1 /*- 2 * BSD LICENSE 3 * 4 * Copyright 2016 6WIND S.A. 5 * Copyright 2016 Mellanox. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * * Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * * Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * * Neither the name of 6WIND S.A. nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 32 */ 33 34 #include <sys/queue.h> 35 #include <string.h> 36 37 /* Verbs header. */ 38 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */ 39 #ifdef PEDANTIC 40 #pragma GCC diagnostic ignored "-Wpedantic" 41 #endif 42 #include <infiniband/verbs.h> 43 #ifdef PEDANTIC 44 #pragma GCC diagnostic error "-Wpedantic" 45 #endif 46 47 #include <rte_ethdev_driver.h> 48 #include <rte_flow.h> 49 #include <rte_flow_driver.h> 50 #include <rte_malloc.h> 51 #include <rte_ip.h> 52 53 #include "mlx5.h" 54 #include "mlx5_defs.h" 55 #include "mlx5_prm.h" 56 57 /* Define minimal priority for control plane flows. */ 58 #define MLX5_CTRL_FLOW_PRIORITY 4 59 60 /* Internet Protocol versions. */ 61 #define MLX5_IPV4 4 62 #define MLX5_IPV6 6 63 64 #ifndef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT 65 struct ibv_counter_set_init_attr { 66 int dummy; 67 }; 68 struct ibv_flow_spec_counter_action { 69 int dummy; 70 }; 71 struct ibv_counter_set { 72 int dummy; 73 }; 74 75 static inline int 76 ibv_destroy_counter_set(struct ibv_counter_set *cs) 77 { 78 (void)cs; 79 return -ENOTSUP; 80 } 81 #endif 82 83 /* Dev ops structure defined in mlx5.c */ 84 extern const struct eth_dev_ops mlx5_dev_ops; 85 extern const struct eth_dev_ops mlx5_dev_ops_isolate; 86 87 static int 88 mlx5_flow_create_eth(const struct rte_flow_item *item, 89 const void *default_mask, 90 void *data); 91 92 static int 93 mlx5_flow_create_vlan(const struct rte_flow_item *item, 94 const void *default_mask, 95 void *data); 96 97 static int 98 mlx5_flow_create_ipv4(const struct rte_flow_item *item, 99 const void *default_mask, 100 void *data); 101 102 static int 103 mlx5_flow_create_ipv6(const struct rte_flow_item *item, 104 const void *default_mask, 105 void *data); 106 107 static int 108 mlx5_flow_create_udp(const struct rte_flow_item *item, 109 const void *default_mask, 110 void *data); 111 112 static int 113 mlx5_flow_create_tcp(const struct rte_flow_item *item, 114 const void *default_mask, 115 void *data); 116 117 static int 118 mlx5_flow_create_vxlan(const struct rte_flow_item *item, 119 const void *default_mask, 120 void *data); 121 122 struct mlx5_flow_parse; 123 124 static void 125 mlx5_flow_create_copy(struct mlx5_flow_parse *parser, void *src, 126 unsigned int size); 127 128 static int 129 mlx5_flow_create_flag_mark(struct mlx5_flow_parse *parser, uint32_t mark_id); 130 131 static int 132 mlx5_flow_create_count(struct priv *priv, struct mlx5_flow_parse *parser); 133 134 /* Hash RX queue types. */ 135 enum hash_rxq_type { 136 HASH_RXQ_TCPV4, 137 HASH_RXQ_UDPV4, 138 HASH_RXQ_IPV4, 139 HASH_RXQ_TCPV6, 140 HASH_RXQ_UDPV6, 141 HASH_RXQ_IPV6, 142 HASH_RXQ_ETH, 143 }; 144 145 /* Initialization data for hash RX queue. */ 146 struct hash_rxq_init { 147 uint64_t hash_fields; /* Fields that participate in the hash. */ 148 uint64_t dpdk_rss_hf; /* Matching DPDK RSS hash fields. */ 149 unsigned int flow_priority; /* Flow priority to use. */ 150 unsigned int ip_version; /* Internet protocol. */ 151 }; 152 153 /* Initialization data for hash RX queues. */ 154 const struct hash_rxq_init hash_rxq_init[] = { 155 [HASH_RXQ_TCPV4] = { 156 .hash_fields = (IBV_RX_HASH_SRC_IPV4 | 157 IBV_RX_HASH_DST_IPV4 | 158 IBV_RX_HASH_SRC_PORT_TCP | 159 IBV_RX_HASH_DST_PORT_TCP), 160 .dpdk_rss_hf = ETH_RSS_NONFRAG_IPV4_TCP, 161 .flow_priority = 0, 162 .ip_version = MLX5_IPV4, 163 }, 164 [HASH_RXQ_UDPV4] = { 165 .hash_fields = (IBV_RX_HASH_SRC_IPV4 | 166 IBV_RX_HASH_DST_IPV4 | 167 IBV_RX_HASH_SRC_PORT_UDP | 168 IBV_RX_HASH_DST_PORT_UDP), 169 .dpdk_rss_hf = ETH_RSS_NONFRAG_IPV4_UDP, 170 .flow_priority = 0, 171 .ip_version = MLX5_IPV4, 172 }, 173 [HASH_RXQ_IPV4] = { 174 .hash_fields = (IBV_RX_HASH_SRC_IPV4 | 175 IBV_RX_HASH_DST_IPV4), 176 .dpdk_rss_hf = (ETH_RSS_IPV4 | 177 ETH_RSS_FRAG_IPV4), 178 .flow_priority = 1, 179 .ip_version = MLX5_IPV4, 180 }, 181 [HASH_RXQ_TCPV6] = { 182 .hash_fields = (IBV_RX_HASH_SRC_IPV6 | 183 IBV_RX_HASH_DST_IPV6 | 184 IBV_RX_HASH_SRC_PORT_TCP | 185 IBV_RX_HASH_DST_PORT_TCP), 186 .dpdk_rss_hf = ETH_RSS_NONFRAG_IPV6_TCP, 187 .flow_priority = 0, 188 .ip_version = MLX5_IPV6, 189 }, 190 [HASH_RXQ_UDPV6] = { 191 .hash_fields = (IBV_RX_HASH_SRC_IPV6 | 192 IBV_RX_HASH_DST_IPV6 | 193 IBV_RX_HASH_SRC_PORT_UDP | 194 IBV_RX_HASH_DST_PORT_UDP), 195 .dpdk_rss_hf = ETH_RSS_NONFRAG_IPV6_UDP, 196 .flow_priority = 0, 197 .ip_version = MLX5_IPV6, 198 }, 199 [HASH_RXQ_IPV6] = { 200 .hash_fields = (IBV_RX_HASH_SRC_IPV6 | 201 IBV_RX_HASH_DST_IPV6), 202 .dpdk_rss_hf = (ETH_RSS_IPV6 | 203 ETH_RSS_FRAG_IPV6), 204 .flow_priority = 1, 205 .ip_version = MLX5_IPV6, 206 }, 207 [HASH_RXQ_ETH] = { 208 .hash_fields = 0, 209 .dpdk_rss_hf = 0, 210 .flow_priority = 2, 211 }, 212 }; 213 214 /* Number of entries in hash_rxq_init[]. */ 215 const unsigned int hash_rxq_init_n = RTE_DIM(hash_rxq_init); 216 217 /** Structure for holding counter stats. */ 218 struct mlx5_flow_counter_stats { 219 uint64_t hits; /**< Number of packets matched by the rule. */ 220 uint64_t bytes; /**< Number of bytes matched by the rule. */ 221 }; 222 223 /** Structure for Drop queue. */ 224 struct mlx5_hrxq_drop { 225 struct ibv_rwq_ind_table *ind_table; /**< Indirection table. */ 226 struct ibv_qp *qp; /**< Verbs queue pair. */ 227 struct ibv_wq *wq; /**< Verbs work queue. */ 228 struct ibv_cq *cq; /**< Verbs completion queue. */ 229 }; 230 231 /* Flows structures. */ 232 struct mlx5_flow { 233 uint64_t hash_fields; /**< Fields that participate in the hash. */ 234 struct ibv_flow_attr *ibv_attr; /**< Pointer to Verbs attributes. */ 235 struct ibv_flow *ibv_flow; /**< Verbs flow. */ 236 struct mlx5_hrxq *hrxq; /**< Hash Rx queues. */ 237 }; 238 239 /* Drop flows structures. */ 240 struct mlx5_flow_drop { 241 struct ibv_flow_attr *ibv_attr; /**< Pointer to Verbs attributes. */ 242 struct ibv_flow *ibv_flow; /**< Verbs flow. */ 243 }; 244 245 struct rte_flow { 246 TAILQ_ENTRY(rte_flow) next; /**< Pointer to the next flow structure. */ 247 uint32_t mark:1; /**< Set if the flow is marked. */ 248 uint32_t drop:1; /**< Drop queue. */ 249 uint16_t queues_n; /**< Number of entries in queue[]. */ 250 uint16_t (*queues)[]; /**< Queues indexes to use. */ 251 struct rte_eth_rss_conf rss_conf; /**< RSS configuration */ 252 uint8_t rss_key[40]; /**< copy of the RSS key. */ 253 struct ibv_counter_set *cs; /**< Holds the counters for the rule. */ 254 struct mlx5_flow_counter_stats counter_stats;/**<The counter stats. */ 255 struct mlx5_flow frxq[RTE_DIM(hash_rxq_init)]; 256 /**< Flow with Rx queue. */ 257 }; 258 259 /** Static initializer for items. */ 260 #define ITEMS(...) \ 261 (const enum rte_flow_item_type []){ \ 262 __VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \ 263 } 264 265 /** Structure to generate a simple graph of layers supported by the NIC. */ 266 struct mlx5_flow_items { 267 /** List of possible actions for these items. */ 268 const enum rte_flow_action_type *const actions; 269 /** Bit-masks corresponding to the possibilities for the item. */ 270 const void *mask; 271 /** 272 * Default bit-masks to use when item->mask is not provided. When 273 * \default_mask is also NULL, the full supported bit-mask (\mask) is 274 * used instead. 275 */ 276 const void *default_mask; 277 /** Bit-masks size in bytes. */ 278 const unsigned int mask_sz; 279 /** 280 * Conversion function from rte_flow to NIC specific flow. 281 * 282 * @param item 283 * rte_flow item to convert. 284 * @param default_mask 285 * Default bit-masks to use when item->mask is not provided. 286 * @param data 287 * Internal structure to store the conversion. 288 * 289 * @return 290 * 0 on success, negative value otherwise. 291 */ 292 int (*convert)(const struct rte_flow_item *item, 293 const void *default_mask, 294 void *data); 295 /** Size in bytes of the destination structure. */ 296 const unsigned int dst_sz; 297 /** List of possible following items. */ 298 const enum rte_flow_item_type *const items; 299 }; 300 301 /** Valid action for this PMD. */ 302 static const enum rte_flow_action_type valid_actions[] = { 303 RTE_FLOW_ACTION_TYPE_DROP, 304 RTE_FLOW_ACTION_TYPE_QUEUE, 305 RTE_FLOW_ACTION_TYPE_MARK, 306 RTE_FLOW_ACTION_TYPE_FLAG, 307 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT 308 RTE_FLOW_ACTION_TYPE_COUNT, 309 #endif 310 RTE_FLOW_ACTION_TYPE_END, 311 }; 312 313 /** Graph of supported items and associated actions. */ 314 static const struct mlx5_flow_items mlx5_flow_items[] = { 315 [RTE_FLOW_ITEM_TYPE_END] = { 316 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH, 317 RTE_FLOW_ITEM_TYPE_VXLAN), 318 }, 319 [RTE_FLOW_ITEM_TYPE_ETH] = { 320 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VLAN, 321 RTE_FLOW_ITEM_TYPE_IPV4, 322 RTE_FLOW_ITEM_TYPE_IPV6), 323 .actions = valid_actions, 324 .mask = &(const struct rte_flow_item_eth){ 325 .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff", 326 .src.addr_bytes = "\xff\xff\xff\xff\xff\xff", 327 .type = -1, 328 }, 329 .default_mask = &rte_flow_item_eth_mask, 330 .mask_sz = sizeof(struct rte_flow_item_eth), 331 .convert = mlx5_flow_create_eth, 332 .dst_sz = sizeof(struct ibv_flow_spec_eth), 333 }, 334 [RTE_FLOW_ITEM_TYPE_VLAN] = { 335 .items = ITEMS(RTE_FLOW_ITEM_TYPE_IPV4, 336 RTE_FLOW_ITEM_TYPE_IPV6), 337 .actions = valid_actions, 338 .mask = &(const struct rte_flow_item_vlan){ 339 .tci = -1, 340 }, 341 .default_mask = &rte_flow_item_vlan_mask, 342 .mask_sz = sizeof(struct rte_flow_item_vlan), 343 .convert = mlx5_flow_create_vlan, 344 .dst_sz = 0, 345 }, 346 [RTE_FLOW_ITEM_TYPE_IPV4] = { 347 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP, 348 RTE_FLOW_ITEM_TYPE_TCP), 349 .actions = valid_actions, 350 .mask = &(const struct rte_flow_item_ipv4){ 351 .hdr = { 352 .src_addr = -1, 353 .dst_addr = -1, 354 .type_of_service = -1, 355 .next_proto_id = -1, 356 }, 357 }, 358 .default_mask = &rte_flow_item_ipv4_mask, 359 .mask_sz = sizeof(struct rte_flow_item_ipv4), 360 .convert = mlx5_flow_create_ipv4, 361 .dst_sz = sizeof(struct ibv_flow_spec_ipv4_ext), 362 }, 363 [RTE_FLOW_ITEM_TYPE_IPV6] = { 364 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP, 365 RTE_FLOW_ITEM_TYPE_TCP), 366 .actions = valid_actions, 367 .mask = &(const struct rte_flow_item_ipv6){ 368 .hdr = { 369 .src_addr = { 370 0xff, 0xff, 0xff, 0xff, 371 0xff, 0xff, 0xff, 0xff, 372 0xff, 0xff, 0xff, 0xff, 373 0xff, 0xff, 0xff, 0xff, 374 }, 375 .dst_addr = { 376 0xff, 0xff, 0xff, 0xff, 377 0xff, 0xff, 0xff, 0xff, 378 0xff, 0xff, 0xff, 0xff, 379 0xff, 0xff, 0xff, 0xff, 380 }, 381 .vtc_flow = -1, 382 .proto = -1, 383 .hop_limits = -1, 384 }, 385 }, 386 .default_mask = &rte_flow_item_ipv6_mask, 387 .mask_sz = sizeof(struct rte_flow_item_ipv6), 388 .convert = mlx5_flow_create_ipv6, 389 .dst_sz = sizeof(struct ibv_flow_spec_ipv6), 390 }, 391 [RTE_FLOW_ITEM_TYPE_UDP] = { 392 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VXLAN), 393 .actions = valid_actions, 394 .mask = &(const struct rte_flow_item_udp){ 395 .hdr = { 396 .src_port = -1, 397 .dst_port = -1, 398 }, 399 }, 400 .default_mask = &rte_flow_item_udp_mask, 401 .mask_sz = sizeof(struct rte_flow_item_udp), 402 .convert = mlx5_flow_create_udp, 403 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp), 404 }, 405 [RTE_FLOW_ITEM_TYPE_TCP] = { 406 .actions = valid_actions, 407 .mask = &(const struct rte_flow_item_tcp){ 408 .hdr = { 409 .src_port = -1, 410 .dst_port = -1, 411 }, 412 }, 413 .default_mask = &rte_flow_item_tcp_mask, 414 .mask_sz = sizeof(struct rte_flow_item_tcp), 415 .convert = mlx5_flow_create_tcp, 416 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp), 417 }, 418 [RTE_FLOW_ITEM_TYPE_VXLAN] = { 419 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH), 420 .actions = valid_actions, 421 .mask = &(const struct rte_flow_item_vxlan){ 422 .vni = "\xff\xff\xff", 423 }, 424 .default_mask = &rte_flow_item_vxlan_mask, 425 .mask_sz = sizeof(struct rte_flow_item_vxlan), 426 .convert = mlx5_flow_create_vxlan, 427 .dst_sz = sizeof(struct ibv_flow_spec_tunnel), 428 }, 429 }; 430 431 /** Structure to pass to the conversion function. */ 432 struct mlx5_flow_parse { 433 uint32_t inner; /**< Set once VXLAN is encountered. */ 434 uint32_t allmulti:1; /**< Set once allmulti dst MAC is encountered. */ 435 uint32_t create:1; 436 /**< Whether resources should remain after a validate. */ 437 uint32_t drop:1; /**< Target is a drop queue. */ 438 uint32_t mark:1; /**< Mark is present in the flow. */ 439 uint32_t count:1; /**< Count is present in the flow. */ 440 uint32_t mark_id; /**< Mark identifier. */ 441 uint16_t queues[RTE_MAX_QUEUES_PER_PORT]; /**< Queues indexes to use. */ 442 uint16_t queues_n; /**< Number of entries in queue[]. */ 443 struct rte_eth_rss_conf rss_conf; /**< RSS configuration */ 444 uint8_t rss_key[40]; /**< copy of the RSS key. */ 445 enum hash_rxq_type layer; /**< Last pattern layer detected. */ 446 struct ibv_counter_set *cs; /**< Holds the counter set for the rule */ 447 struct { 448 struct ibv_flow_attr *ibv_attr; 449 /**< Pointer to Verbs attributes. */ 450 unsigned int offset; 451 /**< Current position or total size of the attribute. */ 452 } queue[RTE_DIM(hash_rxq_init)]; 453 }; 454 455 static const struct rte_flow_ops mlx5_flow_ops = { 456 .validate = mlx5_flow_validate, 457 .create = mlx5_flow_create, 458 .destroy = mlx5_flow_destroy, 459 .flush = mlx5_flow_flush, 460 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT 461 .query = mlx5_flow_query, 462 #else 463 .query = NULL, 464 #endif 465 .isolate = mlx5_flow_isolate, 466 }; 467 468 /* Convert FDIR request to Generic flow. */ 469 struct mlx5_fdir { 470 struct rte_flow_attr attr; 471 struct rte_flow_action actions[2]; 472 struct rte_flow_item items[4]; 473 struct rte_flow_item_eth l2; 474 struct rte_flow_item_eth l2_mask; 475 union { 476 struct rte_flow_item_ipv4 ipv4; 477 struct rte_flow_item_ipv6 ipv6; 478 } l3; 479 union { 480 struct rte_flow_item_udp udp; 481 struct rte_flow_item_tcp tcp; 482 } l4; 483 struct rte_flow_action_queue queue; 484 }; 485 486 /* Verbs specification header. */ 487 struct ibv_spec_header { 488 enum ibv_flow_spec_type type; 489 uint16_t size; 490 }; 491 492 /** 493 * Check support for a given item. 494 * 495 * @param item[in] 496 * Item specification. 497 * @param mask[in] 498 * Bit-masks covering supported fields to compare with spec, last and mask in 499 * \item. 500 * @param size 501 * Bit-Mask size in bytes. 502 * 503 * @return 504 * 0 on success. 505 */ 506 static int 507 mlx5_flow_item_validate(const struct rte_flow_item *item, 508 const uint8_t *mask, unsigned int size) 509 { 510 int ret = 0; 511 512 if (!item->spec && (item->mask || item->last)) 513 return -1; 514 if (item->spec && !item->mask) { 515 unsigned int i; 516 const uint8_t *spec = item->spec; 517 518 for (i = 0; i < size; ++i) 519 if ((spec[i] | mask[i]) != mask[i]) 520 return -1; 521 } 522 if (item->last && !item->mask) { 523 unsigned int i; 524 const uint8_t *spec = item->last; 525 526 for (i = 0; i < size; ++i) 527 if ((spec[i] | mask[i]) != mask[i]) 528 return -1; 529 } 530 if (item->mask) { 531 unsigned int i; 532 const uint8_t *spec = item->mask; 533 534 for (i = 0; i < size; ++i) 535 if ((spec[i] | mask[i]) != mask[i]) 536 return -1; 537 } 538 if (item->spec && item->last) { 539 uint8_t spec[size]; 540 uint8_t last[size]; 541 const uint8_t *apply = mask; 542 unsigned int i; 543 544 if (item->mask) 545 apply = item->mask; 546 for (i = 0; i < size; ++i) { 547 spec[i] = ((const uint8_t *)item->spec)[i] & apply[i]; 548 last[i] = ((const uint8_t *)item->last)[i] & apply[i]; 549 } 550 ret = memcmp(spec, last, size); 551 } 552 return ret; 553 } 554 555 /** 556 * Copy the RSS configuration from the user ones. 557 * 558 * @param priv 559 * Pointer to private structure. 560 * @param parser 561 * Internal parser structure. 562 * @param rss_conf 563 * User RSS configuration to save. 564 * 565 * @return 566 * 0 on success, errno value on failure. 567 */ 568 static int 569 priv_flow_convert_rss_conf(struct priv *priv, 570 struct mlx5_flow_parse *parser, 571 const struct rte_eth_rss_conf *rss_conf) 572 { 573 const struct rte_eth_rss_conf *rss; 574 575 if (rss_conf) { 576 if (rss_conf->rss_hf & MLX5_RSS_HF_MASK) 577 return EINVAL; 578 rss = rss_conf; 579 } else { 580 rss = &priv->rss_conf; 581 } 582 if (rss->rss_key_len > 40) 583 return EINVAL; 584 parser->rss_conf.rss_key_len = rss->rss_key_len; 585 parser->rss_conf.rss_hf = rss->rss_hf; 586 memcpy(parser->rss_key, rss->rss_key, rss->rss_key_len); 587 parser->rss_conf.rss_key = parser->rss_key; 588 return 0; 589 } 590 591 /** 592 * Extract attribute to the parser. 593 * 594 * @param priv 595 * Pointer to private structure. 596 * @param[in] attr 597 * Flow rule attributes. 598 * @param[out] error 599 * Perform verbose error reporting if not NULL. 600 * @param[in, out] parser 601 * Internal parser structure. 602 * 603 * @return 604 * 0 on success, a negative errno value otherwise and rte_errno is set. 605 */ 606 static int 607 priv_flow_convert_attributes(struct priv *priv, 608 const struct rte_flow_attr *attr, 609 struct rte_flow_error *error, 610 struct mlx5_flow_parse *parser) 611 { 612 (void)priv; 613 (void)parser; 614 if (attr->group) { 615 rte_flow_error_set(error, ENOTSUP, 616 RTE_FLOW_ERROR_TYPE_ATTR_GROUP, 617 NULL, 618 "groups are not supported"); 619 return -rte_errno; 620 } 621 if (attr->priority && attr->priority != MLX5_CTRL_FLOW_PRIORITY) { 622 rte_flow_error_set(error, ENOTSUP, 623 RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY, 624 NULL, 625 "priorities are not supported"); 626 return -rte_errno; 627 } 628 if (attr->egress) { 629 rte_flow_error_set(error, ENOTSUP, 630 RTE_FLOW_ERROR_TYPE_ATTR_EGRESS, 631 NULL, 632 "egress is not supported"); 633 return -rte_errno; 634 } 635 if (!attr->ingress) { 636 rte_flow_error_set(error, ENOTSUP, 637 RTE_FLOW_ERROR_TYPE_ATTR_INGRESS, 638 NULL, 639 "only ingress is supported"); 640 return -rte_errno; 641 } 642 return 0; 643 } 644 645 /** 646 * Extract actions request to the parser. 647 * 648 * @param priv 649 * Pointer to private structure. 650 * @param[in] actions 651 * Associated actions (list terminated by the END action). 652 * @param[out] error 653 * Perform verbose error reporting if not NULL. 654 * @param[in, out] parser 655 * Internal parser structure. 656 * 657 * @return 658 * 0 on success, a negative errno value otherwise and rte_errno is set. 659 */ 660 static int 661 priv_flow_convert_actions(struct priv *priv, 662 const struct rte_flow_action actions[], 663 struct rte_flow_error *error, 664 struct mlx5_flow_parse *parser) 665 { 666 /* 667 * Add default RSS configuration necessary for Verbs to create QP even 668 * if no RSS is necessary. 669 */ 670 priv_flow_convert_rss_conf(priv, parser, 671 (const struct rte_eth_rss_conf *) 672 &priv->rss_conf); 673 for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) { 674 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) { 675 continue; 676 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) { 677 parser->drop = 1; 678 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) { 679 const struct rte_flow_action_queue *queue = 680 (const struct rte_flow_action_queue *) 681 actions->conf; 682 uint16_t n; 683 uint16_t found = 0; 684 685 if (!queue || (queue->index > (priv->rxqs_n - 1))) 686 goto exit_action_not_supported; 687 for (n = 0; n < parser->queues_n; ++n) { 688 if (parser->queues[n] == queue->index) { 689 found = 1; 690 break; 691 } 692 } 693 if (parser->queues_n > 1 && !found) { 694 rte_flow_error_set(error, ENOTSUP, 695 RTE_FLOW_ERROR_TYPE_ACTION, 696 actions, 697 "queue action not in RSS queues"); 698 return -rte_errno; 699 } 700 if (!found) { 701 parser->queues_n = 1; 702 parser->queues[0] = queue->index; 703 } 704 } else if (actions->type == RTE_FLOW_ACTION_TYPE_RSS) { 705 const struct rte_flow_action_rss *rss = 706 (const struct rte_flow_action_rss *) 707 actions->conf; 708 uint16_t n; 709 710 if (!rss || !rss->num) { 711 rte_flow_error_set(error, EINVAL, 712 RTE_FLOW_ERROR_TYPE_ACTION, 713 actions, 714 "no valid queues"); 715 return -rte_errno; 716 } 717 if (parser->queues_n == 1) { 718 uint16_t found = 0; 719 720 assert(parser->queues_n); 721 for (n = 0; n < rss->num; ++n) { 722 if (parser->queues[0] == 723 rss->queue[n]) { 724 found = 1; 725 break; 726 } 727 } 728 if (!found) { 729 rte_flow_error_set(error, ENOTSUP, 730 RTE_FLOW_ERROR_TYPE_ACTION, 731 actions, 732 "queue action not in RSS" 733 " queues"); 734 return -rte_errno; 735 } 736 } 737 for (n = 0; n < rss->num; ++n) { 738 if (rss->queue[n] >= priv->rxqs_n) { 739 rte_flow_error_set(error, EINVAL, 740 RTE_FLOW_ERROR_TYPE_ACTION, 741 actions, 742 "queue id > number of" 743 " queues"); 744 return -rte_errno; 745 } 746 } 747 for (n = 0; n < rss->num; ++n) 748 parser->queues[n] = rss->queue[n]; 749 parser->queues_n = rss->num; 750 if (priv_flow_convert_rss_conf(priv, parser, 751 rss->rss_conf)) { 752 rte_flow_error_set(error, EINVAL, 753 RTE_FLOW_ERROR_TYPE_ACTION, 754 actions, 755 "wrong RSS configuration"); 756 return -rte_errno; 757 } 758 } else if (actions->type == RTE_FLOW_ACTION_TYPE_MARK) { 759 const struct rte_flow_action_mark *mark = 760 (const struct rte_flow_action_mark *) 761 actions->conf; 762 763 if (!mark) { 764 rte_flow_error_set(error, EINVAL, 765 RTE_FLOW_ERROR_TYPE_ACTION, 766 actions, 767 "mark must be defined"); 768 return -rte_errno; 769 } else if (mark->id >= MLX5_FLOW_MARK_MAX) { 770 rte_flow_error_set(error, ENOTSUP, 771 RTE_FLOW_ERROR_TYPE_ACTION, 772 actions, 773 "mark must be between 0" 774 " and 16777199"); 775 return -rte_errno; 776 } 777 parser->mark = 1; 778 parser->mark_id = mark->id; 779 } else if (actions->type == RTE_FLOW_ACTION_TYPE_FLAG) { 780 parser->mark = 1; 781 } else if (actions->type == RTE_FLOW_ACTION_TYPE_COUNT && 782 priv->config.flow_counter_en) { 783 parser->count = 1; 784 } else { 785 goto exit_action_not_supported; 786 } 787 } 788 if (parser->drop && parser->mark) 789 parser->mark = 0; 790 if (!parser->queues_n && !parser->drop) { 791 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE, 792 NULL, "no valid action"); 793 return -rte_errno; 794 } 795 return 0; 796 exit_action_not_supported: 797 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION, 798 actions, "action not supported"); 799 return -rte_errno; 800 } 801 802 /** 803 * Validate items. 