1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2016 6WIND S.A.
3 * Copyright 2016 Mellanox Technologies, Ltd
6 #include <netinet/in.h>
13 #include <rte_common.h>
14 #include <rte_ether.h>
15 #include <rte_ethdev_driver.h>
16 #include <rte_eal_paging.h>
18 #include <rte_cycles.h>
19 #include <rte_flow_driver.h>
20 #include <rte_malloc.h>
23 #include <mlx5_glue.h>
24 #include <mlx5_devx_cmds.h>
26 #include <mlx5_malloc.h>
28 #include "mlx5_defs.h"
30 #include "mlx5_flow.h"
31 #include "mlx5_flow_os.h"
32 #include "mlx5_rxtx.h"
33 #include "mlx5_common_os.h"
34 #include "rte_pmd_mlx5.h"
36 struct tunnel_default_miss_ctx {
40 struct rte_flow_action_rss action_rss;
41 struct rte_flow_action_queue miss_queue;
42 struct rte_flow_action_jump miss_jump;
48 flow_tunnel_add_default_miss(struct rte_eth_dev *dev,
49 struct rte_flow *flow,
50 const struct rte_flow_attr *attr,
51 const struct rte_flow_action *app_actions,
53 struct tunnel_default_miss_ctx *ctx,
54 struct rte_flow_error *error);
55 static struct mlx5_flow_tunnel *
56 mlx5_find_tunnel_id(struct rte_eth_dev *dev, uint32_t id);
58 mlx5_flow_tunnel_free(struct rte_eth_dev *dev, struct mlx5_flow_tunnel *tunnel);
60 tunnel_flow_group_to_flow_table(struct rte_eth_dev *dev,
61 const struct mlx5_flow_tunnel *tunnel,
62 uint32_t group, uint32_t *table,
63 struct rte_flow_error *error);
65 static struct mlx5_flow_workspace *mlx5_flow_push_thread_workspace(void);
66 static void mlx5_flow_pop_thread_workspace(void);
69 /** Device flow drivers. */
70 extern const struct mlx5_flow_driver_ops mlx5_flow_verbs_drv_ops;
72 const struct mlx5_flow_driver_ops mlx5_flow_null_drv_ops;
74 const struct mlx5_flow_driver_ops *flow_drv_ops[] = {
75 [MLX5_FLOW_TYPE_MIN] = &mlx5_flow_null_drv_ops,
76 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
77 [MLX5_FLOW_TYPE_DV] = &mlx5_flow_dv_drv_ops,
79 [MLX5_FLOW_TYPE_VERBS] = &mlx5_flow_verbs_drv_ops,
80 [MLX5_FLOW_TYPE_MAX] = &mlx5_flow_null_drv_ops
83 /** Helper macro to build input graph for mlx5_flow_expand_rss(). */
84 #define MLX5_FLOW_EXPAND_RSS_NEXT(...) \
89 /** Node object of input graph for mlx5_flow_expand_rss(). */
90 struct mlx5_flow_expand_node {
91 const int *const next;
93 * List of next node indexes. Index 0 is interpreted as a terminator.
95 const enum rte_flow_item_type type;
96 /**< Pattern item type of current node. */
99 * RSS types bit-field associated with this node
100 * (see ETH_RSS_* definitions).
104 /** Object returned by mlx5_flow_expand_rss(). */
105 struct mlx5_flow_expand_rss {
107 /**< Number of entries @p patterns and @p priorities. */
109 struct rte_flow_item *pattern; /**< Expanded pattern array. */
110 uint32_t priority; /**< Priority offset for each expansion. */
114 static enum rte_flow_item_type
115 mlx5_flow_expand_rss_item_complete(const struct rte_flow_item *item)
117 enum rte_flow_item_type ret = RTE_FLOW_ITEM_TYPE_VOID;
118 uint16_t ether_type = 0;
119 uint16_t ether_type_m;
120 uint8_t ip_next_proto = 0;
121 uint8_t ip_next_proto_m;
123 if (item == NULL || item->spec == NULL)
125 switch (item->type) {
126 case RTE_FLOW_ITEM_TYPE_ETH:
128 ether_type_m = ((const struct rte_flow_item_eth *)
131 ether_type_m = rte_flow_item_eth_mask.type;
132 if (ether_type_m != RTE_BE16(0xFFFF))
134 ether_type = ((const struct rte_flow_item_eth *)
136 if (rte_be_to_cpu_16(ether_type) == RTE_ETHER_TYPE_IPV4)
137 ret = RTE_FLOW_ITEM_TYPE_IPV4;
138 else if (rte_be_to_cpu_16(ether_type) == RTE_ETHER_TYPE_IPV6)
139 ret = RTE_FLOW_ITEM_TYPE_IPV6;
140 else if (rte_be_to_cpu_16(ether_type) == RTE_ETHER_TYPE_VLAN)
141 ret = RTE_FLOW_ITEM_TYPE_VLAN;
143 ret = RTE_FLOW_ITEM_TYPE_END;
145 case RTE_FLOW_ITEM_TYPE_VLAN:
147 ether_type_m = ((const struct rte_flow_item_vlan *)
148 (item->mask))->inner_type;
150 ether_type_m = rte_flow_item_vlan_mask.inner_type;
151 if (ether_type_m != RTE_BE16(0xFFFF))
153 ether_type = ((const struct rte_flow_item_vlan *)
154 (item->spec))->inner_type;
155 if (rte_be_to_cpu_16(ether_type) == RTE_ETHER_TYPE_IPV4)
156 ret = RTE_FLOW_ITEM_TYPE_IPV4;
157 else if (rte_be_to_cpu_16(ether_type) == RTE_ETHER_TYPE_IPV6)
158 ret = RTE_FLOW_ITEM_TYPE_IPV6;
159 else if (rte_be_to_cpu_16(ether_type) == RTE_ETHER_TYPE_VLAN)
160 ret = RTE_FLOW_ITEM_TYPE_VLAN;
162 ret = RTE_FLOW_ITEM_TYPE_END;
164 case RTE_FLOW_ITEM_TYPE_IPV4:
166 ip_next_proto_m = ((const struct rte_flow_item_ipv4 *)
167 (item->mask))->hdr.next_proto_id;
170 rte_flow_item_ipv4_mask.hdr.next_proto_id;
171 if (ip_next_proto_m != 0xFF)
173 ip_next_proto = ((const struct rte_flow_item_ipv4 *)
174 (item->spec))->hdr.next_proto_id;
175 if (ip_next_proto == IPPROTO_UDP)
176 ret = RTE_FLOW_ITEM_TYPE_UDP;
177 else if (ip_next_proto == IPPROTO_TCP)
178 ret = RTE_FLOW_ITEM_TYPE_TCP;
179 else if (ip_next_proto == IPPROTO_IP)
180 ret = RTE_FLOW_ITEM_TYPE_IPV4;
181 else if (ip_next_proto == IPPROTO_IPV6)
182 ret = RTE_FLOW_ITEM_TYPE_IPV6;
184 ret = RTE_FLOW_ITEM_TYPE_END;
186 case RTE_FLOW_ITEM_TYPE_IPV6:
188 ip_next_proto_m = ((const struct rte_flow_item_ipv6 *)
189 (item->mask))->hdr.proto;
192 rte_flow_item_ipv6_mask.hdr.proto;
193 if (ip_next_proto_m != 0xFF)
195 ip_next_proto = ((const struct rte_flow_item_ipv6 *)
196 (item->spec))->hdr.proto;
197 if (ip_next_proto == IPPROTO_UDP)
198 ret = RTE_FLOW_ITEM_TYPE_UDP;
199 else if (ip_next_proto == IPPROTO_TCP)
200 ret = RTE_FLOW_ITEM_TYPE_TCP;
201 else if (ip_next_proto == IPPROTO_IP)
202 ret = RTE_FLOW_ITEM_TYPE_IPV4;
203 else if (ip_next_proto == IPPROTO_IPV6)
204 ret = RTE_FLOW_ITEM_TYPE_IPV6;
206 ret = RTE_FLOW_ITEM_TYPE_END;
209 ret = RTE_FLOW_ITEM_TYPE_VOID;
216 * Expand RSS flows into several possible flows according to the RSS hash
217 * fields requested and the driver capabilities.
220 * Buffer to store the result expansion.
222 * Buffer size in bytes. If 0, @p buf can be NULL.
226 * RSS types to expand (see ETH_RSS_* definitions).
228 * Input graph to expand @p pattern according to @p types.
229 * @param[in] graph_root_index
230 * Index of root node in @p graph, typically 0.
233 * A positive value representing the size of @p buf in bytes regardless of
234 * @p size on success, a negative errno value otherwise and rte_errno is
235 * set, the following errors are defined:
237 * -E2BIG: graph-depth @p graph is too deep.
240 mlx5_flow_expand_rss(struct mlx5_flow_expand_rss *buf, size_t size,
241 const struct rte_flow_item *pattern, uint64_t types,
242 const struct mlx5_flow_expand_node graph[],
243 int graph_root_index)
246 const struct rte_flow_item *item;
247 const struct mlx5_flow_expand_node *node = &graph[graph_root_index];
248 const int *next_node;
249 const int *stack[elt_n];
251 struct rte_flow_item flow_items[elt_n];
254 size_t user_pattern_size = 0;
256 const struct mlx5_flow_expand_node *next = NULL;
257 struct rte_flow_item missed_item;
260 const struct rte_flow_item *last_item = NULL;
262 memset(&missed_item, 0, sizeof(missed_item));
263 lsize = offsetof(struct mlx5_flow_expand_rss, entry) +
264 elt_n * sizeof(buf->entry[0]);
266 buf->entry[0].priority = 0;
267 buf->entry[0].pattern = (void *)&buf->entry[elt_n];
269 addr = buf->entry[0].pattern;
271 for (item = pattern; item->type != RTE_FLOW_ITEM_TYPE_END; item++) {
272 if (item->type != RTE_FLOW_ITEM_TYPE_VOID)
274 for (i = 0; node->next && node->next[i]; ++i) {
275 next = &graph[node->next[i]];
276 if (next->type == item->type)
281 user_pattern_size += sizeof(*item);
283 user_pattern_size += sizeof(*item); /* Handle END item. */
284 lsize += user_pattern_size;
285 /* Copy the user pattern in the first entry of the buffer. */
287 rte_memcpy(addr, pattern, user_pattern_size);
288 addr = (void *)(((uintptr_t)addr) + user_pattern_size);
291 /* Start expanding. */
292 memset(flow_items, 0, sizeof(flow_items));
293 user_pattern_size -= sizeof(*item);
295 * Check if the last valid item has spec set, need complete pattern,
296 * and the pattern can be used for expansion.
298 missed_item.type = mlx5_flow_expand_rss_item_complete(last_item);
299 if (missed_item.type == RTE_FLOW_ITEM_TYPE_END) {
300 /* Item type END indicates expansion is not required. */
303 if (missed_item.type != RTE_FLOW_ITEM_TYPE_VOID) {
306 for (i = 0; node->next && node->next[i]; ++i) {
307 next = &graph[node->next[i]];
308 if (next->type == missed_item.type) {
309 flow_items[0].type = missed_item.type;
310 flow_items[1].type = RTE_FLOW_ITEM_TYPE_END;
316 if (next && missed) {
317 elt = 2; /* missed item + item end. */
319 lsize += elt * sizeof(*item) + user_pattern_size;
320 if ((node->rss_types & types) && lsize <= size) {
321 buf->entry[buf->entries].priority = 1;
322 buf->entry[buf->entries].pattern = addr;
324 rte_memcpy(addr, buf->entry[0].pattern,
326 addr = (void *)(((uintptr_t)addr) + user_pattern_size);
327 rte_memcpy(addr, flow_items, elt * sizeof(*item));
328 addr = (void *)(((uintptr_t)addr) +
329 elt * sizeof(*item));
332 memset(flow_items, 0, sizeof(flow_items));
333 next_node = node->next;
334 stack[stack_pos] = next_node;
335 node = next_node ? &graph[*next_node] : NULL;
337 flow_items[stack_pos].type = node->type;
338 if (node->rss_types & types) {
340 * compute the number of items to copy from the
341 * expansion and copy it.
342 * When the stack_pos is 0, there are 1 element in it,
343 * plus the addition END item.
346 flow_items[stack_pos + 1].type = RTE_FLOW_ITEM_TYPE_END;
347 lsize += elt * sizeof(*item) + user_pattern_size;
349 size_t n = elt * sizeof(*item);
351 buf->entry[buf->entries].priority =
352 stack_pos + 1 + missed;
353 buf->entry[buf->entries].pattern = addr;
355 rte_memcpy(addr, buf->entry[0].pattern,
357 addr = (void *)(((uintptr_t)addr) +
359 rte_memcpy(addr, &missed_item,
360 missed * sizeof(*item));
361 addr = (void *)(((uintptr_t)addr) +
362 missed * sizeof(*item));
363 rte_memcpy(addr, flow_items, n);
364 addr = (void *)(((uintptr_t)addr) + n);
369 next_node = node->next;
370 if (stack_pos++ == elt_n) {
374 stack[stack_pos] = next_node;
375 } else if (*(next_node + 1)) {
376 /* Follow up with the next possibility. */
379 /* Move to the next path. */
381 next_node = stack[--stack_pos];
383 stack[stack_pos] = next_node;
385 node = *next_node ? &graph[*next_node] : NULL;
387 /* no expanded flows but we have missed item, create one rule for it */
388 if (buf->entries == 1 && missed != 0) {
390 lsize += elt * sizeof(*item) + user_pattern_size;
392 buf->entry[buf->entries].priority = 1;
393 buf->entry[buf->entries].pattern = addr;
395 flow_items[0].type = missed_item.type;
396 flow_items[1].type = RTE_FLOW_ITEM_TYPE_END;
397 rte_memcpy(addr, buf->entry[0].pattern,
399 addr = (void *)(((uintptr_t)addr) + user_pattern_size);
400 rte_memcpy(addr, flow_items, elt * sizeof(*item));
406 enum mlx5_expansion {
408 MLX5_EXPANSION_ROOT_OUTER,
409 MLX5_EXPANSION_ROOT_ETH_VLAN,
410 MLX5_EXPANSION_ROOT_OUTER_ETH_VLAN,
411 MLX5_EXPANSION_OUTER_ETH,
412 MLX5_EXPANSION_OUTER_ETH_VLAN,
413 MLX5_EXPANSION_OUTER_VLAN,
414 MLX5_EXPANSION_OUTER_IPV4,
415 MLX5_EXPANSION_OUTER_IPV4_UDP,
416 MLX5_EXPANSION_OUTER_IPV4_TCP,
417 MLX5_EXPANSION_OUTER_IPV6,
418 MLX5_EXPANSION_OUTER_IPV6_UDP,
419 MLX5_EXPANSION_OUTER_IPV6_TCP,
420 MLX5_EXPANSION_VXLAN,
421 MLX5_EXPANSION_VXLAN_GPE,
425 MLX5_EXPANSION_ETH_VLAN,
428 MLX5_EXPANSION_IPV4_UDP,
429 MLX5_EXPANSION_IPV4_TCP,
431 MLX5_EXPANSION_IPV6_UDP,
432 MLX5_EXPANSION_IPV6_TCP,
435 /** Supported expansion of items. */
436 static const struct mlx5_flow_expand_node mlx5_support_expansion[] = {
437 [MLX5_EXPANSION_ROOT] = {
438 .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_ETH,
440 MLX5_EXPANSION_IPV6),
441 .type = RTE_FLOW_ITEM_TYPE_END,
443 [MLX5_EXPANSION_ROOT_OUTER] = {
444 .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_OUTER_ETH,
445 MLX5_EXPANSION_OUTER_IPV4,
446 MLX5_EXPANSION_OUTER_IPV6),
447 .type = RTE_FLOW_ITEM_TYPE_END,
449 [MLX5_EXPANSION_ROOT_ETH_VLAN] = {
450 .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_ETH_VLAN),
451 .type = RTE_FLOW_ITEM_TYPE_END,
453 [MLX5_EXPANSION_ROOT_OUTER_ETH_VLAN] = {
454 .next = MLX5_FLOW_EXPAND_RSS_NEXT
455 (MLX5_EXPANSION_OUTER_ETH_VLAN),
456 .type = RTE_FLOW_ITEM_TYPE_END,
458 [MLX5_EXPANSION_OUTER_ETH] = {
459 .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_OUTER_IPV4,
460 MLX5_EXPANSION_OUTER_IPV6,
461 MLX5_EXPANSION_MPLS),
462 .type = RTE_FLOW_ITEM_TYPE_ETH,
465 [MLX5_EXPANSION_OUTER_ETH_VLAN] = {
466 .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_OUTER_VLAN),
467 .type = RTE_FLOW_ITEM_TYPE_ETH,
470 [MLX5_EXPANSION_OUTER_VLAN] = {
471 .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_OUTER_IPV4,
472 MLX5_EXPANSION_OUTER_IPV6),
473 .type = RTE_FLOW_ITEM_TYPE_VLAN,
475 [MLX5_EXPANSION_OUTER_IPV4] = {
476 .next = MLX5_FLOW_EXPAND_RSS_NEXT
477 (MLX5_EXPANSION_OUTER_IPV4_UDP,
478 MLX5_EXPANSION_OUTER_IPV4_TCP,
481 MLX5_EXPANSION_IPV6),
482 .type = RTE_FLOW_ITEM_TYPE_IPV4,
483 .rss_types = ETH_RSS_IPV4 | ETH_RSS_FRAG_IPV4 |
484 ETH_RSS_NONFRAG_IPV4_OTHER,
486 [MLX5_EXPANSION_OUTER_IPV4_UDP] = {
487 .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_VXLAN,
488 MLX5_EXPANSION_VXLAN_GPE),
489 .type = RTE_FLOW_ITEM_TYPE_UDP,
490 .rss_types = ETH_RSS_NONFRAG_IPV4_UDP,
492 [MLX5_EXPANSION_OUTER_IPV4_TCP] = {
493 .type = RTE_FLOW_ITEM_TYPE_TCP,
494 .rss_types = ETH_RSS_NONFRAG_IPV4_TCP,
496 [MLX5_EXPANSION_OUTER_IPV6] = {
497 .next = MLX5_FLOW_EXPAND_RSS_NEXT
498 (MLX5_EXPANSION_OUTER_IPV6_UDP,
499 MLX5_EXPANSION_OUTER_IPV6_TCP,
501 MLX5_EXPANSION_IPV6),
502 .type = RTE_FLOW_ITEM_TYPE_IPV6,
503 .rss_types = ETH_RSS_IPV6 | ETH_RSS_FRAG_IPV6 |
504 ETH_RSS_NONFRAG_IPV6_OTHER,
506 [MLX5_EXPANSION_OUTER_IPV6_UDP] = {
507 .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_VXLAN,
508 MLX5_EXPANSION_VXLAN_GPE),
509 .type = RTE_FLOW_ITEM_TYPE_UDP,
510 .rss_types = ETH_RSS_NONFRAG_IPV6_UDP,
512 [MLX5_EXPANSION_OUTER_IPV6_TCP] = {
513 .type = RTE_FLOW_ITEM_TYPE_TCP,
514 .rss_types = ETH_RSS_NONFRAG_IPV6_TCP,
516 [MLX5_EXPANSION_VXLAN] = {
517 .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_ETH,
519 MLX5_EXPANSION_IPV6),
520 .type = RTE_FLOW_ITEM_TYPE_VXLAN,
522 [MLX5_EXPANSION_VXLAN_GPE] = {
523 .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_ETH,
525 MLX5_EXPANSION_IPV6),
526 .type = RTE_FLOW_ITEM_TYPE_VXLAN_GPE,
528 [MLX5_EXPANSION_GRE] = {
529 .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4),
530 .type = RTE_FLOW_ITEM_TYPE_GRE,
532 [MLX5_EXPANSION_MPLS] = {
533 .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4,
534 MLX5_EXPANSION_IPV6),
535 .type = RTE_FLOW_ITEM_TYPE_MPLS,
537 [MLX5_EXPANSION_ETH] = {
538 .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4,
539 MLX5_EXPANSION_IPV6),
540 .type = RTE_FLOW_ITEM_TYPE_ETH,
542 [MLX5_EXPANSION_ETH_VLAN] = {
543 .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_VLAN),
544 .type = RTE_FLOW_ITEM_TYPE_ETH,
546 [MLX5_EXPANSION_VLAN] = {
547 .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4,
548 MLX5_EXPANSION_IPV6),
549 .type = RTE_FLOW_ITEM_TYPE_VLAN,
551 [MLX5_EXPANSION_IPV4] = {
552 .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4_UDP,
553 MLX5_EXPANSION_IPV4_TCP),
554 .type = RTE_FLOW_ITEM_TYPE_IPV4,
555 .rss_types = ETH_RSS_IPV4 | ETH_RSS_FRAG_IPV4 |
556 ETH_RSS_NONFRAG_IPV4_OTHER,
558 [MLX5_EXPANSION_IPV4_UDP] = {
559 .type = RTE_FLOW_ITEM_TYPE_UDP,
560 .rss_types = ETH_RSS_NONFRAG_IPV4_UDP,
562 [MLX5_EXPANSION_IPV4_TCP] = {
563 .type = RTE_FLOW_ITEM_TYPE_TCP,
564 .rss_types = ETH_RSS_NONFRAG_IPV4_TCP,
566 [MLX5_EXPANSION_IPV6] = {
567 .next = MLX5_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV6_UDP,
568 MLX5_EXPANSION_IPV6_TCP),
569 .type = RTE_FLOW_ITEM_TYPE_IPV6,
570 .rss_types = ETH_RSS_IPV6 | ETH_RSS_FRAG_IPV6 |
571 ETH_RSS_NONFRAG_IPV6_OTHER,
573 [MLX5_EXPANSION_IPV6_UDP] = {
574 .type = RTE_FLOW_ITEM_TYPE_UDP,
575 .rss_types = ETH_RSS_NONFRAG_IPV6_UDP,
577 [MLX5_EXPANSION_IPV6_TCP] = {
578 .type = RTE_FLOW_ITEM_TYPE_TCP,
579 .rss_types = ETH_RSS_NONFRAG_IPV6_TCP,
583 static struct rte_flow_shared_action *
584 mlx5_shared_action_create(struct rte_eth_dev *dev,
585 const struct rte_flow_shared_action_conf *conf,
586 const struct rte_flow_action *action,
587 struct rte_flow_error *error);
588 static int mlx5_shared_action_destroy
589 (struct rte_eth_dev *dev,
590 struct rte_flow_shared_action *shared_action,
591 struct rte_flow_error *error);
592 static int mlx5_shared_action_update
593 (struct rte_eth_dev *dev,
594 struct rte_flow_shared_action *shared_action,
595 const struct rte_flow_action *action,
596 struct rte_flow_error *error);
597 static int mlx5_shared_action_query
598 (struct rte_eth_dev *dev,
599 const struct rte_flow_shared_action *action,
601 struct rte_flow_error *error);
603 mlx5_flow_tunnel_decap_set(struct rte_eth_dev *dev,
604 struct rte_flow_tunnel *app_tunnel,
605 struct rte_flow_action **actions,
606 uint32_t *num_of_actions,
607 struct rte_flow_error *error);
609 mlx5_flow_tunnel_match(struct rte_eth_dev *dev,
610 struct rte_flow_tunnel *app_tunnel,
611 struct rte_flow_item **items,
612 uint32_t *num_of_items,
613 struct rte_flow_error *error);
615 mlx5_flow_tunnel_item_release(struct rte_eth_dev *dev,
616 struct rte_flow_item *pmd_items,
617 uint32_t num_items, struct rte_flow_error *err);
619 mlx5_flow_tunnel_action_release(struct rte_eth_dev *dev,
620 struct rte_flow_action *pmd_actions,
621 uint32_t num_actions,
622 struct rte_flow_error *err);
624 mlx5_flow_tunnel_get_restore_info(struct rte_eth_dev *dev,
626 struct rte_flow_restore_info *info,
627 struct rte_flow_error *err);
629 static const struct rte_flow_ops mlx5_flow_ops = {
630 .validate = mlx5_flow_validate,
631 .create = mlx5_flow_create,
632 .destroy = mlx5_flow_destroy,
633 .flush = mlx5_flow_flush,
634 .isolate = mlx5_flow_isolate,
635 .query = mlx5_flow_query,
636 .dev_dump = mlx5_flow_dev_dump,
637 .get_aged_flows = mlx5_flow_get_aged_flows,
638 .shared_action_create = mlx5_shared_action_create,
639 .shared_action_destroy = mlx5_shared_action_destroy,
640 .shared_action_update = mlx5_shared_action_update,
641 .shared_action_query = mlx5_shared_action_query,
642 .tunnel_decap_set = mlx5_flow_tunnel_decap_set,
643 .tunnel_match = mlx5_flow_tunnel_match,
644 .tunnel_action_decap_release = mlx5_flow_tunnel_action_release,
645 .tunnel_item_release = mlx5_flow_tunnel_item_release,
646 .get_restore_info = mlx5_flow_tunnel_get_restore_info,
649 /* Tunnel information. */
650 struct mlx5_flow_tunnel_info {
651 uint64_t tunnel; /**< Tunnel bit (see MLX5_FLOW_*). */
652 uint32_t ptype; /**< Tunnel Ptype (see RTE_PTYPE_*). */
655 static struct mlx5_flow_tunnel_info tunnels_info[] = {
657 .tunnel = MLX5_FLOW_LAYER_VXLAN,
658 .ptype = RTE_PTYPE_TUNNEL_VXLAN | RTE_PTYPE_L4_UDP,
661 .tunnel = MLX5_FLOW_LAYER_GENEVE,
662 .ptype = RTE_PTYPE_TUNNEL_GENEVE | RTE_PTYPE_L4_UDP,
665 .tunnel = MLX5_FLOW_LAYER_VXLAN_GPE,
666 .ptype = RTE_PTYPE_TUNNEL_VXLAN_GPE | RTE_PTYPE_L4_UDP,
669 .tunnel = MLX5_FLOW_LAYER_GRE,
670 .ptype = RTE_PTYPE_TUNNEL_GRE,
673 .tunnel = MLX5_FLOW_LAYER_MPLS | MLX5_FLOW_LAYER_OUTER_L4_UDP,
674 .ptype = RTE_PTYPE_TUNNEL_MPLS_IN_UDP | RTE_PTYPE_L4_UDP,
677 .tunnel = MLX5_FLOW_LAYER_MPLS,
678 .ptype = RTE_PTYPE_TUNNEL_MPLS_IN_GRE,
681 .tunnel = MLX5_FLOW_LAYER_NVGRE,
682 .ptype = RTE_PTYPE_TUNNEL_NVGRE,
685 .tunnel = MLX5_FLOW_LAYER_IPIP,
686 .ptype = RTE_PTYPE_TUNNEL_IP,
689 .tunnel = MLX5_FLOW_LAYER_IPV6_ENCAP,
690 .ptype = RTE_PTYPE_TUNNEL_IP,
693 .tunnel = MLX5_FLOW_LAYER_GTP,
694 .ptype = RTE_PTYPE_TUNNEL_GTPU,
698 /* Key of thread specific flow workspace data. */
699 static pthread_key_t key_workspace;
701 /* Thread specific flow workspace data once initialization data. */
702 static pthread_once_t key_workspace_init;
706 * Translate tag ID to register.
709 * Pointer to the Ethernet device structure.
711 * The feature that request the register.
713 * The request register ID.
715 * Error description in case of any.
718 * The request register on success, a negative errno
719 * value otherwise and rte_errno is set.
722 mlx5_flow_get_reg_id(struct rte_eth_dev *dev,
723 enum mlx5_feature_name feature,
725 struct rte_flow_error *error)
727 struct mlx5_priv *priv = dev->data->dev_private;
728 struct mlx5_dev_config *config = &priv->config;
729 enum modify_reg start_reg;
730 bool skip_mtr_reg = false;
733 case MLX5_HAIRPIN_RX:
735 case MLX5_HAIRPIN_TX:
737 case MLX5_METADATA_RX:
738 switch (config->dv_xmeta_en) {
739 case MLX5_XMETA_MODE_LEGACY:
741 case MLX5_XMETA_MODE_META16:
743 case MLX5_XMETA_MODE_META32:
747 case MLX5_METADATA_TX:
749 case MLX5_METADATA_FDB:
750 switch (config->dv_xmeta_en) {
751 case MLX5_XMETA_MODE_LEGACY:
753 case MLX5_XMETA_MODE_META16:
755 case MLX5_XMETA_MODE_META32:
760 switch (config->dv_xmeta_en) {
761 case MLX5_XMETA_MODE_LEGACY:
763 case MLX5_XMETA_MODE_META16:
765 case MLX5_XMETA_MODE_META32:
771 * If meter color and flow match share one register, flow match
772 * should use the meter color register for match.
774 if (priv->mtr_reg_share)
775 return priv->mtr_color_reg;
777 return priv->mtr_color_reg != REG_C_2 ? REG_C_2 :
780 case MLX5_ASO_FLOW_HIT: /* Both features use the same REG_C. */
781 MLX5_ASSERT(priv->mtr_color_reg != REG_NON);
782 return priv->mtr_color_reg;
785 * Metadata COPY_MARK register using is in meter suffix sub
786 * flow while with meter. It's safe to share the same register.
788 return priv->mtr_color_reg != REG_C_2 ? REG_C_2 : REG_C_3;
791 * If meter is enable, it will engage the register for color
792 * match and flow match. If meter color match is not using the
793 * REG_C_2, need to skip the REG_C_x be used by meter color
795 * If meter is disable, free to use all available registers.
797 start_reg = priv->mtr_color_reg != REG_C_2 ? REG_C_2 :
798 (priv->mtr_reg_share ? REG_C_3 : REG_C_4);
799 skip_mtr_reg = !!(priv->mtr_en && start_reg == REG_C_2);
800 if (id > (REG_C_7 - start_reg))
801 return rte_flow_error_set(error, EINVAL,
802 RTE_FLOW_ERROR_TYPE_ITEM,
803 NULL, "invalid tag id");
804 if (config->flow_mreg_c[id + start_reg - REG_C_0] == REG_NON)
805 return rte_flow_error_set(error, ENOTSUP,
806 RTE_FLOW_ERROR_TYPE_ITEM,
807 NULL, "unsupported tag id");
809 * This case means meter is using the REG_C_x great than 2.
810 * Take care not to conflict with meter color REG_C_x.
811 * If the available index REG_C_y >= REG_C_x, skip the
814 if (skip_mtr_reg && config->flow_mreg_c
815 [id + start_reg - REG_C_0] >= priv->mtr_color_reg) {
816 if (id >= (REG_C_7 - start_reg))
817 return rte_flow_error_set(error, EINVAL,
818 RTE_FLOW_ERROR_TYPE_ITEM,
819 NULL, "invalid tag id");
820 if (config->flow_mreg_c
821 [id + 1 + start_reg - REG_C_0] != REG_NON)
822 return config->flow_mreg_c
823 [id + 1 + start_reg - REG_C_0];
824 return rte_flow_error_set(error, ENOTSUP,
825 RTE_FLOW_ERROR_TYPE_ITEM,
826 NULL, "unsupported tag id");
828 return config->flow_mreg_c[id + start_reg - REG_C_0];
831 return rte_flow_error_set(error, EINVAL,
832 RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
833 NULL, "invalid feature name");
837 * Check extensive flow metadata register support.
840 * Pointer to rte_eth_dev structure.
843 * True if device supports extensive flow metadata register, otherwise false.
846 mlx5_flow_ext_mreg_supported(struct rte_eth_dev *dev)
848 struct mlx5_priv *priv = dev->data->dev_private;
849 struct mlx5_dev_config *config = &priv->config;
852 * Having available reg_c can be regarded inclusively as supporting
853 * extensive flow metadata register, which could mean,
854 * - metadata register copy action by modify header.
855 * - 16 modify header actions is supported.
856 * - reg_c's are preserved across different domain (FDB and NIC) on
857 * packet loopback by flow lookup miss.
859 return config->flow_mreg_c[2] != REG_NON;
863 * Verify the @p item specifications (spec, last, mask) are compatible with the
867 * Item specification.
869 * @p item->mask or flow default bit-masks.
870 * @param[in] nic_mask
871 * Bit-masks covering supported fields by the NIC to compare with user mask.
873 * Bit-masks size in bytes.
874 * @param[in] range_accepted
875 * True if range of values is accepted for specific fields, false otherwise.
877 * Pointer to error structure.
880 * 0 on success, a negative errno value otherwise and rte_errno is set.
883 mlx5_flow_item_acceptable(const struct rte_flow_item *item,
885 const uint8_t *nic_mask,
888 struct rte_flow_error *error)
892 MLX5_ASSERT(nic_mask);
893 for (i = 0; i < size; ++i)
894 if ((nic_mask[i] | mask[i]) != nic_mask[i])
895 return rte_flow_error_set(error, ENOTSUP,
896 RTE_FLOW_ERROR_TYPE_ITEM,
898 "mask enables non supported"
900 if (!item->spec && (item->mask || item->last))
901 return rte_flow_error_set(error, EINVAL,
902 RTE_FLOW_ERROR_TYPE_ITEM, item,
903 "mask/last without a spec is not"
905 if (item->spec && item->last && !range_accepted) {
911 for (i = 0; i < size; ++i) {
912 spec[i] = ((const uint8_t *)item->spec)[i] & mask[i];
913 last[i] = ((const uint8_t *)item->last)[i] & mask[i];
915 ret = memcmp(spec, last, size);
917 return rte_flow_error_set(error, EINVAL,
918 RTE_FLOW_ERROR_TYPE_ITEM,
920 "range is not valid");
926 * Adjust the hash fields according to the @p flow information.
