Loading...
1// SPDX-License-Identifier: (LGPL-2.1 OR BSD-2-Clause)
2/* Copyright (c) 2018 Facebook */
3
4#include <byteswap.h>
5#include <endian.h>
6#include <stdio.h>
7#include <stdlib.h>
8#include <string.h>
9#include <fcntl.h>
10#include <unistd.h>
11#include <errno.h>
12#include <sys/utsname.h>
13#include <sys/param.h>
14#include <sys/stat.h>
15#include <linux/kernel.h>
16#include <linux/err.h>
17#include <linux/btf.h>
18#include <gelf.h>
19#include "btf.h"
20#include "bpf.h"
21#include "libbpf.h"
22#include "libbpf_internal.h"
23#include "hashmap.h"
24#include "strset.h"
25
26#define BTF_MAX_NR_TYPES 0x7fffffffU
27#define BTF_MAX_STR_OFFSET 0x7fffffffU
28
29static struct btf_type btf_void;
30
31struct btf {
32 /* raw BTF data in native endianness */
33 void *raw_data;
34 /* raw BTF data in non-native endianness */
35 void *raw_data_swapped;
36 __u32 raw_size;
37 /* whether target endianness differs from the native one */
38 bool swapped_endian;
39
40 /*
41 * When BTF is loaded from an ELF or raw memory it is stored
42 * in a contiguous memory block. The hdr, type_data, and, strs_data
43 * point inside that memory region to their respective parts of BTF
44 * representation:
45 *
46 * +--------------------------------+
47 * | Header | Types | Strings |
48 * +--------------------------------+
49 * ^ ^ ^
50 * | | |
51 * hdr | |
52 * types_data-+ |
53 * strs_data------------+
54 *
55 * If BTF data is later modified, e.g., due to types added or
56 * removed, BTF deduplication performed, etc, this contiguous
57 * representation is broken up into three independently allocated
58 * memory regions to be able to modify them independently.
59 * raw_data is nulled out at that point, but can be later allocated
60 * and cached again if user calls btf__raw_data(), at which point
61 * raw_data will contain a contiguous copy of header, types, and
62 * strings:
63 *
64 * +----------+ +---------+ +-----------+
65 * | Header | | Types | | Strings |
66 * +----------+ +---------+ +-----------+
67 * ^ ^ ^
68 * | | |
69 * hdr | |
70 * types_data----+ |
71 * strset__data(strs_set)-----+
72 *
73 * +----------+---------+-----------+
74 * | Header | Types | Strings |
75 * raw_data----->+----------+---------+-----------+
76 */
77 struct btf_header *hdr;
78
79 void *types_data;
80 size_t types_data_cap; /* used size stored in hdr->type_len */
81
82 /* type ID to `struct btf_type *` lookup index
83 * type_offs[0] corresponds to the first non-VOID type:
84 * - for base BTF it's type [1];
85 * - for split BTF it's the first non-base BTF type.
86 */
87 __u32 *type_offs;
88 size_t type_offs_cap;
89 /* number of types in this BTF instance:
90 * - doesn't include special [0] void type;
91 * - for split BTF counts number of types added on top of base BTF.
92 */
93 __u32 nr_types;
94 /* if not NULL, points to the base BTF on top of which the current
95 * split BTF is based
96 */
97 struct btf *base_btf;
98 /* BTF type ID of the first type in this BTF instance:
99 * - for base BTF it's equal to 1;
100 * - for split BTF it's equal to biggest type ID of base BTF plus 1.
101 */
102 int start_id;
103 /* logical string offset of this BTF instance:
104 * - for base BTF it's equal to 0;
105 * - for split BTF it's equal to total size of base BTF's string section size.
106 */
107 int start_str_off;
108
109 /* only one of strs_data or strs_set can be non-NULL, depending on
110 * whether BTF is in a modifiable state (strs_set is used) or not
111 * (strs_data points inside raw_data)
112 */
113 void *strs_data;
114 /* a set of unique strings */
115 struct strset *strs_set;
116 /* whether strings are already deduplicated */
117 bool strs_deduped;
118
119 /* BTF object FD, if loaded into kernel */
120 int fd;
121
122 /* Pointer size (in bytes) for a target architecture of this BTF */
123 int ptr_sz;
124};
125
126static inline __u64 ptr_to_u64(const void *ptr)
127{
128 return (__u64) (unsigned long) ptr;
129}
130
131/* Ensure given dynamically allocated memory region pointed to by *data* with
132 * capacity of *cap_cnt* elements each taking *elem_sz* bytes has enough
133 * memory to accommodate *add_cnt* new elements, assuming *cur_cnt* elements
134 * are already used. At most *max_cnt* elements can be ever allocated.
135 * If necessary, memory is reallocated and all existing data is copied over,
136 * new pointer to the memory region is stored at *data, new memory region
137 * capacity (in number of elements) is stored in *cap.
138 * On success, memory pointer to the beginning of unused memory is returned.
139 * On error, NULL is returned.
140 */
141void *libbpf_add_mem(void **data, size_t *cap_cnt, size_t elem_sz,
142 size_t cur_cnt, size_t max_cnt, size_t add_cnt)
143{
144 size_t new_cnt;
145 void *new_data;
146
147 if (cur_cnt + add_cnt <= *cap_cnt)
148 return *data + cur_cnt * elem_sz;
149
150 /* requested more than the set limit */
151 if (cur_cnt + add_cnt > max_cnt)
152 return NULL;
153
154 new_cnt = *cap_cnt;
155 new_cnt += new_cnt / 4; /* expand by 25% */
156 if (new_cnt < 16) /* but at least 16 elements */
157 new_cnt = 16;
158 if (new_cnt > max_cnt) /* but not exceeding a set limit */
159 new_cnt = max_cnt;
160 if (new_cnt < cur_cnt + add_cnt) /* also ensure we have enough memory */
161 new_cnt = cur_cnt + add_cnt;
162
163 new_data = libbpf_reallocarray(*data, new_cnt, elem_sz);
164 if (!new_data)
165 return NULL;
166
167 /* zero out newly allocated portion of memory */
168 memset(new_data + (*cap_cnt) * elem_sz, 0, (new_cnt - *cap_cnt) * elem_sz);
169
170 *data = new_data;
171 *cap_cnt = new_cnt;
172 return new_data + cur_cnt * elem_sz;
173}
174
175/* Ensure given dynamically allocated memory region has enough allocated space
176 * to accommodate *need_cnt* elements of size *elem_sz* bytes each
177 */
178int libbpf_ensure_mem(void **data, size_t *cap_cnt, size_t elem_sz, size_t need_cnt)
179{
180 void *p;
181
182 if (need_cnt <= *cap_cnt)
183 return 0;
184
185 p = libbpf_add_mem(data, cap_cnt, elem_sz, *cap_cnt, SIZE_MAX, need_cnt - *cap_cnt);
186 if (!p)
187 return -ENOMEM;
188
189 return 0;
190}
191
192static void *btf_add_type_offs_mem(struct btf *btf, size_t add_cnt)
193{
194 return libbpf_add_mem((void **)&btf->type_offs, &btf->type_offs_cap, sizeof(__u32),
195 btf->nr_types, BTF_MAX_NR_TYPES, add_cnt);
196}
197
198static int btf_add_type_idx_entry(struct btf *btf, __u32 type_off)
199{
200 __u32 *p;
201
202 p = btf_add_type_offs_mem(btf, 1);
203 if (!p)
204 return -ENOMEM;
205
206 *p = type_off;
207 return 0;
208}
209
210static void btf_bswap_hdr(struct btf_header *h)
211{
212 h->magic = bswap_16(h->magic);
213 h->hdr_len = bswap_32(h->hdr_len);
214 h->type_off = bswap_32(h->type_off);
215 h->type_len = bswap_32(h->type_len);
216 h->str_off = bswap_32(h->str_off);
217 h->str_len = bswap_32(h->str_len);
218}
219
220static int btf_parse_hdr(struct btf *btf)
221{
222 struct btf_header *hdr = btf->hdr;
223 __u32 meta_left;
224
225 if (btf->raw_size < sizeof(struct btf_header)) {
226 pr_debug("BTF header not found\n");
227 return -EINVAL;
228 }
229
230 if (hdr->magic == bswap_16(BTF_MAGIC)) {
231 btf->swapped_endian = true;
232 if (bswap_32(hdr->hdr_len) != sizeof(struct btf_header)) {
233 pr_warn("Can't load BTF with non-native endianness due to unsupported header length %u\n",
234 bswap_32(hdr->hdr_len));
235 return -ENOTSUP;
236 }
237 btf_bswap_hdr(hdr);
238 } else if (hdr->magic != BTF_MAGIC) {
239 pr_debug("Invalid BTF magic: %x\n", hdr->magic);
240 return -EINVAL;
241 }
242
243 if (btf->raw_size < hdr->hdr_len) {
244 pr_debug("BTF header len %u larger than data size %u\n",
245 hdr->hdr_len, btf->raw_size);
246 return -EINVAL;
247 }
248
249 meta_left = btf->raw_size - hdr->hdr_len;
250 if (meta_left < (long long)hdr->str_off + hdr->str_len) {
251 pr_debug("Invalid BTF total size: %u\n", btf->raw_size);
252 return -EINVAL;
253 }
254
255 if ((long long)hdr->type_off + hdr->type_len > hdr->str_off) {
256 pr_debug("Invalid BTF data sections layout: type data at %u + %u, strings data at %u + %u\n",
257 hdr->type_off, hdr->type_len, hdr->str_off, hdr->str_len);
258 return -EINVAL;
259 }
260
261 if (hdr->type_off % 4) {
262 pr_debug("BTF type section is not aligned to 4 bytes\n");
263 return -EINVAL;
264 }
265
266 return 0;
267}
268
269static int btf_parse_str_sec(struct btf *btf)
270{
271 const struct btf_header *hdr = btf->hdr;
272 const char *start = btf->strs_data;
273 const char *end = start + btf->hdr->str_len;
274
275 if (btf->base_btf && hdr->str_len == 0)
276 return 0;
277 if (!hdr->str_len || hdr->str_len - 1 > BTF_MAX_STR_OFFSET || end[-1]) {
278 pr_debug("Invalid BTF string section\n");
279 return -EINVAL;
280 }
281 if (!btf->base_btf && start[0]) {
282 pr_debug("Invalid BTF string section\n");
283 return -EINVAL;
284 }
285 return 0;
286}
287
288static int btf_type_size(const struct btf_type *t)
289{
290 const int base_size = sizeof(struct btf_type);
291 __u16 vlen = btf_vlen(t);
292
293 switch (btf_kind(t)) {
294 case BTF_KIND_FWD:
295 case BTF_KIND_CONST:
296 case BTF_KIND_VOLATILE:
297 case BTF_KIND_RESTRICT:
298 case BTF_KIND_PTR:
299 case BTF_KIND_TYPEDEF:
300 case BTF_KIND_FUNC:
301 case BTF_KIND_FLOAT:
302 case BTF_KIND_TYPE_TAG:
303 return base_size;
304 case BTF_KIND_INT:
305 return base_size + sizeof(__u32);
306 case BTF_KIND_ENUM:
307 return base_size + vlen * sizeof(struct btf_enum);
308 case BTF_KIND_ENUM64:
309 return base_size + vlen * sizeof(struct btf_enum64);
310 case BTF_KIND_ARRAY:
311 return base_size + sizeof(struct btf_array);
312 case BTF_KIND_STRUCT:
313 case BTF_KIND_UNION:
314 return base_size + vlen * sizeof(struct btf_member);
315 case BTF_KIND_FUNC_PROTO:
316 return base_size + vlen * sizeof(struct btf_param);
317 case BTF_KIND_VAR:
318 return base_size + sizeof(struct btf_var);
319 case BTF_KIND_DATASEC:
320 return base_size + vlen * sizeof(struct btf_var_secinfo);
321 case BTF_KIND_DECL_TAG:
322 return base_size + sizeof(struct btf_decl_tag);
323 default:
324 pr_debug("Unsupported BTF_KIND:%u\n", btf_kind(t));
325 return -EINVAL;
326 }
327}
328
329static void btf_bswap_type_base(struct btf_type *t)
330{
331 t->name_off = bswap_32(t->name_off);
332 t->info = bswap_32(t->info);
333 t->type = bswap_32(t->type);
334}
335
336static int btf_bswap_type_rest(struct btf_type *t)
337{
338 struct btf_var_secinfo *v;
339 struct btf_enum64 *e64;
340 struct btf_member *m;
341 struct btf_array *a;
342 struct btf_param *p;
343 struct btf_enum *e;
344 __u16 vlen = btf_vlen(t);
345 int i;
346
347 switch (btf_kind(t)) {
348 case BTF_KIND_FWD:
349 case BTF_KIND_CONST:
350 case BTF_KIND_VOLATILE:
351 case BTF_KIND_RESTRICT:
352 case BTF_KIND_PTR:
353 case BTF_KIND_TYPEDEF:
354 case BTF_KIND_FUNC:
355 case BTF_KIND_FLOAT:
356 case BTF_KIND_TYPE_TAG:
357 return 0;
358 case BTF_KIND_INT:
359 *(__u32 *)(t + 1) = bswap_32(*(__u32 *)(t + 1));
360 return 0;
361 case BTF_KIND_ENUM:
362 for (i = 0, e = btf_enum(t); i < vlen; i++, e++) {
363 e->name_off = bswap_32(e->name_off);
364 e->val = bswap_32(e->val);
365 }
366 return 0;
367 case BTF_KIND_ENUM64:
368 for (i = 0, e64 = btf_enum64(t); i < vlen; i++, e64++) {
369 e64->name_off = bswap_32(e64->name_off);
370 e64->val_lo32 = bswap_32(e64->val_lo32);
371 e64->val_hi32 = bswap_32(e64->val_hi32);
372 }
373 return 0;
374 case BTF_KIND_ARRAY:
375 a = btf_array(t);
376 a->type = bswap_32(a->type);
377 a->index_type = bswap_32(a->index_type);
378 a->nelems = bswap_32(a->nelems);
379 return 0;
380 case BTF_KIND_STRUCT:
381 case BTF_KIND_UNION:
382 for (i = 0, m = btf_members(t); i < vlen; i++, m++) {
383 m->name_off = bswap_32(m->name_off);
384 m->type = bswap_32(m->type);
385 m->offset = bswap_32(m->offset);
386 }
387 return 0;
388 case BTF_KIND_FUNC_PROTO:
389 for (i = 0, p = btf_params(t); i < vlen; i++, p++) {
390 p->name_off = bswap_32(p->name_off);
391 p->type = bswap_32(p->type);
392 }
393 return 0;
394 case BTF_KIND_VAR:
395 btf_var(t)->linkage = bswap_32(btf_var(t)->linkage);
396 return 0;
397 case BTF_KIND_DATASEC:
398 for (i = 0, v = btf_var_secinfos(t); i < vlen; i++, v++) {
399 v->type = bswap_32(v->type);
400 v->offset = bswap_32(v->offset);
401 v->size = bswap_32(v->size);
402 }
403 return 0;
404 case BTF_KIND_DECL_TAG:
405 btf_decl_tag(t)->component_idx = bswap_32(btf_decl_tag(t)->component_idx);
406 return 0;
407 default:
408 pr_debug("Unsupported BTF_KIND:%u\n", btf_kind(t));
409 return -EINVAL;
410 }
411}
412
413static int btf_parse_type_sec(struct btf *btf)
414{
415 struct btf_header *hdr = btf->hdr;
416 void *next_type = btf->types_data;
417 void *end_type = next_type + hdr->type_len;
418 int err, type_size;
419
420 while (next_type + sizeof(struct btf_type) <= end_type) {
421 if (btf->swapped_endian)
422 btf_bswap_type_base(next_type);
423
424 type_size = btf_type_size(next_type);
425 if (type_size < 0)
426 return type_size;
427 if (next_type + type_size > end_type) {
428 pr_warn("BTF type [%d] is malformed\n", btf->start_id + btf->nr_types);
429 return -EINVAL;
430 }
431
432 if (btf->swapped_endian && btf_bswap_type_rest(next_type))
433 return -EINVAL;
434
435 err = btf_add_type_idx_entry(btf, next_type - btf->types_data);
436 if (err)
437 return err;
438
439 next_type += type_size;
440 btf->nr_types++;
441 }
442
443 if (next_type != end_type) {
444 pr_warn("BTF types data is malformed\n");
445 return -EINVAL;
446 }
447
448 return 0;
449}
450
451static int btf_validate_str(const struct btf *btf, __u32 str_off, const char *what, __u32 type_id)
452{
453 const char *s;
454
455 s = btf__str_by_offset(btf, str_off);
456 if (!s) {
457 pr_warn("btf: type [%u]: invalid %s (string offset %u)\n", type_id, what, str_off);
458 return -EINVAL;
459 }
460
461 return 0;
462}
463
464static int btf_validate_id(const struct btf *btf, __u32 id, __u32 ctx_id)
465{
466 const struct btf_type *t;
467
468 t = btf__type_by_id(btf, id);
469 if (!t) {
470 pr_warn("btf: type [%u]: invalid referenced type ID %u\n", ctx_id, id);
471 return -EINVAL;
472 }
473
474 return 0;
475}
476
477static int btf_validate_type(const struct btf *btf, const struct btf_type *t, __u32 id)
478{
479 __u32 kind = btf_kind(t);
480 int err, i, n;
481
482 err = btf_validate_str(btf, t->name_off, "type name", id);
483 if (err)
484 return err;
485
486 switch (kind) {
487 case BTF_KIND_UNKN:
488 case BTF_KIND_INT:
489 case BTF_KIND_FWD:
490 case BTF_KIND_FLOAT:
491 break;
492 case BTF_KIND_PTR:
493 case BTF_KIND_TYPEDEF:
494 case BTF_KIND_VOLATILE:
495 case BTF_KIND_CONST:
496 case BTF_KIND_RESTRICT:
497 case BTF_KIND_VAR:
498 case BTF_KIND_DECL_TAG:
499 case BTF_KIND_TYPE_TAG:
500 err = btf_validate_id(btf, t->type, id);
501 if (err)
502 return err;
503 break;
504 case BTF_KIND_ARRAY: {
505 const struct btf_array *a = btf_array(t);
506
507 err = btf_validate_id(btf, a->type, id);
508 err = err ?: btf_validate_id(btf, a->index_type, id);
509 if (err)
510 return err;
511 break;
512 }
513 case BTF_KIND_STRUCT:
514 case BTF_KIND_UNION: {
515 const struct btf_member *m = btf_members(t);
516
517 n = btf_vlen(t);
518 for (i = 0; i < n; i++, m++) {
519 err = btf_validate_str(btf, m->name_off, "field name", id);
520 err = err ?: btf_validate_id(btf, m->type, id);
521 if (err)
522 return err;
523 }
524 break;
525 }
526 case BTF_KIND_ENUM: {
527 const struct btf_enum *m = btf_enum(t);
528
529 n = btf_vlen(t);
530 for (i = 0; i < n; i++, m++) {
531 err = btf_validate_str(btf, m->name_off, "enum name", id);
532 if (err)
533 return err;
534 }
535 break;
536 }
537 case BTF_KIND_ENUM64: {
538 const struct btf_enum64 *m = btf_enum64(t);
539
540 n = btf_vlen(t);
541 for (i = 0; i < n; i++, m++) {
542 err = btf_validate_str(btf, m->name_off, "enum name", id);
543 if (err)
544 return err;
545 }
546 break;
547 }
548 case BTF_KIND_FUNC: {
549 const struct btf_type *ft;
550
551 err = btf_validate_id(btf, t->type, id);
552 if (err)
553 return err;
554 ft = btf__type_by_id(btf, t->type);
555 if (btf_kind(ft) != BTF_KIND_FUNC_PROTO) {
556 pr_warn("btf: type [%u]: referenced type [%u] is not FUNC_PROTO\n", id, t->type);
557 return -EINVAL;
558 }
559 break;
560 }
561 case BTF_KIND_FUNC_PROTO: {
562 const struct btf_param *m = btf_params(t);
563
564 n = btf_vlen(t);
565 for (i = 0; i < n; i++, m++) {
566 err = btf_validate_str(btf, m->name_off, "param name", id);
567 err = err ?: btf_validate_id(btf, m->type, id);
568 if (err)
569 return err;
570 }
571 break;
572 }
573 case BTF_KIND_DATASEC: {
574 const struct btf_var_secinfo *m = btf_var_secinfos(t);
575
576 n = btf_vlen(t);
577 for (i = 0; i < n; i++, m++) {
578 err = btf_validate_id(btf, m->type, id);
579 if (err)
580 return err;
581 }
582 break;
583 }
584 default:
585 pr_warn("btf: type [%u]: unrecognized kind %u\n", id, kind);
586 return -EINVAL;
587 }
588 return 0;
589}
590
591/* Validate basic sanity of BTF. It's intentionally less thorough than
592 * kernel's validation and validates only properties of BTF that libbpf relies
593 * on to be correct (e.g., valid type IDs, valid string offsets, etc)
594 */
595static int btf_sanity_check(const struct btf *btf)
596{
597 const struct btf_type *t;
598 __u32 i, n = btf__type_cnt(btf);
599 int err;
600
601 for (i = 1; i < n; i++) {
602 t = btf_type_by_id(btf, i);
603 err = btf_validate_type(btf, t, i);
604 if (err)
605 return err;
606 }
607 return 0;
608}
609
610__u32 btf__type_cnt(const struct btf *btf)
611{
612 return btf->start_id + btf->nr_types;
613}
614
615const struct btf *btf__base_btf(const struct btf *btf)
616{
617 return btf->base_btf;
618}
619
620/* internal helper returning non-const pointer to a type */
621struct btf_type *btf_type_by_id(const struct btf *btf, __u32 type_id)
622{
623 if (type_id == 0)
624 return &btf_void;
625 if (type_id < btf->start_id)
626 return btf_type_by_id(btf->base_btf, type_id);
627 return btf->types_data + btf->type_offs[type_id - btf->start_id];
628}
629
630const struct btf_type *btf__type_by_id(const struct btf *btf, __u32 type_id)
631{
632 if (type_id >= btf->start_id + btf->nr_types)
633 return errno = EINVAL, NULL;
634 return btf_type_by_id((struct btf *)btf, type_id);
635}
636
637static int determine_ptr_size(const struct btf *btf)
638{
639 static const char * const long_aliases[] = {
640 "long",
641 "long int",
642 "int long",
643 "unsigned long",
644 "long unsigned",
645 "unsigned long int",
646 "unsigned int long",
647 "long unsigned int",
648 "long int unsigned",
649 "int unsigned long",
650 "int long unsigned",
651 };
652 const struct btf_type *t;
653 const char *name;
654 int i, j, n;
655
656 if (btf->base_btf && btf->base_btf->ptr_sz > 0)
657 return btf->base_btf->ptr_sz;
658
659 n = btf__type_cnt(btf);
660 for (i = 1; i < n; i++) {
661 t = btf__type_by_id(btf, i);
662 if (!btf_is_int(t))
663 continue;
664
665 if (t->size != 4 && t->size != 8)
666 continue;
667
668 name = btf__name_by_offset(btf, t->name_off);
669 if (!name)
670 continue;
671
672 for (j = 0; j < ARRAY_SIZE(long_aliases); j++) {
673 if (strcmp(name, long_aliases[j]) == 0)
674 return t->size;
675 }
676 }
677
678 return -1;
679}
680
681static size_t btf_ptr_sz(const struct btf *btf)
682{
683 if (!btf->ptr_sz)
684 ((struct btf *)btf)->ptr_sz = determine_ptr_size(btf);
685 return btf->ptr_sz < 0 ? sizeof(void *) : btf->ptr_sz;
686}
687
688/* Return pointer size this BTF instance assumes. The size is heuristically
689 * determined by looking for 'long' or 'unsigned long' integer type and
690 * recording its size in bytes. If BTF type information doesn't have any such
691 * type, this function returns 0. In the latter case, native architecture's
692 * pointer size is assumed, so will be either 4 or 8, depending on
693 * architecture that libbpf was compiled for. It's possible to override
694 * guessed value by using btf__set_pointer_size() API.
695 */
696size_t btf__pointer_size(const struct btf *btf)
697{
698 if (!btf->ptr_sz)
699 ((struct btf *)btf)->ptr_sz = determine_ptr_size(btf);
700
701 if (btf->ptr_sz < 0)
702 /* not enough BTF type info to guess */
703 return 0;
704
705 return btf->ptr_sz;
706}
707
708/* Override or set pointer size in bytes. Only values of 4 and 8 are
709 * supported.
710 */
711int btf__set_pointer_size(struct btf *btf, size_t ptr_sz)
712{
713 if (ptr_sz != 4 && ptr_sz != 8)
714 return libbpf_err(-EINVAL);
715 btf->ptr_sz = ptr_sz;
716 return 0;
717}
718
719static bool is_host_big_endian(void)
720{
721#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
722 return false;
723#elif __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
724 return true;
725#else
726# error "Unrecognized __BYTE_ORDER__"
727#endif
728}
729
730enum btf_endianness btf__endianness(const struct btf *btf)
731{
732 if (is_host_big_endian())
733 return btf->swapped_endian ? BTF_LITTLE_ENDIAN : BTF_BIG_ENDIAN;
734 else
735 return btf->swapped_endian ? BTF_BIG_ENDIAN : BTF_LITTLE_ENDIAN;
736}
737
738int btf__set_endianness(struct btf *btf, enum btf_endianness endian)
739{
740 if (endian != BTF_LITTLE_ENDIAN && endian != BTF_BIG_ENDIAN)
741 return libbpf_err(-EINVAL);
742
743 btf->swapped_endian = is_host_big_endian() != (endian == BTF_BIG_ENDIAN);
744 if (!btf->swapped_endian) {
745 free(btf->raw_data_swapped);
746 btf->raw_data_swapped = NULL;
747 }
748 return 0;
749}
750
751static bool btf_type_is_void(const struct btf_type *t)
752{
753 return t == &btf_void || btf_is_fwd(t);
754}
755
756static bool btf_type_is_void_or_null(const struct btf_type *t)
757{
758 return !t || btf_type_is_void(t);
759}
760
761#define MAX_RESOLVE_DEPTH 32
762
763__s64 btf__resolve_size(const struct btf *btf, __u32 type_id)
764{
765 const struct btf_array *array;
766 const struct btf_type *t;
767 __u32 nelems = 1;
768 __s64 size = -1;
769 int i;
770
771 t = btf__type_by_id(btf, type_id);
772 for (i = 0; i < MAX_RESOLVE_DEPTH && !btf_type_is_void_or_null(t); i++) {
773 switch (btf_kind(t)) {
774 case BTF_KIND_INT:
775 case BTF_KIND_STRUCT:
776 case BTF_KIND_UNION:
777 case BTF_KIND_ENUM:
778 case BTF_KIND_ENUM64:
779 case BTF_KIND_DATASEC:
780 case BTF_KIND_FLOAT:
781 size = t->size;
782 goto done;
783 case BTF_KIND_PTR:
784 size = btf_ptr_sz(btf);
785 goto done;
786 case BTF_KIND_TYPEDEF:
787 case BTF_KIND_VOLATILE:
788 case BTF_KIND_CONST:
789 case BTF_KIND_RESTRICT:
790 case BTF_KIND_VAR:
791 case BTF_KIND_DECL_TAG:
792 case BTF_KIND_TYPE_TAG:
793 type_id = t->type;
794 break;
795 case BTF_KIND_ARRAY:
796 array = btf_array(t);
797 if (nelems && array->nelems > UINT32_MAX / nelems)
798 return libbpf_err(-E2BIG);
799 nelems *= array->nelems;
800 type_id = array->type;
801 break;
802 default:
803 return libbpf_err(-EINVAL);
804 }
805
806 t = btf__type_by_id(btf, type_id);
807 }
808
809done:
810 if (size < 0)
811 return libbpf_err(-EINVAL);
812 if (nelems && size > UINT32_MAX / nelems)
813 return libbpf_err(-E2BIG);
814
815 return nelems * size;
816}
817
818int btf__align_of(const struct btf *btf, __u32 id)
819{
820 const struct btf_type *t = btf__type_by_id(btf, id);
821 __u16 kind = btf_kind(t);
822
823 switch (kind) {
824 case BTF_KIND_INT:
825 case BTF_KIND_ENUM:
826 case BTF_KIND_ENUM64:
827 case BTF_KIND_FLOAT:
828 return min(btf_ptr_sz(btf), (size_t)t->size);
829 case BTF_KIND_PTR:
830 return btf_ptr_sz(btf);
831 case BTF_KIND_TYPEDEF:
832 case BTF_KIND_VOLATILE:
833 case BTF_KIND_CONST:
834 case BTF_KIND_RESTRICT:
835 case BTF_KIND_TYPE_TAG:
836 return btf__align_of(btf, t->type);
837 case BTF_KIND_ARRAY:
838 return btf__align_of(btf, btf_array(t)->type);
839 case BTF_KIND_STRUCT:
840 case BTF_KIND_UNION: {
841 const struct btf_member *m = btf_members(t);
842 __u16 vlen = btf_vlen(t);
843 int i, max_align = 1, align;
844
845 for (i = 0; i < vlen; i++, m++) {
846 align = btf__align_of(btf, m->type);
847 if (align <= 0)
848 return libbpf_err(align);
849 max_align = max(max_align, align);
850
851 /* if field offset isn't aligned according to field
852 * type's alignment, then struct must be packed
853 */
854 if (btf_member_bitfield_size(t, i) == 0 &&
855 (m->offset % (8 * align)) != 0)
856 return 1;
857 }
858
859 /* if struct/union size isn't a multiple of its alignment,
860 * then struct must be packed
861 */
862 if ((t->size % max_align) != 0)
863 return 1;
864
865 return max_align;
866 }
867 default:
868 pr_warn("unsupported BTF_KIND:%u\n", btf_kind(t));
869 return errno = EINVAL, 0;
870 }
871}
872
873int btf__resolve_type(const struct btf *btf, __u32 type_id)
874{
875 const struct btf_type *t;
876 int depth = 0;
877
878 t = btf__type_by_id(btf, type_id);
879 while (depth < MAX_RESOLVE_DEPTH &&
880 !btf_type_is_void_or_null(t) &&
881 (btf_is_mod(t) || btf_is_typedef(t) || btf_is_var(t))) {
882 type_id = t->type;
883 t = btf__type_by_id(btf, type_id);
884 depth++;
885 }
886
887 if (depth == MAX_RESOLVE_DEPTH || btf_type_is_void_or_null(t))
888 return libbpf_err(-EINVAL);
889
890 return type_id;
891}
892
893__s32 btf__find_by_name(const struct btf *btf, const char *type_name)
894{
895 __u32 i, nr_types = btf__type_cnt(btf);
896
897 if (!strcmp(type_name, "void"))
898 return 0;
899
900 for (i = 1; i < nr_types; i++) {
901 const struct btf_type *t = btf__type_by_id(btf, i);
902 const char *name = btf__name_by_offset(btf, t->name_off);
903
904 if (name && !strcmp(type_name, name))
905 return i;
906 }
907
908 return libbpf_err(-ENOENT);
909}
910
911static __s32 btf_find_by_name_kind(const struct btf *btf, int start_id,
912 const char *type_name, __u32 kind)
913{
914 __u32 i, nr_types = btf__type_cnt(btf);
915
916 if (kind == BTF_KIND_UNKN || !strcmp(type_name, "void"))
917 return 0;
918
919 for (i = start_id; i < nr_types; i++) {
920 const struct btf_type *t = btf__type_by_id(btf, i);
921 const char *name;
922
923 if (btf_kind(t) != kind)
924 continue;
925 name = btf__name_by_offset(btf, t->name_off);
926 if (name && !strcmp(type_name, name))
927 return i;
928 }
929
930 return libbpf_err(-ENOENT);
931}
932
933__s32 btf__find_by_name_kind_own(const struct btf *btf, const char *type_name,
934 __u32 kind)
935{
936 return btf_find_by_name_kind(btf, btf->start_id, type_name, kind);
937}
938
939__s32 btf__find_by_name_kind(const struct btf *btf, const char *type_name,
940 __u32 kind)
941{
942 return btf_find_by_name_kind(btf, 1, type_name, kind);
943}
944
945static bool btf_is_modifiable(const struct btf *btf)
946{
947 return (void *)btf->hdr != btf->raw_data;
948}
949
950void btf__free(struct btf *btf)
951{
952 if (IS_ERR_OR_NULL(btf))
953 return;
954
955 if (btf->fd >= 0)
956 close(btf->fd);
957
958 if (btf_is_modifiable(btf)) {
959 /* if BTF was modified after loading, it will have a split
960 * in-memory representation for header, types, and strings
961 * sections, so we need to free all of them individually. It
962 * might still have a cached contiguous raw data present,
963 * which will be unconditionally freed below.
964 */
965 free(btf->hdr);
966 free(btf->types_data);
967 strset__free(btf->strs_set);
968 }
969 free(btf->raw_data);
970 free(btf->raw_data_swapped);
971 free(btf->type_offs);
972 free(btf);
973}
974
975static struct btf *btf_new_empty(struct btf *base_btf)
976{
977 struct btf *btf;
978
979 btf = calloc(1, sizeof(*btf));
980 if (!btf)
981 return ERR_PTR(-ENOMEM);
982
983 btf->nr_types = 0;
984 btf->start_id = 1;
985 btf->start_str_off = 0;
986 btf->fd = -1;
987 btf->ptr_sz = sizeof(void *);
988 btf->swapped_endian = false;
989
990 if (base_btf) {
991 btf->base_btf = base_btf;
992 btf->start_id = btf__type_cnt(base_btf);
993 btf->start_str_off = base_btf->hdr->str_len;
994 }
995
996 /* +1 for empty string at offset 0 */
997 btf->raw_size = sizeof(struct btf_header) + (base_btf ? 0 : 1);
998 btf->raw_data = calloc(1, btf->raw_size);
999 if (!btf->raw_data) {
1000 free(btf);
1001 return ERR_PTR(-ENOMEM);
1002 }
1003
1004 btf->hdr = btf->raw_data;
1005 btf->hdr->hdr_len = sizeof(struct btf_header);
1006 btf->hdr->magic = BTF_MAGIC;
1007 btf->hdr->version = BTF_VERSION;
1008
1009 btf->types_data = btf->raw_data + btf->hdr->hdr_len;
1010 btf->strs_data = btf->raw_data + btf->hdr->hdr_len;
1011 btf->hdr->str_len = base_btf ? 0 : 1; /* empty string at offset 0 */
1012
1013 return btf;
1014}
1015
1016struct btf *btf__new_empty(void)
1017{
1018 return libbpf_ptr(btf_new_empty(NULL));
1019}
1020
1021struct btf *btf__new_empty_split(struct btf *base_btf)
1022{
1023 return libbpf_ptr(btf_new_empty(base_btf));
1024}
1025
1026static struct btf *btf_new(const void *data, __u32 size, struct btf *base_btf)
1027{
1028 struct btf *btf;
1029 int err;
1030
1031 btf = calloc(1, sizeof(struct btf));
1032 if (!btf)
1033 return ERR_PTR(-ENOMEM);
1034
1035 btf->nr_types = 0;
1036 btf->start_id = 1;
1037 btf->start_str_off = 0;
1038 btf->fd = -1;
1039
1040 if (base_btf) {
1041 btf->base_btf = base_btf;
1042 btf->start_id = btf__type_cnt(base_btf);
1043 btf->start_str_off = base_btf->hdr->str_len;
1044 }
1045
1046 btf->raw_data = malloc(size);
1047 if (!btf->raw_data) {
1048 err = -ENOMEM;
1049 goto done;
1050 }
1051 memcpy(btf->raw_data, data, size);
1052 btf->raw_size = size;
1053
1054 btf->hdr = btf->raw_data;
1055 err = btf_parse_hdr(btf);
1056 if (err)
1057 goto done;
1058
1059 btf->strs_data = btf->raw_data + btf->hdr->hdr_len + btf->hdr->str_off;
1060 btf->types_data = btf->raw_data + btf->hdr->hdr_len + btf->hdr->type_off;
1061
1062 err = btf_parse_str_sec(btf);
1063 err = err ?: btf_parse_type_sec(btf);
1064 err = err ?: btf_sanity_check(btf);
1065 if (err)
1066 goto done;
1067
1068done:
1069 if (err) {
1070 btf__free(btf);
1071 return ERR_PTR(err);
1072 }
1073
1074 return btf;
1075}
1076
1077struct btf *btf__new(const void *data, __u32 size)
1078{
1079 return libbpf_ptr(btf_new(data, size, NULL));
1080}
1081
1082static struct btf *btf_parse_elf(const char *path, struct btf *base_btf,
1083 struct btf_ext **btf_ext)
1084{
1085 Elf_Data *btf_data = NULL, *btf_ext_data = NULL;
1086 int err = 0, fd = -1, idx = 0;
1087 struct btf *btf = NULL;
1088 Elf_Scn *scn = NULL;
1089 Elf *elf = NULL;
1090 GElf_Ehdr ehdr;
1091 size_t shstrndx;
1092
1093 if (elf_version(EV_CURRENT) == EV_NONE) {
1094 pr_warn("failed to init libelf for %s\n", path);
1095 return ERR_PTR(-LIBBPF_ERRNO__LIBELF);
1096 }
1097
1098 fd = open(path, O_RDONLY | O_CLOEXEC);
1099 if (fd < 0) {
1100 err = -errno;
1101 pr_warn("failed to open %s: %s\n", path, strerror(errno));
1102 return ERR_PTR(err);
1103 }
1104
1105 err = -LIBBPF_ERRNO__FORMAT;
1106
1107 elf = elf_begin(fd, ELF_C_READ, NULL);
1108 if (!elf) {
1109 pr_warn("failed to open %s as ELF file\n", path);
1110 goto done;
1111 }
1112 if (!gelf_getehdr(elf, &ehdr)) {
1113 pr_warn("failed to get EHDR from %s\n", path);
1114 goto done;
1115 }
1116
1117 if (elf_getshdrstrndx(elf, &shstrndx)) {
1118 pr_warn("failed to get section names section index for %s\n",
1119 path);
1120 goto done;
1121 }
1122
1123 if (!elf_rawdata(elf_getscn(elf, shstrndx), NULL)) {
1124 pr_warn("failed to get e_shstrndx from %s\n", path);
1125 goto done;
1126 }
1127
1128 while ((scn = elf_nextscn(elf, scn)) != NULL) {
1129 GElf_Shdr sh;
1130 char *name;
1131
1132 idx++;
1133 if (gelf_getshdr(scn, &sh) != &sh) {
1134 pr_warn("failed to get section(%d) header from %s\n",
1135 idx, path);
1136 goto done;
1137 }
1138 name = elf_strptr(elf, shstrndx, sh.sh_name);
1139 if (!name) {
1140 pr_warn("failed to get section(%d) name from %s\n",
1141 idx, path);
1142 goto done;
1143 }
1144 if (strcmp(name, BTF_ELF_SEC) == 0) {
1145 btf_data = elf_getdata(scn, 0);
1146 if (!btf_data) {
1147 pr_warn("failed to get section(%d, %s) data from %s\n",
1148 idx, name, path);
1149 goto done;
1150 }
1151 continue;
1152 } else if (btf_ext && strcmp(name, BTF_EXT_ELF_SEC) == 0) {
1153 btf_ext_data = elf_getdata(scn, 0);
1154 if (!btf_ext_data) {
1155 pr_warn("failed to get section(%d, %s) data from %s\n",
1156 idx, name, path);
1157 goto done;
1158 }
1159 continue;
1160 }
1161 }
1162
1163 if (!btf_data) {
1164 pr_warn("failed to find '%s' ELF section in %s\n", BTF_ELF_SEC, path);
1165 err = -ENODATA;
1166 goto done;
1167 }
1168 btf = btf_new(btf_data->d_buf, btf_data->d_size, base_btf);
1169 err = libbpf_get_error(btf);
1170 if (err)
1171 goto done;
1172
1173 switch (gelf_getclass(elf)) {
1174 case ELFCLASS32:
1175 btf__set_pointer_size(btf, 4);
1176 break;
1177 case ELFCLASS64:
1178 btf__set_pointer_size(btf, 8);
1179 break;
1180 default:
1181 pr_warn("failed to get ELF class (bitness) for %s\n", path);
1182 break;
1183 }
1184
1185 if (btf_ext && btf_ext_data) {
1186 *btf_ext = btf_ext__new(btf_ext_data->d_buf, btf_ext_data->d_size);
1187 err = libbpf_get_error(*btf_ext);
1188 if (err)
1189 goto done;
1190 } else if (btf_ext) {
1191 *btf_ext = NULL;
1192 }
1193done:
1194 if (elf)
1195 elf_end(elf);
1196 close(fd);
1197
1198 if (!err)
1199 return btf;
1200
1201 if (btf_ext)
1202 btf_ext__free(*btf_ext);
1203 btf__free(btf);
1204
1205 return ERR_PTR(err);
1206}
1207
1208struct btf *btf__parse_elf(const char *path, struct btf_ext **btf_ext)
1209{
1210 return libbpf_ptr(btf_parse_elf(path, NULL, btf_ext));
1211}
1212
1213struct btf *btf__parse_elf_split(const char *path, struct btf *base_btf)
1214{
1215 return libbpf_ptr(btf_parse_elf(path, base_btf, NULL));
1216}
1217
1218static struct btf *btf_parse_raw(const char *path, struct btf *base_btf)
1219{
1220 struct btf *btf = NULL;
1221 void *data = NULL;
1222 FILE *f = NULL;
1223 __u16 magic;
1224 int err = 0;
1225 long sz;
1226
1227 f = fopen(path, "rbe");
1228 if (!f) {
1229 err = -errno;
1230 goto err_out;
1231 }
1232
1233 /* check BTF magic */
1234 if (fread(&magic, 1, sizeof(magic), f) < sizeof(magic)) {
1235 err = -EIO;
1236 goto err_out;
1237 }
1238 if (magic != BTF_MAGIC && magic != bswap_16(BTF_MAGIC)) {
1239 /* definitely not a raw BTF */
1240 err = -EPROTO;
1241 goto err_out;
1242 }
1243
1244 /* get file size */
1245 if (fseek(f, 0, SEEK_END)) {
1246 err = -errno;
1247 goto err_out;
1248 }
1249 sz = ftell(f);
1250 if (sz < 0) {
1251 err = -errno;
1252 goto err_out;
1253 }
1254 /* rewind to the start */
1255 if (fseek(f, 0, SEEK_SET)) {
1256 err = -errno;
1257 goto err_out;
1258 }
1259
1260 /* pre-alloc memory and read all of BTF data */
1261 data = malloc(sz);
1262 if (!data) {
1263 err = -ENOMEM;
1264 goto err_out;
1265 }
1266 if (fread(data, 1, sz, f) < sz) {
1267 err = -EIO;
1268 goto err_out;
1269 }
1270
1271 /* finally parse BTF data */
1272 btf = btf_new(data, sz, base_btf);
1273
1274err_out:
1275 free(data);
1276 if (f)
1277 fclose(f);
1278 return err ? ERR_PTR(err) : btf;
1279}
1280
1281struct btf *btf__parse_raw(const char *path)
1282{
1283 return libbpf_ptr(btf_parse_raw(path, NULL));
1284}
1285
1286struct btf *btf__parse_raw_split(const char *path, struct btf *base_btf)
1287{
1288 return libbpf_ptr(btf_parse_raw(path, base_btf));
1289}
1290
1291static struct btf *btf_parse(const char *path, struct btf *base_btf, struct btf_ext **btf_ext)
1292{
1293 struct btf *btf;
1294 int err;
1295
1296 if (btf_ext)
1297 *btf_ext = NULL;
1298
1299 btf = btf_parse_raw(path, base_btf);
1300 err = libbpf_get_error(btf);
1301 if (!err)
1302 return btf;
1303 if (err != -EPROTO)
1304 return ERR_PTR(err);
1305 return btf_parse_elf(path, base_btf, btf_ext);
1306}
1307
1308struct btf *btf__parse(const char *path, struct btf_ext **btf_ext)
1309{
1310 return libbpf_ptr(btf_parse(path, NULL, btf_ext));
1311}
1312
1313struct btf *btf__parse_split(const char *path, struct btf *base_btf)
1314{
1315 return libbpf_ptr(btf_parse(path, base_btf, NULL));
1316}
1317
1318static void *btf_get_raw_data(const struct btf *btf, __u32 *size, bool swap_endian);
1319
1320int btf_load_into_kernel(struct btf *btf, char *log_buf, size_t log_sz, __u32 log_level)
1321{
1322 LIBBPF_OPTS(bpf_btf_load_opts, opts);
1323 __u32 buf_sz = 0, raw_size;
1324 char *buf = NULL, *tmp;
1325 void *raw_data;
1326 int err = 0;
1327
1328 if (btf->fd >= 0)
1329 return libbpf_err(-EEXIST);
1330 if (log_sz && !log_buf)
1331 return libbpf_err(-EINVAL);
1332
1333 /* cache native raw data representation */
1334 raw_data = btf_get_raw_data(btf, &raw_size, false);
1335 if (!raw_data) {
1336 err = -ENOMEM;
1337 goto done;
1338 }
1339 btf->raw_size = raw_size;
1340 btf->raw_data = raw_data;
1341
1342retry_load:
1343 /* if log_level is 0, we won't provide log_buf/log_size to the kernel,
1344 * initially. Only if BTF loading fails, we bump log_level to 1 and
1345 * retry, using either auto-allocated or custom log_buf. This way
1346 * non-NULL custom log_buf provides a buffer just in case, but hopes
1347 * for successful load and no need for log_buf.
1348 */
1349 if (log_level) {
1350 /* if caller didn't provide custom log_buf, we'll keep
1351 * allocating our own progressively bigger buffers for BTF
1352 * verification log
1353 */
1354 if (!log_buf) {
1355 buf_sz = max((__u32)BPF_LOG_BUF_SIZE, buf_sz * 2);
1356 tmp = realloc(buf, buf_sz);
1357 if (!tmp) {
1358 err = -ENOMEM;
1359 goto done;
1360 }
1361 buf = tmp;
1362 buf[0] = '\0';
1363 }
1364
1365 opts.log_buf = log_buf ? log_buf : buf;
1366 opts.log_size = log_buf ? log_sz : buf_sz;
1367 opts.log_level = log_level;
1368 }
1369
1370 btf->fd = bpf_btf_load(raw_data, raw_size, &opts);
1371 if (btf->fd < 0) {
1372 /* time to turn on verbose mode and try again */
1373 if (log_level == 0) {
1374 log_level = 1;
1375 goto retry_load;
1376 }
1377 /* only retry if caller didn't provide custom log_buf, but
1378 * make sure we can never overflow buf_sz
1379 */
1380 if (!log_buf && errno == ENOSPC && buf_sz <= UINT_MAX / 2)
1381 goto retry_load;
1382
1383 err = -errno;
1384 pr_warn("BTF loading error: %d\n", err);
1385 /* don't print out contents of custom log_buf */
1386 if (!log_buf && buf[0])
1387 pr_warn("-- BEGIN BTF LOAD LOG ---\n%s\n-- END BTF LOAD LOG --\n", buf);
1388 }
1389
1390done:
1391 free(buf);
1392 return libbpf_err(err);
1393}
1394
1395int btf__load_into_kernel(struct btf *btf)
1396{
1397 return btf_load_into_kernel(btf, NULL, 0, 0);
1398}
1399
1400int btf__fd(const struct btf *btf)
1401{
1402 return btf->fd;
1403}
1404
1405void btf__set_fd(struct btf *btf, int fd)
1406{
1407 btf->fd = fd;
1408}
1409
1410static const void *btf_strs_data(const struct btf *btf)
1411{
1412 return btf->strs_data ? btf->strs_data : strset__data(btf->strs_set);
1413}
1414
1415static void *btf_get_raw_data(const struct btf *btf, __u32 *size, bool swap_endian)
1416{
1417 struct btf_header *hdr = btf->hdr;
1418 struct btf_type *t;
1419 void *data, *p;
1420 __u32 data_sz;
1421 int i;
1422
1423 data = swap_endian ? btf->raw_data_swapped : btf->raw_data;
1424 if (data) {
1425 *size = btf->raw_size;
1426 return data;
1427 }
1428
1429 data_sz = hdr->hdr_len + hdr->type_len + hdr->str_len;
1430 data = calloc(1, data_sz);
1431 if (!data)
1432 return NULL;
1433 p = data;
1434
1435 memcpy(p, hdr, hdr->hdr_len);
1436 if (swap_endian)
1437 btf_bswap_hdr(p);
1438 p += hdr->hdr_len;
1439
1440 memcpy(p, btf->types_data, hdr->type_len);
1441 if (swap_endian) {
1442 for (i = 0; i < btf->nr_types; i++) {
1443 t = p + btf->type_offs[i];
1444 /* btf_bswap_type_rest() relies on native t->info, so
1445 * we swap base type info after we swapped all the
1446 * additional information
1447 */
1448 if (btf_bswap_type_rest(t))
1449 goto err_out;
1450 btf_bswap_type_base(t);
1451 }
1452 }
1453 p += hdr->type_len;
1454
1455 memcpy(p, btf_strs_data(btf), hdr->str_len);
1456 p += hdr->str_len;
1457
1458 *size = data_sz;
1459 return data;
1460err_out:
1461 free(data);
1462 return NULL;
1463}
1464
1465const void *btf__raw_data(const struct btf *btf_ro, __u32 *size)
1466{
1467 struct btf *btf = (struct btf *)btf_ro;
1468 __u32 data_sz;
1469 void *data;
1470
1471 data = btf_get_raw_data(btf, &data_sz, btf->swapped_endian);
1472 if (!data)
1473 return errno = ENOMEM, NULL;
1474
1475 btf->raw_size = data_sz;
1476 if (btf->swapped_endian)
1477 btf->raw_data_swapped = data;
1478 else
1479 btf->raw_data = data;
1480 *size = data_sz;
1481 return data;
1482}
1483
1484__attribute__((alias("btf__raw_data")))
1485const void *btf__get_raw_data(const struct btf *btf, __u32 *size);
1486
1487const char *btf__str_by_offset(const struct btf *btf, __u32 offset)
1488{
1489 if (offset < btf->start_str_off)
1490 return btf__str_by_offset(btf->base_btf, offset);
1491 else if (offset - btf->start_str_off < btf->hdr->str_len)
1492 return btf_strs_data(btf) + (offset - btf->start_str_off);
1493 else
1494 return errno = EINVAL, NULL;
1495}
1496
1497const char *btf__name_by_offset(const struct btf *btf, __u32 offset)
1498{
1499 return btf__str_by_offset(btf, offset);
1500}
1501
1502struct btf *btf_get_from_fd(int btf_fd, struct btf *base_btf)
1503{
1504 struct bpf_btf_info btf_info;
1505 __u32 len = sizeof(btf_info);
1506 __u32 last_size;
1507 struct btf *btf;
1508 void *ptr;
1509 int err;
1510
1511 /* we won't know btf_size until we call bpf_btf_get_info_by_fd(). so
1512 * let's start with a sane default - 4KiB here - and resize it only if
1513 * bpf_btf_get_info_by_fd() needs a bigger buffer.
