1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2018 Mellanox Technologies, Ltd
5 #include <netinet/in.h>
11 #include <rte_common.h>
12 #include <rte_ether.h>
13 #include <ethdev_driver.h>
15 #include <rte_flow_driver.h>
16 #include <rte_malloc.h>
19 #include <mlx5_glue.h>
21 #include <mlx5_malloc.h>
23 #include "mlx5_defs.h"
25 #include "mlx5_flow.h"
28 #define VERBS_SPEC_INNER(item_flags) \
29 (!!((item_flags) & MLX5_FLOW_LAYER_TUNNEL) ? IBV_FLOW_SPEC_INNER : 0)
31 /* Verbs specification header. */
32 struct ibv_spec_header {
33 enum ibv_flow_spec_type type;
38 * Discover the maximum number of priority available.
41 * Pointer to the Ethernet device structure.
43 * Expected result variants.
45 * Number of entries in @p vprio array.
47 * Number of supported flow priority on success, a negative errno
48 * value otherwise and rte_errno is set.
51 flow_verbs_discover_priorities(struct rte_eth_dev *dev,
52 const uint16_t *vprio, int vprio_n)
54 struct mlx5_priv *priv = dev->data->dev_private;
56 struct ibv_flow_attr attr;
57 struct ibv_flow_spec_eth eth;
58 struct ibv_flow_spec_action_drop drop;
62 .port = (uint8_t)priv->dev_port,
65 .type = IBV_FLOW_SPEC_ETH,
66 .size = sizeof(struct ibv_flow_spec_eth),
69 .size = sizeof(struct ibv_flow_spec_action_drop),
70 .type = IBV_FLOW_SPEC_ACTION_DROP,
73 struct ibv_flow *flow;
74 struct mlx5_hrxq *drop = priv->drop_queue.hrxq;
78 #if defined(HAVE_MLX5DV_DR_DEVX_PORT) || defined(HAVE_MLX5DV_DR_DEVX_PORT_V35)
79 /* If DevX supported, driver must support 16 verbs flow priorities. */
87 for (i = 0; i != vprio_n; i++) {
88 flow_attr.attr.priority = vprio[i] - 1;
89 flow = mlx5_glue->create_flow(drop->qp, &flow_attr.attr);
92 claim_zero(mlx5_glue->destroy_flow(flow));
95 #if defined(HAVE_MLX5DV_DR_DEVX_PORT) || defined(HAVE_MLX5DV_DR_DEVX_PORT_V35)
98 DRV_LOG(INFO, "port %u supported flow priorities:"
99 " 0-%d for ingress or egress root table,"
100 " 0-%d for non-root table or transfer root table.",
101 dev->data->port_id, priority - 2,
102 MLX5_NON_ROOT_FLOW_MAX_PRIO - 1);
107 * Get Verbs flow counter by index.
110 * Pointer to the Ethernet device structure.
112 * mlx5 flow counter index in the container.
114 * mlx5 flow counter pool in the container,
117 * A pointer to the counter, NULL otherwise.
119 static struct mlx5_flow_counter *
120 flow_verbs_counter_get_by_idx(struct rte_eth_dev *dev,
122 struct mlx5_flow_counter_pool **ppool)
124 struct mlx5_priv *priv = dev->data->dev_private;
125 struct mlx5_flow_counter_mng *cmng = &priv->sh->cmng;
126 struct mlx5_flow_counter_pool *pool;
128 idx = (idx - 1) & (MLX5_CNT_SHARED_OFFSET - 1);
129 pool = cmng->pools[idx / MLX5_COUNTERS_PER_POOL];
133 return MLX5_POOL_GET_CNT(pool, idx % MLX5_COUNTERS_PER_POOL);
137 * Create Verbs flow counter with Verbs library.
140 * Pointer to the Ethernet device structure.
141 * @param[in, out] counter
142 * mlx5 flow counter object, contains the counter id,
143 * handle of created Verbs flow counter is returned
144 * in cs field (if counters are supported).
147 * 0 On success else a negative errno value is returned
148 * and rte_errno is set.
151 flow_verbs_counter_create(struct rte_eth_dev *dev,
152 struct mlx5_flow_counter *counter)
154 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42)
155 struct mlx5_priv *priv = dev->data->dev_private;
156 struct ibv_context *ctx = priv->sh->cdev->ctx;
157 struct ibv_counter_set_init_attr init = {
158 .counter_set_id = counter->shared_info.id};
160 counter->dcs_when_free = mlx5_glue->create_counter_set(ctx, &init);
161 if (!counter->dcs_when_free) {
166 #elif defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
167 struct mlx5_priv *priv = dev->data->dev_private;
168 struct ibv_context *ctx = priv->sh->cdev->ctx;
169 struct ibv_counters_init_attr init = {0};
170 struct ibv_counter_attach_attr attach;
173 memset(&attach, 0, sizeof(attach));
174 counter->dcs_when_free = mlx5_glue->create_counters(ctx, &init);
175 if (!counter->dcs_when_free) {
179 attach.counter_desc = IBV_COUNTER_PACKETS;
181 ret = mlx5_glue->attach_counters(counter->dcs_when_free, &attach, NULL);
183 attach.counter_desc = IBV_COUNTER_BYTES;
185 ret = mlx5_glue->attach_counters
186 (counter->dcs_when_free, &attach, NULL);
189 claim_zero(mlx5_glue->destroy_counters(counter->dcs_when_free));
190 counter->dcs_when_free = NULL;
204 * Get a flow counter.
207 * Pointer to the Ethernet device structure.
209 * Counter identifier.
212 * Index to the counter, 0 otherwise and rte_errno is set.
215 flow_verbs_counter_new(struct rte_eth_dev *dev, uint32_t id __rte_unused)
217 struct mlx5_priv *priv = dev->data->dev_private;
218 struct mlx5_flow_counter_mng *cmng = &priv->sh->cmng;
219 struct mlx5_flow_counter_pool *pool = NULL;
220 struct mlx5_flow_counter *cnt = NULL;
221 uint32_t n_valid = cmng->n_valid;
222 uint32_t pool_idx, cnt_idx;
226 for (pool_idx = 0; pool_idx < n_valid; ++pool_idx) {
227 pool = cmng->pools[pool_idx];
230 cnt = TAILQ_FIRST(&pool->counters[0]);
235 struct mlx5_flow_counter_pool **pools;
238 if (n_valid == cmng->n) {
239 /* Resize the container pool array. */
240 size = sizeof(struct mlx5_flow_counter_pool *) *
241 (n_valid + MLX5_CNT_CONTAINER_RESIZE);
242 pools = mlx5_malloc(MLX5_MEM_ZERO, size, 0,
247 memcpy(pools, cmng->pools,
248 sizeof(struct mlx5_flow_counter_pool *) *
250 mlx5_free(cmng->pools);
253 cmng->n += MLX5_CNT_CONTAINER_RESIZE;
255 /* Allocate memory for new pool*/
256 size = sizeof(*pool) + sizeof(*cnt) * MLX5_COUNTERS_PER_POOL;
257 pool = mlx5_malloc(MLX5_MEM_ZERO, size, 0, SOCKET_ID_ANY);
260 for (i = 0; i < MLX5_COUNTERS_PER_POOL; ++i) {
261 cnt = MLX5_POOL_GET_CNT(pool, i);
262 TAILQ_INSERT_HEAD(&pool->counters[0], cnt, next);
264 cnt = MLX5_POOL_GET_CNT(pool, 0);
265 cmng->pools[n_valid] = pool;
269 TAILQ_REMOVE(&pool->counters[0], cnt, next);
270 i = MLX5_CNT_ARRAY_IDX(pool, cnt);
271 cnt_idx = MLX5_MAKE_CNT_IDX(pool_idx, i);
272 /* Create counter with Verbs. */
273 ret = flow_verbs_counter_create(dev, cnt);
275 cnt->dcs_when_active = cnt->dcs_when_free;
280 TAILQ_INSERT_HEAD(&pool->counters[0], cnt, next);
281 /* Some error occurred in Verbs library. */
287 * Release a flow counter.
290 * Pointer to the Ethernet device structure.
292 * Index to the counter handler.
