btf: add BTF types deduplication algorithm

This patch implements BTF types deduplication algorithm. It allows to
greatly compress typical output of pahole's DWARF-to-BTF conversion or
LLVM's compilation output by detecting and collapsing identical types emitted in
isolation per compilation unit. Algorithm also resolves struct/union forward
declarations into concrete BTF types representing referenced struct/union. If
undesired, this resolution can be disabled through specifying corresponding options.

Algorithm itself and its application to Linux kernel's BTF types is
described in details at:
https://facebookmicrosites.github.io/bpf/blog/2018/11/14/btf-enhancement.html

Signed-off-by: Andrii Nakryiko <andriin@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

3 files changed, 1749 insertions, 0 deletions

diff --git a/tools/lib/bpf/btf.c b/tools/lib/bpf/btf.c
index 06bd1a625ff4..e5097be16018 100644
--- a/tools/lib/bpf/btf.c
+++ b/tools/lib/bpf/btf.c
@@ -849,3 +849,1744 @@ __u32 btf_ext__line_info_rec_size(const struct btf_ext *btf_ext)
 {
 	return btf_ext->line_info.rec_size;
 }
+
+struct btf_dedup;
+
+static struct btf_dedup *btf_dedup_new(struct btf *btf, struct btf_ext *btf_ext,
+				       const struct btf_dedup_opts *opts);
+static void btf_dedup_free(struct btf_dedup *d);
+static int btf_dedup_strings(struct btf_dedup *d);
+static int btf_dedup_prim_types(struct btf_dedup *d);
+static int btf_dedup_struct_types(struct btf_dedup *d);
+static int btf_dedup_ref_types(struct btf_dedup *d);
+static int btf_dedup_compact_types(struct btf_dedup *d);
+static int btf_dedup_remap_types(struct btf_dedup *d);
+
+/*
+ * Deduplicate BTF types and strings.
+ *
+ * BTF dedup algorithm takes as an input `struct btf` representing `.BTF` ELF
+ * section with all BTF type descriptors and string data. It overwrites that
+ * memory in-place with deduplicated types and strings without any loss of
+ * information. If optional `struct btf_ext` representing '.BTF.ext' ELF section
+ * is provided, all the strings referenced from .BTF.ext section are honored
+ * and updated to point to the right offsets after deduplication.
+ *
+ * If function returns with error, type/string data might be garbled and should
+ * be discarded.
+ *
+ * More verbose and detailed description of both problem btf_dedup is solving,
+ * as well as solution could be found at:
+ * https://facebookmicrosites.github.io/bpf/blog/2018/11/14/btf-enhancement.html
+ *
+ * Problem description and justification
+ * =====================================
+ *
+ * BTF type information is typically emitted either as a result of conversion
+ * from DWARF to BTF or directly by compiler. In both cases, each compilation
+ * unit contains information about a subset of all the types that are used
+ * in an application. These subsets are frequently overlapping and contain a lot
+ * of duplicated information when later concatenated together into a single
+ * binary. This algorithm ensures that each unique type is represented by single
+ * BTF type descriptor, greatly reducing resulting size of BTF data.
+ *
+ * Compilation unit isolation and subsequent duplication of data is not the only
+ * problem. The same type hierarchy (e.g., struct and all the type that struct
+ * references) in different compilation units can be represented in BTF to
+ * various degrees of completeness (or, rather, incompleteness) due to
+ * struct/union forward declarations.
+ *
+ * Let's take a look at an example, that we'll use to better understand the
+ * problem (and solution). Suppose we have two compilation units, each using
+ * same `struct S`, but each of them having incomplete type information about
+ * struct's fields:
+ *
+ * // CU #1:
+ * struct S;
+ * struct A {
+ *	int a;
+ *	struct A* self;
+ *	struct S* parent;
+ * };
+ * struct B;
+ * struct S {
+ *	struct A* a_ptr;
+ *	struct B* b_ptr;
+ * };
+ *
+ * // CU #2:
+ * struct S;
+ * struct A;
+ * struct B {
+ *	int b;
+ *	struct B* self;
+ *	struct S* parent;
+ * };
+ * struct S {
+ *	struct A* a_ptr;
+ *	struct B* b_ptr;
+ * };
+ *
+ * In case of CU #1, BTF data will know only that `struct B` exist (but no
+ * more), but will know the complete type information about `struct A`. While
+ * for CU #2, it will know full type information about `struct B`, but will
+ * only know about forward declaration of `struct A` (in BTF terms, it will
+ * have `BTF_KIND_FWD` type descriptor with name `B`).
+ *
+ * This compilation unit isolation means that it's possible that there is no
+ * single CU with complete type information describing structs `S`, `A`, and
+ * `B`. Also, we might get tons of duplicated and redundant type information.
+ *
+ * Additional complication we need to keep in mind comes from the fact that
+ * types, in general, can form graphs containing cycles, not just DAGs.
+ *
+ * While algorithm does deduplication, it also merges and resolves type
+ * information (unless disabled throught `struct btf_opts`), whenever possible.
