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.h> 48 #include <rte_flow.h> 49 #include <rte_flow_driver.h> 50 #include <rte_malloc.h> 51 52 #include "mlx5.h" 53 #include "mlx5_prm.h" 54 55 static int 56 mlx5_flow_create_eth(const struct rte_flow_item *item, 57 const void *default_mask, 58 void *data); 59 60 static int 61 mlx5_flow_create_vlan(const struct rte_flow_item *item, 62 const void *default_mask, 63 void *data); 64 65 static int 66 mlx5_flow_create_ipv4(const struct rte_flow_item *item, 67 const void *default_mask, 68 void *data); 69 70 static int 71 mlx5_flow_create_ipv6(const struct rte_flow_item *item, 72 const void *default_mask, 73 void *data); 74 75 static int 76 mlx5_flow_create_udp(const struct rte_flow_item *item, 77 const void *default_mask, 78 void *data); 79 80 static int 81 mlx5_flow_create_tcp(const struct rte_flow_item *item, 82 const void *default_mask, 83 void *data); 84 85 static int 86 mlx5_flow_create_vxlan(const struct rte_flow_item *item, 87 const void *default_mask, 88 void *data); 89 90 struct rte_flow { 91 LIST_ENTRY(rte_flow) next; /**< Pointer to the next flow structure. */ 92 struct ibv_exp_flow_attr *ibv_attr; /**< Pointer to Verbs attributes. */ 93 struct ibv_exp_rwq_ind_table *ind_table; /**< Indirection table. */ 94 struct ibv_qp *qp; /**< Verbs queue pair. */ 95 struct ibv_exp_flow *ibv_flow; /**< Verbs flow. */ 96 struct ibv_exp_wq *wq; /**< Verbs work queue. */ 97 struct ibv_cq *cq; /**< Verbs completion queue. */ 98 struct rxq *rxq; /**< Pointer to the queue, NULL if drop queue. */ 99 uint32_t mark:1; /**< Set if the flow is marked. */ 100 }; 101 102 /** Static initializer for items. */ 103 #define ITEMS(...) \ 104 (const enum rte_flow_item_type []){ \ 105 __VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \ 106 } 107 108 /** Structure to generate a simple graph of layers supported by the NIC. */ 109 struct mlx5_flow_items { 110 /** List of possible actions for these items. */ 111 const enum rte_flow_action_type *const actions; 112 /** Bit-masks corresponding to the possibilities for the item. */ 113 const void *mask; 114 /** 115 * Default bit-masks to use when item->mask is not provided. When 116 * \default_mask is also NULL, the full supported bit-mask (\mask) is 117 * used instead. 118 */ 119 const void *default_mask; 120 /** Bit-masks size in bytes. */ 121 const unsigned int mask_sz; 122 /** 123 * Conversion function from rte_flow to NIC specific flow. 124 * 125 * @param item 126 * rte_flow item to convert. 127 * @param default_mask 128 * Default bit-masks to use when item->mask is not provided. 129 * @param data 130 * Internal structure to store the conversion. 131 * 132 * @return 133 * 0 on success, negative value otherwise. 134 */ 135 int (*convert)(const struct rte_flow_item *item, 136 const void *default_mask, 137 void *data); 138 /** Size in bytes of the destination structure. */ 139 const unsigned int dst_sz; 140 /** List of possible following items. */ 141 const enum rte_flow_item_type *const items; 142 }; 143 144 /** Valid action for this PMD. */ 145 static const enum rte_flow_action_type valid_actions[] = { 146 RTE_FLOW_ACTION_TYPE_DROP, 147 RTE_FLOW_ACTION_TYPE_QUEUE, 148 RTE_FLOW_ACTION_TYPE_MARK, 149 RTE_FLOW_ACTION_TYPE_END, 150 }; 151 152 /** Graph of supported items and associated actions. */ 153 static const struct mlx5_flow_items mlx5_flow_items[] = { 154 [RTE_FLOW_ITEM_TYPE_END] = { 155 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH, 156 RTE_FLOW_ITEM_TYPE_VXLAN), 157 }, 158 [RTE_FLOW_ITEM_TYPE_ETH] = { 159 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VLAN, 160 RTE_FLOW_ITEM_TYPE_IPV4, 161 RTE_FLOW_ITEM_TYPE_IPV6), 162 .actions = valid_actions, 163 .mask = &(const struct rte_flow_item_eth){ 164 .