804 * 805 * @param priv 806 * Pointer to private structure. 807 * @param[in] items 808 * Pattern specification (list terminated by the END pattern item). 809 * @param[out] error 810 * Perform verbose error reporting if not NULL. 811 * @param[in, out] parser 812 * Internal parser structure. 813 * 814 * @return 815 * 0 on success, a negative errno value otherwise and rte_errno is set. 816 */ 817 static int 818 priv_flow_convert_items_validate(struct priv *priv, 819 const struct rte_flow_item items[], 820 struct rte_flow_error *error, 821 struct mlx5_flow_parse *parser) 822 { 823 const struct mlx5_flow_items *cur_item = mlx5_flow_items; 824 unsigned int i; 825 826 (void)priv; 827 /* Initialise the offsets to start after verbs attribute. */ 828 for (i = 0; i != hash_rxq_init_n; ++i) 829 parser->queue[i].offset = sizeof(struct ibv_flow_attr); 830 for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) { 831 const struct mlx5_flow_items *token = NULL; 832 unsigned int n; 833 int err; 834 835 if (items->type == RTE_FLOW_ITEM_TYPE_VOID) 836 continue; 837 for (i = 0; 838 cur_item->items && 839 cur_item->items[i] != RTE_FLOW_ITEM_TYPE_END; 840 ++i) { 841 if (cur_item->items[i] == items->type) { 842 token = &mlx5_flow_items[items->type]; 843 break; 844 } 845 } 846 if (!token) 847 goto exit_item_not_supported; 848 cur_item = token; 849 err = mlx5_flow_item_validate(items, 850 (const uint8_t *)cur_item->mask, 851 cur_item->mask_sz); 852 if (err) 853 goto exit_item_not_supported; 854 if (items->type == RTE_FLOW_ITEM_TYPE_VXLAN) { 855 if (parser->inner) { 856 rte_flow_error_set(error, ENOTSUP, 857 RTE_FLOW_ERROR_TYPE_ITEM, 858 items, 859 "cannot recognize multiple" 860 " VXLAN encapsulations"); 861 return -rte_errno; 862 } 863 parser->inner = IBV_FLOW_SPEC_INNER; 864 } 865 if (parser->drop || parser->queues_n == 1) { 866 parser->queue[HASH_RXQ_ETH].offset += cur_item->dst_sz; 867 } else { 868 for (n = 0; n != hash_rxq_init_n; ++n) 869 parser->queue[n].offset += cur_item->dst_sz; 870 } 871 } 872 if (parser->drop) { 873 parser->queue[HASH_RXQ_ETH].offset += 874 sizeof(struct ibv_flow_spec_action_drop); 875 } 876 if (parser->mark) { 877 for (i = 0; i != hash_rxq_init_n; ++i) 878 parser->queue[i].offset += 879 sizeof(struct ibv_flow_spec_action_tag); 880 } 881 if (parser->count) { 882 unsigned int size = sizeof(struct ibv_flow_spec_counter_action); 883 884 for (i = 0; i != hash_rxq_init_n; ++i) 885 parser->queue[i].offset += size; 886 } 887 return 0; 888 exit_item_not_supported: 889 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM, 890 items, "item not supported"); 891 return -rte_errno; 892 } 893 894 /** 895 * Allocate memory space to store verbs flow attributes. 896 * 897 * @param priv 898 * Pointer to private structure. 899 * @param[in] priority 900 * Flow priority. 901 * @param[in] size 902 * Amount of byte to allocate. 903 * @param[out] error 904 * Perform verbose error reporting if not NULL. 905 * 906 * @return 907 * A verbs flow attribute on success, NULL otherwise. 908 */ 909 static struct ibv_flow_attr* 910 priv_flow_convert_allocate(struct priv *priv, 911 unsigned int priority, 912 unsigned int size, 913 struct rte_flow_error *error) 914 { 915 struct ibv_flow_attr *ibv_attr; 916 917 (void)priv; 918 ibv_attr = rte_calloc(__func__, 1, size, 0); 919 if (!ibv_attr) { 920 rte_flow_error_set(error, ENOMEM, 921 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, 922 NULL, 923 "cannot allocate verbs spec attributes."); 924 return NULL; 925 } 926 ibv_attr->priority = priority; 927 return ibv_attr; 928 } 929 930 /** 931 * Finalise verbs flow attributes. 932 * 933 * @param priv 934 * Pointer to private structure. 935 * @param[in, out] parser 936 * Internal parser structure. 937 */ 938 static void 939 priv_flow_convert_finalise(struct priv *priv, struct mlx5_flow_parse *parser) 940 { 941 const unsigned int ipv4 = 942 hash_rxq_init[parser->layer].ip_version == MLX5_IPV4; 943 const enum hash_rxq_type hmin = ipv4 ? HASH_RXQ_TCPV4 : HASH_RXQ_TCPV6; 944 const enum hash_rxq_type hmax = ipv4 ? HASH_RXQ_IPV4 : HASH_RXQ_IPV6; 945 const enum hash_rxq_type ohmin = ipv4 ? HASH_RXQ_TCPV6 : HASH_RXQ_TCPV4; 946 const enum hash_rxq_type ohmax = ipv4 ? HASH_RXQ_IPV6 : HASH_RXQ_IPV4; 947 const enum hash_rxq_type ip = ipv4 ? HASH_RXQ_IPV4 : HASH_RXQ_IPV6; 948 unsigned int i; 949 950 (void)priv; 951 if (parser->layer == HASH_RXQ_ETH) { 952 goto fill; 953 } else { 954 /* 955 * This layer becomes useless as the pattern define under 956 * layers. 957 */ 958 rte_free(parser->queue[HASH_RXQ_ETH].ibv_attr); 959 parser->queue[HASH_RXQ_ETH].ibv_attr = NULL; 960 } 961 /* Remove opposite kind of layer e.g. IPv6 if the pattern is IPv4. */ 962 for (i = ohmin; i != (ohmax + 1); ++i) { 963 if (!parser->queue[i].ibv_attr) 964 continue; 965 rte_free(parser->queue[i].ibv_attr); 966 parser->queue[i].ibv_attr = NULL; 967 } 968 /* Remove impossible flow according to the RSS configuration. */ 969 if (hash_rxq_init[parser->layer].dpdk_rss_hf & 970 parser->rss_conf.rss_hf) { 971 /* Remove any other flow. */ 972 for (i = hmin; i != (hmax + 1); ++i) { 973 if ((i == parser->layer) || 974 (!parser->queue[i].ibv_attr)) 975 continue; 976 rte_free(parser->queue[i].ibv_attr); 977 parser->queue[i].ibv_attr = NULL; 978 } 979 } else if (!parser->queue[ip].ibv_attr) { 980 /* no RSS possible with the current configuration. */ 981 parser->queues_n = 1; 982 return; 983 } 984 fill: 985 /* 986 * Fill missing layers in verbs specifications, or compute the correct 987 * offset to allocate the memory space for the attributes and 988 * specifications. 989 */ 990 for (i = 0; i != hash_rxq_init_n - 1; ++i) { 991 union { 992 struct ibv_flow_spec_ipv4_ext ipv4; 993 struct ibv_flow_spec_ipv6 ipv6; 994 struct ibv_flow_spec_tcp_udp udp_tcp; 995 } specs; 996 void *dst; 997 uint16_t size; 998 999 if (i == parser->layer) 1000 continue; 1001 if (parser->layer == HASH_RXQ_ETH) { 1002 if (hash_rxq_init[i].ip_version == MLX5_IPV4) { 1003 size = sizeof(struct ibv_flow_spec_ipv4_ext); 1004 specs.ipv4 = (struct ibv_flow_spec_ipv4_ext){ 1005 .type = IBV_FLOW_SPEC_IPV4_EXT, 1006 .size = size, 1007 }; 1008 } else { 1009 size = sizeof(struct ibv_flow_spec_ipv6); 1010 specs.ipv6 = (struct ibv_flow_spec_ipv6){ 1011 .type = IBV_FLOW_SPEC_IPV6, 1012 .size = size, 1013 }; 1014 } 1015 if (parser->queue[i].ibv_attr) { 1016 dst = (void *)((uintptr_t) 1017 parser->queue[i].ibv_attr + 1018 parser->queue[i].offset); 1019 memcpy(dst, &specs, size); 1020 ++parser->queue[i].ibv_attr->num_of_specs; 1021 } 1022 parser->queue[i].offset += size; 1023 } 1024 if ((i == HASH_RXQ_UDPV4) || (i == HASH_RXQ_TCPV4) || 1025 (i == HASH_RXQ_UDPV6) || (i == HASH_RXQ_TCPV6)) { 1026 size = sizeof(struct ibv_flow_spec_tcp_udp); 1027 specs.udp_tcp = (struct ibv_flow_spec_tcp_udp) { 1028 .type = ((i == HASH_RXQ_UDPV4 || 1029 i == HASH_RXQ_UDPV6) ? 1030 IBV_FLOW_SPEC_UDP : 1031 IBV_FLOW_SPEC_TCP), 1032 .size = size, 1033 }; 1034 if (parser->queue[i].ibv_attr) { 1035 dst = (void *)((uintptr_t) 1036 parser->queue[i].ibv_attr + 1037 parser->queue[i].offset); 1038 memcpy(dst, &specs, size); 1039 ++parser->queue[i].ibv_attr->num_of_specs; 1040 } 1041 parser->queue[i].offset += size; 1042 } 1043 } 1044 } 1045 1046 /** 1047 * Validate and convert a flow supported by the NIC. 1048 * 1049 * @param priv 1050 * Pointer to private structure. 1051 * @param[in] attr 1052 * Flow rule attributes. 1053 * @param[in] pattern 1054 * Pattern specification (list terminated by the END pattern item). 1055 * @param[in] actions 1056 * Associated actions (list terminated by the END action). 1057 * @param[out] error 1058 * Perform verbose error reporting if not NULL. 1059 * @param[in, out] parser 1060 * Internal parser structure. 1061 * 1062 * @return 1063 * 0 on success, a negative errno value otherwise and rte_errno is set. 1064 */ 1065 static int 1066 priv_flow_convert(struct priv *priv, 1067 const struct rte_flow_attr *attr, 1068 const struct rte_flow_item items[], 1069 const struct rte_flow_action actions[], 1070 struct rte_flow_error *error, 1071 struct mlx5_flow_parse *parser) 1072 { 1073 const struct mlx5_flow_items *cur_item = mlx5_flow_items; 1074 unsigned int i; 1075 int ret; 1076 1077 /* First step. Validate the attributes, items and actions. */ 1078 *parser = (struct mlx5_flow_parse){ 1079 .create = parser->create, 1080 .layer = HASH_RXQ_ETH, 1081 .mark_id = MLX5_FLOW_MARK_DEFAULT, 1082 }; 1083 ret = priv_flow_convert_attributes(priv, attr, error, parser); 1084 if (ret) 1085 return ret; 1086 ret = priv_flow_convert_actions(priv, actions, error, parser); 1087 if (ret) 1088 return ret; 1089 ret = priv_flow_convert_items_validate(priv, items, error, parser); 1090 if (ret) 1091 return ret; 1092 priv_flow_convert_finalise(priv, parser); 1093 /* 1094 * Second step. 1095 * Allocate the memory space to store verbs specifications. 1096 */ 1097 if (parser->drop || parser->queues_n == 1) { 1098 unsigned int priority = 1099 attr->priority + 1100 hash_rxq_init[HASH_RXQ_ETH].flow_priority; 1101 unsigned int offset = parser->queue[HASH_RXQ_ETH].offset; 1102 1103 parser->queue[HASH_RXQ_ETH].ibv_attr = 1104 priv_flow_convert_allocate(priv, priority, 1105 offset, error); 1106 if (!parser->queue[HASH_RXQ_ETH].ibv_attr) 1107 return ENOMEM; 1108 parser->queue[HASH_RXQ_ETH].offset = 1109 sizeof(struct ibv_flow_attr); 1110 } else { 1111 for (i = 0; i != hash_rxq_init_n; ++i) { 1112 unsigned int priority = 1113 attr->priority + 1114 hash_rxq_init[i].flow_priority; 1115 unsigned int offset; 1116 1117 if (!(parser->rss_conf.rss_hf & 1118 hash_rxq_init[i].dpdk_rss_hf) && 1119 (i != HASH_RXQ_ETH)) 1120 continue; 1121 offset = parser->queue[i].offset; 1122 parser->queue[i].ibv_attr = 1123 priv_flow_convert_allocate(priv, priority, 1124 offset, error); 1125 if (!parser->queue[i].ibv_attr) 1126 goto exit_enomem; 1127 parser->queue[i].offset = sizeof(struct ibv_flow_attr); 1128 } 1129 } 1130 /* Third step. Conversion parse, fill the specifications. */ 1131 parser->inner = 0; 1132 for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) { 1133 if (items->type == RTE_FLOW_ITEM_TYPE_VOID) 1134 continue; 1135 cur_item = &mlx5_flow_items[items->type]; 1136 ret = cur_item->convert(items, 1137 (cur_item->default_mask ? 