Merge branch 'libbpf: support non-mmap()'able data sections'

Andrii Nakryiko says:

====================

Make libbpf more conservative in using BPF_F_MMAPABLE flag with internal BPF
array maps that are backing global data sections. See patch #2 for full
description and justification.

Changes in this dataset support having bpf_spinlock, kptr, rb_tree nodes and
other "special" variables as global variables. Combining this with libbpf's
existing support for multiple custom .data.* sections allows BPF programs to
utilize multiple spinlock/rbtree_node/kptr variables in a pretty natural way
by just putting all such variables into separate data sections (and thus ARRAY
maps).

v1->v2:
  - address Stanislav's feedback, adds acks.
====================

Acked-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

tools/lib/bpf/libbpf.c

@@ -1461,15 +1461,12 @@ static int find_elf_sec_sz(const struct bpf_object *obj, const char *name, __u32
 	return -ENOENT;
 }
 
-static int find_elf_var_offset(const struct bpf_object *obj, const char *name, __u32 *off)
+static Elf64_Sym *find_elf_var_sym(const struct bpf_object *obj, const char *name)
 {
 	Elf_Data *symbols = obj->efile.symbols;
 	const char *sname;
 	size_t si;
 
-	if (!name || !off)
-		return -EINVAL;
-
 	for (si = 0; si < symbols->d_size / sizeof(Elf64_Sym); si++) {
 		Elf64_Sym *sym = elf_sym_by_idx(obj, si);
 
@@ -1483,15 +1480,13 @@ static int find_elf_var_offset(const struct bpf_object *obj, const char *name, _
 		sname = elf_sym_str(obj, sym->st_name);
 		if (!sname) {
 			pr_warn("failed to get sym name string for var %s\n", name);
-			return -EIO;
-		}
-		if (strcmp(name, sname) == 0) {
-			*off = sym->st_value;
-			return 0;
+			return ERR_PTR(-EIO);
 		}
+		if (strcmp(name, sname) == 0)
+			return sym;
 	}
 
-	return -ENOENT;
+	return ERR_PTR(-ENOENT);
 }
 
 static struct bpf_map *bpf_object__add_map(struct bpf_object *obj)
@@ -1582,7 +1577,38 @@ static char *internal_map_name(struct bpf_object *obj, const char *real_name)
 }
 
 static int
-bpf_map_find_btf_info(struct bpf_object *obj, struct bpf_map *map);
+map_fill_btf_type_info(struct bpf_object *obj, struct bpf_map *map);
+
+/* Internal BPF map is mmap()'able only if at least one of corresponding
+ * DATASEC's VARs are to be exposed through BPF skeleton. I.e., it's a GLOBAL
+ * variable and it's not marked as __hidden (which turns it into, effectively,
+ * a STATIC variable).
+ */
+static bool map_is_mmapable(struct bpf_object *obj, struct bpf_map *map)
+{
+	const struct btf_type *t, *vt;
+	struct btf_var_secinfo *vsi;
+	int i, n;
+
+	if (!map->btf_value_type_id)
+		return false;
+
+	t = btf__type_by_id(obj->btf, map->btf_value_type_id);
+	if (!btf_is_datasec(t))
+		return false;
+
+	vsi = btf_var_secinfos(t);
+	for (i = 0, n = btf_vlen(t); i < n; i++, vsi++) {
+		vt = btf__type_by_id(obj->btf, vsi->type);
+		if (!btf_is_var(vt))
+			continue;
+
+		if (btf_var(vt)->linkage != BTF_VAR_STATIC)
+			return true;
+	}
+
+	return false;
+}
 
