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
1082struct btf *btf__new_split(const void *data, __u32 size, struct btf *base_btf)
1083{
1084 return libbpf_ptr(btf_new(data, size, base_btf));
1085}
1086
1087static struct btf *btf_parse_elf(const char *path, struct btf *base_btf,
1088 struct btf_ext **btf_ext)
1089{
1090 Elf_Data *btf_data = NULL, *btf_ext_data = NULL;
1091 int err = 0, fd = -1, idx = 0;
1092 struct btf *btf = NULL;
1093 Elf_Scn *scn = NULL;
1094 Elf *elf = NULL;
1095 GElf_Ehdr ehdr;
1096 size_t shstrndx;
1097
1098 if (elf_version(EV_CURRENT) == EV_NONE) {
1099 pr_warn("failed to init libelf for %s\n", path);
1100 return ERR_PTR(-LIBBPF_ERRNO__LIBELF);
1101 }
1102
1103 fd = open(path, O_RDONLY | O_CLOEXEC);
1104 if (fd < 0) {
1105 err = -errno;
1106 pr_warn("failed to open %s: %s\n", path, strerror(errno));
1107 return ERR_PTR(err);
1108 }
1109
1110 err = -LIBBPF_ERRNO__FORMAT;
1111
1112 elf = elf_begin(fd, ELF_C_READ, NULL);
1113 if (!elf) {
1114 pr_warn("failed to open %s as ELF file\n", path);
1115 goto done;
1116 }
1117 if (!gelf_getehdr(elf, &ehdr)) {
1118 pr_warn("failed to get EHDR from %s\n", path);
1119 goto done;
1120 }
1121
1122 if (elf_getshdrstrndx(elf, &shstrndx)) {
1123 pr_warn("failed to get section names section index for %s\n",
1124 path);
1125 goto done;
1126 }
1127
1128 if (!elf_rawdata(elf_getscn(elf, shstrndx), NULL)) {
1129 pr_warn("failed to get e_shstrndx from %s\n", path);
1130 goto done;
1131 }
1132
1133 while ((scn = elf_nextscn(elf, scn)) != NULL) {
1134 GElf_Shdr sh;
1135 char *name;
1136
1137 idx++;
1138 if (gelf_getshdr(scn, &sh) != &sh) {
1139 pr_warn("failed to get section(%d) header from %s\n",
1140 idx, path);
1141 goto done;
1142 }
1143 name = elf_strptr(elf, shstrndx, sh.sh_name);
1144 if (!name) {
1145 pr_warn("failed to get section(%d) name from %s\n",
1146 idx, path);
1147 goto done;
1148 }
1149 if (strcmp(name, BTF_ELF_SEC) == 0) {
1150 btf_data = elf_getdata(scn, 0);
1151 if (!btf_data) {
1152 pr_warn("failed to get section(%d, %s) data from %s\n",
1153 idx, name, path);
1154 goto done;
1155 }
1156 continue;
1157 } else if (btf_ext && strcmp(name, BTF_EXT_ELF_SEC) == 0) {
1158 btf_ext_data = elf_getdata(scn, 0);
1159 if (!btf_ext_data) {
1160 pr_warn("failed to get section(%d, %s) data from %s\n",
1161 idx, name, path);
1162 goto done;
1163 }
1164 continue;
1165 }
1166 }
1167
1168 if (!btf_data) {
1169 pr_warn("failed to find '%s' ELF section in %s\n", BTF_ELF_SEC, path);
1170 err = -ENODATA;
1171 goto done;
1172 }
1173 btf = btf_new(btf_data->d_buf, btf_data->d_size, base_btf);
1174 err = libbpf_get_error(btf);
1175 if (err)
1176 goto done;
1177
1178 switch (gelf_getclass(elf)) {
1179 case ELFCLASS32:
1180 btf__set_pointer_size(btf, 4);
1181 break;
1182 case ELFCLASS64:
1183 btf__set_pointer_size(btf, 8);
1184 break;
1185 default:
1186 pr_warn("failed to get ELF class (bitness) for %s\n", path);
1187 break;
1188 }
1189
1190 if (btf_ext && btf_ext_data) {
1191 *btf_ext = btf_ext__new(btf_ext_data->d_buf, btf_ext_data->d_size);
1192 err = libbpf_get_error(*btf_ext);
1193 if (err)
1194 goto done;
1195 } else if (btf_ext) {
1196 *btf_ext = NULL;
1197 }
1198done:
1199 if (elf)
1200 elf_end(elf);
1201 close(fd);
1202
1203 if (!err)
1204 return btf;
1205
1206 if (btf_ext)
1207 btf_ext__free(*btf_ext);
1208 btf__free(btf);
1209
1210 return ERR_PTR(err);
1211}
1212
1213struct btf *btf__parse_elf(const char *path, struct btf_ext **btf_ext)
1214{
1215 return libbpf_ptr(btf_parse_elf(path, NULL, btf_ext));
1216}
1217
1218struct btf *btf__parse_elf_split(const char *path, struct btf *base_btf)
1219{
1220 return libbpf_ptr(btf_parse_elf(path, base_btf, NULL));
1221}
1222
1223static struct btf *btf_parse_raw(const char *path, struct btf *base_btf)
1224{
1225 struct btf *btf = NULL;
1226 void *data = NULL;
1227 FILE *f = NULL;
1228 __u16 magic;
1229 int err = 0;
1230 long sz;
1231
1232 f = fopen(path, "rbe");
1233 if (!f) {
1234 err = -errno;
1235 goto err_out;
1236 }
1237
1238 /* check BTF magic */
1239 if (fread(&magic, 1, sizeof(magic), f) < sizeof(magic)) {
1240 err = -EIO;
1241 goto err_out;
1242 }
1243 if (magic != BTF_MAGIC && magic != bswap_16(BTF_MAGIC)) {
1244 /* definitely not a raw BTF */
1245 err = -EPROTO;
1246 goto err_out;
1247 }
1248
1249 /* get file size */
1250 if (fseek(f, 0, SEEK_END)) {
1251 err = -errno;
1252 goto err_out;
1253 }
1254 sz = ftell(f);
1255 if (sz < 0) {
1256 err = -errno;
1257 goto err_out;
1258 }
1259 /* rewind to the start */
1260 if (fseek(f, 0, SEEK_SET)) {
1261 err = -errno;
1262 goto err_out;
1263 }
1264
1265 /* pre-alloc memory and read all of BTF data */
1266 data = malloc(sz);
1267 if (!data) {
1268 err = -ENOMEM;
1269 goto err_out;
1270 }
1271 if (fread(data, 1, sz, f) < sz) {
1272 err = -EIO;
1273 goto err_out;
1274 }
1275
1276 /* finally parse BTF data */
1277 btf = btf_new(data, sz, base_btf);
1278
1279err_out:
1280 free(data);
1281 if (f)
1282 fclose(f);
1283 return err ? ERR_PTR(err) : btf;
1284}
1285
1286struct btf *btf__parse_raw(const char *path)
1287{
1288 return libbpf_ptr(btf_parse_raw(path, NULL));
1289}
1290
1291struct btf *btf__parse_raw_split(const char *path, struct btf *base_btf)
1292{
1293 return libbpf_ptr(btf_parse_raw(path, base_btf));
1294}
1295
1296static struct btf *btf_parse(const char *path, struct btf *base_btf, struct btf_ext **btf_ext)
1297{
1298 struct btf *btf;
1299 int err;
1300
1301 if (btf_ext)
1302 *btf_ext = NULL;
1303
1304 btf = btf_parse_raw(path, base_btf);
1305 err = libbpf_get_error(btf);
1306 if (!err)
1307 return btf;
1308 if (err != -EPROTO)
1309 return ERR_PTR(err);
1310 return btf_parse_elf(path, base_btf, btf_ext);
1311}
1312
1313struct btf *btf__parse(const char *path, struct btf_ext **btf_ext)
1314{
1315 return libbpf_ptr(btf_parse(path, NULL, btf_ext));
1316}
1317
1318struct btf *btf__parse_split(const char *path, struct btf *base_btf)
1319{
1320 return libbpf_ptr(btf_parse(path, base_btf, NULL));
1321}
1322
1323static void *btf_get_raw_data(const struct btf *btf, __u32 *size, bool swap_endian);
1324
1325int btf_load_into_kernel(struct btf *btf,
1326 char *log_buf, size_t log_sz, __u32 log_level,
1327 int token_fd)
1328{
1329 LIBBPF_OPTS(bpf_btf_load_opts, opts);
1330 __u32 buf_sz = 0, raw_size;
1331 char *buf = NULL, *tmp;
1332 void *raw_data;
1333 int err = 0;
1334
1335 if (btf->fd >= 0)
1336 return libbpf_err(-EEXIST);
1337 if (log_sz && !log_buf)
1338 return libbpf_err(-EINVAL);
1339
1340 /* cache native raw data representation */
1341 raw_data = btf_get_raw_data(btf, &raw_size, false);
1342 if (!raw_data) {
1343 err = -ENOMEM;
1344 goto done;
1345 }
1346 btf->raw_size = raw_size;
1347 btf->raw_data = raw_data;
1348
1349retry_load:
1350 /* if log_level is 0, we won't provide log_buf/log_size to the kernel,
1351 * initially. Only if BTF loading fails, we bump log_level to 1 and
1352 * retry, using either auto-allocated or custom log_buf. This way
1353 * non-NULL custom log_buf provides a buffer just in case, but hopes
1354 * for successful load and no need for log_buf.
1355 */
1356 if (log_level) {
1357 /* if caller didn't provide custom log_buf, we'll keep
1358 * allocating our own progressively bigger buffers for BTF
1359 * verification log
1360 */
1361 if (!log_buf) {
1362 buf_sz = max((__u32)BPF_LOG_BUF_SIZE, buf_sz * 2);
1363 tmp = realloc(buf, buf_sz);
1364 if (!tmp) {
1365 err = -ENOMEM;
1366 goto done;
1367 }
1368 buf = tmp;
1369 buf[0] = '\0';
1370 }
1371
1372 opts.log_buf = log_buf ? log_buf : buf;
1373 opts.log_size = log_buf ? log_sz : buf_sz;
1374 opts.log_level = log_level;
1375 }
1376
1377 opts.token_fd = token_fd;
1378 if (token_fd)
1379 opts.btf_flags |= BPF_F_TOKEN_FD;
1380
1381 btf->fd = bpf_btf_load(raw_data, raw_size, &opts);
1382 if (btf->fd < 0) {
1383 /* time to turn on verbose mode and try again */
1384 if (log_level == 0) {
1385 log_level = 1;
1386 goto retry_load;
1387 }
1388 /* only retry if caller didn't provide custom log_buf, but
1389 * make sure we can never overflow buf_sz
1390 */
1391 if (!log_buf && errno == ENOSPC && buf_sz <= UINT_MAX / 2)
1392 goto retry_load;
1393
1394 err = -errno;
1395 pr_warn("BTF loading error: %d\n", err);
1396 /* don't print out contents of custom log_buf */
1397 if (!log_buf && buf[0])
1398 pr_warn("-- BEGIN BTF LOAD LOG ---\n%s\n-- END BTF LOAD LOG --\n", buf);
1399 }
1400
1401done:
1402 free(buf);
1403 return libbpf_err(err);
1404}
1405
1406int btf__load_into_kernel(struct btf *btf)
1407{
1408 return btf_load_into_kernel(btf, NULL, 0, 0, 0);
1409}
1410
1411int btf__fd(const struct btf *btf)
1412{
1413 return btf->fd;
1414}
1415
1416void btf__set_fd(struct btf *btf, int fd)
1417{
1418 btf->fd = fd;
1419}
1420
1421static const void *btf_strs_data(const struct btf *btf)
1422{
1423 return btf->strs_data ? btf->strs_data : strset__data(btf->strs_set);
1424}
1425
1426static void *btf_get_raw_data(const struct btf *btf, __u32 *size, bool swap_endian)
1427{
1428 struct btf_header *hdr = btf->hdr;
1429 struct btf_type *t;
1430 void *data, *p;
1431 __u32 data_sz;
1432 int i;
1433
1434 data = swap_endian ? btf->raw_data_swapped : btf->raw_data;
1435 if (data) {
1436 *size = btf->raw_size;
1437 return data;
1438 }
1439
1440 data_sz = hdr->hdr_len + hdr->type_len + hdr->str_len;
1441 data = calloc(1, data_sz);
1442 if (!data)
1443 return NULL;
1444 p = data;
1445
1446 memcpy(p, hdr, hdr->hdr_len);
1447 if (swap_endian)
1448 btf_bswap_hdr(p);
1449 p += hdr->hdr_len;
1450
1451 memcpy(p, btf->types_data, hdr->type_len);
1452 if (swap_endian) {
1453 for (i = 0; i < btf->nr_types; i++) {
1454 t = p + btf->type_offs[i];
1455 /* btf_bswap_type_rest() relies on native t->info, so
1456 * we swap base type info after we swapped all the
1457 * additional information
1458 */
1459 if (btf_bswap_type_rest(t))
1460 goto err_out;
1461 btf_bswap_type_base(t);
1462 }
1463 }
1464 p += hdr->type_len;
1465
1466 memcpy(p, btf_strs_data(btf), hdr->str_len);
1467 p += hdr->str_len;
1468
1469 *size = data_sz;
1470 return data;
1471err_out:
1472 free(data);
1473 return NULL;
1474}
1475
1476const void *btf__raw_data(const struct btf *btf_ro, __u32 *size)
1477{
1478 struct btf *btf = (struct btf *)btf_ro;
1479 __u32 data_sz;
1480 void *data;
1481
1482 data = btf_get_raw_data(btf, &data_sz, btf->swapped_endian);
1483 if (!data)
1484 return errno = ENOMEM, NULL;
1485
1486 btf->raw_size = data_sz;
1487 if (btf->swapped_endian)
1488 btf->raw_data_swapped = data;
1489 else
1490 btf->raw_data = data;
1491 *size = data_sz;
1492 return data;
1493}
1494
1495__attribute__((alias("btf__raw_data")))
1496const void *btf__get_raw_data(const struct btf *btf, __u32 *size);
1497
1498const char *btf__str_by_offset(const struct btf *btf, __u32 offset)
1499{
1500 if (offset < btf->start_str_off)
1501 return btf__str_by_offset(btf->base_btf, offset);
1502 else if (offset - btf->start_str_off < btf->hdr->str_len)
1503 return btf_strs_data(btf) + (offset - btf->start_str_off);
1504 else
1505 return errno = EINVAL, NULL;
1506}
1507
1508const char *btf__name_by_offset(const struct btf *btf, __u32 offset)
1509{
1510 return btf__str_by_offset(btf, offset);
1511}
1512
1513struct btf *btf_get_from_fd(int btf_fd, struct btf *base_btf)
1514{
1515 struct bpf_btf_info btf_info;
1516 __u32 len = sizeof(btf_info);
1517 __u32 last_size;
1518 struct btf *btf;
1519 void *ptr;
1520 int err;
1521
1522 /* we won't know btf_size until we call bpf_btf_get_info_by_fd(). so
1523 * let's start with a sane default - 4KiB here - and resize it only if
1524 * bpf_btf_get_info_by_fd() needs a bigger buffer.
1525 */
1526 last_size = 4096;
1527 ptr = malloc(last_size);
1528 if (!ptr)
1529 return ERR_PTR(-ENOMEM);
1530
1531 memset(&btf_info, 0, sizeof(btf_info));
1532 btf_info.btf = ptr_to_u64(ptr);
1533 btf_info.btf_size = last_size;
1534 err = bpf_btf_get_info_by_fd(btf_fd, &btf_info, &len);
1535
1536 if (!err && btf_info.btf_size > last_size) {
1537 void *temp_ptr;
1538
1539 last_size = btf_info.btf_size;
1540 temp_ptr = realloc(ptr, last_size);
1541 if (!temp_ptr) {
1542 btf = ERR_PTR(-ENOMEM);
1543 goto exit_free;
1544 }
1545 ptr = temp_ptr;
1546
1547 len = sizeof(btf_info);
1548 memset(&btf_info, 0, sizeof(btf_info));
1549 btf_info.btf = ptr_to_u64(ptr);
1550 btf_info.btf_size = last_size;
1551
1552 err = bpf_btf_get_info_by_fd(btf_fd, &btf_info, &len);
1553 }
1554
1555 if (err || btf_info.btf_size > last_size) {
1556 btf = err ? ERR_PTR(-errno) : ERR_PTR(-E2BIG);
1557 goto exit_free;
1558 }
1559
1560 btf = btf_new(ptr, btf_info.btf_size, base_btf);
1561
1562exit_free:
1563 free(ptr);
1564 return btf;
1565}
1566
1567struct btf *btf__load_from_kernel_by_id_split(__u32 id, struct btf *base_btf)
1568{
1569 struct btf *btf;
1570 int btf_fd;
1571
1572 btf_fd = bpf_btf_get_fd_by_id(id);
1573 if (btf_fd < 0)
1574 return libbpf_err_ptr(-errno);
1575
1576 btf = btf_get_from_fd(btf_fd, base_btf);
1577 close(btf_fd);
1578
1579 return libbpf_ptr(btf);
1580}
1581
1582struct btf *btf__load_from_kernel_by_id(__u32 id)
1583{
1584 return btf__load_from_kernel_by_id_split(id, NULL);
1585}
1586
1587static void btf_invalidate_raw_data(struct btf *btf)
1588{
1589 if (btf->raw_data) {
1590 free(btf->raw_data);
1591 btf->raw_data = NULL;
1592 }
1593 if (btf->raw_data_swapped) {
1594 free(btf->raw_data_swapped);
1595 btf->raw_data_swapped = NULL;
1596 }
1597}
1598
1599/* Ensure BTF is ready to be modified (by splitting into a three memory
1600 * regions for header, types, and strings). Also invalidate cached
1601 * raw_data, if any.
1602 */
1603static int btf_ensure_modifiable(struct btf *btf)
1604{
1605 void *hdr, *types;
1606 struct strset *set = NULL;
1607 int err = -ENOMEM;
1608
1609 if (btf_is_modifiable(btf)) {
1610 /* any BTF modification invalidates raw_data */
1611 btf_invalidate_raw_data(btf);
1612 return 0;
1613 }
1614
1615 /* split raw data into three memory regions */
1616 hdr = malloc(btf->hdr->hdr_len);
1617 types = malloc(btf->hdr->type_len);
1618 if (!hdr || !types)
1619 goto err_out;
1620
1621 memcpy(hdr, btf->hdr, btf->hdr->hdr_len);
1622 memcpy(types, btf->types_data, btf->hdr->type_len);
1623
1624 /* build lookup index for all strings */
1625 set = strset__new(BTF_MAX_STR_OFFSET, btf->strs_data, btf->hdr->str_len);
1626 if (IS_ERR(set)) {
1627 err = PTR_ERR(set);
1628 goto err_out;
1629 }
1630
1631 /* only when everything was successful, update internal state */
1632 btf->hdr = hdr;
1633 btf->types_data = types;
1634 btf->types_data_cap = btf->hdr->type_len;
1635 btf->strs_data = NULL;
1636 btf->strs_set = set;
1637 /* if BTF was created from scratch, all strings are guaranteed to be
1638 * unique and deduplicated
1639 */
1640 if (btf->hdr->str_len == 0)
1641 btf->strs_deduped = true;
1642 if (!btf->base_btf && btf->hdr->str_len == 1)
1643 btf->strs_deduped = true;
1644
1645 /* invalidate raw_data representation */
1646 btf_invalidate_raw_data(btf);
1647
1648 return 0;
1649
1650err_out:
1651 strset__free(set);
1652 free(hdr);
1653 free(types);
1654 return err;
1655}
1656
1657/* Find an offset in BTF string section that corresponds to a given string *s*.
