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 /** IBV supported RSS hash functions combinations */
  46 #define MLX4_IBV_IPV4_HF ( \
  47         IBV_RX_HASH_SRC_IPV4 | \
  48         IBV_RX_HASH_DST_IPV4)
  49 #define MLX4_IBV_IPV6_HF ( \
  50         IBV_RX_HASH_SRC_IPV6 | \
  51         IBV_RX_HASH_DST_IPV6)
  52 #define MLX4_IBV_TCP_HF ( \
  53         IBV_RX_HASH_SRC_PORT_TCP | \
  54         IBV_RX_HASH_DST_PORT_TCP)
  55 #define MLX4_IBV_UDP_HF ( \
  56         IBV_RX_HASH_SRC_PORT_UDP | \
  57         IBV_RX_HASH_DST_PORT_UDP)
  58
  59 /** Supported RSS hash functions combinations */
  60 #define MLX4_RSS_IPV4_HF ( \
  61         ETH_RSS_IPV4 | \
  62         ETH_RSS_FRAG_IPV4 | \
  63         ETH_RSS_NONFRAG_IPV4_OTHER)
  64 #define MLX4_RSS_IPV6_HF ( \
  65         ETH_RSS_IPV6 | \
  66         ETH_RSS_FRAG_IPV6 | \
  67         ETH_RSS_NONFRAG_IPV6_OTHER | \
  68         ETH_RSS_IPV6_EX)
  69 #define MLX4_RSS_IPV4_TCP_HF ( \
  70         ETH_RSS_NONFRAG_IPV4_TCP)
  71 #define MLX4_RSS_IPV6_TCP_HF ( \
  72         ETH_RSS_NONFRAG_IPV6_TCP | \
  73         ETH_RSS_IPV6_TCP_EX)
  74 #define MLX4_RSS_IPV4_UDP_HF ( \
  75         ETH_RSS_NONFRAG_IPV4_UDP)
  76 #define MLX4_RSS_IPV6_UDP_HF ( \
  77         ETH_RSS_NONFRAG_IPV6_UDP | \
  78         ETH_RSS_IPV6_UDP_EX)
  79
  80 /** Static initializer for a list of subsequent item types. */
  81 #define NEXT_ITEM(...) \
  82         (const enum rte_flow_item_type []){ \
  83                 __VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \
  84         }
  85
  86 /** Processor structure associated with a flow item. */
  87 struct mlx4_flow_proc_item {
  88         /** Bit-mask for fields supported by this PMD. */
  89         const void *mask_support;
  90         /** Bit-mask to use when @p item->mask is not provided. */
  91         const void *mask_default;
  92         /** Size in bytes for @p mask_support and @p mask_default. */
  93         const unsigned int mask_sz;
  94         /** Merge a pattern item into a flow rule handle. */
  95         int (*merge)(struct rte_flow *flow,
  96                      const struct rte_flow_item *item,
  97                      const struct mlx4_flow_proc_item *proc,
  98                      struct rte_flow_error *error);
  99         /** Size in bytes of the destination structure. */
 100         const unsigned int dst_sz;
 101         /** List of possible subsequent items. */
 102         const enum rte_flow_item_type *const next_item;
 103 };
 104
 105 /** Shared resources for drop flow rules. */
 106 struct mlx4_drop {
 107         struct ibv_qp *qp; /**< QP target. */
 108         struct ibv_cq *cq; /**< CQ associated with above QP. */
 109         struct priv *priv; /**< Back pointer to private data. */
 110         uint32_t refcnt; /**< Reference count. */
 111 };
 112
 113 /**
 114  * Convert DPDK RSS hash types to their Verbs equivalent.
 115  *
 116  * This function returns the supported (default) set when @p types has
 117  * special value 0.
 118  *
 119  * @param priv
 120  *   Pointer to private structure.
 121  * @param types
 122  *   Hash types in DPDK format (see struct rte_eth_rss_conf).
 123  *
 124  * @return
 125  *   A valid Verbs RSS hash fields mask for mlx4 on success, (uint64_t)-1
 126  *   otherwise and rte_errno is set.
 127  */
 128 uint64_t
 129 mlx4_conv_rss_types(struct priv *priv, uint64_t types)
 130 {
 131         enum { IPV4, IPV6, TCP, UDP, };
 132         static const uint64_t in[] = {
 133                 [IPV4] = (ETH_RSS_IPV4 |
 134                           ETH_RSS_FRAG_IPV4 |
 135                           ETH_RSS_NONFRAG_IPV4_TCP |
 136                           ETH_RSS_NONFRAG_IPV4_UDP |
 137                           ETH_RSS_NONFRAG_IPV4_OTHER),
 138                 [IPV6] = (ETH_RSS_IPV6 |
 139                           ETH_RSS_FRAG_IPV6 |
 140                           ETH_RSS_NONFRAG_IPV6_TCP |
 141                           ETH_RSS_NONFRAG_IPV6_UDP |
 142                           ETH_RSS_NONFRAG_IPV6_OTHER |
 143                           ETH_RSS_IPV6_EX |
 144                           ETH_RSS_IPV6_TCP_EX |
 145                           ETH_RSS_IPV6_UDP_EX),
 146                 [TCP] = (ETH_RSS_NONFRAG_IPV4_TCP |
 147                          ETH_RSS_NONFRAG_IPV6_TCP |
 148                          ETH_RSS_IPV6_TCP_EX),
 149                 [UDP] = (ETH_RSS_NONFRAG_IPV4_UDP |
 150                          ETH_RSS_NONFRAG_IPV6_UDP |
 151                          ETH_RSS_IPV6_UDP_EX),
 152         };
 153         static const uint64_t out[RTE_DIM(in)] = {
 154                 [IPV4] = IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4,
 155                 [IPV6] = IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6,
 156                 [TCP] = IBV_RX_HASH_SRC_PORT_TCP | IBV_RX_HASH_DST_PORT_TCP,
 157                 [UDP] = IBV_RX_HASH_SRC_PORT_UDP | IBV_RX_HASH_DST_PORT_UDP,
 158         };
 159         uint64_t seen = 0;
 160         uint64_t conv = 0;
 161         unsigned int i;
 162
 163         if (!types)
 164                 return priv->hw_rss_sup;
 165         for (i = 0; i != RTE_DIM(in); ++i)
 166                 if (types & in[i]) {
 167                         seen |= types & in[i];
 168                         conv |= out[i];
 169                 }
 170         if ((conv & priv->hw_rss_sup) == conv && !(types & ~seen))
 171                 return conv;
 172         rte_errno = ENOTSUP;
 173         return (uint64_t)-1;
 174 }
 175
 176 /**
 177  * Convert verbs RSS types to their DPDK equivalents.
 178  *
 179  * This function returns a group of RSS DPDK types given their equivalent group
 180  * of verbs types.
 181  * For example both source IPv4 and destination IPv4 verbs types are converted
 182  * into their equivalent RSS group types. If each of these verbs types existed
 183  * exclusively - no conversion would take place.
 184  *
 185  * @param types
 186  *   RSS hash types in verbs format.
 187  *
 188  * @return
 189  *   DPDK RSS hash fields supported by mlx4.