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff", 165 .src.addr_bytes = "\xff\xff\xff\xff\xff\xff", 166 }, 167 .mask_sz = sizeof(struct rte_flow_item_eth), 168 .convert = mlx5_flow_create_eth, 169 .dst_sz = sizeof(struct ibv_exp_flow_spec_eth), 170 }, 171 [RTE_FLOW_ITEM_TYPE_VLAN] = { 172 .items = ITEMS(RTE_FLOW_ITEM_TYPE_IPV4, 173 RTE_FLOW_ITEM_TYPE_IPV6), 174 .actions = valid_actions, 175 .mask = &(const struct rte_flow_item_vlan){ 176 .tci = -1, 177 }, 178 .mask_sz = sizeof(struct rte_flow_item_vlan), 179 .convert = mlx5_flow_create_vlan, 180 .dst_sz = 0, 181 }, 182 [RTE_FLOW_ITEM_TYPE_IPV4] = { 183 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP, 184 RTE_FLOW_ITEM_TYPE_TCP), 185 .actions = valid_actions, 186 .mask = &(const struct rte_flow_item_ipv4){ 187 .hdr = { 188 .src_addr = -1, 189 .dst_addr = -1, 190 .type_of_service = -1, 191 .next_proto_id = -1, 192 }, 193 }, 194 .default_mask = &(const struct rte_flow_item_ipv4){ 195 .hdr = { 196 .src_addr = -1, 197 .dst_addr = -1, 198 }, 199 }, 200 .mask_sz = sizeof(struct rte_flow_item_ipv4), 201 .convert = mlx5_flow_create_ipv4, 202 .dst_sz = sizeof(struct ibv_exp_flow_spec_ipv4_ext), 203 }, 204 [RTE_FLOW_ITEM_TYPE_IPV6] = { 205 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP, 206 RTE_FLOW_ITEM_TYPE_TCP), 207 .actions = valid_actions, 208 .mask = &(const struct rte_flow_item_ipv6){ 209 .hdr = { 210 .src_addr = { 211 0xff, 0xff, 0xff, 0xff, 212 0xff, 0xff, 0xff, 0xff, 213 0xff, 0xff, 0xff, 0xff, 214 0xff, 0xff, 0xff, 0xff, 215 }, 216 .dst_addr = { 217 0xff, 0xff, 0xff, 0xff, 218 0xff, 0xff, 0xff, 0xff, 219 0xff, 0xff, 0xff, 0xff, 220 0xff, 0xff, 0xff, 0xff, 221 }, 222 }, 223 }, 224 .mask_sz = sizeof(struct rte_flow_item_ipv6), 225 .convert = mlx5_flow_create_ipv6, 226 .dst_sz = sizeof(struct ibv_exp_flow_spec_ipv6), 227 }, 228 [RTE_FLOW_ITEM_TYPE_UDP] = { 229 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VXLAN), 230 .actions = valid_actions, 231 .mask = &(const struct rte_flow_item_udp){ 232 .hdr = { 233 .src_port = -1, 234 .dst_port = -1, 235 }, 236 }, 237 .mask_sz = sizeof(struct rte_flow_item_udp), 238 .convert = mlx5_flow_create_udp, 239 .dst_sz = sizeof(struct ibv_exp_flow_spec_tcp_udp), 240 }, 241 [RTE_FLOW_ITEM_TYPE_TCP] = { 242 .actions = valid_actions, 243 .mask = &(const struct rte_flow_item_tcp){ 244 .hdr = { 245 .src_port = -1, 246 .dst_port = -1, 247 }, 248 }, 249 .mask_sz = sizeof(struct rte_flow_item_tcp), 250 .convert = mlx5_flow_create_tcp, 251 .dst_sz = sizeof(struct ibv_exp_flow_spec_tcp_udp), 252 }, 253 [RTE_FLOW_ITEM_TYPE_VXLAN] = { 254 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH), 255 .actions = valid_actions, 256 .mask = &(const struct rte_flow_item_vxlan){ 257 .vni = "\xff\xff\xff", 258 }, 259 .mask_sz = sizeof(struct rte_flow_item_vxlan), 260 .convert = mlx5_flow_create_vxlan, 261 .dst_sz = sizeof(struct ibv_exp_flow_spec_tunnel), 262 }, 263 }; 264 265 /** Structure to pass to the conversion function. */ 266 struct mlx5_flow { 267 struct ibv_exp_flow_attr *ibv_attr; /**< Verbs attribute. */ 268 unsigned int offset; /**< Offset in bytes in the ibv_attr buffer. */ 269 uint32_t inner; /**< Set once VXLAN is encountered. */ 270 }; 271 272 struct mlx5_flow_action { 273 uint32_t queue:1; /**< Target is a receive queue. */ 274 uint32_t drop:1; /**< Target is a drop queue. */ 275 uint32_t mark:1; /**< Mark is present in the flow. */ 276 uint32_t queue_id; /**< Identifier of the queue. */ 277 uint32_t mark_id; /**< Mark identifier. */ 278 }; 279 280 /** 281 * Check support for a given item. 282 * 283 * @param item[in] 284 * Item specification. 285 * @param mask[in] 286 * Bit-masks covering supported fields to compare with spec, last and mask in 287 * \item. 288 * @param size 289 * Bit-Mask size in bytes. 290 * 291 * @return 292 * 0 on success. 293 */ 294 static int 295 mlx5_flow_item_validate(const struct rte_flow_item *item, 296 const uint8_t *mask, unsigned int size) 297 { 298 int ret = 0; 299 300 if (!item->spec && (item->mask || item->last)) 301 return -1; 302 if (item->spec && !item->mask) { 303 unsigned int i; 304 const uint8_t *spec = item->spec; 305 306 for (i = 0; i < size; ++i) 307 if ((spec[i] | mask[i]) != mask[i]) 308 return -1; 309 } 310 if (item->last && !item->mask) { 311 unsigned int i; 312 const uint8_t *spec = item->last; 313 314 for (i = 0; i < size; ++i) 315 if ((spec[i] | mask[i]) != mask[i]) 316 return -1; 317 } 318 if (item->mask) { 319 unsigned int i; 320 const uint8_t *spec = item->mask; 321 322 for (i = 0; i < size; ++i) 323 if ((spec[i] | mask[i]) != mask[i]) 324 return -1; 325 } 326 if (item->spec && item->last) { 327 uint8_t spec[size]; 328 uint8_t last[size]; 329 const uint8_t *apply = mask; 330 unsigned int i; 331 332 if (item->mask) 333 apply = item->mask; 334 for (i = 0; i < size; ++i) { 335 spec[i] = ((const uint8_t *)item->spec)[i] & apply[i]; 336 last[i] = ((const uint8_t *)item->last)[i] & apply[i]; 337 } 338 ret = memcmp(spec, last, size); 339 } 340 return ret; 341 } 342 343 /** 344 * Validate a flow supported by the NIC. 345 * 346 * @param priv 347 * Pointer to private structure. 