xref: /dpdk/drivers/net/mlx5/mlx5_flow.c (revision bb44fb6fe7713ddcd023d5b9bacadf074d68092e)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2016 6WIND S.A.
3  * Copyright 2016 Mellanox Technologies, Ltd
4  */
5 
6 #include <sys/queue.h>
7 #include <stdint.h>
8 #include <string.h>
9 
10 /* Verbs header. */
11 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
12 #ifdef PEDANTIC
13 #pragma GCC diagnostic ignored "-Wpedantic"
14 #endif
15 #include <infiniband/verbs.h>
16 #ifdef PEDANTIC
17 #pragma GCC diagnostic error "-Wpedantic"
18 #endif
19 
20 #include <rte_common.h>
21 #include <rte_ether.h>
22 #include <rte_eth_ctrl.h>
23 #include <rte_ethdev_driver.h>
24 #include <rte_flow.h>
25 #include <rte_flow_driver.h>
26 #include <rte_malloc.h>
27 #include <rte_ip.h>
28 
29 #include "mlx5.h"
30 #include "mlx5_defs.h"
31 #include "mlx5_prm.h"
32 #include "mlx5_glue.h"
33 
34 /* Dev ops structure defined in mlx5.c */
35 extern const struct eth_dev_ops mlx5_dev_ops;
36 extern const struct eth_dev_ops mlx5_dev_ops_isolate;
37 
38 /* Pattern outer Layer bits. */
39 #define MLX5_FLOW_LAYER_OUTER_L2 (1u << 0)
40 #define MLX5_FLOW_LAYER_OUTER_L3_IPV4 (1u << 1)
41 #define MLX5_FLOW_LAYER_OUTER_L3_IPV6 (1u << 2)
42 #define MLX5_FLOW_LAYER_OUTER_L4_UDP (1u << 3)
43 #define MLX5_FLOW_LAYER_OUTER_L4_TCP (1u << 4)
44 #define MLX5_FLOW_LAYER_OUTER_VLAN (1u << 5)
45 
46 /* Pattern inner Layer bits. */
47 #define MLX5_FLOW_LAYER_INNER_L2 (1u << 6)
48 #define MLX5_FLOW_LAYER_INNER_L3_IPV4 (1u << 7)
49 #define MLX5_FLOW_LAYER_INNER_L3_IPV6 (1u << 8)
50 #define MLX5_FLOW_LAYER_INNER_L4_UDP (1u << 9)
51 #define MLX5_FLOW_LAYER_INNER_L4_TCP (1u << 10)
52 #define MLX5_FLOW_LAYER_INNER_VLAN (1u << 11)
53 
54 /* Pattern tunnel Layer bits. */
55 #define MLX5_FLOW_LAYER_VXLAN (1u << 12)
56 #define MLX5_FLOW_LAYER_VXLAN_GPE (1u << 13)
57 #define MLX5_FLOW_LAYER_GRE (1u << 14)
58 #define MLX5_FLOW_LAYER_MPLS (1u << 15)
59 
60 /* Outer Masks. */
61 #define MLX5_FLOW_LAYER_OUTER_L3 \
62 	(MLX5_FLOW_LAYER_OUTER_L3_IPV4 | MLX5_FLOW_LAYER_OUTER_L3_IPV6)
63 #define MLX5_FLOW_LAYER_OUTER_L4 \
64 	(MLX5_FLOW_LAYER_OUTER_L4_UDP | MLX5_FLOW_LAYER_OUTER_L4_TCP)
65 #define MLX5_FLOW_LAYER_OUTER \
66 	(MLX5_FLOW_LAYER_OUTER_L2 | MLX5_FLOW_LAYER_OUTER_L3 | \
67 	 MLX5_FLOW_LAYER_OUTER_L4)
68 
69 /* Tunnel Masks. */
70 #define MLX5_FLOW_LAYER_TUNNEL \
71 	(MLX5_FLOW_LAYER_VXLAN | MLX5_FLOW_LAYER_VXLAN_GPE | \
72 	 MLX5_FLOW_LAYER_GRE | MLX5_FLOW_LAYER_MPLS)
73 
74 /* Inner Masks. */
75 #define MLX5_FLOW_LAYER_INNER_L3 \
76 	(MLX5_FLOW_LAYER_INNER_L3_IPV4 | MLX5_FLOW_LAYER_INNER_L3_IPV6)
77 #define MLX5_FLOW_LAYER_INNER_L4 \
78 	(MLX5_FLOW_LAYER_INNER_L4_UDP | MLX5_FLOW_LAYER_INNER_L4_TCP)
79 #define MLX5_FLOW_LAYER_INNER \
80 	(MLX5_FLOW_LAYER_INNER_L2 | MLX5_FLOW_LAYER_INNER_L3 | \
81 	 MLX5_FLOW_LAYER_INNER_L4)
82 
83 /* Actions that modify the fate of matching traffic. */
84 #define MLX5_FLOW_FATE_DROP (1u << 0)
85 #define MLX5_FLOW_FATE_QUEUE (1u << 1)
86 #define MLX5_FLOW_FATE_RSS (1u << 2)
87 
88 /* Modify a packet. */
89 #define MLX5_FLOW_MOD_FLAG (1u << 0)
90 #define MLX5_FLOW_MOD_MARK (1u << 1)
91 #define MLX5_FLOW_MOD_COUNT (1u << 2)
92 
93 /* possible L3 layers protocols filtering. */
94 #define MLX5_IP_PROTOCOL_TCP 6
95 #define MLX5_IP_PROTOCOL_UDP 17
96 #define MLX5_IP_PROTOCOL_GRE 47
97 #define MLX5_IP_PROTOCOL_MPLS 147
98 
99 /* Priority reserved for default flows. */
100 #define MLX5_FLOW_PRIO_RSVD ((uint32_t)-1)
101 
102 enum mlx5_expansion {
103 	MLX5_EXPANSION_ROOT,
104 	MLX5_EXPANSION_ROOT_OUTER,
105 	MLX5_EXPANSION_OUTER_ETH,
106 	MLX5_EXPANSION_OUTER_IPV4,
107 	MLX5_EXPANSION_OUTER_IPV4_UDP,
108 	MLX5_EXPANSION_OUTER_IPV4_TCP,
109 	MLX5_EXPANSION_OUTER_IPV6,
110 	MLX5_EXPANSION_OUTER_IPV6_UDP,
111 	MLX5_EXPANSION_OUTER_IPV6_TCP,
112 	MLX5_EXPANSION_VXLAN,
113 	MLX5_EXPANSION_VXLAN_GPE,
114 	MLX5_EXPANSION_GRE,
115 	MLX5_EXPANSION_MPLS,
116 	MLX5_EXPANSION_ETH,
117 	MLX5_EXPANSION_IPV4,
118 	MLX5_EXPANSION_IPV4_UDP,
119 	MLX5_EXPANSION_IPV4_TCP,
120 	MLX5_EXPANSION_IPV6,
121 	MLX5_EXPANSION_IPV6_UDP,
122 	MLX5_EXPANSION_IPV6_TCP,
123 };
124 
125 /** Supported expansion of items. */
126 static const struct rte_flow_expand_node mlx5_support_expansion[] = {
127 	[MLX5_EXPANSION_ROOT] = {
128 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_ETH,
129 						 MLX5_EXPANSION_IPV4,
130 						 MLX5_EXPANSION_IPV6),
131 		.type = RTE_FLOW_ITEM_TYPE_END,
132 	},
133 	[MLX5_EXPANSION_ROOT_OUTER] = {
134 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_OUTER_ETH,
135 						 MLX5_EXPANSION_OUTER_IPV4,
136 						 MLX5_EXPANSION_OUTER_IPV6),
137 		.type = RTE_FLOW_ITEM_TYPE_END,
138 	},
139 	[MLX5_EXPANSION_OUTER_ETH] = {
140 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_OUTER_IPV4,
141 						 MLX5_EXPANSION_OUTER_IPV6,
142 						 MLX5_EXPANSION_MPLS),
143 		.type = RTE_FLOW_ITEM_TYPE_ETH,
144 		.rss_types = 0,
145 	},
146 	[MLX5_EXPANSION_OUTER_IPV4] = {
147 		.next = RTE_FLOW_EXPAND_RSS_NEXT
148 			(MLX5_EXPANSION_OUTER_IPV4_UDP,
149 			 MLX5_EXPANSION_OUTER_IPV4_TCP,
150 			 MLX5_EXPANSION_GRE),
151 		.type = RTE_FLOW_ITEM_TYPE_IPV4,
152 		.rss_types = ETH_RSS_IPV4 | ETH_RSS_FRAG_IPV4 |
153 			ETH_RSS_NONFRAG_IPV4_OTHER,
154 	},
155 	[MLX5_EXPANSION_OUTER_IPV4_UDP] = {
156 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_VXLAN,
157 						 MLX5_EXPANSION_VXLAN_GPE),
158 		.type = RTE_FLOW_ITEM_TYPE_UDP,
159 		.rss_types = ETH_RSS_NONFRAG_IPV4_UDP,
160 	},
161 	[MLX5_EXPANSION_OUTER_IPV4_TCP] = {
162 		.type = RTE_FLOW_ITEM_TYPE_TCP,
163 		.rss_types = ETH_RSS_NONFRAG_IPV4_TCP,
164 	},
165 	[MLX5_EXPANSION_OUTER_IPV6] = {
166 		.next = RTE_FLOW_EXPAND_RSS_NEXT
167 			(MLX5_EXPANSION_OUTER_IPV6_UDP,
168 			 MLX5_EXPANSION_OUTER_IPV6_TCP),
169 		.type = RTE_FLOW_ITEM_TYPE_IPV6,
170 		.rss_types = ETH_RSS_IPV6 | ETH_RSS_FRAG_IPV6 |
171 			ETH_RSS_NONFRAG_IPV6_OTHER,
172 	},
173 	[MLX5_EXPANSION_OUTER_IPV6_UDP] = {
174 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_VXLAN,
175 						 MLX5_EXPANSION_VXLAN_GPE),
176 		.type = RTE_FLOW_ITEM_TYPE_UDP,
177 		.rss_types = ETH_RSS_NONFRAG_IPV6_UDP,
178 	},
179 	[MLX5_EXPANSION_OUTER_IPV6_TCP] = {
180 		.type = RTE_FLOW_ITEM_TYPE_TCP,
181 		.rss_types = ETH_RSS_NONFRAG_IPV6_TCP,
182 	},
183 	[MLX5_EXPANSION_VXLAN] = {
184 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_ETH),
185 		.type = RTE_FLOW_ITEM_TYPE_VXLAN,
186 	},
187 	[MLX5_EXPANSION_VXLAN_GPE] = {
188 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_ETH,
189 						 MLX5_EXPANSION_IPV4,
190 						 MLX5_EXPANSION_IPV6),
191 		.type = RTE_FLOW_ITEM_TYPE_VXLAN_GPE,
192 	},
193 	[MLX5_EXPANSION_GRE] = {
194 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4),
195 		.type = RTE_FLOW_ITEM_TYPE_GRE,
196 	},
197 	[MLX5_EXPANSION_MPLS] = {
198 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4,
199 						 MLX5_EXPANSION_IPV6),
200 		.type = RTE_FLOW_ITEM_TYPE_MPLS,
201 	},
202 	[MLX5_EXPANSION_ETH] = {
203 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4,
204 						 MLX5_EXPANSION_IPV6),
205 		.type = RTE_FLOW_ITEM_TYPE_ETH,
206 	},
207 	[MLX5_EXPANSION_IPV4] = {
208 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4_UDP,
209 						 MLX5_EXPANSION_IPV4_TCP),
210 		.type = RTE_FLOW_ITEM_TYPE_IPV4,
211 		.rss_types = ETH_RSS_IPV4 | ETH_RSS_FRAG_IPV4 |
212 			ETH_RSS_NONFRAG_IPV4_OTHER,
213 	},
214 	[MLX5_EXPANSION_IPV4_UDP] = {
215 		.type = RTE_FLOW_ITEM_TYPE_UDP,
216 		.rss_types = ETH_RSS_NONFRAG_IPV4_UDP,
217 	},
218 	[MLX5_EXPANSION_IPV4_TCP] = {
219 		.type = RTE_FLOW_ITEM_TYPE_TCP,
220 		.rss_types = ETH_RSS_NONFRAG_IPV4_TCP,
221 	},
222 	[MLX5_EXPANSION_IPV6] = {
223 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV6_UDP,
224 						 MLX5_EXPANSION_IPV6_TCP),
225 		.type = RTE_FLOW_ITEM_TYPE_IPV6,
226 		.rss_types = ETH_RSS_IPV6 | ETH_RSS_FRAG_IPV6 |
227 			ETH_RSS_NONFRAG_IPV6_OTHER,
228 	},
229 	[MLX5_EXPANSION_IPV6_UDP] = {
230 		.type = RTE_FLOW_ITEM_TYPE_UDP,
231 		.rss_types = ETH_RSS_NONFRAG_IPV6_UDP,
232 	},
233 	[MLX5_EXPANSION_IPV6_TCP] = {
234 		.type = RTE_FLOW_ITEM_TYPE_TCP,
235 		.rss_types = ETH_RSS_NONFRAG_IPV6_TCP,
236 	},
237 };
238 
239 /** Handles information leading to a drop fate. */
240 struct mlx5_flow_verbs {
241 	LIST_ENTRY(mlx5_flow_verbs) next;
242 	unsigned int size; /**< Size of the attribute. */
243 	struct {
244 		struct ibv_flow_attr *attr;
245 		/**< Pointer to the Specification buffer. */
246 		uint8_t *specs; /**< Pointer to the specifications. */
247 	};
248 	struct ibv_flow *flow; /**< Verbs flow pointer. */
249 	struct mlx5_hrxq *hrxq; /**< Hash Rx queue object. */
250 	uint64_t hash_fields; /**< Verbs hash Rx queue hash fields. */
251 };
252 
253 /* Counters information. */
254 struct mlx5_flow_counter {
255 	LIST_ENTRY(mlx5_flow_counter) next; /**< Pointer to the next counter. */
256 	uint32_t shared:1; /**< Share counter ID with other flow rules. */
257 	uint32_t ref_cnt:31; /**< Reference counter. */
258 	uint32_t id; /**< Counter ID. */
259 	struct ibv_counter_set *cs; /**< Holds the counters for the rule. */
260 	uint64_t hits; /**< Number of packets matched by the rule. */
261 	uint64_t bytes; /**< Number of bytes matched by the rule. */
262 };
263 
264 /* Flow structure. */
265 struct rte_flow {
266 	TAILQ_ENTRY(rte_flow) next; /**< Pointer to the next flow structure. */
267 	struct rte_flow_attr attributes; /**< User flow attribute. */
268 	uint32_t l3_protocol_en:1; /**< Protocol filtering requested. */
269 	uint32_t layers;
270 	/**< Bit-fields of present layers see MLX5_FLOW_LAYER_*. */
271 	uint32_t modifier;
272 	/**< Bit-fields of present modifier see MLX5_FLOW_MOD_*. */
273 	uint32_t fate;
274 	/**< Bit-fields of present fate see MLX5_FLOW_FATE_*. */
275 	uint8_t l3_protocol; /**< valid when l3_protocol_en is set. */
276 	LIST_HEAD(verbs, mlx5_flow_verbs) verbs; /**< Verbs flows list. */
277 	struct mlx5_flow_verbs *cur_verbs;
278 	/**< Current Verbs flow structure being filled. */
279 	struct mlx5_flow_counter *counter; /**< Holds Verbs flow counter. */
280 	struct rte_flow_action_rss rss;/**< RSS context. */
281 	uint8_t key[MLX5_RSS_HASH_KEY_LEN]; /**< RSS hash key. */
282 	uint16_t (*queue)[]; /**< Destination queues to redirect traffic to. */
283 };
284 
285 static const struct rte_flow_ops mlx5_flow_ops = {
286 	.validate = mlx5_flow_validate,
287 	.create = mlx5_flow_create,
288 	.destroy = mlx5_flow_destroy,
289 	.flush = mlx5_flow_flush,
290 	.isolate = mlx5_flow_isolate,
291 	.query = mlx5_flow_query,
292 };
293 
294 /* Convert FDIR request to Generic flow. */
295 struct mlx5_fdir {
296 	struct rte_flow_attr attr;
297 	struct rte_flow_action actions[2];
298 	struct rte_flow_item items[4];
299 	struct rte_flow_item_eth l2;
300 	struct rte_flow_item_eth l2_mask;
301 	union {
302 		struct rte_flow_item_ipv4 ipv4;
303 		struct rte_flow_item_ipv6 ipv6;
304 	} l3;
305 	union {
306 		struct rte_flow_item_ipv4 ipv4;
307 		struct rte_flow_item_ipv6 ipv6;
308 	} l3_mask;
309 	union {
310 		struct rte_flow_item_udp udp;
311 		struct rte_flow_item_tcp tcp;
312 	} l4;
313 	union {
314 		struct rte_flow_item_udp udp;
315 		struct rte_flow_item_tcp tcp;
316 	} l4_mask;
317 	struct rte_flow_action_queue queue;
318 };
319 
320 /* Verbs specification header. */
321 struct ibv_spec_header {
322 	enum ibv_flow_spec_type type;
323 	uint16_t size;
324 };
325 
326 /*
327  * Number of sub priorities.
328  * For each kind of pattern matching i.e. L2, L3, L4 to have a correct
329  * matching on the NIC (firmware dependent) L4 most have the higher priority
330  * followed by L3 and ending with L2.
331  */
332 #define MLX5_PRIORITY_MAP_L2 2
333 #define MLX5_PRIORITY_MAP_L3 1
334 #define MLX5_PRIORITY_MAP_L4 0
335 #define MLX5_PRIORITY_MAP_MAX 3
336 
337 /* Map of Verbs to Flow priority with 8 Verbs priorities. */
338 static const uint32_t priority_map_3[][MLX5_PRIORITY_MAP_MAX] = {
339 	{ 0, 1, 2 }, { 2, 3, 4 }, { 5, 6, 7 },
340 };
341 
342 /* Map of Verbs to Flow priority with 16 Verbs priorities. */
343 static const uint32_t priority_map_5[][MLX5_PRIORITY_MAP_MAX] = {
344 	{ 0, 1, 2 }, { 3, 4, 5 }, { 6, 7, 8 },
345 	{ 9, 10, 11 }, { 12, 13, 14 },
346 };
347 
348 /* Tunnel information. */
349 struct mlx5_flow_tunnel_info {
350 	uint32_t tunnel; /**< Tunnel bit (see MLX5_FLOW_*). */
351 	uint32_t ptype; /**< Tunnel Ptype (see RTE_PTYPE_*). */
352 };
353 
354 static struct mlx5_flow_tunnel_info tunnels_info[] = {
355 	{
356 		.tunnel = MLX5_FLOW_LAYER_VXLAN,
357 		.ptype = RTE_PTYPE_TUNNEL_VXLAN | RTE_PTYPE_L4_UDP,
358 	},
359 	{
360 		.tunnel = MLX5_FLOW_LAYER_VXLAN_GPE,
361 		.ptype = RTE_PTYPE_TUNNEL_VXLAN_GPE | RTE_PTYPE_L4_UDP,
362 	},
363 	{
364 		.tunnel = MLX5_FLOW_LAYER_GRE,
365 		.ptype = RTE_PTYPE_TUNNEL_GRE,
366 	},
367 	{
368 		.tunnel = MLX5_FLOW_LAYER_MPLS | MLX5_FLOW_LAYER_OUTER_L4_UDP,
369 		.ptype = RTE_PTYPE_TUNNEL_MPLS_IN_GRE | RTE_PTYPE_L4_UDP,
370 	},
371 	{
372 		.tunnel = MLX5_FLOW_LAYER_MPLS,
373 		.ptype = RTE_PTYPE_TUNNEL_MPLS_IN_GRE,
374 	},
375 };
376 
377 /**
378  * Discover the maximum number of priority available.
