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 priv
 112  *   Pointer to private structure.
 113  * @param rss_hf
 114  *   Hash fields in DPDK format (see struct rte_eth_rss_conf).
 115  *
 116  * @return
 117  *   A valid Verbs RSS hash fields mask for mlx4 on success, (uint64_t)-1
 118  *   otherwise and rte_errno is set.
 119  */
 120 uint64_t
 121 mlx4_conv_rss_hf(struct priv *priv, uint64_t rss_hf)
 122 {
 123         enum { IPV4, IPV6, TCP, UDP, };
 124         const uint64_t in[] = {
 125                 [IPV4] = (ETH_RSS_IPV4 |
 126                           ETH_RSS_FRAG_IPV4 |
 127                           ETH_RSS_NONFRAG_IPV4_TCP |
 128                           ETH_RSS_NONFRAG_IPV4_UDP |
 129                           ETH_RSS_NONFRAG_IPV4_OTHER),
 130                 [IPV6] = (ETH_RSS_IPV6 |
 131                           ETH_RSS_FRAG_IPV6 |
 132                           ETH_RSS_NONFRAG_IPV6_TCP |
 133                           ETH_RSS_NONFRAG_IPV6_UDP |
 134                           ETH_RSS_NONFRAG_IPV6_OTHER |
 135                           ETH_RSS_IPV6_EX |
 136                           ETH_RSS_IPV6_TCP_EX |
 137                           ETH_RSS_IPV6_UDP_EX),
 138                 [TCP] = (ETH_RSS_NONFRAG_IPV4_TCP |
 139                          ETH_RSS_NONFRAG_IPV6_TCP |
 140                          ETH_RSS_IPV6_TCP_EX),
 141                 [UDP] = (ETH_RSS_NONFRAG_IPV4_UDP |
 142                          ETH_RSS_NONFRAG_IPV6_UDP |
 143                          ETH_RSS_IPV6_UDP_EX),
 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 ((conv & priv->hw_rss_sup) == conv) {
 161                 if (rss_hf == (uint64_t)-1)
 162                         return conv;
 163                 if (!(rss_hf & ~seen))
 164                         return conv;
 165         }
 166         rte_errno = ENOTSUP;
 167         return (uint64_t)-1;
 168 }
 169
 170 /**
 171  * Merge Ethernet pattern item into flow rule handle.
 172  *
 173  * Additional mlx4-specific constraints on supported fields:
 174  *
 175  * - No support for partial masks, except in the specific case of matching
 176  *   all multicast traffic (@p spec->dst and @p mask->dst equal to
 177  *   01:00:00:00:00:00).
 178  * - Not providing @p item->spec or providing an empty @p mask->dst is
 179  *   *only* supported if the rule doesn't specify additional matching
 180  *   criteria (i.e. rule is promiscuous-like).
 181  *
 182  * @param[in, out] flow
 183  *   Flow rule handle to update.
 184  * @param[in] item
 185  *   Pattern item to merge.
 186  * @param[in] proc
 187  *   Associated item-processing object.
 188  * @param[out] error
 189  *   Perform verbose error reporting if not NULL.
 190  *
 191  * @return
 192  *   0 on success, a negative errno value otherwise and rte_errno is set.
 193  */
 194 static int
 195 mlx4_flow_merge_eth(struct rte_flow *flow,
 196                     const struct rte_flow_item *item,
 197                     const struct mlx4_flow_proc_item *proc,
 198                     struct rte_flow_error *error)
 199 {
 200         const struct rte_flow_item_eth *spec = item->spec;
 201         const struct rte_flow_item_eth *mask =
 202                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 203         struct ibv_flow_spec_eth *eth;
 204         const char *msg;
 205         unsigned int i;
 206
 207         if (!mask) {
 208                 flow->promisc = 1;
 209         } else {
 210                 uint32_t sum_dst = 0;
 211                 uint32_t sum_src = 0;
 212
 213                 for (i = 0; i != sizeof(mask->dst.addr_bytes); ++i) {
 214                         sum_dst += mask->dst.addr_bytes[i];
 215                         sum_src += mask->src.addr_bytes[i];
 216                 }
 217                 if (sum_src) {
 218                         msg = "mlx4 does not support source MAC matching";
 219                         goto error;
 220                 } else if (!sum_dst) {
 221                         flow->promisc = 1;
 222                 } else if (sum_dst == 1 && mask->dst.addr_bytes[0] == 1) {
 223                         if (!(spec->dst.addr_bytes[0] & 1)) {
 224                                 msg = "mlx4 does not support the explicit"
 225                                         " exclusion of all multicast traffic";
 226                                 goto error;
 227                         }
 228                         flow->allmulti = 1;
 229                 } else if (sum_dst != (UINT8_C(0xff) * ETHER_ADDR_LEN)) {
 230                         msg = "mlx4 does not support matching partial"
 231                                 " Ethernet fields";
 232                         goto error;
 233                 }
 234         }
 235         if (!flow->ibv_attr)
 236                 return 0;
 237         if (flow->promisc) {
 238                 flow->ibv_attr->type = IBV_FLOW_ATTR_ALL_DEFAULT;
 239                 return 0;
 240         }
 241         if (flow->allmulti) {
 242                 flow->ibv_attr->type = IBV_FLOW_ATTR_MC_DEFAULT;
 243                 return 0;
 244         }
 245         ++flow->ibv_attr->num_of_specs;
 246         eth = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 247         *eth = (struct ibv_flow_spec_eth) {
 248                 .type = IBV_FLOW_SPEC_ETH,
 249                 .size = sizeof(*eth),
 250         };
 251         memcpy(eth->val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
 252         memcpy(eth->mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
 253         /* Remove unwanted bits from values. */
 254         for (i = 0; i < ETHER_ADDR_LEN; ++i) {
 255                 eth->val.dst_mac[i] &= eth->mask.dst_mac[i];
 256         }
 257         return 0;
 258 error:
 259         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 260                                   item, msg);
 261 }
 262
 263 /**
 264  * Merge VLAN pattern item into flow rule handle.
 265  *
 266  * Additional mlx4-specific constraints on supported fields:
 267  *
 268  * - Matching *all* VLAN traffic by omitting @p item->spec or providing an
 269  *   empty @p item->mask would also include non-VLAN traffic. Doing so is
 270  *   therefore unsupported.
 271  * - No support for partial masks.
 272  *
 273  * @param[in, out] flow
 274  *   Flow rule handle to update.
 275  * @param[in] item
 276  *   Pattern item to merge.
 277  * @param[in] proc
 278  *   Associated item-processing object.
 279  * @param[out] error
 280  *   Perform verbose error reporting if not NULL.
 281  *
 282  * @return
 283  *   0 on success, a negative errno value otherwise and rte_errno is set.
