lib/librte_ethdev/rte_flow.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2016 6WIND S.A.
   3  * Copyright 2016 Mellanox Technologies, Ltd
   4  */
   5
   6 #include <errno.h>
   7 #include <stddef.h>
   8 #include <stdint.h>
   9 #include <string.h>
  10
  11 #include <rte_common.h>
  12 #include <rte_errno.h>
  13 #include <rte_branch_prediction.h>
  14 #include <rte_string_fns.h>
  15 #include "rte_ethdev.h"
  16 #include "rte_flow_driver.h"
  17 #include "rte_flow.h"
  18
  19 /**
  20  * Flow elements description tables.
  21  */
  22 struct rte_flow_desc_data {
  23         const char *name;
  24         size_t size;
  25 };
  26
  27 /** Generate flow_item[] entry. */
  28 #define MK_FLOW_ITEM(t, s) \
  29         [RTE_FLOW_ITEM_TYPE_ ## t] = { \
  30                 .name = # t, \
  31                 .size = s, \
  32         }
  33
  34 /** Information about known flow pattern items. */
  35 static const struct rte_flow_desc_data rte_flow_desc_item[] = {
  36         MK_FLOW_ITEM(END, 0),
  37         MK_FLOW_ITEM(VOID, 0),
  38         MK_FLOW_ITEM(INVERT, 0),
  39         MK_FLOW_ITEM(ANY, sizeof(struct rte_flow_item_any)),
  40         MK_FLOW_ITEM(PF, 0),
  41         MK_FLOW_ITEM(VF, sizeof(struct rte_flow_item_vf)),
  42         MK_FLOW_ITEM(PHY_PORT, sizeof(struct rte_flow_item_phy_port)),
  43         MK_FLOW_ITEM(PORT_ID, sizeof(struct rte_flow_item_port_id)),
  44         MK_FLOW_ITEM(RAW, sizeof(struct rte_flow_item_raw)),
  45         MK_FLOW_ITEM(ETH, sizeof(struct rte_flow_item_eth)),
  46         MK_FLOW_ITEM(VLAN, sizeof(struct rte_flow_item_vlan)),
  47         MK_FLOW_ITEM(IPV4, sizeof(struct rte_flow_item_ipv4)),
  48         MK_FLOW_ITEM(IPV6, sizeof(struct rte_flow_item_ipv6)),
  49         MK_FLOW_ITEM(ICMP, sizeof(struct rte_flow_item_icmp)),
  50         MK_FLOW_ITEM(UDP, sizeof(struct rte_flow_item_udp)),
  51         MK_FLOW_ITEM(TCP, sizeof(struct rte_flow_item_tcp)),
  52         MK_FLOW_ITEM(SCTP, sizeof(struct rte_flow_item_sctp)),
  53         MK_FLOW_ITEM(VXLAN, sizeof(struct rte_flow_item_vxlan)),
  54         MK_FLOW_ITEM(E_TAG, sizeof(struct rte_flow_item_e_tag)),
  55         MK_FLOW_ITEM(NVGRE, sizeof(struct rte_flow_item_nvgre)),
  56         MK_FLOW_ITEM(MPLS, sizeof(struct rte_flow_item_mpls)),
  57         MK_FLOW_ITEM(GRE, sizeof(struct rte_flow_item_gre)),
  58         MK_FLOW_ITEM(FUZZY, sizeof(struct rte_flow_item_fuzzy)),
  59         MK_FLOW_ITEM(GTP, sizeof(struct rte_flow_item_gtp)),
  60         MK_FLOW_ITEM(GTPC, sizeof(struct rte_flow_item_gtp)),
  61         MK_FLOW_ITEM(GTPU, sizeof(struct rte_flow_item_gtp)),
  62         MK_FLOW_ITEM(ESP, sizeof(struct rte_flow_item_esp)),
  63         MK_FLOW_ITEM(GENEVE, sizeof(struct rte_flow_item_geneve)),
  64         MK_FLOW_ITEM(VXLAN_GPE, sizeof(struct rte_flow_item_vxlan_gpe)),
  65         MK_FLOW_ITEM(ARP_ETH_IPV4, sizeof(struct rte_flow_item_arp_eth_ipv4)),
  66         MK_FLOW_ITEM(IPV6_EXT, sizeof(struct rte_flow_item_ipv6_ext)),
  67         MK_FLOW_ITEM(ICMP6, sizeof(struct rte_flow_item_icmp6)),
  68         MK_FLOW_ITEM(ICMP6_ND_NS, sizeof(struct rte_flow_item_icmp6_nd_ns)),
  69         MK_FLOW_ITEM(ICMP6_ND_NA, sizeof(struct rte_flow_item_icmp6_nd_na)),
  70         MK_FLOW_ITEM(ICMP6_ND_OPT, sizeof(struct rte_flow_item_icmp6_nd_opt)),
  71         MK_FLOW_ITEM(ICMP6_ND_OPT_SLA_ETH,
  72                      sizeof(struct rte_flow_item_icmp6_nd_opt_sla_eth)),
  73         MK_FLOW_ITEM(ICMP6_ND_OPT_TLA_ETH,
  74                      sizeof(struct rte_flow_item_icmp6_nd_opt_tla_eth)),
  75         MK_FLOW_ITEM(MARK, sizeof(struct rte_flow_item_mark)),
  76         MK_FLOW_ITEM(META, sizeof(struct rte_flow_item_meta)),
  77         MK_FLOW_ITEM(GRE_KEY, sizeof(rte_be32_t)),
  78 };
  79
  80 /** Generate flow_action[] entry. */
  81 #define MK_FLOW_ACTION(t, s) \
  82         [RTE_FLOW_ACTION_TYPE_ ## t] = { \
  83                 .name = # t, \
  84                 .size = s, \
  85         }
  86
  87 /** Information about known flow actions. */
  88 static const struct rte_flow_desc_data rte_flow_desc_action[] = {
  89         MK_FLOW_ACTION(END, 0),
  90         MK_FLOW_ACTION(VOID, 0),
  91         MK_FLOW_ACTION(PASSTHRU, 0),
  92         MK_FLOW_ACTION(JUMP, sizeof(struct rte_flow_action_jump)),
  93         MK_FLOW_ACTION(MARK, sizeof(struct rte_flow_action_mark)),
  94         MK_FLOW_ACTION(FLAG, 0),
  95         MK_FLOW_ACTION(QUEUE, sizeof(struct rte_flow_action_queue)),
  96         MK_FLOW_ACTION(DROP, 0),
  97         MK_FLOW_ACTION(COUNT, sizeof(struct rte_flow_action_count)),
  98         MK_FLOW_ACTION(RSS, sizeof(struct rte_flow_action_rss)),
  99         MK_FLOW_ACTION(PF, 0),
 100         MK_FLOW_ACTION(VF, sizeof(struct rte_flow_action_vf)),
 101         MK_FLOW_ACTION(PHY_PORT, sizeof(struct rte_flow_action_phy_port)),
 102         MK_FLOW_ACTION(PORT_ID, sizeof(struct rte_flow_action_port_id)),
 103         MK_FLOW_ACTION(METER, sizeof(struct rte_flow_action_meter)),
 104         MK_FLOW_ACTION(SECURITY, sizeof(struct rte_flow_action_security)),
 105         MK_FLOW_ACTION(OF_SET_MPLS_TTL,
 106                        sizeof(struct rte_flow_action_of_set_mpls_ttl)),
