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