1658 * Returns:
1659 * - >0 offset into string section, if string is found;
1660 * - -ENOENT, if string is not in the string section;
1661 * - <0, on any other error.
1662 */
1663int btf__find_str(struct btf *btf, const char *s)
1664{
1665 int off;
1666
1667 if (btf->base_btf) {
1668 off = btf__find_str(btf->base_btf, s);
1669 if (off != -ENOENT)
1670 return off;
1671 }
1672
1673 /* BTF needs to be in a modifiable state to build string lookup index */
1674 if (btf_ensure_modifiable(btf))
1675 return libbpf_err(-ENOMEM);
1676
1677 off = strset__find_str(btf->strs_set, s);
1678 if (off < 0)
1679 return libbpf_err(off);
1680
1681 return btf->start_str_off + off;
1682}
1683
1684/* Add a string s to the BTF string section.
1685 * Returns:
1686 * - > 0 offset into string section, on success;
1687 * - < 0, on error.
1688 */
1689int btf__add_str(struct btf *btf, const char *s)
1690{
1691 int off;
1692
1693 if (btf->base_btf) {
1694 off = btf__find_str(btf->base_btf, s);
1695 if (off != -ENOENT)
1696 return off;
1697 }
1698
1699 if (btf_ensure_modifiable(btf))
1700 return libbpf_err(-ENOMEM);
1701
1702 off = strset__add_str(btf->strs_set, s);
1703 if (off < 0)
1704 return libbpf_err(off);
1705
1706 btf->hdr->str_len = strset__data_size(btf->strs_set);
1707
1708 return btf->start_str_off + off;
1709}
1710
1711static void *btf_add_type_mem(struct btf *btf, size_t add_sz)
1712{
1713 return libbpf_add_mem(&btf->types_data, &btf->types_data_cap, 1,
1714 btf->hdr->type_len, UINT_MAX, add_sz);
1715}
1716
1717static void btf_type_inc_vlen(struct btf_type *t)
1718{
1719 t->info = btf_type_info(btf_kind(t), btf_vlen(t) + 1, btf_kflag(t));
1720}
1721
1722static int btf_commit_type(struct btf *btf, int data_sz)
1723{
1724 int err;
1725
1726 err = btf_add_type_idx_entry(btf, btf->hdr->type_len);
1727 if (err)
1728 return libbpf_err(err);
1729
1730 btf->hdr->type_len += data_sz;
1731 btf->hdr->str_off += data_sz;
1732 btf->nr_types++;
1733 return btf->start_id + btf->nr_types - 1;
1734}
1735
1736struct btf_pipe {
1737 const struct btf *src;
1738 struct btf *dst;
1739 struct hashmap *str_off_map; /* map string offsets from src to dst */
1740};
1741
1742static int btf_rewrite_str(__u32 *str_off, void *ctx)
1743{
1744 struct btf_pipe *p = ctx;
1745 long mapped_off;
1746 int off, err;
1747
1748 if (!*str_off) /* nothing to do for empty strings */
1749 return 0;
1750
1751 if (p->str_off_map &&
1752 hashmap__find(p->str_off_map, *str_off, &mapped_off)) {
1753 *str_off = mapped_off;
1754 return 0;
1755 }
1756
1757 off = btf__add_str(p->dst, btf__str_by_offset(p->src, *str_off));
1758 if (off < 0)
1759 return off;
1760
1761 /* Remember string mapping from src to dst. It avoids
1762 * performing expensive string comparisons.
1763 */
1764 if (p->str_off_map) {
1765 err = hashmap__append(p->str_off_map, *str_off, off);
1766 if (err)
1767 return err;
1768 }
1769
1770 *str_off = off;
1771 return 0;
1772}
1773
1774int btf__add_type(struct btf *btf, const struct btf *src_btf, const struct btf_type *src_type)
1775{
1776 struct btf_pipe p = { .src = src_btf, .dst = btf };
1777 struct btf_type *t;
1778 int sz, err;
1779
1780 sz = btf_type_size(src_type);
1781 if (sz < 0)
1782 return libbpf_err(sz);
1783
1784 /* deconstruct BTF, if necessary, and invalidate raw_data */
1785 if (btf_ensure_modifiable(btf))
1786 return libbpf_err(-ENOMEM);
1787
1788 t = btf_add_type_mem(btf, sz);
1789 if (!t)
1790 return libbpf_err(-ENOMEM);
1791
1792 memcpy(t, src_type, sz);
1793
1794 err = btf_type_visit_str_offs(t, btf_rewrite_str, &p);
1795 if (err)
1796 return libbpf_err(err);
1797
1798 return btf_commit_type(btf, sz);
1799}
1800
1801static int btf_rewrite_type_ids(__u32 *type_id, void *ctx)
1802{
1803 struct btf *btf = ctx;
1804
1805 if (!*type_id) /* nothing to do for VOID references */
1806 return 0;
1807
1808 /* we haven't updated btf's type count yet, so
1809 * btf->start_id + btf->nr_types - 1 is the type ID offset we should
1810 * add to all newly added BTF types
1811 */
1812 *type_id += btf->start_id + btf->nr_types - 1;
1813 return 0;
1814}
1815
1816static size_t btf_dedup_identity_hash_fn(long key, void *ctx);
1817static bool btf_dedup_equal_fn(long k1, long k2, void *ctx);
1818
1819int btf__add_btf(struct btf *btf, const struct btf *src_btf)
1820{
1821 struct btf_pipe p = { .src = src_btf, .dst = btf };
1822 int data_sz, sz, cnt, i, err, old_strs_len;
1823 __u32 *off;
1824 void *t;
1825
1826 /* appending split BTF isn't supported yet */
1827 if (src_btf->base_btf)
1828 return libbpf_err(-ENOTSUP);
1829
1830 /* deconstruct BTF, if necessary, and invalidate raw_data */
1831 if (btf_ensure_modifiable(btf))
1832 return libbpf_err(-ENOMEM);
1833
1834 /* remember original strings section size if we have to roll back
1835 * partial strings section changes
1836 */
1837 old_strs_len = btf->hdr->str_len;
1838
1839 data_sz = src_btf->hdr->type_len;
1840 cnt = btf__type_cnt(src_btf) - 1;
1841
1842 /* pre-allocate enough memory for new types */
1843 t = btf_add_type_mem(btf, data_sz);
1844 if (!t)
1845 return libbpf_err(-ENOMEM);
1846
1847 /* pre-allocate enough memory for type offset index for new types */
1848 off = btf_add_type_offs_mem(btf, cnt);
1849 if (!off)
1850 return libbpf_err(-ENOMEM);
1851
1852 /* Map the string offsets from src_btf to the offsets from btf to improve performance */
1853 p.str_off_map = hashmap__new(btf_dedup_identity_hash_fn, btf_dedup_equal_fn, NULL);
1854 if (IS_ERR(p.str_off_map))
1855 return libbpf_err(-ENOMEM);
1856
1857 /* bulk copy types data for all types from src_btf */
1858 memcpy(t, src_btf->types_data, data_sz);
1859
1860 for (i = 0; i < cnt; i++) {
1861 sz = btf_type_size(t);
1862 if (sz < 0) {
1863 /* unlikely, has to be corrupted src_btf */
1864 err = sz;
1865 goto err_out;
1866 }
1867
1868 /* fill out type ID to type offset mapping for lookups by type ID */
1869 *off = t - btf->types_data;
1870
1871 /* add, dedup, and remap strings referenced by this BTF type */
1872 err = btf_type_visit_str_offs(t, btf_rewrite_str, &p);
1873 if (err)
1874 goto err_out;
1875
1876 /* remap all type IDs referenced from this BTF type */
1877 err = btf_type_visit_type_ids(t, btf_rewrite_type_ids, btf);
1878 if (err)
1879 goto err_out;
1880
1881 /* go to next type data and type offset index entry */
1882 t += sz;
1883 off++;
1884 }
1885
1886 /* Up until now any of the copied type data was effectively invisible,
1887 * so if we exited early before this point due to error, BTF would be
1888 * effectively unmodified. There would be extra internal memory
1889 * pre-allocated, but it would not be available for querying. But now
1890 * that we've copied and rewritten all the data successfully, we can
1891 * update type count and various internal offsets and sizes to
1892 * "commit" the changes and made them visible to the outside world.
1893 */
1894 btf->hdr->type_len += data_sz;
1895 btf->hdr->str_off += data_sz;
1896 btf->nr_types += cnt;
1897
1898 hashmap__free(p.str_off_map);
1899
1900 /* return type ID of the first added BTF type */
1901 return btf->start_id + btf->nr_types - cnt;
1902err_out:
1903 /* zero out preallocated memory as if it was just allocated with
1904 * libbpf_add_mem()
1905 */
1906 memset(btf->types_data + btf->hdr->type_len, 0, data_sz);
1907 memset(btf->strs_data + old_strs_len, 0, btf->hdr->str_len - old_strs_len);
1908
1909 /* and now restore original strings section size; types data size
1910 * wasn't modified, so doesn't need restoring, see big comment above
1911 */
1912 btf->hdr->str_len = old_strs_len;
1913
1914 hashmap__free(p.str_off_map);
1915
1916 return libbpf_err(err);
1917}
1918
1919/*
1920 * Append new BTF_KIND_INT type with:
1921 * - *name* - non-empty, non-NULL type name;
1922 * - *sz* - power-of-2 (1, 2, 4, ..) size of the type, in bytes;
1923 * - encoding is a combination of BTF_INT_SIGNED, BTF_INT_CHAR, BTF_INT_BOOL.
1924 * Returns:
1925 * - >0, type ID of newly added BTF type;
1926 * - <0, on error.
1927 */
1928int btf__add_int(struct btf *btf, const char *name, size_t byte_sz, int encoding)
1929{
1930 struct btf_type *t;
1931 int sz, name_off;
1932
1933 /* non-empty name */
1934 if (!name || !name[0])
1935 return libbpf_err(-EINVAL);
1936 /* byte_sz must be power of 2 */
1937 if (!byte_sz || (byte_sz & (byte_sz - 1)) || byte_sz > 16)
1938 return libbpf_err(-EINVAL);
1939 if (encoding & ~(BTF_INT_SIGNED | BTF_INT_CHAR | BTF_INT_BOOL))
1940 return libbpf_err(-EINVAL);
1941
1942 /* deconstruct BTF, if necessary, and invalidate raw_data */
1943 if (btf_ensure_modifiable(btf))
1944 return libbpf_err(-ENOMEM);
1945
1946 sz = sizeof(struct btf_type) + sizeof(int);
1947 t = btf_add_type_mem(btf, sz);
1948 if (!t)
1949 return libbpf_err(-ENOMEM);
1950
1951 /* if something goes wrong later, we might end up with an extra string,
1952 * but that shouldn't be a problem, because BTF can't be constructed
1953 * completely anyway and will most probably be just discarded
1954 */
1955 name_off = btf__add_str(btf, name);
1956 if (name_off < 0)
1957 return name_off;
1958
1959 t->name_off = name_off;
1960 t->info = btf_type_info(BTF_KIND_INT, 0, 0);
1961 t->size = byte_sz;
1962 /* set INT info, we don't allow setting legacy bit offset/size */
1963 *(__u32 *)(t + 1) = (encoding << 24) | (byte_sz * 8);
1964
1965 return btf_commit_type(btf, sz);
1966}
1967
1968/*
1969 * Append new BTF_KIND_FLOAT type with:
1970 * - *name* - non-empty, non-NULL type name;
1971 * - *sz* - size of the type, in bytes;
1972 * Returns:
1973 * - >0, type ID of newly added BTF type;
1974 * - <0, on error.
1975 */
1976int btf__add_float(struct btf *btf, const char *name, size_t byte_sz)
1977{
1978 struct btf_type *t;
1979 int sz, name_off;
1980
1981 /* non-empty name */
1982 if (!name || !name[0])
1983 return libbpf_err(-EINVAL);
1984
1985 /* byte_sz must be one of the explicitly allowed values */
1986 if (byte_sz != 2 && byte_sz != 4 && byte_sz != 8 && byte_sz != 12 &&
1987 byte_sz != 16)
1988 return libbpf_err(-EINVAL);
1989
1990 if (btf_ensure_modifiable(btf))
1991 return libbpf_err(-ENOMEM);
1992
1993 sz = sizeof(struct btf_type);
1994 t = btf_add_type_mem(btf, sz);
1995 if (!t)
1996 return libbpf_err(-ENOMEM);
1997
1998 name_off = btf__add_str(btf, name);
1999 if (name_off < 0)
2000 return name_off;
2001
2002 t->name_off = name_off;
2003 t->info = btf_type_info(BTF_KIND_FLOAT, 0, 0);
2004 t->size = byte_sz;
2005
2006 return btf_commit_type(btf, sz);
2007}
2008
2009/* it's completely legal to append BTF types with type IDs pointing forward to
2010 * types that haven't been appended yet, so we only make sure that id looks
2011 * sane, we can't guarantee that ID will always be valid
2012 */
2013static int validate_type_id(int id)
2014{
2015 if (id < 0 || id > BTF_MAX_NR_TYPES)
2016 return -EINVAL;
2017 return 0;
2018}
2019
2020/* generic append function for PTR, TYPEDEF, CONST/VOLATILE/RESTRICT */
2021static int btf_add_ref_kind(struct btf *btf, int kind, const char *name, int ref_type_id)
2022{
2023 struct btf_type *t;
2024 int sz, name_off = 0;
2025
2026 if (validate_type_id(ref_type_id))
2027 return libbpf_err(-EINVAL);
2028
2029 if (btf_ensure_modifiable(btf))
2030 return libbpf_err(-ENOMEM);
2031
2032 sz = sizeof(struct btf_type);
2033 t = btf_add_type_mem(btf, sz);
2034 if (!t)
2035 return libbpf_err(-ENOMEM);
2036
2037 if (name && name[0]) {
2038 name_off = btf__add_str(btf, name);
2039 if (name_off < 0)
2040 return name_off;
2041 }
2042
2043 t->name_off = name_off;
2044 t->info = btf_type_info(kind, 0, 0);
2045 t->type = ref_type_id;
2046
2047 return btf_commit_type(btf, sz);
2048}
2049
2050/*
2051 * Append new BTF_KIND_PTR type with:
2052 * - *ref_type_id* - referenced type ID, it might not exist yet;
2053 * Returns:
2054 * - >0, type ID of newly added BTF type;
2055 * - <0, on error.
2056 */
2057int btf__add_ptr(struct btf *btf, int ref_type_id)
2058{
2059 return btf_add_ref_kind(btf, BTF_KIND_PTR, NULL, ref_type_id);
2060}
2061
2062/*
2063 * Append new BTF_KIND_ARRAY type with:
2064 * - *index_type_id* - type ID of the type describing array index;
2065 * - *elem_type_id* - type ID of the type describing array element;
2066 * - *nr_elems* - the size of the array;
2067 * Returns:
2068 * - >0, type ID of newly added BTF type;
2069 * - <0, on error.
2070 */
2071int btf__add_array(struct btf *btf, int index_type_id, int elem_type_id, __u32 nr_elems)
2072{
2073 struct btf_type *t;
2074 struct btf_array *a;
2075 int sz;
2076
2077 if (validate_type_id(index_type_id) || validate_type_id(elem_type_id))
2078 return libbpf_err(-EINVAL);
2079
2080 if (btf_ensure_modifiable(btf))
2081 return libbpf_err(-ENOMEM);
2082
2083 sz = sizeof(struct btf_type) + sizeof(struct btf_array);
2084 t = btf_add_type_mem(btf, sz);
2085 if (!t)
2086 return libbpf_err(-ENOMEM);
2087
2088 t->name_off = 0;
2089 t->info = btf_type_info(BTF_KIND_ARRAY, 0, 0);
2090 t->size = 0;
2091
2092 a = btf_array(t);
2093 a->type = elem_type_id;
2094 a->index_type = index_type_id;
2095 a->nelems = nr_elems;
2096
2097 return btf_commit_type(btf, sz);
2098}
2099
2100/* generic STRUCT/UNION append function */
2101static int btf_add_composite(struct btf *btf, int kind, const char *name, __u32 bytes_sz)
2102{
2103 struct btf_type *t;
2104 int sz, name_off = 0;
2105
2106 if (btf_ensure_modifiable(btf))
2107 return libbpf_err(-ENOMEM);
2108
2109 sz = sizeof(struct btf_type);
2110 t = btf_add_type_mem(btf, sz);
2111 if (!t)
2112 return libbpf_err(-ENOMEM);
2113
2114 if (name && name[0]) {
2115 name_off = btf__add_str(btf, name);
2116 if (name_off < 0)
2117 return name_off;
2118 }
2119
2120 /* start out with vlen=0 and no kflag; this will be adjusted when
2121 * adding each member
2122 */
2123 t->name_off = name_off;
2124 t->info = btf_type_info(kind, 0, 0);
2125 t->size = bytes_sz;
2126
2127 return btf_commit_type(btf, sz);
2128}
2129
2130/*
2131 * Append new BTF_KIND_STRUCT type with:
2132 * - *name* - name of the struct, can be NULL or empty for anonymous structs;
2133 * - *byte_sz* - size of the struct, in bytes;
2134 *
2135 * Struct initially has no fields in it. Fields can be added by
2136 * btf__add_field() right after btf__add_struct() succeeds.
2137 *
2138 * Returns:
2139 * - >0, type ID of newly added BTF type;
2140 * - <0, on error.
2141 */
2142int btf__add_struct(struct btf *btf, const char *name, __u32 byte_sz)
2143{
2144 return btf_add_composite(btf, BTF_KIND_STRUCT, name, byte_sz);
2145}
2146
2147/*
2148 * Append new BTF_KIND_UNION type with:
2149 * - *name* - name of the union, can be NULL or empty for anonymous union;
2150 * - *byte_sz* - size of the union, in bytes;
2151 *
2152 * Union initially has no fields in it. Fields can be added by
2153 * btf__add_field() right after btf__add_union() succeeds. All fields
2154 * should have *bit_offset* of 0.
2155 *
2156 * Returns:
2157 * - >0, type ID of newly added BTF type;
2158 * - <0, on error.
