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_driver.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                         /* Include inner RSS by default if supported. */
 163                         conv |= priv->hw_rss_sup & IBV_RX_HASH_INNER;
 164                         return conv;
 165                 }
 166                 if (!(rss_hf & ~seen))
 167                         return conv;
 168         }
 169         rte_errno = ENOTSUP;
 170         return (uint64_t)-1;
 171 }
 172
 173 /**
 174  * Merge Ethernet pattern item into flow rule handle.
 175  *
 176  * Additional mlx4-specific constraints on supported fields:
 177  *
 178  * - No support for partial masks, except in the specific case of matching
 179  *   all multicast traffic (@p spec->dst and @p mask->dst equal to
 180  *   01:00:00:00:00:00).
 181  * - Not providing @p item->spec or providing an empty @p mask->dst is
 182  *   *only* supported if the rule doesn't specify additional matching
 183  *   criteria (i.e. rule is promiscuous-like).
 184  *
 185  * @param[in, out] flow
 186  *   Flow rule handle to update.
 187  * @param[in] item
 188  *   Pattern item to merge.
 189  * @param[in] proc
 190  *   Associated item-processing object.
 191  * @param[out] error
 192  *   Perform verbose error reporting if not NULL.
 193  *
 194  * @return
 195  *   0 on success, a negative errno value otherwise and rte_errno is set.
 196  */
 197 static int
 198 mlx4_flow_merge_eth(struct rte_flow *flow,
 199                     const struct rte_flow_item *item,
 200                     const struct mlx4_flow_proc_item *proc,
 201                     struct rte_flow_error *error)
 202 {
 203         const struct rte_flow_item_eth *spec = item->spec;
 204         const struct rte_flow_item_eth *mask =
 205                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 206         struct ibv_flow_spec_eth *eth;
 207         const char *msg;
 208         unsigned int i;
 209
 210         if (!mask) {
 211                 flow->promisc = 1;
 212         } else {
 213                 uint32_t sum_dst = 0;
 214                 uint32_t sum_src = 0;
 215
 216                 for (i = 0; i != sizeof(mask->dst.addr_bytes); ++i) {
 217                         sum_dst += mask->dst.addr_bytes[i];
 218                         sum_src += mask->src.addr_bytes[i];
 219                 }
 220                 if (sum_src) {
 221                         msg = "mlx4 does not support source MAC matching";
 222                         goto error;
 223                 } else if (!sum_dst) {
 224                         flow->promisc = 1;
 225                 } else if (sum_dst == 1 && mask->dst.addr_bytes[0] == 1) {
 226                         if (!(spec->dst.addr_bytes[0] & 1)) {
 227                                 msg = "mlx4 does not support the explicit"
 228                                         " exclusion of all multicast traffic";
 229                                 goto error;
 230                         }
 231                         flow->allmulti = 1;
 232                 } else if (sum_dst != (UINT8_C(0xff) * ETHER_ADDR_LEN)) {
 233                         msg = "mlx4 does not support matching partial"
 234                                 " Ethernet fields";
 235                         goto error;
 236                 }
 237         }
 238         if (!flow->ibv_attr)
 239                 return 0;
 240         if (flow->promisc) {
 241                 flow->ibv_attr->type = IBV_FLOW_ATTR_ALL_DEFAULT;
 242                 return 0;
 243         }
 244         if (flow->allmulti) {
 245                 flow->ibv_attr->type = IBV_FLOW_ATTR_MC_DEFAULT;
 246                 return 0;
 247         }
 248         ++flow->ibv_attr->num_of_specs;
 249         eth = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 250         *eth = (struct ibv_flow_spec_eth) {
 251                 .type = IBV_FLOW_SPEC_ETH,
 252                 .size = sizeof(*eth),
 253         };
 254         memcpy(eth->val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
 255         memcpy(eth->mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
 256         /* Remove unwanted bits from values. */
 257         for (i = 0; i < ETHER_ADDR_LEN; ++i) {
 258                 eth->val.dst_mac[i] &= eth->mask.dst_mac[i];
 259         }
 260         return 0;
 261 error:
 262         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 263                                   item, msg);
 264 }
 265
 266 /**
 267  * Merge VLAN pattern item into flow rule handle.
 268  *
 269  * Additional mlx4-specific constraints on supported fields:
 270  *
 271  * - Matching *all* VLAN traffic by omitting @p item->spec or providing an
 272  *   empty @p item->mask would also include non-VLAN traffic. Doing so is
 273  *   therefore unsupported.
 274  * - No support for partial masks.
 275  *
 276  * @param[in, out] flow
 277  *   Flow rule handle to update.
 278  * @param[in] item
 279  *   Pattern item to merge.
 280  * @param[in] proc
 281  *   Associated item-processing object.
 282  * @param[out] error
 283  *   Perform verbose error reporting if not NULL.
 284  *
 285  * @return
 286  *   0 on success, a negative errno value otherwise and rte_errno is set.
 287  */
 288 static int
 289 mlx4_flow_merge_vlan(struct rte_flow *flow,
 290                      const struct rte_flow_item *item,
 291                      const struct mlx4_flow_proc_item *proc,
 292                      struct rte_flow_error *error)
 293 {
 294         const struct rte_flow_item_vlan *spec = item->spec;
 295         const struct rte_flow_item_vlan *mask =
 296                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 297         struct ibv_flow_spec_eth *eth;
 298         const char *msg;
 299
 300         if (!mask || !mask->tci) {
 301                 msg = "mlx4 cannot match all VLAN traffic while excluding"
 302                         " non-VLAN traffic, TCI VID must be specified";
 303                 goto error;
 304         }
 305         if (mask->tci != RTE_BE16(0x0fff)) {
 306                 msg = "mlx4 does not support partial TCI VID matching";
 307                 goto error;
 308         }
 309         if (!flow->ibv_attr)
 310                 return 0;
 311         eth = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size -
 312                        sizeof(*eth));
 313         eth->val.vlan_tag = spec->tci;
 314         eth->mask.vlan_tag = mask->tci;
 315         eth->val.vlan_tag &= eth->mask.vlan_tag;
 316         return 0;
 317 error:
 318         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 319                                   item, msg);
 320 }
 321
 322 /**
 323  * Merge IPv4 pattern item into flow rule handle.
 324  *
 325  * Additional mlx4-specific constraints on supported fields:
 326  *
 327  * - No support for partial masks.
 328  *
 329  * @param[in, out] flow
 330  *   Flow rule handle to update.
 331  * @param[in] item
 332  *   Pattern item to merge.
 333  * @param[in] proc
 334  *   Associated item-processing object.
 335  * @param[out] error
 336  *   Perform verbose error reporting if not NULL.
 337  *
 338  * @return
 339  *   0 on success, a negative errno value otherwise and rte_errno is set.