+ * E.g., in the example above with two compilation units having partial type
+ * information for structs `A` and `B`, the output of algorithm will emit
+ * a single copy of each BTF type that describes structs `A`, `B`, and `S`
+ * (as well as type information for `int` and pointers), as if they were defined
+ * in a single compilation unit as:
+ *
+ * struct A {
+ *	int a;
+ *	struct A* self;
+ *	struct S* parent;
+ * };
+ * struct B {
+ *	int b;
+ *	struct B* self;
+ *	struct S* parent;
+ * };
+ * struct S {
+ *	struct A* a_ptr;
+ *	struct B* b_ptr;
+ * };
+ *
+ * Algorithm summary
+ * =================
+ *
+ * Algorithm completes its work in 6 separate passes:
+ *
+ * 1. Strings deduplication.
+ * 2. Primitive types deduplication (int, enum, fwd).
+ * 3. Struct/union types deduplication.
+ * 4. Reference types deduplication (pointers, typedefs, arrays, funcs, func
+ *    protos, and const/volatile/restrict modifiers).
+ * 5. Types compaction.
+ * 6. Types remapping.
+ *
+ * Algorithm determines canonical type descriptor, which is a single
+ * representative type for each truly unique type. This canonical type is the
+ * one that will go into final deduplicated BTF type information. For
+ * struct/unions, it is also the type that algorithm will merge additional type
+ * information into (while resolving FWDs), as it discovers it from data in
+ * other CUs. Each input BTF type eventually gets either mapped to itself, if
+ * that type is canonical, or to some other type, if that type is equivalent
+ * and was chosen as canonical representative. This mapping is stored in
+ * `btf_dedup->map` array. This map is also used to record STRUCT/UNION that
+ * FWD type got resolved to.
+ *
+ * To facilitate fast discovery of canonical types, we also maintain canonical
+ * index (`btf_dedup->dedup_table`), which maps type descriptor's signature hash
+ * (i.e., hashed kind, name, size, fields, etc) into a list of canonical types
+ * that match that signature. With sufficiently good choice of type signature
+ * hashing function, we can limit number of canonical types for each unique type
+ * signature to a very small number, allowing to find canonical type for any
+ * duplicated type very quickly.
+ *
+ * Struct/union deduplication is the most critical part and algorithm for
+ * deduplicating structs/unions is described in greater details in comments for
+ * `btf_dedup_is_equiv` function.
+ */
+int btf__dedup(struct btf *btf, struct btf_ext *btf_ext,
+	       const struct btf_dedup_opts *opts)
+{
+	struct btf_dedup *d = btf_dedup_new(btf, btf_ext, opts);
+	int err;
+
+	if (IS_ERR(d)) {
+		pr_debug("btf_dedup_new failed: %ld", PTR_ERR(d));
+		return -EINVAL;
+	}
+
+	err = btf_dedup_strings(d);
+	if (err < 0) {
+		pr_debug("btf_dedup_strings failed:%d\n", err);
+		goto done;
+	}
+	err = btf_dedup_prim_types(d);
+	if (err < 0) {
+		pr_debug("btf_dedup_prim_types failed:%d\n", err);
+		goto done;
+	}
+	err = btf_dedup_struct_types(d);
+	if (err < 0) {
+		pr_debug("btf_dedup_struct_types failed:%d\n", err);
+		goto done;
+	}
+	err = btf_dedup_ref_types(d);
+	if (err < 0) {
+		pr_debug("btf_dedup_ref_types failed:%d\n", err);
+		goto done;
+	}
+	err = btf_dedup_compact_types(d);
+	if (err < 0) {
+		pr_debug("btf_dedup_compact_types failed:%d\n", err);
+		goto done;
+	}
+	err = btf_dedup_remap_types(d);
+	if (err < 0) {
+		pr_debug("btf_dedup_remap_types failed:%d\n", err);
+		goto done;
+	}
+
+done:
+	btf_dedup_free(d);
+	return err;
+}
+
+#define BTF_DEDUP_TABLE_SIZE_LOG 14
+#define BTF_DEDUP_TABLE_MOD ((1 << BTF_DEDUP_TABLE_SIZE_LOG) - 1)
+#define BTF_UNPROCESSED_ID ((__u32)-1)
+#define BTF_IN_PROGRESS_ID ((__u32)-2)
+
+struct btf_dedup_node {
+	struct btf_dedup_node *next;
+	__u32 type_id;
+};
+
+struct btf_dedup {
+	/* .BTF section to be deduped in-place */
+	struct btf *btf;
+	/*
+	 * Optional .BTF.ext section. When provided, any strings referenced
+	 * from it will be taken into account when deduping strings
+	 */
+	struct btf_ext *btf_ext;
+	/*
+	 * This is a map from any type's signature hash to a list of possible
+	 * canonical representative type candidates. Hash collisions are
+	 * ignored, so even types of various kinds can share same list of
+	 * candidates, which is fine because we rely on subsequent
+	 * btf_xxx_equal() checks to authoritatively verify type equality.
