• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 <linux/kernel.h>
17 #include "btf.h"
18 #include "hashmap.h"
19 #include "libbpf.h"
20 #include "libbpf_internal.h"
21 
22 static const char PREFIXES[] = "\t\t\t\t\t\t\t\t\t\t\t\t\t";
23 static const size_t PREFIX_CNT = sizeof(PREFIXES) - 1;
24 
pfx(int lvl)25 static const char *pfx(int lvl)
26 {
27 	return lvl >= PREFIX_CNT ? PREFIXES : &PREFIXES[PREFIX_CNT - lvl];
28 }
29 
30 enum btf_dump_type_order_state {
31 	NOT_ORDERED,
32 	ORDERING,
33 	ORDERED,
34 };
35 
36 enum btf_dump_type_emit_state {
37 	NOT_EMITTED,
38 	EMITTING,
39 	EMITTED,
40 };
41 
42 /* per-type auxiliary state */
43 struct btf_dump_type_aux_state {
44 	/* topological sorting state */
45 	enum btf_dump_type_order_state order_state: 2;
46 	/* emitting state used to determine the need for forward declaration */
47 	enum btf_dump_type_emit_state emit_state: 2;
48 	/* whether forward declaration was already emitted */
49 	__u8 fwd_emitted: 1;
50 	/* whether unique non-duplicate name was already assigned */
51 	__u8 name_resolved: 1;
52 	/* whether type is referenced from any other type */
53 	__u8 referenced: 1;
54 };
55 
56 struct btf_dump {
57 	const struct btf *btf;
58 	const struct btf_ext *btf_ext;
59 	btf_dump_printf_fn_t printf_fn;
60 	struct btf_dump_opts opts;
61 	int ptr_sz;
62 	bool strip_mods;
63 	int last_id;
64 
65 	/* per-type auxiliary state */
66 	struct btf_dump_type_aux_state *type_states;
67 	size_t type_states_cap;
68 	/* per-type optional cached unique name, must be freed, if present */
69 	const char **cached_names;
70 	size_t cached_names_cap;
71 
72 	/* topo-sorted list of dependent type definitions */
73 	__u32 *emit_queue;
74 	int emit_queue_cap;
75 	int emit_queue_cnt;
76 
77 	/*
78 	 * stack of type declarations (e.g., chain of modifiers, arrays,
79 	 * funcs, etc)
80 	 */
81 	__u32 *decl_stack;
82 	int decl_stack_cap;
83 	int decl_stack_cnt;
84 
85 	/* maps struct/union/enum name to a number of name occurrences */
86 	struct hashmap *type_names;
87 	/*
88 	 * maps typedef identifiers and enum value names to a number of such
89 	 * name occurrences
90 	 */
91 	struct hashmap *ident_names;
92 };
93 
str_hash_fn(const void * key,void * ctx)94 static size_t str_hash_fn(const void *key, void *ctx)
95 {
96 	return str_hash(key);
97 }
98 
str_equal_fn(const void * a,const void * b,void * ctx)99 static bool str_equal_fn(const void *a, const void *b, void *ctx)
100 {
101 	return strcmp(a, b) == 0;
102 }
103 
btf_name_of(const struct btf_dump * d,__u32 name_off)104 static const char *btf_name_of(const struct btf_dump *d, __u32 name_off)
105 {
106 	return btf__name_by_offset(d->btf, name_off);
107 }
108 
btf_dump_printf(const struct btf_dump * d,const char * fmt,...)109 static void btf_dump_printf(const struct btf_dump *d, const char *fmt, ...)
110 {
111 	va_list args;
112 
113 	va_start(args, fmt);
114 	d->printf_fn(d->opts.ctx, fmt, args);
115 	va_end(args);
116 }
117 
118 static int btf_dump_mark_referenced(struct btf_dump *d);
119 static int btf_dump_resize(struct btf_dump *d);
120 
btf_dump__new(const struct btf * btf,const struct btf_ext * btf_ext,const struct btf_dump_opts * opts,btf_dump_printf_fn_t printf_fn)121 struct btf_dump *btf_dump__new(const struct btf *btf,
122 			       const struct btf_ext *btf_ext,
123 			       const struct btf_dump_opts *opts,
124 			       btf_dump_printf_fn_t printf_fn)
125 {
126 	struct btf_dump *d;
127 	int err;
128 
129 	d = calloc(1, sizeof(struct btf_dump));
130 	if (!d)
131 		return ERR_PTR(-ENOMEM);
132 
133 	d->btf = btf;
134 	d->btf_ext = btf_ext;
135 	d->printf_fn = printf_fn;
136 	d->opts.ctx = opts ? opts->ctx : NULL;
137 	d->ptr_sz = btf__pointer_size(btf) ? : sizeof(void *);
138 
139 	d->type_names = hashmap__new(str_hash_fn, str_equal_fn, NULL);
140 	if (IS_ERR(d->type_names)) {
141 		err = PTR_ERR(d->type_names);
142 		d->type_names = NULL;
143 		goto err;
144 	}
145 	d->ident_names = hashmap__new(str_hash_fn, str_equal_fn, NULL);
146 	if (IS_ERR(d->ident_names)) {
147 		err = PTR_ERR(d->ident_names);
148 		d->ident_names = NULL;
149 		goto err;
150 	}
151 
152 	err = btf_dump_resize(d);
153 	if (err)
154 		goto err;
155 
156 	return d;
157 err:
158 	btf_dump__free(d);
159 	return ERR_PTR(err);
160 }
161 
btf_dump_resize(struct btf_dump * d)162 static int btf_dump_resize(struct btf_dump *d)
163 {
164 	int err, last_id = btf__get_nr_types(d->btf);
165 
166 	if (last_id <= d->last_id)
167 		return 0;
168 
169 	if (btf_ensure_mem((void **)&d->type_states, &d->type_states_cap,
170 			   sizeof(*d->type_states), last_id + 1))
171 		return -ENOMEM;
172 	if (btf_ensure_mem((void **)&d->cached_names, &d->cached_names_cap,
173 			   sizeof(*d->cached_names), last_id + 1))
174 		return -ENOMEM;
175 
176 	if (d->last_id == 0) {
177 		/* VOID is special */
178 		d->type_states[0].order_state = ORDERED;
179 		d->type_states[0].emit_state = EMITTED;
180 	}
181 
182 	/* eagerly determine referenced types for anon enums */
183 	err = btf_dump_mark_referenced(d);
184 	if (err)
185 		return err;
186 
187 	d->last_id = last_id;
188 	return 0;
189 }
190 
btf_dump_free_names(struct hashmap * map)191 static void btf_dump_free_names(struct hashmap *map)
192 {
193 	size_t bkt;
194 	struct hashmap_entry *cur;
195 
196 	hashmap__for_each_entry(map, cur, bkt)
197 		free((void *)cur->key);
198 
199 	hashmap__free(map);
200 }
201 
btf_dump__free(struct btf_dump * d)202 void btf_dump__free(struct btf_dump *d)
203 {
204 	int i;
205 
206 	if (IS_ERR_OR_NULL(d))
207 		return;
208 
209 	free(d->type_states);
210 	if (d->cached_names) {
211 		/* any set cached name is owned by us and should be freed */
212 		for (i = 0; i <= d->last_id; i++) {
213 			if (d->cached_names[i])
214 				free((void *)d->cached_names[i]);
215 		}
216 	}
217 	free(d->cached_names);
218 	free(d->emit_queue);
219 	free(d->decl_stack);
220 	btf_dump_free_names(d->type_names);
221 	btf_dump_free_names(d->ident_names);
222 
223 	free(d);
224 }
225 
226 static int btf_dump_order_type(struct btf_dump *d, __u32 id, bool through_ptr);
227 static void btf_dump_emit_type(struct btf_dump *d, __u32 id, __u32 cont_id);
228 
229 /*
230  * Dump BTF type in a compilable C syntax, including all the necessary
231  * dependent types, necessary for compilation. If some of the dependent types
232  * were already emitted as part of previous btf_dump__dump_type() invocation
233  * for another type, they won't be emitted again. This API allows callers to
234  * filter out BTF types according to user-defined criterias and emitted only
235  * minimal subset of types, necessary to compile everything. Full struct/union
236  * definitions will still be emitted, even if the only usage is through
237  * pointer and could be satisfied with just a forward declaration.
