drivers/net/mlx4/mlx4_flow.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright 2017 6WIND S.A.
   5  *   Copyright 2017 Mellanox
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of 6WIND S.A. nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 /**
  35  * @file
  36  * Flow API operations for mlx4 driver.
  37  */
  38
  39 #include <arpa/inet.h>
  40 #include <assert.h>
  41 #include <errno.h>
  42 #include <stdalign.h>
  43 #include <stddef.h>
  44 #include <stdint.h>
  45 #include <string.h>
  46 #include <sys/queue.h>
  47
  48 /* Verbs headers do not support -pedantic. */
  49 #ifdef PEDANTIC
  50 #pragma GCC diagnostic ignored "-Wpedantic"
  51 #endif
  52 #include <infiniband/verbs.h>
  53 #ifdef PEDANTIC
  54 #pragma GCC diagnostic error "-Wpedantic"
  55 #endif
  56
  57 #include <rte_byteorder.h>
  58 #include <rte_errno.h>
  59 #include <rte_eth_ctrl.h>
  60 #include <rte_ethdev.h>
  61 #include <rte_ether.h>
  62 #include <rte_flow.h>
  63 #include <rte_flow_driver.h>
  64 #include <rte_malloc.h>
  65
  66 /* PMD headers. */
  67 #include "mlx4.h"
  68 #include "mlx4_flow.h"
  69 #include "mlx4_rxtx.h"
  70 #include "mlx4_utils.h"
  71
  72 /** Static initializer for a list of subsequent item types. */
  73 #define NEXT_ITEM(...) \
  74         (const enum rte_flow_item_type []){ \
  75                 __VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \
  76         }
  77
  78 /** Processor structure associated with a flow item. */
  79 struct mlx4_flow_proc_item {
  80         /** Bit-mask for fields supported by this PMD. */
  81         const void *mask_support;
  82         /** Bit-mask to use when @p item->mask is not provided. */
  83         const void *mask_default;
  84         /** Size in bytes for @p mask_support and @p mask_default. */
  85         const unsigned int mask_sz;
  86         /** Merge a pattern item into a flow rule handle. */
  87         int (*merge)(struct rte_flow *flow,
  88                      const struct rte_flow_item *item,
  89                      const struct mlx4_flow_proc_item *proc,
  90                      struct rte_flow_error *error);
  91         /** Size in bytes of the destination structure. */
  92         const unsigned int dst_sz;
  93         /** List of possible subsequent items. */
  94         const enum rte_flow_item_type *const next_item;
  95 };
  96
  97 /** Shared resources for drop flow rules. */
  98 struct mlx4_drop {
  99         struct ibv_qp *qp; /**< QP target. */
 100         struct ibv_cq *cq; /**< CQ associated with above QP. */
 101         struct priv *priv; /**< Back pointer to private data. */
 102         uint32_t refcnt; /**< Reference count. */
 103 };
 104
 105 /**
 106  * Convert DPDK RSS hash fields to their Verbs equivalent.
 107  *
 108  * This function returns the supported (default) set when @p rss_hf has
 109  * special value (uint64_t)-1.
 110  *
 111  * @param rss_hf
 112  *   Hash fields in DPDK format (see struct rte_eth_rss_conf).
 113  *
 114  * @return
 115  *   A valid Verbs RSS hash fields mask for mlx4 on success, (uint64_t)-1
 116  *   otherwise and rte_errno is set.
 117  */
 118 static uint64_t
 119 mlx4_conv_rss_hf(uint64_t rss_hf)
 120 {
 121         enum { IPV4, IPV6, TCP, UDP, };
 122         const uint64_t in[] = {
 123                 [IPV4] = (ETH_RSS_IPV4 |
 124                           ETH_RSS_FRAG_IPV4 |
 125                           ETH_RSS_NONFRAG_IPV4_TCP |
 126                           ETH_RSS_NONFRAG_IPV4_UDP |
 127                           ETH_RSS_NONFRAG_IPV4_OTHER),
 128                 [IPV6] = (ETH_RSS_IPV6 |
 129                           ETH_RSS_FRAG_IPV6 |
 130                           ETH_RSS_NONFRAG_IPV6_TCP |
 131                           ETH_RSS_NONFRAG_IPV6_UDP |
 132                           ETH_RSS_NONFRAG_IPV6_OTHER |
 133                           ETH_RSS_IPV6_EX |
 134                           ETH_RSS_IPV6_TCP_EX |
 135                           ETH_RSS_IPV6_UDP_EX),
 136                 [TCP] = (ETH_RSS_NONFRAG_IPV4_TCP |
 137                          ETH_RSS_NONFRAG_IPV6_TCP |
 138                          ETH_RSS_IPV6_TCP_EX),
 139                 /*
 140                  * UDP support is temporarily disabled due to an
 141                  * implementation issue in the kernel.
 142                  */
 143                 [UDP] = 0,
 144         };
 145         const uint64_t out[RTE_DIM(in)] = {
 146                 [IPV4] = IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4,
 147                 [IPV6] = IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6,
 148                 [TCP] = IBV_RX_HASH_SRC_PORT_TCP | IBV_RX_HASH_DST_PORT_TCP,
 149                 [UDP] = IBV_RX_HASH_SRC_PORT_UDP | IBV_RX_HASH_DST_PORT_UDP,
 150         };
 151         uint64_t seen = 0;
 152         uint64_t conv = 0;
 153         unsigned int i;
 154
 155         for (i = 0; i != RTE_DIM(in); ++i)
 156                 if (rss_hf & in[i]) {
 157                         seen |= rss_hf & in[i];
 158                         conv |= out[i];
 159                 }
 160         if (rss_hf == (uint64_t)-1)
 161                 return conv;
 162         if (!(rss_hf & ~seen))
 163                 return conv;
 164         rte_errno = ENOTSUP;
 165         return (uint64_t)-1;
 166 }
 167
 168 /**
 169  * Merge Ethernet pattern item into flow rule handle.
 170  *
 171  * Additional mlx4-specific constraints on supported fields:
 172  *
 173  * - No support for partial masks, except in the specific case of matching
 174  *   all multicast traffic (@p spec->dst and @p mask->dst equal to
 175  *   01:00:00:00:00:00).
 176  * - Not providing @p item->spec or providing an empty @p mask->dst is
 177  *   *only* supported if the rule doesn't specify additional matching
 178  *   criteria (i.e. rule is promiscuous-like).
 179  *
 180  * @param[in, out] flow
 181  *   Flow rule handle to update.
 182  * @param[in] item
 183  *   Pattern item to merge.
 184  * @param[in] proc
 185  *   Associated item-processing object.
 186  * @param[out] error
 187  *   Perform verbose error reporting if not NULL.
 188  *
 189  * @return
 190  *   0 on success, a negative errno value otherwise and rte_errno is set.
