1 // SPDX-License-Identifier: (LGPL-2.1 OR BSD-2-Clause)
4 * BTF-to-C type converter.
6 * Copyright (c) 2019 Facebook
14 #include <linux/err.h>
15 #include <linux/btf.h>
16 #include <linux/kernel.h>
20 #include "libbpf_internal.h"
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;
25 static const char *pfx(int lvl)
27 return lvl >= PREFIX_CNT ? PREFIXES : &PREFIXES[PREFIX_CNT - lvl];
30 enum btf_dump_type_order_state {
36 enum btf_dump_type_emit_state {
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 */
50 /* whether unique non-duplicate name was already assigned */
51 __u8 name_resolved: 1;
52 /* whether type is referenced from any other type */
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;
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;
72 /* topo-sorted list of dependent type definitions */
78 * stack of type declarations (e.g., chain of modifiers, arrays,
85 /* maps struct/union/enum name to a number of name occurrences */
86 struct hashmap *type_names;
88 * maps typedef identifiers and enum value names to a number of such
91 struct hashmap *ident_names;
94 static size_t str_hash_fn(const void *key, void *ctx)
99 static bool str_equal_fn(const void *a, const void *b, void *ctx)
101 return strcmp(a, b) == 0;
104 static const char *btf_name_of(const struct btf_dump *d, __u32 name_off)
106 return btf__name_by_offset(d->btf, name_off);
109 static void btf_dump_printf(const struct btf_dump *d, const char *fmt, ...)
114 d->printf_fn(d->opts.ctx, fmt, args);
118 static int btf_dump_mark_referenced(struct btf_dump *d);
119 static int btf_dump_resize(struct btf_dump *d);
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)
129 d = calloc(1, sizeof(struct btf_dump));
131 return libbpf_err_ptr(-ENOMEM);
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 *);
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;
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;
152 err = btf_dump_resize(d);
159 return libbpf_err_ptr(err);
162 static int btf_dump_resize(struct btf_dump *d)
164 int err, last_id = btf__get_nr_types(d->btf);
166 if (last_id <= d->last_id)
169 if (libbpf_ensure_mem((void **)&d->type_states, &d->type_states_cap,
170 sizeof(*d->type_states), last_id + 1))
172 if (libbpf_ensure_mem((void **)&d->cached_names, &d->cached_names_cap,
173 sizeof(*d->cached_names), last_id + 1))
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;
182 /* eagerly determine referenced types for anon enums */
183 err = btf_dump_mark_referenced(d);
187 d->last_id = last_id;
191 void btf_dump__free(struct btf_dump *d)
195 if (IS_ERR_OR_NULL(d))
198 free(d->type_states);
199 if (d->cached_names) {
200 /* any set cached name is owned by us and should be freed */
201 for (i = 0; i <= d->last_id; i++) {
202 if (d->cached_names[i])
203 free((void *)d->cached_names[i]);
206 free(d->cached_names);
209 hashmap__free(d->type_names);
210 hashmap__free(d->ident_names);
215 static int btf_dump_order_type(struct btf_dump *d, __u32 id, bool through_ptr);
216 static void btf_dump_emit_type(struct btf_dump *d, __u32 id, __u32 cont_id);
219 * Dump BTF type in a compilable C syntax, including all the necessary
220 * dependent types, necessary for compilation. If some of the dependent types
221 * were already emitted as part of previous btf_dump__dump_type() invocation
222 * for another type, they won't be emitted again. This API allows callers to
223 * filter out BTF types according to user-defined criterias and emitted only
224 * minimal subset of types, necessary to compile everything. Full struct/union
225 * definitions will still be emitted, even if the only usage is through
226 * pointer and could be satisfied with just a forward declaration.
228 * Dumping is done in two high-level passes:
229 * 1. Topologically sort type definitions to satisfy C rules of compilation.
230 * 2. Emit type definitions in C syntax.
232 * Returns 0 on success; <0, otherwise.
234 int btf_dump__dump_type(struct btf_dump *d, __u32 id)
238 if (id > btf__get_nr_types(d->btf))
239 return libbpf_err(-EINVAL);
241 err = btf_dump_resize(d);
243 return libbpf_err(err);
245 d->emit_queue_cnt = 0;
246 err = btf_dump_order_type(d, id, false);
248 return libbpf_err(err);
250 for (i = 0; i < d->emit_queue_cnt; i++)
251 btf_dump_emit_type(d, d->emit_queue[i], 0 /*top-level*/);
257 * Mark all types that are referenced from any other type. This is used to
258 * determine top-level anonymous enums that need to be emitted as an
259 * independent type declarations.
