xref: /dpdk/drivers/net/mlx5/mlx5_flow.c (revision 3c272b280a50d1d8ecbc848dc01ff82a34232e97)
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->queues_n && !parser->drop) {
790 		rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE,
791 				   NULL, "no valid action");
792 		return -rte_errno;
793 	}
794 	return 0;
795 exit_action_not_supported:
796 	rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
797 			   actions, "action not supported");
798 	return -rte_errno;
799 }
800 
801 /**
802  * Validate items.
803  *
804  * @param priv
805  *   Pointer to private structure.
806  * @param[in] items
807  *   Pattern specification (list terminated by the END pattern item).
808  * @param[out] error
809  *   Perform verbose error reporting if not NULL.
810  * @param[in, out] parser
811  *   Internal parser structure.
812  *
813  * @return
814  *   0 on success, a negative errno value otherwise and rte_errno is set.
815  */
816 static int
817 priv_flow_convert_items_validate(struct priv *priv,
818 				 const struct rte_flow_item items[],
819 				 struct rte_flow_error *error,
820 				 struct mlx5_flow_parse *parser)
821 {
822 	const struct mlx5_flow_items *cur_item = mlx5_flow_items;
823 	unsigned int i;
824 
825 	(void)priv;
826 	/* Initialise the offsets to start after verbs attribute. */
827 	if (parser->drop) {
828 		parser->drop_q.offset = sizeof(struct ibv_flow_attr);
829 	} else {
830 		for (i = 0; i != hash_rxq_init_n; ++i)
831 			parser->queue[i].offset = sizeof(struct ibv_flow_attr);
832 	}
833 	for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
834 		const struct mlx5_flow_items *token = NULL;
835 		unsigned int n;
836 		int err;
837 
838 		if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
839 			continue;
840 		for (i = 0;
841 		     cur_item->items &&
842 		     cur_item->items[i] != RTE_FLOW_ITEM_TYPE_END;
843 		     ++i) {
844 			if (cur_item->items[i] == items->type) {
845 				token = &mlx5_flow_items[items->type];
846 				break;
847 			}
848 		}
849 		if (!token)
850 			goto exit_item_not_supported;
851 		cur_item = token;
852 		err = mlx5_flow_item_validate(items,
853 					      (const uint8_t *)cur_item->mask,
854 					      cur_item->mask_sz);
855 		if (err)
856 			goto exit_item_not_supported;
857 		if (items->type == RTE_FLOW_ITEM_TYPE_VXLAN) {
858 			if (parser->inner) {
859 				rte_flow_error_set(error, ENOTSUP,
860 						   RTE_FLOW_ERROR_TYPE_ITEM,
861 						   items,
862 						   "cannot recognize multiple"
863 						   " VXLAN encapsulations");
864 				return -rte_errno;
865 			}
866 			parser->inner = 1;
867 		}
868 		if (parser->drop) {
869 			parser->drop_q.offset += cur_item->dst_sz;
870 		} else if (parser->queues_n == 1) {
871 			parser->queue[HASH_RXQ_ETH].offset += cur_item->dst_sz;
872 		} else {
873 			for (n = 0; n != hash_rxq_init_n; ++n)
874 				parser->queue[n].offset += cur_item->dst_sz;
875 		}
876 	}
877 	if (parser->mark) {
878 		for (i = 0; i != hash_rxq_init_n; ++i)
879 			parser->queue[i].offset +=
880 				sizeof(struct ibv_flow_spec_action_tag);
881 	}
882 	if (parser->count) {
883 		unsigned int size = sizeof(struct ibv_flow_spec_counter_action);
884 
885 		if (parser->drop) {
886 			parser->drop_q.offset += size;
887 		} else {
888 			for (i = 0; i != hash_rxq_init_n; ++i)
889 				parser->queue[i].offset += size;
890 		}
891 	}
892 	return 0;
893 exit_item_not_supported:
894 	rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
895 			   items, "item not supported");
896 	return -rte_errno;
897 }
898 
899 /**
900  * Allocate memory space to store verbs flow attributes.
901  *
902  * @param priv
903  *   Pointer to private structure.
904  * @param[in] priority
905  *   Flow priority.
906  * @param[in] size
907  *   Amount of byte to allocate.
908  * @param[out] error
909  *   Perform verbose error reporting if not NULL.
910  *
911  * @return
912  *   A verbs flow attribute on success, NULL otherwise.
913  */
914 static struct ibv_flow_attr*
915 priv_flow_convert_allocate(struct priv *priv,
916 			   unsigned int priority,
917 			   unsigned int size,
918 			   struct rte_flow_error *error)
919 {
920 	struct ibv_flow_attr *ibv_attr;
921 
922 	(void)priv;
923 	ibv_attr = rte_calloc(__func__, 1, size, 0);
924 	if (!ibv_attr) {
925 		rte_flow_error_set(error, ENOMEM,
926 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
927 				   NULL,
928 				   "cannot allocate verbs spec attributes.");
929 		return NULL;
930 	}
931 	ibv_attr->priority = priority;
932 	return ibv_attr;
933 }
934 
935 /**
936  * Finalise verbs flow attributes.
937  *
938  * @param priv
939  *   Pointer to private structure.
940  * @param[in, out] parser
941  *   Internal parser structure.
942  */
943 static void
944 priv_flow_convert_finalise(struct priv *priv, struct mlx5_flow_parse *parser)
945 {
946 	const unsigned int ipv4 =
947 		hash_rxq_init[parser->layer].ip_version == MLX5_IPV4;
948 	const enum hash_rxq_type hmin = ipv4 ? HASH_RXQ_TCPV4 : HASH_RXQ_TCPV6;
949 	const enum hash_rxq_type hmax = ipv4 ? HASH_RXQ_IPV4 : HASH_RXQ_IPV6;
950 	const enum hash_rxq_type ohmin = ipv4 ? HASH_RXQ_TCPV6 : HASH_RXQ_TCPV4;
951 	const enum hash_rxq_type ohmax = ipv4 ? HASH_RXQ_IPV6 : HASH_RXQ_IPV4;
952 	const enum hash_rxq_type ip = ipv4 ? HASH_RXQ_IPV4 : HASH_RXQ_IPV6;
953 	unsigned int i;
954 
955 	(void)priv;
956 	if (parser->layer == HASH_RXQ_ETH) {
957 		goto fill;
958 	} else {
959 		/*
960 		 * This layer becomes useless as the pattern define under
961 		 * layers.
962 		 */
963 		rte_free(parser->queue[HASH_RXQ_ETH].ibv_attr);
964 		parser->queue[HASH_RXQ_ETH].ibv_attr = NULL;
965 	}
966 	/* Remove opposite kind of layer e.g. IPv6 if the pattern is IPv4. */
967 	for (i = ohmin; i != (ohmax + 1); ++i) {
968 		if (!parser->queue[i].ibv_attr)
969 			continue;
970 		rte_free(parser->queue[i].ibv_attr);
971 		parser->queue[i].ibv_attr = NULL;
972 	}
973 	/* Remove impossible flow according to the RSS configuration. */
974 	if (hash_rxq_init[parser->layer].dpdk_rss_hf &
975 	    parser->rss_conf.rss_hf) {
976 		/* Remove any other flow. */
977 		for (i = hmin; i != (hmax + 1); ++i) {
978 			if ((i == parser->layer) ||
979 			     (!parser->queue[i].ibv_attr))
980 				continue;
981 			rte_free(parser->queue[i].ibv_attr);
982 			parser->queue[i].ibv_attr = NULL;
983 		}
984 	} else  if (!parser->queue[ip].ibv_attr) {
985 		/* no RSS possible with the current configuration. */
986 		parser->queues_n = 1;
987 		return;
988 	}
989 fill:
990 	/*
991 	 * Fill missing layers in verbs specifications, or compute the correct
992 	 * offset to allocate the memory space for the attributes and
993 	 * specifications.
994 	 */
995 	for (i = 0; i != hash_rxq_init_n - 1; ++i) {
996 		union {
997 			struct ibv_flow_spec_ipv4_ext ipv4;
998 			struct ibv_flow_spec_ipv6 ipv6;
999 			struct ibv_flow_spec_tcp_udp udp_tcp;
1000 		} specs;
1001 		void *dst;
1002 		uint16_t size;
1003 
1004 		if (i == parser->layer)
1005 			continue;
1006 		if (parser->layer == HASH_RXQ_ETH) {
1007 			if (hash_rxq_init[i].ip_version == MLX5_IPV4) {
1008 				size = sizeof(struct ibv_flow_spec_ipv4_ext);
1009 				specs.ipv4 = (struct ibv_flow_spec_ipv4_ext){
1010 					.type = IBV_FLOW_SPEC_IPV4_EXT |
1011 						parser->inner,
1012 					.size = size,
1013 				};
1014 			} else {
1015 				size = sizeof(struct ibv_flow_spec_ipv6);
1016 				specs.ipv6 = (struct ibv_flow_spec_ipv6){
1017 					.type = IBV_FLOW_SPEC_IPV6 |
1018 						parser->inner,
1019 					.size = size,
1020 				};
1021 			}
1022 			if (parser->queue[i].ibv_attr) {
1023 				dst = (void *)((uintptr_t)
1024 					       parser->queue[i].ibv_attr +
1025 					       parser->queue[i].offset);
1026 				memcpy(dst, &specs, size);
1027 				++parser->queue[i].ibv_attr->num_of_specs;
1028 			}
1029 			parser->queue[i].offset += size;
1030 		}
1031 		if ((i == HASH_RXQ_UDPV4) || (i == HASH_RXQ_TCPV4) ||
1032 		    (i == HASH_RXQ_UDPV6) || (i == HASH_RXQ_TCPV6)) {
1033 			size = sizeof(struct ibv_flow_spec_tcp_udp);
1034 			specs.udp_tcp = (struct ibv_flow_spec_tcp_udp) {
1035 				.type = ((i == HASH_RXQ_UDPV4 ||
1036 					  i == HASH_RXQ_UDPV6) ?
1037 					 IBV_FLOW_SPEC_UDP :
1038 					 IBV_FLOW_SPEC_TCP) |
1039 					parser->inner,
1040 				.size = size,
1041 			};
1042 			if (parser->queue[i].ibv_attr) {
1043 				dst = (void *)((uintptr_t)
1044 					       parser->queue[i].ibv_attr +
1045 					       parser->queue[i].offset);
1046 				memcpy(dst, &specs, size);
1047 				++parser->queue[i].ibv_attr->num_of_specs;
1048 			}
1049 			parser->queue[i].offset += size;
1050 		}
1051 	}
1052 }
1053 
1054 /**
1055  * Validate and convert a flow supported by the NIC.
1056  *
1057  * @param priv
1058  *   Pointer to private structure.
1059  * @param[in] attr
1060  *   Flow rule attributes.
1061  * @param[in] pattern
1062  *   Pattern specification (list terminated by the END pattern item).
1063  * @param[in] actions
1064  *   Associated actions (list terminated by the END action).
