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_mbuf.h>
  16 #include <rte_mbuf_dyn.h>
  17 #include "rte_ethdev.h"
  18 #include "rte_flow_driver.h"
  19 #include "rte_flow.h"
  20
  21 /* Mbuf dynamic field name for metadata. */
  22 int rte_flow_dynf_metadata_offs = -1;
  23
  24 /* Mbuf dynamic field flag bit number for metadata. */
  25 uint64_t rte_flow_dynf_metadata_mask;
  26
  27 /**
  28  * Flow elements description tables.
  29  */
  30 struct rte_flow_desc_data {
  31         const char *name;
  32         size_t size;
  33 };
  34
  35 /** Generate flow_item[] entry. */
  36 #define MK_FLOW_ITEM(t, s) \
  37         [RTE_FLOW_ITEM_TYPE_ ## t] = { \
  38                 .name = # t, \
  39                 .size = s, \
  40         }
  41
  42 /** Information about known flow pattern items. */
  43 static const struct rte_flow_desc_data rte_flow_desc_item[] = {
  44         MK_FLOW_ITEM(END, 0),
  45         MK_FLOW_ITEM(VOID, 0),
  46         MK_FLOW_ITEM(INVERT, 0),
  47         MK_FLOW_ITEM(ANY, sizeof(struct rte_flow_item_any)),
  48         MK_FLOW_ITEM(PF, 0),
  49         MK_FLOW_ITEM(VF, sizeof(struct rte_flow_item_vf)),
  50         MK_FLOW_ITEM(PHY_PORT, sizeof(struct rte_flow_item_phy_port)),
  51         MK_FLOW_ITEM(PORT_ID, sizeof(struct rte_flow_item_port_id)),
  52         MK_FLOW_ITEM(RAW, sizeof(struct rte_flow_item_raw)),
  53         MK_FLOW_ITEM(ETH, sizeof(struct rte_flow_item_eth)),
  54         MK_FLOW_ITEM(VLAN, sizeof(struct rte_flow_item_vlan)),
  55         MK_FLOW_ITEM(IPV4, sizeof(struct rte_flow_item_ipv4)),
  56         MK_FLOW_ITEM(IPV6, sizeof(struct rte_flow_item_ipv6)),
  57         MK_FLOW_ITEM(ICMP, sizeof(struct rte_flow_item_icmp)),
  58         MK_FLOW_ITEM(UDP, sizeof(struct rte_flow_item_udp)),
  59         MK_FLOW_ITEM(TCP, sizeof(struct rte_flow_item_tcp)),
  60         MK_FLOW_ITEM(SCTP, sizeof(struct rte_flow_item_sctp)),
  61         MK_FLOW_ITEM(VXLAN, sizeof(struct rte_flow_item_vxlan)),
  62         MK_FLOW_ITEM(E_TAG, sizeof(struct rte_flow_item_e_tag)),
  63         MK_FLOW_ITEM(NVGRE, sizeof(struct rte_flow_item_nvgre)),
  64         MK_FLOW_ITEM(MPLS, sizeof(struct rte_flow_item_mpls)),
  65         MK_FLOW_ITEM(GRE, sizeof(struct rte_flow_item_gre)),
  66         MK_FLOW_ITEM(FUZZY, sizeof(struct rte_flow_item_fuzzy)),
  67         MK_FLOW_ITEM(GTP, sizeof(struct rte_flow_item_gtp)),
  68         MK_FLOW_ITEM(GTPC, sizeof(struct rte_flow_item_gtp)),
  69         MK_FLOW_ITEM(GTPU, sizeof(struct rte_flow_item_gtp)),
  70         MK_FLOW_ITEM(ESP, sizeof(struct rte_flow_item_esp)),
  71         MK_FLOW_ITEM(GENEVE, sizeof(struct rte_flow_item_geneve)),
  72         MK_FLOW_ITEM(VXLAN_GPE, sizeof(struct rte_flow_item_vxlan_gpe)),
  73         MK_FLOW_ITEM(ARP_ETH_IPV4, sizeof(struct rte_flow_item_arp_eth_ipv4)),
  74         MK_FLOW_ITEM(IPV6_EXT, sizeof(struct rte_flow_item_ipv6_ext)),
  75         MK_FLOW_ITEM(ICMP6, sizeof(struct rte_flow_item_icmp6)),
  76         MK_FLOW_ITEM(ICMP6_ND_NS, sizeof(struct rte_flow_item_icmp6_nd_ns)),
  77         MK_FLOW_ITEM(ICMP6_ND_NA, sizeof(struct rte_flow_item_icmp6_nd_na)),
  78         MK_FLOW_ITEM(ICMP6_ND_OPT, sizeof(struct rte_flow_item_icmp6_nd_opt)),
  79         MK_FLOW_ITEM(ICMP6_ND_OPT_SLA_ETH,
  80                      sizeof(struct rte_flow_item_icmp6_nd_opt_sla_eth)),
  81         MK_FLOW_ITEM(ICMP6_ND_OPT_TLA_ETH,
  82                      sizeof(struct rte_flow_item_icmp6_nd_opt_tla_eth)),
  83         MK_FLOW_ITEM(MARK, sizeof(struct rte_flow_item_mark)),
  84         MK_FLOW_ITEM(META, sizeof(struct rte_flow_item_meta)),
  85         MK_FLOW_ITEM(TAG, sizeof(struct rte_flow_item_tag)),
  86         MK_FLOW_ITEM(GRE_KEY, sizeof(rte_be32_t)),
  87         MK_FLOW_ITEM(GTP_PSC, sizeof(struct rte_flow_item_gtp_psc)),
  88         MK_FLOW_ITEM(PPPOES, sizeof(struct rte_flow_item_pppoe)),
  89         MK_FLOW_ITEM(PPPOED, sizeof(struct rte_flow_item_pppoe)),
  90         MK_FLOW_ITEM(PPPOE_PROTO_ID,
  91                         sizeof(struct rte_flow_item_pppoe_proto_id)),
  92         MK_FLOW_ITEM(NSH, sizeof(struct rte_flow_item_nsh)),
  93         MK_FLOW_ITEM(IGMP, sizeof(struct rte_flow_item_igmp)),
  94         MK_FLOW_ITEM(AH, sizeof(struct rte_flow_item_ah)),
  95         MK_FLOW_ITEM(HIGIG2, sizeof(struct rte_flow_item_higig2_hdr)),
  96 };
  97
  98 /** Generate flow_action[] entry. */
  99 #define MK_FLOW_ACTION(t, s) \
 100         [RTE_FLOW_ACTION_TYPE_ ## t] = { \
 101                 .name = # t, \
 102                 .size = s, \
 103         }
 104
 105 /** Information about known flow actions. */
 106 static const struct rte_flow_desc_data rte_flow_desc_action[] = {
 107         MK_FLOW_ACTION(END, 0),
 108         MK_FLOW_ACTION(VOID, 0),
 109         MK_FLOW_ACTION(PASSTHRU, 0),
 110         MK_FLOW_ACTION(JUMP, sizeof(struct rte_flow_action_jump)),
 111         MK_FLOW_ACTION(MARK, sizeof(struct rte_flow_action_mark)),
 112         MK_FLOW_ACTION(FLAG, 0),
 113         MK_FLOW_ACTION(QUEUE, sizeof(struct rte_flow_action_queue)),
 114         MK_FLOW_ACTION(DROP, 0),
 115         MK_FLOW_ACTION(COUNT, sizeof(struct rte_flow_action_count)),
 116         MK_FLOW_ACTION(RSS, sizeof(struct rte_flow_action_rss)),
 117         MK_FLOW_ACTION(PF, 0),
