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_glue.h"
  69 #include "mlx4_flow.h"
  70 #include "mlx4_rxtx.h"
  71 #include "mlx4_utils.h"
  72
  73 /** Static initializer for a list of subsequent item types. */
  74 #define NEXT_ITEM(...) \
  75         (const enum rte_flow_item_type []){ \
  76                 __VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \
  77         }
  78
  79 /** Processor structure associated with a flow item. */
  80 struct mlx4_flow_proc_item {
  81         /** Bit-mask for fields supported by this PMD. */
  82         const void *mask_support;
  83         /** Bit-mask to use when @p item->mask is not provided. */
  84         const void *mask_default;
  85         /** Size in bytes for @p mask_support and @p mask_default. */
  86         const unsigned int mask_sz;
  87         /** Merge a pattern item into a flow rule handle. */
  88         int (*merge)(struct rte_flow *flow,
  89                      const struct rte_flow_item *item,
  90                      const struct mlx4_flow_proc_item *proc,
  91                      struct rte_flow_error *error);
  92         /** Size in bytes of the destination structure. */
  93         const unsigned int dst_sz;
  94         /** List of possible subsequent items. */
  95         const enum rte_flow_item_type *const next_item;
  96 };
  97
  98 /** Shared resources for drop flow rules. */
  99 struct mlx4_drop {
 100         struct ibv_qp *qp; /**< QP target. */
 101         struct ibv_cq *cq; /**< CQ associated with above QP. */
 102         struct priv *priv; /**< Back pointer to private data. */
 103         uint32_t refcnt; /**< Reference count. */
 104 };
 105
 106 /**
 107  * Convert DPDK RSS hash fields to their Verbs equivalent.
 108  *
 109  * This function returns the supported (default) set when @p rss_hf has
 110  * special value (uint64_t)-1.
 111  *
 112  * @param priv
 113  *   Pointer to private structure.
 114  * @param rss_hf
 115  *   Hash fields in DPDK format (see struct rte_eth_rss_conf).
 116  *
 117  * @return
 118  *   A valid Verbs RSS hash fields mask for mlx4 on success, (uint64_t)-1
 119  *   otherwise and rte_errno is set.
 120  */
 121 uint64_t
 122 mlx4_conv_rss_hf(struct priv *priv, uint64_t rss_hf)
 123 {
 124         enum { IPV4, IPV6, TCP, UDP, };
 125         const uint64_t in[] = {
 126                 [IPV4] = (ETH_RSS_IPV4 |
 127                           ETH_RSS_FRAG_IPV4 |
 128                           ETH_RSS_NONFRAG_IPV4_TCP |
 129                           ETH_RSS_NONFRAG_IPV4_UDP |
 130                           ETH_RSS_NONFRAG_IPV4_OTHER),
 131                 [IPV6] = (ETH_RSS_IPV6 |
 132                           ETH_RSS_FRAG_IPV6 |
 133                           ETH_RSS_NONFRAG_IPV6_TCP |
 134                           ETH_RSS_NONFRAG_IPV6_UDP |
 135                           ETH_RSS_NONFRAG_IPV6_OTHER |
 136                           ETH_RSS_IPV6_EX |
 137                           ETH_RSS_IPV6_TCP_EX |
 138                           ETH_RSS_IPV6_UDP_EX),
 139                 [TCP] = (ETH_RSS_NONFRAG_IPV4_TCP |
 140                          ETH_RSS_NONFRAG_IPV6_TCP |
 141                          ETH_RSS_IPV6_TCP_EX),
 142                 [UDP] = (ETH_RSS_NONFRAG_IPV4_UDP |
 143                          ETH_RSS_NONFRAG_IPV6_UDP |
 144                          ETH_RSS_IPV6_UDP_EX),
 145         };
 146         const uint64_t out[RTE_DIM(in)] = {
 147                 [IPV4] = IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4,
 148                 [IPV6] = IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6,
 149                 [TCP] = IBV_RX_HASH_SRC_PORT_TCP | IBV_RX_HASH_DST_PORT_TCP,
 150                 [UDP] = IBV_RX_HASH_SRC_PORT_UDP | IBV_RX_HASH_DST_PORT_UDP,
 151         };
 152         uint64_t seen = 0;
 153         uint64_t conv = 0;
 154         unsigned int i;
 155
 156         for (i = 0; i != RTE_DIM(in); ++i)
 157                 if (rss_hf & in[i]) {
 158                         seen |= rss_hf & in[i];
 159                         conv |= out[i];
 160                 }
 161         if ((conv & priv->hw_rss_sup) == conv) {
 162                 if (rss_hf == (uint64_t)-1) {
 163                         /* Include inner RSS by default if supported. */
 164                         conv |= priv->hw_rss_sup & IBV_RX_HASH_INNER;
 165                         return conv;
 166                 }
 167                 if (!(rss_hf & ~seen))
 168                         return conv;
 169         }
 170         rte_errno = ENOTSUP;
 171         return (uint64_t)-1;
 172 }
 173
 174 /**
 175  * Merge Ethernet pattern item into flow rule handle.
 176  *
 177  * Additional mlx4-specific constraints on supported fields:
 178  *
 179  * - No support for partial masks, except in the specific case of matching
 180  *   all multicast traffic (@p spec->dst and @p mask->dst equal to
 181  *   01:00:00:00:00:00).
 182  * - Not providing @p item->spec or providing an empty @p mask->dst is
 183  *   *only* supported if the rule doesn't specify additional matching
 184  *   criteria (i.e. rule is promiscuous-like).
 185  *
 186  * @param[in, out] flow
 187  *   Flow rule handle to update.
 188  * @param[in] item
 189  *   Pattern item to merge.
 190  * @param[in] proc
 191  *   Associated item-processing object.
 192  * @param[out] error
 193  *   Perform verbose error reporting if not NULL.
 194  *
 195  * @return
 196  *   0 on success, a negative errno value otherwise and rte_errno is set.
 197  */
 198 static int
 199 mlx4_flow_merge_eth(struct rte_flow *flow,
 200                     const struct rte_flow_item *item,
 201                     const struct mlx4_flow_proc_item *proc,
 202                     struct rte_flow_error *error)
 203 {
 204         const struct rte_flow_item_eth *spec = item->spec;
 205         const struct rte_flow_item_eth *mask =
 206                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 207         struct ibv_flow_spec_eth *eth;
 208         const char *msg;
 209         unsigned int i;
 210
 211         if (!mask) {
 212                 flow->promisc = 1;
 213         } else {
 214                 uint32_t sum_dst = 0;
 215                 uint32_t sum_src = 0;
 216
 217                 for (i = 0; i != sizeof(mask->dst.addr_bytes); ++i) {
 218                         sum_dst += mask->dst.addr_bytes[i];
 219                         sum_src += mask->src.addr_bytes[i];
 220                 }
 221                 if (sum_src) {
 222                         msg = "mlx4 does not support source MAC matching";
 223                         goto error;
 224                 } else if (!sum_dst) {
 225                         flow->promisc = 1;
 226                 } else if (sum_dst == 1 && mask->dst.addr_bytes[0] == 1) {
 227                         if (!(spec->dst.addr_bytes[0] & 1)) {
 228                                 msg = "mlx4 does not support the explicit"
 229                                         " exclusion of all multicast traffic";
 230                                 goto error;
 231                         }
 232                         flow->allmulti = 1;
 233                 } else if (sum_dst != (UINT8_C(0xff) * ETHER_ADDR_LEN)) {
 234                         msg = "mlx4 does not support matching partial"
 235                                 " Ethernet fields";
 236                         goto error;
 237                 }
 238         }
 239         if (!flow->ibv_attr)
 240                 return 0;
 241         if (flow->promisc) {
 242                 flow->ibv_attr->type = IBV_FLOW_ATTR_ALL_DEFAULT;
 243                 return 0;
 244         }
 245         if (flow->allmulti) {
 246                 flow->ibv_attr->type = IBV_FLOW_ATTR_MC_DEFAULT;
 247                 return 0;
 248         }
 249         ++flow->ibv_attr->num_of_specs;
 250         eth = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 251         *eth = (struct ibv_flow_spec_eth) {
 252                 .type = IBV_FLOW_SPEC_ETH,
 253                 .size = sizeof(*eth),
 254         };
 255         memcpy(eth->val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
 256         memcpy(eth->mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
 257         /* Remove unwanted bits from values. */
 258         for (i = 0; i < ETHER_ADDR_LEN; ++i) {
 259                 eth->val.dst_mac[i] &= eth->mask.dst_mac[i];
 260         }
 261         return 0;
 262 error:
 263         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 264                                   item, msg);
 265 }
 266
 267 /**
 268  * Merge VLAN pattern item into flow rule handle.