238  *
239  * Dumping is done in two high-level passes:
240  *   1. Topologically sort type definitions to satisfy C rules of compilation.
241  *   2. Emit type definitions in C syntax.
242  *
243  * Returns 0 on success; <0, otherwise.
244  */
btf_dump__dump_type(struct btf_dump * d,__u32 id)245 int btf_dump__dump_type(struct btf_dump *d, __u32 id)
246 {
247 	int err, i;
248 
249 	if (id > btf__get_nr_types(d->btf))
250 		return -EINVAL;
251 
252 	err = btf_dump_resize(d);
253 	if (err)
254 		return err;
255 
256 	d->emit_queue_cnt = 0;
257 	err = btf_dump_order_type(d, id, false);
258 	if (err < 0)
259 		return err;
260 
261 	for (i = 0; i < d->emit_queue_cnt; i++)
262 		btf_dump_emit_type(d, d->emit_queue[i], 0 /*top-level*/);
263 
264 	return 0;
265 }
266 
267 /*
268  * Mark all types that are referenced from any other type. This is used to
269  * determine top-level anonymous enums that need to be emitted as an
270  * independent type declarations.
271  * Anonymous enums come in two flavors: either embedded in a struct's field
272  * definition, in which case they have to be declared inline as part of field
273  * type declaration; or as a top-level anonymous enum, typically used for
274  * declaring global constants. It's impossible to distinguish between two
275  * without knowning whether given enum type was referenced from other type:
276  * top-level anonymous enum won't be referenced by anything, while embedded
277  * one will.
278  */
btf_dump_mark_referenced(struct btf_dump * d)279 static int btf_dump_mark_referenced(struct btf_dump *d)
280 {
281 	int i, j, n = btf__get_nr_types(d->btf);
282 	const struct btf_type *t;
283 	__u16 vlen;
284 
285 	for (i = d->last_id + 1; i <= n; i++) {
286 		t = btf__type_by_id(d->btf, i);
287 		vlen = btf_vlen(t);
288 
289 		switch (btf_kind(t)) {
290 		case BTF_KIND_INT:
291 		case BTF_KIND_ENUM:
292 		case BTF_KIND_FWD:
293 			break;
294 
295 		case BTF_KIND_VOLATILE:
296 		case BTF_KIND_CONST:
297 		case BTF_KIND_RESTRICT:
298 		case BTF_KIND_PTR:
299 		case BTF_KIND_TYPEDEF:
300 		case BTF_KIND_FUNC:
301 		case BTF_KIND_VAR:
302 			d->type_states[t->type].referenced = 1;
303 			break;
304 
305 		case BTF_KIND_ARRAY: {
306 			const struct btf_array *a = btf_array(t);
307 
308 			d->type_states[a->index_type].referenced = 1;
309 			d->type_states[a->type].referenced = 1;
310 			break;
311 		}
312 		case BTF_KIND_STRUCT:
313 		case BTF_KIND_UNION: {
314 			const struct btf_member *m = btf_members(t);
315 
316 			for (j = 0; j < vlen; j++, m++)
317 				d->type_states[m->type].referenced = 1;
318 			break;
319 		}
320 		case BTF_KIND_FUNC_PROTO: {
321 			const struct btf_param *p = btf_params(t);
322 
323 			for (j = 0; j < vlen; j++, p++)
324 				d->type_states[p->type].referenced = 1;
325 			break;
326 		}
327 		case BTF_KIND_DATASEC: {
328 			const struct btf_var_secinfo *v = btf_var_secinfos(t);
329 
330 			for (j = 0; j < vlen; j++, v++)
331 				d->type_states[v->type].referenced = 1;
332 			break;
333 		}
334 		default:
335 			return -EINVAL;
336 		}
337 	}
338 	return 0;
339 }
340 
btf_dump_add_emit_queue_id(struct btf_dump * d,__u32 id)341 static int btf_dump_add_emit_queue_id(struct btf_dump *d, __u32 id)
342 {
343 	__u32 *new_queue;
344 	size_t new_cap;
345 
346 	if (d->emit_queue_cnt >= d->emit_queue_cap) {
347 		new_cap = max(16, d->emit_queue_cap * 3 / 2);
348 		new_queue = libbpf_reallocarray(d->emit_queue, new_cap, sizeof(new_queue[0]));
349 		if (!new_queue)
350 			return -ENOMEM;
351 		d->emit_queue = new_queue;
352 		d->emit_queue_cap = new_cap;
353 	}
354 
355 	d->emit_queue[d->emit_queue_cnt++] = id;
356 	return 0;
357 }
358 
359 /*
360  * Determine order of emitting dependent types and specified type to satisfy
361  * C compilation rules.  This is done through topological sorting with an
362  * additional complication which comes from C rules. The main idea for C is
363  * that if some type is "embedded" into a struct/union, it's size needs to be
364  * known at the time of definition of containing type. E.g., for:
365  *
366  *	struct A {};
367  *	struct B { struct A x; }
368  *
369  * struct A *HAS* to be defined before struct B, because it's "embedded",
370  * i.e., it is part of struct B layout. But in the following case:
371  *
372  *	struct A;
373  *	struct B { struct A *x; }
374  *	struct A {};
375  *
376  * it's enough to just have a forward declaration of struct A at the time of
377  * struct B definition, as struct B has a pointer to struct A, so the size of
378  * field x is known without knowing struct A size: it's sizeof(void *).
379  *
380  * Unfortunately, there are some trickier cases we need to handle, e.g.:
381  *
382  *	struct A {}; // if this was forward-declaration: compilation error
383  *	struct B {
384  *		struct { // anonymous struct
385  *			struct A y;
386  *		} *x;
387  *	};
388  *
389  * In this case, struct B's field x is a pointer, so it's size is known
390  * regardless of the size of (anonymous) struct it points to. But because this
391  * struct is anonymous and thus defined inline inside struct B, *and* it
392  * embeds struct A, compiler requires full definition of struct A to be known
393  * before struct B can be defined. This creates a transitive dependency
394  * between struct A and struct B. If struct A was forward-declared before
395  * struct B definition and fully defined after struct B definition, that would
396  * trigger compilation error.
397  *
398  * All this means that while we are doing topological sorting on BTF type
399  * graph, we need to determine relationships between different types (graph
400  * nodes):
401  *   - weak link (relationship) between X and Y, if Y *CAN* be
402  *   forward-declared at the point of X definition;
403  *   - strong link, if Y *HAS* to be fully-defined before X can be defined.
404  *
405  * The rule is as follows. Given a chain of BTF types from X to Y, if there is
406  * BTF_KIND_PTR type in the chain and at least one non-anonymous type
407  * Z (excluding X, including Y), then link is weak. Otherwise, it's strong.
408  * Weak/strong relationship is determined recursively during DFS traversal and
409  * is returned as a result from btf_dump_order_type().
410  *
411  * btf_dump_order_type() is trying to avoid unnecessary forward declarations,
412  * but it is not guaranteeing that no extraneous forward declarations will be
413  * emitted.
414  *
415  * To avoid extra work, algorithm marks some of BTF types as ORDERED, when
416  * it's done with them, but not for all (e.g., VOLATILE, CONST, RESTRICT,
417  * ARRAY, FUNC_PROTO), as weak/strong semantics for those depends on the
418  * entire graph path, so depending where from one came to that BTF type, it
419  * might cause weak or strong ordering. For types like STRUCT/UNION/INT/ENUM,
420  * once they are processed, there is no need to do it again, so they are
421  * marked as ORDERED. We can mark PTR as ORDERED as well, as it semi-forces
422  * weak link, unless subsequent referenced STRUCT/UNION/ENUM is anonymous. But
423  * in any case, once those are processed, no need to do it again, as the
424  * result won't change.
