Loading...
Note: File does not exist in v3.15.
1// SPDX-License-Identifier: (LGPL-2.1 OR BSD-2-Clause)
2
3/*
4 * BTF-to-C type converter.
5 *
6 * Copyright (c) 2019 Facebook
7 */
8
9#include <stdbool.h>
10#include <stddef.h>
11#include <stdlib.h>
12#include <string.h>
13#include <errno.h>
14#include <linux/err.h>
15#include <linux/btf.h>
16#include "btf.h"
17#include "hashmap.h"
18#include "libbpf.h"
19#include "libbpf_internal.h"
20
21static const char PREFIXES[] = "\t\t\t\t\t\t\t\t\t\t\t\t\t";
22static const size_t PREFIX_CNT = sizeof(PREFIXES) - 1;
23
24static const char *pfx(int lvl)
25{
26 return lvl >= PREFIX_CNT ? PREFIXES : &PREFIXES[PREFIX_CNT - lvl];
27}
28
29enum btf_dump_type_order_state {
30 NOT_ORDERED,
31 ORDERING,
32 ORDERED,
33};
34
35enum btf_dump_type_emit_state {
36 NOT_EMITTED,
37 EMITTING,
38 EMITTED,
39};
40
41/* per-type auxiliary state */
42struct btf_dump_type_aux_state {
43 /* topological sorting state */
44 enum btf_dump_type_order_state order_state: 2;
45 /* emitting state used to determine the need for forward declaration */
46 enum btf_dump_type_emit_state emit_state: 2;
47 /* whether forward declaration was already emitted */
48 __u8 fwd_emitted: 1;
49 /* whether unique non-duplicate name was already assigned */
50 __u8 name_resolved: 1;
51 /* whether type is referenced from any other type */
52 __u8 referenced: 1;
53};
54
55struct btf_dump {
56 const struct btf *btf;
57 const struct btf_ext *btf_ext;
58 btf_dump_printf_fn_t printf_fn;
59 struct btf_dump_opts opts;
60
61 /* per-type auxiliary state */
62 struct btf_dump_type_aux_state *type_states;
63 /* per-type optional cached unique name, must be freed, if present */
64 const char **cached_names;
65
66 /* topo-sorted list of dependent type definitions */
67 __u32 *emit_queue;
68 int emit_queue_cap;
69 int emit_queue_cnt;
70
71 /*
72 * stack of type declarations (e.g., chain of modifiers, arrays,
73 * funcs, etc)
74 */
75 __u32 *decl_stack;
76 int decl_stack_cap;
77 int decl_stack_cnt;
78
79 /* maps struct/union/enum name to a number of name occurrences */
80 struct hashmap *type_names;
81 /*
82 * maps typedef identifiers and enum value names to a number of such
83 * name occurrences
84 */
85 struct hashmap *ident_names;
86};
87
88static size_t str_hash_fn(const void *key, void *ctx)
89{
90 const char *s = key;
91 size_t h = 0;
92
93 while (*s) {
94 h = h * 31 + *s;
95 s++;
96 }
97 return h;
98}
99
100static bool str_equal_fn(const void *a, const void *b, void *ctx)
101{
102 return strcmp(a, b) == 0;
103}
104
105static const char *btf_name_of(const struct btf_dump *d, __u32 name_off)
106{
107 return btf__name_by_offset(d->btf, name_off);
108}
109
110static void btf_dump_printf(const struct btf_dump *d, const char *fmt, ...)
111{
112 va_list args;
113
114 va_start(args, fmt);
115 d->printf_fn(d->opts.ctx, fmt, args);
116 va_end(args);
117}
118
119struct btf_dump *btf_dump__new(const struct btf *btf,
120 const struct btf_ext *btf_ext,
121 const struct btf_dump_opts *opts,
122 btf_dump_printf_fn_t printf_fn)
123{
124 struct btf_dump *d;
125 int err;
126
127 d = calloc(1, sizeof(struct btf_dump));
128 if (!d)
129 return ERR_PTR(-ENOMEM);
130
131 d->btf = btf;
132 d->btf_ext = btf_ext;
133 d->printf_fn = printf_fn;
134 d->opts.ctx = opts ? opts->ctx : NULL;
135
136 d->type_names = hashmap__new(str_hash_fn, str_equal_fn, NULL);
137 if (IS_ERR(d->type_names)) {
138 err = PTR_ERR(d->type_names);
139 d->type_names = NULL;
140 btf_dump__free(d);
141 return ERR_PTR(err);
142 }
143 d->ident_names = hashmap__new(str_hash_fn, str_equal_fn, NULL);
144 if (IS_ERR(d->ident_names)) {
145 err = PTR_ERR(d->ident_names);
146 d->ident_names = NULL;
147 btf_dump__free(d);
148 return ERR_PTR(err);
149 }
150
151 return d;
152}
153
154void btf_dump__free(struct btf_dump *d)
155{
156 int i, cnt;
157
158 if (!d)
159 return;
160
161 free(d->type_states);
162 if (d->cached_names) {
163 /* any set cached name is owned by us and should be freed */
164 for (i = 0, cnt = btf__get_nr_types(d->btf); i <= cnt; i++) {
165 if (d->cached_names[i])
166 free((void *)d->cached_names[i]);
167 }
168 }
169 free(d->cached_names);
170 free(d->emit_queue);
171 free(d->decl_stack);
172 hashmap__free(d->type_names);
173 hashmap__free(d->ident_names);
174
175 free(d);
176}
177
178static int btf_dump_mark_referenced(struct btf_dump *d);
179static int btf_dump_order_type(struct btf_dump *d, __u32 id, bool through_ptr);
180static void btf_dump_emit_type(struct btf_dump *d, __u32 id, __u32 cont_id);
181
182/*
183 * Dump BTF type in a compilable C syntax, including all the necessary
184 * dependent types, necessary for compilation. If some of the dependent types
185 * were already emitted as part of previous btf_dump__dump_type() invocation
186 * for another type, they won't be emitted again. This API allows callers to
187 * filter out BTF types according to user-defined criterias and emitted only
188 * minimal subset of types, necessary to compile everything. Full struct/union
189 * definitions will still be emitted, even if the only usage is through
190 * pointer and could be satisfied with just a forward declaration.
191 *
192 * Dumping is done in two high-level passes:
193 * 1. Topologically sort type definitions to satisfy C rules of compilation.