348 * @param[in] attr 349 * Flow rule attributes. 350 * @param[in] pattern 351 * Pattern specification (list terminated by the END pattern item). 352 * @param[in] actions 353 * Associated actions (list terminated by the END action). 354 * @param[out] error 355 * Perform verbose error reporting if not NULL. 356 * @param[in, out] flow 357 * Flow structure to update. 358 * 359 * @return 360 * 0 on success, a negative errno value otherwise and rte_errno is set. 361 */ 362 static int 363 priv_flow_validate(struct priv *priv, 364 const struct rte_flow_attr *attr, 365 const struct rte_flow_item items[], 366 const struct rte_flow_action actions[], 367 struct rte_flow_error *error, 368 struct mlx5_flow *flow) 369 { 370 const struct mlx5_flow_items *cur_item = mlx5_flow_items; 371 struct mlx5_flow_action action = { 372 .queue = 0, 373 .drop = 0, 374 .mark = 0, 375 }; 376 377 (void)priv; 378 if (attr->group) { 379 rte_flow_error_set(error, ENOTSUP, 380 RTE_FLOW_ERROR_TYPE_ATTR_GROUP, 381 NULL, 382 "groups are not supported"); 383 return -rte_errno; 384 } 385 if (attr->priority) { 386 rte_flow_error_set(error, ENOTSUP, 387 RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY, 388 NULL, 389 "priorities are not supported"); 390 return -rte_errno; 391 } 392 if (attr->egress) { 393 rte_flow_error_set(error, ENOTSUP, 394 RTE_FLOW_ERROR_TYPE_ATTR_EGRESS, 395 NULL, 396 "egress is not supported"); 397 return -rte_errno; 398 } 399 if (!attr->ingress) { 400 rte_flow_error_set(error, ENOTSUP, 401 RTE_FLOW_ERROR_TYPE_ATTR_INGRESS, 402 NULL, 403 "only ingress is supported"); 404 return -rte_errno; 405 } 406 for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) { 407 const struct mlx5_flow_items *token = NULL; 408 unsigned int i; 409 int err; 410 411 if (items->type == RTE_FLOW_ITEM_TYPE_VOID) 412 continue; 413 /* Handle special situation for VLAN. */ 414 if (items->type == RTE_FLOW_ITEM_TYPE_VLAN) { 415 if (((const struct rte_flow_item_vlan *)items)->tci > 416 ETHER_MAX_VLAN_ID) { 417 rte_flow_error_set(error, ENOTSUP, 418 RTE_FLOW_ERROR_TYPE_ITEM, 419 items, 420 "wrong VLAN id value"); 421 return -rte_errno; 422 } 423 } 424 for (i = 0; 425 cur_item->items && 426 cur_item->items[i] != RTE_FLOW_ITEM_TYPE_END; 427 ++i) { 428 if (cur_item->items[i] == items->type) { 429 token = &mlx5_flow_items[items->type]; 430 break; 431 } 432 } 433 if (!token) 434 goto exit_item_not_supported; 435 cur_item = token; 436 err = mlx5_flow_item_validate(items, 437 (const uint8_t *)cur_item->mask, 438 sizeof(cur_item->mask_sz)); 439 if (err) 440 goto exit_item_not_supported; 441 if (flow->ibv_attr && cur_item->convert) { 442 err = cur_item->convert(items, 443 (cur_item->default_mask ? 444 cur_item->default_mask : 445 cur_item->mask), 446 flow); 447 if (err) 448 goto exit_item_not_supported; 449 } 450 flow->offset += cur_item->dst_sz; 451 } 452 for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) { 453 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) { 454 continue; 455 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) { 456 action.drop = 1; 457 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) { 458 const struct rte_flow_action_queue *queue = 459 (const struct rte_flow_action_queue *) 460 actions->conf; 461 462 if (!queue || (queue->index > (priv->rxqs_n - 1))) 463 goto exit_action_not_supported; 464 action.queue = 1; 465 } else if (actions->type == RTE_FLOW_ACTION_TYPE_MARK) { 466 const struct rte_flow_action_mark *mark = 467 (const struct rte_flow_action_mark *) 468 actions->conf; 469 470 if (mark && (mark->id >= MLX5_FLOW_MARK_MAX)) { 471 rte_flow_error_set(error, ENOTSUP, 472 RTE_FLOW_ERROR_TYPE_ACTION, 473 actions, 474 "mark must be between 0" 475 " and 16777199"); 476 return -rte_errno; 477 } 478 action.mark = 1; 479 } else { 480 goto exit_action_not_supported; 481 } 482 } 483 if (action.mark && !flow->ibv_attr) 484 flow->offset += sizeof(struct ibv_exp_flow_spec_action_tag); 485 if (!action.queue && !action.drop) { 486 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE, 487 NULL, "no valid action"); 488 return -rte_errno; 489 } 490 return 0; 491 exit_item_not_supported: 492 