260 * Anonymous enums come in two flavors: either embedded in a struct's field
261 * definition, in which case they have to be declared inline as part of field
262 * type declaration; or as a top-level anonymous enum, typically used for
263 * declaring global constants. It's impossible to distinguish between two
264 * without knowning whether given enum type was referenced from other type:
265 * top-level anonymous enum won't be referenced by anything, while embedded
268 static int btf_dump_mark_referenced(struct btf_dump *d)
270 int i, j, n = btf__get_nr_types(d->btf);
271 const struct btf_type *t;
274 for (i = d->last_id + 1; i <= n; i++) {
275 t = btf__type_by_id(d->btf, i);
278 switch (btf_kind(t)) {
285 case BTF_KIND_VOLATILE:
287 case BTF_KIND_RESTRICT:
289 case BTF_KIND_TYPEDEF:
292 d->type_states[t->type].referenced = 1;
295 case BTF_KIND_ARRAY: {
296 const struct btf_array *a = btf_array(t);
298 d->type_states[a->index_type].referenced = 1;
299 d->type_states[a->type].referenced = 1;
302 case BTF_KIND_STRUCT:
303 case BTF_KIND_UNION: {
304 const struct btf_member *m = btf_members(t);
306 for (j = 0; j < vlen; j++, m++)
307 d->type_states[m->type].referenced = 1;
310 case BTF_KIND_FUNC_PROTO: {
311 const struct btf_param *p = btf_params(t);
313 for (j = 0; j < vlen; j++, p++)
314 d->type_states[p->type].referenced = 1;
317 case BTF_KIND_DATASEC: {
318 const struct btf_var_secinfo *v = btf_var_secinfos(t);
320 for (j = 0; j < vlen; j++, v++)
321 d->type_states[v->type].referenced = 1;
331 static int btf_dump_add_emit_queue_id(struct btf_dump *d, __u32 id)
336 if (d->emit_queue_cnt >= d->emit_queue_cap) {
337 new_cap = max(16, d->emit_queue_cap * 3 / 2);
338 new_queue = libbpf_reallocarray(d->emit_queue, new_cap, sizeof(new_queue[0]));
341 d->emit_queue = new_queue;
342 d->emit_queue_cap = new_cap;
345 d->emit_queue[d->emit_queue_cnt++] = id;
350 * Determine order of emitting dependent types and specified type to satisfy
351 * C compilation rules. This is done through topological sorting with an
352 * additional complication which comes from C rules. The main idea for C is
353 * that if some type is "embedded" into a struct/union, it's size needs to be
354 * known at the time of definition of containing type. E.g., for:
357 * struct B { struct A x; }
359 * struct A *HAS* to be defined before struct B, because it's "embedded",
360 * i.e., it is part of struct B layout. But in the following case:
363 * struct B { struct A *x; }
366 * it's enough to just have a forward declaration of struct A at the time of
367 * struct B definition, as struct B has a pointer to struct A, so the size of
368 * field x is known without knowing struct A size: it's sizeof(void *).
370 * Unfortunately, there are some trickier cases we need to handle, e.g.:
372 * struct A {}; // if this was forward-declaration: compilation error
374 * struct { // anonymous struct
379 * In this case, struct B's field x is a pointer, so it's size is known
380 * regardless of the size of (anonymous) struct it points to. But because this
381 * struct is anonymous and thus defined inline inside struct B, *and* it
382 * embeds struct A, compiler requires full definition of struct A to be known
383 * before struct B can be defined. This creates a transitive dependency
384 * between struct A and struct B. If struct A was forward-declared before
385 * struct B definition and fully defined after struct B definition, that would
386 * trigger compilation error.
388 * All this means that while we are doing topological sorting on BTF type
389 * graph, we need to determine relationships between different types (graph
391 * - weak link (relationship) between X and Y, if Y *CAN* be
392 * forward-declared at the point of X definition;
393 * - strong link, if Y *HAS* to be fully-defined before X can be defined.
395 * The rule is as follows. Given a chain of BTF types from X to Y, if there is
396 * BTF_KIND_PTR type in the chain and at least one non-anonymous type
397 * Z (excluding X, including Y), then link is weak. Otherwise, it's strong.
398 * Weak/strong relationship is determined recursively during DFS traversal and
399 * is returned as a result from btf_dump_order_type().
401 * btf_dump_order_type() is trying to avoid unnecessary forward declarations,
402 * but it is not guaranteeing that no extraneous forward declarations will be
405 * To avoid extra work, algorithm marks some of BTF types as ORDERED, when
406 * it's done with them, but not for all (e.g., VOLATILE, CONST, RESTRICT,
407 * ARRAY, FUNC_PROTO), as weak/strong semantics for those depends on the
408 * entire graph path, so depending where from one came to that BTF type, it
409 * might cause weak or strong ordering. For types like STRUCT/UNION/INT/ENUM,
410 * once they are processed, there is no need to do it again, so they are
411 * marked as ORDERED. We can mark PTR as ORDERED as well, as it semi-forces
412 * weak link, unless subsequent referenced STRUCT/UNION/ENUM is anonymous. But
413 * in any case, once those are processed, no need to do it again, as the
414 * result won't change.
417 * - 1, if type is part of strong link (so there is strong topological
418 * ordering requirements);
419 * - 0, if type is part of weak link (so can be satisfied through forward
421 * - <0, on error (e.g., unsatisfiable type loop detected).
423 static int btf_dump_order_type(struct btf_dump *d, __u32 id, bool through_ptr)
426 * Order state is used to detect strong link cycles, but only for BTF
427 * kinds that are or could be an independent definition (i.e.,
428 * stand-alone fwd decl, enum, typedef, struct, union). Ptrs, arrays,
429 * func_protos, modifiers are just means to get to these definitions.