2159 */
2160int btf__add_union(struct btf *btf, const char *name, __u32 byte_sz)
2161{
2162 return btf_add_composite(btf, BTF_KIND_UNION, name, byte_sz);
2163}
2164
2165static struct btf_type *btf_last_type(struct btf *btf)
2166{
2167 return btf_type_by_id(btf, btf__type_cnt(btf) - 1);
2168}
2169
2170/*
2171 * Append new field for the current STRUCT/UNION type with:
2172 * - *name* - name of the field, can be NULL or empty for anonymous field;
2173 * - *type_id* - type ID for the type describing field type;
2174 * - *bit_offset* - bit offset of the start of the field within struct/union;
2175 * - *bit_size* - bit size of a bitfield, 0 for non-bitfield fields;
2176 * Returns:
2177 * - 0, on success;
2178 * - <0, on error.
2179 */
2180int btf__add_field(struct btf *btf, const char *name, int type_id,
2181 __u32 bit_offset, __u32 bit_size)
2182{
2183 struct btf_type *t;
2184 struct btf_member *m;
2185 bool is_bitfield;
2186 int sz, name_off = 0;
2187
2188 /* last type should be union/struct */
2189 if (btf->nr_types == 0)
2190 return libbpf_err(-EINVAL);
2191 t = btf_last_type(btf);
2192 if (!btf_is_composite(t))
2193 return libbpf_err(-EINVAL);
2194
2195 if (validate_type_id(type_id))
2196 return libbpf_err(-EINVAL);
2197 /* best-effort bit field offset/size enforcement */
2198 is_bitfield = bit_size || (bit_offset % 8 != 0);
2199 if (is_bitfield && (bit_size == 0 || bit_size > 255 || bit_offset > 0xffffff))
2200 return libbpf_err(-EINVAL);
2201
2202 /* only offset 0 is allowed for unions */
2203 if (btf_is_union(t) && bit_offset)
2204 return libbpf_err(-EINVAL);
2205
2206 /* decompose and invalidate raw data */
2207 if (btf_ensure_modifiable(btf))
2208 return libbpf_err(-ENOMEM);
2209
2210 sz = sizeof(struct btf_member);
2211 m = btf_add_type_mem(btf, sz);
2212 if (!m)
2213 return libbpf_err(-ENOMEM);
2214
2215 if (name && name[0]) {
2216 name_off = btf__add_str(btf, name);
2217 if (name_off < 0)
2218 return name_off;
2219 }
2220
2221 m->name_off = name_off;
2222 m->type = type_id;
2223 m->offset = bit_offset | (bit_size << 24);
2224
2225 /* btf_add_type_mem can invalidate t pointer */
2226 t = btf_last_type(btf);
2227 /* update parent type's vlen and kflag */
2228 t->info = btf_type_info(btf_kind(t), btf_vlen(t) + 1, is_bitfield || btf_kflag(t));
2229
2230 btf->hdr->type_len += sz;
2231 btf->hdr->str_off += sz;
2232 return 0;
2233}
2234
2235static int btf_add_enum_common(struct btf *btf, const char *name, __u32 byte_sz,
2236 bool is_signed, __u8 kind)
2237{
2238 struct btf_type *t;
2239 int sz, name_off = 0;
2240
2241 /* byte_sz must be power of 2 */
2242 if (!byte_sz || (byte_sz & (byte_sz - 1)) || byte_sz > 8)
2243 return libbpf_err(-EINVAL);
2244
2245 if (btf_ensure_modifiable(btf))
2246 return libbpf_err(-ENOMEM);
2247
2248 sz = sizeof(struct btf_type);
2249 t = btf_add_type_mem(btf, sz);
2250 if (!t)
2251 return libbpf_err(-ENOMEM);
2252
2253 if (name && name[0]) {
2254 name_off = btf__add_str(btf, name);
2255 if (name_off < 0)
2256 return name_off;
2257 }
2258
2259 /* start out with vlen=0; it will be adjusted when adding enum values */
2260 t->name_off = name_off;
2261 t->info = btf_type_info(kind, 0, is_signed);
2262 t->size = byte_sz;
2263
2264 return btf_commit_type(btf, sz);
2265}
2266
2267/*
2268 * Append new BTF_KIND_ENUM type with:
2269 * - *name* - name of the enum, can be NULL or empty for anonymous enums;
2270 * - *byte_sz* - size of the enum, in bytes.
2271 *
2272 * Enum initially has no enum values in it (and corresponds to enum forward
2273 * declaration). Enumerator values can be added by btf__add_enum_value()
2274 * immediately after btf__add_enum() succeeds.
2275 *
2276 * Returns:
2277 * - >0, type ID of newly added BTF type;
2278 * - <0, on error.
2279 */
2280int btf__add_enum(struct btf *btf, const char *name, __u32 byte_sz)
2281{
2282 /*
2283 * set the signedness to be unsigned, it will change to signed
2284 * if any later enumerator is negative.
2285 */
2286 return btf_add_enum_common(btf, name, byte_sz, false, BTF_KIND_ENUM);
2287}
2288
2289/*
2290 * Append new enum value for the current ENUM type with:
2291 * - *name* - name of the enumerator value, can't be NULL or empty;
2292 * - *value* - integer value corresponding to enum value *name*;
2293 * Returns:
2294 * - 0, on success;
2295 * - <0, on error.
2296 */
2297int btf__add_enum_value(struct btf *btf, const char *name, __s64 value)
2298{
2299 struct btf_type *t;
2300 struct btf_enum *v;
2301 int sz, name_off;
2302
2303 /* last type should be BTF_KIND_ENUM */
2304 if (btf->nr_types == 0)
2305 return libbpf_err(-EINVAL);
2306 t = btf_last_type(btf);
2307 if (!btf_is_enum(t))
2308 return libbpf_err(-EINVAL);
2309
2310 /* non-empty name */
2311 if (!name || !name[0])
2312 return libbpf_err(-EINVAL);
2313 if (value < INT_MIN || value > UINT_MAX)
2314 return libbpf_err(-E2BIG);
2315
2316 /* decompose and invalidate raw data */
2317 if (btf_ensure_modifiable(btf))
2318 return libbpf_err(-ENOMEM);
2319
2320 sz = sizeof(struct btf_enum);
2321 v = btf_add_type_mem(btf, sz);
2322 if (!v)
2323 return libbpf_err(-ENOMEM);
2324
2325 name_off = btf__add_str(btf, name);
2326 if (name_off < 0)
2327 return name_off;
2328
2329 v->name_off = name_off;
2330 v->val = value;
2331
2332 /* update parent type's vlen */
2333 t = btf_last_type(btf);
2334 btf_type_inc_vlen(t);
2335
2336 /* if negative value, set signedness to signed */
2337 if (value < 0)
2338 t->info = btf_type_info(btf_kind(t), btf_vlen(t), true);
2339
2340 btf->hdr->type_len += sz;
2341 btf->hdr->str_off += sz;
2342 return 0;
2343}
2344
2345/*
2346 * Append new BTF_KIND_ENUM64 type with:
2347 * - *name* - name of the enum, can be NULL or empty for anonymous enums;
2348 * - *byte_sz* - size of the enum, in bytes.
2349 * - *is_signed* - whether the enum values are signed or not;
2350 *
2351 * Enum initially has no enum values in it (and corresponds to enum forward
2352 * declaration). Enumerator values can be added by btf__add_enum64_value()
2353 * immediately after btf__add_enum64() succeeds.
2354 *
2355 * Returns:
2356 * - >0, type ID of newly added BTF type;
2357 * - <0, on error.
2358 */
2359int btf__add_enum64(struct btf *btf, const char *name, __u32 byte_sz,
2360 bool is_signed)
2361{
2362 return btf_add_enum_common(btf, name, byte_sz, is_signed,
2363 BTF_KIND_ENUM64);
2364}
2365
2366/*
2367 * Append new enum value for the current ENUM64 type with:
2368 * - *name* - name of the enumerator value, can't be NULL or empty;
2369 * - *value* - integer value corresponding to enum value *name*;
2370 * Returns:
2371 * - 0, on success;
2372 * - <0, on error.
2373 */
2374int btf__add_enum64_value(struct btf *btf, const char *name, __u64 value)
2375{
2376 struct btf_enum64 *v;
2377 struct btf_type *t;
2378 int sz, name_off;
2379
2380 /* last type should be BTF_KIND_ENUM64 */
2381 if (btf->nr_types == 0)
2382 return libbpf_err(-EINVAL);
2383 t = btf_last_type(btf);
2384 if (!btf_is_enum64(t))
2385 return libbpf_err(-EINVAL);
2386
2387 /* non-empty name */
2388 if (!name || !name[0])
2389 return libbpf_err(-EINVAL);
2390
2391 /* decompose and invalidate raw data */
2392 if (btf_ensure_modifiable(btf))
2393 return libbpf_err(-ENOMEM);
2394
2395 sz = sizeof(struct btf_enum64);
2396 v = btf_add_type_mem(btf, sz);
2397 if (!v)
2398 return libbpf_err(-ENOMEM);
2399
2400 name_off = btf__add_str(btf, name);
2401 if (name_off < 0)
2402 return name_off;
2403
2404 v->name_off = name_off;
2405 v->val_lo32 = (__u32)value;
2406 v->val_hi32 = value >> 32;
2407
2408 /* update parent type's vlen */
2409 t = btf_last_type(btf);
2410 btf_type_inc_vlen(t);
2411
2412 btf->hdr->type_len += sz;
2413 btf->hdr->str_off += sz;
2414 return 0;
2415}
2416
2417/*
2418 * Append new BTF_KIND_FWD type with:
2419 * - *name*, non-empty/non-NULL name;
2420 * - *fwd_kind*, kind of forward declaration, one of BTF_FWD_STRUCT,
2421 * BTF_FWD_UNION, or BTF_FWD_ENUM;
2422 * Returns:
2423 * - >0, type ID of newly added BTF type;
2424 * - <0, on error.
2425 */
2426int btf__add_fwd(struct btf *btf, const char *name, enum btf_fwd_kind fwd_kind)
2427{
2428 if (!name || !name[0])
2429 return libbpf_err(-EINVAL);
2430
2431 switch (fwd_kind) {
2432 case BTF_FWD_STRUCT:
2433 case BTF_FWD_UNION: {
2434 struct btf_type *t;
2435 int id;
2436
2437 id = btf_add_ref_kind(btf, BTF_KIND_FWD, name, 0);
2438 if (id <= 0)
2439 return id;
2440 t = btf_type_by_id(btf, id);
2441 t->info = btf_type_info(BTF_KIND_FWD, 0, fwd_kind == BTF_FWD_UNION);
2442 return id;
2443 }
2444 case BTF_FWD_ENUM:
2445 /* enum forward in BTF currently is just an enum with no enum
2446 * values; we also assume a standard 4-byte size for it
2447 */
2448 return btf__add_enum(btf, name, sizeof(int));
2449 default:
2450 return libbpf_err(-EINVAL);
2451 }
2452}
2453
2454/*
2455 * Append new BTF_KING_TYPEDEF type with:
2456 * - *name*, non-empty/non-NULL name;
2457 * - *ref_type_id* - referenced type ID, it might not exist yet;
2458 * Returns:
2459 * - >0, type ID of newly added BTF type;
2460 * - <0, on error.
2461 */
2462int btf__add_typedef(struct btf *btf, const char *name, int ref_type_id)
2463{
2464 if (!name || !name[0])
2465 return libbpf_err(-EINVAL);
2466
2467 return btf_add_ref_kind(btf, BTF_KIND_TYPEDEF, name, ref_type_id);
2468}
2469
2470/*
2471 * Append new BTF_KIND_VOLATILE type with:
2472 * - *ref_type_id* - referenced type ID, it might not exist yet;
2473 * Returns:
2474 * - >0, type ID of newly added BTF type;
2475 * - <0, on error.
2476 */
2477int btf__add_volatile(struct btf *btf, int ref_type_id)
2478{
2479 return btf_add_ref_kind(btf, BTF_KIND_VOLATILE, NULL, ref_type_id);
2480}
2481
2482/*
2483 * Append new BTF_KIND_CONST type with:
2484 * - *ref_type_id* - referenced type ID, it might not exist yet;
2485 * Returns:
2486 * - >0, type ID of newly added BTF type;
2487 * - <0, on error.
2488 */
2489int btf__add_const(struct btf *btf, int ref_type_id)
2490{
2491 return btf_add_ref_kind(btf, BTF_KIND_CONST, NULL, ref_type_id);
2492}
2493
2494/*
2495 * Append new BTF_KIND_RESTRICT type with:
2496 * - *ref_type_id* - referenced type ID, it might not exist yet;
2497 * Returns:
2498 * - >0, type ID of newly added BTF type;
2499 * - <0, on error.
2500 */
2501int btf__add_restrict(struct btf *btf, int ref_type_id)
2502{
2503 return btf_add_ref_kind(btf, BTF_KIND_RESTRICT, NULL, ref_type_id);
2504}
2505
2506/*
2507 * Append new BTF_KIND_TYPE_TAG type with:
2508 * - *value*, non-empty/non-NULL tag value;
2509 * - *ref_type_id* - referenced type ID, it might not exist yet;
2510 * Returns:
2511 * - >0, type ID of newly added BTF type;
2512 * - <0, on error.
2513 */
2514int btf__add_type_tag(struct btf *btf, const char *value, int ref_type_id)
2515{
2516 if (!value || !value[0])
2517 return libbpf_err(-EINVAL);
2518
2519 return btf_add_ref_kind(btf, BTF_KIND_TYPE_TAG, value, ref_type_id);
2520}
2521
2522/*
2523 * Append new BTF_KIND_FUNC type with:
2524 * - *name*, non-empty/non-NULL name;
2525 * - *proto_type_id* - FUNC_PROTO's type ID, it might not exist yet;
2526 * Returns:
2527 * - >0, type ID of newly added BTF type;
2528 * - <0, on error.
2529 */
2530int btf__add_func(struct btf *btf, const char *name,
2531 enum btf_func_linkage linkage, int proto_type_id)
2532{
2533 int id;
2534
2535 if (!name || !name[0])
2536 return libbpf_err(-EINVAL);
2537 if (linkage != BTF_FUNC_STATIC && linkage != BTF_FUNC_GLOBAL &&
2538 linkage != BTF_FUNC_EXTERN)
2539 return libbpf_err(-EINVAL);
2540
2541 id = btf_add_ref_kind(btf, BTF_KIND_FUNC, name, proto_type_id);
2542 if (id > 0) {
2543 struct btf_type *t = btf_type_by_id(btf, id);
2544
2545 t->info = btf_type_info(BTF_KIND_FUNC, linkage, 0);
2546 }
2547 return libbpf_err(id);
2548}
2549
2550/*
2551 * Append new BTF_KIND_FUNC_PROTO with:
2552 * - *ret_type_id* - type ID for return result of a function.
2553 *
2554 * Function prototype initially has no arguments, but they can be added by
2555 * btf__add_func_param() one by one, immediately after
2556 * btf__add_func_proto() succeeded.
2557 *
2558 * Returns:
2559 * - >0, type ID of newly added BTF type;
2560 * - <0, on error.
2561 */
2562int btf__add_func_proto(struct btf *btf, int ret_type_id)
2563{
2564 struct btf_type *t;
2565 int sz;
2566
2567 if (validate_type_id(ret_type_id))
2568 return libbpf_err(-EINVAL);
2569
2570 if (btf_ensure_modifiable(btf))
2571 return libbpf_err(-ENOMEM);
2572
2573 sz = sizeof(struct btf_type);
2574 t = btf_add_type_mem(btf, sz);
2575 if (!t)
2576 return libbpf_err(-ENOMEM);
2577
2578 /* start out with vlen=0; this will be adjusted when adding enum
2579 * values, if necessary
2580 */
2581 t->name_off = 0;
2582 t->info = btf_type_info(BTF_KIND_FUNC_PROTO, 0, 0);
2583 t->type = ret_type_id;
2584
2585 return btf_commit_type(btf, sz);
2586}
2587
2588/*
2589 * Append new function parameter for current FUNC_PROTO type with:
2590 * - *name* - parameter name, can be NULL or empty;
2591 * - *type_id* - type ID describing the type of the parameter.
2592 * Returns:
2593 * - 0, on success;
2594 * - <0, on error.
2595 */
2596int btf__add_func_param(struct btf *btf, const char *name, int type_id)
2597{
2598 struct btf_type *t;
2599 struct btf_param *p;
2600 int sz, name_off = 0;
2601
2602 if (validate_type_id(type_id))
2603 return libbpf_err(-EINVAL);
2604
2605 /* last type should be BTF_KIND_FUNC_PROTO */
2606 if (btf->nr_types == 0)
2607 return libbpf_err(-EINVAL);
2608 t = btf_last_type(btf);
2609 if (!btf_is_func_proto(t))
2610 return libbpf_err(-EINVAL);
2611
2612 /* decompose and invalidate raw data */
2613 if (btf_ensure_modifiable(btf))
2614 return libbpf_err(-ENOMEM);
2615
2616 sz = sizeof(struct btf_param);
2617 p = btf_add_type_mem(btf, sz);
2618 if (!p)
2619 return libbpf_err(-ENOMEM);
2620
2621 if (name && name[0]) {
2622 name_off = btf__add_str(btf, name);
2623 if (name_off < 0)
2624 return name_off;
2625 }
2626
2627 p->name_off = name_off;
2628 p->type = type_id;
2629
2630 /* update parent type's vlen */
2631 t = btf_last_type(btf);
2632 btf_type_inc_vlen(t);
2633
2634 btf->hdr->type_len += sz;
2635 btf->hdr->str_off += sz;
2636 return 0;
2637}
2638
2639/*
2640 * Append new BTF_KIND_VAR type with:
2641 * - *name* - non-empty/non-NULL name;
2642 * - *linkage* - variable linkage, one of BTF_VAR_STATIC,
2643 * BTF_VAR_GLOBAL_ALLOCATED, or BTF_VAR_GLOBAL_EXTERN;
2644 * - *type_id* - type ID of the type describing the type of the variable.
2645 * Returns:
2646 * - >0, type ID of newly added BTF type;
2647 * - <0, on error.
2648 */
2649int btf__add_var(struct btf *btf, const char *name, int linkage, int type_id)
2650{
2651 struct btf_type *t;
2652 struct btf_var *v;
2653 int sz, name_off;
2654
2655 /* non-empty name */
2656 if (!name || !name[0])
2657 return libbpf_err(-EINVAL);
2658 if (linkage != BTF_VAR_STATIC && linkage != BTF_VAR_GLOBAL_ALLOCATED &&
2659 linkage != BTF_VAR_GLOBAL_EXTERN)
2660 return libbpf_err(-EINVAL);
2661 if (validate_type_id(type_id))
2662 return libbpf_err(-EINVAL);
2663
2664 /* deconstruct BTF, if necessary, and invalidate raw_data */
2665 if (btf_ensure_modifiable(btf))
2666 return libbpf_err(-ENOMEM);
2667
2668 sz = sizeof(struct btf_type) + sizeof(struct btf_var);
2669 t = btf_add_type_mem(btf, sz);
2670 if (!t)
2671 return libbpf_err(-ENOMEM);
2672
2673 name_off = btf__add_str(btf, name);
2674 if (name_off < 0)
2675 return name_off;
2676
2677 t->name_off = name_off;
2678 t->info = btf_type_info(BTF_KIND_VAR, 0, 0);
2679 t->type = type_id;
2680
2681 v = btf_var(t);
2682 v->linkage = linkage;
2683
2684 return btf_commit_type(btf, sz);
2685}
2686
2687/*
2688 * Append new BTF_KIND_DATASEC type with:
2689 * - *name* - non-empty/non-NULL name;
2690 * - *byte_sz* - data section size, in bytes.