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM, 493 items, "item not supported"); 494 return -rte_errno; 495 exit_action_not_supported: 496 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION, 497 actions, "action not supported"); 498 return -rte_errno; 499 } 500 501 /** 502 * Validate a flow supported by the NIC. 503 * 504 * @see rte_flow_validate() 505 * @see rte_flow_ops 506 */ 507 int 508 mlx5_flow_validate(struct rte_eth_dev *dev, 509 const struct rte_flow_attr *attr, 510 const struct rte_flow_item items[], 511 const struct rte_flow_action actions[], 512 struct rte_flow_error *error) 513 { 514 struct priv *priv = dev->data->dev_private; 515 int ret; 516 struct mlx5_flow flow = { .offset = sizeof(struct ibv_exp_flow_attr) }; 517 518 priv_lock(priv); 519 ret = priv_flow_validate(priv, attr, items, actions, error, &flow); 520 priv_unlock(priv); 521 return ret; 522 } 523 524 /** 525 * Convert Ethernet item to Verbs specification. 526 * 527 * @param item[in] 528 * Item specification. 529 * @param default_mask[in] 530 * Default bit-masks to use when item->mask is not provided. 531 * @param data[in, out] 532 * User structure. 533 */ 534 static int 535 mlx5_flow_create_eth(const struct rte_flow_item *item, 536 const void *default_mask, 537 void *data) 538 { 539 const struct rte_flow_item_eth *spec = item->spec; 540 const struct rte_flow_item_eth *mask = item->mask; 541 struct mlx5_flow *flow = (struct mlx5_flow *)data; 542 struct ibv_exp_flow_spec_eth *eth; 543 const unsigned int eth_size = sizeof(struct ibv_exp_flow_spec_eth); 544 unsigned int i; 545 546 ++flow->ibv_attr->num_of_specs; 547 flow->ibv_attr->priority = 2; 548 eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset); 549 *eth = (struct ibv_exp_flow_spec_eth) { 550 .type = flow->inner | IBV_EXP_FLOW_SPEC_ETH, 551 .size = eth_size, 552 }; 553 if (!spec) 554 return 0; 555 if (!mask) 556 mask = default_mask; 557 memcpy(eth->val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN); 558 memcpy(eth->val.src_mac, spec->src.addr_bytes, ETHER_ADDR_LEN); 559 memcpy(eth->mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN); 560 memcpy(eth->mask.src_mac, mask->src.addr_bytes, ETHER_ADDR_LEN); 561 /* Remove unwanted bits from values. */ 562 for (i = 0; i < ETHER_ADDR_LEN; ++i) { 563 eth->val.dst_mac[i] &= eth->mask.dst_mac[i]; 564 eth->val.src_mac[i] &= eth->mask.src_mac[i]; 565 } 566 return 0; 567 } 568 569 /** 570 * Convert VLAN item to Verbs specification. 571 * 572 * @param item[in] 573 * Item specification. 574 * @param default_mask[in] 575 * Default bit-masks to use when item->mask is not provided. 576 * @param data[in, out] 577 * User structure. 578 */ 579 static int 580 mlx5_flow_create_vlan(const struct rte_flow_item *item, 581 const void *default_mask, 582 void *data) 583 { 584 const struct rte_flow_item_vlan *spec = item->spec; 585 const struct rte_flow_item_vlan *mask = item->mask; 586 struct mlx5_flow *flow = (struct mlx5_flow *)data; 587 struct ibv_exp_flow_spec_eth *eth; 588 const unsigned int eth_size = sizeof(struct ibv_exp_flow_spec_eth); 589 590 eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset - eth_size); 591 if (!spec) 592 return 0; 593 if (!mask) 594 mask = default_mask; 595 eth->val.vlan_tag = spec->tci; 596 eth->mask.vlan_tag = mask->tci; 597 eth->val.vlan_tag &= eth->mask.vlan_tag; 598 return 0; 599 } 600 601 /** 602 * Convert IPv4 item to Verbs specification. 603 * 604 * @param item[in] 605 * Item specification. 606 * @param default_mask[in] 607 * Default bit-masks to use when item->mask is not provided. 608 * @param data[in, out] 609 * User structure. 610 */ 611 static int 612 mlx5_flow_create_ipv4(const struct rte_flow_item *item, 613 const void *default_mask, 614 void *data) 615 { 616 const struct rte_flow_item_ipv4 *spec = item->spec; 617 const struct rte_flow_item_ipv4 *mask = item->mask; 618 struct mlx5_flow *flow = (struct mlx5_flow *)data; 619 struct ibv_exp_flow_spec_ipv4_ext *ipv4; 620 unsigned int ipv4_size = sizeof(struct ibv_exp_flow_spec_ipv4_ext); 621 622 ++flow->ibv_attr->num_of_specs; 623 flow->ibv_attr->priority = 1; 624 ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->offset); 625 *ipv4 = (struct ibv_exp_flow_spec_ipv4_ext) { 626 .type = flow->inner | IBV_EXP_FLOW_SPEC_IPV4_EXT, 627 .size = ipv4_size, 628 }; 629 if (!spec) 630 return 0; 631 if (!mask) 632 mask = default_mask; 633 ipv4->val = (struct ibv_exp_flow_ipv4_ext_filter){ 634 .src_ip = spec->hdr.src_addr, 635 .dst_ip = spec->hdr.dst_addr, 636 .proto = spec->hdr.next_proto_id, 637 .tos = spec->hdr.type_of_service, 638 }; 639 ipv4->mask = (struct ibv_exp_flow_ipv4_ext_filter){ 640 .src_ip = mask->hdr.src_addr, 641 .dst_ip = mask->hdr.dst_addr, 642 .proto = mask->hdr.next_proto_id, 643 .tos = mask->hdr.type_of_service, 644 }; 645 /* Remove unwanted bits from values. */ 646 ipv4->val.src_ip &= ipv4->mask.src_ip; 647 ipv4->val.dst_ip &= ipv4->mask.dst_ip; 648 ipv4->val.proto &= ipv4->mask.proto; 649 ipv4->val.tos &= ipv4->mask.tos; 650 return 0; 651 } 652 653 /** 654 * Convert IPv6 item to Verbs specification. 655 * 656 * @param item[in] 657 * Item specification. 658 * @param default_mask[in] 659 * Default bit-masks to use when item->mask is not provided. 660 * @param data[in, out] 661 * User structure. 662 */ 663 static int 664 mlx5_flow_create_ipv6(const struct rte_flow_item *item, 665 const void *default_mask, 666 void *data) 667 { 668 const struct rte_flow_item_ipv6 *spec = item->spec; 669 const struct rte_flow_item_ipv6 *mask = item->mask; 670 struct mlx5_flow *flow = (struct mlx5_flow *)data; 671 struct ibv_exp_flow_spec_ipv6 *ipv6; 672 unsigned int ipv6_size = sizeof(struct ibv_exp_flow_spec_ipv6); 673 unsigned int i; 674 675 ++flow->ibv_attr->num_of_specs; 676 flow->ibv_attr->priority = 1; 677 ipv6 = (void *)((uintptr_t)flow->ibv_attr + flow->offset); 678 *ipv6 = (struct ibv_exp_flow_spec_ipv6) { 679 .type = flow->inner | IBV_EXP_FLOW_SPEC_IPV6, 680 .size = ipv6_size, 681 }; 682 if (!spec) 683 return 0; 684 if (!mask) 685 mask = default_mask; 686 memcpy(ipv6->val.src_ip, spec->hdr.src_addr, 687 RTE_DIM(ipv6->val.src_ip)); 688 memcpy(ipv6->val.dst_ip, spec->hdr.dst_addr, 689 RTE_DIM(ipv6->val.dst_ip)); 690 memcpy(ipv6->mask.src_ip, mask->hdr.src_addr, 691 RTE_DIM(ipv6->mask.src_ip)); 692 memcpy(ipv6->mask.dst_ip, mask->hdr.dst_addr, 693 RTE_DIM(ipv6->mask.dst_ip)); 694 /* Remove unwanted bits from values. */ 695 for (i = 0; i < RTE_DIM(ipv6->val.src_ip); ++i) { 696 ipv6->val.src_ip[i] &= ipv6->mask.src_ip[i]; 697 ipv6->val.dst_ip[i] &= ipv6->mask.dst_ip[i]; 698 } 699 return 0; 700 } 701 702 /** 703 * Convert UDP item to Verbs specification. 704 * 705 * @param item[in] 706 * Item specification. 707 * @param default_mask[in] 708 * Default bit-masks to use when item->mask is not provided. 709 * @param data[in, out] 710 * User structure. 711 */ 712 static int 713 mlx5_flow_create_udp(const struct rte_flow_item *item, 714 const void *default_mask, 715 void *data) 716 { 717 const struct rte_flow_item_udp *spec = item->spec; 718 const struct rte_flow_item_udp *mask = item->mask; 719 struct mlx5_flow *flow = (struct mlx5_flow *)data; 720 struct ibv_exp_flow_spec_tcp_udp *udp; 721 unsigned int udp_size = sizeof(struct ibv_exp_flow_spec_tcp_udp); 722 723 ++flow->ibv_attr->num_of_specs; 724 flow->ibv_attr->priority = 0; 725 udp = (void *)((uintptr_t)flow->ibv_attr + flow->offset); 726 *udp = (struct ibv_exp_flow_spec_tcp_udp) { 727 .type = flow->inner | IBV_EXP_FLOW_SPEC_UDP, 728 .size = udp_size, 729 }; 730 if (!spec) 731 return 0; 732 if (!mask) 733 mask = default_mask; 734 udp->val.dst_port = spec->hdr.dst_port; 735 udp->val.src_port = spec->hdr.src_port; 736 udp->mask.dst_port = mask->hdr.dst_port; 737 udp->mask.src_port = mask->hdr.src_port; 738 /* Remove unwanted bits from values. */ 739 udp->val.src_port &= udp->mask.src_port; 740 udp->val.dst_port &= udp->mask.dst_port; 741 return 0; 742 } 743 744 /** 745 * Convert TCP item to Verbs specification. 746 * 747 * @param item[in] 748 * Item specification. 749 * @param default_mask[in] 750 * Default bit-masks to use when item->mask is not provided. 751 * @param data[in, out] 752 * User structure. 