430 * Int/void don't need definitions, they are assumed to be always
431 * properly defined. We also ignore datasec, var, and funcs for now.
432 * So for all non-defining kinds, we never even set ordering state,
433 * for defining kinds we set ORDERING and subsequently ORDERED if it
434 * forms a strong link.
436 struct btf_dump_type_aux_state *tstate = &d->type_states[id];
437 const struct btf_type *t;
441 /* return true, letting typedefs know that it's ok to be emitted */
442 if (tstate->order_state == ORDERED)
445 t = btf__type_by_id(d->btf, id);
447 if (tstate->order_state == ORDERING) {
448 /* type loop, but resolvable through fwd declaration */
449 if (btf_is_composite(t) && through_ptr && t->name_off != 0)
451 pr_warn("unsatisfiable type cycle, id:[%u]\n", id);
455 switch (btf_kind(t)) {
458 tstate->order_state = ORDERED;
462 err = btf_dump_order_type(d, t->type, true);
463 tstate->order_state = ORDERED;
467 return btf_dump_order_type(d, btf_array(t)->type, false);
469 case BTF_KIND_STRUCT:
470 case BTF_KIND_UNION: {
471 const struct btf_member *m = btf_members(t);
473 * struct/union is part of strong link, only if it's embedded
474 * (so no ptr in a path) or it's anonymous (so has to be
475 * defined inline, even if declared through ptr)
477 if (through_ptr && t->name_off != 0)
480 tstate->order_state = ORDERING;
483 for (i = 0; i < vlen; i++, m++) {
484 err = btf_dump_order_type(d, m->type, false);
489 if (t->name_off != 0) {
490 err = btf_dump_add_emit_queue_id(d, id);
495 tstate->order_state = ORDERED;
501 * non-anonymous or non-referenced enums are top-level
502 * declarations and should be emitted. Same logic can be
503 * applied to FWDs, it won't hurt anyways.
505 if (t->name_off != 0 || !tstate->referenced) {
506 err = btf_dump_add_emit_queue_id(d, id);
510 tstate->order_state = ORDERED;
513 case BTF_KIND_TYPEDEF: {
516 is_strong = btf_dump_order_type(d, t->type, through_ptr);
520 /* typedef is similar to struct/union w.r.t. fwd-decls */
521 if (through_ptr && !is_strong)
524 /* typedef is always a named definition */
525 err = btf_dump_add_emit_queue_id(d, id);
529 d->type_states[id].order_state = ORDERED;
532 case BTF_KIND_VOLATILE:
534 case BTF_KIND_RESTRICT:
535 return btf_dump_order_type(d, t->type, through_ptr);
537 case BTF_KIND_FUNC_PROTO: {
538 const struct btf_param *p = btf_params(t);
541 err = btf_dump_order_type(d, t->type, through_ptr);
547 for (i = 0; i < vlen; i++, p++) {
548 err = btf_dump_order_type(d, p->type, through_ptr);
558 case BTF_KIND_DATASEC:
559 d->type_states[id].order_state = ORDERED;
567 static void btf_dump_emit_missing_aliases(struct btf_dump *d, __u32 id,
568 const struct btf_type *t);
570 static void btf_dump_emit_struct_fwd(struct btf_dump *d, __u32 id,
571 const struct btf_type *t);
572 static void btf_dump_emit_struct_def(struct btf_dump *d, __u32 id,
573 const struct btf_type *t, int lvl);
575 static void btf_dump_emit_enum_fwd(struct btf_dump *d, __u32 id,
576 const struct btf_type *t);
577 static void btf_dump_emit_enum_def(struct btf_dump *d, __u32 id,
578 const struct btf_type *t, int lvl);
580 static void btf_dump_emit_fwd_def(struct btf_dump *d, __u32 id,
581 const struct btf_type *t);
583 static void btf_dump_emit_typedef_def(struct btf_dump *d, __u32 id,
584 const struct btf_type *t, int lvl);
586 /* a local view into a shared stack */
592 static void btf_dump_emit_type_decl(struct btf_dump *d, __u32 id,
593 const char *fname, int lvl);
594 static void btf_dump_emit_type_chain(struct btf_dump *d,
595 struct id_stack *decl_stack,
596 const char *fname, int lvl);
598 static const char *btf_dump_type_name(struct btf_dump *d, __u32 id);
599 static const char *btf_dump_ident_name(struct btf_dump *d, __u32 id);
600 static size_t btf_dump_name_dups(struct btf_dump *d, struct hashmap *name_map,
601 const char *orig_name);
603 static bool btf_dump_is_blacklisted(struct btf_dump *d, __u32 id)
605 const struct btf_type *t = btf__type_by_id(d->btf, id);
607 /* __builtin_va_list is a compiler built-in, which causes compilation
608 * errors, when compiling w/ different compiler, then used to compile
609 * original code (e.g., GCC to compile kernel, Clang to use generated
610 * C header from BTF). As it is built-in, it should be already defined
611 * properly internally in compiler.
613 if (t->name_off == 0)
615 return strcmp(btf_name_of(d, t->name_off), "__builtin_va_list") == 0;
619 * Emit C-syntax definitions of types from chains of BTF types.
