1 // SPDX-License-Identifier: (LGPL-2.1 OR BSD-2-Clause)
2 /* Copyright (c) 2018 Facebook */
11 #include <sys/utsname.h>
12 #include <sys/param.h>
14 #include <linux/kernel.h>
15 #include <linux/err.h>
16 #include <linux/btf.h>
21 #include "libbpf_internal.h"
24 /* make sure libbpf doesn't use kernel-only integer typedefs */
25 #pragma GCC poison u8 u16 u32 u64 s8 s16 s32 s64
27 #define BTF_MAX_NR_TYPES 0x7fffffffU
28 #define BTF_MAX_STR_OFFSET 0x7fffffffU
30 static struct btf_type btf_void;
34 struct btf_header *hdr;
37 struct btf_type **types;
46 static inline __u64 ptr_to_u64(const void *ptr)
48 return (__u64) (unsigned long) ptr;
51 static int btf_add_type(struct btf *btf, struct btf_type *t)
53 if (btf->types_size - btf->nr_types < 2) {
54 struct btf_type **new_types;
55 __u32 expand_by, new_size;
57 if (btf->types_size == BTF_MAX_NR_TYPES)
60 expand_by = max(btf->types_size >> 2, 16U);
61 new_size = min(BTF_MAX_NR_TYPES, btf->types_size + expand_by);
63 new_types = realloc(btf->types, sizeof(*new_types) * new_size);
67 if (btf->nr_types == 0)
68 new_types[0] = &btf_void;
70 btf->types = new_types;
71 btf->types_size = new_size;
74 btf->types[++(btf->nr_types)] = t;
79 static int btf_parse_hdr(struct btf *btf)
81 const struct btf_header *hdr = btf->hdr;
84 if (btf->data_size < sizeof(struct btf_header)) {
85 pr_debug("BTF header not found\n");
89 if (hdr->magic != BTF_MAGIC) {
90 pr_debug("Invalid BTF magic:%x\n", hdr->magic);
94 if (hdr->version != BTF_VERSION) {
95 pr_debug("Unsupported BTF version:%u\n", hdr->version);
100 pr_debug("Unsupported BTF flags:%x\n", hdr->flags);
104 meta_left = btf->data_size - sizeof(*hdr);
106 pr_debug("BTF has no data\n");
110 if (meta_left < hdr->type_off) {
111 pr_debug("Invalid BTF type section offset:%u\n", hdr->type_off);
115 if (meta_left < hdr->str_off) {
116 pr_debug("Invalid BTF string section offset:%u\n", hdr->str_off);
120 if (hdr->type_off >= hdr->str_off) {
121 pr_debug("BTF type section offset >= string section offset. No type?\n");
125 if (hdr->type_off & 0x02) {
126 pr_debug("BTF type section is not aligned to 4 bytes\n");
130 btf->nohdr_data = btf->hdr + 1;
135 static int btf_parse_str_sec(struct btf *btf)
137 const struct btf_header *hdr = btf->hdr;
138 const char *start = btf->nohdr_data + hdr->str_off;
139 const char *end = start + btf->hdr->str_len;
141 if (!hdr->str_len || hdr->str_len - 1 > BTF_MAX_STR_OFFSET ||
142 start[0] || end[-1]) {
143 pr_debug("Invalid BTF string section\n");
147 btf->strings = start;
152 static int btf_type_size(struct btf_type *t)
154 int base_size = sizeof(struct btf_type);
155 __u16 vlen = btf_vlen(t);
157 switch (btf_kind(t)) {
160 case BTF_KIND_VOLATILE:
161 case BTF_KIND_RESTRICT:
163 case BTF_KIND_TYPEDEF:
167 return base_size + sizeof(__u32);
169 return base_size + vlen * sizeof(struct btf_enum);
171 return base_size + sizeof(struct btf_array);
172 case BTF_KIND_STRUCT:
174 return base_size + vlen * sizeof(struct btf_member);
175 case BTF_KIND_FUNC_PROTO:
176 return base_size + vlen * sizeof(struct btf_param);
178 return base_size + sizeof(struct btf_var);
179 case BTF_KIND_DATASEC:
180 return base_size + vlen * sizeof(struct btf_var_secinfo);
182 pr_debug("Unsupported BTF_KIND:%u\n", btf_kind(t));
187 static int btf_parse_type_sec(struct btf *btf)
189 struct btf_header *hdr = btf->hdr;
190 void *nohdr_data = btf->nohdr_data;
191 void *next_type = nohdr_data + hdr->type_off;
192 void *end_type = nohdr_data + hdr->str_off;
194 while (next_type < end_type) {
195 struct btf_type *t = next_type;
199 type_size = btf_type_size(t);
202 next_type += type_size;
203 err = btf_add_type(btf, t);
211 __u32 btf__get_nr_types(const struct btf *btf)
213 return btf->nr_types;
216 const struct btf_type *btf__type_by_id(const struct btf *btf, __u32 type_id)
218 if (type_id > btf->nr_types)
221 return btf->types[type_id];
224 static bool btf_type_is_void(const struct btf_type *t)
226 return t == &btf_void || btf_is_fwd(t);
229 static bool btf_type_is_void_or_null(const struct btf_type *t)
231 return !t || btf_type_is_void(t);
234 #define MAX_RESOLVE_DEPTH 32
236 __s64 btf__resolve_size(const struct btf *btf, __u32 type_id)
238 const struct btf_array *array;
239 const struct btf_type *t;
244 t = btf__type_by_id(btf, type_id);
245 for (i = 0; i < MAX_RESOLVE_DEPTH && !btf_type_is_void_or_null(t);
247 switch (btf_kind(t)) {
249 case BTF_KIND_STRUCT:
252 case BTF_KIND_DATASEC:
256 size = sizeof(void *);
258 case BTF_KIND_TYPEDEF:
259 case BTF_KIND_VOLATILE:
261 case BTF_KIND_RESTRICT:
266 array = btf_array(t);
267 if (nelems && array->nelems > UINT32_MAX / nelems)
269 nelems *= array->nelems;
270 type_id = array->type;
276 t = btf__type_by_id(btf, type_id);
282 if (nelems && size > UINT32_MAX / nelems)
285 return nelems * size;
288 int btf__align_of(const struct btf *btf, __u32 id)
290 const struct btf_type *t = btf__type_by_id(btf, id);
291 __u16 kind = btf_kind(t);
296 return min(sizeof(void *), (size_t)t->size);
298 return sizeof(void *);
299 case BTF_KIND_TYPEDEF:
300 case BTF_KIND_VOLATILE:
302 case BTF_KIND_RESTRICT:
303 return btf__align_of(btf, t->type);
305 return btf__align_of(btf, btf_array(t)->type);
306 case BTF_KIND_STRUCT:
307 case BTF_KIND_UNION: {
308 const struct btf_member *m = btf_members(t);
309 __u16 vlen = btf_vlen(t);
310 int i, max_align = 1, align;
312 for (i = 0; i < vlen; i++, m++) {
313 align = btf__align_of(btf, m->type);
316 max_align = max(max_align, align);
322 pr_warn("unsupported BTF_KIND:%u\n", btf_kind(t));
327 int btf__resolve_type(const struct btf *btf, __u32 type_id)
329 const struct btf_type *t;
332 t = btf__type_by_id(btf, type_id);
333 while (depth < MAX_RESOLVE_DEPTH &&
334 !btf_type_is_void_or_null(t) &&
335 (btf_is_mod(t) || btf_is_typedef(t) || btf_is_var(t))) {
337 t = btf__type_by_id(btf, type_id);
341 if (depth == MAX_RESOLVE_DEPTH || btf_type_is_void_or_null(t))
347 __s32 btf__find_by_name(const struct btf *btf, const char *type_name)
351 if (!strcmp(type_name, "void"))
354 for (i = 1; i <= btf->nr_types; i++) {
355 const struct btf_type *t = btf->types[i];
356 const char *name = btf__name_by_offset(btf, t->name_off);
358 if (name && !strcmp(type_name, name))
365 __s32 btf__find_by_name_kind(const struct btf *btf, const char *type_name,
370 if (kind == BTF_KIND_UNKN || !strcmp(type_name, "void"))
373 for (i = 1; i <= btf->nr_types; i++) {
374 const struct btf_type *t = btf->types[i];
377 if (btf_kind(t) != kind)
379 name = btf__name_by_offset(btf, t->name_off);
380 if (name && !strcmp(type_name, name))
387 void btf__free(struct btf *btf)
400 struct btf *btf__new(__u8 *data, __u32 size)
405 btf = calloc(1, sizeof(struct btf));
407 return ERR_PTR(-ENOMEM);
411 btf->data = malloc(size);
417 memcpy(btf->data, data, size);
418 btf->data_size = size;
420 err = btf_parse_hdr(btf);
424 err = btf_parse_str_sec(btf);
428 err = btf_parse_type_sec(btf);
439 static bool btf_check_endianness(const GElf_Ehdr *ehdr)
441 #if __BYTE_ORDER == __LITTLE_ENDIAN
442 return ehdr->e_ident[EI_DATA] == ELFDATA2LSB;
443 #elif __BYTE_ORDER == __BIG_ENDIAN
444 return ehdr->e_ident[EI_DATA] == ELFDATA2MSB;
446 # error "Unrecognized __BYTE_ORDER__"
450 struct btf *btf__parse_elf(const char *path, struct btf_ext **btf_ext)
452 Elf_Data *btf_data = NULL, *btf_ext_data = NULL;
453 int err = 0, fd = -1, idx = 0;
454 struct btf *btf = NULL;
459 if (elf_version(EV_CURRENT) == EV_NONE) {
460 pr_warn("failed to init libelf for %s\n", path);
461 return ERR_PTR(-LIBBPF_ERRNO__LIBELF);
464 fd = open(path, O_RDONLY);
467 pr_warn("failed to open %s: %s\n", path, strerror(errno));
471 err = -LIBBPF_ERRNO__FORMAT;
473 elf = elf_begin(fd, ELF_C_READ, NULL);
475 pr_warn("failed to open %s as ELF file\n", path);
478 if (!gelf_getehdr(elf, &ehdr)) {
479 pr_warn("failed to get EHDR from %s\n", path);
482 if (!btf_check_endianness(&ehdr)) {
483 pr_warn("non-native ELF endianness is not supported\n");
486 if (!elf_rawdata(elf_getscn(elf, ehdr.e_shstrndx), NULL)) {
487 pr_warn("failed to get e_shstrndx from %s\n", path);
491 while ((scn = elf_nextscn(elf, scn)) != NULL) {
496 if (gelf_getshdr(scn, &sh) != &sh) {
497 pr_warn("failed to get section(%d) header from %s\n",
501 name = elf_strptr(elf, ehdr.e_shstrndx, sh.sh_name);
503 pr_warn("failed to get section(%d) name from %s\n",
507 if (strcmp(name, BTF_ELF_SEC) == 0) {
508 btf_data = elf_getdata(scn, 0);
510 pr_warn("failed to get section(%d, %s) data from %s\n",
515 } else if (btf_ext && strcmp(name, BTF_EXT_ELF_SEC) == 0) {
516 btf_ext_data = elf_getdata(scn, 0);
518 pr_warn("failed to get section(%d, %s) data from %s\n",