 340  */
 341 static int
 342 mlx4_flow_merge_ipv4(struct rte_flow *flow,
 343                      const struct rte_flow_item *item,
 344                      const struct mlx4_flow_proc_item *proc,
 345                      struct rte_flow_error *error)
 346 {
 347         const struct rte_flow_item_ipv4 *spec = item->spec;
 348         const struct rte_flow_item_ipv4 *mask =
 349                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 350         struct ibv_flow_spec_ipv4 *ipv4;
 351         const char *msg;
 352
 353         if (mask &&
 354             ((uint32_t)(mask->hdr.src_addr + 1) > UINT32_C(1) ||
 355              (uint32_t)(mask->hdr.dst_addr + 1) > UINT32_C(1))) {
 356                 msg = "mlx4 does not support matching partial IPv4 fields";
 357                 goto error;
 358         }
 359         if (!flow->ibv_attr)
 360                 return 0;
 361         ++flow->ibv_attr->num_of_specs;
 362         ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 363         *ipv4 = (struct ibv_flow_spec_ipv4) {
 364                 .type = IBV_FLOW_SPEC_IPV4,
 365                 .size = sizeof(*ipv4),
 366         };
 367         if (!spec)
 368                 return 0;
 369         ipv4->val = (struct ibv_flow_ipv4_filter) {
 370                 .src_ip = spec->hdr.src_addr,
 371                 .dst_ip = spec->hdr.dst_addr,
 372         };
 373         ipv4->mask = (struct ibv_flow_ipv4_filter) {
 374                 .src_ip = mask->hdr.src_addr,
 375                 .dst_ip = mask->hdr.dst_addr,
 376         };
 377         /* Remove unwanted bits from values. */
 378         ipv4->val.src_ip &= ipv4->mask.src_ip;
 379         ipv4->val.dst_ip &= ipv4->mask.dst_ip;
 380         return 0;
 381 error:
 382         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 383                                   item, msg);
 384 }
 385
 386 /**
 387  * Merge UDP pattern item into flow rule handle.
 388  *
 389  * Additional mlx4-specific constraints on supported fields:
 390  *
 391  * - No support for partial masks.
 392  *
 393  * @param[in, out] flow
 394  *   Flow rule handle to update.
 395  * @param[in] item
 396  *   Pattern item to merge.
 397  * @param[in] proc
 398  *   Associated item-processing object.
 399  * @param[out] error
 400  *   Perform verbose error reporting if not NULL.
 401  *
 402  * @return
 403  *   0 on success, a negative errno value otherwise and rte_errno is set.
 404  */
 405 static int
 406 mlx4_flow_merge_udp(struct rte_flow *flow,
 407                     const struct rte_flow_item *item,
 408                     const struct mlx4_flow_proc_item *proc,
 409                     struct rte_flow_error *error)
 410 {
 411         const struct rte_flow_item_udp *spec = item->spec;
 412         const struct rte_flow_item_udp *mask =
 413                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 414         struct ibv_flow_spec_tcp_udp *udp;
 415         const char *msg;
 416
 417         if (mask &&
 418             ((uint16_t)(mask->hdr.src_port + 1) > UINT16_C(1) ||
 419              (uint16_t)(mask->hdr.dst_port + 1) > UINT16_C(1))) {
 420                 msg = "mlx4 does not support matching partial UDP fields";
 421                 goto error;
 422         }
 423         if (!flow->ibv_attr)
 424                 return 0;
 425         ++flow->ibv_attr->num_of_specs;
 426         udp = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 427         *udp = (struct ibv_flow_spec_tcp_udp) {
 428                 .type = IBV_FLOW_SPEC_UDP,
 429                 .size = sizeof(*udp),
 430         };
 431         if (!spec)
 432                 return 0;
 433         udp->val.dst_port = spec->hdr.dst_port;
 434         udp->val.src_port = spec->hdr.src_port;
 435         udp->mask.dst_port = mask->hdr.dst_port;
 436         udp->mask.src_port = mask->hdr.src_port;
 437         /* Remove unwanted bits from values. */
 438         udp->val.src_port &= udp->mask.src_port;
 439         udp->val.dst_port &= udp->mask.dst_port;
 440         return 0;
 441 error:
 442         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 443                                   item, msg);
 444 }
 445
 446 /**
 447  * Merge TCP pattern item into flow rule handle.
 448  *
 449  * Additional mlx4-specific constraints on supported fields:
 450  *
 451  * - No support for partial masks.
 452  *
 453  * @param[in, out] flow
 454  *   Flow rule handle to update.
 455  * @param[in] item
 456  *   Pattern item to merge.
 457  * @param[in] proc
 458  *   Associated item-processing object.
 459  * @param[out] error
 460  *   Perform verbose error reporting if not NULL.
 461  *
 462  * @return
 463  *   0 on success, a negative errno value otherwise and rte_errno is set.
 464  */
 465 static int
 466 mlx4_flow_merge_tcp(struct rte_flow *flow,
 467                     const struct rte_flow_item *item,
 468                     const struct mlx4_flow_proc_item *proc,
 469                     struct rte_flow_error *error)
 470 {
 471         const struct rte_flow_item_tcp *spec = item->spec;
 472         const struct rte_flow_item_tcp *mask =
 473                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 474         struct ibv_flow_spec_tcp_udp *tcp;
 475         const char *msg;
 476
 477         if (mask &&
 478             ((uint16_t)(mask->hdr.src_port + 1) > UINT16_C(1) ||
 479              (uint16_t)(mask->hdr.dst_port + 1) > UINT16_C(1))) {
 480                 msg = "mlx4 does not support matching partial TCP fields";
 481                 goto error;
 482         }
 483         if (!flow->ibv_attr)
 484                 return 0;
 485         ++flow->ibv_attr->num_of_specs;
 486         tcp = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 487         *tcp = (struct ibv_flow_spec_tcp_udp) {
 488                 .type = IBV_FLOW_SPEC_TCP,
 489                 .size = sizeof(*tcp),
 490         };
 491         if (!spec)
 492                 return 0;
 493         tcp->val.dst_port = spec->hdr.dst_port;
 494         tcp->val.src_port = spec->hdr.src_port;
 495         tcp->mask.dst_port = mask->hdr.dst_port;
 496         tcp->mask.src_port = mask->hdr.src_port;
 497         /* Remove unwanted bits from values. */
 498         tcp->val.src_port &= tcp->mask.src_port;
 499         tcp->val.dst_port &= tcp->mask.dst_port;
 500         return 0;
 501 error:
 502         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 503                                   item, msg);
 504 }
 505
 506 /**
 507  * Perform basic sanity checks on a pattern item.
 508  *
 509  * @param[in] item
 510  *   Item specification.
 511  * @param[in] proc
 512  *   Associated item-processing object.
 513  * @param[out] error
 514  *   Perform verbose error reporting if not NULL.