 118         MK_FLOW_ACTION(VF, sizeof(struct rte_flow_action_vf)),
 119         MK_FLOW_ACTION(PHY_PORT, sizeof(struct rte_flow_action_phy_port)),
 120         MK_FLOW_ACTION(PORT_ID, sizeof(struct rte_flow_action_port_id)),
 121         MK_FLOW_ACTION(METER, sizeof(struct rte_flow_action_meter)),
 122         MK_FLOW_ACTION(SECURITY, sizeof(struct rte_flow_action_security)),
 123         MK_FLOW_ACTION(OF_SET_MPLS_TTL,
 124                        sizeof(struct rte_flow_action_of_set_mpls_ttl)),
 125         MK_FLOW_ACTION(OF_DEC_MPLS_TTL, 0),
 126         MK_FLOW_ACTION(OF_SET_NW_TTL,
 127                        sizeof(struct rte_flow_action_of_set_nw_ttl)),
 128         MK_FLOW_ACTION(OF_DEC_NW_TTL, 0),
 129         MK_FLOW_ACTION(OF_COPY_TTL_OUT, 0),
 130         MK_FLOW_ACTION(OF_COPY_TTL_IN, 0),
 131         MK_FLOW_ACTION(OF_POP_VLAN, 0),
 132         MK_FLOW_ACTION(OF_PUSH_VLAN,
 133                        sizeof(struct rte_flow_action_of_push_vlan)),
 134         MK_FLOW_ACTION(OF_SET_VLAN_VID,
 135                        sizeof(struct rte_flow_action_of_set_vlan_vid)),
 136         MK_FLOW_ACTION(OF_SET_VLAN_PCP,
 137                        sizeof(struct rte_flow_action_of_set_vlan_pcp)),
 138         MK_FLOW_ACTION(OF_POP_MPLS,
 139                        sizeof(struct rte_flow_action_of_pop_mpls)),
 140         MK_FLOW_ACTION(OF_PUSH_MPLS,
 141                        sizeof(struct rte_flow_action_of_push_mpls)),
 142         MK_FLOW_ACTION(VXLAN_ENCAP, sizeof(struct rte_flow_action_vxlan_encap)),
 143         MK_FLOW_ACTION(VXLAN_DECAP, 0),
 144         MK_FLOW_ACTION(NVGRE_ENCAP, sizeof(struct rte_flow_action_vxlan_encap)),
 145         MK_FLOW_ACTION(NVGRE_DECAP, 0),
 146         MK_FLOW_ACTION(RAW_ENCAP, sizeof(struct rte_flow_action_raw_encap)),
 147         MK_FLOW_ACTION(RAW_DECAP, sizeof(struct rte_flow_action_raw_decap)),
 148         MK_FLOW_ACTION(SET_IPV4_SRC,
 149                        sizeof(struct rte_flow_action_set_ipv4)),
 150         MK_FLOW_ACTION(SET_IPV4_DST,
 151                        sizeof(struct rte_flow_action_set_ipv4)),
 152         MK_FLOW_ACTION(SET_IPV6_SRC,
 153                        sizeof(struct rte_flow_action_set_ipv6)),
 154         MK_FLOW_ACTION(SET_IPV6_DST,
 155                        sizeof(struct rte_flow_action_set_ipv6)),
 156         MK_FLOW_ACTION(SET_TP_SRC,
 157                        sizeof(struct rte_flow_action_set_tp)),
 158         MK_FLOW_ACTION(SET_TP_DST,
 159                        sizeof(struct rte_flow_action_set_tp)),
 160         MK_FLOW_ACTION(MAC_SWAP, 0),
 161         MK_FLOW_ACTION(DEC_TTL, 0),
 162         MK_FLOW_ACTION(SET_TTL, sizeof(struct rte_flow_action_set_ttl)),
 163         MK_FLOW_ACTION(SET_MAC_SRC, sizeof(struct rte_flow_action_set_mac)),
 164         MK_FLOW_ACTION(SET_MAC_DST, sizeof(struct rte_flow_action_set_mac)),
 165         MK_FLOW_ACTION(INC_TCP_SEQ, sizeof(rte_be32_t)),
 166         MK_FLOW_ACTION(DEC_TCP_SEQ, sizeof(rte_be32_t)),
 167         MK_FLOW_ACTION(INC_TCP_ACK, sizeof(rte_be32_t)),
 168         MK_FLOW_ACTION(DEC_TCP_ACK, sizeof(rte_be32_t)),
 169         MK_FLOW_ACTION(SET_TAG, sizeof(struct rte_flow_action_set_tag)),
 170         MK_FLOW_ACTION(SET_META, sizeof(struct rte_flow_action_set_meta)),
 171 };
 172
 173 int
 174 rte_flow_dynf_metadata_register(void)
 175 {
 176         int offset;
 177         int flag;
 178
 179         static const struct rte_mbuf_dynfield desc_offs = {
 180                 .name = RTE_MBUF_DYNFIELD_METADATA_NAME,
 181                 .size = sizeof(uint32_t),
 182                 .align = __alignof__(uint32_t),
 183         };
 184         static const struct rte_mbuf_dynflag desc_flag = {
 185                 .name = RTE_MBUF_DYNFLAG_METADATA_NAME,
 186         };
 187
 188         offset = rte_mbuf_dynfield_register(&desc_offs);
 189         if (offset < 0)
 190                 goto error;
 191         flag = rte_mbuf_dynflag_register(&desc_flag);
 192         if (flag < 0)
 193                 goto error;
 194         rte_flow_dynf_metadata_offs = offset;
 195         rte_flow_dynf_metadata_mask = (1ULL << flag);
 196         return 0;
 197
 198 error:
 199         rte_flow_dynf_metadata_offs = -1;
 200         rte_flow_dynf_metadata_mask = 0ULL;
 201         return -rte_errno;
 202 }
 203
 204 static int
 205 flow_err(uint16_t port_id, int ret, struct rte_flow_error *error)
 206 {
 207         if (ret == 0)
 208                 return 0;
 209         if (rte_eth_dev_is_removed(port_id))
 210                 return rte_flow_error_set(error, EIO,
 211                                           RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 212                                           NULL, rte_strerror(EIO));
 213         return ret;
 214 }
 215
 216 /* Get generic flow operations structure from a port. */
 217 const struct rte_flow_ops *
 218 rte_flow_ops_get(uint16_t port_id, struct rte_flow_error *error)
 219 {
 220         struct rte_eth_dev *dev = &rte_eth_devices[port_id];
 221         const struct rte_flow_ops *ops;
 222         int code;
 223
 224         if (unlikely(!rte_eth_dev_is_valid_port(port_id)))
 225                 code = ENODEV;
 226         else if (unlikely(!dev->dev_ops->filter_ctrl ||
 227                           dev->dev_ops->filter_ctrl(dev,