 static int
 bpf_object__init_internal_map(struct bpf_object *obj, enum libbpf_map_type type,
@@ -1614,7 +1640,12 @@ bpf_object__init_internal_map(struct bpf_object *obj, enum libbpf_map_type type,
 	def->max_entries = 1;
 	def->map_flags = type == LIBBPF_MAP_RODATA || type == LIBBPF_MAP_KCONFIG
 			 ? BPF_F_RDONLY_PROG : 0;
-	def->map_flags |= BPF_F_MMAPABLE;
+
+	/* failures are fine because of maps like .rodata.str1.1 */
+	(void) map_fill_btf_type_info(obj, map);
+
+	if (map_is_mmapable(obj, map))
+		def->map_flags |= BPF_F_MMAPABLE;
 
 	pr_debug("map '%s' (global data): at sec_idx %d, offset %zu, flags %x.\n",
 		 map->name, map->sec_idx, map->sec_offset, def->map_flags);
@@ -1631,9 +1662,6 @@ bpf_object__init_internal_map(struct bpf_object *obj, enum libbpf_map_type type,
 		return err;
 	}
 
-	/* failures are fine because of maps like .rodata.str1.1 */
-	(void) bpf_map_find_btf_info(obj, map);
-
 	if (data)
 		memcpy(map->mmaped, data, data_sz);
 
@@ -2545,7 +2573,7 @@ static int bpf_object__init_user_btf_map(struct bpf_object *obj,
 		fill_map_from_def(map->inner_map, &inner_def);
 	}
 
-	err = bpf_map_find_btf_info(obj, map);
+	err = map_fill_btf_type_info(obj, map);
 	if (err)
 		return err;
 
@@ -2850,57 +2878,89 @@ static int compare_vsi_off(const void *_a, const void *_b)
 static int btf_fixup_datasec(struct bpf_object *obj, struct btf *btf,
 			     struct btf_type *t)
 {
-	__u32 size = 0, off = 0, i, vars = btf_vlen(t);
-	const char *name = btf__name_by_offset(btf, t->name_off);
-	const struct btf_type *t_var;
+	__u32 size = 0, i, vars = btf_vlen(t);
+	const char *sec_name = btf__name_by_offset(btf, t->name_off);
 	struct btf_var_secinfo *vsi;
-	const struct btf_var *var;
-	int ret;
+	bool fixup_offsets = false;
+	int err;
 
-	if (!name) {
+	if (!sec_name) {
 		pr_debug("No name found in string section for DATASEC kind.\n");
 		return -ENOENT;
 	}
 
-	/* .extern datasec size and var offsets were set correctly during
-	 * extern collection step, so just skip straight to sorting variables
+	/* Extern-backing datasecs (.ksyms, .kconfig) have their size and
+	 * variable offsets set at the previous step. Further, not every
+	 * extern BTF VAR has corresponding ELF symbol preserved, so we skip
+	 * all fixups altogether for such sections and go straight to sorting
+	 * VARs within their DATASEC.
 	 */
-	if (t->size)
+	if (strcmp(sec_name, KCONFIG_SEC) == 0 || strcmp(sec_name, KSYMS_SEC) == 0)
 		goto sort_vars;
 
-	ret = find_elf_sec_sz(obj, name, &size);
-	if (ret || !size) {
-		pr_debug("Invalid size for section %s: %u bytes\n", name, size);
-		return -ENOENT;
+	/* Clang leaves DATASEC size and VAR offsets as zeroes, so we need to
+	 * fix this up. But BPF static linker already fixes this up and fills
+	 * all the sizes and offsets during static linking. So this step has
+	 * to be optional. But the STV_HIDDEN handling is non-optional for any
+	 * non-extern DATASEC, so the variable fixup loop below handles both
+	 * functions at the same time, paying the cost of BTF VAR <-> ELF
+	 * symbol matching just once.
+	 */
+	if (t->size == 0) {
+		err = find_elf_sec_sz(obj, sec_name, &size);
+		if (err || !size) {
+			pr_debug("sec '%s': failed to determine size from ELF: size %u, err %d\n",
+				 sec_name, size, err);
+			return -ENOENT;
+		}
+
+		t->size = size;
+		fixup_offsets = true;
 	}
 
