drivers/net/mlx4/mlx4_flow.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright 2017 6WIND S.A.
   5  *   Copyright 2017 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 /**
  35  * @file
  36  * Flow API operations for mlx4 driver.
  37  */
  38
  39 #include <arpa/inet.h>
  40 #include <assert.h>
  41 #include <errno.h>
  42 #include <stdalign.h>
  43 #include <stddef.h>
  44 #include <stdint.h>
  45 #include <string.h>
  46 #include <sys/queue.h>
  47
  48 /* Verbs headers do not support -pedantic. */
  49 #ifdef PEDANTIC
  50 #pragma GCC diagnostic ignored "-Wpedantic"
  51 #endif
  52 #include <infiniband/verbs.h>
  53 #ifdef PEDANTIC
  54 #pragma GCC diagnostic error "-Wpedantic"
  55 #endif
  56
  57 #include <rte_byteorder.h>
  58 #include <rte_errno.h>
  59 #include <rte_eth_ctrl.h>
  60 #include <rte_ethdev.h>
  61 #include <rte_ether.h>
  62 #include <rte_flow.h>
  63 #include <rte_flow_driver.h>
  64 #include <rte_malloc.h>
  65
  66 /* PMD headers. */
  67 #include "mlx4.h"
  68 #include "mlx4_flow.h"
  69 #include "mlx4_rxtx.h"
  70 #include "mlx4_utils.h"
  71
  72 /** Static initializer for a list of subsequent item types. */
  73 #define NEXT_ITEM(...) \
  74         (const enum rte_flow_item_type []){ \
  75                 __VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \
  76         }
  77
  78 /** Processor structure associated with a flow item. */
  79 struct mlx4_flow_proc_item {
  80         /** Bit-mask for fields supported by this PMD. */
  81         const void *mask_support;
  82         /** Bit-mask to use when @p item->mask is not provided. */
  83         const void *mask_default;
  84         /** Size in bytes for @p mask_support and @p mask_default. */
  85         const unsigned int mask_sz;
  86         /** Merge a pattern item into a flow rule handle. */
  87         int (*merge)(struct rte_flow *flow,
  88                      const struct rte_flow_item *item,
  89                      const struct mlx4_flow_proc_item *proc,
  90                      struct rte_flow_error *error);
  91         /** Size in bytes of the destination structure. */
  92         const unsigned int dst_sz;
  93         /** List of possible subsequent items. */
  94         const enum rte_flow_item_type *const next_item;
  95 };
  96
  97 /** Shared resources for drop flow rules. */
  98 struct mlx4_drop {
  99         struct ibv_qp *qp; /**< QP target. */
 100         struct ibv_cq *cq; /**< CQ associated with above QP. */
 101         struct priv *priv; /**< Back pointer to private data. */
 102         uint32_t refcnt; /**< Reference count. */
 103 };
 104
 105 /**
 106  * Convert DPDK RSS hash fields to their Verbs equivalent.
 107  *
 108  * @param rss_hf
 109  *   Hash fields in DPDK format (see struct rte_eth_rss_conf).
 110  *
 111  * @return
 112  *   A valid Verbs RSS hash fields mask for mlx4 on success, (uint64_t)-1
 113  *   otherwise and rte_errno is set.
 114  */
 115 static uint64_t
 116 mlx4_conv_rss_hf(uint64_t rss_hf)
 117 {
 118         enum { IPV4, IPV6, TCP, UDP, };
 119         const uint64_t in[] = {
 120                 [IPV4] = (ETH_RSS_IPV4 |
 121                           ETH_RSS_FRAG_IPV4 |
 122                           ETH_RSS_NONFRAG_IPV4_TCP |
 123                           ETH_RSS_NONFRAG_IPV4_UDP |
 124                           ETH_RSS_NONFRAG_IPV4_OTHER),
 125                 [IPV6] = (ETH_RSS_IPV6 |
 126                           ETH_RSS_FRAG_IPV6 |
 127                           ETH_RSS_NONFRAG_IPV6_TCP |
 128                           ETH_RSS_NONFRAG_IPV6_UDP |
 129                           ETH_RSS_NONFRAG_IPV6_OTHER |
 130                           ETH_RSS_IPV6_EX |
 131                           ETH_RSS_IPV6_TCP_EX |
 132                           ETH_RSS_IPV6_UDP_EX),
 133                 [TCP] = (ETH_RSS_NONFRAG_IPV4_TCP |
 134                          ETH_RSS_NONFRAG_IPV6_TCP |
 135                          ETH_RSS_IPV6_TCP_EX),
 136                 /*
 137                  * UDP support is temporarily disabled due to an
 138                  * implementation issue in the kernel.
 139                  */
 140                 [UDP] = 0,
 141         };
 142         const uint64_t out[RTE_DIM(in)] = {
 143                 [IPV4] = IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4,
 144                 [IPV6] = IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6,
 145                 [TCP] = IBV_RX_HASH_SRC_PORT_TCP | IBV_RX_HASH_DST_PORT_TCP,
 146                 [UDP] = IBV_RX_HASH_SRC_PORT_UDP | IBV_RX_HASH_DST_PORT_UDP,
 147         };
 148         uint64_t seen = 0;
 149         uint64_t conv = 0;
 150         unsigned int i;
 151
 152         for (i = 0; i != RTE_DIM(in); ++i)
 153                 if (rss_hf & in[i]) {
 154                         seen |= rss_hf & in[i];
 155                         conv |= out[i];
 156                 }
 157         if (!(rss_hf & ~seen))
 158                 return conv;
 159         rte_errno = ENOTSUP;
 160         return (uint64_t)-1;
 161 }
 162
 163 /**
 164  * Merge Ethernet pattern item into flow rule handle.
 165  *
 166  * Additional mlx4-specific constraints on supported fields:
 167  *
 168  * - No support for partial masks, except in the specific case of matching
 169  *   all multicast traffic (@p spec->dst and @p mask->dst equal to
 170  *   01:00:00:00:00:00).
 171  * - Not providing @p item->spec or providing an empty @p mask->dst is
 172  *   *only* supported if the rule doesn't specify additional matching
 173  *   criteria (i.e. rule is promiscuous-like).
 174  *
 175  * @param[in, out] flow
 176  *   Flow rule handle to update.
 177  * @param[in] item
 178  *   Pattern item to merge.
 179  * @param[in] proc
 180  *   Associated item-processing object.
 181  * @param[out] error
 182  *   Perform verbose error reporting if not NULL.
 183  *
 184  * @return
 185  *   0 on success, a negative errno value otherwise and rte_errno is set.