 107         MK_FLOW_ACTION(OF_DEC_MPLS_TTL, 0),
 108         MK_FLOW_ACTION(OF_SET_NW_TTL,
 109                        sizeof(struct rte_flow_action_of_set_nw_ttl)),
 110         MK_FLOW_ACTION(OF_DEC_NW_TTL, 0),
 111         MK_FLOW_ACTION(OF_COPY_TTL_OUT, 0),
 112         MK_FLOW_ACTION(OF_COPY_TTL_IN, 0),
 113         MK_FLOW_ACTION(OF_POP_VLAN, 0),
 114         MK_FLOW_ACTION(OF_PUSH_VLAN,
 115                        sizeof(struct rte_flow_action_of_push_vlan)),
 116         MK_FLOW_ACTION(OF_SET_VLAN_VID,
 117                        sizeof(struct rte_flow_action_of_set_vlan_vid)),
 118         MK_FLOW_ACTION(OF_SET_VLAN_PCP,
 119                        sizeof(struct rte_flow_action_of_set_vlan_pcp)),
 120         MK_FLOW_ACTION(OF_POP_MPLS,
 121                        sizeof(struct rte_flow_action_of_pop_mpls)),
 122         MK_FLOW_ACTION(OF_PUSH_MPLS,
 123                        sizeof(struct rte_flow_action_of_push_mpls)),
 124         MK_FLOW_ACTION(VXLAN_ENCAP, sizeof(struct rte_flow_action_vxlan_encap)),
 125         MK_FLOW_ACTION(VXLAN_DECAP, 0),
 126         MK_FLOW_ACTION(NVGRE_ENCAP, sizeof(struct rte_flow_action_vxlan_encap)),
 127         MK_FLOW_ACTION(NVGRE_DECAP, 0),
 128         MK_FLOW_ACTION(RAW_ENCAP, sizeof(struct rte_flow_action_raw_encap)),
 129         MK_FLOW_ACTION(RAW_DECAP, sizeof(struct rte_flow_action_raw_decap)),
 130         MK_FLOW_ACTION(SET_IPV4_SRC,
 131                        sizeof(struct rte_flow_action_set_ipv4)),
 132         MK_FLOW_ACTION(SET_IPV4_DST,
 133                        sizeof(struct rte_flow_action_set_ipv4)),
 134         MK_FLOW_ACTION(SET_IPV6_SRC,
 135                        sizeof(struct rte_flow_action_set_ipv6)),
 136         MK_FLOW_ACTION(SET_IPV6_DST,
 137                        sizeof(struct rte_flow_action_set_ipv6)),
 138         MK_FLOW_ACTION(SET_TP_SRC,
 139                        sizeof(struct rte_flow_action_set_tp)),
 140         MK_FLOW_ACTION(SET_TP_DST,
 141                        sizeof(struct rte_flow_action_set_tp)),
 142         MK_FLOW_ACTION(MAC_SWAP, 0),
 143         MK_FLOW_ACTION(DEC_TTL, 0),
 144         MK_FLOW_ACTION(SET_TTL, sizeof(struct rte_flow_action_set_ttl)),
 145         MK_FLOW_ACTION(SET_MAC_SRC, sizeof(struct rte_flow_action_set_mac)),
 146         MK_FLOW_ACTION(SET_MAC_DST, sizeof(struct rte_flow_action_set_mac)),
 147         MK_FLOW_ACTION(INC_TCP_SEQ, sizeof(rte_be32_t)),
 148         MK_FLOW_ACTION(DEC_TCP_SEQ, sizeof(rte_be32_t)),
 149         MK_FLOW_ACTION(INC_TCP_ACK, sizeof(rte_be32_t)),
 150         MK_FLOW_ACTION(DEC_TCP_ACK, sizeof(rte_be32_t)),
 151 };
 152
 153 static int
 154 flow_err(uint16_t port_id, int ret, struct rte_flow_error *error)
 155 {
 156         if (ret == 0)
 157                 return 0;
 158         if (rte_eth_dev_is_removed(port_id))
 159                 return rte_flow_error_set(error, EIO,
 160                                           RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 161                                           NULL, rte_strerror(EIO));
 162         return ret;
 163 }
 164
 165 /* Get generic flow operations structure from a port. */
 166 const struct rte_flow_ops *
 167 rte_flow_ops_get(uint16_t port_id, struct rte_flow_error *error)
 168 {
 169         struct rte_eth_dev *dev = &rte_eth_devices[port_id];
 170         const struct rte_flow_ops *ops;
 171         int code;
 172
 173         if (unlikely(!rte_eth_dev_is_valid_port(port_id)))
 174                 code = ENODEV;
 175         else if (unlikely(!dev->dev_ops->filter_ctrl ||
 176                           dev->dev_ops->filter_ctrl(dev,
 177                                                     RTE_ETH_FILTER_GENERIC,
 178                                                     RTE_ETH_FILTER_GET,
 179                                                     &ops) ||
 180                           !ops))
 181                 code = ENOSYS;
 182         else
 183                 return ops;
 184         rte_flow_error_set(error, code, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 185                            NULL, rte_strerror(code));
 186         return NULL;
 187 }
 188
 189 /* Check whether a flow rule can be created on a given port. */
 190 int
 191 rte_flow_validate(uint16_t port_id,
 192                   const struct rte_flow_attr *attr,
 193                   const struct rte_flow_item pattern[],
 194                   const struct rte_flow_action actions[],
 195                   struct rte_flow_error *error)
 196 {
 197         const struct rte_flow_ops *ops = rte_flow_ops_get(port_id, error);
 198         struct rte_eth_dev *dev = &rte_eth_devices[port_id];
 199
 200         if (unlikely(!ops))
 201                 return -rte_errno;
 202         if (likely(!!ops->validate))
 203                 return flow_err(port_id, ops->validate(dev, attr, pattern,
 204                                                        actions, error), error);
 205         return rte_flow_error_set(error, ENOSYS,
 206                                   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 207                                   NULL, rte_strerror(ENOSYS));