1138 cur_item->default_mask : 1139 cur_item->mask), 1140 parser); 1141 if (ret) { 1142 rte_flow_error_set(error, ret, 1143 RTE_FLOW_ERROR_TYPE_ITEM, 1144 items, "item not supported"); 1145 goto exit_free; 1146 } 1147 } 1148 if (parser->mark) 1149 mlx5_flow_create_flag_mark(parser, parser->mark_id); 1150 if (parser->count && parser->create) { 1151 mlx5_flow_create_count(priv, parser); 1152 if (!parser->cs) 1153 goto exit_count_error; 1154 } 1155 /* 1156 * Last step. Complete missing specification to reach the RSS 1157 * configuration. 1158 */ 1159 if (parser->queues_n > 1) { 1160 priv_flow_convert_finalise(priv, parser); 1161 } else { 1162 /* 1163 * Action queue have their priority overridden with 1164 * Ethernet priority, this priority needs to be adjusted to 1165 * their most specific layer priority. 1166 */ 1167 parser->queue[HASH_RXQ_ETH].ibv_attr->priority = 1168 attr->priority + 1169 hash_rxq_init[parser->layer].flow_priority; 1170 } 1171 exit_free: 1172 /* Only verification is expected, all resources should be released. */ 1173 if (!parser->create) { 1174 for (i = 0; i != hash_rxq_init_n; ++i) { 1175 if (parser->queue[i].ibv_attr) { 1176 rte_free(parser->queue[i].ibv_attr); 1177 parser->queue[i].ibv_attr = NULL; 1178 } 1179 } 1180 } 1181 if (parser->allmulti && 1182 parser->layer == HASH_RXQ_ETH) { 1183 for (i = 0; i != hash_rxq_init_n; ++i) { 1184 if (!parser->queue[i].ibv_attr) 1185 continue; 1186 if (parser->queue[i].ibv_attr->num_of_specs != 1) 1187 break; 1188 parser->queue[i].ibv_attr->type = 1189 IBV_FLOW_ATTR_MC_DEFAULT; 1190 } 1191 } 1192 return ret; 1193 exit_enomem: 1194 for (i = 0; i != hash_rxq_init_n; ++i) { 1195 if (parser->queue[i].ibv_attr) { 1196 rte_free(parser->queue[i].ibv_attr); 1197 parser->queue[i].ibv_attr = NULL; 1198 } 1199 } 1200 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_UNSPECIFIED, 1201 NULL, "cannot allocate verbs spec attributes."); 1202 return ret; 1203 exit_count_error: 1204 rte_flow_error_set(error, EINVAL, RTE_FLOW_ERROR_TYPE_UNSPECIFIED, 1205 NULL, "cannot create counter."); 1206 return rte_errno; 1207 } 1208 1209 /** 1210 * Copy the specification created into the flow. 1211 * 1212 * @param parser 1213 * Internal parser structure. 1214 * @param src 1215 * Create specification. 1216 * @param size 1217 * Size in bytes of the specification to copy. 1218 */ 1219 static void 1220 mlx5_flow_create_copy(struct mlx5_flow_parse *parser, void *src, 1221 unsigned int size) 1222 { 1223 unsigned int i; 1224 void *dst; 1225 1226 for (i = 0; i != hash_rxq_init_n; ++i) { 1227 if (!parser->queue[i].ibv_attr) 1228 continue; 1229 /* Specification must be the same l3 type or none. */ 1230 if (parser->layer == HASH_RXQ_ETH || 1231 (hash_rxq_init[parser->layer].ip_version == 1232 hash_rxq_init[i].ip_version) || 1233 (hash_rxq_init[i].ip_version == 0)) { 1234 dst = (void *)((uintptr_t)parser->queue[i].ibv_attr + 1235 parser->queue[i].offset); 1236 memcpy(dst, src, size); 1237 ++parser->queue[i].ibv_attr->num_of_specs; 1238 parser->queue[i].offset += size; 1239 } 1240 } 1241 } 1242 1243 /** 1244 * Convert Ethernet item to Verbs specification. 1245 * 1246 * @param item[in] 1247 * Item specification. 1248 * @param default_mask[in] 1249 * Default bit-masks to use when item->mask is not provided. 1250 * @param data[in, out] 1251 * User structure. 1252 */ 1253 static int 1254 mlx5_flow_create_eth(const struct rte_flow_item *item, 1255 const void *default_mask, 1256 void *data) 1257 { 1258 const struct rte_flow_item_eth *spec = item->spec; 1259 const struct rte_flow_item_eth *mask = item->mask; 1260 struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data; 1261 const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth); 1262 struct ibv_flow_spec_eth eth = { 1263 .type = parser->inner | IBV_FLOW_SPEC_ETH, 1264 .size = eth_size, 1265 }; 1266 1267 /* Don't update layer for the inner pattern. */ 1268 if (!parser->inner) 1269 parser->layer = HASH_RXQ_ETH; 1270 if (spec) { 1271 unsigned int i; 1272 1273 if (!mask) 1274 mask = default_mask; 1275 memcpy(ð.val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN); 1276 memcpy(ð.val.src_mac, spec->src.addr_bytes, ETHER_ADDR_LEN); 1277 eth.val.ether_type = spec->type; 1278 memcpy(ð.mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN); 1279 memcpy(ð.mask.src_mac, mask->src.addr_bytes, ETHER_ADDR_LEN); 1280 eth.mask.ether_type = mask->type; 1281 /* Remove unwanted bits from values. */ 1282 for (i = 0; i < ETHER_ADDR_LEN; ++i) { 1283 eth.val.dst_mac[i] &= eth.mask.dst_mac[i]; 1284 eth.val.src_mac[i] &= eth.mask.src_mac[i]; 1285 } 1286 eth.val.ether_type &= eth.mask.ether_type; 1287 } 1288 mlx5_flow_create_copy(parser, ð, eth_size); 1289 parser->allmulti = eth.val.dst_mac[0] & 1; 1290 return 0; 1291 } 1292 1293 /** 1294 * Convert VLAN item to Verbs specification. 1295 * 1296 * @param item[in] 1297 * Item specification. 1298 * @param default_mask[in] 1299 * Default bit-masks to use when item->mask is not provided. 1300 * @param data[in, out] 1301 * User structure. 1302 */ 1303 static int 1304 mlx5_flow_create_vlan(const struct rte_flow_item *item, 1305 const void *default_mask, 1306 void *data) 1307 { 1308 const struct rte_flow_item_vlan *spec = item->spec; 1309 const struct rte_flow_item_vlan *mask = item->mask; 1310 struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data; 1311 struct ibv_flow_spec_eth *eth; 1312 const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth); 1313 1314 if (spec) { 1315 unsigned int i; 1316 if (!mask) 1317 mask = default_mask; 1318 1319 for (i = 0; i != hash_rxq_init_n; ++i) { 1320 if (!parser->queue[i].ibv_attr) 1321 continue; 1322 1323 eth = (void *)((uintptr_t)parser->queue[i].ibv_attr + 1324 parser->queue[i].offset - eth_size); 1325 eth->val.vlan_tag = spec->tci; 1326 eth->mask.vlan_tag = mask->tci; 1327 eth->val.vlan_tag &= eth->mask.vlan_tag; 1328 } 1329 } 1330 return 0; 1331 } 1332 1333 /** 1334 * Convert IPv4 item to Verbs specification. 1335 * 1336 * @param item[in] 1337 * Item specification. 1338 * @param default_mask[in] 1339 * Default bit-masks to use when item->mask is not provided. 1340 * @param data[in, out] 1341 * User structure. 1342 */ 1343 static int 1344 mlx5_flow_create_ipv4(const struct rte_flow_item *item, 1345 const void *default_mask, 1346 void *data) 1347 { 1348 const struct rte_flow_item_ipv4 *spec = item->spec; 1349 const struct rte_flow_item_ipv4 *mask = item->mask; 1350 struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data; 1351 unsigned int ipv4_size = sizeof(struct ibv_flow_spec_ipv4_ext); 1352 struct ibv_flow_spec_ipv4_ext ipv4 = { 1353 .type = parser->inner | IBV_FLOW_SPEC_IPV4_EXT, 1354 .size = ipv4_size, 1355 }; 1356 1357 /* Don't update layer for the inner pattern. */ 1358 if (!parser->inner) 1359 parser->layer = HASH_RXQ_IPV4; 1360 if (spec) { 1361 if (!mask) 1362 mask = default_mask; 1363 ipv4.val = (struct ibv_flow_ipv4_ext_filter){ 1364 .src_ip = spec->hdr.src_addr, 1365 .dst_ip = spec->hdr.dst_addr, 1366 .proto = spec->hdr.next_proto_id, 1367 .tos = spec->hdr.type_of_service, 1368 }; 1369 ipv4.mask = (struct ibv_flow_ipv4_ext_filter){ 1370 .src_ip = mask->hdr.src_addr, 1371 .dst_ip = mask->hdr.dst_addr, 1372 .proto = mask->hdr.next_proto_id, 1373 .tos = mask->hdr.type_of_service, 1374 }; 1375 /* Remove unwanted bits from values. */ 1376 ipv4.val.src_ip &= ipv4.mask.src_ip; 1377 ipv4.val.dst_ip &= ipv4.mask.dst_ip; 1378 ipv4.val.proto &= ipv4.mask.proto; 1379 ipv4.val.tos &= ipv4.mask.tos; 1380 } 1381 mlx5_flow_create_copy(parser, &ipv4, ipv4_size); 1382 return 0; 1383 } 1384 1385 /** 1386 * Convert IPv6 item to Verbs specification. 1387 * 1388 * @param item[in] 1389 * Item specification. 1390 * @param default_mask[in] 1391 * Default bit-masks to use when item->mask is not provided. 1392 * @param data[in, out] 1393 * User structure. 1394 */ 1395 static int 1396 mlx5_flow_create_ipv6(const struct rte_flow_item *item, 1397 const void *default_mask, 1398 void *data) 1399 { 1400 const struct rte_flow_item_ipv6 *spec = item->spec; 1401 const struct rte_flow_item_ipv6 *mask = item->mask; 1402 struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data; 1403 unsigned int ipv6_size = sizeof(struct ibv_flow_spec_ipv6); 1404 struct ibv_flow_spec_ipv6 ipv6 = { 1405 .type = parser->inner | IBV_FLOW_SPEC_IPV6, 1406 .size = ipv6_size, 1407 }; 1408 1409 /* Don't update layer for the inner pattern. */ 1410 if (!parser->inner) 1411 parser->layer = HASH_RXQ_IPV6; 1412 if (spec) { 1413 unsigned int i; 1414 uint32_t vtc_flow_val; 1415 uint32_t vtc_flow_mask; 1416 1417 if (!mask) 1418 mask = default_mask; 1419 memcpy(&ipv6.val.src_ip, spec->hdr.src_addr, 1420 RTE_DIM(ipv6.val.src_ip)); 1421 memcpy(&ipv6.val.dst_ip, spec->hdr.dst_addr, 1422 RTE_DIM(ipv6.val.dst_ip)); 1423 memcpy(&ipv6.mask.src_ip, mask->hdr.src_addr, 1424 RTE_DIM(ipv6.mask.src_ip)); 1425 memcpy(&ipv6.mask.dst_ip, mask->hdr.dst_addr, 1426 RTE_DIM(ipv6.mask.dst_ip)); 1427 vtc_flow_val = rte_be_to_cpu_32(spec->hdr.vtc_flow); 1428 vtc_flow_mask = rte_be_to_cpu_32(mask->hdr.vtc_flow); 1429 ipv6.val.flow_label = 1430 rte_cpu_to_be_32((vtc_flow_val & IPV6_HDR_FL_MASK) >> 1431 IPV6_HDR_FL_SHIFT); 1432 ipv6.val.traffic_class = (vtc_flow_val & IPV6_HDR_TC_MASK) >> 1433 IPV6_HDR_TC_SHIFT; 1434 ipv6.val.next_hdr = spec->hdr.proto; 1435 ipv6.val.hop_limit = spec->hdr.hop_limits; 1436 ipv6.mask.flow_label = 1437 rte_cpu_to_be_32((vtc_flow_mask & IPV6_HDR_FL_MASK) >> 1438 IPV6_HDR_FL_SHIFT); 1439 ipv6.mask.traffic_class = (vtc_flow_mask & IPV6_HDR_TC_MASK) >> 1440 IPV6_HDR_TC_SHIFT; 1441 ipv6.mask.next_hdr = mask->hdr.proto; 1442 ipv6.mask.hop_limit = mask->hdr.hop_limits; 1443 /* Remove unwanted bits from values. */ 1444 for (i = 0; i < RTE_DIM(ipv6.val.src_ip); ++i) { 1445 ipv6.val.src_ip[i] &= ipv6.mask.src_ip[i]; 1446 ipv6.val.dst_ip[i] &= ipv6.mask.dst_ip[i]; 1447 } 1448 ipv6.val.flow_label &= ipv6.mask.flow_label; 1449 ipv6.val.traffic_class &= ipv6.mask.traffic_class; 1450 ipv6.val.next_hdr &= ipv6.mask.next_hdr; 1451 ipv6.val.hop_limit &= ipv6.mask.hop_limit; 1452 } 1453 mlx5_flow_create_copy(parser, &ipv6, ipv6_size); 1454 return 0; 1455 } 1456 1457 /** 1458 * Convert UDP item to Verbs specification. 1459 * 1460 * @param item[in] 1461 * Item specification. 1462 * @param default_mask[in] 1463 * Default bit-masks to use when item->mask is not provided. 1464 * @param data[in, out] 1465 * User structure. 1466 */ 1467 static int 1468 mlx5_flow_create_udp(const struct rte_flow_item *item, 1469 const void *default_mask, 1470 void *data) 1471 { 1472 const struct rte_flow_item_udp *spec = item->spec; 1473 const struct rte_flow_item_udp *mask = item->mask; 1474 struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data; 1475 unsigned int udp_size = sizeof(struct ibv_flow_spec_tcp_udp); 1476 struct ibv_flow_spec_tcp_udp udp = { 1477 .type = parser->inner | IBV_FLOW_SPEC_UDP, 1478 .size = udp_size, 1479 }; 1480 1481 /* Don't update layer for the inner pattern. */ 1482 if (!parser->inner) { 1483 if (parser->layer == HASH_RXQ_IPV4) 1484 parser->layer = HASH_RXQ_UDPV4; 1485 else 1486 parser->layer = HASH_RXQ_UDPV6; 1487 } 1488 if (spec) { 1489 if (!mask) 1490 mask = default_mask; 1491 udp.val.dst_port = spec->hdr.dst_port; 1492 udp.val.src_port = spec->hdr.src_port; 1493 udp.mask.dst_port = mask->hdr.dst_port; 1494 udp.mask.src_port = mask->hdr.src_port; 1495 /* Remove unwanted bits from values. */ 1496 udp.val.src_port &= udp.mask.src_port; 1497 udp.val.dst_port &= udp.mask.dst_port; 1498 } 1499 mlx5_flow_create_copy(parser, &udp, udp_size); 1500 return 0; 1501 } 1502 1503 /** 1504 * Convert TCP item to Verbs specification. 