425  *
426  * Returns:
427  *   - 1, if type is part of strong link (so there is strong topological
428  *   ordering requirements);
429  *   - 0, if type is part of weak link (so can be satisfied through forward
430  *   declaration);
431  *   - <0, on error (e.g., unsatisfiable type loop detected).
432  */
btf_dump_order_type(struct btf_dump * d,__u32 id,bool through_ptr)433 static int btf_dump_order_type(struct btf_dump *d, __u32 id, bool through_ptr)
434 {
435 	/*
436 	 * Order state is used to detect strong link cycles, but only for BTF
437 	 * kinds that are or could be an independent definition (i.e.,
438 	 * stand-alone fwd decl, enum, typedef, struct, union). Ptrs, arrays,
439 	 * func_protos, modifiers are just means to get to these definitions.
440 	 * Int/void don't need definitions, they are assumed to be always
441 	 * properly defined.  We also ignore datasec, var, and funcs for now.
442 	 * So for all non-defining kinds, we never even set ordering state,
443 	 * for defining kinds we set ORDERING and subsequently ORDERED if it
444 	 * forms a strong link.
445 	 */
446 	struct btf_dump_type_aux_state *tstate = &d->type_states[id];
447 	const struct btf_type *t;
448 	__u16 vlen;
449 	int err, i;
450 
451 	/* return true, letting typedefs know that it's ok to be emitted */
452 	if (tstate->order_state == ORDERED)
453 		return 1;
454 
455 	t = btf__type_by_id(d->btf, id);
456 
457 	if (tstate->order_state == ORDERING) {
458 		/* type loop, but resolvable through fwd declaration */
459 		if (btf_is_composite(t) && through_ptr && t->name_off != 0)
460 			return 0;
461 		pr_warn("unsatisfiable type cycle, id:[%u]\n", id);
462 		return -ELOOP;
463 	}
464 
465 	switch (btf_kind(t)) {
466 	case BTF_KIND_INT:
467 		tstate->order_state = ORDERED;
468 		return 0;
469 
470 	case BTF_KIND_PTR:
471 		err = btf_dump_order_type(d, t->type, true);
472 		tstate->order_state = ORDERED;
473 		return err;
474 
475 	case BTF_KIND_ARRAY:
476 		return btf_dump_order_type(d, btf_array(t)->type, false);
477 
478 	case BTF_KIND_STRUCT:
479 	case BTF_KIND_UNION: {
480 		const struct btf_member *m = btf_members(t);
481 		/*
482 		 * struct/union is part of strong link, only if it's embedded
483 		 * (so no ptr in a path) or it's anonymous (so has to be
484 		 * defined inline, even if declared through ptr)
485 		 */
486 		if (through_ptr && t->name_off != 0)
487 			return 0;
488 
489 		tstate->order_state = ORDERING;
490 
491 		vlen = btf_vlen(t);
492 		for (i = 0; i < vlen; i++, m++) {
493 			err = btf_dump_order_type(d, m->type, false);
494 			if (err < 0)
495 				return err;
496 		}
497 
498 		if (t->name_off != 0) {
499 			err = btf_dump_add_emit_queue_id(d, id);
500 			if (err < 0)
501 				return err;
502 		}
503 
504 		tstate->order_state = ORDERED;
505 		return 1;
506 	}
507 	case BTF_KIND_ENUM:
508 	case BTF_KIND_FWD:
509 		/*
510 		 * non-anonymous or non-referenced enums are top-level
511 		 * declarations and should be emitted. Same logic can be
512 		 * applied to FWDs, it won't hurt anyways.
513 		 */
514 		if (t->name_off != 0 || !tstate->referenced) {
515 			err = btf_dump_add_emit_queue_id(d, id);
516 			if (err)
517 				return err;
518 		}
519 		tstate->order_state = ORDERED;
520 		return 1;
521 
522 	case BTF_KIND_TYPEDEF: {
523 		int is_strong;
524 
525 		is_strong = btf_dump_order_type(d, t->type, through_ptr);
526 		if (is_strong < 0)
527 			return is_strong;
528 
529 		/* typedef is similar to struct/union w.r.t. fwd-decls */
530 		if (through_ptr && !is_strong)
531 			return 0;
532 
533 		/* typedef is always a named definition */
534 		err = btf_dump_add_emit_queue_id(d, id);
535 		if (err)
536 			return err;
537 
538 		d->type_states[id].order_state = ORDERED;
539 		return 1;
540 	}
541 	case BTF_KIND_VOLATILE:
542 	case BTF_KIND_CONST:
543 	case BTF_KIND_RESTRICT:
544 		return btf_dump_order_type(d, t->type, through_ptr);
545 
546 	case BTF_KIND_FUNC_PROTO: {
547 		const struct btf_param *p = btf_params(t);
548 		bool is_strong;
549 
550 		err = btf_dump_order_type(d, t->type, through_ptr);
551 		if (err < 0)
552 			return err;
553 		is_strong = err > 0;
554 
555 		vlen = btf_vlen(t);
556 		for (i = 0; i < vlen; i++, p++) {
557 			err = btf_dump_order_type(d, p->type, through_ptr);
558 			if (err < 0)
559 				return err;
560 			if (err > 0)
561 				is_strong = true;
562 		}
563 		return is_strong;
564 	}
565 	case BTF_KIND_FUNC:
566 	case BTF_KIND_VAR:
567 	case BTF_KIND_DATASEC:
568 		d->type_states[id].order_state = ORDERED;
569 		return 0;
570 
571 	default:
572 		return -EINVAL;
573 	}
574 }
575 
576 static void btf_dump_emit_missing_aliases(struct btf_dump *d, __u32 id,
577 					  const struct btf_type *t);
578 
579 static void btf_dump_emit_struct_fwd(struct btf_dump *d, __u32 id,
580 				     const struct btf_type *t);
581 static void btf_dump_emit_struct_def(struct btf_dump *d, __u32 id,
582 				     const struct btf_type *t, int lvl);
583 
584 static void btf_dump_emit_enum_fwd(struct btf_dump *d, __u32 id,
585 				   const struct btf_type *t);
586 static void btf_dump_emit_enum_def(struct btf_dump *d, __u32 id,
587 				   const struct btf_type *t, int lvl);
588 
589 static void btf_dump_emit_fwd_def(struct btf_dump *d, __u32 id,
590 				  const struct btf_type *t);
591 
592 static void btf_dump_emit_typedef_def(struct btf_dump *d, __u32 id,
593 				      const struct btf_type *t, int lvl);
594 
595 /* a local view into a shared stack */
596 struct id_stack {
597 	const __u32 *ids;
598 	int cnt;
599 };
600 
601 static void btf_dump_emit_type_decl(struct btf_dump *d, __u32 id,
602 				    const char *fname, int lvl);
603 static void btf_dump_emit_type_chain(struct btf_dump *d,
604 				     struct id_stack *decl_stack,
605 				     const char *fname, int lvl);
606 
607 static const char *btf_dump_type_name(struct btf_dump *d, __u32 id);
608 static const char *btf_dump_ident_name(struct btf_dump *d, __u32 id);
609 static size_t btf_dump_name_dups(struct btf_dump *d, struct hashmap *name_map,
610 				 const char *orig_name);
611 
btf_dump_is_blacklisted(struct btf_dump * d,__u32 id)612 static bool btf_dump_is_blacklisted(struct btf_dump *d, __u32 id)
613 {
614 	const struct btf_type *t = btf__type_by_id(d->btf, id);
615 
616 	/* __builtin_va_list is a compiler built-in, which causes compilation
617 	 * errors, when compiling w/ different compiler, then used to compile
618 	 * original code (e.g., GCC to compile kernel, Clang to use generated
619 	 * C header from BTF). As it is built-in, it should be already defined
620 	 * properly internally in compiler.