928 * @param[in] dev_flow.
929 * Pointer to the mlx5_flow.
931 * 1 when the hash field is for a tunnel item.
932 * @param[in] layer_types
934 * @param[in] hash_fields
938 * The hash fields that should be used.
941 mlx5_flow_hashfields_adjust(struct mlx5_flow_rss_desc *rss_desc,
942 int tunnel __rte_unused, uint64_t layer_types,
943 uint64_t hash_fields)
945 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
946 int rss_request_inner = rss_desc->level >= 2;
948 /* Check RSS hash level for tunnel. */
949 if (tunnel && rss_request_inner)
950 hash_fields |= IBV_RX_HASH_INNER;
951 else if (tunnel || rss_request_inner)
954 /* Check if requested layer matches RSS hash fields. */
955 if (!(rss_desc->types & layer_types))
961 * Lookup and set the ptype in the data Rx part. A single Ptype can be used,
962 * if several tunnel rules are used on this queue, the tunnel ptype will be
966 * Rx queue to update.
969 flow_rxq_tunnel_ptype_update(struct mlx5_rxq_ctrl *rxq_ctrl)
972 uint32_t tunnel_ptype = 0;
974 /* Look up for the ptype to use. */
975 for (i = 0; i != MLX5_FLOW_TUNNEL; ++i) {
976 if (!rxq_ctrl->flow_tunnels_n[i])
979 tunnel_ptype = tunnels_info[i].ptype;
985 rxq_ctrl->rxq.tunnel = tunnel_ptype;
989 * Set the Rx queue flags (Mark/Flag and Tunnel Ptypes) according to the devive
993 * Pointer to the Ethernet device structure.
994 * @param[in] dev_handle
995 * Pointer to device flow handle structure.
998 flow_drv_rxq_flags_set(struct rte_eth_dev *dev,
999 struct mlx5_flow_handle *dev_handle)
1001 struct mlx5_priv *priv = dev->data->dev_private;
1002 const int mark = dev_handle->mark;
1003 const int tunnel = !!(dev_handle->layers & MLX5_FLOW_LAYER_TUNNEL);
1004 struct mlx5_hrxq *hrxq;
1007 if (dev_handle->fate_action != MLX5_FLOW_FATE_QUEUE)
1009 hrxq = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_HRXQ],
1010 dev_handle->rix_hrxq);
1013 for (i = 0; i != hrxq->ind_table->queues_n; ++i) {
1014 int idx = hrxq->ind_table->queues[i];
1015 struct mlx5_rxq_ctrl *rxq_ctrl =
1016 container_of((*priv->rxqs)[idx],
1017 struct mlx5_rxq_ctrl, rxq);
1020 * To support metadata register copy on Tx loopback,
1021 * this must be always enabled (metadata may arive
1022 * from other port - not from local flows only.
1024 if (priv->config.dv_flow_en &&
1025 priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1026 mlx5_flow_ext_mreg_supported(dev)) {
1027 rxq_ctrl->rxq.mark = 1;
1028 rxq_ctrl->flow_mark_n = 1;
1030 rxq_ctrl->rxq.mark = 1;
1031 rxq_ctrl->flow_mark_n++;
1036 /* Increase the counter matching the flow. */
1037 for (j = 0; j != MLX5_FLOW_TUNNEL; ++j) {
1038 if ((tunnels_info[j].tunnel &
1039 dev_handle->layers) ==
1040 tunnels_info[j].tunnel) {
1041 rxq_ctrl->flow_tunnels_n[j]++;
1045 flow_rxq_tunnel_ptype_update(rxq_ctrl);
1051 * Set the Rx queue flags (Mark/Flag and Tunnel Ptypes) for a flow
1054 * Pointer to the Ethernet device structure.
1056 * Pointer to flow structure.
1059 flow_rxq_flags_set(struct rte_eth_dev *dev, struct rte_flow *flow)
1061 struct mlx5_priv *priv = dev->data->dev_private;
1062 uint32_t handle_idx;
1063 struct mlx5_flow_handle *dev_handle;
1065 SILIST_FOREACH(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW], flow->dev_handles,
1066 handle_idx, dev_handle, next)
1067 flow_drv_rxq_flags_set(dev, dev_handle);
1071 * Clear the Rx queue flags (Mark/Flag and Tunnel Ptype) associated with the
1072 * device flow if no other flow uses it with the same kind of request.
1075 * Pointer to Ethernet device.
1076 * @param[in] dev_handle
1077 * Pointer to the device flow handle structure.
1080 flow_drv_rxq_flags_trim(struct rte_eth_dev *dev,
1081 struct mlx5_flow_handle *dev_handle)
1083 struct mlx5_priv *priv = dev->data->dev_private;
1084 const int mark = dev_handle->mark;
1085 const int tunnel = !!(dev_handle->layers & MLX5_FLOW_LAYER_TUNNEL);
1086 struct mlx5_hrxq *hrxq;
1089 if (dev_handle->fate_action != MLX5_FLOW_FATE_QUEUE)
1091 hrxq = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_HRXQ],
1092 dev_handle->rix_hrxq);
1095 MLX5_ASSERT(dev->data->dev_started);
1096 for (i = 0; i != hrxq->ind_table->queues_n; ++i) {
1097 int idx = hrxq->ind_table->queues[i];
1098 struct mlx5_rxq_ctrl *rxq_ctrl =
1099 container_of((*priv->rxqs)[idx],
1100 struct mlx5_rxq_ctrl, rxq);
1102 if (priv->config.dv_flow_en &&
1103 priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1104 mlx5_flow_ext_mreg_supported(dev)) {
1105 rxq_ctrl->rxq.mark = 1;
1106 rxq_ctrl->flow_mark_n = 1;
1108 rxq_ctrl->flow_mark_n--;
1109 rxq_ctrl->rxq.mark = !!rxq_ctrl->flow_mark_n;
1114 /* Decrease the counter matching the flow. */
1115 for (j = 0; j != MLX5_FLOW_TUNNEL; ++j) {
1116 if ((tunnels_info[j].tunnel &
1117 dev_handle->layers) ==
1118 tunnels_info[j].tunnel) {
1119 rxq_ctrl->flow_tunnels_n[j]--;
1123 flow_rxq_tunnel_ptype_update(rxq_ctrl);
1129 * Clear the Rx queue flags (Mark/Flag and Tunnel Ptype) associated with the
1130 * @p flow if no other flow uses it with the same kind of request.
1133 * Pointer to Ethernet device.
1135 * Pointer to the flow.
1138 flow_rxq_flags_trim(struct rte_eth_dev *dev, struct rte_flow *flow)
1140 struct mlx5_priv *priv = dev->data->dev_private;
1141 uint32_t handle_idx;
1142 struct mlx5_flow_handle *dev_handle;
1144 SILIST_FOREACH(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW], flow->dev_handles,
1145 handle_idx, dev_handle, next)
1146 flow_drv_rxq_flags_trim(dev, dev_handle);
1150 * Clear the Mark/Flag and Tunnel ptype information in all Rx queues.
1153 * Pointer to Ethernet device.
1156 flow_rxq_flags_clear(struct rte_eth_dev *dev)
1158 struct mlx5_priv *priv = dev->data->dev_private;
1161 for (i = 0; i != priv->rxqs_n; ++i) {
1162 struct mlx5_rxq_ctrl *rxq_ctrl;
1165 if (!(*priv->rxqs)[i])
1167 rxq_ctrl = container_of((*priv->rxqs)[i],
1168 struct mlx5_rxq_ctrl, rxq);
1169 rxq_ctrl->flow_mark_n = 0;
1170 rxq_ctrl->rxq.mark = 0;
1171 for (j = 0; j != MLX5_FLOW_TUNNEL; ++j)
1172 rxq_ctrl->flow_tunnels_n[j] = 0;
1173 rxq_ctrl->rxq.tunnel = 0;
1178 * Set the Rx queue dynamic metadata (mask and offset) for a flow
1181 * Pointer to the Ethernet device structure.
1184 mlx5_flow_rxq_dynf_metadata_set(struct rte_eth_dev *dev)
1186 struct mlx5_priv *priv = dev->data->dev_private;
1187 struct mlx5_rxq_data *data;
1190 for (i = 0; i != priv->rxqs_n; ++i) {
1191 if (!(*priv->rxqs)[i])
1193 data = (*priv->rxqs)[i];
1194 if (!rte_flow_dynf_metadata_avail()) {
1195 data->dynf_meta = 0;
1196 data->flow_meta_mask = 0;
1197 data->flow_meta_offset = -1;
1199 data->dynf_meta = 1;
1200 data->flow_meta_mask = rte_flow_dynf_metadata_mask;
1201 data->flow_meta_offset = rte_flow_dynf_metadata_offs;
1207 * return a pointer to the desired action in the list of actions.
1209 * @param[in] actions
1210 * The list of actions to search the action in.
1212 * The action to find.
1215 * Pointer to the action in the list, if found. NULL otherwise.
1217 const struct rte_flow_action *
1218 mlx5_flow_find_action(const struct rte_flow_action *actions,
1219 enum rte_flow_action_type action)
1221 if (actions == NULL)
1223 for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++)
1224 if (actions->type == action)
1230 * Validate the flag action.
1232 * @param[in] action_flags
1233 * Bit-fields that holds the actions detected until now.
1235 * Attributes of flow that includes this action.
1237 * Pointer to error structure.
1240 * 0 on success, a negative errno value otherwise and rte_errno is set.
1243 mlx5_flow_validate_action_flag(uint64_t action_flags,
1244 const struct rte_flow_attr *attr,
1245 struct rte_flow_error *error)
1247 if (action_flags & MLX5_FLOW_ACTION_MARK)
1248 return rte_flow_error_set(error, EINVAL,
1249 RTE_FLOW_ERROR_TYPE_ACTION, NULL,
1250 "can't mark and flag in same flow");
1251 if (action_flags & MLX5_FLOW_ACTION_FLAG)
1252 return rte_flow_error_set(error, EINVAL,
1253 RTE_FLOW_ERROR_TYPE_ACTION, NULL,
1255 " actions in same flow");
1257 return rte_flow_error_set(error, ENOTSUP,
1258 RTE_FLOW_ERROR_TYPE_ATTR_EGRESS, NULL,
1259 "flag action not supported for "
1265 * Validate the mark action.
1268 * Pointer to the queue action.
1269 * @param[in] action_flags
1270 * Bit-fields that holds the actions detected until now.
1272 * Attributes of flow that includes this action.
1274 * Pointer to error structure.
1277 * 0 on success, a negative errno value otherwise and rte_errno is set.
1280 mlx5_flow_validate_action_mark(const struct rte_flow_action *action,
1281 uint64_t action_flags,
1282 const struct rte_flow_attr *attr,
1283 struct rte_flow_error *error)
1285 const struct rte_flow_action_mark *mark = action->conf;
1288 return rte_flow_error_set(error, EINVAL,
1289 RTE_FLOW_ERROR_TYPE_ACTION,
1291 "configuration cannot be null");
1292 if (mark->id >= MLX5_FLOW_MARK_MAX)
1293 return rte_flow_error_set(error, EINVAL,
1294 RTE_FLOW_ERROR_TYPE_ACTION_CONF,
1296 "mark id must in 0 <= id < "
1297 RTE_STR(MLX5_FLOW_MARK_MAX));
1298 if (action_flags & MLX5_FLOW_ACTION_FLAG)
1299 return rte_flow_error_set(error, EINVAL,
1300 RTE_FLOW_ERROR_TYPE_ACTION, NULL,
1301 "can't flag and mark in same flow");
1302 if (action_flags & MLX5_FLOW_ACTION_MARK)
1303 return rte_flow_error_set(error, EINVAL,
1304 RTE_FLOW_ERROR_TYPE_ACTION, NULL,
1305 "can't have 2 mark actions in same"
1308 return rte_flow_error_set(error, ENOTSUP,
1309 RTE_FLOW_ERROR_TYPE_ATTR_EGRESS, NULL,
1310 "mark action not supported for "
1316 * Validate the drop action.
1318 * @param[in] action_flags
1319 * Bit-fields that holds the actions detected until now.
1321 * Attributes of flow that includes this action.
1323 * Pointer to error structure.
1326 * 0 on success, a negative errno value otherwise and rte_errno is set.
1329 mlx5_flow_validate_action_drop(uint64_t action_flags __rte_unused,
1330 const struct rte_flow_attr *attr,
1331 struct rte_flow_error *error)
1334 return rte_flow_error_set(error, ENOTSUP,
1335 RTE_FLOW_ERROR_TYPE_ATTR_EGRESS, NULL,
1336 "drop action not supported for "
1342 * Validate the queue action.
1345 * Pointer to the queue action.
1346 * @param[in] action_flags
1347 * Bit-fields that holds the actions detected until now.
1349 * Pointer to the Ethernet device structure.
1351 * Attributes of flow that includes this action.
1353 * Pointer to error structure.
1356 * 0 on success, a negative errno value otherwise and rte_errno is set.
1359 mlx5_flow_validate_action_queue(const struct rte_flow_action *action,
1360 uint64_t action_flags,
1361 struct rte_eth_dev *dev,
1362 const struct rte_flow_attr *attr,
1363 struct rte_flow_error *error)
1365 struct mlx5_priv *priv = dev->data->dev_private;
1366 const struct rte_flow_action_queue *queue = action->conf;
1368 if (action_flags & MLX5_FLOW_FATE_ACTIONS)
1369 return rte_flow_error_set(error, EINVAL,
1370 RTE_FLOW_ERROR_TYPE_ACTION, NULL,
1371 "can't have 2 fate actions in"
1374 return rte_flow_error_set(error, EINVAL,
1375 RTE_FLOW_ERROR_TYPE_ACTION_CONF,
1376 NULL, "No Rx queues configured");
1377 if (queue->index >= priv->rxqs_n)
1378 return rte_flow_error_set(error, EINVAL,
1379 RTE_FLOW_ERROR_TYPE_ACTION_CONF,
1381 "queue index out of range");
1382 if (!(*priv->rxqs)[queue->index])
1383 return rte_flow_error_set(error, EINVAL,
1384 RTE_FLOW_ERROR_TYPE_ACTION_CONF,
1386 "queue is not configured");
1388 return rte_flow_error_set(error, ENOTSUP,
1389 RTE_FLOW_ERROR_TYPE_ATTR_EGRESS, NULL,
1390 "queue action not supported for "
1396 * Validate the rss action.
1399 * Pointer to the Ethernet device structure.
1401 * Pointer to the queue action.
1403 * Pointer to error structure.
1406 * 0 on success, a negative errno value otherwise and rte_errno is set.
1409 mlx5_validate_action_rss(struct rte_eth_dev *dev,
1410 const struct rte_flow_action *action,
1411 struct rte_flow_error *error)
1413 struct mlx5_priv *priv = dev->data->dev_private;
1414 const struct rte_flow_action_rss *rss = action->conf;
1415 enum mlx5_rxq_type rxq_type = MLX5_RXQ_TYPE_UNDEFINED;
1418 if (rss->func != RTE_ETH_HASH_FUNCTION_DEFAULT &&
1419 rss->func != RTE_ETH_HASH_FUNCTION_TOEPLITZ)
1420 return rte_flow_error_set(error, ENOTSUP,
1421 RTE_FLOW_ERROR_TYPE_ACTION_CONF,
1423 "RSS hash function not supported");
1424 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1429 return rte_flow_error_set(error, ENOTSUP,
1430 RTE_FLOW_ERROR_TYPE_ACTION_CONF,
1432 "tunnel RSS is not supported");
1433 /* allow RSS key_len 0 in case of NULL (default) RSS key. */
1434 if (rss->key_len == 0 && rss->key != NULL)
1435 return rte_flow_error_set(error, ENOTSUP,
1436 RTE_FLOW_ERROR_TYPE_ACTION_CONF,
1438 "RSS hash key length 0");
1439 if (rss->key_len > 0 && rss->key_len < MLX5_RSS_HASH_KEY_LEN)
1440 return rte_flow_error_set(error, ENOTSUP,
1441 RTE_FLOW_ERROR_TYPE_ACTION_CONF,
1443 "RSS hash key too small");
1444 if (rss->key_len > MLX5_RSS_HASH_KEY_LEN)
1445 return rte_flow_error_set(error, ENOTSUP,
1446 RTE_FLOW_ERROR_TYPE_ACTION_CONF,
1448 "RSS hash key too large");
1449 if (rss->queue_num > priv->config.ind_table_max_size)
1450 return rte_flow_error_set(error, ENOTSUP,
1451 RTE_FLOW_ERROR_TYPE_ACTION_CONF,
1453 "number of queues too large");
1454 if (rss->types & MLX5_RSS_HF_MASK)
1455 return rte_flow_error_set(error, ENOTSUP,
1456 RTE_FLOW_ERROR_TYPE_ACTION_CONF,
1458 "some RSS protocols are not"
1460 if ((rss->types & (ETH_RSS_L3_SRC_ONLY | ETH_RSS_L3_DST_ONLY)) &&
1461 !(rss->types & ETH_RSS_IP))
1462 return rte_flow_error_set(error, EINVAL,
1463 RTE_FLOW_ERROR_TYPE_ACTION_CONF, NULL,
1464 "L3 partial RSS requested but L3 RSS"
1465 " type not specified");
1466 if ((rss->types & (ETH_RSS_L4_SRC_ONLY | ETH_RSS_L4_DST_ONLY)) &&
1467 !(rss->types & (ETH_RSS_UDP | ETH_RSS_TCP)))
1468 return rte_flow_error_set(error, EINVAL,
1469 RTE_FLOW_ERROR_TYPE_ACTION_CONF, NULL,
1470 "L4 partial RSS requested but L4 RSS"
1471 " type not specified");
1473 return rte_flow_error_set(error, EINVAL,
1474 RTE_FLOW_ERROR_TYPE_ACTION_CONF,
1475 NULL, "No Rx queues configured");
1476 if (!rss->queue_num)
1477 return rte_flow_error_set(error, EINVAL,
1478 RTE_FLOW_ERROR_TYPE_ACTION_CONF,
1479 NULL, "No queues configured");
1480 for (i = 0; i != rss->queue_num; ++i) {
1481 struct mlx5_rxq_ctrl *rxq_ctrl;
1483 if (rss->queue[i] >= priv->rxqs_n)
1484 return rte_flow_error_set
1486 RTE_FLOW_ERROR_TYPE_ACTION_CONF,
1487 &rss->queue[i], "queue index out of range");
1488 if (!(*priv->rxqs)[rss->queue[i]])
1489 return rte_flow_error_set
1490 (error, EINVAL, RTE_FLOW_ERROR_TYPE_ACTION_CONF,
1491 &rss->queue[i], "queue is not configured");
1492 rxq_ctrl = container_of((*priv->rxqs)[rss->queue[i]],
1493 struct mlx5_rxq_ctrl, rxq);
1495 rxq_type = rxq_ctrl->type;
1496 if (rxq_type != rxq_ctrl->type)
1497 return rte_flow_error_set
1498 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION_CONF,
1500 "combining hairpin and regular RSS queues is not supported");
1506 * Validate the rss action.
1509 * Pointer to the queue action.
1510 * @param[in] action_flags
1511 * Bit-fields that holds the actions detected until now.
1513 * Pointer to the Ethernet device structure.
1515 * Attributes of flow that includes this action.
1516 * @param[in] item_flags
1517 * Items that were detected.
1519 * Pointer to error structure.
1522 * 0 on success, a negative errno value otherwise and rte_errno is set.
1525 mlx5_flow_validate_action_rss(const struct rte_flow_action *action,
1526 uint64_t action_flags,
1527 struct rte_eth_dev *dev,
1528 const struct rte_flow_attr *attr,
1529 uint64_t item_flags,
1530 struct rte_flow_error *error)
1532 const struct rte_flow_action_rss *rss = action->conf;
1533 int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
1536 if (action_flags & MLX5_FLOW_FATE_ACTIONS)
1537 return rte_flow_error_set(error, EINVAL,
1538 RTE_FLOW_ERROR_TYPE_ACTION, NULL,
1539 "can't have 2 fate actions"
1541 ret = mlx5_validate_action_rss(dev, action, error);
1545 return rte_flow_error_set(error, ENOTSUP,
1546 RTE_FLOW_ERROR_TYPE_ATTR_EGRESS, NULL,
1547 "rss action not supported for "
1549 if (rss->level > 1 && !tunnel)
1550 return rte_flow_error_set(error, EINVAL,
1551 RTE_FLOW_ERROR_TYPE_ACTION_CONF, NULL,
1552 "inner RSS is not supported for "
1553 "non-tunnel flows");
1554 if ((item_flags & MLX5_FLOW_LAYER_ECPRI) &&
1555 !(item_flags & MLX5_FLOW_LAYER_INNER_L4_UDP)) {
1556 return rte_flow_error_set(error, EINVAL,
1557 RTE_FLOW_ERROR_TYPE_ACTION_CONF, NULL,
1558 "RSS on eCPRI is not supported now");
1564 * Validate the default miss action.
1566 * @param[in] action_flags
1567 * Bit-fields that holds the actions detected until now.
1569 * Pointer to error structure.
1572 * 0 on success, a negative errno value otherwise and rte_errno is set.
1575 mlx5_flow_validate_action_default_miss(uint64_t action_flags,
1576 const struct rte_flow_attr *attr,
1577 struct rte_flow_error *error)
1579 if (action_flags & MLX5_FLOW_FATE_ACTIONS)
1580 return rte_flow_error_set(error, EINVAL,
1581 RTE_FLOW_ERROR_TYPE_ACTION, NULL,
1582 "can't have 2 fate actions in"
1585 return rte_flow_error_set(error, ENOTSUP,
1586 RTE_FLOW_ERROR_TYPE_ATTR_EGRESS, NULL,
1587 "default miss action not supported "
1590 return rte_flow_error_set(error, ENOTSUP,
1591 RTE_FLOW_ERROR_TYPE_ATTR_GROUP, NULL,
1592 "only group 0 is supported");
1594 return rte_flow_error_set(error, ENOTSUP,
1595 RTE_FLOW_ERROR_TYPE_ATTR_TRANSFER,
1596 NULL, "transfer is not supported");
1601 * Validate the count action.
1604 * Pointer to the Ethernet device structure.
1606 * Attributes of flow that includes this action.
1608 * Pointer to error structure.
1611 * 0 on success, a negative errno value otherwise and rte_errno is set.
1614 mlx5_flow_validate_action_count(struct rte_eth_dev *dev __rte_unused,
1615 const struct rte_flow_attr *attr,
1616 struct rte_flow_error *error)
1619 return rte_flow_error_set(error, ENOTSUP,
1620 RTE_FLOW_ERROR_TYPE_ATTR_EGRESS, NULL,
1621 "count action not supported for "
1627 * Verify the @p attributes will be correctly understood by the NIC and store
1628 * them in the @p flow if everything is correct.
1631 * Pointer to the Ethernet device structure.
1632 * @param[in] attributes
1633 * Pointer to flow attributes
1635 * Pointer to error structure.
1638 * 0 on success, a negative errno value otherwise and rte_errno is set.
1641 mlx5_flow_validate_attributes(struct rte_eth_dev *dev,
1642 const struct rte_flow_attr *attributes,
1643 struct rte_flow_error *error)
1645 struct mlx5_priv *priv = dev->data->dev_private;
1646 uint32_t priority_max = priv->config.flow_prio - 1;
1648 if (attributes->group)
1649 return rte_flow_error_set(error, ENOTSUP,
1650 RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
1651 NULL, "groups is not supported");
1652 if (attributes->priority != MLX5_FLOW_PRIO_RSVD &&
1653 attributes->priority >= priority_max)
1654 return rte_flow_error_set(error, ENOTSUP,
1655 RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
1656 NULL, "priority out of range");
1657 if (attributes->egress)
1658 return rte_flow_error_set(error, ENOTSUP,
1659 RTE_FLOW_ERROR_TYPE_ATTR_EGRESS, NULL,
1660 "egress is not supported");
1661 if (attributes->transfer && !priv->config.dv_esw_en)
1662 return rte_flow_error_set(error, ENOTSUP,
1663 RTE_FLOW_ERROR_TYPE_ATTR_TRANSFER,
1664 NULL, "transfer is not supported");
1665 if (!attributes->ingress)
1666 return rte_flow_error_set(error, EINVAL,
1667 RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
1669 "ingress attribute is mandatory");
1674 * Validate ICMP6 item.
1677 * Item specification.
1678 * @param[in] item_flags
1679 * Bit-fields that holds the items detected until now.
1680 * @param[in] ext_vlan_sup
1681 * Whether extended VLAN features are supported or not.
1683 * Pointer to error structure.
1686 * 0 on success, a negative errno value otherwise and rte_errno is set.
1689 mlx5_flow_validate_item_icmp6(const struct rte_flow_item *item,
1690 uint64_t item_flags,
1691 uint8_t target_protocol,
1692 struct rte_flow_error *error)
1694 const struct rte_flow_item_icmp6 *mask = item->mask;
1695 const int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
1696 const uint64_t l3m = tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV6 :
1697 MLX5_FLOW_LAYER_OUTER_L3_IPV6;
1698 const uint64_t l4m = tunnel ? MLX5_FLOW_LAYER_INNER_L4 :
1699 MLX5_FLOW_LAYER_OUTER_L4;
1702 if (target_protocol != 0xFF && target_protocol != IPPROTO_ICMPV6)
1703 return rte_flow_error_set(error, EINVAL,
1704 RTE_FLOW_ERROR_TYPE_ITEM, item,
1705 "protocol filtering not compatible"
1706 " with ICMP6 layer");
1707 if (!(item_flags & l3m))
1708 return rte_flow_error_set(error, EINVAL,
1709 RTE_FLOW_ERROR_TYPE_ITEM, item,
1710 "IPv6 is mandatory to filter on"
1712 if (item_flags & l4m)
1713 return rte_flow_error_set(error, EINVAL,
1714 RTE_FLOW_ERROR_TYPE_ITEM, item,
1715 "multiple L4 layers not supported");
1717 mask = &rte_flow_item_icmp6_mask;
1718 ret = mlx5_flow_item_acceptable
1719 (item, (const uint8_t *)mask,
1720 (const uint8_t *)&rte_flow_item_icmp6_mask,
1721 sizeof(struct rte_flow_item_icmp6),
1722 MLX5_ITEM_RANGE_NOT_ACCEPTED, error);
1729 * Validate ICMP item.
1732 * Item specification.
1733 * @param[in] item_flags
1734 * Bit-fields that holds the items detected until now.
1736 * Pointer to error structure.
1739 * 0 on success, a negative errno value otherwise and rte_errno is set.
1742 mlx5_flow_validate_item_icmp(const struct rte_flow_item *item,
1743 uint64_t item_flags,
1744 uint8_t target_protocol,
1745 struct rte_flow_error *error)
1747 const struct rte_flow_item_icmp *mask = item->mask;
1748 const struct rte_flow_item_icmp nic_mask = {
1749 .hdr.icmp_type = 0xff,
1750 .hdr.icmp_code = 0xff,
1751 .hdr.icmp_ident = RTE_BE16(0xffff),
1752 .hdr.icmp_seq_nb = RTE_BE16(0xffff),
1754 const int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
1755 const uint64_t l3m = tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV4 :
1756 MLX5_FLOW_LAYER_OUTER_L3_IPV4;
1757 const uint64_t l4m = tunnel ? MLX5_FLOW_LAYER_INNER_L4 :
1758 MLX5_FLOW_LAYER_OUTER_L4;
1761 if (target_protocol != 0xFF && target_protocol != IPPROTO_ICMP)
1762 return rte_flow_error_set(error, EINVAL,
1763 RTE_FLOW_ERROR_TYPE_ITEM, item,
1764 "protocol filtering not compatible"
1765 " with ICMP layer");
1766 if (!(item_flags & l3m))
1767 return rte_flow_error_set(error, EINVAL,
1768 RTE_FLOW_ERROR_TYPE_ITEM, item,
1769 "IPv4 is mandatory to filter"
1771 if (item_flags & l4m)
1772 return rte_flow_error_set(error, EINVAL,
1773 RTE_FLOW_ERROR_TYPE_ITEM, item,
1774 "multiple L4 layers not supported");
1777 ret = mlx5_flow_item_acceptable
1778 (item, (const uint8_t *)mask,
1779 (const uint8_t *)&nic_mask,
1780 sizeof(struct rte_flow_item_icmp),
1781 MLX5_ITEM_RANGE_NOT_ACCEPTED, error);
1788 * Validate Ethernet item.
1791 * Item specification.
1792 * @param[in] item_flags
1793 * Bit-fields that holds the items detected until now.
1795 * Pointer to error structure.
1798 * 0 on success, a negative errno value otherwise and rte_errno is set.
1801 mlx5_flow_validate_item_eth(const struct rte_flow_item *item,
1802 uint64_t item_flags, bool ext_vlan_sup,
1803 struct rte_flow_error *error)
1805 const struct rte_flow_item_eth *mask = item->mask;
1806 const struct rte_flow_item_eth nic_mask = {
1807 .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
1808 .src.addr_bytes = "\xff\xff\xff\xff\xff\xff",
1809 .type = RTE_BE16(0xffff),
1810 .has_vlan = ext_vlan_sup ? 1 : 0,
1813 int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
1814 const uint64_t ethm = tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
1815 MLX5_FLOW_LAYER_OUTER_L2;
1817 if (item_flags & ethm)
1818 return rte_flow_error_set(error, ENOTSUP,
1819 RTE_FLOW_ERROR_TYPE_ITEM, item,
1820 "multiple L2 layers not supported");
1821 if ((!tunnel && (item_flags & MLX5_FLOW_LAYER_OUTER_L3)) ||
1822 (tunnel && (item_flags & MLX5_FLOW_LAYER_INNER_L3)))
1823 return rte_flow_error_set(error, EINVAL,
1824 RTE_FLOW_ERROR_TYPE_ITEM, item,
1825 "L2 layer should not follow "
1827 if ((!tunnel && (item_flags & MLX5_FLOW_LAYER_OUTER_VLAN)) ||
1828 (tunnel && (item_flags & MLX5_FLOW_LAYER_INNER_VLAN)))
1829 return rte_flow_error_set(error, EINVAL,
1830 RTE_FLOW_ERROR_TYPE_ITEM, item,
1831 "L2 layer should not follow VLAN");
1833 mask = &rte_flow_item_eth_mask;
1834 ret = mlx5_flow_item_acceptable(item, (const uint8_t *)mask,
1835 (const uint8_t *)&nic_mask,
1836 sizeof(struct rte_flow_item_eth),
1837 MLX5_ITEM_RANGE_NOT_ACCEPTED, error);
1842 * Validate VLAN item.
1845 * Item specification.
1846 * @param[in] item_flags
1847 * Bit-fields that holds the items detected until now.
1849 * Ethernet device flow is being created on.
1851 * Pointer to error structure.
1854 * 0 on success, a negative errno value otherwise and rte_errno is set.
1857 mlx5_flow_validate_item_vlan(const struct rte_flow_item *item,
1858 uint64_t item_flags,
1859 struct rte_eth_dev *dev,
1860 struct rte_flow_error *error)
1862 const struct rte_flow_item_vlan *spec = item->spec;
1863 const struct rte_flow_item_vlan *mask = item->mask;
1864 const struct rte_flow_item_vlan nic_mask = {
1865 .tci = RTE_BE16(UINT16_MAX),
1866 .inner_type = RTE_BE16(UINT16_MAX),
1868 uint16_t vlan_tag = 0;
1869 const int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
1871 const uint64_t l34m = tunnel ? (MLX5_FLOW_LAYER_INNER_L3 |
1872 MLX5_FLOW_LAYER_INNER_L4) :
1873 (MLX5_FLOW_LAYER_OUTER_L3 |
1874 MLX5_FLOW_LAYER_OUTER_L4);
1875 const uint64_t vlanm = tunnel ? MLX5_FLOW_LAYER_INNER_VLAN :
1876 MLX5_FLOW_LAYER_OUTER_VLAN;
1878 if (item_flags & vlanm)
1879 return rte_flow_error_set(error, EINVAL,
1880 RTE_FLOW_ERROR_TYPE_ITEM, item,
1881 "multiple VLAN layers not supported");
1882 else if ((item_flags & l34m) != 0)
1883 return rte_flow_error_set(error, EINVAL,
1884 RTE_FLOW_ERROR_TYPE_ITEM, item,
1885 "VLAN cannot follow L3/L4 layer");
1887 mask = &rte_flow_item_vlan_mask;
1888 ret = mlx5_flow_item_acceptable(item, (const uint8_t *)mask,
1889 (const uint8_t *)&nic_mask,
1890 sizeof(struct rte_flow_item_vlan),
1891 MLX5_ITEM_RANGE_NOT_ACCEPTED, error);
1894 if (!tunnel && mask->tci != RTE_BE16(0x0fff)) {
1895 struct mlx5_priv *priv = dev->data->dev_private;
1897 if (priv->vmwa_context) {
1899 * Non-NULL context means we have a virtual machine
1900 * and SR-IOV enabled, we have to create VLAN interface
1901 * to make hypervisor to setup E-Switch vport
1902 * context correctly. We avoid creating the multiple
1903 * VLAN interfaces, so we cannot support VLAN tag mask.