 284  */
 285 static int
 286 mlx4_flow_merge_vlan(struct rte_flow *flow,
 287                      const struct rte_flow_item *item,
 288                      const struct mlx4_flow_proc_item *proc,
 289                      struct rte_flow_error *error)
 290 {
 291         const struct rte_flow_item_vlan *spec = item->spec;
 292         const struct rte_flow_item_vlan *mask =
 293                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 294         struct ibv_flow_spec_eth *eth;
 295         const char *msg;
 296
 297         if (!mask || !mask->tci) {
 298                 msg = "mlx4 cannot match all VLAN traffic while excluding"
 299                         " non-VLAN traffic, TCI VID must be specified";
 300                 goto error;
 301         }
 302         if (mask->tci != RTE_BE16(0x0fff)) {
 303                 msg = "mlx4 does not support partial TCI VID matching";
 304                 goto error;
 305         }
 306         if (!flow->ibv_attr)
 307                 return 0;
 308         eth = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size -
 309                        sizeof(*eth));
 310         eth->val.vlan_tag = spec->tci;
 311         eth->mask.vlan_tag = mask->tci;
 312         eth->val.vlan_tag &= eth->mask.vlan_tag;
 313         return 0;
 314 error:
 315         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 316                                   item, msg);
 317 }
 318
 319 /**
 320  * Merge IPv4 pattern item into flow rule handle.
 321  *
 322  * Additional mlx4-specific constraints on supported fields:
 323  *
 324  * - No support for partial masks.
 325  *
 326  * @param[in, out] flow
 327  *   Flow rule handle to update.
 328  * @param[in] item
 329  *   Pattern item to merge.
 330  * @param[in] proc
 331  *   Associated item-processing object.
 332  * @param[out] error
 333  *   Perform verbose error reporting if not NULL.
 334  *
 335  * @return
 336  *   0 on success, a negative errno value otherwise and rte_errno is set.
 337  */
 338 static int
 339 mlx4_flow_merge_ipv4(struct rte_flow *flow,
 340                      const struct rte_flow_item *item,
 341                      const struct mlx4_flow_proc_item *proc,
 342                      struct rte_flow_error *error)
 343 {
 344         const struct rte_flow_item_ipv4 *spec = item->spec;
 345         const struct rte_flow_item_ipv4 *mask =
 346                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 347         struct ibv_flow_spec_ipv4 *ipv4;
 348         const char *msg;
 349
 350         if (mask &&
 351             ((uint32_t)(mask->hdr.src_addr + 1) > UINT32_C(1) ||
 352              (uint32_t)(mask->hdr.dst_addr + 1) > UINT32_C(1))) {
 353                 msg = "mlx4 does not support matching partial IPv4 fields";
 354                 goto error;
 355         }
 356         if (!flow->ibv_attr)
 357                 return 0;
 358         ++flow->ibv_attr->num_of_specs;
 359         ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 360         *ipv4 = (struct ibv_flow_spec_ipv4) {
 361                 .type = IBV_FLOW_SPEC_IPV4,
 362                 .size = sizeof(*ipv4),
 363         };
 364         if (!spec)
 365                 return 0;
 366         ipv4->val = (struct ibv_flow_ipv4_filter) {
 367                 .src_ip = spec->hdr.src_addr,
 368                 .dst_ip = spec->hdr.dst_addr,
 369         };
 370         ipv4->mask = (struct ibv_flow_ipv4_filter) {
 371                 .src_ip = mask->hdr.src_addr,
 372                 .dst_ip = mask->hdr.dst_addr,
 373         };
 374         /* Remove unwanted bits from values. */
 375         ipv4->val.src_ip &= ipv4->mask.src_ip;
 376         ipv4->val.dst_ip &= ipv4->mask.dst_ip;
 377         return 0;
 378 error:
 379         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 380                                   item, msg);
 381 }
 382
 383 /**
 384  * Merge UDP pattern item into flow rule handle.
 385  *
 386  * Additional mlx4-specific constraints on supported fields:
 387  *
 388  * - No support for partial masks.
 389  *
 390  * @param[in, out] flow
 391  *   Flow rule handle to update.
 392  * @param[in] item
 393  *   Pattern item to merge.
 394  * @param[in] proc
 395  *   Associated item-processing object.
 396  * @param[out] error
 397  *   Perform verbose error reporting if not NULL.
 398  *
 399  * @return
 400  *   0 on success, a negative errno value otherwise and rte_errno is set.
 401  */
 402 static int
 403 mlx4_flow_merge_udp(struct rte_flow *flow,
 404                     const struct rte_flow_item *item,
 405                     const struct mlx4_flow_proc_item *proc,
 406                     struct rte_flow_error *error)
 407 {
 408         const struct rte_flow_item_udp *spec = item->spec;
 409         const struct rte_flow_item_udp *mask =
 410                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 411         struct ibv_flow_spec_tcp_udp *udp;
 412         const char *msg;
 413
 414         if (mask &&
 415             ((uint16_t)(mask->hdr.src_port + 1) > UINT16_C(1) ||
 416              (uint16_t)(mask->hdr.dst_port + 1) > UINT16_C(1))) {
 417                 msg = "mlx4 does not support matching partial UDP fields";
 418                 goto error;
 419         }
 420         if (!flow->ibv_attr)
 421                 return 0;
 422         ++flow->ibv_attr->num_of_specs;
 423         udp = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 424         *udp = (struct ibv_flow_spec_tcp_udp) {
 425                 .type = IBV_FLOW_SPEC_UDP,
 426                 .size = sizeof(*udp),
 427         };
 428         if (!spec)
 429                 return 0;
 430         udp->val.dst_port = spec->hdr.dst_port;
 431         udp->val.src_port = spec->hdr.src_port;
 432         udp->mask.dst_port = mask->hdr.dst_port;
 433         udp->mask.src_port = mask->hdr.src_port;
 434         /* Remove unwanted bits from values. */
 435         udp->val.src_port &= udp->mask.src_port;
 436         udp->val.dst_port &= udp->mask.dst_port;
 437         return 0;
 438 error:
 439         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 440                                   item, msg);
 441 }
 442
 443 /**
 444  * Merge TCP pattern item into flow rule handle.
 445  *
 446  * Additional mlx4-specific constraints on supported fields:
 447  *
 448  * - No support for partial masks.
 449  *
 450  * @param[in, out] flow
 451  *   Flow rule handle to update.
 452  * @param[in] item
 453  *   Pattern item to merge.
 454  * @param[in] proc
 455  *   Associated item-processing object.
 456  * @param[out] error
 457  *   Perform verbose error reporting if not NULL.
 458  *
 459  * @return
 460  *   0 on success, a negative errno value otherwise and rte_errno is set.