1065  * @param[out] error
1066  *   Perform verbose error reporting if not NULL.
1067  * @param[in, out] parser
1068  *   Internal parser structure.
1069  *
1070  * @return
1071  *   0 on success, a negative errno value otherwise and rte_errno is set.
1072  */
1073 static int
1074 priv_flow_convert(struct priv *priv,
1075 		  const struct rte_flow_attr *attr,
1076 		  const struct rte_flow_item items[],
1077 		  const struct rte_flow_action actions[],
1078 		  struct rte_flow_error *error,
1079 		  struct mlx5_flow_parse *parser)
1080 {
1081 	const struct mlx5_flow_items *cur_item = mlx5_flow_items;
1082 	unsigned int i;
1083 	int ret;
1084 
1085 	/* First step. Validate the attributes, items and actions. */
1086 	*parser = (struct mlx5_flow_parse){
1087 		.create = parser->create,
1088 		.layer = HASH_RXQ_ETH,
1089 		.mark_id = MLX5_FLOW_MARK_DEFAULT,
1090 	};
1091 	ret = priv_flow_convert_attributes(priv, attr, error, parser);
1092 	if (ret)
1093 		return ret;
1094 	ret = priv_flow_convert_actions(priv, actions, error, parser);
1095 	if (ret)
1096 		return ret;
1097 	ret = priv_flow_convert_items_validate(priv, items, error, parser);
1098 	if (ret)
1099 		return ret;
1100 	priv_flow_convert_finalise(priv, parser);
1101 	/*
1102 	 * Second step.
1103 	 * Allocate the memory space to store verbs specifications.
1104 	 */
1105 	if (parser->drop) {
1106 		parser->drop_q.ibv_attr =
1107 			priv_flow_convert_allocate(priv, attr->priority,
1108 						   parser->drop_q.offset,
1109 						   error);
1110 		if (!parser->drop_q.ibv_attr)
1111 			return ENOMEM;
1112 		parser->drop_q.offset = sizeof(struct ibv_flow_attr);
1113 	} else if (parser->queues_n == 1) {
1114 		unsigned int priority =
1115 			attr->priority +
1116 			hash_rxq_init[HASH_RXQ_ETH].flow_priority;
1117 		unsigned int offset = parser->queue[HASH_RXQ_ETH].offset;
1118 
1119 		parser->queue[HASH_RXQ_ETH].ibv_attr =
1120 			priv_flow_convert_allocate(priv, priority,
1121 						   offset, error);
1122 		if (!parser->queue[HASH_RXQ_ETH].ibv_attr)
1123 			return ENOMEM;
1124 		parser->queue[HASH_RXQ_ETH].offset =
1125 			sizeof(struct ibv_flow_attr);
1126 	} else {
1127 		for (i = 0; i != hash_rxq_init_n; ++i) {
1128 			unsigned int priority =
1129 				attr->priority +
1130 				hash_rxq_init[i].flow_priority;
1131 			unsigned int offset;
1132 
1133 			if (!(parser->rss_conf.rss_hf &
1134 			      hash_rxq_init[i].dpdk_rss_hf) &&
1135 			    (i != HASH_RXQ_ETH))
1136 				continue;
1137 			offset = parser->queue[i].offset;
1138 			parser->queue[i].ibv_attr =
1139 				priv_flow_convert_allocate(priv, priority,
1140 							   offset, error);
1141 			if (!parser->queue[i].ibv_attr)
1142 				goto exit_enomem;
1143 			parser->queue[i].offset = sizeof(struct ibv_flow_attr);
1144 		}
1145 	}
1146 	/* Third step. Conversion parse, fill the specifications. */
1147 	parser->inner = 0;
1148 	for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
1149 		if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
1150 			continue;
1151 		cur_item = &mlx5_flow_items[items->type];
1152 		ret = cur_item->convert(items,
1153 					(cur_item->default_mask ?
1154 					 cur_item->default_mask :
1155 					 cur_item->mask),
1156 					parser);
1157 		if (ret) {
1158 			rte_flow_error_set(error, ENOTSUP,
1159 					   RTE_FLOW_ERROR_TYPE_ITEM,
1160 					   items, "item not supported");
1161 			goto exit_free;
1162 		}
1163 	}
1164 	if (parser->mark)
1165 		mlx5_flow_create_flag_mark(parser, parser->mark_id);
1166 	if (parser->count && parser->create) {
1167 		mlx5_flow_create_count(priv, parser);
1168 		if (!parser->cs)
1169 			goto exit_count_error;
1170 	}
1171 	/*
1172 	 * Last step. Complete missing specification to reach the RSS
1173 	 * configuration.
1174 	 */
1175 	if (parser->queues_n > 1)
1176 		priv_flow_convert_finalise(priv, parser);
1177 exit_free:
1178 	/* Only verification is expected, all resources should be released. */
1179 	if (!parser->create) {
1180 		if (parser->drop) {
1181 			rte_free(parser->drop_q.ibv_attr);
1182 			parser->drop_q.ibv_attr = NULL;
1183 		}
1184 		for (i = 0; i != hash_rxq_init_n; ++i) {
1185 			if (parser->queue[i].ibv_attr) {
1186 				rte_free(parser->queue[i].ibv_attr);
1187 				parser->queue[i].ibv_attr = NULL;
1188 			}
1189 		}
1190 	}
1191 	return ret;
1192 exit_enomem:
1193 	for (i = 0; i != hash_rxq_init_n; ++i) {
1194 		if (parser->queue[i].ibv_attr) {
1195 			rte_free(parser->queue[i].ibv_attr);
1196 			parser->queue[i].ibv_attr = NULL;
1197 		}
1198 	}
1199 	rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1200 			   NULL, "cannot allocate verbs spec attributes.");
1201 	return ret;
1202 exit_count_error:
1203 	rte_flow_error_set(error, EINVAL, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1204 			   NULL, "cannot create counter.");
1205 	return rte_errno;
1206 }
1207 
1208 /**
1209  * Copy the specification created into the flow.
1210  *
1211  * @param parser
1212  *   Internal parser structure.
1213  * @param src
1214  *   Create specification.
1215  * @param size
1216  *   Size in bytes of the specification to copy.
1217  */
1218 static void
1219 mlx5_flow_create_copy(struct mlx5_flow_parse *parser, void *src,
1220 		      unsigned int size)
1221 {
1222 	unsigned int i;
1223 	void *dst;
1224 
1225 	if (parser->drop) {
1226 		dst = (void *)((uintptr_t)parser->drop_q.ibv_attr +
1227 				parser->drop_q.offset);
1228 		memcpy(dst, src, size);
1229 		++parser->drop_q.ibv_attr->num_of_specs;
1230 		parser->drop_q.offset += size;
1231 		return;
1232 	}
1233 	for (i = 0; i != hash_rxq_init_n; ++i) {
1234 		if (!parser->queue[i].ibv_attr)
1235 			continue;
1236 		/* Specification must be the same l3 type or none. */
1237 		if (parser->layer == HASH_RXQ_ETH ||
1238 		    (hash_rxq_init[parser->layer].ip_version ==
1239 		     hash_rxq_init[i].ip_version) ||
1240 		    (hash_rxq_init[i].ip_version == 0)) {
1241 			dst = (void *)((uintptr_t)parser->queue[i].ibv_attr +
1242 					parser->queue[i].offset);
1243 			memcpy(dst, src, size);
1244 			++parser->queue[i].ibv_attr->num_of_specs;
1245 			parser->queue[i].offset += size;
1246 		}
1247 	}
1248 }
1249 
1250 /**
1251  * Convert Ethernet item to Verbs specification.
1252  *
1253  * @param item[in]
1254  *   Item specification.
1255  * @param default_mask[in]
1256  *   Default bit-masks to use when item->mask is not provided.
1257  * @param data[in, out]
1258  *   User structure.
1259  */
1260 static int
1261 mlx5_flow_create_eth(const struct rte_flow_item *item,
1262 		     const void *default_mask,
1263 		     void *data)
1264 {
1265 	const struct rte_flow_item_eth *spec = item->spec;
1266 	const struct rte_flow_item_eth *mask = item->mask;
1267 	struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1268 	const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
1269 	struct ibv_flow_spec_eth eth = {
1270 		.type = parser->inner | IBV_FLOW_SPEC_ETH,
1271 		.size = eth_size,
1272 	};
1273 
1274 	parser->layer = HASH_RXQ_ETH;
1275 	if (spec) {
1276 		unsigned int i;
1277 
1278 		if (!mask)
1279 			mask = default_mask;
1280 		memcpy(&eth.val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
1281 		memcpy(&eth.val.src_mac, spec->src.addr_bytes, ETHER_ADDR_LEN);
1282 		eth.val.ether_type = spec->type;
1283 		memcpy(&eth.mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
1284 		memcpy(&eth.mask.src_mac, mask->src.addr_bytes, ETHER_ADDR_LEN);
1285 		eth.mask.ether_type = mask->type;
1286 		/* Remove unwanted bits from values. */
1287 		for (i = 0; i < ETHER_ADDR_LEN; ++i) {
1288 			eth.val.dst_mac[i] &= eth.mask.dst_mac[i];
1289 			eth.val.src_mac[i] &= eth.mask.src_mac[i];
1290 		}
1291 		eth.val.ether_type &= eth.mask.ether_type;
1292 	}
1293 	mlx5_flow_create_copy(parser, &eth, eth_size);
1294 	return 0;
1295 }
1296 
1297 /**
1298  * Convert VLAN item to Verbs specification.
1299  *
1300  * @param item[in]
1301  *   Item specification.
1302  * @param default_mask[in]
1303  *   Default bit-masks to use when item->mask is not provided.
1304  * @param data[in, out]
1305  *   User structure.
1306  */
1307 static int
1308 mlx5_flow_create_vlan(const struct rte_flow_item *item,
1309 		      const void *default_mask,
1310 		      void *data)
1311 {
1312 	const struct rte_flow_item_vlan *spec = item->spec;
1313 	const struct rte_flow_item_vlan *mask = item->mask;
1314 	struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1315 	struct ibv_flow_spec_eth *eth;
1316 	const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
1317 
1318 	if (spec) {
1319 		unsigned int i;
1320 		if (!mask)
1321 			mask = default_mask;
1322 
1323 		if (parser->drop) {
1324 			eth = (void *)((uintptr_t)parser->drop_q.ibv_attr +
1325 				       parser->drop_q.offset - eth_size);
1326 			eth->val.vlan_tag = spec->tci;
1327 			eth->mask.vlan_tag = mask->tci;
1328 			eth->val.vlan_tag &= eth->mask.vlan_tag;
1329 			return 0;
1330 		}
1331 		for (i = 0; i != hash_rxq_init_n; ++i) {
1332 			if (!parser->queue[i].ibv_attr)
1333 				continue;
1334 
1335 			eth = (void *)((uintptr_t)parser->queue[i].ibv_attr +
1336 				       parser->queue[i].offset - eth_size);
1337 			eth->val.vlan_tag = spec->tci;
1338 			eth->mask.vlan_tag = mask->tci;
1339 			eth->val.vlan_tag &= eth->mask.vlan_tag;
1340 		}
1341 	}
1342 	return 0;
1343 }
1344 
1345 /**
1346  * Convert IPv4 item to Verbs specification.