1905 return rte_flow_error_set(error, EINVAL,
1906 RTE_FLOW_ERROR_TYPE_ITEM,
1908 "VLAN tag mask is not"
1909 " supported in virtual"
1914 vlan_tag = spec->tci;
1915 vlan_tag &= mask->tci;
1918 * From verbs perspective an empty VLAN is equivalent
1919 * to a packet without VLAN layer.
1922 return rte_flow_error_set(error, EINVAL,
1923 RTE_FLOW_ERROR_TYPE_ITEM_SPEC,
1925 "VLAN cannot be empty");
1930 * Validate IPV4 item.
1933 * Item specification.
1934 * @param[in] item_flags
1935 * Bit-fields that holds the items detected until now.
1936 * @param[in] last_item
1937 * Previous validated item in the pattern items.
1938 * @param[in] ether_type
1939 * Type in the ethernet layer header (including dot1q).
1940 * @param[in] acc_mask
1941 * Acceptable mask, if NULL default internal default mask
1942 * will be used to check whether item fields are supported.
1943 * @param[in] range_accepted
1944 * True if range of values is accepted for specific fields, false otherwise.
1946 * Pointer to error structure.
1949 * 0 on success, a negative errno value otherwise and rte_errno is set.
1952 mlx5_flow_validate_item_ipv4(const struct rte_flow_item *item,
1953 uint64_t item_flags,
1955 uint16_t ether_type,
1956 const struct rte_flow_item_ipv4 *acc_mask,
1957 bool range_accepted,
1958 struct rte_flow_error *error)
1960 const struct rte_flow_item_ipv4 *mask = item->mask;
1961 const struct rte_flow_item_ipv4 *spec = item->spec;
1962 const struct rte_flow_item_ipv4 nic_mask = {
1964 .src_addr = RTE_BE32(0xffffffff),
1965 .dst_addr = RTE_BE32(0xffffffff),
1966 .type_of_service = 0xff,
1967 .next_proto_id = 0xff,
1970 const int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
1971 const uint64_t l3m = tunnel ? MLX5_FLOW_LAYER_INNER_L3 :
1972 MLX5_FLOW_LAYER_OUTER_L3;
1973 const uint64_t l4m = tunnel ? MLX5_FLOW_LAYER_INNER_L4 :
1974 MLX5_FLOW_LAYER_OUTER_L4;
1976 uint8_t next_proto = 0xFF;
1977 const uint64_t l2_vlan = (MLX5_FLOW_LAYER_L2 |
1978 MLX5_FLOW_LAYER_OUTER_VLAN |
1979 MLX5_FLOW_LAYER_INNER_VLAN);
1981 if ((last_item & l2_vlan) && ether_type &&
1982 ether_type != RTE_ETHER_TYPE_IPV4)
1983 return rte_flow_error_set(error, EINVAL,
1984 RTE_FLOW_ERROR_TYPE_ITEM, item,
1985 "IPv4 cannot follow L2/VLAN layer "
1986 "which ether type is not IPv4");
1987 if (item_flags & MLX5_FLOW_LAYER_IPIP) {
1989 next_proto = mask->hdr.next_proto_id &
1990 spec->hdr.next_proto_id;
1991 if (next_proto == IPPROTO_IPIP || next_proto == IPPROTO_IPV6)
1992 return rte_flow_error_set(error, EINVAL,
1993 RTE_FLOW_ERROR_TYPE_ITEM,
1998 if (item_flags & MLX5_FLOW_LAYER_IPV6_ENCAP)
1999 return rte_flow_error_set(error, EINVAL,
2000 RTE_FLOW_ERROR_TYPE_ITEM, item,
2001 "wrong tunnel type - IPv6 specified "
2002 "but IPv4 item provided");
2003 if (item_flags & l3m)
2004 return rte_flow_error_set(error, ENOTSUP,
2005 RTE_FLOW_ERROR_TYPE_ITEM, item,
2006 "multiple L3 layers not supported");
2007 else if (item_flags & l4m)
2008 return rte_flow_error_set(error, EINVAL,
2009 RTE_FLOW_ERROR_TYPE_ITEM, item,
2010 "L3 cannot follow an L4 layer.");
2011 else if ((item_flags & MLX5_FLOW_LAYER_NVGRE) &&
2012 !(item_flags & MLX5_FLOW_LAYER_INNER_L2))
2013 return rte_flow_error_set(error, EINVAL,
2014 RTE_FLOW_ERROR_TYPE_ITEM, item,
2015 "L3 cannot follow an NVGRE layer.");
2017 mask = &rte_flow_item_ipv4_mask;
2018 else if (mask->hdr.next_proto_id != 0 &&
2019 mask->hdr.next_proto_id != 0xff)
2020 return rte_flow_error_set(error, EINVAL,
2021 RTE_FLOW_ERROR_TYPE_ITEM_MASK, mask,
2022 "partial mask is not supported"
2024 ret = mlx5_flow_item_acceptable(item, (const uint8_t *)mask,
2025 acc_mask ? (const uint8_t *)acc_mask
2026 : (const uint8_t *)&nic_mask,
2027 sizeof(struct rte_flow_item_ipv4),
2028 range_accepted, error);
2035 * Validate IPV6 item.
2038 * Item specification.
2039 * @param[in] item_flags
2040 * Bit-fields that holds the items detected until now.
2041 * @param[in] last_item
2042 * Previous validated item in the pattern items.
2043 * @param[in] ether_type
2044 * Type in the ethernet layer header (including dot1q).
2045 * @param[in] acc_mask
2046 * Acceptable mask, if NULL default internal default mask
2047 * will be used to check whether item fields are supported.
2049 * Pointer to error structure.
2052 * 0 on success, a negative errno value otherwise and rte_errno is set.
2055 mlx5_flow_validate_item_ipv6(const struct rte_flow_item *item,
2056 uint64_t item_flags,
2058 uint16_t ether_type,
2059 const struct rte_flow_item_ipv6 *acc_mask,
2060 struct rte_flow_error *error)
2062 const struct rte_flow_item_ipv6 *mask = item->mask;
2063 const struct rte_flow_item_ipv6 *spec = item->spec;
2064 const struct rte_flow_item_ipv6 nic_mask = {
2067 "\xff\xff\xff\xff\xff\xff\xff\xff"
2068 "\xff\xff\xff\xff\xff\xff\xff\xff",
2070 "\xff\xff\xff\xff\xff\xff\xff\xff"
2071 "\xff\xff\xff\xff\xff\xff\xff\xff",
2072 .vtc_flow = RTE_BE32(0xffffffff),
2076 const int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
2077 const uint64_t l3m = tunnel ? MLX5_FLOW_LAYER_INNER_L3 :
2078 MLX5_FLOW_LAYER_OUTER_L3;
2079 const uint64_t l4m = tunnel ? MLX5_FLOW_LAYER_INNER_L4 :
2080 MLX5_FLOW_LAYER_OUTER_L4;
2082 uint8_t next_proto = 0xFF;
2083 const uint64_t l2_vlan = (MLX5_FLOW_LAYER_L2 |
2084 MLX5_FLOW_LAYER_OUTER_VLAN |
2085 MLX5_FLOW_LAYER_INNER_VLAN);
2087 if ((last_item & l2_vlan) && ether_type &&
2088 ether_type != RTE_ETHER_TYPE_IPV6)
2089 return rte_flow_error_set(error, EINVAL,
2090 RTE_FLOW_ERROR_TYPE_ITEM, item,
2091 "IPv6 cannot follow L2/VLAN layer "
2092 "which ether type is not IPv6");
2093 if (mask && mask->hdr.proto == UINT8_MAX && spec)
2094 next_proto = spec->hdr.proto;
2095 if (item_flags & MLX5_FLOW_LAYER_IPV6_ENCAP) {
2096 if (next_proto == IPPROTO_IPIP || next_proto == IPPROTO_IPV6)
2097 return rte_flow_error_set(error, EINVAL,
2098 RTE_FLOW_ERROR_TYPE_ITEM,
2103 if (next_proto == IPPROTO_HOPOPTS ||
2104 next_proto == IPPROTO_ROUTING ||
2105 next_proto == IPPROTO_FRAGMENT ||
2106 next_proto == IPPROTO_ESP ||
2107 next_proto == IPPROTO_AH ||
2108 next_proto == IPPROTO_DSTOPTS)
2109 return rte_flow_error_set(error, EINVAL,
2110 RTE_FLOW_ERROR_TYPE_ITEM, item,
2111 "IPv6 proto (next header) should "
2112 "not be set as extension header");
2113 if (item_flags & MLX5_FLOW_LAYER_IPIP)
2114 return rte_flow_error_set(error, EINVAL,
2115 RTE_FLOW_ERROR_TYPE_ITEM, item,
2116 "wrong tunnel type - IPv4 specified "
2117 "but IPv6 item provided");
2118 if (item_flags & l3m)
2119 return rte_flow_error_set(error, ENOTSUP,
2120 RTE_FLOW_ERROR_TYPE_ITEM, item,
2121 "multiple L3 layers not supported");
2122 else if (item_flags & l4m)
2123 return rte_flow_error_set(error, EINVAL,
2124 RTE_FLOW_ERROR_TYPE_ITEM, item,
2125 "L3 cannot follow an L4 layer.");
2126 else if ((item_flags & MLX5_FLOW_LAYER_NVGRE) &&
2127 !(item_flags & MLX5_FLOW_LAYER_INNER_L2))
2128 return rte_flow_error_set(error, EINVAL,
2129 RTE_FLOW_ERROR_TYPE_ITEM, item,
2130 "L3 cannot follow an NVGRE layer.");
2132 mask = &rte_flow_item_ipv6_mask;
2133 ret = mlx5_flow_item_acceptable(item, (const uint8_t *)mask,
2134 acc_mask ? (const uint8_t *)acc_mask
2135 : (const uint8_t *)&nic_mask,
2136 sizeof(struct rte_flow_item_ipv6),
2137 MLX5_ITEM_RANGE_NOT_ACCEPTED, error);
2144 * Validate UDP item.
2147 * Item specification.
2148 * @param[in] item_flags
2149 * Bit-fields that holds the items detected until now.
2150 * @param[in] target_protocol
2151 * The next protocol in the previous item.
2152 * @param[in] flow_mask
2153 * mlx5 flow-specific (DV, verbs, etc.) supported header fields mask.
2155 * Pointer to error structure.
2158 * 0 on success, a negative errno value otherwise and rte_errno is set.
2161 mlx5_flow_validate_item_udp(const struct rte_flow_item *item,
2162 uint64_t item_flags,
2163 uint8_t target_protocol,
2164 struct rte_flow_error *error)
2166 const struct rte_flow_item_udp *mask = item->mask;
2167 const int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
2168 const uint64_t l3m = tunnel ? MLX5_FLOW_LAYER_INNER_L3 :
2169 MLX5_FLOW_LAYER_OUTER_L3;
2170 const uint64_t l4m = tunnel ? MLX5_FLOW_LAYER_INNER_L4 :
2171 MLX5_FLOW_LAYER_OUTER_L4;
2174 if (target_protocol != 0xff && target_protocol != IPPROTO_UDP)
2175 return rte_flow_error_set(error, EINVAL,
2176 RTE_FLOW_ERROR_TYPE_ITEM, item,
2177 "protocol filtering not compatible"
2179 if (!(item_flags & l3m))
2180 return rte_flow_error_set(error, EINVAL,
2181 RTE_FLOW_ERROR_TYPE_ITEM, item,
2182 "L3 is mandatory to filter on L4");
2183 if (item_flags & l4m)
2184 return rte_flow_error_set(error, EINVAL,
2185 RTE_FLOW_ERROR_TYPE_ITEM, item,
2186 "multiple L4 layers not supported");
2188 mask = &rte_flow_item_udp_mask;
2189 ret = mlx5_flow_item_acceptable
2190 (item, (const uint8_t *)mask,
2191 (const uint8_t *)&rte_flow_item_udp_mask,
2192 sizeof(struct rte_flow_item_udp), MLX5_ITEM_RANGE_NOT_ACCEPTED,
2200 * Validate TCP item.
2203 * Item specification.
2204 * @param[in] item_flags
2205 * Bit-fields that holds the items detected until now.
2206 * @param[in] target_protocol
2207 * The next protocol in the previous item.
2209 * Pointer to error structure.
2212 * 0 on success, a negative errno value otherwise and rte_errno is set.
2215 mlx5_flow_validate_item_tcp(const struct rte_flow_item *item,
2216 uint64_t item_flags,
2217 uint8_t target_protocol,
2218 const struct rte_flow_item_tcp *flow_mask,
2219 struct rte_flow_error *error)
2221 const struct rte_flow_item_tcp *mask = item->mask;
2222 const int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
2223 const uint64_t l3m = tunnel ? MLX5_FLOW_LAYER_INNER_L3 :
2224 MLX5_FLOW_LAYER_OUTER_L3;
2225 const uint64_t l4m = tunnel ? MLX5_FLOW_LAYER_INNER_L4 :
2226 MLX5_FLOW_LAYER_OUTER_L4;
2229 MLX5_ASSERT(flow_mask);
2230 if (target_protocol != 0xff && target_protocol != IPPROTO_TCP)
2231 return rte_flow_error_set(error, EINVAL,
2232 RTE_FLOW_ERROR_TYPE_ITEM, item,
2233 "protocol filtering not compatible"
2235 if (!(item_flags & l3m))
2236 return rte_flow_error_set(error, EINVAL,
2237 RTE_FLOW_ERROR_TYPE_ITEM, item,
2238 "L3 is mandatory to filter on L4");
2239 if (item_flags & l4m)
2240 return rte_flow_error_set(error, EINVAL,
2241 RTE_FLOW_ERROR_TYPE_ITEM, item,
2242 "multiple L4 layers not supported");
2244 mask = &rte_flow_item_tcp_mask;
2245 ret = mlx5_flow_item_acceptable
2246 (item, (const uint8_t *)mask,
2247 (const uint8_t *)flow_mask,
2248 sizeof(struct rte_flow_item_tcp), MLX5_ITEM_RANGE_NOT_ACCEPTED,
2256 * Validate VXLAN item.
2259 * Item specification.
2260 * @param[in] item_flags
2261 * Bit-fields that holds the items detected until now.
2262 * @param[in] target_protocol
2263 * The next protocol in the previous item.
2265 * Pointer to error structure.
2268 * 0 on success, a negative errno value otherwise and rte_errno is set.
2271 mlx5_flow_validate_item_vxlan(const struct rte_flow_item *item,
2272 uint64_t item_flags,
2273 struct rte_flow_error *error)
2275 const struct rte_flow_item_vxlan *spec = item->spec;
2276 const struct rte_flow_item_vxlan *mask = item->mask;
2281 } id = { .vlan_id = 0, };
2284 if (item_flags & MLX5_FLOW_LAYER_TUNNEL)
2285 return rte_flow_error_set(error, ENOTSUP,
2286 RTE_FLOW_ERROR_TYPE_ITEM, item,
2287 "multiple tunnel layers not"
2290 * Verify only UDPv4 is present as defined in
2291 * https://tools.ietf.org/html/rfc7348
2293 if (!(item_flags & MLX5_FLOW_LAYER_OUTER_L4_UDP))
2294 return rte_flow_error_set(error, EINVAL,
2295 RTE_FLOW_ERROR_TYPE_ITEM, item,
2296 "no outer UDP layer found");
2298 mask = &rte_flow_item_vxlan_mask;
2299 ret = mlx5_flow_item_acceptable
2300 (item, (const uint8_t *)mask,
2301 (const uint8_t *)&rte_flow_item_vxlan_mask,
2302 sizeof(struct rte_flow_item_vxlan),
2303 MLX5_ITEM_RANGE_NOT_ACCEPTED, error);
2307 memcpy(&id.vni[1], spec->vni, 3);
2308 memcpy(&id.vni[1], mask->vni, 3);
2310 if (!(item_flags & MLX5_FLOW_LAYER_OUTER))
2311 return rte_flow_error_set(error, ENOTSUP,
2312 RTE_FLOW_ERROR_TYPE_ITEM, item,
2313 "VXLAN tunnel must be fully defined");
2318 * Validate VXLAN_GPE item.
2321 * Item specification.
2322 * @param[in] item_flags
2323 * Bit-fields that holds the items detected until now.
2325 * Pointer to the private data structure.
2326 * @param[in] target_protocol
2327 * The next protocol in the previous item.
2329 * Pointer to error structure.
2332 * 0 on success, a negative errno value otherwise and rte_errno is set.
2335 mlx5_flow_validate_item_vxlan_gpe(const struct rte_flow_item *item,
2336 uint64_t item_flags,
2337 struct rte_eth_dev *dev,
2338 struct rte_flow_error *error)
2340 struct mlx5_priv *priv = dev->data->dev_private;
2341 const struct rte_flow_item_vxlan_gpe *spec = item->spec;
2342 const struct rte_flow_item_vxlan_gpe *mask = item->mask;
2347 } id = { .vlan_id = 0, };
2349 if (!priv->config.l3_vxlan_en)
2350 return rte_flow_error_set(error, ENOTSUP,
2351 RTE_FLOW_ERROR_TYPE_ITEM, item,
2352 "L3 VXLAN is not enabled by device"
2353 " parameter and/or not configured in"
2355 if (item_flags & MLX5_FLOW_LAYER_TUNNEL)
2356 return rte_flow_error_set(error, ENOTSUP,
2357 RTE_FLOW_ERROR_TYPE_ITEM, item,
2358 "multiple tunnel layers not"
2361 * Verify only UDPv4 is present as defined in
2362 * https://tools.ietf.org/html/rfc7348
2364 if (!(item_flags & MLX5_FLOW_LAYER_OUTER_L4_UDP))
2365 return rte_flow_error_set(error, EINVAL,
2366 RTE_FLOW_ERROR_TYPE_ITEM, item,
2367 "no outer UDP layer found");
2369 mask = &rte_flow_item_vxlan_gpe_mask;
2370 ret = mlx5_flow_item_acceptable
2371 (item, (const uint8_t *)mask,
2372 (const uint8_t *)&rte_flow_item_vxlan_gpe_mask,
2373 sizeof(struct rte_flow_item_vxlan_gpe),
2374 MLX5_ITEM_RANGE_NOT_ACCEPTED, error);
2379 return rte_flow_error_set(error, ENOTSUP,
2380 RTE_FLOW_ERROR_TYPE_ITEM,
2382 "VxLAN-GPE protocol"
2384 memcpy(&id.vni[1], spec->vni, 3);
2385 memcpy(&id.vni[1], mask->vni, 3);
2387 if (!(item_flags & MLX5_FLOW_LAYER_OUTER))
2388 return rte_flow_error_set(error, ENOTSUP,
2389 RTE_FLOW_ERROR_TYPE_ITEM, item,
2390 "VXLAN-GPE tunnel must be fully"
2395 * Validate GRE Key item.
2398 * Item specification.
2399 * @param[in] item_flags
2400 * Bit flags to mark detected items.
2401 * @param[in] gre_item
2402 * Pointer to gre_item
2404 * Pointer to error structure.
2407 * 0 on success, a negative errno value otherwise and rte_errno is set.
2410 mlx5_flow_validate_item_gre_key(const struct rte_flow_item *item,
2411 uint64_t item_flags,
2412 const struct rte_flow_item *gre_item,
2413 struct rte_flow_error *error)
2415 const rte_be32_t *mask = item->mask;
2417 rte_be32_t gre_key_default_mask = RTE_BE32(UINT32_MAX);
2418 const struct rte_flow_item_gre *gre_spec;
2419 const struct rte_flow_item_gre *gre_mask;
2421 if (item_flags & MLX5_FLOW_LAYER_GRE_KEY)
2422 return rte_flow_error_set(error, ENOTSUP,
2423 RTE_FLOW_ERROR_TYPE_ITEM, item,
2424 "Multiple GRE key not support");
2425 if (!(item_flags & MLX5_FLOW_LAYER_GRE))
2426 return rte_flow_error_set(error, ENOTSUP,
2427 RTE_FLOW_ERROR_TYPE_ITEM, item,
2428 "No preceding GRE header");
2429 if (item_flags & MLX5_FLOW_LAYER_INNER)
2430 return rte_flow_error_set(error, ENOTSUP,
2431 RTE_FLOW_ERROR_TYPE_ITEM, item,
2432 "GRE key following a wrong item");
2433 gre_mask = gre_item->mask;
2435 gre_mask = &rte_flow_item_gre_mask;
2436 gre_spec = gre_item->spec;
2437 if (gre_spec && (gre_mask->c_rsvd0_ver & RTE_BE16(0x2000)) &&
2438 !(gre_spec->c_rsvd0_ver & RTE_BE16(0x2000)))
2439 return rte_flow_error_set(error, EINVAL,
2440 RTE_FLOW_ERROR_TYPE_ITEM, item,
2441 "Key bit must be on");
2444 mask = &gre_key_default_mask;
2445 ret = mlx5_flow_item_acceptable
2446 (item, (const uint8_t *)mask,
2447 (const uint8_t *)&gre_key_default_mask,
2448 sizeof(rte_be32_t), MLX5_ITEM_RANGE_NOT_ACCEPTED, error);
2453 * Validate GRE item.
2456 * Item specification.
2457 * @param[in] item_flags
2458 * Bit flags to mark detected items.
2459 * @param[in] target_protocol
2460 * The next protocol in the previous item.
2462 * Pointer to error structure.
2465 * 0 on success, a negative errno value otherwise and rte_errno is set.
2468 mlx5_flow_validate_item_gre(const struct rte_flow_item *item,
2469 uint64_t item_flags,
2470 uint8_t target_protocol,
2471 struct rte_flow_error *error)
2473 const struct rte_flow_item_gre *spec __rte_unused = item->spec;
2474 const struct rte_flow_item_gre *mask = item->mask;
2476 const struct rte_flow_item_gre nic_mask = {
2477 .c_rsvd0_ver = RTE_BE16(0xB000),
2478 .protocol = RTE_BE16(UINT16_MAX),
2481 if (target_protocol != 0xff && target_protocol != IPPROTO_GRE)
2482 return rte_flow_error_set(error, EINVAL,
2483 RTE_FLOW_ERROR_TYPE_ITEM, item,
2484 "protocol filtering not compatible"
2485 " with this GRE layer");
2486 if (item_flags & MLX5_FLOW_LAYER_TUNNEL)
2487 return rte_flow_error_set(error, ENOTSUP,
2488 RTE_FLOW_ERROR_TYPE_ITEM, item,
2489 "multiple tunnel layers not"
2491 if (!(item_flags & MLX5_FLOW_LAYER_OUTER_L3))
2492 return rte_flow_error_set(error, ENOTSUP,
2493 RTE_FLOW_ERROR_TYPE_ITEM, item,
2494 "L3 Layer is missing");
2496 mask = &rte_flow_item_gre_mask;
2497 ret = mlx5_flow_item_acceptable
2498 (item, (const uint8_t *)mask,
2499 (const uint8_t *)&nic_mask,
2500 sizeof(struct rte_flow_item_gre), MLX5_ITEM_RANGE_NOT_ACCEPTED,
2504 #ifndef HAVE_MLX5DV_DR
2505 #ifndef HAVE_IBV_DEVICE_MPLS_SUPPORT
2506 if (spec && (spec->protocol & mask->protocol))
2507 return rte_flow_error_set(error, ENOTSUP,
2508 RTE_FLOW_ERROR_TYPE_ITEM, item,
2509 "without MPLS support the"
2510 " specification cannot be used for"
2518 * Validate Geneve item.
2521 * Item specification.
2522 * @param[in] itemFlags
2523 * Bit-fields that holds the items detected until now.
2525 * Pointer to the private data structure.
2527 * Pointer to error structure.
2530 * 0 on success, a negative errno value otherwise and rte_errno is set.
2534 mlx5_flow_validate_item_geneve(const struct rte_flow_item *item,
2535 uint64_t item_flags,
2536 struct rte_eth_dev *dev,
2537 struct rte_flow_error *error)
2539 struct mlx5_priv *priv = dev->data->dev_private;
2540 const struct rte_flow_item_geneve *spec = item->spec;
2541 const struct rte_flow_item_geneve *mask = item->mask;
2544 uint8_t opt_len = priv->config.hca_attr.geneve_max_opt_len ?
2545 MLX5_GENEVE_OPT_LEN_1 : MLX5_GENEVE_OPT_LEN_0;
2546 const struct rte_flow_item_geneve nic_mask = {
2547 .ver_opt_len_o_c_rsvd0 = RTE_BE16(0x3f80),
2548 .vni = "\xff\xff\xff",
2549 .protocol = RTE_BE16(UINT16_MAX),
2552 if (!priv->config.hca_attr.tunnel_stateless_geneve_rx)
2553 return rte_flow_error_set(error, ENOTSUP,
2554 RTE_FLOW_ERROR_TYPE_ITEM, item,
2555 "L3 Geneve is not enabled by device"
2556 " parameter and/or not configured in"
2558 if (item_flags & MLX5_FLOW_LAYER_TUNNEL)
2559 return rte_flow_error_set(error, ENOTSUP,
2560 RTE_FLOW_ERROR_TYPE_ITEM, item,
2561 "multiple tunnel layers not"
2564 * Verify only UDPv4 is present as defined in
2565 * https://tools.ietf.org/html/rfc7348
2567 if (!(item_flags & MLX5_FLOW_LAYER_OUTER_L4_UDP))
2568 return rte_flow_error_set(error, EINVAL,
2569 RTE_FLOW_ERROR_TYPE_ITEM, item,
2570 "no outer UDP layer found");
2572 mask = &rte_flow_item_geneve_mask;
2573 ret = mlx5_flow_item_acceptable
2574 (item, (const uint8_t *)mask,
2575 (const uint8_t *)&nic_mask,
2576 sizeof(struct rte_flow_item_geneve),
2577 MLX5_ITEM_RANGE_NOT_ACCEPTED, error);
2581 gbhdr = rte_be_to_cpu_16(spec->ver_opt_len_o_c_rsvd0);
2582 if (MLX5_GENEVE_VER_VAL(gbhdr) ||
2583 MLX5_GENEVE_CRITO_VAL(gbhdr) ||
2584 MLX5_GENEVE_RSVD_VAL(gbhdr) || spec->rsvd1)
2585 return rte_flow_error_set(error, ENOTSUP,
2586 RTE_FLOW_ERROR_TYPE_ITEM,
2588 "Geneve protocol unsupported"
2589 " fields are being used");
2590 if (MLX5_GENEVE_OPTLEN_VAL(gbhdr) > opt_len)
2591 return rte_flow_error_set
2593 RTE_FLOW_ERROR_TYPE_ITEM,
2595 "Unsupported Geneve options length");
2597 if (!(item_flags & MLX5_FLOW_LAYER_OUTER))
2598 return rte_flow_error_set
2600 RTE_FLOW_ERROR_TYPE_ITEM, item,
2601 "Geneve tunnel must be fully defined");
2606 * Validate MPLS item.
2609 * Pointer to the rte_eth_dev structure.
2611 * Item specification.
2612 * @param[in] item_flags
2613 * Bit-fields that holds the items detected until now.
2614 * @param[in] prev_layer
2615 * The protocol layer indicated in previous item.
2617 * Pointer to error structure.
2620 * 0 on success, a negative errno value otherwise and rte_errno is set.
2623 mlx5_flow_validate_item_mpls(struct rte_eth_dev *dev __rte_unused,
2624 const struct rte_flow_item *item __rte_unused,
2625 uint64_t item_flags __rte_unused,
2626 uint64_t prev_layer __rte_unused,
2627 struct rte_flow_error *error)
2629 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
2630 const struct rte_flow_item_mpls *mask = item->mask;
2631 struct mlx5_priv *priv = dev->data->dev_private;
2634 if (!priv->config.mpls_en)
2635 return rte_flow_error_set(error, ENOTSUP,
2636 RTE_FLOW_ERROR_TYPE_ITEM, item,
2637 "MPLS not supported or"
2638 " disabled in firmware"
2640 /* MPLS over IP, UDP, GRE is allowed */
2641 if (!(prev_layer & (MLX5_FLOW_LAYER_OUTER_L3 |
2642 MLX5_FLOW_LAYER_OUTER_L4_UDP |
2643 MLX5_FLOW_LAYER_GRE |
2644 MLX5_FLOW_LAYER_GRE_KEY)))
2645 return rte_flow_error_set(error, EINVAL,
2646 RTE_FLOW_ERROR_TYPE_ITEM, item,
2647 "protocol filtering not compatible"
2648 " with MPLS layer");
2649 /* Multi-tunnel isn't allowed but MPLS over GRE is an exception. */
2650 if ((item_flags & MLX5_FLOW_LAYER_TUNNEL) &&
2651 !(item_flags & MLX5_FLOW_LAYER_GRE))
2652 return rte_flow_error_set(error, ENOTSUP,
2653 RTE_FLOW_ERROR_TYPE_ITEM, item,
2654 "multiple tunnel layers not"
2657 mask = &rte_flow_item_mpls_mask;
2658 ret = mlx5_flow_item_acceptable
2659 (item, (const uint8_t *)mask,
2660 (const uint8_t *)&rte_flow_item_mpls_mask,
2661 sizeof(struct rte_flow_item_mpls),
2662 MLX5_ITEM_RANGE_NOT_ACCEPTED, error);
2667 return rte_flow_error_set(error, ENOTSUP,
2668 RTE_FLOW_ERROR_TYPE_ITEM, item,
2669 "MPLS is not supported by Verbs, please"
2675 * Validate NVGRE item.
2678 * Item specification.
2679 * @param[in] item_flags
2680 * Bit flags to mark detected items.
2681 * @param[in] target_protocol
2682 * The next protocol in the previous item.
2684 * Pointer to error structure.
2687 * 0 on success, a negative errno value otherwise and rte_errno is set.
2690 mlx5_flow_validate_item_nvgre(const struct rte_flow_item *item,
2691 uint64_t item_flags,
2692 uint8_t target_protocol,
2693 struct rte_flow_error *error)
2695 const struct rte_flow_item_nvgre *mask = item->mask;
2698 if (target_protocol != 0xff && target_protocol != IPPROTO_GRE)
2699 return rte_flow_error_set(error, EINVAL,
2700 RTE_FLOW_ERROR_TYPE_ITEM, item,
2701 "protocol filtering not compatible"
2702 " with this GRE layer");
2703 if (item_flags & MLX5_FLOW_LAYER_TUNNEL)
2704 return rte_flow_error_set(error, ENOTSUP,
2705 RTE_FLOW_ERROR_TYPE_ITEM, item,
2706 "multiple tunnel layers not"
2708 if (!(item_flags & MLX5_FLOW_LAYER_OUTER_L3))
2709 return rte_flow_error_set(error, ENOTSUP,
2710 RTE_FLOW_ERROR_TYPE_ITEM, item,
2711 "L3 Layer is missing");
2713 mask = &rte_flow_item_nvgre_mask;
2714 ret = mlx5_flow_item_acceptable
2715 (item, (const uint8_t *)mask,
2716 (const uint8_t *)&rte_flow_item_nvgre_mask,
2717 sizeof(struct rte_flow_item_nvgre),
2718 MLX5_ITEM_RANGE_NOT_ACCEPTED, error);
2725 * Validate eCPRI item.
2728 * Item specification.
2729 * @param[in] item_flags
2730 * Bit-fields that holds the items detected until now.
2731 * @param[in] last_item
2732 * Previous validated item in the pattern items.
2733 * @param[in] ether_type
2734 * Type in the ethernet layer header (including dot1q).
2735 * @param[in] acc_mask
2736 * Acceptable mask, if NULL default internal default mask
2737 * will be used to check whether item fields are supported.
2739 * Pointer to error structure.
2742 * 0 on success, a negative errno value otherwise and rte_errno is set.