 461  */
 462 static int
 463 mlx4_flow_merge_tcp(struct rte_flow *flow,
 464                     const struct rte_flow_item *item,
 465                     const struct mlx4_flow_proc_item *proc,
 466                     struct rte_flow_error *error)
 467 {
 468         const struct rte_flow_item_tcp *spec = item->spec;
 469         const struct rte_flow_item_tcp *mask =
 470                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 471         struct ibv_flow_spec_tcp_udp *tcp;
 472         const char *msg;
 473
 474         if (mask &&
 475             ((uint16_t)(mask->hdr.src_port + 1) > UINT16_C(1) ||
 476              (uint16_t)(mask->hdr.dst_port + 1) > UINT16_C(1))) {
 477                 msg = "mlx4 does not support matching partial TCP fields";
 478                 goto error;
 479         }
 480         if (!flow->ibv_attr)
 481                 return 0;
 482         ++flow->ibv_attr->num_of_specs;
 483         tcp = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 484         *tcp = (struct ibv_flow_spec_tcp_udp) {
 485                 .type = IBV_FLOW_SPEC_TCP,
 486                 .size = sizeof(*tcp),
 487         };
 488         if (!spec)
 489                 return 0;
 490         tcp->val.dst_port = spec->hdr.dst_port;
 491         tcp->val.src_port = spec->hdr.src_port;
 492         tcp->mask.dst_port = mask->hdr.dst_port;
 493         tcp->mask.src_port = mask->hdr.src_port;
 494         /* Remove unwanted bits from values. */
 495         tcp->val.src_port &= tcp->mask.src_port;
 496         tcp->val.dst_port &= tcp->mask.dst_port;
 497         return 0;
 498 error:
 499         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 500                                   item, msg);
 501 }
 502
 503 /**
 504  * Perform basic sanity checks on a pattern item.
 505  *
 506  * @param[in] item
 507  *   Item specification.
 508  * @param[in] proc
 509  *   Associated item-processing object.
 510  * @param[out] error
 511  *   Perform verbose error reporting if not NULL.
 512  *
 513  * @return
 514  *   0 on success, a negative errno value otherwise and rte_errno is set.
 515  */
 516 static int
 517 mlx4_flow_item_check(const struct rte_flow_item *item,
 518                      const struct mlx4_flow_proc_item *proc,
 519                      struct rte_flow_error *error)
 520 {
 521         const uint8_t *mask;
 522         unsigned int i;
 523
 524         /* item->last and item->mask cannot exist without item->spec. */
 525         if (!item->spec && (item->mask || item->last))
 526                 return rte_flow_error_set
 527                         (error, EINVAL, RTE_FLOW_ERROR_TYPE_ITEM, item,
 528                          "\"mask\" or \"last\" field provided without a"
 529                          " corresponding \"spec\"");
 530         /* No spec, no mask, no problem. */
 531         if (!item->spec)
 532                 return 0;
 533         mask = item->mask ?
 534                 (const uint8_t *)item->mask :
 535                 (const uint8_t *)proc->mask_default;
 536         assert(mask);
 537         /*
 538          * Single-pass check to make sure that:
 539          * - Mask is supported, no bits are set outside proc->mask_support.
 540          * - Both item->spec and item->last are included in mask.
 541          */
 542         for (i = 0; i != proc->mask_sz; ++i) {
 543                 if (!mask[i])
 544                         continue;
 545                 if ((mask[i] | ((const uint8_t *)proc->mask_support)[i]) !=
 546                     ((const uint8_t *)proc->mask_support)[i])
 547                         return rte_flow_error_set
 548                                 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 549                                  item, "unsupported field found in \"mask\"");
 550                 if (item->last &&
 551                     (((const uint8_t *)item->spec)[i] & mask[i]) !=
 552                     (((const uint8_t *)item->last)[i] & mask[i]))
 553                         return rte_flow_error_set
 554                                 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 555                                  item,
 556                                  "range between \"spec\" and \"last\""
 557                                  " is larger than \"mask\"");
 558         }
 559         return 0;
 560 }
 561
 562 /** Graph of supported items and associated actions. */
 563 static const struct mlx4_flow_proc_item mlx4_flow_proc_item_list[] = {
 564         [RTE_FLOW_ITEM_TYPE_END] = {
 565                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_ETH),
 566         },
 567         [RTE_FLOW_ITEM_TYPE_ETH] = {
 568                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_VLAN,
 569                                        RTE_FLOW_ITEM_TYPE_IPV4),
 570                 .mask_support = &(const struct rte_flow_item_eth){
 571                         /* Only destination MAC can be matched. */
 572                         .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 573                 },
 574                 .mask_default = &rte_flow_item_eth_mask,
 575                 .mask_sz = sizeof(struct rte_flow_item_eth),
 576                 .merge = mlx4_flow_merge_eth,
 577                 .dst_sz = sizeof(struct ibv_flow_spec_eth),
 578         },
 579         [RTE_FLOW_ITEM_TYPE_VLAN] = {
 580                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_IPV4),
 581                 .mask_support = &(const struct rte_flow_item_vlan){
 582                         /* Only TCI VID matching is supported. */
 583                         .tci = RTE_BE16(0x0fff),
 584                 },
 585                 .mask_default = &rte_flow_item_vlan_mask,
 586                 .mask_sz = sizeof(struct rte_flow_item_vlan),
 587                 .merge = mlx4_flow_merge_vlan,
 588                 .dst_sz = 0,
 589         },
 590         [RTE_FLOW_ITEM_TYPE_IPV4] = {
 591                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_UDP,
 592                                        RTE_FLOW_ITEM_TYPE_TCP),
 593                 .mask_support = &(const struct rte_flow_item_ipv4){
 594                         .hdr = {
 595                                 .src_addr = RTE_BE32(0xffffffff),
 596                                 .dst_addr = RTE_BE32(0xffffffff),
 597                         },
 598                 },
 599                 .mask_default = &rte_flow_item_ipv4_mask,
 600                 .mask_sz = sizeof(struct rte_flow_item_ipv4),
 601                 .merge = mlx4_flow_merge_ipv4,
 602                 .dst_sz = sizeof(struct ibv_flow_spec_ipv4),
 603         },
 604         [RTE_FLOW_ITEM_TYPE_UDP] = {
 605                 .mask_support = &(const struct rte_flow_item_udp){
 606                         .hdr = {
 607                                 .src_port = RTE_BE16(0xffff),
 608                                 .dst_port = RTE_BE16(0xffff),
 609                         },
 610                 },
 611                 .mask_default = &rte_flow_item_udp_mask,
 612                 .mask_sz = sizeof(struct rte_flow_item_udp),
 613                 .merge = mlx4_flow_merge_udp,
 614                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 615         },
 616         [RTE_FLOW_ITEM_TYPE_TCP] = {
 617                 .mask_support = &(const struct rte_flow_item_tcp){
 618                         .hdr = {
 619                                 .src_port = RTE_BE16(0xffff),
 620                                 .dst_port = RTE_BE16(0xffff),
 621                         },
 622                 },
 623                 .mask_default = &rte_flow_item_tcp_mask,
 624                 .mask_sz = sizeof(struct rte_flow_item_tcp),
 625                 .merge = mlx4_flow_merge_tcp,
 626                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 627         },
 628 };
 629
 630 /**
 631  * Make sure a flow rule is supported and initialize associated structure.
 632  *
 633  * @param priv
 634  *   Pointer to private structure.
 635  * @param[in] attr
 636  *   Flow rule attributes.
 637  * @param[in] pattern
 638  *   Pattern specification (list terminated by the END pattern item).
 639  * @param[in] actions
 640  *   Associated actions (list terminated by the END action).