+	 */
+	struct btf_dedup_node **dedup_table;
+	/* Canonical types map */
+	__u32 *map;
+	/* Hypothetical mapping, used during type graph equivalence checks */
+	__u32 *hypot_map;
+	__u32 *hypot_list;
+	size_t hypot_cnt;
+	size_t hypot_cap;
+	/* Various option modifying behavior of algorithm */
+	struct btf_dedup_opts opts;
+};
+
+struct btf_str_ptr {
+	const char *str;
+	__u32 new_off;
+	bool used;
+};
+
+struct btf_str_ptrs {
+	struct btf_str_ptr *ptrs;
+	const char *data;
+	__u32 cnt;
+	__u32 cap;
+};
+
+static inline __u32 hash_combine(__u32 h, __u32 value)
+{
+/* 2^31 + 2^29 - 2^25 + 2^22 - 2^19 - 2^16 + 1 */
+#define GOLDEN_RATIO_PRIME 0x9e370001UL
+	return h * 37 + value * GOLDEN_RATIO_PRIME;
+#undef GOLDEN_RATIO_PRIME
+}
+
+#define for_each_hash_node(table, hash, node) \
+	for (node = table[hash & BTF_DEDUP_TABLE_MOD]; node; node = node->next)
+
+static int btf_dedup_table_add(struct btf_dedup *d, __u32 hash, __u32 type_id)
+{
+	struct btf_dedup_node *node = malloc(sizeof(struct btf_dedup_node));
+
+	if (!node)
+		return -ENOMEM;
+	node->type_id = type_id;
+	node->next = d->dedup_table[hash & BTF_DEDUP_TABLE_MOD];
+	d->dedup_table[hash & BTF_DEDUP_TABLE_MOD] = node;
+	return 0;
+}
+
+static int btf_dedup_hypot_map_add(struct btf_dedup *d,
+				   __u32 from_id, __u32 to_id)
+{
+	if (d->hypot_cnt == d->hypot_cap) {
+		__u32 *new_list;
+
+		d->hypot_cap += max(16, d->hypot_cap / 2);
+		new_list = realloc(d->hypot_list, sizeof(__u32) * d->hypot_cap);
+		if (!new_list)
+			return -ENOMEM;
+		d->hypot_list = new_list;
+	}
+	d->hypot_list[d->hypot_cnt++] = from_id;
+	d->hypot_map[from_id] = to_id;
+	return 0;
+}
+
+static void btf_dedup_clear_hypot_map(struct btf_dedup *d)
+{
+	int i;
+
+	for (i = 0; i < d->hypot_cnt; i++)
+		d->hypot_map[d->hypot_list[i]] = BTF_UNPROCESSED_ID;
+	d->hypot_cnt = 0;
+}
+
+static void btf_dedup_table_free(struct btf_dedup *d)
+{
+	struct btf_dedup_node *head, *tmp;
+	int i;
+
+	if (!d->dedup_table)
+		return;
+
+	for (i = 0; i < (1 << BTF_DEDUP_TABLE_SIZE_LOG); i++) {
+		while (d->dedup_table[i]) {
+			tmp = d->dedup_table[i];
+			d->dedup_table[i] = tmp->next;
+			free(tmp);
+		}
+
+		head = d->dedup_table[i];
+		while (head) {
+			tmp = head;
+			head = head->next;
+			free(tmp);
+		}
+	}
+
+	free(d->dedup_table);
+	d->dedup_table = NULL;
+}
+
+static void btf_dedup_free(struct btf_dedup *d)
+{
+	btf_dedup_table_free(d);
+
+	free(d->map);
+	d->map = NULL;
+
+	free(d->hypot_map);
+	d->hypot_map = NULL;
+
+	free(d->hypot_list);
+	d->hypot_list = NULL;
+
+	free(d);
+}
+
+static struct btf_dedup *btf_dedup_new(struct btf *btf, struct btf_ext *btf_ext,
+				       const struct btf_dedup_opts *opts)
+{
+	struct btf_dedup *d = calloc(1, sizeof(struct btf_dedup));
+	int i, err = 0;
+
+	if (!d)
+		return ERR_PTR(-ENOMEM);
+
+	d->btf = btf;
+	d->btf_ext = btf_ext;
+
+	d->dedup_table = calloc(1 << BTF_DEDUP_TABLE_SIZE_LOG,
+				sizeof(struct btf_dedup_node *));
+	if (!d->dedup_table) {
+		err = -ENOMEM;
+		goto done;
+	}
+
+	d->map = malloc(sizeof(__u32) * (1 + btf->nr_types));
+	if (!d->map) {
+		err = -ENOMEM;
+		goto done;
+	}
+	/* special BTF "void" type is made canonical immediately */
+	d->map[0] = 0;
+	for (i = 1; i <= btf->nr_types; i++)
+		d->map[i] = BTF_UNPROCESSED_ID;
+
+	d->hypot_map = malloc(sizeof(__u32) * (1 + btf->nr_types));
+	if (!d->hypot_map) {
+		err = -ENOMEM;
+		goto done;
+	}
+	for (i = 0; i <= btf->nr_types; i++)
+		d->hypot_map[i] = BTF_UNPROCESSED_ID;
+
+	d->opts.dont_resolve_fwds = opts && opts->dont_resolve_fwds;
+
+done:
+	if (err) {
+		btf_dedup_free(d);
+		return ERR_PTR(err);
+	}
+
+	return d;
+}
+
+typedef int (*str_off_fn_t)(__u32 *str_off_ptr, void *ctx);
+
+/*
+ * Iterate over all possible places in .BTF and .BTF.ext that can reference
+ * string and pass pointer to it to a provided callback `fn`.