2745 mlx5_flow_validate_item_ecpri(const struct rte_flow_item *item,
2746 uint64_t item_flags,
2748 uint16_t ether_type,
2749 const struct rte_flow_item_ecpri *acc_mask,
2750 struct rte_flow_error *error)
2752 const struct rte_flow_item_ecpri *mask = item->mask;
2753 const struct rte_flow_item_ecpri nic_mask = {
2757 RTE_BE32(((const struct rte_ecpri_common_hdr) {
2761 .dummy[0] = 0xFFFFFFFF,
2764 const uint64_t outer_l2_vlan = (MLX5_FLOW_LAYER_OUTER_L2 |
2765 MLX5_FLOW_LAYER_OUTER_VLAN);
2766 struct rte_flow_item_ecpri mask_lo;
2768 if (!(last_item & outer_l2_vlan) &&
2769 last_item != MLX5_FLOW_LAYER_OUTER_L4_UDP)
2770 return rte_flow_error_set(error, EINVAL,
2771 RTE_FLOW_ERROR_TYPE_ITEM, item,
2772 "eCPRI can only follow L2/VLAN layer or UDP layer");
2773 if ((last_item & outer_l2_vlan) && ether_type &&
2774 ether_type != RTE_ETHER_TYPE_ECPRI)
2775 return rte_flow_error_set(error, EINVAL,
2776 RTE_FLOW_ERROR_TYPE_ITEM, item,
2777 "eCPRI cannot follow L2/VLAN layer which ether type is not 0xAEFE");
2778 if (item_flags & MLX5_FLOW_LAYER_TUNNEL)
2779 return rte_flow_error_set(error, EINVAL,
2780 RTE_FLOW_ERROR_TYPE_ITEM, item,
2781 "eCPRI with tunnel is not supported right now");
2782 if (item_flags & MLX5_FLOW_LAYER_OUTER_L3)
2783 return rte_flow_error_set(error, ENOTSUP,
2784 RTE_FLOW_ERROR_TYPE_ITEM, item,
2785 "multiple L3 layers not supported");
2786 else if (item_flags & MLX5_FLOW_LAYER_OUTER_L4_TCP)
2787 return rte_flow_error_set(error, EINVAL,
2788 RTE_FLOW_ERROR_TYPE_ITEM, item,
2789 "eCPRI cannot coexist with a TCP layer");
2790 /* In specification, eCPRI could be over UDP layer. */
2791 else if (item_flags & MLX5_FLOW_LAYER_OUTER_L4_UDP)
2792 return rte_flow_error_set(error, EINVAL,
2793 RTE_FLOW_ERROR_TYPE_ITEM, item,
2794 "eCPRI over UDP layer is not yet supported right now");
2795 /* Mask for type field in common header could be zero. */
2797 mask = &rte_flow_item_ecpri_mask;
2798 mask_lo.hdr.common.u32 = rte_be_to_cpu_32(mask->hdr.common.u32);
2799 /* Input mask is in big-endian format. */
2800 if (mask_lo.hdr.common.type != 0 && mask_lo.hdr.common.type != 0xff)
2801 return rte_flow_error_set(error, EINVAL,
2802 RTE_FLOW_ERROR_TYPE_ITEM_MASK, mask,
2803 "partial mask is not supported for protocol");
2804 else if (mask_lo.hdr.common.type == 0 && mask->hdr.dummy[0] != 0)
2805 return rte_flow_error_set(error, EINVAL,
2806 RTE_FLOW_ERROR_TYPE_ITEM_MASK, mask,
2807 "message header mask must be after a type mask");
2808 return mlx5_flow_item_acceptable(item, (const uint8_t *)mask,
2809 acc_mask ? (const uint8_t *)acc_mask
2810 : (const uint8_t *)&nic_mask,
2811 sizeof(struct rte_flow_item_ecpri),
2812 MLX5_ITEM_RANGE_NOT_ACCEPTED, error);
2816 * Release resource related QUEUE/RSS action split.
2819 * Pointer to Ethernet device.
2821 * Flow to release id's from.
2824 flow_mreg_split_qrss_release(struct rte_eth_dev *dev,
2825 struct rte_flow *flow)
2827 struct mlx5_priv *priv = dev->data->dev_private;
2828 uint32_t handle_idx;
2829 struct mlx5_flow_handle *dev_handle;
2831 SILIST_FOREACH(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW], flow->dev_handles,
2832 handle_idx, dev_handle, next)
2833 if (dev_handle->split_flow_id)
2834 mlx5_ipool_free(priv->sh->ipool
2835 [MLX5_IPOOL_RSS_EXPANTION_FLOW_ID],
2836 dev_handle->split_flow_id);
2840 flow_null_validate(struct rte_eth_dev *dev __rte_unused,
2841 const struct rte_flow_attr *attr __rte_unused,
2842 const struct rte_flow_item items[] __rte_unused,
2843 const struct rte_flow_action actions[] __rte_unused,
2844 bool external __rte_unused,
2845 int hairpin __rte_unused,
2846 struct rte_flow_error *error)
2848 return rte_flow_error_set(error, ENOTSUP,
2849 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL, NULL);
2852 static struct mlx5_flow *
2853 flow_null_prepare(struct rte_eth_dev *dev __rte_unused,
2854 const struct rte_flow_attr *attr __rte_unused,
2855 const struct rte_flow_item items[] __rte_unused,
2856 const struct rte_flow_action actions[] __rte_unused,
2857 struct rte_flow_error *error)
2859 rte_flow_error_set(error, ENOTSUP,
2860 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL, NULL);
2865 flow_null_translate(struct rte_eth_dev *dev __rte_unused,
2866 struct mlx5_flow *dev_flow __rte_unused,
2867 const struct rte_flow_attr *attr __rte_unused,
2868 const struct rte_flow_item items[] __rte_unused,
2869 const struct rte_flow_action actions[] __rte_unused,
2870 struct rte_flow_error *error)
2872 return rte_flow_error_set(error, ENOTSUP,
2873 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL, NULL);
2877 flow_null_apply(struct rte_eth_dev *dev __rte_unused,
2878 struct rte_flow *flow __rte_unused,
2879 struct rte_flow_error *error)
2881 return rte_flow_error_set(error, ENOTSUP,
2882 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL, NULL);
2886 flow_null_remove(struct rte_eth_dev *dev __rte_unused,
2887 struct rte_flow *flow __rte_unused)
2892 flow_null_destroy(struct rte_eth_dev *dev __rte_unused,
2893 struct rte_flow *flow __rte_unused)
2898 flow_null_query(struct rte_eth_dev *dev __rte_unused,
2899 struct rte_flow *flow __rte_unused,
2900 const struct rte_flow_action *actions __rte_unused,
2901 void *data __rte_unused,
2902 struct rte_flow_error *error)
2904 return rte_flow_error_set(error, ENOTSUP,
2905 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL, NULL);
2909 flow_null_sync_domain(struct rte_eth_dev *dev __rte_unused,
2910 uint32_t domains __rte_unused,
2911 uint32_t flags __rte_unused)
2916 /* Void driver to protect from null pointer reference. */
2917 const struct mlx5_flow_driver_ops mlx5_flow_null_drv_ops = {
2918 .validate = flow_null_validate,
2919 .prepare = flow_null_prepare,
2920 .translate = flow_null_translate,
2921 .apply = flow_null_apply,
2922 .remove = flow_null_remove,
2923 .destroy = flow_null_destroy,
2924 .query = flow_null_query,
2925 .sync_domain = flow_null_sync_domain,
2929 * Select flow driver type according to flow attributes and device
2933 * Pointer to the dev structure.
2935 * Pointer to the flow attributes.
2938 * flow driver type, MLX5_FLOW_TYPE_MAX otherwise.
2940 static enum mlx5_flow_drv_type
2941 flow_get_drv_type(struct rte_eth_dev *dev, const struct rte_flow_attr *attr)
2943 struct mlx5_priv *priv = dev->data->dev_private;
2944 /* The OS can determine first a specific flow type (DV, VERBS) */
2945 enum mlx5_flow_drv_type type = mlx5_flow_os_get_type();
2947 if (type != MLX5_FLOW_TYPE_MAX)
2949 /* If no OS specific type - continue with DV/VERBS selection */
2950 if (attr->transfer && priv->config.dv_esw_en)
2951 type = MLX5_FLOW_TYPE_DV;
2952 if (!attr->transfer)
2953 type = priv->config.dv_flow_en ? MLX5_FLOW_TYPE_DV :
2954 MLX5_FLOW_TYPE_VERBS;
2958 #define flow_get_drv_ops(type) flow_drv_ops[type]
2961 * Flow driver validation API. This abstracts calling driver specific functions.
2962 * The type of flow driver is determined according to flow attributes.
2965 * Pointer to the dev structure.
2967 * Pointer to the flow attributes.
2969 * Pointer to the list of items.
2970 * @param[in] actions
2971 * Pointer to the list of actions.
2972 * @param[in] external
2973 * This flow rule is created by request external to PMD.
2974 * @param[in] hairpin
2975 * Number of hairpin TX actions, 0 means classic flow.
2977 * Pointer to the error structure.
2980 * 0 on success, a negative errno value otherwise and rte_errno is set.
2983 flow_drv_validate(struct rte_eth_dev *dev,
2984 const struct rte_flow_attr *attr,
2985 const struct rte_flow_item items[],
2986 const struct rte_flow_action actions[],
2987 bool external, int hairpin, struct rte_flow_error *error)
2989 const struct mlx5_flow_driver_ops *fops;
2990 enum mlx5_flow_drv_type type = flow_get_drv_type(dev, attr);
2992 fops = flow_get_drv_ops(type);
2993 return fops->validate(dev, attr, items, actions, external,
2998 * Flow driver preparation API. This abstracts calling driver specific
2999 * functions. Parent flow (rte_flow) should have driver type (drv_type). It
3000 * calculates the size of memory required for device flow, allocates the memory,
3001 * initializes the device flow and returns the pointer.
3004 * This function initializes device flow structure such as dv or verbs in
3005 * struct mlx5_flow. However, it is caller's responsibility to initialize the
3006 * rest. For example, adding returning device flow to flow->dev_flow list and
3007 * setting backward reference to the flow should be done out of this function.
3008 * layers field is not filled either.
3011 * Pointer to the dev structure.
3013 * Pointer to the flow attributes.
3015 * Pointer to the list of items.
3016 * @param[in] actions
3017 * Pointer to the list of actions.
3018 * @param[in] flow_idx
3019 * This memory pool index to the flow.
3021 * Pointer to the error structure.
3024 * Pointer to device flow on success, otherwise NULL and rte_errno is set.
3026 static inline struct mlx5_flow *
3027 flow_drv_prepare(struct rte_eth_dev *dev,
3028 const struct rte_flow *flow,
3029 const struct rte_flow_attr *attr,
3030 const struct rte_flow_item items[],
3031 const struct rte_flow_action actions[],
3033 struct rte_flow_error *error)
3035 const struct mlx5_flow_driver_ops *fops;
3036 enum mlx5_flow_drv_type type = flow->drv_type;
3037 struct mlx5_flow *mlx5_flow = NULL;
3039 MLX5_ASSERT(type > MLX5_FLOW_TYPE_MIN && type < MLX5_FLOW_TYPE_MAX);
3040 fops = flow_get_drv_ops(type);
3041 mlx5_flow = fops->prepare(dev, attr, items, actions, error);
3043 mlx5_flow->flow_idx = flow_idx;
3048 * Flow driver translation API. This abstracts calling driver specific
3049 * functions. Parent flow (rte_flow) should have driver type (drv_type). It
3050 * translates a generic flow into a driver flow. flow_drv_prepare() must
3054 * dev_flow->layers could be filled as a result of parsing during translation
3055 * if needed by flow_drv_apply(). dev_flow->flow->actions can also be filled
3056 * if necessary. As a flow can have multiple dev_flows by RSS flow expansion,
3057 * flow->actions could be overwritten even though all the expanded dev_flows
3058 * have the same actions.
3061 * Pointer to the rte dev structure.
3062 * @param[in, out] dev_flow
3063 * Pointer to the mlx5 flow.
3065 * Pointer to the flow attributes.
3067 * Pointer to the list of items.
3068 * @param[in] actions
3069 * Pointer to the list of actions.
3071 * Pointer to the error structure.
3074 * 0 on success, a negative errno value otherwise and rte_errno is set.
3077 flow_drv_translate(struct rte_eth_dev *dev, struct mlx5_flow *dev_flow,
3078 const struct rte_flow_attr *attr,
3079 const struct rte_flow_item items[],
3080 const struct rte_flow_action actions[],
3081 struct rte_flow_error *error)
3083 const struct mlx5_flow_driver_ops *fops;
3084 enum mlx5_flow_drv_type type = dev_flow->flow->drv_type;
3086 MLX5_ASSERT(type > MLX5_FLOW_TYPE_MIN && type < MLX5_FLOW_TYPE_MAX);
3087 fops = flow_get_drv_ops(type);
3088 return fops->translate(dev, dev_flow, attr, items, actions, error);
3092 * Flow driver apply API. This abstracts calling driver specific functions.
3093 * Parent flow (rte_flow) should have driver type (drv_type). It applies
3094 * translated driver flows on to device. flow_drv_translate() must precede.
3097 * Pointer to Ethernet device structure.
3098 * @param[in, out] flow
3099 * Pointer to flow structure.
3101 * Pointer to error structure.
3104 * 0 on success, a negative errno value otherwise and rte_errno is set.
3107 flow_drv_apply(struct rte_eth_dev *dev, struct rte_flow *flow,
3108 struct rte_flow_error *error)
3110 const struct mlx5_flow_driver_ops *fops;
3111 enum mlx5_flow_drv_type type = flow->drv_type;
3113 MLX5_ASSERT(type > MLX5_FLOW_TYPE_MIN && type < MLX5_FLOW_TYPE_MAX);
3114 fops = flow_get_drv_ops(type);
3115 return fops->apply(dev, flow, error);
3119 * Flow driver destroy API. This abstracts calling driver specific functions.
3120 * Parent flow (rte_flow) should have driver type (drv_type). It removes a flow
3121 * on device and releases resources of the flow.
3124 * Pointer to Ethernet device.
3125 * @param[in, out] flow
3126 * Pointer to flow structure.
3129 flow_drv_destroy(struct rte_eth_dev *dev, struct rte_flow *flow)
3131 const struct mlx5_flow_driver_ops *fops;
3132 enum mlx5_flow_drv_type type = flow->drv_type;
3134 flow_mreg_split_qrss_release(dev, flow);
3135 MLX5_ASSERT(type > MLX5_FLOW_TYPE_MIN && type < MLX5_FLOW_TYPE_MAX);
3136 fops = flow_get_drv_ops(type);
3137 fops->destroy(dev, flow);
3141 * Get RSS action from the action list.
3143 * @param[in] actions
3144 * Pointer to the list of actions.
3147 * Pointer to the RSS action if exist, else return NULL.
3149 static const struct rte_flow_action_rss*
3150 flow_get_rss_action(const struct rte_flow_action actions[])
3152 for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
3153 switch (actions->type) {
3154 case RTE_FLOW_ACTION_TYPE_RSS:
3155 return (const struct rte_flow_action_rss *)
3165 * Get ASO age action by index.
3168 * Pointer to the Ethernet device structure.
3169 * @param[in] age_idx
3170 * Index to the ASO age action.
3173 * The specified ASO age action.
3175 struct mlx5_aso_age_action*
3176 flow_aso_age_get_by_idx(struct rte_eth_dev *dev, uint32_t age_idx)
3178 uint16_t pool_idx = age_idx & UINT16_MAX;
3179 uint16_t offset = (age_idx >> 16) & UINT16_MAX;
3180 struct mlx5_priv *priv = dev->data->dev_private;
3181 struct mlx5_aso_age_mng *mng = priv->sh->aso_age_mng;
3182 struct mlx5_aso_age_pool *pool = mng->pools[pool_idx];
3184 return &pool->actions[offset - 1];
3187 /* maps shared action to translated non shared in some actions array */
3188 struct mlx5_translated_shared_action {
3189 struct rte_flow_shared_action *action; /**< Shared action */
3190 int index; /**< Index in related array of rte_flow_action */
3194 * Translates actions of type RTE_FLOW_ACTION_TYPE_SHARED to related
3195 * non shared action if translation possible.
3196 * This functionality used to run same execution path for both shared & non
3197 * shared actions on flow create. All necessary preparations for shared
3198 * action handling should be preformed on *shared* actions list returned
3202 * Pointer to Ethernet device.
3203 * @param[in] actions
3204 * List of actions to translate.
3205 * @param[out] shared
3206 * List to store translated shared actions.
3207 * @param[in, out] shared_n
3208 * Size of *shared* array. On return should be updated with number of shared
3209 * actions retrieved from the *actions* list.
3210 * @param[out] translated_actions
3211 * List of actions where all shared actions were translated to non shared
3212 * if possible. NULL if no translation took place.
3214 * Pointer to the error structure.
3217 * 0 on success, a negative errno value otherwise and rte_errno is set.
3220 flow_shared_actions_translate(struct rte_eth_dev *dev,
3221 const struct rte_flow_action actions[],
3222 struct mlx5_translated_shared_action *shared,
3224 struct rte_flow_action **translated_actions,
3225 struct rte_flow_error *error)
3227 struct mlx5_priv *priv = dev->data->dev_private;
3228 struct rte_flow_action *translated = NULL;
3229 size_t actions_size;
3232 struct mlx5_translated_shared_action *shared_end = NULL;
3234 for (n = 0; actions[n].type != RTE_FLOW_ACTION_TYPE_END; n++) {
3235 if (actions[n].type != RTE_FLOW_ACTION_TYPE_SHARED)
3237 if (copied_n == *shared_n) {
3238 return rte_flow_error_set
3239 (error, EINVAL, RTE_FLOW_ERROR_TYPE_ACTION_NUM,
3240 NULL, "too many shared actions");
3242 rte_memcpy(&shared[copied_n].action, &actions[n].conf,
3243 sizeof(actions[n].conf));
3244 shared[copied_n].index = n;
3248 *shared_n = copied_n;
3251 actions_size = sizeof(struct rte_flow_action) * n;
3252 translated = mlx5_malloc(MLX5_MEM_ZERO, actions_size, 0, SOCKET_ID_ANY);
3257 memcpy(translated, actions, actions_size);
3258 for (shared_end = shared + copied_n; shared < shared_end; shared++) {
3259 struct mlx5_shared_action_rss *shared_rss;
3260 uint32_t act_idx = (uint32_t)(uintptr_t)shared->action;
3261 uint32_t type = act_idx >> MLX5_SHARED_ACTION_TYPE_OFFSET;
3262 uint32_t idx = act_idx & ((1u << MLX5_SHARED_ACTION_TYPE_OFFSET)
3266 case MLX5_SHARED_ACTION_TYPE_RSS:
3267 shared_rss = mlx5_ipool_get
3268 (priv->sh->ipool[MLX5_IPOOL_RSS_SHARED_ACTIONS], idx);
3269 translated[shared->index].type =
3270 RTE_FLOW_ACTION_TYPE_RSS;
3271 translated[shared->index].conf =
3272 &shared_rss->origin;
3274 case MLX5_SHARED_ACTION_TYPE_AGE:
3275 if (priv->sh->flow_hit_aso_en) {
3276 translated[shared->index].type =
3277 (enum rte_flow_action_type)
3278 MLX5_RTE_FLOW_ACTION_TYPE_AGE;
3279 translated[shared->index].conf =
3280 (void *)(uintptr_t)idx;
3285 mlx5_free(translated);
3286 return rte_flow_error_set
3287 (error, EINVAL, RTE_FLOW_ERROR_TYPE_ACTION,
3288 NULL, "invalid shared action type");
3291 *translated_actions = translated;
3296 * Get Shared RSS action from the action list.
3299 * Pointer to Ethernet device.
3301 * Pointer to the list of actions.
3302 * @param[in] shared_n
3303 * Actions list length.
3306 * The MLX5 RSS action ID if exists, otherwise return 0.
3309 flow_get_shared_rss_action(struct rte_eth_dev *dev,
3310 struct mlx5_translated_shared_action *shared,
3313 struct mlx5_translated_shared_action *shared_end;
3314 struct mlx5_priv *priv = dev->data->dev_private;
3315 struct mlx5_shared_action_rss *shared_rss;
3318 for (shared_end = shared + shared_n; shared < shared_end; shared++) {
3319 uint32_t act_idx = (uint32_t)(uintptr_t)shared->action;
3320 uint32_t type = act_idx >> MLX5_SHARED_ACTION_TYPE_OFFSET;
3321 uint32_t idx = act_idx &
3322 ((1u << MLX5_SHARED_ACTION_TYPE_OFFSET) - 1);
3324 case MLX5_SHARED_ACTION_TYPE_RSS:
3325 shared_rss = mlx5_ipool_get
3326 (priv->sh->ipool[MLX5_IPOOL_RSS_SHARED_ACTIONS],
3328 __atomic_add_fetch(&shared_rss->refcnt, 1,
3339 find_graph_root(const struct rte_flow_item pattern[], uint32_t rss_level)
3341 const struct rte_flow_item *item;
3342 unsigned int has_vlan = 0;
3344 for (item = pattern; item->type != RTE_FLOW_ITEM_TYPE_END; item++) {
3345 if (item->type == RTE_FLOW_ITEM_TYPE_VLAN) {
3351 return rss_level < 2 ? MLX5_EXPANSION_ROOT_ETH_VLAN :
3352 MLX5_EXPANSION_ROOT_OUTER_ETH_VLAN;
3353 return rss_level < 2 ? MLX5_EXPANSION_ROOT :
3354 MLX5_EXPANSION_ROOT_OUTER;
3358 * Get layer flags from the prefix flow.
3360 * Some flows may be split to several subflows, the prefix subflow gets the
3361 * match items and the suffix sub flow gets the actions.
3362 * Some actions need the user defined match item flags to get the detail for
3364 * This function helps the suffix flow to get the item layer flags from prefix
3367 * @param[in] dev_flow
3368 * Pointer the created preifx subflow.
3371 * The layers get from prefix subflow.
3373 static inline uint64_t
3374 flow_get_prefix_layer_flags(struct mlx5_flow *dev_flow)
3376 uint64_t layers = 0;
3379 * Layers bits could be localization, but usually the compiler will
3380 * help to do the optimization work for source code.
3381 * If no decap actions, use the layers directly.
3383 if (!(dev_flow->act_flags & MLX5_FLOW_ACTION_DECAP))
3384 return dev_flow->handle->layers;
3385 /* Convert L3 layers with decap action. */
3386 if (dev_flow->handle->layers & MLX5_FLOW_LAYER_INNER_L3_IPV4)
3387 layers |= MLX5_FLOW_LAYER_OUTER_L3_IPV4;
3388 else if (dev_flow->handle->layers & MLX5_FLOW_LAYER_INNER_L3_IPV6)
3389 layers |= MLX5_FLOW_LAYER_OUTER_L3_IPV6;
3390 /* Convert L4 layers with decap action. */
3391 if (dev_flow->handle->layers & MLX5_FLOW_LAYER_INNER_L4_TCP)
3392 layers |= MLX5_FLOW_LAYER_OUTER_L4_TCP;
3393 else if (dev_flow->handle->layers & MLX5_FLOW_LAYER_INNER_L4_UDP)
3394 layers |= MLX5_FLOW_LAYER_OUTER_L4_UDP;
3399 * Get metadata split action information.
3401 * @param[in] actions
3402 * Pointer to the list of actions.
3404 * Pointer to the return pointer.
3405 * @param[out] qrss_type
3406 * Pointer to the action type to return. RTE_FLOW_ACTION_TYPE_END is returned
3407 * if no QUEUE/RSS is found.
3408 * @param[out] encap_idx
3409 * Pointer to the index of the encap action if exists, otherwise the last
3413 * Total number of actions.
3416 flow_parse_metadata_split_actions_info(const struct rte_flow_action actions[],
3417 const struct rte_flow_action **qrss,
3420 const struct rte_flow_action_raw_encap *raw_encap;
3422 int raw_decap_idx = -1;
3425 for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
3426 switch (actions->type) {
3427 case RTE_FLOW_ACTION_TYPE_VXLAN_ENCAP:
3428 case RTE_FLOW_ACTION_TYPE_NVGRE_ENCAP:
3429 *encap_idx = actions_n;
3431 case RTE_FLOW_ACTION_TYPE_RAW_DECAP:
3432 raw_decap_idx = actions_n;
3434 case RTE_FLOW_ACTION_TYPE_RAW_ENCAP:
3435 raw_encap = actions->conf;
3436 if (raw_encap->size > MLX5_ENCAPSULATION_DECISION_SIZE)
3437 *encap_idx = raw_decap_idx != -1 ?
3438 raw_decap_idx : actions_n;
3440 case RTE_FLOW_ACTION_TYPE_QUEUE:
3441 case RTE_FLOW_ACTION_TYPE_RSS:
3449 if (*encap_idx == -1)
3450 *encap_idx = actions_n;
3451 /* Count RTE_FLOW_ACTION_TYPE_END. */
3452 return actions_n + 1;
3456 * Check meter action from the action list.
3458 * @param[in] actions
3459 * Pointer to the list of actions.
3461 * Pointer to the meter exist flag.
3464 * Total number of actions.
3467 flow_check_meter_action(const struct rte_flow_action actions[], uint32_t *mtr)
3473 for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
3474 switch (actions->type) {
3475 case RTE_FLOW_ACTION_TYPE_METER:
3483 /* Count RTE_FLOW_ACTION_TYPE_END. */
3484 return actions_n + 1;
3488 * Check if the flow should be split due to hairpin.
3489 * The reason for the split is that in current HW we can't
3490 * support encap and push-vlan on Rx, so if a flow contains
3491 * these actions we move it to Tx.
3494 * Pointer to Ethernet device.
3496 * Flow rule attributes.
3497 * @param[in] actions
3498 * Associated actions (list terminated by the END action).
3501 * > 0 the number of actions and the flow should be split,
3502 * 0 when no split required.
3505 flow_check_hairpin_split(struct rte_eth_dev *dev,
3506 const struct rte_flow_attr *attr,
3507 const struct rte_flow_action actions[])
3509 int queue_action = 0;
3512 const struct rte_flow_action_queue *queue;
3513 const struct rte_flow_action_rss *rss;
3514 const struct rte_flow_action_raw_encap *raw_encap;
3515 const struct rte_eth_hairpin_conf *conf;
3519 for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
3520 switch (actions->type) {
3521 case RTE_FLOW_ACTION_TYPE_QUEUE:
3522 queue = actions->conf;
3525 conf = mlx5_rxq_get_hairpin_conf(dev, queue->index);
3526 if (conf != NULL && !!conf->tx_explicit)
3531 case RTE_FLOW_ACTION_TYPE_RSS:
3532 rss = actions->conf;
3533 if (rss == NULL || rss->queue_num == 0)
3535 conf = mlx5_rxq_get_hairpin_conf(dev, rss->queue[0]);
3536 if (conf != NULL && !!conf->tx_explicit)
3541 case RTE_FLOW_ACTION_TYPE_VXLAN_ENCAP:
3542 case RTE_FLOW_ACTION_TYPE_NVGRE_ENCAP:
3543 case RTE_FLOW_ACTION_TYPE_OF_PUSH_VLAN:
3544 case RTE_FLOW_ACTION_TYPE_OF_SET_VLAN_VID:
3545 case RTE_FLOW_ACTION_TYPE_OF_SET_VLAN_PCP:
3549 case RTE_FLOW_ACTION_TYPE_RAW_ENCAP:
3550 raw_encap = actions->conf;
3551 if (raw_encap->size > MLX5_ENCAPSULATION_DECISION_SIZE)
3560 if (split && queue_action)
3565 /* Declare flow create/destroy prototype in advance. */
3567 flow_list_create(struct rte_eth_dev *dev, uint32_t *list,
3568 const struct rte_flow_attr *attr,
3569 const struct rte_flow_item items[],
3570 const struct rte_flow_action actions[],
3571 bool external, struct rte_flow_error *error);
3574 flow_list_destroy(struct rte_eth_dev *dev, uint32_t *list,
3577 struct mlx5_hlist_entry *
3578 flow_dv_mreg_create_cb(struct mlx5_hlist *list, uint64_t key,
3581 struct rte_eth_dev *dev = list->ctx;
3582 struct mlx5_priv *priv = dev->data->dev_private;
3583 struct mlx5_flow_cb_ctx *ctx = cb_ctx;
3584 struct mlx5_flow_mreg_copy_resource *mcp_res;
3585 struct rte_flow_error *error = ctx->error;
3588 uint32_t mark_id = key;
3589 struct rte_flow_attr attr = {
3590 .group = MLX5_FLOW_MREG_CP_TABLE_GROUP,
3593 struct mlx5_rte_flow_item_tag tag_spec = {
3596 struct rte_flow_item items[] = {
3597 [1] = { .type = RTE_FLOW_ITEM_TYPE_END, },
3599 struct rte_flow_action_mark ftag = {
3602 struct mlx5_flow_action_copy_mreg cp_mreg = {
3606 struct rte_flow_action_jump jump = {
3607 .group = MLX5_FLOW_MREG_ACT_TABLE_GROUP,
3609 struct rte_flow_action actions[] = {
3610 [3] = { .type = RTE_FLOW_ACTION_TYPE_END, },
3613 /* Fill the register fileds in the flow. */
3614 ret = mlx5_flow_get_reg_id(dev, MLX5_FLOW_MARK, 0, error);
3618 ret = mlx5_flow_get_reg_id(dev, MLX5_METADATA_RX, 0, error);
3622 /* Provide the full width of FLAG specific value. */
3623 if (mark_id == (priv->sh->dv_regc0_mask & MLX5_FLOW_MARK_DEFAULT))
3624 tag_spec.data = MLX5_FLOW_MARK_DEFAULT;
3625 /* Build a new flow. */
3626 if (mark_id != MLX5_DEFAULT_COPY_ID) {
3627 items[0] = (struct rte_flow_item){
3628 .type = (enum rte_flow_item_type)
3629 MLX5_RTE_FLOW_ITEM_TYPE_TAG,
3632 items[1] = (struct rte_flow_item){
3633 .type = RTE_FLOW_ITEM_TYPE_END,
3635 actions[0] = (struct rte_flow_action){
3636 .type = (enum rte_flow_action_type)
3637 MLX5_RTE_FLOW_ACTION_TYPE_MARK,
3640 actions[1] = (struct rte_flow_action){
3641 .type = (enum rte_flow_action_type)
3642 MLX5_RTE_FLOW_ACTION_TYPE_COPY_MREG,
3645 actions[2] = (struct rte_flow_action){
3646 .type = RTE_FLOW_ACTION_TYPE_JUMP,
3649 actions[3] = (struct rte_flow_action){
3650 .type = RTE_FLOW_ACTION_TYPE_END,
3653 /* Default rule, wildcard match. */
3654 attr.priority = MLX5_FLOW_PRIO_RSVD;
3655 items[0] = (struct rte_flow_item){
3656 .type = RTE_FLOW_ITEM_TYPE_END,
3658 actions[0] = (struct rte_flow_action){
3659 .type = (enum rte_flow_action_type)
3660 MLX5_RTE_FLOW_ACTION_TYPE_COPY_MREG,
3663 actions[1] = (struct rte_flow_action){
3664 .type = RTE_FLOW_ACTION_TYPE_JUMP,
3667 actions[2] = (struct rte_flow_action){
3668 .type = RTE_FLOW_ACTION_TYPE_END,
3671 /* Build a new entry. */
3672 mcp_res = mlx5_ipool_zmalloc(priv->sh->ipool[MLX5_IPOOL_MCP], &idx);
3679 * The copy Flows are not included in any list. There
3680 * ones are referenced from other Flows and can not
3681 * be applied, removed, deleted in ardbitrary order
3682 * by list traversing.
3684 mcp_res->rix_flow = flow_list_create(dev, NULL, &attr, items,
3685 actions, false, error);
3686 if (!mcp_res->rix_flow) {
3687 mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_MCP], idx);
3690 return &mcp_res->hlist_ent;
3694 * Add a flow of copying flow metadata registers in RX_CP_TBL.
3696 * As mark_id is unique, if there's already a registered flow for the mark_id,
3697 * return by increasing the reference counter of the resource. Otherwise, create
3698 * the resource (mcp_res) and flow.
3701 * - If ingress port is ANY and reg_c[1] is mark_id,
3702 * flow_tag := mark_id, reg_b := reg_c[0] and jump to RX_ACT_TBL.
3704 * For default flow (zero mark_id), flow is like,
3705 * - If ingress port is ANY,
3706 * reg_b := reg_c[0] and jump to RX_ACT_TBL.
3709 * Pointer to Ethernet device.
3711 * ID of MARK action, zero means default flow for META.
3713 * Perform verbose error reporting if not NULL.
3716 * Associated resource on success, NULL otherwise and rte_errno is set.
3718 static struct mlx5_flow_mreg_copy_resource *
3719 flow_mreg_add_copy_action(struct rte_eth_dev *dev, uint32_t mark_id,
3720 struct rte_flow_error *error)
3722 struct mlx5_priv *priv = dev->data->dev_private;
3723 struct mlx5_hlist_entry *entry;
3724 struct mlx5_flow_cb_ctx ctx = {
3729 /* Check if already registered. */
3730 MLX5_ASSERT(priv->mreg_cp_tbl);
3731 entry = mlx5_hlist_register(priv->mreg_cp_tbl, mark_id, &ctx);
3734 return container_of(entry, struct mlx5_flow_mreg_copy_resource,
3739 flow_dv_mreg_remove_cb(struct mlx5_hlist *list, struct mlx5_hlist_entry *entry)
3741 struct mlx5_flow_mreg_copy_resource *mcp_res =
3742 container_of(entry, typeof(*mcp_res), hlist_ent);
3743 struct rte_eth_dev *dev = list->ctx;
3744 struct mlx5_priv *priv = dev->data->dev_private;
3746 MLX5_ASSERT(mcp_res->rix_flow);
3747 flow_list_destroy(dev, NULL, mcp_res->rix_flow);
3748 mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_MCP], mcp_res->idx);
3752 * Release flow in RX_CP_TBL.
3755 * Pointer to Ethernet device.
3757 * Parent flow for wich copying is provided.