 641  * @param[out] error
 642  *   Perform verbose error reporting if not NULL.
 643  * @param[in, out] addr
 644  *   Buffer where the resulting flow rule handle pointer must be stored.
 645  *   If NULL, stop processing after validation stage.
 646  *
 647  * @return
 648  *   0 on success, a negative errno value otherwise and rte_errno is set.
 649  */
 650 static int
 651 mlx4_flow_prepare(struct priv *priv,
 652                   const struct rte_flow_attr *attr,
 653                   const struct rte_flow_item pattern[],
 654                   const struct rte_flow_action actions[],
 655                   struct rte_flow_error *error,
 656                   struct rte_flow **addr)
 657 {
 658         const struct rte_flow_item *item;
 659         const struct rte_flow_action *action;
 660         const struct mlx4_flow_proc_item *proc;
 661         struct rte_flow temp = { .ibv_attr_size = sizeof(*temp.ibv_attr) };
 662         struct rte_flow *flow = &temp;
 663         const char *msg = NULL;
 664
 665         if (attr->group)
 666                 return rte_flow_error_set
 667                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 668                          NULL, "groups are not supported");
 669         if (attr->priority > MLX4_FLOW_PRIORITY_LAST)
 670                 return rte_flow_error_set
 671                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 672                          NULL, "maximum priority level is "
 673                          MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST));
 674         if (attr->egress)
 675                 return rte_flow_error_set
 676                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 677                          NULL, "egress is not supported");
 678         if (!attr->ingress)
 679                 return rte_flow_error_set
 680                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 681                          NULL, "only ingress is supported");
 682 fill:
 683         proc = mlx4_flow_proc_item_list;
 684         /* Go over pattern. */
 685         for (item = pattern; item->type; ++item) {
 686                 const struct mlx4_flow_proc_item *next = NULL;
 687                 unsigned int i;
 688                 int err;
 689
 690                 if (item->type == RTE_FLOW_ITEM_TYPE_VOID)
 691                         continue;
 692                 if (item->type == MLX4_FLOW_ITEM_TYPE_INTERNAL) {
 693                         flow->internal = 1;
 694                         continue;
 695                 }
 696                 if (flow->promisc || flow->allmulti) {
 697                         msg = "mlx4 does not support additional matching"
 698                                 " criteria combined with indiscriminate"
 699                                 " matching on Ethernet headers";
 700                         goto exit_item_not_supported;
 701                 }
 702                 for (i = 0; proc->next_item && proc->next_item[i]; ++i) {
 703                         if (proc->next_item[i] == item->type) {
 704                                 next = &mlx4_flow_proc_item_list[item->type];
 705                                 break;
 706                         }
 707                 }
 708                 if (!next)
 709                         goto exit_item_not_supported;
 710                 proc = next;
 711                 /*
 712                  * Perform basic sanity checks only once, while handle is
 713                  * not allocated.
 714                  */
 715                 if (flow == &temp) {
 716                         err = mlx4_flow_item_check(item, proc, error);
 717                         if (err)
 718                                 return err;
 719                 }
 720                 if (proc->merge) {
 721                         err = proc->merge(flow, item, proc, error);
 722                         if (err)
 723                                 return err;
 724                 }
 725                 flow->ibv_attr_size += proc->dst_sz;
 726         }
 727         /* Go over actions list. */
 728         for (action = actions; action->type; ++action) {
 729                 switch (action->type) {
 730                         const struct rte_flow_action_queue *queue;
 731                         const struct rte_flow_action_rss *rss;
 732                         const struct rte_eth_rss_conf *rss_conf;
 733                         unsigned int i;
 734
 735                 case RTE_FLOW_ACTION_TYPE_VOID:
 736                         continue;
 737                 case RTE_FLOW_ACTION_TYPE_DROP:
 738                         flow->drop = 1;
 739                         break;
 740                 case RTE_FLOW_ACTION_TYPE_QUEUE:
 741                         if (flow->rss)
 742                                 break;
 743                         queue = action->conf;
 744                         if (queue->index >= priv->dev->data->nb_rx_queues) {
 745                                 msg = "queue target index beyond number of"
 746                                         " configured Rx queues";
 747                                 goto exit_action_not_supported;
 748                         }
 749                         flow->rss = mlx4_rss_get
 750                                 (priv, 0, mlx4_rss_hash_key_default, 1,
 751                                  &queue->index);
 752                         if (!flow->rss) {
 753                                 msg = "not enough resources for additional"
 754                                         " single-queue RSS context";
 755                                 goto exit_action_not_supported;
 756                         }
 757                         break;
 758                 case RTE_FLOW_ACTION_TYPE_RSS:
 759                         if (flow->rss)
 760                                 break;
 761                         rss = action->conf;
 762                         /* Default RSS configuration if none is provided. */
 763                         rss_conf =
 764                                 rss->rss_conf ?
 765                                 rss->rss_conf :
 766                                 &(struct rte_eth_rss_conf){
 767                                         .rss_key = mlx4_rss_hash_key_default,
 768                                         .rss_key_len = MLX4_RSS_HASH_KEY_SIZE,
 769                                         .rss_hf = -1,
 770                                 };
 771                         /* Sanity checks. */
 772                         for (i = 0; i < rss->num; ++i)
 773                                 if (rss->queue[i] >=
 774                                     priv->dev->data->nb_rx_queues)
 775                                         break;
 776                         if (i != rss->num) {
 777                                 msg = "queue index target beyond number of"
 778                                         " configured Rx queues";
 779                                 goto exit_action_not_supported;
 780                         }
 781                         if (!rte_is_power_of_2(rss->num)) {
 782                                 msg = "for RSS, mlx4 requires the number of"
 783                                         " queues to be a power of two";
 784                                 goto exit_action_not_supported;
 785                         }
 786                         if (rss_conf->rss_key_len !=
 787                             sizeof(flow->rss->key)) {
 788                                 msg = "mlx4 supports exactly one RSS hash key"
 789                                         " length: "