1347  *
1348  * @param item[in]
1349  *   Item specification.
1350  * @param default_mask[in]
1351  *   Default bit-masks to use when item->mask is not provided.
1352  * @param data[in, out]
1353  *   User structure.
1354  */
1355 static int
1356 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
1357 		      const void *default_mask,
1358 		      void *data)
1359 {
1360 	const struct rte_flow_item_ipv4 *spec = item->spec;
1361 	const struct rte_flow_item_ipv4 *mask = item->mask;
1362 	struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1363 	unsigned int ipv4_size = sizeof(struct ibv_flow_spec_ipv4_ext);
1364 	struct ibv_flow_spec_ipv4_ext ipv4 = {
1365 		.type = parser->inner | IBV_FLOW_SPEC_IPV4_EXT,
1366 		.size = ipv4_size,
1367 	};
1368 
1369 	parser->layer = HASH_RXQ_IPV4;
1370 	if (spec) {
1371 		if (!mask)
1372 			mask = default_mask;
1373 		ipv4.val = (struct ibv_flow_ipv4_ext_filter){
1374 			.src_ip = spec->hdr.src_addr,
1375 			.dst_ip = spec->hdr.dst_addr,
1376 			.proto = spec->hdr.next_proto_id,
1377 			.tos = spec->hdr.type_of_service,
1378 		};
1379 		ipv4.mask = (struct ibv_flow_ipv4_ext_filter){
1380 			.src_ip = mask->hdr.src_addr,
1381 			.dst_ip = mask->hdr.dst_addr,
1382 			.proto = mask->hdr.next_proto_id,
1383 			.tos = mask->hdr.type_of_service,
1384 		};
1385 		/* Remove unwanted bits from values. */
1386 		ipv4.val.src_ip &= ipv4.mask.src_ip;
1387 		ipv4.val.dst_ip &= ipv4.mask.dst_ip;
1388 		ipv4.val.proto &= ipv4.mask.proto;
1389 		ipv4.val.tos &= ipv4.mask.tos;
1390 	}
1391 	mlx5_flow_create_copy(parser, &ipv4, ipv4_size);
1392 	return 0;
1393 }
1394 
1395 /**
1396  * Convert IPv6 item to Verbs specification.
1397  *
1398  * @param item[in]
1399  *   Item specification.
1400  * @param default_mask[in]
1401  *   Default bit-masks to use when item->mask is not provided.
1402  * @param data[in, out]
1403  *   User structure.
1404  */
1405 static int
1406 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
1407 		      const void *default_mask,
1408 		      void *data)
1409 {
1410 	const struct rte_flow_item_ipv6 *spec = item->spec;
1411 	const struct rte_flow_item_ipv6 *mask = item->mask;
1412 	struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1413 	unsigned int ipv6_size = sizeof(struct ibv_flow_spec_ipv6);
1414 	struct ibv_flow_spec_ipv6 ipv6 = {
1415 		.type = parser->inner | IBV_FLOW_SPEC_IPV6,
1416 		.size = ipv6_size,
1417 	};
1418 
1419 	parser->layer = HASH_RXQ_IPV6;
1420 	if (spec) {
1421 		unsigned int i;
1422 
1423 		if (!mask)
1424 			mask = default_mask;
1425 		memcpy(&ipv6.val.src_ip, spec->hdr.src_addr,
1426 		       RTE_DIM(ipv6.val.src_ip));
1427 		memcpy(&ipv6.val.dst_ip, spec->hdr.dst_addr,
1428 		       RTE_DIM(ipv6.val.dst_ip));
1429 		memcpy(&ipv6.mask.src_ip, mask->hdr.src_addr,
1430 		       RTE_DIM(ipv6.mask.src_ip));
1431 		memcpy(&ipv6.mask.dst_ip, mask->hdr.dst_addr,
1432 		       RTE_DIM(ipv6.mask.dst_ip));
1433 		ipv6.mask.flow_label = mask->hdr.vtc_flow;
1434 		ipv6.mask.next_hdr = mask->hdr.proto;
1435 		ipv6.mask.hop_limit = mask->hdr.hop_limits;
1436 		/* Remove unwanted bits from values. */
1437 		for (i = 0; i < RTE_DIM(ipv6.val.src_ip); ++i) {
1438 			ipv6.val.src_ip[i] &= ipv6.mask.src_ip[i];
1439 			ipv6.val.dst_ip[i] &= ipv6.mask.dst_ip[i];
1440 		}
1441 		ipv6.val.flow_label &= ipv6.mask.flow_label;
1442 		ipv6.val.next_hdr &= ipv6.mask.next_hdr;
1443 		ipv6.val.hop_limit &= ipv6.mask.hop_limit;
1444 	}
1445 	mlx5_flow_create_copy(parser, &ipv6, ipv6_size);
1446 	return 0;
1447 }
1448 
1449 /**
1450  * Convert UDP item to Verbs specification.
1451  *
1452  * @param item[in]
1453  *   Item specification.
1454  * @param default_mask[in]
1455  *   Default bit-masks to use when item->mask is not provided.
1456  * @param data[in, out]
1457  *   User structure.
1458  */
1459 static int
1460 mlx5_flow_create_udp(const struct rte_flow_item *item,
1461 		     const void *default_mask,
1462 		     void *data)
1463 {
1464 	const struct rte_flow_item_udp *spec = item->spec;
1465 	const struct rte_flow_item_udp *mask = item->mask;
1466 	struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1467 	unsigned int udp_size = sizeof(struct ibv_flow_spec_tcp_udp);
1468 	struct ibv_flow_spec_tcp_udp udp = {
1469 		.type = parser->inner | IBV_FLOW_SPEC_UDP,
1470 		.size = udp_size,
1471 	};
1472 
1473 	if (parser->layer == HASH_RXQ_IPV4)
1474 		parser->layer = HASH_RXQ_UDPV4;
1475 	else
1476 		parser->layer = HASH_RXQ_UDPV6;
1477 	if (spec) {
1478 		if (!mask)
1479 			mask = default_mask;
1480 		udp.val.dst_port = spec->hdr.dst_port;
1481 		udp.val.src_port = spec->hdr.src_port;
1482 		udp.mask.dst_port = mask->hdr.dst_port;
1483 		udp.mask.src_port = mask->hdr.src_port;
1484 		/* Remove unwanted bits from values. */
1485 		udp.val.src_port &= udp.mask.src_port;
1486 		udp.val.dst_port &= udp.mask.dst_port;
1487 	}
1488 	mlx5_flow_create_copy(parser, &udp, udp_size);
1489 	return 0;
1490 }
1491 
1492 /**
1493  * Convert TCP item to Verbs specification.
1494  *
1495  * @param item[in]
1496  *   Item specification.
1497  * @param default_mask[in]
1498  *   Default bit-masks to use when item->mask is not provided.
1499  * @param data[in, out]
1500  *   User structure.
1501  */
1502 static int
1503 mlx5_flow_create_tcp(const struct rte_flow_item *item,
1504 		     const void *default_mask,
1505 		     void *data)
1506 {
1507 	const struct rte_flow_item_tcp *spec = item->spec;
1508 	const struct rte_flow_item_tcp *mask = item->mask;
1509 	struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1510 	unsigned int tcp_size = sizeof(struct ibv_flow_spec_tcp_udp);
1511 	struct ibv_flow_spec_tcp_udp tcp = {
1512 		.type = parser->inner | IBV_FLOW_SPEC_TCP,
1513 		.size = tcp_size,
1514 	};
1515 
1516 	if (parser->layer == HASH_RXQ_IPV4)
1517 		parser->layer = HASH_RXQ_TCPV4;
1518 	else
1519 		parser->layer = HASH_RXQ_TCPV6;
1520 	if (spec) {
1521 		if (!mask)
1522 			mask = default_mask;
1523 		tcp.val.dst_port = spec->hdr.dst_port;
1524 		tcp.val.src_port = spec->hdr.src_port;
1525 		tcp.mask.dst_port = mask->hdr.dst_port;
1526 		tcp.mask.src_port = mask->hdr.src_port;
1527 		/* Remove unwanted bits from values. */
1528 		tcp.val.src_port &= tcp.mask.src_port;
1529 		tcp.val.dst_port &= tcp.mask.dst_port;
1530 	}
1531 	mlx5_flow_create_copy(parser, &tcp, tcp_size);
1532 	return 0;
1533 }
1534 
1535 /**
1536  * Convert VXLAN item to Verbs specification.
1537  *
1538  * @param item[in]
1539  *   Item specification.
1540  * @param default_mask[in]
1541  *   Default bit-masks to use when item->mask is not provided.
1542  * @param data[in, out]
1543  *   User structure.
1544  */
1545 static int
1546 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
1547 		       const void *default_mask,
1548 		       void *data)
1549 {
1550 	const struct rte_flow_item_vxlan *spec = item->spec;
1551 	const struct rte_flow_item_vxlan *mask = item->mask;
1552 	struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1553 	unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
1554 	struct ibv_flow_spec_tunnel vxlan = {
1555 		.type = parser->inner | IBV_FLOW_SPEC_VXLAN_TUNNEL,
1556 		.size = size,
1557 	};
1558 	union vni {
1559 		uint32_t vlan_id;
1560 		uint8_t vni[4];
1561 	} id;
1562 
1563 	id.vni[0] = 0;
1564 	parser->inner = IBV_FLOW_SPEC_INNER;
1565 	if (spec) {
1566 		if (!mask)
1567 			mask = default_mask;
1568 		memcpy(&id.vni[1], spec->vni, 3);
1569 		vxlan.val.tunnel_id = id.vlan_id;
1570 		memcpy(&id.vni[1], mask->vni, 3);
1571 		vxlan.mask.tunnel_id = id.vlan_id;
1572 		/* Remove unwanted bits from values. */
1573 		vxlan.val.tunnel_id &= vxlan.mask.tunnel_id;
1574 	}
1575 	mlx5_flow_create_copy(parser, &vxlan, size);
1576 	return 0;
1577 }
1578 
1579 /**
1580  * Convert mark/flag action to Verbs specification.
1581  *
1582  * @param parser
1583  *   Internal parser structure.
1584  * @param mark_id
1585  *   Mark identifier.
1586  */
1587 static int
1588 mlx5_flow_create_flag_mark(struct mlx5_flow_parse *parser, uint32_t mark_id)
1589 {
1590 	unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
1591 	struct ibv_flow_spec_action_tag tag = {
1592 		.type = IBV_FLOW_SPEC_ACTION_TAG,
1593 		.size = size,
1594 		.tag_id = mlx5_flow_mark_set(mark_id),
1595 	};
1596 
1597 	assert(parser->mark);
1598 	mlx5_flow_create_copy(parser, &tag, size);
1599 	return 0;
1600 }
1601 
1602 /**
1603  * Convert count action to Verbs specification.