3760 flow_mreg_del_copy_action(struct rte_eth_dev *dev,
3761 struct rte_flow *flow)
3763 struct mlx5_flow_mreg_copy_resource *mcp_res;
3764 struct mlx5_priv *priv = dev->data->dev_private;
3766 if (!flow->rix_mreg_copy)
3768 mcp_res = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_MCP],
3769 flow->rix_mreg_copy);
3770 if (!mcp_res || !priv->mreg_cp_tbl)
3772 MLX5_ASSERT(mcp_res->rix_flow);
3773 mlx5_hlist_unregister(priv->mreg_cp_tbl, &mcp_res->hlist_ent);
3774 flow->rix_mreg_copy = 0;
3778 * Remove the default copy action from RX_CP_TBL.
3780 * This functions is called in the mlx5_dev_start(). No thread safe
3784 * Pointer to Ethernet device.
3787 flow_mreg_del_default_copy_action(struct rte_eth_dev *dev)
3789 struct mlx5_hlist_entry *entry;
3790 struct mlx5_priv *priv = dev->data->dev_private;
3792 /* Check if default flow is registered. */
3793 if (!priv->mreg_cp_tbl)
3795 entry = mlx5_hlist_lookup(priv->mreg_cp_tbl,
3796 MLX5_DEFAULT_COPY_ID, NULL);
3799 mlx5_hlist_unregister(priv->mreg_cp_tbl, entry);
3803 * Add the default copy action in in RX_CP_TBL.
3805 * This functions is called in the mlx5_dev_start(). No thread safe
3809 * Pointer to Ethernet device.
3811 * Perform verbose error reporting if not NULL.
3814 * 0 for success, negative value otherwise and rte_errno is set.
3817 flow_mreg_add_default_copy_action(struct rte_eth_dev *dev,
3818 struct rte_flow_error *error)
3820 struct mlx5_priv *priv = dev->data->dev_private;
3821 struct mlx5_flow_mreg_copy_resource *mcp_res;
3823 /* Check whether extensive metadata feature is engaged. */
3824 if (!priv->config.dv_flow_en ||
3825 priv->config.dv_xmeta_en == MLX5_XMETA_MODE_LEGACY ||
3826 !mlx5_flow_ext_mreg_supported(dev) ||
3827 !priv->sh->dv_regc0_mask)
3830 * Add default mreg copy flow may be called multiple time, but
3831 * only be called once in stop. Avoid register it twice.
3833 if (mlx5_hlist_lookup(priv->mreg_cp_tbl, MLX5_DEFAULT_COPY_ID, NULL))
3835 mcp_res = flow_mreg_add_copy_action(dev, MLX5_DEFAULT_COPY_ID, error);
3842 * Add a flow of copying flow metadata registers in RX_CP_TBL.
3844 * All the flow having Q/RSS action should be split by
3845 * flow_mreg_split_qrss_prep() to pass by RX_CP_TBL. A flow in the RX_CP_TBL
3846 * performs the following,
3847 * - CQE->flow_tag := reg_c[1] (MARK)
3848 * - CQE->flow_table_metadata (reg_b) := reg_c[0] (META)
3849 * As CQE's flow_tag is not a register, it can't be simply copied from reg_c[1]
3850 * but there should be a flow per each MARK ID set by MARK action.
3852 * For the aforementioned reason, if there's a MARK action in flow's action
3853 * list, a corresponding flow should be added to the RX_CP_TBL in order to copy
3854 * the MARK ID to CQE's flow_tag like,
3855 * - If reg_c[1] is mark_id,
3856 * flow_tag := mark_id, reg_b := reg_c[0] and jump to RX_ACT_TBL.
3858 * For SET_META action which stores value in reg_c[0], as the destination is
3859 * also a flow metadata register (reg_b), adding a default flow is enough. Zero
3860 * MARK ID means the default flow. The default flow looks like,
3861 * - For all flow, reg_b := reg_c[0] and jump to RX_ACT_TBL.
3864 * Pointer to Ethernet device.
3866 * Pointer to flow structure.
3867 * @param[in] actions
3868 * Pointer to the list of actions.
3870 * Perform verbose error reporting if not NULL.
3873 * 0 on success, negative value otherwise and rte_errno is set.
3876 flow_mreg_update_copy_table(struct rte_eth_dev *dev,
3877 struct rte_flow *flow,
3878 const struct rte_flow_action *actions,
3879 struct rte_flow_error *error)
3881 struct mlx5_priv *priv = dev->data->dev_private;
3882 struct mlx5_dev_config *config = &priv->config;
3883 struct mlx5_flow_mreg_copy_resource *mcp_res;
3884 const struct rte_flow_action_mark *mark;
3886 /* Check whether extensive metadata feature is engaged. */
3887 if (!config->dv_flow_en ||
3888 config->dv_xmeta_en == MLX5_XMETA_MODE_LEGACY ||
3889 !mlx5_flow_ext_mreg_supported(dev) ||
3890 !priv->sh->dv_regc0_mask)
3892 /* Find MARK action. */
3893 for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
3894 switch (actions->type) {
3895 case RTE_FLOW_ACTION_TYPE_FLAG:
3896 mcp_res = flow_mreg_add_copy_action
3897 (dev, MLX5_FLOW_MARK_DEFAULT, error);
3900 flow->rix_mreg_copy = mcp_res->idx;
3902 case RTE_FLOW_ACTION_TYPE_MARK:
3903 mark = (const struct rte_flow_action_mark *)
3906 flow_mreg_add_copy_action(dev, mark->id, error);
3909 flow->rix_mreg_copy = mcp_res->idx;
3918 #define MLX5_MAX_SPLIT_ACTIONS 24
3919 #define MLX5_MAX_SPLIT_ITEMS 24
3922 * Split the hairpin flow.
3923 * Since HW can't support encap and push-vlan on Rx, we move these
3925 * If the count action is after the encap then we also
3926 * move the count action. in this case the count will also measure
3930 * Pointer to Ethernet device.
3931 * @param[in] actions
3932 * Associated actions (list terminated by the END action).
3933 * @param[out] actions_rx
3935 * @param[out] actions_tx
3937 * @param[out] pattern_tx
3938 * The pattern items for the Tx flow.
3939 * @param[out] flow_id
3940 * The flow ID connected to this flow.
3946 flow_hairpin_split(struct rte_eth_dev *dev,
3947 const struct rte_flow_action actions[],
3948 struct rte_flow_action actions_rx[],
3949 struct rte_flow_action actions_tx[],
3950 struct rte_flow_item pattern_tx[],
3953 const struct rte_flow_action_raw_encap *raw_encap;
3954 const struct rte_flow_action_raw_decap *raw_decap;
3955 struct mlx5_rte_flow_action_set_tag *set_tag;
3956 struct rte_flow_action *tag_action;
3957 struct mlx5_rte_flow_item_tag *tag_item;
3958 struct rte_flow_item *item;
3962 for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
3963 switch (actions->type) {
3964 case RTE_FLOW_ACTION_TYPE_VXLAN_ENCAP:
3965 case RTE_FLOW_ACTION_TYPE_NVGRE_ENCAP:
3966 case RTE_FLOW_ACTION_TYPE_OF_PUSH_VLAN:
3967 case RTE_FLOW_ACTION_TYPE_OF_SET_VLAN_VID:
3968 case RTE_FLOW_ACTION_TYPE_OF_SET_VLAN_PCP:
3969 rte_memcpy(actions_tx, actions,
3970 sizeof(struct rte_flow_action));
3973 case RTE_FLOW_ACTION_TYPE_COUNT:
3975 rte_memcpy(actions_tx, actions,
3976 sizeof(struct rte_flow_action));
3979 rte_memcpy(actions_rx, actions,
3980 sizeof(struct rte_flow_action));
3984 case RTE_FLOW_ACTION_TYPE_RAW_ENCAP:
3985 raw_encap = actions->conf;
3986 if (raw_encap->size > MLX5_ENCAPSULATION_DECISION_SIZE) {
3987 memcpy(actions_tx, actions,
3988 sizeof(struct rte_flow_action));
3992 rte_memcpy(actions_rx, actions,
3993 sizeof(struct rte_flow_action));
3997 case RTE_FLOW_ACTION_TYPE_RAW_DECAP:
3998 raw_decap = actions->conf;
3999 if (raw_decap->size < MLX5_ENCAPSULATION_DECISION_SIZE) {
4000 memcpy(actions_tx, actions,
4001 sizeof(struct rte_flow_action));
4004 rte_memcpy(actions_rx, actions,
4005 sizeof(struct rte_flow_action));
4010 rte_memcpy(actions_rx, actions,
4011 sizeof(struct rte_flow_action));
4016 /* Add set meta action and end action for the Rx flow. */
4017 tag_action = actions_rx;
4018 tag_action->type = (enum rte_flow_action_type)
4019 MLX5_RTE_FLOW_ACTION_TYPE_TAG;
4021 rte_memcpy(actions_rx, actions, sizeof(struct rte_flow_action));
4023 set_tag = (void *)actions_rx;
4024 set_tag->id = mlx5_flow_get_reg_id(dev, MLX5_HAIRPIN_RX, 0, NULL);
4025 MLX5_ASSERT(set_tag->id > REG_NON);
4026 set_tag->data = flow_id;
4027 tag_action->conf = set_tag;
4028 /* Create Tx item list. */
4029 rte_memcpy(actions_tx, actions, sizeof(struct rte_flow_action));
4030 addr = (void *)&pattern_tx[2];
4032 item->type = (enum rte_flow_item_type)
4033 MLX5_RTE_FLOW_ITEM_TYPE_TAG;
4034 tag_item = (void *)addr;
4035 tag_item->data = flow_id;
4036 tag_item->id = mlx5_flow_get_reg_id(dev, MLX5_HAIRPIN_TX, 0, NULL);
4037 MLX5_ASSERT(set_tag->id > REG_NON);
4038 item->spec = tag_item;
4039 addr += sizeof(struct mlx5_rte_flow_item_tag);
4040 tag_item = (void *)addr;
4041 tag_item->data = UINT32_MAX;
4042 tag_item->id = UINT16_MAX;
4043 item->mask = tag_item;
4046 item->type = RTE_FLOW_ITEM_TYPE_END;
4051 * The last stage of splitting chain, just creates the subflow
4052 * without any modification.
4055 * Pointer to Ethernet device.
4057 * Parent flow structure pointer.
4058 * @param[in, out] sub_flow
4059 * Pointer to return the created subflow, may be NULL.
4061 * Flow rule attributes.
4063 * Pattern specification (list terminated by the END pattern item).
4064 * @param[in] actions
4065 * Associated actions (list terminated by the END action).
4066 * @param[in] flow_split_info
4067 * Pointer to flow split info structure.
4069 * Perform verbose error reporting if not NULL.
4071 * 0 on success, negative value otherwise
4074 flow_create_split_inner(struct rte_eth_dev *dev,
4075 struct rte_flow *flow,
4076 struct mlx5_flow **sub_flow,
4077 const struct rte_flow_attr *attr,
4078 const struct rte_flow_item items[],
4079 const struct rte_flow_action actions[],
4080 struct mlx5_flow_split_info *flow_split_info,
4081 struct rte_flow_error *error)
4083 struct mlx5_flow *dev_flow;
4085 dev_flow = flow_drv_prepare(dev, flow, attr, items, actions,
4086 flow_split_info->flow_idx, error);
4089 dev_flow->flow = flow;
4090 dev_flow->external = flow_split_info->external;
4091 dev_flow->skip_scale = flow_split_info->skip_scale;
4092 /* Subflow object was created, we must include one in the list. */
4093 SILIST_INSERT(&flow->dev_handles, dev_flow->handle_idx,
4094 dev_flow->handle, next);
4096 * If dev_flow is as one of the suffix flow, some actions in suffix
4097 * flow may need some user defined item layer flags, and pass the
4098 * Metadate rxq mark flag to suffix flow as well.
4100 if (flow_split_info->prefix_layers)
4101 dev_flow->handle->layers = flow_split_info->prefix_layers;
4102 if (flow_split_info->prefix_mark)
4103 dev_flow->handle->mark = 1;
4105 *sub_flow = dev_flow;
4106 return flow_drv_translate(dev, dev_flow, attr, items, actions, error);
4110 * Split the meter flow.
4112 * As meter flow will split to three sub flow, other than meter
4113 * action, the other actions make sense to only meter accepts
4114 * the packet. If it need to be dropped, no other additional
4115 * actions should be take.
4117 * One kind of special action which decapsulates the L3 tunnel
4118 * header will be in the prefix sub flow, as not to take the
4119 * L3 tunnel header into account.
4122 * Pointer to Ethernet device.
4124 * Pattern specification (list terminated by the END pattern item).
4125 * @param[out] sfx_items
4126 * Suffix flow match items (list terminated by the END pattern item).
4127 * @param[in] actions
4128 * Associated actions (list terminated by the END action).
4129 * @param[out] actions_sfx
4130 * Suffix flow actions.
4131 * @param[out] actions_pre
4132 * Prefix flow actions.
4133 * @param[out] pattern_sfx
4134 * The pattern items for the suffix flow.
4135 * @param[out] tag_sfx
4136 * Pointer to suffix flow tag.
4142 flow_meter_split_prep(struct rte_eth_dev *dev,
4143 const struct rte_flow_item items[],
4144 struct rte_flow_item sfx_items[],
4145 const struct rte_flow_action actions[],
4146 struct rte_flow_action actions_sfx[],
4147 struct rte_flow_action actions_pre[])
4149 struct mlx5_priv *priv = dev->data->dev_private;
4150 struct rte_flow_action *tag_action = NULL;
4151 struct rte_flow_item *tag_item;
4152 struct mlx5_rte_flow_action_set_tag *set_tag;
4153 struct rte_flow_error error;
4154 const struct rte_flow_action_raw_encap *raw_encap;
4155 const struct rte_flow_action_raw_decap *raw_decap;
4156 struct mlx5_rte_flow_item_tag *tag_spec;
4157 struct mlx5_rte_flow_item_tag *tag_mask;
4158 uint32_t tag_id = 0;
4159 bool copy_vlan = false;
4161 /* Prepare the actions for prefix and suffix flow. */
4162 for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
4163 struct rte_flow_action **action_cur = NULL;
4165 switch (actions->type) {
4166 case RTE_FLOW_ACTION_TYPE_METER:
4167 /* Add the extra tag action first. */
4168 tag_action = actions_pre;
4169 tag_action->type = (enum rte_flow_action_type)
4170 MLX5_RTE_FLOW_ACTION_TYPE_TAG;
4172 action_cur = &actions_pre;
4174 case RTE_FLOW_ACTION_TYPE_VXLAN_DECAP:
4175 case RTE_FLOW_ACTION_TYPE_NVGRE_DECAP:
4176 action_cur = &actions_pre;
4178 case RTE_FLOW_ACTION_TYPE_RAW_ENCAP:
4179 raw_encap = actions->conf;
4180 if (raw_encap->size < MLX5_ENCAPSULATION_DECISION_SIZE)
4181 action_cur = &actions_pre;
4183 case RTE_FLOW_ACTION_TYPE_RAW_DECAP:
4184 raw_decap = actions->conf;
4185 if (raw_decap->size > MLX5_ENCAPSULATION_DECISION_SIZE)
4186 action_cur = &actions_pre;
4188 case RTE_FLOW_ACTION_TYPE_OF_PUSH_VLAN:
4189 case RTE_FLOW_ACTION_TYPE_OF_SET_VLAN_VID:
4196 action_cur = &actions_sfx;
4197 memcpy(*action_cur, actions, sizeof(struct rte_flow_action));
4200 /* Add end action to the actions. */
4201 actions_sfx->type = RTE_FLOW_ACTION_TYPE_END;
4202 actions_pre->type = RTE_FLOW_ACTION_TYPE_END;
4205 set_tag = (void *)actions_pre;
4206 set_tag->id = mlx5_flow_get_reg_id(dev, MLX5_MTR_SFX, 0, &error);
4207 mlx5_ipool_malloc(priv->sh->ipool[MLX5_IPOOL_RSS_EXPANTION_FLOW_ID],
4209 if (tag_id >= (1 << (sizeof(tag_id) * 8 - MLX5_MTR_COLOR_BITS))) {
4210 DRV_LOG(ERR, "Port %u meter flow id exceed max limit.",
4211 dev->data->port_id);
4212 mlx5_ipool_free(priv->sh->ipool
4213 [MLX5_IPOOL_RSS_EXPANTION_FLOW_ID], tag_id);
4215 } else if (!tag_id) {
4218 set_tag->data = tag_id << MLX5_MTR_COLOR_BITS;
4220 tag_action->conf = set_tag;
4221 /* Prepare the suffix subflow items. */
4222 tag_item = sfx_items++;
4223 for (; items->type != RTE_FLOW_ITEM_TYPE_END; items++) {
4224 int item_type = items->type;
4226 switch (item_type) {
4227 case RTE_FLOW_ITEM_TYPE_PORT_ID:
4228 memcpy(sfx_items, items, sizeof(*sfx_items));
4231 case RTE_FLOW_ITEM_TYPE_VLAN:
4233 memcpy(sfx_items, items, sizeof(*sfx_items));
4235 * Convert to internal match item, it is used
4236 * for vlan push and set vid.
4238 sfx_items->type = (enum rte_flow_item_type)
4239 MLX5_RTE_FLOW_ITEM_TYPE_VLAN;
4247 sfx_items->type = RTE_FLOW_ITEM_TYPE_END;
4249 tag_spec = (struct mlx5_rte_flow_item_tag *)sfx_items;
4250 tag_spec->data = tag_id << MLX5_MTR_COLOR_BITS;
4251 tag_spec->id = mlx5_flow_get_reg_id(dev, MLX5_MTR_SFX, 0, &error);
4252 tag_mask = tag_spec + 1;
4253 tag_mask->data = 0xffffff00;
4254 tag_item->type = (enum rte_flow_item_type)
4255 MLX5_RTE_FLOW_ITEM_TYPE_TAG;
4256 tag_item->spec = tag_spec;
4257 tag_item->last = NULL;
4258 tag_item->mask = tag_mask;
4263 * Split action list having QUEUE/RSS for metadata register copy.
4265 * Once Q/RSS action is detected in user's action list, the flow action
4266 * should be split in order to copy metadata registers, which will happen in
4268 * - CQE->flow_tag := reg_c[1] (MARK)
4269 * - CQE->flow_table_metadata (reg_b) := reg_c[0] (META)
4270 * The Q/RSS action will be performed on RX_ACT_TBL after passing by RX_CP_TBL.
4271 * This is because the last action of each flow must be a terminal action
4272 * (QUEUE, RSS or DROP).
4274 * Flow ID must be allocated to identify actions in the RX_ACT_TBL and it is
4275 * stored and kept in the mlx5_flow structure per each sub_flow.
4277 * The Q/RSS action is replaced with,
4278 * - SET_TAG, setting the allocated flow ID to reg_c[2].
4279 * And the following JUMP action is added at the end,
4280 * - JUMP, to RX_CP_TBL.
4282 * A flow to perform remained Q/RSS action will be created in RX_ACT_TBL by
4283 * flow_create_split_metadata() routine. The flow will look like,
4284 * - If flow ID matches (reg_c[2]), perform Q/RSS.
4287 * Pointer to Ethernet device.
4288 * @param[out] split_actions
4289 * Pointer to store split actions to jump to CP_TBL.
4290 * @param[in] actions
4291 * Pointer to the list of original flow actions.
4293 * Pointer to the Q/RSS action.
4294 * @param[in] actions_n
4295 * Number of original actions.
4297 * Perform verbose error reporting if not NULL.
4300 * non-zero unique flow_id on success, otherwise 0 and
4301 * error/rte_error are set.
4304 flow_mreg_split_qrss_prep(struct rte_eth_dev *dev,
4305 struct rte_flow_action *split_actions,
4306 const struct rte_flow_action *actions,
4307 const struct rte_flow_action *qrss,
4308 int actions_n, struct rte_flow_error *error)
4310 struct mlx5_priv *priv = dev->data->dev_private;
4311 struct mlx5_rte_flow_action_set_tag *set_tag;
4312 struct rte_flow_action_jump *jump;
4313 const int qrss_idx = qrss - actions;
4314 uint32_t flow_id = 0;
4318 * Given actions will be split
4319 * - Replace QUEUE/RSS action with SET_TAG to set flow ID.
4320 * - Add jump to mreg CP_TBL.
4321 * As a result, there will be one more action.
4324 memcpy(split_actions, actions, sizeof(*split_actions) * actions_n);
4325 set_tag = (void *)(split_actions + actions_n);
4327 * If tag action is not set to void(it means we are not the meter
4328 * suffix flow), add the tag action. Since meter suffix flow already
4329 * has the tag added.
4331 if (split_actions[qrss_idx].type != RTE_FLOW_ACTION_TYPE_VOID) {
4333 * Allocate the new subflow ID. This one is unique within
4334 * device and not shared with representors. Otherwise,
4335 * we would have to resolve multi-thread access synch
4336 * issue. Each flow on the shared device is appended
4337 * with source vport identifier, so the resulting
4338 * flows will be unique in the shared (by master and
4339 * representors) domain even if they have coinciding
4342 mlx5_ipool_malloc(priv->sh->ipool
4343 [MLX5_IPOOL_RSS_EXPANTION_FLOW_ID], &flow_id);
4345 return rte_flow_error_set(error, ENOMEM,
4346 RTE_FLOW_ERROR_TYPE_ACTION,
4347 NULL, "can't allocate id "
4348 "for split Q/RSS subflow");
4349 /* Internal SET_TAG action to set flow ID. */
4350 *set_tag = (struct mlx5_rte_flow_action_set_tag){
4353 ret = mlx5_flow_get_reg_id(dev, MLX5_COPY_MARK, 0, error);
4357 /* Construct new actions array. */
4358 /* Replace QUEUE/RSS action. */
4359 split_actions[qrss_idx] = (struct rte_flow_action){
4360 .type = (enum rte_flow_action_type)
4361 MLX5_RTE_FLOW_ACTION_TYPE_TAG,
4365 /* JUMP action to jump to mreg copy table (CP_TBL). */
4366 jump = (void *)(set_tag + 1);
4367 *jump = (struct rte_flow_action_jump){
4368 .group = MLX5_FLOW_MREG_CP_TABLE_GROUP,
4370 split_actions[actions_n - 2] = (struct rte_flow_action){
4371 .type = RTE_FLOW_ACTION_TYPE_JUMP,
4374 split_actions[actions_n - 1] = (struct rte_flow_action){
4375 .type = RTE_FLOW_ACTION_TYPE_END,
4381 * Extend the given action list for Tx metadata copy.
4383 * Copy the given action list to the ext_actions and add flow metadata register
4384 * copy action in order to copy reg_a set by WQE to reg_c[0].
4386 * @param[out] ext_actions
4387 * Pointer to the extended action list.
4388 * @param[in] actions
4389 * Pointer to the list of actions.
4390 * @param[in] actions_n
4391 * Number of actions in the list.
4393 * Perform verbose error reporting if not NULL.
4394 * @param[in] encap_idx
4395 * The encap action inndex.
4398 * 0 on success, negative value otherwise
4401 flow_mreg_tx_copy_prep(struct rte_eth_dev *dev,
4402 struct rte_flow_action *ext_actions,
4403 const struct rte_flow_action *actions,
4404 int actions_n, struct rte_flow_error *error,
4407 struct mlx5_flow_action_copy_mreg *cp_mreg =
4408 (struct mlx5_flow_action_copy_mreg *)
4409 (ext_actions + actions_n + 1);
4412 ret = mlx5_flow_get_reg_id(dev, MLX5_METADATA_RX, 0, error);
4416 ret = mlx5_flow_get_reg_id(dev, MLX5_METADATA_TX, 0, error);
4421 memcpy(ext_actions, actions, sizeof(*ext_actions) * encap_idx);
4422 if (encap_idx == actions_n - 1) {
4423 ext_actions[actions_n - 1] = (struct rte_flow_action){
4424 .type = (enum rte_flow_action_type)
4425 MLX5_RTE_FLOW_ACTION_TYPE_COPY_MREG,
4428 ext_actions[actions_n] = (struct rte_flow_action){
4429 .type = RTE_FLOW_ACTION_TYPE_END,
4432 ext_actions[encap_idx] = (struct rte_flow_action){
4433 .type = (enum rte_flow_action_type)
4434 MLX5_RTE_FLOW_ACTION_TYPE_COPY_MREG,
4437 memcpy(ext_actions + encap_idx + 1, actions + encap_idx,
4438 sizeof(*ext_actions) * (actions_n - encap_idx));
4444 * Check the match action from the action list.
4446 * @param[in] actions
4447 * Pointer to the list of actions.
4449 * Flow rule attributes.
4451 * The action to be check if exist.
4452 * @param[out] match_action_pos
4453 * Pointer to the position of the matched action if exists, otherwise is -1.
4454 * @param[out] qrss_action_pos
4455 * Pointer to the position of the Queue/RSS action if exists, otherwise is -1.
4458 * > 0 the total number of actions.
4459 * 0 if not found match action in action list.
4462 flow_check_match_action(const struct rte_flow_action actions[],
4463 const struct rte_flow_attr *attr,
4464 enum rte_flow_action_type action,
4465 int *match_action_pos, int *qrss_action_pos)
4467 const struct rte_flow_action_sample *sample;
4474 *match_action_pos = -1;
4475 *qrss_action_pos = -1;
4476 for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
4477 if (actions->type == action) {
4479 *match_action_pos = actions_n;
4481 if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE ||
4482 actions->type == RTE_FLOW_ACTION_TYPE_RSS)
4483 *qrss_action_pos = actions_n;
4484 if (actions->type == RTE_FLOW_ACTION_TYPE_JUMP)
4486 if (actions->type == RTE_FLOW_ACTION_TYPE_SAMPLE) {
4487 sample = actions->conf;
4488 ratio = sample->ratio;
4489 sub_type = ((const struct rte_flow_action *)
4490 (sample->actions))->type;
4494 if (flag && action == RTE_FLOW_ACTION_TYPE_SAMPLE && attr->transfer) {
4496 /* JUMP Action not support for Mirroring;
4497 * Mirroring support multi-destination;
4499 if (!jump_flag && sub_type != RTE_FLOW_ACTION_TYPE_END)
4503 /* Count RTE_FLOW_ACTION_TYPE_END. */
4504 return flag ? actions_n + 1 : 0;
4507 #define SAMPLE_SUFFIX_ITEM 2
4510 * Split the sample flow.
4512 * As sample flow will split to two sub flow, sample flow with
4513 * sample action, the other actions will move to new suffix flow.
4515 * Also add unique tag id with tag action in the sample flow,
4516 * the same tag id will be as match in the suffix flow.
4519 * Pointer to Ethernet device.
4521 * FDB egress flow flag.
4522 * @param[out] sfx_items
4523 * Suffix flow match items (list terminated by the END pattern item).
4524 * @param[in] actions
4525 * Associated actions (list terminated by the END action).
4526 * @param[out] actions_sfx
4527 * Suffix flow actions.
4528 * @param[out] actions_pre
4529 * Prefix flow actions.
4530 * @param[in] actions_n
4531 * The total number of actions.
4532 * @param[in] sample_action_pos
4533 * The sample action position.
4534 * @param[in] qrss_action_pos
4535 * The Queue/RSS action position.
4537 * Perform verbose error reporting if not NULL.
4540 * 0 on success, or unique flow_id, a negative errno value
4541 * otherwise and rte_errno is set.
4544 flow_sample_split_prep(struct rte_eth_dev *dev,
4546 struct rte_flow_item sfx_items[],
4547 const struct rte_flow_action actions[],
4548 struct rte_flow_action actions_sfx[],
4549 struct rte_flow_action actions_pre[],
4551 int sample_action_pos,
4552 int qrss_action_pos,
4553 struct rte_flow_error *error)
4555 struct mlx5_priv *priv = dev->data->dev_private;
4556 struct mlx5_rte_flow_action_set_tag *set_tag;
4557 struct mlx5_rte_flow_item_tag *tag_spec;
4558 struct mlx5_rte_flow_item_tag *tag_mask;
4559 uint32_t tag_id = 0;
4563 if (sample_action_pos < 0)
4564 return rte_flow_error_set(error, EINVAL,
4565 RTE_FLOW_ERROR_TYPE_ACTION,
4566 NULL, "invalid position of sample "
4569 /* Prepare the prefix tag action. */
4570 set_tag = (void *)(actions_pre + actions_n + 1);
4571 ret = mlx5_flow_get_reg_id(dev, MLX5_APP_TAG, 0, error);
4575 mlx5_ipool_malloc(priv->sh->ipool
4576 [MLX5_IPOOL_RSS_EXPANTION_FLOW_ID], &tag_id);
4577 set_tag->data = tag_id;
4578 /* Prepare the suffix subflow items. */
4579 tag_spec = (void *)(sfx_items + SAMPLE_SUFFIX_ITEM);
4580 tag_spec->data = tag_id;
4581 tag_spec->id = set_tag->id;
4582 tag_mask = tag_spec + 1;
4583 tag_mask->data = UINT32_MAX;
4584 sfx_items[0] = (struct rte_flow_item){
4585 .type = (enum rte_flow_item_type)
4586 MLX5_RTE_FLOW_ITEM_TYPE_TAG,
4591 sfx_items[1] = (struct rte_flow_item){
4592 .type = (enum rte_flow_item_type)
4593 RTE_FLOW_ITEM_TYPE_END,
4596 /* Prepare the actions for prefix and suffix flow. */
4597 if (qrss_action_pos >= 0 && qrss_action_pos < sample_action_pos) {
4598 index = qrss_action_pos;
4599 /* Put the preceding the Queue/RSS action into prefix flow. */
4601 memcpy(actions_pre, actions,
4602 sizeof(struct rte_flow_action) * index);
4603 /* Put others preceding the sample action into prefix flow. */
4604 if (sample_action_pos > index + 1)
4605 memcpy(actions_pre + index, actions + index + 1,
4606 sizeof(struct rte_flow_action) *
4607 (sample_action_pos - index - 1));
4608 index = sample_action_pos - 1;
4609 /* Put Queue/RSS action into Suffix flow. */
4610 memcpy(actions_sfx, actions + qrss_action_pos,
4611 sizeof(struct rte_flow_action));
4614 index = sample_action_pos;
4616 memcpy(actions_pre, actions,
4617 sizeof(struct rte_flow_action) * index);
4619 /* Add the extra tag action for NIC-RX and E-Switch ingress. */
4621 actions_pre[index++] =
4622 (struct rte_flow_action){
4623 .type = (enum rte_flow_action_type)
4624 MLX5_RTE_FLOW_ACTION_TYPE_TAG,
4628 memcpy(actions_pre + index, actions + sample_action_pos,
4629 sizeof(struct rte_flow_action));
4631 actions_pre[index] = (struct rte_flow_action){
4632 .type = (enum rte_flow_action_type)
4633 RTE_FLOW_ACTION_TYPE_END,
4635 /* Put the actions after sample into Suffix flow. */
4636 memcpy(actions_sfx, actions + sample_action_pos + 1,
4637 sizeof(struct rte_flow_action) *
4638 (actions_n - sample_action_pos - 1));
4643 * The splitting for metadata feature.
4645 * - Q/RSS action on NIC Rx should be split in order to pass by
4646 * the mreg copy table (RX_CP_TBL) and then it jumps to the
4647 * action table (RX_ACT_TBL) which has the split Q/RSS action.
4649 * - All the actions on NIC Tx should have a mreg copy action to
4650 * copy reg_a from WQE to reg_c[0].
4653 * Pointer to Ethernet device.
4655 * Parent flow structure pointer.
4657 * Flow rule attributes.
4659 * Pattern specification (list terminated by the END pattern item).
4660 * @param[in] actions
4661 * Associated actions (list terminated by the END action).
4662 * @param[in] flow_split_info
4663 * Pointer to flow split info structure.
4665 * Perform verbose error reporting if not NULL.
4667 * 0 on success, negative value otherwise
4670 flow_create_split_metadata(struct rte_eth_dev *dev,
4671 struct rte_flow *flow,
4672 const struct rte_flow_attr *attr,
4673 const struct rte_flow_item items[],
4674 const struct rte_flow_action actions[],
4675 struct mlx5_flow_split_info *flow_split_info,
4676 struct rte_flow_error *error)
4678 struct mlx5_priv *priv = dev->data->dev_private;
4679 struct mlx5_dev_config *config = &priv->config;
4680 const struct rte_flow_action *qrss = NULL;
4681 struct rte_flow_action *ext_actions = NULL;
4682 struct mlx5_flow *dev_flow = NULL;
4683 uint32_t qrss_id = 0;
4690 /* Check whether extensive metadata feature is engaged. */
4691 if (!config->dv_flow_en ||
4692 config->dv_xmeta_en == MLX5_XMETA_MODE_LEGACY ||
4693 !mlx5_flow_ext_mreg_supported(dev))
4694 return flow_create_split_inner(dev, flow, NULL, attr, items,
4695 actions, flow_split_info, error);
4696 actions_n = flow_parse_metadata_split_actions_info(actions, &qrss,
4699 /* Exclude hairpin flows from splitting. */
4700 if (qrss->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
4701 const struct rte_flow_action_queue *queue;
4704 if (mlx5_rxq_get_type(dev, queue->index) ==
4705 MLX5_RXQ_TYPE_HAIRPIN)
4707 } else if (qrss->type == RTE_FLOW_ACTION_TYPE_RSS) {
4708 const struct rte_flow_action_rss *rss;
4711 if (mlx5_rxq_get_type(dev, rss->queue[0]) ==
4712 MLX5_RXQ_TYPE_HAIRPIN)
4717 /* Check if it is in meter suffix table. */
4718 mtr_sfx = attr->group == (attr->transfer ?