 191  */
 192 static int
 193 mlx4_flow_merge_eth(struct rte_flow *flow,
 194                     const struct rte_flow_item *item,
 195                     const struct mlx4_flow_proc_item *proc,
 196                     struct rte_flow_error *error)
 197 {
 198         const struct rte_flow_item_eth *spec = item->spec;
 199         const struct rte_flow_item_eth *mask =
 200                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 201         struct ibv_flow_spec_eth *eth;
 202         const char *msg;
 203         unsigned int i;
 204
 205         if (!mask) {
 206                 flow->promisc = 1;
 207         } else {
 208                 uint32_t sum_dst = 0;
 209                 uint32_t sum_src = 0;
 210
 211                 for (i = 0; i != sizeof(mask->dst.addr_bytes); ++i) {
 212                         sum_dst += mask->dst.addr_bytes[i];
 213                         sum_src += mask->src.addr_bytes[i];
 214                 }
 215                 if (sum_src) {
 216                         msg = "mlx4 does not support source MAC matching";
 217                         goto error;
 218                 } else if (!sum_dst) {
 219                         flow->promisc = 1;
 220                 } else if (sum_dst == 1 && mask->dst.addr_bytes[0] == 1) {
 221                         if (!(spec->dst.addr_bytes[0] & 1)) {
 222                                 msg = "mlx4 does not support the explicit"
 223                                         " exclusion of all multicast traffic";
 224                                 goto error;
 225                         }
 226                         flow->allmulti = 1;
 227                 } else if (sum_dst != (UINT8_C(0xff) * ETHER_ADDR_LEN)) {
 228                         msg = "mlx4 does not support matching partial"
 229                                 " Ethernet fields";
 230                         goto error;
 231                 }
 232         }
 233         if (!flow->ibv_attr)
 234                 return 0;
 235         if (flow->promisc) {
 236                 flow->ibv_attr->type = IBV_FLOW_ATTR_ALL_DEFAULT;
 237                 return 0;
 238         }
 239         if (flow->allmulti) {
 240                 flow->ibv_attr->type = IBV_FLOW_ATTR_MC_DEFAULT;
 241                 return 0;
 242         }
 243         ++flow->ibv_attr->num_of_specs;
 244         eth = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 245         *eth = (struct ibv_flow_spec_eth) {
 246                 .type = IBV_FLOW_SPEC_ETH,
 247                 .size = sizeof(*eth),
 248         };
 249         memcpy(eth->val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
 250         memcpy(eth->mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
 251         /* Remove unwanted bits from values. */
 252         for (i = 0; i < ETHER_ADDR_LEN; ++i) {
 253                 eth->val.dst_mac[i] &= eth->mask.dst_mac[i];
 254         }
 255         return 0;
 256 error:
 257         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 258                                   item, msg);
 259 }
 260
 261 /**
 262  * Merge VLAN pattern item into flow rule handle.
 263  *
 264  * Additional mlx4-specific constraints on supported fields:
 265  *
 266  * - Matching *all* VLAN traffic by omitting @p item->spec or providing an
 267  *   empty @p item->mask would also include non-VLAN traffic. Doing so is
 268  *   therefore unsupported.
 269  * - No support for partial masks.
 270  *
 271  * @param[in, out] flow
 272  *   Flow rule handle to update.
 273  * @param[in] item
 274  *   Pattern item to merge.
 275  * @param[in] proc
 276  *   Associated item-processing object.
 277  * @param[out] error
 278  *   Perform verbose error reporting if not NULL.
 279  *
 280  * @return
 281  *   0 on success, a negative errno value otherwise and rte_errno is set.
 282  */
 283 static int
 284 mlx4_flow_merge_vlan(struct rte_flow *flow,
 285                      const struct rte_flow_item *item,
 286                      const struct mlx4_flow_proc_item *proc,
 287                      struct rte_flow_error *error)
 288 {
 289         const struct rte_flow_item_vlan *spec = item->spec;
 290         const struct rte_flow_item_vlan *mask =
 291                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 292         struct ibv_flow_spec_eth *eth;
 293         const char *msg;
 294
 295         if (!mask || !mask->tci) {
 296                 msg = "mlx4 cannot match all VLAN traffic while excluding"
 297                         " non-VLAN traffic, TCI VID must be specified";
 298                 goto error;
 299         }
 300         if (mask->tci != RTE_BE16(0x0fff)) {
 301                 msg = "mlx4 does not support partial TCI VID matching";
 302                 goto error;
 303         }
 304         if (!flow->ibv_attr)
 305                 return 0;
 306         eth = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size -
 307                        sizeof(*eth));
 308         eth->val.vlan_tag = spec->tci;
 309         eth->mask.vlan_tag = mask->tci;
 310         eth->val.vlan_tag &= eth->mask.vlan_tag;
 311         return 0;
 312 error:
 313         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 314                                   item, msg);
 315 }
 316
 317 /**
 318  * Merge IPv4 pattern item into flow rule handle.
 319  *
 320  * Additional mlx4-specific constraints on supported fields:
 321  *
 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_ipv4(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_ipv4 *spec = item->spec;
 343         const struct rte_flow_item_ipv4 *mask =
 344                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 345         struct ibv_flow_spec_ipv4 *ipv4;
 346         const char *msg;
 347
 348         if (mask &&
 349             ((uint32_t)(mask->hdr.src_addr + 1) > UINT32_C(1) ||
 350              (uint32_t)(mask->hdr.dst_addr + 1) > UINT32_C(1))) {
 351                 msg = "mlx4 does not support matching partial IPv4 fields";
 352                 goto error;
 353         }
 354         if (!flow->ibv_attr)
 355                 return 0;
 356         ++flow->ibv_attr->num_of_specs;
 357         ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 358         *ipv4 = (struct ibv_flow_spec_ipv4) {
 359                 .type = IBV_FLOW_SPEC_IPV4,
 360                 .size = sizeof(*ipv4),
 361         };
 362         if (!spec)
 363                 return 0;
 364         ipv4->val = (struct ibv_flow_ipv4_filter) {
 365                 .src_ip = spec->hdr.src_addr,
 366                 .dst_ip = spec->hdr.dst_addr,
 367         };
 368         ipv4->mask = (struct ibv_flow_ipv4_filter) {
 369                 .src_ip = mask->hdr.src_addr,
 370                 .dst_ip = mask->hdr.dst_addr,
 371         };
 372         /* Remove unwanted bits from values. */
 373         ipv4->val.src_ip &= ipv4->mask.src_ip;
 374         ipv4->val.dst_ip &= ipv4->mask.dst_ip;
 375         return 0;
 376 error:
 377         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 378                                   item, msg);
 379 }
 380
 381 /**
 382  * Merge UDP pattern item into flow rule handle.
 383  *
 384  * Additional mlx4-specific constraints on supported fields:
 385  *
 386  * - No support for partial masks.
 387  *
 388  * @param[in, out] flow
 389  *   Flow rule handle to update.
 390  * @param[in] item
 391  *   Pattern item to merge.
 392  * @param[in] proc
 393  *   Associated item-processing object.
 394  * @param[out] error
 395  *   Perform verbose error reporting if not NULL.
 396  *
 397  * @return
 398  *   0 on success, a negative errno value otherwise and rte_errno is set.