 269  *
 270  * Additional mlx4-specific constraints on supported fields:
 271  *
 272  * - Matching *all* VLAN traffic by omitting @p item->spec or providing an
 273  *   empty @p item->mask would also include non-VLAN traffic. Doing so is
 274  *   therefore unsupported.
 275  * - No support for partial masks.
 276  *
 277  * @param[in, out] flow
 278  *   Flow rule handle to update.
 279  * @param[in] item
 280  *   Pattern item to merge.
 281  * @param[in] proc
 282  *   Associated item-processing object.
 283  * @param[out] error
 284  *   Perform verbose error reporting if not NULL.
 285  *
 286  * @return
 287  *   0 on success, a negative errno value otherwise and rte_errno is set.
 288  */
 289 static int
 290 mlx4_flow_merge_vlan(struct rte_flow *flow,
 291                      const struct rte_flow_item *item,
 292                      const struct mlx4_flow_proc_item *proc,
 293                      struct rte_flow_error *error)
 294 {
 295         const struct rte_flow_item_vlan *spec = item->spec;
 296         const struct rte_flow_item_vlan *mask =
 297                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 298         struct ibv_flow_spec_eth *eth;
 299         const char *msg;
 300
 301         if (!mask || !mask->tci) {
 302                 msg = "mlx4 cannot match all VLAN traffic while excluding"
 303                         " non-VLAN traffic, TCI VID must be specified";
 304                 goto error;
 305         }
 306         if (mask->tci != RTE_BE16(0x0fff)) {
 307                 msg = "mlx4 does not support partial TCI VID matching";
 308                 goto error;
 309         }
 310         if (!flow->ibv_attr)
 311                 return 0;
 312         eth = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size -
 313                        sizeof(*eth));
 314         eth->val.vlan_tag = spec->tci;
 315         eth->mask.vlan_tag = mask->tci;
 316         eth->val.vlan_tag &= eth->mask.vlan_tag;
 317         return 0;
 318 error:
 319         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 320                                   item, msg);
 321 }
 322
 323 /**
 324  * Merge IPv4 pattern item into flow rule handle.
 325  *
 326  * Additional mlx4-specific constraints on supported fields:
 327  *
 328  * - No support for partial masks.
 329  *
 330  * @param[in, out] flow
 331  *   Flow rule handle to update.
 332  * @param[in] item
 333  *   Pattern item to merge.
 334  * @param[in] proc
 335  *   Associated item-processing object.
 336  * @param[out] error
 337  *   Perform verbose error reporting if not NULL.
 338  *
 339  * @return
 340  *   0 on success, a negative errno value otherwise and rte_errno is set.
 341  */
 342 static int
 343 mlx4_flow_merge_ipv4(struct rte_flow *flow,
 344                      const struct rte_flow_item *item,
 345                      const struct mlx4_flow_proc_item *proc,
 346                      struct rte_flow_error *error)
 347 {
 348         const struct rte_flow_item_ipv4 *spec = item->spec;
 349         const struct rte_flow_item_ipv4 *mask =
 350                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 351         struct ibv_flow_spec_ipv4 *ipv4;
 352         const char *msg;
 353
 354         if (mask &&
 355             ((uint32_t)(mask->hdr.src_addr + 1) > UINT32_C(1) ||
 356              (uint32_t)(mask->hdr.dst_addr + 1) > UINT32_C(1))) {
 357                 msg = "mlx4 does not support matching partial IPv4 fields";
 358                 goto error;
 359         }
 360         if (!flow->ibv_attr)
 361                 return 0;
 362         ++flow->ibv_attr->num_of_specs;
 363         ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 364         *ipv4 = (struct ibv_flow_spec_ipv4) {
 365                 .type = IBV_FLOW_SPEC_IPV4,
 366                 .size = sizeof(*ipv4),
 367         };
 368         if (!spec)
 369                 return 0;
 370         ipv4->val = (struct ibv_flow_ipv4_filter) {
 371                 .src_ip = spec->hdr.src_addr,
 372                 .dst_ip = spec->hdr.dst_addr,
 373         };
 374         ipv4->mask = (struct ibv_flow_ipv4_filter) {
 375                 .src_ip = mask->hdr.src_addr,
 376                 .dst_ip = mask->hdr.dst_addr,
 377         };
 378         /* Remove unwanted bits from values. */
 379         ipv4->val.src_ip &= ipv4->mask.src_ip;
 380         ipv4->val.dst_ip &= ipv4->mask.dst_ip;
 381         return 0;
 382 error:
 383         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 384                                   item, msg);
 385 }
 386
 387 /**
 388  * Merge UDP pattern item into flow rule handle.
 389  *
 390  * Additional mlx4-specific constraints on supported fields:
 391  *
 392  * - No support for partial masks.
 393  *
 394  * @param[in, out] flow
 395  *   Flow rule handle to update.
 396  * @param[in] item
 397  *   Pattern item to merge.
 398  * @param[in] proc
 399  *   Associated item-processing object.
 400  * @param[out] error
 401  *   Perform verbose error reporting if not NULL.
 402  *
 403  * @return
 404  *   0 on success, a negative errno value otherwise and rte_errno is set.
 405  */
 406 static int
 407 mlx4_flow_merge_udp(struct rte_flow *flow,
 408                     const struct rte_flow_item *item,
 409                     const struct mlx4_flow_proc_item *proc,
 410                     struct rte_flow_error *error)
 411 {
 412         const struct rte_flow_item_udp *spec = item->spec;
 413         const struct rte_flow_item_udp *mask =
 414                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 415         struct ibv_flow_spec_tcp_udp *udp;
 416         const char *msg;
 417
 418         if (mask &&
 419             ((uint16_t)(mask->hdr.src_port + 1) > UINT16_C(1) ||
 420              (uint16_t)(mask->hdr.dst_port + 1) > UINT16_C(1))) {
 421                 msg = "mlx4 does not support matching partial UDP fields";
 422                 goto error;
 423         }
 424         if (!flow->ibv_attr)
 425                 return 0;
 426         ++flow->ibv_attr->num_of_specs;
 427         udp = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 428         *udp = (struct ibv_flow_spec_tcp_udp) {
 429                 .type = IBV_FLOW_SPEC_UDP,
 430                 .size = sizeof(*udp),
 431         };
 432         if (!spec)
 433                 return 0;
 434         udp->val.dst_port = spec->hdr.dst_port;
 435         udp->val.src_port = spec->hdr.src_port;
 436         udp->mask.dst_port = mask->hdr.dst_port;
 437         udp->mask.src_port = mask->hdr.src_port;
 438         /* Remove unwanted bits from values. */
 439         udp->val.src_port &= udp->mask.src_port;
 440         udp->val.dst_port &= udp->mask.dst_port;
 441         return 0;
 442 error:
 443         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 444                                   item, msg);
 445 }
 446
 447 /**
 448  * Merge TCP pattern item into flow rule handle.