4719 (MLX5_FLOW_TABLE_LEVEL_SUFFIX - 1) :
4720 MLX5_FLOW_TABLE_LEVEL_SUFFIX);
4722 * Q/RSS action on NIC Rx should be split in order to pass by
4723 * the mreg copy table (RX_CP_TBL) and then it jumps to the
4724 * action table (RX_ACT_TBL) which has the split Q/RSS action.
4726 act_size = sizeof(struct rte_flow_action) * (actions_n + 1) +
4727 sizeof(struct rte_flow_action_set_tag) +
4728 sizeof(struct rte_flow_action_jump);
4729 ext_actions = mlx5_malloc(MLX5_MEM_ZERO, act_size, 0,
4732 return rte_flow_error_set(error, ENOMEM,
4733 RTE_FLOW_ERROR_TYPE_ACTION,
4734 NULL, "no memory to split "
4737 * If we are the suffix flow of meter, tag already exist.
4738 * Set the tag action to void.
4741 ext_actions[qrss - actions].type =
4742 RTE_FLOW_ACTION_TYPE_VOID;
4744 ext_actions[qrss - actions].type =
4745 (enum rte_flow_action_type)
4746 MLX5_RTE_FLOW_ACTION_TYPE_TAG;
4748 * Create the new actions list with removed Q/RSS action
4749 * and appended set tag and jump to register copy table
4750 * (RX_CP_TBL). We should preallocate unique tag ID here
4751 * in advance, because it is needed for set tag action.
4753 qrss_id = flow_mreg_split_qrss_prep(dev, ext_actions, actions,
4754 qrss, actions_n, error);
4755 if (!mtr_sfx && !qrss_id) {
4759 } else if (attr->egress && !attr->transfer) {
4761 * All the actions on NIC Tx should have a metadata register
4762 * copy action to copy reg_a from WQE to reg_c[meta]
4764 act_size = sizeof(struct rte_flow_action) * (actions_n + 1) +
4765 sizeof(struct mlx5_flow_action_copy_mreg);
4766 ext_actions = mlx5_malloc(MLX5_MEM_ZERO, act_size, 0,
4769 return rte_flow_error_set(error, ENOMEM,
4770 RTE_FLOW_ERROR_TYPE_ACTION,
4771 NULL, "no memory to split "
4773 /* Create the action list appended with copy register. */
4774 ret = flow_mreg_tx_copy_prep(dev, ext_actions, actions,
4775 actions_n, error, encap_idx);
4779 /* Add the unmodified original or prefix subflow. */
4780 ret = flow_create_split_inner(dev, flow, &dev_flow, attr,
4781 items, ext_actions ? ext_actions :
4782 actions, flow_split_info, error);
4785 MLX5_ASSERT(dev_flow);
4787 const struct rte_flow_attr q_attr = {
4788 .group = MLX5_FLOW_MREG_ACT_TABLE_GROUP,
4791 /* Internal PMD action to set register. */
4792 struct mlx5_rte_flow_item_tag q_tag_spec = {
4796 struct rte_flow_item q_items[] = {
4798 .type = (enum rte_flow_item_type)
4799 MLX5_RTE_FLOW_ITEM_TYPE_TAG,
4800 .spec = &q_tag_spec,
4805 .type = RTE_FLOW_ITEM_TYPE_END,
4808 struct rte_flow_action q_actions[] = {
4814 .type = RTE_FLOW_ACTION_TYPE_END,
4817 uint64_t layers = flow_get_prefix_layer_flags(dev_flow);
4820 * Configure the tag item only if there is no meter subflow.
4821 * Since tag is already marked in the meter suffix subflow
4822 * we can just use the meter suffix items as is.
4825 /* Not meter subflow. */
4826 MLX5_ASSERT(!mtr_sfx);
4828 * Put unique id in prefix flow due to it is destroyed
4829 * after suffix flow and id will be freed after there
4830 * is no actual flows with this id and identifier
4831 * reallocation becomes possible (for example, for
4832 * other flows in other threads).
4834 dev_flow->handle->split_flow_id = qrss_id;
4835 ret = mlx5_flow_get_reg_id(dev, MLX5_COPY_MARK, 0,
4839 q_tag_spec.id = ret;
4842 /* Add suffix subflow to execute Q/RSS. */
4843 flow_split_info->prefix_layers = layers;
4844 flow_split_info->prefix_mark = 0;
4845 ret = flow_create_split_inner(dev, flow, &dev_flow,
4846 &q_attr, mtr_sfx ? items :
4848 flow_split_info, error);
4851 /* qrss ID should be freed if failed. */
4853 MLX5_ASSERT(dev_flow);
4858 * We do not destroy the partially created sub_flows in case of error.
4859 * These ones are included into parent flow list and will be destroyed
4860 * by flow_drv_destroy.
4862 mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_RSS_EXPANTION_FLOW_ID],
4864 mlx5_free(ext_actions);
4869 * The splitting for meter feature.
4871 * - The meter flow will be split to two flows as prefix and
4872 * suffix flow. The packets make sense only it pass the prefix
4875 * - Reg_C_5 is used for the packet to match betweend prefix and
4879 * Pointer to Ethernet device.
4881 * Parent flow structure pointer.
4883 * Flow rule attributes.
4885 * Pattern specification (list terminated by the END pattern item).
4886 * @param[in] actions
4887 * Associated actions (list terminated by the END action).
4888 * @param[in] flow_split_info
4889 * Pointer to flow split info structure.
4891 * Perform verbose error reporting if not NULL.
4893 * 0 on success, negative value otherwise
4896 flow_create_split_meter(struct rte_eth_dev *dev,
4897 struct rte_flow *flow,
4898 const struct rte_flow_attr *attr,
4899 const struct rte_flow_item items[],
4900 const struct rte_flow_action actions[],
4901 struct mlx5_flow_split_info *flow_split_info,
4902 struct rte_flow_error *error)
4904 struct mlx5_priv *priv = dev->data->dev_private;
4905 struct rte_flow_action *sfx_actions = NULL;
4906 struct rte_flow_action *pre_actions = NULL;
4907 struct rte_flow_item *sfx_items = NULL;
4908 struct mlx5_flow *dev_flow = NULL;
4909 struct rte_flow_attr sfx_attr = *attr;
4911 uint32_t mtr_tag_id = 0;
4918 actions_n = flow_check_meter_action(actions, &mtr);
4920 /* The five prefix actions: meter, decap, encap, tag, end. */
4921 act_size = sizeof(struct rte_flow_action) * (actions_n + 5) +
4922 sizeof(struct mlx5_rte_flow_action_set_tag);
4923 /* tag, vlan, port id, end. */
4924 #define METER_SUFFIX_ITEM 4
4925 item_size = sizeof(struct rte_flow_item) * METER_SUFFIX_ITEM +
4926 sizeof(struct mlx5_rte_flow_item_tag) * 2;
4927 sfx_actions = mlx5_malloc(MLX5_MEM_ZERO, (act_size + item_size),
4930 return rte_flow_error_set(error, ENOMEM,
4931 RTE_FLOW_ERROR_TYPE_ACTION,
4932 NULL, "no memory to split "
4934 sfx_items = (struct rte_flow_item *)((char *)sfx_actions +
4936 pre_actions = sfx_actions + actions_n;
4937 mtr_tag_id = flow_meter_split_prep(dev, items, sfx_items,
4938 actions, sfx_actions,
4944 /* Add the prefix subflow. */
4945 flow_split_info->prefix_mark = 0;
4946 ret = flow_create_split_inner(dev, flow, &dev_flow,
4947 attr, items, pre_actions,
4948 flow_split_info, error);
4953 dev_flow->handle->split_flow_id = mtr_tag_id;
4954 /* Setting the sfx group atrr. */
4955 sfx_attr.group = sfx_attr.transfer ?
4956 (MLX5_FLOW_TABLE_LEVEL_SUFFIX - 1) :
4957 MLX5_FLOW_TABLE_LEVEL_SUFFIX;
4958 flow_split_info->prefix_layers =
4959 flow_get_prefix_layer_flags(dev_flow);
4960 flow_split_info->prefix_mark = dev_flow->handle->mark;
4962 /* Add the prefix subflow. */
4963 ret = flow_create_split_metadata(dev, flow,
4964 &sfx_attr, sfx_items ?
4966 sfx_actions ? sfx_actions : actions,
4967 flow_split_info, error);
4970 mlx5_free(sfx_actions);
4975 * The splitting for sample feature.
4977 * Once Sample action is detected in the action list, the flow actions should
4978 * be split into prefix sub flow and suffix sub flow.
4980 * The original items remain in the prefix sub flow, all actions preceding the
4981 * sample action and the sample action itself will be copied to the prefix
4982 * sub flow, the actions following the sample action will be copied to the
4983 * suffix sub flow, Queue action always be located in the suffix sub flow.
4985 * In order to make the packet from prefix sub flow matches with suffix sub
4986 * flow, an extra tag action be added into prefix sub flow, and the suffix sub
4987 * flow uses tag item with the unique flow id.
4990 * Pointer to Ethernet device.
4992 * Parent flow structure pointer.
4994 * Flow rule attributes.
4996 * Pattern specification (list terminated by the END pattern item).
4997 * @param[in] actions
4998 * Associated actions (list terminated by the END action).
4999 * @param[in] flow_split_info
5000 * Pointer to flow split info structure.
5002 * Perform verbose error reporting if not NULL.
5004 * 0 on success, negative value otherwise
5007 flow_create_split_sample(struct rte_eth_dev *dev,
5008 struct rte_flow *flow,
5009 const struct rte_flow_attr *attr,
5010 const struct rte_flow_item items[],
5011 const struct rte_flow_action actions[],
5012 struct mlx5_flow_split_info *flow_split_info,
5013 struct rte_flow_error *error)
5015 struct mlx5_priv *priv = dev->data->dev_private;
5016 struct rte_flow_action *sfx_actions = NULL;
5017 struct rte_flow_action *pre_actions = NULL;
5018 struct rte_flow_item *sfx_items = NULL;
5019 struct mlx5_flow *dev_flow = NULL;
5020 struct rte_flow_attr sfx_attr = *attr;
5021 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
5022 struct mlx5_flow_dv_sample_resource *sample_res;
5023 struct mlx5_flow_tbl_data_entry *sfx_tbl_data;
5024 struct mlx5_flow_tbl_resource *sfx_tbl;
5025 union mlx5_flow_tbl_key sfx_table_key;
5029 uint32_t fdb_tx = 0;
5032 int sample_action_pos;
5033 int qrss_action_pos;
5036 if (priv->sampler_en)
5037 actions_n = flow_check_match_action(actions, attr,
5038 RTE_FLOW_ACTION_TYPE_SAMPLE,
5039 &sample_action_pos, &qrss_action_pos);
5041 /* The prefix actions must includes sample, tag, end. */
5042 act_size = sizeof(struct rte_flow_action) * (actions_n * 2 + 1)
5043 + sizeof(struct mlx5_rte_flow_action_set_tag);
5044 item_size = sizeof(struct rte_flow_item) * SAMPLE_SUFFIX_ITEM +
5045 sizeof(struct mlx5_rte_flow_item_tag) * 2;
5046 sfx_actions = mlx5_malloc(MLX5_MEM_ZERO, (act_size +
5047 item_size), 0, SOCKET_ID_ANY);
5049 return rte_flow_error_set(error, ENOMEM,
5050 RTE_FLOW_ERROR_TYPE_ACTION,
5051 NULL, "no memory to split "
5053 /* The representor_id is -1 for uplink. */
5054 fdb_tx = (attr->transfer && priv->representor_id != -1);
5056 sfx_items = (struct rte_flow_item *)((char *)sfx_actions
5058 pre_actions = sfx_actions + actions_n;
5059 tag_id = flow_sample_split_prep(dev, fdb_tx, sfx_items,
5060 actions, sfx_actions,
5061 pre_actions, actions_n,
5063 qrss_action_pos, error);
5064 if (tag_id < 0 || (!fdb_tx && !tag_id)) {
5068 /* Add the prefix subflow. */
5069 ret = flow_create_split_inner(dev, flow, &dev_flow, attr,
5071 flow_split_info, error);
5076 dev_flow->handle->split_flow_id = tag_id;
5077 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
5078 /* Set the sfx group attr. */
5079 sample_res = (struct mlx5_flow_dv_sample_resource *)
5080 dev_flow->dv.sample_res;
5081 sfx_tbl = (struct mlx5_flow_tbl_resource *)
5082 sample_res->normal_path_tbl;
5083 sfx_tbl_data = container_of(sfx_tbl,
5084 struct mlx5_flow_tbl_data_entry, tbl);
5085 sfx_table_key.v64 = sfx_tbl_data->entry.key;
5086 sfx_attr.group = sfx_attr.transfer ?
5087 (sfx_table_key.table_id - 1) :
5088 sfx_table_key.table_id;
5089 flow_split_info->prefix_layers =
5090 flow_get_prefix_layer_flags(dev_flow);
5091 flow_split_info->prefix_mark = dev_flow->handle->mark;
5092 /* Suffix group level already be scaled with factor, set
5093 * skip_scale to 1 to avoid scale again in translation.
5095 flow_split_info->skip_scale = 1;
5098 /* Add the suffix subflow. */
5099 ret = flow_create_split_meter(dev, flow, &sfx_attr,
5100 sfx_items ? sfx_items : items,
5101 sfx_actions ? sfx_actions : actions,
5102 flow_split_info, error);
5105 mlx5_free(sfx_actions);
5110 * Split the flow to subflow set. The splitters might be linked
5111 * in the chain, like this:
5112 * flow_create_split_outer() calls:
5113 * flow_create_split_meter() calls:
5114 * flow_create_split_metadata(meter_subflow_0) calls:
5115 * flow_create_split_inner(metadata_subflow_0)
5116 * flow_create_split_inner(metadata_subflow_1)
5117 * flow_create_split_inner(metadata_subflow_2)
5118 * flow_create_split_metadata(meter_subflow_1) calls:
5119 * flow_create_split_inner(metadata_subflow_0)
5120 * flow_create_split_inner(metadata_subflow_1)
5121 * flow_create_split_inner(metadata_subflow_2)
5123 * This provide flexible way to add new levels of flow splitting.
5124 * The all of successfully created subflows are included to the
5125 * parent flow dev_flow list.
5128 * Pointer to Ethernet device.
5130 * Parent flow structure pointer.
5132 * Flow rule attributes.
5134 * Pattern specification (list terminated by the END pattern item).
5135 * @param[in] actions
5136 * Associated actions (list terminated by the END action).
5137 * @param[in] flow_split_info
5138 * Pointer to flow split info structure.
5140 * Perform verbose error reporting if not NULL.
5142 * 0 on success, negative value otherwise
5145 flow_create_split_outer(struct rte_eth_dev *dev,
5146 struct rte_flow *flow,
5147 const struct rte_flow_attr *attr,
5148 const struct rte_flow_item items[],
5149 const struct rte_flow_action actions[],
5150 struct mlx5_flow_split_info *flow_split_info,
5151 struct rte_flow_error *error)
5155 ret = flow_create_split_sample(dev, flow, attr, items,
5156 actions, flow_split_info, error);
5157 MLX5_ASSERT(ret <= 0);
5161 static struct mlx5_flow_tunnel *
5162 flow_tunnel_from_rule(struct rte_eth_dev *dev,
5163 const struct rte_flow_attr *attr,
5164 const struct rte_flow_item items[],
5165 const struct rte_flow_action actions[])
5167 struct mlx5_flow_tunnel *tunnel;
5169 #pragma GCC diagnostic push
5170 #pragma GCC diagnostic ignored "-Wcast-qual"
5171 if (is_flow_tunnel_match_rule(dev, attr, items, actions))
5172 tunnel = (struct mlx5_flow_tunnel *)items[0].spec;
5173 else if (is_flow_tunnel_steer_rule(dev, attr, items, actions))
5174 tunnel = (struct mlx5_flow_tunnel *)actions[0].conf;
5177 #pragma GCC diagnostic pop
5183 * Adjust flow RSS workspace if needed.
5186 * Pointer to thread flow work space.
5188 * Pointer to RSS descriptor.
5189 * @param[in] nrssq_num
5190 * New RSS queue number.
5193 * 0 on success, -1 otherwise and rte_errno is set.
5196 flow_rss_workspace_adjust(struct mlx5_flow_workspace *wks,
5197 struct mlx5_flow_rss_desc *rss_desc,
5200 if (likely(nrssq_num <= wks->rssq_num))
5202 rss_desc->queue = realloc(rss_desc->queue,
5203 sizeof(*rss_desc->queue) * RTE_ALIGN(nrssq_num, 2));
5204 if (!rss_desc->queue) {
5208 wks->rssq_num = RTE_ALIGN(nrssq_num, 2);
5213 * Create a flow and add it to @p list.
5216 * Pointer to Ethernet device.
5218 * Pointer to a TAILQ flow list. If this parameter NULL,
5219 * no list insertion occurred, flow is just created,
5220 * this is caller's responsibility to track the
5223 * Flow rule attributes.
5225 * Pattern specification (list terminated by the END pattern item).
5226 * @param[in] actions
5227 * Associated actions (list terminated by the END action).
5228 * @param[in] external
5229 * This flow rule is created by request external to PMD.
5231 * Perform verbose error reporting if not NULL.
5234 * A flow index on success, 0 otherwise and rte_errno is set.
5237 flow_list_create(struct rte_eth_dev *dev, uint32_t *list,
5238 const struct rte_flow_attr *attr,
5239 const struct rte_flow_item items[],
5240 const struct rte_flow_action original_actions[],
5241 bool external, struct rte_flow_error *error)
5243 struct mlx5_priv *priv = dev->data->dev_private;
5244 struct rte_flow *flow = NULL;
5245 struct mlx5_flow *dev_flow;
5246 const struct rte_flow_action_rss *rss;
5247 struct mlx5_translated_shared_action
5248 shared_actions[MLX5_MAX_SHARED_ACTIONS];
5249 int shared_actions_n = MLX5_MAX_SHARED_ACTIONS;
5251 struct mlx5_flow_expand_rss buf;
5252 uint8_t buffer[2048];
5255 struct rte_flow_action actions[MLX5_MAX_SPLIT_ACTIONS];
5256 uint8_t buffer[2048];
5259 struct rte_flow_action actions[MLX5_MAX_SPLIT_ACTIONS];
5260 uint8_t buffer[2048];
5261 } actions_hairpin_tx;
5263 struct rte_flow_item items[MLX5_MAX_SPLIT_ITEMS];
5264 uint8_t buffer[2048];
5266 struct mlx5_flow_expand_rss *buf = &expand_buffer.buf;
5267 struct mlx5_flow_rss_desc *rss_desc;
5268 const struct rte_flow_action *p_actions_rx;
5272 struct rte_flow_attr attr_tx = { .priority = 0 };
5273 const struct rte_flow_action *actions;
5274 struct rte_flow_action *translated_actions = NULL;
5275 struct mlx5_flow_tunnel *tunnel;
5276 struct tunnel_default_miss_ctx default_miss_ctx = { 0, };
5277 struct mlx5_flow_workspace *wks = mlx5_flow_push_thread_workspace();
5278 struct mlx5_flow_split_info flow_split_info = {
5279 .external = !!external,
5288 rss_desc = &wks->rss_desc;
5289 ret = flow_shared_actions_translate(dev, original_actions,
5292 &translated_actions, error);
5294 MLX5_ASSERT(translated_actions == NULL);
5297 actions = translated_actions ? translated_actions : original_actions;
5298 p_actions_rx = actions;
5299 hairpin_flow = flow_check_hairpin_split(dev, attr, actions);
5300 ret = flow_drv_validate(dev, attr, items, p_actions_rx,
5301 external, hairpin_flow, error);
5303 goto error_before_hairpin_split;
5304 flow = mlx5_ipool_zmalloc(priv->sh->ipool[MLX5_IPOOL_RTE_FLOW], &idx);
5307 goto error_before_hairpin_split;
5309 if (hairpin_flow > 0) {
5310 if (hairpin_flow > MLX5_MAX_SPLIT_ACTIONS) {
5312 goto error_before_hairpin_split;
5314 flow_hairpin_split(dev, actions, actions_rx.actions,
5315 actions_hairpin_tx.actions, items_tx.items,
5317 p_actions_rx = actions_rx.actions;
5319 flow_split_info.flow_idx = idx;
5320 flow->drv_type = flow_get_drv_type(dev, attr);
5321 MLX5_ASSERT(flow->drv_type > MLX5_FLOW_TYPE_MIN &&
5322 flow->drv_type < MLX5_FLOW_TYPE_MAX);
5323 memset(rss_desc, 0, offsetof(struct mlx5_flow_rss_desc, queue));
5324 rss = flow_get_rss_action(p_actions_rx);
5326 if (flow_rss_workspace_adjust(wks, rss_desc, rss->queue_num))
5329 * The following information is required by
5330 * mlx5_flow_hashfields_adjust() in advance.
5332 rss_desc->level = rss->level;
5333 /* RSS type 0 indicates default RSS type (ETH_RSS_IP). */
5334 rss_desc->types = !rss->types ? ETH_RSS_IP : rss->types;
5336 flow->dev_handles = 0;
5337 if (rss && rss->types) {
5338 unsigned int graph_root;
5340 graph_root = find_graph_root(items, rss->level);
5341 ret = mlx5_flow_expand_rss(buf, sizeof(expand_buffer.buffer),
5343 mlx5_support_expansion, graph_root);
5344 MLX5_ASSERT(ret > 0 &&
5345 (unsigned int)ret < sizeof(expand_buffer.buffer));
5348 buf->entry[0].pattern = (void *)(uintptr_t)items;
5350 rss_desc->shared_rss = flow_get_shared_rss_action(dev, shared_actions,
5352 for (i = 0; i < buf->entries; ++i) {
5353 /* Initialize flow split data. */
5354 flow_split_info.prefix_layers = 0;
5355 flow_split_info.prefix_mark = 0;
5356 flow_split_info.skip_scale = 0;
5358 * The splitter may create multiple dev_flows,
5359 * depending on configuration. In the simplest
5360 * case it just creates unmodified original flow.
5362 ret = flow_create_split_outer(dev, flow, attr,
5363 buf->entry[i].pattern,
5364 p_actions_rx, &flow_split_info,
5368 if (is_flow_tunnel_steer_rule(dev, attr,
5369 buf->entry[i].pattern,
5371 ret = flow_tunnel_add_default_miss(dev, flow, attr,
5377 mlx5_free(default_miss_ctx.queue);
5382 /* Create the tx flow. */
5384 attr_tx.group = MLX5_HAIRPIN_TX_TABLE;
5385 attr_tx.ingress = 0;
5387 dev_flow = flow_drv_prepare(dev, flow, &attr_tx, items_tx.items,
5388 actions_hairpin_tx.actions,
5392 dev_flow->flow = flow;
5393 dev_flow->external = 0;
5394 SILIST_INSERT(&flow->dev_handles, dev_flow->handle_idx,
5395 dev_flow->handle, next);
5396 ret = flow_drv_translate(dev, dev_flow, &attr_tx,
5398 actions_hairpin_tx.actions, error);
5403 * Update the metadata register copy table. If extensive
5404 * metadata feature is enabled and registers are supported
5405 * we might create the extra rte_flow for each unique
5406 * MARK/FLAG action ID.
5408 * The table is updated for ingress Flows only, because
5409 * the egress Flows belong to the different device and
5410 * copy table should be updated in peer NIC Rx domain.
5412 if (attr->ingress &&
5413 (external || attr->group != MLX5_FLOW_MREG_CP_TABLE_GROUP)) {
5414 ret = flow_mreg_update_copy_table(dev, flow, actions, error);
5419 * If the flow is external (from application) OR device is started,
5420 * OR mreg discover, then apply immediately.
5422 if (external || dev->data->dev_started ||
5423 (attr->group == MLX5_FLOW_MREG_CP_TABLE_GROUP &&
5424 attr->priority == MLX5_FLOW_PRIO_RSVD)) {
5425 ret = flow_drv_apply(dev, flow, error);
5430 rte_spinlock_lock(&priv->flow_list_lock);
5431 ILIST_INSERT(priv->sh->ipool[MLX5_IPOOL_RTE_FLOW], list, idx,
5433 rte_spinlock_unlock(&priv->flow_list_lock);
5435 flow_rxq_flags_set(dev, flow);
5436 rte_free(translated_actions);
5437 tunnel = flow_tunnel_from_rule(dev, attr, items, actions);
5440 flow->tunnel_id = tunnel->tunnel_id;
5441 __atomic_add_fetch(&tunnel->refctn, 1, __ATOMIC_RELAXED);
5442 mlx5_free(default_miss_ctx.queue);
5444 mlx5_flow_pop_thread_workspace();
5448 ret = rte_errno; /* Save rte_errno before cleanup. */
5449 flow_mreg_del_copy_action(dev, flow);
5450 flow_drv_destroy(dev, flow);
5451 if (rss_desc->shared_rss)
5452 __atomic_sub_fetch(&((struct mlx5_shared_action_rss *)
5454 (priv->sh->ipool[MLX5_IPOOL_RSS_SHARED_ACTIONS],
5455 rss_desc->shared_rss))->refcnt, 1, __ATOMIC_RELAXED);
5456 mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_RTE_FLOW], idx);
5457 rte_errno = ret; /* Restore rte_errno. */
5460 mlx5_flow_pop_thread_workspace();
5461 error_before_hairpin_split:
5462 rte_free(translated_actions);
5467 * Create a dedicated flow rule on e-switch table 0 (root table), to direct all
5468 * incoming packets to table 1.
5470 * Other flow rules, requested for group n, will be created in
5471 * e-switch table n+1.
5472 * Jump action to e-switch group n will be created to group n+1.
5474 * Used when working in switchdev mode, to utilise advantages of table 1
5478 * Pointer to Ethernet device.
5481 * Pointer to flow on success, NULL otherwise and rte_errno is set.
5484 mlx5_flow_create_esw_table_zero_flow(struct rte_eth_dev *dev)
5486 const struct rte_flow_attr attr = {
5493 const struct rte_flow_item pattern = {
5494 .type = RTE_FLOW_ITEM_TYPE_END,
5496 struct rte_flow_action_jump jump = {
5499 const struct rte_flow_action actions[] = {
5501 .type = RTE_FLOW_ACTION_TYPE_JUMP,
5505 .type = RTE_FLOW_ACTION_TYPE_END,
5508 struct mlx5_priv *priv = dev->data->dev_private;
5509 struct rte_flow_error error;
5511 return (void *)(uintptr_t)flow_list_create(dev, &priv->ctrl_flows,
5513 actions, false, &error);
5517 * Validate a flow supported by the NIC.
5519 * @see rte_flow_validate()
5523 mlx5_flow_validate(struct rte_eth_dev *dev,
5524 const struct rte_flow_attr *attr,
5525 const struct rte_flow_item items[],
5526 const struct rte_flow_action original_actions[],
5527 struct rte_flow_error *error)
5530 struct mlx5_translated_shared_action
5531 shared_actions[MLX5_MAX_SHARED_ACTIONS];
5532 int shared_actions_n = MLX5_MAX_SHARED_ACTIONS;
5533 const struct rte_flow_action *actions;
5534 struct rte_flow_action *translated_actions = NULL;
5535 int ret = flow_shared_actions_translate(dev, original_actions,
5538 &translated_actions, error);
5542 actions = translated_actions ? translated_actions : original_actions;
5543 hairpin_flow = flow_check_hairpin_split(dev, attr, actions);
5544 ret = flow_drv_validate(dev, attr, items, actions,
5545 true, hairpin_flow, error);
5546 rte_free(translated_actions);
5553 * @see rte_flow_create()
5557 mlx5_flow_create(struct rte_eth_dev *dev,
5558 const struct rte_flow_attr *attr,
5559 const struct rte_flow_item items[],
5560 const struct rte_flow_action actions[],
5561 struct rte_flow_error *error)
5563 struct mlx5_priv *priv = dev->data->dev_private;
5566 * If the device is not started yet, it is not allowed to created a
5567 * flow from application. PMD default flows and traffic control flows
5570 if (unlikely(!dev->data->dev_started)) {
5571 DRV_LOG(DEBUG, "port %u is not started when "
5572 "inserting a flow", dev->data->port_id);
5573 rte_flow_error_set(error, ENODEV,
5574 RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
5576 "port not started");
5580 return (void *)(uintptr_t)flow_list_create(dev, &priv->flows,
5581 attr, items, actions, true, error);
5585 * Destroy a flow in a list.
5588 * Pointer to Ethernet device.
5590 * Pointer to the Indexed flow list. If this parameter NULL,
5591 * there is no flow removal from the list. Be noted that as
5592 * flow is add to the indexed list, memory of the indexed
5593 * list points to maybe changed as flow destroyed.
5594 * @param[in] flow_idx
5595 * Index of flow to destroy.
5598 flow_list_destroy(struct rte_eth_dev *dev, uint32_t *list,
5601 struct mlx5_priv *priv = dev->data->dev_private;
5602 struct rte_flow *flow = mlx5_ipool_get(priv->sh->ipool
5603 [MLX5_IPOOL_RTE_FLOW], flow_idx);
5608 * Update RX queue flags only if port is started, otherwise it is
5611 if (dev->data->dev_started)
5612 flow_rxq_flags_trim(dev, flow);
5613 flow_drv_destroy(dev, flow);
5615 rte_spinlock_lock(&priv->flow_list_lock);
5616 ILIST_REMOVE(priv->sh->ipool[MLX5_IPOOL_RTE_FLOW], list,
5617 flow_idx, flow, next);
5618 rte_spinlock_unlock(&priv->flow_list_lock);
5621 struct mlx5_flow_tunnel *tunnel;
5623 tunnel = mlx5_find_tunnel_id(dev, flow->tunnel_id);
5625 if (!__atomic_sub_fetch(&tunnel->refctn, 1, __ATOMIC_RELAXED))
5626 mlx5_flow_tunnel_free(dev, tunnel);
5628 flow_mreg_del_copy_action(dev, flow);
5629 mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_RTE_FLOW], flow_idx);
5633 * Destroy all flows.
5636 * Pointer to Ethernet device.
5638 * Pointer to the Indexed flow list.
5640 * If flushing is called avtively.
5643 mlx5_flow_list_flush(struct rte_eth_dev *dev, uint32_t *list, bool active)
5645 uint32_t num_flushed = 0;
5648 flow_list_destroy(dev, list, *list);
5652 DRV_LOG(INFO, "port %u: %u flows flushed before stopping",
5653 dev->data->port_id, num_flushed);
5658 * Stop all default actions for flows.
5661 * Pointer to Ethernet device.
5664 mlx5_flow_stop_default(struct rte_eth_dev *dev)
5666 flow_mreg_del_default_copy_action(dev);
5667 flow_rxq_flags_clear(dev);
5671 * Start all default actions for flows.
5674 * Pointer to Ethernet device.
5676 * 0 on success, a negative errno value otherwise and rte_errno is set.
5679 mlx5_flow_start_default(struct rte_eth_dev *dev)
5681 struct rte_flow_error error;
5683 /* Make sure default copy action (reg_c[0] -> reg_b) is created. */
5684 return flow_mreg_add_default_copy_action(dev, &error);
5688 * Release key of thread specific flow workspace data.
5691 flow_release_workspace(void *data)
5693 struct mlx5_flow_workspace *wks = data;
5694 struct mlx5_flow_workspace *next;
5698 free(wks->rss_desc.queue);
5705 * Initialize key of thread specific flow workspace data.
5708 flow_alloc_workspace(void)
5710 if (pthread_key_create(&key_workspace, flow_release_workspace))
5711 DRV_LOG(ERR, "Can't create flow workspace data thread key.");
5715 * Get thread specific current flow workspace.
5717 * @return pointer to thread specific flow workspace data, NULL on error.
5719 struct mlx5_flow_workspace*
5720 mlx5_flow_get_thread_workspace(void)
5722 struct mlx5_flow_workspace *data;
5724 data = pthread_getspecific(key_workspace);
5725 MLX5_ASSERT(data && data->inuse);
5726 if (!data || !data->inuse)
5727 DRV_LOG(ERR, "flow workspace not initialized.");
5732 * Allocate and init new flow workspace.
5734 * @return pointer to flow workspace data, NULL on error.
5736 static struct mlx5_flow_workspace*
5737 flow_alloc_thread_workspace(void)
5739 struct mlx5_flow_workspace *data = calloc(1, sizeof(*data));
5742 DRV_LOG(ERR, "Failed to allocate flow workspace "
5746 data->rss_desc.queue = calloc(1,
5747 sizeof(uint16_t) * MLX5_RSSQ_DEFAULT_NUM);
5748 if (!data->rss_desc.queue)
5750 data->rssq_num = MLX5_RSSQ_DEFAULT_NUM;
5753 if (data->rss_desc.queue)
5754 free(data->rss_desc.queue);
5760 * Get new thread specific flow workspace.
5762 * If current workspace inuse, create new one and set as current.
5764 * @return pointer to thread specific flow workspace data, NULL on error.