 790                                         MLX4_STR_EXPAND(MLX4_RSS_HASH_KEY_SIZE);
 791                                 goto exit_action_not_supported;
 792                         }
 793                         for (i = 1; i < rss->num; ++i)
 794                                 if (rss->queue[i] - rss->queue[i - 1] != 1)
 795                                         break;
 796                         if (i != rss->num) {
 797                                 msg = "mlx4 requires RSS contexts to use"
 798                                         " consecutive queue indices only";
 799                                 goto exit_action_not_supported;
 800                         }
 801                         if (rss->queue[0] % rss->num) {
 802                                 msg = "mlx4 requires the first queue of a RSS"
 803                                         " context to be aligned on a multiple"
 804                                         " of the context size";
 805                                 goto exit_action_not_supported;
 806                         }
 807                         flow->rss = mlx4_rss_get
 808                                 (priv,
 809                                  mlx4_conv_rss_hf(priv, rss_conf->rss_hf),
 810                                  rss_conf->rss_key, rss->num, rss->queue);
 811                         if (!flow->rss) {
 812                                 msg = "either invalid parameters or not enough"
 813                                         " resources for additional multi-queue"
 814                                         " RSS context";
 815                                 goto exit_action_not_supported;
 816                         }
 817                         break;
 818                 default:
 819                         goto exit_action_not_supported;
 820                 }
 821         }
 822         if (!flow->rss && !flow->drop)
 823                 return rte_flow_error_set
 824                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 825                          NULL, "no valid action");
 826         /* Validation ends here. */
 827         if (!addr) {
 828                 if (flow->rss)
 829                         mlx4_rss_put(flow->rss);
 830                 return 0;
 831         }
 832         if (flow == &temp) {
 833                 /* Allocate proper handle based on collected data. */
 834                 const struct mlx4_malloc_vec vec[] = {
 835                         {
 836                                 .align = alignof(struct rte_flow),
 837                                 .size = sizeof(*flow),
 838                                 .addr = (void **)&flow,
 839                         },
 840                         {
 841                                 .align = alignof(struct ibv_flow_attr),
 842                                 .size = temp.ibv_attr_size,
 843                                 .addr = (void **)&temp.ibv_attr,
 844                         },
 845                 };
 846
 847                 if (!mlx4_zmallocv(__func__, vec, RTE_DIM(vec)))
 848                         return rte_flow_error_set
 849                                 (error, -rte_errno,
 850                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
 851                                  "flow rule handle allocation failure");
 852                 /* Most fields will be updated by second pass. */
 853                 *flow = (struct rte_flow){
 854                         .ibv_attr = temp.ibv_attr,
 855                         .ibv_attr_size = sizeof(*flow->ibv_attr),
 856                         .rss = temp.rss,
 857                 };
 858                 *flow->ibv_attr = (struct ibv_flow_attr){
 859                         .type = IBV_FLOW_ATTR_NORMAL,
 860                         .size = sizeof(*flow->ibv_attr),
 861                         .priority = attr->priority,
 862                         .port = priv->port,
 863                 };
 864                 goto fill;
 865         }
 866         *addr = flow;
 867         return 0;
 868 exit_item_not_supported:
 869         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 870                                   item, msg ? msg : "item not supported");
 871 exit_action_not_supported:
 872         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 873                                   action, msg ? msg : "action not supported");
 874 }
 875
 876 /**
 877  * Validate a flow supported by the NIC.
 878  *
 879  * @see rte_flow_validate()
 880  * @see rte_flow_ops
 881  */
 882 static int
 883 mlx4_flow_validate(struct rte_eth_dev *dev,
 884                    const struct rte_flow_attr *attr,
 885                    const struct rte_flow_item pattern[],
 886                    const struct rte_flow_action actions[],
 887                    struct rte_flow_error *error)
 888 {
 889         struct priv *priv = dev->data->dev_private;
 890
 891         return mlx4_flow_prepare(priv, attr, pattern, actions, error, NULL);
 892 }
 893
 894 /**
 895  * Get a drop flow rule resources instance.
 896  *
 897  * @param priv
 898  *   Pointer to private structure.
 899  *
 900  * @return
 901  *   Pointer to drop flow resources on success, NULL otherwise and rte_errno
 902  *   is set.
 903  */
 904 static struct mlx4_drop *
 905 mlx4_drop_get(struct priv *priv)
 906 {
 907         struct mlx4_drop *drop = priv->drop;
 908
 909         if (drop) {
 910                 assert(drop->refcnt);
 911                 assert(drop->priv == priv);
 912                 ++drop->refcnt;
 913                 return drop;
 914         }
 915         drop = rte_malloc(__func__, sizeof(*drop), 0);
 916         if (!drop)
 917                 goto error;
 918         *drop = (struct mlx4_drop){
 919                 .priv = priv,
 920                 .refcnt = 1,
 921         };
 922         drop->cq = ibv_create_cq(priv->ctx, 1, NULL, NULL, 0);
 923         if (!drop->cq)
 924                 goto error;
 925         drop->qp = ibv_create_qp(priv->pd,
 926                                  &(struct ibv_qp_init_attr){
 927                                         .send_cq = drop->cq,
 928                                         .recv_cq = drop->cq,
 929                                         .qp_type = IBV_QPT_RAW_PACKET,
 930                                  });
 931         if (!drop->qp)
 932                 goto error;
 933         priv->drop = drop;
 934         return drop;
 935 error:
 936         if (drop->qp)
 937                 claim_zero(ibv_destroy_qp(drop->qp));
 938         if (drop->cq)
 939                 claim_zero(ibv_destroy_cq(drop->cq));
 940         if (drop)
 941                 rte_free(drop);
 942         rte_errno = ENOMEM;
 943         return NULL;
 944 }
 945
 946 /**
 947  * Give back a drop flow rule resources instance.
 948  *
 949  * @param drop
 950  *   Pointer to drop flow rule resources.
 951  */
 952 static void
 953 mlx4_drop_put(struct mlx4_drop *drop)
 954 {
 955         assert(drop->refcnt);
 956         if (--drop->refcnt)
 957                 return;
 958         drop->priv->drop = NULL;
 959         claim_zero(ibv_destroy_qp(drop->qp));
 960         claim_zero(ibv_destroy_cq(drop->cq));
 961         rte_free(drop);
 962 }
 963
 964 /**
 965  * Toggle a configured flow rule.
 966  *
 967  * @param priv
 968  *   Pointer to private structure.
 969  * @param flow
 970  *   Flow rule handle to toggle.
 971  * @param enable
 972  *   Whether associated Verbs flow must be created or removed.
 973  * @param[out] error
 974  *   Perform verbose error reporting if not NULL.
 975  *
 976  * @return
 977  *   0 on success, a negative errno value otherwise and rte_errno is set.