 208 }
 209
 210 /* Create a flow rule on a given port. */
 211 struct rte_flow *
 212 rte_flow_create(uint16_t port_id,
 213                 const struct rte_flow_attr *attr,
 214                 const struct rte_flow_item pattern[],
 215                 const struct rte_flow_action actions[],
 216                 struct rte_flow_error *error)
 217 {
 218         struct rte_eth_dev *dev = &rte_eth_devices[port_id];
 219         struct rte_flow *flow;
 220         const struct rte_flow_ops *ops = rte_flow_ops_get(port_id, error);
 221
 222         if (unlikely(!ops))
 223                 return NULL;
 224         if (likely(!!ops->create)) {
 225                 flow = ops->create(dev, attr, pattern, actions, error);
 226                 if (flow == NULL)
 227                         flow_err(port_id, -rte_errno, error);
 228                 return flow;
 229         }
 230         rte_flow_error_set(error, ENOSYS, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 231                            NULL, rte_strerror(ENOSYS));
 232         return NULL;
 233 }
 234
 235 /* Destroy a flow rule on a given port. */
 236 int
 237 rte_flow_destroy(uint16_t port_id,
 238                  struct rte_flow *flow,
 239                  struct rte_flow_error *error)
 240 {
 241         struct rte_eth_dev *dev = &rte_eth_devices[port_id];
 242         const struct rte_flow_ops *ops = rte_flow_ops_get(port_id, error);
 243
 244         if (unlikely(!ops))
 245                 return -rte_errno;
 246         if (likely(!!ops->destroy))
 247                 return flow_err(port_id, ops->destroy(dev, flow, error),
 248                                 error);
 249         return rte_flow_error_set(error, ENOSYS,
 250                                   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 251                                   NULL, rte_strerror(ENOSYS));
 252 }
 253
 254 /* Destroy all flow rules associated with a port. */
 255 int
 256 rte_flow_flush(uint16_t port_id,
 257                struct rte_flow_error *error)
 258 {
 259         struct rte_eth_dev *dev = &rte_eth_devices[port_id];
 260         const struct rte_flow_ops *ops = rte_flow_ops_get(port_id, error);
 261
 262         if (unlikely(!ops))
 263                 return -rte_errno;
 264         if (likely(!!ops->flush))
 265                 return flow_err(port_id, ops->flush(dev, error), error);
 266         return rte_flow_error_set(error, ENOSYS,
 267                                   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 268                                   NULL, rte_strerror(ENOSYS));
 269 }
 270
 271 /* Query an existing flow rule. */
 272 int
 273 rte_flow_query(uint16_t port_id,
 274                struct rte_flow *flow,
 275                const struct rte_flow_action *action,
 276                void *data,
 277                struct rte_flow_error *error)
 278 {
 279         struct rte_eth_dev *dev = &rte_eth_devices[port_id];
 280         const struct rte_flow_ops *ops = rte_flow_ops_get(port_id, error);
 281
 282         if (!ops)
 283                 return -rte_errno;
 284         if (likely(!!ops->query))
 285                 return flow_err(port_id, ops->query(dev, flow, action, data,
 286                                                     error), error);
 287         return rte_flow_error_set(error, ENOSYS,
 288                                   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 289                                   NULL, rte_strerror(ENOSYS));
 290 }
 291
 292 /* Restrict ingress traffic to the defined flow rules. */
 293 int
 294 rte_flow_isolate(uint16_t port_id,
 295                  int set,
 296                  struct rte_flow_error *error)
 297 {
 298         struct rte_eth_dev *dev = &rte_eth_devices[port_id];
 299         const struct rte_flow_ops *ops = rte_flow_ops_get(port_id, error);
 300
 301         if (!ops)
 302                 return -rte_errno;
 303         if (likely(!!ops->isolate))
 304                 return flow_err(port_id, ops->isolate(dev, set, error), error);
 305         return rte_flow_error_set(error, ENOSYS,
 306                                   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 307                                   NULL, rte_strerror(ENOSYS));
 308 }
 309
 310 /* Initialize flow error structure. */
 311 int
 312 rte_flow_error_set(struct rte_flow_error *error,
 313                    int code,
 314                    enum rte_flow_error_type type,
 315                    const void *cause,
 316                    const char *message)
 317 {
 318         if (error) {
 319                 *error = (struct rte_flow_error){
 320                         .type = type,
 321                         .cause = cause,
 322                         .message = message,
 323                 };
 324         }
 325         rte_errno = code;
 326         return -code;
 327 }
 328
 329 /** Pattern item specification types. */
 330 enum rte_flow_conv_item_spec_type {
 331         RTE_FLOW_CONV_ITEM_SPEC,
 332         RTE_FLOW_CONV_ITEM_LAST,
 333         RTE_FLOW_CONV_ITEM_MASK,
 334 };
 335
 336 /**
 337  * Copy pattern item specification.