 228                                                     RTE_ETH_FILTER_GENERIC,
 229                                                     RTE_ETH_FILTER_GET,
 230                                                     &ops) ||
 231                           !ops))
 232                 code = ENOSYS;
 233         else
 234                 return ops;
 235         rte_flow_error_set(error, code, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 236                            NULL, rte_strerror(code));
 237         return NULL;
 238 }
 239
 240 /* Check whether a flow rule can be created on a given port. */
 241 int
 242 rte_flow_validate(uint16_t port_id,
 243                   const struct rte_flow_attr *attr,
 244                   const struct rte_flow_item pattern[],
 245                   const struct rte_flow_action actions[],
 246                   struct rte_flow_error *error)
 247 {
 248         const struct rte_flow_ops *ops = rte_flow_ops_get(port_id, error);
 249         struct rte_eth_dev *dev = &rte_eth_devices[port_id];
 250
 251         if (unlikely(!ops))
 252                 return -rte_errno;
 253         if (likely(!!ops->validate))
 254                 return flow_err(port_id, ops->validate(dev, attr, pattern,
 255                                                        actions, error), error);
 256         return rte_flow_error_set(error, ENOSYS,
 257                                   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 258                                   NULL, rte_strerror(ENOSYS));
 259 }
 260
 261 /* Create a flow rule on a given port. */
 262 struct rte_flow *
 263 rte_flow_create(uint16_t port_id,
 264                 const struct rte_flow_attr *attr,
 265                 const struct rte_flow_item pattern[],
 266                 const struct rte_flow_action actions[],
 267                 struct rte_flow_error *error)
 268 {
 269         struct rte_eth_dev *dev = &rte_eth_devices[port_id];
 270         struct rte_flow *flow;
 271         const struct rte_flow_ops *ops = rte_flow_ops_get(port_id, error);
 272
 273         if (unlikely(!ops))
 274                 return NULL;
 275         if (likely(!!ops->create)) {
 276                 flow = ops->create(dev, attr, pattern, actions, error);
 277                 if (flow == NULL)
 278                         flow_err(port_id, -rte_errno, error);
 279                 return flow;
 280         }
 281         rte_flow_error_set(error, ENOSYS, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 282                            NULL, rte_strerror(ENOSYS));
 283         return NULL;
 284 }
 285
 286 /* Destroy a flow rule on a given port. */
 287 int
 288 rte_flow_destroy(uint16_t port_id,
 289                  struct rte_flow *flow,
 290                  struct rte_flow_error *error)
 291 {
 292         struct rte_eth_dev *dev = &rte_eth_devices[port_id];
 293         const struct rte_flow_ops *ops = rte_flow_ops_get(port_id, error);
 294
 295         if (unlikely(!ops))
 296                 return -rte_errno;
 297         if (likely(!!ops->destroy))
 298                 return flow_err(port_id, ops->destroy(dev, flow, error),
 299                                 error);
 300         return rte_flow_error_set(error, ENOSYS,
 301                                   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 302                                   NULL, rte_strerror(ENOSYS));
 303 }
 304
 305 /* Destroy all flow rules associated with a port. */
 306 int
 307 rte_flow_flush(uint16_t port_id,
 308                struct rte_flow_error *error)
 309 {
 310         struct rte_eth_dev *dev = &rte_eth_devices[port_id];
 311         const struct rte_flow_ops *ops = rte_flow_ops_get(port_id, error);
 312
 313         if (unlikely(!ops))
 314                 return -rte_errno;
 315         if (likely(!!ops->flush))
 316                 return flow_err(port_id, ops->flush(dev, error), error);
 317         return rte_flow_error_set(error, ENOSYS,
 318                                   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 319                                   NULL, rte_strerror(ENOSYS));
 320 }
 321
 322 /* Query an existing flow rule. */
 323 int
 324 rte_flow_query(uint16_t port_id,
 325                struct rte_flow *flow,
 326                const struct rte_flow_action *action,
 327                void *data,
 328                struct rte_flow_error *error)
 329 {
 330         struct rte_eth_dev *dev = &rte_eth_devices[port_id];
 331         const struct rte_flow_ops *ops = rte_flow_ops_get(port_id, error);
 332
 333         if (!ops)
 334                 return -rte_errno;
 335         if (likely(!!ops->query))
 336                 return flow_err(port_id, ops->query(dev, flow, action, data,
 337                                                     error), error);
 338         return rte_flow_error_set(error, ENOSYS,
 339                                   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 340                                   NULL, rte_strerror(ENOSYS));
 341 }
 342
 343 /* Restrict ingress traffic to the defined flow rules. */
 344 int
 345 rte_flow_isolate(uint16_t port_id,
 346                  int set,
 347                  struct rte_flow_error *error)
 348 {