 190  */
 191 uint64_t
 192 mlx4_ibv_to_rss_types(uint64_t types)
 193 {
 194         enum { IPV4, IPV6, IPV4_TCP, IPV6_TCP, IPV4_UDP, IPV6_UDP};
 195
 196         static const uint64_t in[] = {
 197                 [IPV4] = MLX4_IBV_IPV4_HF,
 198                 [IPV6] = MLX4_IBV_IPV6_HF,
 199                 [IPV4_TCP] = MLX4_IBV_IPV4_HF | MLX4_IBV_TCP_HF,
 200                 [IPV6_TCP] = MLX4_IBV_IPV6_HF | MLX4_IBV_TCP_HF,
 201                 [IPV4_UDP] = MLX4_IBV_IPV4_HF | MLX4_IBV_UDP_HF,
 202                 [IPV6_UDP] = MLX4_IBV_IPV6_HF | MLX4_IBV_UDP_HF,
 203         };
 204         static const uint64_t out[RTE_DIM(in)] = {
 205                 [IPV4] = MLX4_RSS_IPV4_HF,
 206                 [IPV6] = MLX4_RSS_IPV6_HF,
 207                 [IPV4_TCP] = MLX4_RSS_IPV4_HF | MLX4_RSS_IPV4_TCP_HF,
 208                 [IPV6_TCP] = MLX4_RSS_IPV6_HF | MLX4_RSS_IPV6_TCP_HF,
 209                 [IPV4_UDP] = MLX4_RSS_IPV4_HF | MLX4_RSS_IPV4_UDP_HF,
 210                 [IPV6_UDP] = MLX4_RSS_IPV6_HF | MLX4_RSS_IPV6_UDP_HF,
 211         };
 212         uint64_t conv = 0;
 213         unsigned int i;
 214
 215         for (i = 0; i != RTE_DIM(in); ++i)
 216                 if ((types & in[i]) == in[i])
 217                         conv |= out[i];
 218         return conv;
 219 }
 220
 221 /**
 222  * Merge Ethernet pattern item into flow rule handle.
 223  *
 224  * Additional mlx4-specific constraints on supported fields:
 225  *
 226  * - No support for partial masks, except in the specific case of matching
 227  *   all multicast traffic (@p spec->dst and @p mask->dst equal to
 228  *   01:00:00:00:00:00).
 229  * - Not providing @p item->spec or providing an empty @p mask->dst is
 230  *   *only* supported if the rule doesn't specify additional matching
 231  *   criteria (i.e. rule is promiscuous-like).
 232  *
 233  * @param[in, out] flow
 234  *   Flow rule handle to update.
 235  * @param[in] item
 236  *   Pattern item to merge.
 237  * @param[in] proc
 238  *   Associated item-processing object.
 239  * @param[out] error
 240  *   Perform verbose error reporting if not NULL.
 241  *
 242  * @return
 243  *   0 on success, a negative errno value otherwise and rte_errno is set.
 244  */
 245 static int
 246 mlx4_flow_merge_eth(struct rte_flow *flow,
 247                     const struct rte_flow_item *item,
 248                     const struct mlx4_flow_proc_item *proc,
 249                     struct rte_flow_error *error)
 250 {
 251         const struct rte_flow_item_eth *spec = item->spec;
 252         const struct rte_flow_item_eth *mask =
 253                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 254         struct ibv_flow_spec_eth *eth;
 255         const char *msg;
 256         unsigned int i;
 257
 258         if (!mask) {
 259                 flow->promisc = 1;
 260         } else {
 261                 uint32_t sum_dst = 0;
 262                 uint32_t sum_src = 0;
 263
 264                 for (i = 0; i != sizeof(mask->dst.addr_bytes); ++i) {
 265                         sum_dst += mask->dst.addr_bytes[i];
 266                         sum_src += mask->src.addr_bytes[i];
 267                 }
 268                 if (sum_src) {
 269                         msg = "mlx4 does not support source MAC matching";
 270                         goto error;
 271                 } else if (!sum_dst) {
 272                         flow->promisc = 1;
 273                 } else if (sum_dst == 1 && mask->dst.addr_bytes[0] == 1) {
 274                         if (!(spec->dst.addr_bytes[0] & 1)) {
 275                                 msg = "mlx4 does not support the explicit"
 276                                         " exclusion of all multicast traffic";
 277                                 goto error;
 278                         }
 279                         flow->allmulti = 1;
 280                 } else if (sum_dst != (UINT8_C(0xff) * ETHER_ADDR_LEN)) {
 281                         msg = "mlx4 does not support matching partial"
 282                                 " Ethernet fields";
 283                         goto error;
 284                 }
 285         }
 286         if (!flow->ibv_attr)
 287                 return 0;
 288         if (flow->promisc) {
 289                 flow->ibv_attr->type = IBV_FLOW_ATTR_ALL_DEFAULT;
 290                 return 0;
 291         }
 292         if (flow->allmulti) {
 293                 flow->ibv_attr->type = IBV_FLOW_ATTR_MC_DEFAULT;
 294                 return 0;
 295         }
 296         ++flow->ibv_attr->num_of_specs;
 297         eth = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 298         *eth = (struct ibv_flow_spec_eth) {
 299                 .type = IBV_FLOW_SPEC_ETH,
 300                 .size = sizeof(*eth),
 301         };
 302         memcpy(eth->val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
 303         memcpy(eth->mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
 304         /* Remove unwanted bits from values. */
 305         for (i = 0; i < ETHER_ADDR_LEN; ++i) {
 306                 eth->val.dst_mac[i] &= eth->mask.dst_mac[i];
 307         }
 308         return 0;
 309 error:
 310         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 311                                   item, msg);
 312 }
 313
 314 /**
 315  * Merge VLAN pattern item into flow rule handle.
 316  *
 317  * Additional mlx4-specific constraints on supported fields:
 318  *
 319  * - Matching *all* VLAN traffic by omitting @p item->spec or providing an
 320  *   empty @p item->mask would also include non-VLAN traffic. Doing so is
 321  *   therefore unsupported.
 322  * - No support for partial masks.
 323  *
 324  * @param[in, out] flow
 325  *   Flow rule handle to update.
 326  * @param[in] item
 327  *   Pattern item to merge.
 328  * @param[in] proc
 329  *   Associated item-processing object.
 330  * @param[out] error
 331  *   Perform verbose error reporting if not NULL.
 332  *
 333  * @return
 334  *   0 on success, a negative errno value otherwise and rte_errno is set.
 335  */
 336 static int
 337 mlx4_flow_merge_vlan(struct rte_flow *flow,
 338                      const struct rte_flow_item *item,
 339                      const struct mlx4_flow_proc_item *proc,
 340                      struct rte_flow_error *error)
 341 {
 342         const struct rte_flow_item_vlan *spec = item->spec;
 343         const struct rte_flow_item_vlan *mask =
 344                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 345         struct ibv_flow_spec_eth *eth;
 346         const char *msg;
 347
 348         if (!mask || !mask->tci) {
 349                 msg = "mlx4 cannot match all VLAN traffic while excluding"
 350                         " non-VLAN traffic, TCI VID must be specified";
 351                 goto error;
 352         }
 353         if (mask->tci != RTE_BE16(0x0fff)) {
 354                 msg = "mlx4 does not support partial TCI VID matching";
 355                 goto error;
 356         }
 357         if (!flow->ibv_attr)
 358                 return 0;
 359         eth = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size -
 360                        sizeof(*eth));
 361         eth->val.vlan_tag = spec->tci;
 362         eth->mask.vlan_tag = mask->tci;
 363         eth->val.vlan_tag &= eth->mask.vlan_tag;
 364         return 0;
 365 error:
 366         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 367                                   item, msg);
 368 }
 369
 370 /**
 371  * Merge IPv4 pattern item into flow rule handle.
 372  *
 373  * Additional mlx4-specific constraints on supported fields:
 374  *
 375  * - No support for partial masks.
 376  *
 377  * @param[in, out] flow
 378  *   Flow rule handle to update.
 379  * @param[in] item
 380  *   Pattern item to merge.
 381  * @param[in] proc
 382  *   Associated item-processing object.
 383  * @param[out] error
 384  *   Perform verbose error reporting if not NULL.
 385  *
 386  * @return
 387  *   0 on success, a negative errno value otherwise and rte_errno is set.