1604  *
1605  * @param priv
1606  *   Pointer to private structure.
1607  * @param parser
1608  *   Pointer to MLX5 flow parser structure.
1609  *
1610  * @return
1611  *   0 on success, errno value on failure.
1612  */
1613 static int
1614 mlx5_flow_create_count(struct priv *priv __rte_unused,
1615 		       struct mlx5_flow_parse *parser __rte_unused)
1616 {
1617 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
1618 	unsigned int size = sizeof(struct ibv_flow_spec_counter_action);
1619 	struct ibv_counter_set_init_attr init_attr = {0};
1620 	struct ibv_flow_spec_counter_action counter = {
1621 		.type = IBV_FLOW_SPEC_ACTION_COUNT,
1622 		.size = size,
1623 		.counter_set_handle = 0,
1624 	};
1625 
1626 	init_attr.counter_set_id = 0;
1627 	parser->cs = ibv_create_counter_set(priv->ctx, &init_attr);
1628 	if (!parser->cs)
1629 		return EINVAL;
1630 	counter.counter_set_handle = parser->cs->handle;
1631 	mlx5_flow_create_copy(parser, &counter, size);
1632 #endif
1633 	return 0;
1634 }
1635 
1636 /**
1637  * Complete flow rule creation with a drop queue.
1638  *
1639  * @param priv
1640  *   Pointer to private structure.
1641  * @param parser
1642  *   Internal parser structure.
1643  * @param flow
1644  *   Pointer to the rte_flow.
1645  * @param[out] error
1646  *   Perform verbose error reporting if not NULL.
1647  *
1648  * @return
1649  *   0 on success, errno value on failure.
1650  */
1651 static int
1652 priv_flow_create_action_queue_drop(struct priv *priv,
1653 				   struct mlx5_flow_parse *parser,
1654 				   struct rte_flow *flow,
1655 				   struct rte_flow_error *error)
1656 {
1657 	struct ibv_flow_spec_action_drop *drop;
1658 	unsigned int size = sizeof(struct ibv_flow_spec_action_drop);
1659 	int err = 0;
1660 
1661 	assert(priv->pd);
1662 	assert(priv->ctx);
1663 	flow->drop = 1;
1664 	drop = (void *)((uintptr_t)parser->drop_q.ibv_attr +
1665 			parser->drop_q.offset);
1666 	*drop = (struct ibv_flow_spec_action_drop){
1667 			.type = IBV_FLOW_SPEC_ACTION_DROP,
1668 			.size = size,
1669 	};
1670 	++parser->drop_q.ibv_attr->num_of_specs;
1671 	parser->drop_q.offset += size;
1672 	if (!priv->dev->data->dev_started)
1673 		return 0;
1674 	flow->drxq.ibv_attr = parser->drop_q.ibv_attr;
1675 	parser->drop_q.ibv_attr = NULL;
1676 	flow->drxq.ibv_flow = ibv_create_flow(priv->flow_drop_queue->qp,
1677 					      flow->drxq.ibv_attr);
1678 	if (parser->count)
1679 		flow->cs = parser->cs;
1680 	if (!flow->drxq.ibv_flow) {
1681 		rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1682 				   NULL, "flow rule creation failure");
1683 		err = ENOMEM;
1684 		goto error;
1685 	}
1686 	return 0;
1687 error:
1688 	assert(flow);
1689 	if (flow->drxq.ibv_flow) {
1690 		claim_zero(ibv_destroy_flow(flow->drxq.ibv_flow));
1691 		flow->drxq.ibv_flow = NULL;
1692 	}
1693 	if (flow->drxq.ibv_attr) {
1694 		rte_free(flow->drxq.ibv_attr);
1695 		flow->drxq.ibv_attr = NULL;
1696 	}
1697 	if (flow->cs) {
1698 		claim_zero(ibv_destroy_counter_set(flow->cs));
1699 		flow->cs = NULL;
1700 		parser->cs = NULL;
1701 	}
1702 	return err;
1703 }
1704 
1705 /**
1706  * Create hash Rx queues when RSS is enabled.
1707  *
1708  * @param priv
1709  *   Pointer to private structure.
1710  * @param parser
1711  *   Internal parser structure.
1712  * @param flow
1713  *   Pointer to the rte_flow.
1714  * @param[out] error
1715  *   Perform verbose error reporting if not NULL.
1716  *
1717  * @return
1718  *   0 on success, a errno value otherwise and rte_errno is set.
1719  */
1720 static int
1721 priv_flow_create_action_queue_rss(struct priv *priv,
1722 				  struct mlx5_flow_parse *parser,
1723 				  struct rte_flow *flow,
1724 				  struct rte_flow_error *error)
1725 {
1726 	unsigned int i;
1727 
1728 	for (i = 0; i != hash_rxq_init_n; ++i) {
1729 		uint64_t hash_fields;
1730 
1731 		if (!parser->queue[i].ibv_attr)
1732 			continue;
1733 		flow->frxq[i].ibv_attr = parser->queue[i].ibv_attr;
1734 		parser->queue[i].ibv_attr = NULL;
1735 		hash_fields = hash_rxq_init[i].hash_fields;
1736 		flow->frxq[i].hrxq =
1737 			mlx5_priv_hrxq_get(priv,
1738 					   parser->rss_conf.rss_key,
1739 					   parser->rss_conf.rss_key_len,
1740 					   hash_fields,
1741 					   parser->queues,
1742 					   hash_fields ? parser->queues_n : 1);
1743 		if (flow->frxq[i].hrxq)
1744 			continue;
1745 		flow->frxq[i].hrxq =
1746 			mlx5_priv_hrxq_new(priv,
1747 					   parser->rss_conf.rss_key,
1748 					   parser->rss_conf.rss_key_len,
1749 					   hash_fields,
1750 					   parser->queues,
1751 					   hash_fields ? parser->queues_n : 1);
1752 		if (!flow->frxq[i].hrxq) {
1753 			rte_flow_error_set(error, ENOMEM,
1754 					   RTE_FLOW_ERROR_TYPE_HANDLE,
1755 					   NULL, "cannot create hash rxq");
1756 			return ENOMEM;
1757 		}
1758 	}
1759 	return 0;
1760 }
1761 
1762 /**
1763  * Complete flow rule creation.
1764  *
1765  * @param priv
1766  *   Pointer to private structure.
1767  * @param parser
1768  *   Internal parser structure.
1769  * @param flow
1770  *   Pointer to the rte_flow.
1771  * @param[out] error
1772  *   Perform verbose error reporting if not NULL.
1773  *
1774  * @return
1775  *   0 on success, a errno value otherwise and rte_errno is set.
1776  */
1777 static int
1778 priv_flow_create_action_queue(struct priv *priv,
1779 			      struct mlx5_flow_parse *parser,
1780 			      struct rte_flow *flow,
1781 			      struct rte_flow_error *error)
1782 {
1783 	int err = 0;
1784 	unsigned int i;
1785 
1786 	assert(priv->pd);
1787 	assert(priv->ctx);
1788 	assert(!parser->drop);
1789 	err = priv_flow_create_action_queue_rss(priv, parser, flow, error);
1790 	if (err)
1791 		goto error;
1792 	if (parser->count)
1793 		flow->cs = parser->cs;
1794 	if (!priv->dev->data->dev_started)
1795 		return 0;
1796 	for (i = 0; i != hash_rxq_init_n; ++i) {
1797 		if (!flow->frxq[i].hrxq)
1798 			continue;
1799 		flow->frxq[i].ibv_flow =
1800 			ibv_create_flow(flow->frxq[i].hrxq->qp,
1801 					flow->frxq[i].ibv_attr);
1802 		if (!flow->frxq[i].ibv_flow) {
1803 			rte_flow_error_set(error, ENOMEM,
1804 					   RTE_FLOW_ERROR_TYPE_HANDLE,
1805 					   NULL, "flow rule creation failure");
1806 			err = ENOMEM;
1807 			goto error;
1808 		}
1809 		DEBUG("%p type %d QP %p ibv_flow %p",
1810 		      (void *)flow, i,
1811 		      (void *)flow->frxq[i].hrxq,
1812 		      (void *)flow->frxq[i].ibv_flow);
1813 	}
1814 	for (i = 0; i != parser->queues_n; ++i) {
1815 		struct mlx5_rxq_data *q =
1816 			(*priv->rxqs)[parser->queues[i]];
1817 
1818 		q->mark |= parser->mark;
1819 	}
1820 	return 0;
1821 error:
1822 	assert(flow);
1823 	for (i = 0; i != hash_rxq_init_n; ++i) {
1824 		if (flow->frxq[i].ibv_flow) {
1825 			struct ibv_flow *ibv_flow = flow->frxq[i].ibv_flow;
1826 
1827 			claim_zero(ibv_destroy_flow(ibv_flow));
1828 		}
1829 		if (flow->frxq[i].hrxq)
1830 			mlx5_priv_hrxq_release(priv, flow->frxq[i].hrxq);
1831 		if (flow->frxq[i].ibv_attr)
1832 			rte_free(flow->frxq[i].ibv_attr);
1833 	}
1834 	if (flow->cs) {
1835 		claim_zero(ibv_destroy_counter_set(flow->cs));
1836 		flow->cs = NULL;
1837 		parser->cs = NULL;
1838 	}
1839 	return err;
1840 }
1841 
1842 /**
1843  * Convert a flow.
1844  *
1845  * @param priv
1846  *   Pointer to private structure.
1847  * @param list
1848  *   Pointer to a TAILQ flow list.
1849  * @param[in] attr
1850  *   Flow rule attributes.
1851  * @param[in] pattern
1852  *   Pattern specification (list terminated by the END pattern item).
1853  * @param[in] actions
1854  *   Associated actions (list terminated by the END action).
1855  * @param[out] error
1856  *   Perform verbose error reporting if not NULL.
1857  *
1858  * @return
1859  *   A flow on success, NULL otherwise.