2691 *
2692 * Data section is initially empty. Variables info can be added with
2693 * btf__add_datasec_var_info() calls, after btf__add_datasec() succeeds.
2694 *
2695 * Returns:
2696 * - >0, type ID of newly added BTF type;
2697 * - <0, on error.
2698 */
2699int btf__add_datasec(struct btf *btf, const char *name, __u32 byte_sz)
2700{
2701 struct btf_type *t;
2702 int sz, name_off;
2703
2704 /* non-empty name */
2705 if (!name || !name[0])
2706 return libbpf_err(-EINVAL);
2707
2708 if (btf_ensure_modifiable(btf))
2709 return libbpf_err(-ENOMEM);
2710
2711 sz = sizeof(struct btf_type);
2712 t = btf_add_type_mem(btf, sz);
2713 if (!t)
2714 return libbpf_err(-ENOMEM);
2715
2716 name_off = btf__add_str(btf, name);
2717 if (name_off < 0)
2718 return name_off;
2719
2720 /* start with vlen=0, which will be update as var_secinfos are added */
2721 t->name_off = name_off;
2722 t->info = btf_type_info(BTF_KIND_DATASEC, 0, 0);
2723 t->size = byte_sz;
2724
2725 return btf_commit_type(btf, sz);
2726}
2727
2728/*
2729 * Append new data section variable information entry for current DATASEC type:
2730 * - *var_type_id* - type ID, describing type of the variable;
2731 * - *offset* - variable offset within data section, in bytes;
2732 * - *byte_sz* - variable size, in bytes.
2733 *
2734 * Returns:
2735 * - 0, on success;
2736 * - <0, on error.
2737 */
2738int btf__add_datasec_var_info(struct btf *btf, int var_type_id, __u32 offset, __u32 byte_sz)
2739{
2740 struct btf_type *t;
2741 struct btf_var_secinfo *v;
2742 int sz;
2743
2744 /* last type should be BTF_KIND_DATASEC */
2745 if (btf->nr_types == 0)
2746 return libbpf_err(-EINVAL);
2747 t = btf_last_type(btf);
2748 if (!btf_is_datasec(t))
2749 return libbpf_err(-EINVAL);
2750
2751 if (validate_type_id(var_type_id))
2752 return libbpf_err(-EINVAL);
2753
2754 /* decompose and invalidate raw data */
2755 if (btf_ensure_modifiable(btf))
2756 return libbpf_err(-ENOMEM);
2757
2758 sz = sizeof(struct btf_var_secinfo);
2759 v = btf_add_type_mem(btf, sz);
2760 if (!v)
2761 return libbpf_err(-ENOMEM);
2762
2763 v->type = var_type_id;
2764 v->offset = offset;
2765 v->size = byte_sz;
2766
2767 /* update parent type's vlen */
2768 t = btf_last_type(btf);
2769 btf_type_inc_vlen(t);
2770
2771 btf->hdr->type_len += sz;
2772 btf->hdr->str_off += sz;
2773 return 0;
2774}
2775
2776/*
2777 * Append new BTF_KIND_DECL_TAG type with:
2778 * - *value* - non-empty/non-NULL string;
2779 * - *ref_type_id* - referenced type ID, it might not exist yet;
2780 * - *component_idx* - -1 for tagging reference type, otherwise struct/union
2781 * member or function argument index;
2782 * Returns:
2783 * - >0, type ID of newly added BTF type;
2784 * - <0, on error.
2785 */
2786int btf__add_decl_tag(struct btf *btf, const char *value, int ref_type_id,
2787 int component_idx)
2788{
2789 struct btf_type *t;
2790 int sz, value_off;
2791
2792 if (!value || !value[0] || component_idx < -1)
2793 return libbpf_err(-EINVAL);
2794
2795 if (validate_type_id(ref_type_id))
2796 return libbpf_err(-EINVAL);
2797
2798 if (btf_ensure_modifiable(btf))
2799 return libbpf_err(-ENOMEM);
2800
2801 sz = sizeof(struct btf_type) + sizeof(struct btf_decl_tag);
2802 t = btf_add_type_mem(btf, sz);
2803 if (!t)
2804 return libbpf_err(-ENOMEM);
2805
2806 value_off = btf__add_str(btf, value);
2807 if (value_off < 0)
2808 return value_off;
2809
2810 t->name_off = value_off;
2811 t->info = btf_type_info(BTF_KIND_DECL_TAG, 0, false);
2812 t->type = ref_type_id;
2813 btf_decl_tag(t)->component_idx = component_idx;
2814
2815 return btf_commit_type(btf, sz);
2816}
2817
2818struct btf_ext_sec_setup_param {
2819 __u32 off;
2820 __u32 len;
2821 __u32 min_rec_size;
2822 struct btf_ext_info *ext_info;
2823 const char *desc;
2824};
2825
2826static int btf_ext_setup_info(struct btf_ext *btf_ext,
2827 struct btf_ext_sec_setup_param *ext_sec)
2828{
2829 const struct btf_ext_info_sec *sinfo;
2830 struct btf_ext_info *ext_info;
2831 __u32 info_left, record_size;
2832 size_t sec_cnt = 0;
2833 /* The start of the info sec (including the __u32 record_size). */
2834 void *info;
2835
2836 if (ext_sec->len == 0)
2837 return 0;
2838
2839 if (ext_sec->off & 0x03) {
2840 pr_debug(".BTF.ext %s section is not aligned to 4 bytes\n",
2841 ext_sec->desc);
2842 return -EINVAL;
2843 }
2844
2845 info = btf_ext->data + btf_ext->hdr->hdr_len + ext_sec->off;
2846 info_left = ext_sec->len;
2847
2848 if (btf_ext->data + btf_ext->data_size < info + ext_sec->len) {
2849 pr_debug("%s section (off:%u len:%u) is beyond the end of the ELF section .BTF.ext\n",
2850 ext_sec->desc, ext_sec->off, ext_sec->len);
2851 return -EINVAL;
2852 }
2853
2854 /* At least a record size */
2855 if (info_left < sizeof(__u32)) {
2856 pr_debug(".BTF.ext %s record size not found\n", ext_sec->desc);
2857 return -EINVAL;
2858 }
2859
2860 /* The record size needs to meet the minimum standard */
2861 record_size = *(__u32 *)info;
2862 if (record_size < ext_sec->min_rec_size ||
2863 record_size & 0x03) {
2864 pr_debug("%s section in .BTF.ext has invalid record size %u\n",
2865 ext_sec->desc, record_size);
2866 return -EINVAL;
2867 }
2868
2869 sinfo = info + sizeof(__u32);
2870 info_left -= sizeof(__u32);
2871
2872 /* If no records, return failure now so .BTF.ext won't be used. */
2873 if (!info_left) {
2874 pr_debug("%s section in .BTF.ext has no records", ext_sec->desc);
2875 return -EINVAL;
2876 }
2877
2878 while (info_left) {
2879 unsigned int sec_hdrlen = sizeof(struct btf_ext_info_sec);
2880 __u64 total_record_size;
2881 __u32 num_records;
2882
2883 if (info_left < sec_hdrlen) {
2884 pr_debug("%s section header is not found in .BTF.ext\n",
2885 ext_sec->desc);
2886 return -EINVAL;
2887 }
2888
2889 num_records = sinfo->num_info;
2890 if (num_records == 0) {
2891 pr_debug("%s section has incorrect num_records in .BTF.ext\n",
2892 ext_sec->desc);
2893 return -EINVAL;
2894 }
2895
2896 total_record_size = sec_hdrlen + (__u64)num_records * record_size;
2897 if (info_left < total_record_size) {
2898 pr_debug("%s section has incorrect num_records in .BTF.ext\n",
2899 ext_sec->desc);
2900 return -EINVAL;
2901 }
2902
2903 info_left -= total_record_size;
2904 sinfo = (void *)sinfo + total_record_size;
2905 sec_cnt++;
2906 }
2907
2908 ext_info = ext_sec->ext_info;
2909 ext_info->len = ext_sec->len - sizeof(__u32);
2910 ext_info->rec_size = record_size;
2911 ext_info->info = info + sizeof(__u32);
2912 ext_info->sec_cnt = sec_cnt;
2913
2914 return 0;
2915}
2916
2917static int btf_ext_setup_func_info(struct btf_ext *btf_ext)
2918{
2919 struct btf_ext_sec_setup_param param = {
2920 .off = btf_ext->hdr->func_info_off,
2921 .len = btf_ext->hdr->func_info_len,
2922 .min_rec_size = sizeof(struct bpf_func_info_min),
2923 .ext_info = &btf_ext->func_info,
2924 .desc = "func_info"
2925 };
2926
2927 return btf_ext_setup_info(btf_ext, ¶m);
2928}
2929
2930static int btf_ext_setup_line_info(struct btf_ext *btf_ext)
2931{
2932 struct btf_ext_sec_setup_param param = {
2933 .off = btf_ext->hdr->line_info_off,
2934 .len = btf_ext->hdr->line_info_len,
2935 .min_rec_size = sizeof(struct bpf_line_info_min),
2936 .ext_info = &btf_ext->line_info,
2937 .desc = "line_info",
2938 };
2939
2940 return btf_ext_setup_info(btf_ext, ¶m);
2941}
2942
2943static int btf_ext_setup_core_relos(struct btf_ext *btf_ext)
2944{
2945 struct btf_ext_sec_setup_param param = {
2946 .off = btf_ext->hdr->core_relo_off,
2947 .len = btf_ext->hdr->core_relo_len,
2948 .min_rec_size = sizeof(struct bpf_core_relo),
2949 .ext_info = &btf_ext->core_relo_info,
2950 .desc = "core_relo",
2951 };
2952
2953 return btf_ext_setup_info(btf_ext, ¶m);
2954}
2955
2956static int btf_ext_parse_hdr(__u8 *data, __u32 data_size)
2957{
2958 const struct btf_ext_header *hdr = (struct btf_ext_header *)data;
2959
2960 if (data_size < offsetofend(struct btf_ext_header, hdr_len) ||
2961 data_size < hdr->hdr_len) {
2962 pr_debug("BTF.ext header not found");
2963 return -EINVAL;
2964 }
2965
2966 if (hdr->magic == bswap_16(BTF_MAGIC)) {
2967 pr_warn("BTF.ext in non-native endianness is not supported\n");
2968 return -ENOTSUP;
2969 } else if (hdr->magic != BTF_MAGIC) {
2970 pr_debug("Invalid BTF.ext magic:%x\n", hdr->magic);
2971 return -EINVAL;
2972 }
2973
2974 if (hdr->version != BTF_VERSION) {
2975 pr_debug("Unsupported BTF.ext version:%u\n", hdr->version);
2976 return -ENOTSUP;
2977 }
2978
2979 if (hdr->flags) {
2980 pr_debug("Unsupported BTF.ext flags:%x\n", hdr->flags);
2981 return -ENOTSUP;
2982 }
2983
2984 if (data_size == hdr->hdr_len) {
2985 pr_debug("BTF.ext has no data\n");
2986 return -EINVAL;
2987 }
2988
2989 return 0;
2990}
2991
2992void btf_ext__free(struct btf_ext *btf_ext)
2993{
2994 if (IS_ERR_OR_NULL(btf_ext))
2995 return;
2996 free(btf_ext->func_info.sec_idxs);
2997 free(btf_ext->line_info.sec_idxs);
2998 free(btf_ext->core_relo_info.sec_idxs);
2999 free(btf_ext->data);
3000 free(btf_ext);
3001}
3002
3003struct btf_ext *btf_ext__new(const __u8 *data, __u32 size)
3004{
3005 struct btf_ext *btf_ext;
3006 int err;
3007
3008 btf_ext = calloc(1, sizeof(struct btf_ext));
3009 if (!btf_ext)
3010 return libbpf_err_ptr(-ENOMEM);
3011
3012 btf_ext->data_size = size;
3013 btf_ext->data = malloc(size);
3014 if (!btf_ext->data) {
3015 err = -ENOMEM;
3016 goto done;
3017 }
3018 memcpy(btf_ext->data, data, size);
3019
3020 err = btf_ext_parse_hdr(btf_ext->data, size);
3021 if (err)
3022 goto done;
3023
3024 if (btf_ext->hdr->hdr_len < offsetofend(struct btf_ext_header, line_info_len)) {
3025 err = -EINVAL;
3026 goto done;
3027 }
3028
3029 err = btf_ext_setup_func_info(btf_ext);
3030 if (err)
3031 goto done;
3032
3033 err = btf_ext_setup_line_info(btf_ext);
3034 if (err)
3035 goto done;
3036
3037 if (btf_ext->hdr->hdr_len < offsetofend(struct btf_ext_header, core_relo_len))
3038 goto done; /* skip core relos parsing */
3039
3040 err = btf_ext_setup_core_relos(btf_ext);
3041 if (err)
3042 goto done;
3043
3044done:
3045 if (err) {
3046 btf_ext__free(btf_ext);
3047 return libbpf_err_ptr(err);
3048 }
3049
3050 return btf_ext;
3051}
3052
3053const void *btf_ext__raw_data(const struct btf_ext *btf_ext, __u32 *size)
3054{
3055 *size = btf_ext->data_size;
3056 return btf_ext->data;
3057}
3058
3059__attribute__((alias("btf_ext__raw_data")))
3060const void *btf_ext__get_raw_data(const struct btf_ext *btf_ext, __u32 *size);
3061
3062
3063struct btf_dedup;
3064
3065static struct btf_dedup *btf_dedup_new(struct btf *btf, const struct btf_dedup_opts *opts);
3066static void btf_dedup_free(struct btf_dedup *d);
3067static int btf_dedup_prep(struct btf_dedup *d);
3068static int btf_dedup_strings(struct btf_dedup *d);
3069static int btf_dedup_prim_types(struct btf_dedup *d);
3070static int btf_dedup_struct_types(struct btf_dedup *d);
3071static int btf_dedup_ref_types(struct btf_dedup *d);
3072static int btf_dedup_resolve_fwds(struct btf_dedup *d);
3073static int btf_dedup_compact_types(struct btf_dedup *d);
3074static int btf_dedup_remap_types(struct btf_dedup *d);
3075
3076/*
3077 * Deduplicate BTF types and strings.
3078 *
3079 * BTF dedup algorithm takes as an input `struct btf` representing `.BTF` ELF
3080 * section with all BTF type descriptors and string data. It overwrites that
3081 * memory in-place with deduplicated types and strings without any loss of
3082 * information. If optional `struct btf_ext` representing '.BTF.ext' ELF section
3083 * is provided, all the strings referenced from .BTF.ext section are honored
3084 * and updated to point to the right offsets after deduplication.
3085 *
3086 * If function returns with error, type/string data might be garbled and should
3087 * be discarded.
3088 *
3089 * More verbose and detailed description of both problem btf_dedup is solving,
3090 * as well as solution could be found at:
3091 * https://facebookmicrosites.github.io/bpf/blog/2018/11/14/btf-enhancement.html
3092 *
3093 * Problem description and justification
3094 * =====================================
3095 *
3096 * BTF type information is typically emitted either as a result of conversion
3097 * from DWARF to BTF or directly by compiler. In both cases, each compilation
3098 * unit contains information about a subset of all the types that are used
3099 * in an application. These subsets are frequently overlapping and contain a lot
3100 * of duplicated information when later concatenated together into a single
3101 * binary. This algorithm ensures that each unique type is represented by single
3102 * BTF type descriptor, greatly reducing resulting size of BTF data.
3103 *
3104 * Compilation unit isolation and subsequent duplication of data is not the only
3105 * problem. The same type hierarchy (e.g., struct and all the type that struct
3106 * references) in different compilation units can be represented in BTF to
3107 * various degrees of completeness (or, rather, incompleteness) due to
3108 * struct/union forward declarations.
3109 *
3110 * Let's take a look at an example, that we'll use to better understand the
3111 * problem (and solution). Suppose we have two compilation units, each using
3112 * same `struct S`, but each of them having incomplete type information about
3113 * struct's fields:
3114 *
3115 * // CU #1:
3116 * struct S;
3117 * struct A {
3118 * int a;
3119 * struct A* self;
3120 * struct S* parent;
3121 * };
3122 * struct B;
3123 * struct S {
3124 * struct A* a_ptr;
3125 * struct B* b_ptr;
3126 * };
3127 *
3128 * // CU #2:
3129 * struct S;
3130 * struct A;
3131 * struct B {
3132 * int b;
3133 * struct B* self;
3134 * struct S* parent;
3135 * };
3136 * struct S {
3137 * struct A* a_ptr;
3138 * struct B* b_ptr;
3139 * };
3140 *
3141 * In case of CU #1, BTF data will know only that `struct B` exist (but no
3142 * more), but will know the complete type information about `struct A`. While
3143 * for CU #2, it will know full type information about `struct B`, but will
3144 * only know about forward declaration of `struct A` (in BTF terms, it will
3145 * have `BTF_KIND_FWD` type descriptor with name `B`).
3146 *
3147 * This compilation unit isolation means that it's possible that there is no
3148 * single CU with complete type information describing structs `S`, `A`, and
3149 * `B`. Also, we might get tons of duplicated and redundant type information.
3150 *
3151 * Additional complication we need to keep in mind comes from the fact that
3152 * types, in general, can form graphs containing cycles, not just DAGs.
3153 *
3154 * While algorithm does deduplication, it also merges and resolves type
3155 * information (unless disabled throught `struct btf_opts`), whenever possible.
3156 * E.g., in the example above with two compilation units having partial type
3157 * information for structs `A` and `B`, the output of algorithm will emit
3158 * a single copy of each BTF type that describes structs `A`, `B`, and `S`
3159 * (as well as type information for `int` and pointers), as if they were defined
3160 * in a single compilation unit as:
3161 *
3162 * struct A {
3163 * int a;
3164 * struct A* self;
3165 * struct S* parent;
3166 * };
3167 * struct B {
3168 * int b;
3169 * struct B* self;
3170 * struct S* parent;
3171 * };
3172 * struct S {
3173 * struct A* a_ptr;
3174 * struct B* b_ptr;
3175 * };
3176 *
3177 * Algorithm summary
3178 * =================
3179 *
3180 * Algorithm completes its work in 7 separate passes:
3181 *
3182 * 1. Strings deduplication.
3183 * 2. Primitive types deduplication (int, enum, fwd).
3184 * 3. Struct/union types deduplication.
3185 * 4. Resolve unambiguous forward declarations.
3186 * 5. Reference types deduplication (pointers, typedefs, arrays, funcs, func
3187 * protos, and const/volatile/restrict modifiers).
3188 * 6. Types compaction.
3189 * 7. Types remapping.
