tools/lib/bpf/relo_core.c

   1 // SPDX-License-Identifier: (LGPL-2.1 OR BSD-2-Clause)
   2 /* Copyright (c) 2019 Facebook */
   3
   4 #ifdef __KERNEL__
   5 #include <linux/bpf.h>
   6 #include <linux/btf.h>
   7 #include <linux/string.h>
   8 #include <linux/bpf_verifier.h>
   9 #include "relo_core.h"
  10
  11 static const char *btf_kind_str(const struct btf_type *t)
  12 {
  13         return btf_type_str(t);
  14 }
  15
  16 static bool is_ldimm64_insn(struct bpf_insn *insn)
  17 {
  18         return insn->code == (BPF_LD | BPF_IMM | BPF_DW);
  19 }
  20
  21 static const struct btf_type *
  22 skip_mods_and_typedefs(const struct btf *btf, u32 id, u32 *res_id)
  23 {
  24         return btf_type_skip_modifiers(btf, id, res_id);
  25 }
  26
  27 static const char *btf__name_by_offset(const struct btf *btf, u32 offset)
  28 {
  29         return btf_name_by_offset(btf, offset);
  30 }
  31
  32 static s64 btf__resolve_size(const struct btf *btf, u32 type_id)
  33 {
  34         const struct btf_type *t;
  35         int size;
  36
  37         t = btf_type_by_id(btf, type_id);
  38         t = btf_resolve_size(btf, t, &size);
  39         if (IS_ERR(t))
  40                 return PTR_ERR(t);
  41         return size;
  42 }
  43
  44 enum libbpf_print_level {
  45         LIBBPF_WARN,
  46         LIBBPF_INFO,
  47         LIBBPF_DEBUG,
  48 };
  49
  50 #undef pr_warn
  51 #undef pr_info
  52 #undef pr_debug
  53 #define pr_warn(fmt, log, ...)  bpf_log((void *)log, fmt, "", ##__VA_ARGS__)
  54 #define pr_info(fmt, log, ...)  bpf_log((void *)log, fmt, "", ##__VA_ARGS__)
  55 #define pr_debug(fmt, log, ...) bpf_log((void *)log, fmt, "", ##__VA_ARGS__)
  56 #define libbpf_print(level, fmt, ...)   bpf_log((void *)prog_name, fmt, ##__VA_ARGS__)
  57 #else
  58 #include <stdio.h>
  59 #include <string.h>
  60 #include <errno.h>
  61 #include <ctype.h>
  62 #include <linux/err.h>
  63
  64 #include "libbpf.h"
  65 #include "bpf.h"
  66 #include "btf.h"
  67 #include "str_error.h"
  68 #include "libbpf_internal.h"
  69 #endif
  70
  71 static bool is_flex_arr(const struct btf *btf,
  72                         const struct bpf_core_accessor *acc,
  73                         const struct btf_array *arr)
  74 {
  75         const struct btf_type *t;
  76
  77         /* not a flexible array, if not inside a struct or has non-zero size */
  78         if (!acc->name || arr->nelems > 0)
  79                 return false;
  80
  81         /* has to be the last member of enclosing struct */
  82         t = btf_type_by_id(btf, acc->type_id);
  83         return acc->idx == btf_vlen(t) - 1;
  84 }
  85
  86 static const char *core_relo_kind_str(enum bpf_core_relo_kind kind)
  87 {
  88         switch (kind) {
  89         case BPF_CORE_FIELD_BYTE_OFFSET: return "byte_off";
  90         case BPF_CORE_FIELD_BYTE_SIZE: return "byte_sz";
  91         case BPF_CORE_FIELD_EXISTS: return "field_exists";
  92         case BPF_CORE_FIELD_SIGNED: return "signed";
  93         case BPF_CORE_FIELD_LSHIFT_U64: return "lshift_u64";
  94         case BPF_CORE_FIELD_RSHIFT_U64: return "rshift_u64";
  95         case BPF_CORE_TYPE_ID_LOCAL: return "local_type_id";
  96         case BPF_CORE_TYPE_ID_TARGET: return "target_type_id";
  97         case BPF_CORE_TYPE_EXISTS: return "type_exists";
  98         case BPF_CORE_TYPE_MATCHES: return "type_matches";
  99         case BPF_CORE_TYPE_SIZE: return "type_size";
 100         case BPF_CORE_ENUMVAL_EXISTS: return "enumval_exists";
 101         case BPF_CORE_ENUMVAL_VALUE: return "enumval_value";
 102         default: return "unknown";
 103         }
 104 }
 105
 106 static bool core_relo_is_field_based(enum bpf_core_relo_kind kind)
 107 {
 108         switch (kind) {
 109         case BPF_CORE_FIELD_BYTE_OFFSET:
 110         case BPF_CORE_FIELD_BYTE_SIZE:
 111         case BPF_CORE_FIELD_EXISTS:
 112         case BPF_CORE_FIELD_SIGNED:
 113         case BPF_CORE_FIELD_LSHIFT_U64:
 114         case BPF_CORE_FIELD_RSHIFT_U64:
 115                 return true;
 116         default:
 117                 return false;
 118         }
 119 }
 120
 121 static bool core_relo_is_type_based(enum bpf_core_relo_kind kind)
 122 {
 123         switch (kind) {
 124         case BPF_CORE_TYPE_ID_LOCAL:
 125         case BPF_CORE_TYPE_ID_TARGET:
 126         case BPF_CORE_TYPE_EXISTS:
 127         case BPF_CORE_TYPE_MATCHES:
 128         case BPF_CORE_TYPE_SIZE:
 129                 return true;
 130         default:
 131                 return false;
 132         }
 133 }
 134
 135 static bool core_relo_is_enumval_based(enum bpf_core_relo_kind kind)
 136 {
 137         switch (kind) {
 138         case BPF_CORE_ENUMVAL_EXISTS:
 139         case BPF_CORE_ENUMVAL_VALUE:
 140                 return true;
 141         default:
 142                 return false;
 143         }
 144 }
 145
 146 int __bpf_core_types_are_compat(const struct btf *local_btf, __u32 local_id,
 147                                 const struct btf *targ_btf, __u32 targ_id, int level)
 148 {
 149         const struct btf_type *local_type, *targ_type;
 150         int depth = 32; /* max recursion depth */
 151
 152         /* caller made sure that names match (ignoring flavor suffix) */
 153         local_type = btf_type_by_id(local_btf, local_id);
 154         targ_type = btf_type_by_id(targ_btf, targ_id);
 155         if (!btf_kind_core_compat(local_type, targ_type))
 156                 return 0;
 157
 158 recur:
 159         depth--;
 160         if (depth < 0)
 161                 return -EINVAL;
 162
 163         local_type = skip_mods_and_typedefs(local_btf, local_id, &local_id);
 164         targ_type = skip_mods_and_typedefs(targ_btf, targ_id, &targ_id);
 165         if (!local_type || !targ_type)
 166                 return -EINVAL;
 167
 168         if (!btf_kind_core_compat(local_type, targ_type))
 169                 return 0;
 170
 171         switch (btf_kind(local_type)) {
 172         case BTF_KIND_UNKN:
 173         case BTF_KIND_STRUCT:
 174         case BTF_KIND_UNION:
 175         case BTF_KIND_ENUM:
 176         case BTF_KIND_FWD:
 177         case BTF_KIND_ENUM64:
 178                 return 1;
 179         case BTF_KIND_INT:
 180                 /* just reject deprecated bitfield-like integers; all other
 181                  * integers are by default compatible between each other
 182                  */
 183                 return btf_int_offset(local_type) == 0 && btf_int_offset(targ_type) == 0;
 184         case BTF_KIND_PTR:
 185                 local_id = local_type->type;
 186                 targ_id = targ_type->type;
 187                 goto recur;
 188         case BTF_KIND_ARRAY:
 189                 local_id = btf_array(local_type)->type;
 190                 targ_id = btf_array(targ_type)->type;
 191                 goto recur;
 192         case BTF_KIND_FUNC_PROTO: {
 193                 struct btf_param *local_p = btf_params(local_type);
 194                 struct btf_param *targ_p = btf_params(targ_type);
 195                 __u16 local_vlen = btf_vlen(local_type);
 196                 __u16 targ_vlen = btf_vlen(targ_type);
 197                 int i, err;
 198
 199                 if (local_vlen != targ_vlen)
 200                         return 0;
 201
 202                 for (i = 0; i < local_vlen; i++, local_p++, targ_p++) {
 203                         if (level <= 0)
 204                                 return -EINVAL;
 205
 206                         skip_mods_and_typedefs(local_btf, local_p->type, &local_id);
 207                         skip_mods_and_typedefs(targ_btf, targ_p->type, &targ_id);
 208                         err = __bpf_core_types_are_compat(local_btf, local_id, targ_btf, targ_id,
 209                                                           level - 1);
 210                         if (err <= 0)
 211                                 return err;
 212                 }
 213
 214                 /* tail recurse for return type check */
 215                 skip_mods_and_typedefs(local_btf, local_type->type, &local_id);
 216                 skip_mods_and_typedefs(targ_btf, targ_type->type, &targ_id);
 217                 goto recur;
 218         }
 219         default:
 220                 pr_warn("unexpected kind %s relocated, local [%d], target [%d]\n",
 221                         btf_kind_str(local_type), local_id, targ_id);
 222                 return 0;
 223         }
 224 }
 225
 226 /*
 227  * Turn bpf_core_relo into a low- and high-level spec representation,
 228  * validating correctness along the way, as well as calculating resulting
 229  * field bit offset, specified by accessor string. Low-level spec captures
 230  * every single level of nestedness, including traversing anonymous
 231  * struct/union members. High-level one only captures semantically meaningful
 232  * "turning points": named fields and array indicies.