1860  */
1861 static struct rte_flow *
1862 priv_flow_create(struct priv *priv,
1863 		 struct mlx5_flows *list,
1864 		 const struct rte_flow_attr *attr,
1865 		 const struct rte_flow_item items[],
1866 		 const struct rte_flow_action actions[],
1867 		 struct rte_flow_error *error)
1868 {
1869 	struct mlx5_flow_parse parser = { .create = 1, };
1870 	struct rte_flow *flow = NULL;
1871 	unsigned int i;
1872 	int err;
1873 
1874 	err = priv_flow_convert(priv, attr, items, actions, error, &parser);
1875 	if (err)
1876 		goto exit;
1877 	flow = rte_calloc(__func__, 1,
1878 			  sizeof(*flow) + parser.queues_n * sizeof(uint16_t),
1879 			  0);
1880 	if (!flow) {
1881 		rte_flow_error_set(error, ENOMEM,
1882 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1883 				   NULL,
1884 				   "cannot allocate flow memory");
1885 		return NULL;
1886 	}
1887 	/* Copy queues configuration. */
1888 	flow->queues = (uint16_t (*)[])(flow + 1);
1889 	memcpy(flow->queues, parser.queues, parser.queues_n * sizeof(uint16_t));
1890 	flow->queues_n = parser.queues_n;
1891 	/* Copy RSS configuration. */
1892 	flow->rss_conf = parser.rss_conf;
1893 	flow->rss_conf.rss_key = flow->rss_key;
1894 	memcpy(flow->rss_key, parser.rss_key, parser.rss_conf.rss_key_len);
1895 	/* finalise the flow. */
1896 	if (parser.drop)
1897 		err = priv_flow_create_action_queue_drop(priv, &parser, flow,
1898 							 error);
1899 	else
1900 		err = priv_flow_create_action_queue(priv, &parser, flow, error);
1901 	if (err)
1902 		goto exit;
1903 	TAILQ_INSERT_TAIL(list, flow, next);
1904 	DEBUG("Flow created %p", (void *)flow);
1905 	return flow;
1906 exit:
1907 	if (parser.drop) {
1908 		rte_free(parser.drop_q.ibv_attr);
1909 	} else {
1910 		for (i = 0; i != hash_rxq_init_n; ++i) {
1911 			if (parser.queue[i].ibv_attr)
1912 				rte_free(parser.queue[i].ibv_attr);
1913 		}
1914 	}
1915 	rte_free(flow);
1916 	return NULL;
1917 }
1918 
1919 /**
1920  * Validate a flow supported by the NIC.
1921  *
1922  * @see rte_flow_validate()
1923  * @see rte_flow_ops
1924  */
1925 int
1926 mlx5_flow_validate(struct rte_eth_dev *dev,
1927 		   const struct rte_flow_attr *attr,
1928 		   const struct rte_flow_item items[],
1929 		   const struct rte_flow_action actions[],
1930 		   struct rte_flow_error *error)
1931 {
1932 	struct priv *priv = dev->data->dev_private;
1933 	int ret;
1934 	struct mlx5_flow_parse parser = { .create = 0, };
1935 
1936 	priv_lock(priv);
1937 	ret = priv_flow_convert(priv, attr, items, actions, error, &parser);
1938 	priv_unlock(priv);
1939 	return ret;
1940 }
1941 
1942 /**
1943  * Create a flow.
1944  *
1945  * @see rte_flow_create()
1946  * @see rte_flow_ops
1947  */
1948 struct rte_flow *
1949 mlx5_flow_create(struct rte_eth_dev *dev,
1950 		 const struct rte_flow_attr *attr,
1951 		 const struct rte_flow_item items[],
1952 		 const struct rte_flow_action actions[],
1953 		 struct rte_flow_error *error)
1954 {
1955 	struct priv *priv = dev->data->dev_private;
1956 	struct rte_flow *flow;
1957 
1958 	priv_lock(priv);
1959 	flow = priv_flow_create(priv, &priv->flows, attr, items, actions,
1960 				error);
1961 	priv_unlock(priv);
1962 	return flow;
1963 }
1964 
1965 /**
1966  * Destroy a flow.
1967  *
1968  * @param priv
1969  *   Pointer to private structure.
1970  * @param list
1971  *   Pointer to a TAILQ flow list.
1972  * @param[in] flow
1973  *   Flow to destroy.
1974  */
1975 static void
1976 priv_flow_destroy(struct priv *priv,
1977 		  struct mlx5_flows *list,
1978 		  struct rte_flow *flow)
1979 {
1980 	unsigned int i;
1981 
1982 	if (flow->cs) {
1983 		claim_zero(ibv_destroy_counter_set(flow->cs));
1984 		flow->cs = NULL;
1985 	}
1986 	if (flow->drop || !flow->mark)
1987 		goto free;
1988 	for (i = 0; i != flow->queues_n; ++i) {
1989 		struct rte_flow *tmp;
1990 		int mark = 0;
1991 
1992 		/*
1993 		 * To remove the mark from the queue, the queue must not be
1994 		 * present in any other marked flow (RSS or not).
1995 		 */
1996 		TAILQ_FOREACH(tmp, list, next) {
1997 			unsigned int j;
1998 			uint16_t *tqs = NULL;
1999 			uint16_t tq_n = 0;
2000 
2001 			if (!tmp->mark)
2002 				continue;
2003 			for (j = 0; j != hash_rxq_init_n; ++j) {
2004 				if (!tmp->frxq[j].hrxq)
2005 					continue;
2006 				tqs = tmp->frxq[j].hrxq->ind_table->queues;
2007 				tq_n = tmp->frxq[j].hrxq->ind_table->queues_n;
2008 			}
2009 			if (!tq_n)
2010 				continue;
2011 			for (j = 0; (j != tq_n) && !mark; j++)
2012 				if (tqs[j] == (*flow->queues)[i])
2013 					mark = 1;
2014 		}
2015 		(*priv->rxqs)[(*flow->queues)[i]]->mark = mark;
2016 	}
2017 free:
2018 	if (flow->drop) {
2019 		if (flow->drxq.ibv_flow)
2020 			claim_zero(ibv_destroy_flow(flow->drxq.ibv_flow));
2021 		rte_free(flow->drxq.ibv_attr);
2022 	} else {
2023 		for (i = 0; i != hash_rxq_init_n; ++i) {
2024 			struct mlx5_flow *frxq = &flow->frxq[i];
2025 
2026 			if (frxq->ibv_flow)
2027 				claim_zero(ibv_destroy_flow(frxq->ibv_flow));
2028 			if (frxq->hrxq)
2029 				mlx5_priv_hrxq_release(priv, frxq->hrxq);
2030 			if (frxq->ibv_attr)
2031 				rte_free(frxq->ibv_attr);
2032 		}
2033 	}
2034 	TAILQ_REMOVE(list, flow, next);
2035 	DEBUG("Flow destroyed %p", (void *)flow);
2036 	rte_free(flow);
2037 }
2038 
2039 /**
2040  * Destroy all flows.
2041  *
2042  * @param priv
2043  *   Pointer to private structure.
2044  * @param list
2045  *   Pointer to a TAILQ flow list.
2046  */
2047 void
2048 priv_flow_flush(struct priv *priv, struct mlx5_flows *list)
2049 {
2050 	while (!TAILQ_EMPTY(list)) {
2051 		struct rte_flow *flow;
2052 
2053 		flow = TAILQ_FIRST(list);
2054 		priv_flow_destroy(priv, list, flow);
2055 	}
2056 }
2057 
2058 /**
2059  * Create drop queue.
2060  *
2061  * @param priv
2062  *   Pointer to private structure.
2063  *
2064  * @return
2065  *   0 on success.
2066  */
2067 int
2068 priv_flow_create_drop_queue(struct priv *priv)
2069 {
2070 	struct mlx5_hrxq_drop *fdq = NULL;
2071 
2072 	assert(priv->pd);
2073 	assert(priv->ctx);
2074 	fdq = rte_calloc(__func__, 1, sizeof(*fdq), 0);
2075 	if (!fdq) {
2076 		WARN("cannot allocate memory for drop queue");
2077 		goto error;
2078 	}
2079 	fdq->cq = ibv_create_cq(priv->ctx, 1, NULL, NULL, 0);
2080 	if (!fdq->cq) {
2081 		WARN("cannot allocate CQ for drop queue");
2082 		goto error;
2083 	}
2084 	fdq->wq = ibv_create_wq(priv->ctx,
2085 			&(struct ibv_wq_init_attr){
2086 			.wq_type = IBV_WQT_RQ,
2087 			.max_wr = 1,
2088 			.max_sge = 1,
2089 			.pd = priv->pd,
2090 			.cq = fdq->cq,
2091 			});
2092 	if (!fdq->wq) {
2093 		WARN("cannot allocate WQ for drop queue");
2094 		goto error;
2095 	}
2096 	fdq->ind_table = ibv_create_rwq_ind_table(priv->ctx,
2097 			&(struct ibv_rwq_ind_table_init_attr){
2098 			.log_ind_tbl_size = 0,
2099 			.ind_tbl = &fdq->wq,
2100 			.comp_mask = 0,
2101 			});
2102 	if (!fdq->ind_table) {
2103 		WARN("cannot allocate indirection table for drop queue");
2104 		goto error;
2105 	}
2106 	fdq->qp = ibv_create_qp_ex(priv->ctx,
2107 		&(struct ibv_qp_init_attr_ex){
2108 			.qp_type = IBV_QPT_RAW_PACKET,
2109 			.comp_mask =
2110 				IBV_QP_INIT_ATTR_PD |
2111 				IBV_QP_INIT_ATTR_IND_TABLE |
2112 				IBV_QP_INIT_ATTR_RX_HASH,
2113 			.rx_hash_conf = (struct ibv_rx_hash_conf){
2114 				.rx_hash_function =
2115 					IBV_RX_HASH_FUNC_TOEPLITZ,
2116 				.rx_hash_key_len = rss_hash_default_key_len,
2117 				.rx_hash_key = rss_hash_default_key,
2118 				.rx_hash_fields_mask = 0,
2119 				},
2120 			.rwq_ind_tbl = fdq->ind_table,
2121 			.pd = priv->pd
2122 		});
2123 	if (!fdq->qp) {
2124 		WARN("cannot allocate QP for drop queue");
2125 		goto error;
2126 	}
2127 	priv->flow_drop_queue = fdq;
2128 	return 0;
2129 error:
2130 	if (fdq->qp)
2131 		claim_zero(ibv_destroy_qp(fdq->qp));
2132 	if (fdq->ind_table)
2133 		claim_zero(ibv_destroy_rwq_ind_table(fdq->ind_table));
2134 	if (fdq->wq)
2135 		claim_zero(ibv_destroy_wq(fdq->wq));
2136 	if (fdq->cq)
2137 		claim_zero(ibv_destroy_cq(fdq->cq));
2138 	if (fdq)
2139 		rte_free(fdq);
2140 	priv->flow_drop_queue = NULL;
2141 	return -1;
2142 }
2143 
2144 /**
2145  * Delete drop queue.
2146  *
2147  * @param priv
2148  *   Pointer to private structure.
2149  */
2150 void
2151 priv_flow_delete_drop_queue(struct priv *priv)
2152 {
2153 	struct mlx5_hrxq_drop *fdq = priv->flow_drop_queue;
2154 
2155 	if (!fdq)
2156 		return;
2157 	if (fdq->qp)
2158 		claim_zero(ibv_destroy_qp(fdq->qp));
2159 	if (fdq->ind_table)
2160 		claim_zero(ibv_destroy_rwq_ind_table(fdq->ind_table));
2161 	if (fdq->wq)
2162 		claim_zero(ibv_destroy_wq(fdq->wq));
2163 	if (fdq->cq)
2164 		claim_zero(ibv_destroy_cq(fdq->cq));
2165 	rte_free(fdq);
2166 	priv->flow_drop_queue = NULL;
2167 }
2168 
2169 /**
2170  * Remove all flows.