 399  */
 400 static int
 401 mlx4_flow_merge_udp(struct rte_flow *flow,
 402                     const struct rte_flow_item *item,
 403                     const struct mlx4_flow_proc_item *proc,
 404                     struct rte_flow_error *error)
 405 {
 406         const struct rte_flow_item_udp *spec = item->spec;
 407         const struct rte_flow_item_udp *mask =
 408                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 409         struct ibv_flow_spec_tcp_udp *udp;
 410         const char *msg;
 411
 412         if (mask &&
 413             ((uint16_t)(mask->hdr.src_port + 1) > UINT16_C(1) ||
 414              (uint16_t)(mask->hdr.dst_port + 1) > UINT16_C(1))) {
 415                 msg = "mlx4 does not support matching partial UDP fields";
 416                 goto error;
 417         }
 418         if (!flow->ibv_attr)
 419                 return 0;
 420         ++flow->ibv_attr->num_of_specs;
 421         udp = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 422         *udp = (struct ibv_flow_spec_tcp_udp) {
 423                 .type = IBV_FLOW_SPEC_UDP,
 424                 .size = sizeof(*udp),
 425         };
 426         if (!spec)
 427                 return 0;
 428         udp->val.dst_port = spec->hdr.dst_port;
 429         udp->val.src_port = spec->hdr.src_port;
 430         udp->mask.dst_port = mask->hdr.dst_port;
 431         udp->mask.src_port = mask->hdr.src_port;
 432         /* Remove unwanted bits from values. */
 433         udp->val.src_port &= udp->mask.src_port;
 434         udp->val.dst_port &= udp->mask.dst_port;
 435         return 0;
 436 error:
 437         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 438                                   item, msg);
 439 }
 440
 441 /**
 442  * Merge TCP pattern item into flow rule handle.
 443  *
 444  * Additional mlx4-specific constraints on supported fields:
 445  *
 446  * - No support for partial masks.
 447  *
 448  * @param[in, out] flow
 449  *   Flow rule handle to update.
 450  * @param[in] item
 451  *   Pattern item to merge.
 452  * @param[in] proc
 453  *   Associated item-processing object.
 454  * @param[out] error
 455  *   Perform verbose error reporting if not NULL.
 456  *
 457  * @return
 458  *   0 on success, a negative errno value otherwise and rte_errno is set.
 459  */
 460 static int
 461 mlx4_flow_merge_tcp(struct rte_flow *flow,
 462                     const struct rte_flow_item *item,
 463                     const struct mlx4_flow_proc_item *proc,
 464                     struct rte_flow_error *error)
 465 {
 466         const struct rte_flow_item_tcp *spec = item->spec;
 467         const struct rte_flow_item_tcp *mask =
 468                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 469         struct ibv_flow_spec_tcp_udp *tcp;
 470         const char *msg;
 471
 472         if (mask &&
 473             ((uint16_t)(mask->hdr.src_port + 1) > UINT16_C(1) ||
 474              (uint16_t)(mask->hdr.dst_port + 1) > UINT16_C(1))) {
 475                 msg = "mlx4 does not support matching partial TCP fields";
 476                 goto error;
 477         }
 478         if (!flow->ibv_attr)
 479                 return 0;
 480         ++flow->ibv_attr->num_of_specs;
 481         tcp = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 482         *tcp = (struct ibv_flow_spec_tcp_udp) {
 483                 .type = IBV_FLOW_SPEC_TCP,
 484                 .size = sizeof(*tcp),
 485         };
 486         if (!spec)
 487                 return 0;
 488         tcp->val.dst_port = spec->hdr.dst_port;
 489         tcp->val.src_port = spec->hdr.src_port;
 490         tcp->mask.dst_port = mask->hdr.dst_port;
 491         tcp->mask.src_port = mask->hdr.src_port;
 492         /* Remove unwanted bits from values. */
 493         tcp->val.src_port &= tcp->mask.src_port;
 494         tcp->val.dst_port &= tcp->mask.dst_port;
 495         return 0;
 496 error:
 497         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 498                                   item, msg);
 499 }
 500
 501 /**
 502  * Perform basic sanity checks on a pattern item.
 503  *
 504  * @param[in] item
 505  *   Item specification.
 506  * @param[in] proc
 507  *   Associated item-processing object.
 508  * @param[out] error
 509  *   Perform verbose error reporting if not NULL.
 510  *
 511  * @return
 512  *   0 on success, a negative errno value otherwise and rte_errno is set.
 513  */
 514 static int
 515 mlx4_flow_item_check(const struct rte_flow_item *item,
 516                      const struct mlx4_flow_proc_item *proc,
 517                      struct rte_flow_error *error)
 518 {
 519         const uint8_t *mask;
 520         unsigned int i;
 521
 522         /* item->last and item->mask cannot exist without item->spec. */
 523         if (!item->spec && (item->mask || item->last))
 524                 return rte_flow_error_set
 525                         (error, EINVAL, RTE_FLOW_ERROR_TYPE_ITEM, item,
 526                          "\"mask\" or \"last\" field provided without a"
 527                          " corresponding \"spec\"");
 528         /* No spec, no mask, no problem. */
 529         if (!item->spec)
 530                 return 0;
 531         mask = item->mask ?
 532                 (const uint8_t *)item->mask :
 533                 (const uint8_t *)proc->mask_default;
 534         assert(mask);
 535         /*
 536          * Single-pass check to make sure that:
 537          * - Mask is supported, no bits are set outside proc->mask_support.
 538          * - Both item->spec and item->last are included in mask.
 539          */
 540         for (i = 0; i != proc->mask_sz; ++i) {
 541                 if (!mask[i])
 542                         continue;
 543                 if ((mask[i] | ((const uint8_t *)proc->mask_support)[i]) !=
 544                     ((const uint8_t *)proc->mask_support)[i])
 545                         return rte_flow_error_set
 546                                 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 547                                  item, "unsupported field found in \"mask\"");
 548                 if (item->last &&
 549                     (((const uint8_t *)item->spec)[i] & mask[i]) !=
 550                     (((const uint8_t *)item->last)[i] & mask[i]))
 551                         return rte_flow_error_set
 552                                 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 553                                  item,
 554                                  "range between \"spec\" and \"last\""
 555                                  " is larger than \"mask\"");
 556         }
 557         return 0;
 558 }
 559
 560 /** Graph of supported items and associated actions. */
 561 static const struct mlx4_flow_proc_item mlx4_flow_proc_item_list[] = {
 562         [RTE_FLOW_ITEM_TYPE_END] = {
 563                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_ETH),
 564         },
 565         [RTE_FLOW_ITEM_TYPE_ETH] = {
 566                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_VLAN,
 567                                        RTE_FLOW_ITEM_TYPE_IPV4),
 568                 .mask_support = &(const struct rte_flow_item_eth){
 569                         /* Only destination MAC can be matched. */
 570                         .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 571                 },
 572                 .mask_default = &rte_flow_item_eth_mask,
 573                 .mask_sz = sizeof(struct rte_flow_item_eth),
 574                 .merge = mlx4_flow_merge_eth,
 575                 .dst_sz = sizeof(struct ibv_flow_spec_eth),
 576         },
 577         [RTE_FLOW_ITEM_TYPE_VLAN] = {
 578                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_IPV4),
 579                 .mask_support = &(const struct rte_flow_item_vlan){
 580                         /* Only TCI VID matching is supported. */
 581                         .tci = RTE_BE16(0x0fff),
 582                 },
 583                 .mask_default = &rte_flow_item_vlan_mask,
 584                 .mask_sz = sizeof(struct rte_flow_item_vlan),
 585                 .merge = mlx4_flow_merge_vlan,
 586                 .dst_sz = 0,
 587         },
 588         [RTE_FLOW_ITEM_TYPE_IPV4] = {
 589                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_UDP,
 590                                        RTE_FLOW_ITEM_TYPE_TCP),
 591                 .mask_support = &(const struct rte_flow_item_ipv4){
 592                         .hdr = {
 593                                 .src_addr = RTE_BE32(0xffffffff),
 594                                 .dst_addr = RTE_BE32(0xffffffff),
 595                         },
 596                 },
 597                 .mask_default = &rte_flow_item_ipv4_mask,
 598                 .mask_sz = sizeof(struct rte_flow_item_ipv4),
 599                 .merge = mlx4_flow_merge_ipv4,
 600                 .dst_sz = sizeof(struct ibv_flow_spec_ipv4),
 601         },
 602         [RTE_FLOW_ITEM_TYPE_UDP] = {
 603                 .mask_support = &(const struct rte_flow_item_udp){
 604                         .hdr = {
 605                                 .src_port = RTE_BE16(0xffff),
 606                                 .dst_port = RTE_BE16(0xffff),
 607                         },
 608                 },
 609                 .mask_default = &rte_flow_item_udp_mask,
 610                 .mask_sz = sizeof(struct rte_flow_item_udp),
 611                 .merge = mlx4_flow_merge_udp,
 612                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 613         },
 614         [RTE_FLOW_ITEM_TYPE_TCP] = {
 615                 .mask_support = &(const struct rte_flow_item_tcp){
 616                         .hdr = {
 617                                 .src_port = RTE_BE16(0xffff),
 618                                 .dst_port = RTE_BE16(0xffff),
 619                         },
 620                 },
 621                 .mask_default = &rte_flow_item_tcp_mask,
 622                 .mask_sz = sizeof(struct rte_flow_item_tcp),
 623                 .merge = mlx4_flow_merge_tcp,
 624                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 625         },
 626 };
 627
 628 /**
 629  * Make sure a flow rule is supported and initialize associated structure.
