drivers/net/mlx4/mlx4_flow.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2017 6WIND S.A.
   3  * Copyright 2017 Mellanox Technologies, Ltd
   4  */
   5
   6 /**
   7  * @file
   8  * Flow API operations for mlx4 driver.
   9  */
  10
  11 #include <arpa/inet.h>
  12 #include <assert.h>
  13 #include <errno.h>
  14 #include <stdalign.h>
  15 #include <stddef.h>
  16 #include <stdint.h>
  17 #include <string.h>
  18 #include <sys/queue.h>
  19
  20 /* Verbs headers do not support -pedantic. */
  21 #ifdef PEDANTIC
  22 #pragma GCC diagnostic ignored "-Wpedantic"
  23 #endif
  24 #include <infiniband/verbs.h>
  25 #ifdef PEDANTIC
  26 #pragma GCC diagnostic error "-Wpedantic"
  27 #endif
  28
  29 #include <rte_byteorder.h>
  30 #include <rte_errno.h>
  31 #include <rte_eth_ctrl.h>
  32 #include <rte_ethdev_driver.h>
  33 #include <rte_ether.h>
  34 #include <rte_flow.h>
  35 #include <rte_flow_driver.h>
  36 #include <rte_malloc.h>
  37
  38 /* PMD headers. */
  39 #include "mlx4.h"
  40 #include "mlx4_glue.h"
  41 #include "mlx4_flow.h"
  42 #include "mlx4_rxtx.h"
  43 #include "mlx4_utils.h"
  44
  45 /** Static initializer for a list of subsequent item types. */
  46 #define NEXT_ITEM(...) \
  47         (const enum rte_flow_item_type []){ \
  48                 __VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \
  49         }
  50
  51 /** Processor structure associated with a flow item. */
  52 struct mlx4_flow_proc_item {
  53         /** Bit-mask for fields supported by this PMD. */
  54         const void *mask_support;
  55         /** Bit-mask to use when @p item->mask is not provided. */
  56         const void *mask_default;
  57         /** Size in bytes for @p mask_support and @p mask_default. */
  58         const unsigned int mask_sz;
  59         /** Merge a pattern item into a flow rule handle. */
  60         int (*merge)(struct rte_flow *flow,
  61                      const struct rte_flow_item *item,
  62                      const struct mlx4_flow_proc_item *proc,
  63                      struct rte_flow_error *error);
  64         /** Size in bytes of the destination structure. */
  65         const unsigned int dst_sz;
  66         /** List of possible subsequent items. */
  67         const enum rte_flow_item_type *const next_item;
  68 };
  69
  70 /** Shared resources for drop flow rules. */
  71 struct mlx4_drop {
  72         struct ibv_qp *qp; /**< QP target. */
  73         struct ibv_cq *cq; /**< CQ associated with above QP. */
  74         struct priv *priv; /**< Back pointer to private data. */
  75         uint32_t refcnt; /**< Reference count. */
  76 };
  77
  78 /**
  79  * Convert DPDK RSS hash types to their Verbs equivalent.
  80  *
  81  * This function returns the supported (default) set when @p types has
  82  * special value (uint64_t)-1.
  83  *
  84  * @param priv
  85  *   Pointer to private structure.
  86  * @param types
  87  *   Hash types in DPDK format (see struct rte_eth_rss_conf).
  88  *
  89  * @return
  90  *   A valid Verbs RSS hash fields mask for mlx4 on success, (uint64_t)-1
  91  *   otherwise and rte_errno is set.
  92  */
  93 uint64_t
  94 mlx4_conv_rss_types(struct priv *priv, uint64_t types)
  95 {
  96         enum { IPV4, IPV6, TCP, UDP, };
  97         const uint64_t in[] = {
  98                 [IPV4] = (ETH_RSS_IPV4 |
  99                           ETH_RSS_FRAG_IPV4 |
 100                           ETH_RSS_NONFRAG_IPV4_TCP |
 101                           ETH_RSS_NONFRAG_IPV4_UDP |
 102                           ETH_RSS_NONFRAG_IPV4_OTHER),
 103                 [IPV6] = (ETH_RSS_IPV6 |
 104                           ETH_RSS_FRAG_IPV6 |
 105                           ETH_RSS_NONFRAG_IPV6_TCP |
 106                           ETH_RSS_NONFRAG_IPV6_UDP |
 107                           ETH_RSS_NONFRAG_IPV6_OTHER |
 108                           ETH_RSS_IPV6_EX |
 109                           ETH_RSS_IPV6_TCP_EX |
 110                           ETH_RSS_IPV6_UDP_EX),
 111                 [TCP] = (ETH_RSS_NONFRAG_IPV4_TCP |
 112                          ETH_RSS_NONFRAG_IPV6_TCP |
 113                          ETH_RSS_IPV6_TCP_EX),
 114                 [UDP] = (ETH_RSS_NONFRAG_IPV4_UDP |
 115                          ETH_RSS_NONFRAG_IPV6_UDP |
 116                          ETH_RSS_IPV6_UDP_EX),
 117         };
 118         const uint64_t out[RTE_DIM(in)] = {
 119                 [IPV4] = IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4,
 120                 [IPV6] = IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6,
 121                 [TCP] = IBV_RX_HASH_SRC_PORT_TCP | IBV_RX_HASH_DST_PORT_TCP,
 122                 [UDP] = IBV_RX_HASH_SRC_PORT_UDP | IBV_RX_HASH_DST_PORT_UDP,
 123         };
 124         uint64_t seen = 0;
 125         uint64_t conv = 0;
 126         unsigned int i;
 127
 128         if (types == (uint64_t)-1)
 129                 return priv->hw_rss_sup;
 130         for (i = 0; i != RTE_DIM(in); ++i)
 131                 if (types & in[i]) {
 132                         seen |= types & in[i];
 133                         conv |= out[i];
 134                 }
 135         if ((conv & priv->hw_rss_sup) == conv && !(types & ~seen))
 136                 return conv;
 137         rte_errno = ENOTSUP;
 138         return (uint64_t)-1;
 139 }
 140
 141 /**
 142  * Merge Ethernet pattern item into flow rule handle.
 143  *
 144  * Additional mlx4-specific constraints on supported fields:
 145  *
 146  * - No support for partial masks, except in the specific case of matching
 147  *   all multicast traffic (@p spec->dst and @p mask->dst equal to
 148  *   01:00:00:00:00:00).
 149  * - Not providing @p item->spec or providing an empty @p mask->dst is
 150  *   *only* supported if the rule doesn't specify additional matching
 151  *   criteria (i.e. rule is promiscuous-like).
 152  *
 153  * @param[in, out] flow
 154  *   Flow rule handle to update.
 155  * @param[in] item
 156  *   Pattern item to merge.
 157  * @param[in] proc
 158  *   Associated item-processing object.
 159  * @param[out] error
 160  *   Perform verbose error reporting if not NULL.
 161  *
 162  * @return
 163  *   0 on success, a negative errno value otherwise and rte_errno is set.