379  *
380  * @param[in] dev
381  *   Pointer to Ethernet device.
382  *
383  * @return
384  *   number of supported flow priority on success, a negative errno
385  *   value otherwise and rte_errno is set.
386  */
387 int
388 mlx5_flow_discover_priorities(struct rte_eth_dev *dev)
389 {
390 	struct {
391 		struct ibv_flow_attr attr;
392 		struct ibv_flow_spec_eth eth;
393 		struct ibv_flow_spec_action_drop drop;
394 	} flow_attr = {
395 		.attr = {
396 			.num_of_specs = 2,
397 		},
398 		.eth = {
399 			.type = IBV_FLOW_SPEC_ETH,
400 			.size = sizeof(struct ibv_flow_spec_eth),
401 		},
402 		.drop = {
403 			.size = sizeof(struct ibv_flow_spec_action_drop),
404 			.type = IBV_FLOW_SPEC_ACTION_DROP,
405 		},
406 	};
407 	struct ibv_flow *flow;
408 	struct mlx5_hrxq *drop = mlx5_hrxq_drop_new(dev);
409 	uint16_t vprio[] = { 8, 16 };
410 	int i;
411 	int priority = 0;
412 
413 	if (!drop) {
414 		rte_errno = ENOTSUP;
415 		return -rte_errno;
416 	}
417 	for (i = 0; i != RTE_DIM(vprio); i++) {
418 		flow_attr.attr.priority = vprio[i] - 1;
419 		flow = mlx5_glue->create_flow(drop->qp, &flow_attr.attr);
420 		if (!flow)
421 			break;
422 		claim_zero(mlx5_glue->destroy_flow(flow));
423 		priority = vprio[i];
424 	}
425 	switch (priority) {
426 	case 8:
427 		priority = RTE_DIM(priority_map_3);
428 		break;
429 	case 16:
430 		priority = RTE_DIM(priority_map_5);
431 		break;
432 	default:
433 		rte_errno = ENOTSUP;
434 		DRV_LOG(ERR,
435 			"port %u verbs maximum priority: %d expected 8/16",
436 			dev->data->port_id, vprio[i]);
437 		return -rte_errno;
438 	}
439 	mlx5_hrxq_drop_release(dev);
440 	DRV_LOG(INFO, "port %u flow maximum priority: %d",
441 		dev->data->port_id, priority);
442 	return priority;
443 }
444 
445 /**
446  * Adjust flow priority.
447  *
448  * @param dev
449  *   Pointer to Ethernet device.
450  * @param flow
451  *   Pointer to an rte flow.
452  */
453 static void
454 mlx5_flow_adjust_priority(struct rte_eth_dev *dev, struct rte_flow *flow)
455 {
456 	struct priv *priv = dev->data->dev_private;
457 	uint32_t priority = flow->attributes.priority;
458 	uint32_t subpriority = flow->cur_verbs->attr->priority;
459 
460 	switch (priv->config.flow_prio) {
461 	case RTE_DIM(priority_map_3):
462 		priority = priority_map_3[priority][subpriority];
463 		break;
464 	case RTE_DIM(priority_map_5):
465 		priority = priority_map_5[priority][subpriority];
466 		break;
467 	}
468 	flow->cur_verbs->attr->priority = priority;
469 }
470 
471 /**
472  * Get a flow counter.
473  *
474  * @param[in] dev
475  *   Pointer to Ethernet device.
476  * @param[in] shared
477  *   Indicate if this counter is shared with other flows.
478  * @param[in] id
479  *   Counter identifier.
480  *
481  * @return
482  *   A pointer to the counter, NULL otherwise and rte_errno is set.
483  */
484 static struct mlx5_flow_counter *
485 mlx5_flow_counter_new(struct rte_eth_dev *dev, uint32_t shared, uint32_t id)
486 {
487 	struct priv *priv = dev->data->dev_private;
488 	struct mlx5_flow_counter *cnt;
489 
490 	LIST_FOREACH(cnt, &priv->flow_counters, next) {
491 		if (cnt->shared != shared)
492 			continue;
493 		if (cnt->id != id)
494 			continue;
495 		cnt->ref_cnt++;
496 		return cnt;
497 	}
498 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
499 
500 	struct mlx5_flow_counter tmpl = {
501 		.shared = shared,
502 		.id = id,
503 		.cs = mlx5_glue->create_counter_set
504 			(priv->ctx,
505 			 &(struct ibv_counter_set_init_attr){
506 				 .counter_set_id = id,
507 			 }),
508 		.hits = 0,
509 		.bytes = 0,
510 	};
511 
512 	if (!tmpl.cs) {
513 		rte_errno = errno;
514 		return NULL;
515 	}
516 	cnt = rte_calloc(__func__, 1, sizeof(*cnt), 0);
517 	if (!cnt) {
518 		rte_errno = ENOMEM;
519 		return NULL;
520 	}
521 	*cnt = tmpl;
522 	LIST_INSERT_HEAD(&priv->flow_counters, cnt, next);
523 	return cnt;
524 #endif
525 	rte_errno = ENOTSUP;
526 	return NULL;
527 }
528 
529 /**
530  * Release a flow counter.
531  *
532  * @param[in] counter
533  *   Pointer to the counter handler.
534  */
535 static void
536 mlx5_flow_counter_release(struct mlx5_flow_counter *counter)
537 {
538 	if (--counter->ref_cnt == 0) {
539 		claim_zero(mlx5_glue->destroy_counter_set(counter->cs));
540 		LIST_REMOVE(counter, next);
541 		rte_free(counter);
542 	}
543 }
544 
545 /**
546  * Verify the @p attributes will be correctly understood by the NIC and store
547  * them in the @p flow if everything is correct.
548  *
549  * @param[in] dev
550  *   Pointer to Ethernet device.
551  * @param[in] attributes
552  *   Pointer to flow attributes
553  * @param[in, out] flow
554  *   Pointer to the rte_flow structure.
555  * @param[out] error
556  *   Pointer to error structure.
557  *
558  * @return
559  *   0 on success, a negative errno value otherwise and rte_errno is set.
560  */
561 static int
562 mlx5_flow_attributes(struct rte_eth_dev *dev,
563 		     const struct rte_flow_attr *attributes,
564 		     struct rte_flow *flow,
565 		     struct rte_flow_error *error)
566 {
567 	uint32_t priority_max =
568 		((struct priv *)dev->data->dev_private)->config.flow_prio - 1;
569 
570 	if (attributes->group)
571 		return rte_flow_error_set(error, ENOTSUP,
572 					  RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
573 					  NULL,
574 					  "groups is not supported");
575 	if (attributes->priority != MLX5_FLOW_PRIO_RSVD &&
576 	    attributes->priority >= priority_max)
577 		return rte_flow_error_set(error, ENOTSUP,
578 					  RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
579 					  NULL,
580 					  "priority out of range");
581 	if (attributes->egress)
582 		return rte_flow_error_set(error, ENOTSUP,
583 					  RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
584 					  NULL,
585 					  "egress is not supported");
586 	if (attributes->transfer)
587 		return rte_flow_error_set(error, ENOTSUP,
588 					  RTE_FLOW_ERROR_TYPE_ATTR_TRANSFER,
589 					  NULL,
590 					  "transfer is not supported");
591 	if (!attributes->ingress)
592 		return rte_flow_error_set(error, ENOTSUP,
593 					  RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
594 					  NULL,
595 					  "ingress attribute is mandatory");
596 	flow->attributes = *attributes;
597 	if (attributes->priority == MLX5_FLOW_PRIO_RSVD)
598 		flow->attributes.priority = priority_max;
599 	return 0;
600 }
601 
602 /**
603  * Verify the @p item specifications (spec, last, mask) are compatible with the
604  * NIC capabilities.
605  *
606  * @param[in] item
607  *   Item specification.
608  * @param[in] mask
609  *   @p item->mask or flow default bit-masks.
610  * @param[in] nic_mask
611  *   Bit-masks covering supported fields by the NIC to compare with user mask.
612  * @param[in] size
613  *   Bit-masks size in bytes.
614  * @param[out] error
615  *   Pointer to error structure.
616  *
617  * @return
618  *   0 on success, a negative errno value otherwise and rte_errno is set.
619  */
620 static int
621 mlx5_flow_item_acceptable(const struct rte_flow_item *item,
622 			  const uint8_t *mask,
623 			  const uint8_t *nic_mask,
624 			  unsigned int size,
625 			  struct rte_flow_error *error)
626 {
627 	unsigned int i;
628 
629 	assert(nic_mask);
630 	for (i = 0; i < size; ++i)
631 		if ((nic_mask[i] | mask[i]) != nic_mask[i])
632 			return rte_flow_error_set(error, ENOTSUP,
633 						  RTE_FLOW_ERROR_TYPE_ITEM,
634 						  item,
635 						  "mask enables non supported"
636 						  " bits");
637 	if (!item->spec && (item->mask || item->last))
638 		return rte_flow_error_set(error, EINVAL,
639 					  RTE_FLOW_ERROR_TYPE_ITEM,
640 					  item,
641 					  "mask/last without a spec is not"
642 					  " supported");
643 	if (item->spec && item->last) {
644 		uint8_t spec[size];
645 		uint8_t last[size];
646 		unsigned int i;
647 		int ret;
648 
649 		for (i = 0; i < size; ++i) {
650 			spec[i] = ((const uint8_t *)item->spec)[i] & mask[i];
651 			last[i] = ((const uint8_t *)item->last)[i] & mask[i];
652 		}
653 		ret = memcmp(spec, last, size);
654 		if (ret != 0)
655 			return rte_flow_error_set(error, ENOTSUP,
656 						  RTE_FLOW_ERROR_TYPE_ITEM,
657 						  item,
658 						  "range is not supported");
659 	}
660 	return 0;
661 }
662 
663 /**
664  * Add a verbs item specification into @p flow.
665  *
666  * @param[in, out] flow
667  *   Pointer to flow structure.
668  * @param[in] src
669  *   Create specification.
670  * @param[in] size
671  *   Size in bytes of the specification to copy.
672  */
673 static void
674 mlx5_flow_spec_verbs_add(struct rte_flow *flow, void *src, unsigned int size)
675 {
676 	struct mlx5_flow_verbs *verbs = flow->cur_verbs;
677 
678 	if (verbs->specs) {
679 		void *dst;
680 
681 		dst = (void *)(verbs->specs + verbs->size);
682 		memcpy(dst, src, size);
683 		++verbs->attr->num_of_specs;
684 	}
685 	verbs->size += size;
686 }
687 
688 /**
689  * Adjust verbs hash fields according to the @p flow information.
690  *
691  * @param[in, out] flow.
692  *   Pointer to flow structure.
693  * @param[in] tunnel
694  *   1 when the hash field is for a tunnel item.
695  * @param[in] layer_types
696  *   ETH_RSS_* types.
697  * @param[in] hash_fields
698  *   Item hash fields.
699  */
700 static void
701 mlx5_flow_verbs_hashfields_adjust(struct rte_flow *flow,
702 				  int tunnel __rte_unused,
703 				  uint32_t layer_types, uint64_t hash_fields)
704 {
705 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
706 	hash_fields |= (tunnel ? IBV_RX_HASH_INNER : 0);
707 	if (flow->rss.level == 2 && !tunnel)
708 		hash_fields = 0;
709 	else if (flow->rss.level < 2 && tunnel)
710 		hash_fields = 0;
711 #endif
712 	if (!(flow->rss.types & layer_types))
713 		hash_fields = 0;
714 	flow->cur_verbs->hash_fields |= hash_fields;
715 }
716 
717 /**
718  * Convert the @p item into a Verbs specification after ensuring the NIC
719  * will understand and process it correctly.
720  * If the necessary size for the conversion is greater than the @p flow_size,
721  * nothing is written in @p flow, the validation is still performed.
722  *
723  * @param[in] item
724  *   Item specification.
725  * @param[in, out] flow
726  *   Pointer to flow structure.
727  * @param[in] flow_size
728  *   Size in bytes of the available space in @p flow, if too small, nothing is
729  *   written.
730  * @param[out] error
731  *   Pointer to error structure.
732  *
733  * @return
734  *   On success the number of bytes consumed/necessary, if the returned value
735  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
736  *   otherwise another call with this returned memory size should be done.
737  *   On error, a negative errno value is returned and rte_errno is set.
738  */
739 static int
740 mlx5_flow_item_eth(const struct rte_flow_item *item, struct rte_flow *flow,
741 		   const size_t flow_size, struct rte_flow_error *error)
742 {
743 	const struct rte_flow_item_eth *spec = item->spec;
744 	const struct rte_flow_item_eth *mask = item->mask;
745 	const struct rte_flow_item_eth nic_mask = {
746 		.dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
747 		.src.addr_bytes = "\xff\xff\xff\xff\xff\xff",
748 		.type = RTE_BE16(0xffff),
749 	};
750 	const int tunnel = !!(flow->layers & MLX5_FLOW_LAYER_TUNNEL);
751 	const unsigned int size = sizeof(struct ibv_flow_spec_eth);
752 	struct ibv_flow_spec_eth eth = {
753 		.type = IBV_FLOW_SPEC_ETH | (tunnel ? IBV_FLOW_SPEC_INNER : 0),
754 		.size = size,
755 	};
756 	int ret;
757 
758 	if (flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
759 			    MLX5_FLOW_LAYER_OUTER_L2))
760 		return rte_flow_error_set(error, ENOTSUP,
761 					  RTE_FLOW_ERROR_TYPE_ITEM,
762 					  item,
763 					  "L2 layers already configured");
764 	if (!mask)
765 		mask = &rte_flow_item_eth_mask;
766 	ret = mlx5_flow_item_acceptable(item, (const uint8_t *)mask,
767 					(const uint8_t *)&nic_mask,
768 					sizeof(struct rte_flow_item_eth),
769 					error);
770 	if (ret)
771 		return ret;
772 	flow->layers |= tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
773 		MLX5_FLOW_LAYER_OUTER_L2;
774 	if (size > flow_size)
775 		return size;
776 	if (spec) {
777 		unsigned int i;
778 
779 		memcpy(&eth.val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
780 		memcpy(&eth.val.src_mac, spec->src.addr_bytes, ETHER_ADDR_LEN);
781 		eth.val.ether_type = spec->type;
782 		memcpy(&eth.mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
783 		memcpy(&eth.mask.src_mac, mask->src.addr_bytes, ETHER_ADDR_LEN);
784 		eth.mask.ether_type = mask->type;
785 		/* Remove unwanted bits from values. */
786 		for (i = 0; i < ETHER_ADDR_LEN; ++i) {
787 			eth.val.dst_mac[i] &= eth.mask.dst_mac[i];
788 			eth.val.src_mac[i] &= eth.mask.src_mac[i];
789 		}
790 		eth.val.ether_type &= eth.mask.ether_type;
791 	}
792 	flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L2;
793 	mlx5_flow_spec_verbs_add(flow, &eth, size);
794 	return size;
795 }
796 
797 /**
798  * Update the VLAN tag in the Verbs Ethernet specification.
799  *
800  * @param[in, out] attr
801  *   Pointer to Verbs attributes structure.
802  * @param[in] eth
803  *   Verbs structure containing the VLAN information to copy.
804  */
805 static void
806 mlx5_flow_item_vlan_update(struct ibv_flow_attr *attr,
807 			   struct ibv_flow_spec_eth *eth)
808 {
809 	unsigned int i;
810 	const enum ibv_flow_spec_type search = eth->type;
811 	struct ibv_spec_header *hdr = (struct ibv_spec_header *)
812 		((uint8_t *)attr + sizeof(struct ibv_flow_attr));
813 
814 	for (i = 0; i != attr->num_of_specs; ++i) {
815 		if (hdr->type == search) {
816 			struct ibv_flow_spec_eth *e =
817 				(struct ibv_flow_spec_eth *)hdr;
818 
819 			e->val.vlan_tag = eth->val.vlan_tag;
820 			e->mask.vlan_tag = eth->mask.vlan_tag;
821 			e->val.ether_type = eth->val.ether_type;
822 			e->mask.ether_type = eth->mask.ether_type;
823 			break;
824 		}
825 		hdr = (struct ibv_spec_header *)((uint8_t *)hdr + hdr->size);
826 	}
827 }
828 
829 /**
830  * Convert the @p item into @p flow (or by updating the already present
831  * Ethernet Verbs) specification after ensuring the NIC will understand and
832  * process it correctly.
833  * If the necessary size for the conversion is greater than the @p flow_size,
834  * nothing is written in @p flow, the validation is still performed.
835  *
836  * @param[in] item
837  *   Item specification.
838  * @param[in, out] flow
839  *   Pointer to flow structure.
840  * @param[in] flow_size
841  *   Size in bytes of the available space in @p flow, if too small, nothing is
842  *   written.
843  * @param[out] error
844  *   Pointer to error structure.
845  *
846  * @return
847  *   On success the number of bytes consumed/necessary, if the returned value
848  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
849  *   otherwise another call with this returned memory size should be done.
850  *   On error, a negative errno value is returned and rte_errno is set.
851  */
852 static int
853 mlx5_flow_item_vlan(const struct rte_flow_item *item, struct rte_flow *flow,
854 		    const size_t flow_size, struct rte_flow_error *error)
855 {
856 	const struct rte_flow_item_vlan *spec = item->spec;
857 	const struct rte_flow_item_vlan *mask = item->mask;
858 	const struct rte_flow_item_vlan nic_mask = {
859 		.tci = RTE_BE16(0x0fff),
860 		.inner_type = RTE_BE16(0xffff),
861 	};
862 	unsigned int size = sizeof(struct ibv_flow_spec_eth);
863 	const int tunnel = !!(flow->layers & MLX5_FLOW_LAYER_TUNNEL);
864 	struct ibv_flow_spec_eth eth = {
865 		.type = IBV_FLOW_SPEC_ETH | (tunnel ? IBV_FLOW_SPEC_INNER : 0),
866 		.size = size,
867 	};
868 	int ret;
869 	const uint32_t l34m = tunnel ? (MLX5_FLOW_LAYER_INNER_L3 |
870 					MLX5_FLOW_LAYER_INNER_L4) :
871 		(MLX5_FLOW_LAYER_OUTER_L3 | MLX5_FLOW_LAYER_OUTER_L4);
872 	const uint32_t vlanm = tunnel ? MLX5_FLOW_LAYER_INNER_VLAN :
873 		MLX5_FLOW_LAYER_OUTER_VLAN;
874 	const uint32_t l2m = tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
875 		MLX5_FLOW_LAYER_OUTER_L2;
876 
877 	if (flow->layers & vlanm)
878 		return rte_flow_error_set(error, ENOTSUP,
879 					  RTE_FLOW_ERROR_TYPE_ITEM,
880 					  item,
881 					  "VLAN layer already configured");
882 	else if ((flow->layers & l34m) != 0)
883 		return rte_flow_error_set(error, ENOTSUP,
884 					  RTE_FLOW_ERROR_TYPE_ITEM,
885 					  item,
886 					  "L2 layer cannot follow L3/L4 layer");
887 	if (!mask)
888 		mask = &rte_flow_item_vlan_mask;
889 	ret = mlx5_flow_item_acceptable
890 		(item, (const uint8_t *)mask,
891 		 (const uint8_t *)&nic_mask,
892 		 sizeof(struct rte_flow_item_vlan), error);
893 	if (ret)
894 		return ret;
895 	if (spec) {
896 		eth.val.vlan_tag = spec->tci;
897 		eth.mask.vlan_tag = mask->tci;
898 		eth.val.vlan_tag &= eth.mask.vlan_tag;
899 		eth.val.ether_type = spec->inner_type;
900 		eth.mask.ether_type = mask->inner_type;
901 		eth.val.ether_type &= eth.mask.ether_type;
902 	}
903 	/*
904 	 * From verbs perspective an empty VLAN is equivalent
905 	 * to a packet without VLAN layer.