532 btf = btf__new(btf_data->d_buf, btf_data->d_size);
536 if (btf_ext && btf_ext_data) {
537 *btf_ext = btf_ext__new(btf_ext_data->d_buf,
538 btf_ext_data->d_size);
539 if (IS_ERR(*btf_ext))
541 } else if (btf_ext) {
552 * btf is always parsed before btf_ext, so no need to clean up
553 * btf_ext, if btf loading failed
557 if (btf_ext && IS_ERR(*btf_ext)) {
559 err = PTR_ERR(*btf_ext);
565 static int compare_vsi_off(const void *_a, const void *_b)
567 const struct btf_var_secinfo *a = _a;
568 const struct btf_var_secinfo *b = _b;
570 return a->offset - b->offset;
573 static int btf_fixup_datasec(struct bpf_object *obj, struct btf *btf,
576 __u32 size = 0, off = 0, i, vars = btf_vlen(t);
577 const char *name = btf__name_by_offset(btf, t->name_off);
578 const struct btf_type *t_var;
579 struct btf_var_secinfo *vsi;
580 const struct btf_var *var;
584 pr_debug("No name found in string section for DATASEC kind.\n");
588 /* .extern datasec size and var offsets were set correctly during
589 * extern collection step, so just skip straight to sorting variables
594 ret = bpf_object__section_size(obj, name, &size);
595 if (ret || !size || (t->size && t->size != size)) {
596 pr_debug("Invalid size for section %s: %u bytes\n", name, size);
602 for (i = 0, vsi = btf_var_secinfos(t); i < vars; i++, vsi++) {
603 t_var = btf__type_by_id(btf, vsi->type);
604 var = btf_var(t_var);
606 if (!btf_is_var(t_var)) {
607 pr_debug("Non-VAR type seen in section %s\n", name);
611 if (var->linkage == BTF_VAR_STATIC)
614 name = btf__name_by_offset(btf, t_var->name_off);
616 pr_debug("No name found in string section for VAR kind\n");
620 ret = bpf_object__variable_offset(obj, name, &off);
622 pr_debug("No offset found in symbol table for VAR %s\n",
631 qsort(btf_var_secinfos(t), vars, sizeof(*vsi), compare_vsi_off);
635 int btf__finalize_data(struct bpf_object *obj, struct btf *btf)
640 for (i = 1; i <= btf->nr_types; i++) {
641 struct btf_type *t = btf->types[i];
643 /* Loader needs to fix up some of the things compiler
644 * couldn't get its hands on while emitting BTF. This
645 * is section size and global variable offset. We use
646 * the info from the ELF itself for this purpose.
648 if (btf_is_datasec(t)) {
649 err = btf_fixup_datasec(obj, btf, t);
658 int btf__load(struct btf *btf)
660 __u32 log_buf_size = BPF_LOG_BUF_SIZE;
661 char *log_buf = NULL;
667 log_buf = malloc(log_buf_size);
673 btf->fd = bpf_load_btf(btf->data, btf->data_size,
674 log_buf, log_buf_size, false);
677 pr_warn("Error loading BTF: %s(%d)\n", strerror(errno), errno);
679 pr_warn("%s\n", log_buf);
688 int btf__fd(const struct btf *btf)
693 const void *btf__get_raw_data(const struct btf *btf, __u32 *size)
695 *size = btf->data_size;
699 const char *btf__name_by_offset(const struct btf *btf, __u32 offset)
701 if (offset < btf->hdr->str_len)
702 return &btf->strings[offset];
707 int btf__get_from_id(__u32 id, struct btf **btf)
709 struct bpf_btf_info btf_info = { 0 };
710 __u32 len = sizeof(btf_info);
718 btf_fd = bpf_btf_get_fd_by_id(id);
722 /* we won't know btf_size until we call bpf_obj_get_info_by_fd(). so
723 * let's start with a sane default - 4KiB here - and resize it only if
724 * bpf_obj_get_info_by_fd() needs a bigger buffer.
726 btf_info.btf_size = 4096;
727 last_size = btf_info.btf_size;
728 ptr = malloc(last_size);
734 memset(ptr, 0, last_size);
735 btf_info.btf = ptr_to_u64(ptr);
736 err = bpf_obj_get_info_by_fd(btf_fd, &btf_info, &len);
738 if (!err && btf_info.btf_size > last_size) {
741 last_size = btf_info.btf_size;
742 temp_ptr = realloc(ptr, last_size);
748 memset(ptr, 0, last_size);
749 btf_info.btf = ptr_to_u64(ptr);
750 err = bpf_obj_get_info_by_fd(btf_fd, &btf_info, &len);
753 if (err || btf_info.btf_size > last_size) {
758 *btf = btf__new((__u8 *)(long)btf_info.btf, btf_info.btf_size);
771 int btf__get_map_kv_tids(const struct btf *btf, const char *map_name,
772 __u32 expected_key_size, __u32 expected_value_size,
773 __u32 *key_type_id, __u32 *value_type_id)
775 const struct btf_type *container_type;
776 const struct btf_member *key, *value;
777 const size_t max_name = 256;
778 char container_name[max_name];
779 __s64 key_size, value_size;
782 if (snprintf(container_name, max_name, "____btf_map_%s", map_name) ==
784 pr_warn("map:%s length of '____btf_map_%s' is too long\n",
789 container_id = btf__find_by_name(btf, container_name);
790 if (container_id < 0) {
791 pr_debug("map:%s container_name:%s cannot be found in BTF. Missing BPF_ANNOTATE_KV_PAIR?\n",
792 map_name, container_name);
796 container_type = btf__type_by_id(btf, container_id);
797 if (!container_type) {
798 pr_warn("map:%s cannot find BTF type for container_id:%u\n",
799 map_name, container_id);
803 if (!btf_is_struct(container_type) || btf_vlen(container_type) < 2) {
804 pr_warn("map:%s container_name:%s is an invalid container struct\n",
805 map_name, container_name);
809 key = btf_members(container_type);
812 key_size = btf__resolve_size(btf, key->type);
814 pr_warn("map:%s invalid BTF key_type_size\n", map_name);
818 if (expected_key_size != key_size) {
819 pr_warn("map:%s btf_key_type_size:%u != map_def_key_size:%u\n",
820 map_name, (__u32)key_size, expected_key_size);
824 value_size = btf__resolve_size(btf, value->type);
825 if (value_size < 0) {
826 pr_warn("map:%s invalid BTF value_type_size\n", map_name);
830 if (expected_value_size != value_size) {
831 pr_warn("map:%s btf_value_type_size:%u != map_def_value_size:%u\n",
832 map_name, (__u32)value_size, expected_value_size);
836 *key_type_id = key->type;
837 *value_type_id = value->type;
842 struct btf_ext_sec_setup_param {
846 struct btf_ext_info *ext_info;
850 static int btf_ext_setup_info(struct btf_ext *btf_ext,
851 struct btf_ext_sec_setup_param *ext_sec)
853 const struct btf_ext_info_sec *sinfo;
854 struct btf_ext_info *ext_info;
855 __u32 info_left, record_size;
856 /* The start of the info sec (including the __u32 record_size). */
859 if (ext_sec->len == 0)
862 if (ext_sec->off & 0x03) {
863 pr_debug(".BTF.ext %s section is not aligned to 4 bytes\n",
868 info = btf_ext->data + btf_ext->hdr->hdr_len + ext_sec->off;
869 info_left = ext_sec->len;
871 if (btf_ext->data + btf_ext->data_size < info + ext_sec->len) {
872 pr_debug("%s section (off:%u len:%u) is beyond the end of the ELF section .BTF.ext\n",
873 ext_sec->desc, ext_sec->off, ext_sec->len);
877 /* At least a record size */
878 if (info_left < sizeof(__u32)) {
879 pr_debug(".BTF.ext %s record size not found\n", ext_sec->desc);
883 /* The record size needs to meet the minimum standard */
884 record_size = *(__u32 *)info;
885 if (record_size < ext_sec->min_rec_size ||
886 record_size & 0x03) {
887 pr_debug("%s section in .BTF.ext has invalid record size %u\n",
888 ext_sec->desc, record_size);
892 sinfo = info + sizeof(__u32);
893 info_left -= sizeof(__u32);
895 /* If no records, return failure now so .BTF.ext won't be used. */
897 pr_debug("%s section in .BTF.ext has no records", ext_sec->desc);
902 unsigned int sec_hdrlen = sizeof(struct btf_ext_info_sec);
903 __u64 total_record_size;
906 if (info_left < sec_hdrlen) {
907 pr_debug("%s section header is not found in .BTF.ext\n",
912 num_records = sinfo->num_info;
913 if (num_records == 0) {
914 pr_debug("%s section has incorrect num_records in .BTF.ext\n",
919 total_record_size = sec_hdrlen +
920 (__u64)num_records * record_size;
921 if (info_left < total_record_size) {
922 pr_debug("%s section has incorrect num_records in .BTF.ext\n",
927 info_left -= total_record_size;
928 sinfo = (void *)sinfo + total_record_size;
931 ext_info = ext_sec->ext_info;
932 ext_info->len = ext_sec->len - sizeof(__u32);
933 ext_info->rec_size = record_size;
934 ext_info->info = info + sizeof(__u32);
939 static int btf_ext_setup_func_info(struct btf_ext *btf_ext)
941 struct btf_ext_sec_setup_param param = {
942 .off = btf_ext->hdr->func_info_off,
943 .len = btf_ext->hdr->func_info_len,
944 .min_rec_size = sizeof(struct bpf_func_info_min),
945 .ext_info = &btf_ext->func_info,
949 return btf_ext_setup_info(btf_ext, ¶m);
952 static int btf_ext_setup_line_info(struct btf_ext *btf_ext)
954 struct btf_ext_sec_setup_param param = {
955 .off = btf_ext->hdr->line_info_off,
956 .len = btf_ext->hdr->line_info_len,
957 .min_rec_size = sizeof(struct bpf_line_info_min),
958 .ext_info = &btf_ext->line_info,
962 return btf_ext_setup_info(btf_ext, ¶m);
965 static int btf_ext_setup_field_reloc(struct btf_ext *btf_ext)
967 struct btf_ext_sec_setup_param param = {
968 .off = btf_ext->hdr->field_reloc_off,
969 .len = btf_ext->hdr->field_reloc_len,
970 .min_rec_size = sizeof(struct bpf_field_reloc),
971 .ext_info = &btf_ext->field_reloc_info,