5766 static struct mlx5_flow_workspace*
5767 mlx5_flow_push_thread_workspace(void)
5769 struct mlx5_flow_workspace *curr;
5770 struct mlx5_flow_workspace *data;
5772 if (pthread_once(&key_workspace_init, flow_alloc_workspace)) {
5773 DRV_LOG(ERR, "Failed to init flow workspace data thread key.");
5776 curr = pthread_getspecific(key_workspace);
5778 data = flow_alloc_thread_workspace();
5781 } else if (!curr->inuse) {
5783 } else if (curr->next) {
5786 data = flow_alloc_thread_workspace();
5794 /* Set as current workspace */
5795 if (pthread_setspecific(key_workspace, data))
5796 DRV_LOG(ERR, "Failed to set flow workspace to thread.");
5801 * Close current thread specific flow workspace.
5803 * If previous workspace available, set it as current.
5805 * @return pointer to thread specific flow workspace data, NULL on error.
5808 mlx5_flow_pop_thread_workspace(void)
5810 struct mlx5_flow_workspace *data = mlx5_flow_get_thread_workspace();
5815 DRV_LOG(ERR, "Failed to close unused flow workspace.");
5821 if (pthread_setspecific(key_workspace, data->prev))
5822 DRV_LOG(ERR, "Failed to set flow workspace to thread.");
5826 * Verify the flow list is empty
5829 * Pointer to Ethernet device.
5831 * @return the number of flows not released.
5834 mlx5_flow_verify(struct rte_eth_dev *dev)
5836 struct mlx5_priv *priv = dev->data->dev_private;
5837 struct rte_flow *flow;
5841 ILIST_FOREACH(priv->sh->ipool[MLX5_IPOOL_RTE_FLOW], priv->flows, idx,
5843 DRV_LOG(DEBUG, "port %u flow %p still referenced",
5844 dev->data->port_id, (void *)flow);
5851 * Enable default hairpin egress flow.
5854 * Pointer to Ethernet device.
5859 * 0 on success, a negative errno value otherwise and rte_errno is set.
5862 mlx5_ctrl_flow_source_queue(struct rte_eth_dev *dev,
5865 struct mlx5_priv *priv = dev->data->dev_private;
5866 const struct rte_flow_attr attr = {
5870 struct mlx5_rte_flow_item_tx_queue queue_spec = {
5873 struct mlx5_rte_flow_item_tx_queue queue_mask = {
5874 .queue = UINT32_MAX,
5876 struct rte_flow_item items[] = {
5878 .type = (enum rte_flow_item_type)
5879 MLX5_RTE_FLOW_ITEM_TYPE_TX_QUEUE,
5880 .spec = &queue_spec,
5882 .mask = &queue_mask,
5885 .type = RTE_FLOW_ITEM_TYPE_END,
5888 struct rte_flow_action_jump jump = {
5889 .group = MLX5_HAIRPIN_TX_TABLE,
5891 struct rte_flow_action actions[2];
5893 struct rte_flow_error error;
5895 actions[0].type = RTE_FLOW_ACTION_TYPE_JUMP;
5896 actions[0].conf = &jump;
5897 actions[1].type = RTE_FLOW_ACTION_TYPE_END;
5898 flow_idx = flow_list_create(dev, &priv->ctrl_flows,
5899 &attr, items, actions, false, &error);
5902 "Failed to create ctrl flow: rte_errno(%d),"
5903 " type(%d), message(%s)",
5904 rte_errno, error.type,
5905 error.message ? error.message : " (no stated reason)");
5912 * Enable a control flow configured from the control plane.
5915 * Pointer to Ethernet device.
5917 * An Ethernet flow spec to apply.
5919 * An Ethernet flow mask to apply.
5921 * A VLAN flow spec to apply.
5923 * A VLAN flow mask to apply.
5926 * 0 on success, a negative errno value otherwise and rte_errno is set.
5929 mlx5_ctrl_flow_vlan(struct rte_eth_dev *dev,
5930 struct rte_flow_item_eth *eth_spec,
5931 struct rte_flow_item_eth *eth_mask,
5932 struct rte_flow_item_vlan *vlan_spec,
5933 struct rte_flow_item_vlan *vlan_mask)
5935 struct mlx5_priv *priv = dev->data->dev_private;
5936 const struct rte_flow_attr attr = {
5938 .priority = MLX5_FLOW_PRIO_RSVD,
5940 struct rte_flow_item items[] = {
5942 .type = RTE_FLOW_ITEM_TYPE_ETH,
5948 .type = (vlan_spec) ? RTE_FLOW_ITEM_TYPE_VLAN :
5949 RTE_FLOW_ITEM_TYPE_END,
5955 .type = RTE_FLOW_ITEM_TYPE_END,
5958 uint16_t queue[priv->reta_idx_n];
5959 struct rte_flow_action_rss action_rss = {
5960 .func = RTE_ETH_HASH_FUNCTION_DEFAULT,
5962 .types = priv->rss_conf.rss_hf,
5963 .key_len = priv->rss_conf.rss_key_len,
5964 .queue_num = priv->reta_idx_n,
5965 .key = priv->rss_conf.rss_key,
5968 struct rte_flow_action actions[] = {
5970 .type = RTE_FLOW_ACTION_TYPE_RSS,
5971 .conf = &action_rss,
5974 .type = RTE_FLOW_ACTION_TYPE_END,
5978 struct rte_flow_error error;
5981 if (!priv->reta_idx_n || !priv->rxqs_n) {
5984 if (!(dev->data->dev_conf.rxmode.mq_mode & ETH_MQ_RX_RSS_FLAG))
5985 action_rss.types = 0;
5986 for (i = 0; i != priv->reta_idx_n; ++i)
5987 queue[i] = (*priv->reta_idx)[i];
5988 flow_idx = flow_list_create(dev, &priv->ctrl_flows,
5989 &attr, items, actions, false, &error);
5996 * Enable a flow control configured from the control plane.
5999 * Pointer to Ethernet device.
6001 * An Ethernet flow spec to apply.
6003 * An Ethernet flow mask to apply.
6006 * 0 on success, a negative errno value otherwise and rte_errno is set.
6009 mlx5_ctrl_flow(struct rte_eth_dev *dev,
6010 struct rte_flow_item_eth *eth_spec,
6011 struct rte_flow_item_eth *eth_mask)
6013 return mlx5_ctrl_flow_vlan(dev, eth_spec, eth_mask, NULL, NULL);
6017 * Create default miss flow rule matching lacp traffic
6020 * Pointer to Ethernet device.
6022 * An Ethernet flow spec to apply.
6025 * 0 on success, a negative errno value otherwise and rte_errno is set.
6028 mlx5_flow_lacp_miss(struct rte_eth_dev *dev)
6030 struct mlx5_priv *priv = dev->data->dev_private;
6032 * The LACP matching is done by only using ether type since using
6033 * a multicast dst mac causes kernel to give low priority to this flow.
6035 static const struct rte_flow_item_eth lacp_spec = {
6036 .type = RTE_BE16(0x8809),
6038 static const struct rte_flow_item_eth lacp_mask = {
6041 const struct rte_flow_attr attr = {
6044 struct rte_flow_item items[] = {
6046 .type = RTE_FLOW_ITEM_TYPE_ETH,
6051 .type = RTE_FLOW_ITEM_TYPE_END,
6054 struct rte_flow_action actions[] = {
6056 .type = (enum rte_flow_action_type)
6057 MLX5_RTE_FLOW_ACTION_TYPE_DEFAULT_MISS,
6060 .type = RTE_FLOW_ACTION_TYPE_END,
6063 struct rte_flow_error error;
6064 uint32_t flow_idx = flow_list_create(dev, &priv->ctrl_flows,
6065 &attr, items, actions, false, &error);
6075 * @see rte_flow_destroy()
6079 mlx5_flow_destroy(struct rte_eth_dev *dev,
6080 struct rte_flow *flow,
6081 struct rte_flow_error *error __rte_unused)
6083 struct mlx5_priv *priv = dev->data->dev_private;
6085 flow_list_destroy(dev, &priv->flows, (uintptr_t)(void *)flow);
6090 * Destroy all flows.
6092 * @see rte_flow_flush()
6096 mlx5_flow_flush(struct rte_eth_dev *dev,
6097 struct rte_flow_error *error __rte_unused)
6099 struct mlx5_priv *priv = dev->data->dev_private;
6101 mlx5_flow_list_flush(dev, &priv->flows, false);
6108 * @see rte_flow_isolate()
6112 mlx5_flow_isolate(struct rte_eth_dev *dev,
6114 struct rte_flow_error *error)
6116 struct mlx5_priv *priv = dev->data->dev_private;
6118 if (dev->data->dev_started) {
6119 rte_flow_error_set(error, EBUSY,
6120 RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
6122 "port must be stopped first");
6125 priv->isolated = !!enable;
6127 dev->dev_ops = &mlx5_os_dev_ops_isolate;
6129 dev->dev_ops = &mlx5_os_dev_ops;
6131 dev->rx_descriptor_status = mlx5_rx_descriptor_status;
6132 dev->tx_descriptor_status = mlx5_tx_descriptor_status;
6140 * @see rte_flow_query()
6144 flow_drv_query(struct rte_eth_dev *dev,
6146 const struct rte_flow_action *actions,
6148 struct rte_flow_error *error)
6150 struct mlx5_priv *priv = dev->data->dev_private;
6151 const struct mlx5_flow_driver_ops *fops;
6152 struct rte_flow *flow = mlx5_ipool_get(priv->sh->ipool
6153 [MLX5_IPOOL_RTE_FLOW],
6155 enum mlx5_flow_drv_type ftype;
6158 return rte_flow_error_set(error, ENOENT,
6159 RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
6161 "invalid flow handle");
6163 ftype = flow->drv_type;
6164 MLX5_ASSERT(ftype > MLX5_FLOW_TYPE_MIN && ftype < MLX5_FLOW_TYPE_MAX);
6165 fops = flow_get_drv_ops(ftype);
6167 return fops->query(dev, flow, actions, data, error);
6173 * @see rte_flow_query()
6177 mlx5_flow_query(struct rte_eth_dev *dev,
6178 struct rte_flow *flow,
6179 const struct rte_flow_action *actions,
6181 struct rte_flow_error *error)
6185 ret = flow_drv_query(dev, (uintptr_t)(void *)flow, actions, data,
6193 * Manage filter operations.
6196 * Pointer to Ethernet device structure.
6197 * @param filter_type
6200 * Operation to perform.
6202 * Pointer to operation-specific structure.
6205 * 0 on success, a negative errno value otherwise and rte_errno is set.
6208 mlx5_dev_filter_ctrl(struct rte_eth_dev *dev,
6209 enum rte_filter_type filter_type,
6210 enum rte_filter_op filter_op,
6213 switch (filter_type) {
6214 case RTE_ETH_FILTER_GENERIC:
6215 if (filter_op != RTE_ETH_FILTER_GET) {
6219 *(const void **)arg = &mlx5_flow_ops;
6222 DRV_LOG(ERR, "port %u filter type (%d) not supported",
6223 dev->data->port_id, filter_type);
6224 rte_errno = ENOTSUP;
6231 * Create the needed meter and suffix tables.
6234 * Pointer to Ethernet device.
6236 * Pointer to the flow meter.
6239 * Pointer to table set on success, NULL otherwise.
6241 struct mlx5_meter_domains_infos *
6242 mlx5_flow_create_mtr_tbls(struct rte_eth_dev *dev,
6243 const struct mlx5_flow_meter *fm)
6245 const struct mlx5_flow_driver_ops *fops;
6247 fops = flow_get_drv_ops(MLX5_FLOW_TYPE_DV);
6248 return fops->create_mtr_tbls(dev, fm);
6252 * Destroy the meter table set.
6255 * Pointer to Ethernet device.
6257 * Pointer to the meter table set.
6263 mlx5_flow_destroy_mtr_tbls(struct rte_eth_dev *dev,
6264 struct mlx5_meter_domains_infos *tbls)
6266 const struct mlx5_flow_driver_ops *fops;
6268 fops = flow_get_drv_ops(MLX5_FLOW_TYPE_DV);
6269 return fops->destroy_mtr_tbls(dev, tbls);
6273 * Create policer rules.
6276 * Pointer to Ethernet device.
6278 * Pointer to flow meter structure.
6280 * Pointer to flow attributes.
6283 * 0 on success, -1 otherwise.
6286 mlx5_flow_create_policer_rules(struct rte_eth_dev *dev,
6287 struct mlx5_flow_meter *fm,
6288 const struct rte_flow_attr *attr)
6290 const struct mlx5_flow_driver_ops *fops;
6292 fops = flow_get_drv_ops(MLX5_FLOW_TYPE_DV);
6293 return fops->create_policer_rules(dev, fm, attr);
6297 * Destroy policer rules.
6300 * Pointer to flow meter structure.
6302 * Pointer to flow attributes.
6305 * 0 on success, -1 otherwise.
6308 mlx5_flow_destroy_policer_rules(struct rte_eth_dev *dev,
6309 struct mlx5_flow_meter *fm,
6310 const struct rte_flow_attr *attr)
6312 const struct mlx5_flow_driver_ops *fops;
6314 fops = flow_get_drv_ops(MLX5_FLOW_TYPE_DV);
6315 return fops->destroy_policer_rules(dev, fm, attr);
6319 * Allocate a counter.
6322 * Pointer to Ethernet device structure.
6325 * Index to allocated counter on success, 0 otherwise.
6328 mlx5_counter_alloc(struct rte_eth_dev *dev)
6330 const struct mlx5_flow_driver_ops *fops;
6331 struct rte_flow_attr attr = { .transfer = 0 };
6333 if (flow_get_drv_type(dev, &attr) == MLX5_FLOW_TYPE_DV) {
6334 fops = flow_get_drv_ops(MLX5_FLOW_TYPE_DV);
6335 return fops->counter_alloc(dev);
6338 "port %u counter allocate is not supported.",
6339 dev->data->port_id);
6347 * Pointer to Ethernet device structure.
6349 * Index to counter to be free.
6352 mlx5_counter_free(struct rte_eth_dev *dev, uint32_t cnt)
6354 const struct mlx5_flow_driver_ops *fops;
6355 struct rte_flow_attr attr = { .transfer = 0 };
6357 if (flow_get_drv_type(dev, &attr) == MLX5_FLOW_TYPE_DV) {
6358 fops = flow_get_drv_ops(MLX5_FLOW_TYPE_DV);
6359 fops->counter_free(dev, cnt);
6363 "port %u counter free is not supported.",
6364 dev->data->port_id);
6368 * Query counter statistics.
6371 * Pointer to Ethernet device structure.
6373 * Index to counter to query.
6375 * Set to clear counter statistics.
6377 * The counter hits packets number to save.
6379 * The counter hits bytes number to save.
6382 * 0 on success, a negative errno value otherwise.
6385 mlx5_counter_query(struct rte_eth_dev *dev, uint32_t cnt,
6386 bool clear, uint64_t *pkts, uint64_t *bytes)
6388 const struct mlx5_flow_driver_ops *fops;
6389 struct rte_flow_attr attr = { .transfer = 0 };
6391 if (flow_get_drv_type(dev, &attr) == MLX5_FLOW_TYPE_DV) {
6392 fops = flow_get_drv_ops(MLX5_FLOW_TYPE_DV);
6393 return fops->counter_query(dev, cnt, clear, pkts, bytes);
6396 "port %u counter query is not supported.",
6397 dev->data->port_id);
6402 * Allocate a new memory for the counter values wrapped by all the needed
6406 * Pointer to mlx5_dev_ctx_shared object.
6409 * 0 on success, a negative errno value otherwise.
6412 mlx5_flow_create_counter_stat_mem_mng(struct mlx5_dev_ctx_shared *sh)
6414 struct mlx5_devx_mkey_attr mkey_attr;
6415 struct mlx5_counter_stats_mem_mng *mem_mng;
6416 volatile struct flow_counter_stats *raw_data;
6417 int raws_n = MLX5_CNT_CONTAINER_RESIZE + MLX5_MAX_PENDING_QUERIES;
6418 int size = (sizeof(struct flow_counter_stats) *
6419 MLX5_COUNTERS_PER_POOL +
6420 sizeof(struct mlx5_counter_stats_raw)) * raws_n +
6421 sizeof(struct mlx5_counter_stats_mem_mng);
6422 size_t pgsize = rte_mem_page_size();
6426 if (pgsize == (size_t)-1) {
6427 DRV_LOG(ERR, "Failed to get mem page size");
6431 mem = mlx5_malloc(MLX5_MEM_ZERO, size, pgsize, SOCKET_ID_ANY);
6436 mem_mng = (struct mlx5_counter_stats_mem_mng *)(mem + size) - 1;
6437 size = sizeof(*raw_data) * MLX5_COUNTERS_PER_POOL * raws_n;
6438 mem_mng->umem = mlx5_glue->devx_umem_reg(sh->ctx, mem, size,
6439 IBV_ACCESS_LOCAL_WRITE);
6440 if (!mem_mng->umem) {
6445 mkey_attr.addr = (uintptr_t)mem;
6446 mkey_attr.size = size;
6447 mkey_attr.umem_id = mlx5_os_get_umem_id(mem_mng->umem);
6448 mkey_attr.pd = sh->pdn;
6449 mkey_attr.log_entity_size = 0;
6450 mkey_attr.pg_access = 0;
6451 mkey_attr.klm_array = NULL;
6452 mkey_attr.klm_num = 0;
6453 mkey_attr.relaxed_ordering_write = sh->cmng.relaxed_ordering_write;
6454 mkey_attr.relaxed_ordering_read = sh->cmng.relaxed_ordering_read;
6455 mem_mng->dm = mlx5_devx_cmd_mkey_create(sh->ctx, &mkey_attr);
6457 mlx5_glue->devx_umem_dereg(mem_mng->umem);
6462 mem_mng->raws = (struct mlx5_counter_stats_raw *)(mem + size);
6463 raw_data = (volatile struct flow_counter_stats *)mem;
6464 for (i = 0; i < raws_n; ++i) {
6465 mem_mng->raws[i].mem_mng = mem_mng;
6466 mem_mng->raws[i].data = raw_data + i * MLX5_COUNTERS_PER_POOL;
6468 for (i = 0; i < MLX5_MAX_PENDING_QUERIES; ++i)
6469 LIST_INSERT_HEAD(&sh->cmng.free_stat_raws,
6470 mem_mng->raws + MLX5_CNT_CONTAINER_RESIZE + i,
6472 LIST_INSERT_HEAD(&sh->cmng.mem_mngs, mem_mng, next);
6473 sh->cmng.mem_mng = mem_mng;
6478 * Set the statistic memory to the new counter pool.
6481 * Pointer to mlx5_dev_ctx_shared object.
6483 * Pointer to the pool to set the statistic memory.
6486 * 0 on success, a negative errno value otherwise.
6489 mlx5_flow_set_counter_stat_mem(struct mlx5_dev_ctx_shared *sh,
6490 struct mlx5_flow_counter_pool *pool)
6492 struct mlx5_flow_counter_mng *cmng = &sh->cmng;
6493 /* Resize statistic memory once used out. */
6494 if (!(pool->index % MLX5_CNT_CONTAINER_RESIZE) &&
6495 mlx5_flow_create_counter_stat_mem_mng(sh)) {
6496 DRV_LOG(ERR, "Cannot resize counter stat mem.");
6499 rte_spinlock_lock(&pool->sl);
6500 pool->raw = cmng->mem_mng->raws + pool->index %
6501 MLX5_CNT_CONTAINER_RESIZE;
6502 rte_spinlock_unlock(&pool->sl);
6503 pool->raw_hw = NULL;
6507 #define MLX5_POOL_QUERY_FREQ_US 1000000
6510 * Set the periodic procedure for triggering asynchronous batch queries for all
6511 * the counter pools.
6514 * Pointer to mlx5_dev_ctx_shared object.
6517 mlx5_set_query_alarm(struct mlx5_dev_ctx_shared *sh)
6519 uint32_t pools_n, us;
6521 pools_n = __atomic_load_n(&sh->cmng.n_valid, __ATOMIC_RELAXED);
6522 us = MLX5_POOL_QUERY_FREQ_US / pools_n;
6523 DRV_LOG(DEBUG, "Set alarm for %u pools each %u us", pools_n, us);
6524 if (rte_eal_alarm_set(us, mlx5_flow_query_alarm, sh)) {
6525 sh->cmng.query_thread_on = 0;
6526 DRV_LOG(ERR, "Cannot reinitialize query alarm");
6528 sh->cmng.query_thread_on = 1;
6533 * The periodic procedure for triggering asynchronous batch queries for all the
6534 * counter pools. This function is probably called by the host thread.
6537 * The parameter for the alarm process.
6540 mlx5_flow_query_alarm(void *arg)
6542 struct mlx5_dev_ctx_shared *sh = arg;
6544 uint16_t pool_index = sh->cmng.pool_index;
6545 struct mlx5_flow_counter_mng *cmng = &sh->cmng;
6546 struct mlx5_flow_counter_pool *pool;
6549 if (sh->cmng.pending_queries >= MLX5_MAX_PENDING_QUERIES)
6551 rte_spinlock_lock(&cmng->pool_update_sl);
6552 pool = cmng->pools[pool_index];
6553 n_valid = cmng->n_valid;
6554 rte_spinlock_unlock(&cmng->pool_update_sl);
6555 /* Set the statistic memory to the new created pool. */
6556 if ((!pool->raw && mlx5_flow_set_counter_stat_mem(sh, pool)))
6559 /* There is a pool query in progress. */
6562 LIST_FIRST(&sh->cmng.free_stat_raws);
6564 /* No free counter statistics raw memory. */
6567 * Identify the counters released between query trigger and query
6568 * handle more efficiently. The counter released in this gap period
6569 * should wait for a new round of query as the new arrived packets
6570 * will not be taken into account.
6573 ret = mlx5_devx_cmd_flow_counter_query(pool->min_dcs, 0,
6574 MLX5_COUNTERS_PER_POOL,
6576 pool->raw_hw->mem_mng->dm->id,
6580 (uint64_t)(uintptr_t)pool);
6582 DRV_LOG(ERR, "Failed to trigger asynchronous query for dcs ID"
6583 " %d", pool->min_dcs->id);
6584 pool->raw_hw = NULL;
6587 LIST_REMOVE(pool->raw_hw, next);
6588 sh->cmng.pending_queries++;
6590 if (pool_index >= n_valid)
6593 sh->cmng.pool_index = pool_index;
6594 mlx5_set_query_alarm(sh);
6598 * Check and callback event for new aged flow in the counter pool
6601 * Pointer to mlx5_dev_ctx_shared object.
6603 * Pointer to Current counter pool.
6606 mlx5_flow_aging_check(struct mlx5_dev_ctx_shared *sh,
6607 struct mlx5_flow_counter_pool *pool)
6609 struct mlx5_priv *priv;
6610 struct mlx5_flow_counter *cnt;
6611 struct mlx5_age_info *age_info;
6612 struct mlx5_age_param *age_param;
6613 struct mlx5_counter_stats_raw *cur = pool->raw_hw;
6614 struct mlx5_counter_stats_raw *prev = pool->raw;
6615 const uint64_t curr_time = MLX5_CURR_TIME_SEC;
6616 const uint32_t time_delta = curr_time - pool->time_of_last_age_check;
6617 uint16_t expected = AGE_CANDIDATE;
6620 pool->time_of_last_age_check = curr_time;
6621 for (i = 0; i < MLX5_COUNTERS_PER_POOL; ++i) {
6622 cnt = MLX5_POOL_GET_CNT(pool, i);
6623 age_param = MLX5_CNT_TO_AGE(cnt);
6624 if (__atomic_load_n(&age_param->state,
6625 __ATOMIC_RELAXED) != AGE_CANDIDATE)
6627 if (cur->data[i].hits != prev->data[i].hits) {
6628 __atomic_store_n(&age_param->sec_since_last_hit, 0,
6632 if (__atomic_add_fetch(&age_param->sec_since_last_hit,
6634 __ATOMIC_RELAXED) <= age_param->timeout)
6637 * Hold the lock first, or if between the
6638 * state AGE_TMOUT and tailq operation the
6639 * release happened, the release procedure
6640 * may delete a non-existent tailq node.
6642 priv = rte_eth_devices[age_param->port_id].data->dev_private;
6643 age_info = GET_PORT_AGE_INFO(priv);
6644 rte_spinlock_lock(&age_info->aged_sl);
6645 if (__atomic_compare_exchange_n(&age_param->state, &expected,
6648 __ATOMIC_RELAXED)) {
6649 TAILQ_INSERT_TAIL(&age_info->aged_counters, cnt, next);
6650 MLX5_AGE_SET(age_info, MLX5_AGE_EVENT_NEW);
6652 rte_spinlock_unlock(&age_info->aged_sl);
6654 mlx5_age_event_prepare(sh);
6658 * Handler for the HW respond about ready values from an asynchronous batch
6659 * query. This function is probably called by the host thread.
6662 * The pointer to the shared device context.
6663 * @param[in] async_id
6664 * The Devx async ID.
6666 * The status of the completion.
6669 mlx5_flow_async_pool_query_handle(struct mlx5_dev_ctx_shared *sh,
6670 uint64_t async_id, int status)
6672 struct mlx5_flow_counter_pool *pool =
6673 (struct mlx5_flow_counter_pool *)(uintptr_t)async_id;
6674 struct mlx5_counter_stats_raw *raw_to_free;
6675 uint8_t query_gen = pool->query_gen ^ 1;
6676 struct mlx5_flow_counter_mng *cmng = &sh->cmng;
6677 enum mlx5_counter_type cnt_type =
6678 pool->is_aged ? MLX5_COUNTER_TYPE_AGE :
6679 MLX5_COUNTER_TYPE_ORIGIN;
6681 if (unlikely(status)) {
6682 raw_to_free = pool->raw_hw;
6684 raw_to_free = pool->raw;
6686 mlx5_flow_aging_check(sh, pool);
6687 rte_spinlock_lock(&pool->sl);
6688 pool->raw = pool->raw_hw;
6689 rte_spinlock_unlock(&pool->sl);
6690 /* Be sure the new raw counters data is updated in memory. */
6692 if (!TAILQ_EMPTY(&pool->counters[query_gen])) {
6693 rte_spinlock_lock(&cmng->csl[cnt_type]);
6694 TAILQ_CONCAT(&cmng->counters[cnt_type],
6695 &pool->counters[query_gen], next);
6696 rte_spinlock_unlock(&cmng->csl[cnt_type]);
6699 LIST_INSERT_HEAD(&sh->cmng.free_stat_raws, raw_to_free, next);
6700 pool->raw_hw = NULL;
6701 sh->cmng.pending_queries--;
6705 flow_group_to_table(uint32_t port_id, uint32_t group, uint32_t *table,
6706 const struct flow_grp_info *grp_info,
6707 struct rte_flow_error *error)
6709 if (grp_info->transfer && grp_info->external &&
6710 grp_info->fdb_def_rule) {
6711 if (group == UINT32_MAX)
6712 return rte_flow_error_set
6714 RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
6716 "group index not supported");
6721 DRV_LOG(DEBUG, "port %u group=%#x table=%#x", port_id, group, *table);
6726 * Translate the rte_flow group index to HW table value.
6728 * If tunnel offload is disabled, all group ids converted to flow table
6729 * id using the standard method.
6730 * If tunnel offload is enabled, group id can be converted using the
6731 * standard or tunnel conversion method. Group conversion method
6732 * selection depends on flags in `grp_info` parameter:
6733 * - Internal (grp_info.external == 0) groups conversion uses the
6735 * - Group ids in JUMP action converted with the tunnel conversion.
6736 * - Group id in rule attribute conversion depends on a rule type and
6738 * ** non zero group attributes converted with the tunnel method
6739 * ** zero group attribute in non-tunnel rule is converted using the
6740 * standard method - there's only one root table
6741 * ** zero group attribute in steer tunnel rule is converted with the
6742 * standard method - single root table
6743 * ** zero group attribute in match tunnel rule is a special OvS
6744 * case: that value is used for portability reasons. That group
6745 * id is converted with the tunnel conversion method.
6750 * PMD tunnel offload object
6752 * rte_flow group index value.
6755 * @param[in] grp_info
6756 * flags used for conversion
6758 * Pointer to error structure.
6761 * 0 on success, a negative errno value otherwise and rte_errno is set.
6764 mlx5_flow_group_to_table(struct rte_eth_dev *dev,
6765 const struct mlx5_flow_tunnel *tunnel,
6766 uint32_t group, uint32_t *table,
6767 const struct flow_grp_info *grp_info,
6768 struct rte_flow_error *error)
6771 bool standard_translation;
6773 if (!grp_info->skip_scale && grp_info->external &&
6774 group < MLX5_MAX_TABLES_EXTERNAL)
6775 group *= MLX5_FLOW_TABLE_FACTOR;
6776 if (is_tunnel_offload_active(dev)) {
6777 standard_translation = !grp_info->external ||
6778 grp_info->std_tbl_fix;
6780 standard_translation = true;
6783 "port %u group=%u transfer=%d external=%d fdb_def_rule=%d translate=%s",
6784 dev->data->port_id, group, grp_info->transfer,
6785 grp_info->external, grp_info->fdb_def_rule,
6786 standard_translation ? "STANDARD" : "TUNNEL");
6787 if (standard_translation)
6788 ret = flow_group_to_table(dev->data->port_id, group, table,
6791 ret = tunnel_flow_group_to_flow_table(dev, tunnel, group,
6798 * Discover availability of metadata reg_c's.
6800 * Iteratively use test flows to check availability.
6803 * Pointer to the Ethernet device structure.
6806 * 0 on success, a negative errno value otherwise and rte_errno is set.
6809 mlx5_flow_discover_mreg_c(struct rte_eth_dev *dev)
6811 struct mlx5_priv *priv = dev->data->dev_private;
6812 struct mlx5_dev_config *config = &priv->config;
6813 enum modify_reg idx;
6816 /* reg_c[0] and reg_c[1] are reserved. */
6817 config->flow_mreg_c[n++] = REG_C_0;
6818 config->flow_mreg_c[n++] = REG_C_1;
6819 /* Discover availability of other reg_c's. */
6820 for (idx = REG_C_2; idx <= REG_C_7; ++idx) {
6821 struct rte_flow_attr attr = {
6822 .group = MLX5_FLOW_MREG_CP_TABLE_GROUP,
6823 .priority = MLX5_FLOW_PRIO_RSVD,
6826 struct rte_flow_item items[] = {
6828 .type = RTE_FLOW_ITEM_TYPE_END,
6831 struct rte_flow_action actions[] = {
6833 .type = (enum rte_flow_action_type)
6834 MLX5_RTE_FLOW_ACTION_TYPE_COPY_MREG,
6835 .conf = &(struct mlx5_flow_action_copy_mreg){
6841 .type = RTE_FLOW_ACTION_TYPE_JUMP,
6842 .conf = &(struct rte_flow_action_jump){
6843 .group = MLX5_FLOW_MREG_ACT_TABLE_GROUP,
6847 .type = RTE_FLOW_ACTION_TYPE_END,
6851 struct rte_flow *flow;
6852 struct rte_flow_error error;
6854 if (!config->dv_flow_en)
6856 /* Create internal flow, validation skips copy action. */
6857 flow_idx = flow_list_create(dev, NULL, &attr, items,
6858 actions, false, &error);
6859 flow = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_RTE_FLOW],
6863 config->flow_mreg_c[n++] = idx;
6864 flow_list_destroy(dev, NULL, flow_idx);
6866 for (; n < MLX5_MREG_C_NUM; ++n)
6867 config->flow_mreg_c[n] = REG_NON;
6872 * Dump flow raw hw data to file
6875 * The pointer to Ethernet device.
6877 * A pointer to a file for output.
6879 * Perform verbose error reporting if not NULL. PMDs initialize this
6880 * structure in case of error only.
6882 * 0 on success, a nagative value otherwise.
6885 mlx5_flow_dev_dump(struct rte_eth_dev *dev,
6887 struct rte_flow_error *error __rte_unused)
6889 struct mlx5_priv *priv = dev->data->dev_private;
6890 struct mlx5_dev_ctx_shared *sh = priv->sh;
6892 if (!priv->config.dv_flow_en) {
6893 if (fputs("device dv flow disabled\n", file) <= 0)
6897 return mlx5_devx_cmd_flow_dump(sh->fdb_domain, sh->rx_domain,
6898 sh->tx_domain, file);
6902 * Get aged-out flows.
6905 * Pointer to the Ethernet device structure.
6906 * @param[in] context
6907 * The address of an array of pointers to the aged-out flows contexts.
6908 * @param[in] nb_countexts
6909 * The length of context array pointers.
6911 * Perform verbose error reporting if not NULL. Initialized in case of
6915 * how many contexts get in success, otherwise negative errno value.
6916 * if nb_contexts is 0, return the amount of all aged contexts.
6917 * if nb_contexts is not 0 , return the amount of aged flows reported
6918 * in the context array.