1514 */
1515 last_size = 4096;
1516 ptr = malloc(last_size);
1517 if (!ptr)
1518 return ERR_PTR(-ENOMEM);
1519
1520 memset(&btf_info, 0, sizeof(btf_info));
1521 btf_info.btf = ptr_to_u64(ptr);
1522 btf_info.btf_size = last_size;
1523 err = bpf_btf_get_info_by_fd(btf_fd, &btf_info, &len);
1524
1525 if (!err && btf_info.btf_size > last_size) {
1526 void *temp_ptr;
1527
1528 last_size = btf_info.btf_size;
1529 temp_ptr = realloc(ptr, last_size);
1530 if (!temp_ptr) {
1531 btf = ERR_PTR(-ENOMEM);
1532 goto exit_free;
1533 }
1534 ptr = temp_ptr;
1535
1536 len = sizeof(btf_info);
1537 memset(&btf_info, 0, sizeof(btf_info));
1538 btf_info.btf = ptr_to_u64(ptr);
1539 btf_info.btf_size = last_size;
1540
1541 err = bpf_btf_get_info_by_fd(btf_fd, &btf_info, &len);
1542 }
1543
1544 if (err || btf_info.btf_size > last_size) {
1545 btf = err ? ERR_PTR(-errno) : ERR_PTR(-E2BIG);
1546 goto exit_free;
1547 }
1548
1549 btf = btf_new(ptr, btf_info.btf_size, base_btf);
1550
1551exit_free:
1552 free(ptr);
1553 return btf;
1554}
1555
1556struct btf *btf__load_from_kernel_by_id_split(__u32 id, struct btf *base_btf)
1557{
1558 struct btf *btf;
1559 int btf_fd;
1560
1561 btf_fd = bpf_btf_get_fd_by_id(id);
1562 if (btf_fd < 0)
1563 return libbpf_err_ptr(-errno);
1564
1565 btf = btf_get_from_fd(btf_fd, base_btf);
1566 close(btf_fd);
1567
1568 return libbpf_ptr(btf);
1569}
1570
1571struct btf *btf__load_from_kernel_by_id(__u32 id)
1572{
1573 return btf__load_from_kernel_by_id_split(id, NULL);
1574}
1575
1576static void btf_invalidate_raw_data(struct btf *btf)
1577{
1578 if (btf->raw_data) {
1579 free(btf->raw_data);
1580 btf->raw_data = NULL;
1581 }
1582 if (btf->raw_data_swapped) {
1583 free(btf->raw_data_swapped);
1584 btf->raw_data_swapped = NULL;
1585 }
1586}
1587
1588/* Ensure BTF is ready to be modified (by splitting into a three memory
1589 * regions for header, types, and strings). Also invalidate cached
1590 * raw_data, if any.
1591 */
1592static int btf_ensure_modifiable(struct btf *btf)
1593{
1594 void *hdr, *types;
1595 struct strset *set = NULL;
1596 int err = -ENOMEM;
1597
1598 if (btf_is_modifiable(btf)) {
1599 /* any BTF modification invalidates raw_data */
1600 btf_invalidate_raw_data(btf);
1601 return 0;
1602 }
1603
1604 /* split raw data into three memory regions */
1605 hdr = malloc(btf->hdr->hdr_len);
1606 types = malloc(btf->hdr->type_len);
1607 if (!hdr || !types)
1608 goto err_out;
1609
1610 memcpy(hdr, btf->hdr, btf->hdr->hdr_len);
1611 memcpy(types, btf->types_data, btf->hdr->type_len);
1612
1613 /* build lookup index for all strings */
1614 set = strset__new(BTF_MAX_STR_OFFSET, btf->strs_data, btf->hdr->str_len);
1615 if (IS_ERR(set)) {
1616 err = PTR_ERR(set);
1617 goto err_out;
1618 }
1619
1620 /* only when everything was successful, update internal state */
1621 btf->hdr = hdr;
1622 btf->types_data = types;
1623 btf->types_data_cap = btf->hdr->type_len;
1624 btf->strs_data = NULL;
1625 btf->strs_set = set;
1626 /* if BTF was created from scratch, all strings are guaranteed to be
1627 * unique and deduplicated
1628 */
1629 if (btf->hdr->str_len == 0)
1630 btf->strs_deduped = true;
1631 if (!btf->base_btf && btf->hdr->str_len == 1)
1632 btf->strs_deduped = true;
1633
1634 /* invalidate raw_data representation */
1635 btf_invalidate_raw_data(btf);
1636
1637 return 0;
1638
1639err_out:
1640 strset__free(set);
1641 free(hdr);
1642 free(types);
1643 return err;
1644}
1645
1646/* Find an offset in BTF string section that corresponds to a given string *s*.
1647 * Returns:
1648 * - >0 offset into string section, if string is found;
1649 * - -ENOENT, if string is not in the string section;
1650 * - <0, on any other error.
1651 */
1652int btf__find_str(struct btf *btf, const char *s)
1653{
1654 int off;
1655
1656 if (btf->base_btf) {
1657 off = btf__find_str(btf->base_btf, s);
1658 if (off != -ENOENT)
1659 return off;
1660 }
1661
1662 /* BTF needs to be in a modifiable state to build string lookup index */
1663 if (btf_ensure_modifiable(btf))
1664 return libbpf_err(-ENOMEM);
1665
1666 off = strset__find_str(btf->strs_set, s);
1667 if (off < 0)
1668 return libbpf_err(off);
1669
1670 return btf->start_str_off + off;
1671}
1672
1673/* Add a string s to the BTF string section.
1674 * Returns:
1675 * - > 0 offset into string section, on success;
1676 * - < 0, on error.
1677 */
1678int btf__add_str(struct btf *btf, const char *s)
1679{
1680 int off;
1681
1682 if (btf->base_btf) {
1683 off = btf__find_str(btf->base_btf, s);
1684 if (off != -ENOENT)
1685 return off;
1686 }
1687
1688 if (btf_ensure_modifiable(btf))
1689 return libbpf_err(-ENOMEM);
1690
1691 off = strset__add_str(btf->strs_set, s);
1692 if (off < 0)
1693 return libbpf_err(off);
1694
1695 btf->hdr->str_len = strset__data_size(btf->strs_set);
1696
1697 return btf->start_str_off + off;
1698}
1699
1700static void *btf_add_type_mem(struct btf *btf, size_t add_sz)
1701{
1702 return libbpf_add_mem(&btf->types_data, &btf->types_data_cap, 1,
1703 btf->hdr->type_len, UINT_MAX, add_sz);
1704}
1705
1706static void btf_type_inc_vlen(struct btf_type *t)
1707{
1708 t->info = btf_type_info(btf_kind(t), btf_vlen(t) + 1, btf_kflag(t));
1709}
1710
1711static int btf_commit_type(struct btf *btf, int data_sz)
1712{
1713 int err;
1714
1715 err = btf_add_type_idx_entry(btf, btf->hdr->type_len);
1716 if (err)
1717 return libbpf_err(err);
1718
1719 btf->hdr->type_len += data_sz;
1720 btf->hdr->str_off += data_sz;
1721 btf->nr_types++;
1722 return btf->start_id + btf->nr_types - 1;
1723}
1724
1725struct btf_pipe {
1726 const struct btf *src;
1727 struct btf *dst;
1728 struct hashmap *str_off_map; /* map string offsets from src to dst */
1729};
1730
1731static int btf_rewrite_str(__u32 *str_off, void *ctx)
1732{
1733 struct btf_pipe *p = ctx;
1734 long mapped_off;
1735 int off, err;
1736
1737 if (!*str_off) /* nothing to do for empty strings */
1738 return 0;
1739
1740 if (p->str_off_map &&
1741 hashmap__find(p->str_off_map, *str_off, &mapped_off)) {
1742 *str_off = mapped_off;
1743 return 0;
1744 }
1745
1746 off = btf__add_str(p->dst, btf__str_by_offset(p->src, *str_off));
1747 if (off < 0)
1748 return off;
1749
1750 /* Remember string mapping from src to dst. It avoids
1751 * performing expensive string comparisons.
1752 */
1753 if (p->str_off_map) {
1754 err = hashmap__append(p->str_off_map, *str_off, off);
1755 if (err)
1756 return err;
1757 }
1758
1759 *str_off = off;
1760 return 0;
1761}
1762
1763int btf__add_type(struct btf *btf, const struct btf *src_btf, const struct btf_type *src_type)
1764{
1765 struct btf_pipe p = { .src = src_btf, .dst = btf };
1766 struct btf_type *t;
1767 int sz, err;
1768
1769 sz = btf_type_size(src_type);
1770 if (sz < 0)
1771 return libbpf_err(sz);
1772
1773 /* deconstruct BTF, if necessary, and invalidate raw_data */
1774 if (btf_ensure_modifiable(btf))
1775 return libbpf_err(-ENOMEM);
1776
1777 t = btf_add_type_mem(btf, sz);
1778 if (!t)
1779 return libbpf_err(-ENOMEM);
1780
1781 memcpy(t, src_type, sz);
1782
1783 err = btf_type_visit_str_offs(t, btf_rewrite_str, &p);
1784 if (err)
1785 return libbpf_err(err);
1786
1787 return btf_commit_type(btf, sz);
1788}
1789
1790static int btf_rewrite_type_ids(__u32 *type_id, void *ctx)
1791{
1792 struct btf *btf = ctx;
1793
1794 if (!*type_id) /* nothing to do for VOID references */
1795 return 0;
1796
1797 /* we haven't updated btf's type count yet, so
1798 * btf->start_id + btf->nr_types - 1 is the type ID offset we should
1799 * add to all newly added BTF types
1800 */
1801 *type_id += btf->start_id + btf->nr_types - 1;
1802 return 0;
1803}
1804
1805static size_t btf_dedup_identity_hash_fn(long key, void *ctx);
1806static bool btf_dedup_equal_fn(long k1, long k2, void *ctx);
1807
1808int btf__add_btf(struct btf *btf, const struct btf *src_btf)
1809{
1810 struct btf_pipe p = { .src = src_btf, .dst = btf };
1811 int data_sz, sz, cnt, i, err, old_strs_len;
1812 __u32 *off;
1813 void *t;
1814
1815 /* appending split BTF isn't supported yet */
1816 if (src_btf->base_btf)
1817 return libbpf_err(-ENOTSUP);
1818
1819 /* deconstruct BTF, if necessary, and invalidate raw_data */
1820 if (btf_ensure_modifiable(btf))
1821 return libbpf_err(-ENOMEM);
1822
1823 /* remember original strings section size if we have to roll back
1824 * partial strings section changes
1825 */
1826 old_strs_len = btf->hdr->str_len;
1827
1828 data_sz = src_btf->hdr->type_len;
1829 cnt = btf__type_cnt(src_btf) - 1;
1830
1831 /* pre-allocate enough memory for new types */
1832 t = btf_add_type_mem(btf, data_sz);
1833 if (!t)
1834 return libbpf_err(-ENOMEM);
1835
1836 /* pre-allocate enough memory for type offset index for new types */
1837 off = btf_add_type_offs_mem(btf, cnt);
1838 if (!off)
1839 return libbpf_err(-ENOMEM);
1840
1841 /* Map the string offsets from src_btf to the offsets from btf to improve performance */
1842 p.str_off_map = hashmap__new(btf_dedup_identity_hash_fn, btf_dedup_equal_fn, NULL);
1843 if (IS_ERR(p.str_off_map))
1844 return libbpf_err(-ENOMEM);
1845
1846 /* bulk copy types data for all types from src_btf */
1847 memcpy(t, src_btf->types_data, data_sz);
1848
1849 for (i = 0; i < cnt; i++) {
1850 sz = btf_type_size(t);
1851 if (sz < 0) {
1852 /* unlikely, has to be corrupted src_btf */
1853 err = sz;
1854 goto err_out;
1855 }
1856
1857 /* fill out type ID to type offset mapping for lookups by type ID */
1858 *off = t - btf->types_data;
1859
1860 /* add, dedup, and remap strings referenced by this BTF type */
1861 err = btf_type_visit_str_offs(t, btf_rewrite_str, &p);
1862 if (err)
1863 goto err_out;
1864
1865 /* remap all type IDs referenced from this BTF type */
1866 err = btf_type_visit_type_ids(t, btf_rewrite_type_ids, btf);
1867 if (err)
1868 goto err_out;
1869
1870 /* go to next type data and type offset index entry */
1871 t += sz;
1872 off++;
1873 }
1874
1875 /* Up until now any of the copied type data was effectively invisible,
1876 * so if we exited early before this point due to error, BTF would be
1877 * effectively unmodified. There would be extra internal memory
1878 * pre-allocated, but it would not be available for querying. But now
1879 * that we've copied and rewritten all the data successfully, we can
1880 * update type count and various internal offsets and sizes to
1881 * "commit" the changes and made them visible to the outside world.
1882 */
1883 btf->hdr->type_len += data_sz;
1884 btf->hdr->str_off += data_sz;
1885 btf->nr_types += cnt;
1886
1887 hashmap__free(p.str_off_map);
1888
1889 /* return type ID of the first added BTF type */
1890 return btf->start_id + btf->nr_types - cnt;
1891err_out:
1892 /* zero out preallocated memory as if it was just allocated with
1893 * libbpf_add_mem()
1894 */
1895 memset(btf->types_data + btf->hdr->type_len, 0, data_sz);
1896 memset(btf->strs_data + old_strs_len, 0, btf->hdr->str_len - old_strs_len);
1897
1898 /* and now restore original strings section size; types data size
1899 * wasn't modified, so doesn't need restoring, see big comment above
1900 */
1901 btf->hdr->str_len = old_strs_len;
1902
1903 hashmap__free(p.str_off_map);
1904
1905 return libbpf_err(err);
1906}
1907
1908/*
1909 * Append new BTF_KIND_INT type with:
1910 * - *name* - non-empty, non-NULL type name;
1911 * - *sz* - power-of-2 (1, 2, 4, ..) size of the type, in bytes;
1912 * - encoding is a combination of BTF_INT_SIGNED, BTF_INT_CHAR, BTF_INT_BOOL.
1913 * Returns:
1914 * - >0, type ID of newly added BTF type;
1915 * - <0, on error.
1916 */
1917int btf__add_int(struct btf *btf, const char *name, size_t byte_sz, int encoding)
1918{
1919 struct btf_type *t;
1920 int sz, name_off;
1921
1922 /* non-empty name */
1923 if (!name || !name[0])
1924 return libbpf_err(-EINVAL);
1925 /* byte_sz must be power of 2 */
1926 if (!byte_sz || (byte_sz & (byte_sz - 1)) || byte_sz > 16)
1927 return libbpf_err(-EINVAL);
1928 if (encoding & ~(BTF_INT_SIGNED | BTF_INT_CHAR | BTF_INT_BOOL))
1929 return libbpf_err(-EINVAL);
1930
1931 /* deconstruct BTF, if necessary, and invalidate raw_data */
1932 if (btf_ensure_modifiable(btf))
1933 return libbpf_err(-ENOMEM);
1934
1935 sz = sizeof(struct btf_type) + sizeof(int);
1936 t = btf_add_type_mem(btf, sz);
1937 if (!t)
1938 return libbpf_err(-ENOMEM);
1939
1940 /* if something goes wrong later, we might end up with an extra string,
1941 * but that shouldn't be a problem, because BTF can't be constructed
1942 * completely anyway and will most probably be just discarded
1943 */
1944 name_off = btf__add_str(btf, name);
1945 if (name_off < 0)
1946 return name_off;
1947
1948 t->name_off = name_off;
1949 t->info = btf_type_info(BTF_KIND_INT, 0, 0);
1950 t->size = byte_sz;
1951 /* set INT info, we don't allow setting legacy bit offset/size */
1952 *(__u32 *)(t + 1) = (encoding << 24) | (byte_sz * 8);
1953
1954 return btf_commit_type(btf, sz);
1955}
1956
1957/*
1958 * Append new BTF_KIND_FLOAT type with:
1959 * - *name* - non-empty, non-NULL type name;
1960 * - *sz* - size of the type, in bytes;
1961 * Returns:
1962 * - >0, type ID of newly added BTF type;
1963 * - <0, on error.
1964 */
1965int btf__add_float(struct btf *btf, const char *name, size_t byte_sz)
1966{
1967 struct btf_type *t;
1968 int sz, name_off;
1969
1970 /* non-empty name */
1971 if (!name || !name[0])
1972 return libbpf_err(-EINVAL);
1973
1974 /* byte_sz must be one of the explicitly allowed values */
1975 if (byte_sz != 2 && byte_sz != 4 && byte_sz != 8 && byte_sz != 12 &&
1976 byte_sz != 16)
1977 return libbpf_err(-EINVAL);
1978
1979 if (btf_ensure_modifiable(btf))
1980 return libbpf_err(-ENOMEM);
1981
1982 sz = sizeof(struct btf_type);
1983 t = btf_add_type_mem(btf, sz);
1984 if (!t)
1985 return libbpf_err(-ENOMEM);
1986
1987 name_off = btf__add_str(btf, name);
1988 if (name_off < 0)
1989 return name_off;
1990
1991 t->name_off = name_off;
1992 t->info = btf_type_info(BTF_KIND_FLOAT, 0, 0);
1993 t->size = byte_sz;
1994
1995 return btf_commit_type(btf, sz);
1996}
1997
1998/* it's completely legal to append BTF types with type IDs pointing forward to
1999 * types that haven't been appended yet, so we only make sure that id looks
2000 * sane, we can't guarantee that ID will always be valid
2001 */
2002static int validate_type_id(int id)
2003{
2004 if (id < 0 || id > BTF_MAX_NR_TYPES)
2005 return -EINVAL;
2006 return 0;
2007}
2008
2009/* generic append function for PTR, TYPEDEF, CONST/VOLATILE/RESTRICT */
2010static int btf_add_ref_kind(struct btf *btf, int kind, const char *name, int ref_type_id)
2011{
2012 struct btf_type *t;
2013 int sz, name_off = 0;
2014
2015 if (validate_type_id(ref_type_id))
2016 return libbpf_err(-EINVAL);
2017
2018 if (btf_ensure_modifiable(btf))
2019 return libbpf_err(-ENOMEM);
2020
2021 sz = sizeof(struct btf_type);
2022 t = btf_add_type_mem(btf, sz);
2023 if (!t)
2024 return libbpf_err(-ENOMEM);
2025
2026 if (name && name[0]) {
2027 name_off = btf__add_str(btf, name);
2028 if (name_off < 0)
2029 return name_off;
2030 }
2031
2032 t->name_off = name_off;
2033 t->info = btf_type_info(kind, 0, 0);
2034 t->type = ref_type_id;
2035
2036 return btf_commit_type(btf, sz);
2037}
2038
2039/*
2040 * Append new BTF_KIND_PTR type with:
2041 * - *ref_type_id* - referenced type ID, it might not exist yet;
2042 * Returns:
2043 * - >0, type ID of newly added BTF type;
2044 * - <0, on error.
2045 */
2046int btf__add_ptr(struct btf *btf, int ref_type_id)
2047{
2048 return btf_add_ref_kind(btf, BTF_KIND_PTR, NULL, ref_type_id);
2049}
2050
2051/*
2052 * Append new BTF_KIND_ARRAY type with:
2053 * - *index_type_id* - type ID of the type describing array index;
2054 * - *elem_type_id* - type ID of the type describing array element;
2055 * - *nr_elems* - the size of the array;
2056 * Returns:
2057 * - >0, type ID of newly added BTF type;
2058 * - <0, on error.
2059 */
2060int btf__add_array(struct btf *btf, int index_type_id, int elem_type_id, __u32 nr_elems)
2061{
2062 struct btf_type *t;
2063 struct btf_array *a;
2064 int sz;
2065
2066 if (validate_type_id(index_type_id) || validate_type_id(elem_type_id))
2067 return libbpf_err(-EINVAL);
2068
2069 if (btf_ensure_modifiable(btf))
2070 return libbpf_err(-ENOMEM);
2071
2072 sz = sizeof(struct btf_type) + sizeof(struct btf_array);
2073 t = btf_add_type_mem(btf, sz);
2074 if (!t)
2075 return libbpf_err(-ENOMEM);
2076
2077 t->name_off = 0;
2078 t->info = btf_type_info(BTF_KIND_ARRAY, 0, 0);
2079 t->size = 0;
2080
2081 a = btf_array(t);
2082 a->type = elem_type_id;
2083 a->index_type = index_type_id;
2084 a->nelems = nr_elems;
2085
2086 return btf_commit_type(btf, sz);
2087}
2088
2089/* generic STRUCT/UNION append function */
2090static int btf_add_composite(struct btf *btf, int kind, const char *name, __u32 bytes_sz)
2091{
2092 struct btf_type *t;
2093 int sz, name_off = 0;
2094
2095 if (btf_ensure_modifiable(btf))
2096 return libbpf_err(-ENOMEM);
2097
2098 sz = sizeof(struct btf_type);
2099 t = btf_add_type_mem(btf, sz);
2100 if (!t)
2101 return libbpf_err(-ENOMEM);
2102
2103 if (name && name[0]) {
2104 name_off = btf__add_str(btf, name);
2105 if (name_off < 0)
2106 return name_off;
2107 }
2108
2109 /* start out with vlen=0 and no kflag; this will be adjusted when
2110 * adding each member
2111 */
2112 t->name_off = name_off;
2113 t->info = btf_type_info(kind, 0, 0);
2114 t->size = bytes_sz;
2115
2116 return btf_commit_type(btf, sz);
2117}
2118
2119/*
2120 * Append new BTF_KIND_STRUCT type with:
2121 * - *name* - name of the struct, can be NULL or empty for anonymous structs;
2122 * - *byte_sz* - size of the struct, in bytes;
2123 *
2124 * Struct initially has no fields in it. Fields can be added by
2125 * btf__add_field() right after btf__add_struct() succeeds.
2126 *
2127 * Returns:
2128 * - >0, type ID of newly added BTF type;
2129 * - <0, on error.
2130 */
2131int btf__add_struct(struct btf *btf, const char *name, __u32 byte_sz)
2132{
2133 return btf_add_composite(btf, BTF_KIND_STRUCT, name, byte_sz);
2134}
2135
2136/*
2137 * Append new BTF_KIND_UNION type with:
2138 * - *name* - name of the union, can be NULL or empty for anonymous union;
2139 * - *byte_sz* - size of the union, in bytes;
2140 *
2141 * Union initially has no fields in it. Fields can be added by
2142 * btf__add_field() right after btf__add_union() succeeds. All fields
2143 * should have *bit_offset* of 0.
2144 *
2145 * Returns:
2146 * - >0, type ID of newly added BTF type;
2147 * - <0, on error.
2148 */
2149int btf__add_union(struct btf *btf, const char *name, __u32 byte_sz)
2150{
2151 return btf_add_composite(btf, BTF_KIND_UNION, name, byte_sz);
2152}
2153
2154static struct btf_type *btf_last_type(struct btf *btf)
2155{
2156 return btf_type_by_id(btf, btf__type_cnt(btf) - 1);
2157}
2158
2159/*
2160 * Append new field for the current STRUCT/UNION type with:
2161 * - *name* - name of the field, can be NULL or empty for anonymous field;
2162 * - *type_id* - type ID for the type describing field type;
2163 * - *bit_offset* - bit offset of the start of the field within struct/union;
2164 * - *bit_size* - bit size of a bitfield, 0 for non-bitfield fields;
2165 * Returns:
2166 * - 0, on success;
2167 * - <0, on error.
2168 */
2169int btf__add_field(struct btf *btf, const char *name, int type_id,
2170 __u32 bit_offset, __u32 bit_size)
2171{
2172 struct btf_type *t;
2173 struct btf_member *m;
2174 bool is_bitfield;
2175 int sz, name_off = 0;
2176
2177 /* last type should be union/struct */
2178 if (btf->nr_types == 0)
2179 return libbpf_err(-EINVAL);
2180 t = btf_last_type(btf);
2181 if (!btf_is_composite(t))
2182 return libbpf_err(-EINVAL);
2183
2184 if (validate_type_id(type_id))
2185 return libbpf_err(-EINVAL);
2186 /* best-effort bit field offset/size enforcement */
2187 is_bitfield = bit_size || (bit_offset % 8 != 0);
2188 if (is_bitfield && (bit_size == 0 || bit_size > 255 || bit_offset > 0xffffff))
2189 return libbpf_err(-EINVAL);
2190
2191 /* only offset 0 is allowed for unions */
2192 if (btf_is_union(t) && bit_offset)
2193 return libbpf_err(-EINVAL);
2194
2195 /* decompose and invalidate raw data */
2196 if (btf_ensure_modifiable(btf))
2197 return libbpf_err(-ENOMEM);
2198
2199 sz = sizeof(struct btf_member);
2200 m = btf_add_type_mem(btf, sz);
2201 if (!m)
2202 return libbpf_err(-ENOMEM);
2203
2204 if (name && name[0]) {
2205 name_off = btf__add_str(btf, name);
2206 if (name_off < 0)
2207 return name_off;
2208 }
2209
2210 m->name_off = name_off;
2211 m->type = type_id;
2212 m->offset = bit_offset | (bit_size << 24);
2213
2214 /* btf_add_type_mem can invalidate t pointer */
2215 t = btf_last_type(btf);
2216 /* update parent type's vlen and kflag */
2217 t->info = btf_type_info(btf_kind(t), btf_vlen(t) + 1, is_bitfield || btf_kflag(t));
2218
2219 btf->hdr->type_len += sz;
2220 btf->hdr->str_off += sz;
2221 return 0;
2222}
2223
2224static int btf_add_enum_common(struct btf *btf, const char *name, __u32 byte_sz,
2225 bool is_signed, __u8 kind)
2226{
2227 struct btf_type *t;
2228 int sz, name_off = 0;
2229
2230 /* byte_sz must be power of 2 */
2231 if (!byte_sz || (byte_sz & (byte_sz - 1)) || byte_sz > 8)
2232 return libbpf_err(-EINVAL);
2233
2234 if (btf_ensure_modifiable(btf))
2235 return libbpf_err(-ENOMEM);
2236
2237 sz = sizeof(struct btf_type);
2238 t = btf_add_type_mem(btf, sz);
2239 if (!t)
2240 return libbpf_err(-ENOMEM);
2241
2242 if (name && name[0]) {
2243 name_off = btf__add_str(btf, name);
2244 if (name_off < 0)
2245 return name_off;
2246 }
2247
2248 /* start out with vlen=0; it will be adjusted when adding enum values */
2249 t->name_off = name_off;
2250 t->info = btf_type_info(kind, 0, is_signed);
2251 t->size = byte_sz;
2252
2253 return btf_commit_type(btf, sz);
2254}
2255
2256/*
2257 * Append new BTF_KIND_ENUM type with:
2258 * - *name* - name of the enum, can be NULL or empty for anonymous enums;
2259 * - *byte_sz* - size of the enum, in bytes.
2260 *
2261 * Enum initially has no enum values in it (and corresponds to enum forward
2262 * declaration). Enumerator values can be added by btf__add_enum_value()
2263 * immediately after btf__add_enum() succeeds.
2264 *
2265 * Returns:
2266 * - >0, type ID of newly added BTF type;
2267 * - <0, on error.
2268 */
2269int btf__add_enum(struct btf *btf, const char *name, __u32 byte_sz)
2270{
2271 /*
2272 * set the signedness to be unsigned, it will change to signed
2273 * if any later enumerator is negative.
2274 */
2275 return btf_add_enum_common(btf, name, byte_sz, false, BTF_KIND_ENUM);
2276}
2277
2278/*
2279 * Append new enum value for the current ENUM type with:
2280 * - *name* - name of the enumerator value, can't be NULL or empty;
2281 * - *value* - integer value corresponding to enum value *name*;
2282 * Returns:
2283 * - 0, on success;
2284 * - <0, on error.
2285 */
2286int btf__add_enum_value(struct btf *btf, const char *name, __s64 value)
2287{
2288 struct btf_type *t;
2289 struct btf_enum *v;
2290 int sz, name_off;
2291
2292 /* last type should be BTF_KIND_ENUM */
2293 if (btf->nr_types == 0)
2294 return libbpf_err(-EINVAL);
2295 t = btf_last_type(btf);
2296 if (!btf_is_enum(t))
2297 return libbpf_err(-EINVAL);
2298
2299 /* non-empty name */
2300 if (!name || !name[0])
2301 return libbpf_err(-EINVAL);
2302 if (value < INT_MIN || value > UINT_MAX)
2303 return libbpf_err(-E2BIG);
2304
2305 /* decompose and invalidate raw data */
2306 if (btf_ensure_modifiable(btf))
2307 return libbpf_err(-ENOMEM);
2308
2309 sz = sizeof(struct btf_enum);
2310 v = btf_add_type_mem(btf, sz);
2311 if (!v)
2312 return libbpf_err(-ENOMEM);
2313
2314 name_off = btf__add_str(btf, name);
2315 if (name_off < 0)
2316 return name_off;
2317
2318 v->name_off = name_off;
2319 v->val = value;
2320
2321 /* update parent type's vlen */
2322 t = btf_last_type(btf);
2323 btf_type_inc_vlen(t);
2324
2325 /* if negative value, set signedness to signed */
2326 if (value < 0)
2327 t->info = btf_type_info(btf_kind(t), btf_vlen(t), true);
2328
2329 btf->hdr->type_len += sz;
2330 btf->hdr->str_off += sz;
2331 return 0;
2332}
2333
2334/*
2335 * Append new BTF_KIND_ENUM64 type with:
2336 * - *name* - name of the enum, can be NULL or empty for anonymous enums;
2337 * - *byte_sz* - size of the enum, in bytes.
2338 * - *is_signed* - whether the enum values are signed or not;
2339 *
2340 * Enum initially has no enum values in it (and corresponds to enum forward
2341 * declaration). Enumerator values can be added by btf__add_enum64_value()
2342 * immediately after btf__add_enum64() succeeds.
2343 *
2344 * Returns:
2345 * - >0, type ID of newly added BTF type;
2346 * - <0, on error.
2347 */
2348int btf__add_enum64(struct btf *btf, const char *name, __u32 byte_sz,
2349 bool is_signed)
2350{
2351 return btf_add_enum_common(btf, name, byte_sz, is_signed,
2352 BTF_KIND_ENUM64);
2353}
2354
2355/*
2356 * Append new enum value for the current ENUM64 type with:
2357 * - *name* - name of the enumerator value, can't be NULL or empty;
2358 * - *value* - integer value corresponding to enum value *name*;
2359 * Returns:
2360 * - 0, on success;
2361 * - <0, on error.
2362 */
2363int btf__add_enum64_value(struct btf *btf, const char *name, __u64 value)
2364{
2365 struct btf_enum64 *v;
2366 struct btf_type *t;
2367 int sz, name_off;
2368
2369 /* last type should be BTF_KIND_ENUM64 */
2370 if (btf->nr_types == 0)
2371 return libbpf_err(-EINVAL);
2372 t = btf_last_type(btf);
2373 if (!btf_is_enum64(t))
2374 return libbpf_err(-EINVAL);
2375
2376 /* non-empty name */
2377 if (!name || !name[0])
2378 return libbpf_err(-EINVAL);
2379
2380 /* decompose and invalidate raw data */
2381 if (btf_ensure_modifiable(btf))
2382 return libbpf_err(-ENOMEM);
2383
2384 sz = sizeof(struct btf_enum64);
2385 v = btf_add_type_mem(btf, sz);
2386 if (!v)
2387 return libbpf_err(-ENOMEM);
2388
2389 name_off = btf__add_str(btf, name);
2390 if (name_off < 0)
2391 return name_off;
2392
2393 v->name_off = name_off;
2394 v->val_lo32 = (__u32)value;
2395 v->val_hi32 = value >> 32;
2396
2397 /* update parent type's vlen */
2398 t = btf_last_type(btf);
2399 btf_type_inc_vlen(t);
2400
2401 btf->hdr->type_len += sz;
2402 btf->hdr->str_off += sz;
2403 return 0;
2404}
2405
2406/*
2407 * Append new BTF_KIND_FWD type with:
2408 * - *name*, non-empty/non-NULL name;
2409 * - *fwd_kind*, kind of forward declaration, one of BTF_FWD_STRUCT,
2410 * BTF_FWD_UNION, or BTF_FWD_ENUM;
2411 * Returns:
2412 * - >0, type ID of newly added BTF type;
2413 * - <0, on error.
2414 */
2415int btf__add_fwd(struct btf *btf, const char *name, enum btf_fwd_kind fwd_kind)
2416{
2417 if (!name || !name[0])
2418 return libbpf_err(-EINVAL);
2419
2420 switch (fwd_kind) {
2421 case BTF_FWD_STRUCT:
2422 case BTF_FWD_UNION: {
2423 struct btf_type *t;
2424 int id;
2425
2426 id = btf_add_ref_kind(btf, BTF_KIND_FWD, name, 0);
2427 if (id <= 0)
2428 return id;
2429 t = btf_type_by_id(btf, id);
2430 t->info = btf_type_info(BTF_KIND_FWD, 0, fwd_kind == BTF_FWD_UNION);
2431 return id;
2432 }
2433 case BTF_FWD_ENUM:
2434 /* enum forward in BTF currently is just an enum with no enum
2435 * values; we also assume a standard 4-byte size for it
2436 */
2437 return btf__add_enum(btf, name, sizeof(int));
2438 default:
2439 return libbpf_err(-EINVAL);
2440 }
2441}
2442
2443/*
2444 * Append new BTF_KING_TYPEDEF type with:
2445 * - *name*, non-empty/non-NULL name;
2446 * - *ref_type_id* - referenced type ID, it might not exist yet;
2447 * Returns:
2448 * - >0, type ID of newly added BTF type;
2449 * - <0, on error.
2450 */
2451int btf__add_typedef(struct btf *btf, const char *name, int ref_type_id)
2452{
2453 if (!name || !name[0])
2454 return libbpf_err(-EINVAL);
2455
2456 return btf_add_ref_kind(btf, BTF_KIND_TYPEDEF, name, ref_type_id);
2457}
2458
2459/*
2460 * Append new BTF_KIND_VOLATILE type with:
2461 * - *ref_type_id* - referenced type ID, it might not exist yet;
2462 * Returns:
2463 * - >0, type ID of newly added BTF type;
2464 * - <0, on error.
2465 */
2466int btf__add_volatile(struct btf *btf, int ref_type_id)
2467{
2468 return btf_add_ref_kind(btf, BTF_KIND_VOLATILE, NULL, ref_type_id);
2469}
2470
2471/*
2472 * Append new BTF_KIND_CONST type with:
2473 * - *ref_type_id* - referenced type ID, it might not exist yet;
2474 * Returns:
2475 * - >0, type ID of newly added BTF type;
2476 * - <0, on error.
2477 */
2478int btf__add_const(struct btf *btf, int ref_type_id)
2479{
2480 return btf_add_ref_kind(btf, BTF_KIND_CONST, NULL, ref_type_id);
2481}
2482
2483/*
2484 * Append new BTF_KIND_RESTRICT type with:
2485 * - *ref_type_id* - referenced type ID, it might not exist yet;
2486 * Returns:
2487 * - >0, type ID of newly added BTF type;
2488 * - <0, on error.
2489 */
2490int btf__add_restrict(struct btf *btf, int ref_type_id)
2491{
2492 return btf_add_ref_kind(btf, BTF_KIND_RESTRICT, NULL, ref_type_id);
2493}
2494
2495/*
2496 * Append new BTF_KIND_TYPE_TAG type with:
2497 * - *value*, non-empty/non-NULL tag value;
2498 * - *ref_type_id* - referenced type ID, it might not exist yet;
2499 * Returns:
2500 * - >0, type ID of newly added BTF type;
2501 * - <0, on error.
2502 */
2503int btf__add_type_tag(struct btf *btf, const char *value, int ref_type_id)
2504{
2505 if (!value || !value[0])
2506 return libbpf_err(-EINVAL);
2507
2508 return btf_add_ref_kind(btf, BTF_KIND_TYPE_TAG, value, ref_type_id);
2509}
2510
2511/*
2512 * Append new BTF_KIND_FUNC type with:
2513 * - *name*, non-empty/non-NULL name;
2514 * - *proto_type_id* - FUNC_PROTO's type ID, it might not exist yet;
2515 * Returns:
2516 * - >0, type ID of newly added BTF type;
2517 * - <0, on error.
2518 */
2519int btf__add_func(struct btf *btf, const char *name,
2520 enum btf_func_linkage linkage, int proto_type_id)
2521{
2522 int id;
2523
2524 if (!name || !name[0])
2525 return libbpf_err(-EINVAL);
2526 if (linkage != BTF_FUNC_STATIC && linkage != BTF_FUNC_GLOBAL &&
2527 linkage != BTF_FUNC_EXTERN)
2528 return libbpf_err(-EINVAL);
2529
2530 id = btf_add_ref_kind(btf, BTF_KIND_FUNC, name, proto_type_id);
2531 if (id > 0) {
2532 struct btf_type *t = btf_type_by_id(btf, id);
2533
2534 t->info = btf_type_info(BTF_KIND_FUNC, linkage, 0);
2535 }
2536 return libbpf_err(id);
2537}
2538
2539/*
2540 * Append new BTF_KIND_FUNC_PROTO with:
2541 * - *ret_type_id* - type ID for return result of a function.
2542 *
2543 * Function prototype initially has no arguments, but they can be added by
2544 * btf__add_func_param() one by one, immediately after
2545 * btf__add_func_proto() succeeded.
2546 *
2547 * Returns:
2548 * - >0, type ID of newly added BTF type;
2549 * - <0, on error.
2550 */
2551int btf__add_func_proto(struct btf *btf, int ret_type_id)
2552{
2553 struct btf_type *t;
2554 int sz;
2555
2556 if (validate_type_id(ret_type_id))
2557 return libbpf_err(-EINVAL);
2558
2559 if (btf_ensure_modifiable(btf))
2560 return libbpf_err(-ENOMEM);
2561
2562 sz = sizeof(struct btf_type);
2563 t = btf_add_type_mem(btf, sz);
2564 if (!t)
2565 return libbpf_err(-ENOMEM);
2566
2567 /* start out with vlen=0; this will be adjusted when adding enum
2568 * values, if necessary
2569 */
2570 t->name_off = 0;
2571 t->info = btf_type_info(BTF_KIND_FUNC_PROTO, 0, 0);
2572 t->type = ret_type_id;
2573
2574 return btf_commit_type(btf, sz);
2575}
2576
2577/*
2578 * Append new function parameter for current FUNC_PROTO type with:
2579 * - *name* - parameter name, can be NULL or empty;
2580 * - *type_id* - type ID describing the type of the parameter.
2581 * Returns:
2582 * - 0, on success;
2583 * - <0, on error.
2584 */
2585int btf__add_func_param(struct btf *btf, const char *name, int type_id)
2586{
2587 struct btf_type *t;
2588 struct btf_param *p;
2589 int sz, name_off = 0;
2590
2591 if (validate_type_id(type_id))
2592 return libbpf_err(-EINVAL);
2593
2594 /* last type should be BTF_KIND_FUNC_PROTO */
2595 if (btf->nr_types == 0)
2596 return libbpf_err(-EINVAL);
2597 t = btf_last_type(btf);
2598 if (!btf_is_func_proto(t))
2599 return libbpf_err(-EINVAL);
2600
2601 /* decompose and invalidate raw data */
2602 if (btf_ensure_modifiable(btf))
2603 return libbpf_err(-ENOMEM);
2604
2605 sz = sizeof(struct btf_param);
2606 p = btf_add_type_mem(btf, sz);
2607 if (!p)
2608 return libbpf_err(-ENOMEM);
2609
2610 if (name && name[0]) {
2611 name_off = btf__add_str(btf, name);
2612 if (name_off < 0)
2613 return name_off;
2614 }
2615
2616 p->name_off = name_off;
2617 p->type = type_id;
2618
2619 /* update parent type's vlen */
2620 t = btf_last_type(btf);
2621 btf_type_inc_vlen(t);
2622
2623 btf->hdr->type_len += sz;
2624 btf->hdr->str_off += sz;
2625 return 0;
2626}
2627
2628/*
2629 * Append new BTF_KIND_VAR type with:
2630 * - *name* - non-empty/non-NULL name;
2631 * - *linkage* - variable linkage, one of BTF_VAR_STATIC,
2632 * BTF_VAR_GLOBAL_ALLOCATED, or BTF_VAR_GLOBAL_EXTERN;
2633 * - *type_id* - type ID of the type describing the type of the variable.
2634 * Returns:
2635 * - >0, type ID of newly added BTF type;
2636 * - <0, on error.
2637 */
2638int btf__add_var(struct btf *btf, const char *name, int linkage, int type_id)
2639{
2640 struct btf_type *t;
2641 struct btf_var *v;
2642 int sz, name_off;
2643
2644 /* non-empty name */
2645 if (!name || !name[0])
2646 return libbpf_err(-EINVAL);
2647 if (linkage != BTF_VAR_STATIC && linkage != BTF_VAR_GLOBAL_ALLOCATED &&
2648 linkage != BTF_VAR_GLOBAL_EXTERN)
2649 return libbpf_err(-EINVAL);
2650 if (validate_type_id(type_id))
2651 return libbpf_err(-EINVAL);
2652
2653 /* deconstruct BTF, if necessary, and invalidate raw_data */
2654 if (btf_ensure_modifiable(btf))
2655 return libbpf_err(-ENOMEM);
2656
2657 sz = sizeof(struct btf_type) + sizeof(struct btf_var);
2658 t = btf_add_type_mem(btf, sz);
2659 if (!t)
2660 return libbpf_err(-ENOMEM);
2661
2662 name_off = btf__add_str(btf, name);
2663 if (name_off < 0)
2664 return name_off;
2665
2666 t->name_off = name_off;
2667 t->info = btf_type_info(BTF_KIND_VAR, 0, 0);
2668 t->type = type_id;
2669
2670 v = btf_var(t);
2671 v->linkage = linkage;
2672
2673 return btf_commit_type(btf, sz);
2674}
2675
2676/*
2677 * Append new BTF_KIND_DATASEC type with:
2678 * - *name* - non-empty/non-NULL name;
2679 * - *byte_sz* - data section size, in bytes.
2680 *
2681 * Data section is initially empty. Variables info can be added with
2682 * btf__add_datasec_var_info() calls, after btf__add_datasec() succeeds.
2683 *
2684 * Returns:
2685 * - >0, type ID of newly added BTF type;
2686 * - <0, on error.
2687 */
2688int btf__add_datasec(struct btf *btf, const char *name, __u32 byte_sz)
2689{
2690 struct btf_type *t;
2691 int sz, name_off;
2692
2693 /* non-empty name */
2694 if (!name || !name[0])
2695 return libbpf_err(-EINVAL);
2696
2697 if (btf_ensure_modifiable(btf))
2698 return libbpf_err(-ENOMEM);
2699
2700 sz = sizeof(struct btf_type);
2701 t = btf_add_type_mem(btf, sz);
2702 if (!t)
2703 return libbpf_err(-ENOMEM);
2704
2705 name_off = btf__add_str(btf, name);
2706 if (name_off < 0)
2707 return name_off;
2708
2709 /* start with vlen=0, which will be update as var_secinfos are added */
2710 t->name_off = name_off;
2711 t->info = btf_type_info(BTF_KIND_DATASEC, 0, 0);
2712 t->size = byte_sz;
2713
2714 return btf_commit_type(btf, sz);
2715}
2716
2717/*
2718 * Append new data section variable information entry for current DATASEC type:
2719 * - *var_type_id* - type ID, describing type of the variable;
2720 * - *offset* - variable offset within data section, in bytes;
2721 * - *byte_sz* - variable size, in bytes.
2722 *
2723 * Returns:
2724 * - 0, on success;
2725 * - <0, on error.
2726 */
2727int btf__add_datasec_var_info(struct btf *btf, int var_type_id, __u32 offset, __u32 byte_sz)
2728{
2729 struct btf_type *t;
2730 struct btf_var_secinfo *v;
2731 int sz;
2732
2733 /* last type should be BTF_KIND_DATASEC */
2734 if (btf->nr_types == 0)
2735 return libbpf_err(-EINVAL);
2736 t = btf_last_type(btf);
2737 if (!btf_is_datasec(t))
2738 return libbpf_err(-EINVAL);
2739
2740 if (validate_type_id(var_type_id))
2741 return libbpf_err(-EINVAL);
2742
2743 /* decompose and invalidate raw data */
2744 if (btf_ensure_modifiable(btf))
2745 return libbpf_err(-ENOMEM);
2746
2747 sz = sizeof(struct btf_var_secinfo);
2748 v = btf_add_type_mem(btf, sz);
2749 if (!v)
2750 return libbpf_err(-ENOMEM);
2751
2752 v->type = var_type_id;
2753 v->offset = offset;
2754 v->size = byte_sz;
2755
2756 /* update parent type's vlen */
2757 t = btf_last_type(btf);
2758 btf_type_inc_vlen(t);
2759
2760 btf->hdr->type_len += sz;
2761 btf->hdr->str_off += sz;
2762 return 0;
2763}
2764
2765/*
2766 * Append new BTF_KIND_DECL_TAG type with:
2767 * - *value* - non-empty/non-NULL string;
2768 * - *ref_type_id* - referenced type ID, it might not exist yet;
2769 * - *component_idx* - -1 for tagging reference type, otherwise struct/union
2770 * member or function argument index;
2771 * Returns:
2772 * - >0, type ID of newly added BTF type;
2773 * - <0, on error.