972 .desc = "field_reloc",
975 return btf_ext_setup_info(btf_ext, ¶m);
978 static int btf_ext_parse_hdr(__u8 *data, __u32 data_size)
980 const struct btf_ext_header *hdr = (struct btf_ext_header *)data;
982 if (data_size < offsetofend(struct btf_ext_header, hdr_len) ||
983 data_size < hdr->hdr_len) {
984 pr_debug("BTF.ext header not found");
988 if (hdr->magic != BTF_MAGIC) {
989 pr_debug("Invalid BTF.ext magic:%x\n", hdr->magic);
993 if (hdr->version != BTF_VERSION) {
994 pr_debug("Unsupported BTF.ext version:%u\n", hdr->version);
999 pr_debug("Unsupported BTF.ext flags:%x\n", hdr->flags);
1003 if (data_size == hdr->hdr_len) {
1004 pr_debug("BTF.ext has no data\n");
1011 void btf_ext__free(struct btf_ext *btf_ext)
1015 free(btf_ext->data);
1019 struct btf_ext *btf_ext__new(__u8 *data, __u32 size)
1021 struct btf_ext *btf_ext;
1024 err = btf_ext_parse_hdr(data, size);
1026 return ERR_PTR(err);
1028 btf_ext = calloc(1, sizeof(struct btf_ext));
1030 return ERR_PTR(-ENOMEM);
1032 btf_ext->data_size = size;
1033 btf_ext->data = malloc(size);
1034 if (!btf_ext->data) {
1038 memcpy(btf_ext->data, data, size);
1040 if (btf_ext->hdr->hdr_len <
1041 offsetofend(struct btf_ext_header, line_info_len))
1043 err = btf_ext_setup_func_info(btf_ext);
1047 err = btf_ext_setup_line_info(btf_ext);
1051 if (btf_ext->hdr->hdr_len <
1052 offsetofend(struct btf_ext_header, field_reloc_len))
1054 err = btf_ext_setup_field_reloc(btf_ext);
1060 btf_ext__free(btf_ext);
1061 return ERR_PTR(err);
1067 const void *btf_ext__get_raw_data(const struct btf_ext *btf_ext, __u32 *size)
1069 *size = btf_ext->data_size;
1070 return btf_ext->data;
1073 static int btf_ext_reloc_info(const struct btf *btf,
1074 const struct btf_ext_info *ext_info,
1075 const char *sec_name, __u32 insns_cnt,
1076 void **info, __u32 *cnt)
1078 __u32 sec_hdrlen = sizeof(struct btf_ext_info_sec);
1079 __u32 i, record_size, existing_len, records_len;
1080 struct btf_ext_info_sec *sinfo;
1081 const char *info_sec_name;
1085 record_size = ext_info->rec_size;
1086 sinfo = ext_info->info;
1087 remain_len = ext_info->len;
1088 while (remain_len > 0) {
1089 records_len = sinfo->num_info * record_size;
1090 info_sec_name = btf__name_by_offset(btf, sinfo->sec_name_off);
1091 if (strcmp(info_sec_name, sec_name)) {
1092 remain_len -= sec_hdrlen + records_len;
1093 sinfo = (void *)sinfo + sec_hdrlen + records_len;
1097 existing_len = (*cnt) * record_size;
1098 data = realloc(*info, existing_len + records_len);
1102 memcpy(data + existing_len, sinfo->data, records_len);
1103 /* adjust insn_off only, the rest data will be passed
1106 for (i = 0; i < sinfo->num_info; i++) {
1109 insn_off = data + existing_len + (i * record_size);
1110 *insn_off = *insn_off / sizeof(struct bpf_insn) +
1114 *cnt += sinfo->num_info;
1121 int btf_ext__reloc_func_info(const struct btf *btf,
1122 const struct btf_ext *btf_ext,
1123 const char *sec_name, __u32 insns_cnt,
1124 void **func_info, __u32 *cnt)
1126 return btf_ext_reloc_info(btf, &btf_ext->func_info, sec_name,
1127 insns_cnt, func_info, cnt);
1130 int btf_ext__reloc_line_info(const struct btf *btf,
1131 const struct btf_ext *btf_ext,
1132 const char *sec_name, __u32 insns_cnt,
1133 void **line_info, __u32 *cnt)
1135 return btf_ext_reloc_info(btf, &btf_ext->line_info, sec_name,
1136 insns_cnt, line_info, cnt);
1139 __u32 btf_ext__func_info_rec_size(const struct btf_ext *btf_ext)
1141 return btf_ext->func_info.rec_size;
1144 __u32 btf_ext__line_info_rec_size(const struct btf_ext *btf_ext)
1146 return btf_ext->line_info.rec_size;
1151 static struct btf_dedup *btf_dedup_new(struct btf *btf, struct btf_ext *btf_ext,
1152 const struct btf_dedup_opts *opts);
1153 static void btf_dedup_free(struct btf_dedup *d);
1154 static int btf_dedup_strings(struct btf_dedup *d);
1155 static int btf_dedup_prim_types(struct btf_dedup *d);
1156 static int btf_dedup_struct_types(struct btf_dedup *d);
1157 static int btf_dedup_ref_types(struct btf_dedup *d);
1158 static int btf_dedup_compact_types(struct btf_dedup *d);
1159 static int btf_dedup_remap_types(struct btf_dedup *d);
1162 * Deduplicate BTF types and strings.
1164 * BTF dedup algorithm takes as an input `struct btf` representing `.BTF` ELF
1165 * section with all BTF type descriptors and string data. It overwrites that
1166 * memory in-place with deduplicated types and strings without any loss of
1167 * information. If optional `struct btf_ext` representing '.BTF.ext' ELF section
1168 * is provided, all the strings referenced from .BTF.ext section are honored
1169 * and updated to point to the right offsets after deduplication.
1171 * If function returns with error, type/string data might be garbled and should
1174 * More verbose and detailed description of both problem btf_dedup is solving,
1175 * as well as solution could be found at:
1176 * https://facebookmicrosites.github.io/bpf/blog/2018/11/14/btf-enhancement.html
1178 * Problem description and justification
1179 * =====================================
1181 * BTF type information is typically emitted either as a result of conversion
1182 * from DWARF to BTF or directly by compiler. In both cases, each compilation
1183 * unit contains information about a subset of all the types that are used
1184 * in an application. These subsets are frequently overlapping and contain a lot
1185 * of duplicated information when later concatenated together into a single
1186 * binary. This algorithm ensures that each unique type is represented by single
1187 * BTF type descriptor, greatly reducing resulting size of BTF data.
1189 * Compilation unit isolation and subsequent duplication of data is not the only
1190 * problem. The same type hierarchy (e.g., struct and all the type that struct
1191 * references) in different compilation units can be represented in BTF to
1192 * various degrees of completeness (or, rather, incompleteness) due to
1193 * struct/union forward declarations.
1195 * Let's take a look at an example, that we'll use to better understand the
1196 * problem (and solution). Suppose we have two compilation units, each using
1197 * same `struct S`, but each of them having incomplete type information about
1226 * In case of CU #1, BTF data will know only that `struct B` exist (but no
1227 * more), but will know the complete type information about `struct A`. While
1228 * for CU #2, it will know full type information about `struct B`, but will
1229 * only know about forward declaration of `struct A` (in BTF terms, it will
1230 * have `BTF_KIND_FWD` type descriptor with name `B`).
1232 * This compilation unit isolation means that it's possible that there is no
1233 * single CU with complete type information describing structs `S`, `A`, and
1234 * `B`. Also, we might get tons of duplicated and redundant type information.
1236 * Additional complication we need to keep in mind comes from the fact that
1237 * types, in general, can form graphs containing cycles, not just DAGs.
1239 * While algorithm does deduplication, it also merges and resolves type
1240 * information (unless disabled throught `struct btf_opts`), whenever possible.
1241 * E.g., in the example above with two compilation units having partial type
1242 * information for structs `A` and `B`, the output of algorithm will emit
1243 * a single copy of each BTF type that describes structs `A`, `B`, and `S`
1244 * (as well as type information for `int` and pointers), as if they were defined
1245 * in a single compilation unit as:
1265 * Algorithm completes its work in 6 separate passes:
1267 * 1. Strings deduplication.
1268 * 2. Primitive types deduplication (int, enum, fwd).
1269 * 3. Struct/union types deduplication.
1270 * 4. Reference types deduplication (pointers, typedefs, arrays, funcs, func
1271 * protos, and const/volatile/restrict modifiers).
1272 * 5. Types compaction.
1273 * 6. Types remapping.
1275 * Algorithm determines canonical type descriptor, which is a single
1276 * representative type for each truly unique type. This canonical type is the
1277 * one that will go into final deduplicated BTF type information. For
1278 * struct/unions, it is also the type that algorithm will merge additional type
1279 * information into (while resolving FWDs), as it discovers it from data in
1280 * other CUs. Each input BTF type eventually gets either mapped to itself, if
1281 * that type is canonical, or to some other type, if that type is equivalent
1282 * and was chosen as canonical representative. This mapping is stored in
1283 * `btf_dedup->map` array. This map is also used to record STRUCT/UNION that
1284 * FWD type got resolved to.
