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_ethdev_driver.h>
  32 #include <rte_ether.h>
  33 #include <rte_flow.h>
  34 #include <rte_flow_driver.h>
  35 #include <rte_malloc.h>
  36
  37 /* PMD headers. */
  38 #include "mlx4.h"
  39 #include "mlx4_glue.h"
  40 #include "mlx4_flow.h"
  41 #include "mlx4_rxtx.h"
  42 #include "mlx4_utils.h"
  43
  44 /** Static initializer for a list of subsequent item types. */
  45 #define NEXT_ITEM(...) \
  46         (const enum rte_flow_item_type []){ \
  47                 __VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \
  48         }
  49
  50 /** Processor structure associated with a flow item. */
  51 struct mlx4_flow_proc_item {
  52         /** Bit-mask for fields supported by this PMD. */
  53         const void *mask_support;
  54         /** Bit-mask to use when @p item->mask is not provided. */
  55         const void *mask_default;
  56         /** Size in bytes for @p mask_support and @p mask_default. */
  57         const unsigned int mask_sz;
  58         /** Merge a pattern item into a flow rule handle. */
  59         int (*merge)(struct rte_flow *flow,
  60                      const struct rte_flow_item *item,
  61                      const struct mlx4_flow_proc_item *proc,
  62                      struct rte_flow_error *error);
  63         /** Size in bytes of the destination structure. */
  64         const unsigned int dst_sz;
  65         /** List of possible subsequent items. */
  66         const enum rte_flow_item_type *const next_item;
  67 };
  68
  69 /** Shared resources for drop flow rules. */
  70 struct mlx4_drop {
  71         struct ibv_qp *qp; /**< QP target. */
  72         struct ibv_cq *cq; /**< CQ associated with above QP. */
  73         struct mlx4_priv *priv; /**< Back pointer to private data. */
  74         uint32_t refcnt; /**< Reference count. */
  75 };
  76
  77 /**
  78  * Convert supported RSS hash field types between DPDK and Verbs formats.
  79  *
  80  * This function returns the supported (default) set when @p types has
  81  * special value 0.
  82  *
  83  * @param priv
  84  *   Pointer to private structure.
  85  * @param types
  86  *   Depending on @p verbs_to_dpdk, hash types in either DPDK (see struct
  87  *   rte_eth_rss_conf) or Verbs format.
  88  * @param verbs_to_dpdk
  89  *   A zero value converts @p types from DPDK to Verbs, a nonzero value
  90  *   performs the reverse operation.
  91  *
  92  * @return
  93  *   Converted RSS hash fields on success, (uint64_t)-1 otherwise and
  94  *   rte_errno is set.
  95  */
  96 uint64_t
  97 mlx4_conv_rss_types(struct mlx4_priv *priv, uint64_t types, int verbs_to_dpdk)
  98 {
  99         enum {
 100                 INNER,
 101                 IPV4, IPV4_1, IPV4_2, IPV6, IPV6_1, IPV6_2, IPV6_3,
 102                 TCP, UDP,
 103                 IPV4_TCP, IPV4_UDP, IPV6_TCP, IPV6_TCP_1, IPV6_UDP, IPV6_UDP_1,
 104         };
 105         enum {
 106                 VERBS_IPV4 = IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4,
 107                 VERBS_IPV6 = IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6,
 108                 VERBS_TCP = IBV_RX_HASH_SRC_PORT_TCP | IBV_RX_HASH_DST_PORT_TCP,
 109                 VERBS_UDP = IBV_RX_HASH_SRC_PORT_UDP | IBV_RX_HASH_DST_PORT_UDP,
 110         };
 111         static const uint64_t dpdk[] = {
 112                 [INNER] = 0,
 113                 [IPV4] = ETH_RSS_IPV4,
 114                 [IPV4_1] = ETH_RSS_FRAG_IPV4,
 115                 [IPV4_2] = ETH_RSS_NONFRAG_IPV4_OTHER,
 116                 [IPV6] = ETH_RSS_IPV6,
 117                 [IPV6_1] = ETH_RSS_FRAG_IPV6,
 118                 [IPV6_2] = ETH_RSS_NONFRAG_IPV6_OTHER,
 119                 [IPV6_3] = ETH_RSS_IPV6_EX,
 120                 [TCP] = 0,
 121                 [UDP] = 0,
 122                 [IPV4_TCP] = ETH_RSS_NONFRAG_IPV4_TCP,
 123                 [IPV4_UDP] = ETH_RSS_NONFRAG_IPV4_UDP,
 124                 [IPV6_TCP] = ETH_RSS_NONFRAG_IPV6_TCP,
 125                 [IPV6_TCP_1] = ETH_RSS_IPV6_TCP_EX,
 126                 [IPV6_UDP] = ETH_RSS_NONFRAG_IPV6_UDP,
 127                 [IPV6_UDP_1] = ETH_RSS_IPV6_UDP_EX,
 128         };
 129         static const uint64_t verbs[RTE_DIM(dpdk)] = {
 130                 [INNER] = IBV_RX_HASH_INNER,
 131                 [IPV4] = VERBS_IPV4,
 132                 [IPV4_1] = VERBS_IPV4,
 133                 [IPV4_2] = VERBS_IPV4,
 134                 [IPV6] = VERBS_IPV6,
 135                 [IPV6_1] = VERBS_IPV6,
 136                 [IPV6_2] = VERBS_IPV6,
 137                 [IPV6_3] = VERBS_IPV6,
 138                 [TCP] = VERBS_TCP,
 139                 [UDP] = VERBS_UDP,
 140                 [IPV4_TCP] = VERBS_IPV4 | VERBS_TCP,
 141                 [IPV4_UDP] = VERBS_IPV4 | VERBS_UDP,
 142                 [IPV6_TCP] = VERBS_IPV6 | VERBS_TCP,
 143                 [IPV6_TCP_1] = VERBS_IPV6 | VERBS_TCP,
 144                 [IPV6_UDP] = VERBS_IPV6 | VERBS_UDP,
 145                 [IPV6_UDP_1] = VERBS_IPV6 | VERBS_UDP,
 146         };
 147         const uint64_t *in = verbs_to_dpdk ? verbs : dpdk;
 148         const uint64_t *out = verbs_to_dpdk ? dpdk : verbs;
 149         uint64_t seen = 0;
 150         uint64_t conv = 0;
 151         unsigned int i;
 152
 153         if (!types) {
 154                 if (!verbs_to_dpdk)
 155                         return priv->hw_rss_sup;
 156                 types = priv->hw_rss_sup;
 157         }
 158         for (i = 0; i != RTE_DIM(dpdk); ++i)
 159                 if (in[i] && (types & in[i]) == in[i]) {
 160                         seen |= types & in[i];
 161                         conv |= out[i];
 162                 }
 163         if ((verbs_to_dpdk || (conv & priv->hw_rss_sup) == conv) &&
 164             !(types & ~seen))
 165                 return conv;
 166         rte_errno = ENOTSUP;
 167         return (uint64_t)-1;
 168 }
 169
 170 /**
 171  * Merge Ethernet pattern item into flow rule handle.
 172  *
 173  * Additional mlx4-specific constraints on supported fields:
 174  *
 175  * - No support for partial masks, except in the specific case of matching
 176  *   all multicast traffic (@p spec->dst and @p mask->dst equal to
 177  *   01:00:00:00:00:00).
 178  * - Not providing @p item->spec or providing an empty @p mask->dst is
 179  *   *only* supported if the rule doesn't specify additional matching
 180  *   criteria (i.e. rule is promiscuous-like).
 181  *
 182  * @param[in, out] flow
 183  *   Flow rule handle to update.
 184  * @param[in] item
 185  *   Pattern item to merge.
 186  * @param[in] proc
 187  *   Associated item-processing object.
 188  * @param[out] error
 189  *   Perform verbose error reporting if not NULL.
 190  *
 191  * @return
 192  *   0 on success, a negative errno value otherwise and rte_errno is set.
 193  */
 194 static int
 195 mlx4_flow_merge_eth(struct rte_flow *flow,
 196                     const struct rte_flow_item *item,
 197                     const struct mlx4_flow_proc_item *proc,
 198                     struct rte_flow_error *error)
 199 {
 200         const struct rte_flow_item_eth *spec = item->spec;
 201         const struct rte_flow_item_eth *mask =
 202                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 203         struct ibv_flow_spec_eth *eth;
 204         const char *msg;
 205         unsigned int i;
 206
 207         if (mask) {
 208                 uint32_t sum_dst = 0;
 209                 uint32_t sum_src = 0;
 210
 211                 for (i = 0; i != sizeof(mask->dst.addr_bytes); ++i) {
 212                         sum_dst += mask->dst.addr_bytes[i];
 213                         sum_src += mask->src.addr_bytes[i];
 214                 }
 215                 if (sum_src) {
 216                         msg = "mlx4 does not support source MAC matching";
 217                         goto error;
 218                 } else if (!sum_dst) {
 219                         flow->promisc = 1;
 220                 } else if (sum_dst == 1 && mask->dst.addr_bytes[0] == 1) {
 221                         if (!(spec->dst.addr_bytes[0] & 1)) {
 222                                 msg = "mlx4 does not support the explicit"
 223                                         " exclusion of all multicast traffic";
 224                                 goto error;
 225                         }
 226                         flow->allmulti = 1;
 227                 } else if (sum_dst != (UINT8_C(0xff) * RTE_ETHER_ADDR_LEN)) {
 228                         msg = "mlx4 does not support matching partial"
 229                                 " Ethernet fields";
 230                         goto error;
 231                 }
 232         }
 233         if (!flow->ibv_attr)
 234                 return 0;
 235         if (flow->promisc) {
 236                 flow->ibv_attr->type = IBV_FLOW_ATTR_ALL_DEFAULT;
 237                 return 0;
 238         }
 239         if (flow->allmulti) {
 240                 flow->ibv_attr->type = IBV_FLOW_ATTR_MC_DEFAULT;
 241                 return 0;
 242         }
 243         ++flow->ibv_attr->num_of_specs;
 244         eth = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 245         *eth = (struct ibv_flow_spec_eth) {
 246                 .type = IBV_FLOW_SPEC_ETH,
 247                 .size = sizeof(*eth),
 248         };
 249         if (!mask) {
 250                 eth->val.dst_mac[0] = 0xff;
 251                 flow->ibv_attr->type = IBV_FLOW_ATTR_ALL_DEFAULT;
 252                 flow->promisc = 1;
 253                 return 0;
 254         }
 255         memcpy(eth->val.dst_mac, spec->dst.addr_bytes, RTE_ETHER_ADDR_LEN);
 256         memcpy(eth->mask.dst_mac, mask->dst.addr_bytes, RTE_ETHER_ADDR_LEN);
 257         /* Remove unwanted bits from values. */
 258         for (i = 0; i < RTE_ETHER_ADDR_LEN; ++i)
 259                 eth->val.dst_mac[i] &= eth->mask.dst_mac[i];
 260
 261         return 0;
 262 error:
 263         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 264                                   item, msg);
 265 }
 266
 267 /**
 268  * Merge VLAN pattern item into flow rule handle.
 269  *
 270  * Additional mlx4-specific constraints on supported fields:
 271  *
 272  * - Matching *all* VLAN traffic by omitting @p item->spec or providing an
 273  *   empty @p item->mask would also include non-VLAN traffic. Doing so is
 274  *   therefore unsupported.
 275  * - No support for partial masks.
 276  *
 277  * @param[in, out] flow
 278  *   Flow rule handle to update.
 279  * @param[in] item
 280  *   Pattern item to merge.
 281  * @param[in] proc
 282  *   Associated item-processing object.
 283  * @param[out] error
 284  *   Perform verbose error reporting if not NULL.
 285  *
 286  * @return
 287  *   0 on success, a negative errno value otherwise and rte_errno is set.
 288  */
 289 static int
 290 mlx4_flow_merge_vlan(struct rte_flow *flow,
 291                      const struct rte_flow_item *item,
 292                      const struct mlx4_flow_proc_item *proc,
 293                      struct rte_flow_error *error)
 294 {
 295         const struct rte_flow_item_vlan *spec = item->spec;
 296         const struct rte_flow_item_vlan *mask =
 297                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 298         struct ibv_flow_spec_eth *eth;
 299         const char *msg;
 300
 301         if (!mask || !mask->tci) {
 302                 msg = "mlx4 cannot match all VLAN traffic while excluding"
 303                         " non-VLAN traffic, TCI VID must be specified";
 304                 goto error;
 305         }
 306         if (mask->tci != RTE_BE16(0x0fff)) {
 307                 msg = "mlx4 does not support partial TCI VID matching";
 308                 goto error;
 309         }
 310         if (!flow->ibv_attr)
 311                 return 0;
 312         eth = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size -
 313                        sizeof(*eth));
 314         eth->val.vlan_tag = spec->tci;
 315         eth->mask.vlan_tag = mask->tci;
 316         eth->val.vlan_tag &= eth->mask.vlan_tag;
 317         if (flow->ibv_attr->type == IBV_FLOW_ATTR_ALL_DEFAULT)
 318                 flow->ibv_attr->type = IBV_FLOW_ATTR_NORMAL;
 319         return 0;
 320 error:
 321         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 322                                   item, msg);
 323 }
 324
 325 /**
 326  * Merge IPv4 pattern item into flow rule handle.
 327  *
 328  * Additional mlx4-specific constraints on supported fields:
 329  *
 330  * - No support for partial masks.
 331  *
 332  * @param[in, out] flow
 333  *   Flow rule handle to update.
 334  * @param[in] item
 335  *   Pattern item to merge.
 336  * @param[in] proc
 337  *   Associated item-processing object.
 338  * @param[out] error
 339  *   Perform verbose error reporting if not NULL.
 340  *
 341  * @return
 342  *   0 on success, a negative errno value otherwise and rte_errno is set.
 343  */
 344 static int
 345 mlx4_flow_merge_ipv4(struct rte_flow *flow,
 346                      const struct rte_flow_item *item,
 347                      const struct mlx4_flow_proc_item *proc,
 348                      struct rte_flow_error *error)
 349 {
 350         const struct rte_flow_item_ipv4 *spec = item->spec;
 351         const struct rte_flow_item_ipv4 *mask =
 352                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 353         struct ibv_flow_spec_ipv4 *ipv4;
 354         const char *msg;
 355
 356         if (mask &&
 357             ((uint32_t)(mask->hdr.src_addr + 1) > UINT32_C(1) ||
 358              (uint32_t)(mask->hdr.dst_addr + 1) > UINT32_C(1))) {
 359                 msg = "mlx4 does not support matching partial IPv4 fields";
 360                 goto error;
 361         }
 362         if (!flow->ibv_attr)
 363                 return 0;
 364         ++flow->ibv_attr->num_of_specs;
 365         ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 366         *ipv4 = (struct ibv_flow_spec_ipv4) {
 367                 .type = IBV_FLOW_SPEC_IPV4,
 368                 .size = sizeof(*ipv4),
 369         };
 370         if (!spec)
 371                 return 0;
 372         ipv4->val = (struct ibv_flow_ipv4_filter) {
 373                 .src_ip = spec->hdr.src_addr,
 374                 .dst_ip = spec->hdr.dst_addr,
 375         };
 376         ipv4->mask = (struct ibv_flow_ipv4_filter) {
 377                 .src_ip = mask->hdr.src_addr,
 378                 .dst_ip = mask->hdr.dst_addr,
 379         };
 380         /* Remove unwanted bits from values. */
 381         ipv4->val.src_ip &= ipv4->mask.src_ip;
 382         ipv4->val.dst_ip &= ipv4->mask.dst_ip;
 383         return 0;
 384 error:
 385         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 386                                   item, msg);
 387 }
 388
 389 /**
 390  * Merge UDP pattern item into flow rule handle.
 391  *
 392  * Additional mlx4-specific constraints on supported fields:
 393  *
 394  * - No support for partial masks.
 395  * - Due to HW/FW limitation, flow rule priority is not taken into account
 396  *   when matching UDP destination ports, doing is therefore only supported
 397  *   at the highest priority level (0).
 398  *
 399  * @param[in, out] flow
 400  *   Flow rule handle to update.
 401  * @param[in] item
 402  *   Pattern item to merge.
 403  * @param[in] proc
 404  *   Associated item-processing object.
 405  * @param[out] error
 406  *   Perform verbose error reporting if not NULL.
 407  *
 408  * @return
 409  *   0 on success, a negative errno value otherwise and rte_errno is set.
 410  */
 411 static int
 412 mlx4_flow_merge_udp(struct rte_flow *flow,
 413                     const struct rte_flow_item *item,
 414                     const struct mlx4_flow_proc_item *proc,
 415                     struct rte_flow_error *error)
 416 {
 417         const struct rte_flow_item_udp *spec = item->spec;
 418         const struct rte_flow_item_udp *mask =
 419                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 420         struct ibv_flow_spec_tcp_udp *udp;
 421         const char *msg;
 422
 423         if (mask &&
 424             ((uint16_t)(mask->hdr.src_port + 1) > UINT16_C(1) ||
 425              (uint16_t)(mask->hdr.dst_port + 1) > UINT16_C(1))) {
 426                 msg = "mlx4 does not support matching partial UDP fields";
 427                 goto error;
 428         }
 429         if (mask && mask->hdr.dst_port && flow->priority) {
 430                 msg = "combining UDP destination port matching with a nonzero"
 431                         " priority level is not supported";
 432                 goto error;
 433         }
 434         if (!flow->ibv_attr)
 435                 return 0;
 436         ++flow->ibv_attr->num_of_specs;
 437         udp = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 438         *udp = (struct ibv_flow_spec_tcp_udp) {
 439                 .type = IBV_FLOW_SPEC_UDP,
 440                 .size = sizeof(*udp),
 441         };
 442         if (!spec)
 443                 return 0;
 444         udp->val.dst_port = spec->hdr.dst_port;
 445         udp->val.src_port = spec->hdr.src_port;
 446         udp->mask.dst_port = mask->hdr.dst_port;
 447         udp->mask.src_port = mask->hdr.src_port;
 448         /* Remove unwanted bits from values. */
 449         udp->val.src_port &= udp->mask.src_port;
 450         udp->val.dst_port &= udp->mask.dst_port;
 451         return 0;
 452 error:
 453         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 454                                   item, msg);
 455 }
 456
 457 /**
 458  * Merge TCP pattern item into flow rule handle.
 459  *
 460  * Additional mlx4-specific constraints on supported fields:
 461  *
 462  * - No support for partial masks.
 463  *
 464  * @param[in, out] flow
 465  *   Flow rule handle to update.
 466  * @param[in] item
 467  *   Pattern item to merge.
 468  * @param[in] proc
 469  *   Associated item-processing object.
 470  * @param[out] error
 471  *   Perform verbose error reporting if not NULL.
 472  *
 473  * @return
 474  *   0 on success, a negative errno value otherwise and rte_errno is set.
 475  */
 476 static int
 477 mlx4_flow_merge_tcp(struct rte_flow *flow,
 478                     const struct rte_flow_item *item,
 479                     const struct mlx4_flow_proc_item *proc,
 480                     struct rte_flow_error *error)
 481 {
 482         const struct rte_flow_item_tcp *spec = item->spec;
 483         const struct rte_flow_item_tcp *mask =
 484                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 485         struct ibv_flow_spec_tcp_udp *tcp;
 486         const char *msg;
 487
 488         if (mask &&
 489             ((uint16_t)(mask->hdr.src_port + 1) > UINT16_C(1) ||
 490              (uint16_t)(mask->hdr.dst_port + 1) > UINT16_C(1))) {
 491                 msg = "mlx4 does not support matching partial TCP fields";
 492                 goto error;
 493         }
 494         if (!flow->ibv_attr)
 495                 return 0;
 496         ++flow->ibv_attr->num_of_specs;
 497         tcp = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 498         *tcp = (struct ibv_flow_spec_tcp_udp) {
 499                 .type = IBV_FLOW_SPEC_TCP,
 500                 .size = sizeof(*tcp),
 501         };
 502         if (!spec)
 503                 return 0;
 504         tcp->val.dst_port = spec->hdr.dst_port;
 505         tcp->val.src_port = spec->hdr.src_port;
 506         tcp->mask.dst_port = mask->hdr.dst_port;
 507         tcp->mask.src_port = mask->hdr.src_port;
 508         /* Remove unwanted bits from values. */
 509         tcp->val.src_port &= tcp->mask.src_port;
 510         tcp->val.dst_port &= tcp->mask.dst_port;
 511         return 0;
 512 error:
 513         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 514                                   item, msg);
 515 }
 516
 517 /**
 518  * Perform basic sanity checks on a pattern item.
 519  *
 520  * @param[in] item
 521  *   Item specification.
 522  * @param[in] proc
 523  *   Associated item-processing object.
 524  * @param[out] error
 525  *   Perform verbose error reporting if not NULL.
 526  *
 527  * @return
 528  *   0 on success, a negative errno value otherwise and rte_errno is set.
 529  */
 530 static int
 531 mlx4_flow_item_check(const struct rte_flow_item *item,
 532                      const struct mlx4_flow_proc_item *proc,
 533                      struct rte_flow_error *error)
 534 {
 535         const uint8_t *mask;
 536         unsigned int i;
 537
 538         /* item->last and item->mask cannot exist without item->spec. */
 539         if (!item->spec && (item->mask || item->last))
 540                 return rte_flow_error_set
 541                         (error, EINVAL, RTE_FLOW_ERROR_TYPE_ITEM, item,
 542                          "\"mask\" or \"last\" field provided without a"
 543                          " corresponding \"spec\"");
 544         /* No spec, no mask, no problem. */
 545         if (!item->spec)
 546                 return 0;
 547         mask = item->mask ?
 548                 (const uint8_t *)item->mask :
 549                 (const uint8_t *)proc->mask_default;
 550         assert(mask);
 551         /*
 552          * Single-pass check to make sure that:
 553          * - Mask is supported, no bits are set outside proc->mask_support.
 554          * - Both item->spec and item->last are included in mask.
 555          */
 556         for (i = 0; i != proc->mask_sz; ++i) {
 557                 if (!mask[i])
 558                         continue;
 559                 if ((mask[i] | ((const uint8_t *)proc->mask_support)[i]) !=
 560                     ((const uint8_t *)proc->mask_support)[i])
 561                         return rte_flow_error_set
 562                                 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 563                                  item, "unsupported field found in \"mask\"");
 564                 if (item->last &&
 565                     (((const uint8_t *)item->spec)[i] & mask[i]) !=
 566                     (((const uint8_t *)item->last)[i] & mask[i]))
 567                         return rte_flow_error_set
 568                                 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 569                                  item,
 570                                  "range between \"spec\" and \"last\""
 571                                  " is larger than \"mask\"");
 572         }
 573         return 0;
 574 }
 575
 576 /** Graph of supported items and associated actions. */
 577 static const struct mlx4_flow_proc_item mlx4_flow_proc_item_list[] = {
 578         [RTE_FLOW_ITEM_TYPE_END] = {
 579                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_ETH),
 580         },
 581         [RTE_FLOW_ITEM_TYPE_ETH] = {
 582                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_VLAN,
 583                                        RTE_FLOW_ITEM_TYPE_IPV4),
 584                 .mask_support = &(const struct rte_flow_item_eth){
 585                         /* Only destination MAC can be matched. */
 586                         .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 587                 },
 588                 .mask_default = &rte_flow_item_eth_mask,
 589                 .mask_sz = sizeof(struct rte_flow_item_eth),
 590                 .merge = mlx4_flow_merge_eth,
 591                 .dst_sz = sizeof(struct ibv_flow_spec_eth),
 592         },
 593         [RTE_FLOW_ITEM_TYPE_VLAN] = {
 594                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_IPV4),
 595                 .mask_support = &(const struct rte_flow_item_vlan){
 596                         /* Only TCI VID matching is supported. */
 597                         .tci = RTE_BE16(0x0fff),
 598                 },
 599                 .mask_default = &rte_flow_item_vlan_mask,
 600                 .mask_sz = sizeof(struct rte_flow_item_vlan),
 601                 .merge = mlx4_flow_merge_vlan,
 602                 .dst_sz = 0,
 603         },
 604         [RTE_FLOW_ITEM_TYPE_IPV4] = {
 605                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_UDP,
 606                                        RTE_FLOW_ITEM_TYPE_TCP),
 607                 .mask_support = &(const struct rte_flow_item_ipv4){
 608                         .hdr = {
 609                                 .src_addr = RTE_BE32(0xffffffff),
 610                                 .dst_addr = RTE_BE32(0xffffffff),
 611                         },
 612                 },
 613                 .mask_default = &rte_flow_item_ipv4_mask,
 614                 .mask_sz = sizeof(struct rte_flow_item_ipv4),
 615                 .merge = mlx4_flow_merge_ipv4,
 616                 .dst_sz = sizeof(struct ibv_flow_spec_ipv4),
 617         },
 618         [RTE_FLOW_ITEM_TYPE_UDP] = {
 619                 .mask_support = &(const struct rte_flow_item_udp){
 620                         .hdr = {
 621                                 .src_port = RTE_BE16(0xffff),
 622                                 .dst_port = RTE_BE16(0xffff),
 623                         },
 624                 },
 625                 .mask_default = &rte_flow_item_udp_mask,
 626                 .mask_sz = sizeof(struct rte_flow_item_udp),
 627                 .merge = mlx4_flow_merge_udp,
 628                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 629         },
 630         [RTE_FLOW_ITEM_TYPE_TCP] = {
 631                 .mask_support = &(const struct rte_flow_item_tcp){
 632                         .hdr = {
 633                                 .src_port = RTE_BE16(0xffff),
 634                                 .dst_port = RTE_BE16(0xffff),
 635                         },
 636                 },
 637                 .mask_default = &rte_flow_item_tcp_mask,
 638                 .mask_sz = sizeof(struct rte_flow_item_tcp),
 639                 .merge = mlx4_flow_merge_tcp,
 640                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 641         },
 642 };
 643
 644 /**
 645  * Make sure a flow rule is supported and initialize associated structure.
 646  *
 647  * @param priv
 648  *   Pointer to private structure.
 649  * @param[in] attr
 650  *   Flow rule attributes.
 651  * @param[in] pattern
 652  *   Pattern specification (list terminated by the END pattern item).
 653  * @param[in] actions
 654  *   Associated actions (list terminated by the END action).
 655  * @param[out] error
 656  *   Perform verbose error reporting if not NULL.
 657  * @param[in, out] addr
 658  *   Buffer where the resulting flow rule handle pointer must be stored.
 659  *   If NULL, stop processing after validation stage.
 660  *
 661  * @return
 662  *   0 on success, a negative errno value otherwise and rte_errno is set.
 663  */
 664 static int
 665 mlx4_flow_prepare(struct mlx4_priv *priv,
 666                   const struct rte_flow_attr *attr,
 667                   const struct rte_flow_item pattern[],
 668                   const struct rte_flow_action actions[],
 669                   struct rte_flow_error *error,
 670                   struct rte_flow **addr)
 671 {
 672         const struct rte_flow_item *item;
 673         const struct rte_flow_action *action;
 674         const struct mlx4_flow_proc_item *proc;
 675         struct rte_flow temp = { .ibv_attr_size = sizeof(*temp.ibv_attr) };
 676         struct rte_flow *flow = &temp;
 677         const char *msg = NULL;
 678         int overlap;
 679
 680         if (attr->group)
 681                 return rte_flow_error_set
 682                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 683                          NULL, "groups are not supported");
 684         if (attr->priority > MLX4_FLOW_PRIORITY_LAST)
 685                 return rte_flow_error_set
 686                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 687                          NULL, "maximum priority level is "
 688                          MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST));
 689         if (attr->egress)
 690                 return rte_flow_error_set
 691                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 692                          NULL, "egress is not supported");
 693         if (attr->transfer)
 694                 return rte_flow_error_set
 695                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_TRANSFER,
 696                          NULL, "transfer is not supported");
 697         if (!attr->ingress)
 698                 return rte_flow_error_set
 699                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 700                          NULL, "only ingress is supported");
 701 fill:
 702         overlap = 0;
 703         proc = mlx4_flow_proc_item_list;
 704         flow->priority = attr->priority;
 705         /* Go over pattern. */
 706         for (item = pattern; item->type; ++item) {
 707                 const struct mlx4_flow_proc_item *next = NULL;
 708                 unsigned int i;
 709                 int err;
 710
 711                 if (item->type == RTE_FLOW_ITEM_TYPE_VOID)
 712                         continue;
 713                 if (item->type == MLX4_FLOW_ITEM_TYPE_INTERNAL) {
 714                         flow->internal = 1;
 715                         continue;
 716                 }
 717                 if (flow->promisc || flow->allmulti) {
 718                         msg = "mlx4 does not support additional matching"
 719                                 " criteria combined with indiscriminate"
 720                                 " matching on Ethernet headers";
 721                         goto exit_item_not_supported;
 722                 }
 723                 for (i = 0; proc->next_item && proc->next_item[i]; ++i) {
 724                         if (proc->next_item[i] == item->type) {
 725                                 next = &mlx4_flow_proc_item_list[item->type];
 726                                 break;
 727                         }
 728                 }
 729                 if (!next)
 730                         goto exit_item_not_supported;
 731                 proc = next;
 732                 /*
 733                  * Perform basic sanity checks only once, while handle is
 734                  * not allocated.
 735                  */
 736                 if (flow == &temp) {
 737                         err = mlx4_flow_item_check(item, proc, error);
 738                         if (err)
 739                                 return err;
 740                 }
 741                 if (proc->merge) {
 742                         err = proc->merge(flow, item, proc, error);
 743                         if (err)
 744                                 return err;
 745                 }
 746                 flow->ibv_attr_size += proc->dst_sz;
 747         }
 748         /* Go over actions list. */
 749         for (action = actions; action->type; ++action) {
 750                 /* This one may appear anywhere multiple times. */
 751                 if (action->type == RTE_FLOW_ACTION_TYPE_VOID)
 752                         continue;
 753                 /* Fate-deciding actions may appear exactly once. */
 754                 if (overlap) {
 755                         msg = "cannot combine several fate-deciding actions,"
 756                                 " choose between DROP, QUEUE or RSS";
 757                         goto exit_action_not_supported;
 758                 }
 759                 overlap = 1;
 760                 switch (action->type) {
 761                         const struct rte_flow_action_queue *queue;
 762                         const struct rte_flow_action_rss *rss;
 763                         const uint8_t *rss_key;
 764                         uint32_t rss_key_len;
 765                         uint64_t fields;
 766                         unsigned int i;
 767
 768                 case RTE_FLOW_ACTION_TYPE_DROP:
 769                         flow->drop = 1;
 770                         break;
 771                 case RTE_FLOW_ACTION_TYPE_QUEUE:
 772                         if (flow->rss)
 773                                 break;
 774                         queue = action->conf;
 775                         if (queue->index >= ETH_DEV(priv)->data->nb_rx_queues) {
 776                                 msg = "queue target index beyond number of"
 777                                         " configured Rx queues";
 778                                 goto exit_action_not_supported;
 779                         }
 780                         flow->rss = mlx4_rss_get
 781                                 (priv, 0, mlx4_rss_hash_key_default, 1,
 782                                  &queue->index);
 783                         if (!flow->rss) {
 784                                 msg = "not enough resources for additional"
 785                                         " single-queue RSS context";
 786                                 goto exit_action_not_supported;
 787                         }
 788                         break;
 789                 case RTE_FLOW_ACTION_TYPE_RSS:
 790                         if (flow->rss)
 791                                 break;
 792                         rss = action->conf;
 793                         /* Default RSS configuration if none is provided. */
 794                         if (rss->key_len) {
 795                                 rss_key = rss->key;
 796                                 rss_key_len = rss->key_len;
 797                         } else {
 798                                 rss_key = mlx4_rss_hash_key_default;
 799                                 rss_key_len = MLX4_RSS_HASH_KEY_SIZE;
 800                         }
 801                         /* Sanity checks. */
 802                         for (i = 0; i < rss->queue_num; ++i)
 803                                 if (rss->queue[i] >=
 804                                     ETH_DEV(priv)->data->nb_rx_queues)
 805                                         break;
 806                         if (i != rss->queue_num) {
 807                                 msg = "queue index target beyond number of"
 808                                         " configured Rx queues";
 809                                 goto exit_action_not_supported;
 810                         }
 811                         if (!rte_is_power_of_2(rss->queue_num)) {
 812                                 msg = "for RSS, mlx4 requires the number of"
 813                                         " queues to be a power of two";
 814                                 goto exit_action_not_supported;
 815                         }
 816                         if (rss_key_len != sizeof(flow->rss->key)) {
 817                                 msg = "mlx4 supports exactly one RSS hash key"
 818                                         " length: "
 819                                         MLX4_STR_EXPAND(MLX4_RSS_HASH_KEY_SIZE);
 820                                 goto exit_action_not_supported;
 821                         }
 822                         for (i = 1; i < rss->queue_num; ++i)
 823                                 if (rss->queue[i] - rss->queue[i - 1] != 1)
 824                                         break;
 825                         if (i != rss->queue_num) {
 826                                 msg = "mlx4 requires RSS contexts to use"
 827                                         " consecutive queue indices only";
 828                                 goto exit_action_not_supported;
 829                         }
 830                         if (rss->queue[0] % rss->queue_num) {
 831                                 msg = "mlx4 requires the first queue of a RSS"
 832                                         " context to be aligned on a multiple"
 833                                         " of the context size";
 834                                 goto exit_action_not_supported;
 835                         }
 836                         if (rss->func &&
 837                             rss->func != RTE_ETH_HASH_FUNCTION_TOEPLITZ) {
 838                                 msg = "the only supported RSS hash function"
 839                                         " is Toeplitz";
 840                                 goto exit_action_not_supported;
 841                         }
 842                         if (rss->level) {
 843                                 msg = "a nonzero RSS encapsulation level is"
 844                                         " not supported";
 845                                 goto exit_action_not_supported;
 846                         }
 847                         rte_errno = 0;
 848                         fields = mlx4_conv_rss_types(priv, rss->types, 0);
 849                         if (fields == (uint64_t)-1 && rte_errno) {
 850                                 msg = "unsupported RSS hash type requested";
 851                                 goto exit_action_not_supported;
 852                         }
 853                         flow->rss = mlx4_rss_get
 854                                 (priv, fields, rss_key, rss->queue_num,
 855                                  rss->queue);
 856                         if (!flow->rss) {
 857                                 msg = "either invalid parameters or not enough"
 858                                         " resources for additional multi-queue"
 859                                         " RSS context";
 860                                 goto exit_action_not_supported;
 861                         }
 862                         break;
 863                 default:
 864                         goto exit_action_not_supported;
 865                 }
 866         }
 867         /* When fate is unknown, drop traffic. */
 868         if (!overlap)
 869                 flow->drop = 1;
 870         /* Validation ends here. */
 871         if (!addr) {
 872                 if (flow->rss)
 873                         mlx4_rss_put(flow->rss);
 874                 return 0;
 875         }
 876         if (flow == &temp) {
 877                 /* Allocate proper handle based on collected data. */
 878                 const struct mlx4_malloc_vec vec[] = {
 879                         {
 880                                 .align = alignof(struct rte_flow),
 881                                 .size = sizeof(*flow),
 882                                 .addr = (void **)&flow,
 883                         },
 884                         {
 885                                 .align = alignof(struct ibv_flow_attr),
 886                                 .size = temp.ibv_attr_size,
 887                                 .addr = (void **)&temp.ibv_attr,
 888                         },
 889                 };
 890
 891                 if (!mlx4_zmallocv(__func__, vec, RTE_DIM(vec))) {
 892                         if (temp.rss)
 893                                 mlx4_rss_put(temp.rss);
 894                         return rte_flow_error_set
 895                                 (error, -rte_errno,
 896                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
 897                                  "flow rule handle allocation failure");
 898                 }
 899                 /* Most fields will be updated by second pass. */
 900                 *flow = (struct rte_flow){
 901                         .ibv_attr = temp.ibv_attr,
 902                         .ibv_attr_size = sizeof(*flow->ibv_attr),
 903                         .rss = temp.rss,
 904                 };
 905                 *flow->ibv_attr = (struct ibv_flow_attr){
 906                         .type = IBV_FLOW_ATTR_NORMAL,
 907                         .size = sizeof(*flow->ibv_attr),
 908                         .priority = attr->priority,
 909                         .port = priv->port,
 910                 };
 911                 goto fill;
 912         }
 913         *addr = flow;
 914         return 0;
 915 exit_item_not_supported:
 916         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 917                                   item, msg ? msg : "item not supported");
 918 exit_action_not_supported:
 919         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 920                                   action, msg ? msg : "action not supported");
 921 }
 922
 923 /**
 924  * Validate a flow supported by the NIC.
 925  *
 926  * @see rte_flow_validate()
 927  * @see rte_flow_ops
 928  */
 929 static int
 930 mlx4_flow_validate(struct rte_eth_dev *dev,
 931                    const struct rte_flow_attr *attr,
 932                    const struct rte_flow_item pattern[],
 933                    const struct rte_flow_action actions[],
 934                    struct rte_flow_error *error)
 935 {
 936         struct mlx4_priv *priv = dev->data->dev_private;
 937
 938         return mlx4_flow_prepare(priv, attr, pattern, actions, error, NULL);
 939 }
 940
 941 /**
 942  * Get a drop flow rule resources instance.
 943  *
 944  * @param priv
 945  *   Pointer to private structure.
 946  *
 947  * @return
 948  *   Pointer to drop flow resources on success, NULL otherwise and rte_errno
 949  *   is set.
 950  */
 951 static struct mlx4_drop *
 952 mlx4_drop_get(struct mlx4_priv *priv)
 953 {
 954         struct mlx4_drop *drop = priv->drop;
 955
 956         if (drop) {
 957                 assert(drop->refcnt);
 958                 assert(drop->priv == priv);
 959                 ++drop->refcnt;
 960                 return drop;
 961         }
 962         drop = rte_malloc(__func__, sizeof(*drop), 0);
 963         if (!drop)
 964                 goto error;
 965         *drop = (struct mlx4_drop){
 966                 .priv = priv,
 967                 .refcnt = 1,
 968         };
 969         drop->cq = mlx4_glue->create_cq(priv->ctx, 1, NULL, NULL, 0);
 970         if (!drop->cq)
 971                 goto error;
 972         drop->qp = mlx4_glue->create_qp
 973                 (priv->pd,
 974                  &(struct ibv_qp_init_attr){
 975                         .send_cq = drop->cq,
 976                         .recv_cq = drop->cq,
 977                         .qp_type = IBV_QPT_RAW_PACKET,
 978                  });
 979         if (!drop->qp)
 980                 goto error;
 981         priv->drop = drop;
 982         return drop;
 983 error:
 984         if (drop->qp)
 985                 claim_zero(mlx4_glue->destroy_qp(drop->qp));
 986         if (drop->cq)
 987                 claim_zero(mlx4_glue->destroy_cq(drop->cq));
 988         if (drop)
 989                 rte_free(drop);
 990         rte_errno = ENOMEM;
 991         return NULL;
 992 }
 993
 994 /**
 995  * Give back a drop flow rule resources instance.
 996  *
 997  * @param drop
 998  *   Pointer to drop flow rule resources.
 999  */
1000 static void
1001 mlx4_drop_put(struct mlx4_drop *drop)
1002 {
1003         assert(drop->refcnt);
1004         if (--drop->refcnt)
1005                 return;
1006         drop->priv->drop = NULL;
1007         claim_zero(mlx4_glue->destroy_qp(drop->qp));
1008         claim_zero(mlx4_glue->destroy_cq(drop->cq));
1009         rte_free(drop);
1010 }
1011
1012 /**
1013  * Toggle a configured flow rule.
1014  *
1015  * @param priv
1016  *   Pointer to private structure.
1017  * @param flow
1018  *   Flow rule handle to toggle.
1019  * @param enable
1020  *   Whether associated Verbs flow must be created or removed.
1021  * @param[out] error
1022  *   Perform verbose error reporting if not NULL.
1023  *
1024  * @return
1025  *   0 on success, a negative errno value otherwise and rte_errno is set.
1026  */
1027 static int
1028 mlx4_flow_toggle(struct mlx4_priv *priv,
1029                  struct rte_flow *flow,
1030                  int enable,
1031                  struct rte_flow_error *error)
1032 {
1033         struct ibv_qp *qp = NULL;
1034         const char *msg;
1035         int err;
1036
1037         if (!enable) {
1038                 if (!flow->ibv_flow)
1039                         return 0;
1040                 claim_zero(mlx4_glue->destroy_flow(flow->ibv_flow));
1041                 flow->ibv_flow = NULL;
1042                 if (flow->drop)
1043                         mlx4_drop_put(priv->drop);
1044                 else if (flow->rss)
1045                         mlx4_rss_detach(flow->rss);
1046                 return 0;
1047         }
1048         assert(flow->ibv_attr);
1049         if (!flow->internal &&
1050             !priv->isolated &&
1051             flow->ibv_attr->priority == MLX4_FLOW_PRIORITY_LAST) {
1052                 if (flow->ibv_flow) {
1053                         claim_zero(mlx4_glue->destroy_flow(flow->ibv_flow));
1054                         flow->ibv_flow = NULL;
1055                         if (flow->drop)
1056                                 mlx4_drop_put(priv->drop);
1057                         else if (flow->rss)
1058                                 mlx4_rss_detach(flow->rss);
1059                 }
1060                 err = EACCES;
1061                 msg = ("priority level "
1062                        MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST)
1063                        " is reserved when not in isolated mode");
1064                 goto error;
1065         }
1066         if (flow->rss) {
1067                 struct mlx4_rss *rss = flow->rss;
1068                 int missing = 0;
1069                 unsigned int i;
1070
1071                 /* Stop at the first nonexistent target queue. */
1072                 for (i = 0; i != rss->queues; ++i)
1073                         if (rss->queue_id[i] >=
1074                             ETH_DEV(priv)->data->nb_rx_queues ||
1075                             !ETH_DEV(priv)->data->rx_queues[rss->queue_id[i]]) {
1076                                 missing = 1;
1077                                 break;
1078                         }
1079                 if (flow->ibv_flow) {
1080                         if (missing ^ !flow->drop)
1081                                 return 0;
1082                         /* Verbs flow needs updating. */
1083                         claim_zero(mlx4_glue->destroy_flow(flow->ibv_flow));
1084                         flow->ibv_flow = NULL;
1085                         if (flow->drop)
1086                                 mlx4_drop_put(priv->drop);
1087                         else
1088                                 mlx4_rss_detach(rss);
1089                 }
1090                 if (!missing) {
1091                         err = mlx4_rss_attach(rss);
1092                         if (err) {
1093                                 err = -err;
1094                                 msg = "cannot create indirection table or hash"
1095                                         " QP to associate flow rule with";
1096                                 goto error;
1097                         }
1098                         qp = rss->qp;
1099                 }
1100                 /* A missing target queue drops traffic implicitly. */
1101                 flow->drop = missing;
1102         }
1103         if (flow->drop) {
1104                 if (flow->ibv_flow)
1105                         return 0;
1106                 mlx4_drop_get(priv);
1107                 if (!priv->drop) {
1108                         err = rte_errno;
1109                         msg = "resources for drop flow rule cannot be created";
1110                         goto error;
1111                 }
1112                 qp = priv->drop->qp;
1113         }
1114         assert(qp);
1115         if (flow->ibv_flow)
1116                 return 0;
1117         flow->ibv_flow = mlx4_glue->create_flow(qp, flow->ibv_attr);
1118         if (flow->ibv_flow)
1119                 return 0;
1120         if (flow->drop)
1121                 mlx4_drop_put(priv->drop);
1122         else if (flow->rss)
1123                 mlx4_rss_detach(flow->rss);
1124         err = errno;
1125         msg = "flow rule rejected by device";
1126 error:
1127         return rte_flow_error_set
1128                 (error, err, RTE_FLOW_ERROR_TYPE_HANDLE, flow, msg);
1129 }
1130
1131 /**
1132  * Create a flow.
1133  *
1134  * @see rte_flow_create()
1135  * @see rte_flow_ops
1136  */
1137 static struct rte_flow *
1138 mlx4_flow_create(struct rte_eth_dev *dev,
1139                  const struct rte_flow_attr *attr,
1140                  const struct rte_flow_item pattern[],
1141                  const struct rte_flow_action actions[],
1142                  struct rte_flow_error *error)
1143 {
1144         struct mlx4_priv *priv = dev->data->dev_private;
1145         struct rte_flow *flow;
1146         int err;
1147
1148         err = mlx4_flow_prepare(priv, attr, pattern, actions, error, &flow);
1149         if (err)
1150                 return NULL;
1151         err = mlx4_flow_toggle(priv, flow, priv->started, error);
1152         if (!err) {
1153                 struct rte_flow *curr = LIST_FIRST(&priv->flows);
1154
1155                 /* New rules are inserted after internal ones. */
1156                 if (!curr || !curr->internal) {
1157                         LIST_INSERT_HEAD(&priv->flows, flow, next);
1158                 } else {
1159                         while (LIST_NEXT(curr, next) &&
1160                                LIST_NEXT(curr, next)->internal)
1161                                 curr = LIST_NEXT(curr, next);
1162                         LIST_INSERT_AFTER(curr, flow, next);
1163                 }
1164                 return flow;
1165         }
1166         if (flow->rss)
1167                 mlx4_rss_put(flow->rss);
1168         rte_flow_error_set(error, -err, RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1169                            error->message);
1170         rte_free(flow);
1171         return NULL;
1172 }
1173
1174 /**
1175  * Configure isolated mode.
1176  *
1177  * @see rte_flow_isolate()
1178  * @see rte_flow_ops
1179  */
1180 static int
1181 mlx4_flow_isolate(struct rte_eth_dev *dev,
1182                   int enable,
1183                   struct rte_flow_error *error)
1184 {
1185         struct mlx4_priv *priv = dev->data->dev_private;
1186
1187         if (!!enable == !!priv->isolated)
1188                 return 0;
1189         priv->isolated = !!enable;
1190         if (mlx4_flow_sync(priv, error)) {
1191                 priv->isolated = !enable;
1192                 return -rte_errno;
1193         }
1194         return 0;
1195 }
1196
1197 /**
1198  * Destroy a flow rule.
1199  *
1200  * @see rte_flow_destroy()
1201  * @see rte_flow_ops
1202  */
1203 static int
1204 mlx4_flow_destroy(struct rte_eth_dev *dev,
1205                   struct rte_flow *flow,
1206                   struct rte_flow_error *error)
1207 {
1208         struct mlx4_priv *priv = dev->data->dev_private;
1209         int err = mlx4_flow_toggle(priv, flow, 0, error);
1210
1211         if (err)
1212                 return err;
1213         LIST_REMOVE(flow, next);
1214         if (flow->rss)
1215                 mlx4_rss_put(flow->rss);
1216         rte_free(flow);
1217         return 0;
1218 }
1219
1220 /**
1221  * Destroy user-configured flow rules.
1222  *
1223  * This function skips internal flows rules.
1224  *
1225  * @see rte_flow_flush()
1226  * @see rte_flow_ops
1227  */
1228 static int
1229 mlx4_flow_flush(struct rte_eth_dev *dev,
1230                 struct rte_flow_error *error)
1231 {
1232         struct mlx4_priv *priv = dev->data->dev_private;
1233         struct rte_flow *flow = LIST_FIRST(&priv->flows);
1234
1235         while (flow) {
1236                 struct rte_flow *next = LIST_NEXT(flow, next);
1237
1238                 if (!flow->internal)
1239                         mlx4_flow_destroy(dev, flow, error);
1240                 flow = next;
1241         }
1242         return 0;
1243 }
1244
1245 /**
1246  * Helper function to determine the next configured VLAN filter.
1247  *
1248  * @param priv
1249  *   Pointer to private structure.
1250  * @param vlan
1251  *   VLAN ID to use as a starting point.
1252  *
1253  * @return
1254  *   Next configured VLAN ID or a high value (>= 4096) if there is none.
1255  */
1256 static uint16_t
1257 mlx4_flow_internal_next_vlan(struct mlx4_priv *priv, uint16_t vlan)
1258 {
1259         while (vlan < 4096) {
1260                 if (ETH_DEV(priv)->data->vlan_filter_conf.ids[vlan / 64] &
1261                     (UINT64_C(1) << (vlan % 64)))
1262                         return vlan;
1263                 ++vlan;
1264         }
1265         return vlan;
1266 }
1267
1268 /**
1269  * Generate internal flow rules.
1270  *
1271  * Various flow rules are created depending on the mode the device is in:
1272  *
1273  * 1. Promiscuous:
1274  *       port MAC + broadcast + catch-all (VLAN filtering is ignored).
1275  * 2. All multicast:
1276  *       port MAC/VLAN + broadcast + catch-all multicast.
1277  * 3. Otherwise:
1278  *       port MAC/VLAN + broadcast MAC/VLAN.
1279  *
1280  * About MAC flow rules:
1281  *
1282  * - MAC flow rules are generated from @p dev->data->mac_addrs
1283  *   (@p priv->mac array).
1284  * - An additional flow rule for Ethernet broadcasts is also generated.
1285  * - All these are per-VLAN if @p DEV_RX_OFFLOAD_VLAN_FILTER
1286  *   is enabled and VLAN filters are configured.
1287  *
1288  * @param priv
1289  *   Pointer to private structure.
1290  * @param[out] error
1291  *   Perform verbose error reporting if not NULL.
1292  *
1293  * @return
1294  *   0 on success, a negative errno value otherwise and rte_errno is set.
1295  */
1296 static int
1297 mlx4_flow_internal(struct mlx4_priv *priv, struct rte_flow_error *error)
1298 {
1299         struct rte_flow_attr attr = {
1300                 .priority = MLX4_FLOW_PRIORITY_LAST,
1301                 .ingress = 1,
1302         };
1303         struct rte_flow_item_eth eth_spec;
1304         const struct rte_flow_item_eth eth_mask = {
1305                 .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
1306         };
1307         const struct rte_flow_item_eth eth_allmulti = {
1308                 .dst.addr_bytes = "\x01\x00\x00\x00\x00\x00",
1309         };
1310         struct rte_flow_item_vlan vlan_spec;
1311         const struct rte_flow_item_vlan vlan_mask = {
1312                 .tci = RTE_BE16(0x0fff),
1313         };
1314         struct rte_flow_item pattern[] = {
1315                 {
1316                         .type = MLX4_FLOW_ITEM_TYPE_INTERNAL,
1317                 },
1318                 {
1319                         .type = RTE_FLOW_ITEM_TYPE_ETH,
1320                         .spec = &eth_spec,
1321                         .mask = &eth_mask,
1322                 },
1323                 {
1324                         /* Replaced with VLAN if filtering is enabled. */
1325                         .type = RTE_FLOW_ITEM_TYPE_END,
1326                 },
1327                 {
1328                         .type = RTE_FLOW_ITEM_TYPE_END,
1329                 },
1330         };
1331         /*
1332          * Round number of queues down to their previous power of 2 to
1333          * comply with RSS context limitations. Extra queues silently do not
1334          * get RSS by default.
1335          */
1336         uint32_t queues =
1337                 rte_align32pow2(ETH_DEV(priv)->data->nb_rx_queues + 1) >> 1;
1338         uint16_t queue[queues];
1339         struct rte_flow_action_rss action_rss = {
1340                 .func = RTE_ETH_HASH_FUNCTION_DEFAULT,
1341                 .level = 0,
1342                 .types = 0,
1343                 .key_len = MLX4_RSS_HASH_KEY_SIZE,
1344                 .queue_num = queues,
1345                 .key = mlx4_rss_hash_key_default,
1346                 .queue = queue,
1347         };
1348         struct rte_flow_action actions[] = {
1349                 {
1350                         .type = RTE_FLOW_ACTION_TYPE_RSS,
1351                         .conf = &action_rss,
1352                 },
1353                 {
1354                         .type = RTE_FLOW_ACTION_TYPE_END,
1355                 },
1356         };
1357         struct rte_ether_addr *rule_mac = &eth_spec.dst;
1358         rte_be16_t *rule_vlan =
1359                 (ETH_DEV(priv)->data->dev_conf.rxmode.offloads &
1360                  DEV_RX_OFFLOAD_VLAN_FILTER) &&
1361                 !ETH_DEV(priv)->data->promiscuous ?
1362                 &vlan_spec.tci :
1363                 NULL;
1364         uint16_t vlan = 0;
1365         struct rte_flow *flow;
1366         unsigned int i;
1367         int err = 0;
1368
1369         /* Nothing to be done if there are no Rx queues. */
1370         if (!queues)
1371                 goto error;
1372         /* Prepare default RSS configuration. */
1373         for (i = 0; i != queues; ++i)
1374                 queue[i] = i;
1375         /*
1376          * Set up VLAN item if filtering is enabled and at least one VLAN
1377          * filter is configured.
1378          */
1379         if (rule_vlan) {
1380                 vlan = mlx4_flow_internal_next_vlan(priv, 0);
1381                 if (vlan < 4096) {
1382                         pattern[2] = (struct rte_flow_item){
1383                                 .type = RTE_FLOW_ITEM_TYPE_VLAN,
1384                                 .spec = &vlan_spec,
1385                                 .mask = &vlan_mask,
1386                         };
1387 next_vlan:
1388                         *rule_vlan = rte_cpu_to_be_16(vlan);
1389                 } else {
1390                         rule_vlan = NULL;
1391                 }
1392         }
1393         for (i = 0; i != RTE_DIM(priv->mac) + 1; ++i) {
1394                 const struct rte_ether_addr *mac;
1395
1396                 /* Broadcasts are handled by an extra iteration. */
1397                 if (i < RTE_DIM(priv->mac))
1398                         mac = &priv->mac[i];
1399                 else
1400                         mac = &eth_mask.dst;
1401                 if (rte_is_zero_ether_addr(mac))
1402                         continue;
1403                 /* Check if MAC flow rule is already present. */
1404                 for (flow = LIST_FIRST(&priv->flows);
1405                      flow && flow->internal;
1406                      flow = LIST_NEXT(flow, next)) {
1407                         const struct ibv_flow_spec_eth *eth =
1408                                 (const void *)((uintptr_t)flow->ibv_attr +
1409                                                sizeof(*flow->ibv_attr));
1410                         unsigned int j;
1411
1412                         if (!flow->mac)
1413                                 continue;
1414                         assert(flow->ibv_attr->type == IBV_FLOW_ATTR_NORMAL);
1415                         assert(flow->ibv_attr->num_of_specs == 1);
1416                         assert(eth->type == IBV_FLOW_SPEC_ETH);
1417                         assert(flow->rss);
1418                         if (rule_vlan &&
1419                             (eth->val.vlan_tag != *rule_vlan ||
1420                              eth->mask.vlan_tag != RTE_BE16(0x0fff)))
1421                                 continue;
1422                         if (!rule_vlan && eth->mask.vlan_tag)
1423                                 continue;
1424                         for (j = 0; j != sizeof(mac->addr_bytes); ++j)
1425                                 if (eth->val.dst_mac[j] != mac->addr_bytes[j] ||
1426                                     eth->mask.dst_mac[j] != UINT8_C(0xff) ||
1427                                     eth->val.src_mac[j] != UINT8_C(0x00) ||
1428                                     eth->mask.src_mac[j] != UINT8_C(0x00))
1429                                         break;
1430                         if (j != sizeof(mac->addr_bytes))
1431                                 continue;
1432                         if (flow->rss->queues != queues ||
1433                             memcmp(flow->rss->queue_id, action_rss.queue,
1434                                    queues * sizeof(flow->rss->queue_id[0])))
1435                                 continue;
1436                         break;
1437                 }
1438                 if (!flow || !flow->internal) {
1439                         /* Not found, create a new flow rule. */
1440                         memcpy(rule_mac, mac, sizeof(*mac));
1441                         flow = mlx4_flow_create(ETH_DEV(priv), &attr, pattern,
1442                                                 actions, error);
1443                         if (!flow) {
1444                                 err = -rte_errno;
1445                                 goto error;
1446                         }
1447                 }
1448                 flow->select = 1;
1449                 flow->mac = 1;
1450         }
1451         if (rule_vlan) {
1452                 vlan = mlx4_flow_internal_next_vlan(priv, vlan + 1);
1453                 if (vlan < 4096)
1454                         goto next_vlan;
1455         }
1456         /* Take care of promiscuous and all multicast flow rules. */
1457         if (ETH_DEV(priv)->data->promiscuous ||
1458             ETH_DEV(priv)->data->all_multicast) {
1459                 for (flow = LIST_FIRST(&priv->flows);
1460                      flow && flow->internal;
1461                      flow = LIST_NEXT(flow, next)) {
1462                         if (ETH_DEV(priv)->data->promiscuous) {
1463                                 if (flow->promisc)
1464                                         break;
1465                         } else {
1466                                 assert(ETH_DEV(priv)->data->all_multicast);
1467                                 if (flow->allmulti)
1468                                         break;
1469                         }
1470                 }
1471                 if (flow && flow->internal) {
1472                         assert(flow->rss);
1473                         if (flow->rss->queues != queues ||
1474                             memcmp(flow->rss->queue_id, action_rss.queue,
1475                                    queues * sizeof(flow->rss->queue_id[0])))
1476                                 flow = NULL;
1477                 }
1478                 if (!flow || !flow->internal) {
1479                         /* Not found, create a new flow rule. */
1480                         if (ETH_DEV(priv)->data->promiscuous) {
1481                                 pattern[1].spec = NULL;
1482                                 pattern[1].mask = NULL;
1483                         } else {
1484                                 assert(ETH_DEV(priv)->data->all_multicast);
1485                                 pattern[1].spec = &eth_allmulti;
1486                                 pattern[1].mask = &eth_allmulti;
1487                         }
1488                         pattern[2] = pattern[3];
1489                         flow = mlx4_flow_create(ETH_DEV(priv), &attr, pattern,
1490                                                 actions, error);
1491                         if (!flow) {
1492                                 err = -rte_errno;
1493                                 goto error;
1494                         }
1495                 }
1496                 assert(flow->promisc || flow->allmulti);
1497                 flow->select = 1;
1498         }
1499 error:
1500         /* Clear selection and clean up stale internal flow rules. */
1501         flow = LIST_FIRST(&priv->flows);
1502         while (flow && flow->internal) {
1503                 struct rte_flow *next = LIST_NEXT(flow, next);
1504
1505                 if (!flow->select)
1506                         claim_zero(mlx4_flow_destroy(ETH_DEV(priv), flow,
1507                                                      error));
1508                 else
1509                         flow->select = 0;
1510                 flow = next;
1511         }
1512         return err;
1513 }
1514
1515 /**
1516  * Synchronize flow rules.
1517  *
1518  * This function synchronizes flow rules with the state of the device by
1519  * taking into account isolated mode and whether target queues are
1520  * configured.
1521  *
1522  * @param priv
1523  *   Pointer to private structure.
1524  * @param[out] error
1525  *   Perform verbose error reporting if not NULL.
1526  *
1527  * @return
1528  *   0 on success, a negative errno value otherwise and rte_errno is set.
1529  */
1530 int
1531 mlx4_flow_sync(struct mlx4_priv *priv, struct rte_flow_error *error)
1532 {
1533         struct rte_flow *flow;
1534         int ret;
1535
1536         /* Internal flow rules are guaranteed to come first in the list. */
1537         if (priv->isolated) {
1538                 /*
1539                  * Get rid of them in isolated mode, stop at the first
1540                  * non-internal rule found.
1541                  */
1542                 for (flow = LIST_FIRST(&priv->flows);
1543                      flow && flow->internal;
1544                      flow = LIST_FIRST(&priv->flows))
1545                         claim_zero(mlx4_flow_destroy(ETH_DEV(priv), flow,
1546                                                      error));
1547         } else {
1548                 /* Refresh internal rules. */
1549                 ret = mlx4_flow_internal(priv, error);
1550                 if (ret)
1551                         return ret;
1552         }
1553         /* Toggle the remaining flow rules . */
1554         LIST_FOREACH(flow, &priv->flows, next) {
1555                 ret = mlx4_flow_toggle(priv, flow, priv->started, error);
1556                 if (ret)
1557                         return ret;
1558         }
1559         if (!priv->started)
1560                 assert(!priv->drop);
1561         return 0;
1562 }
1563
1564 /**
1565  * Clean up all flow rules.
1566  *
1567  * Unlike mlx4_flow_flush(), this function takes care of all remaining flow
1568  * rules regardless of whether they are internal or user-configured.
1569  *
1570  * @param priv
1571  *   Pointer to private structure.
1572  */
1573 void
1574 mlx4_flow_clean(struct mlx4_priv *priv)
1575 {
1576         struct rte_flow *flow;
1577
1578         while ((flow = LIST_FIRST(&priv->flows)))
1579                 mlx4_flow_destroy(ETH_DEV(priv), flow, NULL);
1580         assert(LIST_EMPTY(&priv->rss));
1581 }
1582
1583 static const struct rte_flow_ops mlx4_flow_ops = {
1584         .validate = mlx4_flow_validate,
1585         .create = mlx4_flow_create,
1586         .destroy = mlx4_flow_destroy,
1587         .flush = mlx4_flow_flush,
1588         .isolate = mlx4_flow_isolate,
1589 };
1590
1591 /**
1592  * Manage filter operations.
1593  *
1594  * @param dev
1595  *   Pointer to Ethernet device structure.
1596  * @param filter_type
1597  *   Filter type.
1598  * @param filter_op
1599  *   Operation to perform.
1600  * @param arg
1601  *   Pointer to operation-specific structure.
1602  *
1603  * @return
1604  *   0 on success, negative errno value otherwise and rte_errno is set.
1605  */
1606 int
1607 mlx4_filter_ctrl(struct rte_eth_dev *dev,
1608                  enum rte_filter_type filter_type,
1609                  enum rte_filter_op filter_op,
1610                  void *arg)
1611 {
1612         switch (filter_type) {
1613         case RTE_ETH_FILTER_GENERIC:
1614                 if (filter_op != RTE_ETH_FILTER_GET)
1615                         break;
1616                 *(const void **)arg = &mlx4_flow_ops;
1617                 return 0;
1618         default:
1619                 ERROR("%p: filter type (%d) not supported",
1620                       (void *)dev, filter_type);
1621                 break;
1622         }
1623         rte_errno = ENOTSUP;
1624         return -rte_errno;
1625 }