 233  * E.g., for this case:
 234  *
 235  *   struct sample {
 236  *       int __unimportant;
 237  *       struct {
 238  *           int __1;
 239  *           int __2;
 240  *           int a[7];
 241  *       };
 242  *   };
 243  *
 244  *   struct sample *s = ...;
 245  *
 246  *   int x = &s->a[3]; // access string = '0:1:2:3'
 247  *
 248  * Low-level spec has 1:1 mapping with each element of access string (it's
 249  * just a parsed access string representation): [0, 1, 2, 3].
 250  *
 251  * High-level spec will capture only 3 points:
 252  *   - initial zero-index access by pointer (&s->... is the same as &s[0]...);
 253  *   - field 'a' access (corresponds to '2' in low-level spec);
 254  *   - array element #3 access (corresponds to '3' in low-level spec).
 255  *
 256  * Type-based relocations (TYPE_EXISTS/TYPE_MATCHES/TYPE_SIZE,
 257  * TYPE_ID_LOCAL/TYPE_ID_TARGET) don't capture any field information. Their
 258  * spec and raw_spec are kept empty.
 259  *
 260  * Enum value-based relocations (ENUMVAL_EXISTS/ENUMVAL_VALUE) use access
 261  * string to specify enumerator's value index that need to be relocated.
 262  */
 263 int bpf_core_parse_spec(const char *prog_name, const struct btf *btf,
 264                         const struct bpf_core_relo *relo,
 265                         struct bpf_core_spec *spec)
 266 {
 267         int access_idx, parsed_len, i;
 268         struct bpf_core_accessor *acc;
 269         const struct btf_type *t;
 270         const char *name, *spec_str;
 271         __u32 id, name_off;
 272         __s64 sz;
 273
 274         spec_str = btf__name_by_offset(btf, relo->access_str_off);
 275         if (str_is_empty(spec_str) || *spec_str == ':')
 276                 return -EINVAL;
 277
 278         memset(spec, 0, sizeof(*spec));
 279         spec->btf = btf;
 280         spec->root_type_id = relo->type_id;
 281         spec->relo_kind = relo->kind;
 282
 283         /* type-based relocations don't have a field access string */
 284         if (core_relo_is_type_based(relo->kind)) {
 285                 if (strcmp(spec_str, "0"))
 286                         return -EINVAL;
 287                 return 0;
 288         }
 289
 290         /* parse spec_str="0:1:2:3:4" into array raw_spec=[0, 1, 2, 3, 4] */
 291         while (*spec_str) {
 292                 if (*spec_str == ':')
 293                         ++spec_str;
 294                 if (sscanf(spec_str, "%d%n", &access_idx, &parsed_len) != 1)
 295                         return -EINVAL;
 296                 if (spec->raw_len == BPF_CORE_SPEC_MAX_LEN)
 297                         return -E2BIG;
 298                 spec_str += parsed_len;
 299                 spec->raw_spec[spec->raw_len++] = access_idx;
 300         }
 301
 302         if (spec->raw_len == 0)
 303                 return -EINVAL;
 304
 305         t = skip_mods_and_typedefs(btf, relo->type_id, &id);
 306         if (!t)
 307                 return -EINVAL;
 308
 309         access_idx = spec->raw_spec[0];
 310         acc = &spec->spec[0];
 311         acc->type_id = id;
 312         acc->idx = access_idx;
 313         spec->len++;
 314
 315         if (core_relo_is_enumval_based(relo->kind)) {
 316                 if (!btf_is_any_enum(t) || spec->raw_len > 1 || access_idx >= btf_vlen(t))
 317                         return -EINVAL;
 318
 319                 /* record enumerator name in a first accessor */
 320                 name_off = btf_is_enum(t) ? btf_enum(t)[access_idx].name_off
 321                                           : btf_enum64(t)[access_idx].name_off;
 322                 acc->name = btf__name_by_offset(btf, name_off);
 323                 return 0;
 324         }
 325
 326         if (!core_relo_is_field_based(relo->kind))
 327                 return -EINVAL;
 328
 329         sz = btf__resolve_size(btf, id);
 330         if (sz < 0)
 331                 return sz;
 332         spec->bit_offset = access_idx * sz * 8;
 333
 334         for (i = 1; i < spec->raw_len; i++) {
 335                 t = skip_mods_and_typedefs(btf, id, &id);
 336                 if (!t)
 337                         return -EINVAL;
 338
 339                 access_idx = spec->raw_spec[i];
 340                 acc = &spec->spec[spec->len];
 341
 342                 if (btf_is_composite(t)) {
 343                         const struct btf_member *m;
 344                         __u32 bit_offset;
 345
 346                         if (access_idx >= btf_vlen(t))
 347                                 return -EINVAL;
 348
 349                         bit_offset = btf_member_bit_offset(t, access_idx);
 350                         spec->bit_offset += bit_offset;
 351
 352                         m = btf_members(t) + access_idx;
 353                         if (m->name_off) {
 354                                 name = btf__name_by_offset(btf, m->name_off);
 355                                 if (str_is_empty(name))
 356                                         return -EINVAL;
 357
 358                                 acc->type_id = id;
 359                                 acc->idx = access_idx;
 360                                 acc->name = name;
 361                                 spec->len++;
 362                         }
 363
 364                         id = m->type;
 365                 } else if (btf_is_array(t)) {
 366                         const struct btf_array *a = btf_array(t);
 367                         bool flex;
 368
 369                         t = skip_mods_and_typedefs(btf, a->type, &id);
 370                         if (!t)
 371                                 return -EINVAL;
 372
 373                         flex = is_flex_arr(btf, acc - 1, a);
 374                         if (!flex && access_idx >= a->nelems)
 375                                 return -EINVAL;
 376
 377                         spec->spec[spec->len].type_id = id;
 378                         spec->spec[spec->len].idx = access_idx;
 379                         spec->len++;
 380
 381                         sz = btf__resolve_size(btf, id);
 382                         if (sz < 0)
 383                                 return sz;
 384                         spec->bit_offset += access_idx * sz * 8;
 385                 } else {
 386                         pr_warn("prog '%s': relo for [%u] %s (at idx %d) captures type [%d] of unexpected kind %s\n",
 387                                 prog_name, relo->type_id, spec_str, i, id, btf_kind_str(t));
 388                         return -EINVAL;
 389                 }
 390         }
 391
 392         return 0;
 393 }
 394
 395 /* Check two types for compatibility for the purpose of field access
 396  * relocation. const/volatile/restrict and typedefs are skipped to ensure we
 397  * are relocating semantically compatible entities:
 398  *   - any two STRUCTs/UNIONs are compatible and can be mixed;
 399  *   - any two FWDs are compatible, if their names match (modulo flavor suffix);
 400  *   - any two PTRs are always compatible;
 401  *   - for ENUMs, names should be the same (ignoring flavor suffix) or at
 402  *     least one of enums should be anonymous;
 403  *   - for ENUMs, check sizes, names are ignored;
 404  *   - for INT, size and signedness are ignored;
 405  *   - any two FLOATs are always compatible;
 406  *   - for ARRAY, dimensionality is ignored, element types are checked for
 407  *     compatibility recursively;
 408  *   - everything else shouldn't be ever a target of relocation.
 409  * These rules are not set in stone and probably will be adjusted as we get
 410  * more experience with using BPF CO-RE relocations.
