drivers/net/mlx5/mlx5_flow.c

   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
  54 static int
  55 mlx5_flow_create_eth(const struct rte_flow_item *item,
  56                      const void *default_mask,
  57                      void *data);
  58
  59 static int
  60 mlx5_flow_create_vlan(const struct rte_flow_item *item,
  61                       const void *default_mask,
  62                       void *data);
  63
  64 static int
  65 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
  66                       const void *default_mask,
  67                       void *data);
  68
  69 static int
  70 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
  71                       const void *default_mask,
  72                       void *data);
  73
  74 static int
  75 mlx5_flow_create_udp(const struct rte_flow_item *item,
  76                      const void *default_mask,
  77                      void *data);
  78
  79 static int
  80 mlx5_flow_create_tcp(const struct rte_flow_item *item,
  81                      const void *default_mask,
  82                      void *data);
  83
  84 struct rte_flow {
  85         LIST_ENTRY(rte_flow) next; /**< Pointer to the next flow structure. */
  86         struct ibv_exp_flow_attr *ibv_attr; /**< Pointer to Verbs attributes. */
  87         struct ibv_exp_rwq_ind_table *ind_table; /**< Indirection table. */
  88         struct ibv_qp *qp; /**< Verbs queue pair. */
  89         struct ibv_exp_flow *ibv_flow; /**< Verbs flow. */
  90         struct ibv_exp_wq *wq; /**< Verbs work queue. */
  91         struct ibv_cq *cq; /**< Verbs completion queue. */
  92         struct rxq *rxq; /**< Pointer to the queue, NULL if drop queue. */
  93 };
  94
  95 /** Static initializer for items. */
  96 #define ITEMS(...) \
  97         (const enum rte_flow_item_type []){ \
  98                 __VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \
  99         }
 100
 101 /** Structure to generate a simple graph of layers supported by the NIC. */
 102 struct mlx5_flow_items {
 103         /** List of possible actions for these items. */
 104         const enum rte_flow_action_type *const actions;
 105         /** Bit-masks corresponding to the possibilities for the item. */
 106         const void *mask;
 107         /** Bit-masks size in bytes. */
 108         const unsigned int mask_sz;
 109         /**
 110          * Conversion function from rte_flow to NIC specific flow.
 111          *
 112          * @param item
 113          *   rte_flow item to convert.
 114          * @param default_mask
 115          *   Default bit-masks to use when item->mask is not provided.
 116          * @param data
 117          *   Internal structure to store the conversion.
 118          *
 119          * @return
 120          *   0 on success, negative value otherwise.
 121          */
 122         int (*convert)(const struct rte_flow_item *item,
 123                        const void *default_mask,
 124                        void *data);
 125         /** Size in bytes of the destination structure. */
 126         const unsigned int dst_sz;
 127         /** List of possible following items.  */
 128         const enum rte_flow_item_type *const items;
 129 };
 130
 131 /** Valid action for this PMD. */
 132 static const enum rte_flow_action_type valid_actions[] = {
 133         RTE_FLOW_ACTION_TYPE_DROP,
 134         RTE_FLOW_ACTION_TYPE_QUEUE,
 135         RTE_FLOW_ACTION_TYPE_END,
 136 };
 137
 138 /** Graph of supported items and associated actions. */
 139 static const struct mlx5_flow_items mlx5_flow_items[] = {
 140         [RTE_FLOW_ITEM_TYPE_END] = {
 141                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH),
 142         },
 143         [RTE_FLOW_ITEM_TYPE_ETH] = {
 144                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VLAN,
 145                                RTE_FLOW_ITEM_TYPE_IPV4,
 146                                RTE_FLOW_ITEM_TYPE_IPV6),
 147                 .actions = valid_actions,
 148                 .mask = &(const struct rte_flow_item_eth){
 149                         .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 150                         .src.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 151                 },
 152                 .mask_sz = sizeof(struct rte_flow_item_eth),
 153                 .convert = mlx5_flow_create_eth,
 154                 .dst_sz = sizeof(struct ibv_exp_flow_spec_eth),
 155         },
 156         [RTE_FLOW_ITEM_TYPE_VLAN] = {
 157                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_IPV4,
 158                                RTE_FLOW_ITEM_TYPE_IPV6),
 159                 .actions = valid_actions,
 160                 .mask = &(const struct rte_flow_item_vlan){
 161                         .tci = -1,
 162                 },
 163                 .mask_sz = sizeof(struct rte_flow_item_vlan),
 164                 .convert = mlx5_flow_create_vlan,
 165                 .dst_sz = 0,
 166         },
 167         [RTE_FLOW_ITEM_TYPE_IPV4] = {
 168                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 169                                RTE_FLOW_ITEM_TYPE_TCP),
 170                 .actions = valid_actions,
 171                 .mask = &(const struct rte_flow_item_ipv4){
 172                         .hdr = {
 173                                 .src_addr = -1,
 174                                 .dst_addr = -1,