621 * High-level handling of determining necessary forward declarations are handled
622 * by btf_dump_emit_type() itself, but all nitty-gritty details of emitting type
623 * declarations/definitions in C syntax are handled by a combo of
624 * btf_dump_emit_type_decl()/btf_dump_emit_type_chain() w/ delegation to
625 * corresponding btf_dump_emit_*_{def,fwd}() functions.
627 * We also keep track of "containing struct/union type ID" to determine when
628 * we reference it from inside and thus can avoid emitting unnecessary forward
631 * This algorithm is designed in such a way, that even if some error occurs
632 * (either technical, e.g., out of memory, or logical, i.e., malformed BTF
633 * that doesn't comply to C rules completely), algorithm will try to proceed
634 * and produce as much meaningful output as possible.
636 static void btf_dump_emit_type(struct btf_dump *d, __u32 id, __u32 cont_id)
638 struct btf_dump_type_aux_state *tstate = &d->type_states[id];
639 bool top_level_def = cont_id == 0;
640 const struct btf_type *t;
643 if (tstate->emit_state == EMITTED)
646 t = btf__type_by_id(d->btf, id);
649 if (tstate->emit_state == EMITTING) {
650 if (tstate->fwd_emitted)
654 case BTF_KIND_STRUCT:
657 * if we are referencing a struct/union that we are
658 * part of - then no need for fwd declaration
662 if (t->name_off == 0) {
663 pr_warn("anonymous struct/union loop, id:[%u]\n",
667 btf_dump_emit_struct_fwd(d, id, t);
668 btf_dump_printf(d, ";\n\n");
669 tstate->fwd_emitted = 1;
671 case BTF_KIND_TYPEDEF:
673 * for typedef fwd_emitted means typedef definition
674 * was emitted, but it can be used only for "weak"
675 * references through pointer only, not for embedding
677 if (!btf_dump_is_blacklisted(d, id)) {
678 btf_dump_emit_typedef_def(d, id, t, 0);
679 btf_dump_printf(d, ";\n\n");
681 tstate->fwd_emitted = 1;
692 /* Emit type alias definitions if necessary */
693 btf_dump_emit_missing_aliases(d, id, t);
695 tstate->emit_state = EMITTED;
699 btf_dump_emit_enum_def(d, id, t, 0);
700 btf_dump_printf(d, ";\n\n");
702 tstate->emit_state = EMITTED;
705 case BTF_KIND_VOLATILE:
707 case BTF_KIND_RESTRICT:
708 btf_dump_emit_type(d, t->type, cont_id);
711 btf_dump_emit_type(d, btf_array(t)->type, cont_id);
714 btf_dump_emit_fwd_def(d, id, t);
715 btf_dump_printf(d, ";\n\n");
716 tstate->emit_state = EMITTED;
718 case BTF_KIND_TYPEDEF:
719 tstate->emit_state = EMITTING;
720 btf_dump_emit_type(d, t->type, id);
722 * typedef can server as both definition and forward
723 * declaration; at this stage someone depends on
724 * typedef as a forward declaration (refers to it
725 * through pointer), so unless we already did it,
726 * emit typedef as a forward declaration
728 if (!tstate->fwd_emitted && !btf_dump_is_blacklisted(d, id)) {
729 btf_dump_emit_typedef_def(d, id, t, 0);
730 btf_dump_printf(d, ";\n\n");
732 tstate->emit_state = EMITTED;
734 case BTF_KIND_STRUCT:
736 tstate->emit_state = EMITTING;
737 /* if it's a top-level struct/union definition or struct/union
738 * is anonymous, then in C we'll be emitting all fields and
739 * their types (as opposed to just `struct X`), so we need to
740 * make sure that all types, referenced from struct/union
741 * members have necessary forward-declarations, where
744 if (top_level_def || t->name_off == 0) {
745 const struct btf_member *m = btf_members(t);
746 __u16 vlen = btf_vlen(t);
749 new_cont_id = t->name_off == 0 ? cont_id : id;
750 for (i = 0; i < vlen; i++, m++)
751 btf_dump_emit_type(d, m->type, new_cont_id);
752 } else if (!tstate->fwd_emitted && id != cont_id) {
753 btf_dump_emit_struct_fwd(d, id, t);
754 btf_dump_printf(d, ";\n\n");
755 tstate->fwd_emitted = 1;
759 btf_dump_emit_struct_def(d, id, t, 0);
760 btf_dump_printf(d, ";\n\n");
761 tstate->emit_state = EMITTED;
763 tstate->emit_state = NOT_EMITTED;
766 case BTF_KIND_FUNC_PROTO: {
767 const struct btf_param *p = btf_params(t);
768 __u16 vlen = btf_vlen(t);
771 btf_dump_emit_type(d, t->type, cont_id);
772 for (i = 0; i < vlen; i++, p++)
773 btf_dump_emit_type(d, p->type, cont_id);
782 static bool btf_is_struct_packed(const struct btf *btf, __u32 id,
783 const struct btf_type *t)
785 const struct btf_member *m;
786 int align, i, bit_sz;
789 align = btf__align_of(btf, id);
790 /* size of a non-packed struct has to be a multiple of its alignment*/
791 if (align && t->size % align)
796 /* all non-bitfield fields have to be naturally aligned */
797 for (i = 0; i < vlen; i++, m++) {
798 align = btf__align_of(btf, m->type);
799 bit_sz = btf_member_bitfield_size(t, i);
800 if (align && bit_sz == 0 && m->offset % (8 * align) != 0)