1505 * 1506 * @param item[in] 1507 * Item specification. 1508 * @param default_mask[in] 1509 * Default bit-masks to use when item->mask is not provided. 1510 * @param data[in, out] 1511 * User structure. 1512 */ 1513 static int 1514 mlx5_flow_create_tcp(const struct rte_flow_item *item, 1515 const void *default_mask, 1516 void *data) 1517 { 1518 const struct rte_flow_item_tcp *spec = item->spec; 1519 const struct rte_flow_item_tcp *mask = item->mask; 1520 struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data; 1521 unsigned int tcp_size = sizeof(struct ibv_flow_spec_tcp_udp); 1522 struct ibv_flow_spec_tcp_udp tcp = { 1523 .type = parser->inner | IBV_FLOW_SPEC_TCP, 1524 .size = tcp_size, 1525 }; 1526 1527 /* Don't update layer for the inner pattern. */ 1528 if (!parser->inner) { 1529 if (parser->layer == HASH_RXQ_IPV4) 1530 parser->layer = HASH_RXQ_TCPV4; 1531 else 1532 parser->layer = HASH_RXQ_TCPV6; 1533 } 1534 if (spec) { 1535 if (!mask) 1536 mask = default_mask; 1537 tcp.val.dst_port = spec->hdr.dst_port; 1538 tcp.val.src_port = spec->hdr.src_port; 1539 tcp.mask.dst_port = mask->hdr.dst_port; 1540 tcp.mask.src_port = mask->hdr.src_port; 1541 /* Remove unwanted bits from values. */ 1542 tcp.val.src_port &= tcp.mask.src_port; 1543 tcp.val.dst_port &= tcp.mask.dst_port; 1544 } 1545 mlx5_flow_create_copy(parser, &tcp, tcp_size); 1546 return 0; 1547 } 1548 1549 /** 1550 * Convert VXLAN item to Verbs specification. 1551 * 1552 * @param item[in] 1553 * Item specification. 1554 * @param default_mask[in] 1555 * Default bit-masks to use when item->mask is not provided. 1556 * @param data[in, out] 1557 * User structure. 1558 */ 1559 static int 1560 mlx5_flow_create_vxlan(const struct rte_flow_item *item, 1561 const void *default_mask, 1562 void *data) 1563 { 1564 const struct rte_flow_item_vxlan *spec = item->spec; 1565 const struct rte_flow_item_vxlan *mask = item->mask; 1566 struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data; 1567 unsigned int size = sizeof(struct ibv_flow_spec_tunnel); 1568 struct ibv_flow_spec_tunnel vxlan = { 1569 .type = parser->inner | IBV_FLOW_SPEC_VXLAN_TUNNEL, 1570 .size = size, 1571 }; 1572 union vni { 1573 uint32_t vlan_id; 1574 uint8_t vni[4]; 1575 } id; 1576 1577 id.vni[0] = 0; 1578 parser->inner = IBV_FLOW_SPEC_INNER; 1579 if (spec) { 1580 if (!mask) 1581 mask = default_mask; 1582 memcpy(&id.vni[1], spec->vni, 3); 1583 vxlan.val.tunnel_id = id.vlan_id; 1584 memcpy(&id.vni[1], mask->vni, 3); 1585 vxlan.mask.tunnel_id = id.vlan_id; 1586 /* Remove unwanted bits from values. */ 1587 vxlan.val.tunnel_id &= vxlan.mask.tunnel_id; 1588 } 1589 /* 1590 * Tunnel id 0 is equivalent as not adding a VXLAN layer, if only this 1591 * layer is defined in the Verbs specification it is interpreted as 1592 * wildcard and all packets will match this rule, if it follows a full 1593 * stack layer (ex: eth / ipv4 / udp), all packets matching the layers 1594 * before will also match this rule. 1595 * To avoid such situation, VNI 0 is currently refused. 1596 */ 1597 if (!vxlan.val.tunnel_id) 1598 return EINVAL; 1599 mlx5_flow_create_copy(parser, &vxlan, size); 1600 return 0; 1601 } 1602 1603 /** 1604 * Convert mark/flag action to Verbs specification. 1605 * 1606 * @param parser 1607 * Internal parser structure. 1608 * @param mark_id 1609 * Mark identifier. 1610 */ 1611 static int 1612 mlx5_flow_create_flag_mark(struct mlx5_flow_parse *parser, uint32_t mark_id) 1613 { 1614 unsigned int size = sizeof(struct ibv_flow_spec_action_tag); 1615 struct ibv_flow_spec_action_tag tag = { 1616 .type = IBV_FLOW_SPEC_ACTION_TAG, 1617 .size = size, 1618 .tag_id = mlx5_flow_mark_set(mark_id), 1619 }; 1620 1621 assert(parser->mark); 1622 mlx5_flow_create_copy(parser, &tag, size); 1623 return 0; 1624 } 1625 1626 /** 1627 * Convert count action to Verbs specification. 1628 * 1629 * @param priv 1630 * Pointer to private structure. 1631 * @param parser 1632 * Pointer to MLX5 flow parser structure. 1633 * 1634 * @return 1635 * 0 on success, errno value on failure. 1636 */ 1637 static int 1638 mlx5_flow_create_count(struct priv *priv __rte_unused, 1639 struct mlx5_flow_parse *parser __rte_unused) 1640 { 1641 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT 1642 unsigned int size = sizeof(struct ibv_flow_spec_counter_action); 1643 struct ibv_counter_set_init_attr init_attr = {0}; 1644 struct ibv_flow_spec_counter_action counter = { 1645 .type = IBV_FLOW_SPEC_ACTION_COUNT, 1646 .size = size, 1647 .counter_set_handle = 0, 1648 }; 1649 1650 init_attr.counter_set_id = 0; 1651 parser->cs = ibv_create_counter_set(priv->ctx, &init_attr); 1652 if (!parser->cs) 1653 return EINVAL; 1654 counter.counter_set_handle = parser->cs->handle; 1655 mlx5_flow_create_copy(parser, &counter, size); 1656 #endif 1657 return 0; 1658 } 1659 1660 /** 1661 * Complete flow rule creation with a drop queue. 1662 * 1663 * @param priv 1664 * Pointer to private structure. 1665 * @param parser 1666 * Internal parser structure. 1667 * @param flow 1668 * Pointer to the rte_flow. 1669 * @param[out] error 1670 * Perform verbose error reporting if not NULL. 1671 * 1672 * @return 1673 * 0 on success, errno value on failure. 1674 */ 1675 static int 1676 priv_flow_create_action_queue_drop(struct priv *priv, 1677 struct mlx5_flow_parse *parser, 1678 struct rte_flow *flow, 1679 struct rte_flow_error *error) 1680 { 1681 struct ibv_flow_spec_action_drop *drop; 1682 unsigned int size = sizeof(struct ibv_flow_spec_action_drop); 1683 int err = 0; 1684 1685 assert(priv->pd); 1686 assert(priv->ctx); 1687 flow->drop = 1; 1688 drop = (void *)((uintptr_t)parser->queue[HASH_RXQ_ETH].ibv_attr + 1689 parser->queue[HASH_RXQ_ETH].offset); 1690 *drop = (struct ibv_flow_spec_action_drop){ 1691 .type = IBV_FLOW_SPEC_ACTION_DROP, 1692 .size = size, 1693 }; 1694 ++parser->queue[HASH_RXQ_ETH].ibv_attr->num_of_specs; 1695 parser->queue[HASH_RXQ_ETH].offset += size; 1696 flow->frxq[HASH_RXQ_ETH].ibv_attr = 1697 parser->queue[HASH_RXQ_ETH].ibv_attr; 1698 if (parser->count) 1699 flow->cs = parser->cs; 1700 if (!priv->dev->data->dev_started) 1701 return 0; 1702 parser->queue[HASH_RXQ_ETH].ibv_attr = NULL; 1703 flow->frxq[HASH_RXQ_ETH].ibv_flow = 1704 ibv_create_flow(priv->flow_drop_queue->qp, 1705 flow->frxq[HASH_RXQ_ETH].ibv_attr); 1706 if (!flow->frxq[HASH_RXQ_ETH].ibv_flow) { 1707 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE, 1708 NULL, "flow rule creation failure"); 1709 err = ENOMEM; 1710 goto error; 1711 } 1712 return 0; 1713 error: 1714 assert(flow); 1715 if (flow->frxq[HASH_RXQ_ETH].ibv_flow) { 1716 claim_zero(ibv_destroy_flow(flow->frxq[HASH_RXQ_ETH].ibv_flow)); 1717 flow->frxq[HASH_RXQ_ETH].ibv_flow = NULL; 1718 } 1719 if (flow->frxq[HASH_RXQ_ETH].ibv_attr) { 1720 rte_free(flow->frxq[HASH_RXQ_ETH].ibv_attr); 1721 flow->frxq[HASH_RXQ_ETH].ibv_attr = NULL; 1722 } 1723 if (flow->cs) { 1724 claim_zero(ibv_destroy_counter_set(flow->cs)); 1725 flow->cs = NULL; 1726 parser->cs = NULL; 1727 } 1728 return err; 1729 } 1730 1731 /** 1732 * Create hash Rx queues when RSS is enabled. 1733 * 1734 * @param priv 1735 * Pointer to private structure. 1736 * @param parser 1737 * Internal parser structure. 1738 * @param flow 1739 * Pointer to the rte_flow. 1740 * @param[out] error 1741 * Perform verbose error reporting if not NULL. 1742 * 1743 * @return 1744 * 0 on success, a errno value otherwise and rte_errno is set. 1745 */ 1746 static int 1747 priv_flow_create_action_queue_rss(struct priv *priv, 1748 struct mlx5_flow_parse *parser, 1749 struct rte_flow *flow, 1750 struct rte_flow_error *error) 1751 { 1752 unsigned int i; 1753 1754 for (i = 0; i != hash_rxq_init_n; ++i) { 1755 uint64_t hash_fields; 1756 1757 if (!parser->queue[i].ibv_attr) 1758 continue; 1759 flow->frxq[i].ibv_attr = parser->queue[i].ibv_attr; 1760 parser->queue[i].ibv_attr = NULL; 1761 hash_fields = hash_rxq_init[i].hash_fields; 1762 if (!priv->dev->data->dev_started) 1763 continue; 1764 flow->frxq[i].hrxq = 1765 mlx5_priv_hrxq_get(priv, 1766 parser->rss_conf.rss_key, 1767 parser->rss_conf.rss_key_len, 1768 hash_fields, 1769 parser->queues, 1770 parser->queues_n); 1771 if (flow->frxq[i].hrxq) 1772 continue; 1773 flow->frxq[i].hrxq = 1774 mlx5_priv_hrxq_new(priv, 1775 parser->rss_conf.rss_key, 1776 parser->rss_conf.rss_key_len, 1777 hash_fields, 1778 parser->queues, 1779 parser->queues_n); 1780 if (!flow->frxq[i].hrxq) { 1781 rte_flow_error_set(error, ENOMEM, 1782 RTE_FLOW_ERROR_TYPE_HANDLE, 1783 NULL, "cannot create hash rxq"); 1784 return ENOMEM; 1785 } 1786 } 1787 return 0; 1788 } 1789 1790 /** 1791 * Complete flow rule creation. 1792 * 1793 * @param priv 1794 * Pointer to private structure. 1795 * @param parser 1796 * Internal parser structure. 1797 * @param flow 1798 * Pointer to the rte_flow. 1799 * @param[out] error 1800 * Perform verbose error reporting if not NULL. 1801 * 1802 * @return 1803 * 0 on success, a errno value otherwise and rte_errno is set. 1804 */ 1805 static int 1806 priv_flow_create_action_queue(struct priv *priv, 1807 struct mlx5_flow_parse *parser, 1808 struct rte_flow *flow, 1809 struct rte_flow_error *error) 1810 { 1811 int err = 0; 1812 unsigned int i; 1813 1814 assert(priv->pd); 1815 assert(priv->ctx); 1816 assert(!parser->drop); 1817 err = priv_flow_create_action_queue_rss(priv, parser, flow, error); 1818 if (err) 1819 goto error; 1820 if (parser->count) 1821 flow->cs = parser->cs; 1822 if (!priv->dev->data->dev_started) 1823 return 0; 1824 for (i = 0; i != hash_rxq_init_n; ++i) { 1825 if (!flow->frxq[i].hrxq) 1826 continue; 1827 flow->frxq[i].ibv_flow = 1828 ibv_create_flow(flow->frxq[i].hrxq->qp, 1829 flow->frxq[i].ibv_attr); 1830 if (!flow->frxq[i].ibv_flow) { 1831 rte_flow_error_set(error, ENOMEM, 1832 RTE_FLOW_ERROR_TYPE_HANDLE, 1833 NULL, "flow rule creation failure"); 1834 err = ENOMEM; 1835 goto error; 1836 } 1837 DEBUG("%p type %d QP %p ibv_flow %p", 1838 (void *)flow, i, 1839 (void *)flow->frxq[i].hrxq, 1840 (void *)flow->frxq[i].ibv_flow); 1841 } 1842 for (i = 0; i != parser->queues_n; ++i) { 1843 struct mlx5_rxq_data *q = 1844 (*priv->rxqs)[parser->queues[i]]; 1845 1846 q->mark |= parser->mark; 1847 } 1848 return 0; 1849 error: 1850 assert(flow); 1851 for (i = 0; i != hash_rxq_init_n; ++i) { 1852 if (flow->frxq[i].ibv_flow) { 1853 struct ibv_flow *ibv_flow = flow->frxq[i].ibv_flow; 1854 1855 claim_zero(ibv_destroy_flow(ibv_flow)); 1856 } 1857 if (flow->frxq[i].hrxq) 1858 mlx5_priv_hrxq_release(priv, flow->frxq[i].hrxq); 1859 if (flow->frxq[i].ibv_attr) 1860 rte_free(flow->frxq[i].ibv_attr); 1861 } 1862 if (flow->cs) { 1863 claim_zero(ibv_destroy_counter_set(flow->cs)); 1864 flow->cs = NULL; 1865 parser->cs = NULL; 1866 } 1867 return err; 1868 } 1869 1870 /** 1871 * Convert a flow. 1872 * 1873 * @param priv 1874 * Pointer to private structure. 