 164  */
 165 static int
 166 mlx4_flow_merge_eth(struct rte_flow *flow,
 167                     const struct rte_flow_item *item,
 168                     const struct mlx4_flow_proc_item *proc,
 169                     struct rte_flow_error *error)
 170 {
 171         const struct rte_flow_item_eth *spec = item->spec;
 172         const struct rte_flow_item_eth *mask =
 173                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 174         struct ibv_flow_spec_eth *eth;
 175         const char *msg;
 176         unsigned int i;
 177
 178         if (!mask) {
 179                 flow->promisc = 1;
 180         } else {
 181                 uint32_t sum_dst = 0;
 182                 uint32_t sum_src = 0;
 183
 184                 for (i = 0; i != sizeof(mask->dst.addr_bytes); ++i) {
 185                         sum_dst += mask->dst.addr_bytes[i];
 186                         sum_src += mask->src.addr_bytes[i];
 187                 }
 188                 if (sum_src) {
 189                         msg = "mlx4 does not support source MAC matching";
 190                         goto error;
 191                 } else if (!sum_dst) {
 192                         flow->promisc = 1;
 193                 } else if (sum_dst == 1 && mask->dst.addr_bytes[0] == 1) {
 194                         if (!(spec->dst.addr_bytes[0] & 1)) {
 195                                 msg = "mlx4 does not support the explicit"
 196                                         " exclusion of all multicast traffic";
 197                                 goto error;
 198                         }
 199                         flow->allmulti = 1;
 200                 } else if (sum_dst != (UINT8_C(0xff) * ETHER_ADDR_LEN)) {
 201                         msg = "mlx4 does not support matching partial"
 202                                 " Ethernet fields";
 203                         goto error;
 204                 }
 205         }
 206         if (!flow->ibv_attr)
 207                 return 0;
 208         if (flow->promisc) {
 209                 flow->ibv_attr->type = IBV_FLOW_ATTR_ALL_DEFAULT;
 210                 return 0;
 211         }
 212         if (flow->allmulti) {
 213                 flow->ibv_attr->type = IBV_FLOW_ATTR_MC_DEFAULT;
 214                 return 0;
 215         }
 216         ++flow->ibv_attr->num_of_specs;
 217         eth = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 218         *eth = (struct ibv_flow_spec_eth) {
 219                 .type = IBV_FLOW_SPEC_ETH,
 220                 .size = sizeof(*eth),
 221         };
 222         memcpy(eth->val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
 223         memcpy(eth->mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
 224         /* Remove unwanted bits from values. */
 225         for (i = 0; i < ETHER_ADDR_LEN; ++i) {
 226                 eth->val.dst_mac[i] &= eth->mask.dst_mac[i];
 227         }
 228         return 0;
 229 error:
 230         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 231                                   item, msg);
 232 }
 233
 234 /**
 235  * Merge VLAN pattern item into flow rule handle.
 236  *
 237  * Additional mlx4-specific constraints on supported fields:
 238  *
 239  * - Matching *all* VLAN traffic by omitting @p item->spec or providing an
 240  *   empty @p item->mask would also include non-VLAN traffic. Doing so is
 241  *   therefore unsupported.
 242  * - No support for partial masks.
 243  *
 244  * @param[in, out] flow
 245  *   Flow rule handle to update.
 246  * @param[in] item
 247  *   Pattern item to merge.
 248  * @param[in] proc
 249  *   Associated item-processing object.
 250  * @param[out] error
 251  *   Perform verbose error reporting if not NULL.
 252  *
 253  * @return
 254  *   0 on success, a negative errno value otherwise and rte_errno is set.
 255  */
 256 static int
 257 mlx4_flow_merge_vlan(struct rte_flow *flow,
 258                      const struct rte_flow_item *item,
 259                      const struct mlx4_flow_proc_item *proc,
 260                      struct rte_flow_error *error)
 261 {
 262         const struct rte_flow_item_vlan *spec = item->spec;
 263         const struct rte_flow_item_vlan *mask =
 264                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 265         struct ibv_flow_spec_eth *eth;
 266         const char *msg;
 267
 268         if (!mask || !mask->tci) {
 269                 msg = "mlx4 cannot match all VLAN traffic while excluding"
 270                         " non-VLAN traffic, TCI VID must be specified";
 271                 goto error;
 272         }
 273         if (mask->tci != RTE_BE16(0x0fff)) {
 274                 msg = "mlx4 does not support partial TCI VID matching";
 275                 goto error;
 276         }
 277         if (!flow->ibv_attr)
 278                 return 0;
 279         eth = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size -
 280                        sizeof(*eth));
 281         eth->val.vlan_tag = spec->tci;
 282         eth->mask.vlan_tag = mask->tci;
 283         eth->val.vlan_tag &= eth->mask.vlan_tag;
 284         return 0;
 285 error:
 286         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 287                                   item, msg);
 288 }
 289
 290 /**
 291  * Merge IPv4 pattern item into flow rule handle.
 292  *
 293  * Additional mlx4-specific constraints on supported fields:
 294  *
 295  * - No support for partial masks.
 296  *
 297  * @param[in, out] flow
 298  *   Flow rule handle to update.
 299  * @param[in] item
 300  *   Pattern item to merge.
 301  * @param[in] proc
 302  *   Associated item-processing object.
 303  * @param[out] error
 304  *   Perform verbose error reporting if not NULL.
 305  *
 306  * @return
 307  *   0 on success, a negative errno value otherwise and rte_errno is set.
 308  */
 309 static int
 310 mlx4_flow_merge_ipv4(struct rte_flow *flow,
 311                      const struct rte_flow_item *item,
 312                      const struct mlx4_flow_proc_item *proc,
 313                      struct rte_flow_error *error)
 314 {
 315         const struct rte_flow_item_ipv4 *spec = item->spec;
 316         const struct rte_flow_item_ipv4 *mask =
 317                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 318         struct ibv_flow_spec_ipv4 *ipv4;
 319         const char *msg;
 320
 321         if (mask &&
 322             ((uint32_t)(mask->hdr.src_addr + 1) > UINT32_C(1) ||
 323              (uint32_t)(mask->hdr.dst_addr + 1) > UINT32_C(1))) {
 324                 msg = "mlx4 does not support matching partial IPv4 fields";
 325                 goto error;
 326         }
 327         if (!flow->ibv_attr)
 328                 return 0;
 329         ++flow->ibv_attr->num_of_specs;
 330         ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 331         *ipv4 = (struct ibv_flow_spec_ipv4) {
 332                 .type = IBV_FLOW_SPEC_IPV4,
 333                 .size = sizeof(*ipv4),
 334         };
 335         if (!spec)
 336                 return 0;
 337         ipv4->val = (struct ibv_flow_ipv4_filter) {
 338                 .src_ip = spec->hdr.src_addr,
 339                 .dst_ip = spec->hdr.dst_addr,
 340         };
 341         ipv4->mask = (struct ibv_flow_ipv4_filter) {
 342                 .src_ip = mask->hdr.src_addr,
 343                 .dst_ip = mask->hdr.dst_addr,
 344         };
 345         /* Remove unwanted bits from values. */
 346         ipv4->val.src_ip &= ipv4->mask.src_ip;
 347         ipv4->val.dst_ip &= ipv4->mask.dst_ip;
 348         return 0;
 349 error:
 350         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 351                                   item, msg);
 352 }
 353
 354 /**
 355  * Merge UDP pattern item into flow rule handle.
 356  *
 357  * Additional mlx4-specific constraints on supported fields:
 358  *
 359  * - No support for partial masks.
 360  * - Due to HW/FW limitation, flow rule priority is not taken into account
 361  *   when matching UDP destination ports, doing is therefore only supported
 362  *   at the highest priority level (0).
 363  *
 364  * @param[in, out] flow
 365  *   Flow rule handle to update.
 366  * @param[in] item
 367  *   Pattern item to merge.
 368  * @param[in] proc
 369  *   Associated item-processing object.
 370  * @param[out] error
 371  *   Perform verbose error reporting if not NULL.
 372  *
 373  * @return
 374  *   0 on success, a negative errno value otherwise and rte_errno is set.