621 	 */
622 	if (t->name_off == 0)
623 		return false;
624 	return strcmp(btf_name_of(d, t->name_off), "__builtin_va_list") == 0;
625 }
626 
627 /*
628  * Emit C-syntax definitions of types from chains of BTF types.
629  *
630  * High-level handling of determining necessary forward declarations are handled
631  * by btf_dump_emit_type() itself, but all nitty-gritty details of emitting type
632  * declarations/definitions in C syntax  are handled by a combo of
633  * btf_dump_emit_type_decl()/btf_dump_emit_type_chain() w/ delegation to
634  * corresponding btf_dump_emit_*_{def,fwd}() functions.
635  *
636  * We also keep track of "containing struct/union type ID" to determine when
637  * we reference it from inside and thus can avoid emitting unnecessary forward
638  * declaration.
639  *
640  * This algorithm is designed in such a way, that even if some error occurs
641  * (either technical, e.g., out of memory, or logical, i.e., malformed BTF
642  * that doesn't comply to C rules completely), algorithm will try to proceed
643  * and produce as much meaningful output as possible.
644  */
btf_dump_emit_type(struct btf_dump * d,__u32 id,__u32 cont_id)645 static void btf_dump_emit_type(struct btf_dump *d, __u32 id, __u32 cont_id)
646 {
647 	struct btf_dump_type_aux_state *tstate = &d->type_states[id];
648 	bool top_level_def = cont_id == 0;
649 	const struct btf_type *t;
650 	__u16 kind;
651 
652 	if (tstate->emit_state == EMITTED)
653 		return;
654 
655 	t = btf__type_by_id(d->btf, id);
656 	kind = btf_kind(t);
657 
658 	if (tstate->emit_state == EMITTING) {
659 		if (tstate->fwd_emitted)
660 			return;
661 
662 		switch (kind) {
663 		case BTF_KIND_STRUCT:
664 		case BTF_KIND_UNION:
665 			/*
666 			 * if we are referencing a struct/union that we are
667 			 * part of - then no need for fwd declaration
668 			 */
669 			if (id == cont_id)
670 				return;
671 			if (t->name_off == 0) {
672 				pr_warn("anonymous struct/union loop, id:[%u]\n",
673 					id);
674 				return;
675 			}
676 			btf_dump_emit_struct_fwd(d, id, t);
677 			btf_dump_printf(d, ";\n\n");
678 			tstate->fwd_emitted = 1;
679 			break;
680 		case BTF_KIND_TYPEDEF:
681 			/*
682 			 * for typedef fwd_emitted means typedef definition
683 			 * was emitted, but it can be used only for "weak"
684 			 * references through pointer only, not for embedding
685 			 */
686 			if (!btf_dump_is_blacklisted(d, id)) {
687 				btf_dump_emit_typedef_def(d, id, t, 0);
688 				btf_dump_printf(d, ";\n\n");
689 			}
690 			tstate->fwd_emitted = 1;
691 			break;
692 		default:
693 			break;
694 		}
695 
696 		return;
697 	}
698 
699 	switch (kind) {
700 	case BTF_KIND_INT:
701 		/* Emit type alias definitions if necessary */
702 		btf_dump_emit_missing_aliases(d, id, t);
703 
704 		tstate->emit_state = EMITTED;
705 		break;
706 	case BTF_KIND_ENUM:
707 		if (top_level_def) {
708 			btf_dump_emit_enum_def(d, id, t, 0);
709 			btf_dump_printf(d, ";\n\n");
710 		}
711 		tstate->emit_state = EMITTED;
712 		break;
713 	case BTF_KIND_PTR:
714 	case BTF_KIND_VOLATILE:
715 	case BTF_KIND_CONST:
716 	case BTF_KIND_RESTRICT:
717 		btf_dump_emit_type(d, t->type, cont_id);
718 		break;
719 	case BTF_KIND_ARRAY:
720 		btf_dump_emit_type(d, btf_array(t)->type, cont_id);
721 		break;
722 	case BTF_KIND_FWD:
723 		btf_dump_emit_fwd_def(d, id, t);
724 		btf_dump_printf(d, ";\n\n");
725 		tstate->emit_state = EMITTED;
726 		break;
727 	case BTF_KIND_TYPEDEF:
728 		tstate->emit_state = EMITTING;
729 		btf_dump_emit_type(d, t->type, id);
730 		/*
731 		 * typedef can server as both definition and forward
732 		 * declaration; at this stage someone depends on
733 		 * typedef as a forward declaration (refers to it
734 		 * through pointer), so unless we already did it,
735 		 * emit typedef as a forward declaration
736 		 */
737 		if (!tstate->fwd_emitted && !btf_dump_is_blacklisted(d, id)) {
738 			btf_dump_emit_typedef_def(d, id, t, 0);
739 			btf_dump_printf(d, ";\n\n");
740 		}
741 		tstate->emit_state = EMITTED;
742 		break;
743 	case BTF_KIND_STRUCT:
744 	case BTF_KIND_UNION:
745 		tstate->emit_state = EMITTING;
746 		/* if it's a top-level struct/union definition or struct/union
747 		 * is anonymous, then in C we'll be emitting all fields and
748 		 * their types (as opposed to just `struct X`), so we need to
749 		 * make sure that all types, referenced from struct/union
750 		 * members have necessary forward-declarations, where
751 		 * applicable
752 		 */
753 		if (top_level_def || t->name_off == 0) {
754 			const struct btf_member *m = btf_members(t);
755 			__u16 vlen = btf_vlen(t);
756 			int i, new_cont_id;
757 
758 			new_cont_id = t->name_off == 0 ? cont_id : id;
759 			for (i = 0; i < vlen; i++, m++)
760 				btf_dump_emit_type(d, m->type, new_cont_id);
761 		} else if (!tstate->fwd_emitted && id != cont_id) {
762 			btf_dump_emit_struct_fwd(d, id, t);
763 			btf_dump_printf(d, ";\n\n");
764 			tstate->fwd_emitted = 1;
765 		}
766 
767 		if (top_level_def) {
768 			btf_dump_emit_struct_def(d, id, t, 0);
769 			btf_dump_printf(d, ";\n\n");
770 			tstate->emit_state = EMITTED;
771 		} else {
772 			tstate->emit_state = NOT_EMITTED;
773 		}
774 		break;
775 	case BTF_KIND_FUNC_PROTO: {
776 		const struct btf_param *p = btf_params(t);
777 		__u16 vlen = btf_vlen(t);
778 		int i;
779 
780 		btf_dump_emit_type(d, t->type, cont_id);
781 		for (i = 0; i < vlen; i++, p++)
782 			btf_dump_emit_type(d, p->type, cont_id);
783 
784 		break;
785 	}
786 	default:
787 		break;
788 	}
789 }
790 
btf_is_struct_packed(const struct btf * btf,__u32 id,const struct btf_type * t)791 static bool btf_is_struct_packed(const struct btf *btf, __u32 id,
792 				 const struct btf_type *t)
793 {
794 	const struct btf_member *m;
795 	int max_align = 1, align, i, bit_sz;
796 	__u16 vlen;
797 
798 	m = btf_members(t);
799 	vlen = btf_vlen(t);
800 	/* all non-bitfield fields have to be naturally aligned */
801 	for (i = 0; i < vlen; i++, m++) {
802 		align = btf__align_of(btf, m->type);
803 		bit_sz = btf_member_bitfield_size(t, i);
804 		if (align && bit_sz == 0 && m->offset % (8 * align) != 0)
805 			return true;
806 		max_align = max(align, max_align);
807 	}
808 	/* size of a non-packed struct has to be a multiple of its alignment */
809 	if (t->size % max_align != 0)
810 		return true;
811 	/*
812 	 * if original struct was marked as packed, but its layout is
813 	 * naturally aligned, we'll detect that it's not packed
814 	 */
815 	return false;
816 }
817 
btf_dump_emit_bit_padding(const struct btf_dump * d,int cur_off,int next_off,int next_align,bool in_bitfield,int lvl)818 static void btf_dump_emit_bit_padding(const struct btf_dump *d,
819 				      int cur_off, int next_off, int next_align,
820 				      bool in_bitfield, int lvl)
821 {
822 	const struct {
823 		const char *name;
824 		int bits;
825 	} pads[] = {
826 		{"long", d->ptr_sz * 8}, {"int", 32}, {"short", 16}, {"char", 8}
827 	};
828 	int new_off, pad_bits, bits, i;
829 	const char *pad_type;
830 
831 	if (cur_off >= next_off)
832 		return; /* no gap */
833 
834 	/* For filling out padding we want to take advantage of
835 	 * natural alignment rules to minimize unnecessary explicit
836 	 * padding. First, we find the largest type (among long, int,
837 	 * short, or char) that can be used to force naturally aligned
838 	 * boundary. Once determined, we'll use such type to fill in
839 	 * the remaining padding gap. In some cases we can rely on
840 	 * compiler filling some gaps, but sometimes we need to force
841 	 * alignment to close natural alignment with markers like
842 	 * `long: 0` (this is always the case for bitfields).  Note
843 	 * that even if struct itself has, let's say 4-byte alignment
844 	 * (i.e., it only uses up to int-aligned types), using `long:
845 	 * X;` explicit padding doesn't actually change struct's
846 	 * overall alignment requirements, but compiler does take into
847 	 * account that type's (long, in this example) natural
848 	 * alignment requirements when adding implicit padding. We use
849 	 * this fact heavily and don't worry about ruining correct
850 	 * struct alignment requirement.