2774 */
2775int btf__add_decl_tag(struct btf *btf, const char *value, int ref_type_id,
2776 int component_idx)
2777{
2778 struct btf_type *t;
2779 int sz, value_off;
2780
2781 if (!value || !value[0] || component_idx < -1)
2782 return libbpf_err(-EINVAL);
2783
2784 if (validate_type_id(ref_type_id))
2785 return libbpf_err(-EINVAL);
2786
2787 if (btf_ensure_modifiable(btf))
2788 return libbpf_err(-ENOMEM);
2789
2790 sz = sizeof(struct btf_type) + sizeof(struct btf_decl_tag);
2791 t = btf_add_type_mem(btf, sz);
2792 if (!t)
2793 return libbpf_err(-ENOMEM);
2794
2795 value_off = btf__add_str(btf, value);
2796 if (value_off < 0)
2797 return value_off;
2798
2799 t->name_off = value_off;
2800 t->info = btf_type_info(BTF_KIND_DECL_TAG, 0, false);
2801 t->type = ref_type_id;
2802 btf_decl_tag(t)->component_idx = component_idx;
2803
2804 return btf_commit_type(btf, sz);
2805}
2806
2807struct btf_ext_sec_setup_param {
2808 __u32 off;
2809 __u32 len;
2810 __u32 min_rec_size;
2811 struct btf_ext_info *ext_info;
2812 const char *desc;
2813};
2814
2815static int btf_ext_setup_info(struct btf_ext *btf_ext,
2816 struct btf_ext_sec_setup_param *ext_sec)
2817{
2818 const struct btf_ext_info_sec *sinfo;
2819 struct btf_ext_info *ext_info;
2820 __u32 info_left, record_size;
2821 size_t sec_cnt = 0;
2822 /* The start of the info sec (including the __u32 record_size). */
2823 void *info;
2824
2825 if (ext_sec->len == 0)
2826 return 0;
2827
2828 if (ext_sec->off & 0x03) {
2829 pr_debug(".BTF.ext %s section is not aligned to 4 bytes\n",
2830 ext_sec->desc);
2831 return -EINVAL;
2832 }
2833
2834 info = btf_ext->data + btf_ext->hdr->hdr_len + ext_sec->off;
2835 info_left = ext_sec->len;
2836
2837 if (btf_ext->data + btf_ext->data_size < info + ext_sec->len) {
2838 pr_debug("%s section (off:%u len:%u) is beyond the end of the ELF section .BTF.ext\n",
2839 ext_sec->desc, ext_sec->off, ext_sec->len);
2840 return -EINVAL;
2841 }
2842
2843 /* At least a record size */
2844 if (info_left < sizeof(__u32)) {
2845 pr_debug(".BTF.ext %s record size not found\n", ext_sec->desc);
2846 return -EINVAL;
2847 }
2848
2849 /* The record size needs to meet the minimum standard */
2850 record_size = *(__u32 *)info;
2851 if (record_size < ext_sec->min_rec_size ||
2852 record_size & 0x03) {
2853 pr_debug("%s section in .BTF.ext has invalid record size %u\n",
2854 ext_sec->desc, record_size);
2855 return -EINVAL;
2856 }
2857
2858 sinfo = info + sizeof(__u32);
2859 info_left -= sizeof(__u32);
2860
2861 /* If no records, return failure now so .BTF.ext won't be used. */
2862 if (!info_left) {
2863 pr_debug("%s section in .BTF.ext has no records", ext_sec->desc);
2864 return -EINVAL;
2865 }
2866
2867 while (info_left) {
2868 unsigned int sec_hdrlen = sizeof(struct btf_ext_info_sec);
2869 __u64 total_record_size;
2870 __u32 num_records;
2871
2872 if (info_left < sec_hdrlen) {
2873 pr_debug("%s section header is not found in .BTF.ext\n",
2874 ext_sec->desc);
2875 return -EINVAL;
2876 }
2877
2878 num_records = sinfo->num_info;
2879 if (num_records == 0) {
2880 pr_debug("%s section has incorrect num_records in .BTF.ext\n",
2881 ext_sec->desc);
2882 return -EINVAL;
2883 }
2884
2885 total_record_size = sec_hdrlen + (__u64)num_records * record_size;
2886 if (info_left < total_record_size) {
2887 pr_debug("%s section has incorrect num_records in .BTF.ext\n",
2888 ext_sec->desc);
2889 return -EINVAL;
2890 }
2891
2892 info_left -= total_record_size;
2893 sinfo = (void *)sinfo + total_record_size;
2894 sec_cnt++;
2895 }
2896
2897 ext_info = ext_sec->ext_info;
2898 ext_info->len = ext_sec->len - sizeof(__u32);
2899 ext_info->rec_size = record_size;
2900 ext_info->info = info + sizeof(__u32);
2901 ext_info->sec_cnt = sec_cnt;
2902
2903 return 0;
2904}
2905
2906static int btf_ext_setup_func_info(struct btf_ext *btf_ext)
2907{
2908 struct btf_ext_sec_setup_param param = {
2909 .off = btf_ext->hdr->func_info_off,
2910 .len = btf_ext->hdr->func_info_len,
2911 .min_rec_size = sizeof(struct bpf_func_info_min),
2912 .ext_info = &btf_ext->func_info,
2913 .desc = "func_info"
2914 };
2915
2916 return btf_ext_setup_info(btf_ext, ¶m);
2917}
2918
2919static int btf_ext_setup_line_info(struct btf_ext *btf_ext)
2920{
2921 struct btf_ext_sec_setup_param param = {
2922 .off = btf_ext->hdr->line_info_off,
2923 .len = btf_ext->hdr->line_info_len,
2924 .min_rec_size = sizeof(struct bpf_line_info_min),
2925 .ext_info = &btf_ext->line_info,
2926 .desc = "line_info",
2927 };
2928
2929 return btf_ext_setup_info(btf_ext, ¶m);
2930}
2931
2932static int btf_ext_setup_core_relos(struct btf_ext *btf_ext)
2933{
2934 struct btf_ext_sec_setup_param param = {
2935 .off = btf_ext->hdr->core_relo_off,
2936 .len = btf_ext->hdr->core_relo_len,
2937 .min_rec_size = sizeof(struct bpf_core_relo),
2938 .ext_info = &btf_ext->core_relo_info,
2939 .desc = "core_relo",
2940 };
2941
2942 return btf_ext_setup_info(btf_ext, ¶m);
2943}
2944
2945static int btf_ext_parse_hdr(__u8 *data, __u32 data_size)
2946{
2947 const struct btf_ext_header *hdr = (struct btf_ext_header *)data;
2948
2949 if (data_size < offsetofend(struct btf_ext_header, hdr_len) ||
2950 data_size < hdr->hdr_len) {
2951 pr_debug("BTF.ext header not found");
2952 return -EINVAL;
2953 }
2954
2955 if (hdr->magic == bswap_16(BTF_MAGIC)) {
2956 pr_warn("BTF.ext in non-native endianness is not supported\n");
2957 return -ENOTSUP;
2958 } else if (hdr->magic != BTF_MAGIC) {
2959 pr_debug("Invalid BTF.ext magic:%x\n", hdr->magic);
2960 return -EINVAL;
2961 }
2962
2963 if (hdr->version != BTF_VERSION) {
2964 pr_debug("Unsupported BTF.ext version:%u\n", hdr->version);
2965 return -ENOTSUP;
2966 }
2967
2968 if (hdr->flags) {
2969 pr_debug("Unsupported BTF.ext flags:%x\n", hdr->flags);
2970 return -ENOTSUP;
2971 }
2972
2973 if (data_size == hdr->hdr_len) {
2974 pr_debug("BTF.ext has no data\n");
2975 return -EINVAL;
2976 }
2977
2978 return 0;
2979}
2980
2981void btf_ext__free(struct btf_ext *btf_ext)
2982{
2983 if (IS_ERR_OR_NULL(btf_ext))
2984 return;
2985 free(btf_ext->func_info.sec_idxs);
2986 free(btf_ext->line_info.sec_idxs);
2987 free(btf_ext->core_relo_info.sec_idxs);
2988 free(btf_ext->data);
2989 free(btf_ext);
2990}
2991
2992struct btf_ext *btf_ext__new(const __u8 *data, __u32 size)
2993{
2994 struct btf_ext *btf_ext;
2995 int err;
2996
2997 btf_ext = calloc(1, sizeof(struct btf_ext));
2998 if (!btf_ext)
2999 return libbpf_err_ptr(-ENOMEM);
3000
3001 btf_ext->data_size = size;
3002 btf_ext->data = malloc(size);
3003 if (!btf_ext->data) {
3004 err = -ENOMEM;
3005 goto done;
3006 }
3007 memcpy(btf_ext->data, data, size);
3008
3009 err = btf_ext_parse_hdr(btf_ext->data, size);
3010 if (err)
3011 goto done;
3012
3013 if (btf_ext->hdr->hdr_len < offsetofend(struct btf_ext_header, line_info_len)) {
3014 err = -EINVAL;
3015 goto done;
3016 }
3017
3018 err = btf_ext_setup_func_info(btf_ext);
3019 if (err)
3020 goto done;
3021
3022 err = btf_ext_setup_line_info(btf_ext);
3023 if (err)
3024 goto done;
3025
3026 if (btf_ext->hdr->hdr_len < offsetofend(struct btf_ext_header, core_relo_len))
3027 goto done; /* skip core relos parsing */
3028
3029 err = btf_ext_setup_core_relos(btf_ext);
3030 if (err)
3031 goto done;
3032
3033done:
3034 if (err) {
3035 btf_ext__free(btf_ext);
3036 return libbpf_err_ptr(err);
3037 }
3038
3039 return btf_ext;
3040}
3041
3042const void *btf_ext__get_raw_data(const struct btf_ext *btf_ext, __u32 *size)
3043{
3044 *size = btf_ext->data_size;
3045 return btf_ext->data;
3046}
3047
3048struct btf_dedup;
3049
3050static struct btf_dedup *btf_dedup_new(struct btf *btf, const struct btf_dedup_opts *opts);
3051static void btf_dedup_free(struct btf_dedup *d);
3052static int btf_dedup_prep(struct btf_dedup *d);
3053static int btf_dedup_strings(struct btf_dedup *d);
3054static int btf_dedup_prim_types(struct btf_dedup *d);
3055static int btf_dedup_struct_types(struct btf_dedup *d);
3056static int btf_dedup_ref_types(struct btf_dedup *d);
3057static int btf_dedup_resolve_fwds(struct btf_dedup *d);
3058static int btf_dedup_compact_types(struct btf_dedup *d);
3059static int btf_dedup_remap_types(struct btf_dedup *d);
3060
3061/*
3062 * Deduplicate BTF types and strings.
3063 *
3064 * BTF dedup algorithm takes as an input `struct btf` representing `.BTF` ELF
3065 * section with all BTF type descriptors and string data. It overwrites that
3066 * memory in-place with deduplicated types and strings without any loss of
3067 * information. If optional `struct btf_ext` representing '.BTF.ext' ELF section
3068 * is provided, all the strings referenced from .BTF.ext section are honored
3069 * and updated to point to the right offsets after deduplication.
3070 *
3071 * If function returns with error, type/string data might be garbled and should
3072 * be discarded.
3073 *
3074 * More verbose and detailed description of both problem btf_dedup is solving,
3075 * as well as solution could be found at:
3076 * https://facebookmicrosites.github.io/bpf/blog/2018/11/14/btf-enhancement.html
3077 *
3078 * Problem description and justification
3079 * =====================================
3080 *
3081 * BTF type information is typically emitted either as a result of conversion
3082 * from DWARF to BTF or directly by compiler. In both cases, each compilation
3083 * unit contains information about a subset of all the types that are used
3084 * in an application. These subsets are frequently overlapping and contain a lot
3085 * of duplicated information when later concatenated together into a single
3086 * binary. This algorithm ensures that each unique type is represented by single
3087 * BTF type descriptor, greatly reducing resulting size of BTF data.
3088 *
3089 * Compilation unit isolation and subsequent duplication of data is not the only
3090 * problem. The same type hierarchy (e.g., struct and all the type that struct
3091 * references) in different compilation units can be represented in BTF to
3092 * various degrees of completeness (or, rather, incompleteness) due to
3093 * struct/union forward declarations.
3094 *
3095 * Let's take a look at an example, that we'll use to better understand the
3096 * problem (and solution). Suppose we have two compilation units, each using
3097 * same `struct S`, but each of them having incomplete type information about
3098 * struct's fields:
3099 *
3100 * // CU #1:
3101 * struct S;
3102 * struct A {
3103 * int a;
3104 * struct A* self;
3105 * struct S* parent;
3106 * };
3107 * struct B;
3108 * struct S {
3109 * struct A* a_ptr;
3110 * struct B* b_ptr;
3111 * };
3112 *
3113 * // CU #2:
3114 * struct S;
3115 * struct A;
3116 * struct B {
3117 * int b;
3118 * struct B* self;
3119 * struct S* parent;
3120 * };
3121 * struct S {
3122 * struct A* a_ptr;
3123 * struct B* b_ptr;
3124 * };
3125 *
3126 * In case of CU #1, BTF data will know only that `struct B` exist (but no
3127 * more), but will know the complete type information about `struct A`. While
3128 * for CU #2, it will know full type information about `struct B`, but will
3129 * only know about forward declaration of `struct A` (in BTF terms, it will
3130 * have `BTF_KIND_FWD` type descriptor with name `B`).
3131 *
3132 * This compilation unit isolation means that it's possible that there is no
3133 * single CU with complete type information describing structs `S`, `A`, and
3134 * `B`. Also, we might get tons of duplicated and redundant type information.
3135 *
3136 * Additional complication we need to keep in mind comes from the fact that
3137 * types, in general, can form graphs containing cycles, not just DAGs.
3138 *
3139 * While algorithm does deduplication, it also merges and resolves type
3140 * information (unless disabled throught `struct btf_opts`), whenever possible.
3141 * E.g., in the example above with two compilation units having partial type
3142 * information for structs `A` and `B`, the output of algorithm will emit
3143 * a single copy of each BTF type that describes structs `A`, `B`, and `S`
3144 * (as well as type information for `int` and pointers), as if they were defined
3145 * in a single compilation unit as:
3146 *
3147 * struct A {
3148 * int a;
3149 * struct A* self;
3150 * struct S* parent;
3151 * };
3152 * struct B {
3153 * int b;
3154 * struct B* self;
3155 * struct S* parent;
3156 * };
3157 * struct S {
3158 * struct A* a_ptr;
3159 * struct B* b_ptr;
3160 * };
3161 *
3162 * Algorithm summary
3163 * =================
3164 *
3165 * Algorithm completes its work in 7 separate passes:
3166 *
3167 * 1. Strings deduplication.
3168 * 2. Primitive types deduplication (int, enum, fwd).
3169 * 3. Struct/union types deduplication.
3170 * 4. Resolve unambiguous forward declarations.
3171 * 5. Reference types deduplication (pointers, typedefs, arrays, funcs, func
3172 * protos, and const/volatile/restrict modifiers).
3173 * 6. Types compaction.
3174 * 7. Types remapping.
3175 *
3176 * Algorithm determines canonical type descriptor, which is a single
3177 * representative type for each truly unique type. This canonical type is the
3178 * one that will go into final deduplicated BTF type information. For
3179 * struct/unions, it is also the type that algorithm will merge additional type
3180 * information into (while resolving FWDs), as it discovers it from data in
3181 * other CUs. Each input BTF type eventually gets either mapped to itself, if
3182 * that type is canonical, or to some other type, if that type is equivalent
3183 * and was chosen as canonical representative. This mapping is stored in
3184 * `btf_dedup->map` array. This map is also used to record STRUCT/UNION that
3185 * FWD type got resolved to.
3186 *
3187 * To facilitate fast discovery of canonical types, we also maintain canonical
3188 * index (`btf_dedup->dedup_table`), which maps type descriptor's signature hash
3189 * (i.e., hashed kind, name, size, fields, etc) into a list of canonical types
3190 * that match that signature. With sufficiently good choice of type signature
3191 * hashing function, we can limit number of canonical types for each unique type
3192 * signature to a very small number, allowing to find canonical type for any
3193 * duplicated type very quickly.
3194 *
3195 * Struct/union deduplication is the most critical part and algorithm for
3196 * deduplicating structs/unions is described in greater details in comments for
3197 * `btf_dedup_is_equiv` function.
3198 */
3199int btf__dedup(struct btf *btf, const struct btf_dedup_opts *opts)
3200{
3201 struct btf_dedup *d;
3202 int err;
3203
3204 if (!OPTS_VALID(opts, btf_dedup_opts))
3205 return libbpf_err(-EINVAL);
3206
3207 d = btf_dedup_new(btf, opts);
3208 if (IS_ERR(d)) {
3209 pr_debug("btf_dedup_new failed: %ld", PTR_ERR(d));
3210 return libbpf_err(-EINVAL);
3211 }
3212
3213 if (btf_ensure_modifiable(btf)) {
3214 err = -ENOMEM;
3215 goto done;
3216 }
3217
3218 err = btf_dedup_prep(d);
3219 if (err) {
3220 pr_debug("btf_dedup_prep failed:%d\n", err);
3221 goto done;
3222 }
3223 err = btf_dedup_strings(d);
3224 if (err < 0) {
3225 pr_debug("btf_dedup_strings failed:%d\n", err);
3226 goto done;
3227 }
3228 err = btf_dedup_prim_types(d);
3229 if (err < 0) {
3230 pr_debug("btf_dedup_prim_types failed:%d\n", err);
3231 goto done;
3232 }
3233 err = btf_dedup_struct_types(d);
3234 if (err < 0) {
3235 pr_debug("btf_dedup_struct_types failed:%d\n", err);
3236 goto done;
3237 }
3238 err = btf_dedup_resolve_fwds(d);
3239 if (err < 0) {
3240 pr_debug("btf_dedup_resolve_fwds failed:%d\n", err);
3241 goto done;
3242 }
3243 err = btf_dedup_ref_types(d);
3244 if (err < 0) {
3245 pr_debug("btf_dedup_ref_types failed:%d\n", err);
3246 goto done;
3247 }
3248 err = btf_dedup_compact_types(d);
3249 if (err < 0) {
3250 pr_debug("btf_dedup_compact_types failed:%d\n", err);
3251 goto done;
3252 }
3253 err = btf_dedup_remap_types(d);
3254 if (err < 0) {
3255 pr_debug("btf_dedup_remap_types failed:%d\n", err);
3256 goto done;
3257 }
3258
3259done:
3260 btf_dedup_free(d);
3261 return libbpf_err(err);
3262}
3263
3264#define BTF_UNPROCESSED_ID ((__u32)-1)
3265#define BTF_IN_PROGRESS_ID ((__u32)-2)
3266
3267struct btf_dedup {
3268 /* .BTF section to be deduped in-place */
3269 struct btf *btf;
3270 /*
3271 * Optional .BTF.ext section. When provided, any strings referenced
3272 * from it will be taken into account when deduping strings
3273 */
3274 struct btf_ext *btf_ext;
3275 /*
3276 * This is a map from any type's signature hash to a list of possible
3277 * canonical representative type candidates. Hash collisions are
3278 * ignored, so even types of various kinds can share same list of
3279 * candidates, which is fine because we rely on subsequent
3280 * btf_xxx_equal() checks to authoritatively verify type equality.
3281 */
3282 struct hashmap *dedup_table;
3283 /* Canonical types map */
3284 __u32 *map;
3285 /* Hypothetical mapping, used during type graph equivalence checks */
3286 __u32 *hypot_map;
3287 __u32 *hypot_list;
3288 size_t hypot_cnt;
3289 size_t hypot_cap;
3290 /* Whether hypothetical mapping, if successful, would need to adjust
3291 * already canonicalized types (due to a new forward declaration to
3292 * concrete type resolution). In such case, during split BTF dedup
3293 * candidate type would still be considered as different, because base
3294 * BTF is considered to be immutable.
3295 */
3296 bool hypot_adjust_canon;
3297 /* Various option modifying behavior of algorithm */
3298 struct btf_dedup_opts opts;
3299 /* temporary strings deduplication state */
3300 struct strset *strs_set;
3301};
3302
3303static long hash_combine(long h, long value)
3304{
3305 return h * 31 + value;
3306}
3307
3308#define for_each_dedup_cand(d, node, hash) \
3309 hashmap__for_each_key_entry(d->dedup_table, node, hash)
3310
3311static int btf_dedup_table_add(struct btf_dedup *d, long hash, __u32 type_id)
3312{
3313 return hashmap__append(d->dedup_table, hash, type_id);
3314}
3315
3316static int btf_dedup_hypot_map_add(struct btf_dedup *d,
3317 __u32 from_id, __u32 to_id)
3318{
3319 if (d->hypot_cnt == d->hypot_cap) {
3320 __u32 *new_list;
3321
3322 d->hypot_cap += max((size_t)16, d->hypot_cap / 2);
3323 new_list = libbpf_reallocarray(d->hypot_list, d->hypot_cap, sizeof(__u32));
3324 if (!new_list)
3325 return -ENOMEM;
3326 d->hypot_list = new_list;
3327 }
3328 d->hypot_list[d->hypot_cnt++] = from_id;
3329 d->hypot_map[from_id] = to_id;
3330 return 0;
3331}
3332
3333static void btf_dedup_clear_hypot_map(struct btf_dedup *d)
3334{
3335 int i;
3336
3337 for (i = 0; i < d->hypot_cnt; i++)
3338 d->hypot_map[d->hypot_list[i]] = BTF_UNPROCESSED_ID;
3339 d->hypot_cnt = 0;
3340 d->hypot_adjust_canon = false;
3341}
3342
3343static void btf_dedup_free(struct btf_dedup *d)
3344{
3345 hashmap__free(d->dedup_table);
3346 d->dedup_table = NULL;
3347
3348 free(d->map);
3349 d->map = NULL;
3350
3351 free(d->hypot_map);
3352 d->hypot_map = NULL;
3353
3354 free(d->hypot_list);
3355 d->hypot_list = NULL;
3356
3357 free(d);
3358}
3359
3360static size_t btf_dedup_identity_hash_fn(long key, void *ctx)
3361{
3362 return key;
3363}
3364
3365static size_t btf_dedup_collision_hash_fn(long key, void *ctx)
3366{
3367 return 0;
3368}
3369
3370static bool btf_dedup_equal_fn(long k1, long k2, void *ctx)
3371{
3372 return k1 == k2;
3373}
3374
3375static struct btf_dedup *btf_dedup_new(struct btf *btf, const struct btf_dedup_opts *opts)
3376{
3377 struct btf_dedup *d = calloc(1, sizeof(struct btf_dedup));
3378 hashmap_hash_fn hash_fn = btf_dedup_identity_hash_fn;
3379 int i, err = 0, type_cnt;
3380
3381 if (!d)
3382 return ERR_PTR(-ENOMEM);
3383
3384 if (OPTS_GET(opts, force_collisions, false))
3385 hash_fn = btf_dedup_collision_hash_fn;
3386
3387 d->btf = btf;
3388 d->btf_ext = OPTS_GET(opts, btf_ext, NULL);
3389
3390 d->dedup_table = hashmap__new(hash_fn, btf_dedup_equal_fn, NULL);
3391 if (IS_ERR(d->dedup_table)) {
3392 err = PTR_ERR(d->dedup_table);
3393 d->dedup_table = NULL;
3394 goto done;
3395 }
3396
3397 type_cnt = btf__type_cnt(btf);
3398 d->map = malloc(sizeof(__u32) * type_cnt);
3399 if (!d->map) {
3400 err = -ENOMEM;
3401 goto done;
3402 }
3403 /* special BTF "void" type is made canonical immediately */
3404 d->map[0] = 0;
3405 for (i = 1; i < type_cnt; i++) {
3406 struct btf_type *t = btf_type_by_id(d->btf, i);
3407
3408 /* VAR and DATASEC are never deduped and are self-canonical */
3409 if (btf_is_var(t) || btf_is_datasec(t))
3410 d->map[i] = i;
3411 else
3412 d->map[i] = BTF_UNPROCESSED_ID;
3413 }
3414
3415 d->hypot_map = malloc(sizeof(__u32) * type_cnt);
3416 if (!d->hypot_map) {
3417 err = -ENOMEM;
3418 goto done;
3419 }
3420 for (i = 0; i < type_cnt; i++)
3421 d->hypot_map[i] = BTF_UNPROCESSED_ID;
3422
3423done:
3424 if (err) {
3425 btf_dedup_free(d);
3426 return ERR_PTR(err);
3427 }
3428
3429 return d;
3430}
3431
3432/*
3433 * Iterate over all possible places in .BTF and .BTF.ext that can reference
3434 * string and pass pointer to it to a provided callback `fn`.
3435 */
3436static int btf_for_each_str_off(struct btf_dedup *d, str_off_visit_fn fn, void *ctx)
3437{
3438 int i, r;
3439
3440 for (i = 0; i < d->btf->nr_types; i++) {
3441 struct btf_type *t = btf_type_by_id(d->btf, d->btf->start_id + i);
3442
3443 r = btf_type_visit_str_offs(t, fn, ctx);
3444 if (r)
3445 return r;
3446 }
3447
3448 if (!d->btf_ext)
3449 return 0;
3450
3451 r = btf_ext_visit_str_offs(d->btf_ext, fn, ctx);
3452 if (r)
3453 return r;
3454
3455 return 0;
3456}
3457
3458static int strs_dedup_remap_str_off(__u32 *str_off_ptr, void *ctx)
3459{
3460 struct btf_dedup *d = ctx;
3461 __u32 str_off = *str_off_ptr;
3462 const char *s;
3463 int off, err;
3464
3465 /* don't touch empty string or string in main BTF */
3466 if (str_off == 0 || str_off < d->btf->start_str_off)
3467 return 0;
3468
3469 s = btf__str_by_offset(d->btf, str_off);
3470 if (d->btf->base_btf) {
3471 err = btf__find_str(d->btf->base_btf, s);
3472 if (err >= 0) {
3473 *str_off_ptr = err;
3474 return 0;
3475 }
3476 if (err != -ENOENT)
3477 return err;
3478 }
3479
3480 off = strset__add_str(d->strs_set, s);
3481 if (off < 0)
3482 return off;
3483
3484 *str_off_ptr = d->btf->start_str_off + off;
3485 return 0;
3486}
3487
3488/*
3489 * Dedup string and filter out those that are not referenced from either .BTF
3490 * or .BTF.ext (if provided) sections.
3491 *
3492 * This is done by building index of all strings in BTF's string section,
3493 * then iterating over all entities that can reference strings (e.g., type
3494 * names, struct field names, .BTF.ext line info, etc) and marking corresponding
3495 * strings as used. After that all used strings are deduped and compacted into
3496 * sequential blob of memory and new offsets are calculated. Then all the string
3497 * references are iterated again and rewritten using new offsets.
3498 */
3499static int btf_dedup_strings(struct btf_dedup *d)
3500{
3501 int err;
3502
3503 if (d->btf->strs_deduped)
3504 return 0;
3505
3506 d->strs_set = strset__new(BTF_MAX_STR_OFFSET, NULL, 0);
3507 if (IS_ERR(d->strs_set)) {
3508 err = PTR_ERR(d->strs_set);
3509 goto err_out;
3510 }
3511
3512 if (!d->btf->base_btf) {
3513 /* insert empty string; we won't be looking it up during strings
3514 * dedup, but it's good to have it for generic BTF string lookups
3515 */
3516 err = strset__add_str(d->strs_set, "");
3517 if (err < 0)
3518 goto err_out;
3519 }
3520
3521 /* remap string offsets */
3522 err = btf_for_each_str_off(d, strs_dedup_remap_str_off, d);
3523 if (err)
3524 goto err_out;
3525
3526 /* replace BTF string data and hash with deduped ones */
3527 strset__free(d->btf->strs_set);
3528 d->btf->hdr->str_len = strset__data_size(d->strs_set);
3529 d->btf->strs_set = d->strs_set;
3530 d->strs_set = NULL;
3531 d->btf->strs_deduped = true;
3532 return 0;
3533
3534err_out:
3535 strset__free(d->strs_set);
3536 d->strs_set = NULL;
3537
3538 return err;
3539}
3540
3541static long btf_hash_common(struct btf_type *t)
3542{
3543 long h;
3544
3545 h = hash_combine(0, t->name_off);
3546 h = hash_combine(h, t->info);
3547 h = hash_combine(h, t->size);
3548 return h;
3549}
3550
3551static bool btf_equal_common(struct btf_type *t1, struct btf_type *t2)
3552{
3553 return t1->name_off == t2->name_off &&
3554 t1->info == t2->info &&
3555 t1->size == t2->size;
3556}
3557
3558/* Calculate type signature hash of INT or TAG. */
3559static long btf_hash_int_decl_tag(struct btf_type *t)
3560{
3561 __u32 info = *(__u32 *)(t + 1);
3562 long h;
3563
3564 h = btf_hash_common(t);
3565 h = hash_combine(h, info);
3566 return h;
3567}
3568
3569/* Check structural equality of two INTs or TAGs. */
3570static bool btf_equal_int_tag(struct btf_type *t1, struct btf_type *t2)
3571{
3572 __u32 info1, info2;
3573
3574 if (!btf_equal_common(t1, t2))
3575 return false;
3576 info1 = *(__u32 *)(t1 + 1);
3577 info2 = *(__u32 *)(t2 + 1);
3578 return info1 == info2;
3579}
3580
3581/* Calculate type signature hash of ENUM/ENUM64. */
3582static long btf_hash_enum(struct btf_type *t)
3583{
3584 long h;
3585
3586 /* don't hash vlen, enum members and size to support enum fwd resolving */
3587 h = hash_combine(0, t->name_off);
3588 return h;
3589}
3590
3591static bool btf_equal_enum_members(struct btf_type *t1, struct btf_type *t2)
3592{
3593 const struct btf_enum *m1, *m2;
3594 __u16 vlen;
3595 int i;
3596
3597 vlen = btf_vlen(t1);
3598 m1 = btf_enum(t1);
3599 m2 = btf_enum(t2);
3600 for (i = 0; i < vlen; i++) {
3601 if (m1->name_off != m2->name_off || m1->val != m2->val)
3602 return false;
3603 m1++;
3604 m2++;
3605 }
3606 return true;
3607}
3608
3609static bool btf_equal_enum64_members(struct btf_type *t1, struct btf_type *t2)
3610{
3611 const struct btf_enum64 *m1, *m2;
3612 __u16 vlen;
3613 int i;
3614
3615 vlen = btf_vlen(t1);
3616 m1 = btf_enum64(t1);
3617 m2 = btf_enum64(t2);
3618 for (i = 0; i < vlen; i++) {
3619 if (m1->name_off != m2->name_off || m1->val_lo32 != m2->val_lo32 ||
3620 m1->val_hi32 != m2->val_hi32)
3621 return false;
3622 m1++;
3623 m2++;
3624 }
3625 return true;
3626}
3627
3628/* Check structural equality of two ENUMs or ENUM64s. */
3629static bool btf_equal_enum(struct btf_type *t1, struct btf_type *t2)
3630{
3631 if (!btf_equal_common(t1, t2))
3632 return false;
3633
3634 /* t1 & t2 kinds are identical because of btf_equal_common */
3635 if (btf_kind(t1) == BTF_KIND_ENUM)
3636 return btf_equal_enum_members(t1, t2);
3637 else
3638 return btf_equal_enum64_members(t1, t2);
3639}
3640
3641static inline bool btf_is_enum_fwd(struct btf_type *t)
3642{
3643 return btf_is_any_enum(t) && btf_vlen(t) == 0;
3644}
3645
3646static bool btf_compat_enum(struct btf_type *t1, struct btf_type *t2)
3647{
3648 if (!btf_is_enum_fwd(t1) && !btf_is_enum_fwd(t2))
3649 return btf_equal_enum(t1, t2);
3650 /* At this point either t1 or t2 or both are forward declarations, thus:
3651 * - skip comparing vlen because it is zero for forward declarations;
3652 * - skip comparing size to allow enum forward declarations
3653 * to be compatible with enum64 full declarations;
3654 * - skip comparing kind for the same reason.
3655 */
3656 return t1->name_off == t2->name_off &&
3657 btf_is_any_enum(t1) && btf_is_any_enum(t2);
3658}
3659
3660/*
3661 * Calculate type signature hash of STRUCT/UNION, ignoring referenced type IDs,
3662 * as referenced type IDs equivalence is established separately during type
3663 * graph equivalence check algorithm.
3664 */
3665static long btf_hash_struct(struct btf_type *t)
3666{
3667 const struct btf_member *member = btf_members(t);
3668 __u32 vlen = btf_vlen(t);
3669 long h = btf_hash_common(t);
3670 int i;
3671
3672 for (i = 0; i < vlen; i++) {
3673 h = hash_combine(h, member->name_off);
3674 h = hash_combine(h, member->offset);
3675 /* no hashing of referenced type ID, it can be unresolved yet */
3676 member++;
3677 }
3678 return h;
3679}
3680
3681/*
3682 * Check structural compatibility of two STRUCTs/UNIONs, ignoring referenced
3683 * type IDs. This check is performed during type graph equivalence check and
3684 * referenced types equivalence is checked separately.
3685 */
3686static bool btf_shallow_equal_struct(struct btf_type *t1, struct btf_type *t2)
3687{
3688 const struct btf_member *m1, *m2;
3689 __u16 vlen;
3690 int i;
3691
3692 if (!btf_equal_common(t1, t2))
3693 return false;
3694
3695 vlen = btf_vlen(t1);
3696 m1 = btf_members(t1);
3697 m2 = btf_members(t2);
3698 for (i = 0; i < vlen; i++) {
3699 if (m1->name_off != m2->name_off || m1->offset != m2->offset)
3700 return false;
3701 m1++;
3702 m2++;
3703 }
3704 return true;
3705}
3706
3707/*
3708 * Calculate type signature hash of ARRAY, including referenced type IDs,
3709 * under assumption that they were already resolved to canonical type IDs and
3710 * are not going to change.
3711 */
3712static long btf_hash_array(struct btf_type *t)
3713{
3714 const struct btf_array *info = btf_array(t);
3715 long h = btf_hash_common(t);
3716
3717 h = hash_combine(h, info->type);
3718 h = hash_combine(h, info->index_type);
3719 h = hash_combine(h, info->nelems);
3720 return h;
3721}
3722
3723/*
3724 * Check exact equality of two ARRAYs, taking into account referenced
3725 * type IDs, under assumption that they were already resolved to canonical
3726 * type IDs and are not going to change.
3727 * This function is called during reference types deduplication to compare
3728 * ARRAY to potential canonical representative.
3729 */
3730static bool btf_equal_array(struct btf_type *t1, struct btf_type *t2)
3731{
3732 const struct btf_array *info1, *info2;
3733
3734 if (!btf_equal_common(t1, t2))
3735 return false;
3736
3737 info1 = btf_array(t1);
3738 info2 = btf_array(t2);
3739 return info1->type == info2->type &&
3740 info1->index_type == info2->index_type &&
3741 info1->nelems == info2->nelems;
3742}
3743
3744/*
3745 * Check structural compatibility of two ARRAYs, ignoring referenced type
3746 * IDs. This check is performed during type graph equivalence check and
3747 * referenced types equivalence is checked separately.
3748 */
3749static bool btf_compat_array(struct btf_type *t1, struct btf_type *t2)
3750{
3751 if (!btf_equal_common(t1, t2))
3752 return false;
3753
3754 return btf_array(t1)->nelems == btf_array(t2)->nelems;
3755}
3756
3757/*
3758 * Calculate type signature hash of FUNC_PROTO, including referenced type IDs,
3759 * under assumption that they were already resolved to canonical type IDs and
3760 * are not going to change.
3761 */
3762static long btf_hash_fnproto(struct btf_type *t)
3763{
3764 const struct btf_param *member = btf_params(t);
3765 __u16 vlen = btf_vlen(t);
3766 long h = btf_hash_common(t);
3767 int i;
3768
3769 for (i = 0; i < vlen; i++) {
3770 h = hash_combine(h, member->name_off);
3771 h = hash_combine(h, member->type);
3772 member++;
3773 }
3774 return h;
3775}
3776
3777/*
3778 * Check exact equality of two FUNC_PROTOs, taking into account referenced
3779 * type IDs, under assumption that they were already resolved to canonical
3780 * type IDs and are not going to change.
3781 * This function is called during reference types deduplication to compare
3782 * FUNC_PROTO to potential canonical representative.
3783 */
3784static bool btf_equal_fnproto(struct btf_type *t1, struct btf_type *t2)
3785{
3786 const struct btf_param *m1, *m2;
3787 __u16 vlen;
3788 int i;
3789
3790 if (!btf_equal_common(t1, t2))
3791 return false;
3792
3793 vlen = btf_vlen(t1);
3794 m1 = btf_params(t1);
3795 m2 = btf_params(t2);
3796 for (i = 0; i < vlen; i++) {
3797 if (m1->name_off != m2->name_off || m1->type != m2->type)
3798 return false;
3799 m1++;
3800 m2++;
3801 }
3802 return true;
3803}
3804
3805/*
3806 * Check structural compatibility of two FUNC_PROTOs, ignoring referenced type
3807 * IDs. This check is performed during type graph equivalence check and
3808 * referenced types equivalence is checked separately.
3809 */
3810static bool btf_compat_fnproto(struct btf_type *t1, struct btf_type *t2)
3811{
3812 const struct btf_param *m1, *m2;
3813 __u16 vlen;
3814 int i;
3815
3816 /* skip return type ID */
3817 if (t1->name_off != t2->name_off || t1->info != t2->info)
3818 return false;
3819
3820 vlen = btf_vlen(t1);
3821 m1 = btf_params(t1);
3822 m2 = btf_params(t2);
3823 for (i = 0; i < vlen; i++) {
3824 if (m1->name_off != m2->name_off)
3825 return false;
3826 m1++;
3827 m2++;
3828 }
3829 return true;
3830}
3831
3832/* Prepare split BTF for deduplication by calculating hashes of base BTF's
3833 * types and initializing the rest of the state (canonical type mapping) for
3834 * the fixed base BTF part.
3835 */
3836static int btf_dedup_prep(struct btf_dedup *d)
3837{
3838 struct btf_type *t;
3839 int type_id;
3840 long h;
3841
3842 if (!d->btf->base_btf)
3843 return 0;
3844
3845 for (type_id = 1; type_id < d->btf->start_id; type_id++) {
3846 t = btf_type_by_id(d->btf, type_id);
3847
3848 /* all base BTF types are self-canonical by definition */
3849 d->map[type_id] = type_id;
3850
3851 switch (btf_kind(t)) {
3852 case BTF_KIND_VAR:
3853 case BTF_KIND_DATASEC:
3854 /* VAR and DATASEC are never hash/deduplicated */
3855 continue;
3856 case BTF_KIND_CONST:
3857 case BTF_KIND_VOLATILE:
3858 case BTF_KIND_RESTRICT:
3859 case BTF_KIND_PTR:
3860 case BTF_KIND_FWD:
3861 case BTF_KIND_TYPEDEF:
3862 case BTF_KIND_FUNC:
3863 case BTF_KIND_FLOAT:
3864 case BTF_KIND_TYPE_TAG:
3865 h = btf_hash_common(t);
3866 break;
3867 case BTF_KIND_INT:
3868 case BTF_KIND_DECL_TAG:
3869 h = btf_hash_int_decl_tag(t);
3870 break;
3871 case BTF_KIND_ENUM:
3872 case BTF_KIND_ENUM64:
3873 h = btf_hash_enum(t);
3874 break;
3875 case BTF_KIND_STRUCT:
3876 case BTF_KIND_UNION:
3877 h = btf_hash_struct(t);
3878 break;
3879 case BTF_KIND_ARRAY:
3880 h = btf_hash_array(t);
3881 break;
3882 case BTF_KIND_FUNC_PROTO:
3883 h = btf_hash_fnproto(t);
3884 break;
3885 default:
3886 pr_debug("unknown kind %d for type [%d]\n", btf_kind(t), type_id);
3887 return -EINVAL;
3888 }
3889 if (btf_dedup_table_add(d, h, type_id))
3890 return -ENOMEM;
3891 }
3892
3893 return 0;
3894}
3895
3896/*
3897 * Deduplicate primitive types, that can't reference other types, by calculating
3898 * their type signature hash and comparing them with any possible canonical
3899 * candidate. If no canonical candidate matches, type itself is marked as
3900 * canonical and is added into `btf_dedup->dedup_table` as another candidate.
3901 */
3902static int btf_dedup_prim_type(struct btf_dedup *d, __u32 type_id)
3903{
3904 struct btf_type *t = btf_type_by_id(d->btf, type_id);
3905 struct hashmap_entry *hash_entry;
3906 struct btf_type *cand;
3907 /* if we don't find equivalent type, then we are canonical */
3908 __u32 new_id = type_id;
3909 __u32 cand_id;
3910 long h;
3911
3912 switch (btf_kind(t)) {
3913 case BTF_KIND_CONST:
3914 case BTF_KIND_VOLATILE:
3915 case BTF_KIND_RESTRICT:
3916 case BTF_KIND_PTR:
3917 case BTF_KIND_TYPEDEF:
3918 case BTF_KIND_ARRAY:
3919 case BTF_KIND_STRUCT:
3920 case BTF_KIND_UNION:
3921 case BTF_KIND_FUNC:
3922 case BTF_KIND_FUNC_PROTO:
3923 case BTF_KIND_VAR:
3924 case BTF_KIND_DATASEC:
3925 case BTF_KIND_DECL_TAG:
3926 case BTF_KIND_TYPE_TAG:
3927 return 0;
3928
3929 case BTF_KIND_INT:
3930 h = btf_hash_int_decl_tag(t);
3931 for_each_dedup_cand(d, hash_entry, h) {
3932 cand_id = hash_entry->value;
3933 cand = btf_type_by_id(d->btf, cand_id);
3934 if (btf_equal_int_tag(t, cand)) {
3935 new_id = cand_id;
3936 break;
3937 }
3938 }
3939 break;
3940
3941 case BTF_KIND_ENUM:
3942 case BTF_KIND_ENUM64:
3943 h = btf_hash_enum(t);
3944 for_each_dedup_cand(d, hash_entry, h) {
3945 cand_id = hash_entry->value;
3946 cand = btf_type_by_id(d->btf, cand_id);
3947 if (btf_equal_enum(t, cand)) {
3948 new_id = cand_id;
3949 break;
3950 }
3951 if (btf_compat_enum(t, cand)) {
3952 if (btf_is_enum_fwd(t)) {
3953 /* resolve fwd to full enum */
3954 new_id = cand_id;
3955 break;
3956 }
3957 /* resolve canonical enum fwd to full enum */
3958 d->map[cand_id] = type_id;
3959 }
3960 }
3961 break;
3962
3963 case BTF_KIND_FWD:
3964 case BTF_KIND_FLOAT:
3965 h = btf_hash_common(t);
3966 for_each_dedup_cand(d, hash_entry, h) {
3967 cand_id = hash_entry->value;
3968 cand = btf_type_by_id(d->btf, cand_id);
3969 if (btf_equal_common(t, cand)) {
3970 new_id = cand_id;
3971 break;
3972 }
3973 }
3974 break;
3975
3976 default:
3977 return -EINVAL;
3978 }
3979
3980 d->map[type_id] = new_id;
3981 if (type_id == new_id && btf_dedup_table_add(d, h, type_id))
3982 return -ENOMEM;
3983
3984 return 0;
3985}
3986
3987static int btf_dedup_prim_types(struct btf_dedup *d)
3988{
3989 int i, err;
3990
3991 for (i = 0; i < d->btf->nr_types; i++) {
3992 err = btf_dedup_prim_type(d, d->btf->start_id + i);
3993 if (err)
3994 return err;
3995 }
3996 return 0;
3997}
3998
3999/*
4000 * Check whether type is already mapped into canonical one (could be to itself).
4001 */
4002static inline bool is_type_mapped(struct btf_dedup *d, uint32_t type_id)
4003{
4004 return d->map[type_id] <= BTF_MAX_NR_TYPES;
4005}
4006
4007/*
4008 * Resolve type ID into its canonical type ID, if any; otherwise return original
4009 * type ID. If type is FWD and is resolved into STRUCT/UNION already, follow
4010 * STRUCT/UNION link and resolve it into canonical type ID as well.
4011 */
4012static inline __u32 resolve_type_id(struct btf_dedup *d, __u32 type_id)
4013{
4014 while (is_type_mapped(d, type_id) && d->map[type_id] != type_id)
4015 type_id = d->map[type_id];
4016 return type_id;
4017}
4018
4019/*
4020 * Resolve FWD to underlying STRUCT/UNION, if any; otherwise return original
4021 * type ID.
4022 */
4023static uint32_t resolve_fwd_id(struct btf_dedup *d, uint32_t type_id)
4024{
4025 __u32 orig_type_id = type_id;
4026
4027 if (!btf_is_fwd(btf__type_by_id(d->btf, type_id)))
4028 return type_id;
4029
4030 while (is_type_mapped(d, type_id) && d->map[type_id] != type_id)
4031 type_id = d->map[type_id];
4032
4033 if (!btf_is_fwd(btf__type_by_id(d->btf, type_id)))
4034 return type_id;
4035
4036 return orig_type_id;
4037}
4038
4039
4040static inline __u16 btf_fwd_kind(struct btf_type *t)
4041{
4042 return btf_kflag(t) ? BTF_KIND_UNION : BTF_KIND_STRUCT;
4043}
4044
4045/* Check if given two types are identical ARRAY definitions */
4046static bool btf_dedup_identical_arrays(struct btf_dedup *d, __u32 id1, __u32 id2)
4047{
4048 struct btf_type *t1, *t2;
4049
4050 t1 = btf_type_by_id(d->btf, id1);
4051 t2 = btf_type_by_id(d->btf, id2);
4052 if (!btf_is_array(t1) || !btf_is_array(t2))
4053 return false;
4054
4055 return btf_equal_array(t1, t2);
4056}
4057
4058/* Check if given two types are identical STRUCT/UNION definitions */
4059static bool btf_dedup_identical_structs(struct btf_dedup *d, __u32 id1, __u32 id2)
4060{
4061 const struct btf_member *m1, *m2;
4062 struct btf_type *t1, *t2;
4063 int n, i;
4064
4065 t1 = btf_type_by_id(d->btf, id1);
4066 t2 = btf_type_by_id(d->btf, id2);
4067
4068 if (!btf_is_composite(t1) || btf_kind(t1) != btf_kind(t2))
4069 return false;
4070
4071 if (!btf_shallow_equal_struct(t1, t2))
4072 return false;
4073
4074 m1 = btf_members(t1);
4075 m2 = btf_members(t2);
4076 for (i = 0, n = btf_vlen(t1); i < n; i++, m1++, m2++) {
4077 if (m1->type != m2->type &&
4078 !btf_dedup_identical_arrays(d, m1->type, m2->type) &&
4079 !btf_dedup_identical_structs(d, m1->type, m2->type))
4080 return false;
4081 }
4082 return true;
4083}
4084
4085/*
4086 * Check equivalence of BTF type graph formed by candidate struct/union (we'll
4087 * call it "candidate graph" in this description for brevity) to a type graph
4088 * formed by (potential) canonical struct/union ("canonical graph" for brevity
4089 * here, though keep in mind that not all types in canonical graph are
4090 * necessarily canonical representatives themselves, some of them might be
4091 * duplicates or its uniqueness might not have been established yet).
4092 * Returns:
4093 * - >0, if type graphs are equivalent;
4094 * - 0, if not equivalent;
4095 * - <0, on error.
4096 *
4097 * Algorithm performs side-by-side DFS traversal of both type graphs and checks
4098 * equivalence of BTF types at each step. If at any point BTF types in candidate
4099 * and canonical graphs are not compatible structurally, whole graphs are
4100 * incompatible. If types are structurally equivalent (i.e., all information
4101 * except referenced type IDs is exactly the same), a mapping from `canon_id` to
4102 * a `cand_id` is recored in hypothetical mapping (`btf_dedup->hypot_map`).
4103 * If a type references other types, then those referenced types are checked
4104 * for equivalence recursively.
4105 *
4106 * During DFS traversal, if we find that for current `canon_id` type we
4107 * already have some mapping in hypothetical map, we check for two possible
4108 * situations:
4109 * - `canon_id` is mapped to exactly the same type as `cand_id`. This will
4110 * happen when type graphs have cycles. In this case we assume those two
4111 * types are equivalent.
4112 * - `canon_id` is mapped to different type. This is contradiction in our
4113 * hypothetical mapping, because same graph in canonical graph corresponds
4114 * to two different types in candidate graph, which for equivalent type
4115 * graphs shouldn't happen. This condition terminates equivalence check
4116 * with negative result.
4117 *
4118 * If type graphs traversal exhausts types to check and find no contradiction,
4119 * then type graphs are equivalent.
4120 *
4121 * When checking types for equivalence, there is one special case: FWD types.
4122 * If FWD type resolution is allowed and one of the types (either from canonical
4123 * or candidate graph) is FWD and other is STRUCT/UNION (depending on FWD's kind
4124 * flag) and their names match, hypothetical mapping is updated to point from
4125 * FWD to STRUCT/UNION. If graphs will be determined as equivalent successfully,
4126 * this mapping will be used to record FWD -> STRUCT/UNION mapping permanently.
4127 *
4128 * Technically, this could lead to incorrect FWD to STRUCT/UNION resolution,
4129 * if there are two exactly named (or anonymous) structs/unions that are
4130 * compatible structurally, one of which has FWD field, while other is concrete
4131 * STRUCT/UNION, but according to C sources they are different structs/unions
4132 * that are referencing different types with the same name. This is extremely
4133 * unlikely to happen, but btf_dedup API allows to disable FWD resolution if
4134 * this logic is causing problems.
4135 *
4136 * Doing FWD resolution means that both candidate and/or canonical graphs can
4137 * consists of portions of the graph that come from multiple compilation units.
4138 * This is due to the fact that types within single compilation unit are always
4139 * deduplicated and FWDs are already resolved, if referenced struct/union
4140 * definiton is available. So, if we had unresolved FWD and found corresponding
4141 * STRUCT/UNION, they will be from different compilation units. This
4142 * consequently means that when we "link" FWD to corresponding STRUCT/UNION,
4143 * type graph will likely have at least two different BTF types that describe
4144 * same type (e.g., most probably there will be two different BTF types for the
4145 * same 'int' primitive type) and could even have "overlapping" parts of type
4146 * graph that describe same subset of types.