906 	 */
907 	if (!eth.mask.vlan_tag)
908 		return rte_flow_error_set(error, EINVAL,
909 					  RTE_FLOW_ERROR_TYPE_ITEM_SPEC,
910 					  item->spec,
911 					  "VLAN cannot be empty");
912 	if (!(flow->layers & l2m)) {
913 		if (size <= flow_size) {
914 			flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L2;
915 			mlx5_flow_spec_verbs_add(flow, &eth, size);
916 		}
917 	} else {
918 		if (flow->cur_verbs)
919 			mlx5_flow_item_vlan_update(flow->cur_verbs->attr,
920 						   &eth);
921 		size = 0; /* Only an update is done in eth specification. */
922 	}
923 	flow->layers |= tunnel ?
924 		(MLX5_FLOW_LAYER_INNER_L2 | MLX5_FLOW_LAYER_INNER_VLAN) :
925 		(MLX5_FLOW_LAYER_OUTER_L2 | MLX5_FLOW_LAYER_OUTER_VLAN);
926 	return size;
927 }
928 
929 /**
930  * Convert the @p item into a Verbs specification after ensuring the NIC
931  * will understand and process it correctly.
932  * If the necessary size for the conversion is greater than the @p flow_size,
933  * nothing is written in @p flow, the validation is still performed.
934  *
935  * @param[in] item
936  *   Item specification.
937  * @param[in, out] flow
938  *   Pointer to flow structure.
939  * @param[in] flow_size
940  *   Size in bytes of the available space in @p flow, if too small, nothing is
941  *   written.
942  * @param[out] error
943  *   Pointer to error structure.
944  *
945  * @return
946  *   On success the number of bytes consumed/necessary, if the returned value
947  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
948  *   otherwise another call with this returned memory size should be done.
949  *   On error, a negative errno value is returned and rte_errno is set.
950  */
951 static int
952 mlx5_flow_item_ipv4(const struct rte_flow_item *item, struct rte_flow *flow,
953 		    const size_t flow_size, struct rte_flow_error *error)
954 {
955 	const struct rte_flow_item_ipv4 *spec = item->spec;
956 	const struct rte_flow_item_ipv4 *mask = item->mask;
957 	const struct rte_flow_item_ipv4 nic_mask = {
958 		.hdr = {
959 			.src_addr = RTE_BE32(0xffffffff),
960 			.dst_addr = RTE_BE32(0xffffffff),
961 			.type_of_service = 0xff,
962 			.next_proto_id = 0xff,
963 		},
964 	};
965 	const int tunnel = !!(flow->layers & MLX5_FLOW_LAYER_TUNNEL);
966 	unsigned int size = sizeof(struct ibv_flow_spec_ipv4_ext);
967 	struct ibv_flow_spec_ipv4_ext ipv4 = {
968 		.type = IBV_FLOW_SPEC_IPV4_EXT |
969 			(tunnel ? IBV_FLOW_SPEC_INNER : 0),
970 		.size = size,
971 	};
972 	int ret;
973 
974 	if (flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L3 :
975 			    MLX5_FLOW_LAYER_OUTER_L3))
976 		return rte_flow_error_set(error, ENOTSUP,
977 					  RTE_FLOW_ERROR_TYPE_ITEM,
978 					  item,
979 					  "multiple L3 layers not supported");
980 	else if (flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L4 :
981 				 MLX5_FLOW_LAYER_OUTER_L4))
982 		return rte_flow_error_set(error, ENOTSUP,
983 					  RTE_FLOW_ERROR_TYPE_ITEM,
984 					  item,
985 					  "L3 cannot follow an L4 layer.");
986 	if (!mask)
987 		mask = &rte_flow_item_ipv4_mask;
988 	ret = mlx5_flow_item_acceptable
989 		(item, (const uint8_t *)mask,
990 		 (const uint8_t *)&nic_mask,
991 		 sizeof(struct rte_flow_item_ipv4), error);
992 	if (ret < 0)
993 		return ret;
994 	flow->layers |= tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV4 :
995 		MLX5_FLOW_LAYER_OUTER_L3_IPV4;
996 	if (spec) {
997 		ipv4.val = (struct ibv_flow_ipv4_ext_filter){
998 			.src_ip = spec->hdr.src_addr,
999 			.dst_ip = spec->hdr.dst_addr,
1000 			.proto = spec->hdr.next_proto_id,
1001 			.tos = spec->hdr.type_of_service,
1002 		};
1003 		ipv4.mask = (struct ibv_flow_ipv4_ext_filter){
1004 			.src_ip = mask->hdr.src_addr,
1005 			.dst_ip = mask->hdr.dst_addr,
1006 			.proto = mask->hdr.next_proto_id,
1007 			.tos = mask->hdr.type_of_service,
1008 		};
1009 		/* Remove unwanted bits from values. */
1010 		ipv4.val.src_ip &= ipv4.mask.src_ip;
1011 		ipv4.val.dst_ip &= ipv4.mask.dst_ip;
1012 		ipv4.val.proto &= ipv4.mask.proto;
1013 		ipv4.val.tos &= ipv4.mask.tos;
1014 	}
1015 	flow->l3_protocol_en = !!ipv4.mask.proto;
1016 	flow->l3_protocol = ipv4.val.proto;
1017 	if (size <= flow_size) {
1018 		mlx5_flow_verbs_hashfields_adjust
1019 			(flow, tunnel,
1020 			 (ETH_RSS_IPV4 | ETH_RSS_FRAG_IPV4 |
1021 			  ETH_RSS_NONFRAG_IPV4_OTHER),
1022 			 (IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4));
1023 		flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L3;
1024 		mlx5_flow_spec_verbs_add(flow, &ipv4, size);
1025 	}
1026 	return size;
1027 }
1028 
1029 /**
1030  * Convert the @p item into a Verbs specification after ensuring the NIC
1031  * will understand and process it correctly.
1032  * If the necessary size for the conversion is greater than the @p flow_size,
1033  * nothing is written in @p flow, the validation is still performed.
1034  *
1035  * @param[in] item
1036  *   Item specification.
1037  * @param[in, out] flow
1038  *   Pointer to flow structure.
1039  * @param[in] flow_size
1040  *   Size in bytes of the available space in @p flow, if too small, nothing is
1041  *   written.
1042  * @param[out] error
1043  *   Pointer to error structure.
1044  *
1045  * @return
1046  *   On success the number of bytes consumed/necessary, if the returned value
1047  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
1048  *   otherwise another call with this returned memory size should be done.
1049  *   On error, a negative errno value is returned and rte_errno is set.
1050  */
1051 static int
1052 mlx5_flow_item_ipv6(const struct rte_flow_item *item, struct rte_flow *flow,
1053 		    const size_t flow_size, struct rte_flow_error *error)
1054 {
1055 	const struct rte_flow_item_ipv6 *spec = item->spec;
1056 	const struct rte_flow_item_ipv6 *mask = item->mask;
1057 	const struct rte_flow_item_ipv6 nic_mask = {
1058 		.hdr = {
1059 			.src_addr =
1060 				"\xff\xff\xff\xff\xff\xff\xff\xff"
1061 				"\xff\xff\xff\xff\xff\xff\xff\xff",
1062 			.dst_addr =
1063 				"\xff\xff\xff\xff\xff\xff\xff\xff"
1064 				"\xff\xff\xff\xff\xff\xff\xff\xff",
1065 			.vtc_flow = RTE_BE32(0xffffffff),
1066 			.proto = 0xff,
1067 			.hop_limits = 0xff,
1068 		},
1069 	};
1070 	const int tunnel = !!(flow->layers & MLX5_FLOW_LAYER_TUNNEL);
1071 	unsigned int size = sizeof(struct ibv_flow_spec_ipv6);
1072 	struct ibv_flow_spec_ipv6 ipv6 = {
1073 		.type = IBV_FLOW_SPEC_IPV6 | (tunnel ? IBV_FLOW_SPEC_INNER : 0),
1074 		.size = size,
1075 	};
1076 	int ret;
1077 
1078 	if (flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L3 :
1079 			    MLX5_FLOW_LAYER_OUTER_L3))
1080 		return rte_flow_error_set(error, ENOTSUP,
1081 					  RTE_FLOW_ERROR_TYPE_ITEM,
1082 					  item,
1083 					  "multiple L3 layers not supported");
1084 	else if (flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L4 :
1085 				 MLX5_FLOW_LAYER_OUTER_L4))
1086 		return rte_flow_error_set(error, ENOTSUP,
1087 					  RTE_FLOW_ERROR_TYPE_ITEM,
1088 					  item,
1089 					  "L3 cannot follow an L4 layer.");
1090 	/*
1091 	 * IPv6 is not recognised by the NIC inside a GRE tunnel.
1092 	 * Such support has to be disabled as the rule will be
1093 	 * accepted.  Issue reproduced with Mellanox OFED 4.3-3.0.2.1 and
1094 	 * Mellanox OFED 4.4-1.0.0.0.
1095 	 */
1096 	if (tunnel && flow->layers & MLX5_FLOW_LAYER_GRE)
1097 		return rte_flow_error_set(error, ENOTSUP,
1098 					  RTE_FLOW_ERROR_TYPE_ITEM,
1099 					  item,
1100 					  "IPv6 inside a GRE tunnel is"
1101 					  " not recognised.");
1102 	if (!mask)
1103 		mask = &rte_flow_item_ipv6_mask;
1104 	ret = mlx5_flow_item_acceptable
1105 		(item, (const uint8_t *)mask,
1106 		 (const uint8_t *)&nic_mask,
1107 		 sizeof(struct rte_flow_item_ipv6), error);
1108 	if (ret < 0)
1109 		return ret;
1110 	flow->layers |= tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV6 :
1111 		MLX5_FLOW_LAYER_OUTER_L3_IPV6;
1112 	if (spec) {
1113 		unsigned int i;
1114 		uint32_t vtc_flow_val;
1115 		uint32_t vtc_flow_mask;
1116 
1117 		memcpy(&ipv6.val.src_ip, spec->hdr.src_addr,
1118 		       RTE_DIM(ipv6.val.src_ip));
1119 		memcpy(&ipv6.val.dst_ip, spec->hdr.dst_addr,
1120 		       RTE_DIM(ipv6.val.dst_ip));
1121 		memcpy(&ipv6.mask.src_ip, mask->hdr.src_addr,
1122 		       RTE_DIM(ipv6.mask.src_ip));
1123 		memcpy(&ipv6.mask.dst_ip, mask->hdr.dst_addr,
1124 		       RTE_DIM(ipv6.mask.dst_ip));
1125 		vtc_flow_val = rte_be_to_cpu_32(spec->hdr.vtc_flow);
1126 		vtc_flow_mask = rte_be_to_cpu_32(mask->hdr.vtc_flow);
1127 		ipv6.val.flow_label =
1128 			rte_cpu_to_be_32((vtc_flow_val & IPV6_HDR_FL_MASK) >>
1129 					 IPV6_HDR_FL_SHIFT);
1130 		ipv6.val.traffic_class = (vtc_flow_val & IPV6_HDR_TC_MASK) >>
1131 					 IPV6_HDR_TC_SHIFT;
1132 		ipv6.val.next_hdr = spec->hdr.proto;
1133 		ipv6.val.hop_limit = spec->hdr.hop_limits;
1134 		ipv6.mask.flow_label =
1135 			rte_cpu_to_be_32((vtc_flow_mask & IPV6_HDR_FL_MASK) >>
1136 					 IPV6_HDR_FL_SHIFT);
1137 		ipv6.mask.traffic_class = (vtc_flow_mask & IPV6_HDR_TC_MASK) >>
1138 					  IPV6_HDR_TC_SHIFT;
1139 		ipv6.mask.next_hdr = mask->hdr.proto;
1140 		ipv6.mask.hop_limit = mask->hdr.hop_limits;
1141 		/* Remove unwanted bits from values. */
1142 		for (i = 0; i < RTE_DIM(ipv6.val.src_ip); ++i) {
1143 			ipv6.val.src_ip[i] &= ipv6.mask.src_ip[i];
1144 			ipv6.val.dst_ip[i] &= ipv6.mask.dst_ip[i];
1145 		}
1146 		ipv6.val.flow_label &= ipv6.mask.flow_label;
1147 		ipv6.val.traffic_class &= ipv6.mask.traffic_class;
1148 		ipv6.val.next_hdr &= ipv6.mask.next_hdr;
1149 		ipv6.val.hop_limit &= ipv6.mask.hop_limit;
1150 	}
1151 	flow->l3_protocol_en = !!ipv6.mask.next_hdr;
1152 	flow->l3_protocol = ipv6.val.next_hdr;
1153 	if (size <= flow_size) {
1154 		mlx5_flow_verbs_hashfields_adjust
1155 			(flow, tunnel,
1156 			 (ETH_RSS_IPV6 | ETH_RSS_NONFRAG_IPV6_OTHER),
1157 			 (IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6));
1158 		flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L3;
1159 		mlx5_flow_spec_verbs_add(flow, &ipv6, size);
1160 	}
1161 	return size;
1162 }
1163 
1164 /**
1165  * Convert the @p item into a Verbs specification after ensuring the NIC
1166  * will understand and process it correctly.
1167  * If the necessary size for the conversion is greater than the @p flow_size,
1168  * nothing is written in @p flow, the validation is still performed.
1169  *
1170  * @param[in] item
1171  *   Item specification.
1172  * @param[in, out] flow
1173  *   Pointer to flow structure.
1174  * @param[in] flow_size
1175  *   Size in bytes of the available space in @p flow, if too small, nothing is
1176  *   written.
1177  * @param[out] error
1178  *   Pointer to error structure.
1179  *
1180  * @return
1181  *   On success the number of bytes consumed/necessary, if the returned value
1182  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
1183  *   otherwise another call with this returned memory size should be done.
1184  *   On error, a negative errno value is returned and rte_errno is set.
1185  */
1186 static int
1187 mlx5_flow_item_udp(const struct rte_flow_item *item, struct rte_flow *flow,
1188 		   const size_t flow_size, struct rte_flow_error *error)
1189 {
1190 	const struct rte_flow_item_udp *spec = item->spec;
1191 	const struct rte_flow_item_udp *mask = item->mask;
1192 	const int tunnel = !!(flow->layers & MLX5_FLOW_LAYER_TUNNEL);
1193 	unsigned int size = sizeof(struct ibv_flow_spec_tcp_udp);
1194 	struct ibv_flow_spec_tcp_udp udp = {
1195 		.type = IBV_FLOW_SPEC_UDP | (tunnel ? IBV_FLOW_SPEC_INNER : 0),
1196 		.size = size,
1197 	};
1198 	int ret;
1199 
1200 	if (flow->l3_protocol_en && flow->l3_protocol != MLX5_IP_PROTOCOL_UDP)
1201 		return rte_flow_error_set(error, ENOTSUP,
1202 					  RTE_FLOW_ERROR_TYPE_ITEM,
1203 					  item,
1204 					  "protocol filtering not compatible"
1205 					  " with UDP layer");
1206 	if (!(flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L3 :
1207 			      MLX5_FLOW_LAYER_OUTER_L3)))
1208 		return rte_flow_error_set(error, ENOTSUP,
1209 					  RTE_FLOW_ERROR_TYPE_ITEM,
1210 					  item,
1211 					  "L3 is mandatory to filter"
1212 					  " on L4");
1213 	if (flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L4 :
1214 			    MLX5_FLOW_LAYER_OUTER_L4))
1215 		return rte_flow_error_set(error, ENOTSUP,
1216 					  RTE_FLOW_ERROR_TYPE_ITEM,
1217 					  item,
1218 					  "L4 layer is already"
1219 					  " present");
1220 	if (!mask)
1221 		mask = &rte_flow_item_udp_mask;
1222 	ret = mlx5_flow_item_acceptable
1223 		(item, (const uint8_t *)mask,
1224 		 (const uint8_t *)&rte_flow_item_udp_mask,
1225 		 sizeof(struct rte_flow_item_udp), error);
1226 	if (ret < 0)
1227 		return ret;
1228 	flow->layers |= tunnel ? MLX5_FLOW_LAYER_INNER_L4_UDP :
1229 		MLX5_FLOW_LAYER_OUTER_L4_UDP;
1230 	if (spec) {
1231 		udp.val.dst_port = spec->hdr.dst_port;
1232 		udp.val.src_port = spec->hdr.src_port;
1233 		udp.mask.dst_port = mask->hdr.dst_port;
1234 		udp.mask.src_port = mask->hdr.src_port;
1235 		/* Remove unwanted bits from values. */
1236 		udp.val.src_port &= udp.mask.src_port;
1237 		udp.val.dst_port &= udp.mask.dst_port;
1238 	}
1239 	if (size <= flow_size) {
1240 		mlx5_flow_verbs_hashfields_adjust(flow, tunnel, ETH_RSS_UDP,
1241 						  (IBV_RX_HASH_SRC_PORT_UDP |
1242 						   IBV_RX_HASH_DST_PORT_UDP));
1243 		flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L4;
1244 		mlx5_flow_spec_verbs_add(flow, &udp, size);
1245 	}
1246 	return size;
1247 }
1248 
1249 /**
1250  * Convert the @p item into a Verbs specification after ensuring the NIC
1251  * will understand and process it correctly.
1252  * If the necessary size for the conversion is greater than the @p flow_size,
1253  * nothing is written in @p flow, the validation is still performed.
1254  *
1255  * @param[in] item
1256  *   Item specification.
1257  * @param[in, out] flow
1258  *   Pointer to flow structure.
1259  * @param[in] flow_size
1260  *   Size in bytes of the available space in @p flow, if too small, nothing is
1261  *   written.
1262  * @param[out] error
1263  *   Pointer to error structure.
1264  *
1265  * @return
1266  *   On success the number of bytes consumed/necessary, if the returned value
1267  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
1268  *   otherwise another call with this returned memory size should be done.
1269  *   On error, a negative errno value is returned and rte_errno is set.