 978  */
 979 static int
 980 mlx4_flow_toggle(struct priv *priv,
 981                  struct rte_flow *flow,
 982                  int enable,
 983                  struct rte_flow_error *error)
 984 {
 985         struct ibv_qp *qp = NULL;
 986         const char *msg;
 987         int err;
 988
 989         if (!enable) {
 990                 if (!flow->ibv_flow)
 991                         return 0;
 992                 claim_zero(ibv_destroy_flow(flow->ibv_flow));
 993                 flow->ibv_flow = NULL;
 994                 if (flow->drop)
 995                         mlx4_drop_put(priv->drop);
 996                 else if (flow->rss)
 997                         mlx4_rss_detach(flow->rss);
 998                 return 0;
 999         }
1000         assert(flow->ibv_attr);
1001         if (!flow->internal &&
1002             !priv->isolated &&
1003             flow->ibv_attr->priority == MLX4_FLOW_PRIORITY_LAST) {
1004                 if (flow->ibv_flow) {
1005                         claim_zero(ibv_destroy_flow(flow->ibv_flow));
1006                         flow->ibv_flow = NULL;
1007                         if (flow->drop)
1008                                 mlx4_drop_put(priv->drop);
1009                         else if (flow->rss)
1010                                 mlx4_rss_detach(flow->rss);
1011                 }
1012                 err = EACCES;
1013                 msg = ("priority level "
1014                        MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST)
1015                        " is reserved when not in isolated mode");
1016                 goto error;
1017         }
1018         if (flow->rss) {
1019                 struct mlx4_rss *rss = flow->rss;
1020                 int missing = 0;
1021                 unsigned int i;
1022
1023                 /* Stop at the first nonexistent target queue. */
1024                 for (i = 0; i != rss->queues; ++i)
1025                         if (rss->queue_id[i] >=
1026                             priv->dev->data->nb_rx_queues ||
1027                             !priv->dev->data->rx_queues[rss->queue_id[i]]) {
1028                                 missing = 1;
1029                                 break;
1030                         }
1031                 if (flow->ibv_flow) {
1032                         if (missing ^ !flow->drop)
1033                                 return 0;
1034                         /* Verbs flow needs updating. */
1035                         claim_zero(ibv_destroy_flow(flow->ibv_flow));
1036                         flow->ibv_flow = NULL;
1037                         if (flow->drop)
1038                                 mlx4_drop_put(priv->drop);
1039                         else
1040                                 mlx4_rss_detach(rss);
1041                 }
1042                 if (!missing) {
1043                         err = mlx4_rss_attach(rss);
1044                         if (err) {
1045                                 err = -err;
1046                                 msg = "cannot create indirection table or hash"
1047                                         " QP to associate flow rule with";
1048                                 goto error;
1049                         }
1050                         qp = rss->qp;
1051                 }
1052                 /* A missing target queue drops traffic implicitly. */
1053                 flow->drop = missing;
1054         }
1055         if (flow->drop) {
1056                 mlx4_drop_get(priv);
1057                 if (!priv->drop) {
1058                         err = rte_errno;
1059                         msg = "resources for drop flow rule cannot be created";
1060                         goto error;
1061                 }
1062                 qp = priv->drop->qp;
1063         }
1064         assert(qp);
1065         if (flow->ibv_flow)
1066                 return 0;
1067         flow->ibv_flow = ibv_create_flow(qp, flow->ibv_attr);
1068         if (flow->ibv_flow)
1069                 return 0;
1070         if (flow->drop)
1071                 mlx4_drop_put(priv->drop);
1072         else if (flow->rss)
1073                 mlx4_rss_detach(flow->rss);
1074         err = errno;
1075         msg = "flow rule rejected by device";
1076 error:
1077         return rte_flow_error_set
1078                 (error, err, RTE_FLOW_ERROR_TYPE_HANDLE, flow, msg);
1079 }
1080
1081 /**
1082  * Create a flow.
1083  *
1084  * @see rte_flow_create()
1085  * @see rte_flow_ops
1086  */
1087 static struct rte_flow *
1088 mlx4_flow_create(struct rte_eth_dev *dev,
1089                  const struct rte_flow_attr *attr,
1090                  const struct rte_flow_item pattern[],
1091                  const struct rte_flow_action actions[],
1092                  struct rte_flow_error *error)
1093 {
1094         struct priv *priv = dev->data->dev_private;
1095         struct rte_flow *flow;
1096         int err;
1097
1098         err = mlx4_flow_prepare(priv, attr, pattern, actions, error, &flow);
1099         if (err)
1100                 return NULL;
1101         err = mlx4_flow_toggle(priv, flow, priv->started, error);
1102         if (!err) {
1103                 struct rte_flow *curr = LIST_FIRST(&priv->flows);
1104
1105                 /* New rules are inserted after internal ones. */
1106                 if (!curr || !curr->internal) {
1107                         LIST_INSERT_HEAD(&priv->flows, flow, next);
1108                 } else {
1109                         while (LIST_NEXT(curr, next) &&
1110                                LIST_NEXT(curr, next)->internal)
1111                                 curr = LIST_NEXT(curr, next);
1112                         LIST_INSERT_AFTER(curr, flow, next);
1113                 }
1114                 return flow;
1115         }
1116         if (flow->rss)
1117                 mlx4_rss_put(flow->rss);
1118         rte_flow_error_set(error, -err, RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1119                            error->message);
1120         rte_free(flow);
1121         return NULL;
1122 }
1123
1124 /**
1125  * Configure isolated mode.
1126  *
1127  * @see rte_flow_isolate()
1128  * @see rte_flow_ops
1129  */
1130 static int
1131 mlx4_flow_isolate(struct rte_eth_dev *dev,
1132                   int enable,
1133                   struct rte_flow_error *error)
1134 {
1135         struct priv *priv = dev->data->dev_private;
1136
1137         if (!!enable == !!priv->isolated)
1138                 return 0;
1139         priv->isolated = !!enable;
1140         if (mlx4_flow_sync(priv, error)) {
1141                 priv->isolated = !enable;
1142                 return -rte_errno;
1143         }
1144         return 0;
1145 }
1146
1147 /**
1148  * Destroy a flow rule.
1149  *
1150  * @see rte_flow_destroy()
1151  * @see rte_flow_ops
1152  */
1153 static int
1154 mlx4_flow_destroy(struct rte_eth_dev *dev,
1155                   struct rte_flow *flow,
1156                   struct rte_flow_error *error)
1157 {
1158         struct priv *priv = dev->data->dev_private;
1159         int err = mlx4_flow_toggle(priv, flow, 0, error);
1160
1161         if (err)
1162                 return err;
1163         LIST_REMOVE(flow, next);
1164         if (flow->rss)
1165                 mlx4_rss_put(flow->rss);
1166         rte_free(flow);
1167         return 0;
1168 }
1169
1170 /**
1171  * Destroy user-configured flow rules.
1172  *
1173  * This function skips internal flows rules.
1174  *
1175  * @see rte_flow_flush()
1176  * @see rte_flow_ops
1177  */
1178 static int
1179 mlx4_flow_flush(struct rte_eth_dev *dev,
1180                 struct rte_flow_error *error)
1181 {
1182         struct priv *priv = dev->data->dev_private;
1183         struct rte_flow *flow = LIST_FIRST(&priv->flows);
1184
1185         while (flow) {
1186                 struct rte_flow *next = LIST_NEXT(flow, next);
1187
1188                 if (!flow->internal)
1189                         mlx4_flow_destroy(dev, flow, error);
1190                 flow = next;
1191         }
1192         return 0;
1193 }
1194
1195 /**
1196  * Helper function to determine the next configured VLAN filter.
1197  *
1198  * @param priv
1199  *   Pointer to private structure.
1200  * @param vlan
1201  *   VLAN ID to use as a starting point.
1202  *
1203  * @return
1204  *   Next configured VLAN ID or a high value (>= 4096) if there is none.
1205  */
1206 static uint16_t
1207 mlx4_flow_internal_next_vlan(struct priv *priv, uint16_t vlan)
1208 {
1209         while (vlan < 4096) {
1210                 if (priv->dev->data->vlan_filter_conf.ids[vlan / 64] &
1211                     (UINT64_C(1) << (vlan % 64)))
1212                         return vlan;
1213                 ++vlan;
1214         }
1215         return vlan;
1216 }
1217
1218 /**
1219  * Generate internal flow rules.