3190 *
3191 * Algorithm determines canonical type descriptor, which is a single
3192 * representative type for each truly unique type. This canonical type is the
3193 * one that will go into final deduplicated BTF type information. For
3194 * struct/unions, it is also the type that algorithm will merge additional type
3195 * information into (while resolving FWDs), as it discovers it from data in
3196 * other CUs. Each input BTF type eventually gets either mapped to itself, if
3197 * that type is canonical, or to some other type, if that type is equivalent
3198 * and was chosen as canonical representative. This mapping is stored in
3199 * `btf_dedup->map` array. This map is also used to record STRUCT/UNION that
3200 * FWD type got resolved to.
3201 *
3202 * To facilitate fast discovery of canonical types, we also maintain canonical
3203 * index (`btf_dedup->dedup_table`), which maps type descriptor's signature hash
3204 * (i.e., hashed kind, name, size, fields, etc) into a list of canonical types
3205 * that match that signature. With sufficiently good choice of type signature
3206 * hashing function, we can limit number of canonical types for each unique type
3207 * signature to a very small number, allowing to find canonical type for any
3208 * duplicated type very quickly.
3209 *
3210 * Struct/union deduplication is the most critical part and algorithm for
3211 * deduplicating structs/unions is described in greater details in comments for
3212 * `btf_dedup_is_equiv` function.
3213 */
3214int btf__dedup(struct btf *btf, const struct btf_dedup_opts *opts)
3215{
3216 struct btf_dedup *d;
3217 int err;
3218
3219 if (!OPTS_VALID(opts, btf_dedup_opts))
3220 return libbpf_err(-EINVAL);
3221
3222 d = btf_dedup_new(btf, opts);
3223 if (IS_ERR(d)) {
3224 pr_debug("btf_dedup_new failed: %ld", PTR_ERR(d));
3225 return libbpf_err(-EINVAL);
3226 }
3227
3228 if (btf_ensure_modifiable(btf)) {
3229 err = -ENOMEM;
3230 goto done;
3231 }
3232
3233 err = btf_dedup_prep(d);
3234 if (err) {
3235 pr_debug("btf_dedup_prep failed:%d\n", err);
3236 goto done;
3237 }
3238 err = btf_dedup_strings(d);
3239 if (err < 0) {
3240 pr_debug("btf_dedup_strings failed:%d\n", err);
3241 goto done;
3242 }
3243 err = btf_dedup_prim_types(d);
3244 if (err < 0) {
3245 pr_debug("btf_dedup_prim_types failed:%d\n", err);
3246 goto done;
3247 }
3248 err = btf_dedup_struct_types(d);
3249 if (err < 0) {
3250 pr_debug("btf_dedup_struct_types failed:%d\n", err);
3251 goto done;
3252 }
3253 err = btf_dedup_resolve_fwds(d);
3254 if (err < 0) {
3255 pr_debug("btf_dedup_resolve_fwds failed:%d\n", err);
3256 goto done;
3257 }
3258 err = btf_dedup_ref_types(d);
3259 if (err < 0) {
3260 pr_debug("btf_dedup_ref_types failed:%d\n", err);
3261 goto done;
3262 }
3263 err = btf_dedup_compact_types(d);
3264 if (err < 0) {
3265 pr_debug("btf_dedup_compact_types failed:%d\n", err);
3266 goto done;
3267 }
3268 err = btf_dedup_remap_types(d);
3269 if (err < 0) {
3270 pr_debug("btf_dedup_remap_types failed:%d\n", err);
3271 goto done;
3272 }
3273
3274done:
3275 btf_dedup_free(d);
3276 return libbpf_err(err);
3277}
3278
3279#define BTF_UNPROCESSED_ID ((__u32)-1)
3280#define BTF_IN_PROGRESS_ID ((__u32)-2)
3281
3282struct btf_dedup {
3283 /* .BTF section to be deduped in-place */
3284 struct btf *btf;
3285 /*
3286 * Optional .BTF.ext section. When provided, any strings referenced
3287 * from it will be taken into account when deduping strings
3288 */
3289 struct btf_ext *btf_ext;
3290 /*
3291 * This is a map from any type's signature hash to a list of possible
3292 * canonical representative type candidates. Hash collisions are
3293 * ignored, so even types of various kinds can share same list of
3294 * candidates, which is fine because we rely on subsequent
3295 * btf_xxx_equal() checks to authoritatively verify type equality.
3296 */
3297 struct hashmap *dedup_table;
3298 /* Canonical types map */
3299 __u32 *map;
3300 /* Hypothetical mapping, used during type graph equivalence checks */
3301 __u32 *hypot_map;
3302 __u32 *hypot_list;
3303 size_t hypot_cnt;
3304 size_t hypot_cap;
3305 /* Whether hypothetical mapping, if successful, would need to adjust
3306 * already canonicalized types (due to a new forward declaration to
3307 * concrete type resolution). In such case, during split BTF dedup
3308 * candidate type would still be considered as different, because base
3309 * BTF is considered to be immutable.
3310 */
3311 bool hypot_adjust_canon;
3312 /* Various option modifying behavior of algorithm */
3313 struct btf_dedup_opts opts;
3314 /* temporary strings deduplication state */
3315 struct strset *strs_set;
3316};
3317
3318static long hash_combine(long h, long value)
3319{
3320 return h * 31 + value;
3321}
3322
3323#define for_each_dedup_cand(d, node, hash) \
3324 hashmap__for_each_key_entry(d->dedup_table, node, hash)
3325
3326static int btf_dedup_table_add(struct btf_dedup *d, long hash, __u32 type_id)
3327{
3328 return hashmap__append(d->dedup_table, hash, type_id);
3329}
3330
3331static int btf_dedup_hypot_map_add(struct btf_dedup *d,
3332 __u32 from_id, __u32 to_id)
3333{
3334 if (d->hypot_cnt == d->hypot_cap) {
3335 __u32 *new_list;
3336
3337 d->hypot_cap += max((size_t)16, d->hypot_cap / 2);
3338 new_list = libbpf_reallocarray(d->hypot_list, d->hypot_cap, sizeof(__u32));
3339 if (!new_list)
3340 return -ENOMEM;
3341 d->hypot_list = new_list;
3342 }
3343 d->hypot_list[d->hypot_cnt++] = from_id;
3344 d->hypot_map[from_id] = to_id;
3345 return 0;
3346}
3347
3348static void btf_dedup_clear_hypot_map(struct btf_dedup *d)
3349{
3350 int i;
3351
3352 for (i = 0; i < d->hypot_cnt; i++)
3353 d->hypot_map[d->hypot_list[i]] = BTF_UNPROCESSED_ID;
3354 d->hypot_cnt = 0;
3355 d->hypot_adjust_canon = false;
3356}
3357
3358static void btf_dedup_free(struct btf_dedup *d)
3359{
3360 hashmap__free(d->dedup_table);
3361 d->dedup_table = NULL;
3362
3363 free(d->map);
3364 d->map = NULL;
3365
3366 free(d->hypot_map);
3367 d->hypot_map = NULL;
3368
3369 free(d->hypot_list);
3370 d->hypot_list = NULL;
3371
3372 free(d);
3373}
3374
3375static size_t btf_dedup_identity_hash_fn(long key, void *ctx)
3376{
3377 return key;
3378}
3379
3380static size_t btf_dedup_collision_hash_fn(long key, void *ctx)
3381{
3382 return 0;
3383}
3384
3385static bool btf_dedup_equal_fn(long k1, long k2, void *ctx)
3386{
3387 return k1 == k2;
3388}
3389
3390static struct btf_dedup *btf_dedup_new(struct btf *btf, const struct btf_dedup_opts *opts)
3391{
3392 struct btf_dedup *d = calloc(1, sizeof(struct btf_dedup));
3393 hashmap_hash_fn hash_fn = btf_dedup_identity_hash_fn;
3394 int i, err = 0, type_cnt;
3395
3396 if (!d)
3397 return ERR_PTR(-ENOMEM);
3398
3399 if (OPTS_GET(opts, force_collisions, false))
3400 hash_fn = btf_dedup_collision_hash_fn;
3401
3402 d->btf = btf;
3403 d->btf_ext = OPTS_GET(opts, btf_ext, NULL);
3404
3405 d->dedup_table = hashmap__new(hash_fn, btf_dedup_equal_fn, NULL);
3406 if (IS_ERR(d->dedup_table)) {
3407 err = PTR_ERR(d->dedup_table);
3408 d->dedup_table = NULL;
3409 goto done;
3410 }
3411
3412 type_cnt = btf__type_cnt(btf);
3413 d->map = malloc(sizeof(__u32) * type_cnt);
3414 if (!d->map) {
3415 err = -ENOMEM;
3416 goto done;
3417 }
3418 /* special BTF "void" type is made canonical immediately */
3419 d->map[0] = 0;
3420 for (i = 1; i < type_cnt; i++) {
3421 struct btf_type *t = btf_type_by_id(d->btf, i);
3422
3423 /* VAR and DATASEC are never deduped and are self-canonical */
3424 if (btf_is_var(t) || btf_is_datasec(t))
3425 d->map[i] = i;
3426 else
3427 d->map[i] = BTF_UNPROCESSED_ID;
3428 }
3429
3430 d->hypot_map = malloc(sizeof(__u32) * type_cnt);
3431 if (!d->hypot_map) {
3432 err = -ENOMEM;
3433 goto done;
3434 }
3435 for (i = 0; i < type_cnt; i++)
3436 d->hypot_map[i] = BTF_UNPROCESSED_ID;
3437
3438done:
3439 if (err) {
3440 btf_dedup_free(d);
3441 return ERR_PTR(err);
3442 }
3443
3444 return d;
3445}
3446
3447/*
3448 * Iterate over all possible places in .BTF and .BTF.ext that can reference
3449 * string and pass pointer to it to a provided callback `fn`.
3450 */
3451static int btf_for_each_str_off(struct btf_dedup *d, str_off_visit_fn fn, void *ctx)
3452{
3453 int i, r;
3454
3455 for (i = 0; i < d->btf->nr_types; i++) {
3456 struct btf_type *t = btf_type_by_id(d->btf, d->btf->start_id + i);
3457
3458 r = btf_type_visit_str_offs(t, fn, ctx);
3459 if (r)
3460 return r;
3461 }
3462
3463 if (!d->btf_ext)
3464 return 0;
3465
3466 r = btf_ext_visit_str_offs(d->btf_ext, fn, ctx);
3467 if (r)
3468 return r;
3469
3470 return 0;
3471}
3472
3473static int strs_dedup_remap_str_off(__u32 *str_off_ptr, void *ctx)
3474{
3475 struct btf_dedup *d = ctx;
3476 __u32 str_off = *str_off_ptr;
3477 const char *s;
3478 int off, err;
3479
3480 /* don't touch empty string or string in main BTF */
3481 if (str_off == 0 || str_off < d->btf->start_str_off)
3482 return 0;
3483
3484 s = btf__str_by_offset(d->btf, str_off);
3485 if (d->btf->base_btf) {
3486 err = btf__find_str(d->btf->base_btf, s);
3487 if (err >= 0) {
3488 *str_off_ptr = err;
3489 return 0;
3490 }
3491 if (err != -ENOENT)
3492 return err;
3493 }
3494
3495 off = strset__add_str(d->strs_set, s);
3496 if (off < 0)
3497 return off;
3498
3499 *str_off_ptr = d->btf->start_str_off + off;
3500 return 0;
3501}
3502
3503/*
3504 * Dedup string and filter out those that are not referenced from either .BTF
3505 * or .BTF.ext (if provided) sections.
3506 *
3507 * This is done by building index of all strings in BTF's string section,
3508 * then iterating over all entities that can reference strings (e.g., type
3509 * names, struct field names, .BTF.ext line info, etc) and marking corresponding
3510 * strings as used. After that all used strings are deduped and compacted into
3511 * sequential blob of memory and new offsets are calculated. Then all the string
3512 * references are iterated again and rewritten using new offsets.
3513 */
3514static int btf_dedup_strings(struct btf_dedup *d)
3515{
3516 int err;
3517
3518 if (d->btf->strs_deduped)
3519 return 0;
3520
3521 d->strs_set = strset__new(BTF_MAX_STR_OFFSET, NULL, 0);
3522 if (IS_ERR(d->strs_set)) {
3523 err = PTR_ERR(d->strs_set);
3524 goto err_out;
3525 }
3526
3527 if (!d->btf->base_btf) {
3528 /* insert empty string; we won't be looking it up during strings
3529 * dedup, but it's good to have it for generic BTF string lookups
3530 */
3531 err = strset__add_str(d->strs_set, "");
3532 if (err < 0)
3533 goto err_out;
3534 }
3535
3536 /* remap string offsets */
3537 err = btf_for_each_str_off(d, strs_dedup_remap_str_off, d);
3538 if (err)
3539 goto err_out;
3540
3541 /* replace BTF string data and hash with deduped ones */
3542 strset__free(d->btf->strs_set);
3543 d->btf->hdr->str_len = strset__data_size(d->strs_set);
3544 d->btf->strs_set = d->strs_set;
3545 d->strs_set = NULL;
3546 d->btf->strs_deduped = true;
3547 return 0;
3548
3549err_out:
3550 strset__free(d->strs_set);
3551 d->strs_set = NULL;
3552
3553 return err;
3554}
3555
3556static long btf_hash_common(struct btf_type *t)
3557{
3558 long h;
3559
3560 h = hash_combine(0, t->name_off);
3561 h = hash_combine(h, t->info);
3562 h = hash_combine(h, t->size);
3563 return h;
3564}
3565
3566static bool btf_equal_common(struct btf_type *t1, struct btf_type *t2)
3567{
3568 return t1->name_off == t2->name_off &&
3569 t1->info == t2->info &&
3570 t1->size == t2->size;
3571}
3572
3573/* Calculate type signature hash of INT or TAG. */
3574static long btf_hash_int_decl_tag(struct btf_type *t)
3575{
3576 __u32 info = *(__u32 *)(t + 1);
3577 long h;
3578
3579 h = btf_hash_common(t);
3580 h = hash_combine(h, info);
3581 return h;
3582}
3583
3584/* Check structural equality of two INTs or TAGs. */
3585static bool btf_equal_int_tag(struct btf_type *t1, struct btf_type *t2)
3586{
3587 __u32 info1, info2;
3588
3589 if (!btf_equal_common(t1, t2))
3590 return false;
3591 info1 = *(__u32 *)(t1 + 1);
3592 info2 = *(__u32 *)(t2 + 1);
3593 return info1 == info2;
3594}
3595
3596/* Calculate type signature hash of ENUM/ENUM64. */
3597static long btf_hash_enum(struct btf_type *t)
3598{
3599 long h;
3600
3601 /* don't hash vlen, enum members and size to support enum fwd resolving */
3602 h = hash_combine(0, t->name_off);
3603 return h;
3604}
3605
3606static bool btf_equal_enum_members(struct btf_type *t1, struct btf_type *t2)
3607{
3608 const struct btf_enum *m1, *m2;
3609 __u16 vlen;
3610 int i;
3611
3612 vlen = btf_vlen(t1);
3613 m1 = btf_enum(t1);
3614 m2 = btf_enum(t2);
3615 for (i = 0; i < vlen; i++) {
3616 if (m1->name_off != m2->name_off || m1->val != m2->val)
3617 return false;
3618 m1++;
3619 m2++;
3620 }
3621 return true;
3622}
3623
3624static bool btf_equal_enum64_members(struct btf_type *t1, struct btf_type *t2)
3625{
3626 const struct btf_enum64 *m1, *m2;
3627 __u16 vlen;
3628 int i;
3629
3630 vlen = btf_vlen(t1);
3631 m1 = btf_enum64(t1);
3632 m2 = btf_enum64(t2);
3633 for (i = 0; i < vlen; i++) {
3634 if (m1->name_off != m2->name_off || m1->val_lo32 != m2->val_lo32 ||
3635 m1->val_hi32 != m2->val_hi32)
3636 return false;
3637 m1++;
3638 m2++;
3639 }
3640 return true;
3641}
3642
3643/* Check structural equality of two ENUMs or ENUM64s. */
3644static bool btf_equal_enum(struct btf_type *t1, struct btf_type *t2)
3645{
3646 if (!btf_equal_common(t1, t2))
3647 return false;
3648
3649 /* t1 & t2 kinds are identical because of btf_equal_common */
3650 if (btf_kind(t1) == BTF_KIND_ENUM)
3651 return btf_equal_enum_members(t1, t2);
3652 else
3653 return btf_equal_enum64_members(t1, t2);
3654}
3655
3656static inline bool btf_is_enum_fwd(struct btf_type *t)
3657{
3658 return btf_is_any_enum(t) && btf_vlen(t) == 0;
3659}
3660
3661static bool btf_compat_enum(struct btf_type *t1, struct btf_type *t2)
3662{
3663 if (!btf_is_enum_fwd(t1) && !btf_is_enum_fwd(t2))
3664 return btf_equal_enum(t1, t2);
3665 /* At this point either t1 or t2 or both are forward declarations, thus:
3666 * - skip comparing vlen because it is zero for forward declarations;
3667 * - skip comparing size to allow enum forward declarations
3668 * to be compatible with enum64 full declarations;
3669 * - skip comparing kind for the same reason.
3670 */
3671 return t1->name_off == t2->name_off &&
3672 btf_is_any_enum(t1) && btf_is_any_enum(t2);
3673}
3674
3675/*
3676 * Calculate type signature hash of STRUCT/UNION, ignoring referenced type IDs,
3677 * as referenced type IDs equivalence is established separately during type
3678 * graph equivalence check algorithm.
3679 */
3680static long btf_hash_struct(struct btf_type *t)
3681{
3682 const struct btf_member *member = btf_members(t);
3683 __u32 vlen = btf_vlen(t);
3684 long h = btf_hash_common(t);
3685 int i;
3686
3687 for (i = 0; i < vlen; i++) {
3688 h = hash_combine(h, member->name_off);
3689 h = hash_combine(h, member->offset);
3690 /* no hashing of referenced type ID, it can be unresolved yet */
3691 member++;
3692 }
3693 return h;
3694}
3695
3696/*
3697 * Check structural compatibility of two STRUCTs/UNIONs, ignoring referenced
3698 * type IDs. This check is performed during type graph equivalence check and
3699 * referenced types equivalence is checked separately.
3700 */
3701static bool btf_shallow_equal_struct(struct btf_type *t1, struct btf_type *t2)
3702{
3703 const struct btf_member *m1, *m2;
3704 __u16 vlen;
3705 int i;
3706
3707 if (!btf_equal_common(t1, t2))
3708 return false;
3709
3710 vlen = btf_vlen(t1);
3711 m1 = btf_members(t1);
3712 m2 = btf_members(t2);
3713 for (i = 0; i < vlen; i++) {
3714 if (m1->name_off != m2->name_off || m1->offset != m2->offset)
3715 return false;
3716 m1++;
3717 m2++;
3718 }
3719 return true;
3720}
3721
3722/*
3723 * Calculate type signature hash of ARRAY, including referenced type IDs,
3724 * under assumption that they were already resolved to canonical type IDs and
3725 * are not going to change.