 630  *
 631  * @param priv
 632  *   Pointer to private structure.
 633  * @param[in] attr
 634  *   Flow rule attributes.
 635  * @param[in] pattern
 636  *   Pattern specification (list terminated by the END pattern item).
 637  * @param[in] actions
 638  *   Associated actions (list terminated by the END action).
 639  * @param[out] error
 640  *   Perform verbose error reporting if not NULL.
 641  * @param[in, out] addr
 642  *   Buffer where the resulting flow rule handle pointer must be stored.
 643  *   If NULL, stop processing after validation stage.
 644  *
 645  * @return
 646  *   0 on success, a negative errno value otherwise and rte_errno is set.
 647  */
 648 static int
 649 mlx4_flow_prepare(struct priv *priv,
 650                   const struct rte_flow_attr *attr,
 651                   const struct rte_flow_item pattern[],
 652                   const struct rte_flow_action actions[],
 653                   struct rte_flow_error *error,
 654                   struct rte_flow **addr)
 655 {
 656         const struct rte_flow_item *item;
 657         const struct rte_flow_action *action;
 658         const struct mlx4_flow_proc_item *proc;
 659         struct rte_flow temp = { .ibv_attr_size = sizeof(*temp.ibv_attr) };
 660         struct rte_flow *flow = &temp;
 661         const char *msg = NULL;
 662
 663         if (attr->group)
 664                 return rte_flow_error_set
 665                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 666                          NULL, "groups are not supported");
 667         if (attr->priority > MLX4_FLOW_PRIORITY_LAST)
 668                 return rte_flow_error_set
 669                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 670                          NULL, "maximum priority level is "
 671                          MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST));
 672         if (attr->egress)
 673                 return rte_flow_error_set
 674                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 675                          NULL, "egress is not supported");
 676         if (!attr->ingress)
 677                 return rte_flow_error_set
 678                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 679                          NULL, "only ingress is supported");
 680 fill:
 681         proc = mlx4_flow_proc_item_list;
 682         /* Go over pattern. */
 683         for (item = pattern; item->type; ++item) {
 684                 const struct mlx4_flow_proc_item *next = NULL;
 685                 unsigned int i;
 686                 int err;
 687
 688                 if (item->type == RTE_FLOW_ITEM_TYPE_VOID)
 689                         continue;
 690                 if (item->type == MLX4_FLOW_ITEM_TYPE_INTERNAL) {
 691                         flow->internal = 1;
 692                         continue;
 693                 }
 694                 if (flow->promisc || flow->allmulti) {
 695                         msg = "mlx4 does not support additional matching"
 696                                 " criteria combined with indiscriminate"
 697                                 " matching on Ethernet headers";
 698                         goto exit_item_not_supported;
 699                 }
 700                 for (i = 0; proc->next_item && proc->next_item[i]; ++i) {
 701                         if (proc->next_item[i] == item->type) {
 702                                 next = &mlx4_flow_proc_item_list[item->type];
 703                                 break;
 704                         }
 705                 }
 706                 if (!next)
 707                         goto exit_item_not_supported;
 708                 proc = next;
 709                 /*
 710                  * Perform basic sanity checks only once, while handle is
 711                  * not allocated.
 712                  */
 713                 if (flow == &temp) {
 714                         err = mlx4_flow_item_check(item, proc, error);
 715                         if (err)
 716                                 return err;
 717                 }
 718                 if (proc->merge) {
 719                         err = proc->merge(flow, item, proc, error);
 720                         if (err)
 721                                 return err;
 722                 }
 723                 flow->ibv_attr_size += proc->dst_sz;
 724         }
 725         /* Go over actions list. */
 726         for (action = actions; action->type; ++action) {
 727                 switch (action->type) {
 728                         const struct rte_flow_action_queue *queue;
 729                         const struct rte_flow_action_rss *rss;
 730                         const struct rte_eth_rss_conf *rss_conf;
 731                         unsigned int i;
 732
 733                 case RTE_FLOW_ACTION_TYPE_VOID:
 734                         continue;
 735                 case RTE_FLOW_ACTION_TYPE_DROP:
 736                         flow->drop = 1;
 737                         break;
 738                 case RTE_FLOW_ACTION_TYPE_QUEUE:
 739                         if (flow->rss)
 740                                 break;
 741                         queue = action->conf;
 742                         if (queue->index >= priv->dev->data->nb_rx_queues) {
 743                                 msg = "queue target index beyond number of"
 744                                         " configured Rx queues";
 745                                 goto exit_action_not_supported;
 746                         }
 747                         flow->rss = mlx4_rss_get
 748                                 (priv, 0, mlx4_rss_hash_key_default, 1,
 749                                  &queue->index);
 750                         if (!flow->rss) {
 751                                 msg = "not enough resources for additional"
 752                                         " single-queue RSS context";
 753                                 goto exit_action_not_supported;
 754                         }
 755                         break;
 756                 case RTE_FLOW_ACTION_TYPE_RSS:
 757                         if (flow->rss)
 758                                 break;
 759                         rss = action->conf;
 760                         /* Default RSS configuration if none is provided. */
 761                         rss_conf =
 762                                 rss->rss_conf ?