 388  */
 389 static int
 390 mlx4_flow_merge_ipv4(struct rte_flow *flow,
 391                      const struct rte_flow_item *item,
 392                      const struct mlx4_flow_proc_item *proc,
 393                      struct rte_flow_error *error)
 394 {
 395         const struct rte_flow_item_ipv4 *spec = item->spec;
 396         const struct rte_flow_item_ipv4 *mask =
 397                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 398         struct ibv_flow_spec_ipv4 *ipv4;
 399         const char *msg;
 400
 401         if (mask &&
 402             ((uint32_t)(mask->hdr.src_addr + 1) > UINT32_C(1) ||
 403              (uint32_t)(mask->hdr.dst_addr + 1) > UINT32_C(1))) {
 404                 msg = "mlx4 does not support matching partial IPv4 fields";
 405                 goto error;
 406         }
 407         if (!flow->ibv_attr)
 408                 return 0;
 409         ++flow->ibv_attr->num_of_specs;
 410         ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 411         *ipv4 = (struct ibv_flow_spec_ipv4) {
 412                 .type = IBV_FLOW_SPEC_IPV4,
 413                 .size = sizeof(*ipv4),
 414         };
 415         if (!spec)
 416                 return 0;
 417         ipv4->val = (struct ibv_flow_ipv4_filter) {
 418                 .src_ip = spec->hdr.src_addr,
 419                 .dst_ip = spec->hdr.dst_addr,
 420         };
 421         ipv4->mask = (struct ibv_flow_ipv4_filter) {
 422                 .src_ip = mask->hdr.src_addr,
 423                 .dst_ip = mask->hdr.dst_addr,
 424         };
 425         /* Remove unwanted bits from values. */
 426         ipv4->val.src_ip &= ipv4->mask.src_ip;
 427         ipv4->val.dst_ip &= ipv4->mask.dst_ip;
 428         return 0;
 429 error:
 430         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 431                                   item, msg);
 432 }
 433
 434 /**
 435  * Merge UDP pattern item into flow rule handle.
 436  *
 437  * Additional mlx4-specific constraints on supported fields:
 438  *
 439  * - No support for partial masks.
 440  * - Due to HW/FW limitation, flow rule priority is not taken into account
 441  *   when matching UDP destination ports, doing is therefore only supported
 442  *   at the highest priority level (0).
 443  *
 444  * @param[in, out] flow
 445  *   Flow rule handle to update.
 446  * @param[in] item
 447  *   Pattern item to merge.
 448  * @param[in] proc
 449  *   Associated item-processing object.
 450  * @param[out] error
 451  *   Perform verbose error reporting if not NULL.
 452  *
 453  * @return
 454  *   0 on success, a negative errno value otherwise and rte_errno is set.
 455  */
 456 static int
 457 mlx4_flow_merge_udp(struct rte_flow *flow,
 458                     const struct rte_flow_item *item,
 459                     const struct mlx4_flow_proc_item *proc,
 460                     struct rte_flow_error *error)
 461 {
 462         const struct rte_flow_item_udp *spec = item->spec;
 463         const struct rte_flow_item_udp *mask =
 464                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 465         struct ibv_flow_spec_tcp_udp *udp;
 466         const char *msg;
 467
 468         if (mask &&
 469             ((uint16_t)(mask->hdr.src_port + 1) > UINT16_C(1) ||
 470              (uint16_t)(mask->hdr.dst_port + 1) > UINT16_C(1))) {
 471                 msg = "mlx4 does not support matching partial UDP fields";
 472                 goto error;
 473         }
 474         if (mask && mask->hdr.dst_port && flow->priority) {
 475                 msg = "combining UDP destination port matching with a nonzero"
 476                         " priority level is not supported";
 477                 goto error;
 478         }
 479         if (!flow->ibv_attr)
 480                 return 0;
 481         ++flow->ibv_attr->num_of_specs;
 482         udp = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 483         *udp = (struct ibv_flow_spec_tcp_udp) {
 484                 .type = IBV_FLOW_SPEC_UDP,
 485                 .size = sizeof(*udp),
 486         };
 487         if (!spec)
 488                 return 0;
 489         udp->val.dst_port = spec->hdr.dst_port;
 490         udp->val.src_port = spec->hdr.src_port;
 491         udp->mask.dst_port = mask->hdr.dst_port;
 492         udp->mask.src_port = mask->hdr.src_port;
 493         /* Remove unwanted bits from values. */
 494         udp->val.src_port &= udp->mask.src_port;
 495         udp->val.dst_port &= udp->mask.dst_port;
 496         return 0;
 497 error:
 498         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 499                                   item, msg);
 500 }
 501
 502 /**
 503  * Merge TCP pattern item into flow rule handle.
 504  *
 505  * Additional mlx4-specific constraints on supported fields:
 506  *
 507  * - No support for partial masks.
 508  *
 509  * @param[in, out] flow
 510  *   Flow rule handle to update.
 511  * @param[in] item
 512  *   Pattern item to merge.
 513  * @param[in] proc
 514  *   Associated item-processing object.
 515  * @param[out] error
 516  *   Perform verbose error reporting if not NULL.
 517  *
 518  * @return
 519  *   0 on success, a negative errno value otherwise and rte_errno is set.
 520  */
 521 static int
 522 mlx4_flow_merge_tcp(struct rte_flow *flow,
 523                     const struct rte_flow_item *item,
 524                     const struct mlx4_flow_proc_item *proc,
 525                     struct rte_flow_error *error)
 526 {
 527         const struct rte_flow_item_tcp *spec = item->spec;
 528         const struct rte_flow_item_tcp *mask =
 529                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 530         struct ibv_flow_spec_tcp_udp *tcp;
 531         const char *msg;
 532
 533         if (mask &&
 534             ((uint16_t)(mask->hdr.src_port + 1) > UINT16_C(1) ||
 535              (uint16_t)(mask->hdr.dst_port + 1) > UINT16_C(1))) {
 536                 msg = "mlx4 does not support matching partial TCP fields";
 537                 goto error;
 538         }
 539         if (!flow->ibv_attr)
 540                 return 0;
 541         ++flow->ibv_attr->num_of_specs;
 542         tcp = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 543         *tcp = (struct ibv_flow_spec_tcp_udp) {
 544                 .type = IBV_FLOW_SPEC_TCP,
 545                 .size = sizeof(*tcp),
 546         };
 547         if (!spec)
 548                 return 0;
 549         tcp->val.dst_port = spec->hdr.dst_port;
 550         tcp->val.src_port = spec->hdr.src_port;
 551         tcp->mask.dst_port = mask->hdr.dst_port;
 552         tcp->mask.src_port = mask->hdr.src_port;
 553         /* Remove unwanted bits from values. */
 554         tcp->val.src_port &= tcp->mask.src_port;
 555         tcp->val.dst_port &= tcp->mask.dst_port;
 556         return 0;
 557 error:
 558         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 559                                   item, msg);
 560 }
 561
 562 /**
 563  * Perform basic sanity checks on a pattern item.
 564  *
 565  * @param[in] item
 566  *   Item specification.
 567  * @param[in] proc
 568  *   Associated item-processing object.
 569  * @param[out] error
 570  *   Perform verbose error reporting if not NULL.
 571  *
 572  * @return
 573  *   0 on success, a negative errno value otherwise and rte_errno is set.
 574  */
 575 static int
 576 mlx4_flow_item_check(const struct rte_flow_item *item,
 577                      const struct mlx4_flow_proc_item *proc,
 578                      struct rte_flow_error *error)
 579 {
 580         const uint8_t *mask;
 581         unsigned int i;
 582
 583         /* item->last and item->mask cannot exist without item->spec. */
 584         if (!item->spec && (item->mask || item->last))
 585                 return rte_flow_error_set
 586                         (error, EINVAL, RTE_FLOW_ERROR_TYPE_ITEM, item,
 587                          "\"mask\" or \"last\" field provided without a"
 588                          " corresponding \"spec\"");
 589         /* No spec, no mask, no problem. */
 590         if (!item->spec)
 591                 return 0;
 592         mask = item->mask ?
 593                 (const uint8_t *)item->mask :
 594                 (const uint8_t *)proc->mask_default;
 595         assert(mask);
 596         /*
 597          * Single-pass check to make sure that:
 598          * - Mask is supported, no bits are set outside proc->mask_support.
 599          * - Both item->spec and item->last are included in mask.