1875 * @param list 1876 * Pointer to a TAILQ flow list. 1877 * @param[in] attr 1878 * Flow rule attributes. 1879 * @param[in] pattern 1880 * Pattern specification (list terminated by the END pattern item). 1881 * @param[in] actions 1882 * Associated actions (list terminated by the END action). 1883 * @param[out] error 1884 * Perform verbose error reporting if not NULL. 1885 * 1886 * @return 1887 * A flow on success, NULL otherwise. 1888 */ 1889 static struct rte_flow * 1890 priv_flow_create(struct priv *priv, 1891 struct mlx5_flows *list, 1892 const struct rte_flow_attr *attr, 1893 const struct rte_flow_item items[], 1894 const struct rte_flow_action actions[], 1895 struct rte_flow_error *error) 1896 { 1897 struct mlx5_flow_parse parser = { .create = 1, }; 1898 struct rte_flow *flow = NULL; 1899 unsigned int i; 1900 int err; 1901 1902 err = priv_flow_convert(priv, attr, items, actions, error, &parser); 1903 if (err) 1904 goto exit; 1905 flow = rte_calloc(__func__, 1, 1906 sizeof(*flow) + parser.queues_n * sizeof(uint16_t), 1907 0); 1908 if (!flow) { 1909 rte_flow_error_set(error, ENOMEM, 1910 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, 1911 NULL, 1912 "cannot allocate flow memory"); 1913 return NULL; 1914 } 1915 /* Copy queues configuration. */ 1916 flow->queues = (uint16_t (*)[])(flow + 1); 1917 memcpy(flow->queues, parser.queues, parser.queues_n * sizeof(uint16_t)); 1918 flow->queues_n = parser.queues_n; 1919 flow->mark = parser.mark; 1920 /* Copy RSS configuration. */ 1921 flow->rss_conf = parser.rss_conf; 1922 flow->rss_conf.rss_key = flow->rss_key; 1923 memcpy(flow->rss_key, parser.rss_key, parser.rss_conf.rss_key_len); 1924 /* finalise the flow. */ 1925 if (parser.drop) 1926 err = priv_flow_create_action_queue_drop(priv, &parser, flow, 1927 error); 1928 else 1929 err = priv_flow_create_action_queue(priv, &parser, flow, error); 1930 if (err) 1931 goto exit; 1932 TAILQ_INSERT_TAIL(list, flow, next); 1933 DEBUG("Flow created %p", (void *)flow); 1934 return flow; 1935 exit: 1936 for (i = 0; i != hash_rxq_init_n; ++i) { 1937 if (parser.queue[i].ibv_attr) 1938 rte_free(parser.queue[i].ibv_attr); 1939 } 1940 rte_free(flow); 1941 return NULL; 1942 } 1943 1944 /** 1945 * Validate a flow supported by the NIC. 1946 * 1947 * @see rte_flow_validate() 1948 * @see rte_flow_ops 1949 */ 1950 int 1951 mlx5_flow_validate(struct rte_eth_dev *dev, 1952 const struct rte_flow_attr *attr, 1953 const struct rte_flow_item items[], 1954 const struct rte_flow_action actions[], 1955 struct rte_flow_error *error) 1956 { 1957 struct priv *priv = dev->data->dev_private; 1958 int ret; 1959 struct mlx5_flow_parse parser = { .create = 0, }; 1960 1961 priv_lock(priv); 1962 ret = priv_flow_convert(priv, attr, items, actions, error, &parser); 1963 priv_unlock(priv); 1964 return ret; 1965 } 1966 1967 /** 1968 * Create a flow. 1969 * 1970 * @see rte_flow_create() 1971 * @see rte_flow_ops 1972 */ 1973 struct rte_flow * 1974 mlx5_flow_create(struct rte_eth_dev *dev, 1975 const struct rte_flow_attr *attr, 1976 const struct rte_flow_item items[], 1977 const struct rte_flow_action actions[], 1978 struct rte_flow_error *error) 1979 { 1980 struct priv *priv = dev->data->dev_private; 1981 struct rte_flow *flow; 1982 1983 priv_lock(priv); 1984 flow = priv_flow_create(priv, &priv->flows, attr, items, actions, 1985 error); 1986 priv_unlock(priv); 1987 return flow; 1988 } 1989 1990 /** 1991 * Destroy a flow. 1992 * 1993 * @param priv 1994 * Pointer to private structure. 1995 * @param list 1996 * Pointer to a TAILQ flow list. 1997 * @param[in] flow 1998 * Flow to destroy. 1999 */ 2000 static void 2001 priv_flow_destroy(struct priv *priv, 2002 struct mlx5_flows *list, 2003 struct rte_flow *flow) 2004 { 2005 unsigned int i; 2006 2007 if (flow->drop || !flow->mark) 2008 goto free; 2009 for (i = 0; i != flow->queues_n; ++i) { 2010 struct rte_flow *tmp; 2011 int mark = 0; 2012 2013 /* 2014 * To remove the mark from the queue, the queue must not be 2015 * present in any other marked flow (RSS or not). 2016 */ 2017 TAILQ_FOREACH(tmp, list, next) { 2018 unsigned int j; 2019 uint16_t *tqs = NULL; 2020 uint16_t tq_n = 0; 2021 2022 if (!tmp->mark) 2023 continue; 2024 for (j = 0; j != hash_rxq_init_n; ++j) { 2025 if (!tmp->frxq[j].hrxq) 2026 continue; 2027 tqs = tmp->frxq[j].hrxq->ind_table->queues; 2028 tq_n = tmp->frxq[j].hrxq->ind_table->queues_n; 2029 } 2030 if (!tq_n) 2031 continue; 2032 for (j = 0; (j != tq_n) && !mark; j++) 2033 if (tqs[j] == (*flow->queues)[i]) 2034 mark = 1; 2035 } 2036 (*priv->rxqs)[(*flow->queues)[i]]->mark = mark; 2037 } 2038 free: 2039 if (flow->drop) { 2040 if (flow->frxq[HASH_RXQ_ETH].ibv_flow) 2041 claim_zero(ibv_destroy_flow 2042 (flow->frxq[HASH_RXQ_ETH].ibv_flow)); 2043 rte_free(flow->frxq[HASH_RXQ_ETH].ibv_attr); 2044 } else { 2045 for (i = 0; i != hash_rxq_init_n; ++i) { 2046 struct mlx5_flow *frxq = &flow->frxq[i]; 2047 2048 if (frxq->ibv_flow) 2049 claim_zero(ibv_destroy_flow(frxq->ibv_flow)); 2050 if (frxq->hrxq) 2051 mlx5_priv_hrxq_release(priv, frxq->hrxq); 2052 if (frxq->ibv_attr) 2053 rte_free(frxq->ibv_attr); 2054 } 2055 } 2056 if (flow->cs) { 2057 claim_zero(ibv_destroy_counter_set(flow->cs)); 2058 flow->cs = NULL; 2059 } 2060 TAILQ_REMOVE(list, flow, next); 2061 DEBUG("Flow destroyed %p", (void *)flow); 2062 rte_free(flow); 2063 } 2064 2065 /** 2066 * Destroy all flows. 2067 * 2068 * @param priv 2069 * Pointer to private structure. 2070 * @param list 2071 * Pointer to a TAILQ flow list. 2072 */ 2073 void 2074 priv_flow_flush(struct priv *priv, struct mlx5_flows *list) 2075 { 2076 while (!TAILQ_EMPTY(list)) { 2077 struct rte_flow *flow; 2078 2079 flow = TAILQ_FIRST(list); 2080 priv_flow_destroy(priv, list, flow); 2081 } 2082 } 2083 2084 /** 2085 * Create drop queue. 2086 * 2087 * @param priv 2088 * Pointer to private structure. 2089 * 2090 * @return 2091 * 0 on success. 2092 */ 2093 int 2094 priv_flow_create_drop_queue(struct priv *priv) 2095 { 2096 struct mlx5_hrxq_drop *fdq = NULL; 2097 2098 assert(priv->pd); 2099 assert(priv->ctx); 2100 fdq = rte_calloc(__func__, 1, sizeof(*fdq), 0); 2101 if (!fdq) { 2102 WARN("cannot allocate memory for drop queue"); 2103 goto error; 2104 } 2105 fdq->cq = ibv_create_cq(priv->ctx, 1, NULL, NULL, 0); 2106 if (!fdq->cq) { 2107 WARN("cannot allocate CQ for drop queue"); 2108 goto error; 2109 } 2110 fdq->wq = ibv_create_wq(priv->ctx, 2111 &(struct ibv_wq_init_attr){ 2112 .wq_type = IBV_WQT_RQ, 2113 .max_wr = 1, 2114 .max_sge = 1, 2115 .pd = priv->pd, 2116 .cq = fdq->cq, 2117 }); 2118 if (!fdq->wq) { 2119 WARN("cannot allocate WQ for drop queue"); 2120 goto error; 2121 } 2122 fdq->ind_table = ibv_create_rwq_ind_table(priv->ctx, 2123 &(struct ibv_rwq_ind_table_init_attr){ 2124 .log_ind_tbl_size = 0, 2125 .ind_tbl = &fdq->wq, 2126 .comp_mask = 0, 2127 }); 2128 if (!fdq->ind_table) { 2129 WARN("cannot allocate indirection table for drop queue"); 2130 goto error; 2131 } 2132 fdq->qp = ibv_create_qp_ex(priv->ctx, 2133 &(struct ibv_qp_init_attr_ex){ 2134 .qp_type = IBV_QPT_RAW_PACKET, 2135 .comp_mask = 2136 IBV_QP_INIT_ATTR_PD | 2137 IBV_QP_INIT_ATTR_IND_TABLE | 2138 IBV_QP_INIT_ATTR_RX_HASH, 2139 .rx_hash_conf = (struct ibv_rx_hash_conf){ 2140 .rx_hash_function = 2141 IBV_RX_HASH_FUNC_TOEPLITZ, 2142 .rx_hash_key_len = rss_hash_default_key_len, 2143 .rx_hash_key = rss_hash_default_key, 2144 .rx_hash_fields_mask = 0, 2145 }, 2146 .rwq_ind_tbl = fdq->ind_table, 2147 .pd = priv->pd 2148 }); 2149 if (!fdq->qp) { 2150 WARN("cannot allocate QP for drop queue"); 2151 goto error; 2152 } 2153 priv->flow_drop_queue = fdq; 2154 return 0; 2155 error: 2156 if (fdq->qp) 2157 claim_zero(ibv_destroy_qp(fdq->qp)); 2158 if (fdq->ind_table) 2159 claim_zero(ibv_destroy_rwq_ind_table(fdq->ind_table)); 2160 if (fdq->wq) 2161 claim_zero(ibv_destroy_wq(fdq->wq)); 2162 if (fdq->cq) 2163 claim_zero(ibv_destroy_cq(fdq->cq)); 2164 if (fdq) 2165 rte_free(fdq); 2166 priv->flow_drop_queue = NULL; 2167 return -1; 2168 } 2169 2170 /** 2171 * Delete drop queue. 2172 * 2173 * @param priv 2174 * Pointer to private structure. 2175 */ 2176 void 2177 priv_flow_delete_drop_queue(struct priv *priv) 2178 { 2179 struct mlx5_hrxq_drop *fdq = priv->flow_drop_queue; 2180 2181 if (!fdq) 2182 return; 2183 if (fdq->qp) 2184 claim_zero(ibv_destroy_qp(fdq->qp)); 2185 if (fdq->ind_table) 2186 claim_zero(ibv_destroy_rwq_ind_table(fdq->ind_table)); 2187 if (fdq->wq) 2188 claim_zero(ibv_destroy_wq(fdq->wq)); 2189 if (fdq->cq) 2190 claim_zero(ibv_destroy_cq(fdq->cq)); 2191 rte_free(fdq); 2192 priv->flow_drop_queue = NULL; 2193 } 2194 2195 /** 2196 * Remove all flows. 2197 * 2198 * @param priv 2199 * Pointer to private structure. 2200 * @param list 2201 * Pointer to a TAILQ flow list. 2202 */ 2203 void 2204 priv_flow_stop(struct priv *priv, struct mlx5_flows *list) 2205 { 2206 struct rte_flow *flow; 2207 2208 TAILQ_FOREACH_REVERSE(flow, list, mlx5_flows, next) { 2209 unsigned int i; 2210 2211 if (flow->drop) { 2212 if (!flow->frxq[HASH_RXQ_ETH].ibv_flow) 2213 continue; 2214 claim_zero(ibv_destroy_flow 2215 (flow->frxq[HASH_RXQ_ETH].ibv_flow)); 2216 flow->frxq[HASH_RXQ_ETH].ibv_flow = NULL; 2217 /* Next flow. */ 2218 continue; 2219 } 2220 if (flow->mark) { 2221 struct mlx5_ind_table_ibv *ind_tbl = NULL; 2222 2223 for (i = 0; i != hash_rxq_init_n; ++i) { 2224 if (!flow->frxq[i].hrxq) 2225 continue; 2226 ind_tbl = flow->frxq[i].hrxq->ind_table; 2227 } 2228 assert(ind_tbl); 2229 for (i = 0; i != ind_tbl->queues_n; ++i) 2230 (*priv->rxqs)[ind_tbl->queues[i]]->mark = 0; 2231 } 2232 for (i = 0; i != hash_rxq_init_n; ++i) { 2233 if (!flow->frxq[i].ibv_flow) 2234 continue; 2235 claim_zero(ibv_destroy_flow(flow->frxq[i].ibv_flow)); 2236 flow->frxq[i].ibv_flow = NULL; 2237 mlx5_priv_hrxq_release(priv, flow->frxq[i].hrxq); 2238 flow->frxq[i].hrxq = NULL; 2239 } 2240 DEBUG("Flow %p removed", (void *)flow); 2241 } 2242 } 2243 2244 /** 2245 * Add all flows. 2246 * 2247 * @param priv 2248 * Pointer to private structure. 