+ */
+static int btf_for_each_str_off(struct btf_dedup *d, str_off_fn_t fn, void *ctx)
+{
+	void *line_data_cur, *line_data_end;
+	int i, j, r, rec_size;
+	struct btf_type *t;
+
+	for (i = 1; i <= d->btf->nr_types; i++) {
+		t = d->btf->types[i];
+		r = fn(&t->name_off, ctx);
+		if (r)
+			return r;
+
+		switch (BTF_INFO_KIND(t->info)) {
+		case BTF_KIND_STRUCT:
+		case BTF_KIND_UNION: {
+			struct btf_member *m = (struct btf_member *)(t + 1);
+			__u16 vlen = BTF_INFO_VLEN(t->info);
+
+			for (j = 0; j < vlen; j++) {
+				r = fn(&m->name_off, ctx);
+				if (r)
+					return r;
+				m++;
+			}
+			break;
+		}
+		case BTF_KIND_ENUM: {
+			struct btf_enum *m = (struct btf_enum *)(t + 1);
+			__u16 vlen = BTF_INFO_VLEN(t->info);
+
+			for (j = 0; j < vlen; j++) {
+				r = fn(&m->name_off, ctx);
+				if (r)
+					return r;
+				m++;
+			}
+			break;
+		}
+		case BTF_KIND_FUNC_PROTO: {
+			struct btf_param *m = (struct btf_param *)(t + 1);
+			__u16 vlen = BTF_INFO_VLEN(t->info);
+
+			for (j = 0; j < vlen; j++) {
+				r = fn(&m->name_off, ctx);
+				if (r)
+					return r;
+				m++;
+			}
+			break;
+		}
+		default:
+			break;
+		}
+	}
+
+	if (!d->btf_ext)
+		return 0;
+
+	line_data_cur = d->btf_ext->line_info.info;
+	line_data_end = d->btf_ext->line_info.info + d->btf_ext->line_info.len;
+	rec_size = d->btf_ext->line_info.rec_size;
+
+	while (line_data_cur < line_data_end) {
+		struct btf_ext_info_sec *sec = line_data_cur;
+		struct bpf_line_info_min *line_info;
+		__u32 num_info = sec->num_info;
+
+		r = fn(&sec->sec_name_off, ctx);
+		if (r)
+			return r;
+
+		line_data_cur += sizeof(struct btf_ext_info_sec);
+		for (i = 0; i < num_info; i++) {
+			line_info = line_data_cur;
+			r = fn(&line_info->file_name_off, ctx);
+			if (r)
+				return r;
+			r = fn(&line_info->line_off, ctx);
+			if (r)
+				return r;
+			line_data_cur += rec_size;
+		}
+	}
+
+	return 0;
+}
+
+static int str_sort_by_content(const void *a1, const void *a2)
+{
+	const struct btf_str_ptr *p1 = a1;
+	const struct btf_str_ptr *p2 = a2;
+
+	return strcmp(p1->str, p2->str);
+}
+
+static int str_sort_by_offset(const void *a1, const void *a2)
+{
+	const struct btf_str_ptr *p1 = a1;
+	const struct btf_str_ptr *p2 = a2;
+
+	if (p1->str != p2->str)
+		return p1->str < p2->str ? -1 : 1;
+	return 0;
+}
+
+static int btf_dedup_str_ptr_cmp(const void *str_ptr, const void *pelem)
+{
+	const struct btf_str_ptr *p = pelem;
+
+	if (str_ptr != p->str)
+		return (const char *)str_ptr < p->str ? -1 : 1;
+	return 0;
+}
+
+static int btf_str_mark_as_used(__u32 *str_off_ptr, void *ctx)
+{
+	struct btf_str_ptrs *strs;
+	struct btf_str_ptr *s;
+
+	if (*str_off_ptr == 0)
+		return 0;
+
+	strs = ctx;
+	s = bsearch(strs->data + *str_off_ptr, strs->ptrs, strs->cnt,
+		    sizeof(struct btf_str_ptr), btf_dedup_str_ptr_cmp);
+	if (!s)
+		return -EINVAL;
+	s->used = true;
+	return 0;
+}
+
+static int btf_str_remap_offset(__u32 *str_off_ptr, void *ctx)
+{
+	struct btf_str_ptrs *strs;
+	struct btf_str_ptr *s;
+
+	if (*str_off_ptr == 0)
+		return 0;
+
+	strs = ctx;
+	s = bsearch(strs->data + *str_off_ptr, strs->ptrs, strs->cnt,
+		    sizeof(struct btf_str_ptr), btf_dedup_str_ptr_cmp);
+	if (!s)
+		return -EINVAL;
+	*str_off_ptr = s->new_off;
+	return 0;
+}
+
+/*
+ * Dedup string and filter out those that are not referenced from either .BTF
+ * or .BTF.ext (if provided) sections.