 449  *
 450  * Additional mlx4-specific constraints on supported fields:
 451  *
 452  * - No support for partial masks.
 453  *
 454  * @param[in, out] flow
 455  *   Flow rule handle to update.
 456  * @param[in] item
 457  *   Pattern item to merge.
 458  * @param[in] proc
 459  *   Associated item-processing object.
 460  * @param[out] error
 461  *   Perform verbose error reporting if not NULL.
 462  *
 463  * @return
 464  *   0 on success, a negative errno value otherwise and rte_errno is set.
 465  */
 466 static int
 467 mlx4_flow_merge_tcp(struct rte_flow *flow,
 468                     const struct rte_flow_item *item,
 469                     const struct mlx4_flow_proc_item *proc,
 470                     struct rte_flow_error *error)
 471 {
 472         const struct rte_flow_item_tcp *spec = item->spec;
 473         const struct rte_flow_item_tcp *mask =
 474                 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
 475         struct ibv_flow_spec_tcp_udp *tcp;
 476         const char *msg;
 477
 478         if (mask &&
 479             ((uint16_t)(mask->hdr.src_port + 1) > UINT16_C(1) ||
 480              (uint16_t)(mask->hdr.dst_port + 1) > UINT16_C(1))) {
 481                 msg = "mlx4 does not support matching partial TCP fields";
 482                 goto error;
 483         }
 484         if (!flow->ibv_attr)
 485                 return 0;
 486         ++flow->ibv_attr->num_of_specs;
 487         tcp = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
 488         *tcp = (struct ibv_flow_spec_tcp_udp) {
 489                 .type = IBV_FLOW_SPEC_TCP,
 490                 .size = sizeof(*tcp),
 491         };
 492         if (!spec)
 493                 return 0;
 494         tcp->val.dst_port = spec->hdr.dst_port;
 495         tcp->val.src_port = spec->hdr.src_port;
 496         tcp->mask.dst_port = mask->hdr.dst_port;
 497         tcp->mask.src_port = mask->hdr.src_port;
 498         /* Remove unwanted bits from values. */
 499         tcp->val.src_port &= tcp->mask.src_port;
 500         tcp->val.dst_port &= tcp->mask.dst_port;
 501         return 0;
 502 error:
 503         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 504                                   item, msg);
 505 }
 506
 507 /**
 508  * Perform basic sanity checks on a pattern item.
 509  *
 510  * @param[in] item
 511  *   Item specification.
 512  * @param[in] proc
 513  *   Associated item-processing object.
 514  * @param[out] error
 515  *   Perform verbose error reporting if not NULL.
 516  *
 517  * @return
 518  *   0 on success, a negative errno value otherwise and rte_errno is set.
 519  */
 520 static int
 521 mlx4_flow_item_check(const struct rte_flow_item *item,
 522                      const struct mlx4_flow_proc_item *proc,
 523                      struct rte_flow_error *error)
 524 {
 525         const uint8_t *mask;
 526         unsigned int i;
 527
 528         /* item->last and item->mask cannot exist without item->spec. */
 529         if (!item->spec && (item->mask || item->last))
 530                 return rte_flow_error_set
 531                         (error, EINVAL, RTE_FLOW_ERROR_TYPE_ITEM, item,
 532                          "\"mask\" or \"last\" field provided without a"
 533                          " corresponding \"spec\"");
 534         /* No spec, no mask, no problem. */
 535         if (!item->spec)
 536                 return 0;
 537         mask = item->mask ?
 538                 (const uint8_t *)item->mask :
 539                 (const uint8_t *)proc->mask_default;
 540         assert(mask);
 541         /*
 542          * Single-pass check to make sure that:
 543          * - Mask is supported, no bits are set outside proc->mask_support.
 544          * - Both item->spec and item->last are included in mask.
 545          */
 546         for (i = 0; i != proc->mask_sz; ++i) {
 547                 if (!mask[i])
 548                         continue;
 549                 if ((mask[i] | ((const uint8_t *)proc->mask_support)[i]) !=
 550                     ((const uint8_t *)proc->mask_support)[i])
 551                         return rte_flow_error_set
 552                                 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 553                                  item, "unsupported field found in \"mask\"");
 554                 if (item->last &&
 555                     (((const uint8_t *)item->spec)[i] & mask[i]) !=
 556                     (((const uint8_t *)item->last)[i] & mask[i]))
 557                         return rte_flow_error_set
 558                                 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 559                                  item,
 560                                  "range between \"spec\" and \"last\""
 561                                  " is larger than \"mask\"");
 562         }
 563         return 0;
 564 }
 565
 566 /** Graph of supported items and associated actions. */
 567 static const struct mlx4_flow_proc_item mlx4_flow_proc_item_list[] = {
 568         [RTE_FLOW_ITEM_TYPE_END] = {
 569                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_ETH),
 570         },
 571         [RTE_FLOW_ITEM_TYPE_ETH] = {
 572                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_VLAN,
 573                                        RTE_FLOW_ITEM_TYPE_IPV4),
 574                 .mask_support = &(const struct rte_flow_item_eth){
 575                         /* Only destination MAC can be matched. */
 576                         .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 577                 },
 578                 .mask_default = &rte_flow_item_eth_mask,
 579                 .mask_sz = sizeof(struct rte_flow_item_eth),
 580                 .merge = mlx4_flow_merge_eth,
 581                 .dst_sz = sizeof(struct ibv_flow_spec_eth),
 582         },
 583         [RTE_FLOW_ITEM_TYPE_VLAN] = {
 584                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_IPV4),
 585                 .mask_support = &(const struct rte_flow_item_vlan){
 586                         /* Only TCI VID matching is supported. */
 587                         .tci = RTE_BE16(0x0fff),
 588                 },
 589                 .mask_default = &rte_flow_item_vlan_mask,
 590                 .mask_sz = sizeof(struct rte_flow_item_vlan),
 591                 .merge = mlx4_flow_merge_vlan,
 592                 .dst_sz = 0,
 593         },
 594         [RTE_FLOW_ITEM_TYPE_IPV4] = {
 595                 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_UDP,
 596                                        RTE_FLOW_ITEM_TYPE_TCP),
 597                 .mask_support = &(const struct rte_flow_item_ipv4){
 598                         .hdr = {
 599                                 .src_addr = RTE_BE32(0xffffffff),
 600                                 .dst_addr = RTE_BE32(0xffffffff),
 601                         },
 602                 },
 603                 .mask_default = &rte_flow_item_ipv4_mask,
 604                 .mask_sz = sizeof(struct rte_flow_item_ipv4),
 605                 .merge = mlx4_flow_merge_ipv4,
 606                 .dst_sz = sizeof(struct ibv_flow_spec_ipv4),
 607         },
 608         [RTE_FLOW_ITEM_TYPE_UDP] = {
 609                 .mask_support = &(const struct rte_flow_item_udp){
 610                         .hdr = {
 611                                 .src_port = RTE_BE16(0xffff),
 612                                 .dst_port = RTE_BE16(0xffff),
 613                         },
 614                 },
 615                 .mask_default = &rte_flow_item_udp_mask,
 616                 .mask_sz = sizeof(struct rte_flow_item_udp),
 617                 .merge = mlx4_flow_merge_udp,
 618                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 619         },
 620         [RTE_FLOW_ITEM_TYPE_TCP] = {
 621                 .mask_support = &(const struct rte_flow_item_tcp){
 622                         .hdr = {
 623                                 .src_port = RTE_BE16(0xffff),
 624                                 .dst_port = RTE_BE16(0xffff),
 625                         },
 626                 },
 627                 .mask_default = &rte_flow_item_tcp_mask,
 628                 .mask_sz = sizeof(struct rte_flow_item_tcp),
 629                 .merge = mlx4_flow_merge_tcp,
 630                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 631         },
 632 };
 633
 634 /**
 635  * Make sure a flow rule is supported and initialize associated structure.