4147 *
4148 * This in turn means that our assumption that each type in canonical graph
4149 * must correspond to exactly one type in candidate graph might not hold
4150 * anymore and will make it harder to detect contradictions using hypothetical
4151 * map. To handle this problem, we allow to follow FWD -> STRUCT/UNION
4152 * resolution only in canonical graph. FWDs in candidate graphs are never
4153 * resolved. To see why it's OK, let's check all possible situations w.r.t. FWDs
4154 * that can occur:
4155 * - Both types in canonical and candidate graphs are FWDs. If they are
4156 * structurally equivalent, then they can either be both resolved to the
4157 * same STRUCT/UNION or not resolved at all. In both cases they are
4158 * equivalent and there is no need to resolve FWD on candidate side.
4159 * - Both types in canonical and candidate graphs are concrete STRUCT/UNION,
4160 * so nothing to resolve as well, algorithm will check equivalence anyway.
4161 * - Type in canonical graph is FWD, while type in candidate is concrete
4162 * STRUCT/UNION. In this case candidate graph comes from single compilation
4163 * unit, so there is exactly one BTF type for each unique C type. After
4164 * resolving FWD into STRUCT/UNION, there might be more than one BTF type
4165 * in canonical graph mapping to single BTF type in candidate graph, but
4166 * because hypothetical mapping maps from canonical to candidate types, it's
4167 * alright, and we still maintain the property of having single `canon_id`
4168 * mapping to single `cand_id` (there could be two different `canon_id`
4169 * mapped to the same `cand_id`, but it's not contradictory).
4170 * - Type in canonical graph is concrete STRUCT/UNION, while type in candidate
4171 * graph is FWD. In this case we are just going to check compatibility of
4172 * STRUCT/UNION and corresponding FWD, and if they are compatible, we'll
4173 * assume that whatever STRUCT/UNION FWD resolves to must be equivalent to
4174 * a concrete STRUCT/UNION from canonical graph. If the rest of type graphs
4175 * turn out equivalent, we'll re-resolve FWD to concrete STRUCT/UNION from
4176 * canonical graph.
4177 */
4178static int btf_dedup_is_equiv(struct btf_dedup *d, __u32 cand_id,
4179 __u32 canon_id)
4180{
4181 struct btf_type *cand_type;
4182 struct btf_type *canon_type;
4183 __u32 hypot_type_id;
4184 __u16 cand_kind;
4185 __u16 canon_kind;
4186 int i, eq;
4187
4188 /* if both resolve to the same canonical, they must be equivalent */
4189 if (resolve_type_id(d, cand_id) == resolve_type_id(d, canon_id))
4190 return 1;
4191
4192 canon_id = resolve_fwd_id(d, canon_id);
4193
4194 hypot_type_id = d->hypot_map[canon_id];
4195 if (hypot_type_id <= BTF_MAX_NR_TYPES) {
4196 if (hypot_type_id == cand_id)
4197 return 1;
4198 /* In some cases compiler will generate different DWARF types
4199 * for *identical* array type definitions and use them for
4200 * different fields within the *same* struct. This breaks type
4201 * equivalence check, which makes an assumption that candidate
4202 * types sub-graph has a consistent and deduped-by-compiler
4203 * types within a single CU. So work around that by explicitly
4204 * allowing identical array types here.
4205 */
4206 if (btf_dedup_identical_arrays(d, hypot_type_id, cand_id))
4207 return 1;
4208 /* It turns out that similar situation can happen with
4209 * struct/union sometimes, sigh... Handle the case where
4210 * structs/unions are exactly the same, down to the referenced
4211 * type IDs. Anything more complicated (e.g., if referenced
4212 * types are different, but equivalent) is *way more*
4213 * complicated and requires a many-to-many equivalence mapping.
4214 */
4215 if (btf_dedup_identical_structs(d, hypot_type_id, cand_id))
4216 return 1;
4217 return 0;
4218 }
4219
4220 if (btf_dedup_hypot_map_add(d, canon_id, cand_id))
4221 return -ENOMEM;
4222
4223 cand_type = btf_type_by_id(d->btf, cand_id);
4224 canon_type = btf_type_by_id(d->btf, canon_id);
4225 cand_kind = btf_kind(cand_type);
4226 canon_kind = btf_kind(canon_type);
4227
4228 if (cand_type->name_off != canon_type->name_off)
4229 return 0;
4230
4231 /* FWD <--> STRUCT/UNION equivalence check, if enabled */
4232 if ((cand_kind == BTF_KIND_FWD || canon_kind == BTF_KIND_FWD)
4233 && cand_kind != canon_kind) {
4234 __u16 real_kind;
4235 __u16 fwd_kind;
4236
4237 if (cand_kind == BTF_KIND_FWD) {
4238 real_kind = canon_kind;
4239 fwd_kind = btf_fwd_kind(cand_type);
4240 } else {
4241 real_kind = cand_kind;
4242 fwd_kind = btf_fwd_kind(canon_type);
4243 /* we'd need to resolve base FWD to STRUCT/UNION */
4244 if (fwd_kind == real_kind && canon_id < d->btf->start_id)
4245 d->hypot_adjust_canon = true;
4246 }
4247 return fwd_kind == real_kind;
4248 }
4249
4250 if (cand_kind != canon_kind)
4251 return 0;
4252
4253 switch (cand_kind) {
4254 case BTF_KIND_INT:
4255 return btf_equal_int_tag(cand_type, canon_type);
4256
4257 case BTF_KIND_ENUM:
4258 case BTF_KIND_ENUM64:
4259 return btf_compat_enum(cand_type, canon_type);
4260
4261 case BTF_KIND_FWD:
4262 case BTF_KIND_FLOAT:
4263 return btf_equal_common(cand_type, canon_type);
4264
4265 case BTF_KIND_CONST:
4266 case BTF_KIND_VOLATILE:
4267 case BTF_KIND_RESTRICT:
4268 case BTF_KIND_PTR:
4269 case BTF_KIND_TYPEDEF:
4270 case BTF_KIND_FUNC:
4271 case BTF_KIND_TYPE_TAG:
4272 if (cand_type->info != canon_type->info)
4273 return 0;
4274 return btf_dedup_is_equiv(d, cand_type->type, canon_type->type);
4275
4276 case BTF_KIND_ARRAY: {
4277 const struct btf_array *cand_arr, *canon_arr;
4278
4279 if (!btf_compat_array(cand_type, canon_type))
4280 return 0;
4281 cand_arr = btf_array(cand_type);
4282 canon_arr = btf_array(canon_type);
4283 eq = btf_dedup_is_equiv(d, cand_arr->index_type, canon_arr->index_type);
4284 if (eq <= 0)
4285 return eq;
4286 return btf_dedup_is_equiv(d, cand_arr->type, canon_arr->type);
4287 }
4288
4289 case BTF_KIND_STRUCT:
4290 case BTF_KIND_UNION: {
4291 const struct btf_member *cand_m, *canon_m;
4292 __u16 vlen;
4293
4294 if (!btf_shallow_equal_struct(cand_type, canon_type))
4295 return 0;
4296 vlen = btf_vlen(cand_type);
4297 cand_m = btf_members(cand_type);
4298 canon_m = btf_members(canon_type);
4299 for (i = 0; i < vlen; i++) {
4300 eq = btf_dedup_is_equiv(d, cand_m->type, canon_m->type);
4301 if (eq <= 0)
4302 return eq;
4303 cand_m++;
4304 canon_m++;
4305 }
4306
4307 return 1;
4308 }
4309
4310 case BTF_KIND_FUNC_PROTO: {
4311 const struct btf_param *cand_p, *canon_p;
4312 __u16 vlen;
4313
4314 if (!btf_compat_fnproto(cand_type, canon_type))
4315 return 0;
4316 eq = btf_dedup_is_equiv(d, cand_type->type, canon_type->type);
4317 if (eq <= 0)
4318 return eq;
4319 vlen = btf_vlen(cand_type);
4320 cand_p = btf_params(cand_type);
4321 canon_p = btf_params(canon_type);
4322 for (i = 0; i < vlen; i++) {
4323 eq = btf_dedup_is_equiv(d, cand_p->type, canon_p->type);
4324 if (eq <= 0)
4325 return eq;
4326 cand_p++;
4327 canon_p++;
4328 }
4329 return 1;
4330 }
4331
4332 default:
4333 return -EINVAL;
4334 }
4335 return 0;
4336}
4337
4338/*
4339 * Use hypothetical mapping, produced by successful type graph equivalence
4340 * check, to augment existing struct/union canonical mapping, where possible.
4341 *
4342 * If BTF_KIND_FWD resolution is allowed, this mapping is also used to record
4343 * FWD -> STRUCT/UNION correspondence as well. FWD resolution is bidirectional:
4344 * it doesn't matter if FWD type was part of canonical graph or candidate one,
4345 * we are recording the mapping anyway. As opposed to carefulness required
4346 * for struct/union correspondence mapping (described below), for FWD resolution
4347 * it's not important, as by the time that FWD type (reference type) will be
4348 * deduplicated all structs/unions will be deduped already anyway.
4349 *
4350 * Recording STRUCT/UNION mapping is purely a performance optimization and is
4351 * not required for correctness. It needs to be done carefully to ensure that
4352 * struct/union from candidate's type graph is not mapped into corresponding
4353 * struct/union from canonical type graph that itself hasn't been resolved into
4354 * canonical representative. The only guarantee we have is that canonical
4355 * struct/union was determined as canonical and that won't change. But any
4356 * types referenced through that struct/union fields could have been not yet
4357 * resolved, so in case like that it's too early to establish any kind of
4358 * correspondence between structs/unions.
4359 *
4360 * No canonical correspondence is derived for primitive types (they are already
4361 * deduplicated completely already anyway) or reference types (they rely on
4362 * stability of struct/union canonical relationship for equivalence checks).
4363 */
4364static void btf_dedup_merge_hypot_map(struct btf_dedup *d)
4365{
4366 __u32 canon_type_id, targ_type_id;
4367 __u16 t_kind, c_kind;
4368 __u32 t_id, c_id;
4369 int i;
4370
4371 for (i = 0; i < d->hypot_cnt; i++) {
4372 canon_type_id = d->hypot_list[i];
4373 targ_type_id = d->hypot_map[canon_type_id];
4374 t_id = resolve_type_id(d, targ_type_id);
4375 c_id = resolve_type_id(d, canon_type_id);
4376 t_kind = btf_kind(btf__type_by_id(d->btf, t_id));
4377 c_kind = btf_kind(btf__type_by_id(d->btf, c_id));
4378 /*
4379 * Resolve FWD into STRUCT/UNION.
4380 * It's ok to resolve FWD into STRUCT/UNION that's not yet
4381 * mapped to canonical representative (as opposed to
4382 * STRUCT/UNION <--> STRUCT/UNION mapping logic below), because
4383 * eventually that struct is going to be mapped and all resolved
4384 * FWDs will automatically resolve to correct canonical
4385 * representative. This will happen before ref type deduping,
4386 * which critically depends on stability of these mapping. This
4387 * stability is not a requirement for STRUCT/UNION equivalence
4388 * checks, though.
4389 */
4390
4391 /* if it's the split BTF case, we still need to point base FWD
4392 * to STRUCT/UNION in a split BTF, because FWDs from split BTF
4393 * will be resolved against base FWD. If we don't point base
4394 * canonical FWD to the resolved STRUCT/UNION, then all the
4395 * FWDs in split BTF won't be correctly resolved to a proper
4396 * STRUCT/UNION.
4397 */
4398 if (t_kind != BTF_KIND_FWD && c_kind == BTF_KIND_FWD)
4399 d->map[c_id] = t_id;
4400
4401 /* if graph equivalence determined that we'd need to adjust
4402 * base canonical types, then we need to only point base FWDs
4403 * to STRUCTs/UNIONs and do no more modifications. For all
4404 * other purposes the type graphs were not equivalent.
4405 */
4406 if (d->hypot_adjust_canon)
4407 continue;
4408
4409 if (t_kind == BTF_KIND_FWD && c_kind != BTF_KIND_FWD)
4410 d->map[t_id] = c_id;
4411
4412 if ((t_kind == BTF_KIND_STRUCT || t_kind == BTF_KIND_UNION) &&
4413 c_kind != BTF_KIND_FWD &&
4414 is_type_mapped(d, c_id) &&
4415 !is_type_mapped(d, t_id)) {
4416 /*
4417 * as a perf optimization, we can map struct/union
4418 * that's part of type graph we just verified for
4419 * equivalence. We can do that for struct/union that has
4420 * canonical representative only, though.
4421 */
4422 d->map[t_id] = c_id;
4423 }
4424 }
4425}
4426
4427/*
4428 * Deduplicate struct/union types.
4429 *
4430 * For each struct/union type its type signature hash is calculated, taking
4431 * into account type's name, size, number, order and names of fields, but
4432 * ignoring type ID's referenced from fields, because they might not be deduped
4433 * completely until after reference types deduplication phase. This type hash
4434 * is used to iterate over all potential canonical types, sharing same hash.
4435 * For each canonical candidate we check whether type graphs that they form
4436 * (through referenced types in fields and so on) are equivalent using algorithm
4437 * implemented in `btf_dedup_is_equiv`. If such equivalence is found and
4438 * BTF_KIND_FWD resolution is allowed, then hypothetical mapping
4439 * (btf_dedup->hypot_map) produced by aforementioned type graph equivalence
4440 * algorithm is used to record FWD -> STRUCT/UNION mapping. It's also used to
4441 * potentially map other structs/unions to their canonical representatives,
4442 * if such relationship hasn't yet been established. This speeds up algorithm
4443 * by eliminating some of the duplicate work.
4444 *
4445 * If no matching canonical representative was found, struct/union is marked
4446 * as canonical for itself and is added into btf_dedup->dedup_table hash map
4447 * for further look ups.
4448 */
4449static int btf_dedup_struct_type(struct btf_dedup *d, __u32 type_id)
4450{
4451 struct btf_type *cand_type, *t;
4452 struct hashmap_entry *hash_entry;
4453 /* if we don't find equivalent type, then we are canonical */
4454 __u32 new_id = type_id;
4455 __u16 kind;
4456 long h;
4457
4458 /* already deduped or is in process of deduping (loop detected) */
4459 if (d->map[type_id] <= BTF_MAX_NR_TYPES)
4460 return 0;
4461
4462 t = btf_type_by_id(d->btf, type_id);
4463 kind = btf_kind(t);
4464
4465 if (kind != BTF_KIND_STRUCT && kind != BTF_KIND_UNION)
4466 return 0;
4467
4468 h = btf_hash_struct(t);
4469 for_each_dedup_cand(d, hash_entry, h) {
4470 __u32 cand_id = hash_entry->value;
4471 int eq;
4472
4473 /*
4474 * Even though btf_dedup_is_equiv() checks for
4475 * btf_shallow_equal_struct() internally when checking two
4476 * structs (unions) for equivalence, we need to guard here
4477 * from picking matching FWD type as a dedup candidate.
4478 * This can happen due to hash collision. In such case just
4479 * relying on btf_dedup_is_equiv() would lead to potentially
4480 * creating a loop (FWD -> STRUCT and STRUCT -> FWD), because
4481 * FWD and compatible STRUCT/UNION are considered equivalent.
4482 */
4483 cand_type = btf_type_by_id(d->btf, cand_id);
4484 if (!btf_shallow_equal_struct(t, cand_type))
4485 continue;
4486
4487 btf_dedup_clear_hypot_map(d);
4488 eq = btf_dedup_is_equiv(d, type_id, cand_id);
4489 if (eq < 0)
4490 return eq;
4491 if (!eq)
4492 continue;
4493 btf_dedup_merge_hypot_map(d);
4494 if (d->hypot_adjust_canon) /* not really equivalent */
4495 continue;
4496 new_id = cand_id;
4497 break;
4498 }
4499
4500 d->map[type_id] = new_id;
4501 if (type_id == new_id && btf_dedup_table_add(d, h, type_id))
4502 return -ENOMEM;
4503
4504 return 0;
4505}
4506
4507static int btf_dedup_struct_types(struct btf_dedup *d)
4508{
4509 int i, err;
4510
4511 for (i = 0; i < d->btf->nr_types; i++) {
4512 err = btf_dedup_struct_type(d, d->btf->start_id + i);
4513 if (err)
4514 return err;
4515 }
4516 return 0;
4517}
4518
4519/*
4520 * Deduplicate reference type.
4521 *
4522 * Once all primitive and struct/union types got deduplicated, we can easily
4523 * deduplicate all other (reference) BTF types. This is done in two steps:
4524 *
4525 * 1. Resolve all referenced type IDs into their canonical type IDs. This
4526 * resolution can be done either immediately for primitive or struct/union types
4527 * (because they were deduped in previous two phases) or recursively for
4528 * reference types. Recursion will always terminate at either primitive or
4529 * struct/union type, at which point we can "unwind" chain of reference types
4530 * one by one. There is no danger of encountering cycles because in C type
4531 * system the only way to form type cycle is through struct/union, so any chain
4532 * of reference types, even those taking part in a type cycle, will inevitably
4533 * reach struct/union at some point.
4534 *
4535 * 2. Once all referenced type IDs are resolved into canonical ones, BTF type
4536 * becomes "stable", in the sense that no further deduplication will cause
4537 * any changes to it. With that, it's now possible to calculate type's signature
4538 * hash (this time taking into account referenced type IDs) and loop over all
4539 * potential canonical representatives. If no match was found, current type
4540 * will become canonical representative of itself and will be added into
4541 * btf_dedup->dedup_table as another possible canonical representative.
4542 */
4543static int btf_dedup_ref_type(struct btf_dedup *d, __u32 type_id)
4544{
4545 struct hashmap_entry *hash_entry;
4546 __u32 new_id = type_id, cand_id;
4547 struct btf_type *t, *cand;
4548 /* if we don't find equivalent type, then we are representative type */
4549 int ref_type_id;
4550 long h;
4551
4552 if (d->map[type_id] == BTF_IN_PROGRESS_ID)
4553 return -ELOOP;
4554 if (d->map[type_id] <= BTF_MAX_NR_TYPES)
4555 return resolve_type_id(d, type_id);
4556
4557 t = btf_type_by_id(d->btf, type_id);
4558 d->map[type_id] = BTF_IN_PROGRESS_ID;
4559
4560 switch (btf_kind(t)) {
4561 case BTF_KIND_CONST:
4562 case BTF_KIND_VOLATILE:
4563 case BTF_KIND_RESTRICT:
4564 case BTF_KIND_PTR:
4565 case BTF_KIND_TYPEDEF:
4566 case BTF_KIND_FUNC:
4567 case BTF_KIND_TYPE_TAG:
4568 ref_type_id = btf_dedup_ref_type(d, t->type);
4569 if (ref_type_id < 0)
4570 return ref_type_id;
4571 t->type = ref_type_id;
4572
4573 h = btf_hash_common(t);
4574 for_each_dedup_cand(d, hash_entry, h) {
4575 cand_id = hash_entry->value;
4576 cand = btf_type_by_id(d->btf, cand_id);
4577 if (btf_equal_common(t, cand)) {
4578 new_id = cand_id;
4579 break;
4580 }
4581 }
4582 break;
4583
4584 case BTF_KIND_DECL_TAG:
4585 ref_type_id = btf_dedup_ref_type(d, t->type);
4586 if (ref_type_id < 0)
4587 return ref_type_id;
4588 t->type = ref_type_id;
4589
4590 h = btf_hash_int_decl_tag(t);
4591 for_each_dedup_cand(d, hash_entry, h) {
4592 cand_id = hash_entry->value;
4593 cand = btf_type_by_id(d->btf, cand_id);
4594 if (btf_equal_int_tag(t, cand)) {
4595 new_id = cand_id;
4596 break;
4597 }
4598 }
4599 break;
4600
4601 case BTF_KIND_ARRAY: {
4602 struct btf_array *info = btf_array(t);
4603
4604 ref_type_id = btf_dedup_ref_type(d, info->type);
4605 if (ref_type_id < 0)
4606 return ref_type_id;
4607 info->type = ref_type_id;
4608
4609 ref_type_id = btf_dedup_ref_type(d, info->index_type);
4610 if (ref_type_id < 0)
4611 return ref_type_id;
4612 info->index_type = ref_type_id;
4613
4614 h = btf_hash_array(t);
4615 for_each_dedup_cand(d, hash_entry, h) {
4616 cand_id = hash_entry->value;
4617 cand = btf_type_by_id(d->btf, cand_id);
4618 if (btf_equal_array(t, cand)) {
4619 new_id = cand_id;
4620 break;
4621 }
4622 }
4623 break;
4624 }
4625
4626 case BTF_KIND_FUNC_PROTO: {
4627 struct btf_param *param;
4628 __u16 vlen;
4629 int i;
4630
4631 ref_type_id = btf_dedup_ref_type(d, t->type);
4632 if (ref_type_id < 0)
4633 return ref_type_id;
4634 t->type = ref_type_id;
4635
4636 vlen = btf_vlen(t);
4637 param = btf_params(t);
4638 for (i = 0; i < vlen; i++) {
4639 ref_type_id = btf_dedup_ref_type(d, param->type);
4640 if (ref_type_id < 0)
4641 return ref_type_id;
4642 param->type = ref_type_id;
4643 param++;
4644 }
4645
4646 h = btf_hash_fnproto(t);
4647 for_each_dedup_cand(d, hash_entry, h) {
4648 cand_id = hash_entry->value;
4649 cand = btf_type_by_id(d->btf, cand_id);
4650 if (btf_equal_fnproto(t, cand)) {
4651 new_id = cand_id;
4652 break;
4653 }
4654 }
4655 break;
4656 }
4657
4658 default:
4659 return -EINVAL;
4660 }
4661
4662 d->map[type_id] = new_id;
4663 if (type_id == new_id && btf_dedup_table_add(d, h, type_id))
4664 return -ENOMEM;
4665
4666 return new_id;
4667}
4668
4669static int btf_dedup_ref_types(struct btf_dedup *d)
4670{
4671 int i, err;
4672
4673 for (i = 0; i < d->btf->nr_types; i++) {
4674 err = btf_dedup_ref_type(d, d->btf->start_id + i);
4675 if (err < 0)
4676 return err;
4677 }
4678 /* we won't need d->dedup_table anymore */
4679 hashmap__free(d->dedup_table);
4680 d->dedup_table = NULL;
4681 return 0;
4682}
4683
4684/*
4685 * Collect a map from type names to type ids for all canonical structs
4686 * and unions. If the same name is shared by several canonical types
4687 * use a special value 0 to indicate this fact.
4688 */
4689static int btf_dedup_fill_unique_names_map(struct btf_dedup *d, struct hashmap *names_map)
4690{
4691 __u32 nr_types = btf__type_cnt(d->btf);
4692 struct btf_type *t;
4693 __u32 type_id;
4694 __u16 kind;
4695 int err;
4696
4697 /*
4698 * Iterate over base and split module ids in order to get all
4699 * available structs in the map.
4700 */
4701 for (type_id = 1; type_id < nr_types; ++type_id) {
4702 t = btf_type_by_id(d->btf, type_id);
4703 kind = btf_kind(t);
4704
4705 if (kind != BTF_KIND_STRUCT && kind != BTF_KIND_UNION)
4706 continue;
4707
4708 /* Skip non-canonical types */
4709 if (type_id != d->map[type_id])
4710 continue;
4711
4712 err = hashmap__add(names_map, t->name_off, type_id);
4713 if (err == -EEXIST)
4714 err = hashmap__set(names_map, t->name_off, 0, NULL, NULL);
4715
4716 if (err)
4717 return err;
4718 }
4719
4720 return 0;
4721}
4722
4723static int btf_dedup_resolve_fwd(struct btf_dedup *d, struct hashmap *names_map, __u32 type_id)
4724{
4725 struct btf_type *t = btf_type_by_id(d->btf, type_id);
4726 enum btf_fwd_kind fwd_kind = btf_kflag(t);
4727 __u16 cand_kind, kind = btf_kind(t);
4728 struct btf_type *cand_t;
4729 uintptr_t cand_id;
4730
4731 if (kind != BTF_KIND_FWD)
4732 return 0;
4733
4734 /* Skip if this FWD already has a mapping */
4735 if (type_id != d->map[type_id])
4736 return 0;
4737
4738 if (!hashmap__find(names_map, t->name_off, &cand_id))
4739 return 0;
4740
4741 /* Zero is a special value indicating that name is not unique */
4742 if (!cand_id)
4743 return 0;
4744
4745 cand_t = btf_type_by_id(d->btf, cand_id);
4746 cand_kind = btf_kind(cand_t);
4747 if ((cand_kind == BTF_KIND_STRUCT && fwd_kind != BTF_FWD_STRUCT) ||
4748 (cand_kind == BTF_KIND_UNION && fwd_kind != BTF_FWD_UNION))
4749 return 0;
4750
4751 d->map[type_id] = cand_id;
4752
4753 return 0;
4754}
4755
4756/*
4757 * Resolve unambiguous forward declarations.
4758 *
4759 * The lion's share of all FWD declarations is resolved during
4760 * `btf_dedup_struct_types` phase when different type graphs are
4761 * compared against each other. However, if in some compilation unit a
4762 * FWD declaration is not a part of a type graph compared against
4763 * another type graph that declaration's canonical type would not be
4764 * changed. Example:
4765 *
4766 * CU #1:
4767 *
4768 * struct foo;
4769 * struct foo *some_global;
4770 *
4771 * CU #2:
4772 *
4773 * struct foo { int u; };
4774 * struct foo *another_global;
4775 *
4776 * After `btf_dedup_struct_types` the BTF looks as follows:
4777 *
4778 * [1] STRUCT 'foo' size=4 vlen=1 ...
4779 * [2] INT 'int' size=4 ...
4780 * [3] PTR '(anon)' type_id=1
4781 * [4] FWD 'foo' fwd_kind=struct
4782 * [5] PTR '(anon)' type_id=4
4783 *
4784 * This pass assumes that such FWD declarations should be mapped to
4785 * structs or unions with identical name in case if the name is not
4786 * ambiguous.
4787 */
4788static int btf_dedup_resolve_fwds(struct btf_dedup *d)
4789{
4790 int i, err;
4791 struct hashmap *names_map;
4792
4793 names_map = hashmap__new(btf_dedup_identity_hash_fn, btf_dedup_equal_fn, NULL);
4794 if (IS_ERR(names_map))
4795 return PTR_ERR(names_map);
4796
4797 err = btf_dedup_fill_unique_names_map(d, names_map);
4798 if (err < 0)
4799 goto exit;
4800
4801 for (i = 0; i < d->btf->nr_types; i++) {
4802 err = btf_dedup_resolve_fwd(d, names_map, d->btf->start_id + i);
4803 if (err < 0)
4804 break;
4805 }
4806
4807exit:
4808 hashmap__free(names_map);
4809 return err;
4810}
4811
4812/*
4813 * Compact types.
4814 *
4815 * After we established for each type its corresponding canonical representative
4816 * type, we now can eliminate types that are not canonical and leave only
4817 * canonical ones layed out sequentially in memory by copying them over
4818 * duplicates. During compaction btf_dedup->hypot_map array is reused to store
4819 * a map from original type ID to a new compacted type ID, which will be used
4820 * during next phase to "fix up" type IDs, referenced from struct/union and
4821 * reference types.
4822 */
4823static int btf_dedup_compact_types(struct btf_dedup *d)
4824{
4825 __u32 *new_offs;
4826 __u32 next_type_id = d->btf->start_id;
4827 const struct btf_type *t;
4828 void *p;
4829 int i, id, len;
4830
4831 /* we are going to reuse hypot_map to store compaction remapping */
4832 d->hypot_map[0] = 0;
4833 /* base BTF types are not renumbered */
4834 for (id = 1; id < d->btf->start_id; id++)
4835 d->hypot_map[id] = id;
4836 for (i = 0, id = d->btf->start_id; i < d->btf->nr_types; i++, id++)
4837 d->hypot_map[id] = BTF_UNPROCESSED_ID;
4838
4839 p = d->btf->types_data;
4840
4841 for (i = 0, id = d->btf->start_id; i < d->btf->nr_types; i++, id++) {
4842 if (d->map[id] != id)
4843 continue;
4844
4845 t = btf__type_by_id(d->btf, id);
4846 len = btf_type_size(t);
4847 if (len < 0)
4848 return len;
4849
4850 memmove(p, t, len);
4851 d->hypot_map[id] = next_type_id;
4852 d->btf->type_offs[next_type_id - d->btf->start_id] = p - d->btf->types_data;
4853 p += len;
4854 next_type_id++;
4855 }
4856
4857 /* shrink struct btf's internal types index and update btf_header */
4858 d->btf->nr_types = next_type_id - d->btf->start_id;
4859 d->btf->type_offs_cap = d->btf->nr_types;
4860 d->btf->hdr->type_len = p - d->btf->types_data;
4861 new_offs = libbpf_reallocarray(d->btf->type_offs, d->btf->type_offs_cap,
4862 sizeof(*new_offs));
4863 if (d->btf->type_offs_cap && !new_offs)
4864 return -ENOMEM;
4865 d->btf->type_offs = new_offs;
4866 d->btf->hdr->str_off = d->btf->hdr->type_len;
4867 d->btf->raw_size = d->btf->hdr->hdr_len + d->btf->hdr->type_len + d->btf->hdr->str_len;
4868 return 0;
4869}
4870
4871/*
4872 * Figure out final (deduplicated and compacted) type ID for provided original
4873 * `type_id` by first resolving it into corresponding canonical type ID and
4874 * then mapping it to a deduplicated type ID, stored in btf_dedup->hypot_map,
4875 * which is populated during compaction phase.
4876 */
4877static int btf_dedup_remap_type_id(__u32 *type_id, void *ctx)
4878{
4879 struct btf_dedup *d = ctx;
4880 __u32 resolved_type_id, new_type_id;
4881
4882 resolved_type_id = resolve_type_id(d, *type_id);
4883 new_type_id = d->hypot_map[resolved_type_id];
4884 if (new_type_id > BTF_MAX_NR_TYPES)
4885 return -EINVAL;
4886
4887 *type_id = new_type_id;
4888 return 0;
4889}
4890
4891/*
4892 * Remap referenced type IDs into deduped type IDs.
4893 *
4894 * After BTF types are deduplicated and compacted, their final type IDs may
4895 * differ from original ones. The map from original to a corresponding
4896 * deduped type ID is stored in btf_dedup->hypot_map and is populated during
4897 * compaction phase. During remapping phase we are rewriting all type IDs
4898 * referenced from any BTF type (e.g., struct fields, func proto args, etc) to
4899 * their final deduped type IDs.
4900 */
4901static int btf_dedup_remap_types(struct btf_dedup *d)
4902{
4903 int i, r;
4904
4905 for (i = 0; i < d->btf->nr_types; i++) {
4906 struct btf_type *t = btf_type_by_id(d->btf, d->btf->start_id + i);
4907
4908 r = btf_type_visit_type_ids(t, btf_dedup_remap_type_id, d);
4909 if (r)
4910 return r;
4911 }
4912
4913 if (!d->btf_ext)
4914 return 0;
4915
4916 r = btf_ext_visit_type_ids(d->btf_ext, btf_dedup_remap_type_id, d);
4917 if (r)
4918 return r;
4919
4920 return 0;
4921}
4922
4923/*
4924 * Probe few well-known locations for vmlinux kernel image and try to load BTF
4925 * data out of it to use for target BTF.
4926 */
4927struct btf *btf__load_vmlinux_btf(void)
4928{
4929 const char *locations[] = {
4930 /* try canonical vmlinux BTF through sysfs first */
4931 "/sys/kernel/btf/vmlinux",
4932 /* fall back to trying to find vmlinux on disk otherwise */
4933 "/boot/vmlinux-%1$s",
4934 "/lib/modules/%1$s/vmlinux-%1$s",
4935 "/lib/modules/%1$s/build/vmlinux",
4936 "/usr/lib/modules/%1$s/kernel/vmlinux",
4937 "/usr/lib/debug/boot/vmlinux-%1$s",
4938 "/usr/lib/debug/boot/vmlinux-%1$s.debug",
4939 "/usr/lib/debug/lib/modules/%1$s/vmlinux",
4940 };
4941 char path[PATH_MAX + 1];
4942 struct utsname buf;
4943 struct btf *btf;
4944 int i, err;
4945
4946 uname(&buf);
4947
4948 for (i = 0; i < ARRAY_SIZE(locations); i++) {
4949 snprintf(path, PATH_MAX, locations[i], buf.release);
4950
4951 if (faccessat(AT_FDCWD, path, R_OK, AT_EACCESS))
4952 continue;
4953
4954 btf = btf__parse(path, NULL);
4955 err = libbpf_get_error(btf);
4956 pr_debug("loading kernel BTF '%s': %d\n", path, err);
4957 if (err)
4958 continue;
4959
4960 return btf;
4961 }
4962
4963 pr_warn("failed to find valid kernel BTF\n");
4964 return libbpf_err_ptr(-ESRCH);
4965}
4966
4967struct btf *libbpf_find_kernel_btf(void) __attribute__((alias("btf__load_vmlinux_btf")));
4968
4969struct btf *btf__load_module_btf(const char *module_name, struct btf *vmlinux_btf)
4970{
4971 char path[80];
4972
4973 snprintf(path, sizeof(path), "/sys/kernel/btf/%s", module_name);
4974 return btf__parse_split(path, vmlinux_btf);
4975}
4976
4977int btf_type_visit_type_ids(struct btf_type *t, type_id_visit_fn visit, void *ctx)
4978{
4979 int i, n, err;
4980
4981 switch (btf_kind(t)) {
4982 case BTF_KIND_INT:
4983 case BTF_KIND_FLOAT:
4984 case BTF_KIND_ENUM:
4985 case BTF_KIND_ENUM64:
4986 return 0;
4987
4988 case BTF_KIND_FWD:
4989 case BTF_KIND_CONST:
4990 case BTF_KIND_VOLATILE:
4991 case BTF_KIND_RESTRICT:
4992 case BTF_KIND_PTR:
4993 case BTF_KIND_TYPEDEF:
4994 case BTF_KIND_FUNC:
4995 case BTF_KIND_VAR:
4996 case BTF_KIND_DECL_TAG:
4997 case BTF_KIND_TYPE_TAG:
4998 return visit(&t->type, ctx);
4999
5000 case BTF_KIND_ARRAY: {
5001 struct btf_array *a = btf_array(t);
5002
5003 err = visit(&a->type, ctx);
5004 err = err ?: visit(&a->index_type, ctx);
5005 return err;
5006 }
5007
5008 case BTF_KIND_STRUCT:
5009 case BTF_KIND_UNION: {
5010 struct btf_member *m = btf_members(t);
5011
5012 for (i = 0, n = btf_vlen(t); i < n; i++, m++) {
5013 err = visit(&m->type, ctx);
5014 if (err)
5015 return err;
5016 }
5017 return 0;
5018 }
5019
5020 case BTF_KIND_FUNC_PROTO: {
5021 struct btf_param *m = btf_params(t);
5022
5023 err = visit(&t->type, ctx);
5024 if (err)
5025 return err;
5026 for (i = 0, n = btf_vlen(t); i < n; i++, m++) {
5027 err = visit(&m->type, ctx);
5028 if (err)
5029 return err;
5030 }
5031 return 0;
5032 }
5033
5034 case BTF_KIND_DATASEC: {
5035 struct btf_var_secinfo *m = btf_var_secinfos(t);
5036
5037 for (i = 0, n = btf_vlen(t); i < n; i++, m++) {
5038 err = visit(&m->type, ctx);
5039 if (err)
5040 return err;
5041 }
5042 return 0;
5043 }
5044
5045 default:
5046 return -EINVAL;
5047 }
5048}
5049
5050int btf_type_visit_str_offs(struct btf_type *t, str_off_visit_fn visit, void *ctx)
5051{
5052 int i, n, err;
5053
5054 err = visit(&t->name_off, ctx);
5055 if (err)
5056 return err;
5057
5058 switch (btf_kind(t)) {
5059 case BTF_KIND_STRUCT:
5060 case BTF_KIND_UNION: {
5061 struct btf_member *m = btf_members(t);
5062
5063 for (i = 0, n = btf_vlen(t); i < n; i++, m++) {
5064 err = visit(&m->name_off, ctx);
5065 if (err)
5066 return err;
5067 }
5068 break;
5069 }
5070 case BTF_KIND_ENUM: {
5071 struct btf_enum *m = btf_enum(t);
5072
5073 for (i = 0, n = btf_vlen(t); i < n; i++, m++) {
5074 err = visit(&m->name_off, ctx);
5075 if (err)
5076 return err;
5077 }
5078 break;
5079 }
5080 case BTF_KIND_ENUM64: {
5081 struct btf_enum64 *m = btf_enum64(t);
5082
5083 for (i = 0, n = btf_vlen(t); i < n; i++, m++) {
5084 err = visit(&m->name_off, ctx);
5085 if (err)
5086 return err;
5087 }
5088 break;
5089 }
5090 case BTF_KIND_FUNC_PROTO: {
5091 struct btf_param *m = btf_params(t);
5092
5093 for (i = 0, n = btf_vlen(t); i < n; i++, m++) {
5094 err = visit(&m->name_off, ctx);
5095 if (err)
5096 return err;
5097 }
5098 break;
5099 }
5100 default:
5101 break;
5102 }
5103
5104 return 0;
5105}
5106
5107int btf_ext_visit_type_ids(struct btf_ext *btf_ext, type_id_visit_fn visit, void *ctx)
5108{
5109 const struct btf_ext_info *seg;
5110 struct btf_ext_info_sec *sec;
5111 int i, err;
5112
5113 seg = &btf_ext->func_info;
5114 for_each_btf_ext_sec(seg, sec) {
5115 struct bpf_func_info_min *rec;
5116
5117 for_each_btf_ext_rec(seg, sec, i, rec) {
5118 err = visit(&rec->type_id, ctx);
5119 if (err < 0)
5120 return err;
5121 }
5122 }
5123
5124 seg = &btf_ext->core_relo_info;
5125 for_each_btf_ext_sec(seg, sec) {
5126 struct bpf_core_relo *rec;
5127
5128 for_each_btf_ext_rec(seg, sec, i, rec) {
5129 err = visit(&rec->type_id, ctx);
5130 if (err < 0)
5131 return err;
5132 }
5133 }
5134
5135 return 0;
5136}
5137
5138int btf_ext_visit_str_offs(struct btf_ext *btf_ext, str_off_visit_fn visit, void *ctx)
5139{
5140 const struct btf_ext_info *seg;
5141 struct btf_ext_info_sec *sec;
5142 int i, err;
5143
5144 seg = &btf_ext->func_info;
5145 for_each_btf_ext_sec(seg, sec) {
5146 err = visit(&sec->sec_name_off, ctx);
5147 if (err)
5148 return err;
5149 }
5150
5151 seg = &btf_ext->line_info;
5152 for_each_btf_ext_sec(seg, sec) {
5153 struct bpf_line_info_min *rec;
5154
5155 err = visit(&sec->sec_name_off, ctx);
5156 if (err)
5157 return err;
5158
5159 for_each_btf_ext_rec(seg, sec, i, rec) {
5160 err = visit(&rec->file_name_off, ctx);
5161 if (err)
5162 return err;
5163 err = visit(&rec->line_off, ctx);
5164 if (err)
5165 return err;
5166 }
5167 }
5168
5169 seg = &btf_ext->core_relo_info;
5170 for_each_btf_ext_sec(seg, sec) {
5171 struct bpf_core_relo *rec;
5172
5173 err = visit(&sec->sec_name_off, ctx);
5174 if (err)
5175 return err;
5176
5177 for_each_btf_ext_rec(seg, sec, i, rec) {
5178 err = visit(&rec->access_str_off, ctx);
5179 if (err)
5180 return err;
5181 }
5182 }
5183
5184 return 0;
5185}
1// SPDX-License-Identifier: (LGPL-2.1 OR BSD-2-Clause)
2/* Copyright (c) 2018 Facebook */
3
4#include <byteswap.h>
5#include <endian.h>
6#include <stdio.h>
7#include <stdlib.h>
8#include <string.h>
9#include <fcntl.h>
10#include <unistd.h>
11#include <errno.h>
12#include <sys/utsname.h>
13#include <sys/param.h>
14#include <sys/stat.h>
15#include <linux/kernel.h>
16#include <linux/err.h>
17#include <linux/btf.h>
18#include <gelf.h>
19#include "btf.h"
20#include "bpf.h"
21#include "libbpf.h"
22#include "libbpf_internal.h"
23#include "hashmap.h"
24#include "strset.h"
25#include "str_error.h"
26
27#define BTF_MAX_NR_TYPES 0x7fffffffU
28#define BTF_MAX_STR_OFFSET 0x7fffffffU
29
30static struct btf_type btf_void;
31
32struct btf {
33 /* raw BTF data in native endianness */
34 void *raw_data;
35 /* raw BTF data in non-native endianness */
36 void *raw_data_swapped;
37 __u32 raw_size;
38 /* whether target endianness differs from the native one */
39 bool swapped_endian;
40
41 /*
42 * When BTF is loaded from an ELF or raw memory it is stored
43 * in a contiguous memory block. The hdr, type_data, and, strs_data
44 * point inside that memory region to their respective parts of BTF
45 * representation:
46 *
47 * +--------------------------------+
48 * | Header | Types | Strings |
49 * +--------------------------------+
50 * ^ ^ ^
51 * | | |
52 * hdr | |
53 * types_data-+ |
54 * strs_data------------+
55 *
56 * If BTF data is later modified, e.g., due to types added or
57 * removed, BTF deduplication performed, etc, this contiguous
58 * representation is broken up into three independently allocated
59 * memory regions to be able to modify them independently.
60 * raw_data is nulled out at that point, but can be later allocated
61 * and cached again if user calls btf__raw_data(), at which point
62 * raw_data will contain a contiguous copy of header, types, and
63 * strings:
64 *
65 * +----------+ +---------+ +-----------+
66 * | Header | | Types | | Strings |
67 * +----------+ +---------+ +-----------+
68 * ^ ^ ^
69 * | | |
70 * hdr | |
71 * types_data----+ |
72 * strset__data(strs_set)-----+
73 *
74 * +----------+---------+-----------+
75 * | Header | Types | Strings |
76 * raw_data----->+----------+---------+-----------+
77 */
78 struct btf_header *hdr;
79
80 void *types_data;
81 size_t types_data_cap; /* used size stored in hdr->type_len */
82
83 /* type ID to `struct btf_type *` lookup index
84 * type_offs[0] corresponds to the first non-VOID type:
85 * - for base BTF it's type [1];
86 * - for split BTF it's the first non-base BTF type.
87 */
88 __u32 *type_offs;
89 size_t type_offs_cap;
90 /* number of types in this BTF instance:
91 * - doesn't include special [0] void type;
92 * - for split BTF counts number of types added on top of base BTF.
93 */
94 __u32 nr_types;
95 /* if not NULL, points to the base BTF on top of which the current
96 * split BTF is based
97 */
98 struct btf *base_btf;
99 /* BTF type ID of the first type in this BTF instance:
100 * - for base BTF it's equal to 1;
101 * - for split BTF it's equal to biggest type ID of base BTF plus 1.
102 */
103 int start_id;
104 /* logical string offset of this BTF instance:
105 * - for base BTF it's equal to 0;
106 * - for split BTF it's equal to total size of base BTF's string section size.
107 */
108 int start_str_off;
109
110 /* only one of strs_data or strs_set can be non-NULL, depending on
111 * whether BTF is in a modifiable state (strs_set is used) or not
112 * (strs_data points inside raw_data)
113 */
114 void *strs_data;
115 /* a set of unique strings */
116 struct strset *strs_set;
117 /* whether strings are already deduplicated */
118 bool strs_deduped;
119
120 /* whether base_btf should be freed in btf_free for this instance */
121 bool owns_base;
122
123 /* BTF object FD, if loaded into kernel */
124 int fd;
125
126 /* Pointer size (in bytes) for a target architecture of this BTF */
127 int ptr_sz;
128};
129
130static inline __u64 ptr_to_u64(const void *ptr)
131{
132 return (__u64) (unsigned long) ptr;
133}
134
135/* Ensure given dynamically allocated memory region pointed to by *data* with
136 * capacity of *cap_cnt* elements each taking *elem_sz* bytes has enough
137 * memory to accommodate *add_cnt* new elements, assuming *cur_cnt* elements
138 * are already used. At most *max_cnt* elements can be ever allocated.
139 * If necessary, memory is reallocated and all existing data is copied over,
140 * new pointer to the memory region is stored at *data, new memory region
141 * capacity (in number of elements) is stored in *cap.
142 * On success, memory pointer to the beginning of unused memory is returned.
143 * On error, NULL is returned.