 600          */
 601         for (i = 0; i != proc->mask_sz; ++i) {
 602                 if (!mask[i])
 603                         continue;
 604                 if ((mask[i] | ((const uint8_t *)proc->mask_support)[i]) !=
 605                     ((const uint8_t *)proc->mask_support)[i])
 606                         return rte_flow_error_set
 607                                 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 608                                  item, "unsupported field found in \"mask\"");
 609                 if (item->last &&
 610                     (((const uint8_t *)item->spec)[i] & mask[i]) !=
 611                     (((const uint8_t *)item->last)[i] & mask[i]))
 612                         return rte_flow_error_set
 613                                 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 614                                  item,
 615                                  "range between \"spec\" and \"last\""
 616                                  " is larger than \"mask\"");
 617         }
 618         return 0;
 619 }
 620
 621 /** Graph of supported items and associated actions. */
 622 static const struct mlx4_flow_proc_item mlx4_flow_proc_item_list[] = {
 623         [RTE_FLOW_ITEM_TYPE_END] = {
 624                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_ETH),
 625         },
 626         [RTE_FLOW_ITEM_TYPE_ETH] = {
 627                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_VLAN,
 628                                        RTE_FLOW_ITEM_TYPE_IPV4),
 629                 .mask_support = &(const struct rte_flow_item_eth){
 630                         /* Only destination MAC can be matched. */
 631                         .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 632                 },
 633                 .mask_default = &rte_flow_item_eth_mask,
 634                 .mask_sz = sizeof(struct rte_flow_item_eth),
 635                 .merge = mlx4_flow_merge_eth,
 636                 .dst_sz = sizeof(struct ibv_flow_spec_eth),
 637         },
 638         [RTE_FLOW_ITEM_TYPE_VLAN] = {
 639                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_IPV4),
 640                 .mask_support = &(const struct rte_flow_item_vlan){
 641                         /* Only TCI VID matching is supported. */
 642                         .tci = RTE_BE16(0x0fff),
 643                 },
 644                 .mask_default = &rte_flow_item_vlan_mask,
 645                 .mask_sz = sizeof(struct rte_flow_item_vlan),
 646                 .merge = mlx4_flow_merge_vlan,
 647                 .dst_sz = 0,
 648         },
 649         [RTE_FLOW_ITEM_TYPE_IPV4] = {
 650                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_UDP,
 651                                        RTE_FLOW_ITEM_TYPE_TCP),
 652                 .mask_support = &(const struct rte_flow_item_ipv4){
 653                         .hdr = {
 654                                 .src_addr = RTE_BE32(0xffffffff),
 655                                 .dst_addr = RTE_BE32(0xffffffff),
 656                         },
 657                 },
 658                 .mask_default = &rte_flow_item_ipv4_mask,
 659                 .mask_sz = sizeof(struct rte_flow_item_ipv4),
 660                 .merge = mlx4_flow_merge_ipv4,
 661                 .dst_sz = sizeof(struct ibv_flow_spec_ipv4),
 662         },
 663         [RTE_FLOW_ITEM_TYPE_UDP] = {
 664                 .mask_support = &(const struct rte_flow_item_udp){
 665                         .hdr = {
 666                                 .src_port = RTE_BE16(0xffff),
 667                                 .dst_port = RTE_BE16(0xffff),
 668                         },
 669                 },
 670                 .mask_default = &rte_flow_item_udp_mask,
 671                 .mask_sz = sizeof(struct rte_flow_item_udp),
 672                 .merge = mlx4_flow_merge_udp,
 673                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 674         },
 675         [RTE_FLOW_ITEM_TYPE_TCP] = {
 676                 .mask_support = &(const struct rte_flow_item_tcp){
 677                         .hdr = {
 678                                 .src_port = RTE_BE16(0xffff),
 679                                 .dst_port = RTE_BE16(0xffff),
 680                         },
 681                 },
 682                 .mask_default = &rte_flow_item_tcp_mask,
 683                 .mask_sz = sizeof(struct rte_flow_item_tcp),
 684                 .merge = mlx4_flow_merge_tcp,
 685                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 686         },
 687 };
 688
 689 /**
 690  * Make sure a flow rule is supported and initialize associated structure.
 691  *
 692  * @param priv
 693  *   Pointer to private structure.
 694  * @param[in] attr
 695  *   Flow rule attributes.
 696  * @param[in] pattern
 697  *   Pattern specification (list terminated by the END pattern item).
 698  * @param[in] actions
 699  *   Associated actions (list terminated by the END action).
 700  * @param[out] error
 701  *   Perform verbose error reporting if not NULL.
 702  * @param[in, out] addr
 703  *   Buffer where the resulting flow rule handle pointer must be stored.
 704  *   If NULL, stop processing after validation stage.
 705  *
 706  * @return
 707  *   0 on success, a negative errno value otherwise and rte_errno is set.
 708  */
 709 static int
 710 mlx4_flow_prepare(struct priv *priv,
 711                   const struct rte_flow_attr *attr,
 712                   const struct rte_flow_item pattern[],
 713                   const struct rte_flow_action actions[],
 714                   struct rte_flow_error *error,
 715                   struct rte_flow **addr)
 716 {
 717         const struct rte_flow_item *item;
 718         const struct rte_flow_action *action;
 719         const struct mlx4_flow_proc_item *proc;
 720         struct rte_flow temp = { .ibv_attr_size = sizeof(*temp.ibv_attr) };
 721         struct rte_flow *flow = &temp;
 722         const char *msg = NULL;
 723         int overlap;
 724
 725         if (attr->group)
 726                 return rte_flow_error_set
 727                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 728                          NULL, "groups are not supported");
 729         if (attr->priority > MLX4_FLOW_PRIORITY_LAST)
 730                 return rte_flow_error_set
 731                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 732                          NULL, "maximum priority level is "
 733                          MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST));
 734         if (attr->egress)
 735                 return rte_flow_error_set
 736                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 737                          NULL, "egress is not supported");
 738         if (attr->transfer)
 739                 return rte_flow_error_set
 740                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_TRANSFER,
 741                          NULL, "transfer is not supported");
 742         if (!attr->ingress)
 743                 return rte_flow_error_set
 744                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 745                          NULL, "only ingress is supported");
 746 fill:
 747         overlap = 0;
 748         proc = mlx4_flow_proc_item_list;
 749         flow->priority = attr->priority;
 750         /* Go over pattern. */
 751         for (item = pattern; item->type; ++item) {
 752                 const struct mlx4_flow_proc_item *next = NULL;
 753                 unsigned int i;
 754                 int err;
 755
 756                 if (item->type == RTE_FLOW_ITEM_TYPE_VOID)
 757                         continue;
 758                 if (item->type == MLX4_FLOW_ITEM_TYPE_INTERNAL) {
 759                         flow->internal = 1;
 760                         continue;
 761                 }
 762                 if (flow->promisc || flow->allmulti) {
 763                         msg = "mlx4 does not support additional matching"
 764                                 " criteria combined with indiscriminate"
 765                                 " matching on Ethernet headers";
 766                         goto exit_item_not_supported;
 767                 }
 768                 for (i = 0; proc->next_item && proc->next_item[i]; ++i) {
 769                         if (proc->next_item[i] == item->type) {
 770                                 next = &mlx4_flow_proc_item_list[item->type];
 771                                 break;
 772                         }
 773                 }
 774                 if (!next)
 775                         goto exit_item_not_supported;
 776                 proc = next;
 777                 /*
 778                  * Perform basic sanity checks only once, while handle is
 779                  * not allocated.