 375  */
 376 static int
 377 mlx4_flow_merge_udp(struct rte_flow *flow,
 378                     const struct rte_flow_item *item,
 379                     const struct mlx4_flow_proc_item *proc,
 380                     struct rte_flow_error *error)
 381 {
 382         const struct rte_flow_item_udp *spec = item->spec;
 383         const struct rte_flow_item_udp *mask =
 384                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 385         struct ibv_flow_spec_tcp_udp *udp;
 386         const char *msg;
 387
 388         if (mask &&
 389             ((uint16_t)(mask->hdr.src_port + 1) > UINT16_C(1) ||
 390              (uint16_t)(mask->hdr.dst_port + 1) > UINT16_C(1))) {
 391                 msg = "mlx4 does not support matching partial UDP fields";
 392                 goto error;
 393         }
 394         if (mask && mask->hdr.dst_port && flow->priority) {
 395                 msg = "combining UDP destination port matching with a nonzero"
 396                         " priority level is not supported";
 397                 goto error;
 398         }
 399         if (!flow->ibv_attr)
 400                 return 0;
 401         ++flow->ibv_attr->num_of_specs;
 402         udp = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 403         *udp = (struct ibv_flow_spec_tcp_udp) {
 404                 .type = IBV_FLOW_SPEC_UDP,
 405                 .size = sizeof(*udp),
 406         };
 407         if (!spec)
 408                 return 0;
 409         udp->val.dst_port = spec->hdr.dst_port;
 410         udp->val.src_port = spec->hdr.src_port;
 411         udp->mask.dst_port = mask->hdr.dst_port;
 412         udp->mask.src_port = mask->hdr.src_port;
 413         /* Remove unwanted bits from values. */
 414         udp->val.src_port &= udp->mask.src_port;
 415         udp->val.dst_port &= udp->mask.dst_port;
 416         return 0;
 417 error:
 418         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 419                                   item, msg);
 420 }
 421
 422 /**
 423  * Merge TCP pattern item into flow rule handle.
 424  *
 425  * Additional mlx4-specific constraints on supported fields:
 426  *
 427  * - No support for partial masks.
 428  *
 429  * @param[in, out] flow
 430  *   Flow rule handle to update.
 431  * @param[in] item
 432  *   Pattern item to merge.
 433  * @param[in] proc
 434  *   Associated item-processing object.
 435  * @param[out] error
 436  *   Perform verbose error reporting if not NULL.
 437  *
 438  * @return
 439  *   0 on success, a negative errno value otherwise and rte_errno is set.
 440  */
 441 static int
 442 mlx4_flow_merge_tcp(struct rte_flow *flow,
 443                     const struct rte_flow_item *item,
 444                     const struct mlx4_flow_proc_item *proc,
 445                     struct rte_flow_error *error)
 446 {
 447         const struct rte_flow_item_tcp *spec = item->spec;
 448         const struct rte_flow_item_tcp *mask =
 449                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 450         struct ibv_flow_spec_tcp_udp *tcp;
 451         const char *msg;
 452
 453         if (mask &&
 454             ((uint16_t)(mask->hdr.src_port + 1) > UINT16_C(1) ||
 455              (uint16_t)(mask->hdr.dst_port + 1) > UINT16_C(1))) {
 456                 msg = "mlx4 does not support matching partial TCP fields";
 457                 goto error;
 458         }
 459         if (!flow->ibv_attr)
 460                 return 0;
 461         ++flow->ibv_attr->num_of_specs;
 462         tcp = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 463         *tcp = (struct ibv_flow_spec_tcp_udp) {
 464                 .type = IBV_FLOW_SPEC_TCP,
 465                 .size = sizeof(*tcp),
 466         };
 467         if (!spec)
 468                 return 0;
 469         tcp->val.dst_port = spec->hdr.dst_port;
 470         tcp->val.src_port = spec->hdr.src_port;
 471         tcp->mask.dst_port = mask->hdr.dst_port;
 472         tcp->mask.src_port = mask->hdr.src_port;
 473         /* Remove unwanted bits from values. */
 474         tcp->val.src_port &= tcp->mask.src_port;
 475         tcp->val.dst_port &= tcp->mask.dst_port;
 476         return 0;
 477 error:
 478         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 479                                   item, msg);
 480 }
 481
 482 /**
 483  * Perform basic sanity checks on a pattern item.
 484  *
 485  * @param[in] item
 486  *   Item specification.
 487  * @param[in] proc
 488  *   Associated item-processing object.
 489  * @param[out] error
 490  *   Perform verbose error reporting if not NULL.
 491  *
 492  * @return
 493  *   0 on success, a negative errno value otherwise and rte_errno is set.
 494  */
 495 static int
 496 mlx4_flow_item_check(const struct rte_flow_item *item,
 497                      const struct mlx4_flow_proc_item *proc,
 498                      struct rte_flow_error *error)
 499 {
 500         const uint8_t *mask;
 501         unsigned int i;
 502
 503         /* item->last and item->mask cannot exist without item->spec. */
 504         if (!item->spec && (item->mask || item->last))
 505                 return rte_flow_error_set
 506                         (error, EINVAL, RTE_FLOW_ERROR_TYPE_ITEM, item,
 507                          "\"mask\" or \"last\" field provided without a"
 508                          " corresponding \"spec\"");
 509         /* No spec, no mask, no problem. */
 510         if (!item->spec)
 511                 return 0;
 512         mask = item->mask ?
 513                 (const uint8_t *)item->mask :
 514                 (const uint8_t *)proc->mask_default;
 515         assert(mask);
 516         /*
 517          * Single-pass check to make sure that:
 518          * - Mask is supported, no bits are set outside proc->mask_support.
 519          * - Both item->spec and item->last are included in mask.
 520          */
 521         for (i = 0; i != proc->mask_sz; ++i) {
 522                 if (!mask[i])
 523                         continue;
 524                 if ((mask[i] | ((const uint8_t *)proc->mask_support)[i]) !=
 525                     ((const uint8_t *)proc->mask_support)[i])
 526                         return rte_flow_error_set
 527                                 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 528                                  item, "unsupported field found in \"mask\"");
 529                 if (item->last &&
 530                     (((const uint8_t *)item->spec)[i] & mask[i]) !=
 531                     (((const uint8_t *)item->last)[i] & mask[i]))
 532                         return rte_flow_error_set
 533                                 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 534                                  item,
 535                                  "range between \"spec\" and \"last\""
 536                                  " is larger than \"mask\"");
 537         }
 538         return 0;
 539 }
 540
 541 /** Graph of supported items and associated actions. */
 542 static const struct mlx4_flow_proc_item mlx4_flow_proc_item_list[] = {
 543         [RTE_FLOW_ITEM_TYPE_END] = {
 544                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_ETH),
 545         },
 546         [RTE_FLOW_ITEM_TYPE_ETH] = {
 547                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_VLAN,
 548                                        RTE_FLOW_ITEM_TYPE_IPV4),
 549                 .mask_support = &(const struct rte_flow_item_eth){
 550                         /* Only destination MAC can be matched. */
 551                         .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 552                 },
 553                 .mask_default = &rte_flow_item_eth_mask,
 554                 .mask_sz = sizeof(struct rte_flow_item_eth),
 555                 .merge = mlx4_flow_merge_eth,
 556                 .dst_sz = sizeof(struct ibv_flow_spec_eth),
 557         },
 558         [RTE_FLOW_ITEM_TYPE_VLAN] = {
 559                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_IPV4),
 560                 .mask_support = &(const struct rte_flow_item_vlan){
 561                         /* Only TCI VID matching is supported. */
 562                         .tci = RTE_BE16(0x0fff),
 563                 },
 564                 .mask_default = &rte_flow_item_vlan_mask,
 565                 .mask_sz = sizeof(struct rte_flow_item_vlan),
 566                 .merge = mlx4_flow_merge_vlan,
 567                 .dst_sz = 0,
 568         },
 569         [RTE_FLOW_ITEM_TYPE_IPV4] = {
 570                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_UDP,
 571                                        RTE_FLOW_ITEM_TYPE_TCP),
 572                 .mask_support = &(const struct rte_flow_item_ipv4){
 573                         .hdr = {
 574                                 .src_addr = RTE_BE32(0xffffffff),
 575                                 .dst_addr = RTE_BE32(0xffffffff),
 576                         },
 577                 },
 578                 .mask_default = &rte_flow_item_ipv4_mask,
 579                 .mask_sz = sizeof(struct rte_flow_item_ipv4),
 580                 .merge = mlx4_flow_merge_ipv4,
 581                 .dst_sz = sizeof(struct ibv_flow_spec_ipv4),
 582         },
 583         [RTE_FLOW_ITEM_TYPE_UDP] = {
 584                 .mask_support = &(const struct rte_flow_item_udp){
 585                         .hdr = {
 586                                 .src_port = RTE_BE16(0xffff),
 587                                 .dst_port = RTE_BE16(0xffff),
 588                         },
 589                 },
 590                 .mask_default = &rte_flow_item_udp_mask,
 591                 .mask_sz = sizeof(struct rte_flow_item_udp),
 592                 .merge = mlx4_flow_merge_udp,
 593                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 594         },
 595         [RTE_FLOW_ITEM_TYPE_TCP] = {
 596                 .mask_support = &(const struct rte_flow_item_tcp){
 597                         .hdr = {
 598                                 .src_port = RTE_BE16(0xffff),
 599                                 .dst_port = RTE_BE16(0xffff),
 600                         },
 601                 },
 602                 .mask_default = &rte_flow_item_tcp_mask,
 603                 .mask_sz = sizeof(struct rte_flow_item_tcp),
 604                 .merge = mlx4_flow_merge_tcp,
 605                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 606         },
 607 };
 608
 609 /**
 610  * Make sure a flow rule is supported and initialize associated structure.