 763                                 rss->rss_conf :
 764                                 &(struct rte_eth_rss_conf){
 765                                         .rss_key = mlx4_rss_hash_key_default,
 766                                         .rss_key_len = MLX4_RSS_HASH_KEY_SIZE,
 767                                         .rss_hf = -1,
 768                                 };
 769                         /* Sanity checks. */
 770                         for (i = 0; i < rss->num; ++i)
 771                                 if (rss->queue[i] >=
 772                                     priv->dev->data->nb_rx_queues)
 773                                         break;
 774                         if (i != rss->num) {
 775                                 msg = "queue index target beyond number of"
 776                                         " configured Rx queues";
 777                                 goto exit_action_not_supported;
 778                         }
 779                         if (!rte_is_power_of_2(rss->num)) {
 780                                 msg = "for RSS, mlx4 requires the number of"
 781                                         " queues to be a power of two";
 782                                 goto exit_action_not_supported;
 783                         }
 784                         if (rss_conf->rss_key_len !=
 785                             sizeof(flow->rss->key)) {
 786                                 msg = "mlx4 supports exactly one RSS hash key"
 787                                         " length: "
 788                                         MLX4_STR_EXPAND(MLX4_RSS_HASH_KEY_SIZE);
 789                                 goto exit_action_not_supported;
 790                         }
 791                         for (i = 1; i < rss->num; ++i)
 792                                 if (rss->queue[i] - rss->queue[i - 1] != 1)
 793                                         break;
 794                         if (i != rss->num) {
 795                                 msg = "mlx4 requires RSS contexts to use"
 796                                         " consecutive queue indices only";
 797                                 goto exit_action_not_supported;
 798                         }
 799                         if (rss->queue[0] % rss->num) {
 800                                 msg = "mlx4 requires the first queue of a RSS"
 801                                         " context to be aligned on a multiple"
 802                                         " of the context size";
 803                                 goto exit_action_not_supported;
 804                         }
 805                         flow->rss = mlx4_rss_get
 806                                 (priv, mlx4_conv_rss_hf(rss_conf->rss_hf),
 807                                  rss_conf->rss_key, rss->num, rss->queue);
 808                         if (!flow->rss) {
 809                                 msg = "either invalid parameters or not enough"
 810                                         " resources for additional multi-queue"
 811                                         " RSS context";
 812                                 goto exit_action_not_supported;
 813                         }
 814                         break;
 815                 default:
 816                         goto exit_action_not_supported;
 817                 }
 818         }
 819         if (!flow->rss && !flow->drop)
 820                 return rte_flow_error_set
 821                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 822                          NULL, "no valid action");
 823         /* Validation ends here. */
 824         if (!addr) {
 825                 if (flow->rss)
 826                         mlx4_rss_put(flow->rss);
 827                 return 0;
 828         }
 829         if (flow == &temp) {
 830                 /* Allocate proper handle based on collected data. */
 831                 const struct mlx4_malloc_vec vec[] = {
 832                         {
 833                                 .align = alignof(struct rte_flow),
 834                                 .size = sizeof(*flow),
 835                                 .addr = (void **)&flow,
 836                         },
 837                         {
 838                                 .align = alignof(struct ibv_flow_attr),
 839                                 .size = temp.ibv_attr_size,
 840                                 .addr = (void **)&temp.ibv_attr,
 841                         },
 842                 };
 843
 844                 if (!mlx4_zmallocv(__func__, vec, RTE_DIM(vec)))
 845                         return rte_flow_error_set
 846                                 (error, -rte_errno,
 847                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
 848                                  "flow rule handle allocation failure");
 849                 /* Most fields will be updated by second pass. */
 850                 *flow = (struct rte_flow){
 851                         .ibv_attr = temp.ibv_attr,
 852                         .ibv_attr_size = sizeof(*flow->ibv_attr),
 853                         .rss = temp.rss,
 854                 };
 855                 *flow->ibv_attr = (struct ibv_flow_attr){
 856                         .type = IBV_FLOW_ATTR_NORMAL,
 857                         .size = sizeof(*flow->ibv_attr),
 858                         .priority = attr->priority,
 859                         .port = priv->port,
 860                 };
 861                 goto fill;
 862         }
 863         *addr = flow;
 864         return 0;
 865 exit_item_not_supported:
 866         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 867                                   item, msg ? msg : "item not supported");
 868 exit_action_not_supported:
 869         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 870                                   action, msg ? msg : "action not supported");
 871 }
 872
 873 /**
 874  * Validate a flow supported by the NIC.
 875  *
 876  * @see rte_flow_validate()
 877  * @see rte_flow_ops
 878  */
 879 static int
 880 mlx4_flow_validate(struct rte_eth_dev *dev,
 881                    const struct rte_flow_attr *attr,
 882                    const struct rte_flow_item pattern[],
 883                    const struct rte_flow_action actions[],
 884                    struct rte_flow_error *error)
 885 {
 886         struct priv *priv = dev->data->dev_private;
 887
 888         return mlx4_flow_prepare(priv, attr, pattern, actions, error, NULL);
 889 }
 890
 891 /**
 892  * Get a drop flow rule resources instance.
 893  *
 894  * @param priv
 895  *   Pointer to private structure.
 896  *
 897  * @return
 898  *   Pointer to drop flow resources on success, NULL otherwise and rte_errno
 899  *   is set.
 900  */
 901 static struct mlx4_drop *
 902 mlx4_drop_get(struct priv *priv)
 903 {
 904         struct mlx4_drop *drop = priv->drop;
 905
 906         if (drop) {
 907                 assert(drop->refcnt);
 908                 assert(drop->priv == priv);
 909                 ++drop->refcnt;
 910                 return drop;
 911         }
 912         drop = rte_malloc(__func__, sizeof(*drop), 0);
 913         if (!drop)
 914                 goto error;
 915         *drop = (struct mlx4_drop){
 916                 .priv = priv,
 917                 .refcnt = 1,
 918         };
 919         drop->cq = ibv_create_cq(priv->ctx, 1, NULL, NULL, 0);
 920         if (!drop->cq)
 921                 goto error;
 922         drop->qp = ibv_create_qp(priv->pd,
 923                                  &(struct ibv_qp_init_attr){
 924                                         .send_cq = drop->cq,
 925                                         .recv_cq = drop->cq,
 926                                         .qp_type = IBV_QPT_RAW_PACKET,
 927                                  });
 928         if (!drop->qp)
 929                 goto error;
 930         priv->drop = drop;
 931         return drop;
 932 error:
 933         if (drop->qp)
 934                 claim_zero(ibv_destroy_qp(drop->qp));
 935         if (drop->cq)
 936                 claim_zero(ibv_destroy_cq(drop->cq));
 937         if (drop)
 938                 rte_free(drop);
 939         rte_errno = ENOMEM;
 940         return NULL;
 941 }
 942
 943 /**
 944  * Give back a drop flow rule resources instance.
 945  *
 946  * @param drop
 947  *   Pointer to drop flow rule resources.
 948  */
 949 static void
 950 mlx4_drop_put(struct mlx4_drop *drop)
 951 {
 952         assert(drop->refcnt);
 953         if (--drop->refcnt)
 954                 return;
 955         drop->priv->drop = NULL;
 956         claim_zero(ibv_destroy_qp(drop->qp));
 957         claim_zero(ibv_destroy_cq(drop->cq));
 958         rte_free(drop);
 959 }
 960
 961 /**
 962  * Toggle a configured flow rule.
 963  *
 964  * @param priv
 965  *   Pointer to private structure.
 966  * @param flow
 967  *   Flow rule handle to toggle.
 968  * @param enable
 969  *   Whether associated Verbs flow must be created or removed.
 970  * @param[out] error
 971  *   Perform verbose error reporting if not NULL.
 972  *
 973  * @return
 974  *   0 on success, a negative errno value otherwise and rte_errno is set.