 780                  */
 781                 if (flow == &temp) {
 782                         err = mlx4_flow_item_check(item, proc, error);
 783                         if (err)
 784                                 return err;
 785                 }
 786                 if (proc->merge) {
 787                         err = proc->merge(flow, item, proc, error);
 788                         if (err)
 789                                 return err;
 790                 }
 791                 flow->ibv_attr_size += proc->dst_sz;
 792         }
 793         /* Go over actions list. */
 794         for (action = actions; action->type; ++action) {
 795                 /* This one may appear anywhere multiple times. */
 796                 if (action->type == RTE_FLOW_ACTION_TYPE_VOID)
 797                         continue;
 798                 /* Fate-deciding actions may appear exactly once. */
 799                 if (overlap) {
 800                         msg = "cannot combine several fate-deciding actions,"
 801                                 " choose between DROP, QUEUE or RSS";
 802                         goto exit_action_not_supported;
 803                 }
 804                 overlap = 1;
 805                 switch (action->type) {
 806                         const struct rte_flow_action_queue *queue;
 807                         const struct rte_flow_action_rss *rss;
 808                         const uint8_t *rss_key;
 809                         uint32_t rss_key_len;
 810                         uint64_t fields;
 811                         unsigned int i;
 812
 813                 case RTE_FLOW_ACTION_TYPE_DROP:
 814                         flow->drop = 1;
 815                         break;
 816                 case RTE_FLOW_ACTION_TYPE_QUEUE:
 817                         if (flow->rss)
 818                                 break;
 819                         queue = action->conf;
 820                         if (queue->index >= priv->dev->data->nb_rx_queues) {
 821                                 msg = "queue target index beyond number of"
 822                                         " configured Rx queues";
 823                                 goto exit_action_not_supported;
 824                         }
 825                         flow->rss = mlx4_rss_get
 826                                 (priv, 0, mlx4_rss_hash_key_default, 1,
 827                                  &queue->index);
 828                         if (!flow->rss) {
 829                                 msg = "not enough resources for additional"
 830                                         " single-queue RSS context";
 831                                 goto exit_action_not_supported;
 832                         }
 833                         break;
 834                 case RTE_FLOW_ACTION_TYPE_RSS:
 835                         if (flow->rss)
 836                                 break;
 837                         rss = action->conf;
 838                         /* Default RSS configuration if none is provided. */
 839                         if (rss->key_len) {
 840                                 rss_key = rss->key;
 841                                 rss_key_len = rss->key_len;
 842                         } else {
 843                                 rss_key = mlx4_rss_hash_key_default;
 844                                 rss_key_len = MLX4_RSS_HASH_KEY_SIZE;
 845                         }
 846                         /* Sanity checks. */
 847                         for (i = 0; i < rss->queue_num; ++i)
 848                                 if (rss->queue[i] >=
 849                                     priv->dev->data->nb_rx_queues)
 850                                         break;
 851                         if (i != rss->queue_num) {
 852                                 msg = "queue index target beyond number of"
 853                                         " configured Rx queues";
 854                                 goto exit_action_not_supported;
 855                         }
 856                         if (!rte_is_power_of_2(rss->queue_num)) {
 857                                 msg = "for RSS, mlx4 requires the number of"
 858                                         " queues to be a power of two";
 859                                 goto exit_action_not_supported;
 860                         }
 861                         if (rss_key_len != sizeof(flow->rss->key)) {
 862                                 msg = "mlx4 supports exactly one RSS hash key"
 863                                         " length: "
 864                                         MLX4_STR_EXPAND(MLX4_RSS_HASH_KEY_SIZE);
 865                                 goto exit_action_not_supported;
 866                         }
 867                         for (i = 1; i < rss->queue_num; ++i)
 868                                 if (rss->queue[i] - rss->queue[i - 1] != 1)
 869                                         break;
 870                         if (i != rss->queue_num) {
 871                                 msg = "mlx4 requires RSS contexts to use"
 872                                         " consecutive queue indices only";
 873                                 goto exit_action_not_supported;
 874                         }
 875                         if (rss->queue[0] % rss->queue_num) {
 876                                 msg = "mlx4 requires the first queue of a RSS"
 877                                         " context to be aligned on a multiple"
 878                                         " of the context size";
 879                                 goto exit_action_not_supported;
 880                         }
 881                         if (rss->func &&
 882                             rss->func != RTE_ETH_HASH_FUNCTION_TOEPLITZ) {
 883                                 msg = "the only supported RSS hash function"
 884                                         " is Toeplitz";
 885                                 goto exit_action_not_supported;
 886                         }
 887                         if (rss->level) {
 888                                 msg = "a nonzero RSS encapsulation level is"
 889                                         " not supported";
 890                                 goto exit_action_not_supported;
 891                         }
 892                         rte_errno = 0;
 893                         fields = mlx4_conv_rss_types(priv, rss->types);
 894                         if (fields == (uint64_t)-1 && rte_errno) {
 895                                 msg = "unsupported RSS hash type requested";
 896                                 goto exit_action_not_supported;
 897                         }
 898                         flow->rss = mlx4_rss_get
 899                                 (priv, fields, rss_key, rss->queue_num,
 900                                  rss->queue);
 901                         if (!flow->rss) {
 902                                 msg = "either invalid parameters or not enough"
 903                                         " resources for additional multi-queue"
 904                                         " RSS context";
 905                                 goto exit_action_not_supported;
 906                         }
 907                         break;
 908                 default:
 909                         goto exit_action_not_supported;
 910                 }
 911         }
 912         /* When fate is unknown, drop traffic. */
 913         if (!overlap)
 914                 flow->drop = 1;
 915         /* Validation ends here. */
 916         if (!addr) {
 917                 if (flow->rss)
 918                         mlx4_rss_put(flow->rss);
 919                 return 0;
 920         }
 921         if (flow == &temp) {
 922                 /* Allocate proper handle based on collected data. */
 923                 const struct mlx4_malloc_vec vec[] = {
 924                         {
 925                                 .align = alignof(struct rte_flow),
 926                                 .size = sizeof(*flow),
 927                                 .addr = (void **)&flow,
 928                         },
 929                         {
 930                                 .align = alignof(struct ibv_flow_attr),
 931                                 .size = temp.ibv_attr_size,
 932                                 .addr = (void **)&temp.ibv_attr,
 933                         },
 934                 };
 935
 936                 if (!mlx4_zmallocv(__func__, vec, RTE_DIM(vec))) {
 937                         if (temp.rss)
 938                                 mlx4_rss_put(temp.rss);
 939                         return rte_flow_error_set
 940                                 (error, -rte_errno,
 941                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
 942                                  "flow rule handle allocation failure");
 943                 }
 944                 /* Most fields will be updated by second pass. */
 945                 *flow = (struct rte_flow){
 946                         .ibv_attr = temp.ibv_attr,
 947                         .ibv_attr_size = sizeof(*flow->ibv_attr),
 948                         .rss = temp.rss,
 949                 };
 950                 *flow->ibv_attr = (struct ibv_flow_attr){
 951                         .type = IBV_FLOW_ATTR_NORMAL,
 952                         .size = sizeof(*flow->ibv_attr),
 953                         .priority = attr->priority,
 954                         .port = priv->port,
 955                 };
 956                 goto fill;
 957         }
 958         *addr = flow;
 959         return 0;
 960 exit_item_not_supported:
 961         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 962                                   item, msg ? msg : "item not supported");
 963 exit_action_not_supported:
 964         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 965                                   action, msg ? msg : "action not supported");
 966 }
 967
 968 /**
 969  * Validate a flow supported by the NIC.
 970  *
 971  * @see rte_flow_validate()
 972  * @see rte_flow_ops
 973  */
 974 static int
 975 mlx4_flow_validate(struct rte_eth_dev *dev,
 976                    const struct rte_flow_attr *attr,
 977                    const struct rte_flow_item pattern[],
 978                    const struct rte_flow_action actions[],
 979                    struct rte_flow_error *error)
 980 {
 981         struct priv *priv = dev->data->dev_private;
 982
 983         return mlx4_flow_prepare(priv, attr, pattern, actions, error, NULL);
 984 }
 985
 986 /**
 987  * Get a drop flow rule resources instance.
 988  *
 989  * @param priv
 990  *   Pointer to private structure.
 991  *
 992  * @return
 993  *   Pointer to drop flow resources on success, NULL otherwise and rte_errno
 994  *   is set.