2171  *
2172  * @param priv
2173  *   Pointer to private structure.
2174  * @param list
2175  *   Pointer to a TAILQ flow list.
2176  */
2177 void
2178 priv_flow_stop(struct priv *priv, struct mlx5_flows *list)
2179 {
2180 	struct rte_flow *flow;
2181 
2182 	TAILQ_FOREACH_REVERSE(flow, list, mlx5_flows, next) {
2183 		unsigned int i;
2184 
2185 		if (flow->drop) {
2186 			if (!flow->drxq.ibv_flow)
2187 				continue;
2188 			claim_zero(ibv_destroy_flow(flow->drxq.ibv_flow));
2189 			flow->drxq.ibv_flow = NULL;
2190 			/* Next flow. */
2191 			continue;
2192 		}
2193 		if (flow->mark) {
2194 			struct mlx5_ind_table_ibv *ind_tbl = NULL;
2195 
2196 			for (i = 0; i != hash_rxq_init_n; ++i) {
2197 				if (!flow->frxq[i].hrxq)
2198 					continue;
2199 				ind_tbl = flow->frxq[i].hrxq->ind_table;
2200 			}
2201 			assert(ind_tbl);
2202 			for (i = 0; i != ind_tbl->queues_n; ++i)
2203 				(*priv->rxqs)[ind_tbl->queues[i]]->mark = 0;
2204 		}
2205 		for (i = 0; i != hash_rxq_init_n; ++i) {
2206 			if (!flow->frxq[i].ibv_flow)
2207 				continue;
2208 			claim_zero(ibv_destroy_flow(flow->frxq[i].ibv_flow));
2209 			flow->frxq[i].ibv_flow = NULL;
2210 			mlx5_priv_hrxq_release(priv, flow->frxq[i].hrxq);
2211 			flow->frxq[i].hrxq = NULL;
2212 		}
2213 		DEBUG("Flow %p removed", (void *)flow);
2214 	}
2215 }
2216 
2217 /**
2218  * Add all flows.
2219  *
2220  * @param priv
2221  *   Pointer to private structure.
2222  * @param list
2223  *   Pointer to a TAILQ flow list.
2224  *
2225  * @return
2226  *   0 on success, a errno value otherwise and rte_errno is set.
2227  */
2228 int
2229 priv_flow_start(struct priv *priv, struct mlx5_flows *list)
2230 {
2231 	struct rte_flow *flow;
2232 
2233 	TAILQ_FOREACH(flow, list, next) {
2234 		unsigned int i;
2235 
2236 		if (flow->drop) {
2237 			flow->drxq.ibv_flow =
2238 				ibv_create_flow(priv->flow_drop_queue->qp,
2239 						flow->drxq.ibv_attr);
2240 			if (!flow->drxq.ibv_flow) {
2241 				DEBUG("Flow %p cannot be applied",
2242 				      (void *)flow);
2243 				rte_errno = EINVAL;
2244 				return rte_errno;
2245 			}
2246 			DEBUG("Flow %p applied", (void *)flow);
2247 			/* Next flow. */
2248 			continue;
2249 		}
2250 		for (i = 0; i != hash_rxq_init_n; ++i) {
2251 			if (!flow->frxq[i].ibv_attr)
2252 				continue;
2253 			flow->frxq[i].hrxq =
2254 				mlx5_priv_hrxq_get(priv, flow->rss_conf.rss_key,
2255 						   flow->rss_conf.rss_key_len,
2256 						   hash_rxq_init[i].hash_fields,
2257 						   (*flow->queues),
2258 						   flow->queues_n);
2259 			if (flow->frxq[i].hrxq)
2260 				goto flow_create;
2261 			flow->frxq[i].hrxq =
2262 				mlx5_priv_hrxq_new(priv, flow->rss_conf.rss_key,
2263 						   flow->rss_conf.rss_key_len,
2264 						   hash_rxq_init[i].hash_fields,
2265 						   (*flow->queues),
2266 						   flow->queues_n);
2267 			if (!flow->frxq[i].hrxq) {
2268 				DEBUG("Flow %p cannot be applied",
2269 				      (void *)flow);
2270 				rte_errno = EINVAL;
2271 				return rte_errno;
2272 			}
2273 flow_create:
2274 			flow->frxq[i].ibv_flow =
2275 				ibv_create_flow(flow->frxq[i].hrxq->qp,
2276 						flow->frxq[i].ibv_attr);
2277 			if (!flow->frxq[i].ibv_flow) {
2278 				DEBUG("Flow %p cannot be applied",
2279 				      (void *)flow);
2280 				rte_errno = EINVAL;
2281 				return rte_errno;
2282 			}
2283 			DEBUG("Flow %p applied", (void *)flow);
2284 		}
2285 		if (!flow->mark)
2286 			continue;
2287 		for (i = 0; i != flow->queues_n; ++i)
2288 			(*priv->rxqs)[(*flow->queues)[i]]->mark = 1;
2289 	}
2290 	return 0;
2291 }
2292 
2293 /**
2294  * Verify the flow list is empty
2295  *
2296  * @param priv
2297  *  Pointer to private structure.
2298  *
2299  * @return the number of flows not released.
2300  */
2301 int
2302 priv_flow_verify(struct priv *priv)
2303 {
2304 	struct rte_flow *flow;
2305 	int ret = 0;
2306 
2307 	TAILQ_FOREACH(flow, &priv->flows, next) {
2308 		DEBUG("%p: flow %p still referenced", (void *)priv,
2309 		      (void *)flow);
2310 		++ret;
2311 	}
2312 	return ret;
2313 }
2314 
2315 /**
2316  * Enable a control flow configured from the control plane.
2317  *
2318  * @param dev
2319  *   Pointer to Ethernet device.
2320  * @param eth_spec
2321  *   An Ethernet flow spec to apply.
2322  * @param eth_mask
2323  *   An Ethernet flow mask to apply.
2324  * @param vlan_spec
2325  *   A VLAN flow spec to apply.
2326  * @param vlan_mask
2327  *   A VLAN flow mask to apply.
2328  *
2329  * @return
2330  *   0 on success.
2331  */
2332 int
2333 mlx5_ctrl_flow_vlan(struct rte_eth_dev *dev,
2334 		    struct rte_flow_item_eth *eth_spec,
2335 		    struct rte_flow_item_eth *eth_mask,
2336 		    struct rte_flow_item_vlan *vlan_spec,
2337 		    struct rte_flow_item_vlan *vlan_mask)
2338 {
2339 	struct priv *priv = dev->data->dev_private;
2340 	const struct rte_flow_attr attr = {
2341 		.ingress = 1,
2342 		.priority = MLX5_CTRL_FLOW_PRIORITY,
2343 	};
2344 	struct rte_flow_item items[] = {
2345 		{
2346 			.type = RTE_FLOW_ITEM_TYPE_ETH,
2347 			.spec = eth_spec,
2348 			.last = NULL,
2349 			.mask = eth_mask,
2350 		},
2351 		{
2352 			.type = (vlan_spec) ? RTE_FLOW_ITEM_TYPE_VLAN :
2353 				RTE_FLOW_ITEM_TYPE_END,
2354 			.spec = vlan_spec,
2355 			.last = NULL,
2356 			.mask = vlan_mask,
2357 		},
2358 		{
2359 			.type = RTE_FLOW_ITEM_TYPE_END,
2360 		},
2361 	};
2362 	struct rte_flow_action actions[] = {
2363 		{
2364 			.type = RTE_FLOW_ACTION_TYPE_RSS,
2365 		},
2366 		{
2367 			.type = RTE_FLOW_ACTION_TYPE_END,
2368 		},
2369 	};
2370 	struct rte_flow *flow;
2371 	struct rte_flow_error error;
2372 	unsigned int i;
2373 	union {
2374 		struct rte_flow_action_rss rss;
2375 		struct {
2376 			const struct rte_eth_rss_conf *rss_conf;
2377 			uint16_t num;
2378 			uint16_t queue[RTE_MAX_QUEUES_PER_PORT];
2379 		} local;
2380 	} action_rss;
2381 
2382 	if (!priv->reta_idx_n)
2383 		return EINVAL;
2384 	for (i = 0; i != priv->reta_idx_n; ++i)
2385 		action_rss.local.queue[i] = (*priv->reta_idx)[i];
2386 	action_rss.local.rss_conf = &priv->rss_conf;
2387 	action_rss.local.num = priv->reta_idx_n;
2388 	actions[0].conf = (const void *)&action_rss.rss;
2389 	flow = priv_flow_create(priv, &priv->ctrl_flows, &attr, items, actions,
2390 				&error);
2391 	if (!flow)
2392 		return rte_errno;
2393 	return 0;
2394 }
2395 
2396 /**
2397  * Enable a flow control configured from the control plane.
2398  *
2399  * @param dev
2400  *   Pointer to Ethernet device.
2401  * @param eth_spec
2402  *   An Ethernet flow spec to apply.
2403  * @param eth_mask
2404  *   An Ethernet flow mask to apply.
2405  *
2406  * @return
2407  *   0 on success.
2408  */
2409 int
2410 mlx5_ctrl_flow(struct rte_eth_dev *dev,
2411 	       struct rte_flow_item_eth *eth_spec,
2412 	       struct rte_flow_item_eth *eth_mask)
2413 {
2414 	return mlx5_ctrl_flow_vlan(dev, eth_spec, eth_mask, NULL, NULL);
2415 }
2416 
2417 /**
2418  * Destroy a flow.
2419  *
2420  * @see rte_flow_destroy()
2421  * @see rte_flow_ops
2422  */
2423 int
2424 mlx5_flow_destroy(struct rte_eth_dev *dev,
2425 		  struct rte_flow *flow,
2426 		  struct rte_flow_error *error)
2427 {
2428 	struct priv *priv = dev->data->dev_private;
2429 
2430 	(void)error;
2431 	priv_lock(priv);
2432 	priv_flow_destroy(priv, &priv->flows, flow);
2433 	priv_unlock(priv);
2434 	return 0;
2435 }
2436 
2437 /**
2438  * Destroy all flows.
2439  *
2440  * @see rte_flow_flush()
2441  * @see rte_flow_ops
2442  */
2443 int
2444 mlx5_flow_flush(struct rte_eth_dev *dev,
2445 		struct rte_flow_error *error)
2446 {
2447 	struct priv *priv = dev->data->dev_private;
2448 
2449 	(void)error;
2450 	priv_lock(priv);
2451 	priv_flow_flush(priv, &priv->flows);
2452 	priv_unlock(priv);
2453 	return 0;
2454 }
2455 
2456 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
2457 /**
2458  * Query flow counter.
2459  *
2460  * @param cs
2461  *   the counter set.
2462  * @param counter_value
2463  *   returned data from the counter.
2464  *
2465  * @return
2466  *   0 on success, a errno value otherwise and rte_errno is set.
