xref: /dpdk/drivers/net/mlx5/mlx5_flow.c (revision b733c60f68f12e064359b27e630305c541a3fbdf)
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 <stdalign.h>
8 #include <stdint.h>
9 #include <string.h>
10 
11 /* Verbs header. */
12 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
13 #ifdef PEDANTIC
14 #pragma GCC diagnostic ignored "-Wpedantic"
15 #endif
16 #include <infiniband/verbs.h>
17 #ifdef PEDANTIC
18 #pragma GCC diagnostic error "-Wpedantic"
19 #endif
20 
21 #include <rte_common.h>
22 #include <rte_ether.h>
23 #include <rte_eth_ctrl.h>
24 #include <rte_ethdev_driver.h>
25 #include <rte_flow.h>
26 #include <rte_flow_driver.h>
27 #include <rte_malloc.h>
28 #include <rte_ip.h>
29 
30 #include "mlx5.h"
31 #include "mlx5_defs.h"
32 #include "mlx5_prm.h"
33 #include "mlx5_glue.h"
34 
35 /* Dev ops structure defined in mlx5.c */
36 extern const struct eth_dev_ops mlx5_dev_ops;
37 extern const struct eth_dev_ops mlx5_dev_ops_isolate;
38 
39 /* Pattern outer Layer bits. */
40 #define MLX5_FLOW_LAYER_OUTER_L2 (1u << 0)
41 #define MLX5_FLOW_LAYER_OUTER_L3_IPV4 (1u << 1)
42 #define MLX5_FLOW_LAYER_OUTER_L3_IPV6 (1u << 2)
43 #define MLX5_FLOW_LAYER_OUTER_L4_UDP (1u << 3)
44 #define MLX5_FLOW_LAYER_OUTER_L4_TCP (1u << 4)
45 #define MLX5_FLOW_LAYER_OUTER_VLAN (1u << 5)
46 
47 /* Pattern inner Layer bits. */
48 #define MLX5_FLOW_LAYER_INNER_L2 (1u << 6)
49 #define MLX5_FLOW_LAYER_INNER_L3_IPV4 (1u << 7)
50 #define MLX5_FLOW_LAYER_INNER_L3_IPV6 (1u << 8)
51 #define MLX5_FLOW_LAYER_INNER_L4_UDP (1u << 9)
52 #define MLX5_FLOW_LAYER_INNER_L4_TCP (1u << 10)
53 #define MLX5_FLOW_LAYER_INNER_VLAN (1u << 11)
54 
55 /* Pattern tunnel Layer bits. */
56 #define MLX5_FLOW_LAYER_VXLAN (1u << 12)
57 #define MLX5_FLOW_LAYER_VXLAN_GPE (1u << 13)
58 #define MLX5_FLOW_LAYER_GRE (1u << 14)
59 #define MLX5_FLOW_LAYER_MPLS (1u << 15)
60 
61 /* Outer Masks. */
62 #define MLX5_FLOW_LAYER_OUTER_L3 \
63 	(MLX5_FLOW_LAYER_OUTER_L3_IPV4 | MLX5_FLOW_LAYER_OUTER_L3_IPV6)
64 #define MLX5_FLOW_LAYER_OUTER_L4 \
65 	(MLX5_FLOW_LAYER_OUTER_L4_UDP | MLX5_FLOW_LAYER_OUTER_L4_TCP)
66 #define MLX5_FLOW_LAYER_OUTER \
67 	(MLX5_FLOW_LAYER_OUTER_L2 | MLX5_FLOW_LAYER_OUTER_L3 | \
68 	 MLX5_FLOW_LAYER_OUTER_L4)
69 
70 /* Tunnel Masks. */
71 #define MLX5_FLOW_LAYER_TUNNEL \
72 	(MLX5_FLOW_LAYER_VXLAN | MLX5_FLOW_LAYER_VXLAN_GPE | \
73 	 MLX5_FLOW_LAYER_GRE | MLX5_FLOW_LAYER_MPLS)
74 
75 /* Inner Masks. */
76 #define MLX5_FLOW_LAYER_INNER_L3 \
77 	(MLX5_FLOW_LAYER_INNER_L3_IPV4 | MLX5_FLOW_LAYER_INNER_L3_IPV6)
78 #define MLX5_FLOW_LAYER_INNER_L4 \
79 	(MLX5_FLOW_LAYER_INNER_L4_UDP | MLX5_FLOW_LAYER_INNER_L4_TCP)
80 #define MLX5_FLOW_LAYER_INNER \
81 	(MLX5_FLOW_LAYER_INNER_L2 | MLX5_FLOW_LAYER_INNER_L3 | \
82 	 MLX5_FLOW_LAYER_INNER_L4)
83 
84 /* Actions that modify the fate of matching traffic. */
85 #define MLX5_FLOW_FATE_DROP (1u << 0)
86 #define MLX5_FLOW_FATE_QUEUE (1u << 1)
87 #define MLX5_FLOW_FATE_RSS (1u << 2)
88 
89 /* Modify a packet. */
90 #define MLX5_FLOW_MOD_FLAG (1u << 0)
91 #define MLX5_FLOW_MOD_MARK (1u << 1)
92 #define MLX5_FLOW_MOD_COUNT (1u << 2)
93 
94 /* possible L3 layers protocols filtering. */
95 #define MLX5_IP_PROTOCOL_TCP 6
96 #define MLX5_IP_PROTOCOL_UDP 17
97 #define MLX5_IP_PROTOCOL_GRE 47
98 #define MLX5_IP_PROTOCOL_MPLS 147
99 
100 /* Priority reserved for default flows. */
101 #define MLX5_FLOW_PRIO_RSVD ((uint32_t)-1)
102 
103 enum mlx5_expansion {
104 	MLX5_EXPANSION_ROOT,
105 	MLX5_EXPANSION_ROOT_OUTER,
106 	MLX5_EXPANSION_ROOT_ETH_VLAN,
107 	MLX5_EXPANSION_ROOT_OUTER_ETH_VLAN,
108 	MLX5_EXPANSION_OUTER_ETH,
109 	MLX5_EXPANSION_OUTER_ETH_VLAN,
110 	MLX5_EXPANSION_OUTER_VLAN,
111 	MLX5_EXPANSION_OUTER_IPV4,
112 	MLX5_EXPANSION_OUTER_IPV4_UDP,
113 	MLX5_EXPANSION_OUTER_IPV4_TCP,
114 	MLX5_EXPANSION_OUTER_IPV6,
115 	MLX5_EXPANSION_OUTER_IPV6_UDP,
116 	MLX5_EXPANSION_OUTER_IPV6_TCP,
117 	MLX5_EXPANSION_VXLAN,
118 	MLX5_EXPANSION_VXLAN_GPE,
119 	MLX5_EXPANSION_GRE,
120 	MLX5_EXPANSION_MPLS,
121 	MLX5_EXPANSION_ETH,
122 	MLX5_EXPANSION_ETH_VLAN,
123 	MLX5_EXPANSION_VLAN,
124 	MLX5_EXPANSION_IPV4,
125 	MLX5_EXPANSION_IPV4_UDP,
126 	MLX5_EXPANSION_IPV4_TCP,
127 	MLX5_EXPANSION_IPV6,
128 	MLX5_EXPANSION_IPV6_UDP,
129 	MLX5_EXPANSION_IPV6_TCP,
130 };
131 
132 /** Supported expansion of items. */
133 static const struct rte_flow_expand_node mlx5_support_expansion[] = {
134 	[MLX5_EXPANSION_ROOT] = {
135 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_ETH,
136 						 MLX5_EXPANSION_IPV4,
137 						 MLX5_EXPANSION_IPV6),
138 		.type = RTE_FLOW_ITEM_TYPE_END,
139 	},
140 	[MLX5_EXPANSION_ROOT_OUTER] = {
141 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_OUTER_ETH,
142 						 MLX5_EXPANSION_OUTER_IPV4,
143 						 MLX5_EXPANSION_OUTER_IPV6),
144 		.type = RTE_FLOW_ITEM_TYPE_END,
145 	},
146 	[MLX5_EXPANSION_ROOT_ETH_VLAN] = {
147 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_ETH_VLAN),
148 		.type = RTE_FLOW_ITEM_TYPE_END,
149 	},
150 	[MLX5_EXPANSION_ROOT_OUTER_ETH_VLAN] = {
151 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_OUTER_ETH_VLAN),
152 		.type = RTE_FLOW_ITEM_TYPE_END,
153 	},
154 	[MLX5_EXPANSION_OUTER_ETH] = {
155 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_OUTER_IPV4,
156 						 MLX5_EXPANSION_OUTER_IPV6,
157 						 MLX5_EXPANSION_MPLS),
158 		.type = RTE_FLOW_ITEM_TYPE_ETH,
159 		.rss_types = 0,
160 	},
161 	[MLX5_EXPANSION_OUTER_ETH_VLAN] = {
162 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_OUTER_VLAN),
163 		.type = RTE_FLOW_ITEM_TYPE_ETH,
164 		.rss_types = 0,
165 	},
166 	[MLX5_EXPANSION_OUTER_VLAN] = {
167 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_OUTER_IPV4,
168 						 MLX5_EXPANSION_OUTER_IPV6),
169 		.type = RTE_FLOW_ITEM_TYPE_VLAN,
170 	},
171 	[MLX5_EXPANSION_OUTER_IPV4] = {
172 		.next = RTE_FLOW_EXPAND_RSS_NEXT
173 			(MLX5_EXPANSION_OUTER_IPV4_UDP,
174 			 MLX5_EXPANSION_OUTER_IPV4_TCP,
175 			 MLX5_EXPANSION_GRE),
176 		.type = RTE_FLOW_ITEM_TYPE_IPV4,
177 		.rss_types = ETH_RSS_IPV4 | ETH_RSS_FRAG_IPV4 |
178 			ETH_RSS_NONFRAG_IPV4_OTHER,
179 	},
180 	[MLX5_EXPANSION_OUTER_IPV4_UDP] = {
181 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_VXLAN,
182 						 MLX5_EXPANSION_VXLAN_GPE),
183 		.type = RTE_FLOW_ITEM_TYPE_UDP,
184 		.rss_types = ETH_RSS_NONFRAG_IPV4_UDP,
185 	},
186 	[MLX5_EXPANSION_OUTER_IPV4_TCP] = {
187 		.type = RTE_FLOW_ITEM_TYPE_TCP,
188 		.rss_types = ETH_RSS_NONFRAG_IPV4_TCP,
189 	},
190 	[MLX5_EXPANSION_OUTER_IPV6] = {
191 		.next = RTE_FLOW_EXPAND_RSS_NEXT
192 			(MLX5_EXPANSION_OUTER_IPV6_UDP,
193 			 MLX5_EXPANSION_OUTER_IPV6_TCP),
194 		.type = RTE_FLOW_ITEM_TYPE_IPV6,
195 		.rss_types = ETH_RSS_IPV6 | ETH_RSS_FRAG_IPV6 |
196 			ETH_RSS_NONFRAG_IPV6_OTHER,
197 	},
198 	[MLX5_EXPANSION_OUTER_IPV6_UDP] = {
199 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_VXLAN,
200 						 MLX5_EXPANSION_VXLAN_GPE),
201 		.type = RTE_FLOW_ITEM_TYPE_UDP,
202 		.rss_types = ETH_RSS_NONFRAG_IPV6_UDP,
203 	},
204 	[MLX5_EXPANSION_OUTER_IPV6_TCP] = {
205 		.type = RTE_FLOW_ITEM_TYPE_TCP,
206 		.rss_types = ETH_RSS_NONFRAG_IPV6_TCP,
207 	},
208 	[MLX5_EXPANSION_VXLAN] = {
209 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_ETH),
210 		.type = RTE_FLOW_ITEM_TYPE_VXLAN,
211 	},
212 	[MLX5_EXPANSION_VXLAN_GPE] = {
213 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_ETH,
214 						 MLX5_EXPANSION_IPV4,
215 						 MLX5_EXPANSION_IPV6),
216 		.type = RTE_FLOW_ITEM_TYPE_VXLAN_GPE,
217 	},
218 	[MLX5_EXPANSION_GRE] = {
219 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4),
220 		.type = RTE_FLOW_ITEM_TYPE_GRE,
221 	},
222 	[MLX5_EXPANSION_MPLS] = {
223 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4,
224 						 MLX5_EXPANSION_IPV6),
225 		.type = RTE_FLOW_ITEM_TYPE_MPLS,
226 	},
227 	[MLX5_EXPANSION_ETH] = {
228 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4,
229 						 MLX5_EXPANSION_IPV6),
230 		.type = RTE_FLOW_ITEM_TYPE_ETH,
231 	},
232 	[MLX5_EXPANSION_ETH_VLAN] = {
233 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_VLAN),
234 		.type = RTE_FLOW_ITEM_TYPE_ETH,
235 	},
236 	[MLX5_EXPANSION_VLAN] = {
237 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4,
238 						 MLX5_EXPANSION_IPV6),
239 		.type = RTE_FLOW_ITEM_TYPE_VLAN,
240 	},
241 	[MLX5_EXPANSION_IPV4] = {
242 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4_UDP,
243 						 MLX5_EXPANSION_IPV4_TCP),
244 		.type = RTE_FLOW_ITEM_TYPE_IPV4,
245 		.rss_types = ETH_RSS_IPV4 | ETH_RSS_FRAG_IPV4 |
246 			ETH_RSS_NONFRAG_IPV4_OTHER,
247 	},
248 	[MLX5_EXPANSION_IPV4_UDP] = {
249 		.type = RTE_FLOW_ITEM_TYPE_UDP,
250 		.rss_types = ETH_RSS_NONFRAG_IPV4_UDP,
251 	},
252 	[MLX5_EXPANSION_IPV4_TCP] = {
253 		.type = RTE_FLOW_ITEM_TYPE_TCP,
254 		.rss_types = ETH_RSS_NONFRAG_IPV4_TCP,
255 	},
256 	[MLX5_EXPANSION_IPV6] = {
257 		.next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV6_UDP,
258 						 MLX5_EXPANSION_IPV6_TCP),
259 		.type = RTE_FLOW_ITEM_TYPE_IPV6,
260 		.rss_types = ETH_RSS_IPV6 | ETH_RSS_FRAG_IPV6 |
261 			ETH_RSS_NONFRAG_IPV6_OTHER,
262 	},
263 	[MLX5_EXPANSION_IPV6_UDP] = {
264 		.type = RTE_FLOW_ITEM_TYPE_UDP,
265 		.rss_types = ETH_RSS_NONFRAG_IPV6_UDP,
266 	},
267 	[MLX5_EXPANSION_IPV6_TCP] = {
268 		.type = RTE_FLOW_ITEM_TYPE_TCP,
269 		.rss_types = ETH_RSS_NONFRAG_IPV6_TCP,
270 	},
271 };
272 
273 /** Handles information leading to a drop fate. */
274 struct mlx5_flow_verbs {
275 	LIST_ENTRY(mlx5_flow_verbs) next;
276 	unsigned int size; /**< Size of the attribute. */
277 	struct {
278 		struct ibv_flow_attr *attr;
279 		/**< Pointer to the Specification buffer. */
280 		uint8_t *specs; /**< Pointer to the specifications. */
281 	};
282 	struct ibv_flow *flow; /**< Verbs flow pointer. */
283 	struct mlx5_hrxq *hrxq; /**< Hash Rx queue object. */
284 	uint64_t hash_fields; /**< Verbs hash Rx queue hash fields. */
285 };
286 
287 /* Counters information. */
288 struct mlx5_flow_counter {
289 	LIST_ENTRY(mlx5_flow_counter) next; /**< Pointer to the next counter. */
290 	uint32_t shared:1; /**< Share counter ID with other flow rules. */
291 	uint32_t ref_cnt:31; /**< Reference counter. */
292 	uint32_t id; /**< Counter ID. */
293 	struct ibv_counter_set *cs; /**< Holds the counters for the rule. */
294 	uint64_t hits; /**< Number of packets matched by the rule. */
295 	uint64_t bytes; /**< Number of bytes matched by the rule. */
296 };
297 
298 /* Flow structure. */
299 struct rte_flow {
300 	TAILQ_ENTRY(rte_flow) next; /**< Pointer to the next flow structure. */
301 	struct rte_flow_attr attributes; /**< User flow attribute. */
302 	uint32_t l3_protocol_en:1; /**< Protocol filtering requested. */
303 	uint32_t layers;
304 	/**< Bit-fields of present layers see MLX5_FLOW_LAYER_*. */
305 	uint32_t modifier;
306 	/**< Bit-fields of present modifier see MLX5_FLOW_MOD_*. */
307 	uint32_t fate;
308 	/**< Bit-fields of present fate see MLX5_FLOW_FATE_*. */
309 	uint8_t l3_protocol; /**< valid when l3_protocol_en is set. */
310 	LIST_HEAD(verbs, mlx5_flow_verbs) verbs; /**< Verbs flows list. */
311 	struct mlx5_flow_verbs *cur_verbs;
312 	/**< Current Verbs flow structure being filled. */
313 	struct mlx5_flow_counter *counter; /**< Holds Verbs flow counter. */
314 	struct rte_flow_action_rss rss;/**< RSS context. */
315 	uint8_t key[MLX5_RSS_HASH_KEY_LEN]; /**< RSS hash key. */
316 	uint16_t (*queue)[]; /**< Destination queues to redirect traffic to. */
317 	void *nl_flow; /**< Netlink flow buffer if relevant. */
318 };
319 
320 static const struct rte_flow_ops mlx5_flow_ops = {
321 	.validate = mlx5_flow_validate,
322 	.create = mlx5_flow_create,
323 	.destroy = mlx5_flow_destroy,
324 	.flush = mlx5_flow_flush,
325 	.isolate = mlx5_flow_isolate,
326 	.query = mlx5_flow_query,
327 };
328 
329 /* Convert FDIR request to Generic flow. */
330 struct mlx5_fdir {
331 	struct rte_flow_attr attr;
332 	struct rte_flow_action actions[2];
333 	struct rte_flow_item items[4];
334 	struct rte_flow_item_eth l2;
335 	struct rte_flow_item_eth l2_mask;
336 	union {
337 		struct rte_flow_item_ipv4 ipv4;
338 		struct rte_flow_item_ipv6 ipv6;
339 	} l3;
340 	union {
341 		struct rte_flow_item_ipv4 ipv4;
342 		struct rte_flow_item_ipv6 ipv6;
343 	} l3_mask;
344 	union {
345 		struct rte_flow_item_udp udp;
346 		struct rte_flow_item_tcp tcp;
347 	} l4;
348 	union {
349 		struct rte_flow_item_udp udp;
350 		struct rte_flow_item_tcp tcp;
351 	} l4_mask;
352 	struct rte_flow_action_queue queue;
353 };
354 
355 /* Verbs specification header. */
356 struct ibv_spec_header {
357 	enum ibv_flow_spec_type type;
358 	uint16_t size;
359 };
360 
361 /*
362  * Number of sub priorities.
363  * For each kind of pattern matching i.e. L2, L3, L4 to have a correct
364  * matching on the NIC (firmware dependent) L4 most have the higher priority
365  * followed by L3 and ending with L2.
366  */
367 #define MLX5_PRIORITY_MAP_L2 2
368 #define MLX5_PRIORITY_MAP_L3 1
369 #define MLX5_PRIORITY_MAP_L4 0
370 #define MLX5_PRIORITY_MAP_MAX 3
371 
372 /* Map of Verbs to Flow priority with 8 Verbs priorities. */
373 static const uint32_t priority_map_3[][MLX5_PRIORITY_MAP_MAX] = {
374 	{ 0, 1, 2 }, { 2, 3, 4 }, { 5, 6, 7 },
375 };
376 
377 /* Map of Verbs to Flow priority with 16 Verbs priorities. */
378 static const uint32_t priority_map_5[][MLX5_PRIORITY_MAP_MAX] = {
379 	{ 0, 1, 2 }, { 3, 4, 5 }, { 6, 7, 8 },
380 	{ 9, 10, 11 }, { 12, 13, 14 },
381 };
382 
383 /* Tunnel information. */
384 struct mlx5_flow_tunnel_info {
385 	uint32_t tunnel; /**< Tunnel bit (see MLX5_FLOW_*). */
386 	uint32_t ptype; /**< Tunnel Ptype (see RTE_PTYPE_*). */
387 };
388 
389 static struct mlx5_flow_tunnel_info tunnels_info[] = {
390 	{
391 		.tunnel = MLX5_FLOW_LAYER_VXLAN,
392 		.ptype = RTE_PTYPE_TUNNEL_VXLAN | RTE_PTYPE_L4_UDP,
393 	},
394 	{
395 		.tunnel = MLX5_FLOW_LAYER_VXLAN_GPE,
396 		.ptype = RTE_PTYPE_TUNNEL_VXLAN_GPE | RTE_PTYPE_L4_UDP,
397 	},
398 	{
399 		.tunnel = MLX5_FLOW_LAYER_GRE,
400 		.ptype = RTE_PTYPE_TUNNEL_GRE,
401 	},
402 	{
403 		.tunnel = MLX5_FLOW_LAYER_MPLS | MLX5_FLOW_LAYER_OUTER_L4_UDP,
404 		.ptype = RTE_PTYPE_TUNNEL_MPLS_IN_GRE | RTE_PTYPE_L4_UDP,
405 	},
406 	{
407 		.tunnel = MLX5_FLOW_LAYER_MPLS,
408 		.ptype = RTE_PTYPE_TUNNEL_MPLS_IN_GRE,
409 	},
410 };
411 
412 /**
413  * Discover the maximum number of priority available.
