Merge branch 'fix-global-subprog-ptr_to_ctx-arg-handling'

Andrii Nakryiko says:

====================
Fix global subprog PTR_TO_CTX arg handling

Fix confusing and incorrect inference of PTR_TO_CTX argument type in BPF
global subprogs. For some program types (iters, tracepoint, any program type
that doesn't have fixed named "canonical" context type) when user uses (in
a correct and valid way) a pointer argument to user-defined anonymous struct
type, verifier will incorrectly assume that it has to be PTR_TO_CTX argument.
While it should be just a PTR_TO_MEM argument with allowed size calculated
from user-provided (even if anonymous) struct.

This did come up in practice and was very confusing to users, so let's prevent
this going forward. We had to do a slight refactoring of
btf_get_prog_ctx_type() to make it easy to support a special s390x KPROBE use
cases. See details in respective patches.

v1->v2:
- special-case typedef bpf_user_pt_regs_t handling for KPROBE programs,
fixing s390x after changes in patch #2.
====================

Link: https://lore.kernel.org/r/[e-mail]
Signed-off-by: Alexei Starovoitov <[e-mail]>

include/linux/btf.h

···
 int register_btf_id_dtor_kfuncs(const struct btf_id_dtor_kfunc *dtors, u32 add_cnt,
 				struct module *owner);
 struct btf_struct_meta *btf_find_struct_meta(const struct btf *btf, u32 btf_id);
-const struct btf_type *
-btf_get_prog_ctx_type(struct bpf_verifier_log *log, const struct btf *btf,
-		      const struct btf_type *t, enum bpf_prog_type prog_type,
-		      int arg);
+bool btf_is_prog_ctx_type(struct bpf_verifier_log *log, const struct btf *btf,
+			   const struct btf_type *t, enum bpf_prog_type prog_type,
+			   int arg);
 int get_kern_ctx_btf_id(struct bpf_verifier_log *log, enum bpf_prog_type prog_type);
 bool btf_types_are_same(const struct btf *btf1, u32 id1,
 			const struct btf *btf2, u32 id2);
···
 {
 	return NULL;
 }
-static inline const struct btf_member *
-btf_get_prog_ctx_type(struct bpf_verifier_log *log, const struct btf *btf,
-		      const struct btf_type *t, enum bpf_prog_type prog_type,
-		      int arg)
+static inline bool
+btf_is_prog_ctx_type(struct bpf_verifier_log *log, const struct btf *btf,
+		     const struct btf_type *t, enum bpf_prog_type prog_type,
+		     int arg)
 {
-	return NULL;
+	return false;
 }
 static inline int get_kern_ctx_btf_id(struct bpf_verifier_log *log,
 				      enum bpf_prog_type prog_type) {

kernel/bpf/btf.c

···
 	return ctx_type->type;
 }
 
-const struct btf_type *
-btf_get_prog_ctx_type(struct bpf_verifier_log *log, const struct btf *btf,
-		      const struct btf_type *t, enum bpf_prog_type prog_type,
-		      int arg)
+bool btf_is_prog_ctx_type(struct bpf_verifier_log *log, const struct btf *btf,
+			  const struct btf_type *t, enum bpf_prog_type prog_type,
+			  int arg)
 {
 	const struct btf_type *ctx_type;
 	const char *tname, *ctx_tname;
 
 	t = btf_type_by_id(btf, t->type);
+
+	/* KPROBE programs allow bpf_user_pt_regs_t typedef, which we need to
+	 * check before we skip all the typedef below.
+	 */
+	if (prog_type == BPF_PROG_TYPE_KPROBE) {
+		while (btf_type_is_modifier(t) && !btf_type_is_typedef(t))
+			t = btf_type_by_id(btf, t->type);
+
+		if (btf_type_is_typedef(t)) {
+			tname = btf_name_by_offset(btf, t->name_off);
+			if (tname && strcmp(tname, "bpf_user_pt_regs_t") == 0)
+				return true;
+		}
+	}
+
 	while (btf_type_is_modifier(t))
 		t = btf_type_by_id(btf, t->type);
 	if (!btf_type_is_struct(t)) {
···
 		 * is not supported yet.
 		 * BPF_PROG_TYPE_RAW_TRACEPOINT is fine.
 		 */
-		return NULL;
+		return false;
 	}
 	tname = btf_name_by_offset(btf, t->name_off);
 	if (!tname) {
 		bpf_log(log, "arg#%d struct doesn't have a name\n", arg);
-		return NULL;
+		return false;
 	}
 