 338  *
 339  * @param[out] buf
 340  *   Output buffer. Can be NULL if @p size is zero.
 341  * @param size
 342  *   Size of @p buf in bytes.
 343  * @param[in] item
 344  *   Pattern item to copy specification from.
 345  * @param type
 346  *   Specification selector for either @p spec, @p last or @p mask.
 347  *
 348  * @return
 349  *   Number of bytes needed to store pattern item specification regardless
 350  *   of @p size. @p buf contents are truncated to @p size if not large
 351  *   enough.
 352  */
 353 static size_t
 354 rte_flow_conv_item_spec(void *buf, const size_t size,
 355                         const struct rte_flow_item *item,
 356                         enum rte_flow_conv_item_spec_type type)
 357 {
 358         size_t off;
 359         const void *data =
 360                 type == RTE_FLOW_CONV_ITEM_SPEC ? item->spec :
 361                 type == RTE_FLOW_CONV_ITEM_LAST ? item->last :
 362                 type == RTE_FLOW_CONV_ITEM_MASK ? item->mask :
 363                 NULL;
 364
 365         switch (item->type) {
 366                 union {
 367                         const struct rte_flow_item_raw *raw;
 368                 } spec;
 369                 union {
 370                         const struct rte_flow_item_raw *raw;
 371                 } last;
 372                 union {
 373                         const struct rte_flow_item_raw *raw;
 374                 } mask;
 375                 union {
 376                         const struct rte_flow_item_raw *raw;
 377                 } src;
 378                 union {
 379                         struct rte_flow_item_raw *raw;
 380                 } dst;
 381                 size_t tmp;
 382
 383         case RTE_FLOW_ITEM_TYPE_RAW:
 384                 spec.raw = item->spec;
 385                 last.raw = item->last ? item->last : item->spec;
 386                 mask.raw = item->mask ? item->mask : &rte_flow_item_raw_mask;
 387                 src.raw = data;
 388                 dst.raw = buf;
 389                 rte_memcpy(dst.raw,
 390                            (&(struct rte_flow_item_raw){
 391                                 .relative = src.raw->relative,
 392                                 .search = src.raw->search,
 393                                 .reserved = src.raw->reserved,
 394                                 .offset = src.raw->offset,
 395                                 .limit = src.raw->limit,
 396                                 .length = src.raw->length,
 397                            }),
 398                            size > sizeof(*dst.raw) ? sizeof(*dst.raw) : size);
 399                 off = sizeof(*dst.raw);
 400                 if (type == RTE_FLOW_CONV_ITEM_SPEC ||
 401                     (type == RTE_FLOW_CONV_ITEM_MASK &&
 402                      ((spec.raw->length & mask.raw->length) >=
 403                       (last.raw->length & mask.raw->length))))
 404                         tmp = spec.raw->length & mask.raw->length;
 405                 else
 406                         tmp = last.raw->length & mask.raw->length;
 407                 if (tmp) {
 408                         off = RTE_ALIGN_CEIL(off, sizeof(*dst.raw->pattern));
 409                         if (size >= off + tmp)
 410                                 dst.raw->pattern = rte_memcpy
 411                                         ((void *)((uintptr_t)dst.raw + off),
 412                                          src.raw->pattern, tmp);
 413                         off += tmp;
 414                 }
 415                 break;
 416         default:
 417                 off = rte_flow_desc_item[item->type].size;
 418                 rte_memcpy(buf, data, (size > off ? off : size));
 419                 break;
 420         }
 421         return off;
 422 }
 423
 424 /**
 425  * Copy action configuration.
 426  *
 427  * @param[out] buf
 428  *   Output buffer. Can be NULL if @p size is zero.
 429  * @param size
 430  *   Size of @p buf in bytes.
 431  * @param[in] action
 432  *   Action to copy configuration from.
 433  *
 434  * @return
 435  *   Number of bytes needed to store pattern item specification regardless
 436  *   of @p size. @p buf contents are truncated to @p size if not large
 437  *   enough.