295 flow_verbs_counter_release(struct rte_eth_dev *dev, uint32_t counter)
297 struct mlx5_flow_counter_pool *pool;
298 struct mlx5_flow_counter *cnt;
300 cnt = flow_verbs_counter_get_by_idx(dev, counter, &pool);
301 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42)
302 claim_zero(mlx5_glue->destroy_counter_set
303 ((struct ibv_counter_set *)cnt->dcs_when_active));
304 #elif defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
305 claim_zero(mlx5_glue->destroy_counters
306 ((struct ibv_counters *)cnt->dcs_when_active));
308 TAILQ_INSERT_HEAD(&pool->counters[0], cnt, next);
312 * Query a flow counter via Verbs library call.
314 * @see rte_flow_query()
318 flow_verbs_counter_query(struct rte_eth_dev *dev __rte_unused,
319 struct rte_flow *flow, void *data,
320 struct rte_flow_error *error)
322 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) || \
323 defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
325 struct mlx5_flow_counter_pool *pool;
326 struct mlx5_flow_counter *cnt = flow_verbs_counter_get_by_idx
327 (dev, flow->counter, &pool);
328 struct rte_flow_query_count *qc = data;
329 uint64_t counters[2] = {0, 0};
330 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42)
331 struct ibv_query_counter_set_attr query_cs_attr = {
332 .dcs_when_free = (struct ibv_counter_set *)
333 cnt->dcs_when_active,
334 .query_flags = IBV_COUNTER_SET_FORCE_UPDATE,
336 struct ibv_counter_set_data query_out = {
338 .outlen = 2 * sizeof(uint64_t),
340 int err = mlx5_glue->query_counter_set(&query_cs_attr,
342 #elif defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
343 int err = mlx5_glue->query_counters
344 ((struct ibv_counters *)cnt->dcs_when_active, counters,
346 IBV_READ_COUNTERS_ATTR_PREFER_CACHED);
349 return rte_flow_error_set
351 RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
353 "cannot read counter");
356 qc->hits = counters[0] - cnt->hits;
357 qc->bytes = counters[1] - cnt->bytes;
359 cnt->hits = counters[0];
360 cnt->bytes = counters[1];
364 return rte_flow_error_set(error, EINVAL,
365 RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
367 "flow does not have counter");
371 return rte_flow_error_set(error, ENOTSUP,
372 RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
374 "counters are not available");
379 * Add a verbs item specification into @p verbs.
382 * Pointer to verbs structure.
384 * Create specification.
386 * Size in bytes of the specification to copy.
389 flow_verbs_spec_add(struct mlx5_flow_verbs_workspace *verbs,
390 void *src, unsigned int size)
396 MLX5_ASSERT(verbs->specs);
397 dst = (void *)(verbs->specs + verbs->size);
398 memcpy(dst, src, size);
399 ++verbs->attr.num_of_specs;
404 * Convert the @p item into a Verbs specification. This function assumes that
405 * the input is valid and that there is space to insert the requested item
408 * @param[in, out] dev_flow
409 * Pointer to dev_flow structure.
411 * Item specification.
412 * @param[in] item_flags
416 flow_verbs_translate_item_eth(struct mlx5_flow *dev_flow,
417 const struct rte_flow_item *item,
420 const struct rte_flow_item_eth *spec = item->spec;
421 const struct rte_flow_item_eth *mask = item->mask;
422 const unsigned int size = sizeof(struct ibv_flow_spec_eth);
423 struct ibv_flow_spec_eth eth = {
424 .type = IBV_FLOW_SPEC_ETH | VERBS_SPEC_INNER(item_flags),
429 mask = &rte_flow_item_eth_mask;
433 memcpy(ð.val.dst_mac, spec->dst.addr_bytes,
435 memcpy(ð.val.src_mac, spec->src.addr_bytes,
437 eth.val.ether_type = spec->type;
438 memcpy(ð.mask.dst_mac, mask->dst.addr_bytes,
440 memcpy(ð.mask.src_mac, mask->src.addr_bytes,
442 eth.mask.ether_type = mask->type;
443 /* Remove unwanted bits from values. */
444 for (i = 0; i < RTE_ETHER_ADDR_LEN; ++i) {
445 eth.val.dst_mac[i] &= eth.mask.dst_mac[i];
446 eth.val.src_mac[i] &= eth.mask.src_mac[i];
448 eth.val.ether_type &= eth.mask.ether_type;
450 flow_verbs_spec_add(&dev_flow->verbs, ð, size);
454 * Update the VLAN tag in the Verbs Ethernet specification.
455 * This function assumes that the input is valid and there is space to add
456 * the requested item.
458 * @param[in, out] attr
459 * Pointer to Verbs attributes structure.
461 * Verbs structure containing the VLAN information to copy.
464 flow_verbs_item_vlan_update(struct ibv_flow_attr *attr,
465 struct ibv_flow_spec_eth *eth)
468 const enum ibv_flow_spec_type search = eth->type;
469 struct ibv_spec_header *hdr = (struct ibv_spec_header *)
470 ((uint8_t *)attr + sizeof(struct ibv_flow_attr));
472 for (i = 0; i != attr->num_of_specs; ++i) {
473 if (hdr->type == search) {
474 struct ibv_flow_spec_eth *e =
475 (struct ibv_flow_spec_eth *)hdr;
477 e->val.vlan_tag = eth->val.vlan_tag;
478 e->mask.vlan_tag = eth->mask.vlan_tag;
479 e->val.ether_type = eth->val.ether_type;
480 e->mask.ether_type = eth->mask.ether_type;
483 hdr = (struct ibv_spec_header *)((uint8_t *)hdr + hdr->size);
488 * Convert the @p item into a Verbs specification. This function assumes that
489 * the input is valid and that there is space to insert the requested item
492 * @param[in, out] dev_flow
493 * Pointer to dev_flow structure.
495 * Item specification.
496 * @param[in] item_flags
500 flow_verbs_translate_item_vlan(struct mlx5_flow *dev_flow,
501 const struct rte_flow_item *item,
504 const struct rte_flow_item_vlan *spec = item->spec;
505 const struct rte_flow_item_vlan *mask = item->mask;
506 unsigned int size = sizeof(struct ibv_flow_spec_eth);
507 const int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
508 struct ibv_flow_spec_eth eth = {
509 .type = IBV_FLOW_SPEC_ETH | VERBS_SPEC_INNER(item_flags),
512 const uint32_t l2m = tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
513 MLX5_FLOW_LAYER_OUTER_L2;
516 mask = &rte_flow_item_vlan_mask;
518 eth.val.vlan_tag = spec->tci;
519 eth.mask.vlan_tag = mask->tci;
520 eth.val.vlan_tag &= eth.mask.vlan_tag;
521 eth.val.ether_type = spec->inner_type;
522 eth.mask.ether_type = mask->inner_type;
523 eth.val.ether_type &= eth.mask.ether_type;
525 if (!(item_flags & l2m))
526 flow_verbs_spec_add(&dev_flow->verbs, ð, size);
528 flow_verbs_item_vlan_update(&dev_flow->verbs.attr, ð);
530 dev_flow->handle->vf_vlan.tag =
531 rte_be_to_cpu_16(spec->tci) & 0x0fff;
535 * Convert the @p item into a Verbs specification. This function assumes that
536 * the input is valid and that there is space to insert the requested item
539 * @param[in, out] dev_flow
540 * Pointer to dev_flow structure.
542 * Item specification.
543 * @param[in] item_flags
547 flow_verbs_translate_item_ipv4(struct mlx5_flow *dev_flow,
548 const struct rte_flow_item *item,
551 const struct rte_flow_item_ipv4 *spec = item->spec;
552 const struct rte_flow_item_ipv4 *mask = item->mask;
553 unsigned int size = sizeof(struct ibv_flow_spec_ipv4_ext);
554 struct ibv_flow_spec_ipv4_ext ipv4 = {
555 .type = IBV_FLOW_SPEC_IPV4_EXT | VERBS_SPEC_INNER(item_flags),
560 mask = &rte_flow_item_ipv4_mask;
562 ipv4.val = (struct ibv_flow_ipv4_ext_filter){
563 .src_ip = spec->hdr.src_addr,
564 .dst_ip = spec->hdr.dst_addr,
565 .proto = spec->hdr.next_proto_id,
566 .tos = spec->hdr.type_of_service,
568 ipv4.mask = (struct ibv_flow_ipv4_ext_filter){
569 .src_ip = mask->hdr.src_addr,
570 .dst_ip = mask->hdr.dst_addr,
571 .proto = mask->hdr.next_proto_id,
572 .tos = mask->hdr.type_of_service,
574 /* Remove unwanted bits from values. */
575 ipv4.val.src_ip &= ipv4.mask.src_ip;
576 ipv4.val.dst_ip &= ipv4.mask.dst_ip;
577 ipv4.val.proto &= ipv4.mask.proto;
578 ipv4.val.tos &= ipv4.mask.tos;
580 flow_verbs_spec_add(&dev_flow->verbs, &ipv4, size);
584 * Convert the @p item into a Verbs specification. This function assumes that
585 * the input is valid and that there is space to insert the requested item
588 * @param[in, out] dev_flow
589 * Pointer to dev_flow structure.