 186  */
 187 static int
 188 mlx4_flow_merge_eth(struct rte_flow *flow,
 189                     const struct rte_flow_item *item,
 190                     const struct mlx4_flow_proc_item *proc,
 191                     struct rte_flow_error *error)
 192 {
 193         const struct rte_flow_item_eth *spec = item->spec;
 194         const struct rte_flow_item_eth *mask =
 195                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 196         struct ibv_flow_spec_eth *eth;
 197         const char *msg;
 198         unsigned int i;
 199
 200         if (!mask) {
 201                 flow->promisc = 1;
 202         } else {
 203                 uint32_t sum_dst = 0;
 204                 uint32_t sum_src = 0;
 205
 206                 for (i = 0; i != sizeof(mask->dst.addr_bytes); ++i) {
 207                         sum_dst += mask->dst.addr_bytes[i];
 208                         sum_src += mask->src.addr_bytes[i];
 209                 }
 210                 if (sum_src) {
 211                         msg = "mlx4 does not support source MAC matching";
 212                         goto error;
 213                 } else if (!sum_dst) {
 214                         flow->promisc = 1;
 215                 } else if (sum_dst == 1 && mask->dst.addr_bytes[0] == 1) {
 216                         if (!(spec->dst.addr_bytes[0] & 1)) {
 217                                 msg = "mlx4 does not support the explicit"
 218                                         " exclusion of all multicast traffic";
 219                                 goto error;
 220                         }
 221                         flow->allmulti = 1;
 222                 } else if (sum_dst != (UINT8_C(0xff) * ETHER_ADDR_LEN)) {
 223                         msg = "mlx4 does not support matching partial"
 224                                 " Ethernet fields";
 225                         goto error;
 226                 }
 227         }
 228         if (!flow->ibv_attr)
 229                 return 0;
 230         if (flow->promisc) {
 231                 flow->ibv_attr->type = IBV_FLOW_ATTR_ALL_DEFAULT;
 232                 return 0;
 233         }
 234         if (flow->allmulti) {
 235                 flow->ibv_attr->type = IBV_FLOW_ATTR_MC_DEFAULT;
 236                 return 0;
 237         }
 238         ++flow->ibv_attr->num_of_specs;
 239         eth = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 240         *eth = (struct ibv_flow_spec_eth) {
 241                 .type = IBV_FLOW_SPEC_ETH,
 242                 .size = sizeof(*eth),
 243         };
 244         memcpy(eth->val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
 245         memcpy(eth->mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
 246         /* Remove unwanted bits from values. */
 247         for (i = 0; i < ETHER_ADDR_LEN; ++i) {
 248                 eth->val.dst_mac[i] &= eth->mask.dst_mac[i];
 249         }
 250         return 0;
 251 error:
 252         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 253                                   item, msg);
 254 }
 255
 256 /**
 257  * Merge VLAN pattern item into flow rule handle.
 258  *
 259  * Additional mlx4-specific constraints on supported fields:
 260  *
 261  * - Matching *all* VLAN traffic by omitting @p item->spec or providing an
 262  *   empty @p item->mask would also include non-VLAN traffic. Doing so is
 263  *   therefore unsupported.
 264  * - No support for partial masks.
 265  *
 266  * @param[in, out] flow
 267  *   Flow rule handle to update.
 268  * @param[in] item
 269  *   Pattern item to merge.
 270  * @param[in] proc
 271  *   Associated item-processing object.
 272  * @param[out] error
 273  *   Perform verbose error reporting if not NULL.
 274  *
 275  * @return
 276  *   0 on success, a negative errno value otherwise and rte_errno is set.
 277  */
 278 static int
 279 mlx4_flow_merge_vlan(struct rte_flow *flow,
 280                      const struct rte_flow_item *item,
 281                      const struct mlx4_flow_proc_item *proc,
 282                      struct rte_flow_error *error)
 283 {
 284         const struct rte_flow_item_vlan *spec = item->spec;
 285         const struct rte_flow_item_vlan *mask =
 286                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 287         struct ibv_flow_spec_eth *eth;
 288         const char *msg;
 289
 290         if (!mask || !mask->tci) {
 291                 msg = "mlx4 cannot match all VLAN traffic while excluding"
 292                         " non-VLAN traffic, TCI VID must be specified";
 293                 goto error;
 294         }
 295         if (mask->tci != RTE_BE16(0x0fff)) {
 296                 msg = "mlx4 does not support partial TCI VID matching";
 297                 goto error;
 298         }
 299         if (!flow->ibv_attr)
 300                 return 0;
 301         eth = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size -
 302                        sizeof(*eth));
 303         eth->val.vlan_tag = spec->tci;
 304         eth->mask.vlan_tag = mask->tci;
 305         eth->val.vlan_tag &= eth->mask.vlan_tag;
 306         return 0;
 307 error:
 308         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 309                                   item, msg);
 310 }
 311
 312 /**
 313  * Merge IPv4 pattern item into flow rule handle.
 314  *
 315  * Additional mlx4-specific constraints on supported fields:
 316  *
 317  * - No support for partial masks.
 318  *
 319  * @param[in, out] flow
 320  *   Flow rule handle to update.
 321  * @param[in] item
 322  *   Pattern item to merge.
 323  * @param[in] proc
 324  *   Associated item-processing object.
 325  * @param[out] error
 326  *   Perform verbose error reporting if not NULL.
 327  *
 328  * @return
 329  *   0 on success, a negative errno value otherwise and rte_errno is set.
 330  */
 331 static int
 332 mlx4_flow_merge_ipv4(struct rte_flow *flow,
 333                      const struct rte_flow_item *item,
 334                      const struct mlx4_flow_proc_item *proc,
 335                      struct rte_flow_error *error)
 336 {
 337         const struct rte_flow_item_ipv4 *spec = item->spec;
 338         const struct rte_flow_item_ipv4 *mask =
 339                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 340         struct ibv_flow_spec_ipv4 *ipv4;
 341         const char *msg;
 342
 343         if (mask &&
 344             ((uint32_t)(mask->hdr.src_addr + 1) > UINT32_C(1) ||
 345              (uint32_t)(mask->hdr.dst_addr + 1) > UINT32_C(1))) {
 346                 msg = "mlx4 does not support matching partial IPv4 fields";
 347                 goto error;
 348         }
 349         if (!flow->ibv_attr)
 350                 return 0;
 351         ++flow->ibv_attr->num_of_specs;
 352         ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 353         *ipv4 = (struct ibv_flow_spec_ipv4) {
 354                 .type = IBV_FLOW_SPEC_IPV4,
 355                 .size = sizeof(*ipv4),
 356         };
 357         if (!spec)
 358                 return 0;
 359         ipv4->val = (struct ibv_flow_ipv4_filter) {
 360                 .src_ip = spec->hdr.src_addr,
 361                 .dst_ip = spec->hdr.dst_addr,
 362         };
 363         ipv4->mask = (struct ibv_flow_ipv4_filter) {
 364                 .src_ip = mask->hdr.src_addr,
 365                 .dst_ip = mask->hdr.dst_addr,
 366         };
 367         /* Remove unwanted bits from values. */
 368         ipv4->val.src_ip &= ipv4->mask.src_ip;
 369         ipv4->val.dst_ip &= ipv4->mask.dst_ip;
 370         return 0;
 371 error:
 372         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 373                                   item, msg);
 374 }
 375
 376 /**
 377  * Merge UDP pattern item into flow rule handle.
 378  *
 379  * Additional mlx4-specific constraints on supported fields:
 380  *
 381  * - No support for partial masks.
 382  *
 383  * @param[in, out] flow
 384  *   Flow rule handle to update.
 385  * @param[in] item
 386  *   Pattern item to merge.
 387  * @param[in] proc
 388  *   Associated item-processing object.
 389  * @param[out] error
 390  *   Perform verbose error reporting if not NULL.
 391  *
 392  * @return
 393  *   0 on success, a negative errno value otherwise and rte_errno is set.