144 */
145void *libbpf_add_mem(void **data, size_t *cap_cnt, size_t elem_sz,
146 size_t cur_cnt, size_t max_cnt, size_t add_cnt)
147{
148 size_t new_cnt;
149 void *new_data;
150
151 if (cur_cnt + add_cnt <= *cap_cnt)
152 return *data + cur_cnt * elem_sz;
153
154 /* requested more than the set limit */
155 if (cur_cnt + add_cnt > max_cnt)
156 return NULL;
157
158 new_cnt = *cap_cnt;
159 new_cnt += new_cnt / 4; /* expand by 25% */
160 if (new_cnt < 16) /* but at least 16 elements */
161 new_cnt = 16;
162 if (new_cnt > max_cnt) /* but not exceeding a set limit */
163 new_cnt = max_cnt;
164 if (new_cnt < cur_cnt + add_cnt) /* also ensure we have enough memory */
165 new_cnt = cur_cnt + add_cnt;
166
167 new_data = libbpf_reallocarray(*data, new_cnt, elem_sz);
168 if (!new_data)
169 return NULL;
170
171 /* zero out newly allocated portion of memory */
172 memset(new_data + (*cap_cnt) * elem_sz, 0, (new_cnt - *cap_cnt) * elem_sz);
173
174 *data = new_data;
175 *cap_cnt = new_cnt;
176 return new_data + cur_cnt * elem_sz;
177}
178
179/* Ensure given dynamically allocated memory region has enough allocated space
180 * to accommodate *need_cnt* elements of size *elem_sz* bytes each
181 */
182int libbpf_ensure_mem(void **data, size_t *cap_cnt, size_t elem_sz, size_t need_cnt)
183{
184 void *p;
185
186 if (need_cnt <= *cap_cnt)
187 return 0;
188
189 p = libbpf_add_mem(data, cap_cnt, elem_sz, *cap_cnt, SIZE_MAX, need_cnt - *cap_cnt);
190 if (!p)
191 return -ENOMEM;
192
193 return 0;
194}
195
196static void *btf_add_type_offs_mem(struct btf *btf, size_t add_cnt)
197{
198 return libbpf_add_mem((void **)&btf->type_offs, &btf->type_offs_cap, sizeof(__u32),
199 btf->nr_types, BTF_MAX_NR_TYPES, add_cnt);
200}
201
202static int btf_add_type_idx_entry(struct btf *btf, __u32 type_off)
203{
204 __u32 *p;
205
206 p = btf_add_type_offs_mem(btf, 1);
207 if (!p)
208 return -ENOMEM;
209
210 *p = type_off;
211 return 0;
212}
213
214static void btf_bswap_hdr(struct btf_header *h)
215{
216 h->magic = bswap_16(h->magic);
217 h->hdr_len = bswap_32(h->hdr_len);
218 h->type_off = bswap_32(h->type_off);
219 h->type_len = bswap_32(h->type_len);
220 h->str_off = bswap_32(h->str_off);
221 h->str_len = bswap_32(h->str_len);
222}
223
224static int btf_parse_hdr(struct btf *btf)
225{
226 struct btf_header *hdr = btf->hdr;
227 __u32 meta_left;
228
229 if (btf->raw_size < sizeof(struct btf_header)) {
230 pr_debug("BTF header not found\n");
231 return -EINVAL;
232 }
233
234 if (hdr->magic == bswap_16(BTF_MAGIC)) {
235 btf->swapped_endian = true;
236 if (bswap_32(hdr->hdr_len) != sizeof(struct btf_header)) {
237 pr_warn("Can't load BTF with non-native endianness due to unsupported header length %u\n",
238 bswap_32(hdr->hdr_len));
239 return -ENOTSUP;
240 }
241 btf_bswap_hdr(hdr);
242 } else if (hdr->magic != BTF_MAGIC) {
243 pr_debug("Invalid BTF magic: %x\n", hdr->magic);
244 return -EINVAL;
245 }
246
247 if (btf->raw_size < hdr->hdr_len) {
248 pr_debug("BTF header len %u larger than data size %u\n",
249 hdr->hdr_len, btf->raw_size);
250 return -EINVAL;
251 }
252
253 meta_left = btf->raw_size - hdr->hdr_len;
254 if (meta_left < (long long)hdr->str_off + hdr->str_len) {
255 pr_debug("Invalid BTF total size: %u\n", btf->raw_size);
256 return -EINVAL;
257 }
258
259 if ((long long)hdr->type_off + hdr->type_len > hdr->str_off) {
260 pr_debug("Invalid BTF data sections layout: type data at %u + %u, strings data at %u + %u\n",
261 hdr->type_off, hdr->type_len, hdr->str_off, hdr->str_len);
262 return -EINVAL;
263 }
264
265 if (hdr->type_off % 4) {
266 pr_debug("BTF type section is not aligned to 4 bytes\n");
267 return -EINVAL;
268 }
269
270 return 0;
271}
272
273static int btf_parse_str_sec(struct btf *btf)
274{
275 const struct btf_header *hdr = btf->hdr;
276 const char *start = btf->strs_data;
277 const char *end = start + btf->hdr->str_len;
278
279 if (btf->base_btf && hdr->str_len == 0)
280 return 0;
281 if (!hdr->str_len || hdr->str_len - 1 > BTF_MAX_STR_OFFSET || end[-1]) {
282 pr_debug("Invalid BTF string section\n");
283 return -EINVAL;
284 }
285 if (!btf->base_btf && start[0]) {
286 pr_debug("Invalid BTF string section\n");
287 return -EINVAL;
288 }
289 return 0;
290}
291
292static int btf_type_size(const struct btf_type *t)
293{
294 const int base_size = sizeof(struct btf_type);
295 __u16 vlen = btf_vlen(t);
296
297 switch (btf_kind(t)) {
298 case BTF_KIND_FWD:
299 case BTF_KIND_CONST:
300 case BTF_KIND_VOLATILE:
301 case BTF_KIND_RESTRICT:
302 case BTF_KIND_PTR:
303 case BTF_KIND_TYPEDEF:
304 case BTF_KIND_FUNC:
305 case BTF_KIND_FLOAT:
306 case BTF_KIND_TYPE_TAG:
307 return base_size;
308 case BTF_KIND_INT:
309 return base_size + sizeof(__u32);
310 case BTF_KIND_ENUM:
311 return base_size + vlen * sizeof(struct btf_enum);
312 case BTF_KIND_ENUM64:
313 return base_size + vlen * sizeof(struct btf_enum64);
314 case BTF_KIND_ARRAY:
315 return base_size + sizeof(struct btf_array);
316 case BTF_KIND_STRUCT:
317 case BTF_KIND_UNION:
318 return base_size + vlen * sizeof(struct btf_member);
319 case BTF_KIND_FUNC_PROTO:
320 return base_size + vlen * sizeof(struct btf_param);
321 case BTF_KIND_VAR:
322 return base_size + sizeof(struct btf_var);
323 case BTF_KIND_DATASEC:
324 return base_size + vlen * sizeof(struct btf_var_secinfo);
325 case BTF_KIND_DECL_TAG:
326 return base_size + sizeof(struct btf_decl_tag);
327 default:
328 pr_debug("Unsupported BTF_KIND:%u\n", btf_kind(t));
329 return -EINVAL;
330 }
331}
332
333static void btf_bswap_type_base(struct btf_type *t)
334{
335 t->name_off = bswap_32(t->name_off);
336 t->info = bswap_32(t->info);
337 t->type = bswap_32(t->type);
338}
339
340static int btf_bswap_type_rest(struct btf_type *t)
341{
342 struct btf_var_secinfo *v;
343 struct btf_enum64 *e64;
344 struct btf_member *m;
345 struct btf_array *a;
346 struct btf_param *p;
347 struct btf_enum *e;
348 __u16 vlen = btf_vlen(t);
349 int i;
350
351 switch (btf_kind(t)) {
352 case BTF_KIND_FWD:
353 case BTF_KIND_CONST:
354 case BTF_KIND_VOLATILE:
355 case BTF_KIND_RESTRICT:
356 case BTF_KIND_PTR:
357 case BTF_KIND_TYPEDEF:
358 case BTF_KIND_FUNC:
359 case BTF_KIND_FLOAT:
360 case BTF_KIND_TYPE_TAG:
361 return 0;
362 case BTF_KIND_INT:
363 *(__u32 *)(t + 1) = bswap_32(*(__u32 *)(t + 1));
364 return 0;
365 case BTF_KIND_ENUM:
366 for (i = 0, e = btf_enum(t); i < vlen; i++, e++) {
367 e->name_off = bswap_32(e->name_off);
368 e->val = bswap_32(e->val);
369 }
370 return 0;
371 case BTF_KIND_ENUM64:
372 for (i = 0, e64 = btf_enum64(t); i < vlen; i++, e64++) {
373 e64->name_off = bswap_32(e64->name_off);
374 e64->val_lo32 = bswap_32(e64->val_lo32);
375 e64->val_hi32 = bswap_32(e64->val_hi32);
376 }
377 return 0;
378 case BTF_KIND_ARRAY:
379 a = btf_array(t);
380 a->type = bswap_32(a->type);
381 a->index_type = bswap_32(a->index_type);
382 a->nelems = bswap_32(a->nelems);
383 return 0;
384 case BTF_KIND_STRUCT:
385 case BTF_KIND_UNION:
386 for (i = 0, m = btf_members(t); i < vlen; i++, m++) {
387 m->name_off = bswap_32(m->name_off);
388 m->type = bswap_32(m->type);
389 m->offset = bswap_32(m->offset);
390 }
391 return 0;
392 case BTF_KIND_FUNC_PROTO:
393 for (i = 0, p = btf_params(t); i < vlen; i++, p++) {
394 p->name_off = bswap_32(p->name_off);
395 p->type = bswap_32(p->type);
396 }
397 return 0;
398 case BTF_KIND_VAR:
399 btf_var(t)->linkage = bswap_32(btf_var(t)->linkage);
400 return 0;
401 case BTF_KIND_DATASEC:
402 for (i = 0, v = btf_var_secinfos(t); i < vlen; i++, v++) {
403 v->type = bswap_32(v->type);
404 v->offset = bswap_32(v->offset);
405 v->size = bswap_32(v->size);
406 }
407 return 0;
408 case BTF_KIND_DECL_TAG:
409 btf_decl_tag(t)->component_idx = bswap_32(btf_decl_tag(t)->component_idx);
410 return 0;
411 default:
412 pr_debug("Unsupported BTF_KIND:%u\n", btf_kind(t));
413 return -EINVAL;
414 }
415}
416
417static int btf_parse_type_sec(struct btf *btf)
418{
419 struct btf_header *hdr = btf->hdr;
420 void *next_type = btf->types_data;
421 void *end_type = next_type + hdr->type_len;
422 int err, type_size;
423
424 while (next_type + sizeof(struct btf_type) <= end_type) {
425 if (btf->swapped_endian)
426 btf_bswap_type_base(next_type);
427
428 type_size = btf_type_size(next_type);
429 if (type_size < 0)
430 return type_size;
431 if (next_type + type_size > end_type) {
432 pr_warn("BTF type [%d] is malformed\n", btf->start_id + btf->nr_types);
433 return -EINVAL;
434 }
435
436 if (btf->swapped_endian && btf_bswap_type_rest(next_type))
437 return -EINVAL;
438
439 err = btf_add_type_idx_entry(btf, next_type - btf->types_data);
440 if (err)
441 return err;
442
443 next_type += type_size;
444 btf->nr_types++;
445 }
446
447 if (next_type != end_type) {
448 pr_warn("BTF types data is malformed\n");
449 return -EINVAL;
450 }
451
452 return 0;
453}
454
455static int btf_validate_str(const struct btf *btf, __u32 str_off, const char *what, __u32 type_id)
456{
457 const char *s;
458
459 s = btf__str_by_offset(btf, str_off);
460 if (!s) {
461 pr_warn("btf: type [%u]: invalid %s (string offset %u)\n", type_id, what, str_off);
462 return -EINVAL;
463 }
464
465 return 0;
466}
467
468static int btf_validate_id(const struct btf *btf, __u32 id, __u32 ctx_id)
469{
470 const struct btf_type *t;
471
472 t = btf__type_by_id(btf, id);
473 if (!t) {
474 pr_warn("btf: type [%u]: invalid referenced type ID %u\n", ctx_id, id);
475 return -EINVAL;
476 }
477
478 return 0;
479}
480
481static int btf_validate_type(const struct btf *btf, const struct btf_type *t, __u32 id)
482{
483 __u32 kind = btf_kind(t);
484 int err, i, n;
485
486 err = btf_validate_str(btf, t->name_off, "type name", id);
487 if (err)
488 return err;
489
490 switch (kind) {
491 case BTF_KIND_UNKN:
492 case BTF_KIND_INT:
493 case BTF_KIND_FWD:
494 case BTF_KIND_FLOAT:
495 break;
496 case BTF_KIND_PTR:
497 case BTF_KIND_TYPEDEF:
498 case BTF_KIND_VOLATILE:
499 case BTF_KIND_CONST:
500 case BTF_KIND_RESTRICT:
501 case BTF_KIND_VAR:
502 case BTF_KIND_DECL_TAG:
503 case BTF_KIND_TYPE_TAG:
504 err = btf_validate_id(btf, t->type, id);
505 if (err)
506 return err;
507 break;
508 case BTF_KIND_ARRAY: {
509 const struct btf_array *a = btf_array(t);
510
511 err = btf_validate_id(btf, a->type, id);
512 err = err ?: btf_validate_id(btf, a->index_type, id);
513 if (err)
514 return err;
515 break;
516 }
517 case BTF_KIND_STRUCT:
518 case BTF_KIND_UNION: {
519 const struct btf_member *m = btf_members(t);
520
521 n = btf_vlen(t);
522 for (i = 0; i < n; i++, m++) {
523 err = btf_validate_str(btf, m->name_off, "field name", id);
524 err = err ?: btf_validate_id(btf, m->type, id);
525 if (err)
526 return err;
527 }
528 break;
529 }
530 case BTF_KIND_ENUM: {
531 const struct btf_enum *m = btf_enum(t);
532
533 n = btf_vlen(t);
534 for (i = 0; i < n; i++, m++) {
535 err = btf_validate_str(btf, m->name_off, "enum name", id);
536 if (err)
537 return err;
538 }
539 break;
540 }
541 case BTF_KIND_ENUM64: {
542 const struct btf_enum64 *m = btf_enum64(t);
543
544 n = btf_vlen(t);
545 for (i = 0; i < n; i++, m++) {
546 err = btf_validate_str(btf, m->name_off, "enum name", id);
547 if (err)
548 return err;
549 }
550 break;
551 }
552 case BTF_KIND_FUNC: {
553 const struct btf_type *ft;
554
555 err = btf_validate_id(btf, t->type, id);
556 if (err)
557 return err;
558 ft = btf__type_by_id(btf, t->type);
559 if (btf_kind(ft) != BTF_KIND_FUNC_PROTO) {
560 pr_warn("btf: type [%u]: referenced type [%u] is not FUNC_PROTO\n", id, t->type);
561 return -EINVAL;
562 }
563 break;
564 }
565 case BTF_KIND_FUNC_PROTO: {
566 const struct btf_param *m = btf_params(t);
567
568 n = btf_vlen(t);
569 for (i = 0; i < n; i++, m++) {
570 err = btf_validate_str(btf, m->name_off, "param name", id);
571 err = err ?: btf_validate_id(btf, m->type, id);
572 if (err)
573 return err;
574 }
575 break;
576 }
577 case BTF_KIND_DATASEC: {
578 const struct btf_var_secinfo *m = btf_var_secinfos(t);
579
580 n = btf_vlen(t);
581 for (i = 0; i < n; i++, m++) {
582 err = btf_validate_id(btf, m->type, id);
583 if (err)
584 return err;
585 }
586 break;
587 }
588 default:
589 pr_warn("btf: type [%u]: unrecognized kind %u\n", id, kind);
590 return -EINVAL;
591 }
592 return 0;
593}
594
595/* Validate basic sanity of BTF. It's intentionally less thorough than
596 * kernel's validation and validates only properties of BTF that libbpf relies
597 * on to be correct (e.g., valid type IDs, valid string offsets, etc)
598 */
599static int btf_sanity_check(const struct btf *btf)
600{
601 const struct btf_type *t;
602 __u32 i, n = btf__type_cnt(btf);
603 int err;
604
605 for (i = btf->start_id; i < n; i++) {
606 t = btf_type_by_id(btf, i);
607 err = btf_validate_type(btf, t, i);
608 if (err)
609 return err;
610 }
611 return 0;
612}
613
614__u32 btf__type_cnt(const struct btf *btf)
615{
616 return btf->start_id + btf->nr_types;
617}
618
619const struct btf *btf__base_btf(const struct btf *btf)
620{
621 return btf->base_btf;
622}
623
624/* internal helper returning non-const pointer to a type */
625struct btf_type *btf_type_by_id(const struct btf *btf, __u32 type_id)
626{
627 if (type_id == 0)
628 return &btf_void;
629 if (type_id < btf->start_id)
630 return btf_type_by_id(btf->base_btf, type_id);
631 return btf->types_data + btf->type_offs[type_id - btf->start_id];
632}
633
634const struct btf_type *btf__type_by_id(const struct btf *btf, __u32 type_id)
635{
636 if (type_id >= btf->start_id + btf->nr_types)
637 return errno = EINVAL, NULL;
638 return btf_type_by_id((struct btf *)btf, type_id);
639}
640
641static int determine_ptr_size(const struct btf *btf)
642{
643 static const char * const long_aliases[] = {
644 "long",
645 "long int",
646 "int long",
647 "unsigned long",
648 "long unsigned",
649 "unsigned long int",
650 "unsigned int long",
651 "long unsigned int",
652 "long int unsigned",
653 "int unsigned long",
654 "int long unsigned",
655 };
656 const struct btf_type *t;
657 const char *name;
658 int i, j, n;
659
660 if (btf->base_btf && btf->base_btf->ptr_sz > 0)
661 return btf->base_btf->ptr_sz;
662
663 n = btf__type_cnt(btf);
664 for (i = 1; i < n; i++) {
665 t = btf__type_by_id(btf, i);
666 if (!btf_is_int(t))
667 continue;
668
669 if (t->size != 4 && t->size != 8)
670 continue;
671
672 name = btf__name_by_offset(btf, t->name_off);
673 if (!name)
674 continue;
675
676 for (j = 0; j < ARRAY_SIZE(long_aliases); j++) {
677 if (strcmp(name, long_aliases[j]) == 0)
678 return t->size;
679 }
680 }
681
682 return -1;
683}
684
685static size_t btf_ptr_sz(const struct btf *btf)
686{
687 if (!btf->ptr_sz)
688 ((struct btf *)btf)->ptr_sz = determine_ptr_size(btf);
689 return btf->ptr_sz < 0 ? sizeof(void *) : btf->ptr_sz;
690}
691
692/* Return pointer size this BTF instance assumes. The size is heuristically
693 * determined by looking for 'long' or 'unsigned long' integer type and
694 * recording its size in bytes. If BTF type information doesn't have any such
695 * type, this function returns 0. In the latter case, native architecture's
696 * pointer size is assumed, so will be either 4 or 8, depending on
697 * architecture that libbpf was compiled for. It's possible to override
698 * guessed value by using btf__set_pointer_size() API.
699 */
700size_t btf__pointer_size(const struct btf *btf)
701{
702 if (!btf->ptr_sz)
703 ((struct btf *)btf)->ptr_sz = determine_ptr_size(btf);
704
705 if (btf->ptr_sz < 0)
706 /* not enough BTF type info to guess */
707 return 0;
708
709 return btf->ptr_sz;
710}
711
712/* Override or set pointer size in bytes. Only values of 4 and 8 are
713 * supported.
714 */
715int btf__set_pointer_size(struct btf *btf, size_t ptr_sz)
716{
717 if (ptr_sz != 4 && ptr_sz != 8)
718 return libbpf_err(-EINVAL);
719 btf->ptr_sz = ptr_sz;
720 return 0;
721}
722
723static bool is_host_big_endian(void)
724{
725#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
726 return false;
727#elif __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
728 return true;
729#else
730# error "Unrecognized __BYTE_ORDER__"
731#endif
732}
733
734enum btf_endianness btf__endianness(const struct btf *btf)
735{
736 if (is_host_big_endian())
737 return btf->swapped_endian ? BTF_LITTLE_ENDIAN : BTF_BIG_ENDIAN;
738 else
739 return btf->swapped_endian ? BTF_BIG_ENDIAN : BTF_LITTLE_ENDIAN;
740}
741
742int btf__set_endianness(struct btf *btf, enum btf_endianness endian)
743{
744 if (endian != BTF_LITTLE_ENDIAN && endian != BTF_BIG_ENDIAN)
745 return libbpf_err(-EINVAL);
746
747 btf->swapped_endian = is_host_big_endian() != (endian == BTF_BIG_ENDIAN);
748 if (!btf->swapped_endian) {
749 free(btf->raw_data_swapped);
750 btf->raw_data_swapped = NULL;
751 }
752 return 0;
753}
754
755static bool btf_type_is_void(const struct btf_type *t)
756{
757 return t == &btf_void || btf_is_fwd(t);
758}
759
760static bool btf_type_is_void_or_null(const struct btf_type *t)
761{
762 return !t || btf_type_is_void(t);
763}
764
765#define MAX_RESOLVE_DEPTH 32
766
767__s64 btf__resolve_size(const struct btf *btf, __u32 type_id)
768{
769 const struct btf_array *array;
770 const struct btf_type *t;
771 __u32 nelems = 1;
772 __s64 size = -1;
773 int i;
774
775 t = btf__type_by_id(btf, type_id);
776 for (i = 0; i < MAX_RESOLVE_DEPTH && !btf_type_is_void_or_null(t); i++) {
777 switch (btf_kind(t)) {
778 case BTF_KIND_INT:
779 case BTF_KIND_STRUCT:
780 case BTF_KIND_UNION:
781 case BTF_KIND_ENUM:
782 case BTF_KIND_ENUM64:
783 case BTF_KIND_DATASEC:
784 case BTF_KIND_FLOAT:
785 size = t->size;
786 goto done;
787 case BTF_KIND_PTR:
788 size = btf_ptr_sz(btf);
789 goto done;
790 case BTF_KIND_TYPEDEF:
791 case BTF_KIND_VOLATILE:
792 case BTF_KIND_CONST:
793 case BTF_KIND_RESTRICT:
794 case BTF_KIND_VAR:
795 case BTF_KIND_DECL_TAG:
796 case BTF_KIND_TYPE_TAG:
797 type_id = t->type;
798 break;
799 case BTF_KIND_ARRAY:
800 array = btf_array(t);
801 if (nelems && array->nelems > UINT32_MAX / nelems)
802 return libbpf_err(-E2BIG);
803 nelems *= array->nelems;
804 type_id = array->type;
805 break;
806 default:
807 return libbpf_err(-EINVAL);
808 }
809
810 t = btf__type_by_id(btf, type_id);
811 }
812
813done:
814 if (size < 0)
815 return libbpf_err(-EINVAL);
816 if (nelems && size > UINT32_MAX / nelems)
817 return libbpf_err(-E2BIG);
818
819 return nelems * size;
820}
821
822int btf__align_of(const struct btf *btf, __u32 id)
823{
824 const struct btf_type *t = btf__type_by_id(btf, id);
825 __u16 kind = btf_kind(t);
826
827 switch (kind) {
828 case BTF_KIND_INT:
829 case BTF_KIND_ENUM:
830 case BTF_KIND_ENUM64:
831 case BTF_KIND_FLOAT:
832 return min(btf_ptr_sz(btf), (size_t)t->size);
833 case BTF_KIND_PTR:
834 return btf_ptr_sz(btf);
835 case BTF_KIND_TYPEDEF:
836 case BTF_KIND_VOLATILE:
837 case BTF_KIND_CONST:
838 case BTF_KIND_RESTRICT:
839 case BTF_KIND_TYPE_TAG:
840 return btf__align_of(btf, t->type);
841 case BTF_KIND_ARRAY:
842 return btf__align_of(btf, btf_array(t)->type);
843 case BTF_KIND_STRUCT:
844 case BTF_KIND_UNION: {
845 const struct btf_member *m = btf_members(t);
846 __u16 vlen = btf_vlen(t);
847 int i, max_align = 1, align;
848
849 for (i = 0; i < vlen; i++, m++) {
850 align = btf__align_of(btf, m->type);
851 if (align <= 0)
852 return libbpf_err(align);
853 max_align = max(max_align, align);
854
855 /* if field offset isn't aligned according to field
856 * type's alignment, then struct must be packed
857 */
858 if (btf_member_bitfield_size(t, i) == 0 &&
859 (m->offset % (8 * align)) != 0)
860 return 1;
861 }
862
863 /* if struct/union size isn't a multiple of its alignment,
864 * then struct must be packed
865 */
866 if ((t->size % max_align) != 0)
867 return 1;
868
869 return max_align;
870 }
871 default:
872 pr_warn("unsupported BTF_KIND:%u\n", btf_kind(t));
873 return errno = EINVAL, 0;
874 }
875}
876
877int btf__resolve_type(const struct btf *btf, __u32 type_id)
878{
879 const struct btf_type *t;
880 int depth = 0;
881
882 t = btf__type_by_id(btf, type_id);
883 while (depth < MAX_RESOLVE_DEPTH &&
884 !btf_type_is_void_or_null(t) &&
885 (btf_is_mod(t) || btf_is_typedef(t) || btf_is_var(t))) {
886 type_id = t->type;
887 t = btf__type_by_id(btf, type_id);
888 depth++;
889 }
890
891 if (depth == MAX_RESOLVE_DEPTH || btf_type_is_void_or_null(t))
892 return libbpf_err(-EINVAL);
893
894 return type_id;
895}
896
897__s32 btf__find_by_name(const struct btf *btf, const char *type_name)
898{
899 __u32 i, nr_types = btf__type_cnt(btf);
900
901 if (!strcmp(type_name, "void"))
902 return 0;
903
904 for (i = 1; i < nr_types; i++) {
905 const struct btf_type *t = btf__type_by_id(btf, i);
906 const char *name = btf__name_by_offset(btf, t->name_off);
907
908 if (name && !strcmp(type_name, name))
909 return i;
910 }
911
912 return libbpf_err(-ENOENT);
913}
914
915static __s32 btf_find_by_name_kind(const struct btf *btf, int start_id,
916 const char *type_name, __u32 kind)
917{
918 __u32 i, nr_types = btf__type_cnt(btf);
919
920 if (kind == BTF_KIND_UNKN || !strcmp(type_name, "void"))
921 return 0;
922
923 for (i = start_id; i < nr_types; i++) {
924 const struct btf_type *t = btf__type_by_id(btf, i);
925 const char *name;
926
927 if (btf_kind(t) != kind)
928 continue;
929 name = btf__name_by_offset(btf, t->name_off);
930 if (name && !strcmp(type_name, name))
931 return i;
932 }
933
934 return libbpf_err(-ENOENT);
935}
936
937__s32 btf__find_by_name_kind_own(const struct btf *btf, const char *type_name,
938 __u32 kind)
939{
940 return btf_find_by_name_kind(btf, btf->start_id, type_name, kind);
941}
942
943__s32 btf__find_by_name_kind(const struct btf *btf, const char *type_name,
944 __u32 kind)
945{
946 return btf_find_by_name_kind(btf, 1, type_name, kind);
947}
948
949static bool btf_is_modifiable(const struct btf *btf)
950{
951 return (void *)btf->hdr != btf->raw_data;
952}
953
954void btf__free(struct btf *btf)
955{
956 if (IS_ERR_OR_NULL(btf))
957 return;
958
959 if (btf->fd >= 0)
960 close(btf->fd);
961
962 if (btf_is_modifiable(btf)) {
963 /* if BTF was modified after loading, it will have a split
964 * in-memory representation for header, types, and strings
965 * sections, so we need to free all of them individually. It
966 * might still have a cached contiguous raw data present,
967 * which will be unconditionally freed below.
968 */
969 free(btf->hdr);
970 free(btf->types_data);
971 strset__free(btf->strs_set);
972 }
973 free(btf->raw_data);
974 free(btf->raw_data_swapped);
975 free(btf->type_offs);
976 if (btf->owns_base)
977 btf__free(btf->base_btf);
978 free(btf);
979}
980
981static struct btf *btf_new_empty(struct btf *base_btf)
982{
983 struct btf *btf;
984
985 btf = calloc(1, sizeof(*btf));
986 if (!btf)
987 return ERR_PTR(-ENOMEM);
988
989 btf->nr_types = 0;
990 btf->start_id = 1;
991 btf->start_str_off = 0;
992 btf->fd = -1;
993 btf->ptr_sz = sizeof(void *);
994 btf->swapped_endian = false;
995
996 if (base_btf) {
997 btf->base_btf = base_btf;
998 btf->start_id = btf__type_cnt(base_btf);
999 btf->start_str_off = base_btf->hdr->str_len;
1000 btf->swapped_endian = base_btf->swapped_endian;
1001 }
1002
1003 /* +1 for empty string at offset 0 */
1004 btf->raw_size = sizeof(struct btf_header) + (base_btf ? 0 : 1);
1005 btf->raw_data = calloc(1, btf->raw_size);
1006 if (!btf->raw_data) {
1007 free(btf);
1008 return ERR_PTR(-ENOMEM);
1009 }
1010
1011 btf->hdr = btf->raw_data;
1012 btf->hdr->hdr_len = sizeof(struct btf_header);
1013 btf->hdr->magic = BTF_MAGIC;
1014 btf->hdr->version = BTF_VERSION;
1015
1016 btf->types_data = btf->raw_data + btf->hdr->hdr_len;
1017 btf->strs_data = btf->raw_data + btf->hdr->hdr_len;
1018 btf->hdr->str_len = base_btf ? 0 : 1; /* empty string at offset 0 */
1019
1020 return btf;
1021}
1022
1023struct btf *btf__new_empty(void)
1024{
1025 return libbpf_ptr(btf_new_empty(NULL));
1026}
1027
1028struct btf *btf__new_empty_split(struct btf *base_btf)
1029{
1030 return libbpf_ptr(btf_new_empty(base_btf));
1031}
1032
1033static struct btf *btf_new(const void *data, __u32 size, struct btf *base_btf)
1034{
1035 struct btf *btf;
1036 int err;
1037
1038 btf = calloc(1, sizeof(struct btf));
1039 if (!btf)
1040 return ERR_PTR(-ENOMEM);
1041
1042 btf->nr_types = 0;
1043 btf->start_id = 1;
1044 btf->start_str_off = 0;
1045 btf->fd = -1;
1046
1047 if (base_btf) {
1048 btf->base_btf = base_btf;
1049 btf->start_id = btf__type_cnt(base_btf);
1050 btf->start_str_off = base_btf->hdr->str_len;
1051 }
1052
1053 btf->raw_data = malloc(size);
1054 if (!btf->raw_data) {
1055 err = -ENOMEM;
1056 goto done;
1057 }
1058 memcpy(btf->raw_data, data, size);
1059 btf->raw_size = size;
1060
1061 btf->hdr = btf->raw_data;
1062 err = btf_parse_hdr(btf);
1063 if (err)
1064 goto done;
1065
1066 btf->strs_data = btf->raw_data + btf->hdr->hdr_len + btf->hdr->str_off;
1067 btf->types_data = btf->raw_data + btf->hdr->hdr_len + btf->hdr->type_off;
1068
1069 err = btf_parse_str_sec(btf);
1070 err = err ?: btf_parse_type_sec(btf);
1071 err = err ?: btf_sanity_check(btf);
1072 if (err)
1073 goto done;
1074
1075done:
1076 if (err) {
1077 btf__free(btf);
1078 return ERR_PTR(err);
1079 }
1080
1081 return btf;
1082}
1083
1084struct btf *btf__new(const void *data, __u32 size)
1085{
1086 return libbpf_ptr(btf_new(data, size, NULL));
1087}
1088
1089struct btf *btf__new_split(const void *data, __u32 size, struct btf *base_btf)
1090{
1091 return libbpf_ptr(btf_new(data, size, base_btf));
1092}
1093
1094struct btf_elf_secs {
1095 Elf_Data *btf_data;
1096 Elf_Data *btf_ext_data;
1097 Elf_Data *btf_base_data;
1098};
1099
1100static int btf_find_elf_sections(Elf *elf, const char *path, struct btf_elf_secs *secs)
1101{
1102 Elf_Scn *scn = NULL;
1103 Elf_Data *data;
1104 GElf_Ehdr ehdr;
1105 size_t shstrndx;
1106 int idx = 0;
1107
1108 if (!gelf_getehdr(elf, &ehdr)) {
1109 pr_warn("failed to get EHDR from %s\n", path);
1110 goto err;
1111 }
1112
1113 if (elf_getshdrstrndx(elf, &shstrndx)) {
1114 pr_warn("failed to get section names section index for %s\n",
1115 path);
1116 goto err;
1117 }
1118
1119 if (!elf_rawdata(elf_getscn(elf, shstrndx), NULL)) {
1120 pr_warn("failed to get e_shstrndx from %s\n", path);
1121 goto err;
1122 }
1123
1124 while ((scn = elf_nextscn(elf, scn)) != NULL) {
1125 Elf_Data **field;
1126 GElf_Shdr sh;
1127 char *name;
1128
1129 idx++;
1130 if (gelf_getshdr(scn, &sh) != &sh) {
1131 pr_warn("failed to get section(%d) header from %s\n",
1132 idx, path);
1133 goto err;
1134 }
1135 name = elf_strptr(elf, shstrndx, sh.sh_name);
1136 if (!name) {
1137 pr_warn("failed to get section(%d) name from %s\n",
1138 idx, path);
1139 goto err;
1140 }
1141
1142 if (strcmp(name, BTF_ELF_SEC) == 0)
1143 field = &secs->btf_data;
1144 else if (strcmp(name, BTF_EXT_ELF_SEC) == 0)
1145 field = &secs->btf_ext_data;
1146 else if (strcmp(name, BTF_BASE_ELF_SEC) == 0)
1147 field = &secs->btf_base_data;
1148 else
1149 continue;
1150
1151 data = elf_getdata(scn, 0);
1152 if (!data) {
1153 pr_warn("failed to get section(%d, %s) data from %s\n",
1154 idx, name, path);
1155 goto err;
1156 }
1157 *field = data;
1158 }
1159
1160 return 0;
1161
1162err:
1163 return -LIBBPF_ERRNO__FORMAT;
1164}
1165
1166static struct btf *btf_parse_elf(const char *path, struct btf *base_btf,
1167 struct btf_ext **btf_ext)
1168{
1169 struct btf_elf_secs secs = {};
1170 struct btf *dist_base_btf = NULL;
1171 struct btf *btf = NULL;
1172 int err = 0, fd = -1;
1173 Elf *elf = NULL;
1174
1175 if (elf_version(EV_CURRENT) == EV_NONE) {
1176 pr_warn("failed to init libelf for %s\n", path);
1177 return ERR_PTR(-LIBBPF_ERRNO__LIBELF);
1178 }
1179
1180 fd = open(path, O_RDONLY | O_CLOEXEC);
1181 if (fd < 0) {
1182 err = -errno;
1183 pr_warn("failed to open %s: %s\n", path, errstr(err));
1184 return ERR_PTR(err);
1185 }
1186
1187 elf = elf_begin(fd, ELF_C_READ, NULL);
1188 if (!elf) {
1189 err = -LIBBPF_ERRNO__FORMAT;
1190 pr_warn("failed to open %s as ELF file\n", path);
1191 goto done;
1192 }
1193
1194 err = btf_find_elf_sections(elf, path, &secs);
1195 if (err)
1196 goto done;
1197
1198 if (!secs.btf_data) {
1199 pr_warn("failed to find '%s' ELF section in %s\n", BTF_ELF_SEC, path);
1200 err = -ENODATA;
1201 goto done;
1202 }
1203
1204 if (secs.btf_base_data) {
1205 dist_base_btf = btf_new(secs.btf_base_data->d_buf, secs.btf_base_data->d_size,
1206 NULL);
1207 if (IS_ERR(dist_base_btf)) {
1208 err = PTR_ERR(dist_base_btf);
1209 dist_base_btf = NULL;
1210 goto done;
1211 }
1212 }
1213
1214 btf = btf_new(secs.btf_data->d_buf, secs.btf_data->d_size,
1215 dist_base_btf ?: base_btf);
1216 if (IS_ERR(btf)) {
1217 err = PTR_ERR(btf);
1218 goto done;
1219 }
1220 if (dist_base_btf && base_btf) {
1221 err = btf__relocate(btf, base_btf);
1222 if (err)
1223 goto done;
1224 btf__free(dist_base_btf);
1225 dist_base_btf = NULL;
1226 }
1227
1228 if (dist_base_btf)
1229 btf->owns_base = true;
1230
1231 switch (gelf_getclass(elf)) {
1232 case ELFCLASS32:
1233 btf__set_pointer_size(btf, 4);
1234 break;
1235 case ELFCLASS64:
1236 btf__set_pointer_size(btf, 8);
1237 break;
1238 default:
1239 pr_warn("failed to get ELF class (bitness) for %s\n", path);
1240 break;
1241 }
1242
1243 if (btf_ext && secs.btf_ext_data) {
1244 *btf_ext = btf_ext__new(secs.btf_ext_data->d_buf, secs.btf_ext_data->d_size);
1245 if (IS_ERR(*btf_ext)) {
1246 err = PTR_ERR(*btf_ext);
1247 goto done;
1248 }
1249 } else if (btf_ext) {
1250 *btf_ext = NULL;
1251 }
1252done:
1253 if (elf)
1254 elf_end(elf);
1255 close(fd);
1256
1257 if (!err)
1258 return btf;
1259
1260 if (btf_ext)
1261 btf_ext__free(*btf_ext);
1262 btf__free(dist_base_btf);
1263 btf__free(btf);
1264
1265 return ERR_PTR(err);
1266}
1267
1268struct btf *btf__parse_elf(const char *path, struct btf_ext **btf_ext)
1269{
1270 return libbpf_ptr(btf_parse_elf(path, NULL, btf_ext));
1271}
1272
1273struct btf *btf__parse_elf_split(const char *path, struct btf *base_btf)
1274{
1275 return libbpf_ptr(btf_parse_elf(path, base_btf, NULL));
1276}
1277
1278static struct btf *btf_parse_raw(const char *path, struct btf *base_btf)
1279{
1280 struct btf *btf = NULL;
1281 void *data = NULL;
1282 FILE *f = NULL;
1283 __u16 magic;
1284 int err = 0;
1285 long sz;
1286
1287 f = fopen(path, "rbe");
1288 if (!f) {
1289 err = -errno;
1290 goto err_out;
1291 }
1292
1293 /* check BTF magic */
1294 if (fread(&magic, 1, sizeof(magic), f) < sizeof(magic)) {
1295 err = -EIO;
1296 goto err_out;
1297 }
1298 if (magic != BTF_MAGIC && magic != bswap_16(BTF_MAGIC)) {
1299 /* definitely not a raw BTF */
1300 err = -EPROTO;
1301 goto err_out;
1302 }
1303
1304 /* get file size */
1305 if (fseek(f, 0, SEEK_END)) {
1306 err = -errno;
1307 goto err_out;
1308 }
1309 sz = ftell(f);
1310 if (sz < 0) {
1311 err = -errno;
1312 goto err_out;
1313 }
1314 /* rewind to the start */
1315 if (fseek(f, 0, SEEK_SET)) {
1316 err = -errno;
1317 goto err_out;
1318 }
1319
1320 /* pre-alloc memory and read all of BTF data */
1321 data = malloc(sz);
1322 if (!data) {
1323 err = -ENOMEM;
1324 goto err_out;
1325 }
1326 if (fread(data, 1, sz, f) < sz) {
1327 err = -EIO;
1328 goto err_out;
1329 }
1330
1331 /* finally parse BTF data */
1332 btf = btf_new(data, sz, base_btf);
1333
1334err_out:
1335 free(data);
1336 if (f)
1337 fclose(f);
1338 return err ? ERR_PTR(err) : btf;
1339}
1340
1341struct btf *btf__parse_raw(const char *path)
1342{
1343 return libbpf_ptr(btf_parse_raw(path, NULL));
1344}
1345
1346struct btf *btf__parse_raw_split(const char *path, struct btf *base_btf)
1347{
1348 return libbpf_ptr(btf_parse_raw(path, base_btf));
1349}
1350
1351static struct btf *btf_parse(const char *path, struct btf *base_btf, struct btf_ext **btf_ext)
1352{
1353 struct btf *btf;
1354 int err;
1355
1356 if (btf_ext)
1357 *btf_ext = NULL;
1358
1359 btf = btf_parse_raw(path, base_btf);
1360 err = libbpf_get_error(btf);
1361 if (!err)
1362 return btf;
1363 if (err != -EPROTO)
1364 return ERR_PTR(err);
1365 return btf_parse_elf(path, base_btf, btf_ext);
1366}
1367
1368struct btf *btf__parse(const char *path, struct btf_ext **btf_ext)
1369{
1370 return libbpf_ptr(btf_parse(path, NULL, btf_ext));
1371}
1372
1373struct btf *btf__parse_split(const char *path, struct btf *base_btf)
1374{
1375 return libbpf_ptr(btf_parse(path, base_btf, NULL));
1376}
1377
1378static void *btf_get_raw_data(const struct btf *btf, __u32 *size, bool swap_endian);
1379
1380int btf_load_into_kernel(struct btf *btf,
1381 char *log_buf, size_t log_sz, __u32 log_level,
1382 int token_fd)
1383{
1384 LIBBPF_OPTS(bpf_btf_load_opts, opts);
1385 __u32 buf_sz = 0, raw_size;
1386 char *buf = NULL, *tmp;
1387 void *raw_data;
1388 int err = 0;
1389
1390 if (btf->fd >= 0)
1391 return libbpf_err(-EEXIST);
1392 if (log_sz && !log_buf)
1393 return libbpf_err(-EINVAL);
1394
1395 /* cache native raw data representation */
1396 raw_data = btf_get_raw_data(btf, &raw_size, false);
1397 if (!raw_data) {
1398 err = -ENOMEM;
1399 goto done;
1400 }
1401 btf->raw_size = raw_size;
1402 btf->raw_data = raw_data;
1403
1404retry_load:
1405 /* if log_level is 0, we won't provide log_buf/log_size to the kernel,
1406 * initially. Only if BTF loading fails, we bump log_level to 1 and
1407 * retry, using either auto-allocated or custom log_buf. This way
1408 * non-NULL custom log_buf provides a buffer just in case, but hopes
1409 * for successful load and no need for log_buf.
1410 */
1411 if (log_level) {
1412 /* if caller didn't provide custom log_buf, we'll keep
1413 * allocating our own progressively bigger buffers for BTF
1414 * verification log
1415 */
1416 if (!log_buf) {
1417 buf_sz = max((__u32)BPF_LOG_BUF_SIZE, buf_sz * 2);
1418 tmp = realloc(buf, buf_sz);
1419 if (!tmp) {
1420 err = -ENOMEM;
1421 goto done;
1422 }
1423 buf = tmp;
1424 buf[0] = '\0';
1425 }
1426
1427 opts.log_buf = log_buf ? log_buf : buf;
1428 opts.log_size = log_buf ? log_sz : buf_sz;
1429 opts.log_level = log_level;
1430 }
1431
1432 opts.token_fd = token_fd;
1433 if (token_fd)
1434 opts.btf_flags |= BPF_F_TOKEN_FD;
1435
1436 btf->fd = bpf_btf_load(raw_data, raw_size, &opts);
1437 if (btf->fd < 0) {
1438 /* time to turn on verbose mode and try again */
1439 if (log_level == 0) {
1440 log_level = 1;
1441 goto retry_load;
1442 }
1443 /* only retry if caller didn't provide custom log_buf, but
1444 * make sure we can never overflow buf_sz
1445 */
1446 if (!log_buf && errno == ENOSPC && buf_sz <= UINT_MAX / 2)
1447 goto retry_load;
1448
1449 err = -errno;
1450 pr_warn("BTF loading error: %s\n", errstr(err));
1451 /* don't print out contents of custom log_buf */
1452 if (!log_buf && buf[0])
1453 pr_warn("-- BEGIN BTF LOAD LOG ---\n%s\n-- END BTF LOAD LOG --\n", buf);
1454 }
1455
1456done:
1457 free(buf);
1458 return libbpf_err(err);
1459}
1460
1461int btf__load_into_kernel(struct btf *btf)
1462{
1463 return btf_load_into_kernel(btf, NULL, 0, 0, 0);
1464}
1465
1466int btf__fd(const struct btf *btf)
1467{
1468 return btf->fd;
1469}
1470
1471void btf__set_fd(struct btf *btf, int fd)
1472{
1473 btf->fd = fd;
1474}
1475
1476static const void *btf_strs_data(const struct btf *btf)
1477{
1478 return btf->strs_data ? btf->strs_data : strset__data(btf->strs_set);
1479}
1480
1481static void *btf_get_raw_data(const struct btf *btf, __u32 *size, bool swap_endian)
1482{
1483 struct btf_header *hdr = btf->hdr;
1484 struct btf_type *t;
1485 void *data, *p;
1486 __u32 data_sz;
1487 int i;
1488
1489 data = swap_endian ? btf->raw_data_swapped : btf->raw_data;
1490 if (data) {
1491 *size = btf->raw_size;
1492 return data;
1493 }
1494
1495 data_sz = hdr->hdr_len + hdr->type_len + hdr->str_len;
1496 data = calloc(1, data_sz);
1497 if (!data)
1498 return NULL;
1499 p = data;
1500
1501 memcpy(p, hdr, hdr->hdr_len);
1502 if (swap_endian)
1503 btf_bswap_hdr(p);
1504 p += hdr->hdr_len;
1505
1506 memcpy(p, btf->types_data, hdr->type_len);
1507 if (swap_endian) {
1508 for (i = 0; i < btf->nr_types; i++) {
1509 t = p + btf->type_offs[i];
1510 /* btf_bswap_type_rest() relies on native t->info, so
1511 * we swap base type info after we swapped all the
1512 * additional information
1513 */
1514 if (btf_bswap_type_rest(t))
1515 goto err_out;
1516 btf_bswap_type_base(t);
1517 }
1518 }
1519 p += hdr->type_len;
1520
1521 memcpy(p, btf_strs_data(btf), hdr->str_len);
1522 p += hdr->str_len;
1523
1524 *size = data_sz;
1525 return data;
1526err_out:
1527 free(data);
1528 return NULL;
1529}
1530
1531const void *btf__raw_data(const struct btf *btf_ro, __u32 *size)
1532{
1533 struct btf *btf = (struct btf *)btf_ro;
1534 __u32 data_sz;
1535 void *data;
1536
1537 data = btf_get_raw_data(btf, &data_sz, btf->swapped_endian);
1538 if (!data)
1539 return errno = ENOMEM, NULL;
1540
1541 btf->raw_size = data_sz;
1542 if (btf->swapped_endian)
1543 btf->raw_data_swapped = data;
1544 else
1545 btf->raw_data = data;
1546 *size = data_sz;
1547 return data;
1548}
1549
1550__attribute__((alias("btf__raw_data")))
1551const void *btf__get_raw_data(const struct btf *btf, __u32 *size);
1552
1553const char *btf__str_by_offset(const struct btf *btf, __u32 offset)
1554{
1555 if (offset < btf->start_str_off)
1556 return btf__str_by_offset(btf->base_btf, offset);
1557 else if (offset - btf->start_str_off < btf->hdr->str_len)
1558 return btf_strs_data(btf) + (offset - btf->start_str_off);
1559 else
1560 return errno = EINVAL, NULL;
1561}
1562
1563const char *btf__name_by_offset(const struct btf *btf, __u32 offset)
1564{
1565 return btf__str_by_offset(btf, offset);
1566}
1567
1568struct btf *btf_get_from_fd(int btf_fd, struct btf *base_btf)
1569{
1570 struct bpf_btf_info btf_info;
1571 __u32 len = sizeof(btf_info);
1572 __u32 last_size;
1573 struct btf *btf;
1574 void *ptr;
1575 int err;
1576
1577 /* we won't know btf_size until we call bpf_btf_get_info_by_fd(). so
1578 * let's start with a sane default - 4KiB here - and resize it only if
1579 * bpf_btf_get_info_by_fd() needs a bigger buffer.
1580 */
1581 last_size = 4096;
1582 ptr = malloc(last_size);
1583 if (!ptr)
1584 return ERR_PTR(-ENOMEM);
1585
1586 memset(&btf_info, 0, sizeof(btf_info));
1587 btf_info.btf = ptr_to_u64(ptr);
1588 btf_info.btf_size = last_size;
1589 err = bpf_btf_get_info_by_fd(btf_fd, &btf_info, &len);
1590
1591 if (!err && btf_info.btf_size > last_size) {
1592 void *temp_ptr;
1593
1594 last_size = btf_info.btf_size;
1595 temp_ptr = realloc(ptr, last_size);
1596 if (!temp_ptr) {
1597 btf = ERR_PTR(-ENOMEM);
1598 goto exit_free;
1599 }
1600 ptr = temp_ptr;
1601
1602 len = sizeof(btf_info);
1603 memset(&btf_info, 0, sizeof(btf_info));
1604 btf_info.btf = ptr_to_u64(ptr);
1605 btf_info.btf_size = last_size;
1606
1607 err = bpf_btf_get_info_by_fd(btf_fd, &btf_info, &len);
1608 }
1609
1610 if (err || btf_info.btf_size > last_size) {
1611 btf = err ? ERR_PTR(-errno) : ERR_PTR(-E2BIG);
1612 goto exit_free;
1613 }
1614
1615 btf = btf_new(ptr, btf_info.btf_size, base_btf);
1616
1617exit_free:
1618 free(ptr);
1619 return btf;
1620}
1621
1622struct btf *btf__load_from_kernel_by_id_split(__u32 id, struct btf *base_btf)
1623{
1624 struct btf *btf;
1625 int btf_fd;
1626
1627 btf_fd = bpf_btf_get_fd_by_id(id);
1628 if (btf_fd < 0)
1629 return libbpf_err_ptr(-errno);
1630
1631 btf = btf_get_from_fd(btf_fd, base_btf);
1632 close(btf_fd);
1633
1634 return libbpf_ptr(btf);
1635}
1636
1637struct btf *btf__load_from_kernel_by_id(__u32 id)
1638{
1639 return btf__load_from_kernel_by_id_split(id, NULL);
1640}
1641
1642static void btf_invalidate_raw_data(struct btf *btf)
1643{
1644 if (btf->raw_data) {
1645 free(btf->raw_data);
1646 btf->raw_data = NULL;
1647 }
1648 if (btf->raw_data_swapped) {
1649 free(btf->raw_data_swapped);
1650 btf->raw_data_swapped = NULL;
1651 }
1652}
1653
1654/* Ensure BTF is ready to be modified (by splitting into a three memory
1655 * regions for header, types, and strings). Also invalidate cached
1656 * raw_data, if any.
1657 */
1658static int btf_ensure_modifiable(struct btf *btf)
1659{
1660 void *hdr, *types;
1661 struct strset *set = NULL;
1662 int err = -ENOMEM;
1663
1664 if (btf_is_modifiable(btf)) {
1665 /* any BTF modification invalidates raw_data */
1666 btf_invalidate_raw_data(btf);
1667 return 0;
1668 }
1669
1670 /* split raw data into three memory regions */
1671 hdr = malloc(btf->hdr->hdr_len);
1672 types = malloc(btf->hdr->type_len);
1673 if (!hdr || !types)
1674 goto err_out;
1675
1676 memcpy(hdr, btf->hdr, btf->hdr->hdr_len);
1677 memcpy(types, btf->types_data, btf->hdr->type_len);
1678
1679 /* build lookup index for all strings */
1680 set = strset__new(BTF_MAX_STR_OFFSET, btf->strs_data, btf->hdr->str_len);
1681 if (IS_ERR(set)) {
1682 err = PTR_ERR(set);
1683 goto err_out;
1684 }
1685
1686 /* only when everything was successful, update internal state */
1687 btf->hdr = hdr;
1688 btf->types_data = types;
1689 btf->types_data_cap = btf->hdr->type_len;
1690 btf->strs_data = NULL;
1691 btf->strs_set = set;
1692 /* if BTF was created from scratch, all strings are guaranteed to be
1693 * unique and deduplicated
1694 */
1695 if (btf->hdr->str_len == 0)
1696 btf->strs_deduped = true;
1697 if (!btf->base_btf && btf->hdr->str_len == 1)
1698 btf->strs_deduped = true;
1699
1700 /* invalidate raw_data representation */
1701 btf_invalidate_raw_data(btf);
1702
1703 return 0;
1704
1705err_out:
1706 strset__free(set);
1707 free(hdr);
1708 free(types);
1709 return err;
1710}
1711
1712/* Find an offset in BTF string section that corresponds to a given string *s*.
1713 * Returns:
1714 * - >0 offset into string section, if string is found;
1715 * - -ENOENT, if string is not in the string section;
1716 * - <0, on any other error.
1717 */
1718int btf__find_str(struct btf *btf, const char *s)
1719{
1720 int off;
1721
1722 if (btf->base_btf) {
1723 off = btf__find_str(btf->base_btf, s);
1724 if (off != -ENOENT)
1725 return off;
1726 }
1727
1728 /* BTF needs to be in a modifiable state to build string lookup index */
1729 if (btf_ensure_modifiable(btf))
1730 return libbpf_err(-ENOMEM);
1731
1732 off = strset__find_str(btf->strs_set, s);
1733 if (off < 0)
1734 return libbpf_err(off);
1735
1736 return btf->start_str_off + off;
1737}
1738
1739/* Add a string s to the BTF string section.
1740 * Returns:
1741 * - > 0 offset into string section, on success;
1742 * - < 0, on error.