194 * 2. Emit type definitions in C syntax.
195 *
196 * Returns 0 on success; <0, otherwise.
197 */
198int btf_dump__dump_type(struct btf_dump *d, __u32 id)
199{
200 int err, i;
201
202 if (id > btf__get_nr_types(d->btf))
203 return -EINVAL;
204
205 /* type states are lazily allocated, as they might not be needed */
206 if (!d->type_states) {
207 d->type_states = calloc(1 + btf__get_nr_types(d->btf),
208 sizeof(d->type_states[0]));
209 if (!d->type_states)
210 return -ENOMEM;
211 d->cached_names = calloc(1 + btf__get_nr_types(d->btf),
212 sizeof(d->cached_names[0]));
213 if (!d->cached_names)
214 return -ENOMEM;
215
216 /* VOID is special */
217 d->type_states[0].order_state = ORDERED;
218 d->type_states[0].emit_state = EMITTED;
219
220 /* eagerly determine referenced types for anon enums */
221 err = btf_dump_mark_referenced(d);
222 if (err)
223 return err;
224 }
225
226 d->emit_queue_cnt = 0;
227 err = btf_dump_order_type(d, id, false);
228 if (err < 0)
229 return err;
230
231 for (i = 0; i < d->emit_queue_cnt; i++)
232 btf_dump_emit_type(d, d->emit_queue[i], 0 /*top-level*/);
233
234 return 0;
235}
236
237/*
238 * Mark all types that are referenced from any other type. This is used to
239 * determine top-level anonymous enums that need to be emitted as an
240 * independent type declarations.
241 * Anonymous enums come in two flavors: either embedded in a struct's field
242 * definition, in which case they have to be declared inline as part of field
243 * type declaration; or as a top-level anonymous enum, typically used for
244 * declaring global constants. It's impossible to distinguish between two
245 * without knowning whether given enum type was referenced from other type:
246 * top-level anonymous enum won't be referenced by anything, while embedded
247 * one will.
248 */
249static int btf_dump_mark_referenced(struct btf_dump *d)
250{
251 int i, j, n = btf__get_nr_types(d->btf);
252 const struct btf_type *t;
253 __u16 vlen;
254
255 for (i = 1; i <= n; i++) {
256 t = btf__type_by_id(d->btf, i);
257 vlen = btf_vlen(t);
258
259 switch (btf_kind(t)) {
260 case BTF_KIND_INT:
261 case BTF_KIND_ENUM:
262 case BTF_KIND_FWD:
263 break;
264
265 case BTF_KIND_VOLATILE:
266 case BTF_KIND_CONST:
267 case BTF_KIND_RESTRICT:
268 case BTF_KIND_PTR:
269 case BTF_KIND_TYPEDEF:
270 case BTF_KIND_FUNC:
271 case BTF_KIND_VAR:
272 d->type_states[t->type].referenced = 1;
273 break;
274
275 case BTF_KIND_ARRAY: {
276 const struct btf_array *a = btf_array(t);
277
278 d->type_states[a->index_type].referenced = 1;
279 d->type_states[a->type].referenced = 1;
280 break;
281 }
282 case BTF_KIND_STRUCT:
283 case BTF_KIND_UNION: {
284 const struct btf_member *m = btf_members(t);
285
286 for (j = 0; j < vlen; j++, m++)
287 d->type_states[m->type].referenced = 1;
288 break;
289 }
290 case BTF_KIND_FUNC_PROTO: {
291 const struct btf_param *p = btf_params(t);
292
293 for (j = 0; j < vlen; j++, p++)
294 d->type_states[p->type].referenced = 1;
295 break;
296 }
297 case BTF_KIND_DATASEC: {
298 const struct btf_var_secinfo *v = btf_var_secinfos(t);
299
300 for (j = 0; j < vlen; j++, v++)
301 d->type_states[v->type].referenced = 1;
302 break;
303 }
304 default:
305 return -EINVAL;
306 }
307 }
308 return 0;
309}
310static int btf_dump_add_emit_queue_id(struct btf_dump *d, __u32 id)
311{
312 __u32 *new_queue;
313 size_t new_cap;
314
315 if (d->emit_queue_cnt >= d->emit_queue_cap) {
316 new_cap = max(16, d->emit_queue_cap * 3 / 2);
317 new_queue = realloc(d->emit_queue,
318 new_cap * sizeof(new_queue[0]));
319 if (!new_queue)
320 return -ENOMEM;
321 d->emit_queue = new_queue;
322 d->emit_queue_cap = new_cap;
323 }
324
325 d->emit_queue[d->emit_queue_cnt++] = id;
326 return 0;
327}
328
329/*
330 * Determine order of emitting dependent types and specified type to satisfy
331 * C compilation rules. This is done through topological sorting with an
332 * additional complication which comes from C rules. The main idea for C is
333 * that if some type is "embedded" into a struct/union, it's size needs to be
334 * known at the time of definition of containing type. E.g., for:
335 *
336 * struct A {};
337 * struct B { struct A x; }
338 *
339 * struct A *HAS* to be defined before struct B, because it's "embedded",
340 * i.e., it is part of struct B layout. But in the following case:
341 *
342 * struct A;
343 * struct B { struct A *x; }
344 * struct A {};
345 *
346 * it's enough to just have a forward declaration of struct A at the time of
347 * struct B definition, as struct B has a pointer to struct A, so the size of
348 * field x is known without knowing struct A size: it's sizeof(void *).
349 *
350 * Unfortunately, there are some trickier cases we need to handle, e.g.:
351 *
352 * struct A {}; // if this was forward-declaration: compilation error
353 * struct B {
354 * struct { // anonymous struct
355 * struct A y;
356 * } *x;
357 * };
358 *
359 * In this case, struct B's field x is a pointer, so it's size is known
360 * regardless of the size of (anonymous) struct it points to. But because this
361 * struct is anonymous and thus defined inline inside struct B, *and* it
362 * embeds struct A, compiler requires full definition of struct A to be known
363 * before struct B can be defined. This creates a transitive dependency
364 * between struct A and struct B. If struct A was forward-declared before
365 * struct B definition and fully defined after struct B definition, that would
366 * trigger compilation error.