 394  */
 395 static int
 396 mlx4_flow_merge_udp(struct rte_flow *flow,
 397                     const struct rte_flow_item *item,
 398                     const struct mlx4_flow_proc_item *proc,
 399                     struct rte_flow_error *error)
 400 {
 401         const struct rte_flow_item_udp *spec = item->spec;
 402         const struct rte_flow_item_udp *mask =
 403                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 404         struct ibv_flow_spec_tcp_udp *udp;
 405         const char *msg;
 406
 407         if (mask &&
 408             ((uint16_t)(mask->hdr.src_port + 1) > UINT16_C(1) ||
 409              (uint16_t)(mask->hdr.dst_port + 1) > UINT16_C(1))) {
 410                 msg = "mlx4 does not support matching partial UDP fields";
 411                 goto error;
 412         }
 413         if (!flow->ibv_attr)
 414                 return 0;
 415         ++flow->ibv_attr->num_of_specs;
 416         udp = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 417         *udp = (struct ibv_flow_spec_tcp_udp) {
 418                 .type = IBV_FLOW_SPEC_UDP,
 419                 .size = sizeof(*udp),
 420         };
 421         if (!spec)
 422                 return 0;
 423         udp->val.dst_port = spec->hdr.dst_port;
 424         udp->val.src_port = spec->hdr.src_port;
 425         udp->mask.dst_port = mask->hdr.dst_port;
 426         udp->mask.src_port = mask->hdr.src_port;
 427         /* Remove unwanted bits from values. */
 428         udp->val.src_port &= udp->mask.src_port;
 429         udp->val.dst_port &= udp->mask.dst_port;
 430         return 0;
 431 error:
 432         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 433                                   item, msg);
 434 }
 435
 436 /**
 437  * Merge TCP pattern item into flow rule handle.
 438  *
 439  * Additional mlx4-specific constraints on supported fields:
 440  *
 441  * - No support for partial masks.
 442  *
 443  * @param[in, out] flow
 444  *   Flow rule handle to update.
 445  * @param[in] item
 446  *   Pattern item to merge.
 447  * @param[in] proc
 448  *   Associated item-processing object.
 449  * @param[out] error
 450  *   Perform verbose error reporting if not NULL.
 451  *
 452  * @return
 453  *   0 on success, a negative errno value otherwise and rte_errno is set.
 454  */
 455 static int
 456 mlx4_flow_merge_tcp(struct rte_flow *flow,
 457                     const struct rte_flow_item *item,
 458                     const struct mlx4_flow_proc_item *proc,
 459                     struct rte_flow_error *error)
 460 {
 461         const struct rte_flow_item_tcp *spec = item->spec;
 462         const struct rte_flow_item_tcp *mask =
 463                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 464         struct ibv_flow_spec_tcp_udp *tcp;
 465         const char *msg;
 466
 467         if (mask &&
 468             ((uint16_t)(mask->hdr.src_port + 1) > UINT16_C(1) ||
 469              (uint16_t)(mask->hdr.dst_port + 1) > UINT16_C(1))) {
 470                 msg = "mlx4 does not support matching partial TCP fields";
 471                 goto error;
 472         }
 473         if (!flow->ibv_attr)
 474                 return 0;
 475         ++flow->ibv_attr->num_of_specs;
 476         tcp = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 477         *tcp = (struct ibv_flow_spec_tcp_udp) {
 478                 .type = IBV_FLOW_SPEC_TCP,
 479                 .size = sizeof(*tcp),
 480         };
 481         if (!spec)
 482                 return 0;
 483         tcp->val.dst_port = spec->hdr.dst_port;
 484         tcp->val.src_port = spec->hdr.src_port;
 485         tcp->mask.dst_port = mask->hdr.dst_port;
 486         tcp->mask.src_port = mask->hdr.src_port;
 487         /* Remove unwanted bits from values. */
 488         tcp->val.src_port &= tcp->mask.src_port;
 489         tcp->val.dst_port &= tcp->mask.dst_port;
 490         return 0;
 491 error:
 492         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 493                                   item, msg);
 494 }
 495
 496 /**
 497  * Perform basic sanity checks on a pattern item.
 498  *
 499  * @param[in] item
 500  *   Item specification.
 501  * @param[in] proc
 502  *   Associated item-processing object.
 503  * @param[out] error
 504  *   Perform verbose error reporting if not NULL.
 505  *
 506  * @return
 507  *   0 on success, a negative errno value otherwise and rte_errno is set.
 508  */
 509 static int
 510 mlx4_flow_item_check(const struct rte_flow_item *item,
 511                      const struct mlx4_flow_proc_item *proc,
 512                      struct rte_flow_error *error)
 513 {
 514         const uint8_t *mask;
 515         unsigned int i;
 516
 517         /* item->last and item->mask cannot exist without item->spec. */
 518         if (!item->spec && (item->mask || item->last))
 519                 return rte_flow_error_set
 520                         (error, EINVAL, RTE_FLOW_ERROR_TYPE_ITEM, item,
 521                          "\"mask\" or \"last\" field provided without a"
 522                          " corresponding \"spec\"");
 523         /* No spec, no mask, no problem. */
 524         if (!item->spec)
 525                 return 0;
 526         mask = item->mask ?
 527                 (const uint8_t *)item->mask :
 528                 (const uint8_t *)proc->mask_default;
 529         assert(mask);
 530         /*
 531          * Single-pass check to make sure that:
 532          * - Mask is supported, no bits are set outside proc->mask_support.
 533          * - Both item->spec and item->last are included in mask.
 534          */
 535         for (i = 0; i != proc->mask_sz; ++i) {
 536                 if (!mask[i])
 537                         continue;
 538                 if ((mask[i] | ((const uint8_t *)proc->mask_support)[i]) !=
 539                     ((const uint8_t *)proc->mask_support)[i])
 540                         return rte_flow_error_set
 541                                 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 542                                  item, "unsupported field found in \"mask\"");
 543                 if (item->last &&
 544                     (((const uint8_t *)item->spec)[i] & mask[i]) !=
 545                     (((const uint8_t *)item->last)[i] & mask[i]))
 546                         return rte_flow_error_set
 547                                 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 548                                  item,
 549                                  "range between \"spec\" and \"last\""
 550                                  " is larger than \"mask\"");
 551         }
 552         return 0;
 553 }
 554
 555 /** Graph of supported items and associated actions. */
 556 static const struct mlx4_flow_proc_item mlx4_flow_proc_item_list[] = {
 557         [RTE_FLOW_ITEM_TYPE_END] = {
 558                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_ETH),
 559         },
 560         [RTE_FLOW_ITEM_TYPE_ETH] = {
 561                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_VLAN,
 562                                        RTE_FLOW_ITEM_TYPE_IPV4),
 563                 .mask_support = &(const struct rte_flow_item_eth){
 564                         /* Only destination MAC can be matched. */
 565                         .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 566                 },
 567                 .mask_default = &rte_flow_item_eth_mask,
 568                 .mask_sz = sizeof(struct rte_flow_item_eth),
 569                 .merge = mlx4_flow_merge_eth,
 570                 .dst_sz = sizeof(struct ibv_flow_spec_eth),
 571         },
 572         [RTE_FLOW_ITEM_TYPE_VLAN] = {
 573                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_IPV4),
 574                 .mask_support = &(const struct rte_flow_item_vlan){
 575                         /* Only TCI VID matching is supported. */
 576                         .tci = RTE_BE16(0x0fff),
 577                 },
 578                 .mask_default = &rte_flow_item_vlan_mask,
 579                 .mask_sz = sizeof(struct rte_flow_item_vlan),
 580                 .merge = mlx4_flow_merge_vlan,
 581                 .dst_sz = 0,
 582         },
 583         [RTE_FLOW_ITEM_TYPE_IPV4] = {
 584                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_UDP,
 585                                        RTE_FLOW_ITEM_TYPE_TCP),
 586                 .mask_support = &(const struct rte_flow_item_ipv4){
 587                         .hdr = {
 588                                 .src_addr = RTE_BE32(0xffffffff),
 589                                 .dst_addr = RTE_BE32(0xffffffff),
 590                         },
 591                 },
 592                 .mask_default = &rte_flow_item_ipv4_mask,
 593                 .mask_sz = sizeof(struct rte_flow_item_ipv4),
 594                 .merge = mlx4_flow_merge_ipv4,
 595                 .dst_sz = sizeof(struct ibv_flow_spec_ipv4),
 596         },
 597         [RTE_FLOW_ITEM_TYPE_UDP] = {
 598                 .mask_support = &(const struct rte_flow_item_udp){
 599                         .hdr = {
 600                                 .src_port = RTE_BE16(0xffff),
 601                                 .dst_port = RTE_BE16(0xffff),
 602                         },
 603                 },
 604                 .mask_default = &rte_flow_item_udp_mask,
 605                 .mask_sz = sizeof(struct rte_flow_item_udp),
 606                 .merge = mlx4_flow_merge_udp,
 607                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 608         },
 609         [RTE_FLOW_ITEM_TYPE_TCP] = {
 610                 .mask_support = &(const struct rte_flow_item_tcp){
 611                         .hdr = {
 612                                 .src_port = RTE_BE16(0xffff),
 613                                 .dst_port = RTE_BE16(0xffff),
 614                         },
 615                 },
 616                 .mask_default = &rte_flow_item_tcp_mask,
 617                 .mask_sz = sizeof(struct rte_flow_item_tcp),
 618                 .merge = mlx4_flow_merge_tcp,
 619                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 620         },
 621 };
 622
 623 /**
 624  * Make sure a flow rule is supported and initialize associated structure.