1743 */
1744int btf__add_str(struct btf *btf, const char *s)
1745{
1746 int off;
1747
1748 if (btf->base_btf) {
1749 off = btf__find_str(btf->base_btf, s);
1750 if (off != -ENOENT)
1751 return off;
1752 }
1753
1754 if (btf_ensure_modifiable(btf))
1755 return libbpf_err(-ENOMEM);
1756
1757 off = strset__add_str(btf->strs_set, s);
1758 if (off < 0)
1759 return libbpf_err(off);
1760
1761 btf->hdr->str_len = strset__data_size(btf->strs_set);
1762
1763 return btf->start_str_off + off;
1764}
1765
1766static void *btf_add_type_mem(struct btf *btf, size_t add_sz)
1767{
1768 return libbpf_add_mem(&btf->types_data, &btf->types_data_cap, 1,
1769 btf->hdr->type_len, UINT_MAX, add_sz);
1770}
1771
1772static void btf_type_inc_vlen(struct btf_type *t)
1773{
1774 t->info = btf_type_info(btf_kind(t), btf_vlen(t) + 1, btf_kflag(t));
1775}
1776
1777static int btf_commit_type(struct btf *btf, int data_sz)
1778{
1779 int err;
1780
1781 err = btf_add_type_idx_entry(btf, btf->hdr->type_len);
1782 if (err)
1783 return libbpf_err(err);
1784
1785 btf->hdr->type_len += data_sz;
1786 btf->hdr->str_off += data_sz;
1787 btf->nr_types++;
1788 return btf->start_id + btf->nr_types - 1;
1789}
1790
1791struct btf_pipe {
1792 const struct btf *src;
1793 struct btf *dst;
1794 struct hashmap *str_off_map; /* map string offsets from src to dst */
1795};
1796
1797static int btf_rewrite_str(struct btf_pipe *p, __u32 *str_off)
1798{
1799 long mapped_off;
1800 int off, err;
1801
1802 if (!*str_off) /* nothing to do for empty strings */
1803 return 0;
1804
1805 if (p->str_off_map &&
1806 hashmap__find(p->str_off_map, *str_off, &mapped_off)) {
1807 *str_off = mapped_off;
1808 return 0;
1809 }
1810
1811 off = btf__add_str(p->dst, btf__str_by_offset(p->src, *str_off));
1812 if (off < 0)
1813 return off;
1814
1815 /* Remember string mapping from src to dst. It avoids
1816 * performing expensive string comparisons.
1817 */
1818 if (p->str_off_map) {
1819 err = hashmap__append(p->str_off_map, *str_off, off);
1820 if (err)
1821 return err;
1822 }
1823
1824 *str_off = off;
1825 return 0;
1826}
1827
1828static int btf_add_type(struct btf_pipe *p, const struct btf_type *src_type)
1829{
1830 struct btf_field_iter it;
1831 struct btf_type *t;
1832 __u32 *str_off;
1833 int sz, err;
1834
1835 sz = btf_type_size(src_type);
1836 if (sz < 0)
1837 return libbpf_err(sz);
1838
1839 /* deconstruct BTF, if necessary, and invalidate raw_data */
1840 if (btf_ensure_modifiable(p->dst))
1841 return libbpf_err(-ENOMEM);
1842
1843 t = btf_add_type_mem(p->dst, sz);
1844 if (!t)
1845 return libbpf_err(-ENOMEM);
1846
1847 memcpy(t, src_type, sz);
1848
1849 err = btf_field_iter_init(&it, t, BTF_FIELD_ITER_STRS);
1850 if (err)
1851 return libbpf_err(err);
1852
1853 while ((str_off = btf_field_iter_next(&it))) {
1854 err = btf_rewrite_str(p, str_off);
1855 if (err)
1856 return libbpf_err(err);
1857 }
1858
1859 return btf_commit_type(p->dst, sz);
1860}
1861
1862int btf__add_type(struct btf *btf, const struct btf *src_btf, const struct btf_type *src_type)
1863{
1864 struct btf_pipe p = { .src = src_btf, .dst = btf };
1865
1866 return btf_add_type(&p, src_type);
1867}
1868
1869static size_t btf_dedup_identity_hash_fn(long key, void *ctx);
1870static bool btf_dedup_equal_fn(long k1, long k2, void *ctx);
1871
1872int btf__add_btf(struct btf *btf, const struct btf *src_btf)
1873{
1874 struct btf_pipe p = { .src = src_btf, .dst = btf };
1875 int data_sz, sz, cnt, i, err, old_strs_len;
1876 __u32 *off;
1877 void *t;
1878
1879 /* appending split BTF isn't supported yet */
1880 if (src_btf->base_btf)
1881 return libbpf_err(-ENOTSUP);
1882
1883 /* deconstruct BTF, if necessary, and invalidate raw_data */
1884 if (btf_ensure_modifiable(btf))
1885 return libbpf_err(-ENOMEM);
1886
1887 /* remember original strings section size if we have to roll back
1888 * partial strings section changes
1889 */
1890 old_strs_len = btf->hdr->str_len;
1891
1892 data_sz = src_btf->hdr->type_len;
1893 cnt = btf__type_cnt(src_btf) - 1;
1894
1895 /* pre-allocate enough memory for new types */
1896 t = btf_add_type_mem(btf, data_sz);
1897 if (!t)
1898 return libbpf_err(-ENOMEM);
1899
1900 /* pre-allocate enough memory for type offset index for new types */
1901 off = btf_add_type_offs_mem(btf, cnt);
1902 if (!off)
1903 return libbpf_err(-ENOMEM);
1904
1905 /* Map the string offsets from src_btf to the offsets from btf to improve performance */
1906 p.str_off_map = hashmap__new(btf_dedup_identity_hash_fn, btf_dedup_equal_fn, NULL);
1907 if (IS_ERR(p.str_off_map))
1908 return libbpf_err(-ENOMEM);
1909
1910 /* bulk copy types data for all types from src_btf */
1911 memcpy(t, src_btf->types_data, data_sz);
1912
1913 for (i = 0; i < cnt; i++) {
1914 struct btf_field_iter it;
1915 __u32 *type_id, *str_off;
1916
1917 sz = btf_type_size(t);
1918 if (sz < 0) {
1919 /* unlikely, has to be corrupted src_btf */
1920 err = sz;
1921 goto err_out;
1922 }
1923
1924 /* fill out type ID to type offset mapping for lookups by type ID */
1925 *off = t - btf->types_data;
1926
1927 /* add, dedup, and remap strings referenced by this BTF type */
1928 err = btf_field_iter_init(&it, t, BTF_FIELD_ITER_STRS);
1929 if (err)
1930 goto err_out;
1931 while ((str_off = btf_field_iter_next(&it))) {
1932 err = btf_rewrite_str(&p, str_off);
1933 if (err)
1934 goto err_out;
1935 }
1936
1937 /* remap all type IDs referenced from this BTF type */
1938 err = btf_field_iter_init(&it, t, BTF_FIELD_ITER_IDS);
1939 if (err)
1940 goto err_out;
1941
1942 while ((type_id = btf_field_iter_next(&it))) {
1943 if (!*type_id) /* nothing to do for VOID references */
1944 continue;
1945
1946 /* we haven't updated btf's type count yet, so
1947 * btf->start_id + btf->nr_types - 1 is the type ID offset we should
1948 * add to all newly added BTF types
1949 */
1950 *type_id += btf->start_id + btf->nr_types - 1;
1951 }
1952
1953 /* go to next type data and type offset index entry */
1954 t += sz;
1955 off++;
1956 }
1957
1958 /* Up until now any of the copied type data was effectively invisible,
1959 * so if we exited early before this point due to error, BTF would be
1960 * effectively unmodified. There would be extra internal memory
1961 * pre-allocated, but it would not be available for querying. But now
1962 * that we've copied and rewritten all the data successfully, we can
1963 * update type count and various internal offsets and sizes to
1964 * "commit" the changes and made them visible to the outside world.
1965 */
1966 btf->hdr->type_len += data_sz;
1967 btf->hdr->str_off += data_sz;
1968 btf->nr_types += cnt;
1969
1970 hashmap__free(p.str_off_map);
1971
1972 /* return type ID of the first added BTF type */
1973 return btf->start_id + btf->nr_types - cnt;
1974err_out:
1975 /* zero out preallocated memory as if it was just allocated with
1976 * libbpf_add_mem()
1977 */
1978 memset(btf->types_data + btf->hdr->type_len, 0, data_sz);
1979 memset(btf->strs_data + old_strs_len, 0, btf->hdr->str_len - old_strs_len);
1980
1981 /* and now restore original strings section size; types data size
1982 * wasn't modified, so doesn't need restoring, see big comment above
1983 */
1984 btf->hdr->str_len = old_strs_len;
1985
1986 hashmap__free(p.str_off_map);
1987
1988 return libbpf_err(err);
1989}
1990
1991/*
1992 * Append new BTF_KIND_INT type with:
1993 * - *name* - non-empty, non-NULL type name;
1994 * - *sz* - power-of-2 (1, 2, 4, ..) size of the type, in bytes;
1995 * - encoding is a combination of BTF_INT_SIGNED, BTF_INT_CHAR, BTF_INT_BOOL.
1996 * Returns:
1997 * - >0, type ID of newly added BTF type;
1998 * - <0, on error.
1999 */
2000int btf__add_int(struct btf *btf, const char *name, size_t byte_sz, int encoding)
2001{
2002 struct btf_type *t;
2003 int sz, name_off;
2004
2005 /* non-empty name */
2006 if (!name || !name[0])
2007 return libbpf_err(-EINVAL);
2008 /* byte_sz must be power of 2 */
2009 if (!byte_sz || (byte_sz & (byte_sz - 1)) || byte_sz > 16)
2010 return libbpf_err(-EINVAL);
2011 if (encoding & ~(BTF_INT_SIGNED | BTF_INT_CHAR | BTF_INT_BOOL))
2012 return libbpf_err(-EINVAL);
2013
2014 /* deconstruct BTF, if necessary, and invalidate raw_data */
2015 if (btf_ensure_modifiable(btf))
2016 return libbpf_err(-ENOMEM);
2017
2018 sz = sizeof(struct btf_type) + sizeof(int);
2019 t = btf_add_type_mem(btf, sz);
2020 if (!t)
2021 return libbpf_err(-ENOMEM);
2022
2023 /* if something goes wrong later, we might end up with an extra string,
2024 * but that shouldn't be a problem, because BTF can't be constructed
2025 * completely anyway and will most probably be just discarded
2026 */
2027 name_off = btf__add_str(btf, name);
2028 if (name_off < 0)
2029 return name_off;
2030
2031 t->name_off = name_off;
2032 t->info = btf_type_info(BTF_KIND_INT, 0, 0);
2033 t->size = byte_sz;
2034 /* set INT info, we don't allow setting legacy bit offset/size */
2035 *(__u32 *)(t + 1) = (encoding << 24) | (byte_sz * 8);
2036
2037 return btf_commit_type(btf, sz);
2038}
2039
2040/*
2041 * Append new BTF_KIND_FLOAT type with:
2042 * - *name* - non-empty, non-NULL type name;
2043 * - *sz* - size of the type, in bytes;
2044 * Returns:
2045 * - >0, type ID of newly added BTF type;
2046 * - <0, on error.
2047 */
2048int btf__add_float(struct btf *btf, const char *name, size_t byte_sz)
2049{
2050 struct btf_type *t;
2051 int sz, name_off;
2052
2053 /* non-empty name */
2054 if (!name || !name[0])
2055 return libbpf_err(-EINVAL);
2056
2057 /* byte_sz must be one of the explicitly allowed values */
2058 if (byte_sz != 2 && byte_sz != 4 && byte_sz != 8 && byte_sz != 12 &&
2059 byte_sz != 16)
2060 return libbpf_err(-EINVAL);
2061
2062 if (btf_ensure_modifiable(btf))
2063 return libbpf_err(-ENOMEM);
2064
2065 sz = sizeof(struct btf_type);
2066 t = btf_add_type_mem(btf, sz);
2067 if (!t)
2068 return libbpf_err(-ENOMEM);
2069
2070 name_off = btf__add_str(btf, name);
2071 if (name_off < 0)
2072 return name_off;
2073
2074 t->name_off = name_off;
2075 t->info = btf_type_info(BTF_KIND_FLOAT, 0, 0);
2076 t->size = byte_sz;
2077
2078 return btf_commit_type(btf, sz);
2079}
2080
2081/* it's completely legal to append BTF types with type IDs pointing forward to
2082 * types that haven't been appended yet, so we only make sure that id looks
2083 * sane, we can't guarantee that ID will always be valid
2084 */
2085static int validate_type_id(int id)
2086{
2087 if (id < 0 || id > BTF_MAX_NR_TYPES)
2088 return -EINVAL;
2089 return 0;
2090}
2091
2092/* generic append function for PTR, TYPEDEF, CONST/VOLATILE/RESTRICT */
2093static int btf_add_ref_kind(struct btf *btf, int kind, const char *name, int ref_type_id)
2094{
2095 struct btf_type *t;
2096 int sz, name_off = 0;
2097
2098 if (validate_type_id(ref_type_id))
2099 return libbpf_err(-EINVAL);
2100
2101 if (btf_ensure_modifiable(btf))
2102 return libbpf_err(-ENOMEM);
2103
2104 sz = sizeof(struct btf_type);
2105 t = btf_add_type_mem(btf, sz);
2106 if (!t)
2107 return libbpf_err(-ENOMEM);
2108
2109 if (name && name[0]) {
2110 name_off = btf__add_str(btf, name);
2111 if (name_off < 0)
2112 return name_off;
2113 }
2114
2115 t->name_off = name_off;
2116 t->info = btf_type_info(kind, 0, 0);
2117 t->type = ref_type_id;
2118
2119 return btf_commit_type(btf, sz);
2120}
2121
2122/*
2123 * Append new BTF_KIND_PTR type with:
2124 * - *ref_type_id* - referenced type ID, it might not exist yet;
2125 * Returns:
2126 * - >0, type ID of newly added BTF type;
2127 * - <0, on error.
2128 */
2129int btf__add_ptr(struct btf *btf, int ref_type_id)
2130{
2131 return btf_add_ref_kind(btf, BTF_KIND_PTR, NULL, ref_type_id);
2132}
2133
2134/*
2135 * Append new BTF_KIND_ARRAY type with:
2136 * - *index_type_id* - type ID of the type describing array index;
2137 * - *elem_type_id* - type ID of the type describing array element;
2138 * - *nr_elems* - the size of the array;
2139 * Returns:
2140 * - >0, type ID of newly added BTF type;
2141 * - <0, on error.
2142 */
2143int btf__add_array(struct btf *btf, int index_type_id, int elem_type_id, __u32 nr_elems)
2144{
2145 struct btf_type *t;
2146 struct btf_array *a;
2147 int sz;
2148
2149 if (validate_type_id(index_type_id) || validate_type_id(elem_type_id))
2150 return libbpf_err(-EINVAL);
2151
2152 if (btf_ensure_modifiable(btf))
2153 return libbpf_err(-ENOMEM);
2154
2155 sz = sizeof(struct btf_type) + sizeof(struct btf_array);
2156 t = btf_add_type_mem(btf, sz);
2157 if (!t)
2158 return libbpf_err(-ENOMEM);
2159
2160 t->name_off = 0;
2161 t->info = btf_type_info(BTF_KIND_ARRAY, 0, 0);
2162 t->size = 0;
2163
2164 a = btf_array(t);
2165 a->type = elem_type_id;
2166 a->index_type = index_type_id;
2167 a->nelems = nr_elems;
2168
2169 return btf_commit_type(btf, sz);
2170}
2171
2172/* generic STRUCT/UNION append function */
2173static int btf_add_composite(struct btf *btf, int kind, const char *name, __u32 bytes_sz)
2174{
2175 struct btf_type *t;
2176 int sz, name_off = 0;
2177
2178 if (btf_ensure_modifiable(btf))
2179 return libbpf_err(-ENOMEM);
2180
2181 sz = sizeof(struct btf_type);
2182 t = btf_add_type_mem(btf, sz);
2183 if (!t)
2184 return libbpf_err(-ENOMEM);
2185
2186 if (name && name[0]) {
2187 name_off = btf__add_str(btf, name);
2188 if (name_off < 0)
2189 return name_off;
2190 }
2191
2192 /* start out with vlen=0 and no kflag; this will be adjusted when
2193 * adding each member
2194 */
2195 t->name_off = name_off;
2196 t->info = btf_type_info(kind, 0, 0);
2197 t->size = bytes_sz;
2198
2199 return btf_commit_type(btf, sz);
2200}
2201
2202/*
2203 * Append new BTF_KIND_STRUCT type with:
2204 * - *name* - name of the struct, can be NULL or empty for anonymous structs;
2205 * - *byte_sz* - size of the struct, in bytes;
2206 *
2207 * Struct initially has no fields in it. Fields can be added by
2208 * btf__add_field() right after btf__add_struct() succeeds.
2209 *
2210 * Returns:
2211 * - >0, type ID of newly added BTF type;
2212 * - <0, on error.
2213 */
2214int btf__add_struct(struct btf *btf, const char *name, __u32 byte_sz)
2215{
2216 return btf_add_composite(btf, BTF_KIND_STRUCT, name, byte_sz);
2217}
2218
2219/*
2220 * Append new BTF_KIND_UNION type with:
2221 * - *name* - name of the union, can be NULL or empty for anonymous union;
2222 * - *byte_sz* - size of the union, in bytes;
2223 *
2224 * Union initially has no fields in it. Fields can be added by
2225 * btf__add_field() right after btf__add_union() succeeds. All fields
2226 * should have *bit_offset* of 0.
2227 *
2228 * Returns:
2229 * - >0, type ID of newly added BTF type;
2230 * - <0, on error.
2231 */
2232int btf__add_union(struct btf *btf, const char *name, __u32 byte_sz)
2233{
2234 return btf_add_composite(btf, BTF_KIND_UNION, name, byte_sz);
2235}
2236
2237static struct btf_type *btf_last_type(struct btf *btf)
2238{
2239 return btf_type_by_id(btf, btf__type_cnt(btf) - 1);
2240}
2241
2242/*
2243 * Append new field for the current STRUCT/UNION type with:
2244 * - *name* - name of the field, can be NULL or empty for anonymous field;
2245 * - *type_id* - type ID for the type describing field type;
2246 * - *bit_offset* - bit offset of the start of the field within struct/union;
2247 * - *bit_size* - bit size of a bitfield, 0 for non-bitfield fields;
2248 * Returns:
2249 * - 0, on success;
2250 * - <0, on error.
2251 */
2252int btf__add_field(struct btf *btf, const char *name, int type_id,
2253 __u32 bit_offset, __u32 bit_size)
2254{
2255 struct btf_type *t;
2256 struct btf_member *m;
2257 bool is_bitfield;
2258 int sz, name_off = 0;
2259
2260 /* last type should be union/struct */
2261 if (btf->nr_types == 0)
2262 return libbpf_err(-EINVAL);
2263 t = btf_last_type(btf);
2264 if (!btf_is_composite(t))
2265 return libbpf_err(-EINVAL);
2266
2267 if (validate_type_id(type_id))
2268 return libbpf_err(-EINVAL);
2269 /* best-effort bit field offset/size enforcement */
2270 is_bitfield = bit_size || (bit_offset % 8 != 0);
2271 if (is_bitfield && (bit_size == 0 || bit_size > 255 || bit_offset > 0xffffff))
2272 return libbpf_err(-EINVAL);
2273
2274 /* only offset 0 is allowed for unions */
2275 if (btf_is_union(t) && bit_offset)
2276 return libbpf_err(-EINVAL);
2277
2278 /* decompose and invalidate raw data */
2279 if (btf_ensure_modifiable(btf))
2280 return libbpf_err(-ENOMEM);
2281
2282 sz = sizeof(struct btf_member);
2283 m = btf_add_type_mem(btf, sz);
2284 if (!m)
2285 return libbpf_err(-ENOMEM);
2286
2287 if (name && name[0]) {
2288 name_off = btf__add_str(btf, name);
2289 if (name_off < 0)
2290 return name_off;
2291 }
2292
2293 m->name_off = name_off;
2294 m->type = type_id;
2295 m->offset = bit_offset | (bit_size << 24);
2296
2297 /* btf_add_type_mem can invalidate t pointer */
2298 t = btf_last_type(btf);
2299 /* update parent type's vlen and kflag */
2300 t->info = btf_type_info(btf_kind(t), btf_vlen(t) + 1, is_bitfield || btf_kflag(t));
2301
2302 btf->hdr->type_len += sz;
2303 btf->hdr->str_off += sz;
2304 return 0;
2305}
2306
2307static int btf_add_enum_common(struct btf *btf, const char *name, __u32 byte_sz,
2308 bool is_signed, __u8 kind)
2309{
2310 struct btf_type *t;
2311 int sz, name_off = 0;
2312
2313 /* byte_sz must be power of 2 */
2314 if (!byte_sz || (byte_sz & (byte_sz - 1)) || byte_sz > 8)
2315 return libbpf_err(-EINVAL);
2316
2317 if (btf_ensure_modifiable(btf))
2318 return libbpf_err(-ENOMEM);
2319
2320 sz = sizeof(struct btf_type);
2321 t = btf_add_type_mem(btf, sz);
2322 if (!t)
2323 return libbpf_err(-ENOMEM);
2324
2325 if (name && name[0]) {
2326 name_off = btf__add_str(btf, name);
2327 if (name_off < 0)
2328 return name_off;
2329 }
2330
2331 /* start out with vlen=0; it will be adjusted when adding enum values */
2332 t->name_off = name_off;
2333 t->info = btf_type_info(kind, 0, is_signed);
2334 t->size = byte_sz;
2335
2336 return btf_commit_type(btf, sz);
2337}
2338
2339/*
2340 * Append new BTF_KIND_ENUM type with:
2341 * - *name* - name of the enum, can be NULL or empty for anonymous enums;
2342 * - *byte_sz* - size of the enum, in bytes.
2343 *
2344 * Enum initially has no enum values in it (and corresponds to enum forward
2345 * declaration). Enumerator values can be added by btf__add_enum_value()
2346 * immediately after btf__add_enum() succeeds.
2347 *
2348 * Returns:
2349 * - >0, type ID of newly added BTF type;
2350 * - <0, on error.
2351 */
2352int btf__add_enum(struct btf *btf, const char *name, __u32 byte_sz)
2353{
2354 /*
2355 * set the signedness to be unsigned, it will change to signed
2356 * if any later enumerator is negative.
2357 */
2358 return btf_add_enum_common(btf, name, byte_sz, false, BTF_KIND_ENUM);
2359}
2360
2361/*
2362 * Append new enum value for the current ENUM type with:
2363 * - *name* - name of the enumerator value, can't be NULL or empty;
2364 * - *value* - integer value corresponding to enum value *name*;
2365 * Returns:
2366 * - 0, on success;
2367 * - <0, on error.
2368 */
2369int btf__add_enum_value(struct btf *btf, const char *name, __s64 value)
2370{
2371 struct btf_type *t;
2372 struct btf_enum *v;
2373 int sz, name_off;
2374
2375 /* last type should be BTF_KIND_ENUM */
2376 if (btf->nr_types == 0)
2377 return libbpf_err(-EINVAL);
2378 t = btf_last_type(btf);
2379 if (!btf_is_enum(t))
2380 return libbpf_err(-EINVAL);
2381
2382 /* non-empty name */
2383 if (!name || !name[0])
2384 return libbpf_err(-EINVAL);
2385 if (value < INT_MIN || value > UINT_MAX)
2386 return libbpf_err(-E2BIG);
2387
2388 /* decompose and invalidate raw data */
2389 if (btf_ensure_modifiable(btf))
2390 return libbpf_err(-ENOMEM);
2391
2392 sz = sizeof(struct btf_enum);
2393 v = btf_add_type_mem(btf, sz);
2394 if (!v)
2395 return libbpf_err(-ENOMEM);
2396
2397 name_off = btf__add_str(btf, name);
2398 if (name_off < 0)
2399 return name_off;
2400
2401 v->name_off = name_off;
2402 v->val = value;
2403
2404 /* update parent type's vlen */
2405 t = btf_last_type(btf);
2406 btf_type_inc_vlen(t);
2407
2408 /* if negative value, set signedness to signed */
2409 if (value < 0)
2410 t->info = btf_type_info(btf_kind(t), btf_vlen(t), true);
2411
2412 btf->hdr->type_len += sz;
2413 btf->hdr->str_off += sz;
2414 return 0;
2415}
2416
2417/*
2418 * Append new BTF_KIND_ENUM64 type with:
2419 * - *name* - name of the enum, can be NULL or empty for anonymous enums;
2420 * - *byte_sz* - size of the enum, in bytes.
2421 * - *is_signed* - whether the enum values are signed or not;
2422 *
2423 * Enum initially has no enum values in it (and corresponds to enum forward
2424 * declaration). Enumerator values can be added by btf__add_enum64_value()
2425 * immediately after btf__add_enum64() succeeds.
2426 *
2427 * Returns:
2428 * - >0, type ID of newly added BTF type;
2429 * - <0, on error.
2430 */
2431int btf__add_enum64(struct btf *btf, const char *name, __u32 byte_sz,
2432 bool is_signed)
2433{
2434 return btf_add_enum_common(btf, name, byte_sz, is_signed,
2435 BTF_KIND_ENUM64);
2436}
2437
2438/*
2439 * Append new enum value for the current ENUM64 type with:
2440 * - *name* - name of the enumerator value, can't be NULL or empty;
2441 * - *value* - integer value corresponding to enum value *name*;
2442 * Returns:
2443 * - 0, on success;
2444 * - <0, on error.
2445 */
2446int btf__add_enum64_value(struct btf *btf, const char *name, __u64 value)
2447{
2448 struct btf_enum64 *v;
2449 struct btf_type *t;
2450 int sz, name_off;
2451
2452 /* last type should be BTF_KIND_ENUM64 */
2453 if (btf->nr_types == 0)
2454 return libbpf_err(-EINVAL);
2455 t = btf_last_type(btf);
2456 if (!btf_is_enum64(t))
2457 return libbpf_err(-EINVAL);
2458
2459 /* non-empty name */
2460 if (!name || !name[0])
2461 return libbpf_err(-EINVAL);
2462
2463 /* decompose and invalidate raw data */
2464 if (btf_ensure_modifiable(btf))
2465 return libbpf_err(-ENOMEM);
2466
2467 sz = sizeof(struct btf_enum64);
2468 v = btf_add_type_mem(btf, sz);
2469 if (!v)
2470 return libbpf_err(-ENOMEM);
2471
2472 name_off = btf__add_str(btf, name);
2473 if (name_off < 0)
2474 return name_off;
2475
2476 v->name_off = name_off;
2477 v->val_lo32 = (__u32)value;
2478 v->val_hi32 = value >> 32;
2479
2480 /* update parent type's vlen */
2481 t = btf_last_type(btf);
2482 btf_type_inc_vlen(t);
2483
2484 btf->hdr->type_len += sz;
2485 btf->hdr->str_off += sz;
2486 return 0;
2487}
2488
2489/*
2490 * Append new BTF_KIND_FWD type with:
2491 * - *name*, non-empty/non-NULL name;
2492 * - *fwd_kind*, kind of forward declaration, one of BTF_FWD_STRUCT,
2493 * BTF_FWD_UNION, or BTF_FWD_ENUM;
2494 * Returns:
2495 * - >0, type ID of newly added BTF type;
2496 * - <0, on error.
2497 */
2498int btf__add_fwd(struct btf *btf, const char *name, enum btf_fwd_kind fwd_kind)
2499{
2500 if (!name || !name[0])
2501 return libbpf_err(-EINVAL);
2502
2503 switch (fwd_kind) {
2504 case BTF_FWD_STRUCT:
2505 case BTF_FWD_UNION: {
2506 struct btf_type *t;
2507 int id;
2508
2509 id = btf_add_ref_kind(btf, BTF_KIND_FWD, name, 0);
2510 if (id <= 0)
2511 return id;
2512 t = btf_type_by_id(btf, id);
2513 t->info = btf_type_info(BTF_KIND_FWD, 0, fwd_kind == BTF_FWD_UNION);
2514 return id;
2515 }
2516 case BTF_FWD_ENUM:
2517 /* enum forward in BTF currently is just an enum with no enum
2518 * values; we also assume a standard 4-byte size for it
2519 */
2520 return btf__add_enum(btf, name, sizeof(int));
2521 default:
2522 return libbpf_err(-EINVAL);
2523 }
2524}
2525
2526/*
2527 * Append new BTF_KING_TYPEDEF type with:
2528 * - *name*, non-empty/non-NULL name;
2529 * - *ref_type_id* - referenced type ID, it might not exist yet;
2530 * Returns:
2531 * - >0, type ID of newly added BTF type;
2532 * - <0, on error.
2533 */
2534int btf__add_typedef(struct btf *btf, const char *name, int ref_type_id)
2535{
2536 if (!name || !name[0])
2537 return libbpf_err(-EINVAL);
2538
2539 return btf_add_ref_kind(btf, BTF_KIND_TYPEDEF, name, ref_type_id);
2540}
2541
2542/*
2543 * Append new BTF_KIND_VOLATILE type with:
2544 * - *ref_type_id* - referenced type ID, it might not exist yet;
2545 * Returns:
2546 * - >0, type ID of newly added BTF type;
2547 * - <0, on error.
2548 */
2549int btf__add_volatile(struct btf *btf, int ref_type_id)
2550{
2551 return btf_add_ref_kind(btf, BTF_KIND_VOLATILE, NULL, ref_type_id);
2552}
2553
2554/*
2555 * Append new BTF_KIND_CONST type with:
2556 * - *ref_type_id* - referenced type ID, it might not exist yet;
2557 * Returns:
2558 * - >0, type ID of newly added BTF type;
2559 * - <0, on error.
2560 */
2561int btf__add_const(struct btf *btf, int ref_type_id)
2562{
2563 return btf_add_ref_kind(btf, BTF_KIND_CONST, NULL, ref_type_id);
2564}
2565
2566/*
2567 * Append new BTF_KIND_RESTRICT type with:
2568 * - *ref_type_id* - referenced type ID, it might not exist yet;
2569 * Returns:
2570 * - >0, type ID of newly added BTF type;
2571 * - <0, on error.
2572 */
2573int btf__add_restrict(struct btf *btf, int ref_type_id)
2574{
2575 return btf_add_ref_kind(btf, BTF_KIND_RESTRICT, NULL, ref_type_id);
2576}
2577
2578/*
2579 * Append new BTF_KIND_TYPE_TAG type with:
2580 * - *value*, non-empty/non-NULL tag value;
2581 * - *ref_type_id* - referenced type ID, it might not exist yet;
2582 * Returns:
2583 * - >0, type ID of newly added BTF type;
2584 * - <0, on error.
2585 */
2586int btf__add_type_tag(struct btf *btf, const char *value, int ref_type_id)
2587{
2588 if (!value || !value[0])
2589 return libbpf_err(-EINVAL);
2590
2591 return btf_add_ref_kind(btf, BTF_KIND_TYPE_TAG, value, ref_type_id);
2592}
2593
2594/*
2595 * Append new BTF_KIND_FUNC type with:
2596 * - *name*, non-empty/non-NULL name;
2597 * - *proto_type_id* - FUNC_PROTO's type ID, it might not exist yet;
2598 * Returns:
2599 * - >0, type ID of newly added BTF type;
2600 * - <0, on error.
2601 */
2602int btf__add_func(struct btf *btf, const char *name,
2603 enum btf_func_linkage linkage, int proto_type_id)
2604{
2605 int id;
2606
2607 if (!name || !name[0])
2608 return libbpf_err(-EINVAL);
2609 if (linkage != BTF_FUNC_STATIC && linkage != BTF_FUNC_GLOBAL &&
2610 linkage != BTF_FUNC_EXTERN)
2611 return libbpf_err(-EINVAL);
2612
2613 id = btf_add_ref_kind(btf, BTF_KIND_FUNC, name, proto_type_id);
2614 if (id > 0) {
2615 struct btf_type *t = btf_type_by_id(btf, id);
2616
2617 t->info = btf_type_info(BTF_KIND_FUNC, linkage, 0);
2618 }
2619 return libbpf_err(id);
2620}
2621
2622/*
2623 * Append new BTF_KIND_FUNC_PROTO with:
2624 * - *ret_type_id* - type ID for return result of a function.
2625 *
2626 * Function prototype initially has no arguments, but they can be added by
2627 * btf__add_func_param() one by one, immediately after
2628 * btf__add_func_proto() succeeded.
2629 *
2630 * Returns:
2631 * - >0, type ID of newly added BTF type;
2632 * - <0, on error.
2633 */
2634int btf__add_func_proto(struct btf *btf, int ret_type_id)
2635{
2636 struct btf_type *t;
2637 int sz;
2638
2639 if (validate_type_id(ret_type_id))
2640 return libbpf_err(-EINVAL);
2641
2642 if (btf_ensure_modifiable(btf))
2643 return libbpf_err(-ENOMEM);
2644
2645 sz = sizeof(struct btf_type);
2646 t = btf_add_type_mem(btf, sz);
2647 if (!t)
2648 return libbpf_err(-ENOMEM);
2649
2650 /* start out with vlen=0; this will be adjusted when adding enum
2651 * values, if necessary
2652 */
2653 t->name_off = 0;
2654 t->info = btf_type_info(BTF_KIND_FUNC_PROTO, 0, 0);
2655 t->type = ret_type_id;
2656
2657 return btf_commit_type(btf, sz);
2658}
2659
2660/*
2661 * Append new function parameter for current FUNC_PROTO type with:
2662 * - *name* - parameter name, can be NULL or empty;
2663 * - *type_id* - type ID describing the type of the parameter.
2664 * Returns:
2665 * - 0, on success;
2666 * - <0, on error.
2667 */
2668int btf__add_func_param(struct btf *btf, const char *name, int type_id)
2669{
2670 struct btf_type *t;
2671 struct btf_param *p;
2672 int sz, name_off = 0;
2673
2674 if (validate_type_id(type_id))
2675 return libbpf_err(-EINVAL);
2676
2677 /* last type should be BTF_KIND_FUNC_PROTO */
2678 if (btf->nr_types == 0)
2679 return libbpf_err(-EINVAL);
2680 t = btf_last_type(btf);
2681 if (!btf_is_func_proto(t))
2682 return libbpf_err(-EINVAL);
2683
2684 /* decompose and invalidate raw data */
2685 if (btf_ensure_modifiable(btf))
2686 return libbpf_err(-ENOMEM);
2687
2688 sz = sizeof(struct btf_param);
2689 p = btf_add_type_mem(btf, sz);
2690 if (!p)
2691 return libbpf_err(-ENOMEM);
2692
2693 if (name && name[0]) {
2694 name_off = btf__add_str(btf, name);
2695 if (name_off < 0)
2696 return name_off;
2697 }
2698
2699 p->name_off = name_off;
2700 p->type = type_id;
2701
2702 /* update parent type's vlen */
2703 t = btf_last_type(btf);
2704 btf_type_inc_vlen(t);
2705
2706 btf->hdr->type_len += sz;
2707 btf->hdr->str_off += sz;
2708 return 0;
2709}
2710
2711/*
2712 * Append new BTF_KIND_VAR type with:
2713 * - *name* - non-empty/non-NULL name;
2714 * - *linkage* - variable linkage, one of BTF_VAR_STATIC,
2715 * BTF_VAR_GLOBAL_ALLOCATED, or BTF_VAR_GLOBAL_EXTERN;
2716 * - *type_id* - type ID of the type describing the type of the variable.
2717 * Returns:
2718 * - >0, type ID of newly added BTF type;
2719 * - <0, on error.
2720 */
2721int btf__add_var(struct btf *btf, const char *name, int linkage, int type_id)
2722{
2723 struct btf_type *t;
2724 struct btf_var *v;
2725 int sz, name_off;
2726
2727 /* non-empty name */
2728 if (!name || !name[0])
2729 return libbpf_err(-EINVAL);
2730 if (linkage != BTF_VAR_STATIC && linkage != BTF_VAR_GLOBAL_ALLOCATED &&
2731 linkage != BTF_VAR_GLOBAL_EXTERN)
2732 return libbpf_err(-EINVAL);
2733 if (validate_type_id(type_id))
2734 return libbpf_err(-EINVAL);
2735
2736 /* deconstruct BTF, if necessary, and invalidate raw_data */
2737 if (btf_ensure_modifiable(btf))
2738 return libbpf_err(-ENOMEM);
2739
2740 sz = sizeof(struct btf_type) + sizeof(struct btf_var);
2741 t = btf_add_type_mem(btf, sz);
2742 if (!t)
2743 return libbpf_err(-ENOMEM);
2744
2745 name_off = btf__add_str(btf, name);
2746 if (name_off < 0)
2747 return name_off;
2748
2749 t->name_off = name_off;
2750 t->info = btf_type_info(BTF_KIND_VAR, 0, 0);
2751 t->type = type_id;
2752
2753 v = btf_var(t);
2754 v->linkage = linkage;
2755
2756 return btf_commit_type(btf, sz);
2757}
2758
2759/*
2760 * Append new BTF_KIND_DATASEC type with:
2761 * - *name* - non-empty/non-NULL name;
2762 * - *byte_sz* - data section size, in bytes.
2763 *
2764 * Data section is initially empty. Variables info can be added with
2765 * btf__add_datasec_var_info() calls, after btf__add_datasec() succeeds.
2766 *
2767 * Returns:
2768 * - >0, type ID of newly added BTF type;
2769 * - <0, on error.
2770 */
2771int btf__add_datasec(struct btf *btf, const char *name, __u32 byte_sz)
2772{
2773 struct btf_type *t;
2774 int sz, name_off;
2775
2776 /* non-empty name */
2777 if (!name || !name[0])
2778 return libbpf_err(-EINVAL);
2779
2780 if (btf_ensure_modifiable(btf))
2781 return libbpf_err(-ENOMEM);
2782
2783 sz = sizeof(struct btf_type);
2784 t = btf_add_type_mem(btf, sz);
2785 if (!t)
2786 return libbpf_err(-ENOMEM);
2787
2788 name_off = btf__add_str(btf, name);
2789 if (name_off < 0)
2790 return name_off;
2791
2792 /* start with vlen=0, which will be update as var_secinfos are added */
2793 t->name_off = name_off;
2794 t->info = btf_type_info(BTF_KIND_DATASEC, 0, 0);
2795 t->size = byte_sz;
2796
2797 return btf_commit_type(btf, sz);
2798}
2799
2800/*
2801 * Append new data section variable information entry for current DATASEC type:
2802 * - *var_type_id* - type ID, describing type of the variable;
2803 * - *offset* - variable offset within data section, in bytes;
2804 * - *byte_sz* - variable size, in bytes.
2805 *
2806 * Returns:
2807 * - 0, on success;
2808 * - <0, on error.
2809 */
2810int btf__add_datasec_var_info(struct btf *btf, int var_type_id, __u32 offset, __u32 byte_sz)
2811{
2812 struct btf_type *t;
2813 struct btf_var_secinfo *v;
2814 int sz;
2815
2816 /* last type should be BTF_KIND_DATASEC */
2817 if (btf->nr_types == 0)
2818 return libbpf_err(-EINVAL);
2819 t = btf_last_type(btf);
2820 if (!btf_is_datasec(t))
2821 return libbpf_err(-EINVAL);
2822
2823 if (validate_type_id(var_type_id))
2824 return libbpf_err(-EINVAL);
2825
2826 /* decompose and invalidate raw data */
2827 if (btf_ensure_modifiable(btf))
2828 return libbpf_err(-ENOMEM);
2829
2830 sz = sizeof(struct btf_var_secinfo);
2831 v = btf_add_type_mem(btf, sz);
2832 if (!v)
2833 return libbpf_err(-ENOMEM);
2834
2835 v->type = var_type_id;
2836 v->offset = offset;
2837 v->size = byte_sz;
2838
2839 /* update parent type's vlen */
2840 t = btf_last_type(btf);
2841 btf_type_inc_vlen(t);
2842
2843 btf->hdr->type_len += sz;
2844 btf->hdr->str_off += sz;
2845 return 0;
2846}
2847
2848/*
2849 * Append new BTF_KIND_DECL_TAG type with:
2850 * - *value* - non-empty/non-NULL string;
2851 * - *ref_type_id* - referenced type ID, it might not exist yet;
2852 * - *component_idx* - -1 for tagging reference type, otherwise struct/union
2853 * member or function argument index;
2854 * Returns:
2855 * - >0, type ID of newly added BTF type;
2856 * - <0, on error.
2857 */
2858int btf__add_decl_tag(struct btf *btf, const char *value, int ref_type_id,
2859 int component_idx)
2860{
2861 struct btf_type *t;
2862 int sz, value_off;
2863
2864 if (!value || !value[0] || component_idx < -1)
2865 return libbpf_err(-EINVAL);
2866
2867 if (validate_type_id(ref_type_id))
2868 return libbpf_err(-EINVAL);
2869
2870 if (btf_ensure_modifiable(btf))
2871 return libbpf_err(-ENOMEM);
2872
2873 sz = sizeof(struct btf_type) + sizeof(struct btf_decl_tag);
2874 t = btf_add_type_mem(btf, sz);
2875 if (!t)
2876 return libbpf_err(-ENOMEM);
2877
2878 value_off = btf__add_str(btf, value);
2879 if (value_off < 0)
2880 return value_off;
2881
2882 t->name_off = value_off;
2883 t->info = btf_type_info(BTF_KIND_DECL_TAG, 0, false);
2884 t->type = ref_type_id;
2885 btf_decl_tag(t)->component_idx = component_idx;
2886
2887 return btf_commit_type(btf, sz);
2888}
2889
2890struct btf_ext_sec_info_param {
2891 __u32 off;
2892 __u32 len;
2893 __u32 min_rec_size;
2894 struct btf_ext_info *ext_info;
2895 const char *desc;
2896};
2897
2898/*
2899 * Parse a single info subsection of the BTF.ext info data:
2900 * - validate subsection structure and elements
2901 * - save info subsection start and sizing details in struct btf_ext
2902 * - endian-independent operation, for calling before byte-swapping
2903 */
2904static int btf_ext_parse_sec_info(struct btf_ext *btf_ext,
2905 struct btf_ext_sec_info_param *ext_sec,
2906 bool is_native)
2907{
2908 const struct btf_ext_info_sec *sinfo;
2909 struct btf_ext_info *ext_info;
2910 __u32 info_left, record_size;
2911 size_t sec_cnt = 0;
2912 void *info;
2913
2914 if (ext_sec->len == 0)
2915 return 0;
2916
2917 if (ext_sec->off & 0x03) {
2918 pr_debug(".BTF.ext %s section is not aligned to 4 bytes\n",
2919 ext_sec->desc);
2920 return -EINVAL;
2921 }
2922
2923 /* The start of the info sec (including the __u32 record_size). */
2924 info = btf_ext->data + btf_ext->hdr->hdr_len + ext_sec->off;
2925 info_left = ext_sec->len;
2926
2927 if (btf_ext->data + btf_ext->data_size < info + ext_sec->len) {
2928 pr_debug("%s section (off:%u len:%u) is beyond the end of the ELF section .BTF.ext\n",
2929 ext_sec->desc, ext_sec->off, ext_sec->len);
2930 return -EINVAL;
2931 }
2932
2933 /* At least a record size */
2934 if (info_left < sizeof(__u32)) {
2935 pr_debug(".BTF.ext %s record size not found\n", ext_sec->desc);
2936 return -EINVAL;
2937 }
2938
2939 /* The record size needs to meet either the minimum standard or, when
2940 * handling non-native endianness data, the exact standard so as
2941 * to allow safe byte-swapping.
2942 */
2943 record_size = is_native ? *(__u32 *)info : bswap_32(*(__u32 *)info);
2944 if (record_size < ext_sec->min_rec_size ||
2945 (!is_native && record_size != ext_sec->min_rec_size) ||
2946 record_size & 0x03) {
2947 pr_debug("%s section in .BTF.ext has invalid record size %u\n",
2948 ext_sec->desc, record_size);
2949 return -EINVAL;
2950 }
2951
2952 sinfo = info + sizeof(__u32);
2953 info_left -= sizeof(__u32);
2954
2955 /* If no records, return failure now so .BTF.ext won't be used. */
2956 if (!info_left) {
2957 pr_debug("%s section in .BTF.ext has no records\n", ext_sec->desc);
2958 return -EINVAL;
2959 }
2960
2961 while (info_left) {
2962 unsigned int sec_hdrlen = sizeof(struct btf_ext_info_sec);
2963 __u64 total_record_size;
2964 __u32 num_records;
2965
2966 if (info_left < sec_hdrlen) {
2967 pr_debug("%s section header is not found in .BTF.ext\n",
2968 ext_sec->desc);
2969 return -EINVAL;
2970 }
2971
2972 num_records = is_native ? sinfo->num_info : bswap_32(sinfo->num_info);
2973 if (num_records == 0) {
2974 pr_debug("%s section has incorrect num_records in .BTF.ext\n",
2975 ext_sec->desc);
2976 return -EINVAL;
2977 }
2978
2979 total_record_size = sec_hdrlen + (__u64)num_records * record_size;
2980 if (info_left < total_record_size) {
2981 pr_debug("%s section has incorrect num_records in .BTF.ext\n",
2982 ext_sec->desc);
2983 return -EINVAL;
2984 }
2985
2986 info_left -= total_record_size;
2987 sinfo = (void *)sinfo + total_record_size;
2988 sec_cnt++;
2989 }
2990
2991 ext_info = ext_sec->ext_info;
2992 ext_info->len = ext_sec->len - sizeof(__u32);
2993 ext_info->rec_size = record_size;
2994 ext_info->info = info + sizeof(__u32);
2995 ext_info->sec_cnt = sec_cnt;
2996
2997 return 0;
2998}
2999
3000/* Parse all info secs in the BTF.ext info data */
3001static int btf_ext_parse_info(struct btf_ext *btf_ext, bool is_native)
3002{
3003 struct btf_ext_sec_info_param func_info = {
3004 .off = btf_ext->hdr->func_info_off,
3005 .len = btf_ext->hdr->func_info_len,
3006 .min_rec_size = sizeof(struct bpf_func_info_min),
3007 .ext_info = &btf_ext->func_info,
3008 .desc = "func_info"
3009 };
3010 struct btf_ext_sec_info_param line_info = {
3011 .off = btf_ext->hdr->line_info_off,
3012 .len = btf_ext->hdr->line_info_len,
3013 .min_rec_size = sizeof(struct bpf_line_info_min),
3014 .ext_info = &btf_ext->line_info,
3015 .desc = "line_info",
3016 };
3017 struct btf_ext_sec_info_param core_relo = {
3018 .off = btf_ext->hdr->core_relo_off,
3019 .len = btf_ext->hdr->core_relo_len,
3020 .min_rec_size = sizeof(struct bpf_core_relo),
3021 .ext_info = &btf_ext->core_relo_info,
3022 .desc = "core_relo",
3023 };
3024 int err;
3025
3026 err = btf_ext_parse_sec_info(btf_ext, &func_info, is_native);
3027 if (err)
3028 return err;
3029
3030 err = btf_ext_parse_sec_info(btf_ext, &line_info, is_native);
3031 if (err)
3032 return err;
3033
3034 if (btf_ext->hdr->hdr_len < offsetofend(struct btf_ext_header, core_relo_len))
3035 return 0; /* skip core relos parsing */
3036
3037 err = btf_ext_parse_sec_info(btf_ext, &core_relo, is_native);
3038 if (err)
3039 return err;
3040
3041 return 0;
3042}
3043
3044/* Swap byte-order of BTF.ext header with any endianness */
3045static void btf_ext_bswap_hdr(struct btf_ext_header *h)
3046{
3047 bool is_native = h->magic == BTF_MAGIC;
3048 __u32 hdr_len;
3049
3050 hdr_len = is_native ? h->hdr_len : bswap_32(h->hdr_len);
3051
3052 h->magic = bswap_16(h->magic);
3053 h->hdr_len = bswap_32(h->hdr_len);
3054 h->func_info_off = bswap_32(h->func_info_off);
3055 h->func_info_len = bswap_32(h->func_info_len);
3056 h->line_info_off = bswap_32(h->line_info_off);
3057 h->line_info_len = bswap_32(h->line_info_len);
3058
3059 if (hdr_len < offsetofend(struct btf_ext_header, core_relo_len))
3060 return;
3061
3062 h->core_relo_off = bswap_32(h->core_relo_off);
3063 h->core_relo_len = bswap_32(h->core_relo_len);
3064}
3065
3066/* Swap byte-order of generic info subsection */
3067static void btf_ext_bswap_info_sec(void *info, __u32 len, bool is_native,
3068 info_rec_bswap_fn bswap_fn)
3069{
3070 struct btf_ext_info_sec *sec;
3071 __u32 info_left, rec_size, *rs;
3072
3073 if (len == 0)
3074 return;
3075
3076 rs = info; /* info record size */
3077 rec_size = is_native ? *rs : bswap_32(*rs);
3078 *rs = bswap_32(*rs);
3079
3080 sec = info + sizeof(__u32); /* info sec #1 */
3081 info_left = len - sizeof(__u32);
3082 while (info_left) {
3083 unsigned int sec_hdrlen = sizeof(struct btf_ext_info_sec);
3084 __u32 i, num_recs;
3085 void *p;
3086
3087 num_recs = is_native ? sec->num_info : bswap_32(sec->num_info);
3088 sec->sec_name_off = bswap_32(sec->sec_name_off);
3089 sec->num_info = bswap_32(sec->num_info);
3090 p = sec->data; /* info rec #1 */
3091 for (i = 0; i < num_recs; i++, p += rec_size)
3092 bswap_fn(p);
3093 sec = p;
3094 info_left -= sec_hdrlen + (__u64)rec_size * num_recs;
3095 }
3096}
3097
3098/*
3099 * Swap byte-order of all info data in a BTF.ext section
3100 * - requires BTF.ext hdr in native endianness
3101 */
3102static void btf_ext_bswap_info(struct btf_ext *btf_ext, void *data)
3103{
3104 const bool is_native = btf_ext->swapped_endian;
3105 const struct btf_ext_header *h = data;
3106 void *info;
3107
3108 /* Swap func_info subsection byte-order */
3109 info = data + h->hdr_len + h->func_info_off;
3110 btf_ext_bswap_info_sec(info, h->func_info_len, is_native,
3111 (info_rec_bswap_fn)bpf_func_info_bswap);
3112
3113 /* Swap line_info subsection byte-order */
3114 info = data + h->hdr_len + h->line_info_off;
3115 btf_ext_bswap_info_sec(info, h->line_info_len, is_native,
3116 (info_rec_bswap_fn)bpf_line_info_bswap);
3117
3118 /* Swap core_relo subsection byte-order (if present) */
3119 if (h->hdr_len < offsetofend(struct btf_ext_header, core_relo_len))
3120 return;
3121
3122 info = data + h->hdr_len + h->core_relo_off;
3123 btf_ext_bswap_info_sec(info, h->core_relo_len, is_native,
3124 (info_rec_bswap_fn)bpf_core_relo_bswap);
3125}
3126
3127/* Parse hdr data and info sections: check and convert to native endianness */
3128static int btf_ext_parse(struct btf_ext *btf_ext)
3129{
3130 __u32 hdr_len, data_size = btf_ext->data_size;
3131 struct btf_ext_header *hdr = btf_ext->hdr;
3132 bool swapped_endian = false;
3133 int err;
3134
3135 if (data_size < offsetofend(struct btf_ext_header, hdr_len)) {
3136 pr_debug("BTF.ext header too short\n");
3137 return -EINVAL;
3138 }
3139
3140 hdr_len = hdr->hdr_len;
3141 if (hdr->magic == bswap_16(BTF_MAGIC)) {
3142 swapped_endian = true;
3143 hdr_len = bswap_32(hdr_len);
3144 } else if (hdr->magic != BTF_MAGIC) {
3145 pr_debug("Invalid BTF.ext magic:%x\n", hdr->magic);
3146 return -EINVAL;
3147 }
3148
3149 /* Ensure known version of structs, current BTF_VERSION == 1 */
3150 if (hdr->version != 1) {
3151 pr_debug("Unsupported BTF.ext version:%u\n", hdr->version);
3152 return -ENOTSUP;
3153 }
3154
3155 if (hdr->flags) {
3156 pr_debug("Unsupported BTF.ext flags:%x\n", hdr->flags);
3157 return -ENOTSUP;
3158 }
3159
3160 if (data_size < hdr_len) {
3161 pr_debug("BTF.ext header not found\n");
3162 return -EINVAL;
3163 } else if (data_size == hdr_len) {
3164 pr_debug("BTF.ext has no data\n");
3165 return -EINVAL;
3166 }
3167
3168 /* Verify mandatory hdr info details present */
3169 if (hdr_len < offsetofend(struct btf_ext_header, line_info_len)) {
3170 pr_warn("BTF.ext header missing func_info, line_info\n");
3171 return -EINVAL;
3172 }
3173
3174 /* Keep hdr native byte-order in memory for introspection */
3175 if (swapped_endian)
3176 btf_ext_bswap_hdr(btf_ext->hdr);
3177
3178 /* Validate info subsections and cache key metadata */
3179 err = btf_ext_parse_info(btf_ext, !swapped_endian);
3180 if (err)
3181 return err;
3182
3183 /* Keep infos native byte-order in memory for introspection */
3184 if (swapped_endian)
3185 btf_ext_bswap_info(btf_ext, btf_ext->data);
3186
3187 /*
3188 * Set btf_ext->swapped_endian only after all header and info data has
3189 * been swapped, helping bswap functions determine if their data are
3190 * in native byte-order when called.