-	t->size = size;
-
 	for (i = 0, vsi = btf_var_secinfos(t); i < vars; i++, vsi++) {
+		const struct btf_type *t_var;
+		struct btf_var *var;
+		const char *var_name;
+		Elf64_Sym *sym;
+
 		t_var = btf__type_by_id(btf, vsi->type);
 		if (!t_var || !btf_is_var(t_var)) {
-			pr_debug("Non-VAR type seen in section %s\n", name);
+			pr_debug("sec '%s': unexpected non-VAR type found\n", sec_name);
 			return -EINVAL;
 		}
 
 		var = btf_var(t_var);
-		if (var->linkage == BTF_VAR_STATIC)
+		if (var->linkage == BTF_VAR_STATIC || var->linkage == BTF_VAR_GLOBAL_EXTERN)
 			continue;
 
-		name = btf__name_by_offset(btf, t_var->name_off);
-		if (!name) {
-			pr_debug("No name found in string section for VAR kind\n");
+		var_name = btf__name_by_offset(btf, t_var->name_off);
+		if (!var_name) {
+			pr_debug("sec '%s': failed to find name of DATASEC's member #%d\n",
+				 sec_name, i);
 			return -ENOENT;
 		}
 
-		ret = find_elf_var_offset(obj, name, &off);
-		if (ret) {
-			pr_debug("No offset found in symbol table for VAR %s\n",
-				 name);
+		sym = find_elf_var_sym(obj, var_name);
+		if (IS_ERR(sym)) {
+			pr_debug("sec '%s': failed to find ELF symbol for VAR '%s'\n",
+				 sec_name, var_name);
 			return -ENOENT;
 		}
 
-		vsi->offset = off;
+		if (fixup_offsets)
+			vsi->offset = sym->st_value;
+
+		/* if variable is a global/weak symbol, but has restricted
+		 * (STV_HIDDEN or STV_INTERNAL) visibility, mark its BTF VAR
+		 * as static. This follows similar logic for functions (BPF
+		 * subprogs) and influences libbpf's further decisions about
+		 * whether to make global data BPF array maps as
+		 * BPF_F_MMAPABLE.
+		 */
+		if (ELF64_ST_VISIBILITY(sym->st_other) == STV_HIDDEN
+		    || ELF64_ST_VISIBILITY(sym->st_other) == STV_INTERNAL)
+			var->linkage = BTF_VAR_STATIC;
 	}
 
 sort_vars:
@@ -2908,13 +2968,16 @@ sort_vars:
 	return 0;
 }
 
-static int btf_finalize_data(struct bpf_object *obj, struct btf *btf)
+static int bpf_object_fixup_btf(struct bpf_object *obj)
 {
-	int err = 0;
-	__u32 i, n = btf__type_cnt(btf);
+	int i, n, err = 0;
 
+	if (!obj->btf)
+		return 0;
+
+	n = btf__type_cnt(obj->btf);
 	for (i = 1; i < n; i++) {
-		struct btf_type *t = btf_type_by_id(btf, i);
+		struct btf_type *t = btf_type_by_id(obj->btf, i);
 
 		/* Loader needs to fix up some of the things compiler
 		 * couldn't get its hands on while emitting BTF. This
@@ -2922,28 +2985,12 @@ static int btf_finalize_data(struct bpf_object *obj, struct btf *btf)
 		 * the info from the ELF itself for this purpose.
 		 */
 		if (btf_is_datasec(t)) {
-			err = btf_fixup_datasec(obj, btf, t);
+			err = btf_fixup_datasec(obj, obj->btf, t);
 			if (err)
-				break;
+				return err;
 		}
 	}
 
-	return libbpf_err(err);
-}
-
-static int bpf_object__finalize_btf(struct bpf_object *obj)
-{
-	int err;
-
-	if (!obj->btf)
-		return 0;
-
-	err = btf_finalize_data(obj, obj->btf);
-	if (err) {
-		pr_warn("Error finalizing %s: %d.\n", BTF_ELF_SEC, err);
-		return err;
-	}
-
 	return 0;
 }
 