1220  *
1221  * Various flow rules are created depending on the mode the device is in:
1222  *
1223  * 1. Promiscuous: port MAC + catch-all (VLAN filtering is ignored).
1224  * 2. All multicast: port MAC/VLAN + catch-all multicast.
1225  * 3. Otherwise: port MAC/VLAN + broadcast MAC/VLAN.
1226  *
1227  * About MAC flow rules:
1228  *
1229  * - MAC flow rules are generated from @p dev->data->mac_addrs
1230  *   (@p priv->mac array).
1231  * - An additional flow rule for Ethernet broadcasts is also generated.
1232  * - All these are per-VLAN if @p dev->data->dev_conf.rxmode.hw_vlan_filter
1233  *   is enabled and VLAN filters are configured.
1234  *
1235  * @param priv
1236  *   Pointer to private structure.
1237  * @param[out] error
1238  *   Perform verbose error reporting if not NULL.
1239  *
1240  * @return
1241  *   0 on success, a negative errno value otherwise and rte_errno is set.
1242  */
1243 static int
1244 mlx4_flow_internal(struct priv *priv, struct rte_flow_error *error)
1245 {
1246         struct rte_flow_attr attr = {
1247                 .priority = MLX4_FLOW_PRIORITY_LAST,
1248                 .ingress = 1,
1249         };
1250         struct rte_flow_item_eth eth_spec;
1251         const struct rte_flow_item_eth eth_mask = {
1252                 .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
1253         };
1254         const struct rte_flow_item_eth eth_allmulti = {
1255                 .dst.addr_bytes = "\x01\x00\x00\x00\x00\x00",
1256         };
1257         struct rte_flow_item_vlan vlan_spec;
1258         const struct rte_flow_item_vlan vlan_mask = {
1259                 .tci = RTE_BE16(0x0fff),
1260         };
1261         struct rte_flow_item pattern[] = {
1262                 {
1263                         .type = MLX4_FLOW_ITEM_TYPE_INTERNAL,
1264                 },
1265                 {
1266                         .type = RTE_FLOW_ITEM_TYPE_ETH,
1267                         .spec = &eth_spec,
1268                         .mask = &eth_mask,
1269                 },
1270                 {
1271                         /* Replaced with VLAN if filtering is enabled. */
1272                         .type = RTE_FLOW_ITEM_TYPE_END,
1273                 },
1274                 {
1275                         .type = RTE_FLOW_ITEM_TYPE_END,
1276                 },
1277         };
1278         /*
1279          * Round number of queues down to their previous power of 2 to
1280          * comply with RSS context limitations. Extra queues silently do not
1281          * get RSS by default.
1282          */
1283         uint32_t queues =
1284                 rte_align32pow2(priv->dev->data->nb_rx_queues + 1) >> 1;
1285         alignas(struct rte_flow_action_rss) uint8_t rss_conf_data
1286                 [offsetof(struct rte_flow_action_rss, queue) +
1287                  sizeof(((struct rte_flow_action_rss *)0)->queue[0]) * queues];
1288         struct rte_flow_action_rss *rss_conf = (void *)rss_conf_data;
1289         struct rte_flow_action actions[] = {
1290                 {
1291                         .type = RTE_FLOW_ACTION_TYPE_RSS,
1292                         .conf = rss_conf,
1293                 },
1294                 {
1295                         .type = RTE_FLOW_ACTION_TYPE_END,
1296                 },
1297         };
1298         struct ether_addr *rule_mac = &eth_spec.dst;
1299         rte_be16_t *rule_vlan =
1300                 priv->dev->data->dev_conf.rxmode.hw_vlan_filter &&
1301                 !priv->dev->data->promiscuous ?
1302                 &vlan_spec.tci :
1303                 NULL;
1304         int broadcast =
1305                 !priv->dev->data->promiscuous &&
1306                 !priv->dev->data->all_multicast;
1307         uint16_t vlan = 0;
1308         struct rte_flow *flow;
1309         unsigned int i;
1310         int err = 0;
1311
1312         /* Nothing to be done if there are no Rx queues. */
1313         if (!queues)
1314                 goto error;
1315         /* Prepare default RSS configuration. */
1316         *rss_conf = (struct rte_flow_action_rss){
1317                 .rss_conf = NULL, /* Rely on default fallback settings. */
1318                 .num = queues,
1319         };
1320         for (i = 0; i != queues; ++i)
1321                 rss_conf->queue[i] = i;
1322         /*
1323          * Set up VLAN item if filtering is enabled and at least one VLAN
1324          * filter is configured.
1325          */
1326         if (rule_vlan) {
1327                 vlan = mlx4_flow_internal_next_vlan(priv, 0);
1328                 if (vlan < 4096) {
1329                         pattern[2] = (struct rte_flow_item){
1330                                 .type = RTE_FLOW_ITEM_TYPE_VLAN,
1331                                 .spec = &vlan_spec,
1332                                 .mask = &vlan_mask,
1333                         };
1334 next_vlan:
1335                         *rule_vlan = rte_cpu_to_be_16(vlan);
1336                 } else {
1337                         rule_vlan = NULL;
1338                 }
1339         }
1340         for (i = 0; i != RTE_DIM(priv->mac) + broadcast; ++i) {
1341                 const struct ether_addr *mac;
1342
1343                 /* Broadcasts are handled by an extra iteration. */
1344                 if (i < RTE_DIM(priv->mac))
1345                         mac = &priv->mac[i];
1346                 else
1347                         mac = &eth_mask.dst;
1348                 if (is_zero_ether_addr(mac))
1349                         continue;
1350                 /* Check if MAC flow rule is already present. */
1351                 for (flow = LIST_FIRST(&priv->flows);
1352                      flow && flow->internal;
1353                      flow = LIST_NEXT(flow, next)) {
1354                         const struct ibv_flow_spec_eth *eth =
1355                                 (const void *)((uintptr_t)flow->ibv_attr +
1356                                                sizeof(*flow->ibv_attr));
1357                         unsigned int j;
1358
1359                         if (!flow->mac)
1360                                 continue;
1361                         assert(flow->ibv_attr->type == IBV_FLOW_ATTR_NORMAL);
1362                         assert(flow->ibv_attr->num_of_specs == 1);
1363                         assert(eth->type == IBV_FLOW_SPEC_ETH);
1364                         assert(flow->rss);
1365                         if (rule_vlan &&
1366                             (eth->val.vlan_tag != *rule_vlan ||
1367                              eth->mask.vlan_tag != RTE_BE16(0x0fff)))
1368                                 continue;
1369                         if (!rule_vlan && eth->mask.vlan_tag)
1370                                 continue;
1371                         for (j = 0; j != sizeof(mac->addr_bytes); ++j)
1372                                 if (eth->val.dst_mac[j] != mac->addr_bytes[j] ||
1373                                     eth->mask.dst_mac[j] != UINT8_C(0xff) ||
1374                                     eth->val.src_mac[j] != UINT8_C(0x00) ||
1375                                     eth->mask.src_mac[j] != UINT8_C(0x00))
1376                                         break;
1377                         if (j != sizeof(mac->addr_bytes))
1378                                 continue;
1379                         if (flow->rss->queues != queues ||
1380                             memcmp(flow->rss->queue_id, rss_conf->queue,