 636  *
 637  * @param priv
 638  *   Pointer to private structure.
 639  * @param[in] attr
 640  *   Flow rule attributes.
 641  * @param[in] pattern
 642  *   Pattern specification (list terminated by the END pattern item).
 643  * @param[in] actions
 644  *   Associated actions (list terminated by the END action).
 645  * @param[out] error
 646  *   Perform verbose error reporting if not NULL.
 647  * @param[in, out] addr
 648  *   Buffer where the resulting flow rule handle pointer must be stored.
 649  *   If NULL, stop processing after validation stage.
 650  *
 651  * @return
 652  *   0 on success, a negative errno value otherwise and rte_errno is set.
 653  */
 654 static int
 655 mlx4_flow_prepare(struct priv *priv,
 656                   const struct rte_flow_attr *attr,
 657                   const struct rte_flow_item pattern[],
 658                   const struct rte_flow_action actions[],
 659                   struct rte_flow_error *error,
 660                   struct rte_flow **addr)
 661 {
 662         const struct rte_flow_item *item;
 663         const struct rte_flow_action *action;
 664         const struct mlx4_flow_proc_item *proc;
 665         struct rte_flow temp = { .ibv_attr_size = sizeof(*temp.ibv_attr) };
 666         struct rte_flow *flow = &temp;
 667         const char *msg = NULL;
 668
 669         if (attr->group)
 670                 return rte_flow_error_set
 671                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 672                          NULL, "groups are not supported");
 673         if (attr->priority > MLX4_FLOW_PRIORITY_LAST)
 674                 return rte_flow_error_set
 675                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 676                          NULL, "maximum priority level is "
 677                          MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST));
 678         if (attr->egress)
 679                 return rte_flow_error_set
 680                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 681                          NULL, "egress is not supported");
 682         if (!attr->ingress)
 683                 return rte_flow_error_set
 684                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 685                          NULL, "only ingress is supported");
 686 fill:
 687         proc = mlx4_flow_proc_item_list;
 688         /* Go over pattern. */
 689         for (item = pattern; item->type; ++item) {
 690                 const struct mlx4_flow_proc_item *next = NULL;
 691                 unsigned int i;
 692                 int err;
 693
 694                 if (item->type == RTE_FLOW_ITEM_TYPE_VOID)
 695                         continue;
 696                 if (item->type == MLX4_FLOW_ITEM_TYPE_INTERNAL) {
 697                         flow->internal = 1;
 698                         continue;
 699                 }
 700                 if (flow->promisc || flow->allmulti) {
 701                         msg = "mlx4 does not support additional matching"
 702                                 " criteria combined with indiscriminate"
 703                                 " matching on Ethernet headers";
 704                         goto exit_item_not_supported;
 705                 }
 706                 for (i = 0; proc->next_item && proc->next_item[i]; ++i) {
 707                         if (proc->next_item[i] == item->type) {
 708                                 next = &mlx4_flow_proc_item_list[item->type];
 709                                 break;
 710                         }
 711                 }
 712                 if (!next)
 713                         goto exit_item_not_supported;
 714                 proc = next;
 715                 /*
 716                  * Perform basic sanity checks only once, while handle is
 717                  * not allocated.
 718                  */
 719                 if (flow == &temp) {
 720                         err = mlx4_flow_item_check(item, proc, error);
 721                         if (err)
 722                                 return err;
 723                 }
 724                 if (proc->merge) {
 725                         err = proc->merge(flow, item, proc, error);
 726                         if (err)
 727                                 return err;
 728                 }
 729                 flow->ibv_attr_size += proc->dst_sz;
 730         }
 731         /* Go over actions list. */
 732         for (action = actions; action->type; ++action) {
 733                 switch (action->type) {
 734                         const struct rte_flow_action_queue *queue;
 735                         const struct rte_flow_action_rss *rss;
 736                         const struct rte_eth_rss_conf *rss_conf;
 737                         unsigned int i;
 738
 739                 case RTE_FLOW_ACTION_TYPE_VOID:
 740                         continue;
 741                 case RTE_FLOW_ACTION_TYPE_DROP:
 742                         flow->drop = 1;
 743                         break;
 744                 case RTE_FLOW_ACTION_TYPE_QUEUE:
 745                         if (flow->rss)
 746                                 break;
 747                         queue = action->conf;
 748                         if (queue->index >= priv->dev->data->nb_rx_queues) {
 749                                 msg = "queue target index beyond number of"
 750                                         " configured Rx queues";
 751                                 goto exit_action_not_supported;
 752                         }
 753                         flow->rss = mlx4_rss_get
 754                                 (priv, 0, mlx4_rss_hash_key_default, 1,
 755                                  &queue->index);
 756                         if (!flow->rss) {
 757                                 msg = "not enough resources for additional"
 758                                         " single-queue RSS context";
 759                                 goto exit_action_not_supported;
 760                         }
 761                         break;
 762                 case RTE_FLOW_ACTION_TYPE_RSS:
 763                         if (flow->rss)
 764                                 break;
 765                         rss = action->conf;
 766                         /* Default RSS configuration if none is provided. */
 767                         rss_conf =
 768                                 rss->rss_conf ?