1286 * To facilitate fast discovery of canonical types, we also maintain canonical
1287 * index (`btf_dedup->dedup_table`), which maps type descriptor's signature hash
1288 * (i.e., hashed kind, name, size, fields, etc) into a list of canonical types
1289 * that match that signature. With sufficiently good choice of type signature
1290 * hashing function, we can limit number of canonical types for each unique type
1291 * signature to a very small number, allowing to find canonical type for any
1292 * duplicated type very quickly.
1294 * Struct/union deduplication is the most critical part and algorithm for
1295 * deduplicating structs/unions is described in greater details in comments for
1296 * `btf_dedup_is_equiv` function.
1298 int btf__dedup(struct btf *btf, struct btf_ext *btf_ext,
1299 const struct btf_dedup_opts *opts)
1301 struct btf_dedup *d = btf_dedup_new(btf, btf_ext, opts);
1305 pr_debug("btf_dedup_new failed: %ld", PTR_ERR(d));
1309 err = btf_dedup_strings(d);
1311 pr_debug("btf_dedup_strings failed:%d\n", err);
1314 err = btf_dedup_prim_types(d);
1316 pr_debug("btf_dedup_prim_types failed:%d\n", err);
1319 err = btf_dedup_struct_types(d);
1321 pr_debug("btf_dedup_struct_types failed:%d\n", err);
1324 err = btf_dedup_ref_types(d);
1326 pr_debug("btf_dedup_ref_types failed:%d\n", err);
1329 err = btf_dedup_compact_types(d);
1331 pr_debug("btf_dedup_compact_types failed:%d\n", err);
1334 err = btf_dedup_remap_types(d);
1336 pr_debug("btf_dedup_remap_types failed:%d\n", err);
1345 #define BTF_UNPROCESSED_ID ((__u32)-1)
1346 #define BTF_IN_PROGRESS_ID ((__u32)-2)
1349 /* .BTF section to be deduped in-place */
1352 * Optional .BTF.ext section. When provided, any strings referenced
1353 * from it will be taken into account when deduping strings
1355 struct btf_ext *btf_ext;
1357 * This is a map from any type's signature hash to a list of possible
1358 * canonical representative type candidates. Hash collisions are
1359 * ignored, so even types of various kinds can share same list of
1360 * candidates, which is fine because we rely on subsequent
1361 * btf_xxx_equal() checks to authoritatively verify type equality.
1363 struct hashmap *dedup_table;
1364 /* Canonical types map */
1366 /* Hypothetical mapping, used during type graph equivalence checks */
1371 /* Various option modifying behavior of algorithm */
1372 struct btf_dedup_opts opts;
1375 struct btf_str_ptr {
1381 struct btf_str_ptrs {
1382 struct btf_str_ptr *ptrs;
1388 static long hash_combine(long h, long value)
1390 return h * 31 + value;
1393 #define for_each_dedup_cand(d, node, hash) \
1394 hashmap__for_each_key_entry(d->dedup_table, node, (void *)hash)
1396 static int btf_dedup_table_add(struct btf_dedup *d, long hash, __u32 type_id)
1398 return hashmap__append(d->dedup_table,
1399 (void *)hash, (void *)(long)type_id);
1402 static int btf_dedup_hypot_map_add(struct btf_dedup *d,
1403 __u32 from_id, __u32 to_id)
1405 if (d->hypot_cnt == d->hypot_cap) {
1408 d->hypot_cap += max((size_t)16, d->hypot_cap / 2);
1409 new_list = realloc(d->hypot_list, sizeof(__u32) * d->hypot_cap);
1412 d->hypot_list = new_list;
1414 d->hypot_list[d->hypot_cnt++] = from_id;
1415 d->hypot_map[from_id] = to_id;
1419 static void btf_dedup_clear_hypot_map(struct btf_dedup *d)
1423 for (i = 0; i < d->hypot_cnt; i++)
1424 d->hypot_map[d->hypot_list[i]] = BTF_UNPROCESSED_ID;
1428 static void btf_dedup_free(struct btf_dedup *d)
1430 hashmap__free(d->dedup_table);
1431 d->dedup_table = NULL;
1437 d->hypot_map = NULL;
1439 free(d->hypot_list);
1440 d->hypot_list = NULL;
1445 static size_t btf_dedup_identity_hash_fn(const void *key, void *ctx)
1450 static size_t btf_dedup_collision_hash_fn(const void *key, void *ctx)
1455 static bool btf_dedup_equal_fn(const void *k1, const void *k2, void *ctx)
1460 static struct btf_dedup *btf_dedup_new(struct btf *btf, struct btf_ext *btf_ext,
1461 const struct btf_dedup_opts *opts)
1463 struct btf_dedup *d = calloc(1, sizeof(struct btf_dedup));
1464 hashmap_hash_fn hash_fn = btf_dedup_identity_hash_fn;
1468 return ERR_PTR(-ENOMEM);
1470 d->opts.dont_resolve_fwds = opts && opts->dont_resolve_fwds;
1471 /* dedup_table_size is now used only to force collisions in tests */
1472 if (opts && opts->dedup_table_size == 1)
1473 hash_fn = btf_dedup_collision_hash_fn;
1476 d->btf_ext = btf_ext;
1478 d->dedup_table = hashmap__new(hash_fn, btf_dedup_equal_fn, NULL);
1479 if (IS_ERR(d->dedup_table)) {
1480 err = PTR_ERR(d->dedup_table);
1481 d->dedup_table = NULL;
1485 d->map = malloc(sizeof(__u32) * (1 + btf->nr_types));
1490 /* special BTF "void" type is made canonical immediately */
1492 for (i = 1; i <= btf->nr_types; i++) {
1493 struct btf_type *t = d->btf->types[i];
1495 /* VAR and DATASEC are never deduped and are self-canonical */
1496 if (btf_is_var(t) || btf_is_datasec(t))
1499 d->map[i] = BTF_UNPROCESSED_ID;
1502 d->hypot_map = malloc(sizeof(__u32) * (1 + btf->nr_types));
1503 if (!d->hypot_map) {
1507 for (i = 0; i <= btf->nr_types; i++)
1508 d->hypot_map[i] = BTF_UNPROCESSED_ID;
1513 return ERR_PTR(err);
1519 typedef int (*str_off_fn_t)(__u32 *str_off_ptr, void *ctx);
1522 * Iterate over all possible places in .BTF and .BTF.ext that can reference
1523 * string and pass pointer to it to a provided callback `fn`.
1525 static int btf_for_each_str_off(struct btf_dedup *d, str_off_fn_t fn, void *ctx)
1527 void *line_data_cur, *line_data_end;
1528 int i, j, r, rec_size;
1531 for (i = 1; i <= d->btf->nr_types; i++) {
1532 t = d->btf->types[i];
1533 r = fn(&t->name_off, ctx);
1537 switch (btf_kind(t)) {
1538 case BTF_KIND_STRUCT:
1539 case BTF_KIND_UNION: {
1540 struct btf_member *m = btf_members(t);
1541 __u16 vlen = btf_vlen(t);
1543 for (j = 0; j < vlen; j++) {
1544 r = fn(&m->name_off, ctx);
1551 case BTF_KIND_ENUM: {
1552 struct btf_enum *m = btf_enum(t);
1553 __u16 vlen = btf_vlen(t);
1555 for (j = 0; j < vlen; j++) {
1556 r = fn(&m->name_off, ctx);
1563 case BTF_KIND_FUNC_PROTO: {
1564 struct btf_param *m = btf_params(t);
1565 __u16 vlen = btf_vlen(t);
1567 for (j = 0; j < vlen; j++) {
1568 r = fn(&m->name_off, ctx);
1583 line_data_cur = d->btf_ext->line_info.info;
1584 line_data_end = d->btf_ext->line_info.info + d->btf_ext->line_info.len;
1585 rec_size = d->btf_ext->line_info.rec_size;
1587 while (line_data_cur < line_data_end) {
1588 struct btf_ext_info_sec *sec = line_data_cur;
1589 struct bpf_line_info_min *line_info;
1590 __u32 num_info = sec->num_info;
1592 r = fn(&sec->sec_name_off, ctx);
1596 line_data_cur += sizeof(struct btf_ext_info_sec);
1597 for (i = 0; i < num_info; i++) {
1598 line_info = line_data_cur;
1599 r = fn(&line_info->file_name_off, ctx);
1602 r = fn(&line_info->line_off, ctx);
1605 line_data_cur += rec_size;
1612 static int str_sort_by_content(const void *a1, const void *a2)
1614 const struct btf_str_ptr *p1 = a1;
1615 const struct btf_str_ptr *p2 = a2;
1617 return strcmp(p1->str, p2->str);
1620 static int str_sort_by_offset(const void *a1, const void *a2)
1622 const struct btf_str_ptr *p1 = a1;
1623 const struct btf_str_ptr *p2 = a2;
1625 if (p1->str != p2->str)
1626 return p1->str < p2->str ? -1 : 1;
1630 static int btf_dedup_str_ptr_cmp(const void *str_ptr, const void *pelem)
1632 const struct btf_str_ptr *p = pelem;
1634 if (str_ptr != p->str)
1635 return (const char *)str_ptr < p->str ? -1 : 1;
1639 static int btf_str_mark_as_used(__u32 *str_off_ptr, void *ctx)
1641 struct btf_str_ptrs *strs;
1642 struct btf_str_ptr *s;
1644 if (*str_off_ptr == 0)
1648 s = bsearch(strs->data + *str_off_ptr, strs->ptrs, strs->cnt,
1649 sizeof(struct btf_str_ptr), btf_dedup_str_ptr_cmp);
1656 static int btf_str_remap_offset(__u32 *str_off_ptr, void *ctx)
1658 struct btf_str_ptrs *strs;
1659 struct btf_str_ptr *s;
1661 if (*str_off_ptr == 0)
1665 s = bsearch(strs->data + *str_off_ptr, strs->ptrs, strs->cnt,
1666 sizeof(struct btf_str_ptr), btf_dedup_str_ptr_cmp);
1669 *str_off_ptr = s->new_off;
1674 * Dedup string and filter out those that are not referenced from either .BTF
1675 * or .BTF.ext (if provided) sections.