 411  */
 412 static int bpf_core_fields_are_compat(const struct btf *local_btf,
 413                                       __u32 local_id,
 414                                       const struct btf *targ_btf,
 415                                       __u32 targ_id)
 416 {
 417         const struct btf_type *local_type, *targ_type;
 418
 419 recur:
 420         local_type = skip_mods_and_typedefs(local_btf, local_id, &local_id);
 421         targ_type = skip_mods_and_typedefs(targ_btf, targ_id, &targ_id);
 422         if (!local_type || !targ_type)
 423                 return -EINVAL;
 424
 425         if (btf_is_composite(local_type) && btf_is_composite(targ_type))
 426                 return 1;
 427         if (!btf_kind_core_compat(local_type, targ_type))
 428                 return 0;
 429
 430         switch (btf_kind(local_type)) {
 431         case BTF_KIND_PTR:
 432         case BTF_KIND_FLOAT:
 433                 return 1;
 434         case BTF_KIND_FWD:
 435         case BTF_KIND_ENUM64:
 436         case BTF_KIND_ENUM: {
 437                 const char *local_name, *targ_name;
 438                 size_t local_len, targ_len;
 439
 440                 local_name = btf__name_by_offset(local_btf,
 441                                                  local_type->name_off);
 442                 targ_name = btf__name_by_offset(targ_btf, targ_type->name_off);
 443                 local_len = bpf_core_essential_name_len(local_name);
 444                 targ_len = bpf_core_essential_name_len(targ_name);
 445                 /* one of them is anonymous or both w/ same flavor-less names */
 446                 return local_len == 0 || targ_len == 0 ||
 447                        (local_len == targ_len &&
 448                         strncmp(local_name, targ_name, local_len) == 0);
 449         }
 450         case BTF_KIND_INT:
 451                 /* just reject deprecated bitfield-like integers; all other
 452                  * integers are by default compatible between each other
 453                  */
 454                 return btf_int_offset(local_type) == 0 &&
 455                        btf_int_offset(targ_type) == 0;
 456         case BTF_KIND_ARRAY:
 457                 local_id = btf_array(local_type)->type;
 458                 targ_id = btf_array(targ_type)->type;
 459                 goto recur;
 460         default:
 461                 return 0;
 462         }
 463 }
 464
 465 /*
 466  * Given single high-level named field accessor in local type, find
 467  * corresponding high-level accessor for a target type. Along the way,
 468  * maintain low-level spec for target as well. Also keep updating target
 469  * bit offset.
 470  *
 471  * Searching is performed through recursive exhaustive enumeration of all
 472  * fields of a struct/union. If there are any anonymous (embedded)
 473  * structs/unions, they are recursively searched as well. If field with
 474  * desired name is found, check compatibility between local and target types,
 475  * before returning result.
 476  *
 477  * 1 is returned, if field is found.
 478  * 0 is returned if no compatible field is found.
 479  * <0 is returned on error.
 480  */
 481 static int bpf_core_match_member(const struct btf *local_btf,
 482                                  const struct bpf_core_accessor *local_acc,
 483                                  const struct btf *targ_btf,
 484                                  __u32 targ_id,
 485                                  struct bpf_core_spec *spec,
 486                                  __u32 *next_targ_id)
 487 {
 488         const struct btf_type *local_type, *targ_type;
 489         const struct btf_member *local_member, *m;
 490         const char *local_name, *targ_name;
 491         __u32 local_id;
 492         int i, n, found;
 493
 494         targ_type = skip_mods_and_typedefs(targ_btf, targ_id, &targ_id);
 495         if (!targ_type)
 496                 return -EINVAL;
 497         if (!btf_is_composite(targ_type))
 498                 return 0;
 499
 500         local_id = local_acc->type_id;
 501         local_type = btf_type_by_id(local_btf, local_id);
 502         local_member = btf_members(local_type) + local_acc->idx;
 503         local_name = btf__name_by_offset(local_btf, local_member->name_off);
 504
 505         n = btf_vlen(targ_type);
 506         m = btf_members(targ_type);
 507         for (i = 0; i < n; i++, m++) {
 508                 __u32 bit_offset;
 509
 510                 bit_offset = btf_member_bit_offset(targ_type, i);
 511
 512                 /* too deep struct/union/array nesting */
 513                 if (spec->raw_len == BPF_CORE_SPEC_MAX_LEN)
 514                         return -E2BIG;
 515
 516                 /* speculate this member will be the good one */
 517                 spec->bit_offset += bit_offset;
 518                 spec->raw_spec[spec->raw_len++] = i;
 519
 520                 targ_name = btf__name_by_offset(targ_btf, m->name_off);
 521                 if (str_is_empty(targ_name)) {
 522                         /* embedded struct/union, we need to go deeper */
 523                         found = bpf_core_match_member(local_btf, local_acc,
 524                                                       targ_btf, m->type,
 525                                                       spec, next_targ_id);
 526                         if (found) /* either found or error */
 527                                 return found;
 528                 } else if (strcmp(local_name, targ_name) == 0) {
 529                         /* matching named field */
 530                         struct bpf_core_accessor *targ_acc;
 531
 532                         targ_acc = &spec->spec[spec->len++];
 533                         targ_acc->type_id = targ_id;
 534                         targ_acc->idx = i;
 535                         targ_acc->name = targ_name;
 536
 537                         *next_targ_id = m->type;
 538                         found = bpf_core_fields_are_compat(local_btf,
 539                                                            local_member->type,
 540                                                            targ_btf, m->type);
 541                         if (!found)
 542                                 spec->len--; /* pop accessor */
 543                         return found;
 544                 }
 545                 /* member turned out not to be what we looked for */
 546                 spec->bit_offset -= bit_offset;
 547                 spec->raw_len--;
 548         }
 549
 550         return 0;
 551 }
 552
 553 /*
 554  * Try to match local spec to a target type and, if successful, produce full
 555  * target spec (high-level, low-level + bit offset).
 556  */
 557 static int bpf_core_spec_match(struct bpf_core_spec *local_spec,
 558                                const struct btf *targ_btf, __u32 targ_id,
 559                                struct bpf_core_spec *targ_spec)
 560 {
 561         const struct btf_type *targ_type;
 562         const struct bpf_core_accessor *local_acc;
 563         struct bpf_core_accessor *targ_acc;
 564         int i, sz, matched;
 565         __u32 name_off;
 566
 567         memset(targ_spec, 0, sizeof(*targ_spec));
 568         targ_spec->btf = targ_btf;
 569         targ_spec->root_type_id = targ_id;
 570         targ_spec->relo_kind = local_spec->relo_kind;
 571
 572         if (core_relo_is_type_based(local_spec->relo_kind)) {
 573                 if (local_spec->relo_kind == BPF_CORE_TYPE_MATCHES)
 574                         return bpf_core_types_match(local_spec->btf,
 575                                                     local_spec->root_type_id,
 576                                                     targ_btf, targ_id);
 577                 else
 578                         return bpf_core_types_are_compat(local_spec->btf,
 579                                                          local_spec->root_type_id,
 580                                                          targ_btf, targ_id);
 581         }
 582
 583         local_acc = &local_spec->spec[0];
 584         targ_acc = &targ_spec->spec[0];
 585
 586         if (core_relo_is_enumval_based(local_spec->relo_kind)) {
 587                 size_t local_essent_len, targ_essent_len;
 588                 const char *targ_name;
 589
 590                 /* has to resolve to an enum */
 591                 targ_type = skip_mods_and_typedefs(targ_spec->btf, targ_id, &targ_id);
 592                 if (!btf_is_any_enum(targ_type))
 593                         return 0;
 594
 595                 local_essent_len = bpf_core_essential_name_len(local_acc->name);
 596
 597                 for (i = 0; i < btf_vlen(targ_type); i++) {
 598                         if (btf_is_enum(targ_type))
 599                                 name_off = btf_enum(targ_type)[i].name_off;
 600                         else
 601                                 name_off = btf_enum64(targ_type)[i].name_off;
 602
 603                         targ_name = btf__name_by_offset(targ_spec->btf, name_off);
 604                         targ_essent_len = bpf_core_essential_name_len(targ_name);
 605                         if (targ_essent_len != local_essent_len)
 606                                 continue;
 607                         if (strncmp(local_acc->name, targ_name, local_essent_len) == 0) {
 608                                 targ_acc->type_id = targ_id;
 609                                 targ_acc->idx = i;
 610                                 targ_acc->name = targ_name;
 611                                 targ_spec->len++;
 612                                 targ_spec->raw_spec[targ_spec->raw_len] = targ_acc->idx;
 613                                 targ_spec->raw_len++;
 614                                 return 1;
 615                         }
 616                 }
 617                 return 0;
 618         }
 619
 620         if (!core_relo_is_field_based(local_spec->relo_kind))
 621                 return -EINVAL;
 622
 623         for (i = 0; i < local_spec->len; i++, local_acc++, targ_acc++) {
 624                 targ_type = skip_mods_and_typedefs(targ_spec->btf, targ_id,
 625                                                    &targ_id);
 626                 if (!targ_type)
 627                         return -EINVAL;
 628
 629                 if (local_acc->name) {
 630                         matched = bpf_core_match_member(local_spec->btf,
 631                                                         local_acc,
 632                                                         targ_btf, targ_id,
 633                                                         targ_spec, &targ_id);