851 	 */
852 	for (i = 0; i < ARRAY_SIZE(pads); i++) {
853 		pad_bits = pads[i].bits;
854 		pad_type = pads[i].name;
855 
856 		new_off = roundup(cur_off, pad_bits);
857 		if (new_off <= next_off)
858 			break;
859 	}
860 
861 	if (new_off > cur_off && new_off <= next_off) {
862 		/* We need explicit `<type>: 0` aligning mark if next
863 		 * field is right on alignment offset and its
864 		 * alignment requirement is less strict than <type>'s
865 		 * alignment (so compiler won't naturally align to the
866 		 * offset we expect), or if subsequent `<type>: X`,
867 		 * will actually completely fit in the remaining hole,
868 		 * making compiler basically ignore `<type>: X`
869 		 * completely.
870 		 */
871 		if (in_bitfield ||
872 		    (new_off == next_off && roundup(cur_off, next_align * 8) != new_off) ||
873 		    (new_off != next_off && next_off - new_off <= new_off - cur_off))
874 			/* but for bitfields we'll emit explicit bit count */
875 			btf_dump_printf(d, "\n%s%s: %d;", pfx(lvl), pad_type,
876 					in_bitfield ? new_off - cur_off : 0);
877 		cur_off = new_off;
878 	}
879 
880 	/* Now we know we start at naturally aligned offset for a chosen
881 	 * padding type (long, int, short, or char), and so the rest is just
882 	 * a straightforward filling of remaining padding gap with full
883 	 * `<type>: sizeof(<type>);` markers, except for the last one, which
884 	 * might need smaller than sizeof(<type>) padding.
885 	 */
886 	while (cur_off != next_off) {
887 		bits = min(next_off - cur_off, pad_bits);
888 		if (bits == pad_bits) {
889 			btf_dump_printf(d, "\n%s%s: %d;", pfx(lvl), pad_type, pad_bits);
890 			cur_off += bits;
891 			continue;
892 		}
893 		/* For the remainder padding that doesn't cover entire
894 		 * pad_type bit length, we pick the smallest necessary type.
895 		 * This is pure aesthetics, we could have just used `long`,
896 		 * but having smallest necessary one communicates better the
897 		 * scale of the padding gap.
898 		 */
899 		for (i = ARRAY_SIZE(pads) - 1; i >= 0; i--) {
900 			pad_type = pads[i].name;
901 			pad_bits = pads[i].bits;
902 			if (pad_bits < bits)
903 				continue;
904 
905 			btf_dump_printf(d, "\n%s%s: %d;", pfx(lvl), pad_type, bits);
906 			cur_off += bits;
907 			break;
908 		}
909 	}
910 }
911 
btf_dump_emit_struct_fwd(struct btf_dump * d,__u32 id,const struct btf_type * t)912 static void btf_dump_emit_struct_fwd(struct btf_dump *d, __u32 id,
913 				     const struct btf_type *t)
914 {
915 	btf_dump_printf(d, "%s %s",
916 			btf_is_struct(t) ? "struct" : "union",
917 			btf_dump_type_name(d, id));
918 }
919 
btf_dump_emit_struct_def(struct btf_dump * d,__u32 id,const struct btf_type * t,int lvl)920 static void btf_dump_emit_struct_def(struct btf_dump *d,
921 				     __u32 id,
922 				     const struct btf_type *t,
923 				     int lvl)
924 {
925 	const struct btf_member *m = btf_members(t);
926 	bool is_struct = btf_is_struct(t);
927 	bool packed, prev_bitfield = false;
928 	int align, i, off = 0;
929 	__u16 vlen = btf_vlen(t);
930 
931 	align = btf__align_of(d->btf, id);
932 	packed = is_struct ? btf_is_struct_packed(d->btf, id, t) : 0;
933 
934 	btf_dump_printf(d, "%s%s%s {",
935 			is_struct ? "struct" : "union",
936 			t->name_off ? " " : "",
937 			btf_dump_type_name(d, id));
938 
939 	for (i = 0; i < vlen; i++, m++) {
940 		const char *fname;
941 		int m_off, m_sz, m_align;
942 		bool in_bitfield;
943 
944 		fname = btf_name_of(d, m->name_off);
945 		m_sz = btf_member_bitfield_size(t, i);
946 		m_off = btf_member_bit_offset(t, i);
947 		m_align = packed ? 1 : btf__align_of(d->btf, m->type);
948 
949 		in_bitfield = prev_bitfield && m_sz != 0;
950 
951 		btf_dump_emit_bit_padding(d, off, m_off, m_align, in_bitfield, lvl + 1);
952 		btf_dump_printf(d, "\n%s", pfx(lvl + 1));
953 		btf_dump_emit_type_decl(d, m->type, fname, lvl + 1);
954 
955 		if (m_sz) {
956 			btf_dump_printf(d, ": %d", m_sz);
957 			off = m_off + m_sz;
958 			prev_bitfield = true;
959 		} else {
960 			m_sz = max((__s64)0, btf__resolve_size(d->btf, m->type));
961 			off = m_off + m_sz * 8;
962 			prev_bitfield = false;
963 		}
964 
965 		btf_dump_printf(d, ";");
966 	}
967 
968 	/* pad at the end, if necessary */
969 	if (is_struct)
970 		btf_dump_emit_bit_padding(d, off, t->size * 8, align, false, lvl + 1);
971 
972 	/*
973 	 * Keep `struct empty {}` on a single line,
974 	 * only print newline when there are regular or padding fields.
975 	 */
976 	if (vlen || t->size) {
977 		btf_dump_printf(d, "\n");
978 		btf_dump_printf(d, "%s}", pfx(lvl));
979 	} else {
980 		btf_dump_printf(d, "}");
981 	}
982 	if (packed)
983 		btf_dump_printf(d, " __attribute__((packed))");
984 }
985 
986 static const char *missing_base_types[][2] = {
987 	/*
988 	 * GCC emits typedefs to its internal __PolyX_t types when compiling Arm
989 	 * SIMD intrinsics. Alias them to standard base types.