2467  */
2468 static int
2469 priv_flow_query_count(struct ibv_counter_set *cs,
2470 		      struct mlx5_flow_counter_stats *counter_stats,
2471 		      struct rte_flow_query_count *query_count,
2472 		      struct rte_flow_error *error)
2473 {
2474 	uint64_t counters[2];
2475 	struct ibv_query_counter_set_attr query_cs_attr = {
2476 		.cs = cs,
2477 		.query_flags = IBV_COUNTER_SET_FORCE_UPDATE,
2478 	};
2479 	struct ibv_counter_set_data query_out = {
2480 		.out = counters,
2481 		.outlen = 2 * sizeof(uint64_t),
2482 	};
2483 	int res = ibv_query_counter_set(&query_cs_attr, &query_out);
2484 
2485 	if (res) {
2486 		rte_flow_error_set(error, -res,
2487 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2488 				   NULL,
2489 				   "cannot read counter");
2490 		return -res;
2491 	}
2492 	query_count->hits_set = 1;
2493 	query_count->bytes_set = 1;
2494 	query_count->hits = counters[0] - counter_stats->hits;
2495 	query_count->bytes = counters[1] - counter_stats->bytes;
2496 	if (query_count->reset) {
2497 		counter_stats->hits = counters[0];
2498 		counter_stats->bytes = counters[1];
2499 	}
2500 	return 0;
2501 }
2502 
2503 /**
2504  * Query a flows.
2505  *
2506  * @see rte_flow_query()
2507  * @see rte_flow_ops
2508  */
2509 int
2510 mlx5_flow_query(struct rte_eth_dev *dev,
2511 		struct rte_flow *flow,
2512 		enum rte_flow_action_type action __rte_unused,
2513 		void *data,
2514 		struct rte_flow_error *error)
2515 {
2516 	struct priv *priv = dev->data->dev_private;
2517 	int res = EINVAL;
2518 
2519 	priv_lock(priv);
2520 	if (flow->cs) {
2521 		res = priv_flow_query_count(flow->cs,
2522 					&flow->counter_stats,
2523 					(struct rte_flow_query_count *)data,
2524 					error);
2525 	} else {
2526 		rte_flow_error_set(error, res,
2527 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2528 				   NULL,
2529 				   "no counter found for flow");
2530 	}
2531 	priv_unlock(priv);
2532 	return -res;
2533 }
2534 #endif
2535 
2536 /**
2537  * Isolated mode.
2538  *
2539  * @see rte_flow_isolate()
2540  * @see rte_flow_ops
2541  */
2542 int
2543 mlx5_flow_isolate(struct rte_eth_dev *dev,
2544 		  int enable,
2545 		  struct rte_flow_error *error)
2546 {
2547 	struct priv *priv = dev->data->dev_private;
2548 
2549 	priv_lock(priv);
2550 	if (dev->data->dev_started) {
2551 		rte_flow_error_set(error, EBUSY,
2552 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2553 				   NULL,
2554 				   "port must be stopped first");
2555 		priv_unlock(priv);
2556 		return -rte_errno;
2557 	}
2558 	priv->isolated = !!enable;
2559 	if (enable)
2560 		priv->dev->dev_ops = &mlx5_dev_ops_isolate;
2561 	else
2562 		priv->dev->dev_ops = &mlx5_dev_ops;
2563 	priv_unlock(priv);
2564 	return 0;
2565 }
2566 
2567 /**
2568  * Convert a flow director filter to a generic flow.
2569  *
2570  * @param priv
2571  *   Private structure.
2572  * @param fdir_filter
2573  *   Flow director filter to add.
2574  * @param attributes
2575  *   Generic flow parameters structure.
2576  *
2577  * @return
2578  *  0 on success, errno value on error.
2579  */
2580 static int
2581 priv_fdir_filter_convert(struct priv *priv,
2582 			 const struct rte_eth_fdir_filter *fdir_filter,
2583 			 struct mlx5_fdir *attributes)
2584 {
2585 	const struct rte_eth_fdir_input *input = &fdir_filter->input;
2586 
2587 	/* Validate queue number. */
2588 	if (fdir_filter->action.rx_queue >= priv->rxqs_n) {
2589 		ERROR("invalid queue number %d", fdir_filter->action.rx_queue);
2590 		return EINVAL;
2591 	}
2592 	/* Validate the behavior. */
2593 	if (fdir_filter->action.behavior != RTE_ETH_FDIR_ACCEPT) {
2594 		ERROR("invalid behavior %d", fdir_filter->action.behavior);
2595 		return ENOTSUP;
2596 	}
2597 	attributes->attr.ingress = 1;
2598 	attributes->items[0] = (struct rte_flow_item) {
2599 		.type = RTE_FLOW_ITEM_TYPE_ETH,
2600 		.spec = &attributes->l2,
2601 	};
2602 	attributes->actions[0] = (struct rte_flow_action){
2603 		.type = RTE_FLOW_ACTION_TYPE_QUEUE,
2604 		.conf = &attributes->queue,
2605 	};
2606 	attributes->queue.index = fdir_filter->action.rx_queue;
2607 	switch (fdir_filter->input.flow_type) {
2608 	case RTE_ETH_FLOW_NONFRAG_IPV4_UDP:
2609 		attributes->l3.ipv4.hdr = (struct ipv4_hdr){
2610 			.src_addr = input->flow.udp4_flow.ip.src_ip,
2611 			.dst_addr = input->flow.udp4_flow.ip.dst_ip,
2612 			.time_to_live = input->flow.udp4_flow.ip.ttl,
2613 			.type_of_service = input->flow.udp4_flow.ip.tos,
2614 			.next_proto_id = input->flow.udp4_flow.ip.proto,
2615 		};
2616 		attributes->l4.udp.hdr = (struct udp_hdr){
2617 			.src_port = input->flow.udp4_flow.src_port,
2618 			.dst_port = input->flow.udp4_flow.dst_port,
2619 		};
2620 		attributes->items[1] = (struct rte_flow_item){
2621 			.type = RTE_FLOW_ITEM_TYPE_IPV4,
2622 			.spec = &attributes->l3,
2623 		};
2624 		attributes->items[2] = (struct rte_flow_item){
2625 			.type = RTE_FLOW_ITEM_TYPE_UDP,
2626 			.spec = &attributes->l4,
2627 		};
2628 		break;
2629 	case RTE_ETH_FLOW_NONFRAG_IPV4_TCP:
2630 		attributes->l3.ipv4.hdr = (struct ipv4_hdr){
2631 			.src_addr = input->flow.tcp4_flow.ip.src_ip,
2632 			.dst_addr = input->flow.tcp4_flow.ip.dst_ip,
2633 			.time_to_live = input->flow.tcp4_flow.ip.ttl,
2634 			.type_of_service = input->flow.tcp4_flow.ip.tos,
2635 			.next_proto_id = input->flow.tcp4_flow.ip.proto,
2636 		};
2637 		attributes->l4.tcp.hdr = (struct tcp_hdr){
2638 			.src_port = input->flow.tcp4_flow.src_port,
2639 			.dst_port = input->flow.tcp4_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_TCP,
2647 			.spec = &attributes->l4,
2648 		};
2649 		break;
2650 	case RTE_ETH_FLOW_NONFRAG_IPV4_OTHER:
2651 		attributes->l3.ipv4.hdr = (struct ipv4_hdr){
2652 			.src_addr = input->flow.ip4_flow.src_ip,
2653 			.dst_addr = input->flow.ip4_flow.dst_ip,
2654 			.time_to_live = input->flow.ip4_flow.ttl,
2655 			.type_of_service = input->flow.ip4_flow.tos,
2656 			.next_proto_id = input->flow.ip4_flow.proto,
2657 		};
2658 		attributes->items[1] = (struct rte_flow_item){
2659 			.type = RTE_FLOW_ITEM_TYPE_IPV4,
2660 			.spec = &attributes->l3,
2661 		};
2662 		break;
2663 	case RTE_ETH_FLOW_NONFRAG_IPV6_UDP:
2664 		attributes->l3.ipv6.hdr = (struct ipv6_hdr){
2665 			.hop_limits = input->flow.udp6_flow.ip.hop_limits,
2666 			.proto = input->flow.udp6_flow.ip.proto,
2667 		};
2668 		memcpy(attributes->l3.ipv6.hdr.src_addr,
2669 		       input->flow.udp6_flow.ip.src_ip,
2670 		       RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2671 		memcpy(attributes->l3.ipv6.hdr.dst_addr,
2672 		       input->flow.udp6_flow.ip.dst_ip,
2673 		       RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2674 		attributes->l4.udp.hdr = (struct udp_hdr){
2675 			.src_port = input->flow.udp6_flow.src_port,
2676 			.dst_port = input->flow.udp6_flow.dst_port,
2677 		};
2678 		attributes->items[1] = (struct rte_flow_item){
2679 			.type = RTE_FLOW_ITEM_TYPE_IPV6,
2680 			.spec = &attributes->l3,
2681 		};
2682 		attributes->items[2] = (struct rte_flow_item){
2683 			.type = RTE_FLOW_ITEM_TYPE_UDP,
2684 			.spec = &attributes->l4,
2685 		};
2686 		break;
2687 	case RTE_ETH_FLOW_NONFRAG_IPV6_TCP:
2688 		attributes->l3.ipv6.hdr = (struct ipv6_hdr){
2689 			.hop_limits = input->flow.tcp6_flow.ip.hop_limits,
2690 			.proto = input->flow.tcp6_flow.ip.proto,
2691 		};
2692 		memcpy(attributes->l3.ipv6.hdr.src_addr,
2693 		       input->flow.tcp6_flow.ip.src_ip,
2694 		       RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2695 		memcpy(attributes->l3.ipv6.hdr.dst_addr,
2696 		       input->flow.tcp6_flow.ip.dst_ip,
2697 		       RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2698 		attributes->l4.tcp.hdr = (struct tcp_hdr){
2699 			.src_port = input->flow.tcp6_flow.src_port,
2700 			.dst_port = input->flow.tcp6_flow.dst_port,
2701 		};
2702 		attributes->items[1] = (struct rte_flow_item){
2703 			.type = RTE_FLOW_ITEM_TYPE_IPV6,
2704 			.spec = &attributes->l3,
2705 		};
2706 		attributes->items[2] = (struct rte_flow_item){
2707 			.type = RTE_FLOW_ITEM_TYPE_UDP,
2708 			.spec = &attributes->l4,
2709 		};
2710 		break;
2711 	case RTE_ETH_FLOW_NONFRAG_IPV6_OTHER:
2712 		attributes->l3.ipv6.hdr = (struct ipv6_hdr){
2713 			.hop_limits = input->flow.ipv6_flow.hop_limits,
2714 			.proto = input->flow.ipv6_flow.proto,
2715 		};
2716 		memcpy(attributes->l3.ipv6.hdr.src_addr,
2717 		       input->flow.ipv6_flow.src_ip,
2718 		       RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2719 		memcpy(attributes->l3.ipv6.hdr.dst_addr,
2720 		       input->flow.ipv6_flow.dst_ip,
2721 		       RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2722 		attributes->items[1] = (struct rte_flow_item){
2723 			.type = RTE_FLOW_ITEM_TYPE_IPV6,
2724 			.spec = &attributes->l3,
2725 		};
2726 		break;
2727 	default:
2728 		ERROR("invalid flow type%d",
2729 		      fdir_filter->input.flow_type);
2730 		return ENOTSUP;
2731 	}
2732 	return 0;
2733 }
2734 
2735 /**
2736  * Add new flow director filter and store it in list.