1270  */
1271 static int
1272 mlx5_flow_item_tcp(const struct rte_flow_item *item, struct rte_flow *flow,
1273 		   const size_t flow_size, struct rte_flow_error *error)
1274 {
1275 	const struct rte_flow_item_tcp *spec = item->spec;
1276 	const struct rte_flow_item_tcp *mask = item->mask;
1277 	const int tunnel = !!(flow->layers & MLX5_FLOW_LAYER_TUNNEL);
1278 	unsigned int size = sizeof(struct ibv_flow_spec_tcp_udp);
1279 	struct ibv_flow_spec_tcp_udp tcp = {
1280 		.type = IBV_FLOW_SPEC_TCP | (tunnel ? IBV_FLOW_SPEC_INNER : 0),
1281 		.size = size,
1282 	};
1283 	int ret;
1284 
1285 	if (flow->l3_protocol_en && flow->l3_protocol != MLX5_IP_PROTOCOL_TCP)
1286 		return rte_flow_error_set(error, ENOTSUP,
1287 					  RTE_FLOW_ERROR_TYPE_ITEM,
1288 					  item,
1289 					  "protocol filtering not compatible"
1290 					  " with TCP layer");
1291 	if (!(flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L3 :
1292 			      MLX5_FLOW_LAYER_OUTER_L3)))
1293 		return rte_flow_error_set(error, ENOTSUP,
1294 					  RTE_FLOW_ERROR_TYPE_ITEM,
1295 					  item,
1296 					  "L3 is mandatory to filter on L4");
1297 	if (flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L4 :
1298 			    MLX5_FLOW_LAYER_OUTER_L4))
1299 		return rte_flow_error_set(error, ENOTSUP,
1300 					  RTE_FLOW_ERROR_TYPE_ITEM,
1301 					  item,
1302 					  "L4 layer is already present");
1303 	if (!mask)
1304 		mask = &rte_flow_item_tcp_mask;
1305 	ret = mlx5_flow_item_acceptable
1306 		(item, (const uint8_t *)mask,
1307 		 (const uint8_t *)&rte_flow_item_tcp_mask,
1308 		 sizeof(struct rte_flow_item_tcp), error);
1309 	if (ret < 0)
1310 		return ret;
1311 	flow->layers |=  tunnel ? MLX5_FLOW_LAYER_INNER_L4_TCP :
1312 		MLX5_FLOW_LAYER_OUTER_L4_TCP;
1313 	if (spec) {
1314 		tcp.val.dst_port = spec->hdr.dst_port;
1315 		tcp.val.src_port = spec->hdr.src_port;
1316 		tcp.mask.dst_port = mask->hdr.dst_port;
1317 		tcp.mask.src_port = mask->hdr.src_port;
1318 		/* Remove unwanted bits from values. */
1319 		tcp.val.src_port &= tcp.mask.src_port;
1320 		tcp.val.dst_port &= tcp.mask.dst_port;
1321 	}
1322 	if (size <= flow_size) {
1323 		mlx5_flow_verbs_hashfields_adjust(flow, tunnel, ETH_RSS_TCP,
1324 						  (IBV_RX_HASH_SRC_PORT_TCP |
1325 						   IBV_RX_HASH_DST_PORT_TCP));
1326 		flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L4;
1327 		mlx5_flow_spec_verbs_add(flow, &tcp, size);
1328 	}
1329 	return size;
1330 }
1331 
1332 /**
1333  * Convert the @p item into a Verbs specification after ensuring the NIC
1334  * will understand and process it correctly.
1335  * If the necessary size for the conversion is greater than the @p flow_size,
1336  * nothing is written in @p flow, the validation is still performed.
1337  *
1338  * @param[in] item
1339  *   Item specification.
1340  * @param[in, out] flow
1341  *   Pointer to flow structure.
1342  * @param[in] flow_size
1343  *   Size in bytes of the available space in @p flow, if too small, nothing is
1344  *   written.
1345  * @param[out] error
1346  *   Pointer to error structure.
1347  *
1348  * @return
1349  *   On success the number of bytes consumed/necessary, if the returned value
1350  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
1351  *   otherwise another call with this returned memory size should be done.
1352  *   On error, a negative errno value is returned and rte_errno is set.
1353  */
1354 static int
1355 mlx5_flow_item_vxlan(const struct rte_flow_item *item, struct rte_flow *flow,
1356 		     const size_t flow_size, struct rte_flow_error *error)
1357 {
1358 	const struct rte_flow_item_vxlan *spec = item->spec;
1359 	const struct rte_flow_item_vxlan *mask = item->mask;
1360 	unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
1361 	struct ibv_flow_spec_tunnel vxlan = {
1362 		.type = IBV_FLOW_SPEC_VXLAN_TUNNEL,
1363 		.size = size,
1364 	};
1365 	int ret;
1366 	union vni {
1367 		uint32_t vlan_id;
1368 		uint8_t vni[4];
1369 	} id = { .vlan_id = 0, };
1370 
1371 	if (flow->layers & MLX5_FLOW_LAYER_TUNNEL)
1372 		return rte_flow_error_set(error, ENOTSUP,
1373 					  RTE_FLOW_ERROR_TYPE_ITEM,
1374 					  item,
1375 					  "a tunnel is already present");
1376 	/*
1377 	 * Verify only UDPv4 is present as defined in
1378 	 * https://tools.ietf.org/html/rfc7348
1379 	 */
1380 	if (!(flow->layers & MLX5_FLOW_LAYER_OUTER_L4_UDP))
1381 		return rte_flow_error_set(error, ENOTSUP,
1382 					  RTE_FLOW_ERROR_TYPE_ITEM,
1383 					  item,
1384 					  "no outer UDP layer found");
1385 	if (!mask)
1386 		mask = &rte_flow_item_vxlan_mask;
1387 	ret = mlx5_flow_item_acceptable
1388 		(item, (const uint8_t *)mask,
1389 		 (const uint8_t *)&rte_flow_item_vxlan_mask,
1390 		 sizeof(struct rte_flow_item_vxlan), error);
1391 	if (ret < 0)
1392 		return ret;
1393 	if (spec) {
1394 		memcpy(&id.vni[1], spec->vni, 3);
1395 		vxlan.val.tunnel_id = id.vlan_id;
1396 		memcpy(&id.vni[1], mask->vni, 3);
1397 		vxlan.mask.tunnel_id = id.vlan_id;
1398 		/* Remove unwanted bits from values. */
1399 		vxlan.val.tunnel_id &= vxlan.mask.tunnel_id;
1400 	}
1401 	/*
1402 	 * Tunnel id 0 is equivalent as not adding a VXLAN layer, if
1403 	 * only this layer is defined in the Verbs specification it is
1404 	 * interpreted as wildcard and all packets will match this
1405 	 * rule, if it follows a full stack layer (ex: eth / ipv4 /
1406 	 * udp), all packets matching the layers before will also
1407 	 * match this rule.  To avoid such situation, VNI 0 is
1408 	 * currently refused.
1409 	 */
1410 	if (!vxlan.val.tunnel_id)
1411 		return rte_flow_error_set(error, EINVAL,
1412 					  RTE_FLOW_ERROR_TYPE_ITEM,
1413 					  item,
1414 					  "VXLAN vni cannot be 0");
1415 	if (!(flow->layers & MLX5_FLOW_LAYER_OUTER))
1416 		return rte_flow_error_set(error, EINVAL,
1417 					  RTE_FLOW_ERROR_TYPE_ITEM,
1418 					  item,
1419 					  "VXLAN tunnel must be fully defined");
1420 	if (size <= flow_size) {
1421 		mlx5_flow_spec_verbs_add(flow, &vxlan, size);
1422 		flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L2;
1423 	}
1424 	flow->layers |= MLX5_FLOW_LAYER_VXLAN;
1425 	return size;
1426 }
1427 
1428 /**
1429  * Convert the @p item into a Verbs specification after ensuring the NIC
1430  * will understand and process it correctly.
1431  * If the necessary size for the conversion is greater than the @p flow_size,
1432  * nothing is written in @p flow, the validation is still performed.
1433  *
1434  * @param dev
1435  *   Pointer to Ethernet device.
1436  * @param[in] item
1437  *   Item specification.
1438  * @param[in, out] flow
1439  *   Pointer to flow structure.
1440  * @param[in] flow_size
1441  *   Size in bytes of the available space in @p flow, if too small, nothing is
1442  *   written.
1443  * @param[out] error
1444  *   Pointer to error structure.
1445  *
1446  * @return
1447  *   On success the number of bytes consumed/necessary, if the returned value
1448  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
1449  *   otherwise another call with this returned memory size should be done.
1450  *   On error, a negative errno value is returned and rte_errno is set.
1451  */
1452 static int
1453 mlx5_flow_item_vxlan_gpe(struct rte_eth_dev *dev,
1454 			 const struct rte_flow_item *item,
1455 			 struct rte_flow *flow, const size_t flow_size,
1456 			 struct rte_flow_error *error)
1457 {
1458 	const struct rte_flow_item_vxlan_gpe *spec = item->spec;
1459 	const struct rte_flow_item_vxlan_gpe *mask = item->mask;
1460 	unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
1461 	struct ibv_flow_spec_tunnel vxlan_gpe = {
1462 		.type = IBV_FLOW_SPEC_VXLAN_TUNNEL,
1463 		.size = size,
1464 	};
1465 	int ret;
1466 	union vni {
1467 		uint32_t vlan_id;
1468 		uint8_t vni[4];
1469 	} id = { .vlan_id = 0, };
1470 
1471 	if (!((struct priv *)dev->data->dev_private)->config.l3_vxlan_en)
1472 		return rte_flow_error_set(error, ENOTSUP,
1473 					  RTE_FLOW_ERROR_TYPE_ITEM,
1474 					  item,
1475 					  "L3 VXLAN is not enabled by device"
1476 					  " parameter and/or not configured in"
1477 					  " firmware");
1478 	if (flow->layers & MLX5_FLOW_LAYER_TUNNEL)
1479 		return rte_flow_error_set(error, ENOTSUP,
1480 					  RTE_FLOW_ERROR_TYPE_ITEM,
1481 					  item,
1482 					  "a tunnel is already present");
1483 	/*
1484 	 * Verify only UDPv4 is present as defined in
1485 	 * https://tools.ietf.org/html/rfc7348
1486 	 */
1487 	if (!(flow->layers & MLX5_FLOW_LAYER_OUTER_L4_UDP))
1488 		return rte_flow_error_set(error, ENOTSUP,
1489 					  RTE_FLOW_ERROR_TYPE_ITEM,
1490 					  item,
1491 					  "no outer UDP layer found");
1492 	if (!mask)
1493 		mask = &rte_flow_item_vxlan_gpe_mask;
1494 	ret = mlx5_flow_item_acceptable
1495 		(item, (const uint8_t *)mask,
1496 		 (const uint8_t *)&rte_flow_item_vxlan_gpe_mask,
1497 		 sizeof(struct rte_flow_item_vxlan_gpe), error);
1498 	if (ret < 0)
1499 		return ret;
1500 	if (spec) {
1501 		memcpy(&id.vni[1], spec->vni, 3);
1502 		vxlan_gpe.val.tunnel_id = id.vlan_id;
1503 		memcpy(&id.vni[1], mask->vni, 3);
1504 		vxlan_gpe.mask.tunnel_id = id.vlan_id;
1505 		if (spec->protocol)
1506 			return rte_flow_error_set
1507 				(error, EINVAL,
1508 				 RTE_FLOW_ERROR_TYPE_ITEM,
1509 				 item,
1510 				 "VxLAN-GPE protocol not supported");
1511 		/* Remove unwanted bits from values. */
1512 		vxlan_gpe.val.tunnel_id &= vxlan_gpe.mask.tunnel_id;
1513 	}
1514 	/*
1515 	 * Tunnel id 0 is equivalent as not adding a VXLAN layer, if only this
1516 	 * layer is defined in the Verbs specification it is interpreted as
1517 	 * wildcard and all packets will match this rule, if it follows a full
1518 	 * stack layer (ex: eth / ipv4 / udp), all packets matching the layers
1519 	 * before will also match this rule.  To avoid such situation, VNI 0
1520 	 * is currently refused.
1521 	 */
1522 	if (!vxlan_gpe.val.tunnel_id)
1523 		return rte_flow_error_set(error, EINVAL,
1524 					  RTE_FLOW_ERROR_TYPE_ITEM,
1525 					  item,
1526 					  "VXLAN-GPE vni cannot be 0");
1527 	if (!(flow->layers & MLX5_FLOW_LAYER_OUTER))
1528 		return rte_flow_error_set(error, EINVAL,
1529 					  RTE_FLOW_ERROR_TYPE_ITEM,
1530 					  item,
1531 					  "VXLAN-GPE tunnel must be fully"
1532 					  " defined");
1533 	if (size <= flow_size) {
1534 		mlx5_flow_spec_verbs_add(flow, &vxlan_gpe, size);
1535 		flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L2;
1536 	}
1537 	flow->layers |= MLX5_FLOW_LAYER_VXLAN_GPE;
1538 	return size;
1539 }
1540 
1541 /**
1542  * Update the protocol in Verbs IPv4/IPv6 spec.
1543  *
1544  * @param[in, out] attr
1545  *   Pointer to Verbs attributes structure.
1546  * @param[in] search
1547  *   Specification type to search in order to update the IP protocol.
1548  * @param[in] protocol
1549  *   Protocol value to set if none is present in the specification.
1550  */
1551 static void
1552 mlx5_flow_item_gre_ip_protocol_update(struct ibv_flow_attr *attr,
1553 				      enum ibv_flow_spec_type search,
1554 				      uint8_t protocol)
1555 {
1556 	unsigned int i;
1557 	struct ibv_spec_header *hdr = (struct ibv_spec_header *)
1558 		((uint8_t *)attr + sizeof(struct ibv_flow_attr));
1559 
1560 	if (!attr)
1561 		return;
1562 	for (i = 0; i != attr->num_of_specs; ++i) {
1563 		if (hdr->type == search) {
1564 			union {
1565 				struct ibv_flow_spec_ipv4_ext *ipv4;
1566 				struct ibv_flow_spec_ipv6 *ipv6;
1567 			} ip;
1568 
1569 			switch (search) {
1570 			case IBV_FLOW_SPEC_IPV4_EXT:
1571 				ip.ipv4 = (struct ibv_flow_spec_ipv4_ext *)hdr;
1572 				if (!ip.ipv4->val.proto) {
1573 					ip.ipv4->val.proto = protocol;
1574 					ip.ipv4->mask.proto = 0xff;
1575 				}
1576 				break;
1577 			case IBV_FLOW_SPEC_IPV6:
1578 				ip.ipv6 = (struct ibv_flow_spec_ipv6 *)hdr;
1579 				if (!ip.ipv6->val.next_hdr) {
1580 					ip.ipv6->val.next_hdr = protocol;
1581 					ip.ipv6->mask.next_hdr = 0xff;
1582 				}
1583 				break;
1584 			default:
1585 				break;
1586 			}
1587 			break;
1588 		}
1589 		hdr = (struct ibv_spec_header *)((uint8_t *)hdr + hdr->size);
1590 	}
1591 }
1592 
1593 /**
1594  * Convert the @p item into a Verbs specification after ensuring the NIC
1595  * will understand and process it correctly.
1596  * It will also update the previous L3 layer with the protocol value matching
1597  * the GRE.
1598  * If the necessary size for the conversion is greater than the @p flow_size,
1599  * nothing is written in @p flow, the validation is still performed.
1600  *
1601  * @param dev
1602  *   Pointer to Ethernet device.
1603  * @param[in] item
1604  *   Item specification.
1605  * @param[in, out] flow
1606  *   Pointer to flow structure.
1607  * @param[in] flow_size
1608  *   Size in bytes of the available space in @p flow, if too small, nothing is
1609  *   written.
1610  * @param[out] error
1611  *   Pointer to error structure.
1612  *
1613  * @return
1614  *   On success the number of bytes consumed/necessary, if the returned value
1615  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
1616  *   otherwise another call with this returned memory size should be done.
1617  *   On error, a negative errno value is returned and rte_errno is set.
1618  */
1619 static int
1620 mlx5_flow_item_gre(const struct rte_flow_item *item,
1621 		   struct rte_flow *flow, const size_t flow_size,
1622 		   struct rte_flow_error *error)
1623 {
1624 	struct mlx5_flow_verbs *verbs = flow->cur_verbs;
1625 	const struct rte_flow_item_gre *spec = item->spec;
1626 	const struct rte_flow_item_gre *mask = item->mask;
1627 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1628 	unsigned int size = sizeof(struct ibv_flow_spec_gre);
1629 	struct ibv_flow_spec_gre tunnel = {
1630 		.type = IBV_FLOW_SPEC_GRE,
1631 		.size = size,
1632 	};
1633 #else
1634 	unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
1635 	struct ibv_flow_spec_tunnel tunnel = {
1636 		.type = IBV_FLOW_SPEC_VXLAN_TUNNEL,
1637 		.size = size,
1638 	};
1639 #endif
1640 	int ret;
1641 
1642 	if (flow->l3_protocol_en && flow->l3_protocol != MLX5_IP_PROTOCOL_GRE)
1643 		return rte_flow_error_set(error, ENOTSUP,
1644 					  RTE_FLOW_ERROR_TYPE_ITEM,
1645 					  item,
1646 					  "protocol filtering not compatible"
1647 					  " with this GRE layer");
1648 	if (flow->layers & MLX5_FLOW_LAYER_TUNNEL)
1649 		return rte_flow_error_set(error, ENOTSUP,
1650 					  RTE_FLOW_ERROR_TYPE_ITEM,
1651 					  item,
1652 					  "a tunnel is already present");
1653 	if (!(flow->layers & MLX5_FLOW_LAYER_OUTER_L3))
1654 		return rte_flow_error_set(error, ENOTSUP,
1655 					  RTE_FLOW_ERROR_TYPE_ITEM,
1656 					  item,
1657 					  "L3 Layer is missing");
1658 	if (!mask)
1659 		mask = &rte_flow_item_gre_mask;
1660 	ret = mlx5_flow_item_acceptable
1661 		(item, (const uint8_t *)mask,
1662 		 (const uint8_t *)&rte_flow_item_gre_mask,
1663 		 sizeof(struct rte_flow_item_gre), error);
1664 	if (ret < 0)
1665 		return ret;
1666 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1667 	if (spec) {
1668 		tunnel.val.c_ks_res0_ver = spec->c_rsvd0_ver;
1669 		tunnel.val.protocol = spec->protocol;
1670 		tunnel.mask.c_ks_res0_ver = mask->c_rsvd0_ver;
1671 		tunnel.mask.protocol = mask->protocol;
1672 		/* Remove unwanted bits from values. */
1673 		tunnel.val.c_ks_res0_ver &= tunnel.mask.c_ks_res0_ver;
1674 		tunnel.val.protocol &= tunnel.mask.protocol;
1675 		tunnel.val.key &= tunnel.mask.key;
1676 	}
1677 #else
1678 	if (spec && (spec->protocol & mask->protocol))
1679 		return rte_flow_error_set(error, ENOTSUP,
1680 					  RTE_FLOW_ERROR_TYPE_ITEM,
1681 					  item,
1682 					  "without MPLS support the"
1683 					  " specification cannot be used for"
1684 					  " filtering");
1685 #endif /* !HAVE_IBV_DEVICE_MPLS_SUPPORT */
1686 	if (size <= flow_size) {
1687 		if (flow->layers & MLX5_FLOW_LAYER_OUTER_L3_IPV4)
1688 			mlx5_flow_item_gre_ip_protocol_update
1689 				(verbs->attr, IBV_FLOW_SPEC_IPV4_EXT,
1690 				 MLX5_IP_PROTOCOL_GRE);
1691 		else
1692 			mlx5_flow_item_gre_ip_protocol_update
1693 				(verbs->attr, IBV_FLOW_SPEC_IPV6,
1694 				 MLX5_IP_PROTOCOL_GRE);
1695 		mlx5_flow_spec_verbs_add(flow, &tunnel, size);
1696 		flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L2;
1697 	}
1698 	flow->layers |= MLX5_FLOW_LAYER_GRE;
1699 	return size;
1700 }
1701 
1702 /**
1703  * Convert the @p item into a Verbs specification after ensuring the NIC
1704  * will understand and process it correctly.
1705  * If the necessary size for the conversion is greater than the @p flow_size,
1706  * nothing is written in @p flow, the validation is still performed.
1707  *
1708  * @param[in] item
1709  *   Item specification.