990 	 */
991 	{ "__Poly8_t",		"unsigned char" },
992 	{ "__Poly16_t",		"unsigned short" },
993 	{ "__Poly64_t",		"unsigned long long" },
994 	{ "__Poly128_t",	"unsigned __int128" },
995 };
996 
btf_dump_emit_missing_aliases(struct btf_dump * d,__u32 id,const struct btf_type * t)997 static void btf_dump_emit_missing_aliases(struct btf_dump *d, __u32 id,
998 					  const struct btf_type *t)
999 {
1000 	const char *name = btf_dump_type_name(d, id);
1001 	int i;
1002 
1003 	for (i = 0; i < ARRAY_SIZE(missing_base_types); i++) {
1004 		if (strcmp(name, missing_base_types[i][0]) == 0) {
1005 			btf_dump_printf(d, "typedef %s %s;\n\n",
1006 					missing_base_types[i][1], name);
1007 			break;
1008 		}
1009 	}
1010 }
1011 
btf_dump_emit_enum_fwd(struct btf_dump * d,__u32 id,const struct btf_type * t)1012 static void btf_dump_emit_enum_fwd(struct btf_dump *d, __u32 id,
1013 				   const struct btf_type *t)
1014 {
1015 	btf_dump_printf(d, "enum %s", btf_dump_type_name(d, id));
1016 }
1017 
btf_dump_emit_enum_def(struct btf_dump * d,__u32 id,const struct btf_type * t,int lvl)1018 static void btf_dump_emit_enum_def(struct btf_dump *d, __u32 id,
1019 				   const struct btf_type *t,
1020 				   int lvl)
1021 {
1022 	const struct btf_enum *v = btf_enum(t);
1023 	__u16 vlen = btf_vlen(t);
1024 	const char *name;
1025 	size_t dup_cnt;
1026 	int i;
1027 
1028 	btf_dump_printf(d, "enum%s%s",
1029 			t->name_off ? " " : "",
1030 			btf_dump_type_name(d, id));
1031 
1032 	if (vlen) {
1033 		btf_dump_printf(d, " {");
1034 		for (i = 0; i < vlen; i++, v++) {
1035 			name = btf_name_of(d, v->name_off);
1036 			/* enumerators share namespace with typedef idents */
1037 			dup_cnt = btf_dump_name_dups(d, d->ident_names, name);
1038 			if (dup_cnt > 1) {
1039 				btf_dump_printf(d, "\n%s%s___%zu = %u,",
1040 						pfx(lvl + 1), name, dup_cnt,
1041 						(__u32)v->val);
1042 			} else {
1043 				btf_dump_printf(d, "\n%s%s = %u,",
1044 						pfx(lvl + 1), name,
1045 						(__u32)v->val);
1046 			}
1047 		}
1048 		btf_dump_printf(d, "\n%s}", pfx(lvl));
1049 	}
1050 }
1051 
btf_dump_emit_fwd_def(struct btf_dump * d,__u32 id,const struct btf_type * t)1052 static void btf_dump_emit_fwd_def(struct btf_dump *d, __u32 id,
1053 				  const struct btf_type *t)
1054 {
1055 	const char *name = btf_dump_type_name(d, id);
1056 
1057 	if (btf_kflag(t))
1058 		btf_dump_printf(d, "union %s", name);
1059 	else
1060 		btf_dump_printf(d, "struct %s", name);
1061 }
1062 
btf_dump_emit_typedef_def(struct btf_dump * d,__u32 id,const struct btf_type * t,int lvl)1063 static void btf_dump_emit_typedef_def(struct btf_dump *d, __u32 id,
1064 				     const struct btf_type *t, int lvl)
1065 {
1066 	const char *name = btf_dump_ident_name(d, id);
1067 
1068 	/*
1069 	 * Old GCC versions are emitting invalid typedef for __gnuc_va_list
1070 	 * pointing to VOID. This generates warnings from btf_dump() and
1071 	 * results in uncompilable header file, so we are fixing it up here
1072 	 * with valid typedef into __builtin_va_list.
1073 	 */
1074 	if (t->type == 0 && strcmp(name, "__gnuc_va_list") == 0) {
1075 		btf_dump_printf(d, "typedef __builtin_va_list __gnuc_va_list");
1076 		return;
1077 	}
1078 
1079 	btf_dump_printf(d, "typedef ");
1080 	btf_dump_emit_type_decl(d, t->type, name, lvl);
1081 }
1082 
btf_dump_push_decl_stack_id(struct btf_dump * d,__u32 id)1083 static int btf_dump_push_decl_stack_id(struct btf_dump *d, __u32 id)
1084 {
1085 	__u32 *new_stack;
1086 	size_t new_cap;
1087 
1088 	if (d->decl_stack_cnt >= d->decl_stack_cap) {
1089 		new_cap = max(16, d->decl_stack_cap * 3 / 2);
1090 		new_stack = libbpf_reallocarray(d->decl_stack, new_cap, sizeof(new_stack[0]));
1091 		if (!new_stack)
1092 			return -ENOMEM;
1093 		d->decl_stack = new_stack;
1094 		d->decl_stack_cap = new_cap;
1095 	}
1096 
1097 	d->decl_stack[d->decl_stack_cnt++] = id;
1098 
1099 	return 0;
1100 }
1101 
1102 /*
1103  * Emit type declaration (e.g., field type declaration in a struct or argument
1104  * declaration in function prototype) in correct C syntax.
1105  *
1106  * For most types it's trivial, but there are few quirky type declaration
1107  * cases worth mentioning:
1108  *   - function prototypes (especially nesting of function prototypes);
1109  *   - arrays;
1110  *   - const/volatile/restrict for pointers vs other types.
1111  *
1112  * For a good discussion of *PARSING* C syntax (as a human), see
1113  * Peter van der Linden's "Expert C Programming: Deep C Secrets",
1114  * Ch.3 "Unscrambling Declarations in C".
1115  *
1116  * It won't help with BTF to C conversion much, though, as it's an opposite
1117  * problem. So we came up with this algorithm in reverse to van der Linden's
1118  * parsing algorithm. It goes from structured BTF representation of type
1119  * declaration to a valid compilable C syntax.
1120  *
1121  * For instance, consider this C typedef:
1122  *	typedef const int * const * arr[10] arr_t;
1123  * It will be represented in BTF with this chain of BTF types:
1124  *	[typedef] -> [array] -> [ptr] -> [const] -> [ptr] -> [const] -> [int]
1125  *
1126  * Notice how [const] modifier always goes before type it modifies in BTF type
1127  * graph, but in C syntax, const/volatile/restrict modifiers are written to
1128  * the right of pointers, but to the left of other types. There are also other
1129  * quirks, like function pointers, arrays of them, functions returning other
1130  * functions, etc.
1131  *
1132  * We handle that by pushing all the types to a stack, until we hit "terminal"
1133  * type (int/enum/struct/union/fwd). Then depending on the kind of a type on
1134  * top of a stack, modifiers are handled differently. Array/function pointers
1135  * have also wildly different syntax and how nesting of them are done. See
1136  * code for authoritative definition.
1137  *
1138  * To avoid allocating new stack for each independent chain of BTF types, we
1139  * share one bigger stack, with each chain working only on its own local view
1140  * of a stack frame. Some care is required to "pop" stack frames after
1141  * processing type declaration chain.