+ *
+ * This is done by building index of all strings in BTF's string section,
+ * then iterating over all entities that can reference strings (e.g., type
+ * names, struct field names, .BTF.ext line info, etc) and marking corresponding
+ * strings as used. After that all used strings are deduped and compacted into
+ * sequential blob of memory and new offsets are calculated. Then all the string
+ * references are iterated again and rewritten using new offsets.
+ */
+static int btf_dedup_strings(struct btf_dedup *d)
+{
+	const struct btf_header *hdr = d->btf->hdr;
+	char *start = (char *)d->btf->nohdr_data + hdr->str_off;
+	char *end = start + d->btf->hdr->str_len;
+	char *p = start, *tmp_strs = NULL;
+	struct btf_str_ptrs strs = {
+		.cnt = 0,
+		.cap = 0,
+		.ptrs = NULL,
+		.data = start,
+	};
+	int i, j, err = 0, grp_idx;
+	bool grp_used;
+
+	/* build index of all strings */
+	while (p < end) {
+		if (strs.cnt + 1 > strs.cap) {
+			struct btf_str_ptr *new_ptrs;
+
+			strs.cap += max(strs.cnt / 2, 16);
+			new_ptrs = realloc(strs.ptrs,
+					   sizeof(strs.ptrs[0]) * strs.cap);
+			if (!new_ptrs) {
+				err = -ENOMEM;
+				goto done;
+			}
+			strs.ptrs = new_ptrs;
+		}
+
+		strs.ptrs[strs.cnt].str = p;
+		strs.ptrs[strs.cnt].used = false;
+
+		p += strlen(p) + 1;
+		strs.cnt++;
+	}
+
+	/* temporary storage for deduplicated strings */
+	tmp_strs = malloc(d->btf->hdr->str_len);
+	if (!tmp_strs) {
+		err = -ENOMEM;
+		goto done;
+	}
+
+	/* mark all used strings */
+	strs.ptrs[0].used = true;
+	err = btf_for_each_str_off(d, btf_str_mark_as_used, &strs);
+	if (err)
+		goto done;
+
+	/* sort strings by context, so that we can identify duplicates */
+	qsort(strs.ptrs, strs.cnt, sizeof(strs.ptrs[0]), str_sort_by_content);
+
+	/*
+	 * iterate groups of equal strings and if any instance in a group was
+	 * referenced, emit single instance and remember new offset
+	 */
+	p = tmp_strs;
+	grp_idx = 0;
+	grp_used = strs.ptrs[0].used;
+	/* iterate past end to avoid code duplication after loop */
+	for (i = 1; i <= strs.cnt; i++) {
+		/*
+		 * when i == strs.cnt, we want to skip string comparison and go
+		 * straight to handling last group of strings (otherwise we'd
+		 * need to handle last group after the loop w/ duplicated code)
+		 */
+		if (i < strs.cnt &&
+		    !strcmp(strs.ptrs[i].str, strs.ptrs[grp_idx].str)) {
+			grp_used = grp_used || strs.ptrs[i].used;
+			continue;
+		}
+
+		/*
+		 * this check would have been required after the loop to handle
+		 * last group of strings, but due to <= condition in a loop
+		 * we avoid that duplication
+		 */
+		if (grp_used) {
+			int new_off = p - tmp_strs;
+			__u32 len = strlen(strs.ptrs[grp_idx].str);
+
+			memmove(p, strs.ptrs[grp_idx].str, len + 1);
+			for (j = grp_idx; j < i; j++)
+				strs.ptrs[j].new_off = new_off;
+			p += len + 1;
+		}
+
+		if (i < strs.cnt) {
+			grp_idx = i;
+			grp_used = strs.ptrs[i].used;
+		}
+	}
+
+	/* replace original strings with deduped ones */
+	d->btf->hdr->str_len = p - tmp_strs;
+	memmove(start, tmp_strs, d->btf->hdr->str_len);
+	end = start + d->btf->hdr->str_len;
+
+	/* restore original order for further binary search lookups */
+	qsort(strs.ptrs, strs.cnt, sizeof(strs.ptrs[0]), str_sort_by_offset);
+
+	/* remap string offsets */
+	err = btf_for_each_str_off(d, btf_str_remap_offset, &strs);
+	if (err)
+		goto done;
+
+	d->btf->hdr->str_len = end - start;
+
+done:
+	free(tmp_strs);
+	free(strs.ptrs);
+	return err;
+}
+
+static __u32 btf_hash_common(struct btf_type *t)
+{
+	__u32 h;
+
+	h = hash_combine(0, t->name_off);
+	h = hash_combine(h, t->info);
+	h = hash_combine(h, t->size);
+	return h;
+}
+
+static bool btf_equal_common(struct btf_type *t1, struct btf_type *t2)
+{
+	return t1->name_off == t2->name_off &&
+	       t1->info == t2->info &&
+	       t1->size == t2->size;
+}
+
+/* Calculate type signature hash of INT. */
+static __u32 btf_hash_int(struct btf_type *t)