414  *
415  * @param[in] dev
416  *   Pointer to Ethernet device.
417  *
418  * @return
419  *   number of supported flow priority on success, a negative errno
420  *   value otherwise and rte_errno is set.
421  */
422 int
423 mlx5_flow_discover_priorities(struct rte_eth_dev *dev)
424 {
425 	struct {
426 		struct ibv_flow_attr attr;
427 		struct ibv_flow_spec_eth eth;
428 		struct ibv_flow_spec_action_drop drop;
429 	} flow_attr = {
430 		.attr = {
431 			.num_of_specs = 2,
432 		},
433 		.eth = {
434 			.type = IBV_FLOW_SPEC_ETH,
435 			.size = sizeof(struct ibv_flow_spec_eth),
436 		},
437 		.drop = {
438 			.size = sizeof(struct ibv_flow_spec_action_drop),
439 			.type = IBV_FLOW_SPEC_ACTION_DROP,
440 		},
441 	};
442 	struct ibv_flow *flow;
443 	struct mlx5_hrxq *drop = mlx5_hrxq_drop_new(dev);
444 	uint16_t vprio[] = { 8, 16 };
445 	int i;
446 	int priority = 0;
447 
448 	if (!drop) {
449 		rte_errno = ENOTSUP;
450 		return -rte_errno;
451 	}
452 	for (i = 0; i != RTE_DIM(vprio); i++) {
453 		flow_attr.attr.priority = vprio[i] - 1;
454 		flow = mlx5_glue->create_flow(drop->qp, &flow_attr.attr);
455 		if (!flow)
456 			break;
457 		claim_zero(mlx5_glue->destroy_flow(flow));
458 		priority = vprio[i];
459 	}
460 	switch (priority) {
461 	case 8:
462 		priority = RTE_DIM(priority_map_3);
463 		break;
464 	case 16:
465 		priority = RTE_DIM(priority_map_5);
466 		break;
467 	default:
468 		rte_errno = ENOTSUP;
469 		DRV_LOG(ERR,
470 			"port %u verbs maximum priority: %d expected 8/16",
471 			dev->data->port_id, vprio[i]);
472 		return -rte_errno;
473 	}
474 	mlx5_hrxq_drop_release(dev);
475 	DRV_LOG(INFO, "port %u flow maximum priority: %d",
476 		dev->data->port_id, priority);
477 	return priority;
478 }
479 
480 /**
481  * Adjust flow priority.
482  *
483  * @param dev
484  *   Pointer to Ethernet device.
485  * @param flow
486  *   Pointer to an rte flow.
487  */
488 static void
489 mlx5_flow_adjust_priority(struct rte_eth_dev *dev, struct rte_flow *flow)
490 {
491 	struct priv *priv = dev->data->dev_private;
492 	uint32_t priority = flow->attributes.priority;
493 	uint32_t subpriority = flow->cur_verbs->attr->priority;
494 
495 	switch (priv->config.flow_prio) {
496 	case RTE_DIM(priority_map_3):
497 		priority = priority_map_3[priority][subpriority];
498 		break;
499 	case RTE_DIM(priority_map_5):
500 		priority = priority_map_5[priority][subpriority];
501 		break;
502 	}
503 	flow->cur_verbs->attr->priority = priority;
504 }
505 
506 /**
507  * Get a flow counter.
508  *
509  * @param[in] dev
510  *   Pointer to Ethernet device.
511  * @param[in] shared
512  *   Indicate if this counter is shared with other flows.
513  * @param[in] id
514  *   Counter identifier.
515  *
516  * @return
517  *   A pointer to the counter, NULL otherwise and rte_errno is set.
518  */
519 static struct mlx5_flow_counter *
520 mlx5_flow_counter_new(struct rte_eth_dev *dev, uint32_t shared, uint32_t id)
521 {
522 	struct priv *priv = dev->data->dev_private;
523 	struct mlx5_flow_counter *cnt;
524 
525 	LIST_FOREACH(cnt, &priv->flow_counters, next) {
526 		if (!cnt->shared || cnt->shared != shared)
527 			continue;
528 		if (cnt->id != id)
529 			continue;
530 		cnt->ref_cnt++;
531 		return cnt;
532 	}
533 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
534 
535 	struct mlx5_flow_counter tmpl = {
536 		.shared = shared,
537 		.id = id,
538 		.cs = mlx5_glue->create_counter_set
539 			(priv->ctx,
540 			 &(struct ibv_counter_set_init_attr){
541 				 .counter_set_id = id,
542 			 }),
543 		.hits = 0,
544 		.bytes = 0,
545 	};
546 
547 	if (!tmpl.cs) {
548 		rte_errno = errno;
549 		return NULL;
550 	}
551 	cnt = rte_calloc(__func__, 1, sizeof(*cnt), 0);
552 	if (!cnt) {
553 		rte_errno = ENOMEM;
554 		return NULL;
555 	}
556 	*cnt = tmpl;
557 	LIST_INSERT_HEAD(&priv->flow_counters, cnt, next);
558 	return cnt;
559 #endif
560 	rte_errno = ENOTSUP;
561 	return NULL;
562 }
563 
564 /**
565  * Release a flow counter.
566  *
567  * @param[in] counter
568  *   Pointer to the counter handler.
569  */
570 static void
571 mlx5_flow_counter_release(struct mlx5_flow_counter *counter)
572 {
573 	if (--counter->ref_cnt == 0) {
574 		claim_zero(mlx5_glue->destroy_counter_set(counter->cs));
575 		LIST_REMOVE(counter, next);
576 		rte_free(counter);
577 	}
578 }
579 
580 /**
581  * Verify the @p attributes will be correctly understood by the NIC and store
582  * them in the @p flow if everything is correct.
583  *
584  * @param[in] dev
585  *   Pointer to Ethernet device.
586  * @param[in] attributes
587  *   Pointer to flow attributes
588  * @param[in, out] flow
589  *   Pointer to the rte_flow structure.
590  * @param[out] error
591  *   Pointer to error structure.
592  *
593  * @return
594  *   0 on success, a negative errno value otherwise and rte_errno is set.
595  */
596 static int
597 mlx5_flow_attributes(struct rte_eth_dev *dev,
598 		     const struct rte_flow_attr *attributes,
599 		     struct rte_flow *flow,
600 		     struct rte_flow_error *error)
601 {
602 	uint32_t priority_max =
603 		((struct priv *)dev->data->dev_private)->config.flow_prio - 1;
604 
605 	if (attributes->group)
606 		return rte_flow_error_set(error, ENOTSUP,
607 					  RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
608 					  NULL,
609 					  "groups is not supported");
610 	if (attributes->priority != MLX5_FLOW_PRIO_RSVD &&
611 	    attributes->priority >= priority_max)
612 		return rte_flow_error_set(error, ENOTSUP,
613 					  RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
614 					  NULL,
615 					  "priority out of range");
616 	if (attributes->egress)
617 		return rte_flow_error_set(error, ENOTSUP,
618 					  RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
619 					  NULL,
620 					  "egress is not supported");
621 	if (attributes->transfer)
622 		return rte_flow_error_set(error, ENOTSUP,
623 					  RTE_FLOW_ERROR_TYPE_ATTR_TRANSFER,
624 					  NULL,
625 					  "transfer is not supported");
626 	if (!attributes->ingress)
627 		return rte_flow_error_set(error, ENOTSUP,
628 					  RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
629 					  NULL,
630 					  "ingress attribute is mandatory");
631 	flow->attributes = *attributes;
632 	if (attributes->priority == MLX5_FLOW_PRIO_RSVD)
633 		flow->attributes.priority = priority_max;
634 	return 0;
635 }
636 
637 /**
638  * Verify the @p item specifications (spec, last, mask) are compatible with the
639  * NIC capabilities.
640  *
641  * @param[in] item
642  *   Item specification.
643  * @param[in] mask
644  *   @p item->mask or flow default bit-masks.
645  * @param[in] nic_mask
646  *   Bit-masks covering supported fields by the NIC to compare with user mask.
647  * @param[in] size
648  *   Bit-masks size in bytes.
649  * @param[out] error
650  *   Pointer to error structure.
651  *
652  * @return
653  *   0 on success, a negative errno value otherwise and rte_errno is set.
654  */
655 static int
656 mlx5_flow_item_acceptable(const struct rte_flow_item *item,
657 			  const uint8_t *mask,
658 			  const uint8_t *nic_mask,
659 			  unsigned int size,
660 			  struct rte_flow_error *error)
661 {
662 	unsigned int i;
663 
664 	assert(nic_mask);
665 	for (i = 0; i < size; ++i)
666 		if ((nic_mask[i] | mask[i]) != nic_mask[i])
667 			return rte_flow_error_set(error, ENOTSUP,
668 						  RTE_FLOW_ERROR_TYPE_ITEM,
669 						  item,
670 						  "mask enables non supported"
671 						  " bits");
672 	if (!item->spec && (item->mask || item->last))
673 		return rte_flow_error_set(error, EINVAL,
674 					  RTE_FLOW_ERROR_TYPE_ITEM,
675 					  item,
676 					  "mask/last without a spec is not"
677 					  " supported");
678 	if (item->spec && item->last) {
679 		uint8_t spec[size];
680 		uint8_t last[size];
681 		unsigned int i;
682 		int ret;
683 
684 		for (i = 0; i < size; ++i) {
685 			spec[i] = ((const uint8_t *)item->spec)[i] & mask[i];
686 			last[i] = ((const uint8_t *)item->last)[i] & mask[i];
687 		}
688 		ret = memcmp(spec, last, size);
689 		if (ret != 0)
690 			return rte_flow_error_set(error, ENOTSUP,
691 						  RTE_FLOW_ERROR_TYPE_ITEM,
692 						  item,
693 						  "range is not supported");
694 	}
695 	return 0;
696 }
697 
698 /**
699  * Add a verbs item specification into @p flow.
700  *
701  * @param[in, out] flow
702  *   Pointer to flow structure.
703  * @param[in] src
704  *   Create specification.
705  * @param[in] size
706  *   Size in bytes of the specification to copy.
707  */
708 static void
709 mlx5_flow_spec_verbs_add(struct rte_flow *flow, void *src, unsigned int size)
710 {
711 	struct mlx5_flow_verbs *verbs = flow->cur_verbs;
712 
713 	if (verbs->specs) {
714 		void *dst;
715 
716 		dst = (void *)(verbs->specs + verbs->size);
717 		memcpy(dst, src, size);
718 		++verbs->attr->num_of_specs;
719 	}
720 	verbs->size += size;
721 }
722 
723 /**
724  * Adjust verbs hash fields according to the @p flow information.
725  *
726  * @param[in, out] flow.
727  *   Pointer to flow structure.
728  * @param[in] tunnel
729  *   1 when the hash field is for a tunnel item.
730  * @param[in] layer_types
731  *   ETH_RSS_* types.
732  * @param[in] hash_fields
733  *   Item hash fields.
734  */
735 static void
736 mlx5_flow_verbs_hashfields_adjust(struct rte_flow *flow,
737 				  int tunnel __rte_unused,
738 				  uint32_t layer_types, uint64_t hash_fields)
739 {
740 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
741 	hash_fields |= (tunnel ? IBV_RX_HASH_INNER : 0);
742 	if (flow->rss.level == 2 && !tunnel)
743 		hash_fields = 0;
744 	else if (flow->rss.level < 2 && tunnel)
745 		hash_fields = 0;
746 #endif
747 	if (!(flow->rss.types & layer_types))
748 		hash_fields = 0;
749 	flow->cur_verbs->hash_fields |= hash_fields;
750 }
751 
752 /**
753  * Convert the @p item into a Verbs specification after ensuring the NIC
754  * will understand and process it correctly.
755  * If the necessary size for the conversion is greater than the @p flow_size,
756  * nothing is written in @p flow, the validation is still performed.
757  *
758  * @param[in] item
759  *   Item specification.
760  * @param[in, out] flow
761  *   Pointer to flow structure.
762  * @param[in] flow_size
763  *   Size in bytes of the available space in @p flow, if too small, nothing is
764  *   written.
765  * @param[out] error
766  *   Pointer to error structure.
767  *
768  * @return
769  *   On success the number of bytes consumed/necessary, if the returned value
770  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
771  *   otherwise another call with this returned memory size should be done.
772  *   On error, a negative errno value is returned and rte_errno is set.
773  */
774 static int
775 mlx5_flow_item_eth(const struct rte_flow_item *item, struct rte_flow *flow,
776 		   const size_t flow_size, struct rte_flow_error *error)
777 {
778 	const struct rte_flow_item_eth *spec = item->spec;
779 	const struct rte_flow_item_eth *mask = item->mask;
780 	const struct rte_flow_item_eth nic_mask = {
781 		.dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
782 		.src.addr_bytes = "\xff\xff\xff\xff\xff\xff",
783 		.type = RTE_BE16(0xffff),
784 	};
785 	const int tunnel = !!(flow->layers & MLX5_FLOW_LAYER_TUNNEL);
786 	const unsigned int size = sizeof(struct ibv_flow_spec_eth);
787 	struct ibv_flow_spec_eth eth = {
788 		.type = IBV_FLOW_SPEC_ETH | (tunnel ? IBV_FLOW_SPEC_INNER : 0),
789 		.size = size,
790 	};
791 	int ret;
792 
793 	if (flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
794 			    MLX5_FLOW_LAYER_OUTER_L2))
795 		return rte_flow_error_set(error, ENOTSUP,
796 					  RTE_FLOW_ERROR_TYPE_ITEM,
797 					  item,
798 					  "L2 layers already configured");
799 	if (!mask)
800 		mask = &rte_flow_item_eth_mask;
801 	ret = mlx5_flow_item_acceptable(item, (const uint8_t *)mask,
802 					(const uint8_t *)&nic_mask,
803 					sizeof(struct rte_flow_item_eth),
804 					error);
805 	if (ret)
806 		return ret;
807 	flow->layers |= tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
808 		MLX5_FLOW_LAYER_OUTER_L2;
809 	if (size > flow_size)
810 		return size;
811 	if (spec) {
812 		unsigned int i;
813 
814 		memcpy(&eth.val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
815 		memcpy(&eth.val.src_mac, spec->src.addr_bytes, ETHER_ADDR_LEN);
816 		eth.val.ether_type = spec->type;
817 		memcpy(&eth.mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
818 		memcpy(&eth.mask.src_mac, mask->src.addr_bytes, ETHER_ADDR_LEN);
819 		eth.mask.ether_type = mask->type;
820 		/* Remove unwanted bits from values. */
821 		for (i = 0; i < ETHER_ADDR_LEN; ++i) {
822 			eth.val.dst_mac[i] &= eth.mask.dst_mac[i];
823 			eth.val.src_mac[i] &= eth.mask.src_mac[i];
824 		}
825 		eth.val.ether_type &= eth.mask.ether_type;
826 	}
827 	flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L2;
828 	mlx5_flow_spec_verbs_add(flow, &eth, size);
829 	return size;
830 }
831 
832 /**
833  * Update the VLAN tag in the Verbs Ethernet specification.
834  *
835  * @param[in, out] attr
836  *   Pointer to Verbs attributes structure.
837  * @param[in] eth
838  *   Verbs structure containing the VLAN information to copy.
839  */
840 static void
841 mlx5_flow_item_vlan_update(struct ibv_flow_attr *attr,
842 			   struct ibv_flow_spec_eth *eth)
843 {
844 	unsigned int i;
845 	const enum ibv_flow_spec_type search = eth->type;
846 	struct ibv_spec_header *hdr = (struct ibv_spec_header *)
847 		((uint8_t *)attr + sizeof(struct ibv_flow_attr));
848 
849 	for (i = 0; i != attr->num_of_specs; ++i) {
850 		if (hdr->type == search) {
851 			struct ibv_flow_spec_eth *e =
852 				(struct ibv_flow_spec_eth *)hdr;
853 
854 			e->val.vlan_tag = eth->val.vlan_tag;
855 			e->mask.vlan_tag = eth->mask.vlan_tag;
856 			e->val.ether_type = eth->val.ether_type;
857 			e->mask.ether_type = eth->mask.ether_type;
858 			break;
859 		}
860 		hdr = (struct ibv_spec_header *)((uint8_t *)hdr + hdr->size);
861 	}
862 }
863 
864 /**
865  * Convert the @p item into @p flow (or by updating the already present
866  * Ethernet Verbs) specification after ensuring the NIC will understand and
867  * process it correctly.
868  * If the necessary size for the conversion is greater than the @p flow_size,
869  * nothing is written in @p flow, the validation is still performed.
870  *
871  * @param[in] item
872  *   Item specification.
873  * @param[in, out] flow
874  *   Pointer to flow structure.
875  * @param[in] flow_size
876  *   Size in bytes of the available space in @p flow, if too small, nothing is
877  *   written.
878  * @param[out] error
879  *   Pointer to error structure.
880  *
881  * @return
882  *   On success the number of bytes consumed/necessary, if the returned value
883  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
884  *   otherwise another call with this returned memory size should be done.
885  *   On error, a negative errno value is returned and rte_errno is set.
886  */
887 static int
888 mlx5_flow_item_vlan(const struct rte_flow_item *item, struct rte_flow *flow,
889 		    const size_t flow_size, struct rte_flow_error *error)
890 {
891 	const struct rte_flow_item_vlan *spec = item->spec;
892 	const struct rte_flow_item_vlan *mask = item->mask;
893 	const struct rte_flow_item_vlan nic_mask = {
894 		.tci = RTE_BE16(0x0fff),
895 		.inner_type = RTE_BE16(0xffff),
896 	};
897 	unsigned int size = sizeof(struct ibv_flow_spec_eth);
898 	const int tunnel = !!(flow->layers & MLX5_FLOW_LAYER_TUNNEL);
899 	struct ibv_flow_spec_eth eth = {
900 		.type = IBV_FLOW_SPEC_ETH | (tunnel ? IBV_FLOW_SPEC_INNER : 0),
901 		.size = size,
902 	};
903 	int ret;
904 	const uint32_t l34m = tunnel ? (MLX5_FLOW_LAYER_INNER_L3 |
905 					MLX5_FLOW_LAYER_INNER_L4) :
906 		(MLX5_FLOW_LAYER_OUTER_L3 | MLX5_FLOW_LAYER_OUTER_L4);
907 	const uint32_t vlanm = tunnel ? MLX5_FLOW_LAYER_INNER_VLAN :
908 		MLX5_FLOW_LAYER_OUTER_VLAN;
909 	const uint32_t l2m = tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
910 		MLX5_FLOW_LAYER_OUTER_L2;
911 
912 	if (flow->layers & vlanm)
913 		return rte_flow_error_set(error, ENOTSUP,
914 					  RTE_FLOW_ERROR_TYPE_ITEM,
915 					  item,
916 					  "VLAN layer already configured");
917 	else if ((flow->layers & l34m) != 0)
918 		return rte_flow_error_set(error, ENOTSUP,
919 					  RTE_FLOW_ERROR_TYPE_ITEM,
920 					  item,
921 					  "L2 layer cannot follow L3/L4 layer");
922 	if (!mask)
923 		mask = &rte_flow_item_vlan_mask;
924 	ret = mlx5_flow_item_acceptable
925 		(item, (const uint8_t *)mask,
926 		 (const uint8_t *)&nic_mask,
927 		 sizeof(struct rte_flow_item_vlan), error);
928 	if (ret)
929 		return ret;
930 	if (spec) {
931 		eth.val.vlan_tag = spec->tci;
932 		eth.mask.vlan_tag = mask->tci;
933 		eth.val.vlan_tag &= eth.mask.vlan_tag;
934 		eth.val.ether_type = spec->inner_type;
935 		eth.mask.ether_type = mask->inner_type;
936 		eth.val.ether_type &= eth.mask.ether_type;
937 	}
938 	/*
939 	 * From verbs perspective an empty VLAN is equivalent
940 	 * to a packet without VLAN layer.