805 * if original struct was marked as packed, but its layout is
806 * naturally aligned, we'll detect that it's not packed
811 static int chip_away_bits(int total, int at_most)
813 return total % at_most ? : at_most;
816 static void btf_dump_emit_bit_padding(const struct btf_dump *d,
817 int cur_off, int m_off, int m_bit_sz,
820 int off_diff = m_off - cur_off;
821 int ptr_bits = d->ptr_sz * 8;
826 if (m_bit_sz == 0 && off_diff < align * 8)
827 /* natural padding will take care of a gap */
830 while (off_diff > 0) {
831 const char *pad_type;
834 if (ptr_bits > 32 && off_diff > 32) {
836 pad_bits = chip_away_bits(off_diff, ptr_bits);
837 } else if (off_diff > 16) {
839 pad_bits = chip_away_bits(off_diff, 32);
840 } else if (off_diff > 8) {
842 pad_bits = chip_away_bits(off_diff, 16);
845 pad_bits = chip_away_bits(off_diff, 8);
847 btf_dump_printf(d, "\n%s%s: %d;", pfx(lvl), pad_type, pad_bits);
848 off_diff -= pad_bits;
852 static void btf_dump_emit_struct_fwd(struct btf_dump *d, __u32 id,
853 const struct btf_type *t)
855 btf_dump_printf(d, "%s %s",
856 btf_is_struct(t) ? "struct" : "union",
857 btf_dump_type_name(d, id));
860 static void btf_dump_emit_struct_def(struct btf_dump *d,
862 const struct btf_type *t,
865 const struct btf_member *m = btf_members(t);
866 bool is_struct = btf_is_struct(t);
867 int align, i, packed, off = 0;
868 __u16 vlen = btf_vlen(t);
870 packed = is_struct ? btf_is_struct_packed(d->btf, id, t) : 0;
872 btf_dump_printf(d, "%s%s%s {",
873 is_struct ? "struct" : "union",
874 t->name_off ? " " : "",
875 btf_dump_type_name(d, id));
877 for (i = 0; i < vlen; i++, m++) {
881 fname = btf_name_of(d, m->name_off);
882 m_sz = btf_member_bitfield_size(t, i);
883 m_off = btf_member_bit_offset(t, i);
884 align = packed ? 1 : btf__align_of(d->btf, m->type);
886 btf_dump_emit_bit_padding(d, off, m_off, m_sz, align, lvl + 1);
887 btf_dump_printf(d, "\n%s", pfx(lvl + 1));
888 btf_dump_emit_type_decl(d, m->type, fname, lvl + 1);
891 btf_dump_printf(d, ": %d", m_sz);
894 m_sz = max((__s64)0, btf__resolve_size(d->btf, m->type));
895 off = m_off + m_sz * 8;
897 btf_dump_printf(d, ";");
900 /* pad at the end, if necessary */
902 align = packed ? 1 : btf__align_of(d->btf, id);
903 btf_dump_emit_bit_padding(d, off, t->size * 8, 0, align,
908 btf_dump_printf(d, "\n");
909 btf_dump_printf(d, "%s}", pfx(lvl));
911 btf_dump_printf(d, " __attribute__((packed))");
914 static const char *missing_base_types[][2] = {
916 * GCC emits typedefs to its internal __PolyX_t types when compiling Arm
917 * SIMD intrinsics. Alias them to standard base types.
919 { "__Poly8_t", "unsigned char" },
920 { "__Poly16_t", "unsigned short" },
921 { "__Poly64_t", "unsigned long long" },
922 { "__Poly128_t", "unsigned __int128" },
925 static void btf_dump_emit_missing_aliases(struct btf_dump *d, __u32 id,
926 const struct btf_type *t)
928 const char *name = btf_dump_type_name(d, id);
931 for (i = 0; i < ARRAY_SIZE(missing_base_types); i++) {
932 if (strcmp(name, missing_base_types[i][0]) == 0) {
933 btf_dump_printf(d, "typedef %s %s;\n\n",
934 missing_base_types[i][1], name);
940 static void btf_dump_emit_enum_fwd(struct btf_dump *d, __u32 id,
941 const struct btf_type *t)
943 btf_dump_printf(d, "enum %s", btf_dump_type_name(d, id));
946 static void btf_dump_emit_enum_def(struct btf_dump *d, __u32 id,
947 const struct btf_type *t,
950 const struct btf_enum *v = btf_enum(t);
951 __u16 vlen = btf_vlen(t);
956 btf_dump_printf(d, "enum%s%s",
957 t->name_off ? " " : "",
958 btf_dump_type_name(d, id));
961 btf_dump_printf(d, " {");
962 for (i = 0; i < vlen; i++, v++) {
963 name = btf_name_of(d, v->name_off);
964 /* enumerators share namespace with typedef idents */
965 dup_cnt = btf_dump_name_dups(d, d->ident_names, name);
967 btf_dump_printf(d, "\n%s%s___%zu = %u,",
968 pfx(lvl + 1), name, dup_cnt,
971 btf_dump_printf(d, "\n%s%s = %u,",
976 btf_dump_printf(d, "\n%s}", pfx(lvl));
980 static void btf_dump_emit_fwd_def(struct btf_dump *d, __u32 id,
981 const struct btf_type *t)
983 const char *name = btf_dump_type_name(d, id);
986 btf_dump_printf(d, "union %s", name);
988 btf_dump_printf(d, "struct %s", name);
991 static void btf_dump_emit_typedef_def(struct btf_dump *d, __u32 id,
992 const struct btf_type *t, int lvl)
994 const char *name = btf_dump_ident_name(d, id);
997 * Old GCC versions are emitting invalid typedef for __gnuc_va_list
998 * pointing to VOID. This generates warnings from btf_dump() and
999 * results in uncompilable header file, so we are fixing it up here
1000 * with valid typedef into __builtin_va_list.