753 */ 754 static int 755 mlx5_flow_create_tcp(const struct rte_flow_item *item, 756 const void *default_mask, 757 void *data) 758 { 759 const struct rte_flow_item_tcp *spec = item->spec; 760 const struct rte_flow_item_tcp *mask = item->mask; 761 struct mlx5_flow *flow = (struct mlx5_flow *)data; 762 struct ibv_exp_flow_spec_tcp_udp *tcp; 763 unsigned int tcp_size = sizeof(struct ibv_exp_flow_spec_tcp_udp); 764 765 ++flow->ibv_attr->num_of_specs; 766 flow->ibv_attr->priority = 0; 767 tcp = (void *)((uintptr_t)flow->ibv_attr + flow->offset); 768 *tcp = (struct ibv_exp_flow_spec_tcp_udp) { 769 .type = flow->inner | IBV_EXP_FLOW_SPEC_TCP, 770 .size = tcp_size, 771 }; 772 if (!spec) 773 return 0; 774 if (!mask) 775 mask = default_mask; 776 tcp->val.dst_port = spec->hdr.dst_port; 777 tcp->val.src_port = spec->hdr.src_port; 778 tcp->mask.dst_port = mask->hdr.dst_port; 779 tcp->mask.src_port = mask->hdr.src_port; 780 /* Remove unwanted bits from values. */ 781 tcp->val.src_port &= tcp->mask.src_port; 782 tcp->val.dst_port &= tcp->mask.dst_port; 783 return 0; 784 } 785 786 /** 787 * Convert VXLAN item to Verbs specification. 788 * 789 * @param item[in] 790 * Item specification. 791 * @param default_mask[in] 792 * Default bit-masks to use when item->mask is not provided. 793 * @param data[in, out] 794 * User structure. 795 */ 796 static int 797 mlx5_flow_create_vxlan(const struct rte_flow_item *item, 798 const void *default_mask, 799 void *data) 800 { 801 const struct rte_flow_item_vxlan *spec = item->spec; 802 const struct rte_flow_item_vxlan *mask = item->mask; 803 struct mlx5_flow *flow = (struct mlx5_flow *)data; 804 struct ibv_exp_flow_spec_tunnel *vxlan; 805 unsigned int size = sizeof(struct ibv_exp_flow_spec_tunnel); 806 union vni { 807 uint32_t vlan_id; 808 uint8_t vni[4]; 809 } id; 810 811 ++flow->ibv_attr->num_of_specs; 812 flow->ibv_attr->priority = 0; 813 id.vni[0] = 0; 814 vxlan = (void *)((uintptr_t)flow->ibv_attr + flow->offset); 815 *vxlan = (struct ibv_exp_flow_spec_tunnel) { 816 .type = flow->inner | IBV_EXP_FLOW_SPEC_VXLAN_TUNNEL, 817 .size = size, 818 }; 819 flow->inner = IBV_EXP_FLOW_SPEC_INNER; 820 if (!spec) 821 return 0; 822 if (!mask) 823 mask = default_mask; 824 memcpy(&id.vni[1], spec->vni, 3); 825 vxlan->val.tunnel_id = id.vlan_id; 826 memcpy(&id.vni[1], mask->vni, 3); 827 vxlan->mask.tunnel_id = id.vlan_id; 828 /* Remove unwanted bits from values. */ 829 vxlan->val.tunnel_id &= vxlan->mask.tunnel_id; 830 return 0; 831 } 832 833 /** 834 * Convert mark/flag action to Verbs specification. 835 * 836 * @param flow 837 * Pointer to MLX5 flow structure. 838 * @param mark_id 839 * Mark identifier. 840 */ 841 static int 842 mlx5_flow_create_flag_mark(struct mlx5_flow *flow, uint32_t mark_id) 843 { 844 struct ibv_exp_flow_spec_action_tag *tag; 845 unsigned int size = sizeof(struct ibv_exp_flow_spec_action_tag); 846 847 tag = (void *)((uintptr_t)flow->ibv_attr + flow->offset); 848 *tag = (struct ibv_exp_flow_spec_action_tag){ 849 .type = IBV_EXP_FLOW_SPEC_ACTION_TAG, 850 .size = size, 851 .tag_id = mlx5_flow_mark_set(mark_id), 852 }; 853 ++flow->ibv_attr->num_of_specs; 854 return 0; 855 } 856 857 /** 858 * Complete flow rule creation. 859 * 860 * @param priv 861 * Pointer to private structure. 862 * @param ibv_attr 863 * Verbs flow attributes. 864 * @param action 865 * Target action structure. 866 * @param[out] error 867 * Perform verbose error reporting if not NULL. 868 * 869 * @return 870 * A flow if the rule could be created. 871 */ 872 static struct rte_flow * 873 priv_flow_create_action_queue(struct priv *priv, 874 struct ibv_exp_flow_attr *ibv_attr, 875 struct mlx5_flow_action *action, 876 struct rte_flow_error *error) 877 { 878 struct rxq_ctrl *rxq; 879 struct rte_flow *rte_flow; 880 881 assert(priv->pd); 882 assert(priv->ctx); 883 rte_flow = rte_calloc(__func__, 1, sizeof(*rte_flow), 0); 884 if (!rte_flow) { 885 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE, 886 NULL, "cannot allocate flow memory"); 887 return NULL; 888 } 889 if (action->drop) { 890 rte_flow->cq = 891 ibv_exp_create_cq(priv->ctx, 1, NULL, NULL, 0, 892 &(struct ibv_exp_cq_init_attr){ 893 .comp_mask = 0, 894 }); 895 if (!rte_flow->cq) { 896 rte_flow_error_set(error, ENOMEM, 897 RTE_FLOW_ERROR_TYPE_HANDLE, 898 NULL, "cannot allocate CQ"); 899 goto error; 900 } 901 rte_flow->wq = ibv_exp_create_wq(priv->ctx, 902 &(struct ibv_exp_wq_init_attr){ 903 .wq_type = IBV_EXP_WQT_RQ, 904 .max_recv_wr = 1, 905 .max_recv_sge = 1, 906 .pd = priv->pd, 907 .cq = rte_flow->cq, 908 }); 909 } else { 910 rxq = container_of((*priv->rxqs)[action->queue_id], 911 struct rxq_ctrl, rxq); 912 rte_flow->rxq = &rxq->rxq; 913 rxq->rxq.mark |= action->mark; 914 rte_flow->wq = rxq->wq; 915 } 916 rte_flow->mark = action->mark; 917 rte_flow->ibv_attr = ibv_attr; 918 rte_flow->ind_table = ibv_exp_create_rwq_ind_table( 919 priv->ctx, 