 975  */
 976 static int
 977 mlx4_flow_toggle(struct priv *priv,
 978                  struct rte_flow *flow,
 979                  int enable,
 980                  struct rte_flow_error *error)
 981 {
 982         struct ibv_qp *qp = NULL;
 983         const char *msg;
 984         int err;
 985
 986         if (!enable) {
 987                 if (!flow->ibv_flow)
 988                         return 0;
 989                 claim_zero(ibv_destroy_flow(flow->ibv_flow));
 990                 flow->ibv_flow = NULL;
 991                 if (flow->drop)
 992                         mlx4_drop_put(priv->drop);
 993                 else if (flow->rss)
 994                         mlx4_rss_detach(flow->rss);
 995                 return 0;
 996         }
 997         assert(flow->ibv_attr);
 998         if (!flow->internal &&
 999             !priv->isolated &&
1000             flow->ibv_attr->priority == MLX4_FLOW_PRIORITY_LAST) {
1001                 if (flow->ibv_flow) {
1002                         claim_zero(ibv_destroy_flow(flow->ibv_flow));
1003                         flow->ibv_flow = NULL;
1004                         if (flow->drop)
1005                                 mlx4_drop_put(priv->drop);
1006                         else if (flow->rss)
1007                                 mlx4_rss_detach(flow->rss);
1008                 }
1009                 err = EACCES;
1010                 msg = ("priority level "
1011                        MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST)
1012                        " is reserved when not in isolated mode");
1013                 goto error;
1014         }
1015         if (flow->rss) {
1016                 struct mlx4_rss *rss = flow->rss;
1017                 int missing = 0;
1018                 unsigned int i;
1019
1020                 /* Stop at the first nonexistent target queue. */
1021                 for (i = 0; i != rss->queues; ++i)
1022                         if (rss->queue_id[i] >=
1023                             priv->dev->data->nb_rx_queues ||
1024                             !priv->dev->data->rx_queues[rss->queue_id[i]]) {
1025                                 missing = 1;
1026                                 break;
1027                         }
1028                 if (flow->ibv_flow) {
1029                         if (missing ^ !flow->drop)
1030                                 return 0;
1031                         /* Verbs flow needs updating. */
1032                         claim_zero(ibv_destroy_flow(flow->ibv_flow));
1033                         flow->ibv_flow = NULL;
1034                         if (flow->drop)
1035                                 mlx4_drop_put(priv->drop);
1036                         else
1037                                 mlx4_rss_detach(rss);
1038                 }
1039                 if (!missing) {
1040                         err = mlx4_rss_attach(rss);
1041                         if (err) {
1042                                 err = -err;
1043                                 msg = "cannot create indirection table or hash"
1044                                         " QP to associate flow rule with";
1045                                 goto error;
1046                         }
1047                         qp = rss->qp;
1048                 }
1049                 /* A missing target queue drops traffic implicitly. */
1050                 flow->drop = missing;
1051         }
1052         if (flow->drop) {
1053                 mlx4_drop_get(priv);
1054                 if (!priv->drop) {
1055                         err = rte_errno;
1056                         msg = "resources for drop flow rule cannot be created";
1057                         goto error;
1058                 }
1059                 qp = priv->drop->qp;
1060         }
1061         assert(qp);
1062         if (flow->ibv_flow)
1063                 return 0;
1064         flow->ibv_flow = ibv_create_flow(qp, flow->ibv_attr);
1065         if (flow->ibv_flow)
1066                 return 0;
1067         if (flow->drop)
1068                 mlx4_drop_put(priv->drop);
1069         else if (flow->rss)
1070                 mlx4_rss_detach(flow->rss);
1071         err = errno;
1072         msg = "flow rule rejected by device";
1073 error:
1074         return rte_flow_error_set
1075                 (error, err, RTE_FLOW_ERROR_TYPE_HANDLE, flow, msg);
1076 }
1077
1078 /**
1079  * Create a flow.
1080  *
1081  * @see rte_flow_create()
1082  * @see rte_flow_ops
1083  */
1084 static struct rte_flow *
1085 mlx4_flow_create(struct rte_eth_dev *dev,
1086                  const struct rte_flow_attr *attr,
1087                  const struct rte_flow_item pattern[],
1088                  const struct rte_flow_action actions[],
1089                  struct rte_flow_error *error)
1090 {
1091         struct priv *priv = dev->data->dev_private;
1092         struct rte_flow *flow;
1093         int err;
1094
1095         err = mlx4_flow_prepare(priv, attr, pattern, actions, error, &flow);
1096         if (err)
1097                 return NULL;
1098         err = mlx4_flow_toggle(priv, flow, priv->started, error);
1099         if (!err) {
1100                 struct rte_flow *curr = LIST_FIRST(&priv->flows);
1101
1102                 /* New rules are inserted after internal ones. */
1103                 if (!curr || !curr->internal) {
1104                         LIST_INSERT_HEAD(&priv->flows, flow, next);
1105                 } else {
1106                         while (LIST_NEXT(curr, next) &&
1107                                LIST_NEXT(curr, next)->internal)
1108                                 curr = LIST_NEXT(curr, next);
1109                         LIST_INSERT_AFTER(curr, flow, next);
1110                 }
1111                 return flow;
1112         }
1113         if (flow->rss)
1114                 mlx4_rss_put(flow->rss);
1115         rte_flow_error_set(error, -err, RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1116                            error->message);
1117         rte_free(flow);
1118         return NULL;
1119 }
1120
1121 /**
1122  * Configure isolated mode.
1123  *
1124  * @see rte_flow_isolate()
1125  * @see rte_flow_ops
1126  */
1127 static int
1128 mlx4_flow_isolate(struct rte_eth_dev *dev,
1129                   int enable,
1130                   struct rte_flow_error *error)
1131 {
1132         struct priv *priv = dev->data->dev_private;
1133
1134         if (!!enable == !!priv->isolated)
1135                 return 0;
1136         priv->isolated = !!enable;
1137         if (mlx4_flow_sync(priv, error)) {
1138                 priv->isolated = !enable;
1139                 return -rte_errno;
1140         }
1141         return 0;
1142 }
1143
1144 /**
1145  * Destroy a flow rule.
1146  *
1147  * @see rte_flow_destroy()
1148  * @see rte_flow_ops
1149  */
1150 static int
1151 mlx4_flow_destroy(struct rte_eth_dev *dev,
1152                   struct rte_flow *flow,
1153                   struct rte_flow_error *error)
1154 {
1155         struct priv *priv = dev->data->dev_private;
1156         int err = mlx4_flow_toggle(priv, flow, 0, error);
1157
1158         if (err)
1159                 return err;
1160         LIST_REMOVE(flow, next);
1161         if (flow->rss)
1162                 mlx4_rss_put(flow->rss);
1163         rte_free(flow);
1164         return 0;
1165 }
1166
1167 /**
1168  * Destroy user-configured flow rules.
1169  *
1170  * This function skips internal flows rules.
1171  *
1172  * @see rte_flow_flush()
1173  * @see rte_flow_ops
1174  */
1175 static int
1176 mlx4_flow_flush(struct rte_eth_dev *dev,
1177                 struct rte_flow_error *error)
1178 {
1179         struct priv *priv = dev->data->dev_private;
1180         struct rte_flow *flow = LIST_FIRST(&priv->flows);
1181
1182         while (flow) {
1183                 struct rte_flow *next = LIST_NEXT(flow, next);
1184
1185                 if (!flow->internal)
1186                         mlx4_flow_destroy(dev, flow, error);
1187                 flow = next;
1188         }
1189         return 0;
1190 }
1191
1192 /**
1193  * Helper function to determine the next configured VLAN filter.