1381                                    queues * sizeof(flow->rss->queue_id[0])))
1382                                 continue;
1383                         break;
1384                 }
1385                 if (!flow || !flow->internal) {
1386                         /* Not found, create a new flow rule. */
1387                         memcpy(rule_mac, mac, sizeof(*mac));
1388                         flow = mlx4_flow_create(priv->dev, &attr, pattern,
1389                                                 actions, error);
1390                         if (!flow) {
1391                                 err = -rte_errno;
1392                                 goto error;
1393                         }
1394                 }
1395                 flow->select = 1;
1396                 flow->mac = 1;
1397         }
1398         if (rule_vlan) {
1399                 vlan = mlx4_flow_internal_next_vlan(priv, vlan + 1);
1400                 if (vlan < 4096)
1401                         goto next_vlan;
1402         }
1403         /* Take care of promiscuous and all multicast flow rules. */
1404         if (!broadcast) {
1405                 for (flow = LIST_FIRST(&priv->flows);
1406                      flow && flow->internal;
1407                      flow = LIST_NEXT(flow, next)) {
1408                         if (priv->dev->data->promiscuous) {
1409                                 if (flow->promisc)
1410                                         break;
1411                         } else {
1412                                 assert(priv->dev->data->all_multicast);
1413                                 if (flow->allmulti)
1414                                         break;
1415                         }
1416                 }
1417                 if (flow && flow->internal) {
1418                         assert(flow->rss);
1419                         if (flow->rss->queues != queues ||
1420                             memcmp(flow->rss->queue_id, rss_conf->queue,
1421                                    queues * sizeof(flow->rss->queue_id[0])))
1422                                 flow = NULL;
1423                 }
1424                 if (!flow || !flow->internal) {
1425                         /* Not found, create a new flow rule. */
1426                         if (priv->dev->data->promiscuous) {
1427                                 pattern[1].spec = NULL;
1428                                 pattern[1].mask = NULL;
1429                         } else {
1430                                 assert(priv->dev->data->all_multicast);
1431                                 pattern[1].spec = &eth_allmulti;
1432                                 pattern[1].mask = &eth_allmulti;
1433                         }
1434                         pattern[2] = pattern[3];
1435                         flow = mlx4_flow_create(priv->dev, &attr, pattern,
1436                                                 actions, error);
1437                         if (!flow) {
1438                                 err = -rte_errno;
1439                                 goto error;
1440                         }
1441                 }
1442                 assert(flow->promisc || flow->allmulti);
1443                 flow->select = 1;
1444         }
1445 error:
1446         /* Clear selection and clean up stale internal flow rules. */
1447         flow = LIST_FIRST(&priv->flows);
1448         while (flow && flow->internal) {
1449                 struct rte_flow *next = LIST_NEXT(flow, next);
1450
1451                 if (!flow->select)
1452                         claim_zero(mlx4_flow_destroy(priv->dev, flow, error));
1453                 else
1454                         flow->select = 0;
1455                 flow = next;
1456         }
1457         return err;
1458 }
1459
1460 /**
1461  * Synchronize flow rules.
1462  *
1463  * This function synchronizes flow rules with the state of the device by
1464  * taking into account isolated mode and whether target queues are
1465  * configured.
1466  *
1467  * @param priv
1468  *   Pointer to private structure.
1469  * @param[out] error
1470  *   Perform verbose error reporting if not NULL.
1471  *
1472  * @return
1473  *   0 on success, a negative errno value otherwise and rte_errno is set.
1474  */
1475 int
1476 mlx4_flow_sync(struct priv *priv, struct rte_flow_error *error)
1477 {
1478         struct rte_flow *flow;
1479         int ret;
1480
1481         /* Internal flow rules are guaranteed to come first in the list. */
1482         if (priv->isolated) {
1483                 /*
1484                  * Get rid of them in isolated mode, stop at the first
1485                  * non-internal rule found.
1486                  */
1487                 for (flow = LIST_FIRST(&priv->flows);
1488                      flow && flow->internal;
1489                      flow = LIST_FIRST(&priv->flows))
1490                         claim_zero(mlx4_flow_destroy(priv->dev, flow, error));
1491         } else {
1492                 /* Refresh internal rules. */
1493                 ret = mlx4_flow_internal(priv, error);
1494                 if (ret)
1495                         return ret;
1496         }
1497         /* Toggle the remaining flow rules . */
1498         LIST_FOREACH(flow, &priv->flows, next) {
1499                 ret = mlx4_flow_toggle(priv, flow, priv->started, error);
1500                 if (ret)
1501                         return ret;
1502         }
1503         if (!priv->started)
1504                 assert(!priv->drop);
1505         return 0;
1506 }
1507
1508 /**
1509  * Clean up all flow rules.
1510  *
1511  * Unlike mlx4_flow_flush(), this function takes care of all remaining flow
1512  * rules regardless of whether they are internal or user-configured.
1513  *
1514  * @param priv
1515  *   Pointer to private structure.
1516  */
1517 void
1518 mlx4_flow_clean(struct priv *priv)
1519 {
1520         struct rte_flow *flow;
1521
1522         while ((flow = LIST_FIRST(&priv->flows)))
1523                 mlx4_flow_destroy(priv->dev, flow, NULL);
1524         assert(LIST_EMPTY(&priv->rss));
1525 }
1526
1527 static const struct rte_flow_ops mlx4_flow_ops = {
1528         .validate = mlx4_flow_validate,
1529         .create = mlx4_flow_create,
1530         .destroy = mlx4_flow_destroy,
1531         .flush = mlx4_flow_flush,
1532         .isolate = mlx4_flow_isolate,
1533 };
1534
1535 /**
1536  * Manage filter operations.
1537  *
1538  * @param dev
1539  *   Pointer to Ethernet device structure.
1540  * @param filter_type
1541  *   Filter type.
1542  * @param filter_op
1543  *   Operation to perform.
1544  * @param arg
1545  *   Pointer to operation-specific structure.
1546  *
1547  * @return
1548  *   0 on success, negative errno value otherwise and rte_errno is set.
1549  */
1550 int
1551 mlx4_filter_ctrl(struct rte_eth_dev *dev,
1552                  enum rte_filter_type filter_type,
1553                  enum rte_filter_op filter_op,
1554                  void *arg)
1555 {
1556         switch (filter_type) {
1557         case RTE_ETH_FILTER_GENERIC:
1558                 if (filter_op != RTE_ETH_FILTER_GET)
1559                         break;
1560                 *(const void **)arg = &mlx4_flow_ops;
1561                 return 0;
1562         default:
1563                 ERROR("%p: filter type (%d) not supported",
1564                       (void *)dev, filter_type);
1565                 break;
1566         }
1567         rte_errno = ENOTSUP;
1568         return -rte_errno;
1569 }