920 &(struct ibv_exp_rwq_ind_table_init_attr){ 921 .pd = priv->pd, 922 .log_ind_tbl_size = 0, 923 .ind_tbl = &rte_flow->wq, 924 .comp_mask = 0, 925 }); 926 if (!rte_flow->ind_table) { 927 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE, 928 NULL, "cannot allocate indirection table"); 929 goto error; 930 } 931 rte_flow->qp = ibv_exp_create_qp( 932 priv->ctx, 933 &(struct ibv_exp_qp_init_attr){ 934 .qp_type = IBV_QPT_RAW_PACKET, 935 .comp_mask = 936 IBV_EXP_QP_INIT_ATTR_PD | 937 IBV_EXP_QP_INIT_ATTR_PORT | 938 IBV_EXP_QP_INIT_ATTR_RX_HASH, 939 .pd = priv->pd, 940 .rx_hash_conf = &(struct ibv_exp_rx_hash_conf){ 941 .rx_hash_function = 942 IBV_EXP_RX_HASH_FUNC_TOEPLITZ, 943 .rx_hash_key_len = rss_hash_default_key_len, 944 .rx_hash_key = rss_hash_default_key, 945 .rx_hash_fields_mask = 0, 946 .rwq_ind_tbl = rte_flow->ind_table, 947 }, 948 .port_num = priv->port, 949 }); 950 if (!rte_flow->qp) { 951 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE, 952 NULL, "cannot allocate QP"); 953 goto error; 954 } 955 rte_flow->ibv_flow = ibv_exp_create_flow(rte_flow->qp, 956 rte_flow->ibv_attr); 957 if (!rte_flow->ibv_flow) { 958 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE, 959 NULL, "flow rule creation failure"); 960 goto error; 961 } 962 return rte_flow; 963 error: 964 assert(rte_flow); 965 if (rte_flow->qp) 966 ibv_destroy_qp(rte_flow->qp); 967 if (rte_flow->ind_table) 968 ibv_exp_destroy_rwq_ind_table(rte_flow->ind_table); 969 if (!rte_flow->rxq && rte_flow->wq) 970 ibv_exp_destroy_wq(rte_flow->wq); 971 if (!rte_flow->rxq && rte_flow->cq) 972 ibv_destroy_cq(rte_flow->cq); 973 rte_free(rte_flow->ibv_attr); 974 rte_free(rte_flow); 975 return NULL; 976 } 977 978 /** 979 * Convert a flow. 980 * 981 * @param priv 982 * Pointer to private structure. 983 * @param[in] attr 984 * Flow rule attributes. 985 * @param[in] pattern 986 * Pattern specification (list terminated by the END pattern item). 987 * @param[in] actions 988 * Associated actions (list terminated by the END action). 989 * @param[out] error 990 * Perform verbose error reporting if not NULL. 991 * 992 * @return 993 * A flow on success, NULL otherwise. 994 */ 995 static struct rte_flow * 996 priv_flow_create(struct priv *priv, 997 const struct rte_flow_attr *attr, 998 const struct rte_flow_item items[], 999 const struct rte_flow_action actions[], 1000 struct rte_flow_error *error) 1001 { 1002 struct rte_flow *rte_flow; 1003 struct mlx5_flow_action action; 1004 struct mlx5_flow flow = { .offset = sizeof(struct ibv_exp_flow_attr), }; 1005 int err; 1006 1007 err = priv_flow_validate(priv, attr, items, actions, error, &flow); 1008 if (err) 1009 goto exit; 1010 flow.ibv_attr = rte_malloc(__func__, flow.offset, 0); 1011 flow.offset = sizeof(struct ibv_exp_flow_attr); 1012 if (!flow.ibv_attr) { 1013 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE, 1014 NULL, "cannot allocate ibv_attr memory"); 1015 goto exit; 1016 } 1017 *flow.ibv_attr = (struct ibv_exp_flow_attr){ 1018 .type = IBV_EXP_FLOW_ATTR_NORMAL, 1019 .size = sizeof(struct ibv_exp_flow_attr), 1020 .priority = attr->priority, 1021 .num_of_specs = 0, 1022 .port = 0, 1023 .flags = 0, 1024 .reserved = 0, 1025 }; 1026 flow.inner = 0; 1027 claim_zero(priv_flow_validate(priv, attr, items, actions, 1028 error, &flow)); 1029 action = (struct mlx5_flow_action){ 1030 .queue = 0, 1031 .drop = 0, 1032 .mark = 0, 1033 .mark_id = MLX5_FLOW_MARK_DEFAULT, 1034 }; 1035 for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) { 1036 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) { 1037 continue; 1038 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) { 1039 action.queue = 1; 1040 action.queue_id = 1041 ((const struct rte_flow_action_queue *) 1042 actions->conf)->index; 1043 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) { 1044 action.drop = 1; 1045 } else if (actions->type == RTE_FLOW_ACTION_TYPE_MARK) { 1046 const struct rte_flow_action_mark *mark = 1047 (const struct rte_flow_action_mark *) 1048 actions->conf; 1049 1050 if (mark) 1051 action.mark_id = mark->id; 1052 action.mark = 1; 1053 } else { 1054 rte_flow_error_set(error, ENOTSUP, 1055 RTE_FLOW_ERROR_TYPE_ACTION, 1056 actions, "unsupported action"); 1057 goto exit; 1058 } 1059 } 1060 if (action.mark) { 1061 mlx5_flow_create_flag_mark(&flow, action.mark_id); 1062 flow.offset += sizeof(struct