1194  *
1195  * @param priv
1196  *   Pointer to private structure.
1197  * @param vlan
1198  *   VLAN ID to use as a starting point.
1199  *
1200  * @return
1201  *   Next configured VLAN ID or a high value (>= 4096) if there is none.
1202  */
1203 static uint16_t
1204 mlx4_flow_internal_next_vlan(struct priv *priv, uint16_t vlan)
1205 {
1206         while (vlan < 4096) {
1207                 if (priv->dev->data->vlan_filter_conf.ids[vlan / 64] &
1208                     (UINT64_C(1) << (vlan % 64)))
1209                         return vlan;
1210                 ++vlan;
1211         }
1212         return vlan;
1213 }
1214
1215 /**
1216  * Generate internal flow rules.
1217  *
1218  * Various flow rules are created depending on the mode the device is in:
1219  *
1220  * 1. Promiscuous: port MAC + catch-all (VLAN filtering is ignored).
1221  * 2. All multicast: port MAC/VLAN + catch-all multicast.
1222  * 3. Otherwise: port MAC/VLAN + broadcast MAC/VLAN.
1223  *
1224  * About MAC flow rules:
1225  *
1226  * - MAC flow rules are generated from @p dev->data->mac_addrs
1227  *   (@p priv->mac array).
1228  * - An additional flow rule for Ethernet broadcasts is also generated.
1229  * - All these are per-VLAN if @p dev->data->dev_conf.rxmode.hw_vlan_filter
1230  *   is enabled and VLAN filters are configured.
1231  *
1232  * @param priv
1233  *   Pointer to private structure.
1234  * @param[out] error
1235  *   Perform verbose error reporting if not NULL.
1236  *
1237  * @return
1238  *   0 on success, a negative errno value otherwise and rte_errno is set.
1239  */
1240 static int
1241 mlx4_flow_internal(struct priv *priv, struct rte_flow_error *error)
1242 {
1243         struct rte_flow_attr attr = {
1244                 .priority = MLX4_FLOW_PRIORITY_LAST,
1245                 .ingress = 1,
1246         };
1247         struct rte_flow_item_eth eth_spec;
1248         const struct rte_flow_item_eth eth_mask = {
1249                 .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
1250         };
1251         const struct rte_flow_item_eth eth_allmulti = {
1252                 .dst.addr_bytes = "\x01\x00\x00\x00\x00\x00",
1253         };
1254         struct rte_flow_item_vlan vlan_spec;
1255         const struct rte_flow_item_vlan vlan_mask = {
1256                 .tci = RTE_BE16(0x0fff),
1257         };
1258         struct rte_flow_item pattern[] = {
1259                 {
1260                         .type = MLX4_FLOW_ITEM_TYPE_INTERNAL,
1261                 },
1262                 {
1263                         .type = RTE_FLOW_ITEM_TYPE_ETH,
1264                         .spec = &eth_spec,
1265                         .mask = &eth_mask,
1266                 },
1267                 {
1268                         /* Replaced with VLAN if filtering is enabled. */
1269                         .type = RTE_FLOW_ITEM_TYPE_END,
1270                 },
1271                 {
1272                         .type = RTE_FLOW_ITEM_TYPE_END,
1273                 },
1274         };
1275         /*
1276          * Round number of queues down to their previous power of 2 to
1277          * comply with RSS context limitations. Extra queues silently do not
1278          * get RSS by default.
1279          */
1280         uint32_t queues =
1281                 rte_align32pow2(priv->dev->data->nb_rx_queues + 1) >> 1;
1282         alignas(struct rte_flow_action_rss) uint8_t rss_conf_data
1283                 [offsetof(struct rte_flow_action_rss, queue) +
1284                  sizeof(((struct rte_flow_action_rss *)0)->queue[0]) * queues];
1285         struct rte_flow_action_rss *rss_conf = (void *)rss_conf_data;
1286         struct rte_flow_action actions[] = {
1287                 {
1288                         .type = RTE_FLOW_ACTION_TYPE_RSS,
1289                         .conf = rss_conf,
1290                 },
1291                 {
1292                         .type = RTE_FLOW_ACTION_TYPE_END,
1293                 },
1294         };
1295         struct ether_addr *rule_mac = &eth_spec.dst;
1296         rte_be16_t *rule_vlan =
1297                 priv->dev->data->dev_conf.rxmode.hw_vlan_filter &&
1298                 !priv->dev->data->promiscuous ?
1299                 &vlan_spec.tci :
1300                 NULL;
1301         int broadcast =
1302                 !priv->dev->data->promiscuous &&
1303                 !priv->dev->data->all_multicast;
1304         uint16_t vlan = 0;
1305         struct rte_flow *flow;
1306         unsigned int i;
1307         int err = 0;
1308
1309         /* Nothing to be done if there are no Rx queues. */
1310         if (!queues)
1311                 goto error;
1312         /* Prepare default RSS configuration. */
1313         *rss_conf = (struct rte_flow_action_rss){
1314                 .rss_conf = NULL, /* Rely on default fallback settings. */
1315                 .num = queues,
1316         };
1317         for (i = 0; i != queues; ++i)
1318                 rss_conf->queue[i] = i;
1319         /*
1320          * Set up VLAN item if filtering is enabled and at least one VLAN
1321          * filter is configured.