 175                         },
 176                 },
 177                 .mask_sz = sizeof(struct rte_flow_item_ipv4),
 178                 .convert = mlx5_flow_create_ipv4,
 179                 .dst_sz = sizeof(struct ibv_exp_flow_spec_ipv4),
 180         },
 181         [RTE_FLOW_ITEM_TYPE_IPV6] = {
 182                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 183                                RTE_FLOW_ITEM_TYPE_TCP),
 184                 .actions = valid_actions,
 185                 .mask = &(const struct rte_flow_item_ipv6){
 186                         .hdr = {
 187                                 .src_addr = {
 188                                         0xff, 0xff, 0xff, 0xff,
 189                                         0xff, 0xff, 0xff, 0xff,
 190                                         0xff, 0xff, 0xff, 0xff,
 191                                         0xff, 0xff, 0xff, 0xff,
 192                                 },
 193                                 .dst_addr = {
 194                                         0xff, 0xff, 0xff, 0xff,
 195                                         0xff, 0xff, 0xff, 0xff,
 196                                         0xff, 0xff, 0xff, 0xff,
 197                                         0xff, 0xff, 0xff, 0xff,
 198                                 },
 199                         },
 200                 },
 201                 .mask_sz = sizeof(struct rte_flow_item_ipv6),
 202                 .convert = mlx5_flow_create_ipv6,
 203                 .dst_sz = sizeof(struct ibv_exp_flow_spec_ipv6),
 204         },
 205         [RTE_FLOW_ITEM_TYPE_UDP] = {
 206                 .actions = valid_actions,
 207                 .mask = &(const struct rte_flow_item_udp){
 208                         .hdr = {
 209                                 .src_port = -1,
 210                                 .dst_port = -1,
 211                         },
 212                 },
 213                 .mask_sz = sizeof(struct rte_flow_item_udp),
 214                 .convert = mlx5_flow_create_udp,
 215                 .dst_sz = sizeof(struct ibv_exp_flow_spec_tcp_udp),
 216         },
 217         [RTE_FLOW_ITEM_TYPE_TCP] = {
 218                 .actions = valid_actions,
 219                 .mask = &(const struct rte_flow_item_tcp){
 220                         .hdr = {
 221                                 .src_port = -1,
 222                                 .dst_port = -1,
 223                         },
 224                 },
 225                 .mask_sz = sizeof(struct rte_flow_item_tcp),
 226                 .convert = mlx5_flow_create_tcp,
 227                 .dst_sz = sizeof(struct ibv_exp_flow_spec_tcp_udp),
 228         },
 229 };
 230
 231 /** Structure to pass to the conversion function. */
 232 struct mlx5_flow {
 233         struct ibv_exp_flow_attr *ibv_attr; /**< Verbs attribute. */
 234         unsigned int offset; /**< Offset in bytes in the ibv_attr buffer. */
 235 };
 236
 237 struct mlx5_flow_action {
 238         uint32_t queue:1; /**< Target is a receive queue. */
 239         uint32_t drop:1; /**< Target is a drop queue. */
 240         uint32_t queue_id; /**< Identifier of the queue. */
 241 };
 242
 243 /**
 244  * Check support for a given item.
 245  *
 246  * @param item[in]
 247  *   Item specification.
 248  * @param mask[in]
 249  *   Bit-masks covering supported fields to compare with spec, last and mask in
 250  *   \item.
 251  * @param size
 252  *   Bit-Mask size in bytes.
 253  *
 254  * @return
 255  *   0 on success.
 256  */
 257 static int
 258 mlx5_flow_item_validate(const struct rte_flow_item *item,
 259                         const uint8_t *mask, unsigned int size)
 260 {
 261         int ret = 0;
 262
 263         if (!item->spec && (item->mask || item->last))
 264                 return -1;
 265         if (item->spec && !item->mask) {
 266                 unsigned int i;
 267                 const uint8_t *spec = item->spec;
 268
 269                 for (i = 0; i < size; ++i)
 270                         if ((spec[i] | mask[i]) != mask[i])
 271                                 return -1;
 272         }
 273         if (item->last && !item->mask) {
 274                 unsigned int i;
 275                 const uint8_t *spec = item->last;
 276
 277                 for (i = 0; i < size; ++i)
 278                         if ((spec[i] | mask[i]) != mask[i])
 279                                 return -1;
 280         }
 281         if (item->mask) {
 282                 unsigned int i;
 283                 const uint8_t *spec = item->mask;
 284
 285                 for (i = 0; i < size; ++i)
 286                         if ((spec[i] | mask[i]) != mask[i])
 287                                 return -1;
 288         }
 289         if (item->spec && item->last) {
 290                 uint8_t spec[size];
 291                 uint8_t last[size];
 292                 const uint8_t *apply = mask;
 293                 unsigned int i;
 294
 295                 if (item->mask)
 296                         apply = item->mask;
 297                 for (i = 0; i < size; ++i) {
 298                         spec[i] = ((const uint8_t *)item->spec)[i] & apply[i];
 299                         last[i] = ((const uint8_t *)item->last)[i] & apply[i];
 300                 }
 301                 ret = memcmp(spec, last, size);
 302         }
 303         return ret;
 304 }
 305
 306 /**
 307  * Validate a flow supported by the NIC.