 634                         if (matched <= 0)
 635                                 return matched;
 636                 } else {
 637                         /* for i=0, targ_id is already treated as array element
 638                          * type (because it's the original struct), for others
 639                          * we should find array element type first
 640                          */
 641                         if (i > 0) {
 642                                 const struct btf_array *a;
 643                                 bool flex;
 644
 645                                 if (!btf_is_array(targ_type))
 646                                         return 0;
 647
 648                                 a = btf_array(targ_type);
 649                                 flex = is_flex_arr(targ_btf, targ_acc - 1, a);
 650                                 if (!flex && local_acc->idx >= a->nelems)
 651                                         return 0;
 652                                 if (!skip_mods_and_typedefs(targ_btf, a->type,
 653                                                             &targ_id))
 654                                         return -EINVAL;
 655                         }
 656
 657                         /* too deep struct/union/array nesting */
 658                         if (targ_spec->raw_len == BPF_CORE_SPEC_MAX_LEN)
 659                                 return -E2BIG;
 660
 661                         targ_acc->type_id = targ_id;
 662                         targ_acc->idx = local_acc->idx;
 663                         targ_acc->name = NULL;
 664                         targ_spec->len++;
 665                         targ_spec->raw_spec[targ_spec->raw_len] = targ_acc->idx;
 666                         targ_spec->raw_len++;
 667
 668                         sz = btf__resolve_size(targ_btf, targ_id);
 669                         if (sz < 0)
 670                                 return sz;
 671                         targ_spec->bit_offset += local_acc->idx * sz * 8;
 672                 }
 673         }
 674
 675         return 1;
 676 }
 677
 678 static int bpf_core_calc_field_relo(const char *prog_name,
 679                                     const struct bpf_core_relo *relo,
 680                                     const struct bpf_core_spec *spec,
 681                                     __u64 *val, __u32 *field_sz, __u32 *type_id,
 682                                     bool *validate)
 683 {
 684         const struct bpf_core_accessor *acc;
 685         const struct btf_type *t;
 686         __u32 byte_off, byte_sz, bit_off, bit_sz, field_type_id;
 687         const struct btf_member *m;
 688         const struct btf_type *mt;
 689         bool bitfield;
 690         __s64 sz;
 691
 692         *field_sz = 0;
 693
 694         if (relo->kind == BPF_CORE_FIELD_EXISTS) {
 695                 *val = spec ? 1 : 0;
 696                 return 0;
 697         }
 698
 699         if (!spec)
 700                 return -EUCLEAN; /* request instruction poisoning */
 701
 702         acc = &spec->spec[spec->len - 1];
 703         t = btf_type_by_id(spec->btf, acc->type_id);
 704
 705         /* a[n] accessor needs special handling */
 706         if (!acc->name) {
 707                 if (relo->kind == BPF_CORE_FIELD_BYTE_OFFSET) {
 708                         *val = spec->bit_offset / 8;
 709                         /* remember field size for load/store mem size */
 710                         sz = btf__resolve_size(spec->btf, acc->type_id);
 711                         if (sz < 0)
 712                                 return -EINVAL;
 713                         *field_sz = sz;
 714                         *type_id = acc->type_id;
 715                 } else if (relo->kind == BPF_CORE_FIELD_BYTE_SIZE) {
 716                         sz = btf__resolve_size(spec->btf, acc->type_id);
 717                         if (sz < 0)
 718                                 return -EINVAL;
 719                         *val = sz;
 720                 } else {
 721                         pr_warn("prog '%s': relo %d at insn #%d can't be applied to array access\n",
 722                                 prog_name, relo->kind, relo->insn_off / 8);
 723                         return -EINVAL;
 724                 }
 725                 if (validate)
 726                         *validate = true;
 727                 return 0;
 728         }
 729
 730         m = btf_members(t) + acc->idx;
 731         mt = skip_mods_and_typedefs(spec->btf, m->type, &field_type_id);
 732         bit_off = spec->bit_offset;
 733         bit_sz = btf_member_bitfield_size(t, acc->idx);
 734
 735         bitfield = bit_sz > 0;
 736         if (bitfield) {
 737                 byte_sz = mt->size;
 738                 byte_off = bit_off / 8 / byte_sz * byte_sz;
 739                 /* figure out smallest int size necessary for bitfield load */
 740                 while (bit_off + bit_sz - byte_off * 8 > byte_sz * 8) {
 741                         if (byte_sz >= 8) {
 742                                 /* bitfield can't be read with 64-bit read */
 743                                 pr_warn("prog '%s': relo %d at insn #%d can't be satisfied for bitfield\n",
 744                                         prog_name, relo->kind, relo->insn_off / 8);
 745                                 return -E2BIG;
 746                         }
 747                         byte_sz *= 2;
 748                         byte_off = bit_off / 8 / byte_sz * byte_sz;
 749                 }
 750         } else {
 751                 sz = btf__resolve_size(spec->btf, field_type_id);
 752                 if (sz < 0)
 753                         return -EINVAL;
 754                 byte_sz = sz;
 755                 byte_off = spec->bit_offset / 8;
 756                 bit_sz = byte_sz * 8;
 757         }
 758
 759         /* for bitfields, all the relocatable aspects are ambiguous and we
 760          * might disagree with compiler, so turn off validation of expected
 761          * value, except for signedness
 762          */
 763         if (validate)
 764                 *validate = !bitfield;
 765
 766         switch (relo->kind) {
 767         case BPF_CORE_FIELD_BYTE_OFFSET:
 768                 *val = byte_off;
 769                 if (!bitfield) {
 770                         *field_sz = byte_sz;
 771                         *type_id = field_type_id;
 772                 }
 773                 break;
 774         case BPF_CORE_FIELD_BYTE_SIZE:
 775                 *val = byte_sz;
 776                 break;
 777         case BPF_CORE_FIELD_SIGNED:
 778                 *val = (btf_is_any_enum(mt) && BTF_INFO_KFLAG(mt->info)) ||
 779                        (btf_int_encoding(mt) & BTF_INT_SIGNED);
 780                 if (validate)
 781                         *validate = true; /* signedness is never ambiguous */
 782                 break;
 783         case BPF_CORE_FIELD_LSHIFT_U64:
 784 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
 785                 *val = 64 - (bit_off + bit_sz - byte_off  * 8);
 786 #else
 787                 *val = (8 - byte_sz) * 8 + (bit_off - byte_off * 8);
 788 #endif
 789                 break;
 790         case BPF_CORE_FIELD_RSHIFT_U64:
 791                 *val = 64 - bit_sz;
 792                 if (validate)
 793                         *validate = true; /* right shift is never ambiguous */
 794                 break;
 795         case BPF_CORE_FIELD_EXISTS:
 796         default:
 797                 return -EOPNOTSUPP;
 798         }
 799
 800         return 0;
 801 }
 802
 803 static int bpf_core_calc_type_relo(const struct bpf_core_relo *relo,
 804                                    const struct bpf_core_spec *spec,
 805                                    __u64 *val, bool *validate)
 806 {
 807         __s64 sz;
 808
 809         /* by default, always check expected value in bpf_insn */
 810         if (validate)
 811                 *validate = true;
 812
 813         /* type-based relos return zero when target type is not found */
 814         if (!spec) {
 815                 *val = 0;
 816                 return 0;
 817         }
 818
 819         switch (relo->kind) {
 820         case BPF_CORE_TYPE_ID_TARGET:
 821                 *val = spec->root_type_id;
 822                 /* type ID, embedded in bpf_insn, might change during linking,
 823                  * so enforcing it is pointless
 824                  */
 825                 if (validate)
 826                         *validate = false;
 827                 break;
 828         case BPF_CORE_TYPE_EXISTS:
 829         case BPF_CORE_TYPE_MATCHES:
 830                 *val = 1;
 831                 break;
 832         case BPF_CORE_TYPE_SIZE:
 833                 sz = btf__resolve_size(spec->btf, spec->root_type_id);
 834                 if (sz < 0)
 835                         return -EINVAL;
 836                 *val = sz;
 837                 break;
 838         case BPF_CORE_TYPE_ID_LOCAL:
 839         /* BPF_CORE_TYPE_ID_LOCAL is handled specially and shouldn't get here */
 840         default:
 841                 return -EOPNOTSUPP;
 842         }
 843
 844         return 0;
 845 }
 846
 847 static int bpf_core_calc_enumval_relo(const struct bpf_core_relo *relo,
 848                                       const struct bpf_core_spec *spec,
 849                                       __u64 *val)
 850 {
 851         const struct btf_type *t;
 852
 853         switch (relo->kind) {
 854         case BPF_CORE_ENUMVAL_EXISTS:
 855                 *val = spec ? 1 : 0;
 856                 break;
 857         case BPF_CORE_ENUMVAL_VALUE:
 858                 if (!spec)
 859                         return -EUCLEAN; /* request instruction poisoning */
 860                 t = btf_type_by_id(spec->btf, spec->spec[0].type_id);
 861                 if (btf_is_enum(t))
 862                         *val = btf_enum(t)[spec->spec[0].idx].val;
 863                 else
 864                         *val = btf_enum64_value(btf_enum64(t) + spec->spec[0].idx);
 865                 break;
 866         default:
 867                 return -EOPNOTSUPP;
 868         }
 869
 870         return 0;
 871 }
 872
 873 /* Calculate original and target relocation values, given local and target
 874  * specs and relocation kind. These values are calculated for each candidate.
 875  * If there are multiple candidates, resulting values should all be consistent
 876  * with each other. Otherwise, libbpf will refuse to proceed due to ambiguity.
 877  * If instruction has to be poisoned, *poison will be set to true.