 515  *
 516  * @return
 517  *   0 on success, a negative errno value otherwise and rte_errno is set.
 518  */
 519 static int
 520 mlx4_flow_item_check(const struct rte_flow_item *item,
 521                      const struct mlx4_flow_proc_item *proc,
 522                      struct rte_flow_error *error)
 523 {
 524         const uint8_t *mask;
 525         unsigned int i;
 526
 527         /* item->last and item->mask cannot exist without item->spec. */
 528         if (!item->spec && (item->mask || item->last))
 529                 return rte_flow_error_set
 530                         (error, EINVAL, RTE_FLOW_ERROR_TYPE_ITEM, item,
 531                          "\"mask\" or \"last\" field provided without a"
 532                          " corresponding \"spec\"");
 533         /* No spec, no mask, no problem. */
 534         if (!item->spec)
 535                 return 0;
 536         mask = item->mask ?
 537                 (const uint8_t *)item->mask :
 538                 (const uint8_t *)proc->mask_default;
 539         assert(mask);
 540         /*
 541          * Single-pass check to make sure that:
 542          * - Mask is supported, no bits are set outside proc->mask_support.
 543          * - Both item->spec and item->last are included in mask.
 544          */
 545         for (i = 0; i != proc->mask_sz; ++i) {
 546                 if (!mask[i])
 547                         continue;
 548                 if ((mask[i] | ((const uint8_t *)proc->mask_support)[i]) !=
 549                     ((const uint8_t *)proc->mask_support)[i])
 550                         return rte_flow_error_set
 551                                 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 552                                  item, "unsupported field found in \"mask\"");
 553                 if (item->last &&
 554                     (((const uint8_t *)item->spec)[i] & mask[i]) !=
 555                     (((const uint8_t *)item->last)[i] & mask[i]))
 556                         return rte_flow_error_set
 557                                 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 558                                  item,
 559                                  "range between \"spec\" and \"last\""
 560                                  " is larger than \"mask\"");
 561         }
 562         return 0;
 563 }
 564
 565 /** Graph of supported items and associated actions. */
 566 static const struct mlx4_flow_proc_item mlx4_flow_proc_item_list[] = {
 567         [RTE_FLOW_ITEM_TYPE_END] = {
 568                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_ETH),
 569         },
 570         [RTE_FLOW_ITEM_TYPE_ETH] = {
 571                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_VLAN,
 572                                        RTE_FLOW_ITEM_TYPE_IPV4),
 573                 .mask_support = &(const struct rte_flow_item_eth){
 574                         /* Only destination MAC can be matched. */
 575                         .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 576                 },
 577                 .mask_default = &rte_flow_item_eth_mask,
 578                 .mask_sz = sizeof(struct rte_flow_item_eth),
 579                 .merge = mlx4_flow_merge_eth,
 580                 .dst_sz = sizeof(struct ibv_flow_spec_eth),
 581         },
 582         [RTE_FLOW_ITEM_TYPE_VLAN] = {
 583                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_IPV4),
 584                 .mask_support = &(const struct rte_flow_item_vlan){
 585                         /* Only TCI VID matching is supported. */
 586                         .tci = RTE_BE16(0x0fff),
 587                 },
 588                 .mask_default = &rte_flow_item_vlan_mask,
 589                 .mask_sz = sizeof(struct rte_flow_item_vlan),
 590                 .merge = mlx4_flow_merge_vlan,
 591                 .dst_sz = 0,
 592         },
 593         [RTE_FLOW_ITEM_TYPE_IPV4] = {
 594                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_UDP,
 595                                        RTE_FLOW_ITEM_TYPE_TCP),
 596                 .mask_support = &(const struct rte_flow_item_ipv4){
 597                         .hdr = {
 598                                 .src_addr = RTE_BE32(0xffffffff),
 599                                 .dst_addr = RTE_BE32(0xffffffff),
 600                         },
 601                 },
 602                 .mask_default = &rte_flow_item_ipv4_mask,
 603                 .mask_sz = sizeof(struct rte_flow_item_ipv4),
 604                 .merge = mlx4_flow_merge_ipv4,
 605                 .dst_sz = sizeof(struct ibv_flow_spec_ipv4),
 606         },
 607         [RTE_FLOW_ITEM_TYPE_UDP] = {
 608                 .mask_support = &(const struct rte_flow_item_udp){
 609                         .hdr = {
 610                                 .src_port = RTE_BE16(0xffff),
 611                                 .dst_port = RTE_BE16(0xffff),
 612                         },
 613                 },
 614                 .mask_default = &rte_flow_item_udp_mask,
 615                 .mask_sz = sizeof(struct rte_flow_item_udp),
 616                 .merge = mlx4_flow_merge_udp,
 617                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 618         },
 619         [RTE_FLOW_ITEM_TYPE_TCP] = {
 620                 .mask_support = &(const struct rte_flow_item_tcp){
 621                         .hdr = {
 622                                 .src_port = RTE_BE16(0xffff),
 623                                 .dst_port = RTE_BE16(0xffff),
 624                         },
 625                 },
 626                 .mask_default = &rte_flow_item_tcp_mask,
 627                 .mask_sz = sizeof(struct rte_flow_item_tcp),
 628                 .merge = mlx4_flow_merge_tcp,
 629                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 630         },
 631 };
 632
 633 /**
 634  * Make sure a flow rule is supported and initialize associated structure.
 635  *
 636  * @param priv
 637  *   Pointer to private structure.
 638  * @param[in] attr
 639  *   Flow rule attributes.
 640  * @param[in] pattern
 641  *   Pattern specification (list terminated by the END pattern item).
 642  * @param[in] actions
 643  *   Associated actions (list terminated by the END action).
 644  * @param[out] error
 645  *   Perform verbose error reporting if not NULL.
 646  * @param[in, out] addr
 647  *   Buffer where the resulting flow rule handle pointer must be stored.
 648  *   If NULL, stop processing after validation stage.
 649  *
 650  * @return
 651  *   0 on success, a negative errno value otherwise and rte_errno is set.