367 *
368 * All this means that while we are doing topological sorting on BTF type
369 * graph, we need to determine relationships between different types (graph
370 * nodes):
371 * - weak link (relationship) between X and Y, if Y *CAN* be
372 * forward-declared at the point of X definition;
373 * - strong link, if Y *HAS* to be fully-defined before X can be defined.
374 *
375 * The rule is as follows. Given a chain of BTF types from X to Y, if there is
376 * BTF_KIND_PTR type in the chain and at least one non-anonymous type
377 * Z (excluding X, including Y), then link is weak. Otherwise, it's strong.
378 * Weak/strong relationship is determined recursively during DFS traversal and
379 * is returned as a result from btf_dump_order_type().
380 *
381 * btf_dump_order_type() is trying to avoid unnecessary forward declarations,
382 * but it is not guaranteeing that no extraneous forward declarations will be
383 * emitted.
384 *
385 * To avoid extra work, algorithm marks some of BTF types as ORDERED, when
386 * it's done with them, but not for all (e.g., VOLATILE, CONST, RESTRICT,
387 * ARRAY, FUNC_PROTO), as weak/strong semantics for those depends on the
388 * entire graph path, so depending where from one came to that BTF type, it
389 * might cause weak or strong ordering. For types like STRUCT/UNION/INT/ENUM,
390 * once they are processed, there is no need to do it again, so they are
391 * marked as ORDERED. We can mark PTR as ORDERED as well, as it semi-forces
392 * weak link, unless subsequent referenced STRUCT/UNION/ENUM is anonymous. But
393 * in any case, once those are processed, no need to do it again, as the
394 * result won't change.
395 *
396 * Returns:
397 * - 1, if type is part of strong link (so there is strong topological
398 * ordering requirements);
399 * - 0, if type is part of weak link (so can be satisfied through forward
400 * declaration);
401 * - <0, on error (e.g., unsatisfiable type loop detected).
402 */
403static int btf_dump_order_type(struct btf_dump *d, __u32 id, bool through_ptr)
404{
405 /*
406 * Order state is used to detect strong link cycles, but only for BTF
407 * kinds that are or could be an independent definition (i.e.,
408 * stand-alone fwd decl, enum, typedef, struct, union). Ptrs, arrays,
409 * func_protos, modifiers are just means to get to these definitions.
410 * Int/void don't need definitions, they are assumed to be always
411 * properly defined. We also ignore datasec, var, and funcs for now.
412 * So for all non-defining kinds, we never even set ordering state,
413 * for defining kinds we set ORDERING and subsequently ORDERED if it
414 * forms a strong link.
415 */
416 struct btf_dump_type_aux_state *tstate = &d->type_states[id];
417 const struct btf_type *t;
418 __u16 vlen;
419 int err, i;
420
421 /* return true, letting typedefs know that it's ok to be emitted */
422 if (tstate->order_state == ORDERED)
423 return 1;
424
425 t = btf__type_by_id(d->btf, id);
426
427 if (tstate->order_state == ORDERING) {
428 /* type loop, but resolvable through fwd declaration */
429 if (btf_is_composite(t) && through_ptr && t->name_off != 0)
430 return 0;
431 pr_warning("unsatisfiable type cycle, id:[%u]\n", id);
432 return -ELOOP;
433 }
434
435 switch (btf_kind(t)) {
436 case BTF_KIND_INT:
437 tstate->order_state = ORDERED;
438 return 0;
439
440 case BTF_KIND_PTR:
441 err = btf_dump_order_type(d, t->type, true);
442 tstate->order_state = ORDERED;
443 return err;
444
445 case BTF_KIND_ARRAY:
446 return btf_dump_order_type(d, btf_array(t)->type, through_ptr);
447
448 case BTF_KIND_STRUCT:
449 case BTF_KIND_UNION: {
450 const struct btf_member *m = btf_members(t);
451 /*
452 * struct/union is part of strong link, only if it's embedded
453 * (so no ptr in a path) or it's anonymous (so has to be
454 * defined inline, even if declared through ptr)
455 */
456 if (through_ptr && t->name_off != 0)
457 return 0;
458
459 tstate->order_state = ORDERING;
460
461 vlen = btf_vlen(t);
462 for (i = 0; i < vlen; i++, m++) {
463 err = btf_dump_order_type(d, m->type, false);
464 if (err < 0)
465 return err;
466 }
467
468 if (t->name_off != 0) {
469 err = btf_dump_add_emit_queue_id(d, id);
470 if (err < 0)
471 return err;
472 }
473
474 tstate->order_state = ORDERED;
475 return 1;
476 }
477 case BTF_KIND_ENUM:
478 case BTF_KIND_FWD:
479 /*
480 * non-anonymous or non-referenced enums are top-level
481 * declarations and should be emitted. Same logic can be
482 * applied to FWDs, it won't hurt anyways.