 878  */
 879 static int bpf_core_calc_relo(const char *prog_name,
 880                               const struct bpf_core_relo *relo,
 881                               int relo_idx,
 882                               const struct bpf_core_spec *local_spec,
 883                               const struct bpf_core_spec *targ_spec,
 884                               struct bpf_core_relo_res *res)
 885 {
 886         int err = -EOPNOTSUPP;
 887
 888         res->orig_val = 0;
 889         res->new_val = 0;
 890         res->poison = false;
 891         res->validate = true;
 892         res->fail_memsz_adjust = false;
 893         res->orig_sz = res->new_sz = 0;
 894         res->orig_type_id = res->new_type_id = 0;
 895
 896         if (core_relo_is_field_based(relo->kind)) {
 897                 err = bpf_core_calc_field_relo(prog_name, relo, local_spec,
 898                                                &res->orig_val, &res->orig_sz,
 899                                                &res->orig_type_id, &res->validate);
 900                 err = err ?: bpf_core_calc_field_relo(prog_name, relo, targ_spec,
 901                                                       &res->new_val, &res->new_sz,
 902                                                       &res->new_type_id, NULL);
 903                 if (err)
 904                         goto done;
 905                 /* Validate if it's safe to adjust load/store memory size.
 906                  * Adjustments are performed only if original and new memory
 907                  * sizes differ.
 908                  */
 909                 res->fail_memsz_adjust = false;
 910                 if (res->orig_sz != res->new_sz) {
 911                         const struct btf_type *orig_t, *new_t;
 912
 913                         orig_t = btf_type_by_id(local_spec->btf, res->orig_type_id);
 914                         new_t = btf_type_by_id(targ_spec->btf, res->new_type_id);
 915
 916                         /* There are two use cases in which it's safe to
 917                          * adjust load/store's mem size:
 918                          *   - reading a 32-bit kernel pointer, while on BPF
 919                          *   size pointers are always 64-bit; in this case
 920                          *   it's safe to "downsize" instruction size due to
 921                          *   pointer being treated as unsigned integer with
 922                          *   zero-extended upper 32-bits;
 923                          *   - reading unsigned integers, again due to
 924                          *   zero-extension is preserving the value correctly.
 925                          *
 926                          * In all other cases it's incorrect to attempt to
 927                          * load/store field because read value will be
 928                          * incorrect, so we poison relocated instruction.
 929                          */
 930                         if (btf_is_ptr(orig_t) && btf_is_ptr(new_t))
 931                                 goto done;
 932                         if (btf_is_int(orig_t) && btf_is_int(new_t) &&
 933                             btf_int_encoding(orig_t) != BTF_INT_SIGNED &&
 934                             btf_int_encoding(new_t) != BTF_INT_SIGNED)
 935                                 goto done;
 936
 937                         /* mark as invalid mem size adjustment, but this will
 938                          * only be checked for LDX/STX/ST insns
 939                          */
 940                         res->fail_memsz_adjust = true;
 941                 }
 942         } else if (core_relo_is_type_based(relo->kind)) {
 943                 err = bpf_core_calc_type_relo(relo, local_spec, &res->orig_val, &res->validate);
 944                 err = err ?: bpf_core_calc_type_relo(relo, targ_spec, &res->new_val, NULL);
 945         } else if (core_relo_is_enumval_based(relo->kind)) {
 946                 err = bpf_core_calc_enumval_relo(relo, local_spec, &res->orig_val);
 947                 err = err ?: bpf_core_calc_enumval_relo(relo, targ_spec, &res->new_val);
 948         }
 949
 950 done:
 951         if (err == -EUCLEAN) {
 952                 /* EUCLEAN is used to signal instruction poisoning request */
 953                 res->poison = true;
 954                 err = 0;
 955         } else if (err == -EOPNOTSUPP) {
 956                 /* EOPNOTSUPP means unknown/unsupported relocation */
 957                 pr_warn("prog '%s': relo #%d: unrecognized CO-RE relocation %s (%d) at insn #%d\n",
 958                         prog_name, relo_idx, core_relo_kind_str(relo->kind),
 959                         relo->kind, relo->insn_off / 8);
 960         }
 961
 962         return err;
 963 }
 964
 965 /*
 966  * Turn instruction for which CO_RE relocation failed into invalid one with
 967  * distinct signature.
 968  */
 969 static void bpf_core_poison_insn(const char *prog_name, int relo_idx,
 970                                  int insn_idx, struct bpf_insn *insn)
 971 {
 972         pr_debug("prog '%s': relo #%d: substituting insn #%d w/ invalid insn\n",
 973                  prog_name, relo_idx, insn_idx);
 974         insn->code = BPF_JMP | BPF_CALL;
 975         insn->dst_reg = 0;
 976         insn->src_reg = 0;
 977         insn->off = 0;
 978         /* if this instruction is reachable (not a dead code),
 979          * verifier will complain with the following message:
 980          * invalid func unknown#195896080
 981          */
 982         insn->imm = 195896080; /* => 0xbad2310 => "bad relo" */
 983 }
 984
 985 static int insn_bpf_size_to_bytes(struct bpf_insn *insn)
 986 {
 987         switch (BPF_SIZE(insn->code)) {
 988         case BPF_DW: return 8;
 989         case BPF_W: return 4;
 990         case BPF_H: return 2;
 991         case BPF_B: return 1;
 992         default: return -1;
 993         }
 994 }
 995
 996 static int insn_bytes_to_bpf_size(__u32 sz)
 997 {
 998         switch (sz) {
 999         case 8: return BPF_DW;
1000         case 4: return BPF_W;
1001         case 2: return BPF_H;
1002         case 1: return BPF_B;
1003         default: return -1;
1004         }
1005 }
1006
1007 /*
1008  * Patch relocatable BPF instruction.
1009  *
1010  * Patched value is determined by relocation kind and target specification.
1011  * For existence relocations target spec will be NULL if field/type is not found.
1012  * Expected insn->imm value is determined using relocation kind and local
1013  * spec, and is checked before patching instruction. If actual insn->imm value
1014  * is wrong, bail out with error.
1015  *
1016  * Currently supported classes of BPF instruction are:
1017  * 1. rX = <imm> (assignment with immediate operand);
1018  * 2. rX += <imm> (arithmetic operations with immediate operand);
1019  * 3. rX = <imm64> (load with 64-bit immediate value);
1020  * 4. rX = *(T *)(rY + <off>), where T is one of {u8, u16, u32, u64};
1021  * 5. *(T *)(rX + <off>) = rY, where T is one of {u8, u16, u32, u64};
1022  * 6. *(T *)(rX + <off>) = <imm>, where T is one of {u8, u16, u32, u64}.