3726 */
3727static long btf_hash_array(struct btf_type *t)
3728{
3729 const struct btf_array *info = btf_array(t);
3730 long h = btf_hash_common(t);
3731
3732 h = hash_combine(h, info->type);
3733 h = hash_combine(h, info->index_type);
3734 h = hash_combine(h, info->nelems);
3735 return h;
3736}
3737
3738/*
3739 * Check exact equality of two ARRAYs, taking into account referenced
3740 * type IDs, under assumption that they were already resolved to canonical
3741 * type IDs and are not going to change.
3742 * This function is called during reference types deduplication to compare
3743 * ARRAY to potential canonical representative.
3744 */
3745static bool btf_equal_array(struct btf_type *t1, struct btf_type *t2)
3746{
3747 const struct btf_array *info1, *info2;
3748
3749 if (!btf_equal_common(t1, t2))
3750 return false;
3751
3752 info1 = btf_array(t1);
3753 info2 = btf_array(t2);
3754 return info1->type == info2->type &&
3755 info1->index_type == info2->index_type &&
3756 info1->nelems == info2->nelems;
3757}
3758
3759/*
3760 * Check structural compatibility of two ARRAYs, ignoring referenced type
3761 * IDs. This check is performed during type graph equivalence check and
3762 * referenced types equivalence is checked separately.
3763 */
3764static bool btf_compat_array(struct btf_type *t1, struct btf_type *t2)
3765{
3766 if (!btf_equal_common(t1, t2))
3767 return false;
3768
3769 return btf_array(t1)->nelems == btf_array(t2)->nelems;
3770}
3771
3772/*
3773 * Calculate type signature hash of FUNC_PROTO, including referenced type IDs,
3774 * under assumption that they were already resolved to canonical type IDs and
3775 * are not going to change.
3776 */
3777static long btf_hash_fnproto(struct btf_type *t)
3778{
3779 const struct btf_param *member = btf_params(t);
3780 __u16 vlen = btf_vlen(t);
3781 long h = btf_hash_common(t);
3782 int i;
3783
3784 for (i = 0; i < vlen; i++) {
3785 h = hash_combine(h, member->name_off);
3786 h = hash_combine(h, member->type);
3787 member++;
3788 }
3789 return h;
3790}
3791
3792/*
3793 * Check exact equality of two FUNC_PROTOs, taking into account referenced
3794 * type IDs, under assumption that they were already resolved to canonical
3795 * type IDs and are not going to change.
3796 * This function is called during reference types deduplication to compare
3797 * FUNC_PROTO to potential canonical representative.
3798 */
3799static bool btf_equal_fnproto(struct btf_type *t1, struct btf_type *t2)
3800{
3801 const struct btf_param *m1, *m2;
3802 __u16 vlen;
3803 int i;
3804
3805 if (!btf_equal_common(t1, t2))
3806 return false;
3807
3808 vlen = btf_vlen(t1);
3809 m1 = btf_params(t1);
3810 m2 = btf_params(t2);
3811 for (i = 0; i < vlen; i++) {
3812 if (m1->name_off != m2->name_off || m1->type != m2->type)
3813 return false;
3814 m1++;
3815 m2++;
3816 }
3817 return true;
3818}
3819
3820/*
3821 * Check structural compatibility of two FUNC_PROTOs, ignoring referenced type
3822 * IDs. This check is performed during type graph equivalence check and
3823 * referenced types equivalence is checked separately.
3824 */
3825static bool btf_compat_fnproto(struct btf_type *t1, struct btf_type *t2)
3826{
3827 const struct btf_param *m1, *m2;
3828 __u16 vlen;
3829 int i;
3830
3831 /* skip return type ID */
3832 if (t1->name_off != t2->name_off || t1->info != t2->info)
3833 return false;
3834
3835 vlen = btf_vlen(t1);
3836 m1 = btf_params(t1);
3837 m2 = btf_params(t2);
3838 for (i = 0; i < vlen; i++) {
3839 if (m1->name_off != m2->name_off)
3840 return false;
3841 m1++;
3842 m2++;
3843 }
3844 return true;
3845}
3846
3847/* Prepare split BTF for deduplication by calculating hashes of base BTF's
3848 * types and initializing the rest of the state (canonical type mapping) for
3849 * the fixed base BTF part.
3850 */
3851static int btf_dedup_prep(struct btf_dedup *d)
3852{
3853 struct btf_type *t;
3854 int type_id;
3855 long h;
3856
3857 if (!d->btf->base_btf)
3858 return 0;
3859
3860 for (type_id = 1; type_id < d->btf->start_id; type_id++) {
3861 t = btf_type_by_id(d->btf, type_id);
3862
3863 /* all base BTF types are self-canonical by definition */
3864 d->map[type_id] = type_id;
3865
3866 switch (btf_kind(t)) {
3867 case BTF_KIND_VAR:
3868 case BTF_KIND_DATASEC:
3869 /* VAR and DATASEC are never hash/deduplicated */
3870 continue;
3871 case BTF_KIND_CONST:
3872 case BTF_KIND_VOLATILE:
3873 case BTF_KIND_RESTRICT:
3874 case BTF_KIND_PTR:
3875 case BTF_KIND_FWD:
3876 case BTF_KIND_TYPEDEF:
3877 case BTF_KIND_FUNC:
3878 case BTF_KIND_FLOAT:
3879 case BTF_KIND_TYPE_TAG:
3880 h = btf_hash_common(t);
3881 break;
3882 case BTF_KIND_INT:
3883 case BTF_KIND_DECL_TAG:
3884 h = btf_hash_int_decl_tag(t);
3885 break;
3886 case BTF_KIND_ENUM:
3887 case BTF_KIND_ENUM64:
3888 h = btf_hash_enum(t);
3889 break;
3890 case BTF_KIND_STRUCT:
3891 case BTF_KIND_UNION:
3892 h = btf_hash_struct(t);
3893 break;
3894 case BTF_KIND_ARRAY:
3895 h = btf_hash_array(t);
3896 break;
3897 case BTF_KIND_FUNC_PROTO:
3898 h = btf_hash_fnproto(t);
3899 break;
3900 default:
3901 pr_debug("unknown kind %d for type [%d]\n", btf_kind(t), type_id);
3902 return -EINVAL;
3903 }
3904 if (btf_dedup_table_add(d, h, type_id))
3905 return -ENOMEM;
3906 }
3907
3908 return 0;
3909}
3910
3911/*
3912 * Deduplicate primitive types, that can't reference other types, by calculating
3913 * their type signature hash and comparing them with any possible canonical
3914 * candidate. If no canonical candidate matches, type itself is marked as
3915 * canonical and is added into `btf_dedup->dedup_table` as another candidate.
3916 */
3917static int btf_dedup_prim_type(struct btf_dedup *d, __u32 type_id)
3918{
3919 struct btf_type *t = btf_type_by_id(d->btf, type_id);
3920 struct hashmap_entry *hash_entry;
3921 struct btf_type *cand;
3922 /* if we don't find equivalent type, then we are canonical */
3923 __u32 new_id = type_id;
3924 __u32 cand_id;
3925 long h;
3926
3927 switch (btf_kind(t)) {
3928 case BTF_KIND_CONST:
3929 case BTF_KIND_VOLATILE:
3930 case BTF_KIND_RESTRICT:
3931 case BTF_KIND_PTR:
3932 case BTF_KIND_TYPEDEF:
3933 case BTF_KIND_ARRAY:
3934 case BTF_KIND_STRUCT:
3935 case BTF_KIND_UNION:
3936 case BTF_KIND_FUNC:
3937 case BTF_KIND_FUNC_PROTO:
3938 case BTF_KIND_VAR:
3939 case BTF_KIND_DATASEC:
3940 case BTF_KIND_DECL_TAG:
3941 case BTF_KIND_TYPE_TAG:
3942 return 0;
3943
3944 case BTF_KIND_INT:
3945 h = btf_hash_int_decl_tag(t);
3946 for_each_dedup_cand(d, hash_entry, h) {
3947 cand_id = hash_entry->value;
3948 cand = btf_type_by_id(d->btf, cand_id);
3949 if (btf_equal_int_tag(t, cand)) {
3950 new_id = cand_id;
3951 break;
3952 }
3953 }
3954 break;
3955
3956 case BTF_KIND_ENUM:
3957 case BTF_KIND_ENUM64:
3958 h = btf_hash_enum(t);
3959 for_each_dedup_cand(d, hash_entry, h) {
3960 cand_id = hash_entry->value;
3961 cand = btf_type_by_id(d->btf, cand_id);
3962 if (btf_equal_enum(t, cand)) {
3963 new_id = cand_id;
3964 break;
3965 }
3966 if (btf_compat_enum(t, cand)) {
3967 if (btf_is_enum_fwd(t)) {
3968 /* resolve fwd to full enum */
3969 new_id = cand_id;
3970 break;
3971 }
3972 /* resolve canonical enum fwd to full enum */
3973 d->map[cand_id] = type_id;
3974 }
3975 }
3976 break;
3977
3978 case BTF_KIND_FWD:
3979 case BTF_KIND_FLOAT:
3980 h = btf_hash_common(t);
3981 for_each_dedup_cand(d, hash_entry, h) {
3982 cand_id = hash_entry->value;
3983 cand = btf_type_by_id(d->btf, cand_id);
3984 if (btf_equal_common(t, cand)) {
3985 new_id = cand_id;
3986 break;
3987 }
3988 }
3989 break;
3990
3991 default:
3992 return -EINVAL;
3993 }
3994
3995 d->map[type_id] = new_id;
3996 if (type_id == new_id && btf_dedup_table_add(d, h, type_id))
3997 return -ENOMEM;
3998
3999 return 0;
4000}
4001
4002static int btf_dedup_prim_types(struct btf_dedup *d)
4003{
4004 int i, err;
4005
4006 for (i = 0; i < d->btf->nr_types; i++) {
4007 err = btf_dedup_prim_type(d, d->btf->start_id + i);
4008 if (err)
4009 return err;
4010 }
4011 return 0;
4012}
4013
4014/*
4015 * Check whether type is already mapped into canonical one (could be to itself).
4016 */
4017static inline bool is_type_mapped(struct btf_dedup *d, uint32_t type_id)
4018{
4019 return d->map[type_id] <= BTF_MAX_NR_TYPES;
4020}
4021
4022/*
4023 * Resolve type ID into its canonical type ID, if any; otherwise return original
4024 * type ID. If type is FWD and is resolved into STRUCT/UNION already, follow
4025 * STRUCT/UNION link and resolve it into canonical type ID as well.
4026 */
4027static inline __u32 resolve_type_id(struct btf_dedup *d, __u32 type_id)
4028{
4029 while (is_type_mapped(d, type_id) && d->map[type_id] != type_id)
4030 type_id = d->map[type_id];
4031 return type_id;
4032}
4033
4034/*
4035 * Resolve FWD to underlying STRUCT/UNION, if any; otherwise return original
4036 * type ID.
4037 */
4038static uint32_t resolve_fwd_id(struct btf_dedup *d, uint32_t type_id)
4039{
4040 __u32 orig_type_id = type_id;
4041
4042 if (!btf_is_fwd(btf__type_by_id(d->btf, type_id)))
4043 return type_id;
4044
4045 while (is_type_mapped(d, type_id) && d->map[type_id] != type_id)
4046 type_id = d->map[type_id];
4047
4048 if (!btf_is_fwd(btf__type_by_id(d->btf, type_id)))
4049 return type_id;
4050
4051 return orig_type_id;
4052}
4053
4054
4055static inline __u16 btf_fwd_kind(struct btf_type *t)
4056{
4057 return btf_kflag(t) ? BTF_KIND_UNION : BTF_KIND_STRUCT;
4058}
4059
4060/* Check if given two types are identical ARRAY definitions */
4061static bool btf_dedup_identical_arrays(struct btf_dedup *d, __u32 id1, __u32 id2)
4062{
4063 struct btf_type *t1, *t2;
4064
4065 t1 = btf_type_by_id(d->btf, id1);
4066 t2 = btf_type_by_id(d->btf, id2);
4067 if (!btf_is_array(t1) || !btf_is_array(t2))
4068 return false;
4069
4070 return btf_equal_array(t1, t2);
4071}
4072
4073/* Check if given two types are identical STRUCT/UNION definitions */
4074static bool btf_dedup_identical_structs(struct btf_dedup *d, __u32 id1, __u32 id2)
4075{
4076 const struct btf_member *m1, *m2;
4077 struct btf_type *t1, *t2;
4078 int n, i;
4079
4080 t1 = btf_type_by_id(d->btf, id1);
4081 t2 = btf_type_by_id(d->btf, id2);
4082
4083 if (!btf_is_composite(t1) || btf_kind(t1) != btf_kind(t2))
4084 return false;
4085
4086 if (!btf_shallow_equal_struct(t1, t2))
4087 return false;
4088
4089 m1 = btf_members(t1);
4090 m2 = btf_members(t2);
4091 for (i = 0, n = btf_vlen(t1); i < n; i++, m1++, m2++) {
4092 if (m1->type != m2->type &&
4093 !btf_dedup_identical_arrays(d, m1->type, m2->type) &&
4094 !btf_dedup_identical_structs(d, m1->type, m2->type))
4095 return false;
4096 }
4097 return true;
4098}
4099
4100/*
4101 * Check equivalence of BTF type graph formed by candidate struct/union (we'll
4102 * call it "candidate graph" in this description for brevity) to a type graph
4103 * formed by (potential) canonical struct/union ("canonical graph" for brevity
4104 * here, though keep in mind that not all types in canonical graph are
4105 * necessarily canonical representatives themselves, some of them might be
4106 * duplicates or its uniqueness might not have been established yet).
4107 * Returns:
4108 * - >0, if type graphs are equivalent;
4109 * - 0, if not equivalent;
4110 * - <0, on error.
4111 *
4112 * Algorithm performs side-by-side DFS traversal of both type graphs and checks
4113 * equivalence of BTF types at each step. If at any point BTF types in candidate
4114 * and canonical graphs are not compatible structurally, whole graphs are
4115 * incompatible. If types are structurally equivalent (i.e., all information
4116 * except referenced type IDs is exactly the same), a mapping from `canon_id` to
4117 * a `cand_id` is recored in hypothetical mapping (`btf_dedup->hypot_map`).
4118 * If a type references other types, then those referenced types are checked
4119 * for equivalence recursively.
4120 *
4121 * During DFS traversal, if we find that for current `canon_id` type we
4122 * already have some mapping in hypothetical map, we check for two possible
4123 * situations:
4124 * - `canon_id` is mapped to exactly the same type as `cand_id`. This will
4125 * happen when type graphs have cycles. In this case we assume those two
4126 * types are equivalent.
4127 * - `canon_id` is mapped to different type. This is contradiction in our
4128 * hypothetical mapping, because same graph in canonical graph corresponds
4129 * to two different types in candidate graph, which for equivalent type
4130 * graphs shouldn't happen. This condition terminates equivalence check
4131 * with negative result.
4132 *
4133 * If type graphs traversal exhausts types to check and find no contradiction,
4134 * then type graphs are equivalent.
4135 *
4136 * When checking types for equivalence, there is one special case: FWD types.
4137 * If FWD type resolution is allowed and one of the types (either from canonical
4138 * or candidate graph) is FWD and other is STRUCT/UNION (depending on FWD's kind
4139 * flag) and their names match, hypothetical mapping is updated to point from
4140 * FWD to STRUCT/UNION. If graphs will be determined as equivalent successfully,
4141 * this mapping will be used to record FWD -> STRUCT/UNION mapping permanently.
4142 *
4143 * Technically, this could lead to incorrect FWD to STRUCT/UNION resolution,
4144 * if there are two exactly named (or anonymous) structs/unions that are
4145 * compatible structurally, one of which has FWD field, while other is concrete
4146 * STRUCT/UNION, but according to C sources they are different structs/unions
4147 * that are referencing different types with the same name. This is extremely
4148 * unlikely to happen, but btf_dedup API allows to disable FWD resolution if
4149 * this logic is causing problems.
4150 *
4151 * Doing FWD resolution means that both candidate and/or canonical graphs can
4152 * consists of portions of the graph that come from multiple compilation units.
4153 * This is due to the fact that types within single compilation unit are always
4154 * deduplicated and FWDs are already resolved, if referenced struct/union
4155 * definiton is available. So, if we had unresolved FWD and found corresponding
4156 * STRUCT/UNION, they will be from different compilation units. This
4157 * consequently means that when we "link" FWD to corresponding STRUCT/UNION,
4158 * type graph will likely have at least two different BTF types that describe
4159 * same type (e.g., most probably there will be two different BTF types for the
4160 * same 'int' primitive type) and could even have "overlapping" parts of type
4161 * graph that describe same subset of types.
4162 *
4163 * This in turn means that our assumption that each type in canonical graph
4164 * must correspond to exactly one type in candidate graph might not hold
4165 * anymore and will make it harder to detect contradictions using hypothetical
4166 * map. To handle this problem, we allow to follow FWD -> STRUCT/UNION
4167 * resolution only in canonical graph. FWDs in candidate graphs are never
4168 * resolved. To see why it's OK, let's check all possible situations w.r.t. FWDs
4169 * that can occur:
4170 * - Both types in canonical and candidate graphs are FWDs. If they are
4171 * structurally equivalent, then they can either be both resolved to the
4172 * same STRUCT/UNION or not resolved at all. In both cases they are
4173 * equivalent and there is no need to resolve FWD on candidate side.
4174 * - Both types in canonical and candidate graphs are concrete STRUCT/UNION,
4175 * so nothing to resolve as well, algorithm will check equivalence anyway.
4176 * - Type in canonical graph is FWD, while type in candidate is concrete
4177 * STRUCT/UNION. In this case candidate graph comes from single compilation
4178 * unit, so there is exactly one BTF type for each unique C type. After
4179 * resolving FWD into STRUCT/UNION, there might be more than one BTF type
4180 * in canonical graph mapping to single BTF type in candidate graph, but
4181 * because hypothetical mapping maps from canonical to candidate types, it's
4182 * alright, and we still maintain the property of having single `canon_id`
4183 * mapping to single `cand_id` (there could be two different `canon_id`
4184 * mapped to the same `cand_id`, but it's not contradictory).
4185 * - Type in canonical graph is concrete STRUCT/UNION, while type in candidate
4186 * graph is FWD. In this case we are just going to check compatibility of
4187 * STRUCT/UNION and corresponding FWD, and if they are compatible, we'll
4188 * assume that whatever STRUCT/UNION FWD resolves to must be equivalent to
4189 * a concrete STRUCT/UNION from canonical graph. If the rest of type graphs
4190 * turn out equivalent, we'll re-resolve FWD to concrete STRUCT/UNION from
4191 * canonical graph.
4192 */
4193static int btf_dedup_is_equiv(struct btf_dedup *d, __u32 cand_id,
4194 __u32 canon_id)
4195{
4196 struct btf_type *cand_type;
4197 struct btf_type *canon_type;
4198 __u32 hypot_type_id;
4199 __u16 cand_kind;
4200 __u16 canon_kind;
4201 int i, eq;
4202
4203 /* if both resolve to the same canonical, they must be equivalent */
4204 if (resolve_type_id(d, cand_id) == resolve_type_id(d, canon_id))
4205 return 1;
4206
4207 canon_id = resolve_fwd_id(d, canon_id);
4208
4209 hypot_type_id = d->hypot_map[canon_id];
4210 if (hypot_type_id <= BTF_MAX_NR_TYPES) {
4211 if (hypot_type_id == cand_id)
4212 return 1;
4213 /* In some cases compiler will generate different DWARF types
4214 * for *identical* array type definitions and use them for
4215 * different fields within the *same* struct. This breaks type
4216 * equivalence check, which makes an assumption that candidate
4217 * types sub-graph has a consistent and deduped-by-compiler
4218 * types within a single CU. So work around that by explicitly
4219 * allowing identical array types here.