@@ -4235,7 +4282,7 @@ bpf_object__collect_prog_relos(struct bpf_object *obj, Elf64_Shdr *shdr, Elf_Dat
 	return 0;
 }
 
-static int bpf_map_find_btf_info(struct bpf_object *obj, struct bpf_map *map)
+static int map_fill_btf_type_info(struct bpf_object *obj, struct bpf_map *map)
 {
 	int id;
 
@@ -7233,7 +7280,7 @@ static struct bpf_object *bpf_object_open(const char *path, const void *obj_buf,
 	err = err ? : bpf_object__check_endianness(obj);
 	err = err ? : bpf_object__elf_collect(obj);
 	err = err ? : bpf_object__collect_externs(obj);
-	err = err ? : bpf_object__finalize_btf(obj);
+	err = err ? : bpf_object_fixup_btf(obj);
 	err = err ? : bpf_object__init_maps(obj, opts);
 	err = err ? : bpf_object_init_progs(obj, opts);
 	err = err ? : bpf_object__collect_relos(obj);

tools/testing/selftests/bpf/prog_tests/skeleton.c

@@ -2,6 +2,7 @@
 /* Copyright (c) 2019 Facebook */
 
 #include <test_progs.h>
+#include <sys/mman.h>
 
 struct s {
 	int a;
@@ -22,7 +23,8 @@ void test_skeleton(void)
 	struct test_skeleton__kconfig *kcfg;
 	const void *elf_bytes;
 	size_t elf_bytes_sz = 0;
-	int i;
+	void *m;
+	int i, fd;
 
 	skel = test_skeleton__open();
 	if (CHECK(!skel, "skel_open", "failed to open skeleton\n"))
@@ -124,6 +126,13 @@ void test_skeleton(void)
 
 	ASSERT_EQ(bss->huge_arr[ARRAY_SIZE(bss->huge_arr) - 1], 123, "huge_arr");
 
+	fd = bpf_map__fd(skel->maps.data_non_mmapable);
+	m = mmap(NULL, getpagesize(), PROT_READ, MAP_SHARED, fd, 0);
+	if (!ASSERT_EQ(m, MAP_FAILED, "unexpected_mmap_success"))
+		munmap(m, getpagesize());
+
+	ASSERT_EQ(bpf_map__map_flags(skel->maps.data_non_mmapable), 0, "non_mmap_flags");
+
 	elf_bytes = test_skeleton__elf_bytes(&elf_bytes_sz);
 	ASSERT_OK_PTR(elf_bytes, "elf_bytes");
 	ASSERT_GE(elf_bytes_sz, 0, "elf_bytes_sz");

tools/testing/selftests/bpf/progs/test_skeleton.c

@@ -53,6 +53,20 @@ int out_mostly_var;
 
 char huge_arr[16 * 1024 * 1024];
 
+/* non-mmapable custom .data section */
+
+struct my_value { int x, y, z; };
+
+__hidden int zero_key SEC(".data.non_mmapable");
+static struct my_value zero_value SEC(".data.non_mmapable");
+
+struct {
+	__uint(type, BPF_MAP_TYPE_ARRAY);
+	__type(key, int);
+	__type(value, struct my_value);
+	__uint(max_entries, 1);
+} my_map SEC(".maps");
+
 SEC("raw_tp/sys_enter")
 int handler(const void *ctx)
 {
@@ -75,6 +89,9 @@ int handler(const void *ctx)
 
 	huge_arr[sizeof(huge_arr) - 1] = 123;
 
+	/* make sure zero_key and zero_value are not optimized out */
+	bpf_map_update_elem(&my_map, &zero_key, &zero_value, BPF_ANY);
+
 	return 0;
 }