1677 * This is done by building index of all strings in BTF's string section,
1678 * then iterating over all entities that can reference strings (e.g., type
1679 * names, struct field names, .BTF.ext line info, etc) and marking corresponding
1680 * strings as used. After that all used strings are deduped and compacted into
1681 * sequential blob of memory and new offsets are calculated. Then all the string
1682 * references are iterated again and rewritten using new offsets.
1684 static int btf_dedup_strings(struct btf_dedup *d)
1686 const struct btf_header *hdr = d->btf->hdr;
1687 char *start = (char *)d->btf->nohdr_data + hdr->str_off;
1688 char *end = start + d->btf->hdr->str_len;
1689 char *p = start, *tmp_strs = NULL;
1690 struct btf_str_ptrs strs = {
1696 int i, j, err = 0, grp_idx;
1699 /* build index of all strings */
1701 if (strs.cnt + 1 > strs.cap) {
1702 struct btf_str_ptr *new_ptrs;
1704 strs.cap += max(strs.cnt / 2, 16U);
1705 new_ptrs = realloc(strs.ptrs,
1706 sizeof(strs.ptrs[0]) * strs.cap);
1711 strs.ptrs = new_ptrs;
1714 strs.ptrs[strs.cnt].str = p;
1715 strs.ptrs[strs.cnt].used = false;
1721 /* temporary storage for deduplicated strings */
1722 tmp_strs = malloc(d->btf->hdr->str_len);
1728 /* mark all used strings */
1729 strs.ptrs[0].used = true;
1730 err = btf_for_each_str_off(d, btf_str_mark_as_used, &strs);
1734 /* sort strings by context, so that we can identify duplicates */
1735 qsort(strs.ptrs, strs.cnt, sizeof(strs.ptrs[0]), str_sort_by_content);
1738 * iterate groups of equal strings and if any instance in a group was
1739 * referenced, emit single instance and remember new offset
1743 grp_used = strs.ptrs[0].used;
1744 /* iterate past end to avoid code duplication after loop */
1745 for (i = 1; i <= strs.cnt; i++) {
1747 * when i == strs.cnt, we want to skip string comparison and go
1748 * straight to handling last group of strings (otherwise we'd
1749 * need to handle last group after the loop w/ duplicated code)
1752 !strcmp(strs.ptrs[i].str, strs.ptrs[grp_idx].str)) {
1753 grp_used = grp_used || strs.ptrs[i].used;
1758 * this check would have been required after the loop to handle
1759 * last group of strings, but due to <= condition in a loop
1760 * we avoid that duplication
1763 int new_off = p - tmp_strs;
1764 __u32 len = strlen(strs.ptrs[grp_idx].str);
1766 memmove(p, strs.ptrs[grp_idx].str, len + 1);
1767 for (j = grp_idx; j < i; j++)
1768 strs.ptrs[j].new_off = new_off;
1774 grp_used = strs.ptrs[i].used;
1778 /* replace original strings with deduped ones */
1779 d->btf->hdr->str_len = p - tmp_strs;
1780 memmove(start, tmp_strs, d->btf->hdr->str_len);
1781 end = start + d->btf->hdr->str_len;
1783 /* restore original order for further binary search lookups */
1784 qsort(strs.ptrs, strs.cnt, sizeof(strs.ptrs[0]), str_sort_by_offset);
1786 /* remap string offsets */
1787 err = btf_for_each_str_off(d, btf_str_remap_offset, &strs);
1791 d->btf->hdr->str_len = end - start;
1799 static long btf_hash_common(struct btf_type *t)
1803 h = hash_combine(0, t->name_off);
1804 h = hash_combine(h, t->info);
1805 h = hash_combine(h, t->size);
1809 static bool btf_equal_common(struct btf_type *t1, struct btf_type *t2)
1811 return t1->name_off == t2->name_off &&
1812 t1->info == t2->info &&
1813 t1->size == t2->size;
1816 /* Calculate type signature hash of INT. */
1817 static long btf_hash_int(struct btf_type *t)
1819 __u32 info = *(__u32 *)(t + 1);
1822 h = btf_hash_common(t);
1823 h = hash_combine(h, info);
1827 /* Check structural equality of two INTs. */
1828 static bool btf_equal_int(struct btf_type *t1, struct btf_type *t2)
1832 if (!btf_equal_common(t1, t2))
1834 info1 = *(__u32 *)(t1 + 1);
1835 info2 = *(__u32 *)(t2 + 1);
1836 return info1 == info2;
1839 /* Calculate type signature hash of ENUM. */
1840 static long btf_hash_enum(struct btf_type *t)
1844 /* don't hash vlen and enum members to support enum fwd resolving */
1845 h = hash_combine(0, t->name_off);
1846 h = hash_combine(h, t->info & ~0xffff);
1847 h = hash_combine(h, t->size);
1851 /* Check structural equality of two ENUMs. */
1852 static bool btf_equal_enum(struct btf_type *t1, struct btf_type *t2)
1854 const struct btf_enum *m1, *m2;
1858 if (!btf_equal_common(t1, t2))
1861 vlen = btf_vlen(t1);
1864 for (i = 0; i < vlen; i++) {
1865 if (m1->name_off != m2->name_off || m1->val != m2->val)
1873 static inline bool btf_is_enum_fwd(struct btf_type *t)
1875 return btf_is_enum(t) && btf_vlen(t) == 0;
1878 static bool btf_compat_enum(struct btf_type *t1, struct btf_type *t2)
1880 if (!btf_is_enum_fwd(t1) && !btf_is_enum_fwd(t2))
1881 return btf_equal_enum(t1, t2);
1882 /* ignore vlen when comparing */
1883 return t1->name_off == t2->name_off &&
1884 (t1->info & ~0xffff) == (t2->info & ~0xffff) &&
1885 t1->size == t2->size;
1889 * Calculate type signature hash of STRUCT/UNION, ignoring referenced type IDs,
1890 * as referenced type IDs equivalence is established separately during type
1891 * graph equivalence check algorithm.
1893 static long btf_hash_struct(struct btf_type *t)
1895 const struct btf_member *member = btf_members(t);
1896 __u32 vlen = btf_vlen(t);
1897 long h = btf_hash_common(t);
1900 for (i = 0; i < vlen; i++) {
1901 h = hash_combine(h, member->name_off);
1902 h = hash_combine(h, member->offset);
1903 /* no hashing of referenced type ID, it can be unresolved yet */
1910 * Check structural compatibility of two FUNC_PROTOs, ignoring referenced type
1911 * IDs. This check is performed during type graph equivalence check and
1912 * referenced types equivalence is checked separately.
1914 static bool btf_shallow_equal_struct(struct btf_type *t1, struct btf_type *t2)
1916 const struct btf_member *m1, *m2;
1920 if (!btf_equal_common(t1, t2))
1923 vlen = btf_vlen(t1);
1924 m1 = btf_members(t1);
1925 m2 = btf_members(t2);
1926 for (i = 0; i < vlen; i++) {
1927 if (m1->name_off != m2->name_off || m1->offset != m2->offset)
1936 * Calculate type signature hash of ARRAY, including referenced type IDs,
1937 * under assumption that they were already resolved to canonical type IDs and
1938 * are not going to change.
1940 static long btf_hash_array(struct btf_type *t)
1942 const struct btf_array *info = btf_array(t);
1943 long h = btf_hash_common(t);
1945 h = hash_combine(h, info->type);
1946 h = hash_combine(h, info->index_type);
1947 h = hash_combine(h, info->nelems);
1952 * Check exact equality of two ARRAYs, taking into account referenced
1953 * type IDs, under assumption that they were already resolved to canonical
1954 * type IDs and are not going to change.
1955 * This function is called during reference types deduplication to compare
1956 * ARRAY to potential canonical representative.
1958 static bool btf_equal_array(struct btf_type *t1, struct btf_type *t2)
1960 const struct btf_array *info1, *info2;
1962 if (!btf_equal_common(t1, t2))
1965 info1 = btf_array(t1);
1966 info2 = btf_array(t2);
1967 return info1->type == info2->type &&
1968 info1->index_type == info2->index_type &&
1969 info1->nelems == info2->nelems;
1973 * Check structural compatibility of two ARRAYs, ignoring referenced type
1974 * IDs. This check is performed during type graph equivalence check and
1975 * referenced types equivalence is checked separately.
1977 static bool btf_compat_array(struct btf_type *t1, struct btf_type *t2)
1979 if (!btf_equal_common(t1, t2))
1982 return btf_array(t1)->nelems == btf_array(t2)->nelems;
1986 * Calculate type signature hash of FUNC_PROTO, including referenced type IDs,
1987 * under assumption that they were already resolved to canonical type IDs and
1988 * are not going to change.
1990 static long btf_hash_fnproto(struct btf_type *t)
1992 const struct btf_param *member = btf_params(t);
1993 __u16 vlen = btf_vlen(t);
1994 long h = btf_hash_common(t);
1997 for (i = 0; i < vlen; i++) {
1998 h = hash_combine(h, member->name_off);
1999 h = hash_combine(h, member->type);
2006 * Check exact equality of two FUNC_PROTOs, taking into account referenced
2007 * type IDs, under assumption that they were already resolved to canonical
2008 * type IDs and are not going to change.
2009 * This function is called during reference types deduplication to compare
2010 * FUNC_PROTO to potential canonical representative.
2012 static bool btf_equal_fnproto(struct btf_type *t1, struct btf_type *t2)
2014 const struct btf_param *m1, *m2;
2018 if (!btf_equal_common(t1, t2))
2021 vlen = btf_vlen(t1);
2022 m1 = btf_params(t1);
2023 m2 = btf_params(t2);
2024 for (i = 0; i < vlen; i++) {
2025 if (m1->name_off != m2->name_off || m1->type != m2->type)
2034 * Check structural compatibility of two FUNC_PROTOs, ignoring referenced type
2035 * IDs. This check is performed during type graph equivalence check and
2036 * referenced types equivalence is checked separately.
2038 static bool btf_compat_fnproto(struct btf_type *t1, struct btf_type *t2)
2040 const struct btf_param *m1, *m2;
2044 /* skip return type ID */
2045 if (t1->name_off != t2->name_off || t1->info != t2->info)
2048 vlen = btf_vlen(t1);
2049 m1 = btf_params(t1);
2050 m2 = btf_params(t2);
2051 for (i = 0; i < vlen; i++) {
2052 if (m1->name_off != m2->name_off)
2061 * Deduplicate primitive types, that can't reference other types, by calculating
2062 * their type signature hash and comparing them with any possible canonical
2063 * candidate. If no canonical candidate matches, type itself is marked as
2064 * canonical and is added into `btf_dedup->dedup_table` as another candidate.