 625  *
 626  * @param priv
 627  *   Pointer to private structure.
 628  * @param[in] attr
 629  *   Flow rule attributes.
 630  * @param[in] pattern
 631  *   Pattern specification (list terminated by the END pattern item).
 632  * @param[in] actions
 633  *   Associated actions (list terminated by the END action).
 634  * @param[out] error
 635  *   Perform verbose error reporting if not NULL.
 636  * @param[in, out] addr
 637  *   Buffer where the resulting flow rule handle pointer must be stored.
 638  *   If NULL, stop processing after validation stage.
 639  *
 640  * @return
 641  *   0 on success, a negative errno value otherwise and rte_errno is set.
 642  */
 643 static int
 644 mlx4_flow_prepare(struct priv *priv,
 645                   const struct rte_flow_attr *attr,
 646                   const struct rte_flow_item pattern[],
 647                   const struct rte_flow_action actions[],
 648                   struct rte_flow_error *error,
 649                   struct rte_flow **addr)
 650 {
 651         const struct rte_flow_item *item;
 652         const struct rte_flow_action *action;
 653         const struct mlx4_flow_proc_item *proc;
 654         struct rte_flow temp = { .ibv_attr_size = sizeof(*temp.ibv_attr) };
 655         struct rte_flow *flow = &temp;
 656         const char *msg = NULL;
 657
 658         if (attr->group)
 659                 return rte_flow_error_set
 660                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 661                          NULL, "groups are not supported");
 662         if (attr->priority > MLX4_FLOW_PRIORITY_LAST)
 663                 return rte_flow_error_set
 664                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 665                          NULL, "maximum priority level is "
 666                          MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST));
 667         if (attr->egress)
 668                 return rte_flow_error_set
 669                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 670                          NULL, "egress is not supported");
 671         if (!attr->ingress)
 672                 return rte_flow_error_set
 673                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 674                          NULL, "only ingress is supported");
 675 fill:
 676         proc = mlx4_flow_proc_item_list;
 677         /* Go over pattern. */
 678         for (item = pattern; item->type; ++item) {
 679                 const struct mlx4_flow_proc_item *next = NULL;
 680                 unsigned int i;
 681                 int err;
 682
 683                 if (item->type == RTE_FLOW_ITEM_TYPE_VOID)
 684                         continue;
 685                 if (item->type == MLX4_FLOW_ITEM_TYPE_INTERNAL) {
 686                         flow->internal = 1;
 687                         continue;
 688                 }
 689                 if (flow->promisc || flow->allmulti) {
 690                         msg = "mlx4 does not support additional matching"
 691                                 " criteria combined with indiscriminate"
 692                                 " matching on Ethernet headers";
 693                         goto exit_item_not_supported;
 694                 }
 695                 for (i = 0; proc->next_item && proc->next_item[i]; ++i) {
 696                         if (proc->next_item[i] == item->type) {
 697                                 next = &mlx4_flow_proc_item_list[item->type];
 698                                 break;
 699                         }
 700                 }
 701                 if (!next)
 702                         goto exit_item_not_supported;
 703                 proc = next;
 704                 /*
 705                  * Perform basic sanity checks only once, while handle is
 706                  * not allocated.
 707                  */
 708                 if (flow == &temp) {
 709                         err = mlx4_flow_item_check(item, proc, error);
 710                         if (err)
 711                                 return err;
 712                 }
 713                 if (proc->merge) {
 714                         err = proc->merge(flow, item, proc, error);
 715                         if (err)
 716                                 return err;
 717                 }
 718                 flow->ibv_attr_size += proc->dst_sz;
 719         }
 720         /* Go over actions list. */
 721         for (action = actions; action->type; ++action) {
 722                 switch (action->type) {
 723                         const struct rte_flow_action_queue *queue;
 724                         const struct rte_flow_action_rss *rss;
 725                         const struct rte_eth_rss_conf *rss_conf;
 726                         unsigned int i;
 727
 728                 case RTE_FLOW_ACTION_TYPE_VOID:
 729                         continue;
 730                 case RTE_FLOW_ACTION_TYPE_DROP:
 731                         flow->drop = 1;
 732                         break;
 733                 case RTE_FLOW_ACTION_TYPE_QUEUE:
 734                         if (flow->rss)
 735                                 break;
 736                         queue = action->conf;
 737                         flow->rss = mlx4_rss_get
 738                                 (priv, 0, mlx4_rss_hash_key_default, 1,
 739                                  &queue->index);
 740                         if (!flow->rss) {
 741                                 msg = "not enough resources for additional"
 742                                         " single-queue RSS context";
 743                                 goto exit_action_not_supported;
 744                         }
 745                         break;
 746                 case RTE_FLOW_ACTION_TYPE_RSS:
 747                         if (flow->rss)
 748                                 break;
 749                         rss = action->conf;
 750                         /* Default RSS configuration if none is provided. */
 751                         rss_conf =
 752                                 rss->rss_conf ?