 349         struct rte_eth_dev *dev = &rte_eth_devices[port_id];
 350         const struct rte_flow_ops *ops = rte_flow_ops_get(port_id, error);
 351
 352         if (!ops)
 353                 return -rte_errno;
 354         if (likely(!!ops->isolate))
 355                 return flow_err(port_id, ops->isolate(dev, set, error), error);
 356         return rte_flow_error_set(error, ENOSYS,
 357                                   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 358                                   NULL, rte_strerror(ENOSYS));
 359 }
 360
 361 /* Initialize flow error structure. */
 362 int
 363 rte_flow_error_set(struct rte_flow_error *error,
 364                    int code,
 365                    enum rte_flow_error_type type,
 366                    const void *cause,
 367                    const char *message)
 368 {
 369         if (error) {
 370                 *error = (struct rte_flow_error){
 371                         .type = type,
 372                         .cause = cause,
 373                         .message = message,
 374                 };
 375         }
 376         rte_errno = code;
 377         return -code;
 378 }
 379
 380 /** Pattern item specification types. */
 381 enum rte_flow_conv_item_spec_type {
 382         RTE_FLOW_CONV_ITEM_SPEC,
 383         RTE_FLOW_CONV_ITEM_LAST,
 384         RTE_FLOW_CONV_ITEM_MASK,
 385 };
 386
 387 /**
 388  * Copy pattern item specification.
 389  *
 390  * @param[out] buf
 391  *   Output buffer. Can be NULL if @p size is zero.
 392  * @param size
 393  *   Size of @p buf in bytes.
 394  * @param[in] item
 395  *   Pattern item to copy specification from.
 396  * @param type
 397  *   Specification selector for either @p spec, @p last or @p mask.
 398  *
 399  * @return
 400  *   Number of bytes needed to store pattern item specification regardless
 401  *   of @p size. @p buf contents are truncated to @p size if not large
 402  *   enough.
 403  */
 404 static size_t
 405 rte_flow_conv_item_spec(void *buf, const size_t size,
 406                         const struct rte_flow_item *item,
 407                         enum rte_flow_conv_item_spec_type type)
 408 {
 409         size_t off;
 410         const void *data =
 411                 type == RTE_FLOW_CONV_ITEM_SPEC ? item->spec :
 412                 type == RTE_FLOW_CONV_ITEM_LAST ? item->last :
 413                 type == RTE_FLOW_CONV_ITEM_MASK ? item->mask :
 414                 NULL;
 415
 416         switch (item->type) {
 417                 union {
 418                         const struct rte_flow_item_raw *raw;
 419                 } spec;
 420                 union {
 421                         const struct rte_flow_item_raw *raw;
 422                 } last;
 423                 union {
 424                         const struct rte_flow_item_raw *raw;
 425                 } mask;
 426                 union {
 427                         const struct rte_flow_item_raw *raw;
 428                 } src;
 429                 union {
 430                         struct rte_flow_item_raw *raw;
 431                 } dst;
 432                 size_t tmp;
 433
 434         case RTE_FLOW_ITEM_TYPE_RAW:
 435                 spec.raw = item->spec;
 436                 last.raw = item->last ? item->last : item->spec;
 437                 mask.raw = item->mask ? item->mask : &rte_flow_item_raw_mask;
 438                 src.raw = data;
 439                 dst.raw = buf;
 440                 rte_memcpy(dst.raw,
 441                            (&(struct rte_flow_item_raw){
 442                                 .relative = src.raw->relative,
 443                                 .search = src.raw->search,
 444                                 .reserved = src.raw->reserved,
 445                                 .offset = src.raw->offset,
 446                                 .limit = src.raw->limit,
 447                                 .length = src.raw->length,
 448                            }),
 449                            size > sizeof(*dst.raw) ? sizeof(*dst.raw) : size);
 450                 off = sizeof(*dst.raw);
 451                 if (type == RTE_FLOW_CONV_ITEM_SPEC ||
 452                     (type == RTE_FLOW_CONV_ITEM_MASK &&
 453                      ((spec.raw->length & mask.raw->length) >=
 454                       (last.raw->length & mask.raw->length))))
 455                         tmp = spec.raw->length & mask.raw->length;
 456                 else
 457                         tmp = last.raw->length & mask.raw->length;
 458                 if (tmp) {
 459                         off = RTE_ALIGN_CEIL(off, sizeof(*dst.raw->pattern));
 460                         if (size >= off + tmp)
 461                                 dst.raw->pattern = rte_memcpy
 462                                         ((void *)((uintptr_t)dst.raw + off),
 463                                          src.raw->pattern, tmp);
 464                         off += tmp;
 465                 }
 466                 break;
 467         default:
 468                 off = rte_flow_desc_item[item->type].size;
 469                 rte_memcpy(buf, data, (size > off ? off : size));
 470                 break;
 471         }
 472         return off;
 473 }
 474
 475 /**
 476  * Copy action configuration.