941 	 */
942 	if (!eth.mask.vlan_tag)
943 		return rte_flow_error_set(error, EINVAL,
944 					  RTE_FLOW_ERROR_TYPE_ITEM_SPEC,
945 					  item->spec,
946 					  "VLAN cannot be empty");
947 	if (!(flow->layers & l2m)) {
948 		if (size <= flow_size) {
949 			flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L2;
950 			mlx5_flow_spec_verbs_add(flow, &eth, size);
951 		}
952 	} else {
953 		if (flow->cur_verbs)
954 			mlx5_flow_item_vlan_update(flow->cur_verbs->attr,
955 						   &eth);
956 		size = 0; /* Only an update is done in eth specification. */
957 	}
958 	flow->layers |= tunnel ?
959 		(MLX5_FLOW_LAYER_INNER_L2 | MLX5_FLOW_LAYER_INNER_VLAN) :
960 		(MLX5_FLOW_LAYER_OUTER_L2 | MLX5_FLOW_LAYER_OUTER_VLAN);
961 	return size;
962 }
963 
964 /**
965  * Convert the @p item into a Verbs specification after ensuring the NIC
966  * will understand and process it correctly.
967  * If the necessary size for the conversion is greater than the @p flow_size,
968  * nothing is written in @p flow, the validation is still performed.
969  *
970  * @param[in] item
971  *   Item specification.
972  * @param[in, out] flow
973  *   Pointer to flow structure.
974  * @param[in] flow_size
975  *   Size in bytes of the available space in @p flow, if too small, nothing is
976  *   written.
977  * @param[out] error
978  *   Pointer to error structure.
979  *
980  * @return
981  *   On success the number of bytes consumed/necessary, if the returned value
982  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
983  *   otherwise another call with this returned memory size should be done.
984  *   On error, a negative errno value is returned and rte_errno is set.
985  */
986 static int
987 mlx5_flow_item_ipv4(const struct rte_flow_item *item, struct rte_flow *flow,
988 		    const size_t flow_size, struct rte_flow_error *error)
989 {
990 	const struct rte_flow_item_ipv4 *spec = item->spec;
991 	const struct rte_flow_item_ipv4 *mask = item->mask;
992 	const struct rte_flow_item_ipv4 nic_mask = {
993 		.hdr = {
994 			.src_addr = RTE_BE32(0xffffffff),
995 			.dst_addr = RTE_BE32(0xffffffff),
996 			.type_of_service = 0xff,
997 			.next_proto_id = 0xff,
998 		},
999 	};
1000 	const int tunnel = !!(flow->layers & MLX5_FLOW_LAYER_TUNNEL);
1001 	unsigned int size = sizeof(struct ibv_flow_spec_ipv4_ext);
1002 	struct ibv_flow_spec_ipv4_ext ipv4 = {
1003 		.type = IBV_FLOW_SPEC_IPV4_EXT |
1004 			(tunnel ? IBV_FLOW_SPEC_INNER : 0),
1005 		.size = size,
1006 	};
1007 	int ret;
1008 
1009 	if (flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L3 :
1010 			    MLX5_FLOW_LAYER_OUTER_L3))
1011 		return rte_flow_error_set(error, ENOTSUP,
1012 					  RTE_FLOW_ERROR_TYPE_ITEM,
1013 					  item,
1014 					  "multiple L3 layers not supported");
1015 	else if (flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L4 :
1016 				 MLX5_FLOW_LAYER_OUTER_L4))
1017 		return rte_flow_error_set(error, ENOTSUP,
1018 					  RTE_FLOW_ERROR_TYPE_ITEM,
1019 					  item,
1020 					  "L3 cannot follow an L4 layer.");
1021 	if (!mask)
1022 		mask = &rte_flow_item_ipv4_mask;
1023 	ret = mlx5_flow_item_acceptable
1024 		(item, (const uint8_t *)mask,
1025 		 (const uint8_t *)&nic_mask,
1026 		 sizeof(struct rte_flow_item_ipv4), error);
1027 	if (ret < 0)
1028 		return ret;
1029 	flow->layers |= tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV4 :
1030 		MLX5_FLOW_LAYER_OUTER_L3_IPV4;
1031 	if (spec) {
1032 		ipv4.val = (struct ibv_flow_ipv4_ext_filter){
1033 			.src_ip = spec->hdr.src_addr,
1034 			.dst_ip = spec->hdr.dst_addr,
1035 			.proto = spec->hdr.next_proto_id,
1036 			.tos = spec->hdr.type_of_service,
1037 		};
1038 		ipv4.mask = (struct ibv_flow_ipv4_ext_filter){
1039 			.src_ip = mask->hdr.src_addr,
1040 			.dst_ip = mask->hdr.dst_addr,
1041 			.proto = mask->hdr.next_proto_id,
1042 			.tos = mask->hdr.type_of_service,
1043 		};
1044 		/* Remove unwanted bits from values. */
1045 		ipv4.val.src_ip &= ipv4.mask.src_ip;
1046 		ipv4.val.dst_ip &= ipv4.mask.dst_ip;
1047 		ipv4.val.proto &= ipv4.mask.proto;
1048 		ipv4.val.tos &= ipv4.mask.tos;
1049 	}
1050 	flow->l3_protocol_en = !!ipv4.mask.proto;
1051 	flow->l3_protocol = ipv4.val.proto;
1052 	if (size <= flow_size) {
1053 		mlx5_flow_verbs_hashfields_adjust
1054 			(flow, tunnel,
1055 			 (ETH_RSS_IPV4 | ETH_RSS_FRAG_IPV4 |
1056 			  ETH_RSS_NONFRAG_IPV4_TCP |
1057 			  ETH_RSS_NONFRAG_IPV4_UDP |
1058 			  ETH_RSS_NONFRAG_IPV4_OTHER),
1059 			 (IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4));
1060 		flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L3;
1061 		mlx5_flow_spec_verbs_add(flow, &ipv4, size);
1062 	}
1063 	return size;
1064 }
1065 
1066 /**
1067  * Convert the @p item into a Verbs specification after ensuring the NIC
1068  * will understand and process it correctly.
1069  * If the necessary size for the conversion is greater than the @p flow_size,
1070  * nothing is written in @p flow, the validation is still performed.
1071  *
1072  * @param[in] item
1073  *   Item specification.
1074  * @param[in, out] flow
1075  *   Pointer to flow structure.
1076  * @param[in] flow_size
1077  *   Size in bytes of the available space in @p flow, if too small, nothing is
1078  *   written.
1079  * @param[out] error
1080  *   Pointer to error structure.
1081  *
1082  * @return
1083  *   On success the number of bytes consumed/necessary, if the returned value
1084  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
1085  *   otherwise another call with this returned memory size should be done.
1086  *   On error, a negative errno value is returned and rte_errno is set.
1087  */
1088 static int
1089 mlx5_flow_item_ipv6(const struct rte_flow_item *item, struct rte_flow *flow,
1090 		    const size_t flow_size, struct rte_flow_error *error)
1091 {
1092 	const struct rte_flow_item_ipv6 *spec = item->spec;
1093 	const struct rte_flow_item_ipv6 *mask = item->mask;
1094 	const struct rte_flow_item_ipv6 nic_mask = {
1095 		.hdr = {
1096 			.src_addr =
1097 				"\xff\xff\xff\xff\xff\xff\xff\xff"
1098 				"\xff\xff\xff\xff\xff\xff\xff\xff",
1099 			.dst_addr =
1100 				"\xff\xff\xff\xff\xff\xff\xff\xff"
1101 				"\xff\xff\xff\xff\xff\xff\xff\xff",
1102 			.vtc_flow = RTE_BE32(0xffffffff),
1103 			.proto = 0xff,
1104 			.hop_limits = 0xff,
1105 		},
1106 	};
1107 	const int tunnel = !!(flow->layers & MLX5_FLOW_LAYER_TUNNEL);
1108 	unsigned int size = sizeof(struct ibv_flow_spec_ipv6);
1109 	struct ibv_flow_spec_ipv6 ipv6 = {
1110 		.type = IBV_FLOW_SPEC_IPV6 | (tunnel ? IBV_FLOW_SPEC_INNER : 0),
1111 		.size = size,
1112 	};
1113 	int ret;
1114 
1115 	if (flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L3 :
1116 			    MLX5_FLOW_LAYER_OUTER_L3))
1117 		return rte_flow_error_set(error, ENOTSUP,
1118 					  RTE_FLOW_ERROR_TYPE_ITEM,
1119 					  item,
1120 					  "multiple L3 layers not supported");
1121 	else if (flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L4 :
1122 				 MLX5_FLOW_LAYER_OUTER_L4))
1123 		return rte_flow_error_set(error, ENOTSUP,
1124 					  RTE_FLOW_ERROR_TYPE_ITEM,
1125 					  item,
1126 					  "L3 cannot follow an L4 layer.");
1127 	/*
1128 	 * IPv6 is not recognised by the NIC inside a GRE tunnel.
1129 	 * Such support has to be disabled as the rule will be
1130 	 * accepted.  Issue reproduced with Mellanox OFED 4.3-3.0.2.1 and
1131 	 * Mellanox OFED 4.4-1.0.0.0.
1132 	 */
1133 	if (tunnel && flow->layers & MLX5_FLOW_LAYER_GRE)
1134 		return rte_flow_error_set(error, ENOTSUP,
1135 					  RTE_FLOW_ERROR_TYPE_ITEM,
1136 					  item,
1137 					  "IPv6 inside a GRE tunnel is"
1138 					  " not recognised.");
1139 	if (!mask)
1140 		mask = &rte_flow_item_ipv6_mask;
1141 	ret = mlx5_flow_item_acceptable
1142 		(item, (const uint8_t *)mask,
1143 		 (const uint8_t *)&nic_mask,
1144 		 sizeof(struct rte_flow_item_ipv6), error);
1145 	if (ret < 0)
1146 		return ret;
1147 	flow->layers |= tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV6 :
1148 		MLX5_FLOW_LAYER_OUTER_L3_IPV6;
1149 	if (spec) {
1150 		unsigned int i;
1151 		uint32_t vtc_flow_val;
1152 		uint32_t vtc_flow_mask;
1153 
1154 		memcpy(&ipv6.val.src_ip, spec->hdr.src_addr,
1155 		       RTE_DIM(ipv6.val.src_ip));
1156 		memcpy(&ipv6.val.dst_ip, spec->hdr.dst_addr,
1157 		       RTE_DIM(ipv6.val.dst_ip));
1158 		memcpy(&ipv6.mask.src_ip, mask->hdr.src_addr,
1159 		       RTE_DIM(ipv6.mask.src_ip));
1160 		memcpy(&ipv6.mask.dst_ip, mask->hdr.dst_addr,
1161 		       RTE_DIM(ipv6.mask.dst_ip));
1162 		vtc_flow_val = rte_be_to_cpu_32(spec->hdr.vtc_flow);
1163 		vtc_flow_mask = rte_be_to_cpu_32(mask->hdr.vtc_flow);
1164 		ipv6.val.flow_label =
1165 			rte_cpu_to_be_32((vtc_flow_val & IPV6_HDR_FL_MASK) >>
1166 					 IPV6_HDR_FL_SHIFT);
1167 		ipv6.val.traffic_class = (vtc_flow_val & IPV6_HDR_TC_MASK) >>
1168 					 IPV6_HDR_TC_SHIFT;
1169 		ipv6.val.next_hdr = spec->hdr.proto;
1170 		ipv6.val.hop_limit = spec->hdr.hop_limits;
1171 		ipv6.mask.flow_label =
1172 			rte_cpu_to_be_32((vtc_flow_mask & IPV6_HDR_FL_MASK) >>
1173 					 IPV6_HDR_FL_SHIFT);
1174 		ipv6.mask.traffic_class = (vtc_flow_mask & IPV6_HDR_TC_MASK) >>
1175 					  IPV6_HDR_TC_SHIFT;
1176 		ipv6.mask.next_hdr = mask->hdr.proto;
1177 		ipv6.mask.hop_limit = mask->hdr.hop_limits;
1178 		/* Remove unwanted bits from values. */
1179 		for (i = 0; i < RTE_DIM(ipv6.val.src_ip); ++i) {
1180 			ipv6.val.src_ip[i] &= ipv6.mask.src_ip[i];
1181 			ipv6.val.dst_ip[i] &= ipv6.mask.dst_ip[i];
1182 		}
1183 		ipv6.val.flow_label &= ipv6.mask.flow_label;
1184 		ipv6.val.traffic_class &= ipv6.mask.traffic_class;
1185 		ipv6.val.next_hdr &= ipv6.mask.next_hdr;
1186 		ipv6.val.hop_limit &= ipv6.mask.hop_limit;
1187 	}
1188 	flow->l3_protocol_en = !!ipv6.mask.next_hdr;
1189 	flow->l3_protocol = ipv6.val.next_hdr;
1190 	if (size <= flow_size) {
1191 		mlx5_flow_verbs_hashfields_adjust
1192 			(flow, tunnel,
1193 			 (ETH_RSS_IPV6 | ETH_RSS_FRAG_IPV6 |
1194 			  ETH_RSS_NONFRAG_IPV6_TCP | ETH_RSS_NONFRAG_IPV6_UDP |
1195 			  ETH_RSS_NONFRAG_IPV6_OTHER | ETH_RSS_IPV6_EX |
1196 			  ETH_RSS_IPV6_TCP_EX | ETH_RSS_IPV6_UDP_EX),
1197 			 (IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6));
1198 		flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L3;
1199 		mlx5_flow_spec_verbs_add(flow, &ipv6, size);
1200 	}
1201 	return size;
1202 }
1203 
1204 /**
1205  * Convert the @p item into a Verbs specification after ensuring the NIC
1206  * will understand and process it correctly.
1207  * If the necessary size for the conversion is greater than the @p flow_size,
1208  * nothing is written in @p flow, the validation is still performed.
1209  *
1210  * @param[in] item
1211  *   Item specification.
1212  * @param[in, out] flow
1213  *   Pointer to flow structure.
1214  * @param[in] flow_size
1215  *   Size in bytes of the available space in @p flow, if too small, nothing is
1216  *   written.
1217  * @param[out] error
1218  *   Pointer to error structure.
1219  *
1220  * @return
1221  *   On success the number of bytes consumed/necessary, if the returned value
1222  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
1223  *   otherwise another call with this returned memory size should be done.
1224  *   On error, a negative errno value is returned and rte_errno is set.
1225  */
1226 static int
1227 mlx5_flow_item_udp(const struct rte_flow_item *item, struct rte_flow *flow,
1228 		   const size_t flow_size, struct rte_flow_error *error)
1229 {
1230 	const struct rte_flow_item_udp *spec = item->spec;
1231 	const struct rte_flow_item_udp *mask = item->mask;
1232 	const int tunnel = !!(flow->layers & MLX5_FLOW_LAYER_TUNNEL);
1233 	unsigned int size = sizeof(struct ibv_flow_spec_tcp_udp);
1234 	struct ibv_flow_spec_tcp_udp udp = {
1235 		.type = IBV_FLOW_SPEC_UDP | (tunnel ? IBV_FLOW_SPEC_INNER : 0),
1236 		.size = size,
1237 	};
1238 	int ret;
1239 
1240 	if (flow->l3_protocol_en && flow->l3_protocol != MLX5_IP_PROTOCOL_UDP)
1241 		return rte_flow_error_set(error, ENOTSUP,
1242 					  RTE_FLOW_ERROR_TYPE_ITEM,
1243 					  item,
1244 					  "protocol filtering not compatible"
1245 					  " with UDP layer");
1246 	if (!(flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L3 :
1247 			      MLX5_FLOW_LAYER_OUTER_L3)))
1248 		return rte_flow_error_set(error, ENOTSUP,
1249 					  RTE_FLOW_ERROR_TYPE_ITEM,
1250 					  item,
1251 					  "L3 is mandatory to filter"
1252 					  " on L4");
1253 	if (flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L4 :
1254 			    MLX5_FLOW_LAYER_OUTER_L4))
1255 		return rte_flow_error_set(error, ENOTSUP,
1256 					  RTE_FLOW_ERROR_TYPE_ITEM,
1257 					  item,
1258 					  "L4 layer is already"
1259 					  " present");
1260 	if (!mask)
1261 		mask = &rte_flow_item_udp_mask;
1262 	ret = mlx5_flow_item_acceptable
1263 		(item, (const uint8_t *)mask,
1264 		 (const uint8_t *)&rte_flow_item_udp_mask,
1265 		 sizeof(struct rte_flow_item_udp), error);
1266 	if (ret < 0)
1267 		return ret;
1268 	flow->layers |= tunnel ? MLX5_FLOW_LAYER_INNER_L4_UDP :
1269 		MLX5_FLOW_LAYER_OUTER_L4_UDP;
1270 	if (spec) {
1271 		udp.val.dst_port = spec->hdr.dst_port;
1272 		udp.val.src_port = spec->hdr.src_port;
1273 		udp.mask.dst_port = mask->hdr.dst_port;
1274 		udp.mask.src_port = mask->hdr.src_port;
1275 		/* Remove unwanted bits from values. */
1276 		udp.val.src_port &= udp.mask.src_port;
1277 		udp.val.dst_port &= udp.mask.dst_port;
1278 	}
1279 	if (size <= flow_size) {
1280 		mlx5_flow_verbs_hashfields_adjust(flow, tunnel, ETH_RSS_UDP,
1281 						  (IBV_RX_HASH_SRC_PORT_UDP |
1282 						   IBV_RX_HASH_DST_PORT_UDP));
1283 		flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L4;
1284 		mlx5_flow_spec_verbs_add(flow, &udp, size);
1285 	}
1286 	return size;
1287 }
1288 
1289 /**
1290  * Convert the @p item into a Verbs specification after ensuring the NIC
1291  * will understand and process it correctly.
1292  * If the necessary size for the conversion is greater than the @p flow_size,
1293  * nothing is written in @p flow, the validation is still performed.
1294  *
1295  * @param[in] item
1296  *   Item specification.
1297  * @param[in, out] flow
1298  *   Pointer to flow structure.
1299  * @param[in] flow_size
1300  *   Size in bytes of the available space in @p flow, if too small, nothing is
1301  *   written.
1302  * @param[out] error
1303  *   Pointer to error structure.
1304  *
1305  * @return
1306  *   On success the number of bytes consumed/necessary, if the returned value
1307  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
1308  *   otherwise another call with this returned memory size should be done.
1309  *   On error, a negative errno value is returned and rte_errno is set.
1310  */
1311 static int
1312 mlx5_flow_item_tcp(const struct rte_flow_item *item, struct rte_flow *flow,
1313 		   const size_t flow_size, struct rte_flow_error *error)
1314 {
1315 	const struct rte_flow_item_tcp *spec = item->spec;
1316 	const struct rte_flow_item_tcp *mask = item->mask;
1317 	const int tunnel = !!(flow->layers & MLX5_FLOW_LAYER_TUNNEL);
1318 	unsigned int size = sizeof(struct ibv_flow_spec_tcp_udp);
1319 	struct ibv_flow_spec_tcp_udp tcp = {
1320 		.type = IBV_FLOW_SPEC_TCP | (tunnel ? IBV_FLOW_SPEC_INNER : 0),
1321 		.size = size,
1322 	};
1323 	int ret;
1324 
1325 	if (flow->l3_protocol_en && flow->l3_protocol != MLX5_IP_PROTOCOL_TCP)
1326 		return rte_flow_error_set(error, ENOTSUP,
1327 					  RTE_FLOW_ERROR_TYPE_ITEM,
1328 					  item,
1329 					  "protocol filtering not compatible"
1330 					  " with TCP layer");
1331 	if (!(flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L3 :
1332 			      MLX5_FLOW_LAYER_OUTER_L3)))
1333 		return rte_flow_error_set(error, ENOTSUP,
1334 					  RTE_FLOW_ERROR_TYPE_ITEM,
1335 					  item,
1336 					  "L3 is mandatory to filter on L4");
1337 	if (flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L4 :
1338 			    MLX5_FLOW_LAYER_OUTER_L4))
1339 		return rte_flow_error_set(error, ENOTSUP,
1340 					  RTE_FLOW_ERROR_TYPE_ITEM,
1341 					  item,
1342 					  "L4 layer is already present");
1343 	if (!mask)
1344 		mask = &rte_flow_item_tcp_mask;
1345 	ret = mlx5_flow_item_acceptable
1346 		(item, (const uint8_t *)mask,
1347 		 (const uint8_t *)&rte_flow_item_tcp_mask,
1348 		 sizeof(struct rte_flow_item_tcp), error);
1349 	if (ret < 0)
1350 		return ret;
1351 	flow->layers |=  tunnel ? MLX5_FLOW_LAYER_INNER_L4_TCP :
1352 		MLX5_FLOW_LAYER_OUTER_L4_TCP;
1353 	if (spec) {
1354 		tcp.val.dst_port = spec->hdr.dst_port;
1355 		tcp.val.src_port = spec->hdr.src_port;
1356 		tcp.mask.dst_port = mask->hdr.dst_port;
1357 		tcp.mask.src_port = mask->hdr.src_port;
1358 		/* Remove unwanted bits from values. */
1359 		tcp.val.src_port &= tcp.mask.src_port;
1360 		tcp.val.dst_port &= tcp.mask.dst_port;
1361 	}
1362 	if (size <= flow_size) {
1363 		mlx5_flow_verbs_hashfields_adjust(flow, tunnel, ETH_RSS_TCP,
1364 						  (IBV_RX_HASH_SRC_PORT_TCP |
1365 						   IBV_RX_HASH_DST_PORT_TCP));
1366 		flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L4;
1367 		mlx5_flow_spec_verbs_add(flow, &tcp, size);
1368 	}
1369 	return size;
1370 }
1371 
1372 /**
1373  * Convert the @p item into a Verbs specification after ensuring the NIC
1374  * will understand and process it correctly.