 995  */
 996 static struct mlx4_drop *
 997 mlx4_drop_get(struct priv *priv)
 998 {
 999         struct mlx4_drop *drop = priv->drop;
1000
1001         if (drop) {
1002                 assert(drop->refcnt);
1003                 assert(drop->priv == priv);
1004                 ++drop->refcnt;
1005                 return drop;
1006         }
1007         drop = rte_malloc(__func__, sizeof(*drop), 0);
1008         if (!drop)
1009                 goto error;
1010         *drop = (struct mlx4_drop){
1011                 .priv = priv,
1012                 .refcnt = 1,
1013         };
1014         drop->cq = mlx4_glue->create_cq(priv->ctx, 1, NULL, NULL, 0);
1015         if (!drop->cq)
1016                 goto error;
1017         drop->qp = mlx4_glue->create_qp
1018                 (priv->pd,
1019                  &(struct ibv_qp_init_attr){
1020                         .send_cq = drop->cq,
1021                         .recv_cq = drop->cq,
1022                         .qp_type = IBV_QPT_RAW_PACKET,
1023                  });
1024         if (!drop->qp)
1025                 goto error;
1026         priv->drop = drop;
1027         return drop;
1028 error:
1029         if (drop->qp)
1030                 claim_zero(mlx4_glue->destroy_qp(drop->qp));
1031         if (drop->cq)
1032                 claim_zero(mlx4_glue->destroy_cq(drop->cq));
1033         if (drop)
1034                 rte_free(drop);
1035         rte_errno = ENOMEM;
1036         return NULL;
1037 }
1038
1039 /**
1040  * Give back a drop flow rule resources instance.
1041  *
1042  * @param drop
1043  *   Pointer to drop flow rule resources.
1044  */
1045 static void
1046 mlx4_drop_put(struct mlx4_drop *drop)
1047 {
1048         assert(drop->refcnt);
1049         if (--drop->refcnt)
1050                 return;
1051         drop->priv->drop = NULL;
1052         claim_zero(mlx4_glue->destroy_qp(drop->qp));
1053         claim_zero(mlx4_glue->destroy_cq(drop->cq));
1054         rte_free(drop);
1055 }
1056
1057 /**
1058  * Toggle a configured flow rule.
1059  *
1060  * @param priv
1061  *   Pointer to private structure.
1062  * @param flow
1063  *   Flow rule handle to toggle.
1064  * @param enable
1065  *   Whether associated Verbs flow must be created or removed.
1066  * @param[out] error
1067  *   Perform verbose error reporting if not NULL.
1068  *
1069  * @return
1070  *   0 on success, a negative errno value otherwise and rte_errno is set.
1071  */
1072 static int
1073 mlx4_flow_toggle(struct priv *priv,
1074                  struct rte_flow *flow,
1075                  int enable,
1076                  struct rte_flow_error *error)
1077 {
1078         struct ibv_qp *qp = NULL;
1079         const char *msg;
1080         int err;
1081
1082         if (!enable) {
1083                 if (!flow->ibv_flow)
1084                         return 0;
1085                 claim_zero(mlx4_glue->destroy_flow(flow->ibv_flow));
1086                 flow->ibv_flow = NULL;
1087                 if (flow->drop)
1088                         mlx4_drop_put(priv->drop);
1089                 else if (flow->rss)
1090                         mlx4_rss_detach(flow->rss);
1091                 return 0;
1092         }
1093         assert(flow->ibv_attr);
1094         if (!flow->internal &&
1095             !priv->isolated &&
1096             flow->ibv_attr->priority == MLX4_FLOW_PRIORITY_LAST) {
1097                 if (flow->ibv_flow) {
1098                         claim_zero(mlx4_glue->destroy_flow(flow->ibv_flow));
1099                         flow->ibv_flow = NULL;
1100                         if (flow->drop)
1101                                 mlx4_drop_put(priv->drop);
1102                         else if (flow->rss)
1103                                 mlx4_rss_detach(flow->rss);
1104                 }
1105                 err = EACCES;
1106                 msg = ("priority level "
1107                        MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST)
1108                        " is reserved when not in isolated mode");
1109                 goto error;
1110         }
1111         if (flow->rss) {
1112                 struct mlx4_rss *rss = flow->rss;
1113                 int missing = 0;
1114                 unsigned int i;
1115
1116                 /* Stop at the first nonexistent target queue. */
1117                 for (i = 0; i != rss->queues; ++i)
1118                         if (rss->queue_id[i] >=
1119                             priv->dev->data->nb_rx_queues ||
1120                             !priv->dev->data->rx_queues[rss->queue_id[i]]) {
1121                                 missing = 1;
1122                                 break;
1123                         }
1124                 if (flow->ibv_flow) {
1125                         if (missing ^ !flow->drop)
1126                                 return 0;
1127                         /* Verbs flow needs updating. */
1128                         claim_zero(mlx4_glue->destroy_flow(flow->ibv_flow));
1129                         flow->ibv_flow = NULL;
1130                         if (flow->drop)
1131                                 mlx4_drop_put(priv->drop);
1132                         else
1133                                 mlx4_rss_detach(rss);
1134                 }
1135                 if (!missing) {
1136                         err = mlx4_rss_attach(rss);
1137                         if (err) {
1138                                 err = -err;
1139                                 msg = "cannot create indirection table or hash"
1140                                         " QP to associate flow rule with";
1141                                 goto error;
1142                         }
1143                         qp = rss->qp;
1144                 }
1145                 /* A missing target queue drops traffic implicitly. */
1146                 flow->drop = missing;
1147         }
1148         if (flow->drop) {
1149                 if (flow->ibv_flow)
1150                         return 0;
1151                 mlx4_drop_get(priv);
1152                 if (!priv->drop) {
1153                         err = rte_errno;
1154                         msg = "resources for drop flow rule cannot be created";
1155                         goto error;
1156                 }
1157                 qp = priv->drop->qp;
1158         }
1159         assert(qp);
1160         if (flow->ibv_flow)
1161                 return 0;
1162         flow->ibv_flow = mlx4_glue->create_flow(qp, flow->ibv_attr);
1163         if (flow->ibv_flow)
1164                 return 0;
1165         if (flow->drop)
1166                 mlx4_drop_put(priv->drop);
1167         else if (flow->rss)
1168                 mlx4_rss_detach(flow->rss);
1169         err = errno;
1170         msg = "flow rule rejected by device";
1171 error:
1172         return rte_flow_error_set
1173                 (error, err, RTE_FLOW_ERROR_TYPE_HANDLE, flow, msg);
1174 }
1175
1176 /**
1177  * Create a flow.
1178  *
1179  * @see rte_flow_create()
1180  * @see rte_flow_ops
1181  */
1182 static struct rte_flow *
1183 mlx4_flow_create(struct rte_eth_dev *dev,
1184                  const struct rte_flow_attr *attr,
1185                  const struct rte_flow_item pattern[],
1186                  const struct rte_flow_action actions[],
1187                  struct rte_flow_error *error)
1188 {
1189         struct priv *priv = dev->data->dev_private;
1190         struct rte_flow *flow;
1191         int err;
1192
1193         err = mlx4_flow_prepare(priv, attr, pattern, actions, error, &flow);
1194         if (err)
1195                 return NULL;
1196         err = mlx4_flow_toggle(priv, flow, priv->started, error);
1197         if (!err) {
1198                 struct rte_flow *curr = LIST_FIRST(&priv->flows);
1199
1200                 /* New rules are inserted after internal ones. */
1201                 if (!curr || !curr->internal) {
1202                         LIST_INSERT_HEAD(&priv->flows, flow, next);
1203                 } else {
1204                         while (LIST_NEXT(curr, next) &&
1205                                LIST_NEXT(curr, next)->internal)
1206                                 curr = LIST_NEXT(curr, next);
1207                         LIST_INSERT_AFTER(curr, flow, next);
1208                 }
1209                 return flow;
1210         }
1211         if (flow->rss)
1212                 mlx4_rss_put(flow->rss);
1213         rte_flow_error_set(error, -err, RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1214                            error->message);
1215         rte_free(flow);
1216         return NULL;
1217 }
1218
1219 /**
1220  * Configure isolated mode.
1221  *
1222  * @see rte_flow_isolate()
1223  * @see rte_flow_ops
1224  */
1225 static int
1226 mlx4_flow_isolate(struct rte_eth_dev *dev,
1227                   int enable,
1228                   struct rte_flow_error *error)
1229 {
1230         struct priv *priv = dev->data->dev_private;
1231
1232         if (!!enable == !!priv->isolated)
1233                 return 0;
1234         priv->isolated = !!enable;
1235         if (mlx4_flow_sync(priv, error)) {
1236                 priv->isolated = !enable;
1237                 return -rte_errno;
1238         }
1239         return 0;
1240 }
1241
1242 /**
1243  * Destroy a flow rule.