1023  */
1024 int bpf_core_patch_insn(const char *prog_name, struct bpf_insn *insn,
1025                         int insn_idx, const struct bpf_core_relo *relo,
1026                         int relo_idx, const struct bpf_core_relo_res *res)
1027 {
1028         __u64 orig_val, new_val;
1029         __u8 class;
1030
1031         class = BPF_CLASS(insn->code);
1032
1033         if (res->poison) {
1034 poison:
1035                 /* poison second part of ldimm64 to avoid confusing error from
1036                  * verifier about "unknown opcode 00"
1037                  */
1038                 if (is_ldimm64_insn(insn))
1039                         bpf_core_poison_insn(prog_name, relo_idx, insn_idx + 1, insn + 1);
1040                 bpf_core_poison_insn(prog_name, relo_idx, insn_idx, insn);
1041                 return 0;
1042         }
1043
1044         orig_val = res->orig_val;
1045         new_val = res->new_val;
1046
1047         switch (class) {
1048         case BPF_ALU:
1049         case BPF_ALU64:
1050                 if (BPF_SRC(insn->code) != BPF_K)
1051                         return -EINVAL;
1052                 if (res->validate && insn->imm != orig_val) {
1053                         pr_warn("prog '%s': relo #%d: unexpected insn #%d (ALU/ALU64) value: got %u, exp %llu -> %llu\n",
1054                                 prog_name, relo_idx,
1055                                 insn_idx, insn->imm, (unsigned long long)orig_val,
1056                                 (unsigned long long)new_val);
1057                         return -EINVAL;
1058                 }
1059                 orig_val = insn->imm;
1060                 insn->imm = new_val;
1061                 pr_debug("prog '%s': relo #%d: patched insn #%d (ALU/ALU64) imm %llu -> %llu\n",
1062                          prog_name, relo_idx, insn_idx,
1063                          (unsigned long long)orig_val, (unsigned long long)new_val);
1064                 break;
1065         case BPF_LDX:
1066         case BPF_ST:
1067         case BPF_STX:
1068                 if (res->validate && insn->off != orig_val) {
1069                         pr_warn("prog '%s': relo #%d: unexpected insn #%d (LDX/ST/STX) value: got %u, exp %llu -> %llu\n",
1070                                 prog_name, relo_idx, insn_idx, insn->off, (unsigned long long)orig_val,
1071                                 (unsigned long long)new_val);
1072                         return -EINVAL;
1073                 }
1074                 if (new_val > SHRT_MAX) {
1075                         pr_warn("prog '%s': relo #%d: insn #%d (LDX/ST/STX) value too big: %llu\n",
1076                                 prog_name, relo_idx, insn_idx, (unsigned long long)new_val);
1077                         return -ERANGE;
1078                 }
1079                 if (res->fail_memsz_adjust) {
1080                         pr_warn("prog '%s': relo #%d: insn #%d (LDX/ST/STX) accesses field incorrectly. "
1081                                 "Make sure you are accessing pointers, unsigned integers, or fields of matching type and size.\n",
1082                                 prog_name, relo_idx, insn_idx);
1083                         goto poison;
1084                 }
1085
1086                 orig_val = insn->off;
1087                 insn->off = new_val;
1088                 pr_debug("prog '%s': relo #%d: patched insn #%d (LDX/ST/STX) off %llu -> %llu\n",
1089                          prog_name, relo_idx, insn_idx, (unsigned long long)orig_val,
1090                          (unsigned long long)new_val);
1091
1092                 if (res->new_sz != res->orig_sz) {
1093                         int insn_bytes_sz, insn_bpf_sz;
1094
1095                         insn_bytes_sz = insn_bpf_size_to_bytes(insn);
1096                         if (insn_bytes_sz != res->orig_sz) {
1097                                 pr_warn("prog '%s': relo #%d: insn #%d (LDX/ST/STX) unexpected mem size: got %d, exp %u\n",
1098                                         prog_name, relo_idx, insn_idx, insn_bytes_sz, res->orig_sz);
1099                                 return -EINVAL;
1100                         }
1101
1102                         insn_bpf_sz = insn_bytes_to_bpf_size(res->new_sz);
1103                         if (insn_bpf_sz < 0) {
1104                                 pr_warn("prog '%s': relo #%d: insn #%d (LDX/ST/STX) invalid new mem size: %u\n",
1105                                         prog_name, relo_idx, insn_idx, res->new_sz);
1106                                 return -EINVAL;
1107                         }
1108
1109                         insn->code = BPF_MODE(insn->code) | insn_bpf_sz | BPF_CLASS(insn->code);
1110                         pr_debug("prog '%s': relo #%d: patched insn #%d (LDX/ST/STX) mem_sz %u -> %u\n",
1111                                  prog_name, relo_idx, insn_idx, res->orig_sz, res->new_sz);
1112                 }
1113                 break;
1114         case BPF_LD: {
1115                 __u64 imm;
1116
1117                 if (!is_ldimm64_insn(insn) ||
1118                     insn[0].src_reg != 0 || insn[0].off != 0 ||
1119                     insn[1].code != 0 || insn[1].dst_reg != 0 ||
1120                     insn[1].src_reg != 0 || insn[1].off != 0) {
1121                         pr_warn("prog '%s': relo #%d: insn #%d (LDIMM64) has unexpected form\n",
1122                                 prog_name, relo_idx, insn_idx);
1123                         return -EINVAL;
1124                 }
1125
1126                 imm = (__u32)insn[0].imm | ((__u64)insn[1].imm << 32);
1127                 if (res->validate && imm != orig_val) {
1128                         pr_warn("prog '%s': relo #%d: unexpected insn #%d (LDIMM64) value: got %llu, exp %llu -> %llu\n",
1129                                 prog_name, relo_idx,
1130                                 insn_idx, (unsigned long long)imm,
1131                                 (unsigned long long)orig_val, (unsigned long long)new_val);
1132                         return -EINVAL;
1133                 }
1134
1135                 insn[0].imm = new_val;
1136                 insn[1].imm = new_val >> 32;
1137                 pr_debug("prog '%s': relo #%d: patched insn #%d (LDIMM64) imm64 %llu -> %llu\n",
1138                          prog_name, relo_idx, insn_idx,
1139                          (unsigned long long)imm, (unsigned long long)new_val);
1140                 break;
1141         }
1142         default:
1143                 pr_warn("prog '%s': relo #%d: trying to relocate unrecognized insn #%d, code:0x%x, src:0x%x, dst:0x%x, off:0x%x, imm:0x%x\n",
1144                         prog_name, relo_idx, insn_idx, insn->code,
1145                         insn->src_reg, insn->dst_reg, insn->off, insn->imm);
1146                 return -EINVAL;
1147         }
1148
1149         return 0;
1150 }
1151
1152 /* Output spec definition in the format:
1153  * [<type-id>] (<type-name>) + <raw-spec> => <offset>@<spec>,
1154  * where <spec> is a C-syntax view of recorded field access, e.g.: x.a[3].b
1155  */
1156 int bpf_core_format_spec(char *buf, size_t buf_sz, const struct bpf_core_spec *spec)
1157 {
1158         const struct btf_type *t;
1159         const char *s;
1160         __u32 type_id;
1161         int i, len = 0;
1162
1163 #define append_buf(fmt, args...)                                \
1164         ({                                                      \
1165                 int r;                                          \
1166                 r = snprintf(buf, buf_sz, fmt, ##args);         \
1167                 len += r;                                       \
1168                 if (r >= buf_sz)                                \
1169                         r = buf_sz;                             \
1170                 buf += r;                                       \
1171                 buf_sz -= r;                                    \
1172         })
1173
1174         type_id = spec->root_type_id;
1175         t = btf_type_by_id(spec->btf, type_id);
1176         s = btf__name_by_offset(spec->btf, t->name_off);
1177
1178         append_buf("<%s> [%u] %s %s",
1179                    core_relo_kind_str(spec->relo_kind),
1180                    type_id, btf_kind_str(t), str_is_empty(s) ? "<anon>" : s);
1181
1182         if (core_relo_is_type_based(spec->relo_kind))
1183                 return len;
1184
1185         if (core_relo_is_enumval_based(spec->relo_kind)) {
1186                 t = skip_mods_and_typedefs(spec->btf, type_id, NULL);
1187                 if (btf_is_enum(t)) {
1188                         const struct btf_enum *e;
1189                         const char *fmt_str;
1190
1191                         e = btf_enum(t) + spec->raw_spec[0];
1192                         s = btf__name_by_offset(spec->btf, e->name_off);
1193                         fmt_str = BTF_INFO_KFLAG(t->info) ? "::%s = %d" : "::%s = %u";
1194                         append_buf(fmt_str, s, e->val);
1195                 } else {
1196                         const struct btf_enum64 *e;
1197                         const char *fmt_str;
1198
1199                         e = btf_enum64(t) + spec->raw_spec[0];
1200                         s = btf__name_by_offset(spec->btf, e->name_off);
1201                         fmt_str = BTF_INFO_KFLAG(t->info) ? "::%s = %lld" : "::%s = %llu";
1202                         append_buf(fmt_str, s, (unsigned long long)btf_enum64_value(e));
1203                 }
1204                 return len;
1205         }
1206
1207         if (core_relo_is_field_based(spec->relo_kind)) {
1208                 for (i = 0; i < spec->len; i++) {
1209                         if (spec->spec[i].name)
1210                                 append_buf(".%s", spec->spec[i].name);
1211                         else if (i > 0 || spec->spec[i].idx > 0)
1212                                 append_buf("[%u]", spec->spec[i].idx);
1213                 }
1214
1215                 append_buf(" (");
1216                 for (i = 0; i < spec->raw_len; i++)
1217                         append_buf("%s%d", i == 0 ? "" : ":", spec->raw_spec[i]);
1218
1219                 if (spec->bit_offset % 8)
1220                         append_buf(" @ offset %u.%u)", spec->bit_offset / 8, spec->bit_offset % 8);
1221                 else
1222                         append_buf(" @ offset %u)", spec->bit_offset / 8);
1223                 return len;
1224         }
1225
1226         return len;
1227 #undef append_buf
1228 }
1229
1230 /*
1231  * Calculate CO-RE relocation target result.
1232  *
1233  * The outline and important points of the algorithm:
1234  * 1. For given local type, find corresponding candidate target types.
1235  *    Candidate type is a type with the same "essential" name, ignoring
1236  *    everything after last triple underscore (___). E.g., `sample`,
1237  *    `sample___flavor_one`, `sample___flavor_another_one`, are all candidates
1238  *    for each other. Names with triple underscore are referred to as
1239  *    "flavors" and are useful, among other things, to allow to
1240  *    specify/support incompatible variations of the same kernel struct, which
1241  *    might differ between different kernel versions and/or build
1242  *    configurations.
1243  *
1244  *    N.B. Struct "flavors" could be generated by bpftool's BTF-to-C
1245  *    converter, when deduplicated BTF of a kernel still contains more than
1246  *    one different types with the same name. In that case, ___2, ___3, etc
1247  *    are appended starting from second name conflict. But start flavors are
1248  *    also useful to be defined "locally", in BPF program, to extract same
1249  *    data from incompatible changes between different kernel
1250  *    versions/configurations. For instance, to handle field renames between
1251  *    kernel versions, one can use two flavors of the struct name with the
1252  *    same common name and use conditional relocations to extract that field,
1253  *    depending on target kernel version.
1254  * 2. For each candidate type, try to match local specification to this
1255  *    candidate target type. Matching involves finding corresponding
1256  *    high-level spec accessors, meaning that all named fields should match,
1257  *    as well as all array accesses should be within the actual bounds. Also,
1258  *    types should be compatible (see bpf_core_fields_are_compat for details).