1710  * @param[in, out] flow
1711  *   Pointer to flow structure.
1712  * @param[in] flow_size
1713  *   Size in bytes of the available space in @p flow, if too small, nothing is
1714  *   written.
1715  * @param[out] error
1716  *   Pointer to error structure.
1717  *
1718  * @return
1719  *   On success the number of bytes consumed/necessary, if the returned value
1720  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
1721  *   otherwise another call with this returned memory size should be done.
1722  *   On error, a negative errno value is returned and rte_errno is set.
1723  */
1724 static int
1725 mlx5_flow_item_mpls(const struct rte_flow_item *item __rte_unused,
1726 		    struct rte_flow *flow __rte_unused,
1727 		    const size_t flow_size __rte_unused,
1728 		    struct rte_flow_error *error)
1729 {
1730 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1731 	const struct rte_flow_item_mpls *spec = item->spec;
1732 	const struct rte_flow_item_mpls *mask = item->mask;
1733 	unsigned int size = sizeof(struct ibv_flow_spec_mpls);
1734 	struct ibv_flow_spec_mpls mpls = {
1735 		.type = IBV_FLOW_SPEC_MPLS,
1736 		.size = size,
1737 	};
1738 	int ret;
1739 
1740 	if (flow->l3_protocol_en && flow->l3_protocol != MLX5_IP_PROTOCOL_MPLS)
1741 		return rte_flow_error_set(error, ENOTSUP,
1742 					  RTE_FLOW_ERROR_TYPE_ITEM,
1743 					  item,
1744 					  "protocol filtering not compatible"
1745 					  " with MPLS layer");
1746 	if (flow->layers & MLX5_FLOW_LAYER_TUNNEL)
1747 		return rte_flow_error_set(error, ENOTSUP,
1748 					  RTE_FLOW_ERROR_TYPE_ITEM,
1749 					  item,
1750 					  "a tunnel is already"
1751 					  " present");
1752 	if (!mask)
1753 		mask = &rte_flow_item_mpls_mask;
1754 	ret = mlx5_flow_item_acceptable
1755 		(item, (const uint8_t *)mask,
1756 		 (const uint8_t *)&rte_flow_item_mpls_mask,
1757 		 sizeof(struct rte_flow_item_mpls), error);
1758 	if (ret < 0)
1759 		return ret;
1760 	if (spec) {
1761 		memcpy(&mpls.val.label, spec, sizeof(mpls.val.label));
1762 		memcpy(&mpls.mask.label, mask, sizeof(mpls.mask.label));
1763 		/* Remove unwanted bits from values.  */
1764 		mpls.val.label &= mpls.mask.label;
1765 	}
1766 	if (size <= flow_size) {
1767 		mlx5_flow_spec_verbs_add(flow, &mpls, size);
1768 		flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L2;
1769 	}
1770 	flow->layers |= MLX5_FLOW_LAYER_MPLS;
1771 	return size;
1772 #endif /* !HAVE_IBV_DEVICE_MPLS_SUPPORT */
1773 	return rte_flow_error_set(error, ENOTSUP,
1774 				  RTE_FLOW_ERROR_TYPE_ITEM,
1775 				  item,
1776 				  "MPLS is not supported by Verbs, please"
1777 				  " update.");
1778 }
1779 
1780 /**
1781  * Convert the @p pattern into a Verbs specifications after ensuring the NIC
1782  * will understand and process it correctly.
1783  * The conversion is performed item per item, each of them is written into
1784  * the @p flow if its size is lesser or equal to @p flow_size.
1785  * Validation and memory consumption computation are still performed until the
1786  * end of @p pattern, unless an error is encountered.
1787  *
1788  * @param[in] pattern
1789  *   Flow pattern.
1790  * @param[in, out] flow
1791  *   Pointer to the rte_flow structure.
1792  * @param[in] flow_size
1793  *   Size in bytes of the available space in @p flow, if too small some
1794  *   garbage may be present.
1795  * @param[out] error
1796  *   Pointer to error structure.
1797  *
1798  * @return
1799  *   On success the number of bytes consumed/necessary, if the returned value
1800  *   is lesser or equal to @p flow_size, the @pattern  has fully been
1801  *   converted, otherwise another call with this returned memory size should
1802  *   be done.
1803  *   On error, a negative errno value is returned and rte_errno is set.
1804  */
1805 static int
1806 mlx5_flow_items(struct rte_eth_dev *dev,
1807 		const struct rte_flow_item pattern[],
1808 		struct rte_flow *flow, const size_t flow_size,
1809 		struct rte_flow_error *error)
1810 {
1811 	int remain = flow_size;
1812 	size_t size = 0;
1813 
1814 	for (; pattern->type != RTE_FLOW_ITEM_TYPE_END; pattern++) {
1815 		int ret = 0;
1816 
1817 		switch (pattern->type) {
1818 		case RTE_FLOW_ITEM_TYPE_VOID:
1819 			break;
1820 		case RTE_FLOW_ITEM_TYPE_ETH:
1821 			ret = mlx5_flow_item_eth(pattern, flow, remain, error);
1822 			break;
1823 		case RTE_FLOW_ITEM_TYPE_VLAN:
1824 			ret = mlx5_flow_item_vlan(pattern, flow, remain, error);
1825 			break;
1826 		case RTE_FLOW_ITEM_TYPE_IPV4:
1827 			ret = mlx5_flow_item_ipv4(pattern, flow, remain, error);
1828 			break;
1829 		case RTE_FLOW_ITEM_TYPE_IPV6:
1830 			ret = mlx5_flow_item_ipv6(pattern, flow, remain, error);
1831 			break;
1832 		case RTE_FLOW_ITEM_TYPE_UDP:
1833 			ret = mlx5_flow_item_udp(pattern, flow, remain, error);
1834 			break;
1835 		case RTE_FLOW_ITEM_TYPE_TCP:
1836 			ret = mlx5_flow_item_tcp(pattern, flow, remain, error);
1837 			break;
1838 		case RTE_FLOW_ITEM_TYPE_VXLAN:
1839 			ret = mlx5_flow_item_vxlan(pattern, flow, remain,
1840 						   error);
1841 			break;
1842 		case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
1843 			ret = mlx5_flow_item_vxlan_gpe(dev, pattern, flow,
1844 						       remain, error);
1845 			break;
1846 		case RTE_FLOW_ITEM_TYPE_GRE:
1847 			ret = mlx5_flow_item_gre(pattern, flow, remain, error);
1848 			break;
1849 		case RTE_FLOW_ITEM_TYPE_MPLS:
1850 			ret = mlx5_flow_item_mpls(pattern, flow, remain, error);
1851 			break;
1852 		default:
1853 			return rte_flow_error_set(error, ENOTSUP,
1854 						  RTE_FLOW_ERROR_TYPE_ITEM,
1855 						  pattern,
1856 						  "item not supported");
1857 		}
1858 		if (ret < 0)
1859 			return ret;
1860 		if (remain > ret)
1861 			remain -= ret;
1862 		else
1863 			remain = 0;
1864 		size += ret;
1865 	}
1866 	if (!flow->layers) {
1867 		const struct rte_flow_item item = {
1868 			.type = RTE_FLOW_ITEM_TYPE_ETH,
1869 		};
1870 
1871 		return mlx5_flow_item_eth(&item, flow, flow_size, error);
1872 	}
1873 	return size;
1874 }
1875 
1876 /**
1877  * Convert the @p action into a Verbs specification after ensuring the NIC
1878  * will understand and process it correctly.
1879  * If the necessary size for the conversion is greater than the @p flow_size,
1880  * nothing is written in @p flow, the validation is still performed.
1881  *
1882  * @param[in] action
1883  *   Action configuration.
1884  * @param[in, out] flow
1885  *   Pointer to flow structure.
1886  * @param[in] flow_size
1887  *   Size in bytes of the available space in @p flow, if too small, nothing is
1888  *   written.
1889  * @param[out] error
1890  *   Pointer to error structure.
1891  *
1892  * @return
1893  *   On success the number of bytes consumed/necessary, if the returned value
1894  *   is lesser or equal to @p flow_size, the @p action has fully been
1895  *   converted, otherwise another call with this returned memory size should
1896  *   be done.
1897  *   On error, a negative errno value is returned and rte_errno is set.
1898  */
1899 static int
1900 mlx5_flow_action_drop(const struct rte_flow_action *action,
1901 		      struct rte_flow *flow, const size_t flow_size,
1902 		      struct rte_flow_error *error)
1903 {
1904 	unsigned int size = sizeof(struct ibv_flow_spec_action_drop);
1905 	struct ibv_flow_spec_action_drop drop = {
1906 			.type = IBV_FLOW_SPEC_ACTION_DROP,
1907 			.size = size,
1908 	};
1909 
1910 	if (flow->fate)
1911 		return rte_flow_error_set(error, ENOTSUP,
1912 					  RTE_FLOW_ERROR_TYPE_ACTION,
1913 					  action,
1914 					  "multiple fate actions are not"
1915 					  " supported");
1916 	if (flow->modifier & (MLX5_FLOW_MOD_FLAG | MLX5_FLOW_MOD_MARK))
1917 		return rte_flow_error_set(error, ENOTSUP,
1918 					  RTE_FLOW_ERROR_TYPE_ACTION,
1919 					  action,
1920 					  "drop is not compatible with"
1921 					  " flag/mark action");
1922 	if (size < flow_size)
1923 		mlx5_flow_spec_verbs_add(flow, &drop, size);
1924 	flow->fate |= MLX5_FLOW_FATE_DROP;
1925 	return size;
1926 }
1927 
1928 /**
1929  * Convert the @p action into @p flow after ensuring the NIC will understand
1930  * and process it correctly.
1931  *
1932  * @param[in] dev
1933  *   Pointer to Ethernet device structure.
1934  * @param[in] action
1935  *   Action configuration.
1936  * @param[in, out] flow
1937  *   Pointer to flow structure.
1938  * @param[out] error
1939  *   Pointer to error structure.
1940  *
1941  * @return
1942  *   0 on success, a negative errno value otherwise and rte_errno is set.
1943  */
1944 static int
1945 mlx5_flow_action_queue(struct rte_eth_dev *dev,
1946 		       const struct rte_flow_action *action,
1947 		       struct rte_flow *flow,
1948 		       struct rte_flow_error *error)
1949 {
1950 	struct priv *priv = dev->data->dev_private;
1951 	const struct rte_flow_action_queue *queue = action->conf;
1952 
1953 	if (flow->fate)
1954 		return rte_flow_error_set(error, ENOTSUP,
1955 					  RTE_FLOW_ERROR_TYPE_ACTION,
1956 					  action,
1957 					  "multiple fate actions are not"
1958 					  " supported");
1959 	if (queue->index >= priv->rxqs_n)
1960 		return rte_flow_error_set(error, EINVAL,
1961 					  RTE_FLOW_ERROR_TYPE_ACTION_CONF,
1962 					  &queue->index,
1963 					  "queue index out of range");
1964 	if (!(*priv->rxqs)[queue->index])
1965 		return rte_flow_error_set(error, EINVAL,
1966 					  RTE_FLOW_ERROR_TYPE_ACTION_CONF,
1967 					  &queue->index,
1968 					  "queue is not configured");
1969 	if (flow->queue)
1970 		(*flow->queue)[0] = queue->index;
1971 	flow->rss.queue_num = 1;
1972 	flow->fate |= MLX5_FLOW_FATE_QUEUE;
1973 	return 0;
1974 }
1975 
1976 /**
1977  * Ensure the @p action will be understood and used correctly by the  NIC.
1978  *
1979  * @param dev
1980  *   Pointer to Ethernet device structure.
1981  * @param action[in]
1982  *   Pointer to flow actions array.
1983  * @param flow[in, out]
1984  *   Pointer to the rte_flow structure.
1985  * @param error[in, out]
1986  *   Pointer to error structure.
1987  *
1988  * @return
1989  *   On success @p flow->queue array and @p flow->rss are filled and valid.
1990  *   On error, a negative errno value is returned and rte_errno is set.
1991  */
1992 static int
1993 mlx5_flow_action_rss(struct rte_eth_dev *dev,
1994 		     const struct rte_flow_action *action,
1995 		     struct rte_flow *flow,
1996 		     struct rte_flow_error *error)
1997 {
1998 	struct priv *priv = dev->data->dev_private;
1999 	const struct rte_flow_action_rss *rss = action->conf;
2000 	unsigned int i;
2001 
2002 	if (flow->fate)
2003 		return rte_flow_error_set(error, ENOTSUP,
2004 					  RTE_FLOW_ERROR_TYPE_ACTION,
2005 					  action,
2006 					  "multiple fate actions are not"
2007 					  " supported");
2008 	if (rss->func != RTE_ETH_HASH_FUNCTION_DEFAULT &&
2009 	    rss->func != RTE_ETH_HASH_FUNCTION_TOEPLITZ)
2010 		return rte_flow_error_set(error, ENOTSUP,
2011 					  RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2012 					  &rss->func,
2013 					  "RSS hash function not supported");
2014 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2015 	if (rss->level > 2)
2016 #else
2017 	if (rss->level > 1)
2018 #endif
2019 		return rte_flow_error_set(error, ENOTSUP,
2020 					  RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2021 					  &rss->level,
2022 					  "tunnel RSS is not supported");
2023 	if (rss->key_len < MLX5_RSS_HASH_KEY_LEN)
2024 		return rte_flow_error_set(error, ENOTSUP,
2025 					  RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2026 					  &rss->key_len,
2027 					  "RSS hash key too small");
2028 	if (rss->key_len > MLX5_RSS_HASH_KEY_LEN)
2029 		return rte_flow_error_set(error, ENOTSUP,
2030 					  RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2031 					  &rss->key_len,
2032 					  "RSS hash key too large");
2033 	if (rss->queue_num > priv->config.ind_table_max_size)
2034 		return rte_flow_error_set(error, ENOTSUP,
2035 					  RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2036 					  &rss->queue_num,
2037 					  "number of queues too large");
2038 	if (rss->types & MLX5_RSS_HF_MASK)
2039 		return rte_flow_error_set(error, ENOTSUP,
2040 					  RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2041 					  &rss->types,
2042 					  "some RSS protocols are not"
2043 					  " supported");
2044 	for (i = 0; i != rss->queue_num; ++i) {
2045 		if (!(*priv->rxqs)[rss->queue[i]])
2046 			return rte_flow_error_set
2047 				(error, EINVAL,
2048 				 RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2049 				 &rss->queue[i],
2050 				 "queue is not configured");
2051 	}
2052 	if (flow->queue)
2053 		memcpy((*flow->queue), rss->queue,
2054 		       rss->queue_num * sizeof(uint16_t));
2055 	flow->rss.queue_num = rss->queue_num;
2056 	memcpy(flow->key, rss->key, MLX5_RSS_HASH_KEY_LEN);
2057 	flow->rss.types = rss->types;
2058 	flow->rss.level = rss->level;
2059 	flow->fate |= MLX5_FLOW_FATE_RSS;
2060 	return 0;
2061 }
2062 
2063 /**
2064  * Convert the @p action into a Verbs specification after ensuring the NIC
2065  * will understand and process it correctly.
2066  * If the necessary size for the conversion is greater than the @p flow_size,
2067  * nothing is written in @p flow, the validation is still performed.
2068  *
2069  * @param[in] action
2070  *   Action configuration.
2071  * @param[in, out] flow
2072  *   Pointer to flow structure.
2073  * @param[in] flow_size
2074  *   Size in bytes of the available space in @p flow, if too small, nothing is
2075  *   written.
2076  * @param[out] error
2077  *   Pointer to error structure.
2078  *
2079  * @return
2080  *   On success the number of bytes consumed/necessary, if the returned value
2081  *   is lesser or equal to @p flow_size, the @p action has fully been
2082  *   converted, otherwise another call with this returned memory size should
2083  *   be done.
2084  *   On error, a negative errno value is returned and rte_errno is set.
2085  */
2086 static int
2087 mlx5_flow_action_flag(const struct rte_flow_action *action,
2088 		      struct rte_flow *flow, const size_t flow_size,
2089 		      struct rte_flow_error *error)
2090 {
2091 	unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
2092 	struct ibv_flow_spec_action_tag tag = {
2093 		.type = IBV_FLOW_SPEC_ACTION_TAG,
2094 		.size = size,
2095 		.tag_id = mlx5_flow_mark_set(MLX5_FLOW_MARK_DEFAULT),
2096 	};
2097 	struct mlx5_flow_verbs *verbs = flow->cur_verbs;
2098 
2099 	if (flow->modifier & MLX5_FLOW_MOD_FLAG)
2100 		return rte_flow_error_set(error, ENOTSUP,
2101 					  RTE_FLOW_ERROR_TYPE_ACTION,
2102 					  action,
2103 					  "flag action already present");
2104 	if (flow->fate & MLX5_FLOW_FATE_DROP)
2105 		return rte_flow_error_set(error, ENOTSUP,
2106 					  RTE_FLOW_ERROR_TYPE_ACTION,
2107 					  action,
2108 					  "flag is not compatible with drop"
2109 					  " action");
2110 	if (flow->modifier & MLX5_FLOW_MOD_MARK)
2111 		size = 0;
2112 	else if (size <= flow_size && verbs)
2113 		mlx5_flow_spec_verbs_add(flow, &tag, size);
2114 	flow->modifier |= MLX5_FLOW_MOD_FLAG;
2115 	return size;
2116 }
2117 
2118 /**
2119  * Update verbs specification to modify the flag to mark.
2120  *
2121  * @param[in, out] verbs
2122  *   Pointer to the mlx5_flow_verbs structure.
2123  * @param[in] mark_id
2124  *   Mark identifier to replace the flag.
2125  */
2126 static void
2127 mlx5_flow_verbs_mark_update(struct mlx5_flow_verbs *verbs, uint32_t mark_id)
2128 {
2129 	struct ibv_spec_header *hdr;
2130 	int i;
2131 
2132 	if (!verbs)
2133 		return;
2134 	/* Update Verbs specification. */
2135 	hdr = (struct ibv_spec_header *)verbs->specs;
2136 	if (!hdr)
2137 		return;
2138 	for (i = 0; i != verbs->attr->num_of_specs; ++i) {
2139 		if (hdr->type == IBV_FLOW_SPEC_ACTION_TAG) {
2140 			struct ibv_flow_spec_action_tag *t =
2141 				(struct ibv_flow_spec_action_tag *)hdr;
2142 
2143 			t->tag_id = mlx5_flow_mark_set(mark_id);
2144 		}
2145 		hdr = (struct ibv_spec_header *)((uintptr_t)hdr + hdr->size);
2146 	}
2147 }
2148 
2149 /**
2150  * Convert the @p action into @p flow (or by updating the already present
2151  * Flag Verbs specification) after ensuring the NIC will understand and
2152  * process it correctly.
2153  * If the necessary size for the conversion is greater than the @p flow_size,
2154  * nothing is written in @p flow, the validation is still performed.
2155  *
2156  * @param[in] action
2157  *   Action configuration.
2158  * @param[in, out] flow
2159  *   Pointer to flow structure.
2160  * @param[in] flow_size
2161  *   Size in bytes of the available space in @p flow, if too small, nothing is
2162  *   written.
2163  * @param[out] error
2164  *   Pointer to error structure.
2165  *
2166  * @return
2167  *   On success the number of bytes consumed/necessary, if the returned value
2168  *   is lesser or equal to @p flow_size, the @p action has fully been
2169  *   converted, otherwise another call with this returned memory size should
2170  *   be done.
2171  *   On error, a negative errno value is returned and rte_errno is set.