 308  *
 309  * @param priv
 310  *   Pointer to private structure.
 311  * @param[in] attr
 312  *   Flow rule attributes.
 313  * @param[in] pattern
 314  *   Pattern specification (list terminated by the END pattern item).
 315  * @param[in] actions
 316  *   Associated actions (list terminated by the END action).
 317  * @param[out] error
 318  *   Perform verbose error reporting if not NULL.
 319  * @param[in, out] flow
 320  *   Flow structure to update.
 321  *
 322  * @return
 323  *   0 on success, a negative errno value otherwise and rte_errno is set.
 324  */
 325 static int
 326 priv_flow_validate(struct priv *priv,
 327                    const struct rte_flow_attr *attr,
 328                    const struct rte_flow_item items[],
 329                    const struct rte_flow_action actions[],
 330                    struct rte_flow_error *error,
 331                    struct mlx5_flow *flow)
 332 {
 333         const struct mlx5_flow_items *cur_item = mlx5_flow_items;
 334         struct mlx5_flow_action action = {
 335                 .queue = 0,
 336                 .drop = 0,
 337         };
 338
 339         (void)priv;
 340         if (attr->group) {
 341                 rte_flow_error_set(error, ENOTSUP,
 342                                    RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 343                                    NULL,
 344                                    "groups are not supported");
 345                 return -rte_errno;
 346         }
 347         if (attr->priority) {
 348                 rte_flow_error_set(error, ENOTSUP,
 349                                    RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 350                                    NULL,
 351                                    "priorities are not supported");
 352                 return -rte_errno;
 353         }
 354         if (attr->egress) {
 355                 rte_flow_error_set(error, ENOTSUP,
 356                                    RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 357                                    NULL,
 358                                    "egress is not supported");
 359                 return -rte_errno;
 360         }
 361         if (!attr->ingress) {
 362                 rte_flow_error_set(error, ENOTSUP,
 363                                    RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 364                                    NULL,
 365                                    "only ingress is supported");
 366                 return -rte_errno;
 367         }
 368         for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
 369                 const struct mlx5_flow_items *token = NULL;
 370                 unsigned int i;
 371                 int err;
 372
 373                 if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
 374                         continue;
 375                 /* Handle special situation for VLAN. */
 376                 if (items->type == RTE_FLOW_ITEM_TYPE_VLAN) {
 377                         if (((const struct rte_flow_item_vlan *)items)->tci >
 378                             ETHER_MAX_VLAN_ID) {
 379                                 rte_flow_error_set(error, ENOTSUP,
 380                                                    RTE_FLOW_ERROR_TYPE_ITEM,
 381                                                    items,
 382                                                    "wrong VLAN id value");
 383                                 return -rte_errno;
 384                         }
 385                 }
 386                 for (i = 0;
 387                      cur_item->items &&
 388                      cur_item->items[i] != RTE_FLOW_ITEM_TYPE_END;
 389                      ++i) {
 390                         if (cur_item->items[i] == items->type) {
 391                                 token = &mlx5_flow_items[items->type];
 392                                 break;
 393                         }
 394                 }
 395                 if (!token)
 396                         goto exit_item_not_supported;
 397                 cur_item = token;
 398                 err = mlx5_flow_item_validate(items,
 399                                               (const uint8_t *)cur_item->mask,
 400                                               sizeof(cur_item->mask_sz));
 401                 if (err)
 402                         goto exit_item_not_supported;
 403                 if (flow->ibv_attr && cur_item->convert) {
 404                         err = cur_item->convert(items, cur_item->mask, flow);
 405                         if (err)
 406                                 goto exit_item_not_supported;
 407                 }
 408                 flow->offset += cur_item->dst_sz;
 409         }
 410         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
 411                 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
 412                         continue;
 413                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
 414                         action.drop = 1;
 415                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
 416                         const struct rte_flow_action_queue *queue =
 417                                 (const struct rte_flow_action_queue *)
 418                                 actions->conf;
 419
 420                         if (!queue || (queue->index > (priv->rxqs_n - 1)))
 421                                 goto exit_action_not_supported;
 422                         action.queue = 1;
 423                 } else {
 424                         goto exit_action_not_supported;
 425                 }
 426         }
 427         if (!action.queue && !action.drop) {
 428                 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE,
 429                                    NULL, "no valid action");
 430                 return -rte_errno;
 431         }
 432         return 0;
 433 exit_item_not_supported:
 434         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 435                            items, "item not supported");
 436         return -rte_errno;
 437 exit_action_not_supported:
 438         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 439                            actions, "action not supported");
 440         return -rte_errno;
 441 }
 442
 443 /**
 444  * Validate a flow supported by the NIC.