 611  *
 612  * @param priv
 613  *   Pointer to private structure.
 614  * @param[in] attr
 615  *   Flow rule attributes.
 616  * @param[in] pattern
 617  *   Pattern specification (list terminated by the END pattern item).
 618  * @param[in] actions
 619  *   Associated actions (list terminated by the END action).
 620  * @param[out] error
 621  *   Perform verbose error reporting if not NULL.
 622  * @param[in, out] addr
 623  *   Buffer where the resulting flow rule handle pointer must be stored.
 624  *   If NULL, stop processing after validation stage.
 625  *
 626  * @return
 627  *   0 on success, a negative errno value otherwise and rte_errno is set.
 628  */
 629 static int
 630 mlx4_flow_prepare(struct priv *priv,
 631                   const struct rte_flow_attr *attr,
 632                   const struct rte_flow_item pattern[],
 633                   const struct rte_flow_action actions[],
 634                   struct rte_flow_error *error,
 635                   struct rte_flow **addr)
 636 {
 637         const struct rte_flow_item *item;
 638         const struct rte_flow_action *action;
 639         const struct mlx4_flow_proc_item *proc;
 640         struct rte_flow temp = { .ibv_attr_size = sizeof(*temp.ibv_attr) };
 641         struct rte_flow *flow = &temp;
 642         const char *msg = NULL;
 643         int overlap;
 644
 645         if (attr->group)
 646                 return rte_flow_error_set
 647                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 648                          NULL, "groups are not supported");
 649         if (attr->priority > MLX4_FLOW_PRIORITY_LAST)
 650                 return rte_flow_error_set
 651                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 652                          NULL, "maximum priority level is "
 653                          MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST));
 654         if (attr->egress)
 655                 return rte_flow_error_set
 656                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 657                          NULL, "egress is not supported");
 658         if (attr->transfer)
 659                 return rte_flow_error_set
 660                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_TRANSFER,
 661                          NULL, "transfer is not supported");
 662         if (!attr->ingress)
 663                 return rte_flow_error_set
 664                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 665                          NULL, "only ingress is supported");
 666 fill:
 667         overlap = 0;
 668         proc = mlx4_flow_proc_item_list;
 669         flow->priority = attr->priority;
 670         /* Go over pattern. */
 671         for (item = pattern; item->type; ++item) {
 672                 const struct mlx4_flow_proc_item *next = NULL;
 673                 unsigned int i;
 674                 int err;
 675
 676                 if (item->type == RTE_FLOW_ITEM_TYPE_VOID)
 677                         continue;
 678                 if (item->type == MLX4_FLOW_ITEM_TYPE_INTERNAL) {
 679                         flow->internal = 1;
 680                         continue;
 681                 }
 682                 if (flow->promisc || flow->allmulti) {
 683                         msg = "mlx4 does not support additional matching"
 684                                 " criteria combined with indiscriminate"
 685                                 " matching on Ethernet headers";
 686                         goto exit_item_not_supported;
 687                 }
 688                 for (i = 0; proc->next_item && proc->next_item[i]; ++i) {
 689                         if (proc->next_item[i] == item->type) {
 690                                 next = &mlx4_flow_proc_item_list[item->type];
 691                                 break;
 692                         }
 693                 }
 694                 if (!next)
 695                         goto exit_item_not_supported;
 696                 proc = next;
 697                 /*
 698                  * Perform basic sanity checks only once, while handle is
 699                  * not allocated.
 700                  */
 701                 if (flow == &temp) {
 702                         err = mlx4_flow_item_check(item, proc, error);
 703                         if (err)
 704                                 return err;
 705                 }
 706                 if (proc->merge) {
 707                         err = proc->merge(flow, item, proc, error);
 708                         if (err)
 709                                 return err;
 710                 }
 711                 flow->ibv_attr_size += proc->dst_sz;
 712         }
 713         /* Go over actions list. */
 714         for (action = actions; action->type; ++action) {
 715                 /* This one may appear anywhere multiple times. */
 716                 if (action->type == RTE_FLOW_ACTION_TYPE_VOID)
 717                         continue;
 718                 /* Fate-deciding actions may appear exactly once. */
 719                 if (overlap) {
 720                         msg = "cannot combine several fate-deciding actions,"
 721                                 " choose between DROP, QUEUE or RSS";
 722                         goto exit_action_not_supported;
 723                 }
 724                 overlap = 1;
 725                 switch (action->type) {
 726                         const struct rte_flow_action_queue *queue;
 727                         const struct rte_flow_action_rss *rss;
 728                         const uint8_t *rss_key;
 729                         uint32_t rss_key_len;
 730                         uint64_t fields;
 731                         unsigned int i;
 732
 733                 case RTE_FLOW_ACTION_TYPE_DROP:
 734                         flow->drop = 1;
 735                         break;
 736                 case RTE_FLOW_ACTION_TYPE_QUEUE:
 737                         if (flow->rss)
 738                                 break;
 739                         queue = action->conf;
 740                         if (queue->index >= priv->dev->data->nb_rx_queues) {
 741                                 msg = "queue target index beyond number of"
 742                                         " configured Rx queues";
 743                                 goto exit_action_not_supported;
 744                         }
 745                         flow->rss = mlx4_rss_get
 746                                 (priv, 0, mlx4_rss_hash_key_default, 1,
 747                                  &queue->index);
 748                         if (!flow->rss) {
 749                                 msg = "not enough resources for additional"
 750                                         " single-queue RSS context";
 751                                 goto exit_action_not_supported;
 752                         }
 753                         break;
 754                 case RTE_FLOW_ACTION_TYPE_RSS:
 755                         if (flow->rss)
 756                                 break;
 757                         rss = action->conf;
 758                         /* Default RSS configuration if none is provided. */
 759                         if (rss->key_len) {
 760                                 rss_key = rss->key;
 761                                 rss_key_len = rss->key_len;
 762                         } else {
 763                                 rss_key = mlx4_rss_hash_key_default;
 764                                 rss_key_len = MLX4_RSS_HASH_KEY_SIZE;
 765                         }
 766                         /* Sanity checks. */
 767                         for (i = 0; i < rss->queue_num; ++i)
 768                                 if (rss->queue[i] >=
 769                                     priv->dev->data->nb_rx_queues)
 770                                         break;
 771                         if (i != rss->queue_num) {
 772                                 msg = "queue index target beyond number of"
 773                                         " configured Rx queues";
 774                                 goto exit_action_not_supported;
 775                         }
 776                         if (!rte_is_power_of_2(rss->queue_num)) {
 777                                 msg = "for RSS, mlx4 requires the number of"