2066 static int btf_dedup_prim_type(struct btf_dedup *d, __u32 type_id)
2068 struct btf_type *t = d->btf->types[type_id];
2069 struct hashmap_entry *hash_entry;
2070 struct btf_type *cand;
2071 /* if we don't find equivalent type, then we are canonical */
2072 __u32 new_id = type_id;
2076 switch (btf_kind(t)) {
2077 case BTF_KIND_CONST:
2078 case BTF_KIND_VOLATILE:
2079 case BTF_KIND_RESTRICT:
2081 case BTF_KIND_TYPEDEF:
2082 case BTF_KIND_ARRAY:
2083 case BTF_KIND_STRUCT:
2084 case BTF_KIND_UNION:
2086 case BTF_KIND_FUNC_PROTO:
2088 case BTF_KIND_DATASEC:
2092 h = btf_hash_int(t);
2093 for_each_dedup_cand(d, hash_entry, h) {
2094 cand_id = (__u32)(long)hash_entry->value;
2095 cand = d->btf->types[cand_id];
2096 if (btf_equal_int(t, cand)) {
2104 h = btf_hash_enum(t);
2105 for_each_dedup_cand(d, hash_entry, h) {
2106 cand_id = (__u32)(long)hash_entry->value;
2107 cand = d->btf->types[cand_id];
2108 if (btf_equal_enum(t, cand)) {
2112 if (d->opts.dont_resolve_fwds)
2114 if (btf_compat_enum(t, cand)) {
2115 if (btf_is_enum_fwd(t)) {
2116 /* resolve fwd to full enum */
2120 /* resolve canonical enum fwd to full enum */
2121 d->map[cand_id] = type_id;
2127 h = btf_hash_common(t);
2128 for_each_dedup_cand(d, hash_entry, h) {
2129 cand_id = (__u32)(long)hash_entry->value;
2130 cand = d->btf->types[cand_id];
2131 if (btf_equal_common(t, cand)) {
2142 d->map[type_id] = new_id;
2143 if (type_id == new_id && btf_dedup_table_add(d, h, type_id))
2149 static int btf_dedup_prim_types(struct btf_dedup *d)
2153 for (i = 1; i <= d->btf->nr_types; i++) {
2154 err = btf_dedup_prim_type(d, i);
2162 * Check whether type is already mapped into canonical one (could be to itself).
2164 static inline bool is_type_mapped(struct btf_dedup *d, uint32_t type_id)
2166 return d->map[type_id] <= BTF_MAX_NR_TYPES;
2170 * Resolve type ID into its canonical type ID, if any; otherwise return original
2171 * type ID. If type is FWD and is resolved into STRUCT/UNION already, follow
2172 * STRUCT/UNION link and resolve it into canonical type ID as well.
2174 static inline __u32 resolve_type_id(struct btf_dedup *d, __u32 type_id)
2176 while (is_type_mapped(d, type_id) && d->map[type_id] != type_id)
2177 type_id = d->map[type_id];
2182 * Resolve FWD to underlying STRUCT/UNION, if any; otherwise return original
2185 static uint32_t resolve_fwd_id(struct btf_dedup *d, uint32_t type_id)
2187 __u32 orig_type_id = type_id;
2189 if (!btf_is_fwd(d->btf->types[type_id]))
2192 while (is_type_mapped(d, type_id) && d->map[type_id] != type_id)
2193 type_id = d->map[type_id];
2195 if (!btf_is_fwd(d->btf->types[type_id]))
2198 return orig_type_id;
2202 static inline __u16 btf_fwd_kind(struct btf_type *t)
2204 return btf_kflag(t) ? BTF_KIND_UNION : BTF_KIND_STRUCT;
2208 * Check equivalence of BTF type graph formed by candidate struct/union (we'll
2209 * call it "candidate graph" in this description for brevity) to a type graph
2210 * formed by (potential) canonical struct/union ("canonical graph" for brevity
2211 * here, though keep in mind that not all types in canonical graph are
2212 * necessarily canonical representatives themselves, some of them might be
2213 * duplicates or its uniqueness might not have been established yet).
2215 * - >0, if type graphs are equivalent;
2216 * - 0, if not equivalent;
2219 * Algorithm performs side-by-side DFS traversal of both type graphs and checks
2220 * equivalence of BTF types at each step. If at any point BTF types in candidate
2221 * and canonical graphs are not compatible structurally, whole graphs are
2222 * incompatible. If types are structurally equivalent (i.e., all information
2223 * except referenced type IDs is exactly the same), a mapping from `canon_id` to
2224 * a `cand_id` is recored in hypothetical mapping (`btf_dedup->hypot_map`).
2225 * If a type references other types, then those referenced types are checked
2226 * for equivalence recursively.
2228 * During DFS traversal, if we find that for current `canon_id` type we
2229 * already have some mapping in hypothetical map, we check for two possible
2231 * - `canon_id` is mapped to exactly the same type as `cand_id`. This will
2232 * happen when type graphs have cycles. In this case we assume those two
2233 * types are equivalent.
2234 * - `canon_id` is mapped to different type. This is contradiction in our
2235 * hypothetical mapping, because same graph in canonical graph corresponds
2236 * to two different types in candidate graph, which for equivalent type
2237 * graphs shouldn't happen. This condition terminates equivalence check
2238 * with negative result.
2240 * If type graphs traversal exhausts types to check and find no contradiction,
2241 * then type graphs are equivalent.
2243 * When checking types for equivalence, there is one special case: FWD types.
2244 * If FWD type resolution is allowed and one of the types (either from canonical
2245 * or candidate graph) is FWD and other is STRUCT/UNION (depending on FWD's kind
2246 * flag) and their names match, hypothetical mapping is updated to point from
2247 * FWD to STRUCT/UNION. If graphs will be determined as equivalent successfully,
2248 * this mapping will be used to record FWD -> STRUCT/UNION mapping permanently.
2250 * Technically, this could lead to incorrect FWD to STRUCT/UNION resolution,
2251 * if there are two exactly named (or anonymous) structs/unions that are
2252 * compatible structurally, one of which has FWD field, while other is concrete
2253 * STRUCT/UNION, but according to C sources they are different structs/unions
2254 * that are referencing different types with the same name. This is extremely
2255 * unlikely to happen, but btf_dedup API allows to disable FWD resolution if
2256 * this logic is causing problems.
2258 * Doing FWD resolution means that both candidate and/or canonical graphs can
2259 * consists of portions of the graph that come from multiple compilation units.
2260 * This is due to the fact that types within single compilation unit are always
2261 * deduplicated and FWDs are already resolved, if referenced struct/union
2262 * definiton is available. So, if we had unresolved FWD and found corresponding
2263 * STRUCT/UNION, they will be from different compilation units. This
2264 * consequently means that when we "link" FWD to corresponding STRUCT/UNION,
2265 * type graph will likely have at least two different BTF types that describe
2266 * same type (e.g., most probably there will be two different BTF types for the
2267 * same 'int' primitive type) and could even have "overlapping" parts of type
2268 * graph that describe same subset of types.
2270 * This in turn means that our assumption that each type in canonical graph
2271 * must correspond to exactly one type in candidate graph might not hold
2272 * anymore and will make it harder to detect contradictions using hypothetical
2273 * map. To handle this problem, we allow to follow FWD -> STRUCT/UNION
2274 * resolution only in canonical graph. FWDs in candidate graphs are never
2275 * resolved. To see why it's OK, let's check all possible situations w.r.t. FWDs
2277 * - Both types in canonical and candidate graphs are FWDs. If they are
2278 * structurally equivalent, then they can either be both resolved to the
2279 * same STRUCT/UNION or not resolved at all. In both cases they are
2280 * equivalent and there is no need to resolve FWD on candidate side.
2281 * - Both types in canonical and candidate graphs are concrete STRUCT/UNION,
2282 * so nothing to resolve as well, algorithm will check equivalence anyway.
2283 * - Type in canonical graph is FWD, while type in candidate is concrete
2284 * STRUCT/UNION. In this case candidate graph comes from single compilation
2285 * unit, so there is exactly one BTF type for each unique C type. After
2286 * resolving FWD into STRUCT/UNION, there might be more than one BTF type
2287 * in canonical graph mapping to single BTF type in candidate graph, but
2288 * because hypothetical mapping maps from canonical to candidate types, it's
2289 * alright, and we still maintain the property of having single `canon_id`
2290 * mapping to single `cand_id` (there could be two different `canon_id`
2291 * mapped to the same `cand_id`, but it's not contradictory).