 445  *
 446  * @see rte_flow_validate()
 447  * @see rte_flow_ops
 448  */
 449 int
 450 mlx5_flow_validate(struct rte_eth_dev *dev,
 451                    const struct rte_flow_attr *attr,
 452                    const struct rte_flow_item items[],
 453                    const struct rte_flow_action actions[],
 454                    struct rte_flow_error *error)
 455 {
 456         struct priv *priv = dev->data->dev_private;
 457         int ret;
 458         struct mlx5_flow flow = { .offset = sizeof(struct ibv_exp_flow_attr) };
 459
 460         priv_lock(priv);
 461         ret = priv_flow_validate(priv, attr, items, actions, error, &flow);
 462         priv_unlock(priv);
 463         return ret;
 464 }
 465
 466 /**
 467  * Convert Ethernet item to Verbs specification.
 468  *
 469  * @param item[in]
 470  *   Item specification.
 471  * @param default_mask[in]
 472  *   Default bit-masks to use when item->mask is not provided.
 473  * @param data[in, out]
 474  *   User structure.
 475  */
 476 static int
 477 mlx5_flow_create_eth(const struct rte_flow_item *item,
 478                      const void *default_mask,
 479                      void *data)
 480 {
 481         const struct rte_flow_item_eth *spec = item->spec;
 482         const struct rte_flow_item_eth *mask = item->mask;
 483         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 484         struct ibv_exp_flow_spec_eth *eth;
 485         const unsigned int eth_size = sizeof(struct ibv_exp_flow_spec_eth);
 486         unsigned int i;
 487
 488         ++flow->ibv_attr->num_of_specs;
 489         flow->ibv_attr->priority = 2;
 490         eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 491         *eth = (struct ibv_exp_flow_spec_eth) {
 492                 .type = IBV_EXP_FLOW_SPEC_ETH,
 493                 .size = eth_size,
 494         };
 495         if (!spec)
 496                 return 0;
 497         if (!mask)
 498                 mask = default_mask;
 499         memcpy(eth->val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
 500         memcpy(eth->val.src_mac, spec->src.addr_bytes, ETHER_ADDR_LEN);
 501         memcpy(eth->mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
 502         memcpy(eth->mask.src_mac, mask->src.addr_bytes, ETHER_ADDR_LEN);
 503         /* Remove unwanted bits from values. */
 504         for (i = 0; i < ETHER_ADDR_LEN; ++i) {
 505                 eth->val.dst_mac[i] &= eth->mask.dst_mac[i];
 506                 eth->val.src_mac[i] &= eth->mask.src_mac[i];
 507         }
 508         return 0;
 509 }
 510
 511 /**
 512  * Convert VLAN item to Verbs specification.
 513  *
 514  * @param item[in]
 515  *   Item specification.
 516  * @param default_mask[in]
 517  *   Default bit-masks to use when item->mask is not provided.
 518  * @param data[in, out]
 519  *   User structure.
 520  */
 521 static int
 522 mlx5_flow_create_vlan(const struct rte_flow_item *item,
 523                       const void *default_mask,
 524                       void *data)
 525 {
 526         const struct rte_flow_item_vlan *spec = item->spec;
 527         const struct rte_flow_item_vlan *mask = item->mask;
 528         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 529         struct ibv_exp_flow_spec_eth *eth;
 530         const unsigned int eth_size = sizeof(struct ibv_exp_flow_spec_eth);
 531
 532         eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset - eth_size);
 533         if (!spec)
 534                 return 0;
 535         if (!mask)
 536                 mask = default_mask;
 537         eth->val.vlan_tag = spec->tci;
 538         eth->mask.vlan_tag = mask->tci;
 539         eth->val.vlan_tag &= eth->mask.vlan_tag;
 540         return 0;
 541 }
 542
 543 /**
 544  * Convert IPv4 item to Verbs specification.
 545  *
 546  * @param item[in]
 547  *   Item specification.
 548  * @param default_mask[in]
 549  *   Default bit-masks to use when item->mask is not provided.
 550  * @param data[in, out]
 551  *   User structure.