 778                                         " queues to be a power of two";
 779                                 goto exit_action_not_supported;
 780                         }
 781                         if (rss_key_len != sizeof(flow->rss->key)) {
 782                                 msg = "mlx4 supports exactly one RSS hash key"
 783                                         " length: "
 784                                         MLX4_STR_EXPAND(MLX4_RSS_HASH_KEY_SIZE);
 785                                 goto exit_action_not_supported;
 786                         }
 787                         for (i = 1; i < rss->queue_num; ++i)
 788                                 if (rss->queue[i] - rss->queue[i - 1] != 1)
 789                                         break;
 790                         if (i != rss->queue_num) {
 791                                 msg = "mlx4 requires RSS contexts to use"
 792                                         " consecutive queue indices only";
 793                                 goto exit_action_not_supported;
 794                         }
 795                         if (rss->queue[0] % rss->queue_num) {
 796                                 msg = "mlx4 requires the first queue of a RSS"
 797                                         " context to be aligned on a multiple"
 798                                         " of the context size";
 799                                 goto exit_action_not_supported;
 800                         }
 801                         if (rss->func &&
 802                             rss->func != RTE_ETH_HASH_FUNCTION_TOEPLITZ) {
 803                                 msg = "the only supported RSS hash function"
 804                                         " is Toeplitz";
 805                                 goto exit_action_not_supported;
 806                         }
 807                         if (rss->level) {
 808                                 msg = "a nonzero RSS encapsulation level is"
 809                                         " not supported";
 810                                 goto exit_action_not_supported;
 811                         }
 812                         rte_errno = 0;
 813                         fields = mlx4_conv_rss_types(priv, rss->types);
 814                         if (fields == (uint64_t)-1 && rte_errno) {
 815                                 msg = "unsupported RSS hash type requested";
 816                                 goto exit_action_not_supported;
 817                         }
 818                         flow->rss = mlx4_rss_get
 819                                 (priv, fields, rss_key, rss->queue_num,
 820                                  rss->queue);
 821                         if (!flow->rss) {
 822                                 msg = "either invalid parameters or not enough"
 823                                         " resources for additional multi-queue"
 824                                         " RSS context";
 825                                 goto exit_action_not_supported;
 826                         }
 827                         break;
 828                 default:
 829                         goto exit_action_not_supported;
 830                 }
 831         }
 832         /* When fate is unknown, drop traffic. */
 833         if (!overlap)
 834                 flow->drop = 1;
 835         /* Validation ends here. */
 836         if (!addr) {
 837                 if (flow->rss)
 838                         mlx4_rss_put(flow->rss);
 839                 return 0;
 840         }
 841         if (flow == &temp) {
 842                 /* Allocate proper handle based on collected data. */
 843                 const struct mlx4_malloc_vec vec[] = {
 844                         {
 845                                 .align = alignof(struct rte_flow),
 846                                 .size = sizeof(*flow),
 847                                 .addr = (void **)&flow,
 848                         },
 849                         {
 850                                 .align = alignof(struct ibv_flow_attr),
 851                                 .size = temp.ibv_attr_size,
 852                                 .addr = (void **)&temp.ibv_attr,
 853                         },
 854                 };
 855
 856                 if (!mlx4_zmallocv(__func__, vec, RTE_DIM(vec))) {
 857                         if (temp.rss)
 858                                 mlx4_rss_put(temp.rss);
 859                         return rte_flow_error_set
 860                                 (error, -rte_errno,
 861                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
 862                                  "flow rule handle allocation failure");
 863                 }
 864                 /* Most fields will be updated by second pass. */
 865                 *flow = (struct rte_flow){
 866                         .ibv_attr = temp.ibv_attr,
 867                         .ibv_attr_size = sizeof(*flow->ibv_attr),
 868                         .rss = temp.rss,
 869                 };
 870                 *flow->ibv_attr = (struct ibv_flow_attr){
 871                         .type = IBV_FLOW_ATTR_NORMAL,
 872                         .size = sizeof(*flow->ibv_attr),
 873                         .priority = attr->priority,
 874                         .port = priv->port,
 875                 };
 876                 goto fill;
 877         }
 878         *addr = flow;
 879         return 0;
 880 exit_item_not_supported:
 881         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 882                                   item, msg ? msg : "item not supported");
 883 exit_action_not_supported:
 884         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 885                                   action, msg ? msg : "action not supported");
 886 }
 887
 888 /**
 889  * Validate a flow supported by the NIC.
 890  *
 891  * @see rte_flow_validate()
 892  * @see rte_flow_ops
 893  */
 894 static int
 895 mlx4_flow_validate(struct rte_eth_dev *dev,
 896                    const struct rte_flow_attr *attr,
 897                    const struct rte_flow_item pattern[],
 898                    const struct rte_flow_action actions[],
 899                    struct rte_flow_error *error)
 900 {
 901         struct priv *priv = dev->data->dev_private;
 902
 903         return mlx4_flow_prepare(priv, attr, pattern, actions, error, NULL);
 904 }
 905
 906 /**
 907  * Get a drop flow rule resources instance.
 908  *
 909  * @param priv
 910  *   Pointer to private structure.
 911  *
 912  * @return
 913  *   Pointer to drop flow resources on success, NULL otherwise and rte_errno
 914  *   is set.
 915  */
 916 static struct mlx4_drop *
 917 mlx4_drop_get(struct priv *priv)
 918 {
 919         struct mlx4_drop *drop = priv->drop;
 920
 921         if (drop) {
 922                 assert(drop->refcnt);
 923                 assert(drop->priv == priv);
 924                 ++drop->refcnt;
 925                 return drop;
 926         }
 927         drop = rte_malloc(__func__, sizeof(*drop), 0);
 928         if (!drop)
 929                 goto error;
 930         *drop = (struct mlx4_drop){
 931                 .priv = priv,
 932                 .refcnt = 1,
 933         };
 934         drop->cq = mlx4_glue->create_cq(priv->ctx, 1, NULL, NULL, 0);
 935         if (!drop->cq)
 936                 goto error;
 937         drop->qp = mlx4_glue->create_qp
 938                 (priv->pd,
 939                  &(struct ibv_qp_init_attr){
 940                         .send_cq = drop->cq,
 941                         .recv_cq = drop->cq,
 942                         .qp_type = IBV_QPT_RAW_PACKET,
 943                  });
 944         if (!drop->qp)
 945                 goto error;
 946         priv->drop = drop;
 947         return drop;
 948 error:
 949         if (drop->qp)
 950                 claim_zero(mlx4_glue->destroy_qp(drop->qp));
 951         if (drop->cq)
 952                 claim_zero(mlx4_glue->destroy_cq(drop->cq));
 953         if (drop)
 954                 rte_free(drop);
 955         rte_errno = ENOMEM;
 956         return NULL;
 957 }
 958
 959 /**
 960  * Give back a drop flow rule resources instance.
 961  *
 962  * @param drop
 963  *   Pointer to drop flow rule resources.