2292 * - Type in canonical graph is concrete STRUCT/UNION, while type in candidate
2293 * graph is FWD. In this case we are just going to check compatibility of
2294 * STRUCT/UNION and corresponding FWD, and if they are compatible, we'll
2295 * assume that whatever STRUCT/UNION FWD resolves to must be equivalent to
2296 * a concrete STRUCT/UNION from canonical graph. If the rest of type graphs
2297 * turn out equivalent, we'll re-resolve FWD to concrete STRUCT/UNION from
2300 static int btf_dedup_is_equiv(struct btf_dedup *d, __u32 cand_id,
2303 struct btf_type *cand_type;
2304 struct btf_type *canon_type;
2305 __u32 hypot_type_id;
2310 /* if both resolve to the same canonical, they must be equivalent */
2311 if (resolve_type_id(d, cand_id) == resolve_type_id(d, canon_id))
2314 canon_id = resolve_fwd_id(d, canon_id);
2316 hypot_type_id = d->hypot_map[canon_id];
2317 if (hypot_type_id <= BTF_MAX_NR_TYPES)
2318 return hypot_type_id == cand_id;
2320 if (btf_dedup_hypot_map_add(d, canon_id, cand_id))
2323 cand_type = d->btf->types[cand_id];
2324 canon_type = d->btf->types[canon_id];
2325 cand_kind = btf_kind(cand_type);
2326 canon_kind = btf_kind(canon_type);
2328 if (cand_type->name_off != canon_type->name_off)
2331 /* FWD <--> STRUCT/UNION equivalence check, if enabled */
2332 if (!d->opts.dont_resolve_fwds
2333 && (cand_kind == BTF_KIND_FWD || canon_kind == BTF_KIND_FWD)
2334 && cand_kind != canon_kind) {
2338 if (cand_kind == BTF_KIND_FWD) {
2339 real_kind = canon_kind;
2340 fwd_kind = btf_fwd_kind(cand_type);
2342 real_kind = cand_kind;
2343 fwd_kind = btf_fwd_kind(canon_type);
2345 return fwd_kind == real_kind;
2348 if (cand_kind != canon_kind)
2351 switch (cand_kind) {
2353 return btf_equal_int(cand_type, canon_type);
2356 if (d->opts.dont_resolve_fwds)
2357 return btf_equal_enum(cand_type, canon_type);
2359 return btf_compat_enum(cand_type, canon_type);
2362 return btf_equal_common(cand_type, canon_type);
2364 case BTF_KIND_CONST:
2365 case BTF_KIND_VOLATILE:
2366 case BTF_KIND_RESTRICT:
2368 case BTF_KIND_TYPEDEF:
2370 if (cand_type->info != canon_type->info)
2372 return btf_dedup_is_equiv(d, cand_type->type, canon_type->type);
2374 case BTF_KIND_ARRAY: {
2375 const struct btf_array *cand_arr, *canon_arr;
2377 if (!btf_compat_array(cand_type, canon_type))
2379 cand_arr = btf_array(cand_type);
2380 canon_arr = btf_array(canon_type);
2381 eq = btf_dedup_is_equiv(d,
2382 cand_arr->index_type, canon_arr->index_type);
2385 return btf_dedup_is_equiv(d, cand_arr->type, canon_arr->type);
2388 case BTF_KIND_STRUCT:
2389 case BTF_KIND_UNION: {
2390 const struct btf_member *cand_m, *canon_m;
2393 if (!btf_shallow_equal_struct(cand_type, canon_type))
2395 vlen = btf_vlen(cand_type);
2396 cand_m = btf_members(cand_type);
2397 canon_m = btf_members(canon_type);
2398 for (i = 0; i < vlen; i++) {
2399 eq = btf_dedup_is_equiv(d, cand_m->type, canon_m->type);
2409 case BTF_KIND_FUNC_PROTO: {
2410 const struct btf_param *cand_p, *canon_p;
2413 if (!btf_compat_fnproto(cand_type, canon_type))
2415 eq = btf_dedup_is_equiv(d, cand_type->type, canon_type->type);
2418 vlen = btf_vlen(cand_type);
2419 cand_p = btf_params(cand_type);
2420 canon_p = btf_params(canon_type);
2421 for (i = 0; i < vlen; i++) {
2422 eq = btf_dedup_is_equiv(d, cand_p->type, canon_p->type);
2438 * Use hypothetical mapping, produced by successful type graph equivalence
2439 * check, to augment existing struct/union canonical mapping, where possible.
2441 * If BTF_KIND_FWD resolution is allowed, this mapping is also used to record
2442 * FWD -> STRUCT/UNION correspondence as well. FWD resolution is bidirectional:
2443 * it doesn't matter if FWD type was part of canonical graph or candidate one,
2444 * we are recording the mapping anyway. As opposed to carefulness required
2445 * for struct/union correspondence mapping (described below), for FWD resolution
2446 * it's not important, as by the time that FWD type (reference type) will be
2447 * deduplicated all structs/unions will be deduped already anyway.
2449 * Recording STRUCT/UNION mapping is purely a performance optimization and is
2450 * not required for correctness. It needs to be done carefully to ensure that
2451 * struct/union from candidate's type graph is not mapped into corresponding
2452 * struct/union from canonical type graph that itself hasn't been resolved into
2453 * canonical representative. The only guarantee we have is that canonical
2454 * struct/union was determined as canonical and that won't change. But any
2455 * types referenced through that struct/union fields could have been not yet
2456 * resolved, so in case like that it's too early to establish any kind of
2457 * correspondence between structs/unions.
2459 * No canonical correspondence is derived for primitive types (they are already
2460 * deduplicated completely already anyway) or reference types (they rely on
2461 * stability of struct/union canonical relationship for equivalence checks).
2463 static void btf_dedup_merge_hypot_map(struct btf_dedup *d)
2465 __u32 cand_type_id, targ_type_id;
2466 __u16 t_kind, c_kind;
2470 for (i = 0; i < d->hypot_cnt; i++) {
2471 cand_type_id = d->hypot_list[i];
2472 targ_type_id = d->hypot_map[cand_type_id];
2473 t_id = resolve_type_id(d, targ_type_id);
2474 c_id = resolve_type_id(d, cand_type_id);
2475 t_kind = btf_kind(d->btf->types[t_id]);
2476 c_kind = btf_kind(d->btf->types[c_id]);
2478 * Resolve FWD into STRUCT/UNION.
2479 * It's ok to resolve FWD into STRUCT/UNION that's not yet
2480 * mapped to canonical representative (as opposed to
2481 * STRUCT/UNION <--> STRUCT/UNION mapping logic below), because
2482 * eventually that struct is going to be mapped and all resolved
2483 * FWDs will automatically resolve to correct canonical
2484 * representative. This will happen before ref type deduping,
2485 * which critically depends on stability of these mapping. This
2486 * stability is not a requirement for STRUCT/UNION equivalence
2489 if (t_kind != BTF_KIND_FWD && c_kind == BTF_KIND_FWD)
2490 d->map[c_id] = t_id;
2491 else if (t_kind == BTF_KIND_FWD && c_kind != BTF_KIND_FWD)
2492 d->map[t_id] = c_id;
2494 if ((t_kind == BTF_KIND_STRUCT || t_kind == BTF_KIND_UNION) &&
2495 c_kind != BTF_KIND_FWD &&
2496 is_type_mapped(d, c_id) &&
2497 !is_type_mapped(d, t_id)) {
2499 * as a perf optimization, we can map struct/union
2500 * that's part of type graph we just verified for
2501 * equivalence. We can do that for struct/union that has
2502 * canonical representative only, though.
2504 d->map[t_id] = c_id;
2510 * Deduplicate struct/union types.
2512 * For each struct/union type its type signature hash is calculated, taking
2513 * into account type's name, size, number, order and names of fields, but
2514 * ignoring type ID's referenced from fields, because they might not be deduped
2515 * completely until after reference types deduplication phase. This type hash
2516 * is used to iterate over all potential canonical types, sharing same hash.
2517 * For each canonical candidate we check whether type graphs that they form
2518 * (through referenced types in fields and so on) are equivalent using algorithm
2519 * implemented in `btf_dedup_is_equiv`. If such equivalence is found and
2520 * BTF_KIND_FWD resolution is allowed, then hypothetical mapping
2521 * (btf_dedup->hypot_map) produced by aforementioned type graph equivalence
2522 * algorithm is used to record FWD -> STRUCT/UNION mapping. It's also used to
2523 * potentially map other structs/unions to their canonical representatives,
2524 * if such relationship hasn't yet been established. This speeds up algorithm
2525 * by eliminating some of the duplicate work.
2527 * If no matching canonical representative was found, struct/union is marked
2528 * as canonical for itself and is added into btf_dedup->dedup_table hash map
2529 * for further look ups.
2531 static int btf_dedup_struct_type(struct btf_dedup *d, __u32 type_id)
2533 struct btf_type *cand_type, *t;
2534 struct hashmap_entry *hash_entry;
2535 /* if we don't find equivalent type, then we are canonical */
2536 __u32 new_id = type_id;
2540 /* already deduped or is in process of deduping (loop detected) */
2541 if (d->map[type_id] <= BTF_MAX_NR_TYPES)
2544 t = d->btf->types[type_id];
2547 if (kind != BTF_KIND_STRUCT && kind != BTF_KIND_UNION)
2550 h = btf_hash_struct(t);
2551 for_each_dedup_cand(d, hash_entry, h) {
2552 __u32 cand_id = (__u32)(long)hash_entry->value;
2556 * Even though btf_dedup_is_equiv() checks for
2557 * btf_shallow_equal_struct() internally when checking two
2558 * structs (unions) for equivalence, we need to guard here
2559 * from picking matching FWD type as a dedup candidate.
2560 * This can happen due to hash collision. In such case just
2561 * relying on btf_dedup_is_equiv() would lead to potentially
2562 * creating a loop (FWD -> STRUCT and STRUCT -> FWD), because
2563 * FWD and compatible STRUCT/UNION are considered equivalent.
2565 cand_type = d->btf->types[cand_id];
2566 if (!btf_shallow_equal_struct(t, cand_type))
2569 btf_dedup_clear_hypot_map(d);
2570 eq = btf_dedup_is_equiv(d, type_id, cand_id);
2576 btf_dedup_merge_hypot_map(d);
2580 d->map[type_id] = new_id;
2581 if (type_id == new_id && btf_dedup_table_add(d, h, type_id))
2587 static int btf_dedup_struct_types(struct btf_dedup *d)
2591 for (i = 1; i <= d->btf->nr_types; i++) {
2592 err = btf_dedup_struct_type(d, i);
2600 * Deduplicate reference type.
2602 * Once all primitive and struct/union types got deduplicated, we can easily
2603 * deduplicate all other (reference) BTF types. This is done in two steps:
2605 * 1. Resolve all referenced type IDs into their canonical type IDs. This
2606 * resolution can be done either immediately for primitive or struct/union types
2607 * (because they were deduped in previous two phases) or recursively for
2608 * reference types. Recursion will always terminate at either primitive or
2609 * struct/union type, at which point we can "unwind" chain of reference types
2610 * one by one. There is no danger of encountering cycles because in C type
2611 * system the only way to form type cycle is through struct/union, so any chain
2612 * of reference types, even those taking part in a type cycle, will inevitably
2613 * reach struct/union at some point.
2615 * 2. Once all referenced type IDs are resolved into canonical ones, BTF type
2616 * becomes "stable", in the sense that no further deduplication will cause
2617 * any changes to it. With that, it's now possible to calculate type's signature
2618 * hash (this time taking into account referenced type IDs) and loop over all
2619 * potential canonical representatives. If no match was found, current type
2620 * will become canonical representative of itself and will be added into
2621 * btf_dedup->dedup_table as another possible canonical representative.