 552  */
 553 static int
 554 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
 555                       const void *default_mask,
 556                       void *data)
 557 {
 558         const struct rte_flow_item_ipv4 *spec = item->spec;
 559         const struct rte_flow_item_ipv4 *mask = item->mask;
 560         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 561         struct ibv_exp_flow_spec_ipv4 *ipv4;
 562         unsigned int ipv4_size = sizeof(struct ibv_exp_flow_spec_ipv4);
 563
 564         ++flow->ibv_attr->num_of_specs;
 565         flow->ibv_attr->priority = 1;
 566         ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 567         *ipv4 = (struct ibv_exp_flow_spec_ipv4) {
 568                 .type = IBV_EXP_FLOW_SPEC_IPV4,
 569                 .size = ipv4_size,
 570         };
 571         if (!spec)
 572                 return 0;
 573         if (!mask)
 574                 mask = default_mask;
 575         ipv4->val = (struct ibv_exp_flow_ipv4_filter){
 576                 .src_ip = spec->hdr.src_addr,
 577                 .dst_ip = spec->hdr.dst_addr,
 578         };
 579         ipv4->mask = (struct ibv_exp_flow_ipv4_filter){
 580                 .src_ip = mask->hdr.src_addr,
 581                 .dst_ip = mask->hdr.dst_addr,
 582         };
 583         /* Remove unwanted bits from values. */
 584         ipv4->val.src_ip &= ipv4->mask.src_ip;
 585         ipv4->val.dst_ip &= ipv4->mask.dst_ip;
 586         return 0;
 587 }
 588
 589 /**
 590  * Convert IPv6 item to Verbs specification.
 591  *
 592  * @param item[in]
 593  *   Item specification.
 594  * @param default_mask[in]
 595  *   Default bit-masks to use when item->mask is not provided.
 596  * @param data[in, out]
 597  *   User structure.
 598  */
 599 static int
 600 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
 601                       const void *default_mask,
 602                       void *data)
 603 {
 604         const struct rte_flow_item_ipv6 *spec = item->spec;
 605         const struct rte_flow_item_ipv6 *mask = item->mask;
 606         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 607         struct ibv_exp_flow_spec_ipv6 *ipv6;
 608         unsigned int ipv6_size = sizeof(struct ibv_exp_flow_spec_ipv6);
 609         unsigned int i;
 610
 611         ++flow->ibv_attr->num_of_specs;
 612         flow->ibv_attr->priority = 1;
 613         ipv6 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 614         *ipv6 = (struct ibv_exp_flow_spec_ipv6) {
 615                 .type = IBV_EXP_FLOW_SPEC_IPV6,
 616                 .size = ipv6_size,
 617         };
 618         if (!spec)
 619                 return 0;
 620         if (!mask)
 621                 mask = default_mask;
 622         memcpy(ipv6->val.src_ip, spec->hdr.src_addr,
 623                RTE_DIM(ipv6->val.src_ip));
 624         memcpy(ipv6->val.dst_ip, spec->hdr.dst_addr,
 625                RTE_DIM(ipv6->val.dst_ip));
 626         memcpy(ipv6->mask.src_ip, mask->hdr.src_addr,
 627                RTE_DIM(ipv6->mask.src_ip));
 628         memcpy(ipv6->mask.dst_ip, mask->hdr.dst_addr,
 629                RTE_DIM(ipv6->mask.dst_ip));
 630         /* Remove unwanted bits from values. */
 631         for (i = 0; i < RTE_DIM(ipv6->val.src_ip); ++i) {
 632                 ipv6->val.src_ip[i] &= ipv6->mask.src_ip[i];
 633                 ipv6->val.dst_ip[i] &= ipv6->mask.dst_ip[i];
 634         }
 635         return 0;
 636 }
 637
 638 /**
 639  * Convert UDP item to Verbs specification.
 640  *
 641  * @param item[in]
 642  *   Item specification.
 643  * @param default_mask[in]
 644  *   Default bit-masks to use when item->mask is not provided.
 645  * @param data[in, out]
 646  *   User structure.
 647  */
 648 static int
 649 mlx5_flow_create_udp(const struct rte_flow_item *item,
 650                      const void *default_mask,
 651                      void *data)
 652 {
 653         const struct rte_flow_item_udp *spec = item->spec;
 654         const struct rte_flow_item_udp *mask = item->mask;
 655         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 656         struct ibv_exp_flow_spec_tcp_udp *udp;
 657         unsigned int udp_size = sizeof(struct ibv_exp_flow_spec_tcp_udp);
 658
 659         ++flow->ibv_attr->num_of_specs;
 660         flow->ibv_attr->priority = 0;
 661         udp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 662         *udp = (struct ibv_exp_flow_spec_tcp_udp) {
 663                 .type = IBV_EXP_FLOW_SPEC_UDP,
 664                 .size = udp_size,
 665         };
 666         if (!spec)
 667                 return 0;
 668         if (!mask)
 669                 mask = default_mask;
 670         udp->val.dst_port = spec->hdr.dst_port;
 671         udp->val.src_port = spec->hdr.src_port;
 672         udp->mask.dst_port = mask->hdr.dst_port;
 673         udp->mask.src_port = mask->hdr.src_port;
 674         /* Remove unwanted bits from values. */
 675         udp->val.src_port &= udp->mask.src_port;
 676         udp->val.dst_port &= udp->mask.dst_port;
 677         return 0;
 678 }
 679
 680 /**
 681  * Convert TCP item to Verbs specification.
 682  *
 683  * @param item[in]
 684  *   Item specification.
 685  * @param default_mask[in]
 686  *   Default bit-masks to use when item->mask is not provided.