1375  * If the necessary size for the conversion is greater than the @p flow_size,
1376  * nothing is written in @p flow, the validation is still performed.
1377  *
1378  * @param[in] item
1379  *   Item specification.
1380  * @param[in, out] flow
1381  *   Pointer to flow structure.
1382  * @param[in] flow_size
1383  *   Size in bytes of the available space in @p flow, if too small, nothing is
1384  *   written.
1385  * @param[out] error
1386  *   Pointer to error structure.
1387  *
1388  * @return
1389  *   On success the number of bytes consumed/necessary, if the returned value
1390  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
1391  *   otherwise another call with this returned memory size should be done.
1392  *   On error, a negative errno value is returned and rte_errno is set.
1393  */
1394 static int
1395 mlx5_flow_item_vxlan(const struct rte_flow_item *item, struct rte_flow *flow,
1396 		     const size_t flow_size, struct rte_flow_error *error)
1397 {
1398 	const struct rte_flow_item_vxlan *spec = item->spec;
1399 	const struct rte_flow_item_vxlan *mask = item->mask;
1400 	unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
1401 	struct ibv_flow_spec_tunnel vxlan = {
1402 		.type = IBV_FLOW_SPEC_VXLAN_TUNNEL,
1403 		.size = size,
1404 	};
1405 	int ret;
1406 	union vni {
1407 		uint32_t vlan_id;
1408 		uint8_t vni[4];
1409 	} id = { .vlan_id = 0, };
1410 
1411 	if (flow->layers & MLX5_FLOW_LAYER_TUNNEL)
1412 		return rte_flow_error_set(error, ENOTSUP,
1413 					  RTE_FLOW_ERROR_TYPE_ITEM,
1414 					  item,
1415 					  "a tunnel is already present");
1416 	/*
1417 	 * Verify only UDPv4 is present as defined in
1418 	 * https://tools.ietf.org/html/rfc7348
1419 	 */
1420 	if (!(flow->layers & MLX5_FLOW_LAYER_OUTER_L4_UDP))
1421 		return rte_flow_error_set(error, ENOTSUP,
1422 					  RTE_FLOW_ERROR_TYPE_ITEM,
1423 					  item,
1424 					  "no outer UDP layer found");
1425 	if (!mask)
1426 		mask = &rte_flow_item_vxlan_mask;
1427 	ret = mlx5_flow_item_acceptable
1428 		(item, (const uint8_t *)mask,
1429 		 (const uint8_t *)&rte_flow_item_vxlan_mask,
1430 		 sizeof(struct rte_flow_item_vxlan), error);
1431 	if (ret < 0)
1432 		return ret;
1433 	if (spec) {
1434 		memcpy(&id.vni[1], spec->vni, 3);
1435 		vxlan.val.tunnel_id = id.vlan_id;
1436 		memcpy(&id.vni[1], mask->vni, 3);
1437 		vxlan.mask.tunnel_id = id.vlan_id;
1438 		/* Remove unwanted bits from values. */
1439 		vxlan.val.tunnel_id &= vxlan.mask.tunnel_id;
1440 	}
1441 	/*
1442 	 * Tunnel id 0 is equivalent as not adding a VXLAN layer, if
1443 	 * only this layer is defined in the Verbs specification it is
1444 	 * interpreted as wildcard and all packets will match this
1445 	 * rule, if it follows a full stack layer (ex: eth / ipv4 /
1446 	 * udp), all packets matching the layers before will also
1447 	 * match this rule.  To avoid such situation, VNI 0 is
1448 	 * currently refused.
1449 	 */
1450 	if (!vxlan.val.tunnel_id)
1451 		return rte_flow_error_set(error, EINVAL,
1452 					  RTE_FLOW_ERROR_TYPE_ITEM,
1453 					  item,
1454 					  "VXLAN vni cannot be 0");
1455 	if (!(flow->layers & MLX5_FLOW_LAYER_OUTER))
1456 		return rte_flow_error_set(error, EINVAL,
1457 					  RTE_FLOW_ERROR_TYPE_ITEM,
1458 					  item,
1459 					  "VXLAN tunnel must be fully defined");
1460 	if (size <= flow_size) {
1461 		mlx5_flow_spec_verbs_add(flow, &vxlan, size);
1462 		flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L2;
1463 	}
1464 	flow->layers |= MLX5_FLOW_LAYER_VXLAN;
1465 	return size;
1466 }
1467 
1468 /**
1469  * Convert the @p item into a Verbs specification after ensuring the NIC
1470  * will understand and process it correctly.
1471  * If the necessary size for the conversion is greater than the @p flow_size,
1472  * nothing is written in @p flow, the validation is still performed.
1473  *
1474  * @param dev
1475  *   Pointer to Ethernet device.
1476  * @param[in] item
1477  *   Item specification.
1478  * @param[in, out] flow
1479  *   Pointer to flow structure.
1480  * @param[in] flow_size
1481  *   Size in bytes of the available space in @p flow, if too small, nothing is
1482  *   written.
1483  * @param[out] error
1484  *   Pointer to error structure.
1485  *
1486  * @return
1487  *   On success the number of bytes consumed/necessary, if the returned value
1488  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
1489  *   otherwise another call with this returned memory size should be done.
1490  *   On error, a negative errno value is returned and rte_errno is set.
1491  */
1492 static int
1493 mlx5_flow_item_vxlan_gpe(struct rte_eth_dev *dev,
1494 			 const struct rte_flow_item *item,
1495 			 struct rte_flow *flow, const size_t flow_size,
1496 			 struct rte_flow_error *error)
1497 {
1498 	const struct rte_flow_item_vxlan_gpe *spec = item->spec;
1499 	const struct rte_flow_item_vxlan_gpe *mask = item->mask;
1500 	unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
1501 	struct ibv_flow_spec_tunnel vxlan_gpe = {
1502 		.type = IBV_FLOW_SPEC_VXLAN_TUNNEL,
1503 		.size = size,
1504 	};
1505 	int ret;
1506 	union vni {
1507 		uint32_t vlan_id;
1508 		uint8_t vni[4];
1509 	} id = { .vlan_id = 0, };
1510 
1511 	if (!((struct priv *)dev->data->dev_private)->config.l3_vxlan_en)
1512 		return rte_flow_error_set(error, ENOTSUP,
1513 					  RTE_FLOW_ERROR_TYPE_ITEM,
1514 					  item,
1515 					  "L3 VXLAN is not enabled by device"
1516 					  " parameter and/or not configured in"
1517 					  " firmware");
1518 	if (flow->layers & MLX5_FLOW_LAYER_TUNNEL)
1519 		return rte_flow_error_set(error, ENOTSUP,
1520 					  RTE_FLOW_ERROR_TYPE_ITEM,
1521 					  item,
1522 					  "a tunnel is already present");
1523 	/*
1524 	 * Verify only UDPv4 is present as defined in
1525 	 * https://tools.ietf.org/html/rfc7348
1526 	 */
1527 	if (!(flow->layers & MLX5_FLOW_LAYER_OUTER_L4_UDP))
1528 		return rte_flow_error_set(error, ENOTSUP,
1529 					  RTE_FLOW_ERROR_TYPE_ITEM,
1530 					  item,
1531 					  "no outer UDP layer found");
1532 	if (!mask)
1533 		mask = &rte_flow_item_vxlan_gpe_mask;
1534 	ret = mlx5_flow_item_acceptable
1535 		(item, (const uint8_t *)mask,
1536 		 (const uint8_t *)&rte_flow_item_vxlan_gpe_mask,
1537 		 sizeof(struct rte_flow_item_vxlan_gpe), error);
1538 	if (ret < 0)
1539 		return ret;
1540 	if (spec) {
1541 		memcpy(&id.vni[1], spec->vni, 3);
1542 		vxlan_gpe.val.tunnel_id = id.vlan_id;
1543 		memcpy(&id.vni[1], mask->vni, 3);
1544 		vxlan_gpe.mask.tunnel_id = id.vlan_id;
1545 		if (spec->protocol)
1546 			return rte_flow_error_set
1547 				(error, EINVAL,
1548 				 RTE_FLOW_ERROR_TYPE_ITEM,
1549 				 item,
1550 				 "VxLAN-GPE protocol not supported");
1551 		/* Remove unwanted bits from values. */
1552 		vxlan_gpe.val.tunnel_id &= vxlan_gpe.mask.tunnel_id;
1553 	}
1554 	/*
1555 	 * Tunnel id 0 is equivalent as not adding a VXLAN layer, if only this
1556 	 * layer is defined in the Verbs specification it is interpreted as
1557 	 * wildcard and all packets will match this rule, if it follows a full
1558 	 * stack layer (ex: eth / ipv4 / udp), all packets matching the layers
1559 	 * before will also match this rule.  To avoid such situation, VNI 0
1560 	 * is currently refused.
1561 	 */
1562 	if (!vxlan_gpe.val.tunnel_id)
1563 		return rte_flow_error_set(error, EINVAL,
1564 					  RTE_FLOW_ERROR_TYPE_ITEM,
1565 					  item,
1566 					  "VXLAN-GPE vni cannot be 0");
1567 	if (!(flow->layers & MLX5_FLOW_LAYER_OUTER))
1568 		return rte_flow_error_set(error, EINVAL,
1569 					  RTE_FLOW_ERROR_TYPE_ITEM,
1570 					  item,
1571 					  "VXLAN-GPE tunnel must be fully"
1572 					  " defined");
1573 	if (size <= flow_size) {
1574 		mlx5_flow_spec_verbs_add(flow, &vxlan_gpe, size);
1575 		flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L2;
1576 	}
1577 	flow->layers |= MLX5_FLOW_LAYER_VXLAN_GPE;
1578 	return size;
1579 }
1580 
1581 /**
1582  * Update the protocol in Verbs IPv4/IPv6 spec.
1583  *
1584  * @param[in, out] attr
1585  *   Pointer to Verbs attributes structure.
1586  * @param[in] search
1587  *   Specification type to search in order to update the IP protocol.
1588  * @param[in] protocol
1589  *   Protocol value to set if none is present in the specification.
1590  */
1591 static void
1592 mlx5_flow_item_gre_ip_protocol_update(struct ibv_flow_attr *attr,
1593 				      enum ibv_flow_spec_type search,
1594 				      uint8_t protocol)
1595 {
1596 	unsigned int i;
1597 	struct ibv_spec_header *hdr = (struct ibv_spec_header *)
1598 		((uint8_t *)attr + sizeof(struct ibv_flow_attr));
1599 
1600 	if (!attr)
1601 		return;
1602 	for (i = 0; i != attr->num_of_specs; ++i) {
1603 		if (hdr->type == search) {
1604 			union {
1605 				struct ibv_flow_spec_ipv4_ext *ipv4;
1606 				struct ibv_flow_spec_ipv6 *ipv6;
1607 			} ip;
1608 
1609 			switch (search) {
1610 			case IBV_FLOW_SPEC_IPV4_EXT:
1611 				ip.ipv4 = (struct ibv_flow_spec_ipv4_ext *)hdr;
1612 				if (!ip.ipv4->val.proto) {
1613 					ip.ipv4->val.proto = protocol;
1614 					ip.ipv4->mask.proto = 0xff;
1615 				}
1616 				break;
1617 			case IBV_FLOW_SPEC_IPV6:
1618 				ip.ipv6 = (struct ibv_flow_spec_ipv6 *)hdr;
1619 				if (!ip.ipv6->val.next_hdr) {
1620 					ip.ipv6->val.next_hdr = protocol;
1621 					ip.ipv6->mask.next_hdr = 0xff;
1622 				}
1623 				break;
1624 			default:
1625 				break;
1626 			}
1627 			break;
1628 		}
1629 		hdr = (struct ibv_spec_header *)((uint8_t *)hdr + hdr->size);
1630 	}
1631 }
1632 
1633 /**
1634  * Convert the @p item into a Verbs specification after ensuring the NIC
1635  * will understand and process it correctly.
1636  * It will also update the previous L3 layer with the protocol value matching
1637  * the GRE.
1638  * If the necessary size for the conversion is greater than the @p flow_size,
1639  * nothing is written in @p flow, the validation is still performed.
1640  *
1641  * @param dev
1642  *   Pointer to Ethernet device.
1643  * @param[in] item
1644  *   Item specification.
1645  * @param[in, out] flow
1646  *   Pointer to flow structure.
1647  * @param[in] flow_size
1648  *   Size in bytes of the available space in @p flow, if too small, nothing is
1649  *   written.
1650  * @param[out] error
1651  *   Pointer to error structure.
1652  *
1653  * @return
1654  *   On success the number of bytes consumed/necessary, if the returned value
1655  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
1656  *   otherwise another call with this returned memory size should be done.
1657  *   On error, a negative errno value is returned and rte_errno is set.
1658  */
1659 static int
1660 mlx5_flow_item_gre(const struct rte_flow_item *item,
1661 		   struct rte_flow *flow, const size_t flow_size,
1662 		   struct rte_flow_error *error)
1663 {
1664 	struct mlx5_flow_verbs *verbs = flow->cur_verbs;
1665 	const struct rte_flow_item_gre *spec = item->spec;
1666 	const struct rte_flow_item_gre *mask = item->mask;
1667 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1668 	unsigned int size = sizeof(struct ibv_flow_spec_gre);
1669 	struct ibv_flow_spec_gre tunnel = {
1670 		.type = IBV_FLOW_SPEC_GRE,
1671 		.size = size,
1672 	};
1673 #else
1674 	unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
1675 	struct ibv_flow_spec_tunnel tunnel = {
1676 		.type = IBV_FLOW_SPEC_VXLAN_TUNNEL,
1677 		.size = size,
1678 	};
1679 #endif
1680 	int ret;
1681 
1682 	if (flow->l3_protocol_en && flow->l3_protocol != MLX5_IP_PROTOCOL_GRE)
1683 		return rte_flow_error_set(error, ENOTSUP,
1684 					  RTE_FLOW_ERROR_TYPE_ITEM,
1685 					  item,
1686 					  "protocol filtering not compatible"
1687 					  " with this GRE layer");
1688 	if (flow->layers & MLX5_FLOW_LAYER_TUNNEL)
1689 		return rte_flow_error_set(error, ENOTSUP,
1690 					  RTE_FLOW_ERROR_TYPE_ITEM,
1691 					  item,
1692 					  "a tunnel is already present");
1693 	if (!(flow->layers & MLX5_FLOW_LAYER_OUTER_L3))
1694 		return rte_flow_error_set(error, ENOTSUP,
1695 					  RTE_FLOW_ERROR_TYPE_ITEM,
1696 					  item,
1697 					  "L3 Layer is missing");
1698 	if (!mask)
1699 		mask = &rte_flow_item_gre_mask;
1700 	ret = mlx5_flow_item_acceptable
1701 		(item, (const uint8_t *)mask,
1702 		 (const uint8_t *)&rte_flow_item_gre_mask,
1703 		 sizeof(struct rte_flow_item_gre), error);
1704 	if (ret < 0)
1705 		return ret;
1706 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1707 	if (spec) {
1708 		tunnel.val.c_ks_res0_ver = spec->c_rsvd0_ver;
1709 		tunnel.val.protocol = spec->protocol;
1710 		tunnel.mask.c_ks_res0_ver = mask->c_rsvd0_ver;
1711 		tunnel.mask.protocol = mask->protocol;
1712 		/* Remove unwanted bits from values. */
1713 		tunnel.val.c_ks_res0_ver &= tunnel.mask.c_ks_res0_ver;
1714 		tunnel.val.protocol &= tunnel.mask.protocol;
1715 		tunnel.val.key &= tunnel.mask.key;
1716 	}
1717 #else
1718 	if (spec && (spec->protocol & mask->protocol))
1719 		return rte_flow_error_set(error, ENOTSUP,
1720 					  RTE_FLOW_ERROR_TYPE_ITEM,
1721 					  item,
1722 					  "without MPLS support the"
1723 					  " specification cannot be used for"
1724 					  " filtering");
1725 #endif /* !HAVE_IBV_DEVICE_MPLS_SUPPORT */
1726 	if (size <= flow_size) {
1727 		if (flow->layers & MLX5_FLOW_LAYER_OUTER_L3_IPV4)
1728 			mlx5_flow_item_gre_ip_protocol_update
1729 				(verbs->attr, IBV_FLOW_SPEC_IPV4_EXT,
1730 				 MLX5_IP_PROTOCOL_GRE);
1731 		else
1732 			mlx5_flow_item_gre_ip_protocol_update
1733 				(verbs->attr, IBV_FLOW_SPEC_IPV6,
1734 				 MLX5_IP_PROTOCOL_GRE);
1735 		mlx5_flow_spec_verbs_add(flow, &tunnel, size);
1736 		flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L2;
1737 	}
1738 	flow->layers |= MLX5_FLOW_LAYER_GRE;
1739 	return size;
1740 }
1741 
1742 /**
1743  * Convert the @p item into a Verbs specification after ensuring the NIC
1744  * will understand and process it correctly.
1745  * If the necessary size for the conversion is greater than the @p flow_size,
1746  * nothing is written in @p flow, the validation is still performed.
1747  *
1748  * @param[in] item
1749  *   Item specification.
1750  * @param[in, out] flow
1751  *   Pointer to flow structure.
1752  * @param[in] flow_size
1753  *   Size in bytes of the available space in @p flow, if too small, nothing is
1754  *   written.
1755  * @param[out] error
1756  *   Pointer to error structure.
1757  *
1758  * @return
1759  *   On success the number of bytes consumed/necessary, if the returned value
1760  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
1761  *   otherwise another call with this returned memory size should be done.
1762  *   On error, a negative errno value is returned and rte_errno is set.
1763  */
1764 static int
1765 mlx5_flow_item_mpls(const struct rte_flow_item *item __rte_unused,
1766 		    struct rte_flow *flow __rte_unused,
1767 		    const size_t flow_size __rte_unused,
1768 		    struct rte_flow_error *error)
1769 {
1770 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1771 	const struct rte_flow_item_mpls *spec = item->spec;
1772 	const struct rte_flow_item_mpls *mask = item->mask;
1773 	unsigned int size = sizeof(struct ibv_flow_spec_mpls);
1774 	struct ibv_flow_spec_mpls mpls = {
1775 		.type = IBV_FLOW_SPEC_MPLS,
1776 		.size = size,
1777 	};
1778 	int ret;
1779 
1780 	if (flow->l3_protocol_en && flow->l3_protocol != MLX5_IP_PROTOCOL_MPLS)
1781 		return rte_flow_error_set(error, ENOTSUP,
1782 					  RTE_FLOW_ERROR_TYPE_ITEM,
1783 					  item,
1784 					  "protocol filtering not compatible"
1785 					  " with MPLS layer");
1786 	/* Multi-tunnel isn't allowed but MPLS over GRE is an exception. */
1787 	if (flow->layers & MLX5_FLOW_LAYER_TUNNEL &&
1788 	    (flow->layers & MLX5_FLOW_LAYER_GRE) != MLX5_FLOW_LAYER_GRE)
1789 		return rte_flow_error_set(error, ENOTSUP,
1790 					  RTE_FLOW_ERROR_TYPE_ITEM,
1791 					  item,
1792 					  "a tunnel is already"
1793 					  " present");
1794 	if (!mask)
1795 		mask = &rte_flow_item_mpls_mask;
1796 	ret = mlx5_flow_item_acceptable
1797 		(item, (const uint8_t *)mask,
1798 		 (const uint8_t *)&rte_flow_item_mpls_mask,
1799 		 sizeof(struct rte_flow_item_mpls), error);
1800 	if (ret < 0)
1801 		return ret;
1802 	if (spec) {
1803 		memcpy(&mpls.val.label, spec, sizeof(mpls.val.label));
1804 		memcpy(&mpls.mask.label, mask, sizeof(mpls.mask.label));
1805 		/* Remove unwanted bits from values.  */
1806 		mpls.val.label &= mpls.mask.label;
1807 	}
1808 	if (size <= flow_size) {
1809 		mlx5_flow_spec_verbs_add(flow, &mpls, size);
1810 		flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L2;
1811 	}
1812 	flow->layers |= MLX5_FLOW_LAYER_MPLS;
1813 	return size;
1814 #endif /* !HAVE_IBV_DEVICE_MPLS_SUPPORT */
1815 	return rte_flow_error_set(error, ENOTSUP,
1816 				  RTE_FLOW_ERROR_TYPE_ITEM,
1817 				  item,
1818 				  "MPLS is not supported by Verbs, please"
1819 				  " update.");
1820 }
1821 
1822 /**
1823  * Convert the @p pattern into a Verbs specifications after ensuring the NIC
1824  * will understand and process it correctly.