 652  */
 653 static int
 654 mlx4_flow_prepare(struct priv *priv,
 655                   const struct rte_flow_attr *attr,
 656                   const struct rte_flow_item pattern[],
 657                   const struct rte_flow_action actions[],
 658                   struct rte_flow_error *error,
 659                   struct rte_flow **addr)
 660 {
 661         const struct rte_flow_item *item;
 662         const struct rte_flow_action *action;
 663         const struct mlx4_flow_proc_item *proc;
 664         struct rte_flow temp = { .ibv_attr_size = sizeof(*temp.ibv_attr) };
 665         struct rte_flow *flow = &temp;
 666         const char *msg = NULL;
 667
 668         if (attr->group)
 669                 return rte_flow_error_set
 670                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 671                          NULL, "groups are not supported");
 672         if (attr->priority > MLX4_FLOW_PRIORITY_LAST)
 673                 return rte_flow_error_set
 674                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 675                          NULL, "maximum priority level is "
 676                          MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST));
 677         if (attr->egress)
 678                 return rte_flow_error_set
 679                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 680                          NULL, "egress is not supported");
 681         if (!attr->ingress)
 682                 return rte_flow_error_set
 683                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 684                          NULL, "only ingress is supported");
 685 fill:
 686         proc = mlx4_flow_proc_item_list;
 687         /* Go over pattern. */
 688         for (item = pattern; item->type; ++item) {
 689                 const struct mlx4_flow_proc_item *next = NULL;
 690                 unsigned int i;
 691                 int err;
 692
 693                 if (item->type == RTE_FLOW_ITEM_TYPE_VOID)
 694                         continue;
 695                 if (item->type == MLX4_FLOW_ITEM_TYPE_INTERNAL) {
 696                         flow->internal = 1;
 697                         continue;
 698                 }
 699                 if (flow->promisc || flow->allmulti) {
 700                         msg = "mlx4 does not support additional matching"
 701                                 " criteria combined with indiscriminate"
 702                                 " matching on Ethernet headers";
 703                         goto exit_item_not_supported;
 704                 }
 705                 for (i = 0; proc->next_item && proc->next_item[i]; ++i) {
 706                         if (proc->next_item[i] == item->type) {
 707                                 next = &mlx4_flow_proc_item_list[item->type];
 708                                 break;
 709                         }
 710                 }
 711                 if (!next)
 712                         goto exit_item_not_supported;
 713                 proc = next;
 714                 /*
 715                  * Perform basic sanity checks only once, while handle is
 716                  * not allocated.
 717                  */
 718                 if (flow == &temp) {
 719                         err = mlx4_flow_item_check(item, proc, error);
 720                         if (err)
 721                                 return err;
 722                 }
 723                 if (proc->merge) {
 724                         err = proc->merge(flow, item, proc, error);
 725                         if (err)
 726                                 return err;
 727                 }
 728                 flow->ibv_attr_size += proc->dst_sz;
 729         }
 730         /* Go over actions list. */
 731         for (action = actions; action->type; ++action) {
 732                 switch (action->type) {
 733                         const struct rte_flow_action_queue *queue;
 734                         const struct rte_flow_action_rss *rss;
 735                         const struct rte_eth_rss_conf *rss_conf;
 736                         unsigned int i;
 737
 738                 case RTE_FLOW_ACTION_TYPE_VOID:
 739                         continue;
 740                 case RTE_FLOW_ACTION_TYPE_DROP:
 741                         flow->drop = 1;
 742                         break;
 743                 case RTE_FLOW_ACTION_TYPE_QUEUE:
 744                         if (flow->rss)
 745                                 break;
 746                         queue = action->conf;
 747                         if (queue->index >= priv->dev->data->nb_rx_queues) {
 748                                 msg = "queue target index beyond number of"
 749                                         " configured Rx queues";
 750                                 goto exit_action_not_supported;
 751                         }
 752                         flow->rss = mlx4_rss_get
 753                                 (priv, 0, mlx4_rss_hash_key_default, 1,
 754                                  &queue->index);
 755                         if (!flow->rss) {
 756                                 msg = "not enough resources for additional"
 757                                         " single-queue RSS context";
 758                                 goto exit_action_not_supported;
 759                         }
 760                         break;
 761                 case RTE_FLOW_ACTION_TYPE_RSS:
 762                         if (flow->rss)
 763                                 break;
 764                         rss = action->conf;
 765                         /* Default RSS configuration if none is provided. */
 766                         rss_conf =
 767                                 rss->rss_conf ?
 768                                 rss->rss_conf :
 769                                 &(struct rte_eth_rss_conf){
 770                                         .rss_key = mlx4_rss_hash_key_default,
 771                                         .rss_key_len = MLX4_RSS_HASH_KEY_SIZE,
 772                                         .rss_hf = -1,
 773                                 };
 774                         /* Sanity checks. */
 775                         for (i = 0; i < rss->num; ++i)
 776                                 if (rss->queue[i] >=
 777                                     priv->dev->data->nb_rx_queues)
 778                                         break;
 779                         if (i != rss->num) {
 780                                 msg = "queue index target beyond number of"
 781                                         " configured Rx queues";
 782                                 goto exit_action_not_supported;
 783                         }
 784                         if (!rte_is_power_of_2(rss->num)) {
 785                                 msg = "for RSS, mlx4 requires the number of"
 786                                         " queues to be a power of two";
 787                                 goto exit_action_not_supported;
 788                         }
 789                         if (rss_conf->rss_key_len !=
 790                             sizeof(flow->rss->key)) {
 791                                 msg = "mlx4 supports exactly one RSS hash key"
 792                                         " length: "
 793                                         MLX4_STR_EXPAND(MLX4_RSS_HASH_KEY_SIZE);
 794                                 goto exit_action_not_supported;
 795                         }
 796                         for (i = 1; i < rss->num; ++i)
 797                                 if (rss->queue[i] - rss->queue[i - 1] != 1)
 798                                         break;
 799                         if (i != rss->num) {
 800                                 msg = "mlx4 requires RSS contexts to use"
 801                                         " consecutive queue indices only";
 802                                 goto exit_action_not_supported;
 803                         }
 804                         if (rss->queue[0] % rss->num) {
 805                                 msg = "mlx4 requires the first queue of a RSS"
 806                                         " context to be aligned on a multiple"
 807                                         " of the context size";
 808                                 goto exit_action_not_supported;
 809                         }
 810                         flow->rss = mlx4_rss_get
 811                                 (priv,
 812                                  mlx4_conv_rss_hf(priv, rss_conf->rss_hf),
 813                                  rss_conf->rss_key, rss->num, rss->queue);
 814                         if (!flow->rss) {
 815                                 msg = "either invalid parameters or not enough"
 816                                         " resources for additional multi-queue"
 817                                         " RSS context";
 818                                 goto exit_action_not_supported;
 819                         }
 820                         break;
 821                 default:
 822                         goto exit_action_not_supported;
 823                 }
 824         }
 825         if (!flow->rss && !flow->drop)
 826                 return rte_flow_error_set
 827                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 828                          NULL, "no valid action");
 829         /* Validation ends here. */
 830         if (!addr) {
 831                 if (flow->rss)
 832                         mlx4_rss_put(flow->rss);
 833                 return 0;
 834         }
 835         if (flow == &temp) {
 836                 /* Allocate proper handle based on collected data. */
 837                 const struct mlx4_malloc_vec vec[] = {
 838                         {
 839                                 .align = alignof(struct rte_flow),
 840                                 .size = sizeof(*flow),
 841                                 .addr = (void **)&flow,
 842                         },
 843                         {
 844                                 .align = alignof(struct ibv_flow_attr),
 845                                 .size = temp.ibv_attr_size,
 846                                 .addr = (void **)&temp.ibv_attr,
 847                         },
 848                 };
 849
 850                 if (!mlx4_zmallocv(__func__, vec, RTE_DIM(vec)))
 851                         return rte_flow_error_set
 852                                 (error, -rte_errno,
 853                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
 854                                  "flow rule handle allocation failure");
 855                 /* Most fields will be updated by second pass. */
 856                 *flow = (struct rte_flow){
 857                         .ibv_attr = temp.ibv_attr,
 858                         .ibv_attr_size = sizeof(*flow->ibv_attr),
 859                         .rss = temp.rss,
 860                 };
 861                 *flow->ibv_attr = (struct ibv_flow_attr){
 862                         .type = IBV_FLOW_ATTR_NORMAL,
 863                         .size = sizeof(*flow->ibv_attr),
 864                         .priority = attr->priority,
 865                         .port = priv->port,
 866                 };
 867                 goto fill;
 868         }
 869         *addr = flow;
 870         return 0;
 871 exit_item_not_supported:
 872         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 873                                   item, msg ? msg : "item not supported");
 874 exit_action_not_supported:
 875         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 876                                   action, msg ? msg : "action not supported");
 877 }
 878
 879 /**
 880  * Validate a flow supported by the NIC.