2737  *
2738  * @param priv
2739  *   Private structure.
2740  * @param fdir_filter
2741  *   Flow director filter to add.
2742  *
2743  * @return
2744  *   0 on success, errno value on failure.
2745  */
2746 static int
2747 priv_fdir_filter_add(struct priv *priv,
2748 		     const struct rte_eth_fdir_filter *fdir_filter)
2749 {
2750 	struct mlx5_fdir attributes = {
2751 		.attr.group = 0,
2752 	};
2753 	struct mlx5_flow_parse parser = {
2754 		.layer = HASH_RXQ_ETH,
2755 	};
2756 	struct rte_flow_error error;
2757 	struct rte_flow *flow;
2758 	int ret;
2759 
2760 	ret = priv_fdir_filter_convert(priv, fdir_filter, &attributes);
2761 	if (ret)
2762 		return -ret;
2763 	ret = priv_flow_convert(priv, &attributes.attr, attributes.items,
2764 				attributes.actions, &error, &parser);
2765 	if (ret)
2766 		return -ret;
2767 	flow = priv_flow_create(priv,
2768 				&priv->flows,
2769 				&attributes.attr,
2770 				attributes.items,
2771 				attributes.actions,
2772 				&error);
2773 	if (flow) {
2774 		TAILQ_INSERT_TAIL(&priv->flows, flow, next);
2775 		DEBUG("FDIR created %p", (void *)flow);
2776 		return 0;
2777 	}
2778 	return ENOTSUP;
2779 }
2780 
2781 /**
2782  * Delete specific filter.
2783  *
2784  * @param priv
2785  *   Private structure.
2786  * @param fdir_filter
2787  *   Filter to be deleted.
2788  *
2789  * @return
2790  *   0 on success, errno value on failure.
2791  */
2792 static int
2793 priv_fdir_filter_delete(struct priv *priv,
2794 			const struct rte_eth_fdir_filter *fdir_filter)
2795 {
2796 	struct mlx5_fdir attributes;
2797 	struct mlx5_flow_parse parser = {
2798 		.create = 1,
2799 		.layer = HASH_RXQ_ETH,
2800 	};
2801 	struct rte_flow_error error;
2802 	struct rte_flow *flow;
2803 	unsigned int i;
2804 	int ret;
2805 
2806 	ret = priv_fdir_filter_convert(priv, fdir_filter, &attributes);
2807 	if (ret)
2808 		return -ret;
2809 	ret = priv_flow_convert(priv, &attributes.attr, attributes.items,
2810 				attributes.actions, &error, &parser);
2811 	if (ret)
2812 		goto exit;
2813 	TAILQ_FOREACH(flow, &priv->flows, next) {
2814 		struct ibv_flow_attr *attr;
2815 		struct ibv_spec_header *attr_h;
2816 		void *spec;
2817 		struct ibv_flow_attr *flow_attr;
2818 		struct ibv_spec_header *flow_h;
2819 		void *flow_spec;
2820 		unsigned int specs_n;
2821 
2822 		if (parser.drop)
2823 			attr = parser.drop_q.ibv_attr;
2824 		else
2825 			attr = parser.queue[HASH_RXQ_ETH].ibv_attr;
2826 		if (flow->drop)
2827 			flow_attr = flow->drxq.ibv_attr;
2828 		else
2829 			flow_attr = flow->frxq[HASH_RXQ_ETH].ibv_attr;
2830 		/* Compare first the attributes. */
2831 		if (memcmp(attr, flow_attr, sizeof(struct ibv_flow_attr)))
2832 			continue;
2833 		if (attr->num_of_specs == 0)
2834 			continue;
2835 		spec = (void *)((uintptr_t)attr +
2836 				sizeof(struct ibv_flow_attr));
2837 		flow_spec = (void *)((uintptr_t)flow_attr +
2838 				     sizeof(struct ibv_flow_attr));
2839 		specs_n = RTE_MIN(attr->num_of_specs, flow_attr->num_of_specs);
2840 		for (i = 0; i != specs_n; ++i) {
2841 			attr_h = spec;
2842 			flow_h = flow_spec;
2843 			if (memcmp(spec, flow_spec,
2844 				   RTE_MIN(attr_h->size, flow_h->size)))
2845 				continue;
2846 			spec = (void *)((uintptr_t)attr + attr_h->size);
2847 			flow_spec = (void *)((uintptr_t)flow_attr +
2848 					     flow_h->size);
2849 		}
2850 		/* At this point, the flow match. */
2851 		break;
2852 	}
2853 	if (flow)
2854 		priv_flow_destroy(priv, &priv->flows, flow);
2855 exit:
2856 	if (parser.drop) {
2857 		rte_free(parser.drop_q.ibv_attr);
2858 	} else {
2859 		for (i = 0; i != hash_rxq_init_n; ++i) {
2860 			if (parser.queue[i].ibv_attr)
2861 				rte_free(parser.queue[i].ibv_attr);
2862 		}
2863 	}
2864 	return -ret;
2865 }
2866 
2867 /**
2868  * Update queue for specific filter.
2869  *
2870  * @param priv
2871  *   Private structure.
2872  * @param fdir_filter
2873  *   Filter to be updated.
2874  *
2875  * @return
2876  *   0 on success, errno value on failure.
2877  */
2878 static int
2879 priv_fdir_filter_update(struct priv *priv,
2880 			const struct rte_eth_fdir_filter *fdir_filter)
2881 {
2882 	int ret;
2883 
2884 	ret = priv_fdir_filter_delete(priv, fdir_filter);
2885 	if (ret)
2886 		return ret;
2887 	ret = priv_fdir_filter_add(priv, fdir_filter);
2888 	return ret;
2889 }
2890 
2891 /**
2892  * Flush all filters.
2893  *
2894  * @param priv
2895  *   Private structure.
2896  */
2897 static void
2898 priv_fdir_filter_flush(struct priv *priv)
2899 {
2900 	priv_flow_flush(priv, &priv->flows);
2901 }
2902 
2903 /**
2904  * Get flow director information.
2905  *
2906  * @param priv
2907  *   Private structure.
2908  * @param[out] fdir_info
2909  *   Resulting flow director information.
2910  */
2911 static void
2912 priv_fdir_info_get(struct priv *priv, struct rte_eth_fdir_info *fdir_info)
2913 {
2914 	struct rte_eth_fdir_masks *mask =
2915 		&priv->dev->data->dev_conf.fdir_conf.mask;
2916 
2917 	fdir_info->mode = priv->dev->data->dev_conf.fdir_conf.mode;
2918 	fdir_info->guarant_spc = 0;
2919 	rte_memcpy(&fdir_info->mask, mask, sizeof(fdir_info->mask));
2920 	fdir_info->max_flexpayload = 0;
2921 	fdir_info->flow_types_mask[0] = 0;
2922 	fdir_info->flex_payload_unit = 0;
2923 	fdir_info->max_flex_payload_segment_num = 0;
2924 	fdir_info->flex_payload_limit = 0;
2925 	memset(&fdir_info->flex_conf, 0, sizeof(fdir_info->flex_conf));
2926 }
2927 
2928 /**
2929  * Deal with flow director operations.
2930  *
2931  * @param priv
2932  *   Pointer to private structure.
2933  * @param filter_op
2934  *   Operation to perform.
2935  * @param arg
2936  *   Pointer to operation-specific structure.
2937  *
2938  * @return
2939  *   0 on success, errno value on failure.
2940  */
2941 static int
2942 priv_fdir_ctrl_func(struct priv *priv, enum rte_filter_op filter_op, void *arg)
2943 {
2944 	enum rte_fdir_mode fdir_mode =
2945 		priv->dev->data->dev_conf.fdir_conf.mode;
2946 	int ret = 0;
2947 
2948 	if (filter_op == RTE_ETH_FILTER_NOP)
2949 		return 0;
2950 	if (fdir_mode != RTE_FDIR_MODE_PERFECT &&
2951 	    fdir_mode != RTE_FDIR_MODE_PERFECT_MAC_VLAN) {
2952 		ERROR("%p: flow director mode %d not supported",
2953 		      (void *)priv, fdir_mode);
2954 		return EINVAL;
2955 	}
2956 	switch (filter_op) {
2957 	case RTE_ETH_FILTER_ADD:
2958 		ret = priv_fdir_filter_add(priv, arg);
2959 		break;
2960 	case RTE_ETH_FILTER_UPDATE:
2961 		ret = priv_fdir_filter_update(priv, arg);
2962 		break;
2963 	case RTE_ETH_FILTER_DELETE:
2964 		ret = priv_fdir_filter_delete(priv, arg);
2965 		break;
2966 	case RTE_ETH_FILTER_FLUSH:
2967 		priv_fdir_filter_flush(priv);
2968 		break;
2969 	case RTE_ETH_FILTER_INFO:
2970 		priv_fdir_info_get(priv, arg);
2971 		break;
2972 	default:
2973 		DEBUG("%p: unknown operation %u", (void *)priv,
2974 		      filter_op);
2975 		ret = EINVAL;
2976 		break;
2977 	}
2978 	return ret;
2979 }
2980 
2981 /**
2982  * Manage filter operations.
2983  *
2984  * @param dev
2985  *   Pointer to Ethernet device structure.
2986  * @param filter_type
2987  *   Filter type.
2988  * @param filter_op
2989  *   Operation to perform.
2990  * @param arg
2991  *   Pointer to operation-specific structure.
2992  *
2993  * @return
2994  *   0 on success, negative errno value on failure.
2995  */
2996 int
2997 mlx5_dev_filter_ctrl(struct rte_eth_dev *dev,
2998 		     enum rte_filter_type filter_type,
2999 		     enum rte_filter_op filter_op,
3000 		     void *arg)
3001 {
3002 	int ret = EINVAL;
3003 	struct priv *priv = dev->data->dev_private;
3004 
3005 	switch (filter_type) {
3006 	case RTE_ETH_FILTER_GENERIC:
3007 		if (filter_op != RTE_ETH_FILTER_GET)
3008 			return -EINVAL;
3009 		*(const void **)arg = &mlx5_flow_ops;
3010 		return 0;
3011 	case RTE_ETH_FILTER_FDIR:
3012 		priv_lock(priv);
3013 		ret = priv_fdir_ctrl_func(priv, filter_op, arg);
3014 		priv_unlock(priv);
3015 		break;
3016 	default:
3017 		ERROR("%p: filter type (%d) not supported",
3018 		      (void *)dev, filter_type);
3019 		break;
3020 	}
3021 	return -ret;
3022 }
3023