1825  * The conversion is performed item per item, each of them is written into
1826  * the @p flow if its size is lesser or equal to @p flow_size.
1827  * Validation and memory consumption computation are still performed until the
1828  * end of @p pattern, unless an error is encountered.
1829  *
1830  * @param[in] pattern
1831  *   Flow pattern.
1832  * @param[in, out] flow
1833  *   Pointer to the rte_flow structure.
1834  * @param[in] flow_size
1835  *   Size in bytes of the available space in @p flow, if too small some
1836  *   garbage may be present.
1837  * @param[out] error
1838  *   Pointer to error structure.
1839  *
1840  * @return
1841  *   On success the number of bytes consumed/necessary, if the returned value
1842  *   is lesser or equal to @p flow_size, the @pattern  has fully been
1843  *   converted, otherwise another call with this returned memory size should
1844  *   be done.
1845  *   On error, a negative errno value is returned and rte_errno is set.
1846  */
1847 static int
1848 mlx5_flow_items(struct rte_eth_dev *dev,
1849 		const struct rte_flow_item pattern[],
1850 		struct rte_flow *flow, const size_t flow_size,
1851 		struct rte_flow_error *error)
1852 {
1853 	int remain = flow_size;
1854 	size_t size = 0;
1855 
1856 	for (; pattern->type != RTE_FLOW_ITEM_TYPE_END; pattern++) {
1857 		int ret = 0;
1858 
1859 		switch (pattern->type) {
1860 		case RTE_FLOW_ITEM_TYPE_VOID:
1861 			break;
1862 		case RTE_FLOW_ITEM_TYPE_ETH:
1863 			ret = mlx5_flow_item_eth(pattern, flow, remain, error);
1864 			break;
1865 		case RTE_FLOW_ITEM_TYPE_VLAN:
1866 			ret = mlx5_flow_item_vlan(pattern, flow, remain, error);
1867 			break;
1868 		case RTE_FLOW_ITEM_TYPE_IPV4:
1869 			ret = mlx5_flow_item_ipv4(pattern, flow, remain, error);
1870 			break;
1871 		case RTE_FLOW_ITEM_TYPE_IPV6:
1872 			ret = mlx5_flow_item_ipv6(pattern, flow, remain, error);
1873 			break;
1874 		case RTE_FLOW_ITEM_TYPE_UDP:
1875 			ret = mlx5_flow_item_udp(pattern, flow, remain, error);
1876 			break;
1877 		case RTE_FLOW_ITEM_TYPE_TCP:
1878 			ret = mlx5_flow_item_tcp(pattern, flow, remain, error);
1879 			break;
1880 		case RTE_FLOW_ITEM_TYPE_VXLAN:
1881 			ret = mlx5_flow_item_vxlan(pattern, flow, remain,
1882 						   error);
1883 			break;
1884 		case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
1885 			ret = mlx5_flow_item_vxlan_gpe(dev, pattern, flow,
1886 						       remain, error);
1887 			break;
1888 		case RTE_FLOW_ITEM_TYPE_GRE:
1889 			ret = mlx5_flow_item_gre(pattern, flow, remain, error);
1890 			break;
1891 		case RTE_FLOW_ITEM_TYPE_MPLS:
1892 			ret = mlx5_flow_item_mpls(pattern, flow, remain, error);
1893 			break;
1894 		default:
1895 			return rte_flow_error_set(error, ENOTSUP,
1896 						  RTE_FLOW_ERROR_TYPE_ITEM,
1897 						  pattern,
1898 						  "item not supported");
1899 		}
1900 		if (ret < 0)
1901 			return ret;
1902 		if (remain > ret)
1903 			remain -= ret;
1904 		else
1905 			remain = 0;
1906 		size += ret;
1907 	}
1908 	if (!flow->layers) {
1909 		const struct rte_flow_item item = {
1910 			.type = RTE_FLOW_ITEM_TYPE_ETH,
1911 		};
1912 
1913 		return mlx5_flow_item_eth(&item, flow, flow_size, error);
1914 	}
1915 	return size;
1916 }
1917 
1918 /**
1919  * Convert the @p action into a Verbs specification after ensuring the NIC
1920  * will understand and process it correctly.
1921  * If the necessary size for the conversion is greater than the @p flow_size,
1922  * nothing is written in @p flow, the validation is still performed.
1923  *
1924  * @param[in] action
1925  *   Action configuration.
1926  * @param[in, out] flow
1927  *   Pointer to flow structure.
1928  * @param[in] flow_size
1929  *   Size in bytes of the available space in @p flow, if too small, nothing is
1930  *   written.
1931  * @param[out] error
1932  *   Pointer to error structure.
1933  *
1934  * @return
1935  *   On success the number of bytes consumed/necessary, if the returned value
1936  *   is lesser or equal to @p flow_size, the @p action has fully been
1937  *   converted, otherwise another call with this returned memory size should
1938  *   be done.
1939  *   On error, a negative errno value is returned and rte_errno is set.
1940  */
1941 static int
1942 mlx5_flow_action_drop(const struct rte_flow_action *action,
1943 		      struct rte_flow *flow, const size_t flow_size,
1944 		      struct rte_flow_error *error)
1945 {
1946 	unsigned int size = sizeof(struct ibv_flow_spec_action_drop);
1947 	struct ibv_flow_spec_action_drop drop = {
1948 			.type = IBV_FLOW_SPEC_ACTION_DROP,
1949 			.size = size,
1950 	};
1951 
1952 	if (flow->fate)
1953 		return rte_flow_error_set(error, ENOTSUP,
1954 					  RTE_FLOW_ERROR_TYPE_ACTION,
1955 					  action,
1956 					  "multiple fate actions are not"
1957 					  " supported");
1958 	if (flow->modifier & (MLX5_FLOW_MOD_FLAG | MLX5_FLOW_MOD_MARK))
1959 		return rte_flow_error_set(error, ENOTSUP,
1960 					  RTE_FLOW_ERROR_TYPE_ACTION,
1961 					  action,
1962 					  "drop is not compatible with"
1963 					  " flag/mark action");
1964 	if (size < flow_size)
1965 		mlx5_flow_spec_verbs_add(flow, &drop, size);
1966 	flow->fate |= MLX5_FLOW_FATE_DROP;
1967 	return size;
1968 }
1969 
1970 /**
1971  * Convert the @p action into @p flow after ensuring the NIC will understand
1972  * and process it correctly.
1973  *
1974  * @param[in] dev
1975  *   Pointer to Ethernet device structure.
1976  * @param[in] action
1977  *   Action configuration.
1978  * @param[in, out] flow
1979  *   Pointer to flow structure.
1980  * @param[out] error
1981  *   Pointer to error structure.
1982  *
1983  * @return
1984  *   0 on success, a negative errno value otherwise and rte_errno is set.
1985  */
1986 static int
1987 mlx5_flow_action_queue(struct rte_eth_dev *dev,
1988 		       const struct rte_flow_action *action,
1989 		       struct rte_flow *flow,
1990 		       struct rte_flow_error *error)
1991 {
1992 	struct priv *priv = dev->data->dev_private;
1993 	const struct rte_flow_action_queue *queue = action->conf;
1994 
1995 	if (flow->fate)
1996 		return rte_flow_error_set(error, ENOTSUP,
1997 					  RTE_FLOW_ERROR_TYPE_ACTION,
1998 					  action,
1999 					  "multiple fate actions are not"
2000 					  " supported");
2001 	if (queue->index >= priv->rxqs_n)
2002 		return rte_flow_error_set(error, EINVAL,
2003 					  RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2004 					  &queue->index,
2005 					  "queue index out of range");
2006 	if (!(*priv->rxqs)[queue->index])
2007 		return rte_flow_error_set(error, EINVAL,
2008 					  RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2009 					  &queue->index,
2010 					  "queue is not configured");
2011 	if (flow->queue)
2012 		(*flow->queue)[0] = queue->index;
2013 	flow->rss.queue_num = 1;
2014 	flow->fate |= MLX5_FLOW_FATE_QUEUE;
2015 	return 0;
2016 }
2017 
2018 /**
2019  * Ensure the @p action will be understood and used correctly by the  NIC.
2020  *
2021  * @param dev
2022  *   Pointer to Ethernet device structure.
2023  * @param action[in]
2024  *   Pointer to flow actions array.
2025  * @param flow[in, out]
2026  *   Pointer to the rte_flow structure.
2027  * @param error[in, out]
2028  *   Pointer to error structure.
2029  *
2030  * @return
2031  *   On success @p flow->queue array and @p flow->rss are filled and valid.
2032  *   On error, a negative errno value is returned and rte_errno is set.
2033  */
2034 static int
2035 mlx5_flow_action_rss(struct rte_eth_dev *dev,
2036 		     const struct rte_flow_action *action,
2037 		     struct rte_flow *flow,
2038 		     struct rte_flow_error *error)
2039 {
2040 	struct priv *priv = dev->data->dev_private;
2041 	const struct rte_flow_action_rss *rss = action->conf;
2042 	unsigned int i;
2043 
2044 	if (flow->fate)
2045 		return rte_flow_error_set(error, ENOTSUP,
2046 					  RTE_FLOW_ERROR_TYPE_ACTION,
2047 					  action,
2048 					  "multiple fate actions are not"
2049 					  " supported");
2050 	if (rss->func != RTE_ETH_HASH_FUNCTION_DEFAULT &&
2051 	    rss->func != RTE_ETH_HASH_FUNCTION_TOEPLITZ)
2052 		return rte_flow_error_set(error, ENOTSUP,
2053 					  RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2054 					  &rss->func,
2055 					  "RSS hash function not supported");
2056 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2057 	if (rss->level > 2)
2058 #else
2059 	if (rss->level > 1)
2060 #endif
2061 		return rte_flow_error_set(error, ENOTSUP,
2062 					  RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2063 					  &rss->level,
2064 					  "tunnel RSS is not supported");
2065 	if (rss->key_len < MLX5_RSS_HASH_KEY_LEN)
2066 		return rte_flow_error_set(error, ENOTSUP,
2067 					  RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2068 					  &rss->key_len,
2069 					  "RSS hash key too small");
2070 	if (rss->key_len > MLX5_RSS_HASH_KEY_LEN)
2071 		return rte_flow_error_set(error, ENOTSUP,
2072 					  RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2073 					  &rss->key_len,
2074 					  "RSS hash key too large");
2075 	if (!rss->queue_num)
2076 		return rte_flow_error_set(error, ENOTSUP,
2077 					  RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2078 					  rss,
2079 					  "no queues were provided for RSS");
2080 	if (rss->queue_num > priv->config.ind_table_max_size)
2081 		return rte_flow_error_set(error, ENOTSUP,
2082 					  RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2083 					  &rss->queue_num,
2084 					  "number of queues too large");
2085 	if (rss->types & MLX5_RSS_HF_MASK)
2086 		return rte_flow_error_set(error, ENOTSUP,
2087 					  RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2088 					  &rss->types,
2089 					  "some RSS protocols are not"
2090 					  " supported");
2091 	for (i = 0; i != rss->queue_num; ++i) {
2092 		if (rss->queue[i] >= priv->rxqs_n)
2093 			return rte_flow_error_set
2094 				(error, EINVAL,
2095 				 RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2096 				 rss,
2097 				 "queue index out of range");
2098 		if (!(*priv->rxqs)[rss->queue[i]])
2099 			return rte_flow_error_set
2100 				(error, EINVAL,
2101 				 RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2102 				 &rss->queue[i],
2103 				 "queue is not configured");
2104 	}
2105 	if (flow->queue)
2106 		memcpy((*flow->queue), rss->queue,
2107 		       rss->queue_num * sizeof(uint16_t));
2108 	flow->rss.queue_num = rss->queue_num;
2109 	memcpy(flow->key, rss->key, MLX5_RSS_HASH_KEY_LEN);
2110 	flow->rss.types = rss->types;
2111 	flow->rss.level = rss->level;
2112 	flow->fate |= MLX5_FLOW_FATE_RSS;
2113 	return 0;
2114 }
2115 
2116 /**
2117  * Convert the @p action into a Verbs specification after ensuring the NIC
2118  * will understand and process it correctly.
2119  * If the necessary size for the conversion is greater than the @p flow_size,
2120  * nothing is written in @p flow, the validation is still performed.
2121  *
2122  * @param[in] action
2123  *   Action configuration.
2124  * @param[in, out] flow
2125  *   Pointer to flow structure.
2126  * @param[in] flow_size
2127  *   Size in bytes of the available space in @p flow, if too small, nothing is
2128  *   written.
2129  * @param[out] error
2130  *   Pointer to error structure.
2131  *
2132  * @return
2133  *   On success the number of bytes consumed/necessary, if the returned value
2134  *   is lesser or equal to @p flow_size, the @p action has fully been
2135  *   converted, otherwise another call with this returned memory size should
2136  *   be done.
2137  *   On error, a negative errno value is returned and rte_errno is set.
2138  */
2139 static int
2140 mlx5_flow_action_flag(const struct rte_flow_action *action,
2141 		      struct rte_flow *flow, const size_t flow_size,
2142 		      struct rte_flow_error *error)
2143 {
2144 	unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
2145 	struct ibv_flow_spec_action_tag tag = {
2146 		.type = IBV_FLOW_SPEC_ACTION_TAG,
2147 		.size = size,
2148 		.tag_id = mlx5_flow_mark_set(MLX5_FLOW_MARK_DEFAULT),
2149 	};
2150 	struct mlx5_flow_verbs *verbs = flow->cur_verbs;
2151 
2152 	if (flow->modifier & MLX5_FLOW_MOD_FLAG)
2153 		return rte_flow_error_set(error, ENOTSUP,
2154 					  RTE_FLOW_ERROR_TYPE_ACTION,
2155 					  action,
2156 					  "flag action already present");
2157 	if (flow->fate & MLX5_FLOW_FATE_DROP)
2158 		return rte_flow_error_set(error, ENOTSUP,
2159 					  RTE_FLOW_ERROR_TYPE_ACTION,
2160 					  action,
2161 					  "flag is not compatible with drop"
2162 					  " action");
2163 	if (flow->modifier & MLX5_FLOW_MOD_MARK)
2164 		size = 0;
2165 	else if (size <= flow_size && verbs)
2166 		mlx5_flow_spec_verbs_add(flow, &tag, size);
2167 	flow->modifier |= MLX5_FLOW_MOD_FLAG;
2168 	return size;
2169 }
2170 
2171 /**
2172  * Update verbs specification to modify the flag to mark.
2173  *
2174  * @param[in, out] verbs
2175  *   Pointer to the mlx5_flow_verbs structure.
2176  * @param[in] mark_id
2177  *   Mark identifier to replace the flag.
2178  */
2179 static void
2180 mlx5_flow_verbs_mark_update(struct mlx5_flow_verbs *verbs, uint32_t mark_id)
2181 {
2182 	struct ibv_spec_header *hdr;
2183 	int i;
2184 
2185 	if (!verbs)
2186 		return;
2187 	/* Update Verbs specification. */
2188 	hdr = (struct ibv_spec_header *)verbs->specs;
2189 	if (!hdr)
2190 		return;
2191 	for (i = 0; i != verbs->attr->num_of_specs; ++i) {
2192 		if (hdr->type == IBV_FLOW_SPEC_ACTION_TAG) {
2193 			struct ibv_flow_spec_action_tag *t =
2194 				(struct ibv_flow_spec_action_tag *)hdr;
2195 
2196 			t->tag_id = mlx5_flow_mark_set(mark_id);
2197 		}
2198 		hdr = (struct ibv_spec_header *)((uintptr_t)hdr + hdr->size);
2199 	}
2200 }
2201 
2202 /**
2203  * Convert the @p action into @p flow (or by updating the already present
2204  * Flag Verbs specification) after ensuring the NIC will understand and
2205  * process it correctly.
2206  * If the necessary size for the conversion is greater than the @p flow_size,
2207  * nothing is written in @p flow, the validation is still performed.
2208  *
2209  * @param[in] action
2210  *   Action configuration.
2211  * @param[in, out] flow
2212  *   Pointer to flow structure.
2213  * @param[in] flow_size
2214  *   Size in bytes of the available space in @p flow, if too small, nothing is
2215  *   written.
2216  * @param[out] error
2217  *   Pointer to error structure.
2218  *
2219  * @return
2220  *   On success the number of bytes consumed/necessary, if the returned value
2221  *   is lesser or equal to @p flow_size, the @p action has fully been
2222  *   converted, otherwise another call with this returned memory size should
2223  *   be done.
2224  *   On error, a negative errno value is returned and rte_errno is set.
2225  */
2226 static int
2227 mlx5_flow_action_mark(const struct rte_flow_action *action,
2228 		      struct rte_flow *flow, const size_t flow_size,
2229 		      struct rte_flow_error *error)
2230 {
2231 	const struct rte_flow_action_mark *mark = action->conf;
2232 	unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
2233 	struct ibv_flow_spec_action_tag tag = {
2234 		.type = IBV_FLOW_SPEC_ACTION_TAG,
2235 		.size = size,
2236 	};
2237 	struct mlx5_flow_verbs *verbs = flow->cur_verbs;
2238 
2239 	if (!mark)
2240 		return rte_flow_error_set(error, EINVAL,
2241 					  RTE_FLOW_ERROR_TYPE_ACTION,
2242 					  action,
2243 					  "configuration cannot be null");
2244 	if (mark->id >= MLX5_FLOW_MARK_MAX)
2245 		return rte_flow_error_set(error, EINVAL,
2246 					  RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2247 					  &mark->id,
2248 					  "mark id must in 0 <= id < "
2249 					  RTE_STR(MLX5_FLOW_MARK_MAX));
2250 	if (flow->modifier & MLX5_FLOW_MOD_MARK)
2251 		return rte_flow_error_set(error, ENOTSUP,
2252 					  RTE_FLOW_ERROR_TYPE_ACTION,
2253 					  action,
2254 					  "mark action already present");
2255 	if (flow->fate & MLX5_FLOW_FATE_DROP)
2256 		return rte_flow_error_set(error, ENOTSUP,
2257 					  RTE_FLOW_ERROR_TYPE_ACTION,
2258 					  action,
2259 					  "mark is not compatible with drop"
2260 					  " action");
2261 	if (flow->modifier & MLX5_FLOW_MOD_FLAG) {
2262 		mlx5_flow_verbs_mark_update(verbs, mark->id);
2263 		size = 0;
2264 	} else if (size <= flow_size) {
2265 		tag.tag_id = mlx5_flow_mark_set(mark->id);
2266 		mlx5_flow_spec_verbs_add(flow, &tag, size);
2267 	}
2268 	flow->modifier |= MLX5_FLOW_MOD_MARK;
2269 	return size;
2270 }
2271 
2272 /**
2273  * Convert the @p action into a Verbs specification after ensuring the NIC
2274  * will understand and process it correctly.
2275  * If the necessary size for the conversion is greater than the @p flow_size,
2276  * nothing is written in @p flow, the validation is still performed.
2277  *
2278  * @param action[in]
2279  *   Action configuration.
2280  * @param flow[in, out]
2281  *   Pointer to flow structure.
2282  * @param flow_size[in]
2283  *   Size in bytes of the available space in @p flow, if too small, nothing is
2284  *   written.