1002 if (t->type == 0 && strcmp(name, "__gnuc_va_list") == 0) {
1003 btf_dump_printf(d, "typedef __builtin_va_list __gnuc_va_list");
1007 btf_dump_printf(d, "typedef ");
1008 btf_dump_emit_type_decl(d, t->type, name, lvl);
1011 static int btf_dump_push_decl_stack_id(struct btf_dump *d, __u32 id)
1016 if (d->decl_stack_cnt >= d->decl_stack_cap) {
1017 new_cap = max(16, d->decl_stack_cap * 3 / 2);
1018 new_stack = libbpf_reallocarray(d->decl_stack, new_cap, sizeof(new_stack[0]));
1021 d->decl_stack = new_stack;
1022 d->decl_stack_cap = new_cap;
1025 d->decl_stack[d->decl_stack_cnt++] = id;
1031 * Emit type declaration (e.g., field type declaration in a struct or argument
1032 * declaration in function prototype) in correct C syntax.
1034 * For most types it's trivial, but there are few quirky type declaration
1035 * cases worth mentioning:
1036 * - function prototypes (especially nesting of function prototypes);
1038 * - const/volatile/restrict for pointers vs other types.
1040 * For a good discussion of *PARSING* C syntax (as a human), see
1041 * Peter van der Linden's "Expert C Programming: Deep C Secrets",
1042 * Ch.3 "Unscrambling Declarations in C".
1044 * It won't help with BTF to C conversion much, though, as it's an opposite
1045 * problem. So we came up with this algorithm in reverse to van der Linden's
1046 * parsing algorithm. It goes from structured BTF representation of type
1047 * declaration to a valid compilable C syntax.
1049 * For instance, consider this C typedef:
1050 * typedef const int * const * arr[10] arr_t;
1051 * It will be represented in BTF with this chain of BTF types:
1052 * [typedef] -> [array] -> [ptr] -> [const] -> [ptr] -> [const] -> [int]
1054 * Notice how [const] modifier always goes before type it modifies in BTF type
1055 * graph, but in C syntax, const/volatile/restrict modifiers are written to
1056 * the right of pointers, but to the left of other types. There are also other
1057 * quirks, like function pointers, arrays of them, functions returning other
1060 * We handle that by pushing all the types to a stack, until we hit "terminal"
1061 * type (int/enum/struct/union/fwd). Then depending on the kind of a type on
1062 * top of a stack, modifiers are handled differently. Array/function pointers
1063 * have also wildly different syntax and how nesting of them are done. See
1064 * code for authoritative definition.
1066 * To avoid allocating new stack for each independent chain of BTF types, we
1067 * share one bigger stack, with each chain working only on its own local view
1068 * of a stack frame. Some care is required to "pop" stack frames after
1069 * processing type declaration chain.
1071 int btf_dump__emit_type_decl(struct btf_dump *d, __u32 id,
1072 const struct btf_dump_emit_type_decl_opts *opts)
1077 if (!OPTS_VALID(opts, btf_dump_emit_type_decl_opts))
1078 return libbpf_err(-EINVAL);
1080 err = btf_dump_resize(d);
1082 return libbpf_err(err);
1084 fname = OPTS_GET(opts, field_name, "");
1085 lvl = OPTS_GET(opts, indent_level, 0);
1086 d->strip_mods = OPTS_GET(opts, strip_mods, false);
1087 btf_dump_emit_type_decl(d, id, fname, lvl);
1088 d->strip_mods = false;
1092 static void btf_dump_emit_type_decl(struct btf_dump *d, __u32 id,
1093 const char *fname, int lvl)
1095 struct id_stack decl_stack;
1096 const struct btf_type *t;
1097 int err, stack_start;
1099 stack_start = d->decl_stack_cnt;
1101 t = btf__type_by_id(d->btf, id);
1102 if (d->strip_mods && btf_is_mod(t))
1105 err = btf_dump_push_decl_stack_id(d, id);
1108 * if we don't have enough memory for entire type decl
1109 * chain, restore stack, emit warning, and try to
1110 * proceed nevertheless
1112 pr_warn("not enough memory for decl stack:%d", err);
1113 d->decl_stack_cnt = stack_start;
1121 switch (btf_kind(t)) {
1123 case BTF_KIND_VOLATILE:
1124 case BTF_KIND_CONST:
1125 case BTF_KIND_RESTRICT:
1126 case BTF_KIND_FUNC_PROTO:
1129 case BTF_KIND_ARRAY:
1130 id = btf_array(t)->type;
1135 case BTF_KIND_STRUCT:
1136 case BTF_KIND_UNION:
1137 case BTF_KIND_TYPEDEF:
1138 case BTF_KIND_FLOAT:
1141 pr_warn("unexpected type in decl chain, kind:%u, id:[%u]\n",
1148 * We might be inside a chain of declarations (e.g., array of function
1149 * pointers returning anonymous (so inlined) structs, having another
1150 * array field). Each of those needs its own "stack frame" to handle
1151 * emitting of declarations. Those stack frames are non-overlapping
1152 * portions of shared btf_dump->decl_stack. To make it a bit nicer to
1153 * handle this set of nested stacks, we create a view corresponding to
1154 * our own "stack frame" and work with it as an independent stack.