 753                                 rss->rss_conf :
 754                                 &(struct rte_eth_rss_conf){
 755                                         .rss_key = mlx4_rss_hash_key_default,
 756                                         .rss_key_len = MLX4_RSS_HASH_KEY_SIZE,
 757                                         .rss_hf = (ETH_RSS_IPV4 |
 758                                                    ETH_RSS_NONFRAG_IPV4_TCP |
 759                                                    ETH_RSS_IPV6 |
 760                                                    ETH_RSS_NONFRAG_IPV6_TCP),
 761                                 };
 762                         /* Sanity checks. */
 763                         if (!rte_is_power_of_2(rss->num)) {
 764                                 msg = "for RSS, mlx4 requires the number of"
 765                                         " queues to be a power of two";
 766                                 goto exit_action_not_supported;
 767                         }
 768                         if (rss_conf->rss_key_len !=
 769                             sizeof(flow->rss->key)) {
 770                                 msg = "mlx4 supports exactly one RSS hash key"
 771                                         " length: "
 772                                         MLX4_STR_EXPAND(MLX4_RSS_HASH_KEY_SIZE);
 773                                 goto exit_action_not_supported;
 774                         }
 775                         for (i = 1; i < rss->num; ++i)
 776                                 if (rss->queue[i] - rss->queue[i - 1] != 1)
 777                                         break;
 778                         if (i != rss->num) {
 779                                 msg = "mlx4 requires RSS contexts to use"
 780                                         " consecutive queue indices only";
 781                                 goto exit_action_not_supported;
 782                         }
 783                         if (rss->queue[0] % rss->num) {
 784                                 msg = "mlx4 requires the first queue of a RSS"
 785                                         " context to be aligned on a multiple"
 786                                         " of the context size";
 787                                 goto exit_action_not_supported;
 788                         }
 789                         flow->rss = mlx4_rss_get
 790                                 (priv, mlx4_conv_rss_hf(rss_conf->rss_hf),
 791                                  rss_conf->rss_key, rss->num, rss->queue);
 792                         if (!flow->rss) {
 793                                 msg = "either invalid parameters or not enough"
 794                                         " resources for additional multi-queue"
 795                                         " RSS context";
 796                                 goto exit_action_not_supported;
 797                         }
 798                         break;
 799                 default:
 800                         goto exit_action_not_supported;
 801                 }
 802         }
 803         if (!flow->rss && !flow->drop)
 804                 return rte_flow_error_set
 805                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 806                          NULL, "no valid action");
 807         /* Validation ends here. */
 808         if (!addr) {
 809                 if (flow->rss)
 810                         mlx4_rss_put(flow->rss);
 811                 return 0;
 812         }
 813         if (flow == &temp) {
 814                 /* Allocate proper handle based on collected data. */
 815                 const struct mlx4_malloc_vec vec[] = {
 816                         {
 817                                 .align = alignof(struct rte_flow),
 818                                 .size = sizeof(*flow),
 819                                 .addr = (void **)&flow,
 820                         },
 821                         {
 822                                 .align = alignof(struct ibv_flow_attr),
 823                                 .size = temp.ibv_attr_size,
 824                                 .addr = (void **)&temp.ibv_attr,
 825                         },
 826                 };
 827
 828                 if (!mlx4_zmallocv(__func__, vec, RTE_DIM(vec)))
 829                         return rte_flow_error_set
 830                                 (error, -rte_errno,
 831                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
 832                                  "flow rule handle allocation failure");
 833                 /* Most fields will be updated by second pass. */
 834                 *flow = (struct rte_flow){
 835                         .ibv_attr = temp.ibv_attr,
 836                         .ibv_attr_size = sizeof(*flow->ibv_attr),
 837                         .rss = temp.rss,
 838                 };
 839                 *flow->ibv_attr = (struct ibv_flow_attr){
 840                         .type = IBV_FLOW_ATTR_NORMAL,
 841                         .size = sizeof(*flow->ibv_attr),
 842                         .priority = attr->priority,
 843                         .port = priv->port,
 844                 };
 845                 goto fill;
 846         }
 847         *addr = flow;
 848         return 0;
 849 exit_item_not_supported:
 850         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 851                                   item, msg ? msg : "item not supported");
 852 exit_action_not_supported:
 853         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 854                                   action, msg ? msg : "action not supported");
 855 }
 856
 857 /**
 858  * Validate a flow supported by the NIC.
 859  *
 860  * @see rte_flow_validate()
 861  * @see rte_flow_ops
 862  */
 863 static int
 864 mlx4_flow_validate(struct rte_eth_dev *dev,
 865                    const struct rte_flow_attr *attr,
 866                    const struct rte_flow_item pattern[],
 867                    const struct rte_flow_action actions[],
 868                    struct rte_flow_error *error)
 869 {
 870         struct priv *priv = dev->data->dev_private;
 871
 872         return mlx4_flow_prepare(priv, attr, pattern, actions, error, NULL);
 873 }
 874
 875 /**
 876  * Get a drop flow rule resources instance.
 877  *
 878  * @param priv
 879  *   Pointer to private structure.
 880  *
 881  * @return
 882  *   Pointer to drop flow resources on success, NULL otherwise and rte_errno
 883  *   is set.
 884  */
 885 static struct mlx4_drop *
 886 mlx4_drop_get(struct priv *priv)
 887 {
 888         struct mlx4_drop *drop = priv->drop;
 889
 890         if (drop) {
 891                 assert(drop->refcnt);
 892                 assert(drop->priv == priv);
 893                 ++drop->refcnt;
 894                 return drop;
 895         }
 896         drop = rte_malloc(__func__, sizeof(*drop), 0);
 897         if (!drop)
 898                 goto error;
 899         *drop = (struct mlx4_drop){
 900                 .priv = priv,
 901                 .refcnt = 1,
 902         };
 903         drop->cq = ibv_create_cq(priv->ctx, 1, NULL, NULL, 0);
 904         if (!drop->cq)
 905                 goto error;
 906         drop->qp = ibv_create_qp(priv->pd,
 907                                  &(struct ibv_qp_init_attr){
 908                                         .send_cq = drop->cq,
 909                                         .recv_cq = drop->cq,
 910                                         .qp_type = IBV_QPT_RAW_PACKET,
 911                                  });
 912         if (!drop->qp)
 913                 goto error;
 914         priv->drop = drop;
 915         return drop;
 916 error:
 917         if (drop->qp)
 918                 claim_zero(ibv_destroy_qp(drop->qp));
 919         if (drop->cq)
 920                 claim_zero(ibv_destroy_cq(drop->cq));
 921         if (drop)
 922                 rte_free(drop);
 923         rte_errno = ENOMEM;
 924         return NULL;
 925 }
 926
 927 /**
 928  * Give back a drop flow rule resources instance.
 929  *
 930  * @param drop
 931  *   Pointer to drop flow rule resources.
 932  */
 933 static void
 934 mlx4_drop_put(struct mlx4_drop *drop)
 935 {
 936         assert(drop->refcnt);
 937         if (--drop->refcnt)
 938                 return;
 939         drop->priv->drop = NULL;
 940         claim_zero(ibv_destroy_qp(drop->qp));
 941         claim_zero(ibv_destroy_cq(drop->cq));
 942         rte_free(drop);
 943 }
 944
 945 /**
 946  * Toggle a configured flow rule.
 947  *
 948  * @param priv
 949  *   Pointer to private structure.
 950  * @param flow
 951  *   Flow rule handle to toggle.
 952  * @param enable
 953  *   Whether associated Verbs flow must be created or removed.
 954  * @param[out] error
 955  *   Perform verbose error reporting if not NULL.
 956  *
 957  * @return
 958  *   0 on success, a negative errno value otherwise and rte_errno is set.