 881  *
 882  * @see rte_flow_validate()
 883  * @see rte_flow_ops
 884  */
 885 static int
 886 mlx4_flow_validate(struct rte_eth_dev *dev,
 887                    const struct rte_flow_attr *attr,
 888                    const struct rte_flow_item pattern[],
 889                    const struct rte_flow_action actions[],
 890                    struct rte_flow_error *error)
 891 {
 892         struct priv *priv = dev->data->dev_private;
 893
 894         return mlx4_flow_prepare(priv, attr, pattern, actions, error, NULL);
 895 }
 896
 897 /**
 898  * Get a drop flow rule resources instance.
 899  *
 900  * @param priv
 901  *   Pointer to private structure.
 902  *
 903  * @return
 904  *   Pointer to drop flow resources on success, NULL otherwise and rte_errno
 905  *   is set.
 906  */
 907 static struct mlx4_drop *
 908 mlx4_drop_get(struct priv *priv)
 909 {
 910         struct mlx4_drop *drop = priv->drop;
 911
 912         if (drop) {
 913                 assert(drop->refcnt);
 914                 assert(drop->priv == priv);
 915                 ++drop->refcnt;
 916                 return drop;
 917         }
 918         drop = rte_malloc(__func__, sizeof(*drop), 0);
 919         if (!drop)
 920                 goto error;
 921         *drop = (struct mlx4_drop){
 922                 .priv = priv,
 923                 .refcnt = 1,
 924         };
 925         drop->cq = ibv_create_cq(priv->ctx, 1, NULL, NULL, 0);
 926         if (!drop->cq)
 927                 goto error;
 928         drop->qp = ibv_create_qp(priv->pd,
 929                                  &(struct ibv_qp_init_attr){
 930                                         .send_cq = drop->cq,
 931                                         .recv_cq = drop->cq,
 932                                         .qp_type = IBV_QPT_RAW_PACKET,
 933                                  });
 934         if (!drop->qp)
 935                 goto error;
 936         priv->drop = drop;
 937         return drop;
 938 error:
 939         if (drop->qp)
 940                 claim_zero(ibv_destroy_qp(drop->qp));
 941         if (drop->cq)
 942                 claim_zero(ibv_destroy_cq(drop->cq));
 943         if (drop)
 944                 rte_free(drop);
 945         rte_errno = ENOMEM;
 946         return NULL;
 947 }
 948
 949 /**
 950  * Give back a drop flow rule resources instance.
 951  *
 952  * @param drop
 953  *   Pointer to drop flow rule resources.
 954  */
 955 static void
 956 mlx4_drop_put(struct mlx4_drop *drop)
 957 {
 958         assert(drop->refcnt);
 959         if (--drop->refcnt)
 960                 return;
 961         drop->priv->drop = NULL;
 962         claim_zero(ibv_destroy_qp(drop->qp));
 963         claim_zero(ibv_destroy_cq(drop->cq));
 964         rte_free(drop);
 965 }
 966
 967 /**
 968  * Toggle a configured flow rule.
 969  *
 970  * @param priv
 971  *   Pointer to private structure.
 972  * @param flow
 973  *   Flow rule handle to toggle.
 974  * @param enable
 975  *   Whether associated Verbs flow must be created or removed.
 976  * @param[out] error
 977  *   Perform verbose error reporting if not NULL.
 978  *
 979  * @return
 980  *   0 on success, a negative errno value otherwise and rte_errno is set.
 981  */
 982 static int
 983 mlx4_flow_toggle(struct priv *priv,
 984                  struct rte_flow *flow,
 985                  int enable,
 986                  struct rte_flow_error *error)
 987 {
 988         struct ibv_qp *qp = NULL;
 989         const char *msg;
 990         int err;
 991
 992         if (!enable) {
 993                 if (!flow->ibv_flow)
 994                         return 0;
 995                 claim_zero(ibv_destroy_flow(flow->ibv_flow));
 996                 flow->ibv_flow = NULL;
 997                 if (flow->drop)
 998                         mlx4_drop_put(priv->drop);
 999                 else if (flow->rss)
1000                         mlx4_rss_detach(flow->rss);
1001                 return 0;
1002         }
1003         assert(flow->ibv_attr);
1004         if (!flow->internal &&
1005             !priv->isolated &&
1006             flow->ibv_attr->priority == MLX4_FLOW_PRIORITY_LAST) {
1007                 if (flow->ibv_flow) {
1008                         claim_zero(ibv_destroy_flow(flow->ibv_flow));
1009                         flow->ibv_flow = NULL;
1010                         if (flow->drop)
1011                                 mlx4_drop_put(priv->drop);
1012                         else if (flow->rss)
1013                                 mlx4_rss_detach(flow->rss);
1014                 }
1015                 err = EACCES;
1016                 msg = ("priority level "
1017                        MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST)
1018                        " is reserved when not in isolated mode");
1019                 goto error;
1020         }
1021         if (flow->rss) {
1022                 struct mlx4_rss *rss = flow->rss;
1023                 int missing = 0;
1024                 unsigned int i;
1025
1026                 /* Stop at the first nonexistent target queue. */
1027                 for (i = 0; i != rss->queues; ++i)
1028                         if (rss->queue_id[i] >=
1029                             priv->dev->data->nb_rx_queues ||
1030                             !priv->dev->data->rx_queues[rss->queue_id[i]]) {
1031                                 missing = 1;
1032                                 break;
1033                         }
1034                 if (flow->ibv_flow) {
1035                         if (missing ^ !flow->drop)
1036                                 return 0;
1037                         /* Verbs flow needs updating. */
1038                         claim_zero(ibv_destroy_flow(flow->ibv_flow));
1039                         flow->ibv_flow = NULL;
1040                         if (flow->drop)
1041                                 mlx4_drop_put(priv->drop);
1042                         else
1043                                 mlx4_rss_detach(rss);
1044                 }
1045                 if (!missing) {
1046                         err = mlx4_rss_attach(rss);
1047                         if (err) {
1048                                 err = -err;
1049                                 msg = "cannot create indirection table or hash"
1050                                         " QP to associate flow rule with";
1051                                 goto error;
1052                         }
1053                         qp = rss->qp;
1054                 }
1055                 /* A missing target queue drops traffic implicitly. */
1056                 flow->drop = missing;
1057         }
1058         if (flow->drop) {
1059                 mlx4_drop_get(priv);
1060                 if (!priv->drop) {
1061                         err = rte_errno;
1062                         msg = "resources for drop flow rule cannot be created";
1063                         goto error;
1064                 }
1065                 qp = priv->drop->qp;
1066         }
1067         assert(qp);
1068         if (flow->ibv_flow)
1069                 return 0;
1070         flow->ibv_flow = ibv_create_flow(qp, flow->ibv_attr);
1071         if (flow->ibv_flow)
1072                 return 0;
1073         if (flow->drop)
1074                 mlx4_drop_put(priv->drop);
1075         else if (flow->rss)
1076                 mlx4_rss_detach(flow->rss);
1077         err = errno;
1078         msg = "flow rule rejected by device";
1079 error:
1080         return rte_flow_error_set
1081                 (error, err, RTE_FLOW_ERROR_TYPE_HANDLE, flow, msg);
1082 }
1083
1084 /**
1085  * Create a flow.
