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 fields to their Verbs equivalent.
  80  *
  81  * This function returns the supported (default) set when @p rss_hf has
  82  * special value (uint64_t)-1.
  83  *
  84  * @param priv
  85  *   Pointer to private structure.
  86  * @param rss_hf
  87  *   Hash fields 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_hf(struct priv *priv, uint64_t rss_hf)
  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 (rss_hf & in[i]) {
 130                         seen |= rss_hf & in[i];
 131                         conv |= out[i];
 132                 }
 133         if ((conv & priv->hw_rss_sup) == conv) {
 134                 if (rss_hf == (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 (!(rss_hf & ~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
 641         if (attr->group)
 642                 return rte_flow_error_set
 643                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 644                          NULL, "groups are not supported");
 645         if (attr->priority > MLX4_FLOW_PRIORITY_LAST)
 646                 return rte_flow_error_set
 647                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 648                          NULL, "maximum priority level is "
 649                          MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST));
 650         if (attr->egress)
 651                 return rte_flow_error_set
 652                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 653                          NULL, "egress is not supported");
 654         if (!attr->ingress)
 655                 return rte_flow_error_set
 656                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 657                          NULL, "only ingress is supported");
 658 fill:
 659         proc = mlx4_flow_proc_item_list;
 660         /* Go over pattern. */
 661         for (item = pattern; item->type; ++item) {
 662                 const struct mlx4_flow_proc_item *next = NULL;
 663                 unsigned int i;
 664                 int err;
 665
 666                 if (item->type == RTE_FLOW_ITEM_TYPE_VOID)
 667                         continue;
 668                 if (item->type == MLX4_FLOW_ITEM_TYPE_INTERNAL) {
 669                         flow->internal = 1;
 670                         continue;
 671                 }
 672                 if (flow->promisc || flow->allmulti) {
 673                         msg = "mlx4 does not support additional matching"
 674                                 " criteria combined with indiscriminate"
 675                                 " matching on Ethernet headers";
 676                         goto exit_item_not_supported;
 677                 }
 678                 for (i = 0; proc->next_item && proc->next_item[i]; ++i) {
 679                         if (proc->next_item[i] == item->type) {
 680                                 next = &mlx4_flow_proc_item_list[item->type];
 681                                 break;
 682                         }
 683                 }
 684                 if (!next)
 685                         goto exit_item_not_supported;
 686                 proc = next;
 687                 /*
 688                  * Perform basic sanity checks only once, while handle is
 689                  * not allocated.
 690                  */
 691                 if (flow == &temp) {
 692                         err = mlx4_flow_item_check(item, proc, error);
 693                         if (err)
 694                                 return err;
 695                 }
 696                 if (proc->merge) {
 697                         err = proc->merge(flow, item, proc, error);
 698                         if (err)
 699                                 return err;
 700                 }
 701                 flow->ibv_attr_size += proc->dst_sz;
 702         }
 703         /* Go over actions list. */
 704         for (action = actions; action->type; ++action) {
 705                 switch (action->type) {
 706                         const struct rte_flow_action_queue *queue;
 707                         const struct rte_flow_action_rss *rss;
 708                         const struct rte_eth_rss_conf *rss_conf;
 709                         uint64_t fields;
 710                         unsigned int i;
 711
 712                 case RTE_FLOW_ACTION_TYPE_VOID:
 713                         continue;
 714                 case RTE_FLOW_ACTION_TYPE_DROP:
 715                         flow->drop = 1;
 716                         break;
 717                 case RTE_FLOW_ACTION_TYPE_QUEUE:
 718                         if (flow->rss)
 719                                 break;
 720                         queue = action->conf;
 721                         if (queue->index >= priv->dev->data->nb_rx_queues) {
 722                                 msg = "queue target index beyond number of"
 723                                         " configured Rx queues";
 724                                 goto exit_action_not_supported;
 725                         }
 726                         flow->rss = mlx4_rss_get
 727                                 (priv, 0, mlx4_rss_hash_key_default, 1,
 728                                  &queue->index);
 729                         if (!flow->rss) {
 730                                 msg = "not enough resources for additional"
 731                                         " single-queue RSS context";
 732                                 goto exit_action_not_supported;
 733                         }
 734                         break;
 735                 case RTE_FLOW_ACTION_TYPE_RSS:
 736                         if (flow->rss)
 737                                 break;
 738                         rss = action->conf;
 739                         /* Default RSS configuration if none is provided. */
 740                         rss_conf =
 741                                 rss->rss_conf ?