1142  */
btf_dump__emit_type_decl(struct btf_dump * d,__u32 id,const struct btf_dump_emit_type_decl_opts * opts)1143 int btf_dump__emit_type_decl(struct btf_dump *d, __u32 id,
1144 			     const struct btf_dump_emit_type_decl_opts *opts)
1145 {
1146 	const char *fname;
1147 	int lvl, err;
1148 
1149 	if (!OPTS_VALID(opts, btf_dump_emit_type_decl_opts))
1150 		return -EINVAL;
1151 
1152 	err = btf_dump_resize(d);
1153 	if (err)
1154 		return -EINVAL;
1155 
1156 	fname = OPTS_GET(opts, field_name, "");
1157 	lvl = OPTS_GET(opts, indent_level, 0);
1158 	d->strip_mods = OPTS_GET(opts, strip_mods, false);
1159 	btf_dump_emit_type_decl(d, id, fname, lvl);
1160 	d->strip_mods = false;
1161 	return 0;
1162 }
1163 
btf_dump_emit_type_decl(struct btf_dump * d,__u32 id,const char * fname,int lvl)1164 static void btf_dump_emit_type_decl(struct btf_dump *d, __u32 id,
1165 				    const char *fname, int lvl)
1166 {
1167 	struct id_stack decl_stack;
1168 	const struct btf_type *t;
1169 	int err, stack_start;
1170 
1171 	stack_start = d->decl_stack_cnt;
1172 	for (;;) {
1173 		t = btf__type_by_id(d->btf, id);
1174 		if (d->strip_mods && btf_is_mod(t))
1175 			goto skip_mod;
1176 
1177 		err = btf_dump_push_decl_stack_id(d, id);
1178 		if (err < 0) {
1179 			/*
1180 			 * if we don't have enough memory for entire type decl
1181 			 * chain, restore stack, emit warning, and try to
1182 			 * proceed nevertheless
1183 			 */
1184 			pr_warn("not enough memory for decl stack:%d", err);
1185 			d->decl_stack_cnt = stack_start;
1186 			return;
1187 		}
1188 skip_mod:
1189 		/* VOID */
1190 		if (id == 0)
1191 			break;
1192 
1193 		switch (btf_kind(t)) {
1194 		case BTF_KIND_PTR:
1195 		case BTF_KIND_VOLATILE:
1196 		case BTF_KIND_CONST:
1197 		case BTF_KIND_RESTRICT:
1198 		case BTF_KIND_FUNC_PROTO:
1199 			id = t->type;
1200 			break;
1201 		case BTF_KIND_ARRAY:
1202 			id = btf_array(t)->type;
1203 			break;
1204 		case BTF_KIND_INT:
1205 		case BTF_KIND_ENUM:
1206 		case BTF_KIND_FWD:
1207 		case BTF_KIND_STRUCT:
1208 		case BTF_KIND_UNION:
1209 		case BTF_KIND_TYPEDEF:
1210 			goto done;
1211 		default:
1212 			pr_warn("unexpected type in decl chain, kind:%u, id:[%u]\n",
1213 				btf_kind(t), id);
1214 			goto done;
1215 		}
1216 	}
1217 done:
1218 	/*
1219 	 * We might be inside a chain of declarations (e.g., array of function
1220 	 * pointers returning anonymous (so inlined) structs, having another
1221 	 * array field). Each of those needs its own "stack frame" to handle
1222 	 * emitting of declarations. Those stack frames are non-overlapping
1223 	 * portions of shared btf_dump->decl_stack. To make it a bit nicer to
1224 	 * handle this set of nested stacks, we create a view corresponding to
1225 	 * our own "stack frame" and work with it as an independent stack.
1226 	 * We'll need to clean up after emit_type_chain() returns, though.
1227 	 */
1228 	decl_stack.ids = d->decl_stack + stack_start;
1229 	decl_stack.cnt = d->decl_stack_cnt - stack_start;
1230 	btf_dump_emit_type_chain(d, &decl_stack, fname, lvl);
1231 	/*
1232 	 * emit_type_chain() guarantees that it will pop its entire decl_stack
1233 	 * frame before returning. But it works with a read-only view into
1234 	 * decl_stack, so it doesn't actually pop anything from the
1235 	 * perspective of shared btf_dump->decl_stack, per se. We need to
1236 	 * reset decl_stack state to how it was before us to avoid it growing
1237 	 * all the time.
1238 	 */
1239 	d->decl_stack_cnt = stack_start;
1240 }
1241 
btf_dump_emit_mods(struct btf_dump * d,struct id_stack * decl_stack)1242 static void btf_dump_emit_mods(struct btf_dump *d, struct id_stack *decl_stack)
1243 {
1244 	const struct btf_type *t;
1245 	__u32 id;
1246 
1247 	while (decl_stack->cnt) {
1248 		id = decl_stack->ids[decl_stack->cnt - 1];
1249 		t = btf__type_by_id(d->btf, id);
1250 
1251 		switch (btf_kind(t)) {
1252 		case BTF_KIND_VOLATILE:
1253 			btf_dump_printf(d, "volatile ");
1254 			break;
1255 		case BTF_KIND_CONST:
1256 			btf_dump_printf(d, "const ");
1257 			break;
1258 		case BTF_KIND_RESTRICT:
1259 			btf_dump_printf(d, "restrict ");
1260 			break;
1261 		default:
1262 			return;
1263 		}
1264 		decl_stack->cnt--;
1265 	}
1266 }
1267 
btf_dump_drop_mods(struct btf_dump * d,struct id_stack * decl_stack)1268 static void btf_dump_drop_mods(struct btf_dump *d, struct id_stack *decl_stack)
1269 {
1270 	const struct btf_type *t;
1271 	__u32 id;
1272 
1273 	while (decl_stack->cnt) {
1274 		id = decl_stack->ids[decl_stack->cnt - 1];
1275 		t = btf__type_by_id(d->btf, id);
1276 		if (!btf_is_mod(t))
1277 			return;
1278 		decl_stack->cnt--;
1279 	}
1280 }
1281 
btf_dump_emit_name(const struct btf_dump * d,const char * name,bool last_was_ptr)1282 static void btf_dump_emit_name(const struct btf_dump *d,
1283 			       const char *name, bool last_was_ptr)
1284 {
1285 	bool separate = name[0] && !last_was_ptr;
1286 
1287 	btf_dump_printf(d, "%s%s", separate ? " " : "", name);
1288 }
1289 
btf_dump_emit_type_chain(struct btf_dump * d,struct id_stack * decls,const char * fname,int lvl)1290 static void btf_dump_emit_type_chain(struct btf_dump *d,
1291 				     struct id_stack *decls,
1292 				     const char *fname, int lvl)
1293 {
1294 	/*
1295 	 * last_was_ptr is used to determine if we need to separate pointer
1296 	 * asterisk (*) from previous part of type signature with space, so
1297 	 * that we get `int ***`, instead of `int * * *`. We default to true
1298 	 * for cases where we have single pointer in a chain. E.g., in ptr ->
1299 	 * func_proto case. func_proto will start a new emit_type_chain call
1300 	 * with just ptr, which should be emitted as (*) or (*<fname>), so we
1301 	 * don't want to prepend space for that last pointer.
1302 	 */
1303 	bool last_was_ptr = true;
1304 	const struct btf_type *t;
1305 	const char *name;
1306 	__u16 kind;
1307 	__u32 id;
1308 
1309 	while (decls->cnt) {
1310 		id = decls->ids[--decls->cnt];
1311 		if (id == 0) {
1312 			/* VOID is a special snowflake */
1313 			btf_dump_emit_mods(d, decls);
1314 			btf_dump_printf(d, "void");
1315 			last_was_ptr = false;
1316 			continue;
1317 		}
1318 
1319 		t = btf__type_by_id(d->btf, id);
1320 		kind = btf_kind(t);
1321 
1322 		switch (kind) {
1323 		case BTF_KIND_INT:
1324 			btf_dump_emit_mods(d, decls);
1325 			name = btf_name_of(d, t->name_off);
1326 			btf_dump_printf(d, "%s", name);
1327 			break;
1328 		case BTF_KIND_STRUCT:
1329 		case BTF_KIND_UNION:
1330 			btf_dump_emit_mods(d, decls);
1331 			/* inline anonymous struct/union */
1332 			if (t->name_off == 0)
1333 				btf_dump_emit_struct_def(d, id, t, lvl);
1334 			else
1335 				btf_dump_emit_struct_fwd(d, id, t);
1336 			break;
1337 		case BTF_KIND_ENUM:
1338 			btf_dump_emit_mods(d, decls);
1339 			/* inline anonymous enum */
1340 			if (t->name_off == 0)
1341 				btf_dump_emit_enum_def(d, id, t, lvl);
1342 			else
1343 				btf_dump_emit_enum_fwd(d, id, t);
1344 			break;
1345 		case BTF_KIND_FWD:
1346 			btf_dump_emit_mods(d, decls);
1347 			btf_dump_emit_fwd_def(d, id, t);
1348 			break;
1349 		case BTF_KIND_TYPEDEF:
1350 			btf_dump_emit_mods(d, decls);
1351 			btf_dump_printf(d, "%s", btf_dump_ident_name(d, id));
1352 			break;
1353 		case BTF_KIND_PTR:
1354 			btf_dump_printf(d, "%s", last_was_ptr ? "*" : " *");
1355 			break;
1356 		case BTF_KIND_VOLATILE:
1357 			btf_dump_printf(d, " volatile");
1358 			break;
1359 		case BTF_KIND_CONST:
1360 			btf_dump_printf(d, " const");
1361 			break;
1362 		case BTF_KIND_RESTRICT:
1363 			btf_dump_printf(d, " restrict");
1364 			break;
1365 		case BTF_KIND_ARRAY: {
1366 			const struct btf_array *a = btf_array(t);
1367 			const struct btf_type *next_t;
1368 			__u32 next_id;
1369 			bool multidim;
1370 			/*
1371 			 * GCC has a bug
1372 			 * (https://gcc.gnu.org/bugzilla/show_bug.cgi?id=8354)
1373 			 * which causes it to emit extra const/volatile
1374 			 * modifiers for an array, if array's element type has
1375 			 * const/volatile modifiers. Clang doesn't do that.