2623 static int btf_dedup_ref_type(struct btf_dedup *d, __u32 type_id)
2625 struct hashmap_entry *hash_entry;
2626 __u32 new_id = type_id, cand_id;
2627 struct btf_type *t, *cand;
2628 /* if we don't find equivalent type, then we are representative type */
2632 if (d->map[type_id] == BTF_IN_PROGRESS_ID)
2634 if (d->map[type_id] <= BTF_MAX_NR_TYPES)
2635 return resolve_type_id(d, type_id);
2637 t = d->btf->types[type_id];
2638 d->map[type_id] = BTF_IN_PROGRESS_ID;
2640 switch (btf_kind(t)) {
2641 case BTF_KIND_CONST:
2642 case BTF_KIND_VOLATILE:
2643 case BTF_KIND_RESTRICT:
2645 case BTF_KIND_TYPEDEF:
2647 ref_type_id = btf_dedup_ref_type(d, t->type);
2648 if (ref_type_id < 0)
2650 t->type = ref_type_id;
2652 h = btf_hash_common(t);
2653 for_each_dedup_cand(d, hash_entry, h) {
2654 cand_id = (__u32)(long)hash_entry->value;
2655 cand = d->btf->types[cand_id];
2656 if (btf_equal_common(t, cand)) {
2663 case BTF_KIND_ARRAY: {
2664 struct btf_array *info = btf_array(t);
2666 ref_type_id = btf_dedup_ref_type(d, info->type);
2667 if (ref_type_id < 0)
2669 info->type = ref_type_id;
2671 ref_type_id = btf_dedup_ref_type(d, info->index_type);
2672 if (ref_type_id < 0)
2674 info->index_type = ref_type_id;
2676 h = btf_hash_array(t);
2677 for_each_dedup_cand(d, hash_entry, h) {
2678 cand_id = (__u32)(long)hash_entry->value;
2679 cand = d->btf->types[cand_id];
2680 if (btf_equal_array(t, cand)) {
2688 case BTF_KIND_FUNC_PROTO: {
2689 struct btf_param *param;
2693 ref_type_id = btf_dedup_ref_type(d, t->type);
2694 if (ref_type_id < 0)
2696 t->type = ref_type_id;
2699 param = btf_params(t);
2700 for (i = 0; i < vlen; i++) {
2701 ref_type_id = btf_dedup_ref_type(d, param->type);
2702 if (ref_type_id < 0)
2704 param->type = ref_type_id;
2708 h = btf_hash_fnproto(t);
2709 for_each_dedup_cand(d, hash_entry, h) {
2710 cand_id = (__u32)(long)hash_entry->value;
2711 cand = d->btf->types[cand_id];
2712 if (btf_equal_fnproto(t, cand)) {
2724 d->map[type_id] = new_id;
2725 if (type_id == new_id && btf_dedup_table_add(d, h, type_id))
2731 static int btf_dedup_ref_types(struct btf_dedup *d)
2735 for (i = 1; i <= d->btf->nr_types; i++) {
2736 err = btf_dedup_ref_type(d, i);
2740 /* we won't need d->dedup_table anymore */
2741 hashmap__free(d->dedup_table);
2742 d->dedup_table = NULL;
2749 * After we established for each type its corresponding canonical representative
2750 * type, we now can eliminate types that are not canonical and leave only
2751 * canonical ones layed out sequentially in memory by copying them over
2752 * duplicates. During compaction btf_dedup->hypot_map array is reused to store
2753 * a map from original type ID to a new compacted type ID, which will be used
2754 * during next phase to "fix up" type IDs, referenced from struct/union and
2757 static int btf_dedup_compact_types(struct btf_dedup *d)
2759 struct btf_type **new_types;
2760 __u32 next_type_id = 1;
2761 char *types_start, *p;
2764 /* we are going to reuse hypot_map to store compaction remapping */
2765 d->hypot_map[0] = 0;
2766 for (i = 1; i <= d->btf->nr_types; i++)
2767 d->hypot_map[i] = BTF_UNPROCESSED_ID;
2769 types_start = d->btf->nohdr_data + d->btf->hdr->type_off;
2772 for (i = 1; i <= d->btf->nr_types; i++) {
2776 len = btf_type_size(d->btf->types[i]);
2780 memmove(p, d->btf->types[i], len);
2781 d->hypot_map[i] = next_type_id;
2782 d->btf->types[next_type_id] = (struct btf_type *)p;
2787 /* shrink struct btf's internal types index and update btf_header */
2788 d->btf->nr_types = next_type_id - 1;
2789 d->btf->types_size = d->btf->nr_types;
2790 d->btf->hdr->type_len = p - types_start;
2791 new_types = realloc(d->btf->types,
2792 (1 + d->btf->nr_types) * sizeof(struct btf_type *));
2795 d->btf->types = new_types;
2797 /* make sure string section follows type information without gaps */
2798 d->btf->hdr->str_off = p - (char *)d->btf->nohdr_data;
2799 memmove(p, d->btf->strings, d->btf->hdr->str_len);
2800 d->btf->strings = p;
2801 p += d->btf->hdr->str_len;
2803 d->btf->data_size = p - (char *)d->btf->data;
2808 * Figure out final (deduplicated and compacted) type ID for provided original
2809 * `type_id` by first resolving it into corresponding canonical type ID and
2810 * then mapping it to a deduplicated type ID, stored in btf_dedup->hypot_map,
2811 * which is populated during compaction phase.
2813 static int btf_dedup_remap_type_id(struct btf_dedup *d, __u32 type_id)
2815 __u32 resolved_type_id, new_type_id;
2817 resolved_type_id = resolve_type_id(d, type_id);
2818 new_type_id = d->hypot_map[resolved_type_id];
2819 if (new_type_id > BTF_MAX_NR_TYPES)
2825 * Remap referenced type IDs into deduped type IDs.
2827 * After BTF types are deduplicated and compacted, their final type IDs may
2828 * differ from original ones. The map from original to a corresponding
2829 * deduped type ID is stored in btf_dedup->hypot_map and is populated during
2830 * compaction phase. During remapping phase we are rewriting all type IDs
2831 * referenced from any BTF type (e.g., struct fields, func proto args, etc) to
2832 * their final deduped type IDs.
2834 static int btf_dedup_remap_type(struct btf_dedup *d, __u32 type_id)
2836 struct btf_type *t = d->btf->types[type_id];
2839 switch (btf_kind(t)) {
2845 case BTF_KIND_CONST:
2846 case BTF_KIND_VOLATILE:
2847 case BTF_KIND_RESTRICT:
2849 case BTF_KIND_TYPEDEF:
2852 r = btf_dedup_remap_type_id(d, t->type);
2858 case BTF_KIND_ARRAY: {
2859 struct btf_array *arr_info = btf_array(t);
2861 r = btf_dedup_remap_type_id(d, arr_info->type);
2865 r = btf_dedup_remap_type_id(d, arr_info->index_type);
2868 arr_info->index_type = r;
2872 case BTF_KIND_STRUCT:
2873 case BTF_KIND_UNION: {
2874 struct btf_member *member = btf_members(t);
2875 __u16 vlen = btf_vlen(t);
2877 for (i = 0; i < vlen; i++) {
2878 r = btf_dedup_remap_type_id(d, member->type);
2887 case BTF_KIND_FUNC_PROTO: {
2888 struct btf_param *param = btf_params(t);
2889 __u16 vlen = btf_vlen(t);
2891 r = btf_dedup_remap_type_id(d, t->type);
2896 for (i = 0; i < vlen; i++) {
2897 r = btf_dedup_remap_type_id(d, param->type);
2906 case BTF_KIND_DATASEC: {
2907 struct btf_var_secinfo *var = btf_var_secinfos(t);
2908 __u16 vlen = btf_vlen(t);
2910 for (i = 0; i < vlen; i++) {
2911 r = btf_dedup_remap_type_id(d, var->type);
2927 static int btf_dedup_remap_types(struct btf_dedup *d)
2931 for (i = 1; i <= d->btf->nr_types; i++) {
2932 r = btf_dedup_remap_type(d, i);
2939 static struct btf *btf_load_raw(const char *path)
2947 if (stat(path, &st))
2948 return ERR_PTR(-errno);
2950 data = malloc(st.st_size);
2952 return ERR_PTR(-ENOMEM);
2954 f = fopen(path, "rb");
2956 btf = ERR_PTR(-errno);
2960 read_cnt = fread(data, 1, st.st_size, f);
2962 if (read_cnt < st.st_size) {
2963 btf = ERR_PTR(-EBADF);
2967 btf = btf__new(data, read_cnt);
2975 * Probe few well-known locations for vmlinux kernel image and try to load BTF
2976 * data out of it to use for target BTF.
2978 struct btf *libbpf_find_kernel_btf(void)
2981 const char *path_fmt;
2984 /* try canonical vmlinux BTF through sysfs first */
2985 { "/sys/kernel/btf/vmlinux", true /* raw BTF */ },
2986 /* fall back to trying to find vmlinux ELF on disk otherwise */
2987 { "/boot/vmlinux-%1$s" },
2988 { "/lib/modules/%1$s/vmlinux-%1$s" },
2989 { "/lib/modules/%1$s/build/vmlinux" },
2990 { "/usr/lib/modules/%1$s/kernel/vmlinux" },
2991 { "/usr/lib/debug/boot/vmlinux-%1$s" },
2992 { "/usr/lib/debug/boot/vmlinux-%1$s.debug" },
2993 { "/usr/lib/debug/lib/modules/%1$s/vmlinux" },
2995 char path[PATH_MAX + 1];
3002 for (i = 0; i < ARRAY_SIZE(locations); i++) {
3003 snprintf(path, PATH_MAX, locations[i].path_fmt, buf.release);
3005 if (access(path, R_OK))
3008 if (locations[i].raw_btf)
3009 btf = btf_load_raw(path);
3011 btf = btf__parse_elf(path, NULL);
3013 pr_debug("loading kernel BTF '%s': %ld\n",
3014 path, IS_ERR(btf) ? PTR_ERR(btf) : 0);
3021 pr_warn("failed to find valid kernel BTF\n");
3022 return ERR_PTR(-ESRCH);