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