 687  * @param data[in, out]
 688  *   User structure.
 689  */
 690 static int
 691 mlx5_flow_create_tcp(const struct rte_flow_item *item,
 692                      const void *default_mask,
 693                      void *data)
 694 {
 695         const struct rte_flow_item_tcp *spec = item->spec;
 696         const struct rte_flow_item_tcp *mask = item->mask;
 697         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 698         struct ibv_exp_flow_spec_tcp_udp *tcp;
 699         unsigned int tcp_size = sizeof(struct ibv_exp_flow_spec_tcp_udp);
 700
 701         ++flow->ibv_attr->num_of_specs;
 702         flow->ibv_attr->priority = 0;
 703         tcp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 704         *tcp = (struct ibv_exp_flow_spec_tcp_udp) {
 705                 .type = IBV_EXP_FLOW_SPEC_TCP,
 706                 .size = tcp_size,
 707         };
 708         if (!spec)
 709                 return 0;
 710         if (!mask)
 711                 mask = default_mask;
 712         tcp->val.dst_port = spec->hdr.dst_port;
 713         tcp->val.src_port = spec->hdr.src_port;
 714         tcp->mask.dst_port = mask->hdr.dst_port;
 715         tcp->mask.src_port = mask->hdr.src_port;
 716         /* Remove unwanted bits from values. */
 717         tcp->val.src_port &= tcp->mask.src_port;
 718         tcp->val.dst_port &= tcp->mask.dst_port;
 719         return 0;
 720 }
 721
 722 /**
 723  * Complete flow rule creation.
 724  *
 725  * @param priv
 726  *   Pointer to private structure.
 727  * @param ibv_attr
 728  *   Verbs flow attributes.
 729  * @param action
 730  *   Target action structure.
 731  * @param[out] error
 732  *   Perform verbose error reporting if not NULL.
 733  *
 734  * @return
 735  *   A flow if the rule could be created.
 736  */
 737 static struct rte_flow *
 738 priv_flow_create_action_queue(struct priv *priv,
 739                               struct ibv_exp_flow_attr *ibv_attr,
 740                               struct mlx5_flow_action *action,
 741                               struct rte_flow_error *error)
 742 {
 743         struct rxq_ctrl *rxq;
 744         struct rte_flow *rte_flow;
 745
 746         assert(priv->pd);
 747         assert(priv->ctx);
 748         rte_flow = rte_calloc(__func__, 1, sizeof(*rte_flow), 0);
 749         if (!rte_flow) {
 750                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 751                                    NULL, "cannot allocate flow memory");
 752                 return NULL;
 753         }
 754         if (action->drop) {
 755                 rte_flow->cq =
 756                         ibv_exp_create_cq(priv->ctx, 1, NULL, NULL, 0,
 757                                           &(struct ibv_exp_cq_init_attr){
 758                                                   .comp_mask = 0,
 759                                           });
 760                 if (!rte_flow->cq) {
 761                         rte_flow_error_set(error, ENOMEM,
 762                                            RTE_FLOW_ERROR_TYPE_HANDLE,
 763                                            NULL, "cannot allocate CQ");
 764                         goto error;
 765                 }
 766                 rte_flow->wq = ibv_exp_create_wq(priv->ctx,
 767                                                  &(struct ibv_exp_wq_init_attr){
 768                                                  .wq_type = IBV_EXP_WQT_RQ,
 769                                                  .max_recv_wr = 1,
 770                                                  .max_recv_sge = 1,
 771                                                  .pd = priv->pd,
 772                                                  .cq = rte_flow->cq,
 773                                                  });
 774         } else {
 775                 rxq = container_of((*priv->rxqs)[action->queue_id],
 776                                    struct rxq_ctrl, rxq);
 777                 rte_flow->rxq = &rxq->rxq;
 778                 rte_flow->wq = rxq->wq;
 779         }
 780         rte_flow->ibv_attr = ibv_attr;
 781         rte_flow->ind_table = ibv_exp_create_rwq_ind_table(
 782                 priv->ctx,
 783                 &(struct ibv_exp_rwq_ind_table_init_attr){
 784                         .pd = priv->pd,
 785                         .log_ind_tbl_size = 0,
 786                         .ind_tbl = &rte_flow->wq,
 787                         .comp_mask = 0,
 788                 });
 789         if (!rte_flow->ind_table) {
 790                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 791                                    NULL, "cannot allocate indirection table");
 792                 goto error;
 793         }
 794         rte_flow->qp = ibv_exp_create_qp(
 795                 priv->ctx,
 796                 &(struct ibv_exp_qp_init_attr){
 797                         .qp_type = IBV_QPT_RAW_PACKET,
 798                         .comp_mask =
 799                                 IBV_EXP_QP_INIT_ATTR_PD |
 800                                 IBV_EXP_QP_INIT_ATTR_PORT |
 801                                 IBV_EXP_QP_INIT_ATTR_RX_HASH,
 802                         .pd = priv->pd,