483 */
484 if (t->name_off != 0 || !tstate->referenced) {
485 err = btf_dump_add_emit_queue_id(d, id);
486 if (err)
487 return err;
488 }
489 tstate->order_state = ORDERED;
490 return 1;
491
492 case BTF_KIND_TYPEDEF: {
493 int is_strong;
494
495 is_strong = btf_dump_order_type(d, t->type, through_ptr);
496 if (is_strong < 0)
497 return is_strong;
498
499 /* typedef is similar to struct/union w.r.t. fwd-decls */
500 if (through_ptr && !is_strong)
501 return 0;
502
503 /* typedef is always a named definition */
504 err = btf_dump_add_emit_queue_id(d, id);
505 if (err)
506 return err;
507
508 d->type_states[id].order_state = ORDERED;
509 return 1;
510 }
511 case BTF_KIND_VOLATILE:
512 case BTF_KIND_CONST:
513 case BTF_KIND_RESTRICT:
514 return btf_dump_order_type(d, t->type, through_ptr);
515
516 case BTF_KIND_FUNC_PROTO: {
517 const struct btf_param *p = btf_params(t);
518 bool is_strong;
519
520 err = btf_dump_order_type(d, t->type, through_ptr);
521 if (err < 0)
522 return err;
523 is_strong = err > 0;
524
525 vlen = btf_vlen(t);
526 for (i = 0; i < vlen; i++, p++) {
527 err = btf_dump_order_type(d, p->type, through_ptr);
528 if (err < 0)
529 return err;
530 if (err > 0)
531 is_strong = true;
532 }
533 return is_strong;
534 }
535 case BTF_KIND_FUNC:
536 case BTF_KIND_VAR:
537 case BTF_KIND_DATASEC:
538 d->type_states[id].order_state = ORDERED;
539 return 0;
540
541 default:
542 return -EINVAL;
543 }
544}
545
546static void btf_dump_emit_struct_fwd(struct btf_dump *d, __u32 id,
547 const struct btf_type *t);
548static void btf_dump_emit_struct_def(struct btf_dump *d, __u32 id,
549 const struct btf_type *t, int lvl);
550
551static void btf_dump_emit_enum_fwd(struct btf_dump *d, __u32 id,
552 const struct btf_type *t);
553static void btf_dump_emit_enum_def(struct btf_dump *d, __u32 id,
554 const struct btf_type *t, int lvl);
555
556static void btf_dump_emit_fwd_def(struct btf_dump *d, __u32 id,
557 const struct btf_type *t);
558
559static void btf_dump_emit_typedef_def(struct btf_dump *d, __u32 id,
560 const struct btf_type *t, int lvl);
561
562/* a local view into a shared stack */
563struct id_stack {
564 const __u32 *ids;
565 int cnt;
566};
567
568static void btf_dump_emit_type_decl(struct btf_dump *d, __u32 id,
569 const char *fname, int lvl);
570static void btf_dump_emit_type_chain(struct btf_dump *d,
571 struct id_stack *decl_stack,
572 const char *fname, int lvl);
573
574static const char *btf_dump_type_name(struct btf_dump *d, __u32 id);
575static const char *btf_dump_ident_name(struct btf_dump *d, __u32 id);
576static size_t btf_dump_name_dups(struct btf_dump *d, struct hashmap *name_map,
577 const char *orig_name);
578
579static bool btf_dump_is_blacklisted(struct btf_dump *d, __u32 id)
580{
581 const struct btf_type *t = btf__type_by_id(d->btf, id);
582
583 /* __builtin_va_list is a compiler built-in, which causes compilation
584 * errors, when compiling w/ different compiler, then used to compile
585 * original code (e.g., GCC to compile kernel, Clang to use generated
586 * C header from BTF). As it is built-in, it should be already defined
587 * properly internally in compiler.
588 */
589 if (t->name_off == 0)
590 return false;
591 return strcmp(btf_name_of(d, t->name_off), "__builtin_va_list") == 0;
592}
593
594/*
595 * Emit C-syntax definitions of types from chains of BTF types.
596 *
597 * High-level handling of determining necessary forward declarations are handled
598 * by btf_dump_emit_type() itself, but all nitty-gritty details of emitting type
599 * declarations/definitions in C syntax are handled by a combo of
600 * btf_dump_emit_type_decl()/btf_dump_emit_type_chain() w/ delegation to
601 * corresponding btf_dump_emit_*_{def,fwd}() functions.
602 *
603 * We also keep track of "containing struct/union type ID" to determine when
604 * we reference it from inside and thus can avoid emitting unnecessary forward
605 * declaration.
606 *
607 * This algorithm is designed in such a way, that even if some error occurs
608 * (either technical, e.g., out of memory, or logical, i.e., malformed BTF
609 * that doesn't comply to C rules completely), algorithm will try to proceed
610 * and produce as much meaningful output as possible.
611 */
612static void btf_dump_emit_type(struct btf_dump *d, __u32 id, __u32 cont_id)
613{
614 struct btf_dump_type_aux_state *tstate = &d->type_states[id];
615 bool top_level_def = cont_id == 0;
616 const struct btf_type *t;
617 __u16 kind;
618
619 if (tstate->emit_state == EMITTED)
620 return;
621
622 t = btf__type_by_id(d->btf, id);
623 kind = btf_kind(t);
624
625 if (tstate->emit_state == EMITTING) {
626 if (tstate->fwd_emitted)
627 return;
628
629 switch (kind) {
630 case BTF_KIND_STRUCT:
631 case BTF_KIND_UNION:
632 /*
633 * if we are referencing a struct/union that we are
634 * part of - then no need for fwd declaration
635 */
636 if (id == cont_id)
637 return;
638 if (t->name_off == 0) {
639 pr_warning("anonymous struct/union loop, id:[%u]\n",
640 id);
641 return;
642 }
643 btf_dump_emit_struct_fwd(d, id, t);
644 btf_dump_printf(d, ";\n\n");
645 tstate->fwd_emitted = 1;
646 break;
647 case BTF_KIND_TYPEDEF:
648 /*
649 * for typedef fwd_emitted means typedef definition
650 * was emitted, but it can be used only for "weak"
651 * references through pointer only, not for embedding
652 */
653 if (!btf_dump_is_blacklisted(d, id)) {
654 btf_dump_emit_typedef_def(d, id, t, 0);
655 btf_dump_printf(d, ";\n\n");
656 };
657 tstate->fwd_emitted = 1;
658 break;
659 default:
660 break;
661 }
662
663 return;
664 }
665
666 switch (kind) {
667 case BTF_KIND_INT:
668 tstate->emit_state = EMITTED;
669 break;
670 case BTF_KIND_ENUM:
671 if (top_level_def) {
672 btf_dump_emit_enum_def(d, id, t, 0);
673 btf_dump_printf(d, ";\n\n");
674 }
675 tstate->emit_state = EMITTED;
676 break;
677 case BTF_KIND_PTR:
678 case BTF_KIND_VOLATILE:
679 case BTF_KIND_CONST:
680 case BTF_KIND_RESTRICT:
681 btf_dump_emit_type(d, t->type, cont_id);
682 break;
683 case BTF_KIND_ARRAY:
684 btf_dump_emit_type(d, btf_array(t)->type, cont_id);
685 break;
686 case BTF_KIND_FWD:
687 btf_dump_emit_fwd_def(d, id, t);
688 btf_dump_printf(d, ";\n\n");
689 tstate->emit_state = EMITTED;
690 break;
691 case BTF_KIND_TYPEDEF:
692 tstate->emit_state = EMITTING;
693 btf_dump_emit_type(d, t->type, id);
694 /*
695 * typedef can server as both definition and forward
696 * declaration; at this stage someone depends on
697 * typedef as a forward declaration (refers to it
698 * through pointer), so unless we already did it,
699 * emit typedef as a forward declaration
700 */
701 if (!tstate->fwd_emitted && !btf_dump_is_blacklisted(d, id)) {
702 btf_dump_emit_typedef_def(d, id, t, 0);
703 btf_dump_printf(d, ";\n\n");
704 }
705 tstate->emit_state = EMITTED;
706 break;
707 case BTF_KIND_STRUCT:
708 case BTF_KIND_UNION:
709 tstate->emit_state = EMITTING;
710 /* if it's a top-level struct/union definition or struct/union
711 * is anonymous, then in C we'll be emitting all fields and
712 * their types (as opposed to just `struct X`), so we need to
713 * make sure that all types, referenced from struct/union
714 * members have necessary forward-declarations, where
715 * applicable
716 */
717 if (top_level_def || t->name_off == 0) {
718 const struct btf_member *m = btf_members(t);
719 __u16 vlen = btf_vlen(t);
720 int i, new_cont_id;
721
722 new_cont_id = t->name_off == 0 ? cont_id : id;
723 for (i = 0; i < vlen; i++, m++)
724 btf_dump_emit_type(d, m->type, new_cont_id);
725 } else if (!tstate->fwd_emitted && id != cont_id) {
726 btf_dump_emit_struct_fwd(d, id, t);
727 btf_dump_printf(d, ";\n\n");
728 tstate->fwd_emitted = 1;
729 }
730
731 if (top_level_def) {
732 btf_dump_emit_struct_def(d, id, t, 0);
733 btf_dump_printf(d, ";\n\n");
734 tstate->emit_state = EMITTED;
735 } else {
736 tstate->emit_state = NOT_EMITTED;
737 }
738 break;
739 case BTF_KIND_FUNC_PROTO: {
740 const struct btf_param *p = btf_params(t);
741 __u16 vlen = btf_vlen(t);
742 int i;
743
744 btf_dump_emit_type(d, t->type, cont_id);
745 for (i = 0; i < vlen; i++, p++)
746 btf_dump_emit_type(d, p->type, cont_id);
747
748 break;
749 }
750 default:
751 break;
752 }
753}
754
755static int btf_align_of(const struct btf *btf, __u32 id)
756{
757 const struct btf_type *t = btf__type_by_id(btf, id);
758 __u16 kind = btf_kind(t);
759
760 switch (kind) {
761 case BTF_KIND_INT:
762 case BTF_KIND_ENUM:
763 return min(sizeof(void *), t->size);
764 case BTF_KIND_PTR:
765 return sizeof(void *);
766 case BTF_KIND_TYPEDEF:
767 case BTF_KIND_VOLATILE:
768 case BTF_KIND_CONST:
769 case BTF_KIND_RESTRICT:
770 return btf_align_of(btf, t->type);
771 case BTF_KIND_ARRAY:
772 return btf_align_of(btf, btf_array(t)->type);
773 case BTF_KIND_STRUCT:
774 case BTF_KIND_UNION: {
775 const struct btf_member *m = btf_members(t);
776 __u16 vlen = btf_vlen(t);
777 int i, align = 1;
778
779 for (i = 0; i < vlen; i++, m++)
780 align = max(align, btf_align_of(btf, m->type));
781
782 return align;
783 }
784 default:
785 pr_warning("unsupported BTF_KIND:%u\n", btf_kind(t));
786 return 1;
787 }
788}
789
790static bool btf_is_struct_packed(const struct btf *btf, __u32 id,
791 const struct btf_type *t)
792{
793 const struct btf_member *m;
794 int align, i, bit_sz;
795 __u16 vlen;
796
797 align = btf_align_of(btf, id);
798 /* size of a non-packed struct has to be a multiple of its alignment*/
799 if (t->size % align)
800 return true;
801
802 m = btf_members(t);
803 vlen = btf_vlen(t);
804 /* all non-bitfield fields have to be naturally aligned */
805 for (i = 0; i < vlen; i++, m++) {
806 align = btf_align_of(btf, m->type);
807 bit_sz = btf_member_bitfield_size(t, i);
808 if (bit_sz == 0 && m->offset % (8 * align) != 0)
809 return true;
810 }
811
812 /*
813 * if original struct was marked as packed, but its layout is
814 * naturally aligned, we'll detect that it's not packed
815 */
816 return false;
817}
818
819static int chip_away_bits(int total, int at_most)
820{
821 return total % at_most ? : at_most;
822}
823
824static void btf_dump_emit_bit_padding(const struct btf_dump *d,
825 int cur_off, int m_off, int m_bit_sz,
826 int align, int lvl)
827{
828 int off_diff = m_off - cur_off;
829 int ptr_bits = sizeof(void *) * 8;
830
831 if (off_diff <= 0)
832 /* no gap */
833 return;
834 if (m_bit_sz == 0 && off_diff < align * 8)
835 /* natural padding will take care of a gap */
836 return;
837
838 while (off_diff > 0) {
839 const char *pad_type;
840 int pad_bits;
841
842 if (ptr_bits > 32 && off_diff > 32) {
843 pad_type = "long";
844 pad_bits = chip_away_bits(off_diff, ptr_bits);
845 } else if (off_diff > 16) {
846 pad_type = "int";
847 pad_bits = chip_away_bits(off_diff, 32);
848 } else if (off_diff > 8) {
849 pad_type = "short";
850 pad_bits = chip_away_bits(off_diff, 16);
851 } else {
852 pad_type = "char";
853 pad_bits = chip_away_bits(off_diff, 8);
854 }
855 btf_dump_printf(d, "\n%s%s: %d;", pfx(lvl), pad_type, pad_bits);
856 off_diff -= pad_bits;
857 }
858}
859
860static void btf_dump_emit_struct_fwd(struct btf_dump *d, __u32 id,
861 const struct btf_type *t)
862{
863 btf_dump_printf(d, "%s %s",
864 btf_is_struct(t) ? "struct" : "union",
865 btf_dump_type_name(d, id));
866}
867