1086  *
1087  * @see rte_flow_create()
1088  * @see rte_flow_ops
1089  */
1090 static struct rte_flow *
1091 mlx4_flow_create(struct rte_eth_dev *dev,
1092                  const struct rte_flow_attr *attr,
1093                  const struct rte_flow_item pattern[],
1094                  const struct rte_flow_action actions[],
1095                  struct rte_flow_error *error)
1096 {
1097         struct priv *priv = dev->data->dev_private;
1098         struct rte_flow *flow;
1099         int err;
1100
1101         err = mlx4_flow_prepare(priv, attr, pattern, actions, error, &flow);
1102         if (err)
1103                 return NULL;
1104         err = mlx4_flow_toggle(priv, flow, priv->started, error);
1105         if (!err) {
1106                 struct rte_flow *curr = LIST_FIRST(&priv->flows);
1107
1108                 /* New rules are inserted after internal ones. */
1109                 if (!curr || !curr->internal) {
1110                         LIST_INSERT_HEAD(&priv->flows, flow, next);
1111                 } else {
1112                         while (LIST_NEXT(curr, next) &&
1113                                LIST_NEXT(curr, next)->internal)
1114                                 curr = LIST_NEXT(curr, next);
1115                         LIST_INSERT_AFTER(curr, flow, next);
1116                 }
1117                 return flow;
1118         }
1119         if (flow->rss)
1120                 mlx4_rss_put(flow->rss);
1121         rte_flow_error_set(error, -err, RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1122                            error->message);
1123         rte_free(flow);
1124         return NULL;
1125 }
1126
1127 /**
1128  * Configure isolated mode.
1129  *
1130  * @see rte_flow_isolate()
1131  * @see rte_flow_ops
1132  */
1133 static int
1134 mlx4_flow_isolate(struct rte_eth_dev *dev,
1135                   int enable,
1136                   struct rte_flow_error *error)
1137 {
1138         struct priv *priv = dev->data->dev_private;
1139
1140         if (!!enable == !!priv->isolated)
1141                 return 0;
1142         priv->isolated = !!enable;
1143         if (mlx4_flow_sync(priv, error)) {
1144                 priv->isolated = !enable;
1145                 return -rte_errno;
1146         }
1147         return 0;
1148 }
1149
1150 /**
1151  * Destroy a flow rule.
1152  *
1153  * @see rte_flow_destroy()
1154  * @see rte_flow_ops
1155  */
1156 static int
1157 mlx4_flow_destroy(struct rte_eth_dev *dev,
1158                   struct rte_flow *flow,
1159                   struct rte_flow_error *error)
1160 {
1161         struct priv *priv = dev->data->dev_private;
1162         int err = mlx4_flow_toggle(priv, flow, 0, error);
1163
1164         if (err)
1165                 return err;
1166         LIST_REMOVE(flow, next);
1167         if (flow->rss)
1168                 mlx4_rss_put(flow->rss);
1169         rte_free(flow);
1170         return 0;
1171 }
1172
1173 /**
1174  * Destroy user-configured flow rules.
1175  *
1176  * This function skips internal flows rules.
1177  *
1178  * @see rte_flow_flush()
1179  * @see rte_flow_ops
1180  */
1181 static int
1182 mlx4_flow_flush(struct rte_eth_dev *dev,
1183                 struct rte_flow_error *error)
1184 {
1185         struct priv *priv = dev->data->dev_private;
1186         struct rte_flow *flow = LIST_FIRST(&priv->flows);
1187
1188         while (flow) {
1189                 struct rte_flow *next = LIST_NEXT(flow, next);
1190
1191                 if (!flow->internal)
1192                         mlx4_flow_destroy(dev, flow, error);
1193                 flow = next;
1194         }
1195         return 0;
1196 }
1197
1198 /**
1199  * Helper function to determine the next configured VLAN filter.
1200  *
1201  * @param priv
1202  *   Pointer to private structure.
1203  * @param vlan
1204  *   VLAN ID to use as a starting point.
1205  *
1206  * @return
1207  *   Next configured VLAN ID or a high value (>= 4096) if there is none.
1208  */
1209 static uint16_t
1210 mlx4_flow_internal_next_vlan(struct priv *priv, uint16_t vlan)
1211 {
1212         while (vlan < 4096) {
1213                 if (priv->dev->data->vlan_filter_conf.ids[vlan / 64] &
1214                     (UINT64_C(1) << (vlan % 64)))
1215                         return vlan;
1216                 ++vlan;
1217         }
1218         return vlan;
1219 }
1220
1221 /**
1222  * Generate internal flow rules.
1223  *
1224  * Various flow rules are created depending on the mode the device is in:
1225  *
1226  * 1. Promiscuous: port MAC + catch-all (VLAN filtering is ignored).
1227  * 2. All multicast: port MAC/VLAN + catch-all multicast.
1228  * 3. Otherwise: port MAC/VLAN + broadcast MAC/VLAN.
1229  *
1230  * About MAC flow rules:
1231  *
1232  * - MAC flow rules are generated from @p dev->data->mac_addrs
1233  *   (@p priv->mac array).