3191 */
3192 btf_ext->swapped_endian = swapped_endian;
3193 return 0;
3194}
3195
3196void btf_ext__free(struct btf_ext *btf_ext)
3197{
3198 if (IS_ERR_OR_NULL(btf_ext))
3199 return;
3200 free(btf_ext->func_info.sec_idxs);
3201 free(btf_ext->line_info.sec_idxs);
3202 free(btf_ext->core_relo_info.sec_idxs);
3203 free(btf_ext->data);
3204 free(btf_ext->data_swapped);
3205 free(btf_ext);
3206}
3207
3208struct btf_ext *btf_ext__new(const __u8 *data, __u32 size)
3209{
3210 struct btf_ext *btf_ext;
3211 int err;
3212
3213 btf_ext = calloc(1, sizeof(struct btf_ext));
3214 if (!btf_ext)
3215 return libbpf_err_ptr(-ENOMEM);
3216
3217 btf_ext->data_size = size;
3218 btf_ext->data = malloc(size);
3219 if (!btf_ext->data) {
3220 err = -ENOMEM;
3221 goto done;
3222 }
3223 memcpy(btf_ext->data, data, size);
3224
3225 err = btf_ext_parse(btf_ext);
3226
3227done:
3228 if (err) {
3229 btf_ext__free(btf_ext);
3230 return libbpf_err_ptr(err);
3231 }
3232
3233 return btf_ext;
3234}
3235
3236static void *btf_ext_raw_data(const struct btf_ext *btf_ext_ro, bool swap_endian)
3237{
3238 struct btf_ext *btf_ext = (struct btf_ext *)btf_ext_ro;
3239 const __u32 data_sz = btf_ext->data_size;
3240 void *data;
3241
3242 /* Return native data (always present) or swapped data if present */
3243 if (!swap_endian)
3244 return btf_ext->data;
3245 else if (btf_ext->data_swapped)
3246 return btf_ext->data_swapped;
3247
3248 /* Recreate missing swapped data, then cache and return */
3249 data = calloc(1, data_sz);
3250 if (!data)
3251 return NULL;
3252 memcpy(data, btf_ext->data, data_sz);
3253
3254 btf_ext_bswap_info(btf_ext, data);
3255 btf_ext_bswap_hdr(data);
3256 btf_ext->data_swapped = data;
3257 return data;
3258}
3259
3260const void *btf_ext__raw_data(const struct btf_ext *btf_ext, __u32 *size)
3261{
3262 void *data;
3263
3264 data = btf_ext_raw_data(btf_ext, btf_ext->swapped_endian);
3265 if (!data)
3266 return errno = ENOMEM, NULL;
3267
3268 *size = btf_ext->data_size;
3269 return data;
3270}
3271
3272__attribute__((alias("btf_ext__raw_data")))
3273const void *btf_ext__get_raw_data(const struct btf_ext *btf_ext, __u32 *size);
3274
3275enum btf_endianness btf_ext__endianness(const struct btf_ext *btf_ext)
3276{
3277 if (is_host_big_endian())
3278 return btf_ext->swapped_endian ? BTF_LITTLE_ENDIAN : BTF_BIG_ENDIAN;
3279 else
3280 return btf_ext->swapped_endian ? BTF_BIG_ENDIAN : BTF_LITTLE_ENDIAN;
3281}
3282
3283int btf_ext__set_endianness(struct btf_ext *btf_ext, enum btf_endianness endian)
3284{
3285 if (endian != BTF_LITTLE_ENDIAN && endian != BTF_BIG_ENDIAN)
3286 return libbpf_err(-EINVAL);
3287
3288 btf_ext->swapped_endian = is_host_big_endian() != (endian == BTF_BIG_ENDIAN);
3289
3290 if (!btf_ext->swapped_endian) {
3291 free(btf_ext->data_swapped);
3292 btf_ext->data_swapped = NULL;
3293 }
3294 return 0;
3295}
3296
3297struct btf_dedup;
3298
3299static struct btf_dedup *btf_dedup_new(struct btf *btf, const struct btf_dedup_opts *opts);
3300static void btf_dedup_free(struct btf_dedup *d);
3301static int btf_dedup_prep(struct btf_dedup *d);
3302static int btf_dedup_strings(struct btf_dedup *d);
3303static int btf_dedup_prim_types(struct btf_dedup *d);
3304static int btf_dedup_struct_types(struct btf_dedup *d);
3305static int btf_dedup_ref_types(struct btf_dedup *d);
3306static int btf_dedup_resolve_fwds(struct btf_dedup *d);
3307static int btf_dedup_compact_types(struct btf_dedup *d);
3308static int btf_dedup_remap_types(struct btf_dedup *d);
3309
3310/*
3311 * Deduplicate BTF types and strings.
3312 *
3313 * BTF dedup algorithm takes as an input `struct btf` representing `.BTF` ELF
3314 * section with all BTF type descriptors and string data. It overwrites that
3315 * memory in-place with deduplicated types and strings without any loss of
3316 * information. If optional `struct btf_ext` representing '.BTF.ext' ELF section
3317 * is provided, all the strings referenced from .BTF.ext section are honored
3318 * and updated to point to the right offsets after deduplication.
3319 *
3320 * If function returns with error, type/string data might be garbled and should
3321 * be discarded.
3322 *
3323 * More verbose and detailed description of both problem btf_dedup is solving,
3324 * as well as solution could be found at:
3325 * https://facebookmicrosites.github.io/bpf/blog/2018/11/14/btf-enhancement.html
3326 *
3327 * Problem description and justification
3328 * =====================================
3329 *
3330 * BTF type information is typically emitted either as a result of conversion
3331 * from DWARF to BTF or directly by compiler. In both cases, each compilation
3332 * unit contains information about a subset of all the types that are used
3333 * in an application. These subsets are frequently overlapping and contain a lot
3334 * of duplicated information when later concatenated together into a single
3335 * binary. This algorithm ensures that each unique type is represented by single
3336 * BTF type descriptor, greatly reducing resulting size of BTF data.
3337 *
3338 * Compilation unit isolation and subsequent duplication of data is not the only
3339 * problem. The same type hierarchy (e.g., struct and all the type that struct
3340 * references) in different compilation units can be represented in BTF to
3341 * various degrees of completeness (or, rather, incompleteness) due to
3342 * struct/union forward declarations.
3343 *
3344 * Let's take a look at an example, that we'll use to better understand the
3345 * problem (and solution). Suppose we have two compilation units, each using
3346 * same `struct S`, but each of them having incomplete type information about
3347 * struct's fields:
3348 *
3349 * // CU #1:
3350 * struct S;
3351 * struct A {
3352 * int a;
3353 * struct A* self;
3354 * struct S* parent;
3355 * };
3356 * struct B;
3357 * struct S {
3358 * struct A* a_ptr;
3359 * struct B* b_ptr;
3360 * };
3361 *
3362 * // CU #2:
3363 * struct S;
3364 * struct A;
3365 * struct B {
3366 * int b;
3367 * struct B* self;
3368 * struct S* parent;
3369 * };
3370 * struct S {
3371 * struct A* a_ptr;
3372 * struct B* b_ptr;
3373 * };
3374 *
3375 * In case of CU #1, BTF data will know only that `struct B` exist (but no
3376 * more), but will know the complete type information about `struct A`. While
3377 * for CU #2, it will know full type information about `struct B`, but will
3378 * only know about forward declaration of `struct A` (in BTF terms, it will
3379 * have `BTF_KIND_FWD` type descriptor with name `B`).
3380 *
3381 * This compilation unit isolation means that it's possible that there is no
3382 * single CU with complete type information describing structs `S`, `A`, and
3383 * `B`. Also, we might get tons of duplicated and redundant type information.
3384 *
3385 * Additional complication we need to keep in mind comes from the fact that
3386 * types, in general, can form graphs containing cycles, not just DAGs.
3387 *
3388 * While algorithm does deduplication, it also merges and resolves type
3389 * information (unless disabled throught `struct btf_opts`), whenever possible.
3390 * E.g., in the example above with two compilation units having partial type
3391 * information for structs `A` and `B`, the output of algorithm will emit
3392 * a single copy of each BTF type that describes structs `A`, `B`, and `S`
3393 * (as well as type information for `int` and pointers), as if they were defined
3394 * in a single compilation unit as:
3395 *
3396 * struct A {
3397 * int a;
3398 * struct A* self;
3399 * struct S* parent;
3400 * };
3401 * struct B {
3402 * int b;
3403 * struct B* self;
3404 * struct S* parent;
3405 * };
3406 * struct S {
3407 * struct A* a_ptr;
3408 * struct B* b_ptr;
3409 * };
3410 *
3411 * Algorithm summary
3412 * =================
3413 *
3414 * Algorithm completes its work in 7 separate passes:
3415 *
3416 * 1. Strings deduplication.
3417 * 2. Primitive types deduplication (int, enum, fwd).
3418 * 3. Struct/union types deduplication.
3419 * 4. Resolve unambiguous forward declarations.
3420 * 5. Reference types deduplication (pointers, typedefs, arrays, funcs, func
3421 * protos, and const/volatile/restrict modifiers).
3422 * 6. Types compaction.
3423 * 7. Types remapping.
3424 *
3425 * Algorithm determines canonical type descriptor, which is a single
3426 * representative type for each truly unique type. This canonical type is the
3427 * one that will go into final deduplicated BTF type information. For
3428 * struct/unions, it is also the type that algorithm will merge additional type
3429 * information into (while resolving FWDs), as it discovers it from data in
3430 * other CUs. Each input BTF type eventually gets either mapped to itself, if
3431 * that type is canonical, or to some other type, if that type is equivalent
3432 * and was chosen as canonical representative. This mapping is stored in
3433 * `btf_dedup->map` array. This map is also used to record STRUCT/UNION that
3434 * FWD type got resolved to.
3435 *
3436 * To facilitate fast discovery of canonical types, we also maintain canonical
3437 * index (`btf_dedup->dedup_table`), which maps type descriptor's signature hash
3438 * (i.e., hashed kind, name, size, fields, etc) into a list of canonical types
3439 * that match that signature. With sufficiently good choice of type signature
3440 * hashing function, we can limit number of canonical types for each unique type
3441 * signature to a very small number, allowing to find canonical type for any
3442 * duplicated type very quickly.
3443 *
3444 * Struct/union deduplication is the most critical part and algorithm for
3445 * deduplicating structs/unions is described in greater details in comments for
3446 * `btf_dedup_is_equiv` function.
3447 */
3448int btf__dedup(struct btf *btf, const struct btf_dedup_opts *opts)
3449{
3450 struct btf_dedup *d;
3451 int err;
3452
3453 if (!OPTS_VALID(opts, btf_dedup_opts))
3454 return libbpf_err(-EINVAL);
3455
3456 d = btf_dedup_new(btf, opts);
3457 if (IS_ERR(d)) {
3458 pr_debug("btf_dedup_new failed: %ld\n", PTR_ERR(d));
3459 return libbpf_err(-EINVAL);
3460 }
3461
3462 if (btf_ensure_modifiable(btf)) {
3463 err = -ENOMEM;
3464 goto done;
3465 }
3466
3467 err = btf_dedup_prep(d);
3468 if (err) {
3469 pr_debug("btf_dedup_prep failed: %s\n", errstr(err));
3470 goto done;
3471 }
3472 err = btf_dedup_strings(d);
3473 if (err < 0) {
3474 pr_debug("btf_dedup_strings failed: %s\n", errstr(err));
3475 goto done;
3476 }
3477 err = btf_dedup_prim_types(d);
3478 if (err < 0) {
3479 pr_debug("btf_dedup_prim_types failed: %s\n", errstr(err));
3480 goto done;
3481 }
3482 err = btf_dedup_struct_types(d);
3483 if (err < 0) {
3484 pr_debug("btf_dedup_struct_types failed: %s\n", errstr(err));
3485 goto done;
3486 }
3487 err = btf_dedup_resolve_fwds(d);
3488 if (err < 0) {
3489 pr_debug("btf_dedup_resolve_fwds failed: %s\n", errstr(err));
3490 goto done;
3491 }
3492 err = btf_dedup_ref_types(d);
3493 if (err < 0) {
3494 pr_debug("btf_dedup_ref_types failed: %s\n", errstr(err));
3495 goto done;
3496 }
3497 err = btf_dedup_compact_types(d);
3498 if (err < 0) {
3499 pr_debug("btf_dedup_compact_types failed: %s\n", errstr(err));
3500 goto done;
3501 }
3502 err = btf_dedup_remap_types(d);
3503 if (err < 0) {
3504 pr_debug("btf_dedup_remap_types failed: %s\n", errstr(err));
3505 goto done;
3506 }
3507
3508done:
3509 btf_dedup_free(d);
3510 return libbpf_err(err);
3511}
3512
3513#define BTF_UNPROCESSED_ID ((__u32)-1)
3514#define BTF_IN_PROGRESS_ID ((__u32)-2)
3515
3516struct btf_dedup {
3517 /* .BTF section to be deduped in-place */
3518 struct btf *btf;
3519 /*
3520 * Optional .BTF.ext section. When provided, any strings referenced
3521 * from it will be taken into account when deduping strings
3522 */
3523 struct btf_ext *btf_ext;
3524 /*
3525 * This is a map from any type's signature hash to a list of possible
3526 * canonical representative type candidates. Hash collisions are
3527 * ignored, so even types of various kinds can share same list of
3528 * candidates, which is fine because we rely on subsequent
3529 * btf_xxx_equal() checks to authoritatively verify type equality.
3530 */
3531 struct hashmap *dedup_table;
3532 /* Canonical types map */
3533 __u32 *map;
3534 /* Hypothetical mapping, used during type graph equivalence checks */
3535 __u32 *hypot_map;
3536 __u32 *hypot_list;
3537 size_t hypot_cnt;
3538 size_t hypot_cap;
3539 /* Whether hypothetical mapping, if successful, would need to adjust
3540 * already canonicalized types (due to a new forward declaration to
3541 * concrete type resolution). In such case, during split BTF dedup
3542 * candidate type would still be considered as different, because base
3543 * BTF is considered to be immutable.
3544 */
3545 bool hypot_adjust_canon;
3546 /* Various option modifying behavior of algorithm */
3547 struct btf_dedup_opts opts;
3548 /* temporary strings deduplication state */
3549 struct strset *strs_set;
3550};
3551
3552static unsigned long hash_combine(unsigned long h, unsigned long value)
3553{
3554 return h * 31 + value;
3555}
3556
3557#define for_each_dedup_cand(d, node, hash) \
3558 hashmap__for_each_key_entry(d->dedup_table, node, hash)
3559
3560static int btf_dedup_table_add(struct btf_dedup *d, long hash, __u32 type_id)
3561{
3562 return hashmap__append(d->dedup_table, hash, type_id);
3563}
3564
3565static int btf_dedup_hypot_map_add(struct btf_dedup *d,
3566 __u32 from_id, __u32 to_id)
3567{
3568 if (d->hypot_cnt == d->hypot_cap) {
3569 __u32 *new_list;
3570
3571 d->hypot_cap += max((size_t)16, d->hypot_cap / 2);
3572 new_list = libbpf_reallocarray(d->hypot_list, d->hypot_cap, sizeof(__u32));
3573 if (!new_list)
3574 return -ENOMEM;
3575 d->hypot_list = new_list;
3576 }
3577 d->hypot_list[d->hypot_cnt++] = from_id;
3578 d->hypot_map[from_id] = to_id;
3579 return 0;
3580}
3581
3582static void btf_dedup_clear_hypot_map(struct btf_dedup *d)
3583{
3584 int i;
3585
3586 for (i = 0; i < d->hypot_cnt; i++)
3587 d->hypot_map[d->hypot_list[i]] = BTF_UNPROCESSED_ID;
3588 d->hypot_cnt = 0;
3589 d->hypot_adjust_canon = false;
3590}
3591
3592static void btf_dedup_free(struct btf_dedup *d)
3593{
3594 hashmap__free(d->dedup_table);
3595 d->dedup_table = NULL;
3596
3597 free(d->map);
3598 d->map = NULL;
3599
3600 free(d->hypot_map);
3601 d->hypot_map = NULL;
3602
3603 free(d->hypot_list);
3604 d->hypot_list = NULL;
3605
3606 free(d);
3607}
3608
3609static size_t btf_dedup_identity_hash_fn(long key, void *ctx)
3610{
3611 return key;
3612}
3613
3614static size_t btf_dedup_collision_hash_fn(long key, void *ctx)
3615{
3616 return 0;
3617}
3618
3619static bool btf_dedup_equal_fn(long k1, long k2, void *ctx)
3620{
3621 return k1 == k2;
3622}
3623
3624static struct btf_dedup *btf_dedup_new(struct btf *btf, const struct btf_dedup_opts *opts)
3625{
3626 struct btf_dedup *d = calloc(1, sizeof(struct btf_dedup));
3627 hashmap_hash_fn hash_fn = btf_dedup_identity_hash_fn;
3628 int i, err = 0, type_cnt;
3629
3630 if (!d)
3631 return ERR_PTR(-ENOMEM);
3632
3633 if (OPTS_GET(opts, force_collisions, false))
3634 hash_fn = btf_dedup_collision_hash_fn;
3635
3636 d->btf = btf;
3637 d->btf_ext = OPTS_GET(opts, btf_ext, NULL);
3638
3639 d->dedup_table = hashmap__new(hash_fn, btf_dedup_equal_fn, NULL);
3640 if (IS_ERR(d->dedup_table)) {
3641 err = PTR_ERR(d->dedup_table);
3642 d->dedup_table = NULL;
3643 goto done;
3644 }
3645
3646 type_cnt = btf__type_cnt(btf);
3647 d->map = malloc(sizeof(__u32) * type_cnt);
3648 if (!d->map) {
3649 err = -ENOMEM;
3650 goto done;
3651 }
3652 /* special BTF "void" type is made canonical immediately */
3653 d->map[0] = 0;
3654 for (i = 1; i < type_cnt; i++) {
3655 struct btf_type *t = btf_type_by_id(d->btf, i);
3656
3657 /* VAR and DATASEC are never deduped and are self-canonical */
3658 if (btf_is_var(t) || btf_is_datasec(t))
3659 d->map[i] = i;
3660 else
3661 d->map[i] = BTF_UNPROCESSED_ID;
3662 }
3663
3664 d->hypot_map = malloc(sizeof(__u32) * type_cnt);
3665 if (!d->hypot_map) {
3666 err = -ENOMEM;
3667 goto done;
3668 }
3669 for (i = 0; i < type_cnt; i++)
3670 d->hypot_map[i] = BTF_UNPROCESSED_ID;
3671
3672done:
3673 if (err) {
3674 btf_dedup_free(d);
3675 return ERR_PTR(err);
3676 }
3677
3678 return d;
3679}
3680
3681/*
3682 * Iterate over all possible places in .BTF and .BTF.ext that can reference
3683 * string and pass pointer to it to a provided callback `fn`.
3684 */
3685static int btf_for_each_str_off(struct btf_dedup *d, str_off_visit_fn fn, void *ctx)
3686{
3687 int i, r;
3688
3689 for (i = 0; i < d->btf->nr_types; i++) {
3690 struct btf_field_iter it;
3691 struct btf_type *t = btf_type_by_id(d->btf, d->btf->start_id + i);
3692 __u32 *str_off;
3693
3694 r = btf_field_iter_init(&it, t, BTF_FIELD_ITER_STRS);
3695 if (r)
3696 return r;
3697
3698 while ((str_off = btf_field_iter_next(&it))) {
3699 r = fn(str_off, ctx);
3700 if (r)
3701 return r;
3702 }
3703 }
3704
3705 if (!d->btf_ext)
3706 return 0;
3707
3708 r = btf_ext_visit_str_offs(d->btf_ext, fn, ctx);
3709 if (r)
3710 return r;
3711
3712 return 0;
3713}
3714
3715static int strs_dedup_remap_str_off(__u32 *str_off_ptr, void *ctx)
3716{
3717 struct btf_dedup *d = ctx;
3718 __u32 str_off = *str_off_ptr;
3719 const char *s;
3720 int off, err;
3721
3722 /* don't touch empty string or string in main BTF */
3723 if (str_off == 0 || str_off < d->btf->start_str_off)
3724 return 0;
3725
3726 s = btf__str_by_offset(d->btf, str_off);
3727 if (d->btf->base_btf) {
3728 err = btf__find_str(d->btf->base_btf, s);
3729 if (err >= 0) {
3730 *str_off_ptr = err;
3731 return 0;
3732 }
3733 if (err != -ENOENT)
3734 return err;
3735 }
3736
3737 off = strset__add_str(d->strs_set, s);
3738 if (off < 0)
3739 return off;
3740
3741 *str_off_ptr = d->btf->start_str_off + off;
3742 return 0;
3743}
3744
3745/*
3746 * Dedup string and filter out those that are not referenced from either .BTF
3747 * or .BTF.ext (if provided) sections.
3748 *
3749 * This is done by building index of all strings in BTF's string section,
3750 * then iterating over all entities that can reference strings (e.g., type
3751 * names, struct field names, .BTF.ext line info, etc) and marking corresponding
3752 * strings as used. After that all used strings are deduped and compacted into
3753 * sequential blob of memory and new offsets are calculated. Then all the string
3754 * references are iterated again and rewritten using new offsets.
3755 */
3756static int btf_dedup_strings(struct btf_dedup *d)
3757{
3758 int err;
3759
3760 if (d->btf->strs_deduped)
3761 return 0;
3762
3763 d->strs_set = strset__new(BTF_MAX_STR_OFFSET, NULL, 0);
3764 if (IS_ERR(d->strs_set)) {
3765 err = PTR_ERR(d->strs_set);
3766 goto err_out;
3767 }
3768
3769 if (!d->btf->base_btf) {
3770 /* insert empty string; we won't be looking it up during strings
3771 * dedup, but it's good to have it for generic BTF string lookups
3772 */
3773 err = strset__add_str(d->strs_set, "");
3774 if (err < 0)
3775 goto err_out;
3776 }
3777
3778 /* remap string offsets */
3779 err = btf_for_each_str_off(d, strs_dedup_remap_str_off, d);
3780 if (err)
3781 goto err_out;
3782
3783 /* replace BTF string data and hash with deduped ones */
3784 strset__free(d->btf->strs_set);
3785 d->btf->hdr->str_len = strset__data_size(d->strs_set);
3786 d->btf->strs_set = d->strs_set;
3787 d->strs_set = NULL;
3788 d->btf->strs_deduped = true;
3789 return 0;
3790
3791err_out:
3792 strset__free(d->strs_set);
3793 d->strs_set = NULL;
3794
3795 return err;
3796}
3797
3798static long btf_hash_common(struct btf_type *t)
3799{
3800 long h;
3801
3802 h = hash_combine(0, t->name_off);
3803 h = hash_combine(h, t->info);
3804 h = hash_combine(h, t->size);
3805 return h;
3806}
3807
3808static bool btf_equal_common(struct btf_type *t1, struct btf_type *t2)
3809{
3810 return t1->name_off == t2->name_off &&
3811 t1->info == t2->info &&
3812 t1->size == t2->size;
3813}
3814
3815/* Calculate type signature hash of INT or TAG. */
3816static long btf_hash_int_decl_tag(struct btf_type *t)
3817{
3818 __u32 info = *(__u32 *)(t + 1);
3819 long h;
3820
3821 h = btf_hash_common(t);
3822 h = hash_combine(h, info);
3823 return h;
3824}
3825
3826/* Check structural equality of two INTs or TAGs. */
3827static bool btf_equal_int_tag(struct btf_type *t1, struct btf_type *t2)
3828{
3829 __u32 info1, info2;
3830
3831 if (!btf_equal_common(t1, t2))
3832 return false;
3833 info1 = *(__u32 *)(t1 + 1);
3834 info2 = *(__u32 *)(t2 + 1);
3835 return info1 == info2;
3836}
3837
3838/* Calculate type signature hash of ENUM/ENUM64. */
3839static long btf_hash_enum(struct btf_type *t)
3840{
3841 long h;
3842
3843 /* don't hash vlen, enum members and size to support enum fwd resolving */
3844 h = hash_combine(0, t->name_off);
3845 return h;
3846}
3847
3848static bool btf_equal_enum_members(struct btf_type *t1, struct btf_type *t2)
3849{
3850 const struct btf_enum *m1, *m2;
3851 __u16 vlen;
3852 int i;
3853
3854 vlen = btf_vlen(t1);
3855 m1 = btf_enum(t1);
3856 m2 = btf_enum(t2);
3857 for (i = 0; i < vlen; i++) {
3858 if (m1->name_off != m2->name_off || m1->val != m2->val)
3859 return false;
3860 m1++;
3861 m2++;
3862 }
3863 return true;
3864}
3865
3866static bool btf_equal_enum64_members(struct btf_type *t1, struct btf_type *t2)
3867{
3868 const struct btf_enum64 *m1, *m2;
3869 __u16 vlen;
3870 int i;
3871
3872 vlen = btf_vlen(t1);
3873 m1 = btf_enum64(t1);
3874 m2 = btf_enum64(t2);
3875 for (i = 0; i < vlen; i++) {
3876 if (m1->name_off != m2->name_off || m1->val_lo32 != m2->val_lo32 ||
3877 m1->val_hi32 != m2->val_hi32)
3878 return false;
3879 m1++;
3880 m2++;
3881 }
3882 return true;
3883}
3884
3885/* Check structural equality of two ENUMs or ENUM64s. */
3886static bool btf_equal_enum(struct btf_type *t1, struct btf_type *t2)
3887{
3888 if (!btf_equal_common(t1, t2))
3889 return false;
3890
3891 /* t1 & t2 kinds are identical because of btf_equal_common */
3892 if (btf_kind(t1) == BTF_KIND_ENUM)
3893 return btf_equal_enum_members(t1, t2);
3894 else
3895 return btf_equal_enum64_members(t1, t2);
3896}
3897
3898static inline bool btf_is_enum_fwd(struct btf_type *t)
3899{
3900 return btf_is_any_enum(t) && btf_vlen(t) == 0;
3901}
3902
3903static bool btf_compat_enum(struct btf_type *t1, struct btf_type *t2)
3904{
3905 if (!btf_is_enum_fwd(t1) && !btf_is_enum_fwd(t2))
3906 return btf_equal_enum(t1, t2);
3907 /* At this point either t1 or t2 or both are forward declarations, thus:
3908 * - skip comparing vlen because it is zero for forward declarations;
3909 * - skip comparing size to allow enum forward declarations
3910 * to be compatible with enum64 full declarations;
3911 * - skip comparing kind for the same reason.
3912 */
3913 return t1->name_off == t2->name_off &&
3914 btf_is_any_enum(t1) && btf_is_any_enum(t2);
3915}
3916
3917/*
3918 * Calculate type signature hash of STRUCT/UNION, ignoring referenced type IDs,
3919 * as referenced type IDs equivalence is established separately during type
3920 * graph equivalence check algorithm.
3921 */
3922static long btf_hash_struct(struct btf_type *t)
3923{
3924 const struct btf_member *member = btf_members(t);
3925 __u32 vlen = btf_vlen(t);
3926 long h = btf_hash_common(t);
3927 int i;
3928
3929 for (i = 0; i < vlen; i++) {
3930 h = hash_combine(h, member->name_off);
3931 h = hash_combine(h, member->offset);
3932 /* no hashing of referenced type ID, it can be unresolved yet */
3933 member++;
3934 }
3935 return h;
3936}
3937
3938/*
3939 * Check structural compatibility of two STRUCTs/UNIONs, ignoring referenced
3940 * type IDs. This check is performed during type graph equivalence check and
3941 * referenced types equivalence is checked separately.
3942 */
3943static bool btf_shallow_equal_struct(struct btf_type *t1, struct btf_type *t2)
3944{
3945 const struct btf_member *m1, *m2;
3946 __u16 vlen;
3947 int i;
3948
3949 if (!btf_equal_common(t1, t2))
3950 return false;
3951
3952 vlen = btf_vlen(t1);
3953 m1 = btf_members(t1);
3954 m2 = btf_members(t2);
3955 for (i = 0; i < vlen; i++) {
3956 if (m1->name_off != m2->name_off || m1->offset != m2->offset)
3957 return false;
3958 m1++;
3959 m2++;
3960 }
3961 return true;
3962}
3963
3964/*
3965 * Calculate type signature hash of ARRAY, including referenced type IDs,
3966 * under assumption that they were already resolved to canonical type IDs and
3967 * are not going to change.
3968 */
3969static long btf_hash_array(struct btf_type *t)
3970{
3971 const struct btf_array *info = btf_array(t);
3972 long h = btf_hash_common(t);
3973
3974 h = hash_combine(h, info->type);
3975 h = hash_combine(h, info->index_type);
3976 h = hash_combine(h, info->nelems);
3977 return h;
3978}
3979
3980/*
3981 * Check exact equality of two ARRAYs, taking into account referenced
3982 * type IDs, under assumption that they were already resolved to canonical
3983 * type IDs and are not going to change.
3984 * This function is called during reference types deduplication to compare
3985 * ARRAY to potential canonical representative.
3986 */
3987static bool btf_equal_array(struct btf_type *t1, struct btf_type *t2)
3988{
3989 const struct btf_array *info1, *info2;
3990
3991 if (!btf_equal_common(t1, t2))
3992 return false;
3993
3994 info1 = btf_array(t1);
3995 info2 = btf_array(t2);
3996 return info1->type == info2->type &&
3997 info1->index_type == info2->index_type &&
3998 info1->nelems == info2->nelems;
3999}
4000
4001/*
4002 * Check structural compatibility of two ARRAYs, ignoring referenced type
4003 * IDs. This check is performed during type graph equivalence check and
4004 * referenced types equivalence is checked separately.
4005 */
4006static bool btf_compat_array(struct btf_type *t1, struct btf_type *t2)
4007{
4008 if (!btf_equal_common(t1, t2))
4009 return false;
4010
4011 return btf_array(t1)->nelems == btf_array(t2)->nelems;
4012}
4013
4014/*
4015 * Calculate type signature hash of FUNC_PROTO, including referenced type IDs,
4016 * under assumption that they were already resolved to canonical type IDs and
4017 * are not going to change.
4018 */
4019static long btf_hash_fnproto(struct btf_type *t)
4020{
4021 const struct btf_param *member = btf_params(t);
4022 __u16 vlen = btf_vlen(t);
4023 long h = btf_hash_common(t);
4024 int i;
4025
4026 for (i = 0; i < vlen; i++) {
4027 h = hash_combine(h, member->name_off);
4028 h = hash_combine(h, member->type);
4029 member++;
4030 }
4031 return h;
4032}
4033
4034/*
4035 * Check exact equality of two FUNC_PROTOs, taking into account referenced
4036 * type IDs, under assumption that they were already resolved to canonical
4037 * type IDs and are not going to change.
4038 * This function is called during reference types deduplication to compare
4039 * FUNC_PROTO to potential canonical representative.
4040 */
4041static bool btf_equal_fnproto(struct btf_type *t1, struct btf_type *t2)
4042{
4043 const struct btf_param *m1, *m2;
4044 __u16 vlen;
4045 int i;
4046
4047 if (!btf_equal_common(t1, t2))
4048 return false;
4049
4050 vlen = btf_vlen(t1);
4051 m1 = btf_params(t1);
4052 m2 = btf_params(t2);
4053 for (i = 0; i < vlen; i++) {
4054 if (m1->name_off != m2->name_off || m1->type != m2->type)
4055 return false;
4056 m1++;
4057 m2++;
4058 }
4059 return true;
4060}
4061
4062/*
4063 * Check structural compatibility of two FUNC_PROTOs, ignoring referenced type
4064 * IDs. This check is performed during type graph equivalence check and
4065 * referenced types equivalence is checked separately.
4066 */
4067static bool btf_compat_fnproto(struct btf_type *t1, struct btf_type *t2)
4068{
4069 const struct btf_param *m1, *m2;
4070 __u16 vlen;
4071 int i;
4072
4073 /* skip return type ID */
4074 if (t1->name_off != t2->name_off || t1->info != t2->info)
4075 return false;
4076
4077 vlen = btf_vlen(t1);
4078 m1 = btf_params(t1);
4079 m2 = btf_params(t2);
4080 for (i = 0; i < vlen; i++) {
4081 if (m1->name_off != m2->name_off)
4082 return false;
4083 m1++;
4084 m2++;
4085 }
4086 return true;
4087}
4088
4089/* Prepare split BTF for deduplication by calculating hashes of base BTF's
4090 * types and initializing the rest of the state (canonical type mapping) for
4091 * the fixed base BTF part.
4092 */
4093static int btf_dedup_prep(struct btf_dedup *d)
4094{
4095 struct btf_type *t;
4096 int type_id;
4097 long h;
4098
4099 if (!d->btf->base_btf)
4100 return 0;
4101
4102 for (type_id = 1; type_id < d->btf->start_id; type_id++) {
4103 t = btf_type_by_id(d->btf, type_id);
4104
4105 /* all base BTF types are self-canonical by definition */
4106 d->map[type_id] = type_id;
4107
4108 switch (btf_kind(t)) {
4109 case BTF_KIND_VAR:
4110 case BTF_KIND_DATASEC:
4111 /* VAR and DATASEC are never hash/deduplicated */
4112 continue;
4113 case BTF_KIND_CONST:
4114 case BTF_KIND_VOLATILE:
4115 case BTF_KIND_RESTRICT:
4116 case BTF_KIND_PTR:
4117 case BTF_KIND_FWD:
4118 case BTF_KIND_TYPEDEF:
4119 case BTF_KIND_FUNC:
4120 case BTF_KIND_FLOAT:
4121 case BTF_KIND_TYPE_TAG:
4122 h = btf_hash_common(t);
4123 break;
4124 case BTF_KIND_INT:
4125 case BTF_KIND_DECL_TAG:
4126 h = btf_hash_int_decl_tag(t);
4127 break;
4128 case BTF_KIND_ENUM:
4129 case BTF_KIND_ENUM64:
4130 h = btf_hash_enum(t);
4131 break;
4132 case BTF_KIND_STRUCT:
4133 case BTF_KIND_UNION:
4134 h = btf_hash_struct(t);
4135 break;
4136 case BTF_KIND_ARRAY:
4137 h = btf_hash_array(t);
4138 break;
4139 case BTF_KIND_FUNC_PROTO:
4140 h = btf_hash_fnproto(t);
4141 break;
4142 default:
4143 pr_debug("unknown kind %d for type [%d]\n", btf_kind(t), type_id);
4144 return -EINVAL;
4145 }
4146 if (btf_dedup_table_add(d, h, type_id))
4147 return -ENOMEM;
4148 }
4149
4150 return 0;
4151}
4152
4153/*
4154 * Deduplicate primitive types, that can't reference other types, by calculating
4155 * their type signature hash and comparing them with any possible canonical
4156 * candidate. If no canonical candidate matches, type itself is marked as
4157 * canonical and is added into `btf_dedup->dedup_table` as another candidate.
4158 */
4159static int btf_dedup_prim_type(struct btf_dedup *d, __u32 type_id)
4160{
4161 struct btf_type *t = btf_type_by_id(d->btf, type_id);
4162 struct hashmap_entry *hash_entry;
4163 struct btf_type *cand;
4164 /* if we don't find equivalent type, then we are canonical */
4165 __u32 new_id = type_id;
4166 __u32 cand_id;
4167 long h;
4168
4169 switch (btf_kind(t)) {
4170 case BTF_KIND_CONST:
4171 case BTF_KIND_VOLATILE:
4172 case BTF_KIND_RESTRICT:
4173 case BTF_KIND_PTR:
4174 case BTF_KIND_TYPEDEF:
4175 case BTF_KIND_ARRAY:
4176 case BTF_KIND_STRUCT:
4177 case BTF_KIND_UNION:
4178 case BTF_KIND_FUNC:
4179 case BTF_KIND_FUNC_PROTO:
4180 case BTF_KIND_VAR:
4181 case BTF_KIND_DATASEC:
4182 case BTF_KIND_DECL_TAG:
4183 case BTF_KIND_TYPE_TAG:
4184 return 0;
4185
4186 case BTF_KIND_INT:
4187 h = btf_hash_int_decl_tag(t);
4188 for_each_dedup_cand(d, hash_entry, h) {
4189 cand_id = hash_entry->value;
4190 cand = btf_type_by_id(d->btf, cand_id);
4191 if (btf_equal_int_tag(t, cand)) {
4192 new_id = cand_id;
4193 break;
4194 }
4195 }
4196 break;
4197
4198 case BTF_KIND_ENUM:
4199 case BTF_KIND_ENUM64:
4200 h = btf_hash_enum(t);
4201 for_each_dedup_cand(d, hash_entry, h) {
4202 cand_id = hash_entry->value;
4203 cand = btf_type_by_id(d->btf, cand_id);
4204 if (btf_equal_enum(t, cand)) {
4205 new_id = cand_id;
4206 break;
4207 }
4208 if (btf_compat_enum(t, cand)) {
4209 if (btf_is_enum_fwd(t)) {
4210 /* resolve fwd to full enum */
4211 new_id = cand_id;
4212 break;
4213 }
4214 /* resolve canonical enum fwd to full enum */
4215 d->map[cand_id] = type_id;
4216 }
4217 }
4218 break;
4219
4220 case BTF_KIND_FWD:
4221 case BTF_KIND_FLOAT:
4222 h = btf_hash_common(t);
4223 for_each_dedup_cand(d, hash_entry, h) {
4224 cand_id = hash_entry->value;
4225 cand = btf_type_by_id(d->btf, cand_id);
4226 if (btf_equal_common(t, cand)) {
4227 new_id = cand_id;
4228 break;
4229 }
4230 }
4231 break;
4232
4233 default:
4234 return -EINVAL;
4235 }
4236
4237 d->map[type_id] = new_id;
4238 if (type_id == new_id && btf_dedup_table_add(d, h, type_id))
4239 return -ENOMEM;
4240
4241 return 0;
4242}
4243
4244static int btf_dedup_prim_types(struct btf_dedup *d)
4245{
4246 int i, err;
4247
4248 for (i = 0; i < d->btf->nr_types; i++) {
4249 err = btf_dedup_prim_type(d, d->btf->start_id + i);
4250 if (err)
4251 return err;
4252 }
4253 return 0;
4254}
4255
4256/*
4257 * Check whether type is already mapped into canonical one (could be to itself).
4258 */
4259static inline bool is_type_mapped(struct btf_dedup *d, uint32_t type_id)
4260{
4261 return d->map[type_id] <= BTF_MAX_NR_TYPES;
4262}
4263
4264/*
4265 * Resolve type ID into its canonical type ID, if any; otherwise return original
4266 * type ID. If type is FWD and is resolved into STRUCT/UNION already, follow
4267 * STRUCT/UNION link and resolve it into canonical type ID as well.
4268 */
4269static inline __u32 resolve_type_id(struct btf_dedup *d, __u32 type_id)
4270{
4271 while (is_type_mapped(d, type_id) && d->map[type_id] != type_id)
4272 type_id = d->map[type_id];
4273 return type_id;
4274}
4275
4276/*
4277 * Resolve FWD to underlying STRUCT/UNION, if any; otherwise return original
4278 * type ID.
4279 */
4280static uint32_t resolve_fwd_id(struct btf_dedup *d, uint32_t type_id)
4281{
4282 __u32 orig_type_id = type_id;
4283
4284 if (!btf_is_fwd(btf__type_by_id(d->btf, type_id)))
4285 return type_id;
4286
4287 while (is_type_mapped(d, type_id) && d->map[type_id] != type_id)
4288 type_id = d->map[type_id];
4289
4290 if (!btf_is_fwd(btf__type_by_id(d->btf, type_id)))
4291 return type_id;
4292
4293 return orig_type_id;
4294}
4295
4296
4297static inline __u16 btf_fwd_kind(struct btf_type *t)
4298{
4299 return btf_kflag(t) ? BTF_KIND_UNION : BTF_KIND_STRUCT;
4300}
4301
4302/* Check if given two types are identical ARRAY definitions */
4303static bool btf_dedup_identical_arrays(struct btf_dedup *d, __u32 id1, __u32 id2)
4304{
4305 struct btf_type *t1, *t2;
4306
4307 t1 = btf_type_by_id(d->btf, id1);
4308 t2 = btf_type_by_id(d->btf, id2);
4309 if (!btf_is_array(t1) || !btf_is_array(t2))
4310 return false;
4311
4312 return btf_equal_array(t1, t2);
4313}
4314
4315/* Check if given two types are identical STRUCT/UNION definitions */
4316static bool btf_dedup_identical_structs(struct btf_dedup *d, __u32 id1, __u32 id2)
4317{
4318 const struct btf_member *m1, *m2;
4319 struct btf_type *t1, *t2;
4320 int n, i;
4321
4322 t1 = btf_type_by_id(d->btf, id1);
4323 t2 = btf_type_by_id(d->btf, id2);
4324
4325 if (!btf_is_composite(t1) || btf_kind(t1) != btf_kind(t2))
4326 return false;
4327
4328 if (!btf_shallow_equal_struct(t1, t2))
4329 return false;
4330
4331 m1 = btf_members(t1);
4332 m2 = btf_members(t2);
4333 for (i = 0, n = btf_vlen(t1); i < n; i++, m1++, m2++) {
4334 if (m1->type != m2->type &&
4335 !btf_dedup_identical_arrays(d, m1->type, m2->type) &&
4336 !btf_dedup_identical_structs(d, m1->type, m2->type))
4337 return false;
4338 }
4339 return true;
4340}
4341
4342/*
4343 * Check equivalence of BTF type graph formed by candidate struct/union (we'll
4344 * call it "candidate graph" in this description for brevity) to a type graph
4345 * formed by (potential) canonical struct/union ("canonical graph" for brevity
4346 * here, though keep in mind that not all types in canonical graph are
4347 * necessarily canonical representatives themselves, some of them might be
4348 * duplicates or its uniqueness might not have been established yet).
4349 * Returns:
4350 * - >0, if type graphs are equivalent;
4351 * - 0, if not equivalent;
4352 * - <0, on error.
4353 *
4354 * Algorithm performs side-by-side DFS traversal of both type graphs and checks
4355 * equivalence of BTF types at each step. If at any point BTF types in candidate
4356 * and canonical graphs are not compatible structurally, whole graphs are
4357 * incompatible. If types are structurally equivalent (i.e., all information
4358 * except referenced type IDs is exactly the same), a mapping from `canon_id` to
4359 * a `cand_id` is recoded in hypothetical mapping (`btf_dedup->hypot_map`).
4360 * If a type references other types, then those referenced types are checked
4361 * for equivalence recursively.
4362 *
4363 * During DFS traversal, if we find that for current `canon_id` type we
4364 * already have some mapping in hypothetical map, we check for two possible
4365 * situations:
4366 * - `canon_id` is mapped to exactly the same type as `cand_id`. This will
4367 * happen when type graphs have cycles. In this case we assume those two
4368 * types are equivalent.
4369 * - `canon_id` is mapped to different type. This is contradiction in our
4370 * hypothetical mapping, because same graph in canonical graph corresponds
4371 * to two different types in candidate graph, which for equivalent type
4372 * graphs shouldn't happen. This condition terminates equivalence check
4373 * with negative result.
4374 *
4375 * If type graphs traversal exhausts types to check and find no contradiction,
4376 * then type graphs are equivalent.
4377 *
4378 * When checking types for equivalence, there is one special case: FWD types.
4379 * If FWD type resolution is allowed and one of the types (either from canonical
4380 * or candidate graph) is FWD and other is STRUCT/UNION (depending on FWD's kind
4381 * flag) and their names match, hypothetical mapping is updated to point from
4382 * FWD to STRUCT/UNION. If graphs will be determined as equivalent successfully,
4383 * this mapping will be used to record FWD -> STRUCT/UNION mapping permanently.
4384 *
4385 * Technically, this could lead to incorrect FWD to STRUCT/UNION resolution,
4386 * if there are two exactly named (or anonymous) structs/unions that are
4387 * compatible structurally, one of which has FWD field, while other is concrete
4388 * STRUCT/UNION, but according to C sources they are different structs/unions
4389 * that are referencing different types with the same name. This is extremely
4390 * unlikely to happen, but btf_dedup API allows to disable FWD resolution if
4391 * this logic is causing problems.
4392 *
4393 * Doing FWD resolution means that both candidate and/or canonical graphs can
4394 * consists of portions of the graph that come from multiple compilation units.