2172  */
2173 static int
2174 mlx5_flow_action_mark(const struct rte_flow_action *action,
2175 		      struct rte_flow *flow, const size_t flow_size,
2176 		      struct rte_flow_error *error)
2177 {
2178 	const struct rte_flow_action_mark *mark = action->conf;
2179 	unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
2180 	struct ibv_flow_spec_action_tag tag = {
2181 		.type = IBV_FLOW_SPEC_ACTION_TAG,
2182 		.size = size,
2183 	};
2184 	struct mlx5_flow_verbs *verbs = flow->cur_verbs;
2185 
2186 	if (!mark)
2187 		return rte_flow_error_set(error, EINVAL,
2188 					  RTE_FLOW_ERROR_TYPE_ACTION,
2189 					  action,
2190 					  "configuration cannot be null");
2191 	if (mark->id >= MLX5_FLOW_MARK_MAX)
2192 		return rte_flow_error_set(error, EINVAL,
2193 					  RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2194 					  &mark->id,
2195 					  "mark id must in 0 <= id < "
2196 					  RTE_STR(MLX5_FLOW_MARK_MAX));
2197 	if (flow->modifier & MLX5_FLOW_MOD_MARK)
2198 		return rte_flow_error_set(error, ENOTSUP,
2199 					  RTE_FLOW_ERROR_TYPE_ACTION,
2200 					  action,
2201 					  "mark action already present");
2202 	if (flow->fate & MLX5_FLOW_FATE_DROP)
2203 		return rte_flow_error_set(error, ENOTSUP,
2204 					  RTE_FLOW_ERROR_TYPE_ACTION,
2205 					  action,
2206 					  "mark is not compatible with drop"
2207 					  " action");
2208 	if (flow->modifier & MLX5_FLOW_MOD_FLAG) {
2209 		mlx5_flow_verbs_mark_update(verbs, mark->id);
2210 		size = 0;
2211 	} else if (size <= flow_size) {
2212 		tag.tag_id = mlx5_flow_mark_set(mark->id);
2213 		mlx5_flow_spec_verbs_add(flow, &tag, size);
2214 	}
2215 	flow->modifier |= MLX5_FLOW_MOD_MARK;
2216 	return size;
2217 }
2218 
2219 /**
2220  * Convert the @p action into a Verbs specification after ensuring the NIC
2221  * will understand and process it correctly.
2222  * If the necessary size for the conversion is greater than the @p flow_size,
2223  * nothing is written in @p flow, the validation is still performed.
2224  *
2225  * @param action[in]
2226  *   Action configuration.
2227  * @param flow[in, out]
2228  *   Pointer to flow structure.
2229  * @param flow_size[in]
2230  *   Size in bytes of the available space in @p flow, if too small, nothing is
2231  *   written.
2232  * @param error[int, out]
2233  *   Pointer to error structure.
2234  *
2235  * @return
2236  *   On success the number of bytes consumed/necessary, if the returned value
2237  *   is lesser or equal to @p flow_size, the @p action has fully been
2238  *   converted, otherwise another call with this returned memory size should
2239  *   be done.
2240  *   On error, a negative errno value is returned and rte_errno is set.
2241  */
2242 static int
2243 mlx5_flow_action_count(struct rte_eth_dev *dev,
2244 		       const struct rte_flow_action *action,
2245 		       struct rte_flow *flow,
2246 		       const size_t flow_size __rte_unused,
2247 		       struct rte_flow_error *error)
2248 {
2249 	const struct rte_flow_action_count *count = action->conf;
2250 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
2251 	unsigned int size = sizeof(struct ibv_flow_spec_counter_action);
2252 	struct ibv_flow_spec_counter_action counter = {
2253 		.type = IBV_FLOW_SPEC_ACTION_COUNT,
2254 		.size = size,
2255 	};
2256 #endif
2257 
2258 	if (!flow->counter) {
2259 		flow->counter = mlx5_flow_counter_new(dev, count->shared,
2260 						      count->id);
2261 		if (!flow->counter)
2262 			return rte_flow_error_set(error, ENOTSUP,
2263 						  RTE_FLOW_ERROR_TYPE_ACTION,
2264 						  action,
2265 						  "cannot get counter"
2266 						  " context.");
2267 	}
2268 	if (!((struct priv *)dev->data->dev_private)->config.flow_counter_en)
2269 		return rte_flow_error_set(error, ENOTSUP,
2270 					  RTE_FLOW_ERROR_TYPE_ACTION,
2271 					  action,
2272 					  "flow counters are not supported.");
2273 	flow->modifier |= MLX5_FLOW_MOD_COUNT;
2274 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
2275 	counter.counter_set_handle = flow->counter->cs->handle;
2276 	if (size <= flow_size)
2277 		mlx5_flow_spec_verbs_add(flow, &counter, size);
2278 	return size;
2279 #endif
2280 	return 0;
2281 }
2282 
2283 /**
2284  * Convert the @p action into @p flow after ensuring the NIC will understand
2285  * and process it correctly.
2286  * The conversion is performed action per action, each of them is written into
2287  * the @p flow if its size is lesser or equal to @p flow_size.
2288  * Validation and memory consumption computation are still performed until the
2289  * end of @p action, unless an error is encountered.
2290  *
2291  * @param[in] dev
2292  *   Pointer to Ethernet device structure.
2293  * @param[in] actions
2294  *   Pointer to flow actions array.
2295  * @param[in, out] flow
2296  *   Pointer to the rte_flow structure.
2297  * @param[in] flow_size
2298  *   Size in bytes of the available space in @p flow, if too small some
2299  *   garbage may be present.
2300  * @param[out] error
2301  *   Pointer to error structure.
2302  *
2303  * @return
2304  *   On success the number of bytes consumed/necessary, if the returned value
2305  *   is lesser or equal to @p flow_size, the @p actions has fully been
2306  *   converted, otherwise another call with this returned memory size should
2307  *   be done.
2308  *   On error, a negative errno value is returned and rte_errno is set.
2309  */
2310 static int
2311 mlx5_flow_actions(struct rte_eth_dev *dev,
2312 		  const struct rte_flow_action actions[],
2313 		  struct rte_flow *flow, const size_t flow_size,
2314 		  struct rte_flow_error *error)
2315 {
2316 	size_t size = 0;
2317 	int remain = flow_size;
2318 	int ret = 0;
2319 
2320 	for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
2321 		switch (actions->type) {
2322 		case RTE_FLOW_ACTION_TYPE_VOID:
2323 			break;
2324 		case RTE_FLOW_ACTION_TYPE_FLAG:
2325 			ret = mlx5_flow_action_flag(actions, flow, remain,
2326 						    error);
2327 			break;
2328 		case RTE_FLOW_ACTION_TYPE_MARK:
2329 			ret = mlx5_flow_action_mark(actions, flow, remain,
2330 						    error);
2331 			break;
2332 		case RTE_FLOW_ACTION_TYPE_DROP:
2333 			ret = mlx5_flow_action_drop(actions, flow, remain,
2334 						    error);
2335 			break;
2336 		case RTE_FLOW_ACTION_TYPE_QUEUE:
2337 			ret = mlx5_flow_action_queue(dev, actions, flow, error);
2338 			break;
2339 		case RTE_FLOW_ACTION_TYPE_RSS:
2340 			ret = mlx5_flow_action_rss(dev, actions, flow, error);
2341 			break;
2342 		case RTE_FLOW_ACTION_TYPE_COUNT:
2343 			ret = mlx5_flow_action_count(dev, actions, flow, remain,
2344 						     error);
2345 			break;
2346 		default:
2347 			return rte_flow_error_set(error, ENOTSUP,
2348 						  RTE_FLOW_ERROR_TYPE_ACTION,
2349 						  actions,
2350 						  "action not supported");
2351 		}
2352 		if (ret < 0)
2353 			return ret;
2354 		if (remain > ret)
2355 			remain -= ret;
2356 		else
2357 			remain = 0;
2358 		size += ret;
2359 	}
2360 	if (!flow->fate)
2361 		return rte_flow_error_set(error, ENOTSUP,
2362 					  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2363 					  NULL,
2364 					  "no fate action found");
2365 	return size;
2366 }
2367 
2368 /**
2369  * Convert the @p attributes, @p pattern, @p action, into an flow for the NIC
2370  * after ensuring the NIC will understand and process it correctly.
2371  * The conversion is only performed item/action per item/action, each of
2372  * them is written into the @p flow if its size is lesser or equal to @p
2373  * flow_size.
2374  * Validation and memory consumption computation are still performed until the
2375  * end, unless an error is encountered.
2376  *
2377  * @param[in] dev
2378  *   Pointer to Ethernet device.
2379  * @param[in, out] flow
2380  *   Pointer to flow structure.
2381  * @param[in] flow_size
2382  *   Size in bytes of the available space in @p flow, if too small some
2383  *   garbage may be present.
2384  * @param[in] attributes
2385  *   Flow rule attributes.
2386  * @param[in] pattern
2387  *   Pattern specification (list terminated by the END pattern item).
2388  * @param[in] actions
2389  *   Associated actions (list terminated by the END action).
2390  * @param[out] error
2391  *   Perform verbose error reporting if not NULL.
2392  *
2393  * @return
2394  *   On success the number of bytes consumed/necessary, if the returned value
2395  *   is lesser or equal to @p flow_size, the flow has fully been converted and
2396  *   can be applied, otherwise another call with this returned memory size
2397  *   should be done.
2398  *   On error, a negative errno value is returned and rte_errno is set.
2399  */
2400 static int
2401 mlx5_flow_merge(struct rte_eth_dev *dev, struct rte_flow *flow,
2402 		const size_t flow_size,
2403 		const struct rte_flow_attr *attributes,
2404 		const struct rte_flow_item pattern[],
2405 		const struct rte_flow_action actions[],
2406 		struct rte_flow_error *error)
2407 {
2408 	struct rte_flow local_flow = { .layers = 0, };
2409 	size_t size = sizeof(*flow);
2410 	union {
2411 		struct rte_flow_expand_rss buf;
2412 		uint8_t buffer[2048];
2413 	} expand_buffer;
2414 	struct rte_flow_expand_rss *buf = &expand_buffer.buf;
2415 	struct mlx5_flow_verbs *original_verbs = NULL;
2416 	size_t original_verbs_size = 0;
2417 	uint32_t original_layers = 0;
2418 	int expanded_pattern_idx = 0;
2419 	int ret;
2420 	uint32_t i;
2421 
2422 	if (size > flow_size)
2423 		flow = &local_flow;
2424 	ret = mlx5_flow_attributes(dev, attributes, flow, error);
2425 	if (ret < 0)
2426 		return ret;
2427 	ret = mlx5_flow_actions(dev, actions, &local_flow, 0, error);
2428 	if (ret < 0)
2429 		return ret;
2430 	if (local_flow.rss.types) {
2431 		ret = rte_flow_expand_rss(buf, sizeof(expand_buffer.buffer),
2432 					  pattern, local_flow.rss.types,
2433 					  mlx5_support_expansion,
2434 					  local_flow.rss.level < 2 ?
2435 					  MLX5_EXPANSION_ROOT :
2436 					  MLX5_EXPANSION_ROOT_OUTER);
2437 		assert(ret > 0 &&
2438 		       (unsigned int)ret < sizeof(expand_buffer.buffer));
2439 	} else {
2440 		buf->entries = 1;
2441 		buf->entry[0].pattern = (void *)(uintptr_t)pattern;
2442 	}
2443 	size += RTE_ALIGN_CEIL(local_flow.rss.queue_num * sizeof(uint16_t),
2444 			       sizeof(void *));
2445 	if (size <= flow_size)
2446 		flow->queue = (void *)(flow + 1);
2447 	LIST_INIT(&flow->verbs);
2448 	flow->layers = 0;
2449 	flow->modifier = 0;
2450 	flow->fate = 0;
2451 	for (i = 0; i != buf->entries; ++i) {
2452 		size_t off = size;
2453 		size_t off2;
2454 
2455 		flow->layers = original_layers;
2456 		size += sizeof(struct ibv_flow_attr) +
2457 			sizeof(struct mlx5_flow_verbs);
2458 		off2 = size;
2459 		if (size < flow_size) {
2460 			flow->cur_verbs = (void *)((uintptr_t)flow + off);
2461 			flow->cur_verbs->attr = (void *)(flow->cur_verbs + 1);
2462 			flow->cur_verbs->specs =
2463 				(void *)(flow->cur_verbs->attr + 1);
2464 		}
2465 		/* First iteration convert the pattern into Verbs. */
2466 		if (i == 0) {
2467 			/* Actions don't need to be converted several time. */
2468 			ret = mlx5_flow_actions(dev, actions, flow,
2469 						(size < flow_size) ?
2470 						flow_size - size : 0,
2471 						error);
2472 			if (ret < 0)
2473 				return ret;
2474 			size += ret;
2475 		} else {
2476 			/*
2477 			 * Next iteration means the pattern has already been
2478 			 * converted and an expansion is necessary to match
2479 			 * the user RSS request.  For that only the expanded
2480 			 * items will be converted, the common part with the
2481 			 * user pattern are just copied into the next buffer
2482 			 * zone.
2483 			 */
2484 			size += original_verbs_size;
2485 			if (size < flow_size) {
2486 				rte_memcpy(flow->cur_verbs->attr,
2487 					   original_verbs->attr,
2488 					   original_verbs_size +
2489 					   sizeof(struct ibv_flow_attr));
2490 				flow->cur_verbs->size = original_verbs_size;
2491 			}
2492 		}
2493 		ret = mlx5_flow_items
2494 			(dev,
2495 			 (const struct rte_flow_item *)
2496 			 &buf->entry[i].pattern[expanded_pattern_idx],
2497 			 flow,
2498 			 (size < flow_size) ? flow_size - size : 0, error);
2499 		if (ret < 0)
2500 			return ret;
2501 		size += ret;
2502 		if (size <= flow_size) {
2503 			mlx5_flow_adjust_priority(dev, flow);
2504 			LIST_INSERT_HEAD(&flow->verbs, flow->cur_verbs, next);
2505 		}
2506 		/*
2507 		 * Keep a pointer of the first verbs conversion and the layers
2508 		 * it has encountered.
2509 		 */
2510 		if (i == 0) {
2511 			original_verbs = flow->cur_verbs;
2512 			original_verbs_size = size - off2;
2513 			original_layers = flow->layers;
2514 			/*
2515 			 * move the index of the expanded pattern to the
2516 			 * first item not addressed yet.
2517 			 */
2518 			if (pattern->type == RTE_FLOW_ITEM_TYPE_END) {
2519 				expanded_pattern_idx++;
2520 			} else {
2521 				const struct rte_flow_item *item = pattern;
2522 
2523 				for (item = pattern;
2524 				     item->type != RTE_FLOW_ITEM_TYPE_END;
2525 				     ++item)
2526 					expanded_pattern_idx++;
2527 			}
2528 		}
2529 	}
2530 	/* Restore the origin layers in the flow. */
2531 	flow->layers = original_layers;
2532 	return size;
2533 }
2534 
2535 /**
2536  * Lookup and set the ptype in the data Rx part.  A single Ptype can be used,
2537  * if several tunnel rules are used on this queue, the tunnel ptype will be
2538  * cleared.
2539  *
2540  * @param rxq_ctrl
2541  *   Rx queue to update.
2542  */
2543 static void
2544 mlx5_flow_rxq_tunnel_ptype_update(struct mlx5_rxq_ctrl *rxq_ctrl)
2545 {
2546 	unsigned int i;
2547 	uint32_t tunnel_ptype = 0;
2548 
2549 	/* Look up for the ptype to use. */
2550 	for (i = 0; i != MLX5_FLOW_TUNNEL; ++i) {
2551 		if (!rxq_ctrl->flow_tunnels_n[i])
2552 			continue;
2553 		if (!tunnel_ptype) {
2554 			tunnel_ptype = tunnels_info[i].ptype;
2555 		} else {
2556 			tunnel_ptype = 0;
2557 			break;
2558 		}
2559 	}
2560 	rxq_ctrl->rxq.tunnel = tunnel_ptype;
2561 }
2562 
2563 /**
2564  * Set the Rx queue flags (Mark/Flag and Tunnel Ptypes) according to the flow.
2565  *
2566  * @param[in] dev
2567  *   Pointer to Ethernet device.
2568  * @param[in] flow
2569  *   Pointer to flow structure.
2570  */
2571 static void
2572 mlx5_flow_rxq_flags_set(struct rte_eth_dev *dev, struct rte_flow *flow)
2573 {
2574 	struct priv *priv = dev->data->dev_private;
2575 	const int mark = !!(flow->modifier &
2576 			    (MLX5_FLOW_MOD_FLAG | MLX5_FLOW_MOD_MARK));
2577 	const int tunnel = !!(flow->layers & MLX5_FLOW_LAYER_TUNNEL);
2578 	unsigned int i;
2579 
2580 	for (i = 0; i != flow->rss.queue_num; ++i) {
2581 		int idx = (*flow->queue)[i];
2582 		struct mlx5_rxq_ctrl *rxq_ctrl =
2583 			container_of((*priv->rxqs)[idx],
2584 				     struct mlx5_rxq_ctrl, rxq);
2585 
2586 		if (mark) {
2587 			rxq_ctrl->rxq.mark = 1;
2588 			rxq_ctrl->flow_mark_n++;
2589 		}
2590 		if (tunnel) {
2591 			unsigned int j;
2592 
2593 			/* Increase the counter matching the flow. */
2594 			for (j = 0; j != MLX5_FLOW_TUNNEL; ++j) {
2595 				if ((tunnels_info[j].tunnel & flow->layers) ==
2596 				    tunnels_info[j].tunnel) {
2597 					rxq_ctrl->flow_tunnels_n[j]++;
2598 					break;
2599 				}
2600 			}
2601 			mlx5_flow_rxq_tunnel_ptype_update(rxq_ctrl);
2602 		}
2603 	}
2604 }
2605 
2606 /**
2607  * Clear the Rx queue flags (Mark/Flag and Tunnel Ptype) associated with the
2608  * @p flow if no other flow uses it with the same kind of request.
2609  *
2610  * @param dev
2611  *   Pointer to Ethernet device.
2612  * @param[in] flow
2613  *   Pointer to the flow.
2614  */
2615 static void
2616 mlx5_flow_rxq_flags_trim(struct rte_eth_dev *dev, struct rte_flow *flow)
2617 {
2618 	struct priv *priv = dev->data->dev_private;
2619 	const int mark = !!(flow->modifier &
2620 			    (MLX5_FLOW_MOD_FLAG | MLX5_FLOW_MOD_MARK));
2621 	const int tunnel = !!(flow->layers & MLX5_FLOW_LAYER_TUNNEL);
2622 	unsigned int i;
2623 
2624 	assert(dev->data->dev_started);
2625 	for (i = 0; i != flow->rss.queue_num; ++i) {
2626 		int idx = (*flow->queue)[i];
2627 		struct mlx5_rxq_ctrl *rxq_ctrl =
2628 			container_of((*priv->rxqs)[idx],
2629 				     struct mlx5_rxq_ctrl, rxq);
2630 
2631 		if (mark) {
2632 			rxq_ctrl->flow_mark_n--;
2633 			rxq_ctrl->rxq.mark = !!rxq_ctrl->flow_mark_n;
2634 		}
2635 		if (tunnel) {
2636 			unsigned int j;
2637 
2638 			/* Decrease the counter matching the flow. */
2639 			for (j = 0; j != MLX5_FLOW_TUNNEL; ++j) {
2640 				if ((tunnels_info[j].tunnel & flow->layers) ==
2641 				    tunnels_info[j].tunnel) {
2642 					rxq_ctrl->flow_tunnels_n[j]--;
2643 					break;
2644 				}
2645 			}
2646 			mlx5_flow_rxq_tunnel_ptype_update(rxq_ctrl);
2647 		}
2648 	}
2649 }
2650 
2651 /**
2652  * Clear the Mark/Flag and Tunnel ptype information in all Rx queues.