1234  * - An additional flow rule for Ethernet broadcasts is also generated.
1235  * - All these are per-VLAN if @p DEV_RX_OFFLOAD_VLAN_FILTER
1236  *   is enabled and VLAN filters are configured.
1237  *
1238  * @param priv
1239  *   Pointer to private structure.
1240  * @param[out] error
1241  *   Perform verbose error reporting if not NULL.
1242  *
1243  * @return
1244  *   0 on success, a negative errno value otherwise and rte_errno is set.
1245  */
1246 static int
1247 mlx4_flow_internal(struct priv *priv, struct rte_flow_error *error)
1248 {
1249         struct rte_flow_attr attr = {
1250                 .priority = MLX4_FLOW_PRIORITY_LAST,
1251                 .ingress = 1,
1252         };
1253         struct rte_flow_item_eth eth_spec;
1254         const struct rte_flow_item_eth eth_mask = {
1255                 .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
1256         };
1257         const struct rte_flow_item_eth eth_allmulti = {
1258                 .dst.addr_bytes = "\x01\x00\x00\x00\x00\x00",
1259         };
1260         struct rte_flow_item_vlan vlan_spec;
1261         const struct rte_flow_item_vlan vlan_mask = {
1262                 .tci = RTE_BE16(0x0fff),
1263         };
1264         struct rte_flow_item pattern[] = {
1265                 {
1266                         .type = MLX4_FLOW_ITEM_TYPE_INTERNAL,
1267                 },
1268                 {
1269                         .type = RTE_FLOW_ITEM_TYPE_ETH,
1270                         .spec = &eth_spec,
1271                         .mask = &eth_mask,
1272                 },
1273                 {
1274                         /* Replaced with VLAN if filtering is enabled. */
1275                         .type = RTE_FLOW_ITEM_TYPE_END,
1276                 },
1277                 {
1278                         .type = RTE_FLOW_ITEM_TYPE_END,
1279                 },
1280         };
1281         /*
1282          * Round number of queues down to their previous power of 2 to
1283          * comply with RSS context limitations. Extra queues silently do not
1284          * get RSS by default.
1285          */
1286         uint32_t queues =
1287                 rte_align32pow2(priv->dev->data->nb_rx_queues + 1) >> 1;
1288         alignas(struct rte_flow_action_rss) uint8_t rss_conf_data
1289                 [offsetof(struct rte_flow_action_rss, queue) +
1290                  sizeof(((struct rte_flow_action_rss *)0)->queue[0]) * queues];
1291         struct rte_flow_action_rss *rss_conf = (void *)rss_conf_data;
1292         struct rte_flow_action actions[] = {
1293                 {
1294                         .type = RTE_FLOW_ACTION_TYPE_RSS,
1295                         .conf = rss_conf,
1296                 },
1297                 {
1298                         .type = RTE_FLOW_ACTION_TYPE_END,
1299                 },
1300         };
1301         struct ether_addr *rule_mac = &eth_spec.dst;
1302         rte_be16_t *rule_vlan =
1303                 (priv->dev->data->dev_conf.rxmode.offloads &
1304                  DEV_RX_OFFLOAD_VLAN_FILTER) &&
1305                 !priv->dev->data->promiscuous ?
1306                 &vlan_spec.tci :
1307                 NULL;
1308         int broadcast =
1309                 !priv->dev->data->promiscuous &&
1310                 !priv->dev->data->all_multicast;
1311         uint16_t vlan = 0;
1312         struct rte_flow *flow;
1313         unsigned int i;
1314         int err = 0;
1315
1316         /* Nothing to be done if there are no Rx queues. */
1317         if (!queues)
1318                 goto error;
1319         /* Prepare default RSS configuration. */
1320         *rss_conf = (struct rte_flow_action_rss){
1321                 .rss_conf = NULL, /* Rely on default fallback settings. */
1322                 .num = queues,
1323         };
1324         for (i = 0; i != queues; ++i)
1325                 rss_conf->queue[i] = i;
1326         /*
1327          * Set up VLAN item if filtering is enabled and at least one VLAN
1328          * filter is configured.
1329          */
1330         if (rule_vlan) {
1331                 vlan = mlx4_flow_internal_next_vlan(priv, 0);
1332                 if (vlan < 4096) {
1333                         pattern[2] = (struct rte_flow_item){
1334                                 .type = RTE_FLOW_ITEM_TYPE_VLAN,
1335                                 .spec = &vlan_spec,
1336                                 .mask = &vlan_mask,
1337                         };
1338 next_vlan:
1339                         *rule_vlan = rte_cpu_to_be_16(vlan);
1340                 } else {
1341                         rule_vlan = NULL;
1342                 }
1343         }
1344         for (i = 0; i != RTE_DIM(priv->mac) + broadcast; ++i) {
1345                 const struct ether_addr *mac;
1346
1347                 /* Broadcasts are handled by an extra iteration. */
1348                 if (i < RTE_DIM(priv->mac))
1349                         mac = &priv->mac[i];
1350                 else
1351                         mac = &eth_mask.dst;
1352                 if (is_zero_ether_addr(mac))
1353                         continue;
1354                 /* Check if MAC flow rule is already present. */
1355                 for (flow = LIST_FIRST(&priv->flows);
1356                      flow && flow->internal;
1357                      flow = LIST_NEXT(flow, next)) {
1358                         const struct ibv_flow_spec_eth *eth =
1359                                 (const void *)((uintptr_t)flow->ibv_attr +
1360                                                sizeof(*flow->ibv_attr));
1361                         unsigned int j;
1362
1363                         if (!flow->mac)
1364                                 continue;
1365                         assert(flow->ibv_attr->type == IBV_FLOW_ATTR_NORMAL);
1366                         assert(flow->ibv_attr->num_of_specs == 1);
1367                         assert(eth->type == IBV_FLOW_SPEC_ETH);
1368                         assert(flow->rss);
1369                         if (rule_vlan &&
1370                             (eth->val.vlan_tag != *rule_vlan ||
1371                              eth->mask.vlan_tag != RTE_BE16(0x0fff)))
1372                                 continue;
1373                         if (!rule_vlan && eth->mask.vlan_tag)
1374                                 continue;
1375                         for (j = 0; j != sizeof(mac->addr_bytes); ++j)
1376                                 if (eth->val.dst_mac[j] != mac->addr_bytes[j] ||
1377                                     eth->mask.dst_mac[j] != UINT8_C(0xff) ||
1378                                     eth->val.src_mac[j] != UINT8_C(0x00) ||
1379                                     eth->mask.src_mac[j] != UINT8_C(0x00))
1380                                         break;
1381                         if (j != sizeof(mac->addr_bytes))
1382                                 continue;
1383                         if (flow->rss->queues != queues ||
1384                             memcmp(flow->rss->queue_id, rss_conf->queue,
1385                                    queues * sizeof(flow->rss->queue_id[0])))
1386                                 continue;
1387                         break;
1388                 }
1389                 if (!flow || !flow->internal) {