591 * Item specification.
592 * @param[in] item_flags
596 flow_verbs_translate_item_ipv6(struct mlx5_flow *dev_flow,
597 const struct rte_flow_item *item,
600 const struct rte_flow_item_ipv6 *spec = item->spec;
601 const struct rte_flow_item_ipv6 *mask = item->mask;
602 unsigned int size = sizeof(struct ibv_flow_spec_ipv6);
603 struct ibv_flow_spec_ipv6 ipv6 = {
604 .type = IBV_FLOW_SPEC_IPV6 | VERBS_SPEC_INNER(item_flags),
609 mask = &rte_flow_item_ipv6_mask;
612 uint32_t vtc_flow_val;
613 uint32_t vtc_flow_mask;
615 memcpy(&ipv6.val.src_ip, spec->hdr.src_addr,
616 RTE_DIM(ipv6.val.src_ip));
617 memcpy(&ipv6.val.dst_ip, spec->hdr.dst_addr,
618 RTE_DIM(ipv6.val.dst_ip));
619 memcpy(&ipv6.mask.src_ip, mask->hdr.src_addr,
620 RTE_DIM(ipv6.mask.src_ip));
621 memcpy(&ipv6.mask.dst_ip, mask->hdr.dst_addr,
622 RTE_DIM(ipv6.mask.dst_ip));
623 vtc_flow_val = rte_be_to_cpu_32(spec->hdr.vtc_flow);
624 vtc_flow_mask = rte_be_to_cpu_32(mask->hdr.vtc_flow);
625 ipv6.val.flow_label =
626 rte_cpu_to_be_32((vtc_flow_val & RTE_IPV6_HDR_FL_MASK) >>
627 RTE_IPV6_HDR_FL_SHIFT);
628 ipv6.val.traffic_class = (vtc_flow_val & RTE_IPV6_HDR_TC_MASK) >>
629 RTE_IPV6_HDR_TC_SHIFT;
630 ipv6.val.next_hdr = spec->hdr.proto;
631 ipv6.mask.flow_label =
632 rte_cpu_to_be_32((vtc_flow_mask & RTE_IPV6_HDR_FL_MASK) >>
633 RTE_IPV6_HDR_FL_SHIFT);
634 ipv6.mask.traffic_class = (vtc_flow_mask & RTE_IPV6_HDR_TC_MASK) >>
635 RTE_IPV6_HDR_TC_SHIFT;
636 ipv6.mask.next_hdr = mask->hdr.proto;
637 /* Remove unwanted bits from values. */
638 for (i = 0; i < RTE_DIM(ipv6.val.src_ip); ++i) {
639 ipv6.val.src_ip[i] &= ipv6.mask.src_ip[i];
640 ipv6.val.dst_ip[i] &= ipv6.mask.dst_ip[i];
642 ipv6.val.flow_label &= ipv6.mask.flow_label;
643 ipv6.val.traffic_class &= ipv6.mask.traffic_class;
644 ipv6.val.next_hdr &= ipv6.mask.next_hdr;
646 flow_verbs_spec_add(&dev_flow->verbs, &ipv6, size);
650 * Convert the @p item into a Verbs specification. This function assumes that
651 * the input is valid and that there is space to insert the requested item
654 * @param[in, out] dev_flow
655 * Pointer to dev_flow structure.
657 * Item specification.
658 * @param[in] item_flags
662 flow_verbs_translate_item_tcp(struct mlx5_flow *dev_flow,
663 const struct rte_flow_item *item,
664 uint64_t item_flags __rte_unused)
666 const struct rte_flow_item_tcp *spec = item->spec;
667 const struct rte_flow_item_tcp *mask = item->mask;
668 unsigned int size = sizeof(struct ibv_flow_spec_tcp_udp);
669 struct ibv_flow_spec_tcp_udp tcp = {
670 .type = IBV_FLOW_SPEC_TCP | VERBS_SPEC_INNER(item_flags),
675 mask = &rte_flow_item_tcp_mask;
677 tcp.val.dst_port = spec->hdr.dst_port;
678 tcp.val.src_port = spec->hdr.src_port;
679 tcp.mask.dst_port = mask->hdr.dst_port;
680 tcp.mask.src_port = mask->hdr.src_port;
681 /* Remove unwanted bits from values. */
682 tcp.val.src_port &= tcp.mask.src_port;
683 tcp.val.dst_port &= tcp.mask.dst_port;
685 flow_verbs_spec_add(&dev_flow->verbs, &tcp, size);
689 * Convert the @p item into a Verbs specification. This function assumes that
690 * the input is valid and that there is space to insert the requested item
693 * @param[in, out] dev_flow
694 * Pointer to dev_flow structure.
696 * Item specification.
697 * @param[in] item_flags
701 flow_verbs_translate_item_udp(struct mlx5_flow *dev_flow,
702 const struct rte_flow_item *item,
703 uint64_t item_flags __rte_unused)
705 const struct rte_flow_item_udp *spec = item->spec;
706 const struct rte_flow_item_udp *mask = item->mask;
707 unsigned int size = sizeof(struct ibv_flow_spec_tcp_udp);
708 struct ibv_flow_spec_tcp_udp udp = {
709 .type = IBV_FLOW_SPEC_UDP | VERBS_SPEC_INNER(item_flags),
714 mask = &rte_flow_item_udp_mask;
716 udp.val.dst_port = spec->hdr.dst_port;
717 udp.val.src_port = spec->hdr.src_port;
718 udp.mask.dst_port = mask->hdr.dst_port;
719 udp.mask.src_port = mask->hdr.src_port;
720 /* Remove unwanted bits from values. */
721 udp.val.src_port &= udp.mask.src_port;
722 udp.val.dst_port &= udp.mask.dst_port;
725 while (item->type == RTE_FLOW_ITEM_TYPE_VOID)
727 if (!(udp.val.dst_port & udp.mask.dst_port)) {
728 switch ((item)->type) {
729 case RTE_FLOW_ITEM_TYPE_VXLAN:
730 udp.val.dst_port = htons(MLX5_UDP_PORT_VXLAN);
731 udp.mask.dst_port = 0xffff;
733 case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
734 udp.val.dst_port = htons(MLX5_UDP_PORT_VXLAN_GPE);
735 udp.mask.dst_port = 0xffff;
737 case RTE_FLOW_ITEM_TYPE_MPLS:
738 udp.val.dst_port = htons(MLX5_UDP_PORT_MPLS);
739 udp.mask.dst_port = 0xffff;
746 flow_verbs_spec_add(&dev_flow->verbs, &udp, size);
750 * Convert the @p item into a Verbs specification. This function assumes that
751 * the input is valid and that there is space to insert the requested item
754 * @param[in, out] dev_flow
755 * Pointer to dev_flow structure.
757 * Item specification.
758 * @param[in] item_flags
762 flow_verbs_translate_item_vxlan(struct mlx5_flow *dev_flow,
763 const struct rte_flow_item *item,
764 uint64_t item_flags __rte_unused)
766 const struct rte_flow_item_vxlan *spec = item->spec;
767 const struct rte_flow_item_vxlan *mask = item->mask;
768 unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
769 struct ibv_flow_spec_tunnel vxlan = {
770 .type = IBV_FLOW_SPEC_VXLAN_TUNNEL,
776 } id = { .vlan_id = 0, };
779 mask = &rte_flow_item_vxlan_mask;
781 memcpy(&id.vni[1], spec->vni, 3);
782 vxlan.val.tunnel_id = id.vlan_id;
783 memcpy(&id.vni[1], mask->vni, 3);
784 vxlan.mask.tunnel_id = id.vlan_id;
785 /* Remove unwanted bits from values. */
786 vxlan.val.tunnel_id &= vxlan.mask.tunnel_id;
788 flow_verbs_spec_add(&dev_flow->verbs, &vxlan, size);
792 * Convert the @p item into a Verbs specification. This function assumes that
793 * the input is valid and that there is space to insert the requested item
796 * @param[in, out] dev_flow
797 * Pointer to dev_flow structure.