1155 * We'll need to clean up after emit_type_chain() returns, though.
1157 decl_stack.ids = d->decl_stack + stack_start;
1158 decl_stack.cnt = d->decl_stack_cnt - stack_start;
1159 btf_dump_emit_type_chain(d, &decl_stack, fname, lvl);
1161 * emit_type_chain() guarantees that it will pop its entire decl_stack
1162 * frame before returning. But it works with a read-only view into
1163 * decl_stack, so it doesn't actually pop anything from the
1164 * perspective of shared btf_dump->decl_stack, per se. We need to
1165 * reset decl_stack state to how it was before us to avoid it growing
1168 d->decl_stack_cnt = stack_start;
1171 static void btf_dump_emit_mods(struct btf_dump *d, struct id_stack *decl_stack)
1173 const struct btf_type *t;
1176 while (decl_stack->cnt) {
1177 id = decl_stack->ids[decl_stack->cnt - 1];
1178 t = btf__type_by_id(d->btf, id);
1180 switch (btf_kind(t)) {
1181 case BTF_KIND_VOLATILE:
1182 btf_dump_printf(d, "volatile ");
1184 case BTF_KIND_CONST:
1185 btf_dump_printf(d, "const ");
1187 case BTF_KIND_RESTRICT:
1188 btf_dump_printf(d, "restrict ");
1197 static void btf_dump_drop_mods(struct btf_dump *d, struct id_stack *decl_stack)
1199 const struct btf_type *t;
1202 while (decl_stack->cnt) {
1203 id = decl_stack->ids[decl_stack->cnt - 1];
1204 t = btf__type_by_id(d->btf, id);
1211 static void btf_dump_emit_name(const struct btf_dump *d,
1212 const char *name, bool last_was_ptr)
1214 bool separate = name[0] && !last_was_ptr;
1216 btf_dump_printf(d, "%s%s", separate ? " " : "", name);
1219 static void btf_dump_emit_type_chain(struct btf_dump *d,
1220 struct id_stack *decls,
1221 const char *fname, int lvl)
1224 * last_was_ptr is used to determine if we need to separate pointer
1225 * asterisk (*) from previous part of type signature with space, so
1226 * that we get `int ***`, instead of `int * * *`. We default to true
1227 * for cases where we have single pointer in a chain. E.g., in ptr ->
1228 * func_proto case. func_proto will start a new emit_type_chain call
1229 * with just ptr, which should be emitted as (*) or (*<fname>), so we
1230 * don't want to prepend space for that last pointer.
1232 bool last_was_ptr = true;
1233 const struct btf_type *t;
1238 while (decls->cnt) {
1239 id = decls->ids[--decls->cnt];
1241 /* VOID is a special snowflake */
1242 btf_dump_emit_mods(d, decls);
1243 btf_dump_printf(d, "void");
1244 last_was_ptr = false;
1248 t = btf__type_by_id(d->btf, id);
1253 case BTF_KIND_FLOAT:
1254 btf_dump_emit_mods(d, decls);
1255 name = btf_name_of(d, t->name_off);
1256 btf_dump_printf(d, "%s", name);
1258 case BTF_KIND_STRUCT:
1259 case BTF_KIND_UNION:
1260 btf_dump_emit_mods(d, decls);
1261 /* inline anonymous struct/union */
1262 if (t->name_off == 0)
1263 btf_dump_emit_struct_def(d, id, t, lvl);
1265 btf_dump_emit_struct_fwd(d, id, t);
1268 btf_dump_emit_mods(d, decls);
1269 /* inline anonymous enum */
1270 if (t->name_off == 0)
1271 btf_dump_emit_enum_def(d, id, t, lvl);
1273 btf_dump_emit_enum_fwd(d, id, t);
1276 btf_dump_emit_mods(d, decls);
1277 btf_dump_emit_fwd_def(d, id, t);
1279 case BTF_KIND_TYPEDEF:
1280 btf_dump_emit_mods(d, decls);
1281 btf_dump_printf(d, "%s", btf_dump_ident_name(d, id));
1284 btf_dump_printf(d, "%s", last_was_ptr ? "*" : " *");
1286 case BTF_KIND_VOLATILE:
1287 btf_dump_printf(d, " volatile");
1289 case BTF_KIND_CONST:
1290 btf_dump_printf(d, " const");
1292 case BTF_KIND_RESTRICT:
1293 btf_dump_printf(d, " restrict");
1295 case BTF_KIND_ARRAY: {
1296 const struct btf_array *a = btf_array(t);
1297 const struct btf_type *next_t;
1302 * (https://gcc.gnu.org/bugzilla/show_bug.cgi?id=8354)
1303 * which causes it to emit extra const/volatile
1304 * modifiers for an array, if array's element type has
1305 * const/volatile modifiers. Clang doesn't do that.