1390                         /* Not found, create a new flow rule. */
1391                         memcpy(rule_mac, mac, sizeof(*mac));
1392                         flow = mlx4_flow_create(priv->dev, &attr, pattern,
1393                                                 actions, error);
1394                         if (!flow) {
1395                                 err = -rte_errno;
1396                                 goto error;
1397                         }
1398                 }
1399                 flow->select = 1;
1400                 flow->mac = 1;
1401         }
1402         if (rule_vlan) {
1403                 vlan = mlx4_flow_internal_next_vlan(priv, vlan + 1);
1404                 if (vlan < 4096)
1405                         goto next_vlan;
1406         }
1407         /* Take care of promiscuous and all multicast flow rules. */
1408         if (!broadcast) {
1409                 for (flow = LIST_FIRST(&priv->flows);
1410                      flow && flow->internal;
1411                      flow = LIST_NEXT(flow, next)) {
1412                         if (priv->dev->data->promiscuous) {
1413                                 if (flow->promisc)
1414                                         break;
1415                         } else {
1416                                 assert(priv->dev->data->all_multicast);
1417                                 if (flow->allmulti)
1418                                         break;
1419                         }
1420                 }
1421                 if (flow && flow->internal) {
1422                         assert(flow->rss);
1423                         if (flow->rss->queues != queues ||
1424                             memcmp(flow->rss->queue_id, rss_conf->queue,
1425                                    queues * sizeof(flow->rss->queue_id[0])))
1426                                 flow = NULL;
1427                 }
1428                 if (!flow || !flow->internal) {
1429                         /* Not found, create a new flow rule. */
1430                         if (priv->dev->data->promiscuous) {
1431                                 pattern[1].spec = NULL;
1432                                 pattern[1].mask = NULL;
1433                         } else {
1434                                 assert(priv->dev->data->all_multicast);
1435                                 pattern[1].spec = &eth_allmulti;
1436                                 pattern[1].mask = &eth_allmulti;
1437                         }
1438                         pattern[2] = pattern[3];
1439                         flow = mlx4_flow_create(priv->dev, &attr, pattern,
1440                                                 actions, error);
1441                         if (!flow) {
1442                                 err = -rte_errno;
1443                                 goto error;
1444                         }
1445                 }
1446                 assert(flow->promisc || flow->allmulti);
1447                 flow->select = 1;
1448         }
1449 error:
1450         /* Clear selection and clean up stale internal flow rules. */
1451         flow = LIST_FIRST(&priv->flows);
1452         while (flow && flow->internal) {
1453                 struct rte_flow *next = LIST_NEXT(flow, next);
1454
1455                 if (!flow->select)
1456                         claim_zero(mlx4_flow_destroy(priv->dev, flow, error));
1457                 else
1458                         flow->select = 0;
1459                 flow = next;
1460         }
1461         return err;
1462 }
1463
1464 /**
1465  * Synchronize flow rules.
1466  *
1467  * This function synchronizes flow rules with the state of the device by
1468  * taking into account isolated mode and whether target queues are
1469  * configured.
1470  *
1471  * @param priv
1472  *   Pointer to private structure.
1473  * @param[out] error
1474  *   Perform verbose error reporting if not NULL.
1475  *
1476  * @return
1477  *   0 on success, a negative errno value otherwise and rte_errno is set.
1478  */
1479 int
1480 mlx4_flow_sync(struct priv *priv, struct rte_flow_error *error)
1481 {
1482         struct rte_flow *flow;
1483         int ret;
1484
1485         /* Internal flow rules are guaranteed to come first in the list. */
1486         if (priv->isolated) {
1487                 /*
1488                  * Get rid of them in isolated mode, stop at the first
1489                  * non-internal rule found.
1490                  */
1491                 for (flow = LIST_FIRST(&priv->flows);
1492                      flow && flow->internal;
1493                      flow = LIST_FIRST(&priv->flows))
1494                         claim_zero(mlx4_flow_destroy(priv->dev, flow, error));
1495         } else {
1496                 /* Refresh internal rules. */
1497                 ret = mlx4_flow_internal(priv, error);
1498                 if (ret)
1499                         return ret;
1500         }
1501         /* Toggle the remaining flow rules . */
1502         LIST_FOREACH(flow, &priv->flows, next) {
1503                 ret = mlx4_flow_toggle(priv, flow, priv->started, error);
1504                 if (ret)
1505                         return ret;
1506         }
1507         if (!priv->started)
1508                 assert(!priv->drop);
1509         return 0;
1510 }
1511
1512 /**
1513  * Clean up all flow rules.
1514  *
1515  * Unlike mlx4_flow_flush(), this function takes care of all remaining flow
1516  * rules regardless of whether they are internal or user-configured.
1517  *
1518  * @param priv
1519  *   Pointer to private structure.
1520  */
1521 void
1522 mlx4_flow_clean(struct priv *priv)
1523 {
1524         struct rte_flow *flow;
1525
1526         while ((flow = LIST_FIRST(&priv->flows)))
1527                 mlx4_flow_destroy(priv->dev, flow, NULL);
1528         assert(LIST_EMPTY(&priv->rss));
1529 }
1530
1531 static const struct rte_flow_ops mlx4_flow_ops = {
1532         .validate = mlx4_flow_validate,
1533         .create = mlx4_flow_create,
1534         .destroy = mlx4_flow_destroy,
1535         .flush = mlx4_flow_flush,
1536         .isolate = mlx4_flow_isolate,
1537 };
1538
1539 /**
1540  * Manage filter operations.
1541  *
1542  * @param dev
1543  *   Pointer to Ethernet device structure.
1544  * @param filter_type
1545  *   Filter type.
1546  * @param filter_op
1547  *   Operation to perform.
1548  * @param arg
1549  *   Pointer to operation-specific structure.
1550  *
1551  * @return
1552  *   0 on success, negative errno value otherwise and rte_errno is set.
1553  */
1554 int
1555 mlx4_filter_ctrl(struct rte_eth_dev *dev,
1556                  enum rte_filter_type filter_type,
1557                  enum rte_filter_op filter_op,
1558                  void *arg)
1559 {
1560         switch (filter_type) {
1561         case RTE_ETH_FILTER_GENERIC:
1562                 if (filter_op != RTE_ETH_FILTER_GET)
1563                         break;
1564                 *(const void **)arg = &mlx4_flow_ops;
1565                 return 0;
1566         default:
1567                 ERROR("%p: filter type (%d) not supported",
1568                       (void *)dev, filter_type);
1569                 break;
1570         }
1571         rte_errno = ENOTSUP;
1572         return -rte_errno;
1573 }