 964  */
 965 static void
 966 mlx4_drop_put(struct mlx4_drop *drop)
 967 {
 968         assert(drop->refcnt);
 969         if (--drop->refcnt)
 970                 return;
 971         drop->priv->drop = NULL;
 972         claim_zero(mlx4_glue->destroy_qp(drop->qp));
 973         claim_zero(mlx4_glue->destroy_cq(drop->cq));
 974         rte_free(drop);
 975 }
 976
 977 /**
 978  * Toggle a configured flow rule.
 979  *
 980  * @param priv
 981  *   Pointer to private structure.
 982  * @param flow
 983  *   Flow rule handle to toggle.
 984  * @param enable
 985  *   Whether associated Verbs flow must be created or removed.
 986  * @param[out] error
 987  *   Perform verbose error reporting if not NULL.
 988  *
 989  * @return
 990  *   0 on success, a negative errno value otherwise and rte_errno is set.
 991  */
 992 static int
 993 mlx4_flow_toggle(struct priv *priv,
 994                  struct rte_flow *flow,
 995                  int enable,
 996                  struct rte_flow_error *error)
 997 {
 998         struct ibv_qp *qp = NULL;
 999         const char *msg;
1000         int err;
1001
1002         if (!enable) {
1003                 if (!flow->ibv_flow)
1004                         return 0;
1005                 claim_zero(mlx4_glue->destroy_flow(flow->ibv_flow));
1006                 flow->ibv_flow = NULL;
1007                 if (flow->drop)
1008                         mlx4_drop_put(priv->drop);
1009                 else if (flow->rss)
1010                         mlx4_rss_detach(flow->rss);
1011                 return 0;
1012         }
1013         assert(flow->ibv_attr);
1014         if (!flow->internal &&
1015             !priv->isolated &&
1016             flow->ibv_attr->priority == MLX4_FLOW_PRIORITY_LAST) {
1017                 if (flow->ibv_flow) {
1018                         claim_zero(mlx4_glue->destroy_flow(flow->ibv_flow));
1019                         flow->ibv_flow = NULL;
1020                         if (flow->drop)
1021                                 mlx4_drop_put(priv->drop);
1022                         else if (flow->rss)
1023                                 mlx4_rss_detach(flow->rss);
1024                 }
1025                 err = EACCES;
1026                 msg = ("priority level "
1027                        MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST)
1028                        " is reserved when not in isolated mode");
1029                 goto error;
1030         }
1031         if (flow->rss) {
1032                 struct mlx4_rss *rss = flow->rss;
1033                 int missing = 0;
1034                 unsigned int i;
1035
1036                 /* Stop at the first nonexistent target queue. */
1037                 for (i = 0; i != rss->queues; ++i)
1038                         if (rss->queue_id[i] >=
1039                             priv->dev->data->nb_rx_queues ||
1040                             !priv->dev->data->rx_queues[rss->queue_id[i]]) {
1041                                 missing = 1;
1042                                 break;
1043                         }
1044                 if (flow->ibv_flow) {
1045                         if (missing ^ !flow->drop)
1046                                 return 0;
1047                         /* Verbs flow needs updating. */
1048                         claim_zero(mlx4_glue->destroy_flow(flow->ibv_flow));
1049                         flow->ibv_flow = NULL;
1050                         if (flow->drop)
1051                                 mlx4_drop_put(priv->drop);
1052                         else
1053                                 mlx4_rss_detach(rss);
1054                 }
1055                 if (!missing) {
1056                         err = mlx4_rss_attach(rss);
1057                         if (err) {
1058                                 err = -err;
1059                                 msg = "cannot create indirection table or hash"
1060                                         " QP to associate flow rule with";
1061                                 goto error;
1062                         }
1063                         qp = rss->qp;
1064                 }
1065                 /* A missing target queue drops traffic implicitly. */
1066                 flow->drop = missing;
1067         }
1068         if (flow->drop) {
1069                 if (flow->ibv_flow)
1070                         return 0;
1071                 mlx4_drop_get(priv);
1072                 if (!priv->drop) {
1073                         err = rte_errno;
1074                         msg = "resources for drop flow rule cannot be created";
1075                         goto error;
1076                 }
1077                 qp = priv->drop->qp;
1078         }
1079         assert(qp);
1080         if (flow->ibv_flow)
1081                 return 0;
1082         flow->ibv_flow = mlx4_glue->create_flow(qp, flow->ibv_attr);
1083         if (flow->ibv_flow)
1084                 return 0;
1085         if (flow->drop)
1086                 mlx4_drop_put(priv->drop);
1087         else if (flow->rss)
1088                 mlx4_rss_detach(flow->rss);
1089         err = errno;
1090         msg = "flow rule rejected by device";
1091 error:
1092         return rte_flow_error_set
1093                 (error, err, RTE_FLOW_ERROR_TYPE_HANDLE, flow, msg);
1094 }
1095
1096 /**
1097  * Create a flow.
1098  *
1099  * @see rte_flow_create()
1100  * @see rte_flow_ops
1101  */
1102 static struct rte_flow *
1103 mlx4_flow_create(struct rte_eth_dev *dev,
1104                  const struct rte_flow_attr *attr,
1105                  const struct rte_flow_item pattern[],
1106                  const struct rte_flow_action actions[],
1107                  struct rte_flow_error *error)
1108 {
1109         struct priv *priv = dev->data->dev_private;
1110         struct rte_flow *flow;
1111         int err;
1112
1113         err = mlx4_flow_prepare(priv, attr, pattern, actions, error, &flow);
1114         if (err)
1115                 return NULL;
1116         err = mlx4_flow_toggle(priv, flow, priv->started, error);
1117         if (!err) {
1118                 struct rte_flow *curr = LIST_FIRST(&priv->flows);
1119
1120                 /* New rules are inserted after internal ones. */
1121                 if (!curr || !curr->internal) {
1122                         LIST_INSERT_HEAD(&priv->flows, flow, next);
1123                 } else {
1124                         while (LIST_NEXT(curr, next) &&
1125                                LIST_NEXT(curr, next)->internal)
1126                                 curr = LIST_NEXT(curr, next);
1127                         LIST_INSERT_AFTER(curr, flow, next);
1128                 }
1129                 return flow;
1130         }
1131         if (flow->rss)
1132                 mlx4_rss_put(flow->rss);
1133         rte_flow_error_set(error, -err, RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1134                            error->message);
1135         rte_free(flow);
1136         return NULL;
1137 }
1138
1139 /**
1140  * Configure isolated mode.
1141  *
1142  * @see rte_flow_isolate()
1143  * @see rte_flow_ops
1144  */
1145 static int
1146 mlx4_flow_isolate(struct rte_eth_dev *dev,
1147                   int enable,
1148                   struct rte_flow_error *error)
1149 {
1150         struct priv *priv = dev->data->dev_private;
1151
1152         if (!!enable == !!priv->isolated)
1153                 return 0;
1154         priv->isolated = !!enable;
1155         if (mlx4_flow_sync(priv, error)) {
1156                 priv->isolated = !enable;
1157                 return -rte_errno;
1158         }
1159         return 0;
1160 }
1161
1162 /**
1163  * Destroy a flow rule.
1164  *
1165  * @see rte_flow_destroy()
1166  * @see rte_flow_ops
1167  */
1168 static int
1169 mlx4_flow_destroy(struct rte_eth_dev *dev,
1170                   struct rte_flow *flow,
1171                   struct rte_flow_error *error)
1172 {
1173         struct priv *priv = dev->data->dev_private;
1174         int err = mlx4_flow_toggle(priv, flow, 0, error);
1175
1176         if (err)
1177                 return err;
1178         LIST_REMOVE(flow, next);
1179         if (flow->rss)
1180                 mlx4_rss_put(flow->rss);
1181         rte_free(flow);
1182         return 0;
1183 }
1184
1185 /**
1186  * Destroy user-configured flow rules.
1187  *
1188  * This function skips internal flows rules.