 959  */
 960 static int
 961 mlx4_flow_toggle(struct priv *priv,
 962                  struct rte_flow *flow,
 963                  int enable,
 964                  struct rte_flow_error *error)
 965 {
 966         struct ibv_qp *qp = NULL;
 967         const char *msg;
 968         int err;
 969
 970         if (!enable) {
 971                 if (!flow->ibv_flow)
 972                         return 0;
 973                 claim_zero(ibv_destroy_flow(flow->ibv_flow));
 974                 flow->ibv_flow = NULL;
 975                 if (flow->drop)
 976                         mlx4_drop_put(priv->drop);
 977                 else if (flow->rss)
 978                         mlx4_rss_detach(flow->rss);
 979                 return 0;
 980         }
 981         assert(flow->ibv_attr);
 982         if (!flow->internal &&
 983             !priv->isolated &&
 984             flow->ibv_attr->priority == MLX4_FLOW_PRIORITY_LAST) {
 985                 if (flow->ibv_flow) {
 986                         claim_zero(ibv_destroy_flow(flow->ibv_flow));
 987                         flow->ibv_flow = NULL;
 988                         if (flow->drop)
 989                                 mlx4_drop_put(priv->drop);
 990                         else if (flow->rss)
 991                                 mlx4_rss_detach(flow->rss);
 992                 }
 993                 err = EACCES;
 994                 msg = ("priority level "
 995                        MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST)
 996                        " is reserved when not in isolated mode");
 997                 goto error;
 998         }
 999         if (flow->rss) {
1000                 struct mlx4_rss *rss = flow->rss;
1001                 int missing = 0;
1002                 unsigned int i;
1003
1004                 /* Stop at the first nonexistent target queue. */
1005                 for (i = 0; i != rss->queues; ++i)
1006                         if (rss->queue_id[i] >=
1007                             priv->dev->data->nb_rx_queues ||
1008                             !priv->dev->data->rx_queues[rss->queue_id[i]]) {
1009                                 missing = 1;
1010                                 break;
1011                         }
1012                 if (flow->ibv_flow) {
1013                         if (missing ^ !flow->drop)
1014                                 return 0;
1015                         /* Verbs flow needs updating. */
1016                         claim_zero(ibv_destroy_flow(flow->ibv_flow));
1017                         flow->ibv_flow = NULL;
1018                         if (flow->drop)
1019                                 mlx4_drop_put(priv->drop);
1020                         else
1021                                 mlx4_rss_detach(rss);
1022                 }
1023                 if (!missing) {
1024                         err = mlx4_rss_attach(rss);
1025                         if (err) {
1026                                 err = -err;
1027                                 msg = "cannot create indirection table or hash"
1028                                         " QP to associate flow rule with";
1029                                 goto error;
1030                         }
1031                         qp = rss->qp;
1032                 }
1033                 /* A missing target queue drops traffic implicitly. */
1034                 flow->drop = missing;
1035         }
1036         if (flow->drop) {
1037                 mlx4_drop_get(priv);
1038                 if (!priv->drop) {
1039                         err = rte_errno;
1040                         msg = "resources for drop flow rule cannot be created";
1041                         goto error;
1042                 }
1043                 qp = priv->drop->qp;
1044         }
1045         assert(qp);
1046         if (flow->ibv_flow)
1047                 return 0;
1048         flow->ibv_flow = ibv_create_flow(qp, flow->ibv_attr);
1049         if (flow->ibv_flow)
1050                 return 0;
1051         if (flow->drop)
1052                 mlx4_drop_put(priv->drop);
1053         else if (flow->rss)
1054                 mlx4_rss_detach(flow->rss);
1055         err = errno;
1056         msg = "flow rule rejected by device";
1057 error:
1058         return rte_flow_error_set
1059                 (error, err, RTE_FLOW_ERROR_TYPE_HANDLE, flow, msg);
1060 }
1061
1062 /**
1063  * Create a flow.
1064  *
1065  * @see rte_flow_create()
1066  * @see rte_flow_ops
1067  */
1068 static struct rte_flow *
1069 mlx4_flow_create(struct rte_eth_dev *dev,
1070                  const struct rte_flow_attr *attr,
1071                  const struct rte_flow_item pattern[],
1072                  const struct rte_flow_action actions[],
1073                  struct rte_flow_error *error)
1074 {
1075         struct priv *priv = dev->data->dev_private;
1076         struct rte_flow *flow;
1077         int err;
1078
1079         err = mlx4_flow_prepare(priv, attr, pattern, actions, error, &flow);
1080         if (err)
1081                 return NULL;
1082         err = mlx4_flow_toggle(priv, flow, priv->started, error);
1083         if (!err) {
1084                 struct rte_flow *curr = LIST_FIRST(&priv->flows);
1085
1086                 /* New rules are inserted after internal ones. */
1087                 if (!curr || !curr->internal) {
1088                         LIST_INSERT_HEAD(&priv->flows, flow, next);
1089                 } else {
1090                         while (LIST_NEXT(curr, next) &&
1091                                LIST_NEXT(curr, next)->internal)
1092                                 curr = LIST_NEXT(curr, next);
1093                         LIST_INSERT_AFTER(curr, flow, next);
1094                 }
1095                 return flow;
1096         }
1097         if (flow->rss)
1098                 mlx4_rss_put(flow->rss);
1099         rte_flow_error_set(error, -err, RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1100                            error->message);
1101         rte_free(flow);
1102         return NULL;
1103 }
1104
1105 /**
1106  * Configure isolated mode.
1107  *
1108  * @see rte_flow_isolate()
1109  * @see rte_flow_ops
1110  */
1111 static int
1112 mlx4_flow_isolate(struct rte_eth_dev *dev,
1113                   int enable,
1114                   struct rte_flow_error *error)
1115 {
1116         struct priv *priv = dev->data->dev_private;
1117
1118         if (!!enable == !!priv->isolated)
1119                 return 0;
1120         priv->isolated = !!enable;
1121         if (mlx4_flow_sync(priv, error)) {
1122                 priv->isolated = !enable;
1123                 return -rte_errno;
1124         }
1125         return 0;
1126 }
1127
1128 /**
1129  * Destroy a flow rule.
1130  *
1131  * @see rte_flow_destroy()
1132  * @see rte_flow_ops
1133  */
1134 static int
1135 mlx4_flow_destroy(struct rte_eth_dev *dev,
1136                   struct rte_flow *flow,
1137                   struct rte_flow_error *error)
1138 {
1139         struct priv *priv = dev->data->dev_private;
1140         int err = mlx4_flow_toggle(priv, flow, 0, error);
1141
1142         if (err)
1143                 return err;
1144         LIST_REMOVE(flow, next);
1145         if (flow->rss)
1146                 mlx4_rss_put(flow->rss);
1147         rte_free(flow);
1148         return 0;
1149 }
1150
1151 /**
1152  * Destroy user-configured flow rules.
1153  *
1154  * This function skips internal flows rules.
1155  *
1156  * @see rte_flow_flush()
1157  * @see rte_flow_ops
1158  */
1159 static int
1160 mlx4_flow_flush(struct rte_eth_dev *dev,
1161                 struct rte_flow_error *error)
1162 {
1163         struct priv *priv = dev->data->dev_private;
1164         struct rte_flow *flow = LIST_FIRST(&priv->flows);
1165
1166         while (flow) {
1167                 struct rte_flow *next = LIST_NEXT(flow, next);
1168
1169                 if (!flow->internal)
1170                         mlx4_flow_destroy(dev, flow, error);
1171                 flow = next;
1172         }
1173         return 0;
1174 }
1175
1176 /**
1177  * Helper function to determine the next configured VLAN filter.