2249 * @param list 2250 * Pointer to a TAILQ flow list. 2251 * 2252 * @return 2253 * 0 on success, a errno value otherwise and rte_errno is set. 2254 */ 2255 int 2256 priv_flow_start(struct priv *priv, struct mlx5_flows *list) 2257 { 2258 struct rte_flow *flow; 2259 2260 TAILQ_FOREACH(flow, list, next) { 2261 unsigned int i; 2262 2263 if (flow->drop) { 2264 flow->frxq[HASH_RXQ_ETH].ibv_flow = 2265 ibv_create_flow 2266 (priv->flow_drop_queue->qp, 2267 flow->frxq[HASH_RXQ_ETH].ibv_attr); 2268 if (!flow->frxq[HASH_RXQ_ETH].ibv_flow) { 2269 DEBUG("Flow %p cannot be applied", 2270 (void *)flow); 2271 rte_errno = EINVAL; 2272 return rte_errno; 2273 } 2274 DEBUG("Flow %p applied", (void *)flow); 2275 /* Next flow. */ 2276 continue; 2277 } 2278 for (i = 0; i != hash_rxq_init_n; ++i) { 2279 if (!flow->frxq[i].ibv_attr) 2280 continue; 2281 flow->frxq[i].hrxq = 2282 mlx5_priv_hrxq_get(priv, flow->rss_conf.rss_key, 2283 flow->rss_conf.rss_key_len, 2284 hash_rxq_init[i].hash_fields, 2285 (*flow->queues), 2286 flow->queues_n); 2287 if (flow->frxq[i].hrxq) 2288 goto flow_create; 2289 flow->frxq[i].hrxq = 2290 mlx5_priv_hrxq_new(priv, flow->rss_conf.rss_key, 2291 flow->rss_conf.rss_key_len, 2292 hash_rxq_init[i].hash_fields, 2293 (*flow->queues), 2294 flow->queues_n); 2295 if (!flow->frxq[i].hrxq) { 2296 DEBUG("Flow %p cannot be applied", 2297 (void *)flow); 2298 rte_errno = EINVAL; 2299 return rte_errno; 2300 } 2301 flow_create: 2302 flow->frxq[i].ibv_flow = 2303 ibv_create_flow(flow->frxq[i].hrxq->qp, 2304 flow->frxq[i].ibv_attr); 2305 if (!flow->frxq[i].ibv_flow) { 2306 DEBUG("Flow %p cannot be applied", 2307 (void *)flow); 2308 rte_errno = EINVAL; 2309 return rte_errno; 2310 } 2311 DEBUG("Flow %p applied", (void *)flow); 2312 } 2313 if (!flow->mark) 2314 continue; 2315 for (i = 0; i != flow->queues_n; ++i) 2316 (*priv->rxqs)[(*flow->queues)[i]]->mark = 1; 2317 } 2318 return 0; 2319 } 2320 2321 /** 2322 * Verify the flow list is empty 2323 * 2324 * @param priv 2325 * Pointer to private structure. 2326 * 2327 * @return the number of flows not released. 2328 */ 2329 int 2330 priv_flow_verify(struct priv *priv) 2331 { 2332 struct rte_flow *flow; 2333 int ret = 0; 2334 2335 TAILQ_FOREACH(flow, &priv->flows, next) { 2336 DEBUG("%p: flow %p still referenced", (void *)priv, 2337 (void *)flow); 2338 ++ret; 2339 } 2340 return ret; 2341 } 2342 2343 /** 2344 * Enable a control flow configured from the control plane. 2345 * 2346 * @param dev 2347 * Pointer to Ethernet device. 2348 * @param eth_spec 2349 * An Ethernet flow spec to apply. 2350 * @param eth_mask 2351 * An Ethernet flow mask to apply. 2352 * @param vlan_spec 2353 * A VLAN flow spec to apply. 2354 * @param vlan_mask 2355 * A VLAN flow mask to apply. 2356 * 2357 * @return 2358 * 0 on success. 2359 */ 2360 int 2361 mlx5_ctrl_flow_vlan(struct rte_eth_dev *dev, 2362 struct rte_flow_item_eth *eth_spec, 2363 struct rte_flow_item_eth *eth_mask, 2364 struct rte_flow_item_vlan *vlan_spec, 2365 struct rte_flow_item_vlan *vlan_mask) 2366 { 2367 struct priv *priv = dev->data->dev_private; 2368 const struct rte_flow_attr attr = { 2369 .ingress = 1, 2370 .priority = MLX5_CTRL_FLOW_PRIORITY, 2371 }; 2372 struct rte_flow_item items[] = { 2373 { 2374 .type = RTE_FLOW_ITEM_TYPE_ETH, 2375 .spec = eth_spec, 2376 .last = NULL, 2377 .mask = eth_mask, 2378 }, 2379 { 2380 .type = (vlan_spec) ? RTE_FLOW_ITEM_TYPE_VLAN : 2381 RTE_FLOW_ITEM_TYPE_END, 2382 .spec = vlan_spec, 2383 .last = NULL, 2384 .mask = vlan_mask, 2385 }, 2386 { 2387 .type = RTE_FLOW_ITEM_TYPE_END, 2388 }, 2389 }; 2390 struct rte_flow_action actions[] = { 2391 { 2392 .type = RTE_FLOW_ACTION_TYPE_RSS, 2393 }, 2394 { 2395 .type = RTE_FLOW_ACTION_TYPE_END, 2396 }, 2397 }; 2398 struct rte_flow *flow; 2399 struct rte_flow_error error; 2400 unsigned int i; 2401 union { 2402 struct rte_flow_action_rss rss; 2403 struct { 2404 const struct rte_eth_rss_conf *rss_conf; 2405 uint16_t num; 2406 uint16_t queue[RTE_MAX_QUEUES_PER_PORT]; 2407 } local; 2408 } action_rss; 2409 2410 if (!priv->reta_idx_n) 2411 return EINVAL; 2412 for (i = 0; i != priv->reta_idx_n; ++i) 2413 action_rss.local.queue[i] = (*priv->reta_idx)[i]; 2414 action_rss.local.rss_conf = &priv->rss_conf; 2415 action_rss.local.num = priv->reta_idx_n; 2416 actions[0].conf = (const void *)&action_rss.rss; 2417 flow = priv_flow_create(priv, &priv->ctrl_flows, &attr, items, actions, 2418 &error); 2419 if (!flow) 2420 return rte_errno; 2421 return 0; 2422 } 2423 2424 /** 2425 * Enable a flow control configured from the control plane. 2426 * 2427 * @param dev 2428 * Pointer to Ethernet device. 2429 * @param eth_spec 2430 * An Ethernet flow spec to apply. 2431 * @param eth_mask 2432 * An Ethernet flow mask to apply. 2433 * 2434 * @return 2435 * 0 on success. 2436 */ 2437 int 2438 mlx5_ctrl_flow(struct rte_eth_dev *dev, 2439 struct rte_flow_item_eth *eth_spec, 2440 struct rte_flow_item_eth *eth_mask) 2441 { 2442 return mlx5_ctrl_flow_vlan(dev, eth_spec, eth_mask, NULL, NULL); 2443 } 2444 2445 /** 2446 * Destroy a flow. 2447 * 2448 * @see rte_flow_destroy() 2449 * @see rte_flow_ops 2450 */ 2451 int 2452 mlx5_flow_destroy(struct rte_eth_dev *dev, 2453 struct rte_flow *flow, 2454 struct rte_flow_error *error) 2455 { 2456 struct priv *priv = dev->data->dev_private; 2457 2458 (void)error; 2459 priv_lock(priv); 2460 priv_flow_destroy(priv, &priv->flows, flow); 2461 priv_unlock(priv); 2462 return 0; 2463 } 2464 2465 /** 2466 * Destroy all flows. 2467 * 2468 * @see rte_flow_flush() 2469 * @see rte_flow_ops 2470 */ 2471 int 2472 mlx5_flow_flush(struct rte_eth_dev *dev, 2473 struct rte_flow_error *error) 2474 { 2475 struct priv *priv = dev->data->dev_private; 2476 2477 (void)error; 2478 priv_lock(priv); 2479 priv_flow_flush(priv, &priv->flows); 2480 priv_unlock(priv); 2481 return 0; 2482 } 2483 2484 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT 2485 /** 2486 * Query flow counter. 2487 * 2488 * @param cs 2489 * the counter set. 2490 * @param counter_value 2491 * returned data from the counter. 2492 * 2493 * @return 2494 * 0 on success, a errno value otherwise and rte_errno is set. 2495 */ 2496 static int 2497 priv_flow_query_count(struct ibv_counter_set *cs, 2498 struct mlx5_flow_counter_stats *counter_stats, 2499 struct rte_flow_query_count *query_count, 2500 struct rte_flow_error *error) 2501 { 2502 uint64_t counters[2]; 2503 struct ibv_query_counter_set_attr query_cs_attr = { 2504 .cs = cs, 2505 .query_flags = IBV_COUNTER_SET_FORCE_UPDATE, 2506 }; 2507 struct ibv_counter_set_data query_out = { 2508 .out = counters, 2509 .outlen = 2 * sizeof(uint64_t), 2510 }; 2511 int res = ibv_query_counter_set(&query_cs_attr, &query_out); 2512 2513 if (res) { 2514 rte_flow_error_set(error, -res, 2515 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, 2516 NULL, 2517 "cannot read counter"); 2518 return -res; 2519 } 2520 query_count->hits_set = 1; 2521 query_count->bytes_set = 1; 2522 query_count->hits = counters[0] - counter_stats->hits; 2523 query_count->bytes = counters[1] - counter_stats->bytes; 2524 if (query_count->reset) { 2525 counter_stats->hits = counters[0]; 2526 counter_stats->bytes = counters[1]; 2527 } 2528 return 0; 2529 } 2530 2531 /** 2532 * Query a flows. 2533 * 2534 * @see rte_flow_query() 2535 * @see rte_flow_ops 2536 */ 2537 int 2538 mlx5_flow_query(struct rte_eth_dev *dev, 2539 struct rte_flow *flow, 2540 enum rte_flow_action_type action __rte_unused, 2541 void *data, 2542 struct rte_flow_error *error) 2543 { 2544 struct priv *priv = dev->data->dev_private; 2545 int res = EINVAL; 2546 2547 priv_lock(priv); 2548 if (flow->cs) { 2549 res = priv_flow_query_count(flow->cs, 2550 &flow->counter_stats, 2551 (struct rte_flow_query_count *)data, 2552 error); 2553 } else { 2554 rte_flow_error_set(error, res, 2555 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, 2556 NULL, 2557 "no counter found for flow"); 2558 } 2559 priv_unlock(priv); 2560 return -res; 2561 } 2562 #endif 2563 2564 /** 2565 * Isolated mode. 2566 * 2567 * @see rte_flow_isolate() 2568 * @see rte_flow_ops 2569 */ 2570 int 2571 mlx5_flow_isolate(struct rte_eth_dev *dev, 2572 int enable, 2573 struct rte_flow_error *error) 2574 { 2575 struct priv *priv = dev->data->dev_private; 2576 2577 priv_lock(priv); 2578 if (dev->data->dev_started) { 2579 rte_flow_error_set(error, EBUSY, 2580 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, 2581 NULL, 2582 "port must be stopped first"); 2583 priv_unlock(priv); 2584 return -rte_errno; 2585 } 2586 priv->isolated = !!enable; 2587 if (enable) 2588 priv->dev->dev_ops = &mlx5_dev_ops_isolate; 2589 else 2590 priv->dev->dev_ops = &mlx5_dev_ops; 2591 priv_unlock(priv); 2592 return 0; 2593 } 2594 2595 /** 2596 * Convert a flow director filter to a generic flow. 2597 * 2598 * @param priv 2599 * Private structure. 2600 * @param fdir_filter 2601 * Flow director filter to add. 2602 * @param attributes 2603 * Generic flow parameters structure. 2604 * 2605 * @return 2606 * 0 on success, errno value on error. 2607 */ 2608 static int 2609 priv_fdir_filter_convert(struct priv *priv, 2610 const struct rte_eth_fdir_filter *fdir_filter, 2611 struct mlx5_fdir *attributes) 2612 { 2613 const struct rte_eth_fdir_input *input = &fdir_filter->input; 2614 2615 /* Validate queue number. */ 2616 if (fdir_filter->action.rx_queue >= priv->rxqs_n) { 2617 ERROR("invalid queue number %d", fdir_filter->action.rx_queue); 2618 return EINVAL; 2619 } 2620 attributes->attr.ingress = 1; 2621 attributes->items[0] = (struct rte_flow_item) { 2622 .type = RTE_FLOW_ITEM_TYPE_ETH, 2623 .spec = &attributes->l2, 2624 .mask = &attributes->l2_mask, 2625 }; 2626 switch (fdir_filter->action.behavior) { 2627 case RTE_ETH_FDIR_ACCEPT: 2628 attributes->actions[0] = (struct rte_flow_action){ 2629 .type = RTE_FLOW_ACTION_TYPE_QUEUE, 2630 .conf = &attributes->queue, 2631 }; 2632 break; 2633 case RTE_ETH_FDIR_REJECT: 2634 attributes->actions[0] = (struct rte_flow_action){ 2635 .type = RTE_FLOW_ACTION_TYPE_DROP, 2636 }; 2637 break; 2638 default: 2639 ERROR("invalid behavior %d", fdir_filter->action.behavior); 2640 return ENOTSUP; 2641 } 2642 attributes->queue.index = fdir_filter->action.rx_queue; 2643 switch (fdir_filter->input.flow_type) { 2644 case RTE_ETH_FLOW_NONFRAG_IPV4_UDP: 2645 attributes->l3.ipv4.hdr = (struct ipv4_hdr){ 2646 .src_addr = input->flow.udp4_flow.ip.src_ip, 2647 .dst_addr = input->flow.udp4_flow.ip.dst_ip, 2648 .time_to_live = input->flow.udp4_flow.ip.ttl, 2649 .type_of_service = input->flow.udp4_flow.ip.tos, 2650 .next_proto_id = input->flow.udp4_flow.ip.proto, 2651 }; 2652 attributes->l4.udp.hdr = (struct udp_hdr){ 2653 .src_port = input->flow.udp4_flow.src_port, 2654 .dst_port = input->flow.udp4_flow.dst_port, 2655 }; 2656 attributes->items[1] = (struct rte_flow_item){ 2657 .type = RTE_FLOW_ITEM_TYPE_IPV4, 2658 .spec = &attributes->l3, 2659 }; 2660 attributes->items[2] = (struct rte_flow_item){ 2661 .type = RTE_FLOW_ITEM_TYPE_UDP, 2662 .spec = &attributes->l4, 2663 }; 2664 break; 2665 case RTE_ETH_FLOW_NONFRAG_IPV4_TCP: 2666 attributes->l3.ipv4.hdr = (struct ipv4_hdr){ 2667 .src_addr = input->flow.tcp4_flow.ip.src_ip, 2668 .dst_addr = input->flow.tcp4_flow.ip.dst_ip, 2669 .time_to_live = input->flow.tcp4_flow.ip.ttl, 2670 .type_of_service = input->flow.tcp4_flow.ip.tos, 2671 .next_proto_id = input->flow.tcp4_flow.ip.proto, 2672 }; 2673 attributes->l4.tcp.hdr = (struct tcp_hdr){ 2674 .src_port = input->flow.tcp4_flow.src_port, 2675 .dst_port = input->flow.tcp4_flow.dst_port, 2676 }; 2677 attributes->items[1] = (struct rte_flow_item){ 2678 .type = RTE_FLOW_ITEM_TYPE_IPV4, 2679 .spec = &attributes->l3, 2680 }; 2681 attributes->items[2] = (struct rte_flow_item){ 2682 .type = RTE_FLOW_ITEM_TYPE_TCP, 2683 .spec = &attributes->l4, 2684 }; 2685 break; 2686 case RTE_ETH_FLOW_NONFRAG_IPV4_OTHER: 2687 attributes->l3.ipv4.hdr = (struct ipv4_hdr){ 2688 .src_addr = input->flow.ip4_flow.src_ip, 2689 .dst_addr = input->flow.ip4_flow.dst_ip, 2690 .time_to_live = input->flow.ip4_flow.ttl, 2691 .type_of_service = input->flow.ip4_flow.tos, 2692 .next_proto_id = input->flow.ip4_flow.proto, 2693 }; 2694 attributes->items[1] = (struct rte_flow_item){ 2695 .type = RTE_FLOW_ITEM_TYPE_IPV4, 2696 .spec = &attributes->l3, 2697 }; 2698 break; 2699 case RTE_ETH_FLOW_NONFRAG_IPV6_UDP: 2700 attributes->l3.ipv6.hdr = (struct ipv6_hdr){ 2701 .hop_limits = input->flow.udp6_flow.ip.hop_limits, 2702 .proto = input->flow.udp6_flow.ip.proto, 2703 }; 2704 memcpy(attributes->l3.ipv6.hdr.src_addr, 2705 input->flow.udp6_flow.ip.src_ip, 2706 RTE_DIM(attributes->l3.ipv6.hdr.src_addr)); 2707 memcpy(attributes->l3.ipv6.hdr.dst_addr, 2708 input->flow.udp6_flow.ip.dst_ip, 2709 RTE_DIM(attributes->l3.ipv6.hdr.src_addr)); 2710 attributes->l4.udp.hdr = (struct udp_hdr){ 2711 .src_port = input->flow.udp6_flow.src_port, 2712 .dst_port = input->flow.udp6_flow.dst_port, 2713 }; 2714 attributes->items[1] = (struct rte_flow_item){ 2715 .type = RTE_FLOW_ITEM_TYPE_IPV6, 2716 .spec = &attributes->l3, 2717 }; 2718 attributes->items[2] = (struct rte_flow_item){ 2719 .type = RTE_FLOW_ITEM_TYPE_UDP, 2720 .spec = &attributes->l4, 2721 }; 2722 break; 2723 case RTE_ETH_FLOW_NONFRAG_IPV6_TCP: 2724 attributes->l3.ipv6.hdr = (struct ipv6_hdr){ 2725 .hop_limits = input->flow.tcp6_flow.ip.hop_limits, 2726 .proto = input->flow.tcp6_flow.ip.proto, 2727 }; 2728 memcpy(attributes->l3.ipv6.hdr.src_addr, 2729 input->flow.tcp6_flow.ip.src_ip, 2730 RTE_DIM(attributes->l3.ipv6.hdr.src_addr)); 2731 memcpy(attributes->l3.ipv6.hdr.dst_addr, 2732 input->flow.tcp6_flow.ip.dst_ip, 2733 RTE_DIM(attributes->l3.ipv6.hdr.src_addr)); 2734 attributes->l4.tcp.hdr = (struct tcp_hdr){ 2735 .src_port = input->flow.tcp6_flow.src_port, 2736 .dst_port = input->flow.tcp6_flow.dst_port, 2737 }; 2738 attributes->items[1] = (struct rte_flow_item){ 2739 .type = RTE_FLOW_ITEM_TYPE_IPV6, 2740 .spec = &attributes->l3, 2741 }; 2742 attributes->items[2] = (struct rte_flow_item){ 2743 .type = RTE_FLOW_ITEM_TYPE_TCP, 2744 .spec = &attributes->l4, 2745 }; 2746 break; 2747 case RTE_ETH_FLOW_NONFRAG_IPV6_OTHER: 2748 attributes->l3.ipv6.hdr = (struct ipv6_hdr){ 2749 .hop_limits = input->flow.ipv6_flow.hop_limits, 2750 .proto = input->flow.ipv6_flow.proto, 2751 }; 2752 memcpy(attributes->l3.ipv6.hdr.src_addr, 2753 input->flow.ipv6_flow.src_ip, 2754 RTE_DIM(attributes->l3.ipv6.hdr.src_addr)); 2755 memcpy(attributes->l3.ipv6.hdr.dst_addr, 2756 input->flow.ipv6_flow.dst_ip, 2757 RTE_DIM(attributes->l3.ipv6.hdr.src_addr)); 2758 attributes->items[1] = (struct rte_flow_item){ 2759 .type = RTE_FLOW_ITEM_TYPE_IPV6, 2760 .spec = &attributes->l3, 2761 }; 2762 break; 2763 default: 2764 ERROR("invalid flow type%d", 2765 fdir_filter->input.flow_type); 2766 return ENOTSUP; 2767 } 2768 return 0; 2769 } 2770 2771 /** 2772 * Add new flow director filter and store it in list. 2773 * 2774 * @param priv 2775 * Private structure. 2776 * @param fdir_filter 2777 * Flow director filter to add. 2778 * 2779 * @return 2780 * 0 on success, errno value on failure. 2781 */ 2782 static int 2783 priv_fdir_filter_add(struct priv *priv, 2784 const struct rte_eth_fdir_filter *fdir_filter) 2785 { 2786 struct mlx5_fdir attributes = { 2787 .attr.group = 0, 2788 .l2_mask = { 2789 .dst.addr_bytes = "\x00\x00\x00\x00\x00\x00", 2790 .src.addr_bytes = "\x00\x00\x00\x00\x00\x00", 2791 .type = 0, 2792 }, 2793 }; 2794 struct mlx5_flow_parse parser = { 2795 .layer = HASH_RXQ_ETH, 2796 }; 2797 struct rte_flow_error error; 2798 struct rte_flow *flow; 2799 int ret; 2800 2801 ret = priv_fdir_filter_convert(priv, fdir_filter, &attributes); 2802 if (ret) 2803 return -ret; 2804 ret = priv_flow_convert(priv, &attributes.attr, attributes.items, 2805 attributes.actions, &error, &parser); 2806 if (ret) 2807 return -ret; 2808 flow = priv_flow_create(priv, 2809 &priv->flows, 2810 &attributes.attr, 2811 attributes.items, 2812 attributes.actions, 2813 &error); 2814 if (flow) { 2815 DEBUG("FDIR created %p", (void *)flow); 2816 return 0; 2817 } 2818 return ENOTSUP; 2819 } 2820 2821 /** 2822 * Delete specific filter. 2823 * 2824 * @param priv 2825 * Private structure. 2826 * @param fdir_filter 2827 * Filter to be deleted. 2828 * 2829 * @return 2830 * 0 on success, errno value on failure. 2831 */ 2832 static int 2833 priv_fdir_filter_delete(struct priv *priv, 2834 const struct rte_eth_fdir_filter *fdir_filter) 2835 { 2836 struct mlx5_fdir attributes = { 2837 .attr.group = 0, 2838 }; 2839 struct mlx5_flow_parse parser = { 2840 .create = 1, 2841 .layer = HASH_RXQ_ETH, 2842 }; 2843 struct rte_flow_error error; 2844 struct rte_flow *flow; 2845 unsigned int i; 2846 int ret; 2847 2848 ret = priv_fdir_filter_convert(priv, fdir_filter, &attributes); 2849 if (ret) 2850 return -ret; 2851 ret = priv_flow_convert(priv, &attributes.attr, attributes.items, 2852 attributes.actions, &error, &parser); 2853 if (ret) 2854 goto exit; 2855 /* 2856 * Special case for drop action which is only set in the 2857 * specifications when the flow is created. In this situation the 2858 * drop specification is missing. 2859 */ 2860 if (parser.drop) { 2861 struct ibv_flow_spec_action_drop *drop; 2862 2863 drop = (void *)((uintptr_t)parser.queue[HASH_RXQ_ETH].ibv_attr + 2864 parser.queue[HASH_RXQ_ETH].offset); 2865 *drop = (struct ibv_flow_spec_action_drop){ 2866 .type = IBV_FLOW_SPEC_ACTION_DROP, 2867 .size = sizeof(struct ibv_flow_spec_action_drop), 2868 }; 2869 parser.queue[HASH_RXQ_ETH].ibv_attr->num_of_specs++; 2870 } 2871 TAILQ_FOREACH(flow, &priv->flows, next) { 2872 struct ibv_flow_attr *attr; 2873 struct ibv_spec_header *attr_h; 2874 void *spec; 2875 struct ibv_flow_attr *flow_attr; 2876 struct ibv_spec_header *flow_h; 2877 void *flow_spec; 2878 unsigned int specs_n; 2879 2880 attr = parser.queue[HASH_RXQ_ETH].ibv_attr; 2881 flow_attr = flow->frxq[HASH_RXQ_ETH].ibv_attr; 2882 /* Compare first the attributes. */ 2883 if (memcmp(attr, flow_attr, sizeof(struct ibv_flow_attr))) 2884 continue; 2885 if (attr->num_of_specs == 0) 2886 continue; 2887 spec = (void *)((uintptr_t)attr + 2888 sizeof(struct ibv_flow_attr)); 2889 flow_spec = (void *)((uintptr_t)flow_attr + 2890 sizeof(struct ibv_flow_attr)); 2891 specs_n = RTE_MIN(attr->num_of_specs, flow_attr->num_of_specs); 2892 for (i = 0; i != specs_n; ++i) { 2893 attr_h = spec; 2894 flow_h = flow_spec; 2895 if (memcmp(spec, flow_spec, 2896 RTE_MIN(attr_h->size, flow_h->size))) 2897 goto wrong_flow; 2898 spec = (void *)((uintptr_t)spec + attr_h->size); 2899 flow_spec = (void *)((uintptr_t)flow_spec + 2900 flow_h->size); 2901 } 2902 /* At this point, the flow match. */ 2903 break; 2904 wrong_flow: 2905 /* The flow does not match. */ 2906 continue; 2907 } 2908 if (flow) 2909 priv_flow_destroy(priv, &priv->flows, flow); 2910 exit: 2911 for (i = 0; i != hash_rxq_init_n; ++i) { 2912 if (parser.queue[i].ibv_attr) 2913 rte_free(parser.queue[i].ibv_attr); 2914 } 2915 return -ret; 2916 } 2917 2918 /** 2919 * Update queue for specific filter. 2920 * 2921 * @param priv 2922 * Private structure. 2923 * @param fdir_filter 2924 * Filter to be updated. 2925 * 2926 * @return 2927 * 0 on success, errno value on failure. 2928 */ 2929 static int 2930 priv_fdir_filter_update(struct priv *priv, 2931 const struct rte_eth_fdir_filter *fdir_filter) 2932 { 2933 int ret; 2934 2935 ret = priv_fdir_filter_delete(priv, fdir_filter); 2936 if (ret) 2937 return ret; 2938 ret = priv_fdir_filter_add(priv, fdir_filter); 2939 return ret; 2940 } 2941 2942 /** 2943 * Flush all filters. 2944 * 2945 * @param priv 2946 * Private structure. 2947 */ 2948 static void 2949 priv_fdir_filter_flush(struct priv *priv) 2950 { 2951 priv_flow_flush(priv, &priv->flows); 2952 } 2953 2954 /** 2955 * Get flow director information. 2956 * 2957 * @param priv 2958 * Private structure. 2959 * @param[out] fdir_info 2960 * Resulting flow director information. 2961 */ 2962 static void 2963 priv_fdir_info_get(struct priv *priv, struct rte_eth_fdir_info *fdir_info) 2964 { 2965 struct rte_eth_fdir_masks *mask = 2966 &priv->dev->data->dev_conf.fdir_conf.mask; 2967 2968 fdir_info->mode = priv->dev->data->dev_conf.fdir_conf.mode; 2969 fdir_info->guarant_spc = 0; 2970 rte_memcpy(&fdir_info->mask, mask, sizeof(fdir_info->mask)); 2971 fdir_info->max_flexpayload = 0; 2972 fdir_info->flow_types_mask[0] = 0; 2973 fdir_info->flex_payload_unit = 0; 2974 fdir_info->max_flex_payload_segment_num = 0; 2975 fdir_info->flex_payload_limit = 0; 2976 memset(&fdir_info->flex_conf, 0, sizeof(fdir_info->flex_conf)); 2977 } 2978 2979 /** 2980 * Deal with flow director operations. 2981 * 2982 * @param priv 2983 * Pointer to private structure. 2984 * @param filter_op 2985 * Operation to perform. 2986 * @param arg 2987 * Pointer to operation-specific structure. 2988 * 2989 * @return 2990 * 0 on success, errno value on failure. 2991 */ 2992 static int 2993 priv_fdir_ctrl_func(struct priv *priv, enum rte_filter_op filter_op, void *arg) 2994 { 2995 enum rte_fdir_mode fdir_mode = 2996 priv->dev->data->dev_conf.fdir_conf.mode; 2997 int ret = 0; 2998 2999 if (filter_op == RTE_ETH_FILTER_NOP) 3000 return 0; 3001 if (fdir_mode != RTE_FDIR_MODE_PERFECT && 3002 fdir_mode != RTE_FDIR_MODE_PERFECT_MAC_VLAN) { 3003 ERROR("%p: flow director mode %d not supported", 3004 (void *)priv, fdir_mode); 3005 return EINVAL; 3006 } 3007 switch (filter_op) { 3008 case RTE_ETH_FILTER_ADD: 3009 ret = priv_fdir_filter_add(priv, arg); 3010 break; 3011 case RTE_ETH_FILTER_UPDATE: 3012 ret = priv_fdir_filter_update(priv, arg); 3013 break; 3014 case RTE_ETH_FILTER_DELETE: 3015 ret = priv_fdir_filter_delete(priv, arg); 3016 break; 3017 case RTE_ETH_FILTER_FLUSH: 3018 priv_fdir_filter_flush(priv); 3019 break; 3020 case RTE_ETH_FILTER_INFO: 3021 priv_fdir_info_get(priv, arg); 3022 break; 3023 default: 3024 DEBUG("%p: unknown operation %u", (void *)priv, 3025 filter_op); 3026 ret = EINVAL; 3027 break; 3028 } 3029 return ret; 3030 } 3031 3032 /** 3033 * Manage filter operations. 3034 * 3035 * @param dev 3036 * Pointer to Ethernet device structure. 3037 * @param filter_type 3038 * Filter type. 3039 * @param filter_op 3040 * Operation to perform. 3041 * @param arg 3042 * Pointer to operation-specific structure. 3043 * 3044 * @return 3045 * 0 on success, negative errno value on failure. 3046 */ 3047 int 3048 mlx5_dev_filter_ctrl(struct rte_eth_dev *dev, 3049 enum rte_filter_type filter_type, 3050 enum rte_filter_op filter_op, 3051 void *arg) 3052 { 3053 int ret = EINVAL; 3054 struct priv *priv = dev->data->dev_private; 3055 3056 switch (filter_type) { 3057 case RTE_ETH_FILTER_GENERIC: 3058 if (filter_op != RTE_ETH_FILTER_GET) 3059 return -EINVAL; 3060 *(const void **)arg = &mlx5_flow_ops; 3061 return 0; 3062 case RTE_ETH_FILTER_FDIR: 3063 priv_lock(priv); 3064 ret = priv_fdir_ctrl_func(priv, filter_op, arg); 3065 priv_unlock(priv); 3066 break; 3067 default: 3068 ERROR("%p: filter type (%d) not supported", 3069 (void *)dev, filter_type); 3070 break; 3071 } 3072 return -ret; 3073 } 3074