868static void btf_dump_emit_struct_def(struct btf_dump *d,
869 __u32 id,
870 const struct btf_type *t,
871 int lvl)
872{
873 const struct btf_member *m = btf_members(t);
874 bool is_struct = btf_is_struct(t);
875 int align, i, packed, off = 0;
876 __u16 vlen = btf_vlen(t);
877
878 packed = is_struct ? btf_is_struct_packed(d->btf, id, t) : 0;
879 align = packed ? 1 : btf_align_of(d->btf, id);
880
881 btf_dump_printf(d, "%s%s%s {",
882 is_struct ? "struct" : "union",
883 t->name_off ? " " : "",
884 btf_dump_type_name(d, id));
885
886 for (i = 0; i < vlen; i++, m++) {
887 const char *fname;
888 int m_off, m_sz;
889
890 fname = btf_name_of(d, m->name_off);
891 m_sz = btf_member_bitfield_size(t, i);
892 m_off = btf_member_bit_offset(t, i);
893 align = packed ? 1 : btf_align_of(d->btf, m->type);
894
895 btf_dump_emit_bit_padding(d, off, m_off, m_sz, align, lvl + 1);
896 btf_dump_printf(d, "\n%s", pfx(lvl + 1));
897 btf_dump_emit_type_decl(d, m->type, fname, lvl + 1);
898
899 if (m_sz) {
900 btf_dump_printf(d, ": %d", m_sz);
901 off = m_off + m_sz;
902 } else {
903 m_sz = max(0, btf__resolve_size(d->btf, m->type));
904 off = m_off + m_sz * 8;
905 }
906 btf_dump_printf(d, ";");
907 }
908
909 if (vlen)
910 btf_dump_printf(d, "\n");
911 btf_dump_printf(d, "%s}", pfx(lvl));
912 if (packed)
913 btf_dump_printf(d, " __attribute__((packed))");
914}
915
916static void btf_dump_emit_enum_fwd(struct btf_dump *d, __u32 id,
917 const struct btf_type *t)
918{
919 btf_dump_printf(d, "enum %s", btf_dump_type_name(d, id));
920}
921
922static void btf_dump_emit_enum_def(struct btf_dump *d, __u32 id,
923 const struct btf_type *t,
924 int lvl)
925{
926 const struct btf_enum *v = btf_enum(t);
927 __u16 vlen = btf_vlen(t);
928 const char *name;
929 size_t dup_cnt;
930 int i;
931
932 btf_dump_printf(d, "enum%s%s",
933 t->name_off ? " " : "",
934 btf_dump_type_name(d, id));
935
936 if (vlen) {
937 btf_dump_printf(d, " {");
938 for (i = 0; i < vlen; i++, v++) {
939 name = btf_name_of(d, v->name_off);
940 /* enumerators share namespace with typedef idents */
941 dup_cnt = btf_dump_name_dups(d, d->ident_names, name);
942 if (dup_cnt > 1) {
943 btf_dump_printf(d, "\n%s%s___%zu = %d,",
944 pfx(lvl + 1), name, dup_cnt,
945 (__s32)v->val);
946 } else {
947 btf_dump_printf(d, "\n%s%s = %d,",
948 pfx(lvl + 1), name,
949 (__s32)v->val);
950 }
951 }
952 btf_dump_printf(d, "\n%s}", pfx(lvl));
953 }
954}
955
956static void btf_dump_emit_fwd_def(struct btf_dump *d, __u32 id,
957 const struct btf_type *t)
958{
959 const char *name = btf_dump_type_name(d, id);
960
961 if (btf_kflag(t))
962 btf_dump_printf(d, "union %s", name);
963 else
964 btf_dump_printf(d, "struct %s", name);
965}
966
967static void btf_dump_emit_typedef_def(struct btf_dump *d, __u32 id,
968 const struct btf_type *t, int lvl)
969{
970 const char *name = btf_dump_ident_name(d, id);
971
972 btf_dump_printf(d, "typedef ");
973 btf_dump_emit_type_decl(d, t->type, name, lvl);
974}
975
976static int btf_dump_push_decl_stack_id(struct btf_dump *d, __u32 id)
977{
978 __u32 *new_stack;
979 size_t new_cap;
980
981 if (d->decl_stack_cnt >= d->decl_stack_cap) {
982 new_cap = max(16, d->decl_stack_cap * 3 / 2);
983 new_stack = realloc(d->decl_stack,
984 new_cap * sizeof(new_stack[0]));
985 if (!new_stack)
986 return -ENOMEM;
987 d->decl_stack = new_stack;
988 d->decl_stack_cap = new_cap;
989 }
990
991 d->decl_stack[d->decl_stack_cnt++] = id;
992
993 return 0;
994}
995
996/*
997 * Emit type declaration (e.g., field type declaration in a struct or argument
998 * declaration in function prototype) in correct C syntax.
999 *
1000 * For most types it's trivial, but there are few quirky type declaration
1001 * cases worth mentioning:
1002 * - function prototypes (especially nesting of function prototypes);
1003 * - arrays;
1004 * - const/volatile/restrict for pointers vs other types.
1005 *
1006 * For a good discussion of *PARSING* C syntax (as a human), see
1007 * Peter van der Linden's "Expert C Programming: Deep C Secrets",
1008 * Ch.3 "Unscrambling Declarations in C".
1009 *
1010 * It won't help with BTF to C conversion much, though, as it's an opposite
1011 * problem. So we came up with this algorithm in reverse to van der Linden's
1012 * parsing algorithm. It goes from structured BTF representation of type
1013 * declaration to a valid compilable C syntax.
1014 *
1015 * For instance, consider this C typedef:
1016 * typedef const int * const * arr[10] arr_t;
1017 * It will be represented in BTF with this chain of BTF types:
1018 * [typedef] -> [array] -> [ptr] -> [const] -> [ptr] -> [const] -> [int]
1019 *
1020 * Notice how [const] modifier always goes before type it modifies in BTF type
1021 * graph, but in C syntax, const/volatile/restrict modifiers are written to
1022 * the right of pointers, but to the left of other types. There are also other
1023 * quirks, like function pointers, arrays of them, functions returning other
1024 * functions, etc.
1025 *
1026 * We handle that by pushing all the types to a stack, until we hit "terminal"
1027 * type (int/enum/struct/union/fwd). Then depending on the kind of a type on
1028 * top of a stack, modifiers are handled differently. Array/function pointers
1029 * have also wildly different syntax and how nesting of them are done. See
1030 * code for authoritative definition.
1031 *
1032 * To avoid allocating new stack for each independent chain of BTF types, we
1033 * share one bigger stack, with each chain working only on its own local view
1034 * of a stack frame. Some care is required to "pop" stack frames after
1035 * processing type declaration chain.