 477  *
 478  * @param[out] buf
 479  *   Output buffer. Can be NULL if @p size is zero.
 480  * @param size
 481  *   Size of @p buf in bytes.
 482  * @param[in] action
 483  *   Action to copy configuration from.
 484  *
 485  * @return
 486  *   Number of bytes needed to store pattern item specification regardless
 487  *   of @p size. @p buf contents are truncated to @p size if not large
 488  *   enough.
 489  */
 490 static size_t
 491 rte_flow_conv_action_conf(void *buf, const size_t size,
 492                           const struct rte_flow_action *action)
 493 {
 494         size_t off;
 495
 496         switch (action->type) {
 497                 union {
 498                         const struct rte_flow_action_rss *rss;
 499                         const struct rte_flow_action_vxlan_encap *vxlan_encap;
 500                         const struct rte_flow_action_nvgre_encap *nvgre_encap;
 501                 } src;
 502                 union {
 503                         struct rte_flow_action_rss *rss;
 504                         struct rte_flow_action_vxlan_encap *vxlan_encap;
 505                         struct rte_flow_action_nvgre_encap *nvgre_encap;
 506                 } dst;
 507                 size_t tmp;
 508                 int ret;
 509
 510         case RTE_FLOW_ACTION_TYPE_RSS:
 511                 src.rss = action->conf;
 512                 dst.rss = buf;
 513                 rte_memcpy(dst.rss,
 514                            (&(struct rte_flow_action_rss){
 515                                 .func = src.rss->func,
 516                                 .level = src.rss->level,
 517                                 .types = src.rss->types,
 518                                 .key_len = src.rss->key_len,
 519                                 .queue_num = src.rss->queue_num,
 520                            }),
 521                            size > sizeof(*dst.rss) ? sizeof(*dst.rss) : size);
 522                 off = sizeof(*dst.rss);
 523                 if (src.rss->key_len) {
 524                         off = RTE_ALIGN_CEIL(off, sizeof(*dst.rss->key));
 525                         tmp = sizeof(*src.rss->key) * src.rss->key_len;
 526                         if (size >= off + tmp)
 527                                 dst.rss->key = rte_memcpy
 528                                         ((void *)((uintptr_t)dst.rss + off),
 529                                          src.rss->key, tmp);
 530                         off += tmp;
 531                 }
 532                 if (src.rss->queue_num) {
 533                         off = RTE_ALIGN_CEIL(off, sizeof(*dst.rss->queue));
 534                         tmp = sizeof(*src.rss->queue) * src.rss->queue_num;
 535                         if (size >= off + tmp)
 536                                 dst.rss->queue = rte_memcpy
 537                                         ((void *)((uintptr_t)dst.rss + off),
 538                                          src.rss->queue, tmp);
 539                         off += tmp;
 540                 }
 541                 break;
 542         case RTE_FLOW_ACTION_TYPE_VXLAN_ENCAP:
 543         case RTE_FLOW_ACTION_TYPE_NVGRE_ENCAP:
 544                 src.vxlan_encap = action->conf;
 545                 dst.vxlan_encap = buf;
 546                 RTE_BUILD_BUG_ON(sizeof(*src.vxlan_encap) !=
 547                                  sizeof(*src.nvgre_encap) ||
 548                                  offsetof(struct rte_flow_action_vxlan_encap,
 549                                           definition) !=
 550                                  offsetof(struct rte_flow_action_nvgre_encap,
 551                                           definition));
 552                 off = sizeof(*dst.vxlan_encap);
 553                 if (src.vxlan_encap->definition) {
 554                         off = RTE_ALIGN_CEIL
 555                                 (off, sizeof(*dst.vxlan_encap->definition));
 556                         ret = rte_flow_conv
 557                                 (RTE_FLOW_CONV_OP_PATTERN,
 558                                  (void *)((uintptr_t)dst.vxlan_encap + off),
 559                                  size > off ? size - off : 0,
 560                                  src.vxlan_encap->definition, NULL);
 561                         if (ret < 0)
 562                                 return 0;
 563                         if (size >= off + ret)
 564                                 dst.vxlan_encap->definition =
 565                                         (void *)((uintptr_t)dst.vxlan_encap +
 566                                                  off);
 567                         off += ret;
 568                 }
 569                 break;
 570         default:
 571                 off = rte_flow_desc_action[action->type].size;
 572                 rte_memcpy(buf, action->conf, (size > off ? off : size));
 573                 break;
 574         }
 575         return off;
 576 }
 577
 578 /**
 579  * Copy a list of pattern items.
 580  *
 581  * @param[out] dst
 582  *   Destination buffer. Can be NULL if @p size is zero.
 583  * @param size
 584  *   Size of @p dst in bytes.
 585  * @param[in] src
 586  *   Source pattern items.
 587  * @param num
 588  *   Maximum number of pattern items to process from @p src or 0 to process
 589  *   the entire list. In both cases, processing stops after
 590  *   RTE_FLOW_ITEM_TYPE_END is encountered.
 591  * @param[out] error
 592  *   Perform verbose error reporting if not NULL.
 593  *
 594  * @return
 595  *   A positive value representing the number of bytes needed to store
 596  *   pattern items regardless of @p size on success (@p buf contents are
 597  *   truncated to @p size if not large enough), a negative errno value
 598  *   otherwise and rte_errno is set.