2653  *
2654  * @param dev
2655  *   Pointer to Ethernet device.
2656  */
2657 static void
2658 mlx5_flow_rxq_flags_clear(struct rte_eth_dev *dev)
2659 {
2660 	struct priv *priv = dev->data->dev_private;
2661 	unsigned int i;
2662 	unsigned int idx;
2663 
2664 	for (idx = 0, i = 0; idx != priv->rxqs_n; ++i) {
2665 		struct mlx5_rxq_ctrl *rxq_ctrl;
2666 		unsigned int j;
2667 
2668 		if (!(*priv->rxqs)[idx])
2669 			continue;
2670 		rxq_ctrl = container_of((*priv->rxqs)[idx],
2671 					struct mlx5_rxq_ctrl, rxq);
2672 		rxq_ctrl->flow_mark_n = 0;
2673 		rxq_ctrl->rxq.mark = 0;
2674 		for (j = 0; j != MLX5_FLOW_TUNNEL; ++j)
2675 			rxq_ctrl->flow_tunnels_n[j] = 0;
2676 		rxq_ctrl->rxq.tunnel = 0;
2677 		++idx;
2678 	}
2679 }
2680 
2681 /**
2682  * Validate a flow supported by the NIC.
2683  *
2684  * @see rte_flow_validate()
2685  * @see rte_flow_ops
2686  */
2687 int
2688 mlx5_flow_validate(struct rte_eth_dev *dev,
2689 		   const struct rte_flow_attr *attr,
2690 		   const struct rte_flow_item items[],
2691 		   const struct rte_flow_action actions[],
2692 		   struct rte_flow_error *error)
2693 {
2694 	int ret = mlx5_flow_merge(dev, NULL, 0, attr, items, actions, error);
2695 
2696 	if (ret < 0)
2697 		return ret;
2698 	return 0;
2699 }
2700 
2701 /**
2702  * Remove the flow.
2703  *
2704  * @param[in] dev
2705  *   Pointer to Ethernet device.
2706  * @param[in, out] flow
2707  *   Pointer to flow structure.
2708  */
2709 static void
2710 mlx5_flow_remove(struct rte_eth_dev *dev, struct rte_flow *flow)
2711 {
2712 	struct mlx5_flow_verbs *verbs;
2713 
2714 	LIST_FOREACH(verbs, &flow->verbs, next) {
2715 		if (verbs->flow) {
2716 			claim_zero(mlx5_glue->destroy_flow(verbs->flow));
2717 			verbs->flow = NULL;
2718 		}
2719 		if (verbs->hrxq) {
2720 			if (flow->fate & MLX5_FLOW_FATE_DROP)
2721 				mlx5_hrxq_drop_release(dev);
2722 			else
2723 				mlx5_hrxq_release(dev, verbs->hrxq);
2724 			verbs->hrxq = NULL;
2725 		}
2726 	}
2727 	if (flow->counter) {
2728 		mlx5_flow_counter_release(flow->counter);
2729 		flow->counter = NULL;
2730 	}
2731 }
2732 
2733 /**
2734  * Apply the flow.
2735  *
2736  * @param[in] dev
2737  *   Pointer to Ethernet device structure.
2738  * @param[in, out] flow
2739  *   Pointer to flow structure.
2740  * @param[out] error
2741  *   Pointer to error structure.
2742  *
2743  * @return
2744  *   0 on success, a negative errno value otherwise and rte_errno is set.
2745  */
2746 static int
2747 mlx5_flow_apply(struct rte_eth_dev *dev, struct rte_flow *flow,
2748 		struct rte_flow_error *error)
2749 {
2750 	struct mlx5_flow_verbs *verbs;
2751 	int err;
2752 
2753 	LIST_FOREACH(verbs, &flow->verbs, next) {
2754 		if (flow->fate & MLX5_FLOW_FATE_DROP) {
2755 			verbs->hrxq = mlx5_hrxq_drop_new(dev);
2756 			if (!verbs->hrxq) {
2757 				rte_flow_error_set
2758 					(error, errno,
2759 					 RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2760 					 NULL,
2761 					 "cannot get drop hash queue");
2762 				goto error;
2763 			}
2764 		} else {
2765 			struct mlx5_hrxq *hrxq;
2766 
2767 			hrxq = mlx5_hrxq_get(dev, flow->key,
2768 					     MLX5_RSS_HASH_KEY_LEN,
2769 					     verbs->hash_fields,
2770 					     (*flow->queue),
2771 					     flow->rss.queue_num);
2772 			if (!hrxq)
2773 				hrxq = mlx5_hrxq_new(dev, flow->key,
2774 						     MLX5_RSS_HASH_KEY_LEN,
2775 						     verbs->hash_fields,
2776 						     (*flow->queue),
2777 						     flow->rss.queue_num);
2778 			if (!hrxq) {
2779 				rte_flow_error_set
2780 					(error, rte_errno,
2781 					 RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2782 					 NULL,
2783 					 "cannot get hash queue");
2784 				goto error;
2785 			}
2786 			verbs->hrxq = hrxq;
2787 		}
2788 		verbs->flow =
2789 			mlx5_glue->create_flow(verbs->hrxq->qp, verbs->attr);
2790 		if (!verbs->flow) {
2791 			rte_flow_error_set(error, errno,
2792 					   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2793 					   NULL,
2794 					   "hardware refuses to create flow");
2795 			goto error;
2796 		}
2797 	}
2798 	return 0;
2799 error:
2800 	err = rte_errno; /* Save rte_errno before cleanup. */
2801 	LIST_FOREACH(verbs, &flow->verbs, next) {
2802 		if (verbs->hrxq) {
2803 			if (flow->fate & MLX5_FLOW_FATE_DROP)
2804 				mlx5_hrxq_drop_release(dev);
2805 			else
2806 				mlx5_hrxq_release(dev, verbs->hrxq);
2807 			verbs->hrxq = NULL;
2808 		}
2809 	}
2810 	rte_errno = err; /* Restore rte_errno. */
2811 	return -rte_errno;
2812 }
2813 
2814 /**
2815  * Create a flow and add it to @p list.
2816  *
2817  * @param dev
2818  *   Pointer to Ethernet device.
2819  * @param list
2820  *   Pointer to a TAILQ flow list.
2821  * @param[in] attr
2822  *   Flow rule attributes.
2823  * @param[in] items
2824  *   Pattern specification (list terminated by the END pattern item).
2825  * @param[in] actions
2826  *   Associated actions (list terminated by the END action).
2827  * @param[out] error
2828  *   Perform verbose error reporting if not NULL.
2829  *
2830  * @return
2831  *   A flow on success, NULL otherwise and rte_errno is set.
2832  */
2833 static struct rte_flow *
2834 mlx5_flow_list_create(struct rte_eth_dev *dev,
2835 		      struct mlx5_flows *list,
2836 		      const struct rte_flow_attr *attr,
2837 		      const struct rte_flow_item items[],
2838 		      const struct rte_flow_action actions[],
2839 		      struct rte_flow_error *error)
2840 {
2841 	struct rte_flow *flow = NULL;
2842 	size_t size = 0;
2843 	int ret;
2844 
2845 	ret = mlx5_flow_merge(dev, flow, size, attr, items, actions, error);
2846 	if (ret < 0)
2847 		return NULL;
2848 	size = ret;
2849 	flow = rte_calloc(__func__, 1, size, 0);
2850 	if (!flow) {
2851 		rte_flow_error_set(error, ENOMEM,
2852 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2853 				   NULL,
2854 				   "not enough memory to create flow");
2855 		return NULL;
2856 	}
2857 	ret = mlx5_flow_merge(dev, flow, size, attr, items, actions, error);
2858 	if (ret < 0) {
2859 		rte_free(flow);
2860 		return NULL;
2861 	}
2862 	assert((size_t)ret == size);
2863 	if (dev->data->dev_started) {
2864 		ret = mlx5_flow_apply(dev, flow, error);
2865 		if (ret < 0) {
2866 			ret = rte_errno; /* Save rte_errno before cleanup. */
2867 			if (flow) {
2868 				mlx5_flow_remove(dev, flow);
2869 				rte_free(flow);
2870 			}
2871 			rte_errno = ret; /* Restore rte_errno. */
2872 			return NULL;
2873 		}
2874 	}
2875 	TAILQ_INSERT_TAIL(list, flow, next);
2876 	mlx5_flow_rxq_flags_set(dev, flow);
2877 	return flow;
2878 }
2879 
2880 /**
2881  * Create a flow.
2882  *
2883  * @see rte_flow_create()
2884  * @see rte_flow_ops
2885  */
2886 struct rte_flow *
2887 mlx5_flow_create(struct rte_eth_dev *dev,
2888 		 const struct rte_flow_attr *attr,
2889 		 const struct rte_flow_item items[],
2890 		 const struct rte_flow_action actions[],
2891 		 struct rte_flow_error *error)
2892 {
2893 	return mlx5_flow_list_create
2894 		(dev, &((struct priv *)dev->data->dev_private)->flows,
2895 		 attr, items, actions, error);
2896 }
2897 
2898 /**
2899  * Destroy a flow in a list.
2900  *
2901  * @param dev
2902  *   Pointer to Ethernet device.
2903  * @param list
2904  *   Pointer to a TAILQ flow list.
2905  * @param[in] flow
2906  *   Flow to destroy.
2907  */
2908 static void
2909 mlx5_flow_list_destroy(struct rte_eth_dev *dev, struct mlx5_flows *list,
2910 		       struct rte_flow *flow)
2911 {
2912 	mlx5_flow_remove(dev, flow);
2913 	TAILQ_REMOVE(list, flow, next);
2914 	/*
2915 	 * Update RX queue flags only if port is started, otherwise it is
2916 	 * already clean.
2917 	 */
2918 	if (dev->data->dev_started)
2919 		mlx5_flow_rxq_flags_trim(dev, flow);
2920 	rte_free(flow);
2921 }
2922 
2923 /**
2924  * Destroy all flows.
2925  *
2926  * @param dev
2927  *   Pointer to Ethernet device.
2928  * @param list
2929  *   Pointer to a TAILQ flow list.
2930  */
2931 void
2932 mlx5_flow_list_flush(struct rte_eth_dev *dev, struct mlx5_flows *list)
2933 {
2934 	while (!TAILQ_EMPTY(list)) {
2935 		struct rte_flow *flow;
2936 
2937 		flow = TAILQ_FIRST(list);
2938 		mlx5_flow_list_destroy(dev, list, flow);
2939 	}
2940 }
2941 
2942 /**
2943  * Remove all flows.
2944  *
2945  * @param dev
2946  *   Pointer to Ethernet device.
2947  * @param list
2948  *   Pointer to a TAILQ flow list.
2949  */
2950 void
2951 mlx5_flow_stop(struct rte_eth_dev *dev, struct mlx5_flows *list)
2952 {
2953 	struct rte_flow *flow;
2954 
2955 	TAILQ_FOREACH_REVERSE(flow, list, mlx5_flows, next)
2956 		mlx5_flow_remove(dev, flow);
2957 	mlx5_flow_rxq_flags_clear(dev);
2958 }
2959 
2960 /**
2961  * Add all flows.
2962  *
2963  * @param dev
2964  *   Pointer to Ethernet device.
2965  * @param list
2966  *   Pointer to a TAILQ flow list.
2967  *
2968  * @return
2969  *   0 on success, a negative errno value otherwise and rte_errno is set.
2970  */
2971 int
2972 mlx5_flow_start(struct rte_eth_dev *dev, struct mlx5_flows *list)
2973 {
2974 	struct rte_flow *flow;
2975 	struct rte_flow_error error;
2976 	int ret = 0;
2977 
2978 	TAILQ_FOREACH(flow, list, next) {
2979 		ret = mlx5_flow_apply(dev, flow, &error);
2980 		if (ret < 0)
2981 			goto error;
2982 		mlx5_flow_rxq_flags_set(dev, flow);
2983 	}
2984 	return 0;
2985 error:
2986 	ret = rte_errno; /* Save rte_errno before cleanup. */
2987 	mlx5_flow_stop(dev, list);
2988 	rte_errno = ret; /* Restore rte_errno. */
2989 	return -rte_errno;
2990 }
2991 
2992 /**
2993  * Verify the flow list is empty
2994  *
2995  * @param dev
2996  *  Pointer to Ethernet device.
2997  *
2998  * @return the number of flows not released.
2999  */
3000 int
3001 mlx5_flow_verify(struct rte_eth_dev *dev)
3002 {
3003 	struct priv *priv = dev->data->dev_private;
3004 	struct rte_flow *flow;
3005 	int ret = 0;
3006 
3007 	TAILQ_FOREACH(flow, &priv->flows, next) {
3008 		DRV_LOG(DEBUG, "port %u flow %p still referenced",
3009 			dev->data->port_id, (void *)flow);
3010 		++ret;
3011 	}
3012 	return ret;
3013 }
3014 
3015 /**
3016  * Enable a control flow configured from the control plane.
3017  *
3018  * @param dev
3019  *   Pointer to Ethernet device.
3020  * @param eth_spec
3021  *   An Ethernet flow spec to apply.
3022  * @param eth_mask
3023  *   An Ethernet flow mask to apply.
3024  * @param vlan_spec
3025  *   A VLAN flow spec to apply.
3026  * @param vlan_mask
3027  *   A VLAN flow mask to apply.
3028  *
3029  * @return
3030  *   0 on success, a negative errno value otherwise and rte_errno is set.
3031  */
3032 int
3033 mlx5_ctrl_flow_vlan(struct rte_eth_dev *dev,
3034 		    struct rte_flow_item_eth *eth_spec,
3035 		    struct rte_flow_item_eth *eth_mask,
3036 		    struct rte_flow_item_vlan *vlan_spec,
3037 		    struct rte_flow_item_vlan *vlan_mask)
3038 {
3039 	struct priv *priv = dev->data->dev_private;
3040 	const struct rte_flow_attr attr = {
3041 		.ingress = 1,
3042 		.priority = MLX5_FLOW_PRIO_RSVD,
3043 	};
3044 	struct rte_flow_item items[] = {
3045 		{
3046 			.type = RTE_FLOW_ITEM_TYPE_ETH,
3047 			.spec = eth_spec,
3048 			.last = NULL,
3049 			.mask = eth_mask,
3050 		},
3051 		{
3052 			.type = (vlan_spec) ? RTE_FLOW_ITEM_TYPE_VLAN :
3053 				RTE_FLOW_ITEM_TYPE_END,
3054 			.spec = vlan_spec,
3055 			.last = NULL,
3056 			.mask = vlan_mask,
3057 		},
3058 		{
3059 			.type = RTE_FLOW_ITEM_TYPE_END,
3060 		},
3061 	};
3062 	uint16_t queue[priv->reta_idx_n];
3063 	struct rte_flow_action_rss action_rss = {
3064 		.func = RTE_ETH_HASH_FUNCTION_DEFAULT,
3065 		.level = 0,
3066 		.types = priv->rss_conf.rss_hf,
3067 		.key_len = priv->rss_conf.rss_key_len,
3068 		.queue_num = priv->reta_idx_n,
3069 		.key = priv->rss_conf.rss_key,
3070 		.queue = queue,
3071 	};
3072 	struct rte_flow_action actions[] = {
3073 		{
3074 			.type = RTE_FLOW_ACTION_TYPE_RSS,
3075 			.conf = &action_rss,
3076 		},
3077 		{
3078 			.type = RTE_FLOW_ACTION_TYPE_END,
3079 		},
3080 	};
3081 	struct rte_flow *flow;
3082 	struct rte_flow_error error;
3083 	unsigned int i;
3084 
3085 	if (!priv->reta_idx_n) {
3086 		rte_errno = EINVAL;
3087 		return -rte_errno;
3088 	}
3089 	for (i = 0; i != priv->reta_idx_n; ++i)
3090 		queue[i] = (*priv->reta_idx)[i];
3091 	flow = mlx5_flow_list_create(dev, &priv->ctrl_flows, &attr, items,
3092 				     actions, &error);
3093 	if (!flow)
3094 		return -rte_errno;
3095 	return 0;
3096 }
3097 
3098 /**
3099  * Enable a flow control configured from the control plane.
3100  *
3101  * @param dev
3102  *   Pointer to Ethernet device.
3103  * @param eth_spec
3104  *   An Ethernet flow spec to apply.
3105  * @param eth_mask
3106  *   An Ethernet flow mask to apply.
3107  *
3108  * @return
3109  *   0 on success, a negative errno value otherwise and rte_errno is set.
3110  */
3111 int
3112 mlx5_ctrl_flow(struct rte_eth_dev *dev,
3113 	       struct rte_flow_item_eth *eth_spec,
3114 	       struct rte_flow_item_eth *eth_mask)
3115 {
3116 	return mlx5_ctrl_flow_vlan(dev, eth_spec, eth_mask, NULL, NULL);
3117 }
3118 
3119 /**
3120  * Destroy a flow.
3121  *
3122  * @see rte_flow_destroy()
3123  * @see rte_flow_ops
3124  */
3125 int
3126 mlx5_flow_destroy(struct rte_eth_dev *dev,
3127 		  struct rte_flow *flow,
3128 		  struct rte_flow_error *error __rte_unused)
3129 {
3130 	struct priv *priv = dev->data->dev_private;
3131 
3132 	mlx5_flow_list_destroy(dev, &priv->flows, flow);
3133 	return 0;
3134 }
3135 
3136 /**
3137  * Destroy all flows.
3138  *
3139  * @see rte_flow_flush()
3140  * @see rte_flow_ops
3141  */
3142 int
3143 mlx5_flow_flush(struct rte_eth_dev *dev,
3144 		struct rte_flow_error *error __rte_unused)
3145 {
3146 	struct priv *priv = dev->data->dev_private;
3147 
3148 	mlx5_flow_list_flush(dev, &priv->flows);
3149 	return 0;
3150 }
3151 
3152 /**
3153  * Isolated mode.
3154  *
3155  * @see rte_flow_isolate()
3156  * @see rte_flow_ops
3157  */
3158 int
3159 mlx5_flow_isolate(struct rte_eth_dev *dev,
3160 		  int enable,
3161 		  struct rte_flow_error *error)
3162 {
3163 	struct priv *priv = dev->data->dev_private;
3164 
3165 	if (dev->data->dev_started) {
3166 		rte_flow_error_set(error, EBUSY,
3167 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
3168 				   NULL,
3169 				   "port must be stopped first");
3170 		return -rte_errno;
3171 	}
3172 	priv->isolated = !!enable;
3173 	if (enable)
3174 		dev->dev_ops = &mlx5_dev_ops_isolate;
3175 	else
3176 		dev->dev_ops = &mlx5_dev_ops;
3177 	return 0;
3178 }
3179 
3180 /**
3181  * Query flow counter.
3182  *
3183  * @param flow
3184  *   Pointer to the flow.
3185  *
3186  * @return
3187  *   0 on success, a negative errno value otherwise and rte_errno is set.