2285  * @param error[int, out]
2286  *   Pointer to error structure.
2287  *
2288  * @return
2289  *   On success the number of bytes consumed/necessary, if the returned value
2290  *   is lesser or equal to @p flow_size, the @p action has fully been
2291  *   converted, otherwise another call with this returned memory size should
2292  *   be done.
2293  *   On error, a negative errno value is returned and rte_errno is set.
2294  */
2295 static int
2296 mlx5_flow_action_count(struct rte_eth_dev *dev,
2297 		       const struct rte_flow_action *action,
2298 		       struct rte_flow *flow,
2299 		       const size_t flow_size __rte_unused,
2300 		       struct rte_flow_error *error)
2301 {
2302 	const struct rte_flow_action_count *count = action->conf;
2303 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
2304 	unsigned int size = sizeof(struct ibv_flow_spec_counter_action);
2305 	struct ibv_flow_spec_counter_action counter = {
2306 		.type = IBV_FLOW_SPEC_ACTION_COUNT,
2307 		.size = size,
2308 	};
2309 #endif
2310 
2311 	if (!flow->counter) {
2312 		flow->counter = mlx5_flow_counter_new(dev, count->shared,
2313 						      count->id);
2314 		if (!flow->counter)
2315 			return rte_flow_error_set(error, ENOTSUP,
2316 						  RTE_FLOW_ERROR_TYPE_ACTION,
2317 						  action,
2318 						  "cannot get counter"
2319 						  " context.");
2320 	}
2321 	if (!((struct priv *)dev->data->dev_private)->config.flow_counter_en)
2322 		return rte_flow_error_set(error, ENOTSUP,
2323 					  RTE_FLOW_ERROR_TYPE_ACTION,
2324 					  action,
2325 					  "flow counters are not supported.");
2326 	flow->modifier |= MLX5_FLOW_MOD_COUNT;
2327 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
2328 	counter.counter_set_handle = flow->counter->cs->handle;
2329 	if (size <= flow_size)
2330 		mlx5_flow_spec_verbs_add(flow, &counter, size);
2331 	return size;
2332 #endif
2333 	return 0;
2334 }
2335 
2336 /**
2337  * Convert the @p action into @p flow after ensuring the NIC will understand
2338  * and process it correctly.
2339  * The conversion is performed action per action, each of them is written into
2340  * the @p flow if its size is lesser or equal to @p flow_size.
2341  * Validation and memory consumption computation are still performed until the
2342  * end of @p action, unless an error is encountered.
2343  *
2344  * @param[in] dev
2345  *   Pointer to Ethernet device structure.
2346  * @param[in] actions
2347  *   Pointer to flow actions array.
2348  * @param[in, out] flow
2349  *   Pointer to the rte_flow structure.
2350  * @param[in] flow_size
2351  *   Size in bytes of the available space in @p flow, if too small some
2352  *   garbage may be present.
2353  * @param[out] error
2354  *   Pointer to error structure.
2355  *
2356  * @return
2357  *   On success the number of bytes consumed/necessary, if the returned value
2358  *   is lesser or equal to @p flow_size, the @p actions has fully been
2359  *   converted, otherwise another call with this returned memory size should
2360  *   be done.
2361  *   On error, a negative errno value is returned and rte_errno is set.
2362  */
2363 static int
2364 mlx5_flow_actions(struct rte_eth_dev *dev,
2365 		  const struct rte_flow_action actions[],
2366 		  struct rte_flow *flow, const size_t flow_size,
2367 		  struct rte_flow_error *error)
2368 {
2369 	size_t size = 0;
2370 	int remain = flow_size;
2371 	int ret = 0;
2372 
2373 	for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
2374 		switch (actions->type) {
2375 		case RTE_FLOW_ACTION_TYPE_VOID:
2376 			break;
2377 		case RTE_FLOW_ACTION_TYPE_FLAG:
2378 			ret = mlx5_flow_action_flag(actions, flow, remain,
2379 						    error);
2380 			break;
2381 		case RTE_FLOW_ACTION_TYPE_MARK:
2382 			ret = mlx5_flow_action_mark(actions, flow, remain,
2383 						    error);
2384 			break;
2385 		case RTE_FLOW_ACTION_TYPE_DROP:
2386 			ret = mlx5_flow_action_drop(actions, flow, remain,
2387 						    error);
2388 			break;
2389 		case RTE_FLOW_ACTION_TYPE_QUEUE:
2390 			ret = mlx5_flow_action_queue(dev, actions, flow, error);
2391 			break;
2392 		case RTE_FLOW_ACTION_TYPE_RSS:
2393 			ret = mlx5_flow_action_rss(dev, actions, flow, error);
2394 			break;
2395 		case RTE_FLOW_ACTION_TYPE_COUNT:
2396 			ret = mlx5_flow_action_count(dev, actions, flow, remain,
2397 						     error);
2398 			break;
2399 		default:
2400 			return rte_flow_error_set(error, ENOTSUP,
2401 						  RTE_FLOW_ERROR_TYPE_ACTION,
2402 						  actions,
2403 						  "action not supported");
2404 		}
2405 		if (ret < 0)
2406 			return ret;
2407 		if (remain > ret)
2408 			remain -= ret;
2409 		else
2410 			remain = 0;
2411 		size += ret;
2412 	}
2413 	if (!flow->fate)
2414 		return rte_flow_error_set(error, ENOTSUP,
2415 					  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2416 					  NULL,
2417 					  "no fate action found");
2418 	return size;
2419 }
2420 
2421 /**
2422  * Validate flow rule and fill flow structure accordingly.
2423  *
2424  * @param dev
2425  *   Pointer to Ethernet device.
2426  * @param[out] flow
2427  *   Pointer to flow structure.
2428  * @param flow_size
2429  *   Size of allocated space for @p flow.
2430  * @param[in] attr
2431  *   Flow rule attributes.
2432  * @param[in] pattern
2433  *   Pattern specification (list terminated by the END pattern item).
2434  * @param[in] actions
2435  *   Associated actions (list terminated by the END action).
2436  * @param[out] error
2437  *   Perform verbose error reporting if not NULL.
2438  *
2439  * @return
2440  *   A positive value representing the size of the flow object in bytes
2441  *   regardless of @p flow_size on success, a negative errno value otherwise
2442  *   and rte_errno is set.
2443  */
2444 static int
2445 mlx5_flow_merge_switch(struct rte_eth_dev *dev,
2446 		       struct rte_flow *flow,
2447 		       size_t flow_size,
2448 		       const struct rte_flow_attr *attr,
2449 		       const struct rte_flow_item pattern[],
2450 		       const struct rte_flow_action actions[],
2451 		       struct rte_flow_error *error)
2452 {
2453 	unsigned int n = mlx5_dev_to_port_id(dev->device, NULL, 0);
2454 	uint16_t port_id[!n + n];
2455 	struct mlx5_nl_flow_ptoi ptoi[!n + n + 1];
2456 	size_t off = RTE_ALIGN_CEIL(sizeof(*flow), alignof(max_align_t));
2457 	unsigned int i;
2458 	unsigned int own = 0;
2459 	int ret;
2460 
2461 	/* At least one port is needed when no switch domain is present. */
2462 	if (!n) {
2463 		n = 1;
2464 		port_id[0] = dev->data->port_id;
2465 	} else {
2466 		n = RTE_MIN(mlx5_dev_to_port_id(dev->device, port_id, n), n);
2467 	}
2468 	for (i = 0; i != n; ++i) {
2469 		struct rte_eth_dev_info dev_info;
2470 
2471 		rte_eth_dev_info_get(port_id[i], &dev_info);
2472 		if (port_id[i] == dev->data->port_id)
2473 			own = i;
2474 		ptoi[i].port_id = port_id[i];
2475 		ptoi[i].ifindex = dev_info.if_index;
2476 	}
2477 	/* Ensure first entry of ptoi[] is the current device. */
2478 	if (own) {
2479 		ptoi[n] = ptoi[0];
2480 		ptoi[0] = ptoi[own];
2481 		ptoi[own] = ptoi[n];
2482 	}
2483 	/* An entry with zero ifindex terminates ptoi[]. */
2484 	ptoi[n].port_id = 0;
2485 	ptoi[n].ifindex = 0;
2486 	if (flow_size < off)
2487 		flow_size = 0;
2488 	ret = mlx5_nl_flow_transpose((uint8_t *)flow + off,
2489 				     flow_size ? flow_size - off : 0,
2490 				     ptoi, attr, pattern, actions, error);
2491 	if (ret < 0)
2492 		return ret;
2493 	if (flow_size) {
2494 		*flow = (struct rte_flow){
2495 			.attributes = *attr,
2496 			.nl_flow = (uint8_t *)flow + off,
2497 		};
2498 		/*
2499 		 * Generate a reasonably unique handle based on the address
2500 		 * of the target buffer.
2501 		 *
2502 		 * This is straightforward on 32-bit systems where the flow
2503 		 * pointer can be used directly. Otherwise, its least
2504 		 * significant part is taken after shifting it by the
2505 		 * previous power of two of the pointed buffer size.
2506 		 */
2507 		if (sizeof(flow) <= 4)
2508 			mlx5_nl_flow_brand(flow->nl_flow, (uintptr_t)flow);
2509 		else
2510 			mlx5_nl_flow_brand
2511 				(flow->nl_flow,
2512 				 (uintptr_t)flow >>
2513 				 rte_log2_u32(rte_align32prevpow2(flow_size)));
2514 	}
2515 	return off + ret;
2516 }
2517 
2518 static unsigned int
2519 mlx5_find_graph_root(const struct rte_flow_item pattern[], uint32_t rss_level)
2520 {
2521 	const struct rte_flow_item *item;
2522 	unsigned int has_vlan = 0;
2523 
2524 	for (item = pattern; item->type != RTE_FLOW_ITEM_TYPE_END; item++) {
2525 		if (item->type == RTE_FLOW_ITEM_TYPE_VLAN) {
2526 			has_vlan = 1;
2527 			break;
2528 		}
2529 	}
2530 	if (has_vlan)
2531 		return rss_level < 2 ? MLX5_EXPANSION_ROOT_ETH_VLAN :
2532 				       MLX5_EXPANSION_ROOT_OUTER_ETH_VLAN;
2533 	return rss_level < 2 ? MLX5_EXPANSION_ROOT :
2534 			       MLX5_EXPANSION_ROOT_OUTER;
2535 }
2536 
2537 /**
2538  * Convert the @p attributes, @p pattern, @p action, into an flow for the NIC
2539  * after ensuring the NIC will understand and process it correctly.
2540  * The conversion is only performed item/action per item/action, each of
2541  * them is written into the @p flow if its size is lesser or equal to @p
2542  * flow_size.
2543  * Validation and memory consumption computation are still performed until the
2544  * end, unless an error is encountered.
2545  *
2546  * @param[in] dev
2547  *   Pointer to Ethernet device.
2548  * @param[in, out] flow
2549  *   Pointer to flow structure.
2550  * @param[in] flow_size
2551  *   Size in bytes of the available space in @p flow, if too small some
2552  *   garbage may be present.
2553  * @param[in] attributes
2554  *   Flow rule attributes.
2555  * @param[in] pattern
2556  *   Pattern specification (list terminated by the END pattern item).
2557  * @param[in] actions
2558  *   Associated actions (list terminated by the END action).
2559  * @param[out] error
2560  *   Perform verbose error reporting if not NULL.
2561  *
2562  * @return
2563  *   On success the number of bytes consumed/necessary, if the returned value
2564  *   is lesser or equal to @p flow_size, the flow has fully been converted and
2565  *   can be applied, otherwise another call with this returned memory size
2566  *   should be done.
2567  *   On error, a negative errno value is returned and rte_errno is set.
2568  */
2569 static int
2570 mlx5_flow_merge(struct rte_eth_dev *dev, struct rte_flow *flow,
2571 		const size_t flow_size,
2572 		const struct rte_flow_attr *attributes,
2573 		const struct rte_flow_item pattern[],
2574 		const struct rte_flow_action actions[],
2575 		struct rte_flow_error *error)
2576 {
2577 	struct rte_flow local_flow = { .layers = 0, };
2578 	size_t size = sizeof(*flow);
2579 	union {
2580 		struct rte_flow_expand_rss buf;
2581 		uint8_t buffer[2048];
2582 	} expand_buffer;
2583 	struct rte_flow_expand_rss *buf = &expand_buffer.buf;
2584 	struct mlx5_flow_verbs *original_verbs = NULL;
2585 	size_t original_verbs_size = 0;
2586 	uint32_t original_layers = 0;
2587 	int expanded_pattern_idx = 0;
2588 	int ret;
2589 	uint32_t i;
2590 
2591 	if (attributes->transfer)
2592 		return mlx5_flow_merge_switch(dev, flow, flow_size,
2593 					      attributes, pattern,
2594 					      actions, error);
2595 	if (size > flow_size)
2596 		flow = &local_flow;
2597 	ret = mlx5_flow_attributes(dev, attributes, flow, error);
2598 	if (ret < 0)
2599 		return ret;
2600 	ret = mlx5_flow_actions(dev, actions, &local_flow, 0, error);
2601 	if (ret < 0)
2602 		return ret;
2603 	if (local_flow.rss.types) {
2604 		unsigned int graph_root;
2605 
2606 		graph_root = mlx5_find_graph_root(pattern,
2607 						  local_flow.rss.level);
2608 		ret = rte_flow_expand_rss(buf, sizeof(expand_buffer.buffer),
2609 					  pattern, local_flow.rss.types,
2610 					  mlx5_support_expansion,
2611 					  graph_root);
2612 		assert(ret > 0 &&
2613 		       (unsigned int)ret < sizeof(expand_buffer.buffer));
2614 	} else {
2615 		buf->entries = 1;
2616 		buf->entry[0].pattern = (void *)(uintptr_t)pattern;
2617 	}
2618 	size += RTE_ALIGN_CEIL(local_flow.rss.queue_num * sizeof(uint16_t),
2619 			       sizeof(void *));
2620 	if (size <= flow_size)
2621 		flow->queue = (void *)(flow + 1);
2622 	LIST_INIT(&flow->verbs);
2623 	flow->layers = 0;
2624 	flow->modifier = 0;
2625 	flow->fate = 0;
2626 	for (i = 0; i != buf->entries; ++i) {
2627 		size_t off = size;
2628 		size_t off2;
2629 
2630 		flow->layers = original_layers;
2631 		size += sizeof(struct ibv_flow_attr) +
2632 			sizeof(struct mlx5_flow_verbs);
2633 		off2 = size;
2634 		if (size < flow_size) {
2635 			flow->cur_verbs = (void *)((uintptr_t)flow + off);
2636 			flow->cur_verbs->attr = (void *)(flow->cur_verbs + 1);
2637 			flow->cur_verbs->specs =
2638 				(void *)(flow->cur_verbs->attr + 1);
2639 		}
2640 		/* First iteration convert the pattern into Verbs. */
2641 		if (i == 0) {
2642 			/* Actions don't need to be converted several time. */
2643 			ret = mlx5_flow_actions(dev, actions, flow,
2644 						(size < flow_size) ?
2645 						flow_size - size : 0,
2646 						error);
2647 			if (ret < 0)
2648 				return ret;
2649 			size += ret;
2650 		} else {
2651 			/*
2652 			 * Next iteration means the pattern has already been
2653 			 * converted and an expansion is necessary to match
2654 			 * the user RSS request.  For that only the expanded
2655 			 * items will be converted, the common part with the
2656 			 * user pattern are just copied into the next buffer
2657 			 * zone.
2658 			 */
2659 			size += original_verbs_size;
2660 			if (size < flow_size) {
2661 				rte_memcpy(flow->cur_verbs->attr,
2662 					   original_verbs->attr,
2663 					   original_verbs_size +
2664 					   sizeof(struct ibv_flow_attr));
2665 				flow->cur_verbs->size = original_verbs_size;
2666 			}
2667 		}
2668 		ret = mlx5_flow_items
2669 			(dev,
2670 			 (const struct rte_flow_item *)
2671 			 &buf->entry[i].pattern[expanded_pattern_idx],
2672 			 flow,
2673 			 (size < flow_size) ? flow_size - size : 0, error);
2674 		if (ret < 0)
2675 			return ret;
2676 		size += ret;
2677 		if (size <= flow_size) {
2678 			mlx5_flow_adjust_priority(dev, flow);
2679 			LIST_INSERT_HEAD(&flow->verbs, flow->cur_verbs, next);
2680 		}
2681 		/*
2682 		 * Keep a pointer of the first verbs conversion and the layers
2683 		 * it has encountered.
2684 		 */
2685 		if (i == 0) {
2686 			original_verbs = flow->cur_verbs;
2687 			original_verbs_size = size - off2;
2688 			original_layers = flow->layers;
2689 			/*
2690 			 * move the index of the expanded pattern to the
2691 			 * first item not addressed yet.
2692 			 */
2693 			if (pattern->type == RTE_FLOW_ITEM_TYPE_END) {
2694 				expanded_pattern_idx++;
2695 			} else {
2696 				const struct rte_flow_item *item = pattern;
2697 
2698 				for (item = pattern;
2699 				     item->type != RTE_FLOW_ITEM_TYPE_END;
2700 				     ++item)
2701 					expanded_pattern_idx++;
2702 			}
2703 		}
2704 	}
2705 	/* Restore the origin layers in the flow. */
2706 	flow->layers = original_layers;
2707 	return size;
2708 }
2709 
2710 /**
2711  * Lookup and set the ptype in the data Rx part.  A single Ptype can be used,
2712  * if several tunnel rules are used on this queue, the tunnel ptype will be
2713  * cleared.
2714  *
2715  * @param rxq_ctrl
2716  *   Rx queue to update.
2717  */
2718 static void
2719 mlx5_flow_rxq_tunnel_ptype_update(struct mlx5_rxq_ctrl *rxq_ctrl)
2720 {
2721 	unsigned int i;
2722 	uint32_t tunnel_ptype = 0;
2723 
2724 	/* Look up for the ptype to use. */
2725 	for (i = 0; i != MLX5_FLOW_TUNNEL; ++i) {
2726 		if (!rxq_ctrl->flow_tunnels_n[i])
2727 			continue;
2728 		if (!tunnel_ptype) {
2729 			tunnel_ptype = tunnels_info[i].ptype;
2730 		} else {
2731 			tunnel_ptype = 0;
2732 			break;
2733 		}
2734 	}
2735 	rxq_ctrl->rxq.tunnel = tunnel_ptype;
2736 }
2737 
2738 /**
2739  * Set the Rx queue flags (Mark/Flag and Tunnel Ptypes) according to the flow.
2740  *
2741  * @param[in] dev
2742  *   Pointer to Ethernet device.
2743  * @param[in] flow
2744  *   Pointer to flow structure.
2745  */
2746 static void
2747 mlx5_flow_rxq_flags_set(struct rte_eth_dev *dev, struct rte_flow *flow)
2748 {
2749 	struct priv *priv = dev->data->dev_private;
2750 	const int mark = !!(flow->modifier &
2751 			    (MLX5_FLOW_MOD_FLAG | MLX5_FLOW_MOD_MARK));
2752 	const int tunnel = !!(flow->layers & MLX5_FLOW_LAYER_TUNNEL);
2753 	unsigned int i;
2754 
2755 	for (i = 0; i != flow->rss.queue_num; ++i) {
2756 		int idx = (*flow->queue)[i];
2757 		struct mlx5_rxq_ctrl *rxq_ctrl =
2758 			container_of((*priv->rxqs)[idx],
2759 				     struct mlx5_rxq_ctrl, rxq);
2760 
2761 		if (mark) {
2762 			rxq_ctrl->rxq.mark = 1;
2763 			rxq_ctrl->flow_mark_n++;
2764 		}
2765 		if (tunnel) {
2766 			unsigned int j;
2767 
2768 			/* Increase the counter matching the flow. */
2769 			for (j = 0; j != MLX5_FLOW_TUNNEL; ++j) {
2770 				if ((tunnels_info[j].tunnel & flow->layers) ==
2771 				    tunnels_info[j].tunnel) {
2772 					rxq_ctrl->flow_tunnels_n[j]++;
2773 					break;
2774 				}
2775 			}
2776 			mlx5_flow_rxq_tunnel_ptype_update(rxq_ctrl);
2777 		}
2778 	}
2779 }
2780 
2781 /**
2782  * Clear the Rx queue flags (Mark/Flag and Tunnel Ptype) associated with the
2783  * @p flow if no other flow uses it with the same kind of request.
2784  *
2785  * @param dev
2786  *   Pointer to Ethernet device.
2787  * @param[in] flow
2788  *   Pointer to the flow.