 599  */
 600 static int
 601 rte_flow_conv_pattern(struct rte_flow_item *dst,
 602                       const size_t size,
 603                       const struct rte_flow_item *src,
 604                       unsigned int num,
 605                       struct rte_flow_error *error)
 606 {
 607         uintptr_t data = (uintptr_t)dst;
 608         size_t off;
 609         size_t ret;
 610         unsigned int i;
 611
 612         for (i = 0, off = 0; !num || i != num; ++i, ++src, ++dst) {
 613                 if ((size_t)src->type >= RTE_DIM(rte_flow_desc_item) ||
 614                     !rte_flow_desc_item[src->type].name)
 615                         return rte_flow_error_set
 616                                 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM, src,
 617                                  "cannot convert unknown item type");
 618                 if (size >= off + sizeof(*dst))
 619                         *dst = (struct rte_flow_item){
 620                                 .type = src->type,
 621                         };
 622                 off += sizeof(*dst);
 623                 if (!src->type)
 624                         num = i + 1;
 625         }
 626         num = i;
 627         src -= num;
 628         dst -= num;
 629         do {
 630                 if (src->spec) {
 631                         off = RTE_ALIGN_CEIL(off, sizeof(double));
 632                         ret = rte_flow_conv_item_spec
 633                                 ((void *)(data + off),
 634                                  size > off ? size - off : 0, src,
 635                                  RTE_FLOW_CONV_ITEM_SPEC);
 636                         if (size && size >= off + ret)
 637                                 dst->spec = (void *)(data + off);
 638                         off += ret;
 639
 640                 }
 641                 if (src->last) {
 642                         off = RTE_ALIGN_CEIL(off, sizeof(double));
 643                         ret = rte_flow_conv_item_spec
 644                                 ((void *)(data + off),
 645                                  size > off ? size - off : 0, src,
 646                                  RTE_FLOW_CONV_ITEM_LAST);
 647                         if (size && size >= off + ret)
 648                                 dst->last = (void *)(data + off);
 649                         off += ret;
 650                 }
 651                 if (src->mask) {
 652                         off = RTE_ALIGN_CEIL(off, sizeof(double));
 653                         ret = rte_flow_conv_item_spec
 654                                 ((void *)(data + off),
 655                                  size > off ? size - off : 0, src,
 656                                  RTE_FLOW_CONV_ITEM_MASK);
 657                         if (size && size >= off + ret)
 658                                 dst->mask = (void *)(data + off);
 659                         off += ret;
 660                 }
 661                 ++src;
 662                 ++dst;
 663         } while (--num);
 664         return off;
 665 }
 666
 667 /**
 668  * Copy a list of actions.
 669  *
 670  * @param[out] dst
 671  *   Destination buffer. Can be NULL if @p size is zero.
 672  * @param size
 673  *   Size of @p dst in bytes.
 674  * @param[in] src
 675  *   Source actions.
 676  * @param num
 677  *   Maximum number of actions to process from @p src or 0 to process the
 678  *   entire list. In both cases, processing stops after
 679  *   RTE_FLOW_ACTION_TYPE_END is encountered.
 680  * @param[out] error
 681  *   Perform verbose error reporting if not NULL.
 682  *
 683  * @return
 684  *   A positive value representing the number of bytes needed to store
 685  *   actions regardless of @p size on success (@p buf contents are truncated
 686  *   to @p size if not large enough), a negative errno value otherwise and
 687  *   rte_errno is set.
 688  */
 689 static int
 690 rte_flow_conv_actions(struct rte_flow_action *dst,
 691                       const size_t size,
 692                       const struct rte_flow_action *src,
 693                       unsigned int num,
 694                       struct rte_flow_error *error)
 695 {
 696         uintptr_t data = (uintptr_t)dst;
 697         size_t off;
 698         size_t ret;
 699         unsigned int i;
 700
 701         for (i = 0, off = 0; !num || i != num; ++i, ++src, ++dst) {
 702                 if ((size_t)src->type >= RTE_DIM(rte_flow_desc_action) ||
 703                     !rte_flow_desc_action[src->type].name)
 704                         return rte_flow_error_set
 705                                 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 706                                  src, "cannot convert unknown action type");
 707                 if (size >= off + sizeof(*dst))
 708                         *dst = (struct rte_flow_action){
 709                                 .type = src->type,
 710                         };
 711                 off += sizeof(*dst);
 712                 if (!src->type)
 713                         num = i + 1;
 714         }
 715         num = i;
 716         src -= num;
 717         dst -= num;
 718         do {
 719                 if (src->conf) {
 720                         off = RTE_ALIGN_CEIL(off, sizeof(double));
 721                         ret = rte_flow_conv_action_conf
 722                                 ((void *)(data + off),
 723                                  size > off ? size - off : 0, src);
 724                         if (size && size >= off + ret)
 725                                 dst->conf = (void *)(data + off);
 726                         off += ret;
 727                 }
 728                 ++src;
 729                 ++dst;
 730         } while (--num);
 731         return off;
 732 }
 733
 734 /**
 735  * Copy flow rule components.
 736  *
 737  * This comprises the flow rule descriptor itself, attributes, pattern and
 738  * actions list. NULL components in @p src are skipped.
 739  *
 740  * @param[out] dst
 741  *   Destination buffer. Can be NULL if @p size is zero.
 742  * @param size
 743  *   Size of @p dst in bytes.
 744  * @param[in] src
 745  *   Source flow rule descriptor.
 746  * @param[out] error
 747  *   Perform verbose error reporting if not NULL.
 748  *
 749  * @return
 750  *   A positive value representing the number of bytes needed to store all
 751  *   components including the descriptor regardless of @p size on success
 752  *   (@p buf contents are truncated to @p size if not large enough), a
 753  *   negative errno value otherwise and rte_errno is set.