1376 			 * In general, it doesn't seem very meaningful to have
1377 			 * a const/volatile modifier for array, so we are
1378 			 * going to silently skip them here.
1379 			 */
1380 			btf_dump_drop_mods(d, decls);
1381 
1382 			if (decls->cnt == 0) {
1383 				btf_dump_emit_name(d, fname, last_was_ptr);
1384 				btf_dump_printf(d, "[%u]", a->nelems);
1385 				return;
1386 			}
1387 
1388 			next_id = decls->ids[decls->cnt - 1];
1389 			next_t = btf__type_by_id(d->btf, next_id);
1390 			multidim = btf_is_array(next_t);
1391 			/* we need space if we have named non-pointer */
1392 			if (fname[0] && !last_was_ptr)
1393 				btf_dump_printf(d, " ");
1394 			/* no parentheses for multi-dimensional array */
1395 			if (!multidim)
1396 				btf_dump_printf(d, "(");
1397 			btf_dump_emit_type_chain(d, decls, fname, lvl);
1398 			if (!multidim)
1399 				btf_dump_printf(d, ")");
1400 			btf_dump_printf(d, "[%u]", a->nelems);
1401 			return;
1402 		}
1403 		case BTF_KIND_FUNC_PROTO: {
1404 			const struct btf_param *p = btf_params(t);
1405 			__u16 vlen = btf_vlen(t);
1406 			int i;
1407 
1408 			/*
1409 			 * GCC emits extra volatile qualifier for
1410 			 * __attribute__((noreturn)) function pointers. Clang
1411 			 * doesn't do it. It's a GCC quirk for backwards
1412 			 * compatibility with code written for GCC <2.5. So,
1413 			 * similarly to extra qualifiers for array, just drop
1414 			 * them, instead of handling them.
1415 			 */
1416 			btf_dump_drop_mods(d, decls);
1417 			if (decls->cnt) {
1418 				btf_dump_printf(d, " (");
1419 				btf_dump_emit_type_chain(d, decls, fname, lvl);
1420 				btf_dump_printf(d, ")");
1421 			} else {
1422 				btf_dump_emit_name(d, fname, last_was_ptr);
1423 			}
1424 			btf_dump_printf(d, "(");
1425 			/*
1426 			 * Clang for BPF target generates func_proto with no
1427 			 * args as a func_proto with a single void arg (e.g.,
1428 			 * `int (*f)(void)` vs just `int (*f)()`). We are
1429 			 * going to pretend there are no args for such case.
1430 			 */
1431 			if (vlen == 1 && p->type == 0) {
1432 				btf_dump_printf(d, ")");
1433 				return;
1434 			}
1435 
1436 			for (i = 0; i < vlen; i++, p++) {
1437 				if (i > 0)
1438 					btf_dump_printf(d, ", ");
1439 
1440 				/* last arg of type void is vararg */
1441 				if (i == vlen - 1 && p->type == 0) {
1442 					btf_dump_printf(d, "...");
1443 					break;
1444 				}
1445 
1446 				name = btf_name_of(d, p->name_off);
1447 				btf_dump_emit_type_decl(d, p->type, name, lvl);
1448 			}
1449 
1450 			btf_dump_printf(d, ")");
1451 			return;
1452 		}
1453 		default:
1454 			pr_warn("unexpected type in decl chain, kind:%u, id:[%u]\n",
1455 				kind, id);
1456 			return;
1457 		}
1458 
1459 		last_was_ptr = kind == BTF_KIND_PTR;
1460 	}
1461 
1462 	btf_dump_emit_name(d, fname, last_was_ptr);
1463 }
1464 
1465 /* return number of duplicates (occurrences) of a given name */
btf_dump_name_dups(struct btf_dump * d,struct hashmap * name_map,const char * orig_name)1466 static size_t btf_dump_name_dups(struct btf_dump *d, struct hashmap *name_map,
1467 				 const char *orig_name)
1468 {
1469 	char *old_name, *new_name;
1470 	size_t dup_cnt = 0;
1471 	int err;
1472 
1473 	new_name = strdup(orig_name);
1474 	if (!new_name)
1475 		return 1;
1476 
1477 	hashmap__find(name_map, orig_name, (void **)&dup_cnt);
1478 	dup_cnt++;
1479 
1480 	err = hashmap__set(name_map, new_name, (void *)dup_cnt,
1481 			   (const void **)&old_name, NULL);
1482 	if (err)
1483 		free(new_name);
1484 
1485 	free(old_name);
1486 
1487 	return dup_cnt;
1488 }
1489 
btf_dump_resolve_name(struct btf_dump * d,__u32 id,struct hashmap * name_map)1490 static const char *btf_dump_resolve_name(struct btf_dump *d, __u32 id,
1491 					 struct hashmap *name_map)
1492 {
1493 	struct btf_dump_type_aux_state *s = &d->type_states[id];
1494 	const struct btf_type *t = btf__type_by_id(d->btf, id);
1495 	const char *orig_name = btf_name_of(d, t->name_off);
1496 	const char **cached_name = &d->cached_names[id];
1497 	size_t dup_cnt;
1498 
1499 	if (t->name_off == 0)
1500 		return "";
1501 
1502 	if (s->name_resolved)
1503 		return *cached_name ? *cached_name : orig_name;
1504 
1505 	if (btf_is_fwd(t) || (btf_is_enum(t) && btf_vlen(t) == 0)) {
1506 		s->name_resolved = 1;
1507 		return orig_name;
1508 	}
1509 
1510 	dup_cnt = btf_dump_name_dups(d, name_map, orig_name);
1511 	if (dup_cnt > 1) {
1512 		const size_t max_len = 256;
1513 		char new_name[max_len];
1514 
1515 		snprintf(new_name, max_len, "%s___%zu", orig_name, dup_cnt);
1516 		*cached_name = strdup(new_name);
1517 	}
1518 
1519 	s->name_resolved = 1;
1520 	return *cached_name ? *cached_name : orig_name;
1521 }
1522 
btf_dump_type_name(struct btf_dump * d,__u32 id)1523 static const char *btf_dump_type_name(struct btf_dump *d, __u32 id)
1524 {
1525 	return btf_dump_resolve_name(d, id, d->type_names);
1526 }
1527 
btf_dump_ident_name(struct btf_dump * d,__u32 id)1528 static const char *btf_dump_ident_name(struct btf_dump *d, __u32 id)
1529 {
1530 	return btf_dump_resolve_name(d, id, d->ident_names);
1531 }
1532