1322          */
1323         if (rule_vlan) {
1324                 vlan = mlx4_flow_internal_next_vlan(priv, 0);
1325                 if (vlan < 4096) {
1326                         pattern[2] = (struct rte_flow_item){
1327                                 .type = RTE_FLOW_ITEM_TYPE_VLAN,
1328                                 .spec = &vlan_spec,
1329                                 .mask = &vlan_mask,
1330                         };
1331 next_vlan:
1332                         *rule_vlan = rte_cpu_to_be_16(vlan);
1333                 } else {
1334                         rule_vlan = NULL;
1335                 }
1336         }
1337         for (i = 0; i != RTE_DIM(priv->mac) + broadcast; ++i) {
1338                 const struct ether_addr *mac;
1339
1340                 /* Broadcasts are handled by an extra iteration. */
1341                 if (i < RTE_DIM(priv->mac))
1342                         mac = &priv->mac[i];
1343                 else
1344                         mac = &eth_mask.dst;
1345                 if (is_zero_ether_addr(mac))
1346                         continue;
1347                 /* Check if MAC flow rule is already present. */
1348                 for (flow = LIST_FIRST(&priv->flows);
1349                      flow && flow->internal;
1350                      flow = LIST_NEXT(flow, next)) {
1351                         const struct ibv_flow_spec_eth *eth =
1352                                 (const void *)((uintptr_t)flow->ibv_attr +
1353                                                sizeof(*flow->ibv_attr));
1354                         unsigned int j;
1355
1356                         if (!flow->mac)
1357                                 continue;
1358                         assert(flow->ibv_attr->type == IBV_FLOW_ATTR_NORMAL);
1359                         assert(flow->ibv_attr->num_of_specs == 1);
1360                         assert(eth->type == IBV_FLOW_SPEC_ETH);
1361                         assert(flow->rss);
1362                         if (rule_vlan &&
1363                             (eth->val.vlan_tag != *rule_vlan ||
1364                              eth->mask.vlan_tag != RTE_BE16(0x0fff)))
1365                                 continue;
1366                         if (!rule_vlan && eth->mask.vlan_tag)
1367                                 continue;
1368                         for (j = 0; j != sizeof(mac->addr_bytes); ++j)
1369                                 if (eth->val.dst_mac[j] != mac->addr_bytes[j] ||
1370                                     eth->mask.dst_mac[j] != UINT8_C(0xff) ||
1371                                     eth->val.src_mac[j] != UINT8_C(0x00) ||
1372                                     eth->mask.src_mac[j] != UINT8_C(0x00))
1373                                         break;
1374                         if (j != sizeof(mac->addr_bytes))
1375                                 continue;
1376                         if (flow->rss->queues != queues ||
1377                             memcmp(flow->rss->queue_id, rss_conf->queue,
1378                                    queues * sizeof(flow->rss->queue_id[0])))
1379                                 continue;
1380                         break;
1381                 }
1382                 if (!flow || !flow->internal) {
1383                         /* Not found, create a new flow rule. */
1384                         memcpy(rule_mac, mac, sizeof(*mac));
1385                         flow = mlx4_flow_create(priv->dev, &attr, pattern,
1386                                                 actions, error);
1387                         if (!flow) {
1388                                 err = -rte_errno;
1389                                 goto error;
1390                         }
1391                 }
1392                 flow->select = 1;
1393                 flow->mac = 1;
1394         }
1395         if (rule_vlan) {
1396                 vlan = mlx4_flow_internal_next_vlan(priv, vlan + 1);
1397                 if (vlan < 4096)
1398                         goto next_vlan;
1399         }
1400         /* Take care of promiscuous and all multicast flow rules. */
1401         if (!broadcast) {
1402                 for (flow = LIST_FIRST(&priv->flows);
1403                      flow && flow->internal;
1404                      flow = LIST_NEXT(flow, next)) {
1405                         if (priv->dev->data->promiscuous) {
1406                                 if (flow->promisc)
1407                                         break;
1408                         } else {
1409                                 assert(priv->dev->data->all_multicast);
1410                                 if (flow->allmulti)
1411                                         break;
1412                         }
1413                 }
1414                 if (flow && flow->internal) {
1415                         assert(flow->rss);
1416                         if (flow->rss->queues != queues ||
1417                             memcmp(flow->rss->queue_id, rss_conf->queue,
1418                                    queues * sizeof(flow->rss->queue_id[0])))
1419                                 flow = NULL;
1420                 }
1421                 if (!flow || !flow->internal) {
1422                         /* Not found, create a new flow rule. */
1423                         if (priv->dev->data->promiscuous) {
1424                                 pattern[1].spec = NULL;
1425                                 pattern[1].mask = NULL;
1426                         } else {
1427                                 assert(priv->dev->data->all_multicast);
1428                                 pattern[1].spec = &eth_allmulti;
1429                                 pattern[1].mask = &eth_allmulti;
1430                         }
1431                         pattern[2] = pattern[3];
1432                         flow = mlx4_flow_create(priv->dev, &attr, pattern,
1433                                                 actions, error);
1434                         if (!flow) {
1435                                 err = -rte_errno;
1436                                 goto error;
1437                         }
1438                 }
1439                 assert(flow->promisc || flow->allmulti);
1440                 flow->select = 1;
1441         }
1442 error:
1443         /* Clear selection and clean up stale internal flow rules. */
1444         flow = LIST_FIRST(&priv->flows);
1445         while (flow && flow->internal) {
1446                 struct rte_flow *next = LIST_NEXT(flow, next);
1447
1448                 if (!flow->select)
1449                         claim_zero(mlx4_flow_destroy(priv->dev, flow, error));
1450                 else
1451                         flow->select = 0;
1452                 flow = next;
1453         }
1454         return err;
1455 }
1456
1457 /**
1458  * Synchronize flow rules.
1459  *
1460  * This function synchronizes flow rules with the state of the device by
1461  * taking into account isolated mode and whether target queues are
1462  * configured.
1463  *
1464  * @param priv
1465  *   Pointer to private structure.
1466  * @param[out] error
1467  *   Perform verbose error reporting if not NULL.
1468  *
1469  * @return
1470  *   0 on success, a negative errno value otherwise and rte_errno is set.
1471  */
1472 int
1473 mlx4_flow_sync(struct priv *priv, struct rte_flow_error *error)
1474 {
1475         struct rte_flow *flow;
1476         int ret;
1477
1478         /* Internal flow rules are guaranteed to come first in the list. */
1479         if (priv->isolated) {
1480                 /*
1481                  * Get rid of them in isolated mode, stop at the first
1482                  * non-internal rule found.
1483                  */
1484                 for (flow = LIST_FIRST(&priv->flows);
1485                      flow && flow->internal;
1486                      flow = LIST_FIRST(&priv->flows))
1487                         claim_zero(mlx4_flow_destroy(priv->dev, flow, error));
1488         } else {
1489                 /* Refresh internal rules. */
1490                 ret = mlx4_flow_internal(priv, error);
1491                 if (ret)
1492                         return ret;
1493         }
1494         /* Toggle the remaining flow rules . */
1495         LIST_FOREACH(flow, &priv->flows, next) {
1496                 ret = mlx4_flow_toggle(priv, flow, priv->started, error);
1497                 if (ret)
1498                         return ret;
1499         }
1500         if (!priv->started)
1501                 assert(!priv->drop);
1502         return 0;
1503 }
1504
1505 /**
1506  * Clean up all flow rules.
1507  *
1508  * Unlike mlx4_flow_flush(), this function takes care of all remaining flow
1509  * rules regardless of whether they are internal or user-configured.
1510  *
1511  * @param priv
1512  *   Pointer to private structure.
1513  */
1514 void
1515 mlx4_flow_clean(struct priv *priv)
1516 {
1517         struct rte_flow *flow;
1518
1519         while ((flow = LIST_FIRST(&priv->flows)))
1520                 mlx4_flow_destroy(priv->dev, flow, NULL);
1521         assert(LIST_EMPTY(&priv->rss));
1522 }
1523
1524 static const struct rte_flow_ops mlx4_flow_ops = {
1525         .validate = mlx4_flow_validate,
1526         .create = mlx4_flow_create,
1527         .destroy = mlx4_flow_destroy,
1528         .flush = mlx4_flow_flush,
1529         .isolate = mlx4_flow_isolate,
1530 };
1531
1532 /**
1533  * Manage filter operations.
1534  *
1535  * @param dev
1536  *   Pointer to Ethernet device structure.
1537  * @param filter_type
1538  *   Filter type.
1539  * @param filter_op
1540  *   Operation to perform.
1541  * @param arg
1542  *   Pointer to operation-specific structure.
1543  *
1544  * @return
1545  *   0 on success, negative errno value otherwise and rte_errno is set.
1546  */
1547 int
1548 mlx4_filter_ctrl(struct rte_eth_dev *dev,
1549                  enum rte_filter_type filter_type,
1550                  enum rte_filter_op filter_op,
1551                  void *arg)
1552 {
1553         switch (filter_type) {
1554         case RTE_ETH_FILTER_GENERIC:
1555                 if (filter_op != RTE_ETH_FILTER_GET)
1556                         break;
1557                 *(const void **)arg = &mlx4_flow_ops;
1558                 return 0;
1559         default:
1560                 ERROR("%p: filter type (%d) not supported",
1561                       (void *)dev, filter_type);
1562                 break;
1563         }
1564         rte_errno = ENOTSUP;
1565         return -rte_errno;
1566 }