6921 mlx5_flow_get_aged_flows(struct rte_eth_dev *dev, void **contexts,
6922 uint32_t nb_contexts, struct rte_flow_error *error)
6924 const struct mlx5_flow_driver_ops *fops;
6925 struct rte_flow_attr attr = { .transfer = 0 };
6927 if (flow_get_drv_type(dev, &attr) == MLX5_FLOW_TYPE_DV) {
6928 fops = flow_get_drv_ops(MLX5_FLOW_TYPE_DV);
6929 return fops->get_aged_flows(dev, contexts, nb_contexts,
6933 "port %u get aged flows is not supported.",
6934 dev->data->port_id);
6938 /* Wrapper for driver action_validate op callback */
6940 flow_drv_action_validate(struct rte_eth_dev *dev,
6941 const struct rte_flow_shared_action_conf *conf,
6942 const struct rte_flow_action *action,
6943 const struct mlx5_flow_driver_ops *fops,
6944 struct rte_flow_error *error)
6946 static const char err_msg[] = "shared action validation unsupported";
6948 if (!fops->action_validate) {
6949 DRV_LOG(ERR, "port %u %s.", dev->data->port_id, err_msg);
6950 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
6954 return fops->action_validate(dev, conf, action, error);
6958 * Destroys the shared action by handle.
6961 * Pointer to Ethernet device structure.
6963 * Handle for the shared action to be destroyed.
6965 * Perform verbose error reporting if not NULL. PMDs initialize this
6966 * structure in case of error only.
6969 * 0 on success, a negative errno value otherwise and rte_errno is set.
6971 * @note: wrapper for driver action_create op callback.
6974 mlx5_shared_action_destroy(struct rte_eth_dev *dev,
6975 struct rte_flow_shared_action *action,
6976 struct rte_flow_error *error)
6978 static const char err_msg[] = "shared action destruction unsupported";
6979 struct rte_flow_attr attr = { .transfer = 0 };
6980 const struct mlx5_flow_driver_ops *fops =
6981 flow_get_drv_ops(flow_get_drv_type(dev, &attr));
6983 if (!fops->action_destroy) {
6984 DRV_LOG(ERR, "port %u %s.", dev->data->port_id, err_msg);
6985 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
6989 return fops->action_destroy(dev, action, error);
6992 /* Wrapper for driver action_destroy op callback */
6994 flow_drv_action_update(struct rte_eth_dev *dev,
6995 struct rte_flow_shared_action *action,
6996 const void *action_conf,
6997 const struct mlx5_flow_driver_ops *fops,
6998 struct rte_flow_error *error)
7000 static const char err_msg[] = "shared action update unsupported";
7002 if (!fops->action_update) {
7003 DRV_LOG(ERR, "port %u %s.", dev->data->port_id, err_msg);
7004 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
7008 return fops->action_update(dev, action, action_conf, error);
7011 /* Wrapper for driver action_destroy op callback */
7013 flow_drv_action_query(struct rte_eth_dev *dev,
7014 const struct rte_flow_shared_action *action,
7016 const struct mlx5_flow_driver_ops *fops,
7017 struct rte_flow_error *error)
7019 static const char err_msg[] = "shared action query unsupported";
7021 if (!fops->action_query) {
7022 DRV_LOG(ERR, "port %u %s.", dev->data->port_id, err_msg);
7023 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
7027 return fops->action_query(dev, action, data, error);
7031 * Create shared action for reuse in multiple flow rules.
7034 * Pointer to Ethernet device structure.
7036 * Action configuration for shared action creation.
7038 * Perform verbose error reporting if not NULL. PMDs initialize this
7039 * structure in case of error only.
7041 * A valid handle in case of success, NULL otherwise and rte_errno is set.
7043 static struct rte_flow_shared_action *
7044 mlx5_shared_action_create(struct rte_eth_dev *dev,
7045 const struct rte_flow_shared_action_conf *conf,
7046 const struct rte_flow_action *action,
7047 struct rte_flow_error *error)
7049 static const char err_msg[] = "shared action creation unsupported";
7050 struct rte_flow_attr attr = { .transfer = 0 };
7051 const struct mlx5_flow_driver_ops *fops =
7052 flow_get_drv_ops(flow_get_drv_type(dev, &attr));
7054 if (flow_drv_action_validate(dev, conf, action, fops, error))
7056 if (!fops->action_create) {
7057 DRV_LOG(ERR, "port %u %s.", dev->data->port_id, err_msg);
7058 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
7062 return fops->action_create(dev, conf, action, error);
7066 * Updates inplace the shared action configuration pointed by *action* handle
7067 * with the configuration provided as *action* argument.
7068 * The update of the shared action configuration effects all flow rules reusing
7069 * the action via handle.
7072 * Pointer to Ethernet device structure.
7073 * @param[in] shared_action
7074 * Handle for the shared action to be updated.
7076 * Action specification used to modify the action pointed by handle.
7077 * *action* should be of same type with the action pointed by the *action*
7078 * handle argument, otherwise considered as invalid.
7080 * Perform verbose error reporting if not NULL. PMDs initialize this
7081 * structure in case of error only.
7084 * 0 on success, a negative errno value otherwise and rte_errno is set.
7087 mlx5_shared_action_update(struct rte_eth_dev *dev,
7088 struct rte_flow_shared_action *shared_action,
7089 const struct rte_flow_action *action,
7090 struct rte_flow_error *error)
7092 struct rte_flow_attr attr = { .transfer = 0 };
7093 const struct mlx5_flow_driver_ops *fops =
7094 flow_get_drv_ops(flow_get_drv_type(dev, &attr));
7097 ret = flow_drv_action_validate(dev, NULL, action, fops, error);
7100 return flow_drv_action_update(dev, shared_action, action->conf, fops,
7105 * Query the shared action by handle.
7107 * This function allows retrieving action-specific data such as counters.
7108 * Data is gathered by special action which may be present/referenced in
7109 * more than one flow rule definition.
7111 * \see RTE_FLOW_ACTION_TYPE_COUNT
7114 * Pointer to Ethernet device structure.
7116 * Handle for the shared action to query.
7117 * @param[in, out] data
7118 * Pointer to storage for the associated query data type.
7120 * Perform verbose error reporting if not NULL. PMDs initialize this
7121 * structure in case of error only.
7124 * 0 on success, a negative errno value otherwise and rte_errno is set.
7127 mlx5_shared_action_query(struct rte_eth_dev *dev,
7128 const struct rte_flow_shared_action *action,
7130 struct rte_flow_error *error)
7132 struct rte_flow_attr attr = { .transfer = 0 };
7133 const struct mlx5_flow_driver_ops *fops =
7134 flow_get_drv_ops(flow_get_drv_type(dev, &attr));
7136 return flow_drv_action_query(dev, action, data, fops, error);
7140 * Destroy all shared actions.
7143 * Pointer to Ethernet device.
7146 * 0 on success, a negative errno value otherwise and rte_errno is set.
7149 mlx5_shared_action_flush(struct rte_eth_dev *dev)
7151 struct rte_flow_error error;
7152 struct mlx5_priv *priv = dev->data->dev_private;
7153 struct mlx5_shared_action_rss *action;
7157 ILIST_FOREACH(priv->sh->ipool[MLX5_IPOOL_RSS_SHARED_ACTIONS],
7158 priv->rss_shared_actions, idx, action, next) {
7159 ret |= mlx5_shared_action_destroy(dev,
7160 (struct rte_flow_shared_action *)(uintptr_t)idx, &error);
7165 #ifndef HAVE_MLX5DV_DR
7166 #define MLX5_DOMAIN_SYNC_FLOW ((1 << 0) | (1 << 1))
7168 #define MLX5_DOMAIN_SYNC_FLOW \
7169 (MLX5DV_DR_DOMAIN_SYNC_FLAGS_SW | MLX5DV_DR_DOMAIN_SYNC_FLAGS_HW)
7172 int rte_pmd_mlx5_sync_flow(uint16_t port_id, uint32_t domains)
7174 struct rte_eth_dev *dev = &rte_eth_devices[port_id];
7175 const struct mlx5_flow_driver_ops *fops;
7177 struct rte_flow_attr attr = { .transfer = 0 };
7179 fops = flow_get_drv_ops(flow_get_drv_type(dev, &attr));
7180 ret = fops->sync_domain(dev, domains, MLX5_DOMAIN_SYNC_FLOW);
7187 * tunnel offload functionalilty is defined for DV environment only
7189 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
7191 union tunnel_offload_mark {
7194 uint32_t app_reserve:8;
7195 uint32_t table_id:15;
7196 uint32_t transfer:1;
7197 uint32_t _unused_:8;
7202 mlx5_access_tunnel_offload_db
7203 (struct rte_eth_dev *dev,
7204 bool (*match)(struct rte_eth_dev *,
7205 struct mlx5_flow_tunnel *, const void *),
7206 void (*hit)(struct rte_eth_dev *, struct mlx5_flow_tunnel *, void *),
7207 void (*miss)(struct rte_eth_dev *, void *),
7208 void *ctx, bool lock_op);
7211 flow_tunnel_add_default_miss(struct rte_eth_dev *dev,
7212 struct rte_flow *flow,
7213 const struct rte_flow_attr *attr,
7214 const struct rte_flow_action *app_actions,
7216 struct tunnel_default_miss_ctx *ctx,
7217 struct rte_flow_error *error)
7219 struct mlx5_priv *priv = dev->data->dev_private;
7220 struct mlx5_flow *dev_flow;
7221 struct rte_flow_attr miss_attr = *attr;
7222 const struct mlx5_flow_tunnel *tunnel = app_actions[0].conf;
7223 const struct rte_flow_item miss_items[2] = {
7225 .type = RTE_FLOW_ITEM_TYPE_ETH,
7231 .type = RTE_FLOW_ITEM_TYPE_END,
7237 union tunnel_offload_mark mark_id;
7238 struct rte_flow_action_mark miss_mark;
7239 struct rte_flow_action miss_actions[3] = {
7240 [0] = { .type = RTE_FLOW_ACTION_TYPE_MARK, .conf = &miss_mark },
7241 [2] = { .type = RTE_FLOW_ACTION_TYPE_END, .conf = NULL }
7243 const struct rte_flow_action_jump *jump_data;
7244 uint32_t i, flow_table = 0; /* prevent compilation warning */
7245 struct flow_grp_info grp_info = {
7247 .transfer = attr->transfer,
7248 .fdb_def_rule = !!priv->fdb_def_rule,
7253 if (!attr->transfer) {
7256 miss_actions[1].type = RTE_FLOW_ACTION_TYPE_RSS;
7257 q_size = priv->reta_idx_n * sizeof(ctx->queue[0]);
7258 ctx->queue = mlx5_malloc(MLX5_MEM_SYS | MLX5_MEM_ZERO, q_size,
7261 return rte_flow_error_set
7263 RTE_FLOW_ERROR_TYPE_ACTION_CONF,
7264 NULL, "invalid default miss RSS");
7265 ctx->action_rss.func = RTE_ETH_HASH_FUNCTION_DEFAULT,
7266 ctx->action_rss.level = 0,
7267 ctx->action_rss.types = priv->rss_conf.rss_hf,
7268 ctx->action_rss.key_len = priv->rss_conf.rss_key_len,
7269 ctx->action_rss.queue_num = priv->reta_idx_n,
7270 ctx->action_rss.key = priv->rss_conf.rss_key,
7271 ctx->action_rss.queue = ctx->queue;
7272 if (!priv->reta_idx_n || !priv->rxqs_n)
7273 return rte_flow_error_set
7275 RTE_FLOW_ERROR_TYPE_ACTION_CONF,
7276 NULL, "invalid port configuration");
7277 if (!(dev->data->dev_conf.rxmode.mq_mode & ETH_MQ_RX_RSS_FLAG))
7278 ctx->action_rss.types = 0;
7279 for (i = 0; i != priv->reta_idx_n; ++i)
7280 ctx->queue[i] = (*priv->reta_idx)[i];
7282 miss_actions[1].type = RTE_FLOW_ACTION_TYPE_JUMP;
7283 ctx->miss_jump.group = MLX5_TNL_MISS_FDB_JUMP_GRP;
7285 miss_actions[1].conf = (typeof(miss_actions[1].conf))ctx->raw;
7286 for (; app_actions->type != RTE_FLOW_ACTION_TYPE_JUMP; app_actions++);
7287 jump_data = app_actions->conf;
7288 miss_attr.priority = MLX5_TNL_MISS_RULE_PRIORITY;
7289 miss_attr.group = jump_data->group;
7290 ret = mlx5_flow_group_to_table(dev, tunnel, jump_data->group,
7291 &flow_table, &grp_info, error);
7293 return rte_flow_error_set(error, EINVAL,
7294 RTE_FLOW_ERROR_TYPE_ACTION_CONF,
7295 NULL, "invalid tunnel id");
7296 mark_id.app_reserve = 0;
7297 mark_id.table_id = tunnel_flow_tbl_to_id(flow_table);
7298 mark_id.transfer = !!attr->transfer;
7299 mark_id._unused_ = 0;
7300 miss_mark.id = mark_id.val;
7301 dev_flow = flow_drv_prepare(dev, flow, &miss_attr,
7302 miss_items, miss_actions, flow_idx, error);
7305 dev_flow->flow = flow;
7306 dev_flow->external = true;
7307 dev_flow->tunnel = tunnel;
7308 /* Subflow object was created, we must include one in the list. */
7309 SILIST_INSERT(&flow->dev_handles, dev_flow->handle_idx,
7310 dev_flow->handle, next);
7312 "port %u tunnel type=%d id=%u miss rule priority=%u group=%u",
7313 dev->data->port_id, tunnel->app_tunnel.type,
7314 tunnel->tunnel_id, miss_attr.priority, miss_attr.group);
7315 ret = flow_drv_translate(dev, dev_flow, &miss_attr, miss_items,
7316 miss_actions, error);
7318 ret = flow_mreg_update_copy_table(dev, flow, miss_actions,
7324 static const struct mlx5_flow_tbl_data_entry *
7325 tunnel_mark_decode(struct rte_eth_dev *dev, uint32_t mark)
7327 struct mlx5_priv *priv = dev->data->dev_private;
7328 struct mlx5_dev_ctx_shared *sh = priv->sh;
7329 struct mlx5_hlist_entry *he;
7330 union tunnel_offload_mark mbits = { .val = mark };
7331 union mlx5_flow_tbl_key table_key = {
7333 .table_id = tunnel_id_to_flow_tbl(mbits.table_id),
7335 .domain = !!mbits.transfer,
7339 he = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64, NULL);
7341 container_of(he, struct mlx5_flow_tbl_data_entry, entry) : NULL;
7345 mlx5_flow_tunnel_grp2tbl_remove_cb(struct mlx5_hlist *list,
7346 struct mlx5_hlist_entry *entry)
7348 struct mlx5_dev_ctx_shared *sh = list->ctx;
7349 struct tunnel_tbl_entry *tte = container_of(entry, typeof(*tte), hash);
7351 mlx5_ipool_free(sh->ipool[MLX5_IPOOL_TNL_TBL_ID],
7352 tunnel_flow_tbl_to_id(tte->flow_table));
7356 static struct mlx5_hlist_entry *
7357 mlx5_flow_tunnel_grp2tbl_create_cb(struct mlx5_hlist *list,
7358 uint64_t key __rte_unused,
7359 void *ctx __rte_unused)
7361 struct mlx5_dev_ctx_shared *sh = list->ctx;
7362 struct tunnel_tbl_entry *tte;
7364 tte = mlx5_malloc(MLX5_MEM_SYS | MLX5_MEM_ZERO,
7369 mlx5_ipool_malloc(sh->ipool[MLX5_IPOOL_TNL_TBL_ID],
7371 if (tte->flow_table >= MLX5_MAX_TABLES) {
7372 DRV_LOG(ERR, "Tunnel TBL ID %d exceed max limit.",
7374 mlx5_ipool_free(sh->ipool[MLX5_IPOOL_TNL_TBL_ID],
7377 } else if (!tte->flow_table) {
7380 tte->flow_table = tunnel_id_to_flow_tbl(tte->flow_table);
7389 tunnel_flow_group_to_flow_table(struct rte_eth_dev *dev,
7390 const struct mlx5_flow_tunnel *tunnel,
7391 uint32_t group, uint32_t *table,
7392 struct rte_flow_error *error)
7394 struct mlx5_hlist_entry *he;
7395 struct tunnel_tbl_entry *tte;
7396 union tunnel_tbl_key key = {
7397 .tunnel_id = tunnel ? tunnel->tunnel_id : 0,
7400 struct mlx5_flow_tunnel_hub *thub = mlx5_tunnel_hub(dev);
7401 struct mlx5_hlist *group_hash;
7403 group_hash = tunnel ? tunnel->groups : thub->groups;
7404 he = mlx5_hlist_register(group_hash, key.val, NULL);
7406 return rte_flow_error_set(error, EINVAL,
7407 RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
7409 "tunnel group index not supported");
7410 tte = container_of(he, typeof(*tte), hash);
7411 *table = tte->flow_table;
7412 DRV_LOG(DEBUG, "port %u tunnel %u group=%#x table=%#x",
7413 dev->data->port_id, key.tunnel_id, group, *table);
7418 mlx5_flow_tunnel_free(struct rte_eth_dev *dev,
7419 struct mlx5_flow_tunnel *tunnel)
7421 struct mlx5_priv *priv = dev->data->dev_private;
7422 struct mlx5_indexed_pool *ipool;
7424 DRV_LOG(DEBUG, "port %u release pmd tunnel id=0x%x",
7425 dev->data->port_id, tunnel->tunnel_id);
7426 mlx5_hlist_destroy(tunnel->groups);
7427 ipool = priv->sh->ipool[MLX5_IPOOL_TUNNEL_ID];
7428 mlx5_ipool_free(ipool, tunnel->tunnel_id);
7432 mlx5_access_tunnel_offload_db
7433 (struct rte_eth_dev *dev,
7434 bool (*match)(struct rte_eth_dev *,
7435 struct mlx5_flow_tunnel *, const void *),
7436 void (*hit)(struct rte_eth_dev *, struct mlx5_flow_tunnel *, void *),
7437 void (*miss)(struct rte_eth_dev *, void *),
7438 void *ctx, bool lock_op)
7440 bool verdict = false;
7441 struct mlx5_flow_tunnel_hub *thub = mlx5_tunnel_hub(dev);
7442 struct mlx5_flow_tunnel *tunnel;
7444 rte_spinlock_lock(&thub->sl);
7445 LIST_FOREACH(tunnel, &thub->tunnels, chain) {
7446 verdict = match(dev, tunnel, (const void *)ctx);
7451 rte_spinlock_unlock(&thub->sl);
7453 hit(dev, tunnel, ctx);
7454 if (!verdict && miss)
7457 rte_spinlock_unlock(&thub->sl);
7462 struct tunnel_db_find_tunnel_id_ctx {
7464 struct mlx5_flow_tunnel *tunnel;
7468 find_tunnel_id_match(struct rte_eth_dev *dev,
7469 struct mlx5_flow_tunnel *tunnel, const void *x)
7471 const struct tunnel_db_find_tunnel_id_ctx *ctx = x;
7474 return tunnel->tunnel_id == ctx->tunnel_id;
7478 find_tunnel_id_hit(struct rte_eth_dev *dev,
7479 struct mlx5_flow_tunnel *tunnel, void *x)
7481 struct tunnel_db_find_tunnel_id_ctx *ctx = x;
7483 ctx->tunnel = tunnel;
7486 static struct mlx5_flow_tunnel *
7487 mlx5_find_tunnel_id(struct rte_eth_dev *dev, uint32_t id)
7489 struct tunnel_db_find_tunnel_id_ctx ctx = {
7493 mlx5_access_tunnel_offload_db(dev, find_tunnel_id_match,
7494 find_tunnel_id_hit, NULL, &ctx, true);
7499 static struct mlx5_flow_tunnel *
7500 mlx5_flow_tunnel_allocate(struct rte_eth_dev *dev,
7501 const struct rte_flow_tunnel *app_tunnel)
7503 struct mlx5_priv *priv = dev->data->dev_private;
7504 struct mlx5_indexed_pool *ipool;
7505 struct mlx5_flow_tunnel *tunnel;
7508 ipool = priv->sh->ipool[MLX5_IPOOL_TUNNEL_ID];
7509 tunnel = mlx5_ipool_zmalloc(ipool, &id);
7512 if (id >= MLX5_MAX_TUNNELS) {
7513 mlx5_ipool_free(ipool, id);
7514 DRV_LOG(ERR, "Tunnel ID %d exceed max limit.", id);
7517 tunnel->groups = mlx5_hlist_create("tunnel groups", 1024, 0, 0,
7518 mlx5_flow_tunnel_grp2tbl_create_cb,
7520 mlx5_flow_tunnel_grp2tbl_remove_cb);
7521 if (!tunnel->groups) {
7522 mlx5_ipool_free(ipool, id);
7525 tunnel->groups->ctx = priv->sh;
7526 /* initiate new PMD tunnel */
7527 memcpy(&tunnel->app_tunnel, app_tunnel, sizeof(*app_tunnel));
7528 tunnel->tunnel_id = id;
7529 tunnel->action.type = (typeof(tunnel->action.type))
7530 MLX5_RTE_FLOW_ACTION_TYPE_TUNNEL_SET;
7531 tunnel->action.conf = tunnel;
7532 tunnel->item.type = (typeof(tunnel->item.type))
7533 MLX5_RTE_FLOW_ITEM_TYPE_TUNNEL;
7534 tunnel->item.spec = tunnel;
7535 tunnel->item.last = NULL;
7536 tunnel->item.mask = NULL;
7538 DRV_LOG(DEBUG, "port %u new pmd tunnel id=0x%x",
7539 dev->data->port_id, tunnel->tunnel_id);
7544 struct tunnel_db_get_tunnel_ctx {
7545 const struct rte_flow_tunnel *app_tunnel;
7546 struct mlx5_flow_tunnel *tunnel;
7549 static bool get_tunnel_match(struct rte_eth_dev *dev,
7550 struct mlx5_flow_tunnel *tunnel, const void *x)
7552 const struct tunnel_db_get_tunnel_ctx *ctx = x;
7555 return !memcmp(ctx->app_tunnel, &tunnel->app_tunnel,
7556 sizeof(*ctx->app_tunnel));
7559 static void get_tunnel_hit(struct rte_eth_dev *dev,
7560 struct mlx5_flow_tunnel *tunnel, void *x)
7562 /* called under tunnel spinlock protection */
7563 struct tunnel_db_get_tunnel_ctx *ctx = x;
7567 ctx->tunnel = tunnel;
7570 static void get_tunnel_miss(struct rte_eth_dev *dev, void *x)
7572 /* called under tunnel spinlock protection */
7573 struct mlx5_flow_tunnel_hub *thub = mlx5_tunnel_hub(dev);
7574 struct tunnel_db_get_tunnel_ctx *ctx = x;
7576 rte_spinlock_unlock(&thub->sl);
7577 ctx->tunnel = mlx5_flow_tunnel_allocate(dev, ctx->app_tunnel);
7578 ctx->tunnel->refctn = 1;
7579 rte_spinlock_lock(&thub->sl);
7581 LIST_INSERT_HEAD(&thub->tunnels, ctx->tunnel, chain);
7586 mlx5_get_flow_tunnel(struct rte_eth_dev *dev,
7587 const struct rte_flow_tunnel *app_tunnel,
7588 struct mlx5_flow_tunnel **tunnel)
7590 struct tunnel_db_get_tunnel_ctx ctx = {
7591 .app_tunnel = app_tunnel,
7594 mlx5_access_tunnel_offload_db(dev, get_tunnel_match, get_tunnel_hit,
7595 get_tunnel_miss, &ctx, true);
7596 *tunnel = ctx.tunnel;
7597 return ctx.tunnel ? 0 : -ENOMEM;
7600 void mlx5_release_tunnel_hub(struct mlx5_dev_ctx_shared *sh, uint16_t port_id)
7602 struct mlx5_flow_tunnel_hub *thub = sh->tunnel_hub;
7606 if (!LIST_EMPTY(&thub->tunnels))
7607 DRV_LOG(WARNING, "port %u tunnels present\n", port_id);
7608 mlx5_hlist_destroy(thub->groups);
7612 int mlx5_alloc_tunnel_hub(struct mlx5_dev_ctx_shared *sh)
7615 struct mlx5_flow_tunnel_hub *thub;
7617 thub = mlx5_malloc(MLX5_MEM_SYS | MLX5_MEM_ZERO, sizeof(*thub),
7621 LIST_INIT(&thub->tunnels);
7622 rte_spinlock_init(&thub->sl);
7623 thub->groups = mlx5_hlist_create("flow groups", MLX5_MAX_TABLES, 0,
7624 0, mlx5_flow_tunnel_grp2tbl_create_cb,
7626 mlx5_flow_tunnel_grp2tbl_remove_cb);
7627 if (!thub->groups) {
7631 thub->groups->ctx = sh;
7632 sh->tunnel_hub = thub;
7638 mlx5_hlist_destroy(thub->groups);
7645 mlx5_flow_tunnel_validate(struct rte_eth_dev *dev,
7646 struct rte_flow_tunnel *tunnel,
7647 const char *err_msg)
7650 if (!is_tunnel_offload_active(dev)) {
7651 err_msg = "tunnel offload was not activated";
7653 } else if (!tunnel) {
7654 err_msg = "no application tunnel";
7658 switch (tunnel->type) {
7660 err_msg = "unsupported tunnel type";
7662 case RTE_FLOW_ITEM_TYPE_VXLAN:
7671 mlx5_flow_tunnel_decap_set(struct rte_eth_dev *dev,
7672 struct rte_flow_tunnel *app_tunnel,
7673 struct rte_flow_action **actions,
7674 uint32_t *num_of_actions,
7675 struct rte_flow_error *error)
7678 struct mlx5_flow_tunnel *tunnel;
7679 const char *err_msg = NULL;
7680 bool verdict = mlx5_flow_tunnel_validate(dev, app_tunnel, err_msg);
7683 return rte_flow_error_set(error, EINVAL,
7684 RTE_FLOW_ERROR_TYPE_ACTION_CONF, NULL,
7686 ret = mlx5_get_flow_tunnel(dev, app_tunnel, &tunnel);
7688 return rte_flow_error_set(error, ret,
7689 RTE_FLOW_ERROR_TYPE_ACTION_CONF, NULL,
7690 "failed to initialize pmd tunnel");
7692 *actions = &tunnel->action;
7693 *num_of_actions = 1;
7698 mlx5_flow_tunnel_match(struct rte_eth_dev *dev,
7699 struct rte_flow_tunnel *app_tunnel,
7700 struct rte_flow_item **items,
7701 uint32_t *num_of_items,
7702 struct rte_flow_error *error)
7705 struct mlx5_flow_tunnel *tunnel;
7706 const char *err_msg = NULL;
7707 bool verdict = mlx5_flow_tunnel_validate(dev, app_tunnel, err_msg);
7710 return rte_flow_error_set(error, EINVAL,
7711 RTE_FLOW_ERROR_TYPE_HANDLE, NULL,
7713 ret = mlx5_get_flow_tunnel(dev, app_tunnel, &tunnel);
7715 return rte_flow_error_set(error, ret,
7716 RTE_FLOW_ERROR_TYPE_HANDLE, NULL,
7717 "failed to initialize pmd tunnel");
7719 *items = &tunnel->item;
7724 struct tunnel_db_element_release_ctx {
7725 struct rte_flow_item *items;
7726 struct rte_flow_action *actions;
7727 uint32_t num_elements;
7728 struct rte_flow_error *error;
7733 tunnel_element_release_match(struct rte_eth_dev *dev,
7734 struct mlx5_flow_tunnel *tunnel, const void *x)
7736 const struct tunnel_db_element_release_ctx *ctx = x;
7739 if (ctx->num_elements != 1)
7741 else if (ctx->items)
7742 return ctx->items == &tunnel->item;
7743 else if (ctx->actions)
7744 return ctx->actions == &tunnel->action;
7750 tunnel_element_release_hit(struct rte_eth_dev *dev,
7751 struct mlx5_flow_tunnel *tunnel, void *x)
7753 struct tunnel_db_element_release_ctx *ctx = x;
7755 if (!__atomic_sub_fetch(&tunnel->refctn, 1, __ATOMIC_RELAXED))
7756 mlx5_flow_tunnel_free(dev, tunnel);
7760 tunnel_element_release_miss(struct rte_eth_dev *dev, void *x)
7762 struct tunnel_db_element_release_ctx *ctx = x;
7764 ctx->ret = rte_flow_error_set(ctx->error, EINVAL,
7765 RTE_FLOW_ERROR_TYPE_HANDLE, NULL,
7766 "invalid argument");
7770 mlx5_flow_tunnel_item_release(struct rte_eth_dev *dev,
7771 struct rte_flow_item *pmd_items,
7772 uint32_t num_items, struct rte_flow_error *err)
7774 struct tunnel_db_element_release_ctx ctx = {
7777 .num_elements = num_items,
7781 mlx5_access_tunnel_offload_db(dev, tunnel_element_release_match,
7782 tunnel_element_release_hit,
7783 tunnel_element_release_miss, &ctx, false);
7789 mlx5_flow_tunnel_action_release(struct rte_eth_dev *dev,
7790 struct rte_flow_action *pmd_actions,
7791 uint32_t num_actions, struct rte_flow_error *err)
7793 struct tunnel_db_element_release_ctx ctx = {
7795 .actions = pmd_actions,
7796 .num_elements = num_actions,
7800 mlx5_access_tunnel_offload_db(dev, tunnel_element_release_match,
7801 tunnel_element_release_hit,
7802 tunnel_element_release_miss, &ctx, false);
7808 mlx5_flow_tunnel_get_restore_info(struct rte_eth_dev *dev,
7810 struct rte_flow_restore_info *info,
7811 struct rte_flow_error *err)
7813 uint64_t ol_flags = m->ol_flags;
7814 const struct mlx5_flow_tbl_data_entry *tble;
7815 const uint64_t mask = PKT_RX_FDIR | PKT_RX_FDIR_ID;
7817 if (!is_tunnel_offload_active(dev)) {
7822 if ((ol_flags & mask) != mask)
7824 tble = tunnel_mark_decode(dev, m->hash.fdir.hi);
7826 DRV_LOG(DEBUG, "port %u invalid miss tunnel mark %#x",
7827 dev->data->port_id, m->hash.fdir.hi);
7830 MLX5_ASSERT(tble->tunnel);
7831 memcpy(&info->tunnel, &tble->tunnel->app_tunnel, sizeof(info->tunnel));
7832 info->group_id = tble->group_id;
7833 info->flags = RTE_FLOW_RESTORE_INFO_TUNNEL |
7834 RTE_FLOW_RESTORE_INFO_GROUP_ID |
7835 RTE_FLOW_RESTORE_INFO_ENCAPSULATED;
7840 return rte_flow_error_set(err, EINVAL,
7841 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
7842 "failed to get restore info");
7845 #else /* HAVE_IBV_FLOW_DV_SUPPORT */
7847 mlx5_flow_tunnel_decap_set(__rte_unused struct rte_eth_dev *dev,
7848 __rte_unused struct rte_flow_tunnel *app_tunnel,
7849 __rte_unused struct rte_flow_action **actions,
7850 __rte_unused uint32_t *num_of_actions,
7851 __rte_unused struct rte_flow_error *error)
7857 mlx5_flow_tunnel_match(__rte_unused struct rte_eth_dev *dev,
7858 __rte_unused struct rte_flow_tunnel *app_tunnel,
7859 __rte_unused struct rte_flow_item **items,
7860 __rte_unused uint32_t *num_of_items,
7861 __rte_unused struct rte_flow_error *error)
7867 mlx5_flow_tunnel_item_release(__rte_unused struct rte_eth_dev *dev,
7868 __rte_unused struct rte_flow_item *pmd_items,
7869 __rte_unused uint32_t num_items,
7870 __rte_unused struct rte_flow_error *err)
7876 mlx5_flow_tunnel_action_release(__rte_unused struct rte_eth_dev *dev,
7877 __rte_unused struct rte_flow_action *pmd_action,
7878 __rte_unused uint32_t num_actions,
7879 __rte_unused struct rte_flow_error *err)
7885 mlx5_flow_tunnel_get_restore_info(__rte_unused struct rte_eth_dev *dev,
7886 __rte_unused struct rte_mbuf *m,
7887 __rte_unused struct rte_flow_restore_info *i,
7888 __rte_unused struct rte_flow_error *err)
7894 flow_tunnel_add_default_miss(__rte_unused struct rte_eth_dev *dev,
7895 __rte_unused struct rte_flow *flow,
7896 __rte_unused const struct rte_flow_attr *attr,
7897 __rte_unused const struct rte_flow_action *actions,
7898 __rte_unused uint32_t flow_idx,
7899 __rte_unused struct tunnel_default_miss_ctx *ctx,
7900 __rte_unused struct rte_flow_error *error)
7905 static struct mlx5_flow_tunnel *
7906 mlx5_find_tunnel_id(__rte_unused struct rte_eth_dev *dev,
7907 __rte_unused uint32_t id)
7913 mlx5_flow_tunnel_free(__rte_unused struct rte_eth_dev *dev,
7914 __rte_unused struct mlx5_flow_tunnel *tunnel)
7919 tunnel_flow_group_to_flow_table(__rte_unused struct rte_eth_dev *dev,
7920 __rte_unused const struct mlx5_flow_tunnel *t,
7921 __rte_unused uint32_t group,
7922 __rte_unused uint32_t *table,
7923 struct rte_flow_error *error)
7925 return rte_flow_error_set(error, ENOTSUP,
7926 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
7927 "tunnel offload requires DV support");
7931 mlx5_release_tunnel_hub(__rte_unused struct mlx5_dev_ctx_shared *sh,
7932 __rte_unused uint16_t port_id)
7935 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */