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 mlx4_priv *priv; /**< Back pointer to private data. */
  75         uint32_t refcnt; /**< Reference count. */
  76 };
  77
  78 /**
  79  * Convert supported RSS hash field types between DPDK and Verbs formats.
  80  *
  81  * This function returns the supported (default) set when @p types has
  82  * special value 0.
  83  *
  84  * @param priv
  85  *   Pointer to private structure.
  86  * @param types
  87  *   Depending on @p verbs_to_dpdk, hash types in either DPDK (see struct
  88  *   rte_eth_rss_conf) or Verbs format.
  89  * @param verbs_to_dpdk
  90  *   A zero value converts @p types from DPDK to Verbs, a nonzero value
  91  *   performs the reverse operation.
  92  *
  93  * @return
  94  *   Converted RSS hash fields on success, (uint64_t)-1 otherwise and
  95  *   rte_errno is set.
  96  */
  97 uint64_t
  98 mlx4_conv_rss_types(struct mlx4_priv *priv, uint64_t types, int verbs_to_dpdk)
  99 {
 100         enum {
 101                 INNER,
 102                 IPV4, IPV4_1, IPV4_2, IPV6, IPV6_1, IPV6_2, IPV6_3,
 103                 TCP, UDP,
 104                 IPV4_TCP, IPV4_UDP, IPV6_TCP, IPV6_TCP_1, IPV6_UDP, IPV6_UDP_1,
 105         };
 106         enum {
 107                 VERBS_IPV4 = IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4,
 108                 VERBS_IPV6 = IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6,
 109                 VERBS_TCP = IBV_RX_HASH_SRC_PORT_TCP | IBV_RX_HASH_DST_PORT_TCP,
 110                 VERBS_UDP = IBV_RX_HASH_SRC_PORT_UDP | IBV_RX_HASH_DST_PORT_UDP,
 111         };
 112         static const uint64_t dpdk[] = {
 113                 [INNER] = 0,
 114                 [IPV4] = ETH_RSS_IPV4,
 115                 [IPV4_1] = ETH_RSS_FRAG_IPV4,
 116                 [IPV4_2] = ETH_RSS_NONFRAG_IPV4_OTHER,
 117                 [IPV6] = ETH_RSS_IPV6,
 118                 [IPV6_1] = ETH_RSS_FRAG_IPV6,
 119                 [IPV6_2] = ETH_RSS_NONFRAG_IPV6_OTHER,
 120                 [IPV6_3] = ETH_RSS_IPV6_EX,
 121                 [TCP] = 0,
 122                 [UDP] = 0,
 123                 [IPV4_TCP] = ETH_RSS_NONFRAG_IPV4_TCP,
 124                 [IPV4_UDP] = ETH_RSS_NONFRAG_IPV4_UDP,
 125                 [IPV6_TCP] = ETH_RSS_NONFRAG_IPV6_TCP,
 126                 [IPV6_TCP_1] = ETH_RSS_IPV6_TCP_EX,
 127                 [IPV6_UDP] = ETH_RSS_NONFRAG_IPV6_UDP,
 128                 [IPV6_UDP_1] = ETH_RSS_IPV6_UDP_EX,
 129         };
 130         static const uint64_t verbs[RTE_DIM(dpdk)] = {
 131                 [INNER] = IBV_RX_HASH_INNER,
 132                 [IPV4] = VERBS_IPV4,
 133                 [IPV4_1] = VERBS_IPV4,
 134                 [IPV4_2] = VERBS_IPV4,
 135                 [IPV6] = VERBS_IPV6,
 136                 [IPV6_1] = VERBS_IPV6,
 137                 [IPV6_2] = VERBS_IPV6,
 138                 [IPV6_3] = VERBS_IPV6,
 139                 [TCP] = VERBS_TCP,
 140                 [UDP] = VERBS_UDP,
 141                 [IPV4_TCP] = VERBS_IPV4 | VERBS_TCP,
 142                 [IPV4_UDP] = VERBS_IPV4 | VERBS_UDP,
 143                 [IPV6_TCP] = VERBS_IPV6 | VERBS_TCP,
 144                 [IPV6_TCP_1] = VERBS_IPV6 | VERBS_TCP,
 145                 [IPV6_UDP] = VERBS_IPV6 | VERBS_UDP,
 146                 [IPV6_UDP_1] = VERBS_IPV6 | VERBS_UDP,
 147         };
 148         const uint64_t *in = verbs_to_dpdk ? verbs : dpdk;
 149         const uint64_t *out = verbs_to_dpdk ? dpdk : verbs;
 150         uint64_t seen = 0;
 151         uint64_t conv = 0;
 152         unsigned int i;
 153
 154         if (!types) {
 155                 if (!verbs_to_dpdk)
 156                         return priv->hw_rss_sup;
 157                 types = priv->hw_rss_sup;
 158         }
 159         for (i = 0; i != RTE_DIM(dpdk); ++i)
 160                 if (in[i] && (types & in[i]) == in[i]) {
 161                         seen |= types & in[i];
 162                         conv |= out[i];
 163                 }
 164         if ((verbs_to_dpdk || (conv & priv->hw_rss_sup) == conv) &&
 165             !(types & ~seen))
 166                 return conv;
 167         rte_errno = ENOTSUP;
 168         return (uint64_t)-1;
 169 }
 170
 171 /**
 172  * Merge Ethernet pattern item into flow rule handle.
 173  *
 174  * Additional mlx4-specific constraints on supported fields:
 175  *
 176  * - No support for partial masks, except in the specific case of matching
 177  *   all multicast traffic (@p spec->dst and @p mask->dst equal to
 178  *   01:00:00:00:00:00).
 179  * - Not providing @p item->spec or providing an empty @p mask->dst is
 180  *   *only* supported if the rule doesn't specify additional matching
 181  *   criteria (i.e. rule is promiscuous-like).
 182  *
 183  * @param[in, out] flow
 184  *   Flow rule handle to update.
 185  * @param[in] item
 186  *   Pattern item to merge.
 187  * @param[in] proc
 188  *   Associated item-processing object.
 189  * @param[out] error
 190  *   Perform verbose error reporting if not NULL.
 191  *
 192  * @return
 193  *   0 on success, a negative errno value otherwise and rte_errno is set.
 194  */
 195 static int
 196 mlx4_flow_merge_eth(struct rte_flow *flow,
 197                     const struct rte_flow_item *item,
 198                     const struct mlx4_flow_proc_item *proc,
 199                     struct rte_flow_error *error)
 200 {
 201         const struct rte_flow_item_eth *spec = item->spec;
 202         const struct rte_flow_item_eth *mask =
 203                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 204         struct ibv_flow_spec_eth *eth;
 205         const char *msg;
 206         unsigned int i;
 207
 208         if (mask) {
 209                 uint32_t sum_dst = 0;
 210                 uint32_t sum_src = 0;
 211
 212                 for (i = 0; i != sizeof(mask->dst.addr_bytes); ++i) {
 213                         sum_dst += mask->dst.addr_bytes[i];
 214                         sum_src += mask->src.addr_bytes[i];
 215                 }
 216                 if (sum_src) {
 217                         msg = "mlx4 does not support source MAC matching";
 218                         goto error;
 219                 } else if (!sum_dst) {
 220                         flow->promisc = 1;
 221                 } else if (sum_dst == 1 && mask->dst.addr_bytes[0] == 1) {
 222                         if (!(spec->dst.addr_bytes[0] & 1)) {
 223                                 msg = "mlx4 does not support the explicit"
 224                                         " exclusion of all multicast traffic";
 225                                 goto error;
 226                         }
 227                         flow->allmulti = 1;
 228                 } else if (sum_dst != (UINT8_C(0xff) * ETHER_ADDR_LEN)) {
 229                         msg = "mlx4 does not support matching partial"
 230                                 " Ethernet fields";
 231                         goto error;
 232                 }
 233         }
 234         if (!flow->ibv_attr)
 235                 return 0;
 236         if (flow->promisc) {
 237                 flow->ibv_attr->type = IBV_FLOW_ATTR_ALL_DEFAULT;
 238                 return 0;
 239         }
 240         if (flow->allmulti) {
 241                 flow->ibv_attr->type = IBV_FLOW_ATTR_MC_DEFAULT;
 242                 return 0;
 243         }
 244         ++flow->ibv_attr->num_of_specs;
 245         eth = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 246         *eth = (struct ibv_flow_spec_eth) {
 247                 .type = IBV_FLOW_SPEC_ETH,
 248                 .size = sizeof(*eth),
 249         };
 250         if (!mask) {
 251                 eth->val.dst_mac[0] = 0xff;
 252                 flow->ibv_attr->type = IBV_FLOW_ATTR_ALL_DEFAULT;
 253                 flow->promisc = 1;
 254                 return 0;
 255         }
 256         memcpy(eth->val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
 257         memcpy(eth->mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
 258         /* Remove unwanted bits from values. */
 259         for (i = 0; i < ETHER_ADDR_LEN; ++i) {
 260                 eth->val.dst_mac[i] &= eth->mask.dst_mac[i];
 261         }
 262         return 0;
 263 error:
 264         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 265                                   item, msg);
 266 }
 267
 268 /**
 269  * Merge VLAN pattern item into flow rule handle.
 270  *
 271  * Additional mlx4-specific constraints on supported fields:
 272  *
 273  * - Matching *all* VLAN traffic by omitting @p item->spec or providing an
 274  *   empty @p item->mask would also include non-VLAN traffic. Doing so is
 275  *   therefore unsupported.
 276  * - No support for partial masks.
 277  *
 278  * @param[in, out] flow
 279  *   Flow rule handle to update.
 280  * @param[in] item
 281  *   Pattern item to merge.
 282  * @param[in] proc
 283  *   Associated item-processing object.
 284  * @param[out] error
 285  *   Perform verbose error reporting if not NULL.
 286  *
 287  * @return
 288  *   0 on success, a negative errno value otherwise and rte_errno is set.
 289  */
 290 static int
 291 mlx4_flow_merge_vlan(struct rte_flow *flow,
 292                      const struct rte_flow_item *item,
 293                      const struct mlx4_flow_proc_item *proc,
 294                      struct rte_flow_error *error)
 295 {
 296         const struct rte_flow_item_vlan *spec = item->spec;
 297         const struct rte_flow_item_vlan *mask =
 298                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 299         struct ibv_flow_spec_eth *eth;
 300         const char *msg;
 301
 302         if (!mask || !mask->tci) {
 303                 msg = "mlx4 cannot match all VLAN traffic while excluding"
 304                         " non-VLAN traffic, TCI VID must be specified";
 305                 goto error;
 306         }
 307         if (mask->tci != RTE_BE16(0x0fff)) {
 308                 msg = "mlx4 does not support partial TCI VID matching";
 309                 goto error;
 310         }
 311         if (!flow->ibv_attr)
 312                 return 0;
 313         eth = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size -
 314                        sizeof(*eth));
 315         eth->val.vlan_tag = spec->tci;
 316         eth->mask.vlan_tag = mask->tci;
 317         eth->val.vlan_tag &= eth->mask.vlan_tag;
 318         if (flow->ibv_attr->type == IBV_FLOW_ATTR_ALL_DEFAULT)
 319                 flow->ibv_attr->type = IBV_FLOW_ATTR_NORMAL;
 320         return 0;
 321 error:
 322         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 323                                   item, msg);
 324 }
 325
 326 /**
 327  * Merge IPv4 pattern item into flow rule handle.
 328  *
 329  * Additional mlx4-specific constraints on supported fields:
 330  *
 331  * - No support for partial masks.
 332  *
 333  * @param[in, out] flow
 334  *   Flow rule handle to update.
 335  * @param[in] item
 336  *   Pattern item to merge.
 337  * @param[in] proc
 338  *   Associated item-processing object.
 339  * @param[out] error
 340  *   Perform verbose error reporting if not NULL.
 341  *
 342  * @return
 343  *   0 on success, a negative errno value otherwise and rte_errno is set.
 344  */
 345 static int
 346 mlx4_flow_merge_ipv4(struct rte_flow *flow,
 347                      const struct rte_flow_item *item,
 348                      const struct mlx4_flow_proc_item *proc,
 349                      struct rte_flow_error *error)
 350 {
 351         const struct rte_flow_item_ipv4 *spec = item->spec;
 352         const struct rte_flow_item_ipv4 *mask =
 353                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 354         struct ibv_flow_spec_ipv4 *ipv4;
 355         const char *msg;
 356
 357         if (mask &&
 358             ((uint32_t)(mask->hdr.src_addr + 1) > UINT32_C(1) ||
 359              (uint32_t)(mask->hdr.dst_addr + 1) > UINT32_C(1))) {
 360                 msg = "mlx4 does not support matching partial IPv4 fields";
 361                 goto error;
 362         }
 363         if (!flow->ibv_attr)
 364                 return 0;
 365         ++flow->ibv_attr->num_of_specs;
 366         ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 367         *ipv4 = (struct ibv_flow_spec_ipv4) {
 368                 .type = IBV_FLOW_SPEC_IPV4,
 369                 .size = sizeof(*ipv4),
 370         };
 371         if (!spec)
 372                 return 0;
 373         ipv4->val = (struct ibv_flow_ipv4_filter) {
 374                 .src_ip = spec->hdr.src_addr,
 375                 .dst_ip = spec->hdr.dst_addr,
 376         };
 377         ipv4->mask = (struct ibv_flow_ipv4_filter) {
 378                 .src_ip = mask->hdr.src_addr,
 379                 .dst_ip = mask->hdr.dst_addr,
 380         };
 381         /* Remove unwanted bits from values. */
 382         ipv4->val.src_ip &= ipv4->mask.src_ip;
 383         ipv4->val.dst_ip &= ipv4->mask.dst_ip;
 384         return 0;
 385 error:
 386         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 387                                   item, msg);
 388 }
 389
 390 /**
 391  * Merge UDP pattern item into flow rule handle.
 392  *
 393  * Additional mlx4-specific constraints on supported fields:
 394  *
 395  * - No support for partial masks.
 396  * - Due to HW/FW limitation, flow rule priority is not taken into account
 397  *   when matching UDP destination ports, doing is therefore only supported
 398  *   at the highest priority level (0).
 399  *
 400  * @param[in, out] flow
 401  *   Flow rule handle to update.
 402  * @param[in] item
 403  *   Pattern item to merge.
 404  * @param[in] proc
 405  *   Associated item-processing object.
 406  * @param[out] error
 407  *   Perform verbose error reporting if not NULL.
 408  *
 409  * @return
 410  *   0 on success, a negative errno value otherwise and rte_errno is set.
 411  */
 412 static int
 413 mlx4_flow_merge_udp(struct rte_flow *flow,
 414                     const struct rte_flow_item *item,
 415                     const struct mlx4_flow_proc_item *proc,
 416                     struct rte_flow_error *error)
 417 {
 418         const struct rte_flow_item_udp *spec = item->spec;
 419         const struct rte_flow_item_udp *mask =
 420                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 421         struct ibv_flow_spec_tcp_udp *udp;
 422         const char *msg;
 423
 424         if (mask &&
 425             ((uint16_t)(mask->hdr.src_port + 1) > UINT16_C(1) ||
 426              (uint16_t)(mask->hdr.dst_port + 1) > UINT16_C(1))) {
 427                 msg = "mlx4 does not support matching partial UDP fields";
 428                 goto error;
 429         }
 430         if (mask && mask->hdr.dst_port && flow->priority) {
 431                 msg = "combining UDP destination port matching with a nonzero"
 432                         " priority level is not supported";
 433                 goto error;
 434         }
 435         if (!flow->ibv_attr)
 436                 return 0;
 437         ++flow->ibv_attr->num_of_specs;
 438         udp = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 439         *udp = (struct ibv_flow_spec_tcp_udp) {
 440                 .type = IBV_FLOW_SPEC_UDP,
 441                 .size = sizeof(*udp),
 442         };
 443         if (!spec)
 444                 return 0;
 445         udp->val.dst_port = spec->hdr.dst_port;
 446         udp->val.src_port = spec->hdr.src_port;
 447         udp->mask.dst_port = mask->hdr.dst_port;
 448         udp->mask.src_port = mask->hdr.src_port;
 449         /* Remove unwanted bits from values. */
 450         udp->val.src_port &= udp->mask.src_port;
 451         udp->val.dst_port &= udp->mask.dst_port;
 452         return 0;
 453 error:
 454         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 455                                   item, msg);
 456 }
 457
 458 /**
 459  * Merge TCP pattern item into flow rule handle.
 460  *
 461  * Additional mlx4-specific constraints on supported fields:
 462  *
 463  * - No support for partial masks.
 464  *
 465  * @param[in, out] flow
 466  *   Flow rule handle to update.
 467  * @param[in] item
 468  *   Pattern item to merge.
 469  * @param[in] proc
 470  *   Associated item-processing object.
 471  * @param[out] error
 472  *   Perform verbose error reporting if not NULL.
 473  *
 474  * @return
 475  *   0 on success, a negative errno value otherwise and rte_errno is set.
 476  */
 477 static int
 478 mlx4_flow_merge_tcp(struct rte_flow *flow,
 479                     const struct rte_flow_item *item,
 480                     const struct mlx4_flow_proc_item *proc,
 481                     struct rte_flow_error *error)
 482 {
 483         const struct rte_flow_item_tcp *spec = item->spec;
 484         const struct rte_flow_item_tcp *mask =
 485                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 486         struct ibv_flow_spec_tcp_udp *tcp;
 487         const char *msg;
 488
 489         if (mask &&
 490             ((uint16_t)(mask->hdr.src_port + 1) > UINT16_C(1) ||
 491              (uint16_t)(mask->hdr.dst_port + 1) > UINT16_C(1))) {
 492                 msg = "mlx4 does not support matching partial TCP fields";
 493                 goto error;
 494         }
 495         if (!flow->ibv_attr)
 496                 return 0;
 497         ++flow->ibv_attr->num_of_specs;
 498         tcp = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 499         *tcp = (struct ibv_flow_spec_tcp_udp) {
 500                 .type = IBV_FLOW_SPEC_TCP,
 501                 .size = sizeof(*tcp),
 502         };
 503         if (!spec)
 504                 return 0;
 505         tcp->val.dst_port = spec->hdr.dst_port;
 506         tcp->val.src_port = spec->hdr.src_port;
 507         tcp->mask.dst_port = mask->hdr.dst_port;
 508         tcp->mask.src_port = mask->hdr.src_port;
 509         /* Remove unwanted bits from values. */
 510         tcp->val.src_port &= tcp->mask.src_port;
 511         tcp->val.dst_port &= tcp->mask.dst_port;
 512         return 0;
 513 error:
 514         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 515                                   item, msg);
 516 }
 517
 518 /**
 519  * Perform basic sanity checks on a pattern item.
 520  *
 521  * @param[in] item
 522  *   Item specification.
 523  * @param[in] proc
 524  *   Associated item-processing object.
 525  * @param[out] error
 526  *   Perform verbose error reporting if not NULL.
 527  *
 528  * @return
 529  *   0 on success, a negative errno value otherwise and rte_errno is set.
 530  */
 531 static int
 532 mlx4_flow_item_check(const struct rte_flow_item *item,
 533                      const struct mlx4_flow_proc_item *proc,
 534                      struct rte_flow_error *error)
 535 {
 536         const uint8_t *mask;
 537         unsigned int i;
 538
 539         /* item->last and item->mask cannot exist without item->spec. */
 540         if (!item->spec && (item->mask || item->last))
 541                 return rte_flow_error_set
 542                         (error, EINVAL, RTE_FLOW_ERROR_TYPE_ITEM, item,
 543                          "\"mask\" or \"last\" field provided without a"
 544                          " corresponding \"spec\"");
 545         /* No spec, no mask, no problem. */
 546         if (!item->spec)
 547                 return 0;
 548         mask = item->mask ?
 549                 (const uint8_t *)item->mask :
 550                 (const uint8_t *)proc->mask_default;
 551         assert(mask);
 552         /*
 553          * Single-pass check to make sure that:
 554          * - Mask is supported, no bits are set outside proc->mask_support.
 555          * - Both item->spec and item->last are included in mask.
 556          */
 557         for (i = 0; i != proc->mask_sz; ++i) {
 558                 if (!mask[i])
 559                         continue;
 560                 if ((mask[i] | ((const uint8_t *)proc->mask_support)[i]) !=
 561                     ((const uint8_t *)proc->mask_support)[i])
 562                         return rte_flow_error_set
 563                                 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 564                                  item, "unsupported field found in \"mask\"");
 565                 if (item->last &&
 566                     (((const uint8_t *)item->spec)[i] & mask[i]) !=
 567                     (((const uint8_t *)item->last)[i] & mask[i]))
 568                         return rte_flow_error_set
 569                                 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 570                                  item,
 571                                  "range between \"spec\" and \"last\""
 572                                  " is larger than \"mask\"");
 573         }
 574         return 0;
 575 }
 576
 577 /** Graph of supported items and associated actions. */
 578 static const struct mlx4_flow_proc_item mlx4_flow_proc_item_list[] = {
 579         [RTE_FLOW_ITEM_TYPE_END] = {
 580                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_ETH),
 581         },
 582         [RTE_FLOW_ITEM_TYPE_ETH] = {
 583                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_VLAN,
 584                                        RTE_FLOW_ITEM_TYPE_IPV4),
 585                 .mask_support = &(const struct rte_flow_item_eth){
 586                         /* Only destination MAC can be matched. */
 587                         .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 588                 },
 589                 .mask_default = &rte_flow_item_eth_mask,
 590                 .mask_sz = sizeof(struct rte_flow_item_eth),
 591                 .merge = mlx4_flow_merge_eth,
 592                 .dst_sz = sizeof(struct ibv_flow_spec_eth),
 593         },
 594         [RTE_FLOW_ITEM_TYPE_VLAN] = {
 595                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_IPV4),
 596                 .mask_support = &(const struct rte_flow_item_vlan){
 597                         /* Only TCI VID matching is supported. */
 598                         .tci = RTE_BE16(0x0fff),
 599                 },
 600                 .mask_default = &rte_flow_item_vlan_mask,
 601                 .mask_sz = sizeof(struct rte_flow_item_vlan),
 602                 .merge = mlx4_flow_merge_vlan,
 603                 .dst_sz = 0,
 604         },
 605         [RTE_FLOW_ITEM_TYPE_IPV4] = {
 606                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_UDP,
 607                                        RTE_FLOW_ITEM_TYPE_TCP),
 608                 .mask_support = &(const struct rte_flow_item_ipv4){
 609                         .hdr = {
 610                                 .src_addr = RTE_BE32(0xffffffff),
 611                                 .dst_addr = RTE_BE32(0xffffffff),
 612                         },
 613                 },
 614                 .mask_default = &rte_flow_item_ipv4_mask,
 615                 .mask_sz = sizeof(struct rte_flow_item_ipv4),
 616                 .merge = mlx4_flow_merge_ipv4,
 617                 .dst_sz = sizeof(struct ibv_flow_spec_ipv4),
 618         },
 619         [RTE_FLOW_ITEM_TYPE_UDP] = {
 620                 .mask_support = &(const struct rte_flow_item_udp){
 621                         .hdr = {
 622                                 .src_port = RTE_BE16(0xffff),
 623                                 .dst_port = RTE_BE16(0xffff),
 624                         },
 625                 },
 626                 .mask_default = &rte_flow_item_udp_mask,
 627                 .mask_sz = sizeof(struct rte_flow_item_udp),
 628                 .merge = mlx4_flow_merge_udp,
 629                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 630         },
 631         [RTE_FLOW_ITEM_TYPE_TCP] = {
 632                 .mask_support = &(const struct rte_flow_item_tcp){
 633                         .hdr = {
 634                                 .src_port = RTE_BE16(0xffff),
 635                                 .dst_port = RTE_BE16(0xffff),
 636                         },
 637                 },
 638                 .mask_default = &rte_flow_item_tcp_mask,
 639                 .mask_sz = sizeof(struct rte_flow_item_tcp),
 640                 .merge = mlx4_flow_merge_tcp,
 641                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 642         },
 643 };
 644
 645 /**
 646  * Make sure a flow rule is supported and initialize associated structure.
 647  *
 648  * @param priv
 649  *   Pointer to private structure.
 650  * @param[in] attr
 651  *   Flow rule attributes.
 652  * @param[in] pattern
 653  *   Pattern specification (list terminated by the END pattern item).
 654  * @param[in] actions
 655  *   Associated actions (list terminated by the END action).
 656  * @param[out] error
 657  *   Perform verbose error reporting if not NULL.
 658  * @param[in, out] addr
 659  *   Buffer where the resulting flow rule handle pointer must be stored.
 660  *   If NULL, stop processing after validation stage.
 661  *
 662  * @return
 663  *   0 on success, a negative errno value otherwise and rte_errno is set.
 664  */
 665 static int
 666 mlx4_flow_prepare(struct mlx4_priv *priv,
 667                   const struct rte_flow_attr *attr,
 668                   const struct rte_flow_item pattern[],
 669                   const struct rte_flow_action actions[],
 670                   struct rte_flow_error *error,
 671                   struct rte_flow **addr)
 672 {
 673         const struct rte_flow_item *item;
 674         const struct rte_flow_action *action;
 675         const struct mlx4_flow_proc_item *proc;
 676         struct rte_flow temp = { .ibv_attr_size = sizeof(*temp.ibv_attr) };
 677         struct rte_flow *flow = &temp;
 678         const char *msg = NULL;
 679         int overlap;
 680
 681         if (attr->group)
 682                 return rte_flow_error_set
 683                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 684                          NULL, "groups are not supported");
 685         if (attr->priority > MLX4_FLOW_PRIORITY_LAST)
 686                 return rte_flow_error_set
 687                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 688                          NULL, "maximum priority level is "
 689                          MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST));
 690         if (attr->egress)
 691                 return rte_flow_error_set
 692                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 693                          NULL, "egress is not supported");
 694         if (attr->transfer)
 695                 return rte_flow_error_set
 696                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_TRANSFER,
 697                          NULL, "transfer is not supported");
 698         if (!attr->ingress)
 699                 return rte_flow_error_set
 700                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 701                          NULL, "only ingress is supported");
 702 fill:
 703         overlap = 0;
 704         proc = mlx4_flow_proc_item_list;
 705         flow->priority = attr->priority;
 706         /* Go over pattern. */
 707         for (item = pattern; item->type; ++item) {
 708                 const struct mlx4_flow_proc_item *next = NULL;
 709                 unsigned int i;
 710                 int err;
 711
 712                 if (item->type == RTE_FLOW_ITEM_TYPE_VOID)
 713                         continue;
 714                 if (item->type == MLX4_FLOW_ITEM_TYPE_INTERNAL) {
 715                         flow->internal = 1;
 716                         continue;
 717                 }
 718                 if (flow->promisc || flow->allmulti) {
 719                         msg = "mlx4 does not support additional matching"
 720                                 " criteria combined with indiscriminate"
 721                                 " matching on Ethernet headers";
 722                         goto exit_item_not_supported;
 723                 }
 724                 for (i = 0; proc->next_item && proc->next_item[i]; ++i) {
 725                         if (proc->next_item[i] == item->type) {
 726                                 next = &mlx4_flow_proc_item_list[item->type];
 727                                 break;
 728                         }
 729                 }
 730                 if (!next)
 731                         goto exit_item_not_supported;
 732                 proc = next;
 733                 /*
 734                  * Perform basic sanity checks only once, while handle is
 735                  * not allocated.
 736                  */
 737                 if (flow == &temp) {
 738                         err = mlx4_flow_item_check(item, proc, error);
 739                         if (err)
 740                                 return err;
 741                 }
 742                 if (proc->merge) {
 743                         err = proc->merge(flow, item, proc, error);
 744                         if (err)
 745                                 return err;
 746                 }
 747                 flow->ibv_attr_size += proc->dst_sz;
 748         }
 749         /* Go over actions list. */
 750         for (action = actions; action->type; ++action) {
 751                 /* This one may appear anywhere multiple times. */
 752                 if (action->type == RTE_FLOW_ACTION_TYPE_VOID)
 753                         continue;
 754                 /* Fate-deciding actions may appear exactly once. */
 755                 if (overlap) {
 756                         msg = "cannot combine several fate-deciding actions,"
 757                                 " choose between DROP, QUEUE or RSS";
 758                         goto exit_action_not_supported;
 759                 }
 760                 overlap = 1;
 761                 switch (action->type) {
 762                         const struct rte_flow_action_queue *queue;
 763                         const struct rte_flow_action_rss *rss;
 764                         const uint8_t *rss_key;
 765                         uint32_t rss_key_len;
 766                         uint64_t fields;
 767                         unsigned int i;
 768
 769                 case RTE_FLOW_ACTION_TYPE_DROP:
 770                         flow->drop = 1;
 771                         break;
 772                 case RTE_FLOW_ACTION_TYPE_QUEUE:
 773                         if (flow->rss)
 774                                 break;
 775                         queue = action->conf;
 776                         if (queue->index >= ETH_DEV(priv)->data->nb_rx_queues) {
 777                                 msg = "queue target index beyond number of"
 778                                         " configured Rx queues";
 779                                 goto exit_action_not_supported;
 780                         }
 781                         flow->rss = mlx4_rss_get
 782                                 (priv, 0, mlx4_rss_hash_key_default, 1,
 783                                  &queue->index);
 784                         if (!flow->rss) {
 785                                 msg = "not enough resources for additional"
 786                                         " single-queue RSS context";
 787                                 goto exit_action_not_supported;
 788                         }
 789                         break;
 790                 case RTE_FLOW_ACTION_TYPE_RSS:
 791                         if (flow->rss)
 792                                 break;
 793                         rss = action->conf;
 794                         /* Default RSS configuration if none is provided. */
 795                         if (rss->key_len) {
 796                                 rss_key = rss->key;
 797                                 rss_key_len = rss->key_len;
 798                         } else {
 799                                 rss_key = mlx4_rss_hash_key_default;
 800                                 rss_key_len = MLX4_RSS_HASH_KEY_SIZE;
 801                         }
 802                         /* Sanity checks. */
 803                         for (i = 0; i < rss->queue_num; ++i)
 804                                 if (rss->queue[i] >=
 805                                     ETH_DEV(priv)->data->nb_rx_queues)
 806                                         break;
 807                         if (i != rss->queue_num) {
 808                                 msg = "queue index target beyond number of"
 809                                         " configured Rx queues";
 810                                 goto exit_action_not_supported;
 811                         }
 812                         if (!rte_is_power_of_2(rss->queue_num)) {
 813                                 msg = "for RSS, mlx4 requires the number of"
 814                                         " queues to be a power of two";
 815                                 goto exit_action_not_supported;
 816                         }
 817                         if (rss_key_len != sizeof(flow->rss->key)) {
 818                                 msg = "mlx4 supports exactly one RSS hash key"
 819                                         " length: "
 820                                         MLX4_STR_EXPAND(MLX4_RSS_HASH_KEY_SIZE);
 821                                 goto exit_action_not_supported;
 822                         }
 823                         for (i = 1; i < rss->queue_num; ++i)
 824                                 if (rss->queue[i] - rss->queue[i - 1] != 1)
 825                                         break;
 826                         if (i != rss->queue_num) {
 827                                 msg = "mlx4 requires RSS contexts to use"
 828                                         " consecutive queue indices only";
 829                                 goto exit_action_not_supported;
 830                         }
 831                         if (rss->queue[0] % rss->queue_num) {
 832                                 msg = "mlx4 requires the first queue of a RSS"
 833                                         " context to be aligned on a multiple"
 834                                         " of the context size";
 835                                 goto exit_action_not_supported;
 836                         }
 837                         if (rss->func &&
 838                             rss->func != RTE_ETH_HASH_FUNCTION_TOEPLITZ) {
 839                                 msg = "the only supported RSS hash function"
 840                                         " is Toeplitz";
 841                                 goto exit_action_not_supported;
 842                         }
 843                         if (rss->level) {
 844                                 msg = "a nonzero RSS encapsulation level is"
 845                                         " not supported";
 846                                 goto exit_action_not_supported;
 847                         }
 848                         rte_errno = 0;
 849                         fields = mlx4_conv_rss_types(priv, rss->types, 0);
 850                         if (fields == (uint64_t)-1 && rte_errno) {
 851                                 msg = "unsupported RSS hash type requested";
 852                                 goto exit_action_not_supported;
 853                         }
 854                         flow->rss = mlx4_rss_get
 855                                 (priv, fields, rss_key, rss->queue_num,
 856                                  rss->queue);
 857                         if (!flow->rss) {
 858                                 msg = "either invalid parameters or not enough"
 859                                         " resources for additional multi-queue"
 860                                         " RSS context";
 861                                 goto exit_action_not_supported;
 862                         }
 863                         break;
 864                 default:
 865                         goto exit_action_not_supported;
 866                 }
 867         }
 868         /* When fate is unknown, drop traffic. */
 869         if (!overlap)
 870                 flow->drop = 1;
 871         /* Validation ends here. */
 872         if (!addr) {
 873                 if (flow->rss)
 874                         mlx4_rss_put(flow->rss);
 875                 return 0;
 876         }
 877         if (flow == &temp) {
 878                 /* Allocate proper handle based on collected data. */
 879                 const struct mlx4_malloc_vec vec[] = {
 880                         {
 881                                 .align = alignof(struct rte_flow),
 882                                 .size = sizeof(*flow),
 883                                 .addr = (void **)&flow,
 884                         },
 885                         {
 886                                 .align = alignof(struct ibv_flow_attr),
 887                                 .size = temp.ibv_attr_size,
 888                                 .addr = (void **)&temp.ibv_attr,
 889                         },
 890                 };
 891
 892                 if (!mlx4_zmallocv(__func__, vec, RTE_DIM(vec))) {
 893                         if (temp.rss)
 894                                 mlx4_rss_put(temp.rss);
 895                         return rte_flow_error_set
 896                                 (error, -rte_errno,
 897                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
 898                                  "flow rule handle allocation failure");
 899                 }
 900                 /* Most fields will be updated by second pass. */
 901                 *flow = (struct rte_flow){
 902                         .ibv_attr = temp.ibv_attr,
 903                         .ibv_attr_size = sizeof(*flow->ibv_attr),
 904                         .rss = temp.rss,
 905                 };
 906                 *flow->ibv_attr = (struct ibv_flow_attr){
 907                         .type = IBV_FLOW_ATTR_NORMAL,
 908                         .size = sizeof(*flow->ibv_attr),
 909                         .priority = attr->priority,
 910                         .port = priv->port,
 911                 };
 912                 goto fill;
 913         }
 914         *addr = flow;
 915         return 0;
 916 exit_item_not_supported:
 917         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 918                                   item, msg ? msg : "item not supported");
 919 exit_action_not_supported:
 920         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 921                                   action, msg ? msg : "action not supported");
 922 }
 923
 924 /**
 925  * Validate a flow supported by the NIC.
 926  *
 927  * @see rte_flow_validate()
 928  * @see rte_flow_ops
 929  */
 930 static int
 931 mlx4_flow_validate(struct rte_eth_dev *dev,
 932                    const struct rte_flow_attr *attr,
 933                    const struct rte_flow_item pattern[],
 934                    const struct rte_flow_action actions[],
 935                    struct rte_flow_error *error)
 936 {
 937         struct mlx4_priv *priv = dev->data->dev_private;
 938
 939         return mlx4_flow_prepare(priv, attr, pattern, actions, error, NULL);
 940 }
 941
 942 /**
 943  * Get a drop flow rule resources instance.
 944  *
 945  * @param priv
 946  *   Pointer to private structure.
 947  *
 948  * @return
 949  *   Pointer to drop flow resources on success, NULL otherwise and rte_errno
 950  *   is set.
 951  */
 952 static struct mlx4_drop *
 953 mlx4_drop_get(struct mlx4_priv *priv)
 954 {
 955         struct mlx4_drop *drop = priv->drop;
 956
 957         if (drop) {
 958                 assert(drop->refcnt);
 959                 assert(drop->priv == priv);
 960                 ++drop->refcnt;
 961                 return drop;
 962         }
 963         drop = rte_malloc(__func__, sizeof(*drop), 0);
 964         if (!drop)
 965                 goto error;
 966         *drop = (struct mlx4_drop){
 967                 .priv = priv,
 968                 .refcnt = 1,
 969         };
 970         drop->cq = mlx4_glue->create_cq(priv->ctx, 1, NULL, NULL, 0);
 971         if (!drop->cq)
 972                 goto error;
 973         drop->qp = mlx4_glue->create_qp
 974                 (priv->pd,
 975                  &(struct ibv_qp_init_attr){
 976                         .send_cq = drop->cq,
 977                         .recv_cq = drop->cq,
 978                         .qp_type = IBV_QPT_RAW_PACKET,
 979                  });
 980         if (!drop->qp)
 981                 goto error;
 982         priv->drop = drop;
 983         return drop;
 984 error:
 985         if (drop->qp)
 986                 claim_zero(mlx4_glue->destroy_qp(drop->qp));
 987         if (drop->cq)
 988                 claim_zero(mlx4_glue->destroy_cq(drop->cq));
 989         if (drop)
 990                 rte_free(drop);
 991         rte_errno = ENOMEM;
 992         return NULL;
 993 }
 994
 995 /**
 996  * Give back a drop flow rule resources instance.
 997  *
 998  * @param drop
 999  *   Pointer to drop flow rule resources.
1000  */
1001 static void
1002 mlx4_drop_put(struct mlx4_drop *drop)
1003 {
1004         assert(drop->refcnt);
1005         if (--drop->refcnt)
1006                 return;
1007         drop->priv->drop = NULL;
1008         claim_zero(mlx4_glue->destroy_qp(drop->qp));
1009         claim_zero(mlx4_glue->destroy_cq(drop->cq));
1010         rte_free(drop);
1011 }
1012
1013 /**
1014  * Toggle a configured flow rule.
1015  *
1016  * @param priv
1017  *   Pointer to private structure.
1018  * @param flow
1019  *   Flow rule handle to toggle.
1020  * @param enable
1021  *   Whether associated Verbs flow must be created or removed.
1022  * @param[out] error
1023  *   Perform verbose error reporting if not NULL.
1024  *
1025  * @return
1026  *   0 on success, a negative errno value otherwise and rte_errno is set.
1027  */
1028 static int
1029 mlx4_flow_toggle(struct mlx4_priv *priv,
1030                  struct rte_flow *flow,
1031                  int enable,
1032                  struct rte_flow_error *error)
1033 {
1034         struct ibv_qp *qp = NULL;
1035         const char *msg;
1036         int err;
1037
1038         if (!enable) {
1039                 if (!flow->ibv_flow)
1040                         return 0;
1041                 claim_zero(mlx4_glue->destroy_flow(flow->ibv_flow));
1042                 flow->ibv_flow = NULL;
1043                 if (flow->drop)
1044                         mlx4_drop_put(priv->drop);
1045                 else if (flow->rss)
1046                         mlx4_rss_detach(flow->rss);
1047                 return 0;
1048         }
1049         assert(flow->ibv_attr);
1050         if (!flow->internal &&
1051             !priv->isolated &&
1052             flow->ibv_attr->priority == MLX4_FLOW_PRIORITY_LAST) {
1053                 if (flow->ibv_flow) {
1054                         claim_zero(mlx4_glue->destroy_flow(flow->ibv_flow));
1055                         flow->ibv_flow = NULL;
1056                         if (flow->drop)
1057                                 mlx4_drop_put(priv->drop);
1058                         else if (flow->rss)
1059                                 mlx4_rss_detach(flow->rss);
1060                 }
1061                 err = EACCES;
1062                 msg = ("priority level "
1063                        MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST)
1064                        " is reserved when not in isolated mode");
1065                 goto error;
1066         }
1067         if (flow->rss) {
1068                 struct mlx4_rss *rss = flow->rss;
1069                 int missing = 0;
1070                 unsigned int i;
1071
1072                 /* Stop at the first nonexistent target queue. */
1073                 for (i = 0; i != rss->queues; ++i)
1074                         if (rss->queue_id[i] >=
1075                             ETH_DEV(priv)->data->nb_rx_queues ||
1076                             !ETH_DEV(priv)->data->rx_queues[rss->queue_id[i]]) {
1077                                 missing = 1;
1078                                 break;
1079                         }
1080                 if (flow->ibv_flow) {
1081                         if (missing ^ !flow->drop)
1082                                 return 0;
1083                         /* Verbs flow needs updating. */
1084                         claim_zero(mlx4_glue->destroy_flow(flow->ibv_flow));
1085                         flow->ibv_flow = NULL;
1086                         if (flow->drop)
1087                                 mlx4_drop_put(priv->drop);
1088                         else
1089                                 mlx4_rss_detach(rss);
1090                 }
1091                 if (!missing) {
1092                         err = mlx4_rss_attach(rss);
1093                         if (err) {
1094                                 err = -err;
1095                                 msg = "cannot create indirection table or hash"
1096                                         " QP to associate flow rule with";
1097                                 goto error;
1098                         }
1099                         qp = rss->qp;
1100                 }
1101                 /* A missing target queue drops traffic implicitly. */
1102                 flow->drop = missing;
1103         }
1104         if (flow->drop) {
1105                 if (flow->ibv_flow)
1106                         return 0;
1107                 mlx4_drop_get(priv);
1108                 if (!priv->drop) {
1109                         err = rte_errno;
1110                         msg = "resources for drop flow rule cannot be created";
1111                         goto error;
1112                 }
1113                 qp = priv->drop->qp;
1114         }
1115         assert(qp);
1116         if (flow->ibv_flow)
1117                 return 0;
1118         flow->ibv_flow = mlx4_glue->create_flow(qp, flow->ibv_attr);
1119         if (flow->ibv_flow)
1120                 return 0;
1121         if (flow->drop)
1122                 mlx4_drop_put(priv->drop);
1123         else if (flow->rss)
1124                 mlx4_rss_detach(flow->rss);
1125         err = errno;
1126         msg = "flow rule rejected by device";
1127 error:
1128         return rte_flow_error_set
1129                 (error, err, RTE_FLOW_ERROR_TYPE_HANDLE, flow, msg);
1130 }
1131
1132 /**
1133  * Create a flow.
1134  *
1135  * @see rte_flow_create()
1136  * @see rte_flow_ops
1137  */
1138 static struct rte_flow *
1139 mlx4_flow_create(struct rte_eth_dev *dev,
1140                  const struct rte_flow_attr *attr,
1141                  const struct rte_flow_item pattern[],
1142                  const struct rte_flow_action actions[],
1143                  struct rte_flow_error *error)
1144 {
1145         struct mlx4_priv *priv = dev->data->dev_private;
1146         struct rte_flow *flow;
1147         int err;
1148
1149         err = mlx4_flow_prepare(priv, attr, pattern, actions, error, &flow);
1150         if (err)
1151                 return NULL;
1152         err = mlx4_flow_toggle(priv, flow, priv->started, error);
1153         if (!err) {
1154                 struct rte_flow *curr = LIST_FIRST(&priv->flows);
1155
1156                 /* New rules are inserted after internal ones. */
1157                 if (!curr || !curr->internal) {
1158                         LIST_INSERT_HEAD(&priv->flows, flow, next);
1159                 } else {
1160                         while (LIST_NEXT(curr, next) &&
1161                                LIST_NEXT(curr, next)->internal)
1162                                 curr = LIST_NEXT(curr, next);
1163                         LIST_INSERT_AFTER(curr, flow, next);
1164                 }
1165                 return flow;
1166         }
1167         if (flow->rss)
1168                 mlx4_rss_put(flow->rss);
1169         rte_flow_error_set(error, -err, RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1170                            error->message);
1171         rte_free(flow);
1172         return NULL;
1173 }
1174
1175 /**
1176  * Configure isolated mode.
1177  *
1178  * @see rte_flow_isolate()
1179  * @see rte_flow_ops
1180  */
1181 static int
1182 mlx4_flow_isolate(struct rte_eth_dev *dev,
1183                   int enable,
1184                   struct rte_flow_error *error)
1185 {
1186         struct mlx4_priv *priv = dev->data->dev_private;
1187
1188         if (!!enable == !!priv->isolated)
1189                 return 0;
1190         priv->isolated = !!enable;
1191         if (mlx4_flow_sync(priv, error)) {
1192                 priv->isolated = !enable;
1193                 return -rte_errno;
1194         }
1195         return 0;
1196 }
1197
1198 /**
1199  * Destroy a flow rule.
1200  *
1201  * @see rte_flow_destroy()
1202  * @see rte_flow_ops
1203  */
1204 static int
1205 mlx4_flow_destroy(struct rte_eth_dev *dev,
1206                   struct rte_flow *flow,
1207                   struct rte_flow_error *error)
1208 {
1209         struct mlx4_priv *priv = dev->data->dev_private;
1210         int err = mlx4_flow_toggle(priv, flow, 0, error);
1211
1212         if (err)
1213                 return err;
1214         LIST_REMOVE(flow, next);
1215         if (flow->rss)
1216                 mlx4_rss_put(flow->rss);
1217         rte_free(flow);
1218         return 0;
1219 }
1220
1221 /**
1222  * Destroy user-configured flow rules.
1223  *
1224  * This function skips internal flows rules.
1225  *
1226  * @see rte_flow_flush()
1227  * @see rte_flow_ops
1228  */
1229 static int
1230 mlx4_flow_flush(struct rte_eth_dev *dev,
1231                 struct rte_flow_error *error)
1232 {
1233         struct mlx4_priv *priv = dev->data->dev_private;
1234         struct rte_flow *flow = LIST_FIRST(&priv->flows);
1235
1236         while (flow) {
1237                 struct rte_flow *next = LIST_NEXT(flow, next);
1238
1239                 if (!flow->internal)
1240                         mlx4_flow_destroy(dev, flow, error);
1241                 flow = next;
1242         }
1243         return 0;
1244 }
1245
1246 /**
1247  * Helper function to determine the next configured VLAN filter.
1248  *
1249  * @param priv
1250  *   Pointer to private structure.
1251  * @param vlan
1252  *   VLAN ID to use as a starting point.
1253  *
1254  * @return
1255  *   Next configured VLAN ID or a high value (>= 4096) if there is none.
1256  */
1257 static uint16_t
1258 mlx4_flow_internal_next_vlan(struct mlx4_priv *priv, uint16_t vlan)
1259 {
1260         while (vlan < 4096) {
1261                 if (ETH_DEV(priv)->data->vlan_filter_conf.ids[vlan / 64] &
1262                     (UINT64_C(1) << (vlan % 64)))
1263                         return vlan;
1264                 ++vlan;
1265         }
1266         return vlan;
1267 }
1268
1269 /**
1270  * Generate internal flow rules.
1271  *
1272  * Various flow rules are created depending on the mode the device is in:
1273  *
1274  * 1. Promiscuous:
1275  *       port MAC + broadcast + catch-all (VLAN filtering is ignored).
1276  * 2. All multicast:
1277  *       port MAC/VLAN + broadcast + catch-all multicast.
1278  * 3. Otherwise:
1279  *       port MAC/VLAN + broadcast MAC/VLAN.
1280  *
1281  * About MAC flow rules:
1282  *
1283  * - MAC flow rules are generated from @p dev->data->mac_addrs
1284  *   (@p priv->mac array).
1285  * - An additional flow rule for Ethernet broadcasts is also generated.
1286  * - All these are per-VLAN if @p DEV_RX_OFFLOAD_VLAN_FILTER
1287  *   is enabled and VLAN filters are configured.
1288  *
1289  * @param priv
1290  *   Pointer to private structure.
1291  * @param[out] error
1292  *   Perform verbose error reporting if not NULL.
1293  *
1294  * @return
1295  *   0 on success, a negative errno value otherwise and rte_errno is set.
1296  */
1297 static int
1298 mlx4_flow_internal(struct mlx4_priv *priv, struct rte_flow_error *error)
1299 {
1300         struct rte_flow_attr attr = {
1301                 .priority = MLX4_FLOW_PRIORITY_LAST,
1302                 .ingress = 1,
1303         };
1304         struct rte_flow_item_eth eth_spec;
1305         const struct rte_flow_item_eth eth_mask = {
1306                 .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
1307         };
1308         const struct rte_flow_item_eth eth_allmulti = {
1309                 .dst.addr_bytes = "\x01\x00\x00\x00\x00\x00",
1310         };
1311         struct rte_flow_item_vlan vlan_spec;
1312         const struct rte_flow_item_vlan vlan_mask = {
1313                 .tci = RTE_BE16(0x0fff),
1314         };
1315         struct rte_flow_item pattern[] = {
1316                 {
1317                         .type = MLX4_FLOW_ITEM_TYPE_INTERNAL,
1318                 },
1319                 {
1320                         .type = RTE_FLOW_ITEM_TYPE_ETH,
1321                         .spec = &eth_spec,
1322                         .mask = &eth_mask,
1323                 },
1324                 {
1325                         /* Replaced with VLAN if filtering is enabled. */
1326                         .type = RTE_FLOW_ITEM_TYPE_END,
1327                 },
1328                 {
1329                         .type = RTE_FLOW_ITEM_TYPE_END,
1330                 },
1331         };
1332         /*
1333          * Round number of queues down to their previous power of 2 to
1334          * comply with RSS context limitations. Extra queues silently do not
1335          * get RSS by default.
1336          */
1337         uint32_t queues =
1338                 rte_align32pow2(ETH_DEV(priv)->data->nb_rx_queues + 1) >> 1;
1339         uint16_t queue[queues];
1340         struct rte_flow_action_rss action_rss = {
1341                 .func = RTE_ETH_HASH_FUNCTION_DEFAULT,
1342                 .level = 0,
1343                 .types = 0,
1344                 .key_len = MLX4_RSS_HASH_KEY_SIZE,
1345                 .queue_num = queues,
1346                 .key = mlx4_rss_hash_key_default,
1347                 .queue = queue,
1348         };
1349         struct rte_flow_action actions[] = {
1350                 {
1351                         .type = RTE_FLOW_ACTION_TYPE_RSS,
1352                         .conf = &action_rss,
1353                 },
1354                 {
1355                         .type = RTE_FLOW_ACTION_TYPE_END,
1356                 },
1357         };
1358         struct ether_addr *rule_mac = &eth_spec.dst;
1359         rte_be16_t *rule_vlan =
1360                 (ETH_DEV(priv)->data->dev_conf.rxmode.offloads &
1361                  DEV_RX_OFFLOAD_VLAN_FILTER) &&
1362                 !ETH_DEV(priv)->data->promiscuous ?
1363                 &vlan_spec.tci :
1364                 NULL;
1365         uint16_t vlan = 0;
1366         struct rte_flow *flow;
1367         unsigned int i;
1368         int err = 0;
1369
1370         /* Nothing to be done if there are no Rx queues. */
1371         if (!queues)
1372                 goto error;
1373         /* Prepare default RSS configuration. */
1374         for (i = 0; i != queues; ++i)
1375                 queue[i] = i;
1376         /*
1377          * Set up VLAN item if filtering is enabled and at least one VLAN
1378          * filter is configured.
1379          */
1380         if (rule_vlan) {
1381                 vlan = mlx4_flow_internal_next_vlan(priv, 0);
1382                 if (vlan < 4096) {
1383                         pattern[2] = (struct rte_flow_item){
1384                                 .type = RTE_FLOW_ITEM_TYPE_VLAN,
1385                                 .spec = &vlan_spec,
1386                                 .mask = &vlan_mask,
1387                         };
1388 next_vlan:
1389                         *rule_vlan = rte_cpu_to_be_16(vlan);
1390                 } else {
1391                         rule_vlan = NULL;
1392                 }
1393         }
1394         for (i = 0; i != RTE_DIM(priv->mac) + 1; ++i) {
1395                 const struct ether_addr *mac;
1396
1397                 /* Broadcasts are handled by an extra iteration. */
1398                 if (i < RTE_DIM(priv->mac))
1399                         mac = &priv->mac[i];
1400                 else
1401                         mac = &eth_mask.dst;
1402                 if (is_zero_ether_addr(mac))
1403                         continue;
1404                 /* Check if MAC flow rule is already present. */
1405                 for (flow = LIST_FIRST(&priv->flows);
1406                      flow && flow->internal;
1407                      flow = LIST_NEXT(flow, next)) {
1408                         const struct ibv_flow_spec_eth *eth =
1409                                 (const void *)((uintptr_t)flow->ibv_attr +
1410                                                sizeof(*flow->ibv_attr));
1411                         unsigned int j;
1412
1413                         if (!flow->mac)
1414                                 continue;
1415                         assert(flow->ibv_attr->type == IBV_FLOW_ATTR_NORMAL);
1416                         assert(flow->ibv_attr->num_of_specs == 1);
1417                         assert(eth->type == IBV_FLOW_SPEC_ETH);
1418                         assert(flow->rss);
1419                         if (rule_vlan &&
1420                             (eth->val.vlan_tag != *rule_vlan ||
1421                              eth->mask.vlan_tag != RTE_BE16(0x0fff)))
1422                                 continue;
1423                         if (!rule_vlan && eth->mask.vlan_tag)
1424                                 continue;
1425                         for (j = 0; j != sizeof(mac->addr_bytes); ++j)
1426                                 if (eth->val.dst_mac[j] != mac->addr_bytes[j] ||
1427                                     eth->mask.dst_mac[j] != UINT8_C(0xff) ||
1428                                     eth->val.src_mac[j] != UINT8_C(0x00) ||
1429                                     eth->mask.src_mac[j] != UINT8_C(0x00))
1430                                         break;
1431                         if (j != sizeof(mac->addr_bytes))
1432                                 continue;
1433                         if (flow->rss->queues != queues ||
1434                             memcmp(flow->rss->queue_id, action_rss.queue,
1435                                    queues * sizeof(flow->rss->queue_id[0])))
1436                                 continue;
1437                         break;
1438                 }
1439                 if (!flow || !flow->internal) {
1440                         /* Not found, create a new flow rule. */
1441                         memcpy(rule_mac, mac, sizeof(*mac));
1442                         flow = mlx4_flow_create(ETH_DEV(priv), &attr, pattern,
1443                                                 actions, error);
1444                         if (!flow) {
1445                                 err = -rte_errno;
1446                                 goto error;
1447                         }
1448                 }
1449                 flow->select = 1;
1450                 flow->mac = 1;
1451         }
1452         if (rule_vlan) {
1453                 vlan = mlx4_flow_internal_next_vlan(priv, vlan + 1);
1454                 if (vlan < 4096)
1455                         goto next_vlan;
1456         }
1457         /* Take care of promiscuous and all multicast flow rules. */
1458         if (ETH_DEV(priv)->data->promiscuous ||
1459             ETH_DEV(priv)->data->all_multicast) {
1460                 for (flow = LIST_FIRST(&priv->flows);
1461                      flow && flow->internal;
1462                      flow = LIST_NEXT(flow, next)) {
1463                         if (ETH_DEV(priv)->data->promiscuous) {
1464                                 if (flow->promisc)
1465                                         break;
1466                         } else {
1467                                 assert(ETH_DEV(priv)->data->all_multicast);
1468                                 if (flow->allmulti)
1469                                         break;
1470                         }
1471                 }
1472                 if (flow && flow->internal) {
1473                         assert(flow->rss);
1474                         if (flow->rss->queues != queues ||
1475                             memcmp(flow->rss->queue_id, action_rss.queue,
1476                                    queues * sizeof(flow->rss->queue_id[0])))
1477                                 flow = NULL;
1478                 }
1479                 if (!flow || !flow->internal) {
1480                         /* Not found, create a new flow rule. */
1481                         if (ETH_DEV(priv)->data->promiscuous) {
1482                                 pattern[1].spec = NULL;
1483                                 pattern[1].mask = NULL;
1484                         } else {
1485                                 assert(ETH_DEV(priv)->data->all_multicast);
1486                                 pattern[1].spec = &eth_allmulti;
1487                                 pattern[1].mask = &eth_allmulti;
1488                         }
1489                         pattern[2] = pattern[3];
1490                         flow = mlx4_flow_create(ETH_DEV(priv), &attr, pattern,
1491                                                 actions, error);
1492                         if (!flow) {
1493                                 err = -rte_errno;
1494                                 goto error;
1495                         }
1496                 }
1497                 assert(flow->promisc || flow->allmulti);
1498                 flow->select = 1;
1499         }
1500 error:
1501         /* Clear selection and clean up stale internal flow rules. */
1502         flow = LIST_FIRST(&priv->flows);
1503         while (flow && flow->internal) {
1504                 struct rte_flow *next = LIST_NEXT(flow, next);
1505
1506                 if (!flow->select)
1507                         claim_zero(mlx4_flow_destroy(ETH_DEV(priv), flow,
1508                                                      error));
1509                 else
1510                         flow->select = 0;
1511                 flow = next;
1512         }
1513         return err;
1514 }
1515
1516 /**
1517  * Synchronize flow rules.
1518  *
1519  * This function synchronizes flow rules with the state of the device by
1520  * taking into account isolated mode and whether target queues are
1521  * configured.
1522  *
1523  * @param priv
1524  *   Pointer to private structure.
1525  * @param[out] error
1526  *   Perform verbose error reporting if not NULL.
1527  *
1528  * @return
1529  *   0 on success, a negative errno value otherwise and rte_errno is set.
1530  */
1531 int
1532 mlx4_flow_sync(struct mlx4_priv *priv, struct rte_flow_error *error)
1533 {
1534         struct rte_flow *flow;
1535         int ret;
1536
1537         /* Internal flow rules are guaranteed to come first in the list. */
1538         if (priv->isolated) {
1539                 /*
1540                  * Get rid of them in isolated mode, stop at the first
1541                  * non-internal rule found.
1542                  */
1543                 for (flow = LIST_FIRST(&priv->flows);
1544                      flow && flow->internal;
1545                      flow = LIST_FIRST(&priv->flows))
1546                         claim_zero(mlx4_flow_destroy(ETH_DEV(priv), flow,
1547                                                      error));
1548         } else {
1549                 /* Refresh internal rules. */
1550                 ret = mlx4_flow_internal(priv, error);
1551                 if (ret)
1552                         return ret;
1553         }
1554         /* Toggle the remaining flow rules . */
1555         LIST_FOREACH(flow, &priv->flows, next) {
1556                 ret = mlx4_flow_toggle(priv, flow, priv->started, error);
1557                 if (ret)
1558                         return ret;
1559         }
1560         if (!priv->started)
1561                 assert(!priv->drop);
1562         return 0;
1563 }
1564
1565 /**
1566  * Clean up all flow rules.
1567  *
1568  * Unlike mlx4_flow_flush(), this function takes care of all remaining flow
1569  * rules regardless of whether they are internal or user-configured.
1570  *
1571  * @param priv
1572  *   Pointer to private structure.
1573  */
1574 void
1575 mlx4_flow_clean(struct mlx4_priv *priv)
1576 {
1577         struct rte_flow *flow;
1578
1579         while ((flow = LIST_FIRST(&priv->flows)))
1580                 mlx4_flow_destroy(ETH_DEV(priv), flow, NULL);
1581         assert(LIST_EMPTY(&priv->rss));
1582 }
1583
1584 static const struct rte_flow_ops mlx4_flow_ops = {
1585         .validate = mlx4_flow_validate,
1586         .create = mlx4_flow_create,
1587         .destroy = mlx4_flow_destroy,
1588         .flush = mlx4_flow_flush,
1589         .isolate = mlx4_flow_isolate,
1590 };
1591
1592 /**
1593  * Manage filter operations.
1594  *
1595  * @param dev
1596  *   Pointer to Ethernet device structure.
1597  * @param filter_type
1598  *   Filter type.
1599  * @param filter_op
1600  *   Operation to perform.
1601  * @param arg
1602  *   Pointer to operation-specific structure.
1603  *
1604  * @return
1605  *   0 on success, negative errno value otherwise and rte_errno is set.
1606  */
1607 int
1608 mlx4_filter_ctrl(struct rte_eth_dev *dev,
1609                  enum rte_filter_type filter_type,
1610                  enum rte_filter_op filter_op,
1611                  void *arg)
1612 {
1613         switch (filter_type) {
1614         case RTE_ETH_FILTER_GENERIC:
1615                 if (filter_op != RTE_ETH_FILTER_GET)
1616                         break;
1617                 *(const void **)arg = &mlx4_flow_ops;
1618                 return 0;
1619         default:
1620                 ERROR("%p: filter type (%d) not supported",
1621                       (void *)dev, filter_type);
1622                 break;
1623         }
1624         rte_errno = ENOTSUP;
1625         return -rte_errno;
1626 }