4220 */
4221 if (btf_dedup_identical_arrays(d, hypot_type_id, cand_id))
4222 return 1;
4223 /* It turns out that similar situation can happen with
4224 * struct/union sometimes, sigh... Handle the case where
4225 * structs/unions are exactly the same, down to the referenced
4226 * type IDs. Anything more complicated (e.g., if referenced
4227 * types are different, but equivalent) is *way more*
4228 * complicated and requires a many-to-many equivalence mapping.
4229 */
4230 if (btf_dedup_identical_structs(d, hypot_type_id, cand_id))
4231 return 1;
4232 return 0;
4233 }
4234
4235 if (btf_dedup_hypot_map_add(d, canon_id, cand_id))
4236 return -ENOMEM;
4237
4238 cand_type = btf_type_by_id(d->btf, cand_id);
4239 canon_type = btf_type_by_id(d->btf, canon_id);
4240 cand_kind = btf_kind(cand_type);
4241 canon_kind = btf_kind(canon_type);
4242
4243 if (cand_type->name_off != canon_type->name_off)
4244 return 0;
4245
4246 /* FWD <--> STRUCT/UNION equivalence check, if enabled */
4247 if ((cand_kind == BTF_KIND_FWD || canon_kind == BTF_KIND_FWD)
4248 && cand_kind != canon_kind) {
4249 __u16 real_kind;
4250 __u16 fwd_kind;
4251
4252 if (cand_kind == BTF_KIND_FWD) {
4253 real_kind = canon_kind;
4254 fwd_kind = btf_fwd_kind(cand_type);
4255 } else {
4256 real_kind = cand_kind;
4257 fwd_kind = btf_fwd_kind(canon_type);
4258 /* we'd need to resolve base FWD to STRUCT/UNION */
4259 if (fwd_kind == real_kind && canon_id < d->btf->start_id)
4260 d->hypot_adjust_canon = true;
4261 }
4262 return fwd_kind == real_kind;
4263 }
4264
4265 if (cand_kind != canon_kind)
4266 return 0;
4267
4268 switch (cand_kind) {
4269 case BTF_KIND_INT:
4270 return btf_equal_int_tag(cand_type, canon_type);
4271
4272 case BTF_KIND_ENUM:
4273 case BTF_KIND_ENUM64:
4274 return btf_compat_enum(cand_type, canon_type);
4275
4276 case BTF_KIND_FWD:
4277 case BTF_KIND_FLOAT:
4278 return btf_equal_common(cand_type, canon_type);
4279
4280 case BTF_KIND_CONST:
4281 case BTF_KIND_VOLATILE:
4282 case BTF_KIND_RESTRICT:
4283 case BTF_KIND_PTR:
4284 case BTF_KIND_TYPEDEF:
4285 case BTF_KIND_FUNC:
4286 case BTF_KIND_TYPE_TAG:
4287 if (cand_type->info != canon_type->info)
4288 return 0;
4289 return btf_dedup_is_equiv(d, cand_type->type, canon_type->type);
4290
4291 case BTF_KIND_ARRAY: {
4292 const struct btf_array *cand_arr, *canon_arr;
4293
4294 if (!btf_compat_array(cand_type, canon_type))
4295 return 0;
4296 cand_arr = btf_array(cand_type);
4297 canon_arr = btf_array(canon_type);
4298 eq = btf_dedup_is_equiv(d, cand_arr->index_type, canon_arr->index_type);
4299 if (eq <= 0)
4300 return eq;
4301 return btf_dedup_is_equiv(d, cand_arr->type, canon_arr->type);
4302 }
4303
4304 case BTF_KIND_STRUCT:
4305 case BTF_KIND_UNION: {
4306 const struct btf_member *cand_m, *canon_m;
4307 __u16 vlen;
4308
4309 if (!btf_shallow_equal_struct(cand_type, canon_type))
4310 return 0;
4311 vlen = btf_vlen(cand_type);
4312 cand_m = btf_members(cand_type);
4313 canon_m = btf_members(canon_type);
4314 for (i = 0; i < vlen; i++) {
4315 eq = btf_dedup_is_equiv(d, cand_m->type, canon_m->type);
4316 if (eq <= 0)
4317 return eq;
4318 cand_m++;
4319 canon_m++;
4320 }
4321
4322 return 1;
4323 }
4324
4325 case BTF_KIND_FUNC_PROTO: {
4326 const struct btf_param *cand_p, *canon_p;
4327 __u16 vlen;
4328
4329 if (!btf_compat_fnproto(cand_type, canon_type))
4330 return 0;
4331 eq = btf_dedup_is_equiv(d, cand_type->type, canon_type->type);
4332 if (eq <= 0)
4333 return eq;
4334 vlen = btf_vlen(cand_type);
4335 cand_p = btf_params(cand_type);
4336 canon_p = btf_params(canon_type);
4337 for (i = 0; i < vlen; i++) {
4338 eq = btf_dedup_is_equiv(d, cand_p->type, canon_p->type);
4339 if (eq <= 0)
4340 return eq;
4341 cand_p++;
4342 canon_p++;
4343 }
4344 return 1;
4345 }
4346
4347 default:
4348 return -EINVAL;
4349 }
4350 return 0;
4351}
4352
4353/*
4354 * Use hypothetical mapping, produced by successful type graph equivalence
4355 * check, to augment existing struct/union canonical mapping, where possible.
4356 *
4357 * If BTF_KIND_FWD resolution is allowed, this mapping is also used to record
4358 * FWD -> STRUCT/UNION correspondence as well. FWD resolution is bidirectional:
4359 * it doesn't matter if FWD type was part of canonical graph or candidate one,
4360 * we are recording the mapping anyway. As opposed to carefulness required
4361 * for struct/union correspondence mapping (described below), for FWD resolution
4362 * it's not important, as by the time that FWD type (reference type) will be
4363 * deduplicated all structs/unions will be deduped already anyway.
4364 *
4365 * Recording STRUCT/UNION mapping is purely a performance optimization and is
4366 * not required for correctness. It needs to be done carefully to ensure that
4367 * struct/union from candidate's type graph is not mapped into corresponding
4368 * struct/union from canonical type graph that itself hasn't been resolved into
4369 * canonical representative. The only guarantee we have is that canonical
4370 * struct/union was determined as canonical and that won't change. But any
4371 * types referenced through that struct/union fields could have been not yet
4372 * resolved, so in case like that it's too early to establish any kind of
4373 * correspondence between structs/unions.
4374 *
4375 * No canonical correspondence is derived for primitive types (they are already
4376 * deduplicated completely already anyway) or reference types (they rely on
4377 * stability of struct/union canonical relationship for equivalence checks).
4378 */
4379static void btf_dedup_merge_hypot_map(struct btf_dedup *d)
4380{
4381 __u32 canon_type_id, targ_type_id;
4382 __u16 t_kind, c_kind;
4383 __u32 t_id, c_id;
4384 int i;
4385
4386 for (i = 0; i < d->hypot_cnt; i++) {
4387 canon_type_id = d->hypot_list[i];
4388 targ_type_id = d->hypot_map[canon_type_id];
4389 t_id = resolve_type_id(d, targ_type_id);
4390 c_id = resolve_type_id(d, canon_type_id);
4391 t_kind = btf_kind(btf__type_by_id(d->btf, t_id));
4392 c_kind = btf_kind(btf__type_by_id(d->btf, c_id));
4393 /*
4394 * Resolve FWD into STRUCT/UNION.
4395 * It's ok to resolve FWD into STRUCT/UNION that's not yet
4396 * mapped to canonical representative (as opposed to
4397 * STRUCT/UNION <--> STRUCT/UNION mapping logic below), because
4398 * eventually that struct is going to be mapped and all resolved
4399 * FWDs will automatically resolve to correct canonical
4400 * representative. This will happen before ref type deduping,
4401 * which critically depends on stability of these mapping. This
4402 * stability is not a requirement for STRUCT/UNION equivalence
4403 * checks, though.
4404 */
4405
4406 /* if it's the split BTF case, we still need to point base FWD
4407 * to STRUCT/UNION in a split BTF, because FWDs from split BTF
4408 * will be resolved against base FWD. If we don't point base
4409 * canonical FWD to the resolved STRUCT/UNION, then all the
4410 * FWDs in split BTF won't be correctly resolved to a proper
4411 * STRUCT/UNION.
4412 */
4413 if (t_kind != BTF_KIND_FWD && c_kind == BTF_KIND_FWD)
4414 d->map[c_id] = t_id;
4415
4416 /* if graph equivalence determined that we'd need to adjust
4417 * base canonical types, then we need to only point base FWDs
4418 * to STRUCTs/UNIONs and do no more modifications. For all
4419 * other purposes the type graphs were not equivalent.
4420 */
4421 if (d->hypot_adjust_canon)
4422 continue;
4423
4424 if (t_kind == BTF_KIND_FWD && c_kind != BTF_KIND_FWD)
4425 d->map[t_id] = c_id;
4426
4427 if ((t_kind == BTF_KIND_STRUCT || t_kind == BTF_KIND_UNION) &&
4428 c_kind != BTF_KIND_FWD &&
4429 is_type_mapped(d, c_id) &&
4430 !is_type_mapped(d, t_id)) {
4431 /*
4432 * as a perf optimization, we can map struct/union
4433 * that's part of type graph we just verified for
4434 * equivalence. We can do that for struct/union that has
4435 * canonical representative only, though.
4436 */
4437 d->map[t_id] = c_id;
4438 }
4439 }
4440}
4441
4442/*
4443 * Deduplicate struct/union types.
4444 *
4445 * For each struct/union type its type signature hash is calculated, taking
4446 * into account type's name, size, number, order and names of fields, but
4447 * ignoring type ID's referenced from fields, because they might not be deduped
4448 * completely until after reference types deduplication phase. This type hash
4449 * is used to iterate over all potential canonical types, sharing same hash.
4450 * For each canonical candidate we check whether type graphs that they form
4451 * (through referenced types in fields and so on) are equivalent using algorithm
4452 * implemented in `btf_dedup_is_equiv`. If such equivalence is found and
4453 * BTF_KIND_FWD resolution is allowed, then hypothetical mapping
4454 * (btf_dedup->hypot_map) produced by aforementioned type graph equivalence
4455 * algorithm is used to record FWD -> STRUCT/UNION mapping. It's also used to
4456 * potentially map other structs/unions to their canonical representatives,
4457 * if such relationship hasn't yet been established. This speeds up algorithm
4458 * by eliminating some of the duplicate work.
4459 *
4460 * If no matching canonical representative was found, struct/union is marked
4461 * as canonical for itself and is added into btf_dedup->dedup_table hash map
4462 * for further look ups.
4463 */
4464static int btf_dedup_struct_type(struct btf_dedup *d, __u32 type_id)
4465{
4466 struct btf_type *cand_type, *t;
4467 struct hashmap_entry *hash_entry;
4468 /* if we don't find equivalent type, then we are canonical */
4469 __u32 new_id = type_id;
4470 __u16 kind;
4471 long h;
4472
4473 /* already deduped or is in process of deduping (loop detected) */
4474 if (d->map[type_id] <= BTF_MAX_NR_TYPES)
4475 return 0;
4476
4477 t = btf_type_by_id(d->btf, type_id);
4478 kind = btf_kind(t);
4479
4480 if (kind != BTF_KIND_STRUCT && kind != BTF_KIND_UNION)
4481 return 0;
4482
4483 h = btf_hash_struct(t);
4484 for_each_dedup_cand(d, hash_entry, h) {
4485 __u32 cand_id = hash_entry->value;
4486 int eq;
4487
4488 /*
4489 * Even though btf_dedup_is_equiv() checks for
4490 * btf_shallow_equal_struct() internally when checking two
4491 * structs (unions) for equivalence, we need to guard here
4492 * from picking matching FWD type as a dedup candidate.
4493 * This can happen due to hash collision. In such case just
4494 * relying on btf_dedup_is_equiv() would lead to potentially
4495 * creating a loop (FWD -> STRUCT and STRUCT -> FWD), because
4496 * FWD and compatible STRUCT/UNION are considered equivalent.
4497 */
4498 cand_type = btf_type_by_id(d->btf, cand_id);
4499 if (!btf_shallow_equal_struct(t, cand_type))
4500 continue;
4501
4502 btf_dedup_clear_hypot_map(d);
4503 eq = btf_dedup_is_equiv(d, type_id, cand_id);
4504 if (eq < 0)
4505 return eq;
4506 if (!eq)
4507 continue;
4508 btf_dedup_merge_hypot_map(d);
4509 if (d->hypot_adjust_canon) /* not really equivalent */
4510 continue;
4511 new_id = cand_id;
4512 break;
4513 }
4514
4515 d->map[type_id] = new_id;
4516 if (type_id == new_id && btf_dedup_table_add(d, h, type_id))
4517 return -ENOMEM;
4518
4519 return 0;
4520}
4521
4522static int btf_dedup_struct_types(struct btf_dedup *d)
4523{
4524 int i, err;
4525
4526 for (i = 0; i < d->btf->nr_types; i++) {
4527 err = btf_dedup_struct_type(d, d->btf->start_id + i);
4528 if (err)
4529 return err;
4530 }
4531 return 0;
4532}
4533
4534/*
4535 * Deduplicate reference type.
4536 *
4537 * Once all primitive and struct/union types got deduplicated, we can easily
4538 * deduplicate all other (reference) BTF types. This is done in two steps:
4539 *
4540 * 1. Resolve all referenced type IDs into their canonical type IDs. This
4541 * resolution can be done either immediately for primitive or struct/union types
4542 * (because they were deduped in previous two phases) or recursively for
4543 * reference types. Recursion will always terminate at either primitive or
4544 * struct/union type, at which point we can "unwind" chain of reference types
4545 * one by one. There is no danger of encountering cycles because in C type
4546 * system the only way to form type cycle is through struct/union, so any chain
4547 * of reference types, even those taking part in a type cycle, will inevitably
4548 * reach struct/union at some point.
4549 *
4550 * 2. Once all referenced type IDs are resolved into canonical ones, BTF type
4551 * becomes "stable", in the sense that no further deduplication will cause
4552 * any changes to it. With that, it's now possible to calculate type's signature
4553 * hash (this time taking into account referenced type IDs) and loop over all
4554 * potential canonical representatives. If no match was found, current type
4555 * will become canonical representative of itself and will be added into
4556 * btf_dedup->dedup_table as another possible canonical representative.
4557 */
4558static int btf_dedup_ref_type(struct btf_dedup *d, __u32 type_id)
4559{
4560 struct hashmap_entry *hash_entry;
4561 __u32 new_id = type_id, cand_id;
4562 struct btf_type *t, *cand;
4563 /* if we don't find equivalent type, then we are representative type */
4564 int ref_type_id;
4565 long h;
4566
4567 if (d->map[type_id] == BTF_IN_PROGRESS_ID)
4568 return -ELOOP;
4569 if (d->map[type_id] <= BTF_MAX_NR_TYPES)
4570 return resolve_type_id(d, type_id);
4571
4572 t = btf_type_by_id(d->btf, type_id);
4573 d->map[type_id] = BTF_IN_PROGRESS_ID;
4574
4575 switch (btf_kind(t)) {
4576 case BTF_KIND_CONST:
4577 case BTF_KIND_VOLATILE:
4578 case BTF_KIND_RESTRICT:
4579 case BTF_KIND_PTR:
4580 case BTF_KIND_TYPEDEF:
4581 case BTF_KIND_FUNC:
4582 case BTF_KIND_TYPE_TAG:
4583 ref_type_id = btf_dedup_ref_type(d, t->type);
4584 if (ref_type_id < 0)
4585 return ref_type_id;
4586 t->type = ref_type_id;
4587
4588 h = btf_hash_common(t);
4589 for_each_dedup_cand(d, hash_entry, h) {
4590 cand_id = hash_entry->value;
4591 cand = btf_type_by_id(d->btf, cand_id);
4592 if (btf_equal_common(t, cand)) {
4593 new_id = cand_id;
4594 break;
4595 }
4596 }
4597 break;
4598
4599 case BTF_KIND_DECL_TAG:
4600 ref_type_id = btf_dedup_ref_type(d, t->type);
4601 if (ref_type_id < 0)
4602 return ref_type_id;
4603 t->type = ref_type_id;
4604
4605 h = btf_hash_int_decl_tag(t);
4606 for_each_dedup_cand(d, hash_entry, h) {
4607 cand_id = hash_entry->value;
4608 cand = btf_type_by_id(d->btf, cand_id);
4609 if (btf_equal_int_tag(t, cand)) {
4610 new_id = cand_id;
4611 break;
4612 }
4613 }
4614 break;
4615
4616 case BTF_KIND_ARRAY: {
4617 struct btf_array *info = btf_array(t);
4618
4619 ref_type_id = btf_dedup_ref_type(d, info->type);
4620 if (ref_type_id < 0)
4621 return ref_type_id;
4622 info->type = ref_type_id;
4623
4624 ref_type_id = btf_dedup_ref_type(d, info->index_type);
4625 if (ref_type_id < 0)
4626 return ref_type_id;
4627 info->index_type = ref_type_id;
4628
4629 h = btf_hash_array(t);
4630 for_each_dedup_cand(d, hash_entry, h) {
4631 cand_id = hash_entry->value;
4632 cand = btf_type_by_id(d->btf, cand_id);
4633 if (btf_equal_array(t, cand)) {
4634 new_id = cand_id;
4635 break;
4636 }
4637 }
4638 break;
4639 }
4640
4641 case BTF_KIND_FUNC_PROTO: {
4642 struct btf_param *param;
4643 __u16 vlen;
4644 int i;
4645
4646 ref_type_id = btf_dedup_ref_type(d, t->type);
4647 if (ref_type_id < 0)
4648 return ref_type_id;
4649 t->type = ref_type_id;
4650
4651 vlen = btf_vlen(t);
4652 param = btf_params(t);
4653 for (i = 0; i < vlen; i++) {
4654 ref_type_id = btf_dedup_ref_type(d, param->type);
4655 if (ref_type_id < 0)
4656 return ref_type_id;
4657 param->type = ref_type_id;
4658 param++;
4659 }
4660
4661 h = btf_hash_fnproto(t);
4662 for_each_dedup_cand(d, hash_entry, h) {
4663 cand_id = hash_entry->value;
4664 cand = btf_type_by_id(d->btf, cand_id);
4665 if (btf_equal_fnproto(t, cand)) {
4666 new_id = cand_id;
4667 break;
4668 }
4669 }
4670 break;
4671 }
4672
4673 default:
4674 return -EINVAL;
4675 }
4676
4677 d->map[type_id] = new_id;
4678 if (type_id == new_id && btf_dedup_table_add(d, h, type_id))
4679 return -ENOMEM;
4680
4681 return new_id;
4682}
4683
4684static int btf_dedup_ref_types(struct btf_dedup *d)
4685{
4686 int i, err;
4687
4688 for (i = 0; i < d->btf->nr_types; i++) {
4689 err = btf_dedup_ref_type(d, d->btf->start_id + i);
4690 if (err < 0)
4691 return err;
4692 }
4693 /* we won't need d->dedup_table anymore */
4694 hashmap__free(d->dedup_table);
4695 d->dedup_table = NULL;
4696 return 0;
4697}
4698
4699/*
4700 * Collect a map from type names to type ids for all canonical structs
4701 * and unions. If the same name is shared by several canonical types
4702 * use a special value 0 to indicate this fact.