ibv_exp_flow_spec_action_tag); 1063 } 1064 rte_flow = priv_flow_create_action_queue(priv, flow.ibv_attr, 1065 &action, error); 1066 return rte_flow; 1067 exit: 1068 rte_free(flow.ibv_attr); 1069 return NULL; 1070 } 1071 1072 /** 1073 * Create a flow. 1074 * 1075 * @see rte_flow_create() 1076 * @see rte_flow_ops 1077 */ 1078 struct rte_flow * 1079 mlx5_flow_create(struct rte_eth_dev *dev, 1080 const struct rte_flow_attr *attr, 1081 const struct rte_flow_item items[], 1082 const struct rte_flow_action actions[], 1083 struct rte_flow_error *error) 1084 { 1085 struct priv *priv = dev->data->dev_private; 1086 struct rte_flow *flow; 1087 1088 priv_lock(priv); 1089 flow = priv_flow_create(priv, attr, items, actions, error); 1090 if (flow) { 1091 LIST_INSERT_HEAD(&priv->flows, flow, next); 1092 DEBUG("Flow created %p", (void *)flow); 1093 } 1094 priv_unlock(priv); 1095 return flow; 1096 } 1097 1098 /** 1099 * Destroy a flow. 1100 * 1101 * @param priv 1102 * Pointer to private structure. 1103 * @param[in] flow 1104 * Flow to destroy. 1105 */ 1106 static void 1107 priv_flow_destroy(struct priv *priv, 1108 struct rte_flow *flow) 1109 { 1110 (void)priv; 1111 LIST_REMOVE(flow, next); 1112 if (flow->ibv_flow) 1113 claim_zero(ibv_exp_destroy_flow(flow->ibv_flow)); 1114 if (flow->qp) 1115 claim_zero(ibv_destroy_qp(flow->qp)); 1116 if (flow->ind_table) 1117 claim_zero(ibv_exp_destroy_rwq_ind_table(flow->ind_table)); 1118 if (!flow->rxq && flow->wq) 1119 claim_zero(ibv_exp_destroy_wq(flow->wq)); 1120 if (!flow->rxq && flow->cq) 1121 claim_zero(ibv_destroy_cq(flow->cq)); 1122 if (flow->mark) { 1123 struct rte_flow *tmp; 1124 uint32_t mark_n = 0; 1125 1126 for (tmp = LIST_FIRST(&priv->flows); 1127 tmp; 1128 tmp = LIST_NEXT(tmp, next)) { 1129 if ((flow->rxq == tmp->rxq) && tmp->mark) 1130 ++mark_n; 1131 } 1132 flow->rxq->mark = !!mark_n; 1133 } 1134 rte_free(flow->ibv_attr); 1135 DEBUG("Flow destroyed %p", (void *)flow); 1136 rte_free(flow); 1137 } 1138 1139 /** 1140 * Destroy a flow. 1141 * 1142 * @see rte_flow_destroy() 1143 * @see rte_flow_ops 1144 */ 1145 int 1146 mlx5_flow_destroy(struct rte_eth_dev *dev, 1147 struct rte_flow *flow, 1148 struct rte_flow_error *error) 1149 { 1150 struct priv *priv = dev->data->dev_private; 1151 1152 (void)error; 1153 priv_lock(priv); 1154 priv_flow_destroy(priv, flow); 1155 priv_unlock(priv); 1156 return 0; 1157 } 1158 1159 /** 1160 * Destroy all flows. 1161 * 1162 * @param priv 1163 * Pointer to private structure. 1164 */ 1165 static void 1166 priv_flow_flush(struct priv *priv) 1167 { 1168 while (!LIST_EMPTY(&priv->flows)) { 1169 struct rte_flow *flow; 1170 1171 flow = LIST_FIRST(&priv->flows); 1172 priv_flow_destroy(priv, flow); 1173 } 1174 } 1175 1176 /** 1177 * Destroy all flows. 1178 * 1179 * @see rte_flow_flush() 1180 * @see rte_flow_ops 1181 */ 1182 int 1183 mlx5_flow_flush(struct rte_eth_dev *dev, 1184 struct rte_flow_error *error) 1185 { 1186 struct priv *priv = dev->data->dev_private; 1187 1188 (void)error; 1189 priv_lock(priv); 1190 priv_flow_flush(priv); 1191 priv_unlock(priv); 1192 return 0; 1193 } 1194 1195 /** 1196 * Remove all flows. 1197 * 1198 * Called by dev_stop() to remove all flows. 1199 * 1200 * @param priv 1201 * Pointer to private structure. 1202 */ 1203 void 1204 priv_flow_stop(struct priv *priv) 1205 { 1206 struct rte_flow *flow; 1207 1208 for (flow = LIST_FIRST(&priv->flows); 1209 flow; 1210 flow = LIST_NEXT(flow, next)) { 1211 claim_zero(ibv_exp_destroy_flow(flow->ibv_flow)); 1212 flow->ibv_flow = NULL; 1213 if (flow->mark) 1214 flow->rxq->mark = 0; 1215 DEBUG("Flow %p removed", (void *)flow); 1216 } 1217 } 1218 1219 /** 1220 * Add all flows. 1221 * 1222 * @param priv 1223 * Pointer to private structure. 1224 * 1225 * @return 1226 * 0 on success, a errno value otherwise and rte_errno is set. 1227 */ 1228 int 1229 priv_flow_start(struct priv *priv) 1230 { 1231 struct rte_flow *flow; 1232 1233 for (flow = LIST_FIRST(&priv->flows); 1234 flow; 1235 flow = LIST_NEXT(flow, next)) { 1236 flow->ibv_flow = ibv_exp_create_flow(flow->qp, 1237 flow->ibv_attr); 1238 if (!flow->ibv_flow) { 1239 DEBUG("Flow %p cannot be applied", (void *)flow); 1240 rte_errno = EINVAL; 1241 return rte_errno; 1242 } 1243 DEBUG("Flow %p applied", (void *)flow); 1244 if (flow->rxq) 1245 flow->rxq->mark |= flow->mark; 1246 } 1247 return 0; 1248 } 1249