 803                         .rx_hash_conf = &(struct ibv_exp_rx_hash_conf){
 804                                 .rx_hash_function =
 805                                         IBV_EXP_RX_HASH_FUNC_TOEPLITZ,
 806                                 .rx_hash_key_len = rss_hash_default_key_len,
 807                                 .rx_hash_key = rss_hash_default_key,
 808                                 .rx_hash_fields_mask = 0,
 809                                 .rwq_ind_tbl = rte_flow->ind_table,
 810                         },
 811                         .port_num = priv->port,
 812                 });
 813         if (!rte_flow->qp) {
 814                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 815                                    NULL, "cannot allocate QP");
 816                 goto error;
 817         }
 818         rte_flow->ibv_flow = ibv_exp_create_flow(rte_flow->qp,
 819                                                  rte_flow->ibv_attr);
 820         if (!rte_flow->ibv_flow) {
 821                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 822                                    NULL, "flow rule creation failure");
 823                 goto error;
 824         }
 825         return rte_flow;
 826 error:
 827         assert(rte_flow);
 828         if (rte_flow->qp)
 829                 ibv_destroy_qp(rte_flow->qp);
 830         if (rte_flow->ind_table)
 831                 ibv_exp_destroy_rwq_ind_table(rte_flow->ind_table);
 832         if (!rte_flow->rxq && rte_flow->wq)
 833                 ibv_exp_destroy_wq(rte_flow->wq);
 834         if (!rte_flow->rxq && rte_flow->cq)
 835                 ibv_destroy_cq(rte_flow->cq);
 836         rte_free(rte_flow->ibv_attr);
 837         rte_free(rte_flow);
 838         return NULL;
 839 }
 840
 841 /**
 842  * Convert a flow.
 843  *
 844  * @param priv
 845  *   Pointer to private structure.
 846  * @param[in] attr
 847  *   Flow rule attributes.
 848  * @param[in] pattern
 849  *   Pattern specification (list terminated by the END pattern item).
 850  * @param[in] actions
 851  *   Associated actions (list terminated by the END action).
 852  * @param[out] error
 853  *   Perform verbose error reporting if not NULL.
 854  *
 855  * @return
 856  *   A flow on success, NULL otherwise.
 857  */
 858 static struct rte_flow *
 859 priv_flow_create(struct priv *priv,
 860                  const struct rte_flow_attr *attr,
 861                  const struct rte_flow_item items[],
 862                  const struct rte_flow_action actions[],
 863                  struct rte_flow_error *error)
 864 {
 865         struct rte_flow *rte_flow;
 866         struct mlx5_flow_action action;
 867         struct mlx5_flow flow = { .offset = sizeof(struct ibv_exp_flow_attr), };
 868         int err;
 869
 870         err = priv_flow_validate(priv, attr, items, actions, error, &flow);
 871         if (err)
 872                 goto exit;
 873         flow.ibv_attr = rte_malloc(__func__, flow.offset, 0);
 874         flow.offset = sizeof(struct ibv_exp_flow_attr);
 875         if (!flow.ibv_attr) {
 876                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 877                                    NULL, "cannot allocate ibv_attr memory");
 878                 goto exit;
 879         }
 880         *flow.ibv_attr = (struct ibv_exp_flow_attr){
 881                 .type = IBV_EXP_FLOW_ATTR_NORMAL,
 882                 .size = sizeof(struct ibv_exp_flow_attr),
 883                 .priority = attr->priority,
 884                 .num_of_specs = 0,
 885                 .port = 0,
 886                 .flags = 0,
 887                 .reserved = 0,
 888         };
 889         claim_zero(priv_flow_validate(priv, attr, items, actions,
 890                                       error, &flow));
 891         action = (struct mlx5_flow_action){
 892                 .queue = 0,
 893                 .drop = 0,
 894         };
 895         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
 896                 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
 897                         continue;
 898                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
 899                         action.queue = 1;
 900                         action.queue_id =
 901                                 ((const struct rte_flow_action_queue *)
 902                                  actions->conf)->index;
 903                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
 904                         action.drop = 1;
 905                 } else {
 906                         rte_flow_error_set(error, ENOTSUP,
 907                                            RTE_FLOW_ERROR_TYPE_ACTION,
 908                                            actions, "unsupported action");
 909                         goto exit;
 910                 }
 911         }
 912         rte_flow = priv_flow_create_action_queue(priv, flow.ibv_attr,
 913                                                  &action, error);
 914         return rte_flow;
 915 exit:
 916         rte_free(flow.ibv_attr);
 917         return NULL;
 918 }
 919
 920 /**
 921  * Create a flow.