799 * Item specification.
800 * @param[in] item_flags
804 flow_verbs_translate_item_vxlan_gpe(struct mlx5_flow *dev_flow,
805 const struct rte_flow_item *item,
806 uint64_t item_flags __rte_unused)
808 const struct rte_flow_item_vxlan_gpe *spec = item->spec;
809 const struct rte_flow_item_vxlan_gpe *mask = item->mask;
810 unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
811 struct ibv_flow_spec_tunnel vxlan_gpe = {
812 .type = IBV_FLOW_SPEC_VXLAN_TUNNEL,
818 } id = { .vlan_id = 0, };
821 mask = &rte_flow_item_vxlan_gpe_mask;
823 memcpy(&id.vni[1], spec->vni, 3);
824 vxlan_gpe.val.tunnel_id = id.vlan_id;
825 memcpy(&id.vni[1], mask->vni, 3);
826 vxlan_gpe.mask.tunnel_id = id.vlan_id;
827 /* Remove unwanted bits from values. */
828 vxlan_gpe.val.tunnel_id &= vxlan_gpe.mask.tunnel_id;
830 flow_verbs_spec_add(&dev_flow->verbs, &vxlan_gpe, size);
834 * Update the protocol in Verbs IPv4/IPv6 spec.
836 * @param[in, out] attr
837 * Pointer to Verbs attributes structure.
839 * Specification type to search in order to update the IP protocol.
840 * @param[in] protocol
841 * Protocol value to set if none is present in the specification.
844 flow_verbs_item_gre_ip_protocol_update(struct ibv_flow_attr *attr,
845 enum ibv_flow_spec_type search,
849 struct ibv_spec_header *hdr = (struct ibv_spec_header *)
850 ((uint8_t *)attr + sizeof(struct ibv_flow_attr));
854 for (i = 0; i != attr->num_of_specs; ++i) {
855 if (hdr->type == search) {
857 struct ibv_flow_spec_ipv4_ext *ipv4;
858 struct ibv_flow_spec_ipv6 *ipv6;
862 case IBV_FLOW_SPEC_IPV4_EXT:
863 ip.ipv4 = (struct ibv_flow_spec_ipv4_ext *)hdr;
864 if (!ip.ipv4->val.proto) {
865 ip.ipv4->val.proto = protocol;
866 ip.ipv4->mask.proto = 0xff;
869 case IBV_FLOW_SPEC_IPV6:
870 ip.ipv6 = (struct ibv_flow_spec_ipv6 *)hdr;
871 if (!ip.ipv6->val.next_hdr) {
872 ip.ipv6->val.next_hdr = protocol;
873 ip.ipv6->mask.next_hdr = 0xff;
881 hdr = (struct ibv_spec_header *)((uint8_t *)hdr + hdr->size);
886 * Convert the @p item into a Verbs specification. This function assumes that
887 * the input is valid and that there is space to insert the requested item
890 * @param[in, out] dev_flow
891 * Pointer to dev_flow structure.
893 * Item specification.
894 * @param[in] item_flags
898 flow_verbs_translate_item_gre(struct mlx5_flow *dev_flow,
899 const struct rte_flow_item *item __rte_unused,
902 struct mlx5_flow_verbs_workspace *verbs = &dev_flow->verbs;
903 #ifndef HAVE_IBV_DEVICE_MPLS_SUPPORT
904 unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
905 struct ibv_flow_spec_tunnel tunnel = {
906 .type = IBV_FLOW_SPEC_VXLAN_TUNNEL,
910 const struct rte_flow_item_gre *spec = item->spec;
911 const struct rte_flow_item_gre *mask = item->mask;
912 unsigned int size = sizeof(struct ibv_flow_spec_gre);
913 struct ibv_flow_spec_gre tunnel = {
914 .type = IBV_FLOW_SPEC_GRE,
919 mask = &rte_flow_item_gre_mask;
921 tunnel.val.c_ks_res0_ver = spec->c_rsvd0_ver;
922 tunnel.val.protocol = spec->protocol;
923 tunnel.mask.c_ks_res0_ver = mask->c_rsvd0_ver;
924 tunnel.mask.protocol = mask->protocol;
925 /* Remove unwanted bits from values. */
926 tunnel.val.c_ks_res0_ver &= tunnel.mask.c_ks_res0_ver;
927 tunnel.val.protocol &= tunnel.mask.protocol;
928 tunnel.val.key &= tunnel.mask.key;
931 if (item_flags & MLX5_FLOW_LAYER_OUTER_L3_IPV4)
932 flow_verbs_item_gre_ip_protocol_update(&verbs->attr,
933 IBV_FLOW_SPEC_IPV4_EXT,
936 flow_verbs_item_gre_ip_protocol_update(&verbs->attr,
939 flow_verbs_spec_add(verbs, &tunnel, size);
943 * Convert the @p action into a Verbs specification. This function assumes that
944 * the input is valid and that there is space to insert the requested action
945 * into the flow. This function also return the action that was added.
947 * @param[in, out] dev_flow
948 * Pointer to dev_flow structure.
950 * Item specification.
951 * @param[in] item_flags
955 flow_verbs_translate_item_mpls(struct mlx5_flow *dev_flow __rte_unused,
956 const struct rte_flow_item *item __rte_unused,
957 uint64_t item_flags __rte_unused)
959 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
960 const struct rte_flow_item_mpls *spec = item->spec;
961 const struct rte_flow_item_mpls *mask = item->mask;
962 unsigned int size = sizeof(struct ibv_flow_spec_mpls);
963 struct ibv_flow_spec_mpls mpls = {
964 .type = IBV_FLOW_SPEC_MPLS,
969 mask = &rte_flow_item_mpls_mask;
971 memcpy(&mpls.val.label, spec, sizeof(mpls.val.label));
972 memcpy(&mpls.mask.label, mask, sizeof(mpls.mask.label));
973 /* Remove unwanted bits from values. */
974 mpls.val.label &= mpls.mask.label;
976 flow_verbs_spec_add(&dev_flow->verbs, &mpls, size);
981 * Convert the @p action into a Verbs specification. This function assumes that
982 * the input is valid and that there is space to insert the requested action
985 * @param[in] dev_flow
986 * Pointer to mlx5_flow.
988 * Action configuration.
991 flow_verbs_translate_action_drop
992 (struct mlx5_flow *dev_flow,
993 const struct rte_flow_action *action __rte_unused)
995 unsigned int size = sizeof(struct ibv_flow_spec_action_drop);
996 struct ibv_flow_spec_action_drop drop = {
997 .type = IBV_FLOW_SPEC_ACTION_DROP,
1001 flow_verbs_spec_add(&dev_flow->verbs, &drop, size);
1005 * Convert the @p action into a Verbs specification. This function assumes that
1006 * the input is valid and that there is space to insert the requested action
1009 * @param[in] rss_desc
1010 * Pointer to mlx5_flow_rss_desc.
1012 * Action configuration.
1015 flow_verbs_translate_action_queue(struct mlx5_flow_rss_desc *rss_desc,
1016 const struct rte_flow_action *action)
1018 const struct rte_flow_action_queue *queue = action->conf;
1020 rss_desc->queue[0] = queue->index;
1021 rss_desc->queue_num = 1;
1025 * Convert the @p action into a Verbs specification. This function assumes that
1026 * the input is valid and that there is space to insert the requested action
1029 * @param[in] rss_desc
1030 * Pointer to mlx5_flow_rss_desc.
1032 * Action configuration.