2789  */
2790 static void
2791 mlx5_flow_rxq_flags_trim(struct rte_eth_dev *dev, struct rte_flow *flow)
2792 {
2793 	struct priv *priv = dev->data->dev_private;
2794 	const int mark = !!(flow->modifier &
2795 			    (MLX5_FLOW_MOD_FLAG | MLX5_FLOW_MOD_MARK));
2796 	const int tunnel = !!(flow->layers & MLX5_FLOW_LAYER_TUNNEL);
2797 	unsigned int i;
2798 
2799 	assert(dev->data->dev_started);
2800 	for (i = 0; i != flow->rss.queue_num; ++i) {
2801 		int idx = (*flow->queue)[i];
2802 		struct mlx5_rxq_ctrl *rxq_ctrl =
2803 			container_of((*priv->rxqs)[idx],
2804 				     struct mlx5_rxq_ctrl, rxq);
2805 
2806 		if (mark) {
2807 			rxq_ctrl->flow_mark_n--;
2808 			rxq_ctrl->rxq.mark = !!rxq_ctrl->flow_mark_n;
2809 		}
2810 		if (tunnel) {
2811 			unsigned int j;
2812 
2813 			/* Decrease the counter matching the flow. */
2814 			for (j = 0; j != MLX5_FLOW_TUNNEL; ++j) {
2815 				if ((tunnels_info[j].tunnel & flow->layers) ==
2816 				    tunnels_info[j].tunnel) {
2817 					rxq_ctrl->flow_tunnels_n[j]--;
2818 					break;
2819 				}
2820 			}
2821 			mlx5_flow_rxq_tunnel_ptype_update(rxq_ctrl);
2822 		}
2823 	}
2824 }
2825 
2826 /**
2827  * Clear the Mark/Flag and Tunnel ptype information in all Rx queues.
2828  *
2829  * @param dev
2830  *   Pointer to Ethernet device.
2831  */
2832 static void
2833 mlx5_flow_rxq_flags_clear(struct rte_eth_dev *dev)
2834 {
2835 	struct priv *priv = dev->data->dev_private;
2836 	unsigned int i;
2837 
2838 	for (i = 0; i != priv->rxqs_n; ++i) {
2839 		struct mlx5_rxq_ctrl *rxq_ctrl;
2840 		unsigned int j;
2841 
2842 		if (!(*priv->rxqs)[i])
2843 			continue;
2844 		rxq_ctrl = container_of((*priv->rxqs)[i],
2845 					struct mlx5_rxq_ctrl, rxq);
2846 		rxq_ctrl->flow_mark_n = 0;
2847 		rxq_ctrl->rxq.mark = 0;
2848 		for (j = 0; j != MLX5_FLOW_TUNNEL; ++j)
2849 			rxq_ctrl->flow_tunnels_n[j] = 0;
2850 		rxq_ctrl->rxq.tunnel = 0;
2851 	}
2852 }
2853 
2854 /**
2855  * Validate a flow supported by the NIC.
2856  *
2857  * @see rte_flow_validate()
2858  * @see rte_flow_ops
2859  */
2860 int
2861 mlx5_flow_validate(struct rte_eth_dev *dev,
2862 		   const struct rte_flow_attr *attr,
2863 		   const struct rte_flow_item items[],
2864 		   const struct rte_flow_action actions[],
2865 		   struct rte_flow_error *error)
2866 {
2867 	int ret = mlx5_flow_merge(dev, NULL, 0, attr, items, actions, error);
2868 
2869 	if (ret < 0)
2870 		return ret;
2871 	return 0;
2872 }
2873 
2874 /**
2875  * Remove the flow.
2876  *
2877  * @param[in] dev
2878  *   Pointer to Ethernet device.
2879  * @param[in, out] flow
2880  *   Pointer to flow structure.
2881  */
2882 static void
2883 mlx5_flow_remove(struct rte_eth_dev *dev, struct rte_flow *flow)
2884 {
2885 	struct priv *priv = dev->data->dev_private;
2886 	struct mlx5_flow_verbs *verbs;
2887 
2888 	if (flow->nl_flow && priv->mnl_socket)
2889 		mlx5_nl_flow_destroy(priv->mnl_socket, flow->nl_flow, NULL);
2890 	LIST_FOREACH(verbs, &flow->verbs, next) {
2891 		if (verbs->flow) {
2892 			claim_zero(mlx5_glue->destroy_flow(verbs->flow));
2893 			verbs->flow = NULL;
2894 		}
2895 		if (verbs->hrxq) {
2896 			if (flow->fate & MLX5_FLOW_FATE_DROP)
2897 				mlx5_hrxq_drop_release(dev);
2898 			else
2899 				mlx5_hrxq_release(dev, verbs->hrxq);
2900 			verbs->hrxq = NULL;
2901 		}
2902 	}
2903 	if (flow->counter) {
2904 		mlx5_flow_counter_release(flow->counter);
2905 		flow->counter = NULL;
2906 	}
2907 }
2908 
2909 /**
2910  * Apply the flow.
2911  *
2912  * @param[in] dev
2913  *   Pointer to Ethernet device structure.
2914  * @param[in, out] flow
2915  *   Pointer to flow structure.
2916  * @param[out] error
2917  *   Pointer to error structure.
2918  *
2919  * @return
2920  *   0 on success, a negative errno value otherwise and rte_errno is set.
2921  */
2922 static int
2923 mlx5_flow_apply(struct rte_eth_dev *dev, struct rte_flow *flow,
2924 		struct rte_flow_error *error)
2925 {
2926 	struct priv *priv = dev->data->dev_private;
2927 	struct mlx5_flow_verbs *verbs;
2928 	int err;
2929 
2930 	LIST_FOREACH(verbs, &flow->verbs, next) {
2931 		if (flow->fate & MLX5_FLOW_FATE_DROP) {
2932 			verbs->hrxq = mlx5_hrxq_drop_new(dev);
2933 			if (!verbs->hrxq) {
2934 				rte_flow_error_set
2935 					(error, errno,
2936 					 RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2937 					 NULL,
2938 					 "cannot get drop hash queue");
2939 				goto error;
2940 			}
2941 		} else {
2942 			struct mlx5_hrxq *hrxq;
2943 
2944 			hrxq = mlx5_hrxq_get(dev, flow->key,
2945 					     MLX5_RSS_HASH_KEY_LEN,
2946 					     verbs->hash_fields,
2947 					     (*flow->queue),
2948 					     flow->rss.queue_num);
2949 			if (!hrxq)
2950 				hrxq = mlx5_hrxq_new(dev, flow->key,
2951 						     MLX5_RSS_HASH_KEY_LEN,
2952 						     verbs->hash_fields,
2953 						     (*flow->queue),
2954 						     flow->rss.queue_num,
2955 						     !!(flow->layers &
2956 						      MLX5_FLOW_LAYER_TUNNEL));
2957 			if (!hrxq) {
2958 				rte_flow_error_set
2959 					(error, rte_errno,
2960 					 RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2961 					 NULL,
2962 					 "cannot get hash queue");
2963 				goto error;
2964 			}
2965 			verbs->hrxq = hrxq;
2966 		}
2967 		verbs->flow =
2968 			mlx5_glue->create_flow(verbs->hrxq->qp, verbs->attr);
2969 		if (!verbs->flow) {
2970 			rte_flow_error_set(error, errno,
2971 					   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2972 					   NULL,
2973 					   "hardware refuses to create flow");
2974 			goto error;
2975 		}
2976 	}
2977 	if (flow->nl_flow &&
2978 	    priv->mnl_socket &&
2979 	    mlx5_nl_flow_create(priv->mnl_socket, flow->nl_flow, error))
2980 		goto error;
2981 	return 0;
2982 error:
2983 	err = rte_errno; /* Save rte_errno before cleanup. */
2984 	LIST_FOREACH(verbs, &flow->verbs, next) {
2985 		if (verbs->hrxq) {
2986 			if (flow->fate & MLX5_FLOW_FATE_DROP)
2987 				mlx5_hrxq_drop_release(dev);
2988 			else
2989 				mlx5_hrxq_release(dev, verbs->hrxq);
2990 			verbs->hrxq = NULL;
2991 		}
2992 	}
2993 	rte_errno = err; /* Restore rte_errno. */
2994 	return -rte_errno;
2995 }
2996 
2997 /**
2998  * Create a flow and add it to @p list.
2999  *
3000  * @param dev
3001  *   Pointer to Ethernet device.
3002  * @param list
3003  *   Pointer to a TAILQ flow list.
3004  * @param[in] attr
3005  *   Flow rule attributes.
3006  * @param[in] items
3007  *   Pattern specification (list terminated by the END pattern item).
3008  * @param[in] actions
3009  *   Associated actions (list terminated by the END action).
3010  * @param[out] error
3011  *   Perform verbose error reporting if not NULL.
3012  *
3013  * @return
3014  *   A flow on success, NULL otherwise and rte_errno is set.
3015  */
3016 static struct rte_flow *
3017 mlx5_flow_list_create(struct rte_eth_dev *dev,
3018 		      struct mlx5_flows *list,
3019 		      const struct rte_flow_attr *attr,
3020 		      const struct rte_flow_item items[],
3021 		      const struct rte_flow_action actions[],
3022 		      struct rte_flow_error *error)
3023 {
3024 	struct rte_flow *flow = NULL;
3025 	size_t size = 0;
3026 	int ret;
3027 
3028 	ret = mlx5_flow_merge(dev, flow, size, attr, items, actions, error);
3029 	if (ret < 0)
3030 		return NULL;
3031 	size = ret;
3032 	flow = rte_calloc(__func__, 1, size, 0);
3033 	if (!flow) {
3034 		rte_flow_error_set(error, ENOMEM,
3035 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
3036 				   NULL,
3037 				   "not enough memory to create flow");
3038 		return NULL;
3039 	}
3040 	ret = mlx5_flow_merge(dev, flow, size, attr, items, actions, error);
3041 	if (ret < 0) {
3042 		rte_free(flow);
3043 		return NULL;
3044 	}
3045 	assert((size_t)ret == size);
3046 	if (dev->data->dev_started) {
3047 		ret = mlx5_flow_apply(dev, flow, error);
3048 		if (ret < 0) {
3049 			ret = rte_errno; /* Save rte_errno before cleanup. */
3050 			if (flow) {
3051 				mlx5_flow_remove(dev, flow);
3052 				rte_free(flow);
3053 			}
3054 			rte_errno = ret; /* Restore rte_errno. */
3055 			return NULL;
3056 		}
3057 	}
3058 	TAILQ_INSERT_TAIL(list, flow, next);
3059 	mlx5_flow_rxq_flags_set(dev, flow);
3060 	return flow;
3061 }
3062 
3063 /**
3064  * Create a flow.
3065  *
3066  * @see rte_flow_create()
3067  * @see rte_flow_ops
3068  */
3069 struct rte_flow *
3070 mlx5_flow_create(struct rte_eth_dev *dev,
3071 		 const struct rte_flow_attr *attr,
3072 		 const struct rte_flow_item items[],
3073 		 const struct rte_flow_action actions[],
3074 		 struct rte_flow_error *error)
3075 {
3076 	return mlx5_flow_list_create
3077 		(dev, &((struct priv *)dev->data->dev_private)->flows,
3078 		 attr, items, actions, error);
3079 }
3080 
3081 /**
3082  * Destroy a flow in a list.
3083  *
3084  * @param dev
3085  *   Pointer to Ethernet device.
3086  * @param list
3087  *   Pointer to a TAILQ flow list.
3088  * @param[in] flow
3089  *   Flow to destroy.
3090  */
3091 static void
3092 mlx5_flow_list_destroy(struct rte_eth_dev *dev, struct mlx5_flows *list,
3093 		       struct rte_flow *flow)
3094 {
3095 	mlx5_flow_remove(dev, flow);
3096 	TAILQ_REMOVE(list, flow, next);
3097 	/*
3098 	 * Update RX queue flags only if port is started, otherwise it is
3099 	 * already clean.
3100 	 */
3101 	if (dev->data->dev_started)
3102 		mlx5_flow_rxq_flags_trim(dev, flow);
3103 	rte_free(flow);
3104 }
3105 
3106 /**
3107  * Destroy all flows.
3108  *
3109  * @param dev
3110  *   Pointer to Ethernet device.
3111  * @param list
3112  *   Pointer to a TAILQ flow list.
3113  */
3114 void
3115 mlx5_flow_list_flush(struct rte_eth_dev *dev, struct mlx5_flows *list)
3116 {
3117 	while (!TAILQ_EMPTY(list)) {
3118 		struct rte_flow *flow;
3119 
3120 		flow = TAILQ_FIRST(list);
3121 		mlx5_flow_list_destroy(dev, list, flow);
3122 	}
3123 }
3124 
3125 /**
3126  * Remove all flows.
3127  *
3128  * @param dev
3129  *   Pointer to Ethernet device.
3130  * @param list
3131  *   Pointer to a TAILQ flow list.
3132  */
3133 void
3134 mlx5_flow_stop(struct rte_eth_dev *dev, struct mlx5_flows *list)
3135 {
3136 	struct rte_flow *flow;
3137 
3138 	TAILQ_FOREACH_REVERSE(flow, list, mlx5_flows, next)
3139 		mlx5_flow_remove(dev, flow);
3140 	mlx5_flow_rxq_flags_clear(dev);
3141 }
3142 
3143 /**
3144  * Add all flows.
3145  *
3146  * @param dev
3147  *   Pointer to Ethernet device.
3148  * @param list
3149  *   Pointer to a TAILQ flow list.
3150  *
3151  * @return
3152  *   0 on success, a negative errno value otherwise and rte_errno is set.
3153  */
3154 int
3155 mlx5_flow_start(struct rte_eth_dev *dev, struct mlx5_flows *list)
3156 {
3157 	struct rte_flow *flow;
3158 	struct rte_flow_error error;
3159 	int ret = 0;
3160 
3161 	TAILQ_FOREACH(flow, list, next) {
3162 		ret = mlx5_flow_apply(dev, flow, &error);
3163 		if (ret < 0)
3164 			goto error;
3165 		mlx5_flow_rxq_flags_set(dev, flow);
3166 	}
3167 	return 0;
3168 error:
3169 	ret = rte_errno; /* Save rte_errno before cleanup. */
3170 	mlx5_flow_stop(dev, list);
3171 	rte_errno = ret; /* Restore rte_errno. */
3172 	return -rte_errno;
3173 }
3174 
3175 /**
3176  * Verify the flow list is empty
3177  *
3178  * @param dev
3179  *  Pointer to Ethernet device.
3180  *
3181  * @return the number of flows not released.
3182  */
3183 int
3184 mlx5_flow_verify(struct rte_eth_dev *dev)
3185 {
3186 	struct priv *priv = dev->data->dev_private;
3187 	struct rte_flow *flow;
3188 	int ret = 0;
3189 
3190 	TAILQ_FOREACH(flow, &priv->flows, next) {
3191 		DRV_LOG(DEBUG, "port %u flow %p still referenced",
3192 			dev->data->port_id, (void *)flow);
3193 		++ret;
3194 	}
3195 	return ret;
3196 }
3197 
3198 /**
3199  * Enable a control flow configured from the control plane.
3200  *
3201  * @param dev
3202  *   Pointer to Ethernet device.
3203  * @param eth_spec
3204  *   An Ethernet flow spec to apply.
3205  * @param eth_mask
3206  *   An Ethernet flow mask to apply.
3207  * @param vlan_spec
3208  *   A VLAN flow spec to apply.
3209  * @param vlan_mask
3210  *   A VLAN flow mask to apply.
3211  *
3212  * @return
3213  *   0 on success, a negative errno value otherwise and rte_errno is set.
3214  */
3215 int
3216 mlx5_ctrl_flow_vlan(struct rte_eth_dev *dev,
3217 		    struct rte_flow_item_eth *eth_spec,
3218 		    struct rte_flow_item_eth *eth_mask,
3219 		    struct rte_flow_item_vlan *vlan_spec,
3220 		    struct rte_flow_item_vlan *vlan_mask)
3221 {
3222 	struct priv *priv = dev->data->dev_private;
3223 	const struct rte_flow_attr attr = {
3224 		.ingress = 1,
3225 		.priority = MLX5_FLOW_PRIO_RSVD,
3226 	};
3227 	struct rte_flow_item items[] = {
3228 		{
3229 			.type = RTE_FLOW_ITEM_TYPE_ETH,
3230 			.spec = eth_spec,
3231 			.last = NULL,
3232 			.mask = eth_mask,
3233 		},
3234 		{
3235 			.type = (vlan_spec) ? RTE_FLOW_ITEM_TYPE_VLAN :
3236 				RTE_FLOW_ITEM_TYPE_END,
3237 			.spec = vlan_spec,
3238 			.last = NULL,
3239 			.mask = vlan_mask,
3240 		},
3241 		{
3242 			.type = RTE_FLOW_ITEM_TYPE_END,
3243 		},
3244 	};
3245 	uint16_t queue[priv->reta_idx_n];
3246 	struct rte_flow_action_rss action_rss = {
3247 		.func = RTE_ETH_HASH_FUNCTION_DEFAULT,
3248 		.level = 0,
3249 		.types = priv->rss_conf.rss_hf,
3250 		.key_len = priv->rss_conf.rss_key_len,
3251 		.queue_num = priv->reta_idx_n,
3252 		.key = priv->rss_conf.rss_key,
3253 		.queue = queue,
3254 	};
3255 	struct rte_flow_action actions[] = {
3256 		{
3257 			.type = RTE_FLOW_ACTION_TYPE_RSS,
3258 			.conf = &action_rss,
3259 		},
3260 		{
3261 			.type = RTE_FLOW_ACTION_TYPE_END,
3262 		},
3263 	};
3264 	struct rte_flow *flow;
3265 	struct rte_flow_error error;
3266 	unsigned int i;
3267 
3268 	if (!priv->reta_idx_n) {
3269 		rte_errno = EINVAL;
3270 		return -rte_errno;
3271 	}
3272 	for (i = 0; i != priv->reta_idx_n; ++i)
3273 		queue[i] = (*priv->reta_idx)[i];
3274 	flow = mlx5_flow_list_create(dev, &priv->ctrl_flows, &attr, items,
3275 				     actions, &error);
3276 	if (!flow)
3277 		return -rte_errno;
3278 	return 0;
3279 }
3280 
3281 /**
3282  * Enable a flow control configured from the control plane.
3283  *
3284  * @param dev
3285  *   Pointer to Ethernet device.
3286  * @param eth_spec
3287  *   An Ethernet flow spec to apply.
3288  * @param eth_mask
3289  *   An Ethernet flow mask to apply.
3290  *
3291  * @return
3292  *   0 on success, a negative errno value otherwise and rte_errno is set.
3293  */
3294 int
3295 mlx5_ctrl_flow(struct rte_eth_dev *dev,
3296 	       struct rte_flow_item_eth *eth_spec,
3297 	       struct rte_flow_item_eth *eth_mask)
3298 {
3299 	return mlx5_ctrl_flow_vlan(dev, eth_spec, eth_mask, NULL, NULL);
3300 }
3301 
3302 /**
3303  * Destroy a flow.
3304  *
3305  * @see rte_flow_destroy()
3306  * @see rte_flow_ops
3307  */
3308 int
3309 mlx5_flow_destroy(struct rte_eth_dev *dev,
3310 		  struct rte_flow *flow,
3311 		  struct rte_flow_error *error __rte_unused)
3312 {
3313 	struct priv *priv = dev->data->dev_private;
3314 
3315 	mlx5_flow_list_destroy(dev, &priv->flows, flow);
3316 	return 0;
3317 }
3318 
3319 /**
3320  * Destroy all flows.
3321  *
3322  * @see rte_flow_flush()
3323  * @see rte_flow_ops
3324  */
3325 int
3326 mlx5_flow_flush(struct rte_eth_dev *dev,
3327 		struct rte_flow_error *error __rte_unused)
3328 {
3329 	struct priv *priv = dev->data->dev_private;
3330 
3331 	mlx5_flow_list_flush(dev, &priv->flows);
3332 	return 0;
3333 }
3334 
3335 /**
3336  * Isolated mode.
3337  *
3338  * @see rte_flow_isolate()
3339  * @see rte_flow_ops
3340  */
3341 int
3342 mlx5_flow_isolate(struct rte_eth_dev *dev,
3343 		  int enable,
3344 		  struct rte_flow_error *error)
3345 {
3346 	struct priv *priv = dev->data->dev_private;
3347 
3348 	if (dev->data->dev_started) {
3349 		rte_flow_error_set(error, EBUSY,
3350 				   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
3351 				   NULL,
3352 				   "port must be stopped first");
3353 		return -rte_errno;
3354 	}
3355 	priv->isolated = !!enable;
3356 	if (enable)
3357 		dev->dev_ops = &mlx5_dev_ops_isolate;
3358 	else
3359 		dev->dev_ops = &mlx5_dev_ops;
3360 	return 0;
3361 }
3362 
3363 /**
3364  * Query flow counter.