1244  *
1245  * @see rte_flow_destroy()
1246  * @see rte_flow_ops
1247  */
1248 static int
1249 mlx4_flow_destroy(struct rte_eth_dev *dev,
1250                   struct rte_flow *flow,
1251                   struct rte_flow_error *error)
1252 {
1253         struct priv *priv = dev->data->dev_private;
1254         int err = mlx4_flow_toggle(priv, flow, 0, error);
1255
1256         if (err)
1257                 return err;
1258         LIST_REMOVE(flow, next);
1259         if (flow->rss)
1260                 mlx4_rss_put(flow->rss);
1261         rte_free(flow);
1262         return 0;
1263 }
1264
1265 /**
1266  * Destroy user-configured flow rules.
1267  *
1268  * This function skips internal flows rules.
1269  *
1270  * @see rte_flow_flush()
1271  * @see rte_flow_ops
1272  */
1273 static int
1274 mlx4_flow_flush(struct rte_eth_dev *dev,
1275                 struct rte_flow_error *error)
1276 {
1277         struct priv *priv = dev->data->dev_private;
1278         struct rte_flow *flow = LIST_FIRST(&priv->flows);
1279
1280         while (flow) {
1281                 struct rte_flow *next = LIST_NEXT(flow, next);
1282
1283                 if (!flow->internal)
1284                         mlx4_flow_destroy(dev, flow, error);
1285                 flow = next;
1286         }
1287         return 0;
1288 }
1289
1290 /**
1291  * Helper function to determine the next configured VLAN filter.
1292  *
1293  * @param priv
1294  *   Pointer to private structure.
1295  * @param vlan
1296  *   VLAN ID to use as a starting point.
1297  *
1298  * @return
1299  *   Next configured VLAN ID or a high value (>= 4096) if there is none.
1300  */
1301 static uint16_t
1302 mlx4_flow_internal_next_vlan(struct priv *priv, uint16_t vlan)
1303 {
1304         while (vlan < 4096) {
1305                 if (priv->dev->data->vlan_filter_conf.ids[vlan / 64] &
1306                     (UINT64_C(1) << (vlan % 64)))
1307                         return vlan;
1308                 ++vlan;
1309         }
1310         return vlan;
1311 }
1312
1313 /**
1314  * Generate internal flow rules.
1315  *
1316  * Various flow rules are created depending on the mode the device is in:
1317  *
1318  * 1. Promiscuous:
1319  *       port MAC + broadcast + catch-all (VLAN filtering is ignored).
1320  * 2. All multicast:
1321  *       port MAC/VLAN + broadcast + catch-all multicast.
1322  * 3. Otherwise:
1323  *       port MAC/VLAN + broadcast MAC/VLAN.
1324  *
1325  * About MAC flow rules:
1326  *
1327  * - MAC flow rules are generated from @p dev->data->mac_addrs
1328  *   (@p priv->mac array).
1329  * - An additional flow rule for Ethernet broadcasts is also generated.
1330  * - All these are per-VLAN if @p DEV_RX_OFFLOAD_VLAN_FILTER
1331  *   is enabled and VLAN filters are configured.
1332  *
1333  * @param priv
1334  *   Pointer to private structure.
1335  * @param[out] error
1336  *   Perform verbose error reporting if not NULL.
1337  *
1338  * @return
1339  *   0 on success, a negative errno value otherwise and rte_errno is set.
1340  */
1341 static int
1342 mlx4_flow_internal(struct priv *priv, struct rte_flow_error *error)
1343 {
1344         struct rte_flow_attr attr = {
1345                 .priority = MLX4_FLOW_PRIORITY_LAST,
1346                 .ingress = 1,
1347         };
1348         struct rte_flow_item_eth eth_spec;
1349         const struct rte_flow_item_eth eth_mask = {
1350                 .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
1351         };
1352         const struct rte_flow_item_eth eth_allmulti = {
1353                 .dst.addr_bytes = "\x01\x00\x00\x00\x00\x00",
1354         };
1355         struct rte_flow_item_vlan vlan_spec;
1356         const struct rte_flow_item_vlan vlan_mask = {
1357                 .tci = RTE_BE16(0x0fff),
1358         };
1359         struct rte_flow_item pattern[] = {
1360                 {
1361                         .type = MLX4_FLOW_ITEM_TYPE_INTERNAL,
1362                 },
1363                 {
1364                         .type = RTE_FLOW_ITEM_TYPE_ETH,
1365                         .spec = &eth_spec,
1366                         .mask = &eth_mask,
1367                 },
1368                 {
1369                         /* Replaced with VLAN if filtering is enabled. */
1370                         .type = RTE_FLOW_ITEM_TYPE_END,
1371                 },
1372                 {
1373                         .type = RTE_FLOW_ITEM_TYPE_END,
1374                 },
1375         };
1376         /*
1377          * Round number of queues down to their previous power of 2 to
1378          * comply with RSS context limitations. Extra queues silently do not
1379          * get RSS by default.
1380          */
1381         uint32_t queues =
1382                 rte_align32pow2(priv->dev->data->nb_rx_queues + 1) >> 1;
1383         uint16_t queue[queues];
1384         struct rte_flow_action_rss action_rss = {
1385                 .func = RTE_ETH_HASH_FUNCTION_DEFAULT,
1386                 .level = 0,
1387                 .types = 0,
1388                 .key_len = MLX4_RSS_HASH_KEY_SIZE,
1389                 .queue_num = queues,
1390                 .key = mlx4_rss_hash_key_default,
1391                 .queue = queue,
1392         };
1393         struct rte_flow_action actions[] = {
1394                 {
1395                         .type = RTE_FLOW_ACTION_TYPE_RSS,
1396                         .conf = &action_rss,
1397                 },
1398                 {
1399                         .type = RTE_FLOW_ACTION_TYPE_END,
1400                 },
1401         };
1402         struct ether_addr *rule_mac = &eth_spec.dst;
1403         rte_be16_t *rule_vlan =
1404                 (priv->dev->data->dev_conf.rxmode.offloads &
1405                  DEV_RX_OFFLOAD_VLAN_FILTER) &&
1406                 !priv->dev->data->promiscuous ?
1407                 &vlan_spec.tci :
1408                 NULL;
1409         uint16_t vlan = 0;
1410         struct rte_flow *flow;
1411         unsigned int i;
1412         int err = 0;
1413
1414         /* Nothing to be done if there are no Rx queues. */
1415         if (!queues)
1416                 goto error;
1417         /* Prepare default RSS configuration. */
1418         for (i = 0; i != queues; ++i)
1419                 queue[i] = i;
1420         /*
1421          * Set up VLAN item if filtering is enabled and at least one VLAN
1422          * filter is configured.