 438  */
 439 static size_t
 440 rte_flow_conv_action_conf(void *buf, const size_t size,
 441                           const struct rte_flow_action *action)
 442 {
 443         size_t off;
 444
 445         switch (action->type) {
 446                 union {
 447                         const struct rte_flow_action_rss *rss;
 448                         const struct rte_flow_action_vxlan_encap *vxlan_encap;
 449                         const struct rte_flow_action_nvgre_encap *nvgre_encap;
 450                 } src;
 451                 union {
 452                         struct rte_flow_action_rss *rss;
 453                         struct rte_flow_action_vxlan_encap *vxlan_encap;
 454                         struct rte_flow_action_nvgre_encap *nvgre_encap;
 455                 } dst;
 456                 size_t tmp;
 457                 int ret;
 458
 459         case RTE_FLOW_ACTION_TYPE_RSS:
 460                 src.rss = action->conf;
 461                 dst.rss = buf;
 462                 rte_memcpy(dst.rss,
 463                            (&(struct rte_flow_action_rss){
 464                                 .func = src.rss->func,
 465                                 .level = src.rss->level,
 466                                 .types = src.rss->types,
 467                                 .key_len = src.rss->key_len,
 468                                 .queue_num = src.rss->queue_num,
 469                            }),
 470                            size > sizeof(*dst.rss) ? sizeof(*dst.rss) : size);
 471                 off = sizeof(*dst.rss);
 472                 if (src.rss->key_len) {
 473                         off = RTE_ALIGN_CEIL(off, sizeof(*dst.rss->key));
 474                         tmp = sizeof(*src.rss->key) * src.rss->key_len;
 475                         if (size >= off + tmp)
 476                                 dst.rss->key = rte_memcpy
 477                                         ((void *)((uintptr_t)dst.rss + off),
 478                                          src.rss->key, tmp);
 479                         off += tmp;
 480                 }
 481                 if (src.rss->queue_num) {
 482                         off = RTE_ALIGN_CEIL(off, sizeof(*dst.rss->queue));
 483                         tmp = sizeof(*src.rss->queue) * src.rss->queue_num;
 484                         if (size >= off + tmp)
 485                                 dst.rss->queue = rte_memcpy
 486                                         ((void *)((uintptr_t)dst.rss + off),
 487                                          src.rss->queue, tmp);
 488                         off += tmp;
 489                 }
 490                 break;
 491         case RTE_FLOW_ACTION_TYPE_VXLAN_ENCAP:
 492         case RTE_FLOW_ACTION_TYPE_NVGRE_ENCAP:
 493                 src.vxlan_encap = action->conf;
 494                 dst.vxlan_encap = buf;
 495                 RTE_BUILD_BUG_ON(sizeof(*src.vxlan_encap) !=
 496                                  sizeof(*src.nvgre_encap) ||
 497                                  offsetof(struct rte_flow_action_vxlan_encap,
 498                                           definition) !=
 499                                  offsetof(struct rte_flow_action_nvgre_encap,
 500                                           definition));
 501                 off = sizeof(*dst.vxlan_encap);
 502                 if (src.vxlan_encap->definition) {
 503                         off = RTE_ALIGN_CEIL
 504                                 (off, sizeof(*dst.vxlan_encap->definition));
 505                         ret = rte_flow_conv
 506                                 (RTE_FLOW_CONV_OP_PATTERN,
 507                                  (void *)((uintptr_t)dst.vxlan_encap + off),
 508                                  size > off ? size - off : 0,
 509                                  src.vxlan_encap->definition, NULL);
 510                         if (ret < 0)
 511                                 return 0;
 512                         if (size >= off + ret)
 513                                 dst.vxlan_encap->definition =
 514                                         (void *)((uintptr_t)dst.vxlan_encap +
 515                                                  off);
 516                         off += ret;
 517                 }
 518                 break;
 519         default:
 520                 off = rte_flow_desc_action[action->type].size;
 521                 rte_memcpy(buf, action->conf, (size > off ? off : size));
 522                 break;
 523         }
 524         return off;
 525 }
 526
 527 /**
 528  * Copy a list of pattern items.
 529  *
 530  * @param[out] dst
 531  *   Destination buffer. Can be NULL if @p size is zero.
 532  * @param size
 533  *   Size of @p dst in bytes.
 534  * @param[in] src
 535  *   Source pattern items.
 536  * @param num
 537  *   Maximum number of pattern items to process from @p src or 0 to process
 538  *   the entire list. In both cases, processing stops after
 539  *   RTE_FLOW_ITEM_TYPE_END is encountered.
 540  * @param[out] error
 541  *   Perform verbose error reporting if not NULL.
 542  *
 543  * @return
 544  *   A positive value representing the number of bytes needed to store
 545  *   pattern items regardless of @p size on success (@p buf contents are
 546  *   truncated to @p size if not large enough), a negative errno value
 547  *   otherwise and rte_errno is set.
 548  */
 549 static int
 550 rte_flow_conv_pattern(struct rte_flow_item *dst,
 551                       const size_t size,
 552                       const struct rte_flow_item *src,
 553                       unsigned int num,
 554                       struct rte_flow_error *error)
 555 {
 556         uintptr_t data = (uintptr_t)dst;
 557         size_t off;
 558         size_t ret;
 559         unsigned int i;
 560
 561         for (i = 0, off = 0; !num || i != num; ++i, ++src, ++dst) {
 562                 if ((size_t)src->type >= RTE_DIM(rte_flow_desc_item) ||
 563                     !rte_flow_desc_item[src->type].name)
 564                         return rte_flow_error_set
 565                                 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM, src,
 566                                  "cannot convert unknown item type");
 567                 if (size >= off + sizeof(*dst))
 568                         *dst = (struct rte_flow_item){
 569                                 .type = src->type,
 570                         };
 571                 off += sizeof(*dst);
 572                 if (!src->type)
 573                         num = i + 1;
 574         }
 575         num = i;
 576         src -= num;
 577         dst -= num;
 578         do {
 579                 if (src->spec) {
 580                         off = RTE_ALIGN_CEIL(off, sizeof(double));
 581                         ret = rte_flow_conv_item_spec
 582                                 ((void *)(data + off),
 583                                  size > off ? size - off : 0, src,
 584                                  RTE_FLOW_CONV_ITEM_SPEC);
 585                         if (size && size >= off + ret)
 586                                 dst->spec = (void *)(data + off);
 587                         off += ret;
 588
 589                 }
 590                 if (src->last) {
 591                         off = RTE_ALIGN_CEIL(off, sizeof(double));
 592                         ret = rte_flow_conv_item_spec
 593                                 ((void *)(data + off),
 594                                  size > off ? size - off : 0, src,
 595                                  RTE_FLOW_CONV_ITEM_LAST);
 596                         if (size && size >= off + ret)
 597                                 dst->last = (void *)(data + off);
 598                         off += ret;
 599                 }
 600                 if (src->mask) {
 601                         off = RTE_ALIGN_CEIL(off, sizeof(double));
 602                         ret = rte_flow_conv_item_spec
 603                                 ((void *)(data + off),
 604                                  size > off ? size - off : 0, src,
 605                                  RTE_FLOW_CONV_ITEM_MASK);
 606                         if (size && size >= off + ret)
 607                                 dst->mask = (void *)(data + off);
 608                         off += ret;
 609                 }
 610                 ++src;
 611                 ++dst;
 612         } while (--num);
 613         return off;
 614 }
 615
 616 /**
 617  * Copy a list of actions.