1189  *
1190  * @see rte_flow_flush()
1191  * @see rte_flow_ops
1192  */
1193 static int
1194 mlx4_flow_flush(struct rte_eth_dev *dev,
1195                 struct rte_flow_error *error)
1196 {
1197         struct priv *priv = dev->data->dev_private;
1198         struct rte_flow *flow = LIST_FIRST(&priv->flows);
1199
1200         while (flow) {
1201                 struct rte_flow *next = LIST_NEXT(flow, next);
1202
1203                 if (!flow->internal)
1204                         mlx4_flow_destroy(dev, flow, error);
1205                 flow = next;
1206         }
1207         return 0;
1208 }
1209
1210 /**
1211  * Helper function to determine the next configured VLAN filter.
1212  *
1213  * @param priv
1214  *   Pointer to private structure.
1215  * @param vlan
1216  *   VLAN ID to use as a starting point.
1217  *
1218  * @return
1219  *   Next configured VLAN ID or a high value (>= 4096) if there is none.
1220  */
1221 static uint16_t
1222 mlx4_flow_internal_next_vlan(struct priv *priv, uint16_t vlan)
1223 {
1224         while (vlan < 4096) {
1225                 if (priv->dev->data->vlan_filter_conf.ids[vlan / 64] &
1226                     (UINT64_C(1) << (vlan % 64)))
1227                         return vlan;
1228                 ++vlan;
1229         }
1230         return vlan;
1231 }
1232
1233 /**
1234  * Generate internal flow rules.
1235  *
1236  * Various flow rules are created depending on the mode the device is in:
1237  *
1238  * 1. Promiscuous:
1239  *       port MAC + broadcast + catch-all (VLAN filtering is ignored).
1240  * 2. All multicast:
1241  *       port MAC/VLAN + broadcast + catch-all multicast.
1242  * 3. Otherwise:
1243  *       port MAC/VLAN + broadcast MAC/VLAN.
1244  *
1245  * About MAC flow rules:
1246  *
1247  * - MAC flow rules are generated from @p dev->data->mac_addrs
1248  *   (@p priv->mac array).
1249  * - An additional flow rule for Ethernet broadcasts is also generated.
1250  * - All these are per-VLAN if @p DEV_RX_OFFLOAD_VLAN_FILTER
1251  *   is enabled and VLAN filters are configured.
1252  *
1253  * @param priv
1254  *   Pointer to private structure.
1255  * @param[out] error
1256  *   Perform verbose error reporting if not NULL.
1257  *
1258  * @return
1259  *   0 on success, a negative errno value otherwise and rte_errno is set.
1260  */
1261 static int
1262 mlx4_flow_internal(struct priv *priv, struct rte_flow_error *error)
1263 {
1264         struct rte_flow_attr attr = {
1265                 .priority = MLX4_FLOW_PRIORITY_LAST,
1266                 .ingress = 1,
1267         };
1268         struct rte_flow_item_eth eth_spec;
1269         const struct rte_flow_item_eth eth_mask = {
1270                 .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
1271         };
1272         const struct rte_flow_item_eth eth_allmulti = {
1273                 .dst.addr_bytes = "\x01\x00\x00\x00\x00\x00",
1274         };
1275         struct rte_flow_item_vlan vlan_spec;
1276         const struct rte_flow_item_vlan vlan_mask = {
1277                 .tci = RTE_BE16(0x0fff),
1278         };
1279         struct rte_flow_item pattern[] = {
1280                 {
1281                         .type = MLX4_FLOW_ITEM_TYPE_INTERNAL,
1282                 },
1283                 {
1284                         .type = RTE_FLOW_ITEM_TYPE_ETH,
1285                         .spec = &eth_spec,
1286                         .mask = &eth_mask,
1287                 },
1288                 {
1289                         /* Replaced with VLAN if filtering is enabled. */
1290                         .type = RTE_FLOW_ITEM_TYPE_END,
1291                 },
1292                 {
1293                         .type = RTE_FLOW_ITEM_TYPE_END,
1294                 },
1295         };
1296         /*
1297          * Round number of queues down to their previous power of 2 to
1298          * comply with RSS context limitations. Extra queues silently do not
1299          * get RSS by default.
1300          */
1301         uint32_t queues =
1302                 rte_align32pow2(priv->dev->data->nb_rx_queues + 1) >> 1;
1303         uint16_t queue[queues];
1304         struct rte_flow_action_rss action_rss = {
1305                 .func = RTE_ETH_HASH_FUNCTION_DEFAULT,
1306                 .level = 0,
1307                 .types = -1,
1308                 .key_len = MLX4_RSS_HASH_KEY_SIZE,
1309                 .queue_num = queues,
1310                 .key = mlx4_rss_hash_key_default,
1311                 .queue = queue,
1312         };
1313         struct rte_flow_action actions[] = {
1314                 {
1315                         .type = RTE_FLOW_ACTION_TYPE_RSS,
1316                         .conf = &action_rss,
1317                 },
1318                 {
1319                         .type = RTE_FLOW_ACTION_TYPE_END,
1320                 },
1321         };
1322         struct ether_addr *rule_mac = &eth_spec.dst;
1323         rte_be16_t *rule_vlan =
1324                 (priv->dev->data->dev_conf.rxmode.offloads &
1325                  DEV_RX_OFFLOAD_VLAN_FILTER) &&
1326                 !priv->dev->data->promiscuous ?
1327                 &vlan_spec.tci :
1328                 NULL;
1329         uint16_t vlan = 0;
1330         struct rte_flow *flow;
1331         unsigned int i;
1332         int err = 0;
1333
1334         /* Nothing to be done if there are no Rx queues. */
1335         if (!queues)
1336                 goto error;
1337         /* Prepare default RSS configuration. */
1338         for (i = 0; i != queues; ++i)
1339                 queue[i] = i;
1340         /*
1341          * Set up VLAN item if filtering is enabled and at least one VLAN
1342          * filter is configured.