 742                                 rss->rss_conf :
 743                                 &(struct rte_eth_rss_conf){
 744                                         .rss_key = mlx4_rss_hash_key_default,
 745                                         .rss_key_len = MLX4_RSS_HASH_KEY_SIZE,
 746                                         .rss_hf = -1,
 747                                 };
 748                         /* Sanity checks. */
 749                         for (i = 0; i < rss->num; ++i)
 750                                 if (rss->queue[i] >=
 751                                     priv->dev->data->nb_rx_queues)
 752                                         break;
 753                         if (i != rss->num) {
 754                                 msg = "queue index target beyond number of"
 755                                         " configured Rx queues";
 756                                 goto exit_action_not_supported;
 757                         }
 758                         if (!rte_is_power_of_2(rss->num)) {
 759                                 msg = "for RSS, mlx4 requires the number of"
 760                                         " queues to be a power of two";
 761                                 goto exit_action_not_supported;
 762                         }
 763                         if (rss_conf->rss_key_len !=
 764                             sizeof(flow->rss->key)) {
 765                                 msg = "mlx4 supports exactly one RSS hash key"
 766                                         " length: "
 767                                         MLX4_STR_EXPAND(MLX4_RSS_HASH_KEY_SIZE);
 768                                 goto exit_action_not_supported;
 769                         }
 770                         for (i = 1; i < rss->num; ++i)
 771                                 if (rss->queue[i] - rss->queue[i - 1] != 1)
 772                                         break;
 773                         if (i != rss->num) {
 774                                 msg = "mlx4 requires RSS contexts to use"
 775                                         " consecutive queue indices only";
 776                                 goto exit_action_not_supported;
 777                         }
 778                         if (rss->queue[0] % rss->num) {
 779                                 msg = "mlx4 requires the first queue of a RSS"
 780                                         " context to be aligned on a multiple"
 781                                         " of the context size";
 782                                 goto exit_action_not_supported;
 783                         }
 784                         rte_errno = 0;
 785                         fields = mlx4_conv_rss_hf(priv, rss_conf->rss_hf);
 786                         if (fields == (uint64_t)-1 && rte_errno) {
 787                                 msg = "unsupported RSS hash type requested";
 788                                 goto exit_action_not_supported;
 789                         }
 790                         flow->rss = mlx4_rss_get
 791                                 (priv, fields, rss_conf->rss_key, rss->num,
 792                                  rss->queue);
 793                         if (!flow->rss) {
 794                                 msg = "either invalid parameters or not enough"
 795                                         " resources for additional multi-queue"
 796                                         " RSS context";
 797                                 goto exit_action_not_supported;
 798                         }
 799                         break;
 800                 default:
 801                         goto exit_action_not_supported;
 802                 }
 803         }
 804         if (!flow->rss && !flow->drop)
 805                 return rte_flow_error_set
 806                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 807                          NULL, "no valid action");
 808         /* Validation ends here. */
 809         if (!addr) {
 810                 if (flow->rss)
 811                         mlx4_rss_put(flow->rss);
 812                 return 0;
 813         }
 814         if (flow == &temp) {
 815                 /* Allocate proper handle based on collected data. */
 816                 const struct mlx4_malloc_vec vec[] = {
 817                         {
 818                                 .align = alignof(struct rte_flow),
 819                                 .size = sizeof(*flow),
 820                                 .addr = (void **)&flow,
 821                         },
 822                         {
 823                                 .align = alignof(struct ibv_flow_attr),
 824                                 .size = temp.ibv_attr_size,
 825                                 .addr = (void **)&temp.ibv_attr,
 826                         },
 827                 };
 828
 829                 if (!mlx4_zmallocv(__func__, vec, RTE_DIM(vec))) {
 830                         if (temp.rss)
 831                                 mlx4_rss_put(temp.rss);
 832                         return rte_flow_error_set
 833                                 (error, -rte_errno,
 834                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
 835                                  "flow rule handle allocation failure");
 836                 }
 837                 /* Most fields will be updated by second pass. */
 838                 *flow = (struct rte_flow){
 839                         .ibv_attr = temp.ibv_attr,
 840                         .ibv_attr_size = sizeof(*flow->ibv_attr),
 841                         .rss = temp.rss,
 842                 };
 843                 *flow->ibv_attr = (struct ibv_flow_attr){
 844                         .type = IBV_FLOW_ATTR_NORMAL,
 845                         .size = sizeof(*flow->ibv_attr),
 846                         .priority = attr->priority,
 847                         .port = priv->port,
 848                 };
 849                 goto fill;
 850         }
 851         *addr = flow;
 852         return 0;
 853 exit_item_not_supported:
 854         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 855                                   item, msg ? msg : "item not supported");
 856 exit_action_not_supported:
 857         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 858                                   action, msg ? msg : "action not supported");
 859 }
 860
 861 /**
 862  * Validate a flow supported by the NIC.
 863  *
 864  * @see rte_flow_validate()
 865  * @see rte_flow_ops
 866  */
 867 static int
 868 mlx4_flow_validate(struct rte_eth_dev *dev,
 869                    const struct rte_flow_attr *attr,
 870                    const struct rte_flow_item pattern[],
 871                    const struct rte_flow_action actions[],
 872                    struct rte_flow_error *error)
 873 {
 874         struct priv *priv = dev->data->dev_private;
 875
 876         return mlx4_flow_prepare(priv, attr, pattern, actions, error, NULL);
 877 }
 878
 879 /**
 880  * Get a drop flow rule resources instance.
 881  *
 882  * @param priv
 883  *   Pointer to private structure.
 884  *
 885  * @return
 886  *   Pointer to drop flow resources on success, NULL otherwise and rte_errno
 887  *   is set.
 888  */
 889 static struct mlx4_drop *
 890 mlx4_drop_get(struct priv *priv)
 891 {
 892         struct mlx4_drop *drop = priv->drop;
 893
 894         if (drop) {
 895                 assert(drop->refcnt);
 896                 assert(drop->priv == priv);
 897                 ++drop->refcnt;
 898                 return drop;
 899         }
 900         drop = rte_malloc(__func__, sizeof(*drop), 0);
 901         if (!drop)
 902                 goto error;
 903         *drop = (struct mlx4_drop){
 904                 .priv = priv,
 905                 .refcnt = 1,
 906         };
 907         drop->cq = mlx4_glue->create_cq(priv->ctx, 1, NULL, NULL, 0);
 908         if (!drop->cq)
 909                 goto error;
 910         drop->qp = mlx4_glue->create_qp
 911                 (priv->pd,
 912                  &(struct ibv_qp_init_attr){
 913                         .send_cq = drop->cq,
 914                         .recv_cq = drop->cq,
 915                         .qp_type = IBV_QPT_RAW_PACKET,
 916                  });
 917         if (!drop->qp)
 918                 goto error;
 919         priv->drop = drop;
 920         return drop;
 921 error:
 922         if (drop->qp)
 923                 claim_zero(mlx4_glue->destroy_qp(drop->qp));
 924         if (drop->cq)
 925                 claim_zero(mlx4_glue->destroy_cq(drop->cq));
 926         if (drop)
 927                 rte_free(drop);
 928         rte_errno = ENOMEM;
 929         return NULL;
 930 }
 931
 932 /**
 933  * Give back a drop flow rule resources instance.
 934  *
 935  * @param drop
 936  *   Pointer to drop flow rule resources.
 937  */
 938 static void
 939 mlx4_drop_put(struct mlx4_drop *drop)
 940 {
 941         assert(drop->refcnt);
 942         if (--drop->refcnt)
 943                 return;
 944         drop->priv->drop = NULL;
 945         claim_zero(mlx4_glue->destroy_qp(drop->qp));
 946         claim_zero(mlx4_glue->destroy_cq(drop->cq));
 947         rte_free(drop);
 948 }
 949
 950 /**
 951  * Toggle a configured flow rule.
 952  *
 953  * @param priv
 954  *   Pointer to private structure.
 955  * @param flow
 956  *   Flow rule handle to toggle.
 957  * @param enable
 958  *   Whether associated Verbs flow must be created or removed.
 959  * @param[out] error
 960  *   Perform verbose error reporting if not NULL.
 961  *
 962  * @return
 963  *   0 on success, a negative errno value otherwise and rte_errno is set.
 964  */
 965 static int
 966 mlx4_flow_toggle(struct priv *priv,
 967                  struct rte_flow *flow,
 968                  int enable,
 969                  struct rte_flow_error *error)
 970 {
 971         struct ibv_qp *qp = NULL;
 972         const char *msg;
 973         int err;
 974
 975         if (!enable) {
 976                 if (!flow->ibv_flow)
 977                         return 0;
 978                 claim_zero(mlx4_glue->destroy_flow(flow->ibv_flow));
 979                 flow->ibv_flow = NULL;
 980                 if (flow->drop)
 981                         mlx4_drop_put(priv->drop);
 982                 else if (flow->rss)
 983                         mlx4_rss_detach(flow->rss);
 984                 return 0;
 985         }
 986         assert(flow->ibv_attr);
 987         if (!flow->internal &&
 988             !priv->isolated &&
 989             flow->ibv_attr->priority == MLX4_FLOW_PRIORITY_LAST) {
 990                 if (flow->ibv_flow) {
 991                         claim_zero(mlx4_glue->destroy_flow(flow->ibv_flow));
 992                         flow->ibv_flow = NULL;
 993                         if (flow->drop)
 994                                 mlx4_drop_put(priv->drop);
 995                         else if (flow->rss)
 996                                 mlx4_rss_detach(flow->rss);
 997                 }
 998                 err = EACCES;
 999                 msg = ("priority level "
1000                        MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST)
1001                        " is reserved when not in isolated mode");
1002                 goto error;
1003         }
1004         if (flow->rss) {
1005                 struct mlx4_rss *rss = flow->rss;
1006                 int missing = 0;
1007                 unsigned int i;
1008
1009                 /* Stop at the first nonexistent target queue. */
1010                 for (i = 0; i != rss->queues; ++i)
1011                         if (rss->queue_id[i] >=
1012                             priv->dev->data->nb_rx_queues ||
1013                             !priv->dev->data->rx_queues[rss->queue_id[i]]) {
1014                                 missing = 1;
1015                                 break;
1016                         }
1017                 if (flow->ibv_flow) {
1018                         if (missing ^ !flow->drop)
1019                                 return 0;
1020                         /* Verbs flow needs updating. */
1021                         claim_zero(mlx4_glue->destroy_flow(flow->ibv_flow));
1022                         flow->ibv_flow = NULL;
1023                         if (flow->drop)
1024                                 mlx4_drop_put(priv->drop);
1025                         else
1026                                 mlx4_rss_detach(rss);
1027                 }
1028                 if (!missing) {
1029                         err = mlx4_rss_attach(rss);
1030                         if (err) {
1031                                 err = -err;
1032                                 msg = "cannot create indirection table or hash"
1033                                         " QP to associate flow rule with";
1034                                 goto error;
1035                         }
1036                         qp = rss->qp;
1037                 }
1038                 /* A missing target queue drops traffic implicitly. */
1039                 flow->drop = missing;
1040         }
1041         if (flow->drop) {
1042                 if (flow->ibv_flow)
1043                         return 0;
1044                 mlx4_drop_get(priv);
1045                 if (!priv->drop) {
1046                         err = rte_errno;
1047                         msg = "resources for drop flow rule cannot be created";
1048                         goto error;
1049                 }
1050                 qp = priv->drop->qp;
1051         }
1052         assert(qp);
1053         if (flow->ibv_flow)
1054                 return 0;
1055         flow->ibv_flow = mlx4_glue->create_flow(qp, flow->ibv_attr);
1056         if (flow->ibv_flow)
1057                 return 0;
1058         if (flow->drop)
1059                 mlx4_drop_put(priv->drop);
1060         else if (flow->rss)
1061                 mlx4_rss_detach(flow->rss);
1062         err = errno;
1063         msg = "flow rule rejected by device";
1064 error:
1065         return rte_flow_error_set
1066                 (error, err, RTE_FLOW_ERROR_TYPE_HANDLE, flow, msg);
1067 }
1068
1069 /**
1070  * Create a flow.
1071  *
1072  * @see rte_flow_create()
1073  * @see rte_flow_ops
1074  */
1075 static struct rte_flow *
1076 mlx4_flow_create(struct rte_eth_dev *dev,
1077                  const struct rte_flow_attr *attr,
1078                  const struct rte_flow_item pattern[],
1079                  const struct rte_flow_action actions[],
1080                  struct rte_flow_error *error)
1081 {
1082         struct priv *priv = dev->data->dev_private;
1083         struct rte_flow *flow;
1084         int err;
1085
1086         err = mlx4_flow_prepare(priv, attr, pattern, actions, error, &flow);
1087         if (err)
1088                 return NULL;
1089         err = mlx4_flow_toggle(priv, flow, priv->started, error);
1090         if (!err) {
1091                 struct rte_flow *curr = LIST_FIRST(&priv->flows);
1092
1093                 /* New rules are inserted after internal ones. */
1094                 if (!curr || !curr->internal) {
1095                         LIST_INSERT_HEAD(&priv->flows, flow, next);
1096                 } else {
1097                         while (LIST_NEXT(curr, next) &&
1098                                LIST_NEXT(curr, next)->internal)
1099                                 curr = LIST_NEXT(curr, next);
1100                         LIST_INSERT_AFTER(curr, flow, next);
1101                 }
1102                 return flow;
1103         }
1104         if (flow->rss)
1105                 mlx4_rss_put(flow->rss);
1106         rte_flow_error_set(error, -err, RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1107                            error->message);
1108         rte_free(flow);
1109         return NULL;
1110 }
1111
1112 /**
1113  * Configure isolated mode.
1114  *
1115  * @see rte_flow_isolate()
1116  * @see rte_flow_ops
1117  */
1118 static int
1119 mlx4_flow_isolate(struct rte_eth_dev *dev,
1120                   int enable,
1121                   struct rte_flow_error *error)
1122 {
1123         struct priv *priv = dev->data->dev_private;
1124
1125         if (!!enable == !!priv->isolated)
1126                 return 0;
1127         priv->isolated = !!enable;
1128         if (mlx4_flow_sync(priv, error)) {
1129                 priv->isolated = !enable;
1130                 return -rte_errno;
1131         }
1132         return 0;
1133 }
1134
1135 /**
1136  * Destroy a flow rule.
1137  *
1138  * @see rte_flow_destroy()
1139  * @see rte_flow_ops
1140  */
1141 static int
1142 mlx4_flow_destroy(struct rte_eth_dev *dev,
1143                   struct rte_flow *flow,
1144                   struct rte_flow_error *error)
1145 {
1146         struct priv *priv = dev->data->dev_private;
1147         int err = mlx4_flow_toggle(priv, flow, 0, error);
1148
1149         if (err)
1150                 return err;
1151         LIST_REMOVE(flow, next);
1152         if (flow->rss)
1153                 mlx4_rss_put(flow->rss);
1154         rte_free(flow);
1155         return 0;
1156 }
1157
1158 /**
1159  * Destroy user-configured flow rules.
1160  *
1161  * This function skips internal flows rules.
1162  *
1163  * @see rte_flow_flush()
1164  * @see rte_flow_ops
1165  */
1166 static int
1167 mlx4_flow_flush(struct rte_eth_dev *dev,
1168                 struct rte_flow_error *error)
1169 {
1170         struct priv *priv = dev->data->dev_private;
1171         struct rte_flow *flow = LIST_FIRST(&priv->flows);
1172
1173         while (flow) {
1174                 struct rte_flow *next = LIST_NEXT(flow, next);
1175
1176                 if (!flow->internal)
1177                         mlx4_flow_destroy(dev, flow, error);
1178                 flow = next;
1179         }
1180         return 0;
1181 }
1182
1183 /**
1184  * Helper function to determine the next configured VLAN filter.
1185  *
1186  * @param priv
1187  *   Pointer to private structure.
1188  * @param vlan
1189  *   VLAN ID to use as a starting point.
1190  *
1191  * @return
1192  *   Next configured VLAN ID or a high value (>= 4096) if there is none.
1193  */
1194 static uint16_t
1195 mlx4_flow_internal_next_vlan(struct priv *priv, uint16_t vlan)
1196 {
1197         while (vlan < 4096) {
1198                 if (priv->dev->data->vlan_filter_conf.ids[vlan / 64] &
1199                     (UINT64_C(1) << (vlan % 64)))
1200                         return vlan;
1201                 ++vlan;
1202         }
1203         return vlan;
1204 }
1205
1206 /**
1207  * Generate internal flow rules.
1208  *
1209  * Various flow rules are created depending on the mode the device is in:
1210  *
1211  * 1. Promiscuous:
1212  *       port MAC + broadcast + catch-all (VLAN filtering is ignored).
1213  * 2. All multicast:
1214  *       port MAC/VLAN + broadcast + catch-all multicast.
1215  * 3. Otherwise:
1216  *       port MAC/VLAN + broadcast MAC/VLAN.
1217  *
1218  * About MAC flow rules:
1219  *
1220  * - MAC flow rules are generated from @p dev->data->mac_addrs
1221  *   (@p priv->mac array).
1222  * - An additional flow rule for Ethernet broadcasts is also generated.
1223  * - All these are per-VLAN if @p DEV_RX_OFFLOAD_VLAN_FILTER
1224  *   is enabled and VLAN filters are configured.
1225  *
1226  * @param priv
1227  *   Pointer to private structure.
1228  * @param[out] error
1229  *   Perform verbose error reporting if not NULL.
1230  *
1231  * @return
1232  *   0 on success, a negative errno value otherwise and rte_errno is set.
1233  */
1234 static int
1235 mlx4_flow_internal(struct priv *priv, struct rte_flow_error *error)
1236 {
1237         struct rte_flow_attr attr = {
1238                 .priority = MLX4_FLOW_PRIORITY_LAST,
1239                 .ingress = 1,
1240         };
1241         struct rte_flow_item_eth eth_spec;
1242         const struct rte_flow_item_eth eth_mask = {
1243                 .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
1244         };
1245         const struct rte_flow_item_eth eth_allmulti = {
1246                 .dst.addr_bytes = "\x01\x00\x00\x00\x00\x00",
1247         };
1248         struct rte_flow_item_vlan vlan_spec;
1249         const struct rte_flow_item_vlan vlan_mask = {
1250                 .tci = RTE_BE16(0x0fff),
1251         };
1252         struct rte_flow_item pattern[] = {
1253                 {
1254                         .type = MLX4_FLOW_ITEM_TYPE_INTERNAL,
1255                 },
1256                 {
1257                         .type = RTE_FLOW_ITEM_TYPE_ETH,
1258                         .spec = &eth_spec,
1259                         .mask = &eth_mask,
1260                 },
1261                 {
1262                         /* Replaced with VLAN if filtering is enabled. */
1263                         .type = RTE_FLOW_ITEM_TYPE_END,
1264                 },
1265                 {
1266                         .type = RTE_FLOW_ITEM_TYPE_END,
1267                 },
1268         };
1269         /*
1270          * Round number of queues down to their previous power of 2 to
1271          * comply with RSS context limitations. Extra queues silently do not
1272          * get RSS by default.
1273          */
1274         uint32_t queues =
1275                 rte_align32pow2(priv->dev->data->nb_rx_queues + 1) >> 1;
1276         alignas(struct rte_flow_action_rss) uint8_t rss_conf_data
1277                 [offsetof(struct rte_flow_action_rss, queue) +
1278                  sizeof(((struct rte_flow_action_rss *)0)->queue[0]) * queues];
1279         struct rte_flow_action_rss *rss_conf = (void *)rss_conf_data;
1280         struct rte_flow_action actions[] = {
1281                 {
1282                         .type = RTE_FLOW_ACTION_TYPE_RSS,
1283                         .conf = rss_conf,
1284                 },
1285                 {
1286                         .type = RTE_FLOW_ACTION_TYPE_END,
1287                 },
1288         };
1289         struct ether_addr *rule_mac = &eth_spec.dst;
1290         rte_be16_t *rule_vlan =
1291                 (priv->dev->data->dev_conf.rxmode.offloads &
1292                  DEV_RX_OFFLOAD_VLAN_FILTER) &&
1293                 !priv->dev->data->promiscuous ?
1294                 &vlan_spec.tci :
1295                 NULL;
1296         uint16_t vlan = 0;
1297         struct rte_flow *flow;
1298         unsigned int i;
1299         int err = 0;
1300
1301         /* Nothing to be done if there are no Rx queues. */
1302         if (!queues)
1303                 goto error;
1304         /* Prepare default RSS configuration. */
1305         *rss_conf = (struct rte_flow_action_rss){
1306                 .rss_conf = NULL, /* Rely on default fallback settings. */
1307                 .num = queues,
1308         };
1309         for (i = 0; i != queues; ++i)
1310                 rss_conf->queue[i] = i;
1311         /*
1312          * Set up VLAN item if filtering is enabled and at least one VLAN
1313          * filter is configured.
1314          */
1315         if (rule_vlan) {
1316                 vlan = mlx4_flow_internal_next_vlan(priv, 0);
1317                 if (vlan < 4096) {
1318                         pattern[2] = (struct rte_flow_item){
1319                                 .type = RTE_FLOW_ITEM_TYPE_VLAN,
1320                                 .spec = &vlan_spec,
1321                                 .mask = &vlan_mask,
1322                         };
1323 next_vlan:
1324                         *rule_vlan = rte_cpu_to_be_16(vlan);
1325                 } else {
1326                         rule_vlan = NULL;
1327                 }
1328         }
1329         for (i = 0; i != RTE_DIM(priv->mac) + 1; ++i) {
1330                 const struct ether_addr *mac;
1331
1332                 /* Broadcasts are handled by an extra iteration. */
1333                 if (i < RTE_DIM(priv->mac))
1334                         mac = &priv->mac[i];
1335                 else
1336                         mac = &eth_mask.dst;
1337                 if (is_zero_ether_addr(mac))
1338                         continue;
1339                 /* Check if MAC flow rule is already present. */
1340                 for (flow = LIST_FIRST(&priv->flows);
1341                      flow && flow->internal;
1342                      flow = LIST_NEXT(flow, next)) {
1343                         const struct ibv_flow_spec_eth *eth =
1344                                 (const void *)((uintptr_t)flow->ibv_attr +
1345                                                sizeof(*flow->ibv_attr));
1346                         unsigned int j;
1347
1348                         if (!flow->mac)
1349                                 continue;
1350                         assert(flow->ibv_attr->type == IBV_FLOW_ATTR_NORMAL);
1351                         assert(flow->ibv_attr->num_of_specs == 1);
1352                         assert(eth->type == IBV_FLOW_SPEC_ETH);
1353                         assert(flow->rss);
1354                         if (rule_vlan &&
1355                             (eth->val.vlan_tag != *rule_vlan ||
1356                              eth->mask.vlan_tag != RTE_BE16(0x0fff)))
1357                                 continue;
1358                         if (!rule_vlan && eth->mask.vlan_tag)
1359                                 continue;
1360                         for (j = 0; j != sizeof(mac->addr_bytes); ++j)
1361                                 if (eth->val.dst_mac[j] != mac->addr_bytes[j] ||
1362                                     eth->mask.dst_mac[j] != UINT8_C(0xff) ||
1363                                     eth->val.src_mac[j] != UINT8_C(0x00) ||
1364                                     eth->mask.src_mac[j] != UINT8_C(0x00))
1365                                         break;
1366                         if (j != sizeof(mac->addr_bytes))
1367                                 continue;
1368                         if (flow->rss->queues != queues ||
1369                             memcmp(flow->rss->queue_id, rss_conf->queue,
1370                                    queues * sizeof(flow->rss->queue_id[0])))
1371                                 continue;
1372                         break;
1373                 }
1374                 if (!flow || !flow->internal) {
1375                         /* Not found, create a new flow rule. */
1376                         memcpy(rule_mac, mac, sizeof(*mac));
1377                         flow = mlx4_flow_create(priv->dev, &attr, pattern,
1378                                                 actions, error);
1379                         if (!flow) {
1380                                 err = -rte_errno;
1381                                 goto error;
1382                         }
1383                 }
1384                 flow->select = 1;
1385                 flow->mac = 1;
1386         }
1387         if (rule_vlan) {
1388                 vlan = mlx4_flow_internal_next_vlan(priv, vlan + 1);
1389                 if (vlan < 4096)
1390                         goto next_vlan;
1391         }
1392         /* Take care of promiscuous and all multicast flow rules. */
1393         if (priv->dev->data->promiscuous || priv->dev->data->all_multicast) {
1394                 for (flow = LIST_FIRST(&priv->flows);
1395                      flow && flow->internal;
1396                      flow = LIST_NEXT(flow, next)) {
1397                         if (priv->dev->data->promiscuous) {
1398                                 if (flow->promisc)
1399                                         break;
1400                         } else {
1401                                 assert(priv->dev->data->all_multicast);
1402                                 if (flow->allmulti)
1403                                         break;
1404                         }
1405                 }
1406                 if (flow && flow->internal) {
1407                         assert(flow->rss);
1408                         if (flow->rss->queues != queues ||
1409                             memcmp(flow->rss->queue_id, rss_conf->queue,
1410                                    queues * sizeof(flow->rss->queue_id[0])))
1411                                 flow = NULL;
1412                 }
1413                 if (!flow || !flow->internal) {
1414                         /* Not found, create a new flow rule. */
1415                         if (priv->dev->data->promiscuous) {
1416                                 pattern[1].spec = NULL;
1417                                 pattern[1].mask = NULL;
1418                         } else {
1419                                 assert(priv->dev->data->all_multicast);
1420                                 pattern[1].spec = &eth_allmulti;
1421                                 pattern[1].mask = &eth_allmulti;
1422                         }
1423                         pattern[2] = pattern[3];
1424                         flow = mlx4_flow_create(priv->dev, &attr, pattern,
1425                                                 actions, error);
1426                         if (!flow) {
1427                                 err = -rte_errno;
1428                                 goto error;
1429                         }
1430                 }
1431                 assert(flow->promisc || flow->allmulti);
1432                 flow->select = 1;
1433         }
1434 error:
1435         /* Clear selection and clean up stale internal flow rules. */
1436         flow = LIST_FIRST(&priv->flows);
1437         while (flow && flow->internal) {
1438                 struct rte_flow *next = LIST_NEXT(flow, next);
1439
1440                 if (!flow->select)
1441                         claim_zero(mlx4_flow_destroy(priv->dev, flow, error));
1442                 else
1443                         flow->select = 0;
1444                 flow = next;
1445         }
1446         return err;
1447 }
1448
1449 /**
1450  * Synchronize flow rules.
1451  *
1452  * This function synchronizes flow rules with the state of the device by
1453  * taking into account isolated mode and whether target queues are
1454  * configured.
1455  *
1456  * @param priv
1457  *   Pointer to private structure.
1458  * @param[out] error
1459  *   Perform verbose error reporting if not NULL.
1460  *
1461  * @return
1462  *   0 on success, a negative errno value otherwise and rte_errno is set.
1463  */
1464 int
1465 mlx4_flow_sync(struct priv *priv, struct rte_flow_error *error)
1466 {
1467         struct rte_flow *flow;
1468         int ret;
1469
1470         /* Internal flow rules are guaranteed to come first in the list. */
1471         if (priv->isolated) {
1472                 /*
1473                  * Get rid of them in isolated mode, stop at the first
1474                  * non-internal rule found.
1475                  */
1476                 for (flow = LIST_FIRST(&priv->flows);
1477                      flow && flow->internal;
1478                      flow = LIST_FIRST(&priv->flows))
1479                         claim_zero(mlx4_flow_destroy(priv->dev, flow, error));
1480         } else {
1481                 /* Refresh internal rules. */
1482                 ret = mlx4_flow_internal(priv, error);
1483                 if (ret)
1484                         return ret;
1485         }
1486         /* Toggle the remaining flow rules . */
1487         LIST_FOREACH(flow, &priv->flows, next) {
1488                 ret = mlx4_flow_toggle(priv, flow, priv->started, error);
1489                 if (ret)
1490                         return ret;
1491         }
1492         if (!priv->started)
1493                 assert(!priv->drop);
1494         return 0;
1495 }
1496
1497 /**
1498  * Clean up all flow rules.
1499  *
1500  * Unlike mlx4_flow_flush(), this function takes care of all remaining flow
1501  * rules regardless of whether they are internal or user-configured.
1502  *
1503  * @param priv
1504  *   Pointer to private structure.
1505  */
1506 void
1507 mlx4_flow_clean(struct priv *priv)
1508 {
1509         struct rte_flow *flow;
1510
1511         while ((flow = LIST_FIRST(&priv->flows)))
1512                 mlx4_flow_destroy(priv->dev, flow, NULL);
1513         assert(LIST_EMPTY(&priv->rss));
1514 }
1515
1516 static const struct rte_flow_ops mlx4_flow_ops = {
1517         .validate = mlx4_flow_validate,
1518         .create = mlx4_flow_create,
1519         .destroy = mlx4_flow_destroy,
1520         .flush = mlx4_flow_flush,
1521         .isolate = mlx4_flow_isolate,
1522 };
1523
1524 /**
1525  * Manage filter operations.
1526  *
1527  * @param dev
1528  *   Pointer to Ethernet device structure.
1529  * @param filter_type
1530  *   Filter type.
1531  * @param filter_op
1532  *   Operation to perform.
1533  * @param arg
1534  *   Pointer to operation-specific structure.
1535  *
1536  * @return
1537  *   0 on success, negative errno value otherwise and rte_errno is set.
1538  */
1539 int
1540 mlx4_filter_ctrl(struct rte_eth_dev *dev,
1541                  enum rte_filter_type filter_type,
1542                  enum rte_filter_op filter_op,
1543                  void *arg)
1544 {
1545         switch (filter_type) {
1546         case RTE_ETH_FILTER_GENERIC:
1547                 if (filter_op != RTE_ETH_FILTER_GET)
1548                         break;
1549                 *(const void **)arg = &mlx4_flow_ops;
1550                 return 0;
1551         default:
1552                 ERROR("%p: filter type (%d) not supported",
1553                       (void *)dev, filter_type);
1554                 break;
1555         }
1556         rte_errno = ENOTSUP;
1557         return -rte_errno;
1558 }