+{
+	__u32 info = *(__u32 *)(t + 1);
+	__u32 h;
+
+	h = btf_hash_common(t);
+	h = hash_combine(h, info);
+	return h;
+}
+
+/* Check structural equality of two INTs. */
+static bool btf_equal_int(struct btf_type *t1, struct btf_type *t2)
+{
+	__u32 info1, info2;
+
+	if (!btf_equal_common(t1, t2))
+		return false;
+	info1 = *(__u32 *)(t1 + 1);
+	info2 = *(__u32 *)(t2 + 1);
+	return info1 == info2;
+}
+
+/* Calculate type signature hash of ENUM. */
+static __u32 btf_hash_enum(struct btf_type *t)
+{
+	struct btf_enum *member = (struct btf_enum *)(t + 1);
+	__u32 vlen = BTF_INFO_VLEN(t->info);
+	__u32 h = btf_hash_common(t);
+	int i;
+
+	for (i = 0; i < vlen; i++) {
+		h = hash_combine(h, member->name_off);
+		h = hash_combine(h, member->val);
+		member++;
+	}
+	return h;
+}
+
+/* Check structural equality of two ENUMs. */
+static bool btf_equal_enum(struct btf_type *t1, struct btf_type *t2)
+{
+	struct btf_enum *m1, *m2;
+	__u16 vlen;
+	int i;
+
+	if (!btf_equal_common(t1, t2))
+		return false;
+
+	vlen = BTF_INFO_VLEN(t1->info);
+	m1 = (struct btf_enum *)(t1 + 1);
+	m2 = (struct btf_enum *)(t2 + 1);
+	for (i = 0; i < vlen; i++) {
+		if (m1->name_off != m2->name_off || m1->val != m2->val)
+			return false;
+		m1++;
+		m2++;
+	}
+	return true;
+}
+
+/*
+ * Calculate type signature hash of STRUCT/UNION, ignoring referenced type IDs,
+ * as referenced type IDs equivalence is established separately during type
+ * graph equivalence check algorithm.
+ */
+static __u32 btf_hash_struct(struct btf_type *t)
+{
+	struct btf_member *member = (struct btf_member *)(t + 1);
+	__u32 vlen = BTF_INFO_VLEN(t->info);
+	__u32 h = btf_hash_common(t);
+	int i;
+
+	for (i = 0; i < vlen; i++) {
+		h = hash_combine(h, member->name_off);
+		h = hash_combine(h, member->offset);
+		/* no hashing of referenced type ID, it can be unresolved yet */
+		member++;
+	}
+	return h;
+}
+
+/*
+ * Check structural compatibility of two FUNC_PROTOs, ignoring referenced type
+ * IDs. This check is performed during type graph equivalence check and
+ * referenced types equivalence is checked separately.
+ */
+static bool btf_equal_struct(struct btf_type *t1, struct btf_type *t2)
+{
+	struct btf_member *m1, *m2;
+	__u16 vlen;
+	int i;
+
+	if (!btf_equal_common(t1, t2))
+		return false;
+
+	vlen = BTF_INFO_VLEN(t1->info);
+	m1 = (struct btf_member *)(t1 + 1);
+	m2 = (struct btf_member *)(t2 + 1);
+	for (i = 0; i < vlen; i++) {
+		if (m1->name_off != m2->name_off || m1->offset != m2->offset)
+			return false;
+		m1++;
+		m2++;
+	}
+	return true;
+}
+
+/*
+ * Calculate type signature hash of ARRAY, including referenced type IDs,
+ * under assumption that they were already resolved to canonical type IDs and
+ * are not going to change.
+ */
+static __u32 btf_hash_array(struct btf_type *t)
+{
+	struct btf_array *info = (struct btf_array *)(t + 1);
+	__u32 h = btf_hash_common(t);
+
+	h = hash_combine(h, info->type);
+	h = hash_combine(h, info->index_type);
+	h = hash_combine(h, info->nelems);
+	return h;
+}
+
+/*
+ * Check exact equality of two ARRAYs, taking into account referenced
+ * type IDs, under assumption that they were already resolved to canonical
+ * type IDs and are not going to change.
+ * This function is called during reference types deduplication to compare
+ * ARRAY to potential canonical representative.
+ */
+static bool btf_equal_array(struct btf_type *t1, struct btf_type *t2)
+{
+	struct btf_array *info1, *info2;
+
+	if (!btf_equal_common(t1, t2))
+		return false;
+
+	info1 = (struct btf_array *)(t1 + 1);
+	info2 = (struct btf_array *)(t2 + 1);
+	return info1->type == info2->type &&
+	       info1->index_type == info2->index_type &&
+	       info1->nelems == info2->nelems;
+}
+
+/*
+ * Check structural compatibility of two ARRAYs, ignoring referenced type
+ * IDs. This check is performed during type graph equivalence check and
+ * referenced types equivalence is checked separately.