 618  *
 619  * @param[out] dst
 620  *   Destination buffer. Can be NULL if @p size is zero.
 621  * @param size
 622  *   Size of @p dst in bytes.
 623  * @param[in] src
 624  *   Source actions.
 625  * @param num
 626  *   Maximum number of actions to process from @p src or 0 to process the
 627  *   entire list. In both cases, processing stops after
 628  *   RTE_FLOW_ACTION_TYPE_END is encountered.
 629  * @param[out] error
 630  *   Perform verbose error reporting if not NULL.
 631  *
 632  * @return
 633  *   A positive value representing the number of bytes needed to store
 634  *   actions regardless of @p size on success (@p buf contents are truncated
 635  *   to @p size if not large enough), a negative errno value otherwise and
 636  *   rte_errno is set.
 637  */
 638 static int
 639 rte_flow_conv_actions(struct rte_flow_action *dst,
 640                       const size_t size,
 641                       const struct rte_flow_action *src,
 642                       unsigned int num,
 643                       struct rte_flow_error *error)
 644 {
 645         uintptr_t data = (uintptr_t)dst;
 646         size_t off;
 647         size_t ret;
 648         unsigned int i;
 649
 650         for (i = 0, off = 0; !num || i != num; ++i, ++src, ++dst) {
 651                 if ((size_t)src->type >= RTE_DIM(rte_flow_desc_action) ||
 652                     !rte_flow_desc_action[src->type].name)
 653                         return rte_flow_error_set
 654                                 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 655                                  src, "cannot convert unknown action type");
 656                 if (size >= off + sizeof(*dst))
 657                         *dst = (struct rte_flow_action){
 658                                 .type = src->type,
 659                         };
 660                 off += sizeof(*dst);
 661                 if (!src->type)
 662                         num = i + 1;
 663         }
 664         num = i;
 665         src -= num;
 666         dst -= num;
 667         do {
 668                 if (src->conf) {
 669                         off = RTE_ALIGN_CEIL(off, sizeof(double));
 670                         ret = rte_flow_conv_action_conf
 671                                 ((void *)(data + off),
 672                                  size > off ? size - off : 0, src);
 673                         if (size && size >= off + ret)
 674                                 dst->conf = (void *)(data + off);
 675                         off += ret;
 676                 }
 677                 ++src;
 678                 ++dst;
 679         } while (--num);
 680         return off;
 681 }
 682
 683 /**
 684  * Copy flow rule components.
 685  *
 686  * This comprises the flow rule descriptor itself, attributes, pattern and
 687  * actions list. NULL components in @p src are skipped.
 688  *
 689  * @param[out] dst
 690  *   Destination buffer. Can be NULL if @p size is zero.
 691  * @param size
 692  *   Size of @p dst in bytes.
 693  * @param[in] src
 694  *   Source flow rule descriptor.
 695  * @param[out] error
 696  *   Perform verbose error reporting if not NULL.
 697  *
 698  * @return
 699  *   A positive value representing the number of bytes needed to store all
 700  *   components including the descriptor regardless of @p size on success
 701  *   (@p buf contents are truncated to @p size if not large enough), a
 702  *   negative errno value otherwise and rte_errno is set.
 703  */
 704 static int
 705 rte_flow_conv_rule(struct rte_flow_conv_rule *dst,
 706                    const size_t size,
 707                    const struct rte_flow_conv_rule *src,
 708                    struct rte_flow_error *error)
 709 {
 710         size_t off;
 711         int ret;
 712
 713         rte_memcpy(dst,
 714                    (&(struct rte_flow_conv_rule){
 715                         .attr = NULL,
 716                         .pattern = NULL,
 717                         .actions = NULL,
 718                    }),
 719                    size > sizeof(*dst) ? sizeof(*dst) : size);
 720         off = sizeof(*dst);
 721         if (src->attr_ro) {
 722                 off = RTE_ALIGN_CEIL(off, sizeof(double));
 723                 if (size && size >= off + sizeof(*dst->attr))
 724                         dst->attr = rte_memcpy
 725                                 ((void *)((uintptr_t)dst + off),
 726                                  src->attr_ro, sizeof(*dst->attr));
 727                 off += sizeof(*dst->attr);
 728         }
 729         if (src->pattern_ro) {
 730                 off = RTE_ALIGN_CEIL(off, sizeof(double));
 731                 ret = rte_flow_conv_pattern((void *)((uintptr_t)dst + off),
 732                                             size > off ? size - off : 0,
 733                                             src->pattern_ro, 0, error);
 734                 if (ret < 0)
 735                         return ret;
 736                 if (size && size >= off + (size_t)ret)
 737                         dst->pattern = (void *)((uintptr_t)dst + off);
 738                 off += ret;
 739         }
 740         if (src->actions_ro) {
 741                 off = RTE_ALIGN_CEIL(off, sizeof(double));
 742                 ret = rte_flow_conv_actions((void *)((uintptr_t)dst + off),
 743                                             size > off ? size - off : 0,
 744                                             src->actions_ro, 0, error);
 745                 if (ret < 0)
 746                         return ret;
 747                 if (size >= off + (size_t)ret)
 748                         dst->actions = (void *)((uintptr_t)dst + off);
 749                 off += ret;
 750         }
 751         return off;
 752 }
 753
 754 /**
 755  * Retrieve the name of a pattern item/action type.