1423          */
1424         if (rule_vlan) {
1425                 vlan = mlx4_flow_internal_next_vlan(priv, 0);
1426                 if (vlan < 4096) {
1427                         pattern[2] = (struct rte_flow_item){
1428                                 .type = RTE_FLOW_ITEM_TYPE_VLAN,
1429                                 .spec = &vlan_spec,
1430                                 .mask = &vlan_mask,
1431                         };
1432 next_vlan:
1433                         *rule_vlan = rte_cpu_to_be_16(vlan);
1434                 } else {
1435                         rule_vlan = NULL;
1436                 }
1437         }
1438         for (i = 0; i != RTE_DIM(priv->mac) + 1; ++i) {
1439                 const struct ether_addr *mac;
1440
1441                 /* Broadcasts are handled by an extra iteration. */
1442                 if (i < RTE_DIM(priv->mac))
1443                         mac = &priv->mac[i];
1444                 else
1445                         mac = &eth_mask.dst;
1446                 if (is_zero_ether_addr(mac))
1447                         continue;
1448                 /* Check if MAC flow rule is already present. */
1449                 for (flow = LIST_FIRST(&priv->flows);
1450                      flow && flow->internal;
1451                      flow = LIST_NEXT(flow, next)) {
1452                         const struct ibv_flow_spec_eth *eth =
1453                                 (const void *)((uintptr_t)flow->ibv_attr +
1454                                                sizeof(*flow->ibv_attr));
1455                         unsigned int j;
1456
1457                         if (!flow->mac)
1458                                 continue;
1459                         assert(flow->ibv_attr->type == IBV_FLOW_ATTR_NORMAL);
1460                         assert(flow->ibv_attr->num_of_specs == 1);
1461                         assert(eth->type == IBV_FLOW_SPEC_ETH);
1462                         assert(flow->rss);
1463                         if (rule_vlan &&
1464                             (eth->val.vlan_tag != *rule_vlan ||
1465                              eth->mask.vlan_tag != RTE_BE16(0x0fff)))
1466                                 continue;
1467                         if (!rule_vlan && eth->mask.vlan_tag)
1468                                 continue;
1469                         for (j = 0; j != sizeof(mac->addr_bytes); ++j)
1470                                 if (eth->val.dst_mac[j] != mac->addr_bytes[j] ||
1471                                     eth->mask.dst_mac[j] != UINT8_C(0xff) ||
1472                                     eth->val.src_mac[j] != UINT8_C(0x00) ||
1473                                     eth->mask.src_mac[j] != UINT8_C(0x00))
1474                                         break;
1475                         if (j != sizeof(mac->addr_bytes))
1476                                 continue;
1477                         if (flow->rss->queues != queues ||
1478                             memcmp(flow->rss->queue_id, action_rss.queue,
1479                                    queues * sizeof(flow->rss->queue_id[0])))
1480                                 continue;
1481                         break;
1482                 }
1483                 if (!flow || !flow->internal) {
1484                         /* Not found, create a new flow rule. */
1485                         memcpy(rule_mac, mac, sizeof(*mac));
1486                         flow = mlx4_flow_create(priv->dev, &attr, pattern,
1487                                                 actions, error);
1488                         if (!flow) {
1489                                 err = -rte_errno;
1490                                 goto error;
1491                         }
1492                 }
1493                 flow->select = 1;
1494                 flow->mac = 1;
1495         }
1496         if (rule_vlan) {
1497                 vlan = mlx4_flow_internal_next_vlan(priv, vlan + 1);
1498                 if (vlan < 4096)
1499                         goto next_vlan;
1500         }
1501         /* Take care of promiscuous and all multicast flow rules. */
1502         if (priv->dev->data->promiscuous || priv->dev->data->all_multicast) {
1503                 for (flow = LIST_FIRST(&priv->flows);
1504                      flow && flow->internal;
1505                      flow = LIST_NEXT(flow, next)) {
1506                         if (priv->dev->data->promiscuous) {
1507                                 if (flow->promisc)
1508                                         break;
1509                         } else {
1510                                 assert(priv->dev->data->all_multicast);
1511                                 if (flow->allmulti)
1512                                         break;
1513                         }
1514                 }
1515                 if (flow && flow->internal) {
1516                         assert(flow->rss);
1517                         if (flow->rss->queues != queues ||
1518                             memcmp(flow->rss->queue_id, action_rss.queue,
1519                                    queues * sizeof(flow->rss->queue_id[0])))
1520                                 flow = NULL;
1521                 }
1522                 if (!flow || !flow->internal) {
1523                         /* Not found, create a new flow rule. */
1524                         if (priv->dev->data->promiscuous) {
1525                                 pattern[1].spec = NULL;
1526                                 pattern[1].mask = NULL;
1527                         } else {
1528                                 assert(priv->dev->data->all_multicast);
1529                                 pattern[1].spec = &eth_allmulti;
1530                                 pattern[1].mask = &eth_allmulti;
1531                         }
1532                         pattern[2] = pattern[3];
1533                         flow = mlx4_flow_create(priv->dev, &attr, pattern,
1534                                                 actions, error);
1535                         if (!flow) {
1536                                 err = -rte_errno;
1537                                 goto error;
1538                         }
1539                 }
1540                 assert(flow->promisc || flow->allmulti);
1541                 flow->select = 1;
1542         }
1543 error:
1544         /* Clear selection and clean up stale internal flow rules. */
1545         flow = LIST_FIRST(&priv->flows);
1546         while (flow && flow->internal) {
1547                 struct rte_flow *next = LIST_NEXT(flow, next);
1548
1549                 if (!flow->select)
1550                         claim_zero(mlx4_flow_destroy(priv->dev, flow, error));
1551                 else
1552                         flow->select = 0;
1553                 flow = next;
1554         }
1555         return err;
1556 }
1557
1558 /**
1559  * Synchronize flow rules.
1560  *
1561  * This function synchronizes flow rules with the state of the device by
1562  * taking into account isolated mode and whether target queues are
1563  * configured.
1564  *
1565  * @param priv
1566  *   Pointer to private structure.
1567  * @param[out] error
1568  *   Perform verbose error reporting if not NULL.
1569  *
1570  * @return
1571  *   0 on success, a negative errno value otherwise and rte_errno is set.
1572  */
1573 int
1574 mlx4_flow_sync(struct priv *priv, struct rte_flow_error *error)
1575 {
1576         struct rte_flow *flow;
1577         int ret;
1578
1579         /* Internal flow rules are guaranteed to come first in the list. */
1580         if (priv->isolated) {
1581                 /*
1582                  * Get rid of them in isolated mode, stop at the first
1583                  * non-internal rule found.
1584                  */
1585                 for (flow = LIST_FIRST(&priv->flows);
1586                      flow && flow->internal;
1587                      flow = LIST_FIRST(&priv->flows))
1588                         claim_zero(mlx4_flow_destroy(priv->dev, flow, error));
1589         } else {
1590                 /* Refresh internal rules. */
1591                 ret = mlx4_flow_internal(priv, error);
1592                 if (ret)
1593                         return ret;
1594         }
1595         /* Toggle the remaining flow rules . */
1596         LIST_FOREACH(flow, &priv->flows, next) {
1597                 ret = mlx4_flow_toggle(priv, flow, priv->started, error);
1598                 if (ret)
1599                         return ret;
1600         }
1601         if (!priv->started)
1602                 assert(!priv->drop);
1603         return 0;
1604 }
1605
1606 /**
1607  * Clean up all flow rules.
1608  *
1609  * Unlike mlx4_flow_flush(), this function takes care of all remaining flow
1610  * rules regardless of whether they are internal or user-configured.
1611  *
1612  * @param priv
1613  *   Pointer to private structure.
1614  */
1615 void
1616 mlx4_flow_clean(struct priv *priv)
1617 {
1618         struct rte_flow *flow;
1619
1620         while ((flow = LIST_FIRST(&priv->flows)))
1621                 mlx4_flow_destroy(priv->dev, flow, NULL);
1622         assert(LIST_EMPTY(&priv->rss));
1623 }
1624
1625 static const struct rte_flow_ops mlx4_flow_ops = {
1626         .validate = mlx4_flow_validate,
1627         .create = mlx4_flow_create,
1628         .destroy = mlx4_flow_destroy,
1629         .flush = mlx4_flow_flush,
1630         .isolate = mlx4_flow_isolate,
1631 };
1632
1633 /**
1634  * Manage filter operations.
1635  *
1636  * @param dev
1637  *   Pointer to Ethernet device structure.
1638  * @param filter_type
1639  *   Filter type.
1640  * @param filter_op
1641  *   Operation to perform.
1642  * @param arg
1643  *   Pointer to operation-specific structure.
1644  *
1645  * @return
1646  *   0 on success, negative errno value otherwise and rte_errno is set.
1647  */
1648 int
1649 mlx4_filter_ctrl(struct rte_eth_dev *dev,
1650                  enum rte_filter_type filter_type,
1651                  enum rte_filter_op filter_op,
1652                  void *arg)
1653 {
1654         switch (filter_type) {
1655         case RTE_ETH_FILTER_GENERIC:
1656                 if (filter_op != RTE_ETH_FILTER_GET)
1657                         break;
1658                 *(const void **)arg = &mlx4_flow_ops;
1659                 return 0;
1660         default:
1661                 ERROR("%p: filter type (%d) not supported",
1662                       (void *)dev, filter_type);
1663                 break;
1664         }
1665         rte_errno = ENOTSUP;
1666         return -rte_errno;
1667 }