1343          */
1344         if (rule_vlan) {
1345                 vlan = mlx4_flow_internal_next_vlan(priv, 0);
1346                 if (vlan < 4096) {
1347                         pattern[2] = (struct rte_flow_item){
1348                                 .type = RTE_FLOW_ITEM_TYPE_VLAN,
1349                                 .spec = &vlan_spec,
1350                                 .mask = &vlan_mask,
1351                         };
1352 next_vlan:
1353                         *rule_vlan = rte_cpu_to_be_16(vlan);
1354                 } else {
1355                         rule_vlan = NULL;
1356                 }
1357         }
1358         for (i = 0; i != RTE_DIM(priv->mac) + 1; ++i) {
1359                 const struct ether_addr *mac;
1360
1361                 /* Broadcasts are handled by an extra iteration. */
1362                 if (i < RTE_DIM(priv->mac))
1363                         mac = &priv->mac[i];
1364                 else
1365                         mac = &eth_mask.dst;
1366                 if (is_zero_ether_addr(mac))
1367                         continue;
1368                 /* Check if MAC flow rule is already present. */
1369                 for (flow = LIST_FIRST(&priv->flows);
1370                      flow && flow->internal;
1371                      flow = LIST_NEXT(flow, next)) {
1372                         const struct ibv_flow_spec_eth *eth =
1373                                 (const void *)((uintptr_t)flow->ibv_attr +
1374                                                sizeof(*flow->ibv_attr));
1375                         unsigned int j;
1376
1377                         if (!flow->mac)
1378                                 continue;
1379                         assert(flow->ibv_attr->type == IBV_FLOW_ATTR_NORMAL);
1380                         assert(flow->ibv_attr->num_of_specs == 1);
1381                         assert(eth->type == IBV_FLOW_SPEC_ETH);
1382                         assert(flow->rss);
1383                         if (rule_vlan &&
1384                             (eth->val.vlan_tag != *rule_vlan ||
1385                              eth->mask.vlan_tag != RTE_BE16(0x0fff)))
1386                                 continue;
1387                         if (!rule_vlan && eth->mask.vlan_tag)
1388                                 continue;
1389                         for (j = 0; j != sizeof(mac->addr_bytes); ++j)
1390                                 if (eth->val.dst_mac[j] != mac->addr_bytes[j] ||
1391                                     eth->mask.dst_mac[j] != UINT8_C(0xff) ||
1392                                     eth->val.src_mac[j] != UINT8_C(0x00) ||
1393                                     eth->mask.src_mac[j] != UINT8_C(0x00))
1394                                         break;
1395                         if (j != sizeof(mac->addr_bytes))
1396                                 continue;
1397                         if (flow->rss->queues != queues ||
1398                             memcmp(flow->rss->queue_id, action_rss.queue,
1399                                    queues * sizeof(flow->rss->queue_id[0])))
1400                                 continue;
1401                         break;
1402                 }
1403                 if (!flow || !flow->internal) {
1404                         /* Not found, create a new flow rule. */
1405                         memcpy(rule_mac, mac, sizeof(*mac));
1406                         flow = mlx4_flow_create(priv->dev, &attr, pattern,
1407                                                 actions, error);
1408                         if (!flow) {
1409                                 err = -rte_errno;
1410                                 goto error;
1411                         }
1412                 }
1413                 flow->select = 1;
1414                 flow->mac = 1;
1415         }
1416         if (rule_vlan) {
1417                 vlan = mlx4_flow_internal_next_vlan(priv, vlan + 1);
1418                 if (vlan < 4096)
1419                         goto next_vlan;
1420         }
1421         /* Take care of promiscuous and all multicast flow rules. */
1422         if (priv->dev->data->promiscuous || priv->dev->data->all_multicast) {
1423                 for (flow = LIST_FIRST(&priv->flows);
1424                      flow && flow->internal;
1425                      flow = LIST_NEXT(flow, next)) {
1426                         if (priv->dev->data->promiscuous) {
1427                                 if (flow->promisc)
1428                                         break;
1429                         } else {
1430                                 assert(priv->dev->data->all_multicast);
1431                                 if (flow->allmulti)
1432                                         break;
1433                         }
1434                 }
1435                 if (flow && flow->internal) {
1436                         assert(flow->rss);
1437                         if (flow->rss->queues != queues ||
1438                             memcmp(flow->rss->queue_id, action_rss.queue,
1439                                    queues * sizeof(flow->rss->queue_id[0])))
1440                                 flow = NULL;
1441                 }
1442                 if (!flow || !flow->internal) {
1443                         /* Not found, create a new flow rule. */
1444                         if (priv->dev->data->promiscuous) {
1445                                 pattern[1].spec = NULL;
1446                                 pattern[1].mask = NULL;
1447                         } else {
1448                                 assert(priv->dev->data->all_multicast);
1449                                 pattern[1].spec = &eth_allmulti;
1450                                 pattern[1].mask = &eth_allmulti;
1451                         }
1452                         pattern[2] = pattern[3];
1453                         flow = mlx4_flow_create(priv->dev, &attr, pattern,
1454                                                 actions, error);
1455                         if (!flow) {
1456                                 err = -rte_errno;
1457                                 goto error;
1458                         }
1459                 }
1460                 assert(flow->promisc || flow->allmulti);
1461                 flow->select = 1;
1462         }
1463 error:
1464         /* Clear selection and clean up stale internal flow rules. */
1465         flow = LIST_FIRST(&priv->flows);
1466         while (flow && flow->internal) {
1467                 struct rte_flow *next = LIST_NEXT(flow, next);
1468
1469                 if (!flow->select)
1470                         claim_zero(mlx4_flow_destroy(priv->dev, flow, error));
1471                 else
1472                         flow->select = 0;
1473                 flow = next;
1474         }
1475         return err;
1476 }
1477
1478 /**
1479  * Synchronize flow rules.
1480  *
1481  * This function synchronizes flow rules with the state of the device by
1482  * taking into account isolated mode and whether target queues are
1483  * configured.
1484  *
1485  * @param priv
1486  *   Pointer to private structure.
1487  * @param[out] error
1488  *   Perform verbose error reporting if not NULL.
1489  *
1490  * @return
1491  *   0 on success, a negative errno value otherwise and rte_errno is set.
1492  */
1493 int
1494 mlx4_flow_sync(struct priv *priv, struct rte_flow_error *error)
1495 {
1496         struct rte_flow *flow;
1497         int ret;
1498
1499         /* Internal flow rules are guaranteed to come first in the list. */
1500         if (priv->isolated) {
1501                 /*
1502                  * Get rid of them in isolated mode, stop at the first
1503                  * non-internal rule found.
1504                  */
1505                 for (flow = LIST_FIRST(&priv->flows);
1506                      flow && flow->internal;
1507                      flow = LIST_FIRST(&priv->flows))
1508                         claim_zero(mlx4_flow_destroy(priv->dev, flow, error));
1509         } else {
1510                 /* Refresh internal rules. */
1511                 ret = mlx4_flow_internal(priv, error);
1512                 if (ret)
1513                         return ret;
1514         }
1515         /* Toggle the remaining flow rules . */
1516         LIST_FOREACH(flow, &priv->flows, next) {
1517                 ret = mlx4_flow_toggle(priv, flow, priv->started, error);
1518                 if (ret)
1519                         return ret;
1520         }
1521         if (!priv->started)
1522                 assert(!priv->drop);
1523         return 0;
1524 }
1525
1526 /**
1527  * Clean up all flow rules.
1528  *
1529  * Unlike mlx4_flow_flush(), this function takes care of all remaining flow
1530  * rules regardless of whether they are internal or user-configured.
1531  *
1532  * @param priv
1533  *   Pointer to private structure.
1534  */
1535 void
1536 mlx4_flow_clean(struct priv *priv)
1537 {
1538         struct rte_flow *flow;
1539
1540         while ((flow = LIST_FIRST(&priv->flows)))
1541                 mlx4_flow_destroy(priv->dev, flow, NULL);
1542         assert(LIST_EMPTY(&priv->rss));
1543 }
1544
1545 static const struct rte_flow_ops mlx4_flow_ops = {
1546         .validate = mlx4_flow_validate,
1547         .create = mlx4_flow_create,
1548         .destroy = mlx4_flow_destroy,
1549         .flush = mlx4_flow_flush,
1550         .isolate = mlx4_flow_isolate,
1551 };
1552
1553 /**
1554  * Manage filter operations.
1555  *
1556  * @param dev
1557  *   Pointer to Ethernet device structure.
1558  * @param filter_type
1559  *   Filter type.
1560  * @param filter_op
1561  *   Operation to perform.
1562  * @param arg
1563  *   Pointer to operation-specific structure.
1564  *
1565  * @return
1566  *   0 on success, negative errno value otherwise and rte_errno is set.
1567  */
1568 int
1569 mlx4_filter_ctrl(struct rte_eth_dev *dev,
1570                  enum rte_filter_type filter_type,
1571                  enum rte_filter_op filter_op,
1572                  void *arg)
1573 {
1574         switch (filter_type) {
1575         case RTE_ETH_FILTER_GENERIC:
1576                 if (filter_op != RTE_ETH_FILTER_GET)
1577                         break;
1578                 *(const void **)arg = &mlx4_flow_ops;
1579                 return 0;
1580         default:
1581                 ERROR("%p: filter type (%d) not supported",
1582                       (void *)dev, filter_type);
1583                 break;
1584         }
1585         rte_errno = ENOTSUP;
1586         return -rte_errno;
1587 }