 769                                 rss->rss_conf :
 770                                 &(struct rte_eth_rss_conf){
 771                                         .rss_key = mlx4_rss_hash_key_default,
 772                                         .rss_key_len = MLX4_RSS_HASH_KEY_SIZE,
 773                                         .rss_hf = -1,
 774                                 };
 775                         /* Sanity checks. */
 776                         for (i = 0; i < rss->num; ++i)
 777                                 if (rss->queue[i] >=
 778                                     priv->dev->data->nb_rx_queues)
 779                                         break;
 780                         if (i != rss->num) {
 781                                 msg = "queue index target beyond number of"
 782                                         " configured Rx queues";
 783                                 goto exit_action_not_supported;
 784                         }
 785                         if (!rte_is_power_of_2(rss->num)) {
 786                                 msg = "for RSS, mlx4 requires the number of"
 787                                         " queues to be a power of two";
 788                                 goto exit_action_not_supported;
 789                         }
 790                         if (rss_conf->rss_key_len !=
 791                             sizeof(flow->rss->key)) {
 792                                 msg = "mlx4 supports exactly one RSS hash key"
 793                                         " length: "
 794                                         MLX4_STR_EXPAND(MLX4_RSS_HASH_KEY_SIZE);
 795                                 goto exit_action_not_supported;
 796                         }
 797                         for (i = 1; i < rss->num; ++i)
 798                                 if (rss->queue[i] - rss->queue[i - 1] != 1)
 799                                         break;
 800                         if (i != rss->num) {
 801                                 msg = "mlx4 requires RSS contexts to use"
 802                                         " consecutive queue indices only";
 803                                 goto exit_action_not_supported;
 804                         }
 805                         if (rss->queue[0] % rss->num) {
 806                                 msg = "mlx4 requires the first queue of a RSS"
 807                                         " context to be aligned on a multiple"
 808                                         " of the context size";
 809                                 goto exit_action_not_supported;
 810                         }
 811                         flow->rss = mlx4_rss_get
 812                                 (priv,
 813                                  mlx4_conv_rss_hf(priv, rss_conf->rss_hf),
 814                                  rss_conf->rss_key, rss->num, rss->queue);
 815                         if (!flow->rss) {
 816                                 msg = "either invalid parameters or not enough"
 817                                         " resources for additional multi-queue"
 818                                         " RSS context";
 819                                 goto exit_action_not_supported;
 820                         }
 821                         break;
 822                 default:
 823                         goto exit_action_not_supported;
 824                 }
 825         }
 826         if (!flow->rss && !flow->drop)
 827                 return rte_flow_error_set
 828                         (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 829                          NULL, "no valid action");
 830         /* Validation ends here. */
 831         if (!addr) {
 832                 if (flow->rss)
 833                         mlx4_rss_put(flow->rss);
 834                 return 0;
 835         }
 836         if (flow == &temp) {
 837                 /* Allocate proper handle based on collected data. */
 838                 const struct mlx4_malloc_vec vec[] = {
 839                         {
 840                                 .align = alignof(struct rte_flow),
 841                                 .size = sizeof(*flow),
 842                                 .addr = (void **)&flow,
 843                         },
 844                         {
 845                                 .align = alignof(struct ibv_flow_attr),
 846                                 .size = temp.ibv_attr_size,
 847                                 .addr = (void **)&temp.ibv_attr,
 848                         },
 849                 };
 850
 851                 if (!mlx4_zmallocv(__func__, vec, RTE_DIM(vec)))
 852                         return rte_flow_error_set
 853                                 (error, -rte_errno,
 854                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
 855                                  "flow rule handle allocation failure");
 856                 /* Most fields will be updated by second pass. */
 857                 *flow = (struct rte_flow){
 858                         .ibv_attr = temp.ibv_attr,
 859                         .ibv_attr_size = sizeof(*flow->ibv_attr),
 860                         .rss = temp.rss,
 861                 };
 862                 *flow->ibv_attr = (struct ibv_flow_attr){
 863                         .type = IBV_FLOW_ATTR_NORMAL,
 864                         .size = sizeof(*flow->ibv_attr),
 865                         .priority = attr->priority,
 866                         .port = priv->port,
 867                 };
 868                 goto fill;
 869         }
 870         *addr = flow;
 871         return 0;
 872 exit_item_not_supported:
 873         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 874                                   item, msg ? msg : "item not supported");
 875 exit_action_not_supported:
 876         return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 877                                   action, msg ? msg : "action not supported");
 878 }
 879
 880 /**
 881  * Validate a flow supported by the NIC.
 882  *
 883  * @see rte_flow_validate()
 884  * @see rte_flow_ops
 885  */
 886 static int
 887 mlx4_flow_validate(struct rte_eth_dev *dev,
 888                    const struct rte_flow_attr *attr,
 889                    const struct rte_flow_item pattern[],
 890                    const struct rte_flow_action actions[],
 891                    struct rte_flow_error *error)
 892 {
 893         struct priv *priv = dev->data->dev_private;
 894
 895         return mlx4_flow_prepare(priv, attr, pattern, actions, error, NULL);
 896 }
 897
 898 /**
 899  * Get a drop flow rule resources instance.
 900  *
 901  * @param priv
 902  *   Pointer to private structure.
 903  *
 904  * @return
 905  *   Pointer to drop flow resources on success, NULL otherwise and rte_errno
 906  *   is set.
 907  */
 908 static struct mlx4_drop *
 909 mlx4_drop_get(struct priv *priv)
 910 {
 911         struct mlx4_drop *drop = priv->drop;
 912
 913         if (drop) {
 914                 assert(drop->refcnt);
 915                 assert(drop->priv == priv);
 916                 ++drop->refcnt;
 917                 return drop;
 918         }
 919         drop = rte_malloc(__func__, sizeof(*drop), 0);
 920         if (!drop)
 921                 goto error;
 922         *drop = (struct mlx4_drop){
 923                 .priv = priv,
 924                 .refcnt = 1,
 925         };
 926         drop->cq = mlx4_glue->create_cq(priv->ctx, 1, NULL, NULL, 0);
 927         if (!drop->cq)
 928                 goto error;
 929         drop->qp = mlx4_glue->create_qp
 930                 (priv->pd,
 931                  &(struct ibv_qp_init_attr){
 932                         .send_cq = drop->cq,
 933                         .recv_cq = drop->cq,
 934                         .qp_type = IBV_QPT_RAW_PACKET,
 935                  });
 936         if (!drop->qp)
 937                 goto error;
 938         priv->drop = drop;
 939         return drop;
 940 error:
 941         if (drop->qp)
 942                 claim_zero(mlx4_glue->destroy_qp(drop->qp));
 943         if (drop->cq)
 944                 claim_zero(mlx4_glue->destroy_cq(drop->cq));
 945         if (drop)
 946                 rte_free(drop);
 947         rte_errno = ENOMEM;
 948         return NULL;
 949 }
 950
 951 /**
 952  * Give back a drop flow rule resources instance.
 953  *
 954  * @param drop
 955  *   Pointer to drop flow rule resources.
 956  */
 957 static void
 958 mlx4_drop_put(struct mlx4_drop *drop)
 959 {
 960         assert(drop->refcnt);
 961         if (--drop->refcnt)
 962                 return;
 963         drop->priv->drop = NULL;
 964         claim_zero(mlx4_glue->destroy_qp(drop->qp));
 965         claim_zero(mlx4_glue->destroy_cq(drop->cq));
 966         rte_free(drop);
 967 }
 968
 969 /**
 970  * Toggle a configured flow rule.
 971  *
 972  * @param priv
 973  *   Pointer to private structure.
 974  * @param flow
 975  *   Flow rule handle to toggle.
 976  * @param enable
 977  *   Whether associated Verbs flow must be created or removed.
 978  * @param[out] error
 979  *   Perform verbose error reporting if not NULL.
 980  *
 981  * @return
 982  *   0 on success, a negative errno value otherwise and rte_errno is set.