+ */
+static bool btf_compat_array(struct btf_type *t1, struct btf_type *t2)
+{
+	struct btf_array *info1, *info2;
+
+	if (!btf_equal_common(t1, t2))
+		return false;
+
+	info1 = (struct btf_array *)(t1 + 1);
+	info2 = (struct btf_array *)(t2 + 1);
+	return info1->nelems == info2->nelems;
+}
+
+/*
+ * Calculate type signature hash of FUNC_PROTO, including referenced type IDs,
+ * under assumption that they were already resolved to canonical type IDs and
+ * are not going to change.
+ */
+static inline __u32 btf_hash_fnproto(struct btf_type *t)
+{
+	struct btf_param *member = (struct btf_param *)(t + 1);
+	__u16 vlen = BTF_INFO_VLEN(t->info);
+	__u32 h = btf_hash_common(t);
+	int i;
+
+	for (i = 0; i < vlen; i++) {
+		h = hash_combine(h, member->name_off);
+		h = hash_combine(h, member->type);
+		member++;
+	}
+	return h;
+}
+
+/*
+ * Check exact equality of two FUNC_PROTOs, taking into account referenced
+ * type IDs, under assumption that they were already resolved to canonical
+ * type IDs and are not going to change.
+ * This function is called during reference types deduplication to compare
+ * FUNC_PROTO to potential canonical representative.
+ */
+static inline bool btf_equal_fnproto(struct btf_type *t1, struct btf_type *t2)
+{
+	struct btf_param *m1, *m2;
+	__u16 vlen;
+	int i;
+
+	if (!btf_equal_common(t1, t2))
+		return false;
+
+	vlen = BTF_INFO_VLEN(t1->info);
+	m1 = (struct btf_param *)(t1 + 1);
+	m2 = (struct btf_param *)(t2 + 1);
+	for (i = 0; i < vlen; i++) {
+		if (m1->name_off != m2->name_off || m1->type != m2->type)
+			return false;
+		m1++;
+		m2++;
+	}
+	return true;
+}
+
+/*
+ * Check structural compatibility of two FUNC_PROTOs, ignoring referenced type
+ * IDs. This check is performed during type graph equivalence check and
+ * referenced types equivalence is checked separately.
+ */
+static inline bool btf_compat_fnproto(struct btf_type *t1, struct btf_type *t2)
+{
+	struct btf_param *m1, *m2;
+	__u16 vlen;
+	int i;
+
+	/* skip return type ID */
+	if (t1->name_off != t2->name_off || t1->info != t2->info)
+		return false;
+
+	vlen = BTF_INFO_VLEN(t1->info);
+	m1 = (struct btf_param *)(t1 + 1);
+	m2 = (struct btf_param *)(t2 + 1);
+	for (i = 0; i < vlen; i++) {
+		if (m1->name_off != m2->name_off)
+			return false;
+		m1++;
+		m2++;
+	}
+	return true;
+}
+
+/*
+ * Deduplicate primitive types, that can't reference other types, by calculating
+ * their type signature hash and comparing them with any possible canonical
+ * candidate. If no canonical candidate matches, type itself is marked as
+ * canonical and is added into `btf_dedup->dedup_table` as another candidate.
+ */
+static int btf_dedup_prim_type(struct btf_dedup *d, __u32 type_id)
+{
+	struct btf_type *t = d->btf->types[type_id];
+	struct btf_type *cand;
+	struct btf_dedup_node *cand_node;
+	/* if we don't find equivalent type, then we are canonical */
+	__u32 new_id = type_id;
+	__u32 h;
+
+	switch (BTF_INFO_KIND(t->info)) {
+	case BTF_KIND_CONST:
+	case BTF_KIND_VOLATILE:
+	case BTF_KIND_RESTRICT:
+	case BTF_KIND_PTR:
+	case BTF_KIND_TYPEDEF:
+	case BTF_KIND_ARRAY:
+	case BTF_KIND_STRUCT:
+	case BTF_KIND_UNION:
+	case BTF_KIND_FUNC:
+	case BTF_KIND_FUNC_PROTO:
+		return 0;
+
+	case BTF_KIND_INT:
+		h = btf_hash_int(t);
+		for_each_hash_node(d->dedup_table, h, cand_node) {
+			cand = d->btf->types[cand_node->type_id];
+			if (btf_equal_int(t, cand)) {
+				new_id = cand_node->type_id;
+				break;
+			}
+		}
+		break;
+
+	case BTF_KIND_ENUM:
+		h = btf_hash_enum(t);
+		for_each_hash_node(d->dedup_table, h, cand_node) {
+			cand = d->btf->types[cand_node->type_id];
+			if (btf_equal_enum(t, cand)) {
+				new_id = cand_node->type_id;
+				break;
+			}
+		}
+		break;
+
+	case BTF_KIND_FWD:
+		h = btf_hash_common(t);
+		for_each_hash_node(d->dedup_table, h, cand_node) {
+			cand = d->btf->types[cand_node->type_id];
+			if (btf_equal_common(t, cand)) {
+				new_id = cand_node->type_id;
+				break;
+			}
+		}
+		break;
+
+	default:
+		return -EINVAL;
+	}
+
+	d->map[type_id] = new_id;
+	if (type_id == new_id && btf_dedup_table_add(d, h, type_id))
+		return -ENOMEM;
+
+	return 0;
+}
+
+static int btf_dedup_prim_types(struct btf_dedup *d)
+{
+	int i, err;
+
+	for (i = 1; i <= d->btf->nr_types; i++) {
+		err = btf_dedup_prim_type(d, i);
+		if (err)
+			return err;
+	}
+	return 0;
+}
+
+/*
+ * Check whether type is already mapped into canonical one (could be to itself).