1035 flow_verbs_translate_action_rss(struct mlx5_flow_rss_desc *rss_desc,
1036 const struct rte_flow_action *action)
1038 const struct rte_flow_action_rss *rss = action->conf;
1039 const uint8_t *rss_key;
1041 memcpy(rss_desc->queue, rss->queue, rss->queue_num * sizeof(uint16_t));
1042 rss_desc->queue_num = rss->queue_num;
1043 /* NULL RSS key indicates default RSS key. */
1044 rss_key = !rss->key ? rss_hash_default_key : rss->key;
1045 memcpy(rss_desc->key, rss_key, MLX5_RSS_HASH_KEY_LEN);
1047 * rss->level and rss.types should be set in advance when expanding
1053 * Convert the @p action into a Verbs specification. This function assumes that
1054 * the input is valid and that there is space to insert the requested action
1057 * @param[in] dev_flow
1058 * Pointer to mlx5_flow.
1060 * Action configuration.
1063 flow_verbs_translate_action_flag
1064 (struct mlx5_flow *dev_flow,
1065 const struct rte_flow_action *action __rte_unused)
1067 unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
1068 struct ibv_flow_spec_action_tag tag = {
1069 .type = IBV_FLOW_SPEC_ACTION_TAG,
1071 .tag_id = mlx5_flow_mark_set(MLX5_FLOW_MARK_DEFAULT),
1074 flow_verbs_spec_add(&dev_flow->verbs, &tag, size);
1078 * Convert the @p action into a Verbs specification. This function assumes that
1079 * the input is valid and that there is space to insert the requested action
1082 * @param[in] dev_flow
1083 * Pointer to mlx5_flow.
1085 * Action configuration.
1088 flow_verbs_translate_action_mark(struct mlx5_flow *dev_flow,
1089 const struct rte_flow_action *action)
1091 const struct rte_flow_action_mark *mark = action->conf;
1092 unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
1093 struct ibv_flow_spec_action_tag tag = {
1094 .type = IBV_FLOW_SPEC_ACTION_TAG,
1096 .tag_id = mlx5_flow_mark_set(mark->id),
1099 flow_verbs_spec_add(&dev_flow->verbs, &tag, size);
1103 * Convert the @p action into a Verbs specification. This function assumes that
1104 * the input is valid and that there is space to insert the requested action
1108 * Pointer to the Ethernet device structure.
1110 * Action configuration.
1111 * @param[in] dev_flow
1112 * Pointer to mlx5_flow.
1114 * Pointer to error structure.
1117 * 0 On success else a negative errno value is returned and rte_errno is set.
1120 flow_verbs_translate_action_count(struct mlx5_flow *dev_flow,
1121 const struct rte_flow_action *action,
1122 struct rte_eth_dev *dev,
1123 struct rte_flow_error *error)
1125 const struct rte_flow_action_count *count = action->conf;
1126 struct rte_flow *flow = dev_flow->flow;
1127 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) || \
1128 defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1129 struct mlx5_flow_counter_pool *pool;
1130 struct mlx5_flow_counter *cnt = NULL;
1131 unsigned int size = sizeof(struct ibv_flow_spec_counter_action);
1132 struct ibv_flow_spec_counter_action counter = {
1133 .type = IBV_FLOW_SPEC_ACTION_COUNT,
1138 if (!flow->counter) {
1139 flow->counter = flow_verbs_counter_new(dev, count->id);
1141 return rte_flow_error_set(error, rte_errno,
1142 RTE_FLOW_ERROR_TYPE_ACTION,
1144 "cannot get counter"
1147 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42)
1148 cnt = flow_verbs_counter_get_by_idx(dev, flow->counter, &pool);
1149 counter.counter_set_handle =
1150 ((struct ibv_counter_set *)cnt->dcs_when_active)->handle;
1151 flow_verbs_spec_add(&dev_flow->verbs, &counter, size);
1152 #elif defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1153 cnt = flow_verbs_counter_get_by_idx(dev, flow->counter, &pool);
1154 counter.counters = (struct ibv_counters *)cnt->dcs_when_active;
1155 flow_verbs_spec_add(&dev_flow->verbs, &counter, size);
1161 * Internal validation function. For validating both actions and items.
1164 * Pointer to the Ethernet device structure.
1166 * Pointer to the flow attributes.
1168 * Pointer to the list of items.
1169 * @param[in] actions
1170 * Pointer to the list of actions.
1171 * @param[in] external
1172 * This flow rule is created by request external to PMD.
1173 * @param[in] hairpin
1174 * Number of hairpin TX actions, 0 means classic flow.
1176 * Pointer to the error structure.
1179 * 0 on success, a negative errno value otherwise and rte_errno is set.
1182 flow_verbs_validate(struct rte_eth_dev *dev,
1183 const struct rte_flow_attr *attr,
1184 const struct rte_flow_item items[],
1185 const struct rte_flow_action actions[],
1186 bool external __rte_unused,
1187 int hairpin __rte_unused,
1188 struct rte_flow_error *error)
1191 uint64_t action_flags = 0;
1192 uint64_t item_flags = 0;
1193 uint64_t last_item = 0;
1194 uint8_t next_protocol = 0xff;
1195 uint16_t ether_type = 0;
1196 bool is_empty_vlan = false;
1197 uint16_t udp_dport = 0;
1201 ret = mlx5_flow_validate_attributes(dev, attr, error);
1204 for (; items->type != RTE_FLOW_ITEM_TYPE_END; items++) {
1205 int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
1208 switch (items->type) {
1209 case RTE_FLOW_ITEM_TYPE_VOID:
1211 case RTE_FLOW_ITEM_TYPE_ETH:
1212 ret = mlx5_flow_validate_item_eth(items, item_flags,
1216 last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
1217 MLX5_FLOW_LAYER_OUTER_L2;
1218 if (items->mask != NULL && items->spec != NULL) {
1220 ((const struct rte_flow_item_eth *)
1223 ((const struct rte_flow_item_eth *)
1225 if (ether_type == RTE_BE16(RTE_ETHER_TYPE_VLAN))
1226 is_empty_vlan = true;
1227 ether_type = rte_be_to_cpu_16(ether_type);
1232 case RTE_FLOW_ITEM_TYPE_VLAN:
1233 ret = mlx5_flow_validate_item_vlan(items, item_flags,
1237 last_item = tunnel ? (MLX5_FLOW_LAYER_INNER_L2 |
1238 MLX5_FLOW_LAYER_INNER_VLAN) :
1239 (MLX5_FLOW_LAYER_OUTER_L2 |
1240 MLX5_FLOW_LAYER_OUTER_VLAN);
1241 if (items->mask != NULL && items->spec != NULL) {
1243 ((const struct rte_flow_item_vlan *)
1244 items->spec)->inner_type;
1246 ((const struct rte_flow_item_vlan *)
1247 items->mask)->inner_type;
1248 ether_type = rte_be_to_cpu_16(ether_type);
1252 is_empty_vlan = false;
1254 case RTE_FLOW_ITEM_TYPE_IPV4:
1255 ret = mlx5_flow_validate_item_ipv4
1257 last_item, ether_type, NULL,
1258 MLX5_ITEM_RANGE_NOT_ACCEPTED,
1262 last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV4 :
1263 MLX5_FLOW_LAYER_OUTER_L3_IPV4;
1264 if (items->mask != NULL &&
1265 ((const struct rte_flow_item_ipv4 *)
1266 items->mask)->hdr.next_proto_id) {
1268 ((const struct rte_flow_item_ipv4 *)
1269 (items->spec))->hdr.next_proto_id;
1271 ((const struct rte_flow_item_ipv4 *)
1272 (items->mask))->hdr.next_proto_id;
1274 /* Reset for inner layer. */
1275 next_protocol = 0xff;
1278 case RTE_FLOW_ITEM_TYPE_IPV6:
1279 ret = mlx5_flow_validate_item_ipv6(items, item_flags,
1285 last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV6 :
1286 MLX5_FLOW_LAYER_OUTER_L3_IPV6;
1287 if (items->mask != NULL &&
1288 ((const struct rte_flow_item_ipv6 *)
1289 items->mask)->hdr.proto) {
1291 ((const struct rte_flow_item_ipv6 *)
1292 items->spec)->hdr.proto;
1294 ((const struct rte_flow_item_ipv6 *)
1295 items->mask)->hdr.proto;
1297 /* Reset for inner layer. */
1298 next_protocol = 0xff;
1301 case RTE_FLOW_ITEM_TYPE_UDP:
1302 ret = mlx5_flow_validate_item_udp(items, item_flags,
1305 const struct rte_flow_item_udp *spec = items->spec;
1306 const struct rte_flow_item_udp *mask = items->mask;
1308 mask = &rte_flow_item_udp_mask;
1310 udp_dport = rte_be_to_cpu_16
1311 (spec->hdr.dst_port &
1312 mask->hdr.dst_port);
1316 last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L4_UDP :
1317 MLX5_FLOW_LAYER_OUTER_L4_UDP;
1319 case RTE_FLOW_ITEM_TYPE_TCP:
1320 ret = mlx5_flow_validate_item_tcp
1323 &rte_flow_item_tcp_mask,
1327 last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L4_TCP :
1328 MLX5_FLOW_LAYER_OUTER_L4_TCP;
1330 case RTE_FLOW_ITEM_TYPE_VXLAN:
1331 ret = mlx5_flow_validate_item_vxlan(dev, udp_dport,
1336 last_item = MLX5_FLOW_LAYER_VXLAN;
1338 case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
1339 ret = mlx5_flow_validate_item_vxlan_gpe(items,
1344 last_item = MLX5_FLOW_LAYER_VXLAN_GPE;
1346 case RTE_FLOW_ITEM_TYPE_GRE:
1347 ret = mlx5_flow_validate_item_gre(items, item_flags,
1348 next_protocol, error);
1351 last_item = MLX5_FLOW_LAYER_GRE;
1353 case RTE_FLOW_ITEM_TYPE_MPLS:
1354 ret = mlx5_flow_validate_item_mpls(dev, items,
1359 last_item = MLX5_FLOW_LAYER_MPLS;
1361 case RTE_FLOW_ITEM_TYPE_ICMP:
1362 case RTE_FLOW_ITEM_TYPE_ICMP6:
1363 return rte_flow_error_set(error, ENOTSUP,
1364 RTE_FLOW_ERROR_TYPE_ITEM,
1366 "item not supported");
1368 return rte_flow_error_set(error, ENOTSUP,
1369 RTE_FLOW_ERROR_TYPE_ITEM,
1370 NULL, "item not supported");
1372 item_flags |= last_item;
1375 return rte_flow_error_set(error, ENOTSUP,
1376 RTE_FLOW_ERROR_TYPE_ITEM, NULL,
1377 "VLAN matching without vid specification is not supported");
1378 for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
1379 switch (actions->type) {
1380 case RTE_FLOW_ACTION_TYPE_VOID:
1382 case RTE_FLOW_ACTION_TYPE_FLAG:
1383 ret = mlx5_flow_validate_action_flag(action_flags,
1388 action_flags |= MLX5_FLOW_ACTION_FLAG;
1390 case RTE_FLOW_ACTION_TYPE_MARK:
1391 ret = mlx5_flow_validate_action_mark(actions,
1397 action_flags |= MLX5_FLOW_ACTION_MARK;
1399 case RTE_FLOW_ACTION_TYPE_DROP:
1400 ret = mlx5_flow_validate_action_drop(action_flags,
1405 action_flags |= MLX5_FLOW_ACTION_DROP;
1407 case RTE_FLOW_ACTION_TYPE_QUEUE:
1408 ret = mlx5_flow_validate_action_queue(actions,
1414 action_flags |= MLX5_FLOW_ACTION_QUEUE;
1416 case RTE_FLOW_ACTION_TYPE_RSS:
1417 ret = mlx5_flow_validate_action_rss(actions,
1423 action_flags |= MLX5_FLOW_ACTION_RSS;
1425 case RTE_FLOW_ACTION_TYPE_COUNT:
1426 ret = mlx5_flow_validate_action_count(dev, attr, error);
1429 action_flags |= MLX5_FLOW_ACTION_COUNT;
1432 return rte_flow_error_set(error, ENOTSUP,
1433 RTE_FLOW_ERROR_TYPE_ACTION,
1435 "action not supported");
1439 * Validate the drop action mutual exclusion with other actions.
1440 * Drop action is mutually-exclusive with any other action, except for
1443 if ((action_flags & MLX5_FLOW_ACTION_DROP) &&
1444 (action_flags & ~(MLX5_FLOW_ACTION_DROP | MLX5_FLOW_ACTION_COUNT)))
1445 return rte_flow_error_set(error, EINVAL,
1446 RTE_FLOW_ERROR_TYPE_ACTION, NULL,
1447 "Drop action is mutually-exclusive "
1448 "with any other action, except for "
1450 if (!(action_flags & MLX5_FLOW_FATE_ACTIONS))
1451 return rte_flow_error_set(error, EINVAL,
1452 RTE_FLOW_ERROR_TYPE_ACTION, actions,
1453 "no fate action is found");
1458 * Calculate the required bytes that are needed for the action part of the verbs
1461 * @param[in] actions
1462 * Pointer to the list of actions.
1465 * The size of the memory needed for all actions.
1468 flow_verbs_get_actions_size(const struct rte_flow_action actions[])
1472 for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
1473 switch (actions->type) {
1474 case RTE_FLOW_ACTION_TYPE_VOID:
1476 case RTE_FLOW_ACTION_TYPE_FLAG:
1477 size += sizeof(struct ibv_flow_spec_action_tag);
1479 case RTE_FLOW_ACTION_TYPE_MARK:
1480 size += sizeof(struct ibv_flow_spec_action_tag);
1482 case RTE_FLOW_ACTION_TYPE_DROP:
1483 size += sizeof(struct ibv_flow_spec_action_drop);
1485 case RTE_FLOW_ACTION_TYPE_QUEUE:
1487 case RTE_FLOW_ACTION_TYPE_RSS:
1489 case RTE_FLOW_ACTION_TYPE_COUNT:
1490 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) || \
1491 defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1492 size += sizeof(struct ibv_flow_spec_counter_action);
1503 * Calculate the required bytes that are needed for the item part of the verbs
1507 * Pointer to the list of items.
1510 * The size of the memory needed for all items.
1513 flow_verbs_get_items_size(const struct rte_flow_item items[])
1517 for (; items->type != RTE_FLOW_ITEM_TYPE_END; items++) {
1518 switch (items->type) {
1519 case RTE_FLOW_ITEM_TYPE_VOID:
1521 case RTE_FLOW_ITEM_TYPE_ETH:
1522 size += sizeof(struct ibv_flow_spec_eth);
1524 case RTE_FLOW_ITEM_TYPE_VLAN:
1525 size += sizeof(struct ibv_flow_spec_eth);
1527 case RTE_FLOW_ITEM_TYPE_IPV4:
1528 size += sizeof(struct ibv_flow_spec_ipv4_ext);
1530 case RTE_FLOW_ITEM_TYPE_IPV6:
1531 size += sizeof(struct ibv_flow_spec_ipv6);
1533 case RTE_FLOW_ITEM_TYPE_UDP:
1534 size += sizeof(struct ibv_flow_spec_tcp_udp);
1536 case RTE_FLOW_ITEM_TYPE_TCP:
1537 size += sizeof(struct ibv_flow_spec_tcp_udp);
1539 case RTE_FLOW_ITEM_TYPE_VXLAN:
1540 size += sizeof(struct ibv_flow_spec_tunnel);
1542 case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
1543 size += sizeof(struct ibv_flow_spec_tunnel);
1545 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1546 case RTE_FLOW_ITEM_TYPE_GRE:
1547 size += sizeof(struct ibv_flow_spec_gre);
1549 case RTE_FLOW_ITEM_TYPE_MPLS:
1550 size += sizeof(struct ibv_flow_spec_mpls);
1553 case RTE_FLOW_ITEM_TYPE_GRE:
1554 size += sizeof(struct ibv_flow_spec_tunnel);
1565 * Internal preparation function. Allocate mlx5_flow with the required size.
1566 * The required size is calculate based on the actions and items. This function
1567 * also returns the detected actions and items for later use.
1570 * Pointer to Ethernet device.
1572 * Pointer to the flow attributes.
1574 * Pointer to the list of items.
1575 * @param[in] actions
1576 * Pointer to the list of actions.
1578 * Pointer to the error structure.