 983  */
 984 static int
 985 mlx4_flow_toggle(struct priv *priv,
 986                  struct rte_flow *flow,
 987                  int enable,
 988                  struct rte_flow_error *error)
 989 {
 990         struct ibv_qp *qp = NULL;
 991         const char *msg;
 992         int err;
 993
 994         if (!enable) {
 995                 if (!flow->ibv_flow)
 996                         return 0;
 997                 claim_zero(mlx4_glue->destroy_flow(flow->ibv_flow));
 998                 flow->ibv_flow = NULL;
 999                 if (flow->drop)
1000                         mlx4_drop_put(priv->drop);
1001                 else if (flow->rss)
1002                         mlx4_rss_detach(flow->rss);
1003                 return 0;
1004         }
1005         assert(flow->ibv_attr);
1006         if (!flow->internal &&
1007             !priv->isolated &&
1008             flow->ibv_attr->priority == MLX4_FLOW_PRIORITY_LAST) {
1009                 if (flow->ibv_flow) {
1010                         claim_zero(mlx4_glue->destroy_flow(flow->ibv_flow));
1011                         flow->ibv_flow = NULL;
1012                         if (flow->drop)
1013                                 mlx4_drop_put(priv->drop);
1014                         else if (flow->rss)
1015                                 mlx4_rss_detach(flow->rss);
1016                 }
1017                 err = EACCES;
1018                 msg = ("priority level "
1019                        MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST)
1020                        " is reserved when not in isolated mode");
1021                 goto error;
1022         }
1023         if (flow->rss) {
1024                 struct mlx4_rss *rss = flow->rss;
1025                 int missing = 0;
1026                 unsigned int i;
1027
1028                 /* Stop at the first nonexistent target queue. */
1029                 for (i = 0; i != rss->queues; ++i)
1030                         if (rss->queue_id[i] >=
1031                             priv->dev->data->nb_rx_queues ||
1032                             !priv->dev->data->rx_queues[rss->queue_id[i]]) {
1033                                 missing = 1;
1034                                 break;
1035                         }
1036                 if (flow->ibv_flow) {
1037                         if (missing ^ !flow->drop)
1038                                 return 0;
1039                         /* Verbs flow needs updating. */
1040                         claim_zero(mlx4_glue->destroy_flow(flow->ibv_flow));
1041                         flow->ibv_flow = NULL;
1042                         if (flow->drop)
1043                                 mlx4_drop_put(priv->drop);
1044                         else
1045                                 mlx4_rss_detach(rss);
1046                 }
1047                 if (!missing) {
1048                         err = mlx4_rss_attach(rss);
1049                         if (err) {
1050                                 err = -err;
1051                                 msg = "cannot create indirection table or hash"
1052                                         " QP to associate flow rule with";
1053                                 goto error;
1054                         }
1055                         qp = rss->qp;
1056                 }
1057                 /* A missing target queue drops traffic implicitly. */
1058                 flow->drop = missing;
1059         }
1060         if (flow->drop) {
1061                 mlx4_drop_get(priv);
1062                 if (!priv->drop) {
1063                         err = rte_errno;
1064                         msg = "resources for drop flow rule cannot be created";
1065                         goto error;
1066                 }
1067                 qp = priv->drop->qp;
1068         }
1069         assert(qp);
1070         if (flow->ibv_flow)
1071                 return 0;
1072         flow->ibv_flow = mlx4_glue->create_flow(qp, flow->ibv_attr);
1073         if (flow->ibv_flow)
1074                 return 0;
1075         if (flow->drop)
1076                 mlx4_drop_put(priv->drop);
1077         else if (flow->rss)
1078                 mlx4_rss_detach(flow->rss);
1079         err = errno;
1080         msg = "flow rule rejected by device";
1081 error:
1082         return rte_flow_error_set
1083                 (error, err, RTE_FLOW_ERROR_TYPE_HANDLE, flow, msg);
1084 }
1085
1086 /**
1087  * Create a flow.
1088  *
1089  * @see rte_flow_create()
1090  * @see rte_flow_ops
1091  */
1092 static struct rte_flow *
1093 mlx4_flow_create(struct rte_eth_dev *dev,
1094                  const struct rte_flow_attr *attr,
1095                  const struct rte_flow_item pattern[],
1096                  const struct rte_flow_action actions[],
1097                  struct rte_flow_error *error)
1098 {
1099         struct priv *priv = dev->data->dev_private;
1100         struct rte_flow *flow;
1101         int err;
1102
1103         err = mlx4_flow_prepare(priv, attr, pattern, actions, error, &flow);
1104         if (err)
1105                 return NULL;
1106         err = mlx4_flow_toggle(priv, flow, priv->started, error);
1107         if (!err) {
1108                 struct rte_flow *curr = LIST_FIRST(&priv->flows);
1109
1110                 /* New rules are inserted after internal ones. */
1111                 if (!curr || !curr->internal) {
1112                         LIST_INSERT_HEAD(&priv->flows, flow, next);
1113                 } else {
1114                         while (LIST_NEXT(curr, next) &&
1115                                LIST_NEXT(curr, next)->internal)
1116                                 curr = LIST_NEXT(curr, next);
1117                         LIST_INSERT_AFTER(curr, flow, next);
1118                 }
1119                 return flow;
1120         }
1121         if (flow->rss)
1122                 mlx4_rss_put(flow->rss);
1123         rte_flow_error_set(error, -err, RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1124                            error->message);
1125         rte_free(flow);
1126         return NULL;
1127 }
1128
1129 /**
1130  * Configure isolated mode.
1131  *
1132  * @see rte_flow_isolate()
1133  * @see rte_flow_ops
1134  */
1135 static int
1136 mlx4_flow_isolate(struct rte_eth_dev *dev,
1137                   int enable,
1138                   struct rte_flow_error *error)
1139 {
1140         struct priv *priv = dev->data->dev_private;
1141
1142         if (!!enable == !!priv->isolated)
1143                 return 0;
1144         priv->isolated = !!enable;
1145         if (mlx4_flow_sync(priv, error)) {
1146                 priv->isolated = !enable;
1147                 return -rte_errno;
1148         }
1149         return 0;
1150 }
1151
1152 /**
1153  * Destroy a flow rule.
1154  *
1155  * @see rte_flow_destroy()
1156  * @see rte_flow_ops
1157  */
1158 static int
1159 mlx4_flow_destroy(struct rte_eth_dev *dev,
1160                   struct rte_flow *flow,
1161                   struct rte_flow_error *error)
1162 {
1163         struct priv *priv = dev->data->dev_private;
1164         int err = mlx4_flow_toggle(priv, flow, 0, error);
1165
1166         if (err)
1167                 return err;
1168         LIST_REMOVE(flow, next);
1169         if (flow->rss)
1170                 mlx4_rss_put(flow->rss);
1171         rte_free(flow);
1172         return 0;
1173 }
1174
1175 /**
1176  * Destroy user-configured flow rules.
1177  *
1178  * This function skips internal flows rules.
1179  *
1180  * @see rte_flow_flush()
1181  * @see rte_flow_ops
1182  */
1183 static int
1184 mlx4_flow_flush(struct rte_eth_dev *dev,
1185                 struct rte_flow_error *error)
1186 {
1187         struct priv *priv = dev->data->dev_private;
1188         struct rte_flow *flow = LIST_FIRST(&priv->flows);
1189
1190         while (flow) {
1191                 struct rte_flow *next = LIST_NEXT(flow, next);
1192
1193                 if (!flow->internal)
1194                         mlx4_flow_destroy(dev, flow, error);
1195                 flow = next;
1196         }
1197         return 0;
1198 }
1199
1200 /**
1201  * Helper function to determine the next configured VLAN filter.