 754  */
 755 static int
 756 rte_flow_conv_rule(struct rte_flow_conv_rule *dst,
 757                    const size_t size,
 758                    const struct rte_flow_conv_rule *src,
 759                    struct rte_flow_error *error)
 760 {
 761         size_t off;
 762         int ret;
 763
 764         rte_memcpy(dst,
 765                    (&(struct rte_flow_conv_rule){
 766                         .attr = NULL,
 767                         .pattern = NULL,
 768                         .actions = NULL,
 769                    }),
 770                    size > sizeof(*dst) ? sizeof(*dst) : size);
 771         off = sizeof(*dst);
 772         if (src->attr_ro) {
 773                 off = RTE_ALIGN_CEIL(off, sizeof(double));
 774                 if (size && size >= off + sizeof(*dst->attr))
 775                         dst->attr = rte_memcpy
 776                                 ((void *)((uintptr_t)dst + off),
 777                                  src->attr_ro, sizeof(*dst->attr));
 778                 off += sizeof(*dst->attr);
 779         }
 780         if (src->pattern_ro) {
 781                 off = RTE_ALIGN_CEIL(off, sizeof(double));
 782                 ret = rte_flow_conv_pattern((void *)((uintptr_t)dst + off),
 783                                             size > off ? size - off : 0,
 784                                             src->pattern_ro, 0, error);
 785                 if (ret < 0)
 786                         return ret;
 787                 if (size && size >= off + (size_t)ret)
 788                         dst->pattern = (void *)((uintptr_t)dst + off);
 789                 off += ret;
 790         }
 791         if (src->actions_ro) {
 792                 off = RTE_ALIGN_CEIL(off, sizeof(double));
 793                 ret = rte_flow_conv_actions((void *)((uintptr_t)dst + off),
 794                                             size > off ? size - off : 0,
 795                                             src->actions_ro, 0, error);
 796                 if (ret < 0)
 797                         return ret;
 798                 if (size >= off + (size_t)ret)
 799                         dst->actions = (void *)((uintptr_t)dst + off);
 800                 off += ret;
 801         }
 802         return off;
 803 }
 804
 805 /**
 806  * Retrieve the name of a pattern item/action type.
 807  *
 808  * @param is_action
 809  *   Nonzero when @p src represents an action type instead of a pattern item
 810  *   type.
 811  * @param is_ptr
 812  *   Nonzero to write string address instead of contents into @p dst.
 813  * @param[out] dst
 814  *   Destination buffer. Can be NULL if @p size is zero.
 815  * @param size
 816  *   Size of @p dst in bytes.
 817  * @param[in] src
 818  *   Depending on @p is_action, source pattern item or action type cast as a
 819  *   pointer.
 820  * @param[out] error
 821  *   Perform verbose error reporting if not NULL.
 822  *
 823  * @return
 824  *   A positive value representing the number of bytes needed to store the
 825  *   name or its address regardless of @p size on success (@p buf contents
 826  *   are truncated to @p size if not large enough), a negative errno value
 827  *   otherwise and rte_errno is set.
 828  */
 829 static int
 830 rte_flow_conv_name(int is_action,
 831                    int is_ptr,
 832                    char *dst,
 833                    const size_t size,
 834                    const void *src,
 835                    struct rte_flow_error *error)
 836 {
 837         struct desc_info {
 838                 const struct rte_flow_desc_data *data;
 839                 size_t num;
 840         };
 841         static const struct desc_info info_rep[2] = {
 842                 { rte_flow_desc_item, RTE_DIM(rte_flow_desc_item), },
 843                 { rte_flow_desc_action, RTE_DIM(rte_flow_desc_action), },
 844         };
 845         const struct desc_info *const info = &info_rep[!!is_action];
 846         unsigned int type = (uintptr_t)src;
 847
 848         if (type >= info->num)
 849                 return rte_flow_error_set
 850                         (error, EINVAL, RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
 851                          "unknown object type to retrieve the name of");
 852         if (!is_ptr)
 853                 return strlcpy(dst, info->data[type].name, size);
 854         if (size >= sizeof(const char **))
 855                 *((const char **)dst) = info->data[type].name;
 856         return sizeof(const char **);
 857 }
 858
 859 /** Helper function to convert flow API objects. */
 860 int
 861 rte_flow_conv(enum rte_flow_conv_op op,
 862               void *dst,
 863               size_t size,
 864               const void *src,
 865               struct rte_flow_error *error)
 866 {
 867         switch (op) {
 868                 const struct rte_flow_attr *attr;
 869
 870         case RTE_FLOW_CONV_OP_NONE:
 871                 return 0;
 872         case RTE_FLOW_CONV_OP_ATTR:
 873                 attr = src;
 874                 if (size > sizeof(*attr))
 875                         size = sizeof(*attr);
 876                 rte_memcpy(dst, attr, size);
 877                 return sizeof(*attr);
 878         case RTE_FLOW_CONV_OP_ITEM:
 879                 return rte_flow_conv_pattern(dst, size, src, 1, error);
 880         case RTE_FLOW_CONV_OP_ACTION:
 881                 return rte_flow_conv_actions(dst, size, src, 1, error);
 882         case RTE_FLOW_CONV_OP_PATTERN:
 883                 return rte_flow_conv_pattern(dst, size, src, 0, error);
 884         case RTE_FLOW_CONV_OP_ACTIONS:
 885                 return rte_flow_conv_actions(dst, size, src, 0, error);
 886         case RTE_FLOW_CONV_OP_RULE:
 887                 return rte_flow_conv_rule(dst, size, src, error);
 888         case RTE_FLOW_CONV_OP_ITEM_NAME:
 889                 return rte_flow_conv_name(0, 0, dst, size, src, error);
 890         case RTE_FLOW_CONV_OP_ACTION_NAME:
 891                 return rte_flow_conv_name(1, 0, dst, size, src, error);
 892         case RTE_FLOW_CONV_OP_ITEM_NAME_PTR:
 893                 return rte_flow_conv_name(0, 1, dst, size, src, error);
 894         case RTE_FLOW_CONV_OP_ACTION_NAME_PTR:
 895                 return rte_flow_conv_name(1, 1, dst, size, src, error);
 896         }
 897         return rte_flow_error_set
 898                 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
 899                  "unknown object conversion operation");
 900 }
 901
 902 /** Store a full rte_flow description. */
 903 size_t
 904 rte_flow_copy(struct rte_flow_desc *desc, size_t len,
 905               const struct rte_flow_attr *attr,
 906               const struct rte_flow_item *items,
 907               const struct rte_flow_action *actions)
 908 {
 909         /*
 910          * Overlap struct rte_flow_conv with struct rte_flow_desc in order
 911          * to convert the former to the latter without wasting space.
 912          */
 913         struct rte_flow_conv_rule *dst =
 914                 len ?