1581 * Pointer to mlx5_flow object on success, otherwise NULL and rte_errno
1584 static struct mlx5_flow *
1585 flow_verbs_prepare(struct rte_eth_dev *dev,
1586 const struct rte_flow_attr *attr __rte_unused,
1587 const struct rte_flow_item items[],
1588 const struct rte_flow_action actions[],
1589 struct rte_flow_error *error)
1592 uint32_t handle_idx = 0;
1593 struct mlx5_flow *dev_flow;
1594 struct mlx5_flow_handle *dev_handle;
1595 struct mlx5_priv *priv = dev->data->dev_private;
1596 struct mlx5_flow_workspace *wks = mlx5_flow_get_thread_workspace();
1599 size += flow_verbs_get_actions_size(actions);
1600 size += flow_verbs_get_items_size(items);
1601 if (size > MLX5_VERBS_MAX_SPEC_ACT_SIZE) {
1602 rte_flow_error_set(error, E2BIG,
1603 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1604 "Verbs spec/action size too large");
1607 /* In case of corrupting the memory. */
1608 if (wks->flow_idx >= MLX5_NUM_MAX_DEV_FLOWS) {
1609 rte_flow_error_set(error, ENOSPC,
1610 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1611 "not free temporary device flow");
1614 dev_handle = mlx5_ipool_zmalloc(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW],
1617 rte_flow_error_set(error, ENOMEM,
1618 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1619 "not enough memory to create flow handle");
1622 MLX5_ASSERT(wks->flow_idx + 1 < RTE_DIM(wks->flows));
1623 dev_flow = &wks->flows[wks->flow_idx++];
1624 dev_flow->handle = dev_handle;
1625 dev_flow->handle_idx = handle_idx;
1626 /* Memcpy is used, only size needs to be cleared to 0. */
1627 dev_flow->verbs.size = 0;
1628 dev_flow->verbs.attr.num_of_specs = 0;
1629 dev_flow->ingress = attr->ingress;
1630 dev_flow->hash_fields = 0;
1631 /* Need to set transfer attribute: not supported in Verbs mode. */
1636 * Fill the flow with verb spec.
1639 * Pointer to Ethernet device.
1640 * @param[in, out] dev_flow
1641 * Pointer to the mlx5 flow.
1643 * Pointer to the flow attributes.
1645 * Pointer to the list of items.
1646 * @param[in] actions
1647 * Pointer to the list of actions.
1649 * Pointer to the error structure.
1652 * 0 on success, else a negative errno value otherwise and rte_errno is set.
1655 flow_verbs_translate(struct rte_eth_dev *dev,
1656 struct mlx5_flow *dev_flow,
1657 const struct rte_flow_attr *attr,
1658 const struct rte_flow_item items[],
1659 const struct rte_flow_action actions[],
1660 struct rte_flow_error *error)
1662 uint64_t item_flags = 0;
1663 uint64_t action_flags = 0;
1664 uint64_t priority = attr->priority;
1665 uint32_t subpriority = 0;
1666 struct mlx5_priv *priv = dev->data->dev_private;
1667 struct mlx5_flow_workspace *wks = mlx5_flow_get_thread_workspace();
1668 struct mlx5_flow_rss_desc *rss_desc;
1671 rss_desc = &wks->rss_desc;
1672 if (priority == MLX5_FLOW_LOWEST_PRIO_INDICATOR)
1673 priority = priv->sh->flow_max_priority - 1;
1674 for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
1677 switch (actions->type) {
1678 case RTE_FLOW_ACTION_TYPE_VOID:
1680 case RTE_FLOW_ACTION_TYPE_FLAG:
1681 flow_verbs_translate_action_flag(dev_flow, actions);
1682 action_flags |= MLX5_FLOW_ACTION_FLAG;
1683 dev_flow->handle->mark = 1;
1685 case RTE_FLOW_ACTION_TYPE_MARK:
1686 flow_verbs_translate_action_mark(dev_flow, actions);
1687 action_flags |= MLX5_FLOW_ACTION_MARK;
1688 dev_flow->handle->mark = 1;
1690 case RTE_FLOW_ACTION_TYPE_DROP:
1691 flow_verbs_translate_action_drop(dev_flow, actions);
1692 action_flags |= MLX5_FLOW_ACTION_DROP;
1693 dev_flow->handle->fate_action = MLX5_FLOW_FATE_DROP;
1695 case RTE_FLOW_ACTION_TYPE_QUEUE:
1696 flow_verbs_translate_action_queue(rss_desc, actions);
1697 action_flags |= MLX5_FLOW_ACTION_QUEUE;
1698 dev_flow->handle->fate_action = MLX5_FLOW_FATE_QUEUE;
1700 case RTE_FLOW_ACTION_TYPE_RSS:
1701 flow_verbs_translate_action_rss(rss_desc, actions);
1702 action_flags |= MLX5_FLOW_ACTION_RSS;
1703 dev_flow->handle->fate_action = MLX5_FLOW_FATE_QUEUE;
1705 case RTE_FLOW_ACTION_TYPE_COUNT:
1706 ret = flow_verbs_translate_action_count(dev_flow,
1711 action_flags |= MLX5_FLOW_ACTION_COUNT;
1714 return rte_flow_error_set(error, ENOTSUP,
1715 RTE_FLOW_ERROR_TYPE_ACTION,
1717 "action not supported");
1720 dev_flow->act_flags = action_flags;
1721 for (; items->type != RTE_FLOW_ITEM_TYPE_END; items++) {
1722 int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
1724 switch (items->type) {
1725 case RTE_FLOW_ITEM_TYPE_VOID:
1727 case RTE_FLOW_ITEM_TYPE_ETH:
1728 flow_verbs_translate_item_eth(dev_flow, items,
1730 subpriority = MLX5_PRIORITY_MAP_L2;
1731 item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
1732 MLX5_FLOW_LAYER_OUTER_L2;
1734 case RTE_FLOW_ITEM_TYPE_VLAN:
1735 flow_verbs_translate_item_vlan(dev_flow, items,
1737 subpriority = MLX5_PRIORITY_MAP_L2;
1738 item_flags |= tunnel ? (MLX5_FLOW_LAYER_INNER_L2 |
1739 MLX5_FLOW_LAYER_INNER_VLAN) :
1740 (MLX5_FLOW_LAYER_OUTER_L2 |
1741 MLX5_FLOW_LAYER_OUTER_VLAN);
1743 case RTE_FLOW_ITEM_TYPE_IPV4:
1744 flow_verbs_translate_item_ipv4(dev_flow, items,
1746 subpriority = MLX5_PRIORITY_MAP_L3;
1747 dev_flow->hash_fields |=
1748 mlx5_flow_hashfields_adjust
1750 MLX5_IPV4_LAYER_TYPES,
1751 MLX5_IPV4_IBV_RX_HASH);
1752 item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV4 :
1753 MLX5_FLOW_LAYER_OUTER_L3_IPV4;
1755 case RTE_FLOW_ITEM_TYPE_IPV6:
1756 flow_verbs_translate_item_ipv6(dev_flow, items,
1758 subpriority = MLX5_PRIORITY_MAP_L3;
1759 dev_flow->hash_fields |=
1760 mlx5_flow_hashfields_adjust
1762 MLX5_IPV6_LAYER_TYPES,
1763 MLX5_IPV6_IBV_RX_HASH);
1764 item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV6 :
1765 MLX5_FLOW_LAYER_OUTER_L3_IPV6;
1767 case RTE_FLOW_ITEM_TYPE_TCP:
1768 flow_verbs_translate_item_tcp(dev_flow, items,
1770 subpriority = MLX5_PRIORITY_MAP_L4;
1771 if (dev_flow->hash_fields != 0)
1772 dev_flow->hash_fields |=
1773 mlx5_flow_hashfields_adjust
1774 (rss_desc, tunnel, RTE_ETH_RSS_TCP,
1775 (IBV_RX_HASH_SRC_PORT_TCP |
1776 IBV_RX_HASH_DST_PORT_TCP));
1777 item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L4_TCP :
1778 MLX5_FLOW_LAYER_OUTER_L4_TCP;
1780 case RTE_FLOW_ITEM_TYPE_UDP:
1781 flow_verbs_translate_item_udp(dev_flow, items,
1783 subpriority = MLX5_PRIORITY_MAP_L4;
1784 if (dev_flow->hash_fields != 0)
1785 dev_flow->hash_fields |=
1786 mlx5_flow_hashfields_adjust
1787 (rss_desc, tunnel, RTE_ETH_RSS_UDP,
1788 (IBV_RX_HASH_SRC_PORT_UDP |
1789 IBV_RX_HASH_DST_PORT_UDP));
1790 item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L4_UDP :
1791 MLX5_FLOW_LAYER_OUTER_L4_UDP;
1793 case RTE_FLOW_ITEM_TYPE_VXLAN:
1794 flow_verbs_translate_item_vxlan(dev_flow, items,
1796 subpriority = MLX5_TUNNEL_PRIO_GET(rss_desc);
1797 item_flags |= MLX5_FLOW_LAYER_VXLAN;
1799 case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
1800 flow_verbs_translate_item_vxlan_gpe(dev_flow, items,
1802 subpriority = MLX5_TUNNEL_PRIO_GET(rss_desc);
1803 item_flags |= MLX5_FLOW_LAYER_VXLAN_GPE;
1805 case RTE_FLOW_ITEM_TYPE_GRE:
1806 flow_verbs_translate_item_gre(dev_flow, items,
1808 subpriority = MLX5_TUNNEL_PRIO_GET(rss_desc);
1809 item_flags |= MLX5_FLOW_LAYER_GRE;
1811 case RTE_FLOW_ITEM_TYPE_MPLS:
1812 flow_verbs_translate_item_mpls(dev_flow, items,
1814 subpriority = MLX5_TUNNEL_PRIO_GET(rss_desc);
1815 item_flags |= MLX5_FLOW_LAYER_MPLS;
1818 return rte_flow_error_set(error, ENOTSUP,
1819 RTE_FLOW_ERROR_TYPE_ITEM,
1820 NULL, "item not supported");
1823 dev_flow->handle->layers = item_flags;
1824 /* Other members of attr will be ignored. */
1825 dev_flow->verbs.attr.priority =
1826 mlx5_flow_adjust_priority(dev, priority, subpriority);
1827 dev_flow->verbs.attr.port = (uint8_t)priv->dev_port;
1832 * Remove the flow from the NIC but keeps it in memory.