1202  *
1203  * @param priv
1204  *   Pointer to private structure.
1205  * @param vlan
1206  *   VLAN ID to use as a starting point.
1207  *
1208  * @return
1209  *   Next configured VLAN ID or a high value (>= 4096) if there is none.
1210  */
1211 static uint16_t
1212 mlx4_flow_internal_next_vlan(struct priv *priv, uint16_t vlan)
1213 {
1214         while (vlan < 4096) {
1215                 if (priv->dev->data->vlan_filter_conf.ids[vlan / 64] &
1216                     (UINT64_C(1) << (vlan % 64)))
1217                         return vlan;
1218                 ++vlan;
1219         }
1220         return vlan;
1221 }
1222
1223 /**
1224  * Generate internal flow rules.
1225  *
1226  * Various flow rules are created depending on the mode the device is in:
1227  *
1228  * 1. Promiscuous:
1229  *       port MAC + broadcast + catch-all (VLAN filtering is ignored).
1230  * 2. All multicast:
1231  *       port MAC/VLAN + broadcast + catch-all multicast.
1232  * 3. Otherwise:
1233  *       port MAC/VLAN + broadcast MAC/VLAN.
1234  *
1235  * About MAC flow rules:
1236  *
1237  * - MAC flow rules are generated from @p dev->data->mac_addrs
1238  *   (@p priv->mac array).
1239  * - An additional flow rule for Ethernet broadcasts is also generated.
1240  * - All these are per-VLAN if @p DEV_RX_OFFLOAD_VLAN_FILTER
1241  *   is enabled and VLAN filters are configured.
1242  *
1243  * @param priv
1244  *   Pointer to private structure.
1245  * @param[out] error
1246  *   Perform verbose error reporting if not NULL.
1247  *
1248  * @return
1249  *   0 on success, a negative errno value otherwise and rte_errno is set.
1250  */
1251 static int
1252 mlx4_flow_internal(struct priv *priv, struct rte_flow_error *error)
1253 {
1254         struct rte_flow_attr attr = {
1255                 .priority = MLX4_FLOW_PRIORITY_LAST,
1256                 .ingress = 1,
1257         };
1258         struct rte_flow_item_eth eth_spec;
1259         const struct rte_flow_item_eth eth_mask = {
1260                 .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
1261         };
1262         const struct rte_flow_item_eth eth_allmulti = {
1263                 .dst.addr_bytes = "\x01\x00\x00\x00\x00\x00",
1264         };
1265         struct rte_flow_item_vlan vlan_spec;
1266         const struct rte_flow_item_vlan vlan_mask = {
1267                 .tci = RTE_BE16(0x0fff),
1268         };
1269         struct rte_flow_item pattern[] = {
1270                 {
1271                         .type = MLX4_FLOW_ITEM_TYPE_INTERNAL,
1272                 },
1273                 {
1274                         .type = RTE_FLOW_ITEM_TYPE_ETH,
1275                         .spec = &eth_spec,
1276                         .mask = &eth_mask,
1277                 },
1278                 {
1279                         /* Replaced with VLAN if filtering is enabled. */
1280                         .type = RTE_FLOW_ITEM_TYPE_END,
1281                 },
1282                 {
1283                         .type = RTE_FLOW_ITEM_TYPE_END,
1284                 },
1285         };
1286         /*
1287          * Round number of queues down to their previous power of 2 to
1288          * comply with RSS context limitations. Extra queues silently do not
1289          * get RSS by default.
1290          */
1291         uint32_t queues =
1292                 rte_align32pow2(priv->dev->data->nb_rx_queues + 1) >> 1;
1293         alignas(struct rte_flow_action_rss) uint8_t rss_conf_data
1294                 [offsetof(struct rte_flow_action_rss, queue) +
1295                  sizeof(((struct rte_flow_action_rss *)0)->queue[0]) * queues];
1296         struct rte_flow_action_rss *rss_conf = (void *)rss_conf_data;
1297         struct rte_flow_action actions[] = {
1298                 {
1299                         .type = RTE_FLOW_ACTION_TYPE_RSS,
1300                         .conf = rss_conf,
1301                 },
1302                 {
1303                         .type = RTE_FLOW_ACTION_TYPE_END,
1304                 },
1305         };
1306         struct ether_addr *rule_mac = &eth_spec.dst;
1307         rte_be16_t *rule_vlan =
1308                 (priv->dev->data->dev_conf.rxmode.offloads &
1309                  DEV_RX_OFFLOAD_VLAN_FILTER) &&
1310                 !priv->dev->data->promiscuous ?
1311                 &vlan_spec.tci :
1312                 NULL;
1313         uint16_t vlan = 0;
1314         struct rte_flow *flow;
1315         unsigned int i;
1316         int err = 0;
1317
1318         /* Nothing to be done if there are no Rx queues. */
1319         if (!queues)
1320                 goto error;
1321         /* Prepare default RSS configuration. */
1322         *rss_conf = (struct rte_flow_action_rss){
1323                 .rss_conf = NULL, /* Rely on default fallback settings. */
1324                 .num = queues,
1325         };
1326         for (i = 0; i != queues; ++i)
1327                 rss_conf->queue[i] = i;
1328         /*
1329          * Set up VLAN item if filtering is enabled and at least one VLAN
1330          * filter is configured.