4703 */
4704static int btf_dedup_fill_unique_names_map(struct btf_dedup *d, struct hashmap *names_map)
4705{
4706 __u32 nr_types = btf__type_cnt(d->btf);
4707 struct btf_type *t;
4708 __u32 type_id;
4709 __u16 kind;
4710 int err;
4711
4712 /*
4713 * Iterate over base and split module ids in order to get all
4714 * available structs in the map.
4715 */
4716 for (type_id = 1; type_id < nr_types; ++type_id) {
4717 t = btf_type_by_id(d->btf, type_id);
4718 kind = btf_kind(t);
4719
4720 if (kind != BTF_KIND_STRUCT && kind != BTF_KIND_UNION)
4721 continue;
4722
4723 /* Skip non-canonical types */
4724 if (type_id != d->map[type_id])
4725 continue;
4726
4727 err = hashmap__add(names_map, t->name_off, type_id);
4728 if (err == -EEXIST)
4729 err = hashmap__set(names_map, t->name_off, 0, NULL, NULL);
4730
4731 if (err)
4732 return err;
4733 }
4734
4735 return 0;
4736}
4737
4738static int btf_dedup_resolve_fwd(struct btf_dedup *d, struct hashmap *names_map, __u32 type_id)
4739{
4740 struct btf_type *t = btf_type_by_id(d->btf, type_id);
4741 enum btf_fwd_kind fwd_kind = btf_kflag(t);
4742 __u16 cand_kind, kind = btf_kind(t);
4743 struct btf_type *cand_t;
4744 uintptr_t cand_id;
4745
4746 if (kind != BTF_KIND_FWD)
4747 return 0;
4748
4749 /* Skip if this FWD already has a mapping */
4750 if (type_id != d->map[type_id])
4751 return 0;
4752
4753 if (!hashmap__find(names_map, t->name_off, &cand_id))
4754 return 0;
4755
4756 /* Zero is a special value indicating that name is not unique */
4757 if (!cand_id)
4758 return 0;
4759
4760 cand_t = btf_type_by_id(d->btf, cand_id);
4761 cand_kind = btf_kind(cand_t);
4762 if ((cand_kind == BTF_KIND_STRUCT && fwd_kind != BTF_FWD_STRUCT) ||
4763 (cand_kind == BTF_KIND_UNION && fwd_kind != BTF_FWD_UNION))
4764 return 0;
4765
4766 d->map[type_id] = cand_id;
4767
4768 return 0;
4769}
4770
4771/*
4772 * Resolve unambiguous forward declarations.
4773 *
4774 * The lion's share of all FWD declarations is resolved during
4775 * `btf_dedup_struct_types` phase when different type graphs are
4776 * compared against each other. However, if in some compilation unit a
4777 * FWD declaration is not a part of a type graph compared against
4778 * another type graph that declaration's canonical type would not be
4779 * changed. Example:
4780 *
4781 * CU #1:
4782 *
4783 * struct foo;
4784 * struct foo *some_global;
4785 *
4786 * CU #2:
4787 *
4788 * struct foo { int u; };
4789 * struct foo *another_global;
4790 *
4791 * After `btf_dedup_struct_types` the BTF looks as follows:
4792 *
4793 * [1] STRUCT 'foo' size=4 vlen=1 ...
4794 * [2] INT 'int' size=4 ...
4795 * [3] PTR '(anon)' type_id=1
4796 * [4] FWD 'foo' fwd_kind=struct
4797 * [5] PTR '(anon)' type_id=4
4798 *
4799 * This pass assumes that such FWD declarations should be mapped to
4800 * structs or unions with identical name in case if the name is not
4801 * ambiguous.
4802 */
4803static int btf_dedup_resolve_fwds(struct btf_dedup *d)
4804{
4805 int i, err;
4806 struct hashmap *names_map;
4807
4808 names_map = hashmap__new(btf_dedup_identity_hash_fn, btf_dedup_equal_fn, NULL);
4809 if (IS_ERR(names_map))
4810 return PTR_ERR(names_map);
4811
4812 err = btf_dedup_fill_unique_names_map(d, names_map);
4813 if (err < 0)
4814 goto exit;
4815
4816 for (i = 0; i < d->btf->nr_types; i++) {
4817 err = btf_dedup_resolve_fwd(d, names_map, d->btf->start_id + i);
4818 if (err < 0)
4819 break;
4820 }
4821
4822exit:
4823 hashmap__free(names_map);
4824 return err;
4825}
4826
4827/*
4828 * Compact types.
4829 *
4830 * After we established for each type its corresponding canonical representative
4831 * type, we now can eliminate types that are not canonical and leave only
4832 * canonical ones layed out sequentially in memory by copying them over
4833 * duplicates. During compaction btf_dedup->hypot_map array is reused to store
4834 * a map from original type ID to a new compacted type ID, which will be used
4835 * during next phase to "fix up" type IDs, referenced from struct/union and
4836 * reference types.
4837 */
4838static int btf_dedup_compact_types(struct btf_dedup *d)
4839{
4840 __u32 *new_offs;
4841 __u32 next_type_id = d->btf->start_id;
4842 const struct btf_type *t;
4843 void *p;
4844 int i, id, len;
4845
4846 /* we are going to reuse hypot_map to store compaction remapping */
4847 d->hypot_map[0] = 0;
4848 /* base BTF types are not renumbered */
4849 for (id = 1; id < d->btf->start_id; id++)
4850 d->hypot_map[id] = id;
4851 for (i = 0, id = d->btf->start_id; i < d->btf->nr_types; i++, id++)
4852 d->hypot_map[id] = BTF_UNPROCESSED_ID;
4853
4854 p = d->btf->types_data;
4855
4856 for (i = 0, id = d->btf->start_id; i < d->btf->nr_types; i++, id++) {
4857 if (d->map[id] != id)
4858 continue;
4859
4860 t = btf__type_by_id(d->btf, id);
4861 len = btf_type_size(t);
4862 if (len < 0)
4863 return len;
4864
4865 memmove(p, t, len);
4866 d->hypot_map[id] = next_type_id;
4867 d->btf->type_offs[next_type_id - d->btf->start_id] = p - d->btf->types_data;
4868 p += len;
4869 next_type_id++;
4870 }
4871
4872 /* shrink struct btf's internal types index and update btf_header */
4873 d->btf->nr_types = next_type_id - d->btf->start_id;
4874 d->btf->type_offs_cap = d->btf->nr_types;
4875 d->btf->hdr->type_len = p - d->btf->types_data;
4876 new_offs = libbpf_reallocarray(d->btf->type_offs, d->btf->type_offs_cap,
4877 sizeof(*new_offs));
4878 if (d->btf->type_offs_cap && !new_offs)
4879 return -ENOMEM;
4880 d->btf->type_offs = new_offs;
4881 d->btf->hdr->str_off = d->btf->hdr->type_len;
4882 d->btf->raw_size = d->btf->hdr->hdr_len + d->btf->hdr->type_len + d->btf->hdr->str_len;
4883 return 0;
4884}
4885
4886/*
4887 * Figure out final (deduplicated and compacted) type ID for provided original
4888 * `type_id` by first resolving it into corresponding canonical type ID and
4889 * then mapping it to a deduplicated type ID, stored in btf_dedup->hypot_map,
4890 * which is populated during compaction phase.
4891 */
4892static int btf_dedup_remap_type_id(__u32 *type_id, void *ctx)
4893{
4894 struct btf_dedup *d = ctx;
4895 __u32 resolved_type_id, new_type_id;
4896
4897 resolved_type_id = resolve_type_id(d, *type_id);
4898 new_type_id = d->hypot_map[resolved_type_id];
4899 if (new_type_id > BTF_MAX_NR_TYPES)
4900 return -EINVAL;
4901
4902 *type_id = new_type_id;
4903 return 0;
4904}
4905
4906/*
4907 * Remap referenced type IDs into deduped type IDs.
4908 *
4909 * After BTF types are deduplicated and compacted, their final type IDs may
4910 * differ from original ones. The map from original to a corresponding
4911 * deduped type ID is stored in btf_dedup->hypot_map and is populated during
4912 * compaction phase. During remapping phase we are rewriting all type IDs
4913 * referenced from any BTF type (e.g., struct fields, func proto args, etc) to
4914 * their final deduped type IDs.
4915 */
4916static int btf_dedup_remap_types(struct btf_dedup *d)
4917{
4918 int i, r;
4919
4920 for (i = 0; i < d->btf->nr_types; i++) {
4921 struct btf_type *t = btf_type_by_id(d->btf, d->btf->start_id + i);
4922
4923 r = btf_type_visit_type_ids(t, btf_dedup_remap_type_id, d);
4924 if (r)
4925 return r;
4926 }
4927
4928 if (!d->btf_ext)
4929 return 0;
4930
4931 r = btf_ext_visit_type_ids(d->btf_ext, btf_dedup_remap_type_id, d);
4932 if (r)
4933 return r;
4934
4935 return 0;
4936}
4937
4938/*
4939 * Probe few well-known locations for vmlinux kernel image and try to load BTF
4940 * data out of it to use for target BTF.
4941 */
4942struct btf *btf__load_vmlinux_btf(void)
4943{
4944 const char *sysfs_btf_path = "/sys/kernel/btf/vmlinux";
4945 /* fall back locations, trying to find vmlinux on disk */
4946 const char *locations[] = {
4947 "/boot/vmlinux-%1$s",
4948 "/lib/modules/%1$s/vmlinux-%1$s",
4949 "/lib/modules/%1$s/build/vmlinux",
4950 "/usr/lib/modules/%1$s/kernel/vmlinux",
4951 "/usr/lib/debug/boot/vmlinux-%1$s",
4952 "/usr/lib/debug/boot/vmlinux-%1$s.debug",
4953 "/usr/lib/debug/lib/modules/%1$s/vmlinux",
4954 };
4955 char path[PATH_MAX + 1];
4956 struct utsname buf;
4957 struct btf *btf;
4958 int i, err;
4959
4960 /* is canonical sysfs location accessible? */
4961 if (faccessat(AT_FDCWD, sysfs_btf_path, F_OK, AT_EACCESS) < 0) {
4962 pr_warn("kernel BTF is missing at '%s', was CONFIG_DEBUG_INFO_BTF enabled?\n",
4963 sysfs_btf_path);
4964 } else {
4965 btf = btf__parse(sysfs_btf_path, NULL);
4966 if (!btf) {
4967 err = -errno;
4968 pr_warn("failed to read kernel BTF from '%s': %d\n", sysfs_btf_path, err);
4969 return libbpf_err_ptr(err);
4970 }
4971 pr_debug("loaded kernel BTF from '%s'\n", sysfs_btf_path);
4972 return btf;
4973 }
4974
4975 /* try fallback locations */
4976 uname(&buf);
4977 for (i = 0; i < ARRAY_SIZE(locations); i++) {
4978 snprintf(path, PATH_MAX, locations[i], buf.release);
4979
4980 if (faccessat(AT_FDCWD, path, R_OK, AT_EACCESS))
4981 continue;
4982
4983 btf = btf__parse(path, NULL);
4984 err = libbpf_get_error(btf);
4985 pr_debug("loading kernel BTF '%s': %d\n", path, err);
4986 if (err)
4987 continue;
4988
4989 return btf;
4990 }
4991
4992 pr_warn("failed to find valid kernel BTF\n");
4993 return libbpf_err_ptr(-ESRCH);
4994}
4995
4996struct btf *libbpf_find_kernel_btf(void) __attribute__((alias("btf__load_vmlinux_btf")));
4997
4998struct btf *btf__load_module_btf(const char *module_name, struct btf *vmlinux_btf)
4999{
5000 char path[80];
5001
5002 snprintf(path, sizeof(path), "/sys/kernel/btf/%s", module_name);
5003 return btf__parse_split(path, vmlinux_btf);
5004}
5005
5006int btf_type_visit_type_ids(struct btf_type *t, type_id_visit_fn visit, void *ctx)
5007{
5008 int i, n, err;
5009
5010 switch (btf_kind(t)) {
5011 case BTF_KIND_INT:
5012 case BTF_KIND_FLOAT:
5013 case BTF_KIND_ENUM:
5014 case BTF_KIND_ENUM64:
5015 return 0;
5016
5017 case BTF_KIND_FWD:
5018 case BTF_KIND_CONST:
5019 case BTF_KIND_VOLATILE:
5020 case BTF_KIND_RESTRICT:
5021 case BTF_KIND_PTR:
5022 case BTF_KIND_TYPEDEF:
5023 case BTF_KIND_FUNC:
5024 case BTF_KIND_VAR:
5025 case BTF_KIND_DECL_TAG:
5026 case BTF_KIND_TYPE_TAG:
5027 return visit(&t->type, ctx);
5028
5029 case BTF_KIND_ARRAY: {
5030 struct btf_array *a = btf_array(t);
5031
5032 err = visit(&a->type, ctx);
5033 err = err ?: visit(&a->index_type, ctx);
5034 return err;
5035 }
5036
5037 case BTF_KIND_STRUCT:
5038 case BTF_KIND_UNION: {
5039 struct btf_member *m = btf_members(t);
5040
5041 for (i = 0, n = btf_vlen(t); i < n; i++, m++) {
5042 err = visit(&m->type, ctx);
5043 if (err)
5044 return err;
5045 }
5046 return 0;
5047 }
5048
5049 case BTF_KIND_FUNC_PROTO: {
5050 struct btf_param *m = btf_params(t);
5051
5052 err = visit(&t->type, ctx);
5053 if (err)
5054 return err;
5055 for (i = 0, n = btf_vlen(t); i < n; i++, m++) {
5056 err = visit(&m->type, ctx);
5057 if (err)
5058 return err;
5059 }
5060 return 0;
5061 }
5062
5063 case BTF_KIND_DATASEC: {
5064 struct btf_var_secinfo *m = btf_var_secinfos(t);
5065
5066 for (i = 0, n = btf_vlen(t); i < n; i++, m++) {
5067 err = visit(&m->type, ctx);
5068 if (err)
5069 return err;
5070 }
5071 return 0;
5072 }
5073
5074 default:
5075 return -EINVAL;
5076 }
5077}
5078
5079int btf_type_visit_str_offs(struct btf_type *t, str_off_visit_fn visit, void *ctx)
5080{
5081 int i, n, err;
5082
5083 err = visit(&t->name_off, ctx);
5084 if (err)
5085 return err;
5086
5087 switch (btf_kind(t)) {
5088 case BTF_KIND_STRUCT:
5089 case BTF_KIND_UNION: {
5090 struct btf_member *m = btf_members(t);
5091
5092 for (i = 0, n = btf_vlen(t); i < n; i++, m++) {
5093 err = visit(&m->name_off, ctx);
5094 if (err)
5095 return err;
5096 }
5097 break;
5098 }
5099 case BTF_KIND_ENUM: {
5100 struct btf_enum *m = btf_enum(t);
5101
5102 for (i = 0, n = btf_vlen(t); i < n; i++, m++) {
5103 err = visit(&m->name_off, ctx);
5104 if (err)
5105 return err;
5106 }
5107 break;
5108 }
5109 case BTF_KIND_ENUM64: {
5110 struct btf_enum64 *m = btf_enum64(t);
5111
5112 for (i = 0, n = btf_vlen(t); i < n; i++, m++) {
5113 err = visit(&m->name_off, ctx);
5114 if (err)
5115 return err;
5116 }
5117 break;
5118 }
5119 case BTF_KIND_FUNC_PROTO: {
5120 struct btf_param *m = btf_params(t);
5121
5122 for (i = 0, n = btf_vlen(t); i < n; i++, m++) {
5123 err = visit(&m->name_off, ctx);
5124 if (err)
5125 return err;
5126 }
5127 break;
5128 }
5129 default:
5130 break;
5131 }
5132
5133 return 0;
5134}
5135
5136int btf_ext_visit_type_ids(struct btf_ext *btf_ext, type_id_visit_fn visit, void *ctx)
5137{
5138 const struct btf_ext_info *seg;
5139 struct btf_ext_info_sec *sec;
5140 int i, err;
5141
5142 seg = &btf_ext->func_info;
5143 for_each_btf_ext_sec(seg, sec) {
5144 struct bpf_func_info_min *rec;
5145
5146 for_each_btf_ext_rec(seg, sec, i, rec) {
5147 err = visit(&rec->type_id, ctx);
5148 if (err < 0)
5149 return err;
5150 }
5151 }
5152
5153 seg = &btf_ext->core_relo_info;
5154 for_each_btf_ext_sec(seg, sec) {
5155 struct bpf_core_relo *rec;
5156
5157 for_each_btf_ext_rec(seg, sec, i, rec) {
5158 err = visit(&rec->type_id, ctx);
5159 if (err < 0)
5160 return err;
5161 }
5162 }
5163
5164 return 0;
5165}
5166
5167int btf_ext_visit_str_offs(struct btf_ext *btf_ext, str_off_visit_fn visit, void *ctx)
5168{
5169 const struct btf_ext_info *seg;
5170 struct btf_ext_info_sec *sec;
5171 int i, err;
5172
5173 seg = &btf_ext->func_info;
5174 for_each_btf_ext_sec(seg, sec) {
5175 err = visit(&sec->sec_name_off, ctx);
5176 if (err)
5177 return err;
5178 }
5179
5180 seg = &btf_ext->line_info;
5181 for_each_btf_ext_sec(seg, sec) {
5182 struct bpf_line_info_min *rec;
5183
5184 err = visit(&sec->sec_name_off, ctx);
5185 if (err)
5186 return err;
5187
5188 for_each_btf_ext_rec(seg, sec, i, rec) {
5189 err = visit(&rec->file_name_off, ctx);
5190 if (err)
5191 return err;
5192 err = visit(&rec->line_off, ctx);
5193 if (err)
5194 return err;
5195 }
5196 }
5197
5198 seg = &btf_ext->core_relo_info;
5199 for_each_btf_ext_sec(seg, sec) {
5200 struct bpf_core_relo *rec;
5201
5202 err = visit(&sec->sec_name_off, ctx);
5203 if (err)
5204 return err;
5205
5206 for_each_btf_ext_rec(seg, sec, i, rec) {
5207 err = visit(&rec->access_str_off, ctx);
5208 if (err)
5209 return err;
5210 }
5211 }
5212
5213 return 0;
5214}
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}