1835 * Pointer to the Ethernet device structure.
1836 * @param[in, out] flow
1837 * Pointer to flow structure.
1840 flow_verbs_remove(struct rte_eth_dev *dev, struct rte_flow *flow)
1842 struct mlx5_priv *priv = dev->data->dev_private;
1843 struct mlx5_flow_handle *handle;
1844 uint32_t handle_idx;
1848 SILIST_FOREACH(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW], flow->dev_handles,
1849 handle_idx, handle, next) {
1850 if (handle->drv_flow) {
1851 claim_zero(mlx5_glue->destroy_flow(handle->drv_flow));
1852 handle->drv_flow = NULL;
1854 /* hrxq is union, don't touch it only the flag is set. */
1855 if (handle->rix_hrxq &&
1856 handle->fate_action == MLX5_FLOW_FATE_QUEUE) {
1857 mlx5_hrxq_release(dev, handle->rix_hrxq);
1858 handle->rix_hrxq = 0;
1860 if (handle->vf_vlan.tag && handle->vf_vlan.created)
1861 mlx5_vlan_vmwa_release(dev, &handle->vf_vlan);
1866 * Remove the flow from the NIC and the memory.
1869 * Pointer to the Ethernet device structure.
1870 * @param[in, out] flow
1871 * Pointer to flow structure.
1874 flow_verbs_destroy(struct rte_eth_dev *dev, struct rte_flow *flow)
1876 struct mlx5_priv *priv = dev->data->dev_private;
1877 struct mlx5_flow_handle *handle;
1881 flow_verbs_remove(dev, flow);
1882 while (flow->dev_handles) {
1883 uint32_t tmp_idx = flow->dev_handles;
1885 handle = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW],
1889 flow->dev_handles = handle->next.next;
1890 mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW],
1893 if (flow->counter) {
1894 flow_verbs_counter_release(dev, flow->counter);
1900 * Apply the flow to the NIC.
1903 * Pointer to the Ethernet device structure.
1904 * @param[in, out] flow
1905 * Pointer to flow structure.
1907 * Pointer to error structure.
1910 * 0 on success, a negative errno value otherwise and rte_errno is set.
1913 flow_verbs_apply(struct rte_eth_dev *dev, struct rte_flow *flow,
1914 struct rte_flow_error *error)
1916 struct mlx5_priv *priv = dev->data->dev_private;
1917 struct mlx5_flow_handle *handle;
1918 struct mlx5_flow *dev_flow;
1919 struct mlx5_hrxq *hrxq;
1920 uint32_t dev_handles;
1923 struct mlx5_flow_workspace *wks = mlx5_flow_get_thread_workspace();
1926 for (idx = wks->flow_idx - 1; idx >= 0; idx--) {
1927 dev_flow = &wks->flows[idx];
1928 handle = dev_flow->handle;
1929 if (handle->fate_action == MLX5_FLOW_FATE_DROP) {
1930 MLX5_ASSERT(priv->drop_queue.hrxq);
1931 hrxq = priv->drop_queue.hrxq;
1934 struct mlx5_flow_rss_desc *rss_desc = &wks->rss_desc;
1936 MLX5_ASSERT(rss_desc->queue_num);
1937 rss_desc->key_len = MLX5_RSS_HASH_KEY_LEN;
1938 rss_desc->hash_fields = dev_flow->hash_fields;
1939 rss_desc->tunnel = !!(handle->layers &
1940 MLX5_FLOW_LAYER_TUNNEL);
1941 rss_desc->shared_rss = 0;
1942 hrxq_idx = mlx5_hrxq_get(dev, rss_desc);
1943 hrxq = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_HRXQ],
1948 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1949 "cannot get hash queue");
1952 handle->rix_hrxq = hrxq_idx;
1955 handle->drv_flow = mlx5_glue->create_flow
1956 (hrxq->qp, &dev_flow->verbs.attr);
1957 if (!handle->drv_flow) {
1958 rte_flow_error_set(error, errno,
1959 RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1961 "hardware refuses to create flow");
1964 if (priv->vmwa_context &&
1965 handle->vf_vlan.tag && !handle->vf_vlan.created) {
1967 * The rule contains the VLAN pattern.
1968 * For VF we are going to create VLAN
1969 * interface to make hypervisor set correct
1970 * e-Switch vport context.
1972 mlx5_vlan_vmwa_acquire(dev, &handle->vf_vlan);
1977 err = rte_errno; /* Save rte_errno before cleanup. */
1978 SILIST_FOREACH(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW], flow->dev_handles,
1979 dev_handles, handle, next) {
1980 /* hrxq is union, don't touch it only the flag is set. */
1981 if (handle->rix_hrxq &&
1982 handle->fate_action == MLX5_FLOW_FATE_QUEUE) {
1983 mlx5_hrxq_release(dev, handle->rix_hrxq);
1984 handle->rix_hrxq = 0;
1986 if (handle->vf_vlan.tag && handle->vf_vlan.created)
1987 mlx5_vlan_vmwa_release(dev, &handle->vf_vlan);
1989 rte_errno = err; /* Restore rte_errno. */
1996 * @see rte_flow_query()
2000 flow_verbs_query(struct rte_eth_dev *dev,
2001 struct rte_flow *flow,
2002 const struct rte_flow_action *actions,
2004 struct rte_flow_error *error)
2008 for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
2009 switch (actions->type) {
2010 case RTE_FLOW_ACTION_TYPE_VOID:
2012 case RTE_FLOW_ACTION_TYPE_COUNT:
2013 ret = flow_verbs_counter_query(dev, flow, data, error);
2016 return rte_flow_error_set(error, ENOTSUP,
2017 RTE_FLOW_ERROR_TYPE_ACTION,
2019 "action not supported");
2026 flow_verbs_sync_domain(struct rte_eth_dev *dev, uint32_t domains,
2030 RTE_SET_USED(domains);
2031 RTE_SET_USED(flags);
2036 const struct mlx5_flow_driver_ops mlx5_flow_verbs_drv_ops = {
2037 .validate = flow_verbs_validate,
2038 .prepare = flow_verbs_prepare,
2039 .translate = flow_verbs_translate,
2040 .apply = flow_verbs_apply,
2041 .remove = flow_verbs_remove,
2042 .destroy = flow_verbs_destroy,
2043 .query = flow_verbs_query,
2044 .sync_domain = flow_verbs_sync_domain,
2045 .discover_priorities = flow_verbs_discover_priorities,