1259  * 3. It is supported and expected that there might be multiple flavors
1260  *    matching the spec. As long as all the specs resolve to the same set of
1261  *    offsets across all candidates, there is no error. If there is any
1262  *    ambiguity, CO-RE relocation will fail. This is necessary to accommodate
1263  *    imperfection of BTF deduplication, which can cause slight duplication of
1264  *    the same BTF type, if some directly or indirectly referenced (by
1265  *    pointer) type gets resolved to different actual types in different
1266  *    object files. If such a situation occurs, deduplicated BTF will end up
1267  *    with two (or more) structurally identical types, which differ only in
1268  *    types they refer to through pointer. This should be OK in most cases and
1269  *    is not an error.
1270  * 4. Candidate types search is performed by linearly scanning through all
1271  *    types in target BTF. It is anticipated that this is overall more
1272  *    efficient memory-wise and not significantly worse (if not better)
1273  *    CPU-wise compared to prebuilding a map from all local type names to
1274  *    a list of candidate type names. It's also sped up by caching resolved
1275  *    list of matching candidates per each local "root" type ID, that has at
1276  *    least one bpf_core_relo associated with it. This list is shared
1277  *    between multiple relocations for the same type ID and is updated as some
1278  *    of the candidates are pruned due to structural incompatibility.
1279  */
1280 int bpf_core_calc_relo_insn(const char *prog_name,
1281                             const struct bpf_core_relo *relo,
1282                             int relo_idx,
1283                             const struct btf *local_btf,
1284                             struct bpf_core_cand_list *cands,
1285                             struct bpf_core_spec *specs_scratch,
1286                             struct bpf_core_relo_res *targ_res)
1287 {
1288         struct bpf_core_spec *local_spec = &specs_scratch[0];
1289         struct bpf_core_spec *cand_spec = &specs_scratch[1];
1290         struct bpf_core_spec *targ_spec = &specs_scratch[2];
1291         struct bpf_core_relo_res cand_res;
1292         const struct btf_type *local_type;
1293         const char *local_name;
1294         __u32 local_id;
1295         char spec_buf[256];
1296         int i, j, err;
1297
1298         local_id = relo->type_id;
1299         local_type = btf_type_by_id(local_btf, local_id);
1300         local_name = btf__name_by_offset(local_btf, local_type->name_off);
1301         if (!local_name)
1302                 return -EINVAL;
1303
1304         err = bpf_core_parse_spec(prog_name, local_btf, relo, local_spec);
1305         if (err) {
1306                 const char *spec_str;
1307
1308                 spec_str = btf__name_by_offset(local_btf, relo->access_str_off);
1309                 pr_warn("prog '%s': relo #%d: parsing [%d] %s %s + %s failed: %d\n",
1310                         prog_name, relo_idx, local_id, btf_kind_str(local_type),
1311                         str_is_empty(local_name) ? "<anon>" : local_name,
1312                         spec_str ?: "<?>", err);
1313                 return -EINVAL;
1314         }
1315
1316         bpf_core_format_spec(spec_buf, sizeof(spec_buf), local_spec);
1317         pr_debug("prog '%s': relo #%d: %s\n", prog_name, relo_idx, spec_buf);
1318
1319         /* TYPE_ID_LOCAL relo is special and doesn't need candidate search */
1320         if (relo->kind == BPF_CORE_TYPE_ID_LOCAL) {
1321                 /* bpf_insn's imm value could get out of sync during linking */
1322                 memset(targ_res, 0, sizeof(*targ_res));
1323                 targ_res->validate = false;
1324                 targ_res->poison = false;
1325                 targ_res->orig_val = local_spec->root_type_id;
1326                 targ_res->new_val = local_spec->root_type_id;
1327                 return 0;
1328         }
1329
1330         /* libbpf doesn't support candidate search for anonymous types */
1331         if (str_is_empty(local_name)) {
1332                 pr_warn("prog '%s': relo #%d: <%s> (%d) relocation doesn't support anonymous types\n",
1333                         prog_name, relo_idx, core_relo_kind_str(relo->kind), relo->kind);
1334                 return -EOPNOTSUPP;
1335         }
1336
1337         for (i = 0, j = 0; i < cands->len; i++) {
1338                 err = bpf_core_spec_match(local_spec, cands->cands[i].btf,
1339                                           cands->cands[i].id, cand_spec);
1340                 if (err < 0) {
1341                         bpf_core_format_spec(spec_buf, sizeof(spec_buf), cand_spec);
1342                         pr_warn("prog '%s': relo #%d: error matching candidate #%d %s: %d\n ",
1343                                 prog_name, relo_idx, i, spec_buf, err);
1344                         return err;
1345                 }
1346
1347                 bpf_core_format_spec(spec_buf, sizeof(spec_buf), cand_spec);
1348                 pr_debug("prog '%s': relo #%d: %s candidate #%d %s\n", prog_name,
1349                          relo_idx, err == 0 ? "non-matching" : "matching", i, spec_buf);
1350
1351                 if (err == 0)
1352                         continue;
1353
1354                 err = bpf_core_calc_relo(prog_name, relo, relo_idx, local_spec, cand_spec, &cand_res);
1355                 if (err)
1356                         return err;
1357
1358                 if (j == 0) {
1359                         *targ_res = cand_res;
1360                         *targ_spec = *cand_spec;
1361                 } else if (cand_spec->bit_offset != targ_spec->bit_offset) {
1362                         /* if there are many field relo candidates, they
1363                          * should all resolve to the same bit offset
1364                          */
1365                         pr_warn("prog '%s': relo #%d: field offset ambiguity: %u != %u\n",
1366                                 prog_name, relo_idx, cand_spec->bit_offset,
1367                                 targ_spec->bit_offset);
1368                         return -EINVAL;
1369                 } else if (cand_res.poison != targ_res->poison ||
1370                            cand_res.new_val != targ_res->new_val) {
1371                         /* all candidates should result in the same relocation
1372                          * decision and value, otherwise it's dangerous to
1373                          * proceed due to ambiguity
1374                          */
1375                         pr_warn("prog '%s': relo #%d: relocation decision ambiguity: %s %llu != %s %llu\n",
1376                                 prog_name, relo_idx,
1377                                 cand_res.poison ? "failure" : "success",
1378                                 (unsigned long long)cand_res.new_val,
1379                                 targ_res->poison ? "failure" : "success",
1380                                 (unsigned long long)targ_res->new_val);
1381                         return -EINVAL;
1382                 }
1383
1384                 cands->cands[j++] = cands->cands[i];
1385         }
1386
1387         /*
1388          * For BPF_CORE_FIELD_EXISTS relo or when used BPF program has field
1389          * existence checks or kernel version/config checks, it's expected
1390          * that we might not find any candidates. In this case, if field
1391          * wasn't found in any candidate, the list of candidates shouldn't
1392          * change at all, we'll just handle relocating appropriately,
1393          * depending on relo's kind.
1394          */
1395         if (j > 0)
1396                 cands->len = j;
1397
1398         /*
1399          * If no candidates were found, it might be both a programmer error,
1400          * as well as expected case, depending whether instruction w/
1401          * relocation is guarded in some way that makes it unreachable (dead
1402          * code) if relocation can't be resolved. This is handled in
1403          * bpf_core_patch_insn() uniformly by replacing that instruction with
1404          * BPF helper call insn (using invalid helper ID). If that instruction
1405          * is indeed unreachable, then it will be ignored and eliminated by
1406          * verifier. If it was an error, then verifier will complain and point
1407          * to a specific instruction number in its log.