3365  *
3366  * @param flow
3367  *   Pointer to the flow.
3368  *
3369  * @return
3370  *   0 on success, a negative errno value otherwise and rte_errno is set.
3371  */
3372 static int
3373 mlx5_flow_query_count(struct rte_flow *flow __rte_unused,
3374 		      void *data __rte_unused,
3375 		      struct rte_flow_error *error)
3376 {
3377 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
3378 	if (flow->modifier & MLX5_FLOW_MOD_COUNT) {
3379 		struct rte_flow_query_count *qc = data;
3380 		uint64_t counters[2] = {0, 0};
3381 		struct ibv_query_counter_set_attr query_cs_attr = {
3382 			.cs = flow->counter->cs,
3383 			.query_flags = IBV_COUNTER_SET_FORCE_UPDATE,
3384 		};
3385 		struct ibv_counter_set_data query_out = {
3386 			.out = counters,
3387 			.outlen = 2 * sizeof(uint64_t),
3388 		};
3389 		int err = mlx5_glue->query_counter_set(&query_cs_attr,
3390 						       &query_out);
3391 
3392 		if (err)
3393 			return rte_flow_error_set
3394 				(error, err,
3395 				 RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
3396 				 NULL,
3397 				 "cannot read counter");
3398 		qc->hits_set = 1;
3399 		qc->bytes_set = 1;
3400 		qc->hits = counters[0] - flow->counter->hits;
3401 		qc->bytes = counters[1] - flow->counter->bytes;
3402 		if (qc->reset) {
3403 			flow->counter->hits = counters[0];
3404 			flow->counter->bytes = counters[1];
3405 		}
3406 		return 0;
3407 	}
3408 	return rte_flow_error_set(error, ENOTSUP,
3409 				  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
3410 				  NULL,
3411 				  "flow does not have counter");
3412 #endif
3413 	return rte_flow_error_set(error, ENOTSUP,
3414 				  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
3415 				  NULL,
3416 				  "counters are not available");
3417 }
3418 
3419 /**
3420  * Query a flows.
3421  *
3422  * @see rte_flow_query()
3423  * @see rte_flow_ops
3424  */
3425 int
3426 mlx5_flow_query(struct rte_eth_dev *dev __rte_unused,
3427 		struct rte_flow *flow,
3428 		const struct rte_flow_action *actions,
3429 		void *data,
3430 		struct rte_flow_error *error)
3431 {
3432 	int ret = 0;
3433 
3434 	for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
3435 		switch (actions->type) {
3436 		case RTE_FLOW_ACTION_TYPE_VOID:
3437 			break;
3438 		case RTE_FLOW_ACTION_TYPE_COUNT:
3439 			ret = mlx5_flow_query_count(flow, data, error);
3440 			break;
3441 		default:
3442 			return rte_flow_error_set(error, ENOTSUP,
3443 						  RTE_FLOW_ERROR_TYPE_ACTION,
3444 						  actions,
3445 						  "action not supported");
3446 		}
3447 		if (ret < 0)
3448 			return ret;
3449 	}
3450 	return 0;
3451 }
3452 
3453 /**
3454  * Convert a flow director filter to a generic flow.
3455  *
3456  * @param dev
3457  *   Pointer to Ethernet device.
3458  * @param fdir_filter
3459  *   Flow director filter to add.
3460  * @param attributes
3461  *   Generic flow parameters structure.
3462  *
3463  * @return
3464  *   0 on success, a negative errno value otherwise and rte_errno is set.
3465  */
3466 static int
3467 mlx5_fdir_filter_convert(struct rte_eth_dev *dev,
3468 			 const struct rte_eth_fdir_filter *fdir_filter,
3469 			 struct mlx5_fdir *attributes)
3470 {
3471 	struct priv *priv = dev->data->dev_private;
3472 	const struct rte_eth_fdir_input *input = &fdir_filter->input;
3473 	const struct rte_eth_fdir_masks *mask =
3474 		&dev->data->dev_conf.fdir_conf.mask;
3475 
3476 	/* Validate queue number. */
3477 	if (fdir_filter->action.rx_queue >= priv->rxqs_n) {
3478 		DRV_LOG(ERR, "port %u invalid queue number %d",
3479 			dev->data->port_id, fdir_filter->action.rx_queue);
3480 		rte_errno = EINVAL;
3481 		return -rte_errno;
3482 	}
3483 	attributes->attr.ingress = 1;
3484 	attributes->items[0] = (struct rte_flow_item) {
3485 		.type = RTE_FLOW_ITEM_TYPE_ETH,
3486 		.spec = &attributes->l2,
3487 		.mask = &attributes->l2_mask,
3488 	};
3489 	switch (fdir_filter->action.behavior) {
3490 	case RTE_ETH_FDIR_ACCEPT:
3491 		attributes->actions[0] = (struct rte_flow_action){
3492 			.type = RTE_FLOW_ACTION_TYPE_QUEUE,
3493 			.conf = &attributes->queue,
3494 		};
3495 		break;
3496 	case RTE_ETH_FDIR_REJECT:
3497 		attributes->actions[0] = (struct rte_flow_action){
3498 			.type = RTE_FLOW_ACTION_TYPE_DROP,
3499 		};
3500 		break;
3501 	default:
3502 		DRV_LOG(ERR, "port %u invalid behavior %d",
3503 			dev->data->port_id,
3504 			fdir_filter->action.behavior);
3505 		rte_errno = ENOTSUP;
3506 		return -rte_errno;
3507 	}
3508 	attributes->queue.index = fdir_filter->action.rx_queue;
3509 	/* Handle L3. */
3510 	switch (fdir_filter->input.flow_type) {
3511 	case RTE_ETH_FLOW_NONFRAG_IPV4_UDP:
3512 	case RTE_ETH_FLOW_NONFRAG_IPV4_TCP:
3513 	case RTE_ETH_FLOW_NONFRAG_IPV4_OTHER:
3514 		attributes->l3.ipv4.hdr = (struct ipv4_hdr){
3515 			.src_addr = input->flow.ip4_flow.src_ip,
3516 			.dst_addr = input->flow.ip4_flow.dst_ip,
3517 			.time_to_live = input->flow.ip4_flow.ttl,
3518 			.type_of_service = input->flow.ip4_flow.tos,
3519 			.next_proto_id = input->flow.ip4_flow.proto,
3520 		};
3521 		attributes->l3_mask.ipv4.hdr = (struct ipv4_hdr){
3522 			.src_addr = mask->ipv4_mask.src_ip,
3523 			.dst_addr = mask->ipv4_mask.dst_ip,
3524 			.time_to_live = mask->ipv4_mask.ttl,
3525 			.type_of_service = mask->ipv4_mask.tos,
3526 			.next_proto_id = mask->ipv4_mask.proto,
3527 		};
3528 		attributes->items[1] = (struct rte_flow_item){
3529 			.type = RTE_FLOW_ITEM_TYPE_IPV4,
3530 			.spec = &attributes->l3,
3531 			.mask = &attributes->l3_mask,
3532 		};
3533 		break;
3534 	case RTE_ETH_FLOW_NONFRAG_IPV6_UDP:
3535 	case RTE_ETH_FLOW_NONFRAG_IPV6_TCP:
3536 	case RTE_ETH_FLOW_NONFRAG_IPV6_OTHER:
3537 		attributes->l3.ipv6.hdr = (struct ipv6_hdr){
3538 			.hop_limits = input->flow.ipv6_flow.hop_limits,
3539 			.proto = input->flow.ipv6_flow.proto,
3540 		};
3541 
3542 		memcpy(attributes->l3.ipv6.hdr.src_addr,
3543 		       input->flow.ipv6_flow.src_ip,
3544 		       RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
3545 		memcpy(attributes->l3.ipv6.hdr.dst_addr,
3546 		       input->flow.ipv6_flow.dst_ip,
3547 		       RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
3548 		memcpy(attributes->l3_mask.ipv6.hdr.src_addr,
3549 		       mask->ipv6_mask.src_ip,
3550 		       RTE_DIM(attributes->l3_mask.ipv6.hdr.src_addr));
3551 		memcpy(attributes->l3_mask.ipv6.hdr.dst_addr,
3552 		       mask->ipv6_mask.dst_ip,
3553 		       RTE_DIM(attributes->l3_mask.ipv6.hdr.src_addr));
3554 		attributes->items[1] = (struct rte_flow_item){
3555 			.type = RTE_FLOW_ITEM_TYPE_IPV6,
3556 			.spec = &attributes->l3,
3557 			.mask = &attributes->l3_mask,
3558 		};
3559 		break;
3560 	default:
3561 		DRV_LOG(ERR, "port %u invalid flow type%d",
3562 			dev->data->port_id, fdir_filter->input.flow_type);
3563 		rte_errno = ENOTSUP;
3564 		return -rte_errno;
3565 	}
3566 	/* Handle L4. */
3567 	switch (fdir_filter->input.flow_type) {
3568 	case RTE_ETH_FLOW_NONFRAG_IPV4_UDP:
3569 		attributes->l4.udp.hdr = (struct udp_hdr){
3570 			.src_port = input->flow.udp4_flow.src_port,
3571 			.dst_port = input->flow.udp4_flow.dst_port,
3572 		};
3573 		attributes->l4_mask.udp.hdr = (struct udp_hdr){
3574 			.src_port = mask->src_port_mask,
3575 			.dst_port = mask->dst_port_mask,
3576 		};
3577 		attributes->items[2] = (struct rte_flow_item){
3578 			.type = RTE_FLOW_ITEM_TYPE_UDP,
3579 			.spec = &attributes->l4,
3580 			.mask = &attributes->l4_mask,
3581 		};
3582 		break;
3583 	case RTE_ETH_FLOW_NONFRAG_IPV4_TCP:
3584 		attributes->l4.tcp.hdr = (struct tcp_hdr){
3585 			.src_port = input->flow.tcp4_flow.src_port,
3586 			.dst_port = input->flow.tcp4_flow.dst_port,
3587 		};
3588 		attributes->l4_mask.tcp.hdr = (struct tcp_hdr){
3589 			.src_port = mask->src_port_mask,
3590 			.dst_port = mask->dst_port_mask,
3591 		};
3592 		attributes->items[2] = (struct rte_flow_item){
3593 			.type = RTE_FLOW_ITEM_TYPE_TCP,
3594 			.spec = &attributes->l4,
3595 			.mask = &attributes->l4_mask,
3596 		};
3597 		break;
3598 	case RTE_ETH_FLOW_NONFRAG_IPV6_UDP:
3599 		attributes->l4.udp.hdr = (struct udp_hdr){
3600 			.src_port = input->flow.udp6_flow.src_port,
3601 			.dst_port = input->flow.udp6_flow.dst_port,
3602 		};
3603 		attributes->l4_mask.udp.hdr = (struct udp_hdr){
3604 			.src_port = mask->src_port_mask,
3605 			.dst_port = mask->dst_port_mask,
3606 		};
3607 		attributes->items[2] = (struct rte_flow_item){
3608 			.type = RTE_FLOW_ITEM_TYPE_UDP,
3609 			.spec = &attributes->l4,
3610 			.mask = &attributes->l4_mask,
3611 		};
3612 		break;
3613 	case RTE_ETH_FLOW_NONFRAG_IPV6_TCP:
3614 		attributes->l4.tcp.hdr = (struct tcp_hdr){
3615 			.src_port = input->flow.tcp6_flow.src_port,
3616 			.dst_port = input->flow.tcp6_flow.dst_port,
3617 		};
3618 		attributes->l4_mask.tcp.hdr = (struct tcp_hdr){
3619 			.src_port = mask->src_port_mask,
3620 			.dst_port = mask->dst_port_mask,
3621 		};
3622 		attributes->items[2] = (struct rte_flow_item){
3623 			.type = RTE_FLOW_ITEM_TYPE_TCP,
3624 			.spec = &attributes->l4,
3625 			.mask = &attributes->l4_mask,
3626 		};
3627 		break;
3628 	case RTE_ETH_FLOW_NONFRAG_IPV4_OTHER:
3629 	case RTE_ETH_FLOW_NONFRAG_IPV6_OTHER:
3630 		break;
3631 	default:
3632 		DRV_LOG(ERR, "port %u invalid flow type%d",
3633 			dev->data->port_id, fdir_filter->input.flow_type);
3634 		rte_errno = ENOTSUP;
3635 		return -rte_errno;
3636 	}
3637 	return 0;
3638 }
3639 
3640 /**
3641  * Add new flow director filter and store it in list.
3642  *
3643  * @param dev
3644  *   Pointer to Ethernet device.
3645  * @param fdir_filter
3646  *   Flow director filter to add.
3647  *
3648  * @return
3649  *   0 on success, a negative errno value otherwise and rte_errno is set.
3650  */
3651 static int
3652 mlx5_fdir_filter_add(struct rte_eth_dev *dev,
3653 		     const struct rte_eth_fdir_filter *fdir_filter)
3654 {
3655 	struct priv *priv = dev->data->dev_private;
3656 	struct mlx5_fdir attributes = {
3657 		.attr.group = 0,
3658 		.l2_mask = {
3659 			.dst.addr_bytes = "\x00\x00\x00\x00\x00\x00",
3660 			.src.addr_bytes = "\x00\x00\x00\x00\x00\x00",
3661 			.type = 0,
3662 		},
3663 	};
3664 	struct rte_flow_error error;
3665 	struct rte_flow *flow;
3666 	int ret;
3667 
3668 	ret = mlx5_fdir_filter_convert(dev, fdir_filter, &attributes);
3669 	if (ret)
3670 		return ret;
3671 	flow = mlx5_flow_list_create(dev, &priv->flows, &attributes.attr,
3672 				     attributes.items, attributes.actions,
3673 				     &error);
3674 	if (flow) {
3675 		DRV_LOG(DEBUG, "port %u FDIR created %p", dev->data->port_id,
3676 			(void *)flow);
3677 		return 0;
3678 	}
3679 	return -rte_errno;
3680 }
3681 
3682 /**
3683  * Delete specific filter.
3684  *
3685  * @param dev
3686  *   Pointer to Ethernet device.
3687  * @param fdir_filter
3688  *   Filter to be deleted.
3689  *
3690  * @return
3691  *   0 on success, a negative errno value otherwise and rte_errno is set.
3692  */
3693 static int
3694 mlx5_fdir_filter_delete(struct rte_eth_dev *dev __rte_unused,
3695 			const struct rte_eth_fdir_filter *fdir_filter
3696 			__rte_unused)
3697 {
3698 	rte_errno = ENOTSUP;
3699 	return -rte_errno;
3700 }
3701 
3702 /**
3703  * Update queue for specific filter.
3704  *
3705  * @param dev
3706  *   Pointer to Ethernet device.
3707  * @param fdir_filter
3708  *   Filter to be updated.
3709  *
3710  * @return
3711  *   0 on success, a negative errno value otherwise and rte_errno is set.
3712  */
3713 static int
3714 mlx5_fdir_filter_update(struct rte_eth_dev *dev,
3715 			const struct rte_eth_fdir_filter *fdir_filter)
3716 {
3717 	int ret;
3718 
3719 	ret = mlx5_fdir_filter_delete(dev, fdir_filter);
3720 	if (ret)
3721 		return ret;
3722 	return mlx5_fdir_filter_add(dev, fdir_filter);
3723 }
3724 
3725 /**
3726  * Flush all filters.
3727  *
3728  * @param dev
3729  *   Pointer to Ethernet device.
3730  */
3731 static void
3732 mlx5_fdir_filter_flush(struct rte_eth_dev *dev)
3733 {
3734 	struct priv *priv = dev->data->dev_private;
3735 
3736 	mlx5_flow_list_flush(dev, &priv->flows);
3737 }
3738 
3739 /**
3740  * Get flow director information.
3741  *
3742  * @param dev
3743  *   Pointer to Ethernet device.
3744  * @param[out] fdir_info
3745  *   Resulting flow director information.
3746  */
3747 static void
3748 mlx5_fdir_info_get(struct rte_eth_dev *dev, struct rte_eth_fdir_info *fdir_info)
3749 {
3750 	struct rte_eth_fdir_masks *mask =
3751 		&dev->data->dev_conf.fdir_conf.mask;
3752 
3753 	fdir_info->mode = dev->data->dev_conf.fdir_conf.mode;
3754 	fdir_info->guarant_spc = 0;
3755 	rte_memcpy(&fdir_info->mask, mask, sizeof(fdir_info->mask));
3756 	fdir_info->max_flexpayload = 0;
3757 	fdir_info->flow_types_mask[0] = 0;
3758 	fdir_info->flex_payload_unit = 0;
3759 	fdir_info->max_flex_payload_segment_num = 0;
3760 	fdir_info->flex_payload_limit = 0;
3761 	memset(&fdir_info->flex_conf, 0, sizeof(fdir_info->flex_conf));
3762 }
3763 
3764 /**
3765  * Deal with flow director operations.
3766  *
3767  * @param dev
3768  *   Pointer to Ethernet device.
3769  * @param filter_op
3770  *   Operation to perform.
3771  * @param arg
3772  *   Pointer to operation-specific structure.
3773  *
3774  * @return
3775  *   0 on success, a negative errno value otherwise and rte_errno is set.
3776  */
3777 static int
3778 mlx5_fdir_ctrl_func(struct rte_eth_dev *dev, enum rte_filter_op filter_op,
3779 		    void *arg)
3780 {
3781 	enum rte_fdir_mode fdir_mode =
3782 		dev->data->dev_conf.fdir_conf.mode;
3783 
3784 	if (filter_op == RTE_ETH_FILTER_NOP)
3785 		return 0;
3786 	if (fdir_mode != RTE_FDIR_MODE_PERFECT &&
3787 	    fdir_mode != RTE_FDIR_MODE_PERFECT_MAC_VLAN) {
3788 		DRV_LOG(ERR, "port %u flow director mode %d not supported",
3789 			dev->data->port_id, fdir_mode);
3790 		rte_errno = EINVAL;
3791 		return -rte_errno;
3792 	}
3793 	switch (filter_op) {
3794 	case RTE_ETH_FILTER_ADD:
3795 		return mlx5_fdir_filter_add(dev, arg);
3796 	case RTE_ETH_FILTER_UPDATE:
3797 		return mlx5_fdir_filter_update(dev, arg);
3798 	case RTE_ETH_FILTER_DELETE:
3799 		return mlx5_fdir_filter_delete(dev, arg);
3800 	case RTE_ETH_FILTER_FLUSH:
3801 		mlx5_fdir_filter_flush(dev);
3802 		break;
3803 	case RTE_ETH_FILTER_INFO:
3804 		mlx5_fdir_info_get(dev, arg);
3805 		break;
3806 	default:
3807 		DRV_LOG(DEBUG, "port %u unknown operation %u",
3808 			dev->data->port_id, filter_op);
3809 		rte_errno = EINVAL;
3810 		return -rte_errno;
3811 	}
3812 	return 0;
3813 }
3814 
3815 /**
3816  * Manage filter operations.
3817  *
3818  * @param dev
3819  *   Pointer to Ethernet device structure.
3820  * @param filter_type
3821  *   Filter type.
3822  * @param filter_op
3823  *   Operation to perform.
3824  * @param arg
3825  *   Pointer to operation-specific structure.
3826  *
3827  * @return
3828  *   0 on success, a negative errno value otherwise and rte_errno is set.
3829  */
3830 int
3831 mlx5_dev_filter_ctrl(struct rte_eth_dev *dev,
3832 		     enum rte_filter_type filter_type,
3833 		     enum rte_filter_op filter_op,
3834 		     void *arg)
3835 {
3836 	switch (filter_type) {
3837 	case RTE_ETH_FILTER_GENERIC:
3838 		if (filter_op != RTE_ETH_FILTER_GET) {
3839 			rte_errno = EINVAL;
3840 			return -rte_errno;
3841 		}
3842 		*(const void **)arg = &mlx5_flow_ops;
3843 		return 0;
3844 	case RTE_ETH_FILTER_FDIR:
3845 		return mlx5_fdir_ctrl_func(dev, filter_op, arg);
3846 	default:
3847 		DRV_LOG(ERR, "port %u filter type (%d) not supported",
3848 			dev->data->port_id, filter_type);
3849 		rte_errno = ENOTSUP;
3850 		return -rte_errno;
3851 	}
3852 	return 0;
3853 }
3854