1178  *
1179  * @param priv
1180  *   Pointer to private structure.
1181  * @param vlan
1182  *   VLAN ID to use as a starting point.
1183  *
1184  * @return
1185  *   Next configured VLAN ID or a high value (>= 4096) if there is none.
1186  */
1187 static uint16_t
1188 mlx4_flow_internal_next_vlan(struct priv *priv, uint16_t vlan)
1189 {
1190         while (vlan < 4096) {
1191                 if (priv->dev->data->vlan_filter_conf.ids[vlan / 64] &
1192                     (UINT64_C(1) << (vlan % 64)))
1193                         return vlan;
1194                 ++vlan;
1195         }
1196         return vlan;
1197 }
1198
1199 /**
1200  * Generate internal flow rules.
1201  *
1202  * Various flow rules are created depending on the mode the device is in:
1203  *
1204  * 1. Promiscuous: port MAC + catch-all (VLAN filtering is ignored).
1205  * 2. All multicast: port MAC/VLAN + catch-all multicast.
1206  * 3. Otherwise: port MAC/VLAN + broadcast MAC/VLAN.
1207  *
1208  * About MAC flow rules:
1209  *
1210  * - MAC flow rules are generated from @p dev->data->mac_addrs
1211  *   (@p priv->mac array).
1212  * - An additional flow rule for Ethernet broadcasts is also generated.
1213  * - All these are per-VLAN if @p dev->data->dev_conf.rxmode.hw_vlan_filter
1214  *   is enabled and VLAN filters are configured.
1215  *
1216  * @param priv
1217  *   Pointer to private structure.
1218  * @param[out] error
1219  *   Perform verbose error reporting if not NULL.
1220  *
1221  * @return
1222  *   0 on success, a negative errno value otherwise and rte_errno is set.
1223  */
1224 static int
1225 mlx4_flow_internal(struct priv *priv, struct rte_flow_error *error)
1226 {
1227         struct rte_flow_attr attr = {
1228                 .priority = MLX4_FLOW_PRIORITY_LAST,
1229                 .ingress = 1,
1230         };
1231         struct rte_flow_item_eth eth_spec;
1232         const struct rte_flow_item_eth eth_mask = {
1233                 .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
1234         };
1235         const struct rte_flow_item_eth eth_allmulti = {
1236                 .dst.addr_bytes = "\x01\x00\x00\x00\x00\x00",
1237         };
1238         struct rte_flow_item_vlan vlan_spec;
1239         const struct rte_flow_item_vlan vlan_mask = {
1240                 .tci = RTE_BE16(0x0fff),
1241         };
1242         struct rte_flow_item pattern[] = {
1243                 {
1244                         .type = MLX4_FLOW_ITEM_TYPE_INTERNAL,
1245                 },
1246                 {
1247                         .type = RTE_FLOW_ITEM_TYPE_ETH,
1248                         .spec = &eth_spec,
1249                         .mask = &eth_mask,
1250                 },
1251                 {
1252                         /* Replaced with VLAN if filtering is enabled. */
1253                         .type = RTE_FLOW_ITEM_TYPE_END,
1254                 },
1255                 {
1256                         .type = RTE_FLOW_ITEM_TYPE_END,
1257                 },
1258         };
1259         /*
1260          * Round number of queues down to their previous power of 2 to
1261          * comply with RSS context limitations. Extra queues silently do not
1262          * get RSS by default.
1263          */
1264         uint32_t queues =
1265                 rte_align32pow2(priv->dev->data->nb_rx_queues + 1) >> 1;
1266         alignas(struct rte_flow_action_rss) uint8_t rss_conf_data
1267                 [offsetof(struct rte_flow_action_rss, queue) +
1268                  sizeof(((struct rte_flow_action_rss *)0)->queue[0]) * queues];
1269         struct rte_flow_action_rss *rss_conf = (void *)rss_conf_data;
1270         struct rte_flow_action actions[] = {
1271                 {
1272                         .type = RTE_FLOW_ACTION_TYPE_RSS,
1273                         .conf = rss_conf,
1274                 },
1275                 {
1276                         .type = RTE_FLOW_ACTION_TYPE_END,
1277                 },
1278         };
1279         struct ether_addr *rule_mac = &eth_spec.dst;
1280         rte_be16_t *rule_vlan =
1281                 priv->dev->data->dev_conf.rxmode.hw_vlan_filter &&
1282                 !priv->dev->data->promiscuous ?
1283                 &vlan_spec.tci :
1284                 NULL;
1285         int broadcast =
1286                 !priv->dev->data->promiscuous &&
1287                 !priv->dev->data->all_multicast;
1288         uint16_t vlan = 0;
1289         struct rte_flow *flow;
1290         unsigned int i;
1291         int err = 0;
1292
1293         /* Nothing to be done if there are no Rx queues. */
1294         if (!queues)
1295                 goto error;
1296         /* Prepare default RSS configuration. */
1297         *rss_conf = (struct rte_flow_action_rss){
1298                 .rss_conf = NULL, /* Rely on default fallback settings. */
1299                 .num = queues,
1300         };
1301         for (i = 0; i != queues; ++i)
1302                 rss_conf->queue[i] = i;
1303         /*
1304          * Set up VLAN item if filtering is enabled and at least one VLAN
1305          * filter is configured.