3188  */
3189 static int
3190 mlx5_flow_query_count(struct rte_flow *flow __rte_unused,
3191 		      void *data __rte_unused,
3192 		      struct rte_flow_error *error)
3193 {
3194 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
3195 	struct rte_flow_query_count *qc = data;
3196 	uint64_t counters[2] = {0, 0};
3197 	struct ibv_query_counter_set_attr query_cs_attr = {
3198 		.cs = flow->counter->cs,
3199 		.query_flags = IBV_COUNTER_SET_FORCE_UPDATE,
3200 	};
3201 	struct ibv_counter_set_data query_out = {
3202 		.out = counters,
3203 		.outlen = 2 * sizeof(uint64_t),
3204 	};
3205 	int err = mlx5_glue->query_counter_set(&query_cs_attr, &query_out);
3206 
3207 	if (err)
3208 		return rte_flow_error_set(error, err,
3209 					  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
3210 					  NULL,
3211 					  "cannot read counter");
3212 	qc->hits_set = 1;
3213 	qc->bytes_set = 1;
3214 	qc->hits = counters[0] - flow->counter->hits;
3215 	qc->bytes = counters[1] - flow->counter->bytes;
3216 	if (qc->reset) {
3217 		flow->counter->hits = counters[0];
3218 		flow->counter->bytes = counters[1];
3219 	}
3220 	return 0;
3221 #endif
3222 	return rte_flow_error_set(error, ENOTSUP,
3223 				  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
3224 				  NULL,
3225 				  "counters are not available");
3226 }
3227 
3228 /**
3229  * Query a flows.
3230  *
3231  * @see rte_flow_query()
3232  * @see rte_flow_ops
3233  */
3234 int
3235 mlx5_flow_query(struct rte_eth_dev *dev __rte_unused,
3236 		struct rte_flow *flow,
3237 		const struct rte_flow_action *actions,
3238 		void *data,
3239 		struct rte_flow_error *error)
3240 {
3241 	int ret = 0;
3242 
3243 	for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
3244 		switch (actions->type) {
3245 		case RTE_FLOW_ACTION_TYPE_VOID:
3246 			break;
3247 		case RTE_FLOW_ACTION_TYPE_COUNT:
3248 			ret = mlx5_flow_query_count(flow, data, error);
3249 			break;
3250 		default:
3251 			return rte_flow_error_set(error, ENOTSUP,
3252 						  RTE_FLOW_ERROR_TYPE_ACTION,
3253 						  actions,
3254 						  "action not supported");
3255 		}
3256 		if (ret < 0)
3257 			return ret;
3258 	}
3259 	return 0;
3260 }
3261 
3262 /**
3263  * Convert a flow director filter to a generic flow.
3264  *
3265  * @param dev
3266  *   Pointer to Ethernet device.
3267  * @param fdir_filter
3268  *   Flow director filter to add.
3269  * @param attributes
3270  *   Generic flow parameters structure.
3271  *
3272  * @return
3273  *   0 on success, a negative errno value otherwise and rte_errno is set.
3274  */
3275 static int
3276 mlx5_fdir_filter_convert(struct rte_eth_dev *dev,
3277 			 const struct rte_eth_fdir_filter *fdir_filter,
3278 			 struct mlx5_fdir *attributes)
3279 {
3280 	struct priv *priv = dev->data->dev_private;
3281 	const struct rte_eth_fdir_input *input = &fdir_filter->input;
3282 	const struct rte_eth_fdir_masks *mask =
3283 		&dev->data->dev_conf.fdir_conf.mask;
3284 
3285 	/* Validate queue number. */
3286 	if (fdir_filter->action.rx_queue >= priv->rxqs_n) {
3287 		DRV_LOG(ERR, "port %u invalid queue number %d",
3288 			dev->data->port_id, fdir_filter->action.rx_queue);
3289 		rte_errno = EINVAL;
3290 		return -rte_errno;
3291 	}
3292 	attributes->attr.ingress = 1;
3293 	attributes->items[0] = (struct rte_flow_item) {
3294 		.type = RTE_FLOW_ITEM_TYPE_ETH,
3295 		.spec = &attributes->l2,
3296 		.mask = &attributes->l2_mask,
3297 	};
3298 	switch (fdir_filter->action.behavior) {
3299 	case RTE_ETH_FDIR_ACCEPT:
3300 		attributes->actions[0] = (struct rte_flow_action){
3301 			.type = RTE_FLOW_ACTION_TYPE_QUEUE,
3302 			.conf = &attributes->queue,
3303 		};
3304 		break;
3305 	case RTE_ETH_FDIR_REJECT:
3306 		attributes->actions[0] = (struct rte_flow_action){
3307 			.type = RTE_FLOW_ACTION_TYPE_DROP,
3308 		};
3309 		break;
3310 	default:
3311 		DRV_LOG(ERR, "port %u invalid behavior %d",
3312 			dev->data->port_id,
3313 			fdir_filter->action.behavior);
3314 		rte_errno = ENOTSUP;
3315 		return -rte_errno;
3316 	}
3317 	attributes->queue.index = fdir_filter->action.rx_queue;
3318 	/* Handle L3. */
3319 	switch (fdir_filter->input.flow_type) {
3320 	case RTE_ETH_FLOW_NONFRAG_IPV4_UDP:
3321 	case RTE_ETH_FLOW_NONFRAG_IPV4_TCP:
3322 	case RTE_ETH_FLOW_NONFRAG_IPV4_OTHER:
3323 		attributes->l3.ipv4.hdr = (struct ipv4_hdr){
3324 			.src_addr = input->flow.ip4_flow.src_ip,
3325 			.dst_addr = input->flow.ip4_flow.dst_ip,
3326 			.time_to_live = input->flow.ip4_flow.ttl,
3327 			.type_of_service = input->flow.ip4_flow.tos,
3328 			.next_proto_id = input->flow.ip4_flow.proto,
3329 		};
3330 		attributes->l3_mask.ipv4.hdr = (struct ipv4_hdr){
3331 			.src_addr = mask->ipv4_mask.src_ip,
3332 			.dst_addr = mask->ipv4_mask.dst_ip,
3333 			.time_to_live = mask->ipv4_mask.ttl,
3334 			.type_of_service = mask->ipv4_mask.tos,
3335 			.next_proto_id = mask->ipv4_mask.proto,
3336 		};
3337 		attributes->items[1] = (struct rte_flow_item){
3338 			.type = RTE_FLOW_ITEM_TYPE_IPV4,
3339 			.spec = &attributes->l3,
3340 			.mask = &attributes->l3_mask,
3341 		};
3342 		break;
3343 	case RTE_ETH_FLOW_NONFRAG_IPV6_UDP:
3344 	case RTE_ETH_FLOW_NONFRAG_IPV6_TCP:
3345 	case RTE_ETH_FLOW_NONFRAG_IPV6_OTHER:
3346 		attributes->l3.ipv6.hdr = (struct ipv6_hdr){
3347 			.hop_limits = input->flow.ipv6_flow.hop_limits,
3348 			.proto = input->flow.ipv6_flow.proto,
3349 		};
3350 
3351 		memcpy(attributes->l3.ipv6.hdr.src_addr,
3352 		       input->flow.ipv6_flow.src_ip,
3353 		       RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
3354 		memcpy(attributes->l3.ipv6.hdr.dst_addr,
3355 		       input->flow.ipv6_flow.dst_ip,
3356 		       RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
3357 		memcpy(attributes->l3_mask.ipv6.hdr.src_addr,
3358 		       mask->ipv6_mask.src_ip,
3359 		       RTE_DIM(attributes->l3_mask.ipv6.hdr.src_addr));
3360 		memcpy(attributes->l3_mask.ipv6.hdr.dst_addr,
3361 		       mask->ipv6_mask.dst_ip,
3362 		       RTE_DIM(attributes->l3_mask.ipv6.hdr.src_addr));
3363 		attributes->items[1] = (struct rte_flow_item){
3364 			.type = RTE_FLOW_ITEM_TYPE_IPV6,
3365 			.spec = &attributes->l3,
3366 			.mask = &attributes->l3_mask,
3367 		};
3368 		break;
3369 	default:
3370 		DRV_LOG(ERR, "port %u invalid flow type%d",
3371 			dev->data->port_id, fdir_filter->input.flow_type);
3372 		rte_errno = ENOTSUP;
3373 		return -rte_errno;
3374 	}
3375 	/* Handle L4. */
3376 	switch (fdir_filter->input.flow_type) {
3377 	case RTE_ETH_FLOW_NONFRAG_IPV4_UDP:
3378 		attributes->l4.udp.hdr = (struct udp_hdr){
3379 			.src_port = input->flow.udp4_flow.src_port,
3380 			.dst_port = input->flow.udp4_flow.dst_port,
3381 		};
3382 		attributes->l4_mask.udp.hdr = (struct udp_hdr){
3383 			.src_port = mask->src_port_mask,
3384 			.dst_port = mask->dst_port_mask,
3385 		};
3386 		attributes->items[2] = (struct rte_flow_item){
3387 			.type = RTE_FLOW_ITEM_TYPE_UDP,
3388 			.spec = &attributes->l4,
3389 			.mask = &attributes->l4_mask,
3390 		};
3391 		break;
3392 	case RTE_ETH_FLOW_NONFRAG_IPV4_TCP:
3393 		attributes->l4.tcp.hdr = (struct tcp_hdr){
3394 			.src_port = input->flow.tcp4_flow.src_port,
3395 			.dst_port = input->flow.tcp4_flow.dst_port,
3396 		};
3397 		attributes->l4_mask.tcp.hdr = (struct tcp_hdr){
3398 			.src_port = mask->src_port_mask,
3399 			.dst_port = mask->dst_port_mask,
3400 		};
3401 		attributes->items[2] = (struct rte_flow_item){
3402 			.type = RTE_FLOW_ITEM_TYPE_TCP,
3403 			.spec = &attributes->l4,
3404 			.mask = &attributes->l4_mask,
3405 		};
3406 		break;
3407 	case RTE_ETH_FLOW_NONFRAG_IPV6_UDP:
3408 		attributes->l4.udp.hdr = (struct udp_hdr){
3409 			.src_port = input->flow.udp6_flow.src_port,
3410 			.dst_port = input->flow.udp6_flow.dst_port,
3411 		};
3412 		attributes->l4_mask.udp.hdr = (struct udp_hdr){
3413 			.src_port = mask->src_port_mask,
3414 			.dst_port = mask->dst_port_mask,
3415 		};
3416 		attributes->items[2] = (struct rte_flow_item){
3417 			.type = RTE_FLOW_ITEM_TYPE_UDP,
3418 			.spec = &attributes->l4,
3419 			.mask = &attributes->l4_mask,
3420 		};
3421 		break;
3422 	case RTE_ETH_FLOW_NONFRAG_IPV6_TCP:
3423 		attributes->l4.tcp.hdr = (struct tcp_hdr){
3424 			.src_port = input->flow.tcp6_flow.src_port,
3425 			.dst_port = input->flow.tcp6_flow.dst_port,
3426 		};
3427 		attributes->l4_mask.tcp.hdr = (struct tcp_hdr){
3428 			.src_port = mask->src_port_mask,
3429 			.dst_port = mask->dst_port_mask,
3430 		};
3431 		attributes->items[2] = (struct rte_flow_item){
3432 			.type = RTE_FLOW_ITEM_TYPE_TCP,
3433 			.spec = &attributes->l4,
3434 			.mask = &attributes->l4_mask,
3435 		};
3436 		break;
3437 	case RTE_ETH_FLOW_NONFRAG_IPV4_OTHER:
3438 	case RTE_ETH_FLOW_NONFRAG_IPV6_OTHER:
3439 		break;
3440 	default:
3441 		DRV_LOG(ERR, "port %u invalid flow type%d",
3442 			dev->data->port_id, fdir_filter->input.flow_type);
3443 		rte_errno = ENOTSUP;
3444 		return -rte_errno;
3445 	}
3446 	return 0;
3447 }
3448 
3449 /**
3450  * Add new flow director filter and store it in list.
3451  *
3452  * @param dev
3453  *   Pointer to Ethernet device.
3454  * @param fdir_filter
3455  *   Flow director filter to add.
3456  *
3457  * @return
3458  *   0 on success, a negative errno value otherwise and rte_errno is set.
3459  */
3460 static int
3461 mlx5_fdir_filter_add(struct rte_eth_dev *dev,
3462 		     const struct rte_eth_fdir_filter *fdir_filter)
3463 {
3464 	struct priv *priv = dev->data->dev_private;
3465 	struct mlx5_fdir attributes = {
3466 		.attr.group = 0,
3467 		.l2_mask = {
3468 			.dst.addr_bytes = "\x00\x00\x00\x00\x00\x00",
3469 			.src.addr_bytes = "\x00\x00\x00\x00\x00\x00",
3470 			.type = 0,
3471 		},
3472 	};
3473 	struct rte_flow_error error;
3474 	struct rte_flow *flow;
3475 	int ret;
3476 
3477 	ret = mlx5_fdir_filter_convert(dev, fdir_filter, &attributes);
3478 	if (ret)
3479 		return ret;
3480 	flow = mlx5_flow_list_create(dev, &priv->flows, &attributes.attr,
3481 				     attributes.items, attributes.actions,
3482 				     &error);
3483 	if (flow) {
3484 		DRV_LOG(DEBUG, "port %u FDIR created %p", dev->data->port_id,
3485 			(void *)flow);
3486 		return 0;
3487 	}
3488 	return -rte_errno;
3489 }
3490 
3491 /**
3492  * Delete specific filter.
3493  *
3494  * @param dev
3495  *   Pointer to Ethernet device.
3496  * @param fdir_filter
3497  *   Filter to be deleted.
3498  *
3499  * @return
3500  *   0 on success, a negative errno value otherwise and rte_errno is set.
3501  */
3502 static int
3503 mlx5_fdir_filter_delete(struct rte_eth_dev *dev __rte_unused,
3504 			const struct rte_eth_fdir_filter *fdir_filter
3505 			__rte_unused)
3506 {
3507 	rte_errno = ENOTSUP;
3508 	return -rte_errno;
3509 }
3510 
3511 /**
3512  * Update queue for specific filter.
3513  *
3514  * @param dev
3515  *   Pointer to Ethernet device.
3516  * @param fdir_filter
3517  *   Filter to be updated.
3518  *
3519  * @return
3520  *   0 on success, a negative errno value otherwise and rte_errno is set.
3521  */
3522 static int
3523 mlx5_fdir_filter_update(struct rte_eth_dev *dev,
3524 			const struct rte_eth_fdir_filter *fdir_filter)
3525 {
3526 	int ret;
3527 
3528 	ret = mlx5_fdir_filter_delete(dev, fdir_filter);
3529 	if (ret)
3530 		return ret;
3531 	return mlx5_fdir_filter_add(dev, fdir_filter);
3532 }
3533 
3534 /**
3535  * Flush all filters.
3536  *
3537  * @param dev
3538  *   Pointer to Ethernet device.
3539  */
3540 static void
3541 mlx5_fdir_filter_flush(struct rte_eth_dev *dev)
3542 {
3543 	struct priv *priv = dev->data->dev_private;
3544 
3545 	mlx5_flow_list_flush(dev, &priv->flows);
3546 }
3547 
3548 /**
3549  * Get flow director information.
3550  *
3551  * @param dev
3552  *   Pointer to Ethernet device.
3553  * @param[out] fdir_info
3554  *   Resulting flow director information.
3555  */
3556 static void
3557 mlx5_fdir_info_get(struct rte_eth_dev *dev, struct rte_eth_fdir_info *fdir_info)
3558 {
3559 	struct rte_eth_fdir_masks *mask =
3560 		&dev->data->dev_conf.fdir_conf.mask;
3561 
3562 	fdir_info->mode = dev->data->dev_conf.fdir_conf.mode;
3563 	fdir_info->guarant_spc = 0;
3564 	rte_memcpy(&fdir_info->mask, mask, sizeof(fdir_info->mask));
3565 	fdir_info->max_flexpayload = 0;
3566 	fdir_info->flow_types_mask[0] = 0;
3567 	fdir_info->flex_payload_unit = 0;
3568 	fdir_info->max_flex_payload_segment_num = 0;
3569 	fdir_info->flex_payload_limit = 0;
3570 	memset(&fdir_info->flex_conf, 0, sizeof(fdir_info->flex_conf));
3571 }
3572 
3573 /**
3574  * Deal with flow director operations.
3575  *
3576  * @param dev
3577  *   Pointer to Ethernet device.
3578  * @param filter_op
3579  *   Operation to perform.
3580  * @param arg
3581  *   Pointer to operation-specific structure.
3582  *
3583  * @return
3584  *   0 on success, a negative errno value otherwise and rte_errno is set.
3585  */
3586 static int
3587 mlx5_fdir_ctrl_func(struct rte_eth_dev *dev, enum rte_filter_op filter_op,
3588 		    void *arg)
3589 {
3590 	enum rte_fdir_mode fdir_mode =
3591 		dev->data->dev_conf.fdir_conf.mode;
3592 
3593 	if (filter_op == RTE_ETH_FILTER_NOP)
3594 		return 0;
3595 	if (fdir_mode != RTE_FDIR_MODE_PERFECT &&
3596 	    fdir_mode != RTE_FDIR_MODE_PERFECT_MAC_VLAN) {
3597 		DRV_LOG(ERR, "port %u flow director mode %d not supported",
3598 			dev->data->port_id, fdir_mode);
3599 		rte_errno = EINVAL;
3600 		return -rte_errno;
3601 	}
3602 	switch (filter_op) {
3603 	case RTE_ETH_FILTER_ADD:
3604 		return mlx5_fdir_filter_add(dev, arg);
3605 	case RTE_ETH_FILTER_UPDATE:
3606 		return mlx5_fdir_filter_update(dev, arg);
3607 	case RTE_ETH_FILTER_DELETE:
3608 		return mlx5_fdir_filter_delete(dev, arg);
3609 	case RTE_ETH_FILTER_FLUSH:
3610 		mlx5_fdir_filter_flush(dev);
3611 		break;
3612 	case RTE_ETH_FILTER_INFO:
3613 		mlx5_fdir_info_get(dev, arg);
3614 		break;
3615 	default:
3616 		DRV_LOG(DEBUG, "port %u unknown operation %u",
3617 			dev->data->port_id, filter_op);
3618 		rte_errno = EINVAL;
3619 		return -rte_errno;
3620 	}
3621 	return 0;
3622 }
3623 
3624 /**
3625  * Manage filter operations.
3626  *
3627  * @param dev
3628  *   Pointer to Ethernet device structure.
3629  * @param filter_type
3630  *   Filter type.
3631  * @param filter_op
3632  *   Operation to perform.
3633  * @param arg
3634  *   Pointer to operation-specific structure.
3635  *
3636  * @return
3637  *   0 on success, a negative errno value otherwise and rte_errno is set.
3638  */
3639 int
3640 mlx5_dev_filter_ctrl(struct rte_eth_dev *dev,
3641 		     enum rte_filter_type filter_type,
3642 		     enum rte_filter_op filter_op,
3643 		     void *arg)
3644 {
3645 	switch (filter_type) {
3646 	case RTE_ETH_FILTER_GENERIC:
3647 		if (filter_op != RTE_ETH_FILTER_GET) {
3648 			rte_errno = EINVAL;
3649 			return -rte_errno;
3650 		}
3651 		*(const void **)arg = &mlx5_flow_ops;
3652 		return 0;
3653 	case RTE_ETH_FILTER_FDIR:
3654 		return mlx5_fdir_ctrl_func(dev, filter_op, arg);
3655 	default:
3656 		DRV_LOG(ERR, "port %u filter type (%d) not supported",
3657 			dev->data->port_id, filter_type);
3658 		rte_errno = ENOTSUP;
3659 		return -rte_errno;
3660 	}
3661 	return 0;
3662 }
3663