 922  *
 923  * @see rte_flow_create()
 924  * @see rte_flow_ops
 925  */
 926 struct rte_flow *
 927 mlx5_flow_create(struct rte_eth_dev *dev,
 928                  const struct rte_flow_attr *attr,
 929                  const struct rte_flow_item items[],
 930                  const struct rte_flow_action actions[],
 931                  struct rte_flow_error *error)
 932 {
 933         struct priv *priv = dev->data->dev_private;
 934         struct rte_flow *flow;
 935
 936         priv_lock(priv);
 937         flow = priv_flow_create(priv, attr, items, actions, error);
 938         if (flow) {
 939                 LIST_INSERT_HEAD(&priv->flows, flow, next);
 940                 DEBUG("Flow created %p", (void *)flow);
 941         }
 942         priv_unlock(priv);
 943         return flow;
 944 }
 945
 946 /**
 947  * Destroy a flow.
 948  *
 949  * @param priv
 950  *   Pointer to private structure.
 951  * @param[in] flow
 952  *   Flow to destroy.
 953  */
 954 static void
 955 priv_flow_destroy(struct priv *priv,
 956                   struct rte_flow *flow)
 957 {
 958         (void)priv;
 959         LIST_REMOVE(flow, next);
 960         if (flow->ibv_flow)
 961                 claim_zero(ibv_exp_destroy_flow(flow->ibv_flow));
 962         if (flow->qp)
 963                 claim_zero(ibv_destroy_qp(flow->qp));
 964         if (flow->ind_table)
 965                 claim_zero(ibv_exp_destroy_rwq_ind_table(flow->ind_table));
 966         if (!flow->rxq && flow->wq)
 967                 claim_zero(ibv_exp_destroy_wq(flow->wq));
 968         if (!flow->rxq && flow->cq)
 969                 claim_zero(ibv_destroy_cq(flow->cq));
 970         rte_free(flow->ibv_attr);
 971         DEBUG("Flow destroyed %p", (void *)flow);
 972         rte_free(flow);
 973 }
 974
 975 /**
 976  * Destroy a flow.
 977  *
 978  * @see rte_flow_destroy()
 979  * @see rte_flow_ops
 980  */
 981 int
 982 mlx5_flow_destroy(struct rte_eth_dev *dev,
 983                   struct rte_flow *flow,
 984                   struct rte_flow_error *error)
 985 {
 986         struct priv *priv = dev->data->dev_private;
 987
 988         (void)error;
 989         priv_lock(priv);
 990         priv_flow_destroy(priv, flow);
 991         priv_unlock(priv);
 992         return 0;
 993 }
 994
 995 /**
 996  * Destroy all flows.
 997  *
 998  * @param priv
 999  *   Pointer to private structure.
1000  */
1001 static void
1002 priv_flow_flush(struct priv *priv)
1003 {
1004         while (!LIST_EMPTY(&priv->flows)) {
1005                 struct rte_flow *flow;
1006
1007                 flow = LIST_FIRST(&priv->flows);
1008                 priv_flow_destroy(priv, flow);
1009         }
1010 }
1011
1012 /**
1013  * Destroy all flows.
1014  *
1015  * @see rte_flow_flush()
1016  * @see rte_flow_ops
1017  */
1018 int
1019 mlx5_flow_flush(struct rte_eth_dev *dev,
1020                 struct rte_flow_error *error)
1021 {
1022         struct priv *priv = dev->data->dev_private;
1023
1024         (void)error;
1025         priv_lock(priv);
1026         priv_flow_flush(priv);
1027         priv_unlock(priv);
1028         return 0;
1029 }
1030
1031 /**
1032  * Remove all flows.
1033  *
1034  * Called by dev_stop() to remove all flows.
1035  *
1036  * @param priv
1037  *   Pointer to private structure.
1038  */
1039 void
1040 priv_flow_stop(struct priv *priv)
1041 {
1042         struct rte_flow *flow;
1043
1044         for (flow = LIST_FIRST(&priv->flows);
1045              flow;
1046              flow = LIST_NEXT(flow, next)) {
1047                 claim_zero(ibv_exp_destroy_flow(flow->ibv_flow));
1048                 flow->ibv_flow = NULL;
1049                 DEBUG("Flow %p removed", (void *)flow);
1050         }
1051 }
1052
1053 /**
1054  * Add all flows.
1055  *
1056  * @param priv
1057  *   Pointer to private structure.
1058  *
1059  * @return
1060  *   0 on success, a errno value otherwise and rte_errno is set.
1061  */
1062 int
1063 priv_flow_start(struct priv *priv)
1064 {
1065         struct rte_flow *flow;
1066
1067         for (flow = LIST_FIRST(&priv->flows);
1068              flow;
1069              flow = LIST_NEXT(flow, next)) {
1070                 flow->ibv_flow = ibv_exp_create_flow(flow->qp,
1071                                                      flow->ibv_attr);
1072                 if (!flow->ibv_flow) {
1073                         DEBUG("Flow %p cannot be applied", (void *)flow);
1074                         rte_errno = EINVAL;
1075                         return rte_errno;
1076                 }
1077                 DEBUG("Flow %p applied", (void *)flow);
1078         }
1079         return 0;
1080 }