1036 */
1037static void btf_dump_emit_type_decl(struct btf_dump *d, __u32 id,
1038 const char *fname, int lvl)
1039{
1040 struct id_stack decl_stack;
1041 const struct btf_type *t;
1042 int err, stack_start;
1043
1044 stack_start = d->decl_stack_cnt;
1045 for (;;) {
1046 err = btf_dump_push_decl_stack_id(d, id);
1047 if (err < 0) {
1048 /*
1049 * if we don't have enough memory for entire type decl
1050 * chain, restore stack, emit warning, and try to
1051 * proceed nevertheless
1052 */
1053 pr_warning("not enough memory for decl stack:%d", err);
1054 d->decl_stack_cnt = stack_start;
1055 return;
1056 }
1057
1058 /* VOID */
1059 if (id == 0)
1060 break;
1061
1062 t = btf__type_by_id(d->btf, id);
1063 switch (btf_kind(t)) {
1064 case BTF_KIND_PTR:
1065 case BTF_KIND_VOLATILE:
1066 case BTF_KIND_CONST:
1067 case BTF_KIND_RESTRICT:
1068 case BTF_KIND_FUNC_PROTO:
1069 id = t->type;
1070 break;
1071 case BTF_KIND_ARRAY:
1072 id = btf_array(t)->type;
1073 break;
1074 case BTF_KIND_INT:
1075 case BTF_KIND_ENUM:
1076 case BTF_KIND_FWD:
1077 case BTF_KIND_STRUCT:
1078 case BTF_KIND_UNION:
1079 case BTF_KIND_TYPEDEF:
1080 goto done;
1081 default:
1082 pr_warning("unexpected type in decl chain, kind:%u, id:[%u]\n",
1083 btf_kind(t), id);
1084 goto done;
1085 }
1086 }
1087done:
1088 /*
1089 * We might be inside a chain of declarations (e.g., array of function
1090 * pointers returning anonymous (so inlined) structs, having another
1091 * array field). Each of those needs its own "stack frame" to handle
1092 * emitting of declarations. Those stack frames are non-overlapping
1093 * portions of shared btf_dump->decl_stack. To make it a bit nicer to
1094 * handle this set of nested stacks, we create a view corresponding to
1095 * our own "stack frame" and work with it as an independent stack.
1096 * We'll need to clean up after emit_type_chain() returns, though.
1097 */
1098 decl_stack.ids = d->decl_stack + stack_start;
1099 decl_stack.cnt = d->decl_stack_cnt - stack_start;
1100 btf_dump_emit_type_chain(d, &decl_stack, fname, lvl);
1101 /*
1102 * emit_type_chain() guarantees that it will pop its entire decl_stack
1103 * frame before returning. But it works with a read-only view into
1104 * decl_stack, so it doesn't actually pop anything from the
1105 * perspective of shared btf_dump->decl_stack, per se. We need to
1106 * reset decl_stack state to how it was before us to avoid it growing
1107 * all the time.
1108 */
1109 d->decl_stack_cnt = stack_start;
1110}
1111
1112static void btf_dump_emit_mods(struct btf_dump *d, struct id_stack *decl_stack)
1113{
1114 const struct btf_type *t;
1115 __u32 id;
1116
1117 while (decl_stack->cnt) {
1118 id = decl_stack->ids[decl_stack->cnt - 1];
1119 t = btf__type_by_id(d->btf, id);
1120
1121 switch (btf_kind(t)) {
1122 case BTF_KIND_VOLATILE:
1123 btf_dump_printf(d, "volatile ");
1124 break;
1125 case BTF_KIND_CONST:
1126 btf_dump_printf(d, "const ");
1127 break;
1128 case BTF_KIND_RESTRICT:
1129 btf_dump_printf(d, "restrict ");
1130 break;
1131 default:
1132 return;
1133 }
1134 decl_stack->cnt--;
1135 }
1136}
1137
1138static void btf_dump_emit_name(const struct btf_dump *d,
1139 const char *name, bool last_was_ptr)
1140{
1141 bool separate = name[0] && !last_was_ptr;
1142
1143 btf_dump_printf(d, "%s%s", separate ? " " : "", name);
1144}
1145
1146static void btf_dump_emit_type_chain(struct btf_dump *d,
1147 struct id_stack *decls,
1148 const char *fname, int lvl)
1149{
1150 /*
1151 * last_was_ptr is used to determine if we need to separate pointer
1152 * asterisk (*) from previous part of type signature with space, so
1153 * that we get `int ***`, instead of `int * * *`. We default to true
1154 * for cases where we have single pointer in a chain. E.g., in ptr ->
1155 * func_proto case. func_proto will start a new emit_type_chain call
1156 * with just ptr, which should be emitted as (*) or (*<fname>), so we
1157 * don't want to prepend space for that last pointer.
1158 */
1159 bool last_was_ptr = true;
1160 const struct btf_type *t;
1161 const char *name;
1162 __u16 kind;
1163 __u32 id;
1164
1165 while (decls->cnt) {
1166 id = decls->ids[--decls->cnt];
1167 if (id == 0) {
1168 /* VOID is a special snowflake */
1169 btf_dump_emit_mods(d, decls);
1170 btf_dump_printf(d, "void");
1171 last_was_ptr = false;
1172 continue;
1173 }
1174
1175 t = btf__type_by_id(d->btf, id);
1176 kind = btf_kind(t);
1177
1178 switch (kind) {
1179 case BTF_KIND_INT:
1180 btf_dump_emit_mods(d, decls);
1181 name = btf_name_of(d, t->name_off);
1182 btf_dump_printf(d, "%s", name);
1183 break;
1184 case BTF_KIND_STRUCT:
1185 case BTF_KIND_UNION:
1186 btf_dump_emit_mods(d, decls);
1187 /* inline anonymous struct/union */
1188 if (t->name_off == 0)
1189 btf_dump_emit_struct_def(d, id, t, lvl);
1190 else
1191 btf_dump_emit_struct_fwd(d, id, t);
1192 break;
1193 case BTF_KIND_ENUM:
1194 btf_dump_emit_mods(d, decls);
1195 /* inline anonymous enum */
1196 if (t->name_off == 0)
1197 btf_dump_emit_enum_def(d, id, t, lvl);
1198 else
1199 btf_dump_emit_enum_fwd(d, id, t);
1200 break;
1201 case BTF_KIND_FWD:
1202 btf_dump_emit_mods(d, decls);
1203 btf_dump_emit_fwd_def(d, id, t);
1204 break;
1205 case BTF_KIND_TYPEDEF:
1206 btf_dump_emit_mods(d, decls);
1207 btf_dump_printf(d, "%s", btf_dump_ident_name(d, id));
1208 break;
1209 case BTF_KIND_PTR:
1210 btf_dump_printf(d, "%s", last_was_ptr ? "*" : " *");
1211 break;
1212 case BTF_KIND_VOLATILE:
1213 btf_dump_printf(d, " volatile");
1214 break;
1215 case BTF_KIND_CONST:
1216 btf_dump_printf(d, " const");
1217 break;
1218 case BTF_KIND_RESTRICT:
1219 btf_dump_printf(d, " restrict");
1220 break;
1221 case BTF_KIND_ARRAY: {
1222 const struct btf_array *a = btf_array(t);
1223 const struct btf_type *next_t;
1224 __u32 next_id;
1225 bool multidim;
1226 /*
1227 * GCC has a bug
1228 * (https://gcc.gnu.org/bugzilla/show_bug.cgi?id=8354)
1229 * which causes it to emit extra const/volatile
1230 * modifiers for an array, if array's element type has
1231 * const/volatile modifiers. Clang doesn't do that.