 	ctx_type = find_canonical_prog_ctx_type(prog_type);
 	if (!ctx_type) {
 		bpf_log(log, "btf_vmlinux is malformed\n");
 		/* should not happen */
-		return NULL;
+		return false;
 	}
 again:
 	ctx_tname = btf_name_by_offset(btf_vmlinux, ctx_type->name_off);
 	if (!ctx_tname) {
 		/* should not happen */
 		bpf_log(log, "Please fix kernel include/linux/bpf_types.h\n");
-		return NULL;
+		return false;
 	}
+	/* program types without named context types work only with arg:ctx tag */
+	if (ctx_tname[0] == '\0')
+		return false;
 	/* only compare that prog's ctx type name is the same as
 	 * kernel expects. No need to compare field by field.
 	 * It's ok for bpf prog to do:
···
 	 * { // no fields of skb are ever used }
 	 */
 	if (strcmp(ctx_tname, "__sk_buff") == 0 && strcmp(tname, "sk_buff") == 0)
-		return ctx_type;
+		return true;
 	if (strcmp(ctx_tname, "xdp_md") == 0 && strcmp(tname, "xdp_buff") == 0)
-		return ctx_type;
+		return true;
 	if (strcmp(ctx_tname, tname)) {
 		/* bpf_user_pt_regs_t is a typedef, so resolve it to
 		 * underlying struct and check name again
 		 */
 		if (!btf_type_is_modifier(ctx_type))
-			return NULL;
+			return false;
 		while (btf_type_is_modifier(ctx_type))
 			ctx_type = btf_type_by_id(btf_vmlinux, ctx_type->type);
 		goto again;
 	}
-	return ctx_type;
+	return true;
 }
 
 /* forward declarations for arch-specific underlying types of
···
 				     enum bpf_prog_type prog_type,
 				     int arg)
 {
-	if (!btf_get_prog_ctx_type(log, btf, t, prog_type, arg))
+	if (!btf_is_prog_ctx_type(log, btf, t, prog_type, arg))
 		return -ENOENT;
 	return find_kern_ctx_type_id(prog_type);
 }
···
 		if (!btf_type_is_ptr(t))
 			goto skip_pointer;
 
-		if ((tags & ARG_TAG_CTX) || btf_get_prog_ctx_type(log, btf, t, prog_type, i)) {
+		if ((tags & ARG_TAG_CTX) || btf_is_prog_ctx_type(log, btf, t, prog_type, i)) {
 			if (tags & ~ARG_TAG_CTX) {
 				bpf_log(log, "arg#%d has invalid combination of tags\n", i);
 				return -EINVAL;

kernel/bpf/verifier.c

···
 	 * type to our caller. When a set of conditions hold in the BTF type of
 	 * arguments, we resolve it to a known kfunc_ptr_arg_type.
 	 */
-	if (btf_get_prog_ctx_type(&env->log, meta->btf, t, resolve_prog_type(env->prog), argno))
+	if (btf_is_prog_ctx_type(&env->log, meta->btf, t, resolve_prog_type(env->prog), argno))
 		return KF_ARG_PTR_TO_CTX;
 
 	if (is_kfunc_arg_alloc_obj(meta->btf, &args[argno]))

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

···
 	return kprobe_typedef_ctx_subprog(ctx);
 }
 
+/* s390x defines:
+ *
+ * typedef user_pt_regs bpf_user_pt_regs_t;
+ * typedef struct { ... } user_pt_regs;
+ *
+ * And so "canonical" underlying struct type is anonymous.
+ * So on s390x only valid ways to have PTR_TO_CTX argument in global subprogs
+ * are:
+ *   - bpf_user_pt_regs_t *ctx (typedef);
+ *   - struct bpf_user_pt_regs_t *ctx (backwards compatible struct hack);
+ *   - void *ctx __arg_ctx (arg:ctx tag)
+ *
+ * Other architectures also allow using underlying struct types (e.g.,
+ * `struct pt_regs *ctx` for x86-64)
+ */
+#ifndef bpf_target_s390
+
 #define pt_regs_struct_t typeof(*(__PT_REGS_CAST((struct pt_regs *)NULL)))
 
 __weak int kprobe_struct_ctx_subprog(pt_regs_struct_t *ctx)
···
 {
 	return kprobe_struct_ctx_subprog(ctx);
 }
+
+#endif
 
 /* this is current hack to make this work on old kernels */
 struct bpf_user_pt_regs_t {};

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

···
 	return subprog_nullable_ptr_bad(&x);
 }
 
+typedef struct {
+	int x;
+} user_struct_t;
+
+__noinline __weak int subprog_user_anon_mem(user_struct_t *t)
+{
+	return t ? t->x : 0;
+}
+
+SEC("?tracepoint")
+__failure __log_level(2)
+__msg("invalid bpf_context access")
+__msg("Caller passes invalid args into func#1 ('subprog_user_anon_mem')")
+int anon_user_mem_invalid(void *ctx)
+{
+	/* can't pass PTR_TO_CTX as user memory */
+	return subprog_user_anon_mem(ctx);
+}
+
+SEC("?tracepoint")
+__success __log_level(2)
+__msg("Func#1 ('subprog_user_anon_mem') is safe for any args that match its prototype")
+int anon_user_mem_valid(void *ctx)
+{
+	user_struct_t t = { .x = 42 };
+
+	return subprog_user_anon_mem(&t);
+}
+
 __noinline __weak int subprog_nonnull_ptr_good(int *p1 __arg_nonnull, int *p2 __arg_nonnull)
 {
 	return (*p1) * (*p2); /* good, no need for NULL checks */