 915                 (void *)((uintptr_t)desc +
 916                          (offsetof(struct rte_flow_desc, actions) -
 917                           offsetof(struct rte_flow_conv_rule, actions))) :
 918                 NULL;
 919         size_t dst_size =
 920                 len > sizeof(*desc) - sizeof(*dst) ?
 921                 len - (sizeof(*desc) - sizeof(*dst)) :
 922                 0;
 923         struct rte_flow_conv_rule src = {
 924                 .attr_ro = NULL,
 925                 .pattern_ro = items,
 926                 .actions_ro = actions,
 927         };
 928         int ret;
 929
 930         RTE_BUILD_BUG_ON(sizeof(struct rte_flow_desc) <
 931                          sizeof(struct rte_flow_conv_rule));
 932         if (dst_size &&
 933             (&dst->pattern != &desc->items ||
 934              &dst->actions != &desc->actions ||
 935              (uintptr_t)(dst + 1) != (uintptr_t)(desc + 1))) {
 936                 rte_errno = EINVAL;
 937                 return 0;
 938         }
 939         ret = rte_flow_conv(RTE_FLOW_CONV_OP_RULE, dst, dst_size, &src, NULL);
 940         if (ret < 0)
 941                 return 0;
 942         ret += sizeof(*desc) - sizeof(*dst);
 943         rte_memcpy(desc,
 944                    (&(struct rte_flow_desc){
 945                         .size = ret,
 946                         .attr = *attr,
 947                         .items = dst_size ? dst->pattern : NULL,
 948                         .actions = dst_size ? dst->actions : NULL,
 949                    }),
 950                    len > sizeof(*desc) ? sizeof(*desc) : len);
 951         return ret;
 952 }
 953
 954 /**
 955  * Expand RSS flows into several possible flows according to the RSS hash
 956  * fields requested and the driver capabilities.
 957  */
 958 int
 959 rte_flow_expand_rss(struct rte_flow_expand_rss *buf, size_t size,
 960                     const struct rte_flow_item *pattern, uint64_t types,
 961                     const struct rte_flow_expand_node graph[],
 962                     int graph_root_index)
 963 {
 964         const int elt_n = 8;
 965         const struct rte_flow_item *item;
 966         const struct rte_flow_expand_node *node = &graph[graph_root_index];
 967         const int *next_node;
 968         const int *stack[elt_n];
 969         int stack_pos = 0;
 970         struct rte_flow_item flow_items[elt_n];
 971         unsigned int i;
 972         size_t lsize;
 973         size_t user_pattern_size = 0;
 974         void *addr = NULL;
 975
 976         lsize = offsetof(struct rte_flow_expand_rss, entry) +
 977                 elt_n * sizeof(buf->entry[0]);
 978         if (lsize <= size) {
 979                 buf->entry[0].priority = 0;
 980                 buf->entry[0].pattern = (void *)&buf->entry[elt_n];
 981                 buf->entries = 0;
 982                 addr = buf->entry[0].pattern;
 983         }
 984         for (item = pattern; item->type != RTE_FLOW_ITEM_TYPE_END; item++) {
 985                 const struct rte_flow_expand_node *next = NULL;
 986
 987                 for (i = 0; node->next && node->next[i]; ++i) {
 988                         next = &graph[node->next[i]];
 989                         if (next->type == item->type)
 990                                 break;
 991                 }
 992                 if (next)
 993                         node = next;
 994                 user_pattern_size += sizeof(*item);
 995         }
 996         user_pattern_size += sizeof(*item); /* Handle END item. */
 997         lsize += user_pattern_size;
 998         /* Copy the user pattern in the first entry of the buffer. */
 999         if (lsize <= size) {
1000                 rte_memcpy(addr, pattern, user_pattern_size);
1001                 addr = (void *)(((uintptr_t)addr) + user_pattern_size);
1002                 buf->entries = 1;
1003         }
1004         /* Start expanding. */
1005         memset(flow_items, 0, sizeof(flow_items));
1006         user_pattern_size -= sizeof(*item);
1007         next_node = node->next;
1008         stack[stack_pos] = next_node;
1009         node = next_node ? &graph[*next_node] : NULL;
1010         while (node) {
1011                 flow_items[stack_pos].type = node->type;
1012                 if (node->rss_types & types) {
1013                         /*
1014                          * compute the number of items to copy from the
1015                          * expansion and copy it.
1016                          * When the stack_pos is 0, there are 1 element in it,
1017                          * plus the addition END item.
1018                          */
1019                         int elt = stack_pos + 2;
1020
1021                         flow_items[stack_pos + 1].type = RTE_FLOW_ITEM_TYPE_END;
1022                         lsize += elt * sizeof(*item) + user_pattern_size;
1023                         if (lsize <= size) {
1024                                 size_t n = elt * sizeof(*item);
1025
1026                                 buf->entry[buf->entries].priority =
1027                                         stack_pos + 1;
1028                                 buf->entry[buf->entries].pattern = addr;
1029                                 buf->entries++;
1030                                 rte_memcpy(addr, buf->entry[0].pattern,
1031                                            user_pattern_size);
1032                                 addr = (void *)(((uintptr_t)addr) +
1033                                                 user_pattern_size);
1034                                 rte_memcpy(addr, flow_items, n);
1035                                 addr = (void *)(((uintptr_t)addr) + n);
1036                         }
1037                 }
1038                 /* Go deeper. */
1039                 if (node->next) {
1040                         next_node = node->next;
1041                         if (stack_pos++ == elt_n) {
1042                                 rte_errno = E2BIG;
1043                                 return -rte_errno;
1044                         }
1045                         stack[stack_pos] = next_node;
1046                 } else if (*(next_node + 1)) {
1047                         /* Follow up with the next possibility. */
1048                         ++next_node;
1049                 } else {
1050                         /* Move to the next path. */
1051                         if (stack_pos)
1052                                 next_node = stack[--stack_pos];
1053                         next_node++;
1054                         stack[stack_pos] = next_node;
1055                 }
1056                 node = *next_node ? &graph[*next_node] : NULL;
1057         };
1058         return lsize;
1059 }