+ */
+static inline bool is_type_mapped(struct btf_dedup *d, uint32_t type_id)
+{
+	return d->map[type_id] <= BTF_MAX_TYPE;
+}
+
+/*
+ * Resolve type ID into its canonical type ID, if any; otherwise return original
+ * type ID. If type is FWD and is resolved into STRUCT/UNION already, follow
+ * STRUCT/UNION link and resolve it into canonical type ID as well.
+ */
+static inline __u32 resolve_type_id(struct btf_dedup *d, __u32 type_id)
+{
+	while (is_type_mapped(d, type_id) && d->map[type_id] != type_id)
+		type_id = d->map[type_id];
+	return type_id;
+}
+
+/*
+ * Resolve FWD to underlying STRUCT/UNION, if any; otherwise return original
+ * type ID.
+ */
+static uint32_t resolve_fwd_id(struct btf_dedup *d, uint32_t type_id)
+{
+	__u32 orig_type_id = type_id;
+
+	if (BTF_INFO_KIND(d->btf->types[type_id]->info) != BTF_KIND_FWD)
+		return type_id;
+
+	while (is_type_mapped(d, type_id) && d->map[type_id] != type_id)
+		type_id = d->map[type_id];
+
+	if (BTF_INFO_KIND(d->btf->types[type_id]->info) != BTF_KIND_FWD)
+		return type_id;
+
+	return orig_type_id;
+}
+
+
+static inline __u16 btf_fwd_kind(struct btf_type *t)
+{
+	return BTF_INFO_KFLAG(t->info) ? BTF_KIND_UNION : BTF_KIND_STRUCT;
+}
+
+/*
+ * Check equivalence of BTF type graph formed by candidate struct/union (we'll
+ * call it "candidate graph" in this description for brevity) to a type graph
+ * formed by (potential) canonical struct/union ("canonical graph" for brevity
+ * here, though keep in mind that not all types in canonical graph are
+ * necessarily canonical representatives themselves, some of them might be
+ * duplicates or its uniqueness might not have been established yet).
+ * Returns:
+ *  - >0, if type graphs are equivalent;
+ *  -  0, if not equivalent;
+ *  - <0, on error.
+ *
+ * Algorithm performs side-by-side DFS traversal of both type graphs and checks
+ * equivalence of BTF types at each step. If at any point BTF types in candidate
+ * and canonical graphs are not compatible structurally, whole graphs are
+ * incompatible. If types are structurally equivalent (i.e., all information
+ * except referenced type IDs is exactly the same), a mapping from `canon_id` to
+ * a `cand_id` is recored in hypothetical mapping (`btf_dedup->hypot_map`).
+ * If a type references other types, then those referenced types are checked
+ * for equivalence recursively.
+ *
+ * During DFS traversal, if we find that for current `canon_id` type we
+ * already have some mapping in hypothetical map, we check for two possible
+ * situations:
+ *   - `canon_id` is mapped to exactly the same type as `cand_id`. This will
+ *     happen when type graphs have cycles. In this case we assume those two
+ *     types are equivalent.
+ *   - `canon_id` is mapped to different type. This is contradiction in our
+ *     hypothetical mapping, because same graph in canonical graph corresponds
+ *     to two different types in candidate graph, which for equivalent type
+ *     graphs shouldn't happen. This condition terminates equivalence check
+ *     with negative result.
+ *
+ * If type graphs traversal exhausts types to check and find no contradiction,
+ * then type graphs are equivalent.
+ *
+ * When checking types for equivalence, there is one special case: FWD types.
+ * If FWD type resolution is allowed and one of the types (either from canonical
+ * or candidate graph) is FWD and other is STRUCT/UNION (depending on FWD's kind
+ * flag) and their names match, hypothetical mapping is updated to point from
+ * FWD to STRUCT/UNION. If graphs will be determined as equivalent successfully,
+ * this mapping will be used to record FWD -> STRUCT/UNION mapping permanently.
+ *
+ * Technically, this could lead to incorrect FWD to STRUCT/UNION resolution,
+ * if there are two exactly named (or anonymous) structs/unions that are
+ * compatible structurally, one of which has FWD field, while other is concrete
+ * STRUCT/UNION, but according to C sources they are different structs/unions
+ * that are referencing different types with the same name. This is extremely
+ * unlikely to happen, but btf_dedup API allows to disable FWD resolution if
+ * this logic is causing problems.
+ *
+ * Doing FWD resolution means that both candidate and/or canonical graphs can
+ * consists of portions of the graph that come from multiple compilation units.
+ * This is due to the fact that types within single compilation unit are always