 756  *
 757  * @param is_action
 758  *   Nonzero when @p src represents an action type instead of a pattern item
 759  *   type.
 760  * @param is_ptr
 761  *   Nonzero to write string address instead of contents into @p dst.
 762  * @param[out] dst
 763  *   Destination buffer. Can be NULL if @p size is zero.
 764  * @param size
 765  *   Size of @p dst in bytes.
 766  * @param[in] src
 767  *   Depending on @p is_action, source pattern item or action type cast as a
 768  *   pointer.
 769  * @param[out] error
 770  *   Perform verbose error reporting if not NULL.
 771  *
 772  * @return
 773  *   A positive value representing the number of bytes needed to store the
 774  *   name or its address regardless of @p size on success (@p buf contents
 775  *   are truncated to @p size if not large enough), a negative errno value
 776  *   otherwise and rte_errno is set.
 777  */
 778 static int
 779 rte_flow_conv_name(int is_action,
 780                    int is_ptr,
 781                    char *dst,
 782                    const size_t size,
 783                    const void *src,
 784                    struct rte_flow_error *error)
 785 {
 786         struct desc_info {
 787                 const struct rte_flow_desc_data *data;
 788                 size_t num;
 789         };
 790         static const struct desc_info info_rep[2] = {
 791                 { rte_flow_desc_item, RTE_DIM(rte_flow_desc_item), },
 792                 { rte_flow_desc_action, RTE_DIM(rte_flow_desc_action), },
 793         };
 794         const struct desc_info *const info = &info_rep[!!is_action];
 795         unsigned int type = (uintptr_t)src;
 796
 797         if (type >= info->num)
 798                 return rte_flow_error_set
 799                         (error, EINVAL, RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
 800                          "unknown object type to retrieve the name of");
 801         if (!is_ptr)
 802                 return strlcpy(dst, info->data[type].name, size);
 803         if (size >= sizeof(const char **))
 804                 *((const char **)dst) = info->data[type].name;
 805         return sizeof(const char **);
 806 }
 807
 808 /** Helper function to convert flow API objects. */
 809 int
 810 rte_flow_conv(enum rte_flow_conv_op op,
 811               void *dst,
 812               size_t size,
 813               const void *src,
 814               struct rte_flow_error *error)
 815 {
 816         switch (op) {
 817                 const struct rte_flow_attr *attr;
 818
 819         case RTE_FLOW_CONV_OP_NONE:
 820                 return 0;
 821         case RTE_FLOW_CONV_OP_ATTR:
 822                 attr = src;
 823                 if (size > sizeof(*attr))
 824                         size = sizeof(*attr);
 825                 rte_memcpy(dst, attr, size);
 826                 return sizeof(*attr);
 827         case RTE_FLOW_CONV_OP_ITEM:
 828                 return rte_flow_conv_pattern(dst, size, src, 1, error);
 829         case RTE_FLOW_CONV_OP_ACTION:
 830                 return rte_flow_conv_actions(dst, size, src, 1, error);
 831         case RTE_FLOW_CONV_OP_PATTERN:
 832                 return rte_flow_conv_pattern(dst, size, src, 0, error);
 833         case RTE_FLOW_CONV_OP_ACTIONS:
 834                 return rte_flow_conv_actions(dst, size, src, 0, error);
 835         case RTE_FLOW_CONV_OP_RULE:
 836                 return rte_flow_conv_rule(dst, size, src, error);
 837         case RTE_FLOW_CONV_OP_ITEM_NAME:
 838                 return rte_flow_conv_name(0, 0, dst, size, src, error);
 839         case RTE_FLOW_CONV_OP_ACTION_NAME:
 840                 return rte_flow_conv_name(1, 0, dst, size, src, error);
 841         case RTE_FLOW_CONV_OP_ITEM_NAME_PTR:
 842                 return rte_flow_conv_name(0, 1, dst, size, src, error);
 843         case RTE_FLOW_CONV_OP_ACTION_NAME_PTR:
 844                 return rte_flow_conv_name(1, 1, dst, size, src, error);
 845         }
 846         return rte_flow_error_set
 847                 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
 848                  "unknown object conversion operation");
 849 }
 850
 851 /** Store a full rte_flow description. */
 852 size_t
 853 rte_flow_copy(struct rte_flow_desc *desc, size_t len,
 854               const struct rte_flow_attr *attr,
 855               const struct rte_flow_item *items,
 856               const struct rte_flow_action *actions)
 857 {
 858         /*
 859          * Overlap struct rte_flow_conv with struct rte_flow_desc in order
 860          * to convert the former to the latter without wasting space.
 861          */
 862         struct rte_flow_conv_rule *dst =
 863                 len ?
 864                 (void *)((uintptr_t)desc +
 865                          (offsetof(struct rte_flow_desc, actions) -
 866                           offsetof(struct rte_flow_conv_rule, actions))) :
 867                 NULL;
 868         size_t dst_size =
 869                 len > sizeof(*desc) - sizeof(*dst) ?
