_conf.mode;
2966 	int ret = 0;
2967 
2968 	if (filter_op == RTE_ETH_FILTER_NOP)
2969 		return 0;
2970 	if (fdir_mode != RTE_FDIR_MODE_PERFECT &&
2971 	    fdir_mode != RTE_FDIR_MODE_PERFECT_MAC_VLAN) {
2972 		ERROR("%p: flow director mode %d not supported",
2973 		      (void *)priv, fdir_mode);
2974 		return EINVAL;
2975 	}
2976 	switch (filter_op) {
2977 	case RTE_ETH_FILTER_ADD:
2978 		ret = priv_fdir_filter_add(priv, arg);
2979 		break;
2980 	case RTE_ETH_FILTER_UPDATE:
2981 		ret = priv_fdir_filter_update(priv, arg);
2982 		break;
2983 	case RTE_ETH_FILTER_DELETE:
2984 		ret = priv_fdir_filter_delete(priv, arg);
2985 		break;
2986 	case RTE_ETH_FILTER_FLUSH:
2987 		priv_fdir_filter_flush(priv);
2988 		break;
2989 	case RTE_ETH_FILTER_INFO:
2990 		priv_fdir_info_get(priv, arg);
2991 		break;
2992 	default:
2993 		DEBUG("%p: unknown operation %u", (void *)priv,
2994 		      filter_op);
2995 		ret = EINVAL;
2996 		break;
2997 	}
2998 	return ret;
2999 }
3000 
3001 /**
3002  * Manage filter operations.
3003  *
3004  * @param dev
3005  *   Pointer to Ethernet device structure.
3006  * @param filter_type
3007  *   Filter type.
3008  * @param filter_op
3009  *   Operation to perform.
3010  * @param arg
3011  *   Pointer to operation-specific structure.
3012  *
3013  * @return
3014  *   0 on success, negative errno value on failure.
3015  */
3016 int
3017 mlx5_dev_filter_ctrl(struct rte_eth_dev *dev,
3018 		     enum rte_filter_type filter_type,
3019 		     enum rte_filter_op filter_op,
3020 		     void *arg)
3021 {
3022 	int ret = EINVAL;
3023 	struct priv *priv = dev->data->dev_private;
3024 
3025 	switch (filter_type) {
3026 	case RTE_ETH_FILTER_GENERIC:
3027 		if (filter_op != RTE_ETH_FILTER_GET)
3028 			return -EINVAL;
3029 		*(const void **)arg = &mlx5_flow_ops;
3030 		return 0;
3031 	case RTE_ETH_FILTER_FDIR:
3032 		priv_lock(priv);
3033 		ret = priv_fdir_ctrl_func(priv, filter_op, arg);
3034 		priv_unlock(priv);
3035 		break;
3036 	default:
3037 		ERROR("%p: filter type (%d) not supported",
3038 		      (void *)dev, filter_type);
3039 		break;
3040 	}
3041 	return -ret;
3042 }
3043