1331          */
1332         if (rule_vlan) {
1333                 vlan = mlx4_flow_internal_next_vlan(priv, 0);
1334                 if (vlan < 4096) {
1335                         pattern[2] = (struct rte_flow_item){
1336                                 .type = RTE_FLOW_ITEM_TYPE_VLAN,
1337                                 .spec = &vlan_spec,
1338                                 .mask = &vlan_mask,
1339                         };
1340 next_vlan:
1341                         *rule_vlan = rte_cpu_to_be_16(vlan);
1342                 } else {
1343                         rule_vlan = NULL;
1344                 }
1345         }
1346         for (i = 0; i != RTE_DIM(priv->mac) + 1; ++i) {
1347                 const struct ether_addr *mac;
1348
1349                 /* Broadcasts are handled by an extra iteration. */
1350                 if (i < RTE_DIM(priv->mac))
1351                         mac = &priv->mac[i];
1352                 else
1353                         mac = &eth_mask.dst;
1354                 if (is_zero_ether_addr(mac))
1355                         continue;
1356                 /* Check if MAC flow rule is already present. */
1357                 for (flow = LIST_FIRST(&priv->flows);
1358                      flow && flow->internal;
1359                      flow = LIST_NEXT(flow, next)) {
1360                         const struct ibv_flow_spec_eth *eth =
1361                                 (const void *)((uintptr_t)flow->ibv_attr +
1362                                                sizeof(*flow->ibv_attr));
1363                         unsigned int j;
1364
1365                         if (!flow->mac)
1366                                 continue;
1367                         assert(flow->ibv_attr->type == IBV_FLOW_ATTR_NORMAL);
1368                         assert(flow->ibv_attr->num_of_specs == 1);
1369                         assert(eth->type == IBV_FLOW_SPEC_ETH);
1370                         assert(flow->rss);
1371                         if (rule_vlan &&
1372                             (eth->val.vlan_tag != *rule_vlan ||
1373                              eth->mask.vlan_tag != RTE_BE16(0x0fff)))
1374                                 continue;
1375                         if (!rule_vlan && eth->mask.vlan_tag)
1376                                 continue;
1377                         for (j = 0; j != sizeof(mac->addr_bytes); ++j)
1378                                 if (eth->val.dst_mac[j] != mac->addr_bytes[j] ||
1379                                     eth->mask.dst_mac[j] != UINT8_C(0xff) ||
1380                                     eth->val.src_mac[j] != UINT8_C(0x00) ||
1381                                     eth->mask.src_mac[j] != UINT8_C(0x00))
1382                                         break;
1383                         if (j != sizeof(mac->addr_bytes))
1384                                 continue;
1385                         if (flow->rss->queues != queues ||
1386                             memcmp(flow->rss->queue_id, rss_conf->queue,
1387                                    queues * sizeof(flow->rss->queue_id[0])))
1388                                 continue;
1389                         break;
1390                 }
1391                 if (!flow || !flow->internal) {
1392                         /* Not found, create a new flow rule. */
1393                         memcpy(rule_mac, mac, sizeof(*mac));
1394                         flow = mlx4_flow_create(priv->dev, &attr, pattern,
1395                                                 actions, error);
1396                         if (!flow) {
1397                                 err = -rte_errno;
1398                                 goto error;
1399                         }
1400                 }
1401                 flow->select = 1;
1402                 flow->mac = 1;
1403         }
1404         if (rule_vlan) {
1405                 vlan = mlx4_flow_internal_next_vlan(priv, vlan + 1);
1406                 if (vlan < 4096)
1407                         goto next_vlan;
1408         }
1409         /* Take care of promiscuous and all multicast flow rules. */
1410         if (priv->dev->data->promiscuous || priv->dev->data->all_multicast) {
1411                 for (flow = LIST_FIRST(&priv->flows);
1412                      flow && flow->internal;
1413                      flow = LIST_NEXT(flow, next)) {
1414                         if (priv->dev->data->promiscuous) {
1415                                 if (flow->promisc)
1416                                         break;
1417                         } else {
1418                                 assert(priv->dev->data->all_multicast);
1419                                 if (flow->allmulti)
1420                                         break;
1421                         }
1422                 }
1423                 if (flow && flow->internal) {
1424                         assert(flow->rss);
1425                         if (flow->rss->queues != queues ||
1426                             memcmp(flow->rss->queue_id, rss_conf->queue,
1427                                    queues * sizeof(flow->rss->queue_id[0])))
1428                                 flow = NULL;
1429                 }
1430                 if (!flow || !flow->internal) {
1431                         /* Not found, create a new flow rule. */
1432                         if (priv->dev->data->promiscuous) {
1433                                 pattern[1].spec = NULL;
1434                                 pattern[1].mask = NULL;
1435                         } else {
1436                                 assert(priv->dev->data->all_multicast);
1437                                 pattern[1].spec = &eth_allmulti;
1438                                 pattern[1].mask = &eth_allmulti;
1439                         }
1440                         pattern[2] = pattern[3];
1441                         flow = mlx4_flow_create(priv->dev, &attr, pattern,
1442                                                 actions, error);
1443                         if (!flow) {
1444                                 err = -rte_errno;
1445                                 goto error;
1446                         }
1447                 }
1448                 assert(flow->promisc || flow->allmulti);
1449                 flow->select = 1;
1450         }
1451 error:
1452         /* Clear selection and clean up stale internal flow rules. */
1453         flow = LIST_FIRST(&priv->flows);
1454         while (flow && flow->internal) {
1455                 struct rte_flow *next = LIST_NEXT(flow, next);
1456
1457                 if (!flow->select)
1458                         claim_zero(mlx4_flow_destroy(priv->dev, flow, error));
1459                 else
1460                         flow->select = 0;
1461                 flow = next;
1462         }
1463         return err;
1464 }
1465
1466 /**
1467  * Synchronize flow rules.
1468  *
1469  * This function synchronizes flow rules with the state of the device by
1470  * taking into account isolated mode and whether target queues are
1471  * configured.
1472  *
1473  * @param priv
1474  *   Pointer to private structure.
1475  * @param[out] error
1476  *   Perform verbose error reporting if not NULL.
1477  *
1478  * @return
1479  *   0 on success, a negative errno value otherwise and rte_errno is set.
1480  */
1481 int
1482 mlx4_flow_sync(struct priv *priv, struct rte_flow_error *error)
1483 {
1484         struct rte_flow *flow;
1485         int ret;
1486
1487         /* Internal flow rules are guaranteed to come first in the list. */
1488         if (priv->isolated) {
1489                 /*
1490                  * Get rid of them in isolated mode, stop at the first
1491                  * non-internal rule found.
1492                  */
1493                 for (flow = LIST_FIRST(&priv->flows);
1494                      flow && flow->internal;
1495                      flow = LIST_FIRST(&priv->flows))
1496                         claim_zero(mlx4_flow_destroy(priv->dev, flow, error));
1497         } else {
1498                 /* Refresh internal rules. */
1499                 ret = mlx4_flow_internal(priv, error);
1500                 if (ret)
1501                         return ret;
1502         }
1503         /* Toggle the remaining flow rules . */
1504         LIST_FOREACH(flow, &priv->flows, next) {
1505                 ret = mlx4_flow_toggle(priv, flow, priv->started, error);
1506                 if (ret)
1507                         return ret;
1508         }
1509         if (!priv->started)
1510                 assert(!priv->drop);
1511         return 0;
1512 }
1513
1514 /**
1515  * Clean up all flow rules.
1516  *
1517  * Unlike mlx4_flow_flush(), this function takes care of all remaining flow
1518  * rules regardless of whether they are internal or user-configured.
1519  *
1520  * @param priv
1521  *   Pointer to private structure.
1522  */
1523 void
1524 mlx4_flow_clean(struct priv *priv)
1525 {
1526         struct rte_flow *flow;
1527
1528         while ((flow = LIST_FIRST(&priv->flows)))
1529                 mlx4_flow_destroy(priv->dev, flow, NULL);
1530         assert(LIST_EMPTY(&priv->rss));
1531 }
1532
1533 static const struct rte_flow_ops mlx4_flow_ops = {
1534         .validate = mlx4_flow_validate,
1535         .create = mlx4_flow_create,
1536         .destroy = mlx4_flow_destroy,
1537         .flush = mlx4_flow_flush,
1538         .isolate = mlx4_flow_isolate,
1539 };
1540
1541 /**
1542  * Manage filter operations.
1543  *
1544  * @param dev
1545  *   Pointer to Ethernet device structure.
1546  * @param filter_type
1547  *   Filter type.
1548  * @param filter_op
1549  *   Operation to perform.
1550  * @param arg
1551  *   Pointer to operation-specific structure.
1552  *
1553  * @return
1554  *   0 on success, negative errno value otherwise and rte_errno is set.
1555  */
1556 int
1557 mlx4_filter_ctrl(struct rte_eth_dev *dev,
1558                  enum rte_filter_type filter_type,
1559                  enum rte_filter_op filter_op,
1560                  void *arg)
1561 {
1562         switch (filter_type) {
1563         case RTE_ETH_FILTER_GENERIC:
1564                 if (filter_op != RTE_ETH_FILTER_GET)
1565                         break;
1566                 *(const void **)arg = &mlx4_flow_ops;
1567                 return 0;
1568         default:
1569                 ERROR("%p: filter type (%d) not supported",
1570                       (void *)dev, filter_type);
1571                 break;
1572         }
1573         rte_errno = ENOTSUP;
1574         return -rte_errno;
1575 }