 870                 len - (sizeof(*desc) - sizeof(*dst)) :
 871                 0;
 872         struct rte_flow_conv_rule src = {
 873                 .attr_ro = NULL,
 874                 .pattern_ro = items,
 875                 .actions_ro = actions,
 876         };
 877         int ret;
 878
 879         RTE_BUILD_BUG_ON(sizeof(struct rte_flow_desc) <
 880                          sizeof(struct rte_flow_conv_rule));
 881         if (dst_size &&
 882             (&dst->pattern != &desc->items ||
 883              &dst->actions != &desc->actions ||
 884              (uintptr_t)(dst + 1) != (uintptr_t)(desc + 1))) {
 885                 rte_errno = EINVAL;
 886                 return 0;
 887         }
 888         ret = rte_flow_conv(RTE_FLOW_CONV_OP_RULE, dst, dst_size, &src, NULL);
 889         if (ret < 0)
 890                 return 0;
 891         ret += sizeof(*desc) - sizeof(*dst);
 892         rte_memcpy(desc,
 893                    (&(struct rte_flow_desc){
 894                         .size = ret,
 895                         .attr = *attr,
 896                         .items = dst_size ? dst->pattern : NULL,
 897                         .actions = dst_size ? dst->actions : NULL,
 898                    }),
 899                    len > sizeof(*desc) ? sizeof(*desc) : len);
 900         return ret;
 901 }
 902
 903 /**
 904  * Expand RSS flows into several possible flows according to the RSS hash
 905  * fields requested and the driver capabilities.
 906  */
 907 int
 908 rte_flow_expand_rss(struct rte_flow_expand_rss *buf, size_t size,
 909                     const struct rte_flow_item *pattern, uint64_t types,
 910                     const struct rte_flow_expand_node graph[],
 911                     int graph_root_index)
 912 {
 913         const int elt_n = 8;
 914         const struct rte_flow_item *item;
 915         const struct rte_flow_expand_node *node = &graph[graph_root_index];
 916         const int *next_node;
 917         const int *stack[elt_n];
 918         int stack_pos = 0;
 919         struct rte_flow_item flow_items[elt_n];
 920         unsigned int i;
 921         size_t lsize;
 922         size_t user_pattern_size = 0;
 923         void *addr = NULL;
 924
 925         lsize = offsetof(struct rte_flow_expand_rss, entry) +
 926                 elt_n * sizeof(buf->entry[0]);
 927         if (lsize <= size) {
 928                 buf->entry[0].priority = 0;
 929                 buf->entry[0].pattern = (void *)&buf->entry[elt_n];
 930                 buf->entries = 0;
 931                 addr = buf->entry[0].pattern;
 932         }
 933         for (item = pattern; item->type != RTE_FLOW_ITEM_TYPE_END; item++) {
 934                 const struct rte_flow_expand_node *next = NULL;
 935
 936                 for (i = 0; node->next && node->next[i]; ++i) {
 937                         next = &graph[node->next[i]];
 938                         if (next->type == item->type)
 939                                 break;
 940                 }
 941                 if (next)
 942                         node = next;
 943                 user_pattern_size += sizeof(*item);
 944         }
 945         user_pattern_size += sizeof(*item); /* Handle END item. */
 946         lsize += user_pattern_size;
 947         /* Copy the user pattern in the first entry of the buffer. */
 948         if (lsize <= size) {
 949                 rte_memcpy(addr, pattern, user_pattern_size);
 950                 addr = (void *)(((uintptr_t)addr) + user_pattern_size);
 951                 buf->entries = 1;
 952         }
 953         /* Start expanding. */
 954         memset(flow_items, 0, sizeof(flow_items));
 955         user_pattern_size -= sizeof(*item);
 956         next_node = node->next;
 957         stack[stack_pos] = next_node;
 958         node = next_node ? &graph[*next_node] : NULL;
 959         while (node) {
 960                 flow_items[stack_pos].type = node->type;
 961                 if (node->rss_types & types) {
 962                         /*
 963                          * compute the number of items to copy from the
 964                          * expansion and copy it.
 965                          * When the stack_pos is 0, there are 1 element in it,
 966                          * plus the addition END item.
 967                          */
 968                         int elt = stack_pos + 2;
 969
 970                         flow_items[stack_pos + 1].type = RTE_FLOW_ITEM_TYPE_END;
 971                         lsize += elt * sizeof(*item) + user_pattern_size;
 972                         if (lsize <= size) {
 973                                 size_t n = elt * sizeof(*item);
 974
 975                                 buf->entry[buf->entries].priority =
 976                                         stack_pos + 1;
 977                                 buf->entry[buf->entries].pattern = addr;
 978                                 buf->entries++;
 979                                 rte_memcpy(addr, buf->entry[0].pattern,
 980                                            user_pattern_size);
 981                                 addr = (void *)(((uintptr_t)addr) +
 982                                                 user_pattern_size);
 983                                 rte_memcpy(addr, flow_items, n);
 984                                 addr = (void *)(((uintptr_t)addr) + n);
 985                         }
 986                 }
 987                 /* Go deeper. */
 988                 if (node->next) {
 989                         next_node = node->next;
 990                         if (stack_pos++ == elt_n) {
 991                                 rte_errno = E2BIG;
 992                                 return -rte_errno;
 993                         }
 994                         stack[stack_pos] = next_node;
 995                 } else if (*(next_node + 1)) {
 996                         /* Follow up with the next possibility. */
 997                         ++next_node;
 998                 } else {
 999                         /* Move to the next path. */
1000                         if (stack_pos)
1001                                 next_node = stack[--stack_pos];
1002                         next_node++;
1003                         stack[stack_pos] = next_node;
1004                 }
1005                 node = *next_node ? &graph[*next_node] : NULL;
1006         };
1007         return lsize;
1008 }