1306 * In general, it doesn't seem very meaningful to have
1307 * a const/volatile modifier for array, so we are
1308 * going to silently skip them here.
1310 btf_dump_drop_mods(d, decls);
1312 if (decls->cnt == 0) {
1313 btf_dump_emit_name(d, fname, last_was_ptr);
1314 btf_dump_printf(d, "[%u]", a->nelems);
1318 next_id = decls->ids[decls->cnt - 1];
1319 next_t = btf__type_by_id(d->btf, next_id);
1320 multidim = btf_is_array(next_t);
1321 /* we need space if we have named non-pointer */
1322 if (fname[0] && !last_was_ptr)
1323 btf_dump_printf(d, " ");
1324 /* no parentheses for multi-dimensional array */
1326 btf_dump_printf(d, "(");
1327 btf_dump_emit_type_chain(d, decls, fname, lvl);
1329 btf_dump_printf(d, ")");
1330 btf_dump_printf(d, "[%u]", a->nelems);
1333 case BTF_KIND_FUNC_PROTO: {
1334 const struct btf_param *p = btf_params(t);
1335 __u16 vlen = btf_vlen(t);
1339 * GCC emits extra volatile qualifier for
1340 * __attribute__((noreturn)) function pointers. Clang
1341 * doesn't do it. It's a GCC quirk for backwards
1342 * compatibility with code written for GCC <2.5. So,
1343 * similarly to extra qualifiers for array, just drop
1344 * them, instead of handling them.
1346 btf_dump_drop_mods(d, decls);
1348 btf_dump_printf(d, " (");
1349 btf_dump_emit_type_chain(d, decls, fname, lvl);
1350 btf_dump_printf(d, ")");
1352 btf_dump_emit_name(d, fname, last_was_ptr);
1354 btf_dump_printf(d, "(");
1356 * Clang for BPF target generates func_proto with no
1357 * args as a func_proto with a single void arg (e.g.,
1358 * `int (*f)(void)` vs just `int (*f)()`). We are
1359 * going to pretend there are no args for such case.
1361 if (vlen == 1 && p->type == 0) {
1362 btf_dump_printf(d, ")");
1366 for (i = 0; i < vlen; i++, p++) {
1368 btf_dump_printf(d, ", ");
1370 /* last arg of type void is vararg */
1371 if (i == vlen - 1 && p->type == 0) {
1372 btf_dump_printf(d, "...");
1376 name = btf_name_of(d, p->name_off);
1377 btf_dump_emit_type_decl(d, p->type, name, lvl);
1380 btf_dump_printf(d, ")");
1384 pr_warn("unexpected type in decl chain, kind:%u, id:[%u]\n",
1389 last_was_ptr = kind == BTF_KIND_PTR;
1392 btf_dump_emit_name(d, fname, last_was_ptr);
1395 /* return number of duplicates (occurrences) of a given name */
1396 static size_t btf_dump_name_dups(struct btf_dump *d, struct hashmap *name_map,
1397 const char *orig_name)
1401 hashmap__find(name_map, orig_name, (void **)&dup_cnt);
1403 hashmap__set(name_map, orig_name, (void *)dup_cnt, NULL, NULL);
1408 static const char *btf_dump_resolve_name(struct btf_dump *d, __u32 id,
1409 struct hashmap *name_map)
1411 struct btf_dump_type_aux_state *s = &d->type_states[id];
1412 const struct btf_type *t = btf__type_by_id(d->btf, id);
1413 const char *orig_name = btf_name_of(d, t->name_off);
1414 const char **cached_name = &d->cached_names[id];
1417 if (t->name_off == 0)
1420 if (s->name_resolved)
1421 return *cached_name ? *cached_name : orig_name;
1423 dup_cnt = btf_dump_name_dups(d, name_map, orig_name);
1425 const size_t max_len = 256;
1426 char new_name[max_len];
1428 snprintf(new_name, max_len, "%s___%zu", orig_name, dup_cnt);
1429 *cached_name = strdup(new_name);
1432 s->name_resolved = 1;
1433 return *cached_name ? *cached_name : orig_name;
1436 static const char *btf_dump_type_name(struct btf_dump *d, __u32 id)
1438 return btf_dump_resolve_name(d, id, d->type_names);
1441 static const char *btf_dump_ident_name(struct btf_dump *d, __u32 id)
1443 return btf_dump_resolve_name(d, id, d->ident_names);