1408          */
1409         if (j == 0) {
1410                 pr_debug("prog '%s': relo #%d: no matching targets found\n",
1411                          prog_name, relo_idx);
1412
1413                 /* calculate single target relo result explicitly */
1414                 err = bpf_core_calc_relo(prog_name, relo, relo_idx, local_spec, NULL, targ_res);
1415                 if (err)
1416                         return err;
1417         }
1418
1419         return 0;
1420 }
1421
1422 static bool bpf_core_names_match(const struct btf *local_btf, size_t local_name_off,
1423                                  const struct btf *targ_btf, size_t targ_name_off)
1424 {
1425         const char *local_n, *targ_n;
1426         size_t local_len, targ_len;
1427
1428         local_n = btf__name_by_offset(local_btf, local_name_off);
1429         targ_n = btf__name_by_offset(targ_btf, targ_name_off);
1430
1431         if (str_is_empty(targ_n))
1432                 return str_is_empty(local_n);
1433
1434         targ_len = bpf_core_essential_name_len(targ_n);
1435         local_len = bpf_core_essential_name_len(local_n);
1436
1437         return targ_len == local_len && strncmp(local_n, targ_n, local_len) == 0;
1438 }
1439
1440 static int bpf_core_enums_match(const struct btf *local_btf, const struct btf_type *local_t,
1441                                 const struct btf *targ_btf, const struct btf_type *targ_t)
1442 {
1443         __u16 local_vlen = btf_vlen(local_t);
1444         __u16 targ_vlen = btf_vlen(targ_t);
1445         int i, j;
1446
1447         if (local_t->size != targ_t->size)
1448                 return 0;
1449
1450         if (local_vlen > targ_vlen)
1451                 return 0;
1452
1453         /* iterate over the local enum's variants and make sure each has
1454          * a symbolic name correspondent in the target
1455          */
1456         for (i = 0; i < local_vlen; i++) {
1457                 bool matched = false;
1458                 __u32 local_n_off, targ_n_off;
1459
1460                 local_n_off = btf_is_enum(local_t) ? btf_enum(local_t)[i].name_off :
1461                                                      btf_enum64(local_t)[i].name_off;
1462
1463                 for (j = 0; j < targ_vlen; j++) {
1464                         targ_n_off = btf_is_enum(targ_t) ? btf_enum(targ_t)[j].name_off :
1465                                                            btf_enum64(targ_t)[j].name_off;
1466
1467                         if (bpf_core_names_match(local_btf, local_n_off, targ_btf, targ_n_off)) {
1468                                 matched = true;
1469                                 break;
1470                         }
1471                 }
1472
1473                 if (!matched)
1474                         return 0;
1475         }
1476         return 1;
1477 }
1478
1479 static int bpf_core_composites_match(const struct btf *local_btf, const struct btf_type *local_t,
1480                                      const struct btf *targ_btf, const struct btf_type *targ_t,
1481                                      bool behind_ptr, int level)
1482 {
1483         const struct btf_member *local_m = btf_members(local_t);
1484         __u16 local_vlen = btf_vlen(local_t);
1485         __u16 targ_vlen = btf_vlen(targ_t);
1486         int i, j, err;
1487
1488         if (local_vlen > targ_vlen)
1489                 return 0;
1490
1491         /* check that all local members have a match in the target */
1492         for (i = 0; i < local_vlen; i++, local_m++) {
1493                 const struct btf_member *targ_m = btf_members(targ_t);
1494                 bool matched = false;
1495
1496                 for (j = 0; j < targ_vlen; j++, targ_m++) {
1497                         if (!bpf_core_names_match(local_btf, local_m->name_off,
1498                                                   targ_btf, targ_m->name_off))
1499                                 continue;
1500
1501                         err = __bpf_core_types_match(local_btf, local_m->type, targ_btf,
1502                                                      targ_m->type, behind_ptr, level - 1);
1503                         if (err > 0) {
1504                                 matched = true;
1505                                 break;
1506                         }
1507                 }
1508
1509                 if (!matched)
1510                         return 0;
1511         }
1512         return 1;
1513 }
1514
1515 /* Check that two types "match".
1516  *
1517  * The matching relation is defined as follows:
1518  * - modifiers and typedefs are stripped (and, hence, effectively ignored)
1519  * - generally speaking types need to be of same kind (struct vs. struct, union
1520  *   vs. union, etc.)
1521  *   - exceptions are struct/union behind a pointer which could also match a
1522  *     forward declaration of a struct or union, respectively, and enum vs.
1523  *     enum64 (see below)
1524  * Then, depending on type:
1525  * - integers:
1526  *   - match if size and signedness match
1527  * - arrays & pointers:
1528  *   - target types are recursively matched
1529  * - structs & unions:
1530  *   - local members need to exist in target with the same name
1531  *   - for each member we recursively check match unless it is already behind a
1532  *     pointer, in which case we only check matching names and compatible kind
1533  * - enums:
1534  *   - local variants have to have a match in target by symbolic name (but not
1535  *     numeric value)
1536  *   - size has to match (but enum may match enum64 and vice versa)
1537  * - function pointers:
1538  *   - number and position of arguments in local type has to match target
1539  *   - for each argument and the return value we recursively check match
1540  */
1541 int __bpf_core_types_match(const struct btf *local_btf, __u32 local_id, const struct btf *targ_btf,
1542                            __u32 targ_id, bool behind_ptr, int level)
1543 {
1544         const struct btf_type *local_t, *targ_t;
1545         int depth = 32; /* max recursion depth */
1546         __u16 local_k, targ_k;
1547
1548         if (level <= 0)
1549                 return -EINVAL;
1550
1551         local_t = btf_type_by_id(local_btf, local_id);
1552         targ_t = btf_type_by_id(targ_btf, targ_id);
1553
1554 recur:
1555         depth--;
1556         if (depth < 0)
1557                 return -EINVAL;
1558
1559         local_t = skip_mods_and_typedefs(local_btf, local_id, &local_id);
1560         targ_t = skip_mods_and_typedefs(targ_btf, targ_id, &targ_id);
1561         if (!local_t || !targ_t)
1562                 return -EINVAL;
1563
1564         if (!bpf_core_names_match(local_btf, local_t->name_off, targ_btf, targ_t->name_off))
1565                 return 0;
1566
1567         local_k = btf_kind(local_t);
1568         targ_k = btf_kind(targ_t);
1569
1570         switch (local_k) {
1571         case BTF_KIND_UNKN:
1572                 return local_k == targ_k;
1573         case BTF_KIND_FWD: {
1574                 bool local_f = BTF_INFO_KFLAG(local_t->info);
1575
1576                 if (behind_ptr) {
1577                         if (local_k == targ_k)
1578                                 return local_f == BTF_INFO_KFLAG(targ_t->info);
1579
1580                         /* for forward declarations kflag dictates whether the
1581                          * target is a struct (0) or union (1)
1582                          */
1583                         return (targ_k == BTF_KIND_STRUCT && !local_f) ||
1584                                (targ_k == BTF_KIND_UNION && local_f);
1585                 } else {
1586                         if (local_k != targ_k)
1587                                 return 0;
1588
1589                         /* match if the forward declaration is for the same kind */
1590                         return local_f == BTF_INFO_KFLAG(targ_t->info);
1591                 }
1592         }
1593         case BTF_KIND_ENUM:
1594         case BTF_KIND_ENUM64:
1595                 if (!btf_is_any_enum(targ_t))
1596                         return 0;
1597
1598                 return bpf_core_enums_match(local_btf, local_t, targ_btf, targ_t);
1599         case BTF_KIND_STRUCT:
1600         case BTF_KIND_UNION:
1601                 if (behind_ptr) {
1602                         bool targ_f = BTF_INFO_KFLAG(targ_t->info);
1603
1604                         if (local_k == targ_k)
1605                                 return 1;
1606
1607                         if (targ_k != BTF_KIND_FWD)
1608                                 return 0;
1609
1610                         return (local_k == BTF_KIND_UNION) == targ_f;
1611                 } else {
1612                         if (local_k != targ_k)
1613                                 return 0;
1614
1615                         return bpf_core_composites_match(local_btf, local_t, targ_btf, targ_t,
1616                                                          behind_ptr, level);
1617                 }
1618         case BTF_KIND_INT: {
1619                 __u8 local_sgn;
1620                 __u8 targ_sgn;
1621
1622                 if (local_k != targ_k)
1623                         return 0;
1624
1625                 local_sgn = btf_int_encoding(local_t) & BTF_INT_SIGNED;
1626                 targ_sgn = btf_int_encoding(targ_t) & BTF_INT_SIGNED;
1627
1628                 return local_t->size == targ_t->size && local_sgn == targ_sgn;
1629         }
1630         case BTF_KIND_PTR:
1631                 if (local_k != targ_k)
1632                         return 0;
1633
1634                 behind_ptr = true;
1635
1636                 local_id = local_t->type;
1637                 targ_id = targ_t->type;
1638                 goto recur;
1639         case BTF_KIND_ARRAY: {
1640                 const struct btf_array *local_array = btf_array(local_t);
1641                 const struct btf_array *targ_array = btf_array(targ_t);
1642
1643                 if (local_k != targ_k)
1644                         return 0;
1645
1646                 if (local_array->nelems != targ_array->nelems)
1647                         return 0;
1648
1649                 local_id = local_array->type;
1650                 targ_id = targ_array->type;
1651                 goto recur;
1652         }
1653         case BTF_KIND_FUNC_PROTO: {
1654                 struct btf_param *local_p = btf_params(local_t);
1655                 struct btf_param *targ_p = btf_params(targ_t);
1656                 __u16 local_vlen = btf_vlen(local_t);
1657                 __u16 targ_vlen = btf_vlen(targ_t);
1658                 int i, err;
1659
1660                 if (local_k != targ_k)
1661                         return 0;
1662
1663                 if (local_vlen != targ_vlen)
1664                         return 0;
1665
1666                 for (i = 0; i < local_vlen; i++, local_p++, targ_p++) {
1667                         err = __bpf_core_types_match(local_btf, local_p->type, targ_btf,
1668                                                      targ_p->type, behind_ptr, level - 1);
1669                         if (err <= 0)
1670                                 return err;
1671                 }
1672
1673                 /* tail recurse for return type check */
1674                 local_id = local_t->type;
1675                 targ_id = targ_t->type;
1676                 goto recur;
1677         }
1678         default:
1679                 pr_warn("unexpected kind %s relocated, local [%d], target [%d]\n",
1680                         btf_kind_str(local_t), local_id, targ_id);
1681                 return 0;
1682         }
1683 }