1306          */
1307         if (rule_vlan) {
1308                 vlan = mlx4_flow_internal_next_vlan(priv, 0);
1309                 if (vlan < 4096) {
1310                         pattern[2] = (struct rte_flow_item){
1311                                 .type = RTE_FLOW_ITEM_TYPE_VLAN,
1312                                 .spec = &vlan_spec,
1313                                 .mask = &vlan_mask,
1314                         };
1315 next_vlan:
1316                         *rule_vlan = rte_cpu_to_be_16(vlan);
1317                 } else {
1318                         rule_vlan = NULL;
1319                 }
1320         }
1321         for (i = 0; i != RTE_DIM(priv->mac) + broadcast; ++i) {
1322                 const struct ether_addr *mac;
1323
1324                 /* Broadcasts are handled by an extra iteration. */
1325                 if (i < RTE_DIM(priv->mac))
1326                         mac = &priv->mac[i];
1327                 else
1328                         mac = &eth_mask.dst;
1329                 if (is_zero_ether_addr(mac))
1330                         continue;
1331                 /* Check if MAC flow rule is already present. */
1332                 for (flow = LIST_FIRST(&priv->flows);
1333                      flow && flow->internal;
1334                      flow = LIST_NEXT(flow, next)) {
1335                         const struct ibv_flow_spec_eth *eth =
1336                                 (const void *)((uintptr_t)flow->ibv_attr +
1337                                                sizeof(*flow->ibv_attr));
1338                         unsigned int j;
1339
1340                         if (!flow->mac)
1341                                 continue;
1342                         assert(flow->ibv_attr->type == IBV_FLOW_ATTR_NORMAL);
1343                         assert(flow->ibv_attr->num_of_specs == 1);
1344                         assert(eth->type == IBV_FLOW_SPEC_ETH);
1345                         assert(flow->rss);
1346                         if (rule_vlan &&
1347                             (eth->val.vlan_tag != *rule_vlan ||
1348                              eth->mask.vlan_tag != RTE_BE16(0x0fff)))
1349                                 continue;
1350                         if (!rule_vlan && eth->mask.vlan_tag)
1351                                 continue;
1352                         for (j = 0; j != sizeof(mac->addr_bytes); ++j)
1353                                 if (eth->val.dst_mac[j] != mac->addr_bytes[j] ||
1354                                     eth->mask.dst_mac[j] != UINT8_C(0xff) ||
1355                                     eth->val.src_mac[j] != UINT8_C(0x00) ||
1356                                     eth->mask.src_mac[j] != UINT8_C(0x00))
1357                                         break;
1358                         if (j != sizeof(mac->addr_bytes))
1359                                 continue;
1360                         if (flow->rss->queues != queues ||
1361                             memcmp(flow->rss->queue_id, rss_conf->queue,
1362                                    queues * sizeof(flow->rss->queue_id[0])))
1363                                 continue;
1364                         break;
1365                 }
1366                 if (!flow || !flow->internal) {
1367                         /* Not found, create a new flow rule. */
1368                         memcpy(rule_mac, mac, sizeof(*mac));
1369                         flow = mlx4_flow_create(priv->dev, &attr, pattern,
1370                                                 actions, error);
1371                         if (!flow) {
1372                                 err = -rte_errno;
1373                                 goto error;
1374                         }
1375                 }
1376                 flow->select = 1;
1377                 flow->mac = 1;
1378         }
1379         if (rule_vlan) {
1380                 vlan = mlx4_flow_internal_next_vlan(priv, vlan + 1);
1381                 if (vlan < 4096)
1382                         goto next_vlan;
1383         }
1384         /* Take care of promiscuous and all multicast flow rules. */
1385         if (!broadcast) {
1386                 for (flow = LIST_FIRST(&priv->flows);
1387                      flow && flow->internal;
1388                      flow = LIST_NEXT(flow, next)) {
1389                         if (priv->dev->data->promiscuous) {
1390                                 if (flow->promisc)
1391                                         break;
1392                         } else {
1393                                 assert(priv->dev->data->all_multicast);
1394                                 if (flow->allmulti)
1395                                         break;
1396                         }
1397                 }
1398                 if (flow && flow->internal) {
1399                         assert(flow->rss);
1400                         if (flow->rss->queues != queues ||
1401                             memcmp(flow->rss->queue_id, rss_conf->queue,
1402                                    queues * sizeof(flow->rss->queue_id[0])))
1403                                 flow = NULL;
1404                 }
1405                 if (!flow || !flow->internal) {
1406                         /* Not found, create a new flow rule. */
1407                         if (priv->dev->data->promiscuous) {
1408                                 pattern[1].spec = NULL;
1409                                 pattern[1].mask = NULL;
1410                         } else {
1411                                 assert(priv->dev->data->all_multicast);
1412                                 pattern[1].spec = &eth_allmulti;
1413                                 pattern[1].mask = &eth_allmulti;
1414                         }
1415                         pattern[2] = pattern[3];
1416                         flow = mlx4_flow_create(priv->dev, &attr, pattern,
1417                                                 actions, error);
1418                         if (!flow) {
1419                                 err = -rte_errno;
1420                                 goto error;
1421                         }
1422                 }
1423                 assert(flow->promisc || flow->allmulti);
1424                 flow->select = 1;
1425         }
1426 error:
1427         /* Clear selection and clean up stale internal flow rules. */
1428         flow = LIST_FIRST(&priv->flows);
1429         while (flow && flow->internal) {
1430                 struct rte_flow *next = LIST_NEXT(flow, next);
1431
1432                 if (!flow->select)
1433                         claim_zero(mlx4_flow_destroy(priv->dev, flow, error));
1434                 else
1435                         flow->select = 0;
1436                 flow = next;
1437         }
1438         return err;
1439 }
1440
1441 /**
1442  * Synchronize flow rules.
1443  *
1444  * This function synchronizes flow rules with the state of the device by
1445  * taking into account isolated mode and whether target queues are
1446  * configured.
1447  *
1448  * @param priv
1449  *   Pointer to private structure.
1450  * @param[out] error
1451  *   Perform verbose error reporting if not NULL.
1452  *
1453  * @return
1454  *   0 on success, a negative errno value otherwise and rte_errno is set.
1455  */
1456 int
1457 mlx4_flow_sync(struct priv *priv, struct rte_flow_error *error)
1458 {
1459         struct rte_flow *flow;
1460         int ret;
1461
1462         /* Internal flow rules are guaranteed to come first in the list. */
1463         if (priv->isolated) {
1464                 /*
1465                  * Get rid of them in isolated mode, stop at the first
1466                  * non-internal rule found.
1467                  */
1468                 for (flow = LIST_FIRST(&priv->flows);
1469                      flow && flow->internal;
1470                      flow = LIST_FIRST(&priv->flows))
1471                         claim_zero(mlx4_flow_destroy(priv->dev, flow, error));
1472         } else {
1473                 /* Refresh internal rules. */
1474                 ret = mlx4_flow_internal(priv, error);
1475                 if (ret)
1476                         return ret;
1477         }
1478         /* Toggle the remaining flow rules . */
1479         LIST_FOREACH(flow, &priv->flows, next) {
1480                 ret = mlx4_flow_toggle(priv, flow, priv->started, error);
1481                 if (ret)
1482                         return ret;
1483         }
1484         if (!priv->started)
1485                 assert(!priv->drop);
1486         return 0;
1487 }
1488
1489 /**
1490  * Clean up all flow rules.
1491  *
1492  * Unlike mlx4_flow_flush(), this function takes care of all remaining flow
1493  * rules regardless of whether they are internal or user-configured.
1494  *
1495  * @param priv
1496  *   Pointer to private structure.
1497  */
1498 void
1499 mlx4_flow_clean(struct priv *priv)
1500 {
1501         struct rte_flow *flow;
1502
1503         while ((flow = LIST_FIRST(&priv->flows)))
1504                 mlx4_flow_destroy(priv->dev, flow, NULL);
1505         assert(LIST_EMPTY(&priv->rss));
1506 }
1507
1508 static const struct rte_flow_ops mlx4_flow_ops = {
1509         .validate = mlx4_flow_validate,
1510         .create = mlx4_flow_create,
1511         .destroy = mlx4_flow_destroy,
1512         .flush = mlx4_flow_flush,
1513         .isolate = mlx4_flow_isolate,
1514 };
1515
1516 /**
1517  * Manage filter operations.
1518  *
1519  * @param dev
1520  *   Pointer to Ethernet device structure.
1521  * @param filter_type
1522  *   Filter type.
1523  * @param filter_op
1524  *   Operation to perform.
1525  * @param arg
1526  *   Pointer to operation-specific structure.
1527  *
1528  * @return
1529  *   0 on success, negative errno value otherwise and rte_errno is set.
1530  */
1531 int
1532 mlx4_filter_ctrl(struct rte_eth_dev *dev,
1533                  enum rte_filter_type filter_type,
1534                  enum rte_filter_op filter_op,
1535                  void *arg)
1536 {
1537         switch (filter_type) {
1538         case RTE_ETH_FILTER_GENERIC:
1539                 if (filter_op != RTE_ETH_FILTER_GET)
1540                         break;
1541                 *(const void **)arg = &mlx4_flow_ops;
1542                 return 0;
1543         default:
1544                 ERROR("%p: filter type (%d) not supported",
1545                       (void *)dev, filter_type);
1546                 break;
1547         }
1548         rte_errno = ENOTSUP;
1549         return -rte_errno;
1550 }