4395 * This is due to the fact that types within single compilation unit are always
4396 * deduplicated and FWDs are already resolved, if referenced struct/union
4397 * definition is available. So, if we had unresolved FWD and found corresponding
4398 * STRUCT/UNION, they will be from different compilation units. This
4399 * consequently means that when we "link" FWD to corresponding STRUCT/UNION,
4400 * type graph will likely have at least two different BTF types that describe
4401 * same type (e.g., most probably there will be two different BTF types for the
4402 * same 'int' primitive type) and could even have "overlapping" parts of type
4403 * graph that describe same subset of types.
4404 *
4405 * This in turn means that our assumption that each type in canonical graph
4406 * must correspond to exactly one type in candidate graph might not hold
4407 * anymore and will make it harder to detect contradictions using hypothetical
4408 * map. To handle this problem, we allow to follow FWD -> STRUCT/UNION
4409 * resolution only in canonical graph. FWDs in candidate graphs are never
4410 * resolved. To see why it's OK, let's check all possible situations w.r.t. FWDs
4411 * that can occur:
4412 * - Both types in canonical and candidate graphs are FWDs. If they are
4413 * structurally equivalent, then they can either be both resolved to the
4414 * same STRUCT/UNION or not resolved at all. In both cases they are
4415 * equivalent and there is no need to resolve FWD on candidate side.
4416 * - Both types in canonical and candidate graphs are concrete STRUCT/UNION,
4417 * so nothing to resolve as well, algorithm will check equivalence anyway.
4418 * - Type in canonical graph is FWD, while type in candidate is concrete
4419 * STRUCT/UNION. In this case candidate graph comes from single compilation
4420 * unit, so there is exactly one BTF type for each unique C type. After
4421 * resolving FWD into STRUCT/UNION, there might be more than one BTF type
4422 * in canonical graph mapping to single BTF type in candidate graph, but
4423 * because hypothetical mapping maps from canonical to candidate types, it's
4424 * alright, and we still maintain the property of having single `canon_id`
4425 * mapping to single `cand_id` (there could be two different `canon_id`
4426 * mapped to the same `cand_id`, but it's not contradictory).
4427 * - Type in canonical graph is concrete STRUCT/UNION, while type in candidate
4428 * graph is FWD. In this case we are just going to check compatibility of
4429 * STRUCT/UNION and corresponding FWD, and if they are compatible, we'll
4430 * assume that whatever STRUCT/UNION FWD resolves to must be equivalent to
4431 * a concrete STRUCT/UNION from canonical graph. If the rest of type graphs
4432 * turn out equivalent, we'll re-resolve FWD to concrete STRUCT/UNION from
4433 * canonical graph.
4434 */
4435static int btf_dedup_is_equiv(struct btf_dedup *d, __u32 cand_id,
4436 __u32 canon_id)
4437{
4438 struct btf_type *cand_type;
4439 struct btf_type *canon_type;
4440 __u32 hypot_type_id;
4441 __u16 cand_kind;
4442 __u16 canon_kind;
4443 int i, eq;
4444
4445 /* if both resolve to the same canonical, they must be equivalent */
4446 if (resolve_type_id(d, cand_id) == resolve_type_id(d, canon_id))
4447 return 1;
4448
4449 canon_id = resolve_fwd_id(d, canon_id);
4450
4451 hypot_type_id = d->hypot_map[canon_id];
4452 if (hypot_type_id <= BTF_MAX_NR_TYPES) {
4453 if (hypot_type_id == cand_id)
4454 return 1;
4455 /* In some cases compiler will generate different DWARF types
4456 * for *identical* array type definitions and use them for
4457 * different fields within the *same* struct. This breaks type
4458 * equivalence check, which makes an assumption that candidate
4459 * types sub-graph has a consistent and deduped-by-compiler
4460 * types within a single CU. So work around that by explicitly
4461 * allowing identical array types here.
4462 */
4463 if (btf_dedup_identical_arrays(d, hypot_type_id, cand_id))
4464 return 1;
4465 /* It turns out that similar situation can happen with
4466 * struct/union sometimes, sigh... Handle the case where
4467 * structs/unions are exactly the same, down to the referenced
4468 * type IDs. Anything more complicated (e.g., if referenced
4469 * types are different, but equivalent) is *way more*
4470 * complicated and requires a many-to-many equivalence mapping.
4471 */
4472 if (btf_dedup_identical_structs(d, hypot_type_id, cand_id))
4473 return 1;
4474 return 0;
4475 }
4476
4477 if (btf_dedup_hypot_map_add(d, canon_id, cand_id))
4478 return -ENOMEM;
4479
4480 cand_type = btf_type_by_id(d->btf, cand_id);
4481 canon_type = btf_type_by_id(d->btf, canon_id);
4482 cand_kind = btf_kind(cand_type);
4483 canon_kind = btf_kind(canon_type);
4484
4485 if (cand_type->name_off != canon_type->name_off)
4486 return 0;
4487
4488 /* FWD <--> STRUCT/UNION equivalence check, if enabled */
4489 if ((cand_kind == BTF_KIND_FWD || canon_kind == BTF_KIND_FWD)
4490 && cand_kind != canon_kind) {
4491 __u16 real_kind;
4492 __u16 fwd_kind;
4493
4494 if (cand_kind == BTF_KIND_FWD) {
4495 real_kind = canon_kind;
4496 fwd_kind = btf_fwd_kind(cand_type);
4497 } else {
4498 real_kind = cand_kind;
4499 fwd_kind = btf_fwd_kind(canon_type);
4500 /* we'd need to resolve base FWD to STRUCT/UNION */
4501 if (fwd_kind == real_kind && canon_id < d->btf->start_id)
4502 d->hypot_adjust_canon = true;
4503 }
4504 return fwd_kind == real_kind;
4505 }
4506
4507 if (cand_kind != canon_kind)
4508 return 0;
4509
4510 switch (cand_kind) {
4511 case BTF_KIND_INT:
4512 return btf_equal_int_tag(cand_type, canon_type);
4513
4514 case BTF_KIND_ENUM:
4515 case BTF_KIND_ENUM64:
4516 return btf_compat_enum(cand_type, canon_type);
4517
4518 case BTF_KIND_FWD:
4519 case BTF_KIND_FLOAT:
4520 return btf_equal_common(cand_type, canon_type);
4521
4522 case BTF_KIND_CONST:
4523 case BTF_KIND_VOLATILE:
4524 case BTF_KIND_RESTRICT:
4525 case BTF_KIND_PTR:
4526 case BTF_KIND_TYPEDEF:
4527 case BTF_KIND_FUNC:
4528 case BTF_KIND_TYPE_TAG:
4529 if (cand_type->info != canon_type->info)
4530 return 0;
4531 return btf_dedup_is_equiv(d, cand_type->type, canon_type->type);
4532
4533 case BTF_KIND_ARRAY: {
4534 const struct btf_array *cand_arr, *canon_arr;
4535
4536 if (!btf_compat_array(cand_type, canon_type))
4537 return 0;
4538 cand_arr = btf_array(cand_type);
4539 canon_arr = btf_array(canon_type);
4540 eq = btf_dedup_is_equiv(d, cand_arr->index_type, canon_arr->index_type);
4541 if (eq <= 0)
4542 return eq;
4543 return btf_dedup_is_equiv(d, cand_arr->type, canon_arr->type);
4544 }
4545
4546 case BTF_KIND_STRUCT:
4547 case BTF_KIND_UNION: {
4548 const struct btf_member *cand_m, *canon_m;
4549 __u16 vlen;
4550
4551 if (!btf_shallow_equal_struct(cand_type, canon_type))
4552 return 0;
4553 vlen = btf_vlen(cand_type);
4554 cand_m = btf_members(cand_type);
4555 canon_m = btf_members(canon_type);
4556 for (i = 0; i < vlen; i++) {
4557 eq = btf_dedup_is_equiv(d, cand_m->type, canon_m->type);
4558 if (eq <= 0)
4559 return eq;
4560 cand_m++;
4561 canon_m++;
4562 }
4563
4564 return 1;
4565 }
4566
4567 case BTF_KIND_FUNC_PROTO: {
4568 const struct btf_param *cand_p, *canon_p;
4569 __u16 vlen;
4570
4571 if (!btf_compat_fnproto(cand_type, canon_type))
4572 return 0;
4573 eq = btf_dedup_is_equiv(d, cand_type->type, canon_type->type);
4574 if (eq <= 0)
4575 return eq;
4576 vlen = btf_vlen(cand_type);
4577 cand_p = btf_params(cand_type);
4578 canon_p = btf_params(canon_type);
4579 for (i = 0; i < vlen; i++) {
4580 eq = btf_dedup_is_equiv(d, cand_p->type, canon_p->type);
4581 if (eq <= 0)
4582 return eq;
4583 cand_p++;
4584 canon_p++;
4585 }
4586 return 1;
4587 }
4588
4589 default:
4590 return -EINVAL;
4591 }
4592 return 0;
4593}
4594
4595/*
4596 * Use hypothetical mapping, produced by successful type graph equivalence
4597 * check, to augment existing struct/union canonical mapping, where possible.
4598 *
4599 * If BTF_KIND_FWD resolution is allowed, this mapping is also used to record
4600 * FWD -> STRUCT/UNION correspondence as well. FWD resolution is bidirectional:
4601 * it doesn't matter if FWD type was part of canonical graph or candidate one,
4602 * we are recording the mapping anyway. As opposed to carefulness required
4603 * for struct/union correspondence mapping (described below), for FWD resolution
4604 * it's not important, as by the time that FWD type (reference type) will be
4605 * deduplicated all structs/unions will be deduped already anyway.
4606 *
4607 * Recording STRUCT/UNION mapping is purely a performance optimization and is
4608 * not required for correctness. It needs to be done carefully to ensure that
4609 * struct/union from candidate's type graph is not mapped into corresponding
4610 * struct/union from canonical type graph that itself hasn't been resolved into
4611 * canonical representative. The only guarantee we have is that canonical
4612 * struct/union was determined as canonical and that won't change. But any
4613 * types referenced through that struct/union fields could have been not yet
4614 * resolved, so in case like that it's too early to establish any kind of
4615 * correspondence between structs/unions.
4616 *
4617 * No canonical correspondence is derived for primitive types (they are already
4618 * deduplicated completely already anyway) or reference types (they rely on
4619 * stability of struct/union canonical relationship for equivalence checks).
4620 */
4621static void btf_dedup_merge_hypot_map(struct btf_dedup *d)
4622{
4623 __u32 canon_type_id, targ_type_id;
4624 __u16 t_kind, c_kind;
4625 __u32 t_id, c_id;
4626 int i;
4627
4628 for (i = 0; i < d->hypot_cnt; i++) {
4629 canon_type_id = d->hypot_list[i];
4630 targ_type_id = d->hypot_map[canon_type_id];
4631 t_id = resolve_type_id(d, targ_type_id);
4632 c_id = resolve_type_id(d, canon_type_id);
4633 t_kind = btf_kind(btf__type_by_id(d->btf, t_id));
4634 c_kind = btf_kind(btf__type_by_id(d->btf, c_id));
4635 /*
4636 * Resolve FWD into STRUCT/UNION.
4637 * It's ok to resolve FWD into STRUCT/UNION that's not yet
4638 * mapped to canonical representative (as opposed to
4639 * STRUCT/UNION <--> STRUCT/UNION mapping logic below), because
4640 * eventually that struct is going to be mapped and all resolved
4641 * FWDs will automatically resolve to correct canonical
4642 * representative. This will happen before ref type deduping,
4643 * which critically depends on stability of these mapping. This
4644 * stability is not a requirement for STRUCT/UNION equivalence
4645 * checks, though.
4646 */
4647
4648 /* if it's the split BTF case, we still need to point base FWD
4649 * to STRUCT/UNION in a split BTF, because FWDs from split BTF
4650 * will be resolved against base FWD. If we don't point base
4651 * canonical FWD to the resolved STRUCT/UNION, then all the
4652 * FWDs in split BTF won't be correctly resolved to a proper
4653 * STRUCT/UNION.
4654 */
4655 if (t_kind != BTF_KIND_FWD && c_kind == BTF_KIND_FWD)
4656 d->map[c_id] = t_id;
4657
4658 /* if graph equivalence determined that we'd need to adjust
4659 * base canonical types, then we need to only point base FWDs
4660 * to STRUCTs/UNIONs and do no more modifications. For all
4661 * other purposes the type graphs were not equivalent.
4662 */
4663 if (d->hypot_adjust_canon)
4664 continue;
4665
4666 if (t_kind == BTF_KIND_FWD && c_kind != BTF_KIND_FWD)
4667 d->map[t_id] = c_id;
4668
4669 if ((t_kind == BTF_KIND_STRUCT || t_kind == BTF_KIND_UNION) &&
4670 c_kind != BTF_KIND_FWD &&
4671 is_type_mapped(d, c_id) &&
4672 !is_type_mapped(d, t_id)) {
4673 /*
4674 * as a perf optimization, we can map struct/union
4675 * that's part of type graph we just verified for
4676 * equivalence. We can do that for struct/union that has
4677 * canonical representative only, though.
4678 */
4679 d->map[t_id] = c_id;
4680 }
4681 }
4682}
4683
4684/*
4685 * Deduplicate struct/union types.
4686 *
4687 * For each struct/union type its type signature hash is calculated, taking
4688 * into account type's name, size, number, order and names of fields, but
4689 * ignoring type ID's referenced from fields, because they might not be deduped
4690 * completely until after reference types deduplication phase. This type hash
4691 * is used to iterate over all potential canonical types, sharing same hash.
4692 * For each canonical candidate we check whether type graphs that they form
4693 * (through referenced types in fields and so on) are equivalent using algorithm
4694 * implemented in `btf_dedup_is_equiv`. If such equivalence is found and
4695 * BTF_KIND_FWD resolution is allowed, then hypothetical mapping
4696 * (btf_dedup->hypot_map) produced by aforementioned type graph equivalence
4697 * algorithm is used to record FWD -> STRUCT/UNION mapping. It's also used to
4698 * potentially map other structs/unions to their canonical representatives,
4699 * if such relationship hasn't yet been established. This speeds up algorithm
4700 * by eliminating some of the duplicate work.
4701 *
4702 * If no matching canonical representative was found, struct/union is marked
4703 * as canonical for itself and is added into btf_dedup->dedup_table hash map
4704 * for further look ups.
4705 */
4706static int btf_dedup_struct_type(struct btf_dedup *d, __u32 type_id)
4707{
4708 struct btf_type *cand_type, *t;
4709 struct hashmap_entry *hash_entry;
4710 /* if we don't find equivalent type, then we are canonical */
4711 __u32 new_id = type_id;
4712 __u16 kind;
4713 long h;
4714
4715 /* already deduped or is in process of deduping (loop detected) */
4716 if (d->map[type_id] <= BTF_MAX_NR_TYPES)
4717 return 0;
4718
4719 t = btf_type_by_id(d->btf, type_id);
4720 kind = btf_kind(t);
4721
4722 if (kind != BTF_KIND_STRUCT && kind != BTF_KIND_UNION)
4723 return 0;
4724
4725 h = btf_hash_struct(t);
4726 for_each_dedup_cand(d, hash_entry, h) {
4727 __u32 cand_id = hash_entry->value;
4728 int eq;
4729
4730 /*
4731 * Even though btf_dedup_is_equiv() checks for
4732 * btf_shallow_equal_struct() internally when checking two
4733 * structs (unions) for equivalence, we need to guard here
4734 * from picking matching FWD type as a dedup candidate.
4735 * This can happen due to hash collision. In such case just
4736 * relying on btf_dedup_is_equiv() would lead to potentially
4737 * creating a loop (FWD -> STRUCT and STRUCT -> FWD), because
4738 * FWD and compatible STRUCT/UNION are considered equivalent.
4739 */
4740 cand_type = btf_type_by_id(d->btf, cand_id);
4741 if (!btf_shallow_equal_struct(t, cand_type))
4742 continue;
4743
4744 btf_dedup_clear_hypot_map(d);
4745 eq = btf_dedup_is_equiv(d, type_id, cand_id);
4746 if (eq < 0)
4747 return eq;
4748 if (!eq)
4749 continue;
4750 btf_dedup_merge_hypot_map(d);
4751 if (d->hypot_adjust_canon) /* not really equivalent */
4752 continue;
4753 new_id = cand_id;
4754 break;
4755 }
4756
4757 d->map[type_id] = new_id;
4758 if (type_id == new_id && btf_dedup_table_add(d, h, type_id))
4759 return -ENOMEM;
4760
4761 return 0;
4762}
4763
4764static int btf_dedup_struct_types(struct btf_dedup *d)
4765{
4766 int i, err;
4767
4768 for (i = 0; i < d->btf->nr_types; i++) {
4769 err = btf_dedup_struct_type(d, d->btf->start_id + i);
4770 if (err)
4771 return err;
4772 }
4773 return 0;
4774}
4775
4776/*
4777 * Deduplicate reference type.
4778 *
4779 * Once all primitive and struct/union types got deduplicated, we can easily
4780 * deduplicate all other (reference) BTF types. This is done in two steps:
4781 *
4782 * 1. Resolve all referenced type IDs into their canonical type IDs. This
4783 * resolution can be done either immediately for primitive or struct/union types
4784 * (because they were deduped in previous two phases) or recursively for
4785 * reference types. Recursion will always terminate at either primitive or
4786 * struct/union type, at which point we can "unwind" chain of reference types
4787 * one by one. There is no danger of encountering cycles because in C type
4788 * system the only way to form type cycle is through struct/union, so any chain
4789 * of reference types, even those taking part in a type cycle, will inevitably
4790 * reach struct/union at some point.
4791 *
4792 * 2. Once all referenced type IDs are resolved into canonical ones, BTF type
4793 * becomes "stable", in the sense that no further deduplication will cause
4794 * any changes to it. With that, it's now possible to calculate type's signature
4795 * hash (this time taking into account referenced type IDs) and loop over all
4796 * potential canonical representatives. If no match was found, current type
4797 * will become canonical representative of itself and will be added into
4798 * btf_dedup->dedup_table as another possible canonical representative.
4799 */
4800static int btf_dedup_ref_type(struct btf_dedup *d, __u32 type_id)
4801{
4802 struct hashmap_entry *hash_entry;
4803 __u32 new_id = type_id, cand_id;
4804 struct btf_type *t, *cand;
4805 /* if we don't find equivalent type, then we are representative type */
4806 int ref_type_id;
4807 long h;
4808
4809 if (d->map[type_id] == BTF_IN_PROGRESS_ID)
4810 return -ELOOP;
4811 if (d->map[type_id] <= BTF_MAX_NR_TYPES)
4812 return resolve_type_id(d, type_id);
4813
4814 t = btf_type_by_id(d->btf, type_id);
4815 d->map[type_id] = BTF_IN_PROGRESS_ID;
4816
4817 switch (btf_kind(t)) {
4818 case BTF_KIND_CONST:
4819 case BTF_KIND_VOLATILE:
4820 case BTF_KIND_RESTRICT:
4821 case BTF_KIND_PTR:
4822 case BTF_KIND_TYPEDEF:
4823 case BTF_KIND_FUNC:
4824 case BTF_KIND_TYPE_TAG:
4825 ref_type_id = btf_dedup_ref_type(d, t->type);
4826 if (ref_type_id < 0)
4827 return ref_type_id;
4828 t->type = ref_type_id;
4829
4830 h = btf_hash_common(t);
4831 for_each_dedup_cand(d, hash_entry, h) {
4832 cand_id = hash_entry->value;
4833 cand = btf_type_by_id(d->btf, cand_id);
4834 if (btf_equal_common(t, cand)) {
4835 new_id = cand_id;
4836 break;
4837 }
4838 }
4839 break;
4840
4841 case BTF_KIND_DECL_TAG:
4842 ref_type_id = btf_dedup_ref_type(d, t->type);
4843 if (ref_type_id < 0)
4844 return ref_type_id;
4845 t->type = ref_type_id;
4846
4847 h = btf_hash_int_decl_tag(t);
4848 for_each_dedup_cand(d, hash_entry, h) {
4849 cand_id = hash_entry->value;
4850 cand = btf_type_by_id(d->btf, cand_id);
4851 if (btf_equal_int_tag(t, cand)) {
4852 new_id = cand_id;
4853 break;
4854 }
4855 }
4856 break;
4857
4858 case BTF_KIND_ARRAY: {
4859 struct btf_array *info = btf_array(t);
4860
4861 ref_type_id = btf_dedup_ref_type(d, info->type);
4862 if (ref_type_id < 0)
4863 return ref_type_id;
4864 info->type = ref_type_id;
4865
4866 ref_type_id = btf_dedup_ref_type(d, info->index_type);
4867 if (ref_type_id < 0)
4868 return ref_type_id;
4869 info->index_type = ref_type_id;
4870
4871 h = btf_hash_array(t);
4872 for_each_dedup_cand(d, hash_entry, h) {
4873 cand_id = hash_entry->value;
4874 cand = btf_type_by_id(d->btf, cand_id);
4875 if (btf_equal_array(t, cand)) {
4876 new_id = cand_id;
4877 break;
4878 }
4879 }
4880 break;
4881 }
4882
4883 case BTF_KIND_FUNC_PROTO: {
4884 struct btf_param *param;
4885 __u16 vlen;
4886 int i;
4887
4888 ref_type_id = btf_dedup_ref_type(d, t->type);
4889 if (ref_type_id < 0)
4890 return ref_type_id;
4891 t->type = ref_type_id;
4892
4893 vlen = btf_vlen(t);
4894 param = btf_params(t);
4895 for (i = 0; i < vlen; i++) {
4896 ref_type_id = btf_dedup_ref_type(d, param->type);
4897 if (ref_type_id < 0)
4898 return ref_type_id;
4899 param->type = ref_type_id;
4900 param++;
4901 }
4902
4903 h = btf_hash_fnproto(t);
4904 for_each_dedup_cand(d, hash_entry, h) {
4905 cand_id = hash_entry->value;
4906 cand = btf_type_by_id(d->btf, cand_id);
4907 if (btf_equal_fnproto(t, cand)) {
4908 new_id = cand_id;
4909 break;
4910 }
4911 }
4912 break;
4913 }
4914
4915 default:
4916 return -EINVAL;
4917 }
4918
4919 d->map[type_id] = new_id;
4920 if (type_id == new_id && btf_dedup_table_add(d, h, type_id))
4921 return -ENOMEM;
4922
4923 return new_id;
4924}
4925
4926static int btf_dedup_ref_types(struct btf_dedup *d)
4927{
4928 int i, err;
4929
4930 for (i = 0; i < d->btf->nr_types; i++) {
4931 err = btf_dedup_ref_type(d, d->btf->start_id + i);
4932 if (err < 0)
4933 return err;
4934 }
4935 /* we won't need d->dedup_table anymore */
4936 hashmap__free(d->dedup_table);
4937 d->dedup_table = NULL;
4938 return 0;
4939}
4940
4941/*
4942 * Collect a map from type names to type ids for all canonical structs
4943 * and unions. If the same name is shared by several canonical types
4944 * use a special value 0 to indicate this fact.
4945 */
4946static int btf_dedup_fill_unique_names_map(struct btf_dedup *d, struct hashmap *names_map)
4947{
4948 __u32 nr_types = btf__type_cnt(d->btf);
4949 struct btf_type *t;
4950 __u32 type_id;
4951 __u16 kind;
4952 int err;
4953
4954 /*
4955 * Iterate over base and split module ids in order to get all
4956 * available structs in the map.
4957 */
4958 for (type_id = 1; type_id < nr_types; ++type_id) {
4959 t = btf_type_by_id(d->btf, type_id);
4960 kind = btf_kind(t);
4961
4962 if (kind != BTF_KIND_STRUCT && kind != BTF_KIND_UNION)
4963 continue;
4964
4965 /* Skip non-canonical types */
4966 if (type_id != d->map[type_id])
4967 continue;
4968
4969 err = hashmap__add(names_map, t->name_off, type_id);
4970 if (err == -EEXIST)
4971 err = hashmap__set(names_map, t->name_off, 0, NULL, NULL);
4972
4973 if (err)
4974 return err;
4975 }
4976
4977 return 0;
4978}
4979
4980static int btf_dedup_resolve_fwd(struct btf_dedup *d, struct hashmap *names_map, __u32 type_id)
4981{
4982 struct btf_type *t = btf_type_by_id(d->btf, type_id);
4983 enum btf_fwd_kind fwd_kind = btf_kflag(t);
4984 __u16 cand_kind, kind = btf_kind(t);
4985 struct btf_type *cand_t;
4986 uintptr_t cand_id;
4987
4988 if (kind != BTF_KIND_FWD)
4989 return 0;
4990
4991 /* Skip if this FWD already has a mapping */
4992 if (type_id != d->map[type_id])
4993 return 0;
4994
4995 if (!hashmap__find(names_map, t->name_off, &cand_id))
4996 return 0;
4997
4998 /* Zero is a special value indicating that name is not unique */
4999 if (!cand_id)
5000 return 0;
5001
5002 cand_t = btf_type_by_id(d->btf, cand_id);
5003 cand_kind = btf_kind(cand_t);
5004 if ((cand_kind == BTF_KIND_STRUCT && fwd_kind != BTF_FWD_STRUCT) ||
5005 (cand_kind == BTF_KIND_UNION && fwd_kind != BTF_FWD_UNION))
5006 return 0;
5007
5008 d->map[type_id] = cand_id;
5009
5010 return 0;
5011}
5012
5013/*
5014 * Resolve unambiguous forward declarations.
5015 *
5016 * The lion's share of all FWD declarations is resolved during
5017 * `btf_dedup_struct_types` phase when different type graphs are
5018 * compared against each other. However, if in some compilation unit a
5019 * FWD declaration is not a part of a type graph compared against
5020 * another type graph that declaration's canonical type would not be
5021 * changed. Example:
5022 *
5023 * CU #1:
5024 *
5025 * struct foo;
5026 * struct foo *some_global;
5027 *
5028 * CU #2:
5029 *
5030 * struct foo { int u; };
5031 * struct foo *another_global;
5032 *
5033 * After `btf_dedup_struct_types` the BTF looks as follows:
5034 *
5035 * [1] STRUCT 'foo' size=4 vlen=1 ...
5036 * [2] INT 'int' size=4 ...
5037 * [3] PTR '(anon)' type_id=1
5038 * [4] FWD 'foo' fwd_kind=struct
5039 * [5] PTR '(anon)' type_id=4
5040 *
5041 * This pass assumes that such FWD declarations should be mapped to
5042 * structs or unions with identical name in case if the name is not
5043 * ambiguous.
5044 */
5045static int btf_dedup_resolve_fwds(struct btf_dedup *d)
5046{
5047 int i, err;
5048 struct hashmap *names_map;
5049
5050 names_map = hashmap__new(btf_dedup_identity_hash_fn, btf_dedup_equal_fn, NULL);
5051 if (IS_ERR(names_map))
5052 return PTR_ERR(names_map);
5053
5054 err = btf_dedup_fill_unique_names_map(d, names_map);
5055 if (err < 0)
5056 goto exit;
5057
5058 for (i = 0; i < d->btf->nr_types; i++) {
5059 err = btf_dedup_resolve_fwd(d, names_map, d->btf->start_id + i);
5060 if (err < 0)
5061 break;
5062 }
5063
5064exit:
5065 hashmap__free(names_map);
5066 return err;
5067}
5068
5069/*
5070 * Compact types.
5071 *
5072 * After we established for each type its corresponding canonical representative
5073 * type, we now can eliminate types that are not canonical and leave only
5074 * canonical ones layed out sequentially in memory by copying them over
5075 * duplicates. During compaction btf_dedup->hypot_map array is reused to store
5076 * a map from original type ID to a new compacted type ID, which will be used
5077 * during next phase to "fix up" type IDs, referenced from struct/union and
5078 * reference types.
5079 */
5080static int btf_dedup_compact_types(struct btf_dedup *d)
5081{
5082 __u32 *new_offs;
5083 __u32 next_type_id = d->btf->start_id;
5084 const struct btf_type *t;
5085 void *p;
5086 int i, id, len;
5087
5088 /* we are going to reuse hypot_map to store compaction remapping */
5089 d->hypot_map[0] = 0;
5090 /* base BTF types are not renumbered */
5091 for (id = 1; id < d->btf->start_id; id++)
5092 d->hypot_map[id] = id;
5093 for (i = 0, id = d->btf->start_id; i < d->btf->nr_types; i++, id++)
5094 d->hypot_map[id] = BTF_UNPROCESSED_ID;
5095
5096 p = d->btf->types_data;
5097
5098 for (i = 0, id = d->btf->start_id; i < d->btf->nr_types; i++, id++) {
5099 if (d->map[id] != id)
5100 continue;
5101
5102 t = btf__type_by_id(d->btf, id);
5103 len = btf_type_size(t);
5104 if (len < 0)
5105 return len;
5106
5107 memmove(p, t, len);
5108 d->hypot_map[id] = next_type_id;
5109 d->btf->type_offs[next_type_id - d->btf->start_id] = p - d->btf->types_data;
5110 p += len;
5111 next_type_id++;
5112 }
5113
5114 /* shrink struct btf's internal types index and update btf_header */
5115 d->btf->nr_types = next_type_id - d->btf->start_id;
5116 d->btf->type_offs_cap = d->btf->nr_types;
5117 d->btf->hdr->type_len = p - d->btf->types_data;
5118 new_offs = libbpf_reallocarray(d->btf->type_offs, d->btf->type_offs_cap,
5119 sizeof(*new_offs));
5120 if (d->btf->type_offs_cap && !new_offs)
5121 return -ENOMEM;
5122 d->btf->type_offs = new_offs;
5123 d->btf->hdr->str_off = d->btf->hdr->type_len;
5124 d->btf->raw_size = d->btf->hdr->hdr_len + d->btf->hdr->type_len + d->btf->hdr->str_len;
5125 return 0;
5126}
5127
5128/*
5129 * Figure out final (deduplicated and compacted) type ID for provided original
5130 * `type_id` by first resolving it into corresponding canonical type ID and
5131 * then mapping it to a deduplicated type ID, stored in btf_dedup->hypot_map,
5132 * which is populated during compaction phase.
5133 */
5134static int btf_dedup_remap_type_id(__u32 *type_id, void *ctx)
5135{
5136 struct btf_dedup *d = ctx;
5137 __u32 resolved_type_id, new_type_id;
5138
5139 resolved_type_id = resolve_type_id(d, *type_id);
5140 new_type_id = d->hypot_map[resolved_type_id];
5141 if (new_type_id > BTF_MAX_NR_TYPES)
5142 return -EINVAL;
5143
5144 *type_id = new_type_id;
5145 return 0;
5146}
5147
5148/*
5149 * Remap referenced type IDs into deduped type IDs.
5150 *
5151 * After BTF types are deduplicated and compacted, their final type IDs may
5152 * differ from original ones. The map from original to a corresponding
5153 * deduped type ID is stored in btf_dedup->hypot_map and is populated during
5154 * compaction phase. During remapping phase we are rewriting all type IDs
5155 * referenced from any BTF type (e.g., struct fields, func proto args, etc) to
5156 * their final deduped type IDs.
5157 */
5158static int btf_dedup_remap_types(struct btf_dedup *d)
5159{
5160 int i, r;
5161
5162 for (i = 0; i < d->btf->nr_types; i++) {
5163 struct btf_type *t = btf_type_by_id(d->btf, d->btf->start_id + i);
5164 struct btf_field_iter it;
5165 __u32 *type_id;
5166
5167 r = btf_field_iter_init(&it, t, BTF_FIELD_ITER_IDS);
5168 if (r)
5169 return r;
5170
5171 while ((type_id = btf_field_iter_next(&it))) {
5172 __u32 resolved_id, new_id;
5173
5174 resolved_id = resolve_type_id(d, *type_id);
5175 new_id = d->hypot_map[resolved_id];
5176 if (new_id > BTF_MAX_NR_TYPES)
5177 return -EINVAL;
5178
5179 *type_id = new_id;
5180 }
5181 }
5182
5183 if (!d->btf_ext)
5184 return 0;
5185
5186 r = btf_ext_visit_type_ids(d->btf_ext, btf_dedup_remap_type_id, d);
5187 if (r)
5188 return r;
5189
5190 return 0;
5191}
5192
5193/*
5194 * Probe few well-known locations for vmlinux kernel image and try to load BTF
5195 * data out of it to use for target BTF.
5196 */
5197struct btf *btf__load_vmlinux_btf(void)
5198{
5199 const char *sysfs_btf_path = "/sys/kernel/btf/vmlinux";
5200 /* fall back locations, trying to find vmlinux on disk */
5201 const char *locations[] = {
5202 "/boot/vmlinux-%1$s",
5203 "/lib/modules/%1$s/vmlinux-%1$s",
5204 "/lib/modules/%1$s/build/vmlinux",
5205 "/usr/lib/modules/%1$s/kernel/vmlinux",
5206 "/usr/lib/debug/boot/vmlinux-%1$s",
5207 "/usr/lib/debug/boot/vmlinux-%1$s.debug",
5208 "/usr/lib/debug/lib/modules/%1$s/vmlinux",
5209 };
5210 char path[PATH_MAX + 1];
5211 struct utsname buf;
5212 struct btf *btf;
5213 int i, err;
5214
5215 /* is canonical sysfs location accessible? */
5216 if (faccessat(AT_FDCWD, sysfs_btf_path, F_OK, AT_EACCESS) < 0) {
5217 pr_warn("kernel BTF is missing at '%s', was CONFIG_DEBUG_INFO_BTF enabled?\n",
5218 sysfs_btf_path);
5219 } else {
5220 btf = btf__parse(sysfs_btf_path, NULL);
5221 if (!btf) {
5222 err = -errno;
5223 pr_warn("failed to read kernel BTF from '%s': %s\n",
5224 sysfs_btf_path, errstr(err));
5225 return libbpf_err_ptr(err);
5226 }
5227 pr_debug("loaded kernel BTF from '%s'\n", sysfs_btf_path);
5228 return btf;
5229 }
5230
5231 /* try fallback locations */
5232 uname(&buf);
5233 for (i = 0; i < ARRAY_SIZE(locations); i++) {
5234 snprintf(path, PATH_MAX, locations[i], buf.release);
5235
5236 if (faccessat(AT_FDCWD, path, R_OK, AT_EACCESS))
5237 continue;
5238
5239 btf = btf__parse(path, NULL);
5240 err = libbpf_get_error(btf);
5241 pr_debug("loading kernel BTF '%s': %s\n", path, errstr(err));
5242 if (err)
5243 continue;
5244
5245 return btf;
5246 }
5247
5248 pr_warn("failed to find valid kernel BTF\n");
5249 return libbpf_err_ptr(-ESRCH);
5250}
5251
5252struct btf *libbpf_find_kernel_btf(void) __attribute__((alias("btf__load_vmlinux_btf")));
5253
5254struct btf *btf__load_module_btf(const char *module_name, struct btf *vmlinux_btf)
5255{
5256 char path[80];
5257
5258 snprintf(path, sizeof(path), "/sys/kernel/btf/%s", module_name);
5259 return btf__parse_split(path, vmlinux_btf);
5260}
5261
5262int btf_ext_visit_type_ids(struct btf_ext *btf_ext, type_id_visit_fn visit, void *ctx)
5263{
5264 const struct btf_ext_info *seg;
5265 struct btf_ext_info_sec *sec;
5266 int i, err;
5267
5268 seg = &btf_ext->func_info;
5269 for_each_btf_ext_sec(seg, sec) {
5270 struct bpf_func_info_min *rec;
5271
5272 for_each_btf_ext_rec(seg, sec, i, rec) {
5273 err = visit(&rec->type_id, ctx);
5274 if (err < 0)
5275 return err;
5276 }
5277 }
5278
5279 seg = &btf_ext->core_relo_info;
5280 for_each_btf_ext_sec(seg, sec) {
5281 struct bpf_core_relo *rec;
5282
5283 for_each_btf_ext_rec(seg, sec, i, rec) {
5284 err = visit(&rec->type_id, ctx);
5285 if (err < 0)
5286 return err;
5287 }
5288 }
5289
5290 return 0;
5291}
5292
5293int btf_ext_visit_str_offs(struct btf_ext *btf_ext, str_off_visit_fn visit, void *ctx)
5294{
5295 const struct btf_ext_info *seg;
5296 struct btf_ext_info_sec *sec;
5297 int i, err;
5298
5299 seg = &btf_ext->func_info;
5300 for_each_btf_ext_sec(seg, sec) {
5301 err = visit(&sec->sec_name_off, ctx);
5302 if (err)
5303 return err;
5304 }
5305
5306 seg = &btf_ext->line_info;
5307 for_each_btf_ext_sec(seg, sec) {
5308 struct bpf_line_info_min *rec;
5309
5310 err = visit(&sec->sec_name_off, ctx);
5311 if (err)
5312 return err;
5313
5314 for_each_btf_ext_rec(seg, sec, i, rec) {
5315 err = visit(&rec->file_name_off, ctx);
5316 if (err)
5317 return err;
5318 err = visit(&rec->line_off, ctx);
5319 if (err)
5320 return err;
5321 }
5322 }
5323
5324 seg = &btf_ext->core_relo_info;
5325 for_each_btf_ext_sec(seg, sec) {
5326 struct bpf_core_relo *rec;
5327
5328 err = visit(&sec->sec_name_off, ctx);
5329 if (err)
5330 return err;
5331
5332 for_each_btf_ext_rec(seg, sec, i, rec) {
5333 err = visit(&rec->access_str_off, ctx);
5334 if (err)
5335 return err;
5336 }
5337 }
5338
5339 return 0;
5340}
5341
5342struct btf_distill {
5343 struct btf_pipe pipe;
5344 int *id_map;
5345 unsigned int split_start_id;
5346 unsigned int split_start_str;
5347 int diff_id;
5348};
5349
5350static int btf_add_distilled_type_ids(struct btf_distill *dist, __u32 i)
5351{
5352 struct btf_type *split_t = btf_type_by_id(dist->pipe.src, i);
5353 struct btf_field_iter it;
5354 __u32 *id;
5355 int err;
5356
5357 err = btf_field_iter_init(&it, split_t, BTF_FIELD_ITER_IDS);
5358 if (err)
5359 return err;
5360 while ((id = btf_field_iter_next(&it))) {
5361 struct btf_type *base_t;
5362
5363 if (!*id)
5364 continue;
5365 /* split BTF id, not needed */
5366 if (*id >= dist->split_start_id)
5367 continue;
5368 /* already added ? */
5369 if (dist->id_map[*id] > 0)
5370 continue;
5371
5372 /* only a subset of base BTF types should be referenced from
5373 * split BTF; ensure nothing unexpected is referenced.
5374 */
5375 base_t = btf_type_by_id(dist->pipe.src, *id);
5376 switch (btf_kind(base_t)) {
5377 case BTF_KIND_INT:
5378 case BTF_KIND_FLOAT:
5379 case BTF_KIND_FWD:
5380 case BTF_KIND_ARRAY:
5381 case BTF_KIND_STRUCT:
5382 case BTF_KIND_UNION:
5383 case BTF_KIND_TYPEDEF:
5384 case BTF_KIND_ENUM:
5385 case BTF_KIND_ENUM64:
5386 case BTF_KIND_PTR:
5387 case BTF_KIND_CONST:
5388 case BTF_KIND_RESTRICT:
5389 case BTF_KIND_VOLATILE:
5390 case BTF_KIND_FUNC_PROTO:
5391 case BTF_KIND_TYPE_TAG:
5392 dist->id_map[*id] = *id;
5393 break;
5394 default:
5395 pr_warn("unexpected reference to base type[%u] of kind [%u] when creating distilled base BTF.\n",
5396 *id, btf_kind(base_t));
5397 return -EINVAL;
5398 }
5399 /* If a base type is used, ensure types it refers to are
5400 * marked as used also; so for example if we find a PTR to INT
5401 * we need both the PTR and INT.
5402 *
5403 * The only exception is named struct/unions, since distilled
5404 * base BTF composite types have no members.
5405 */
5406 if (btf_is_composite(base_t) && base_t->name_off)
5407 continue;
5408 err = btf_add_distilled_type_ids(dist, *id);
5409 if (err)
5410 return err;
5411 }
5412 return 0;
5413}
5414
5415static int btf_add_distilled_types(struct btf_distill *dist)
5416{
5417 bool adding_to_base = dist->pipe.dst->start_id == 1;
5418 int id = btf__type_cnt(dist->pipe.dst);
5419 struct btf_type *t;
5420 int i, err = 0;
5421
5422
5423 /* Add types for each of the required references to either distilled
5424 * base or split BTF, depending on type characteristics.
5425 */
5426 for (i = 1; i < dist->split_start_id; i++) {
5427 const char *name;
5428 int kind;
5429
5430 if (!dist->id_map[i])
5431 continue;
5432 t = btf_type_by_id(dist->pipe.src, i);
5433 kind = btf_kind(t);
5434 name = btf__name_by_offset(dist->pipe.src, t->name_off);
5435
5436 switch (kind) {
5437 case BTF_KIND_INT:
5438 case BTF_KIND_FLOAT:
5439 case BTF_KIND_FWD:
5440 /* Named int, float, fwd are added to base. */
5441 if (!adding_to_base)
5442 continue;
5443 err = btf_add_type(&dist->pipe, t);
5444 break;
5445 case BTF_KIND_STRUCT:
5446 case BTF_KIND_UNION:
5447 /* Named struct/union are added to base as 0-vlen
5448 * struct/union of same size. Anonymous struct/unions
5449 * are added to split BTF as-is.
5450 */
5451 if (adding_to_base) {
5452 if (!t->name_off)
5453 continue;
5454 err = btf_add_composite(dist->pipe.dst, kind, name, t->size);
5455 } else {
5456 if (t->name_off)
5457 continue;
5458 err = btf_add_type(&dist->pipe, t);
5459 }
5460 break;
5461 case BTF_KIND_ENUM:
5462 case BTF_KIND_ENUM64:
5463 /* Named enum[64]s are added to base as a sized
5464 * enum; relocation will match with appropriately-named
5465 * and sized enum or enum64.
5466 *
5467 * Anonymous enums are added to split BTF as-is.
5468 */
5469 if (adding_to_base) {
5470 if (!t->name_off)
5471 continue;
5472 err = btf__add_enum(dist->pipe.dst, name, t->size);
5473 } else {
5474 if (t->name_off)
5475 continue;
5476 err = btf_add_type(&dist->pipe, t);
5477 }
5478 break;
5479 case BTF_KIND_ARRAY:
5480 case BTF_KIND_TYPEDEF:
5481 case BTF_KIND_PTR:
5482 case BTF_KIND_CONST:
5483 case BTF_KIND_RESTRICT:
5484 case BTF_KIND_VOLATILE:
5485 case BTF_KIND_FUNC_PROTO:
5486 case BTF_KIND_TYPE_TAG:
5487 /* All other types are added to split BTF. */
5488 if (adding_to_base)
5489 continue;
5490 err = btf_add_type(&dist->pipe, t);
5491 break;
5492 default:
5493 pr_warn("unexpected kind when adding base type '%s'[%u] of kind [%u] to distilled base BTF.\n",
5494 name, i, kind);
5495 return -EINVAL;
5496
5497 }
5498 if (err < 0)
5499 break;
5500 dist->id_map[i] = id++;
5501 }
5502 return err;
5503}
5504
5505/* Split BTF ids without a mapping will be shifted downwards since distilled
5506 * base BTF is smaller than the original base BTF. For those that have a
5507 * mapping (either to base or updated split BTF), update the id based on
5508 * that mapping.
5509 */
5510static int btf_update_distilled_type_ids(struct btf_distill *dist, __u32 i)
5511{
5512 struct btf_type *t = btf_type_by_id(dist->pipe.dst, i);
5513 struct btf_field_iter it;
5514 __u32 *id;
5515 int err;
5516
5517 err = btf_field_iter_init(&it, t, BTF_FIELD_ITER_IDS);
5518 if (err)
5519 return err;
5520 while ((id = btf_field_iter_next(&it))) {
5521 if (dist->id_map[*id])
5522 *id = dist->id_map[*id];
5523 else if (*id >= dist->split_start_id)
5524 *id -= dist->diff_id;
5525 }
5526 return 0;
5527}
5528
5529/* Create updated split BTF with distilled base BTF; distilled base BTF
5530 * consists of BTF information required to clarify the types that split
5531 * BTF refers to, omitting unneeded details. Specifically it will contain
5532 * base types and memberless definitions of named structs, unions and enumerated
5533 * types. Associated reference types like pointers, arrays and anonymous
5534 * structs, unions and enumerated types will be added to split BTF.
5535 * Size is recorded for named struct/unions to help guide matching to the
5536 * target base BTF during later relocation.
5537 *
5538 * The only case where structs, unions or enumerated types are fully represented
5539 * is when they are anonymous; in such cases, the anonymous type is added to
5540 * split BTF in full.
5541 *
5542 * We return newly-created split BTF where the split BTF refers to a newly-created
5543 * distilled base BTF. Both must be freed separately by the caller.
5544 */
5545int btf__distill_base(const struct btf *src_btf, struct btf **new_base_btf,
5546 struct btf **new_split_btf)
5547{
5548 struct btf *new_base = NULL, *new_split = NULL;
5549 const struct btf *old_base;
5550 unsigned int n = btf__type_cnt(src_btf);
5551 struct btf_distill dist = {};
5552 struct btf_type *t;
5553 int i, err = 0;
5554
5555 /* src BTF must be split BTF. */
5556 old_base = btf__base_btf(src_btf);
5557 if (!new_base_btf || !new_split_btf || !old_base)
5558 return libbpf_err(-EINVAL);
5559
5560 new_base = btf__new_empty();
5561 if (!new_base)
5562 return libbpf_err(-ENOMEM);
5563
5564 btf__set_endianness(new_base, btf__endianness(src_btf));
5565
5566 dist.id_map = calloc(n, sizeof(*dist.id_map));
5567 if (!dist.id_map) {
5568 err = -ENOMEM;
5569 goto done;
5570 }
5571 dist.pipe.src = src_btf;
5572 dist.pipe.dst = new_base;
5573 dist.pipe.str_off_map = hashmap__new(btf_dedup_identity_hash_fn, btf_dedup_equal_fn, NULL);
5574 if (IS_ERR(dist.pipe.str_off_map)) {
5575 err = -ENOMEM;
5576 goto done;
5577 }
5578 dist.split_start_id = btf__type_cnt(old_base);
5579 dist.split_start_str = old_base->hdr->str_len;
5580
5581 /* Pass over src split BTF; generate the list of base BTF type ids it
5582 * references; these will constitute our distilled BTF set to be
5583 * distributed over base and split BTF as appropriate.
5584 */
5585 for (i = src_btf->start_id; i < n; i++) {
5586 err = btf_add_distilled_type_ids(&dist, i);
5587 if (err < 0)
5588 goto done;
5589 }
5590 /* Next add types for each of the required references to base BTF and split BTF
5591 * in turn.
5592 */
5593 err = btf_add_distilled_types(&dist);
5594 if (err < 0)
5595 goto done;
5596
5597 /* Create new split BTF with distilled base BTF as its base; the final
5598 * state is split BTF with distilled base BTF that represents enough
5599 * about its base references to allow it to be relocated with the base
5600 * BTF available.
5601 */
5602 new_split = btf__new_empty_split(new_base);
5603 if (!new_split) {
5604 err = -errno;
5605 goto done;
5606 }
5607 dist.pipe.dst = new_split;
5608 /* First add all split types */
5609 for (i = src_btf->start_id; i < n; i++) {
5610 t = btf_type_by_id(src_btf, i);
5611 err = btf_add_type(&dist.pipe, t);
5612 if (err < 0)
5613 goto done;
5614 }
5615 /* Now add distilled types to split BTF that are not added to base. */
5616 err = btf_add_distilled_types(&dist);
5617 if (err < 0)
5618 goto done;
5619
5620 /* All split BTF ids will be shifted downwards since there are less base
5621 * BTF ids in distilled base BTF.
5622 */
5623 dist.diff_id = dist.split_start_id - btf__type_cnt(new_base);
5624
5625 n = btf__type_cnt(new_split);
5626 /* Now update base/split BTF ids. */
5627 for (i = 1; i < n; i++) {
5628 err = btf_update_distilled_type_ids(&dist, i);
5629 if (err < 0)
5630 break;
5631 }
5632done:
5633 free(dist.id_map);
5634 hashmap__free(dist.pipe.str_off_map);
5635 if (err) {
5636 btf__free(new_split);
5637 btf__free(new_base);
5638 return libbpf_err(err);
5639 }
5640 *new_base_btf = new_base;
5641 *new_split_btf = new_split;
5642
5643 return 0;
5644}
5645
5646const struct btf_header *btf_header(const struct btf *btf)
5647{
5648 return btf->hdr;
5649}
5650
5651void btf_set_base_btf(struct btf *btf, const struct btf *base_btf)
5652{
5653 btf->base_btf = (struct btf *)base_btf;
5654 btf->start_id = btf__type_cnt(base_btf);
5655 btf->start_str_off = base_btf->hdr->str_len;
5656}
5657
5658int btf__relocate(struct btf *btf, const struct btf *base_btf)
5659{
5660 int err = btf_relocate(btf, base_btf, NULL);
5661
5662 if (!err)
5663 btf->owns_base = false;
5664 return libbpf_err(err);
5665}