1232 * In general, it doesn't seem very meaningful to have
1233 * a const/volatile modifier for array, so we are
1234 * going to silently skip them here.
1235 */
1236 while (decls->cnt) {
1237 next_id = decls->ids[decls->cnt - 1];
1238 next_t = btf__type_by_id(d->btf, next_id);
1239 if (btf_is_mod(next_t))
1240 decls->cnt--;
1241 else
1242 break;
1243 }
1244
1245 if (decls->cnt == 0) {
1246 btf_dump_emit_name(d, fname, last_was_ptr);
1247 btf_dump_printf(d, "[%u]", a->nelems);
1248 return;
1249 }
1250
1251 next_id = decls->ids[decls->cnt - 1];
1252 next_t = btf__type_by_id(d->btf, next_id);
1253 multidim = btf_is_array(next_t);
1254 /* we need space if we have named non-pointer */
1255 if (fname[0] && !last_was_ptr)
1256 btf_dump_printf(d, " ");
1257 /* no parentheses for multi-dimensional array */
1258 if (!multidim)
1259 btf_dump_printf(d, "(");
1260 btf_dump_emit_type_chain(d, decls, fname, lvl);
1261 if (!multidim)
1262 btf_dump_printf(d, ")");
1263 btf_dump_printf(d, "[%u]", a->nelems);
1264 return;
1265 }
1266 case BTF_KIND_FUNC_PROTO: {
1267 const struct btf_param *p = btf_params(t);
1268 __u16 vlen = btf_vlen(t);
1269 int i;
1270
1271 btf_dump_emit_mods(d, decls);
1272 if (decls->cnt) {
1273 btf_dump_printf(d, " (");
1274 btf_dump_emit_type_chain(d, decls, fname, lvl);
1275 btf_dump_printf(d, ")");
1276 } else {
1277 btf_dump_emit_name(d, fname, last_was_ptr);
1278 }
1279 btf_dump_printf(d, "(");
1280 /*
1281 * Clang for BPF target generates func_proto with no
1282 * args as a func_proto with a single void arg (e.g.,
1283 * `int (*f)(void)` vs just `int (*f)()`). We are
1284 * going to pretend there are no args for such case.
1285 */
1286 if (vlen == 1 && p->type == 0) {
1287 btf_dump_printf(d, ")");
1288 return;
1289 }
1290
1291 for (i = 0; i < vlen; i++, p++) {
1292 if (i > 0)
1293 btf_dump_printf(d, ", ");
1294
1295 /* last arg of type void is vararg */
1296 if (i == vlen - 1 && p->type == 0) {
1297 btf_dump_printf(d, "...");
1298 break;
1299 }
1300
1301 name = btf_name_of(d, p->name_off);
1302 btf_dump_emit_type_decl(d, p->type, name, lvl);
1303 }
1304
1305 btf_dump_printf(d, ")");
1306 return;
1307 }
1308 default:
1309 pr_warning("unexpected type in decl chain, kind:%u, id:[%u]\n",
1310 kind, id);
1311 return;
1312 }
1313
1314 last_was_ptr = kind == BTF_KIND_PTR;
1315 }
1316
1317 btf_dump_emit_name(d, fname, last_was_ptr);
1318}
1319
1320/* return number of duplicates (occurrences) of a given name */
1321static size_t btf_dump_name_dups(struct btf_dump *d, struct hashmap *name_map,
1322 const char *orig_name)
1323{
1324 size_t dup_cnt = 0;
1325
1326 hashmap__find(name_map, orig_name, (void **)&dup_cnt);
1327 dup_cnt++;
1328 hashmap__set(name_map, orig_name, (void *)dup_cnt, NULL, NULL);
1329
1330 return dup_cnt;
1331}
1332
1333static const char *btf_dump_resolve_name(struct btf_dump *d, __u32 id,
1334 struct hashmap *name_map)
1335{
1336 struct btf_dump_type_aux_state *s = &d->type_states[id];
1337 const struct btf_type *t = btf__type_by_id(d->btf, id);
1338 const char *orig_name = btf_name_of(d, t->name_off);
1339 const char **cached_name = &d->cached_names[id];
1340 size_t dup_cnt;
1341
1342 if (t->name_off == 0)
1343 return "";
1344
1345 if (s->name_resolved)
1346 return *cached_name ? *cached_name : orig_name;
1347
1348 dup_cnt = btf_dump_name_dups(d, name_map, orig_name);
1349 if (dup_cnt > 1) {
1350 const size_t max_len = 256;
1351 char new_name[max_len];
1352
1353 snprintf(new_name, max_len, "%s___%zu", orig_name, dup_cnt);
1354 *cached_name = strdup(new_name);
1355 }
1356
1357 s->name_resolved = 1;
1358 return *cached_name ? *cached_name : orig_name;
1359}
1360
1361static const char *btf_dump_type_name(struct btf_dump *d, __u32 id)
1362{
1363 return btf_dump_resolve_name(d, id, d->type_names);
1364}
1365
1366static const char *btf_dump_ident_name(struct btf_dump *d, __u32 id)
1367{
1368 return btf_dump_resolve_name(d, id, d->ident_names);
1369}