bpf: Teach verifier that trusted PTR_TO_BTF_ID pointers are non-NULL

In reg_type_not_null(), we currently assume that a pointer may be NULL
if it has the PTR_MAYBE_NULL modifier, or if it doesn't belong to one of
several base type of pointers that are never NULL-able. For example,
PTR_TO_CTX, PTR_TO_MAP_VALUE, etc.

It turns out that in some cases, PTR_TO_BTF_ID can never be NULL as
well, though we currently don't specify it. For example, if you had the
following program:

SEC("tc")
long example_refcnt_fail(void *ctx)
{
	struct bpf_cpumask *mask1, *mask2;

	mask1 = bpf_cpumask_create();
	mask2 = bpf_cpumask_create();

        if (!mask1 || !mask2)
		goto error_release;

	bpf_cpumask_test_cpu(0, (const struct cpumask *)mask1);
	bpf_cpumask_test_cpu(0, (const struct cpumask *)mask2);

error_release:
	if (mask1)
		bpf_cpumask_release(mask1);
	if (mask2)
		bpf_cpumask_release(mask2);
	return ret;
}

The verifier will incorrectly fail to load the program, thinking
(unintuitively) that we have a possibly-unreleased reference if the mask
is NULL, because we (correctly) don't issue a bpf_cpumask_release() on
the NULL path.

The reason the verifier gets confused is due to the fact that we don't
explicitly tell the verifier that trusted PTR_TO_BTF_ID pointers can
never be NULL. Basically, if we successfully get past the if check
(meaning both pointers go from ptr_or_null_bpf_cpumask to
ptr_bpf_cpumask), the verifier will correctly assume that the references
need to be dropped on any possible branch that leads to program exit.
However, it will _incorrectly_ think that the ptr == NULL branch is
possible, and will erroneously detect it as a branch on which we failed
to drop the reference.

The solution is of course to teach the verifier that trusted
PTR_TO_BTF_ID pointers can never be NULL, so that it doesn't incorrectly
think it's possible for the reference to be present on the ptr == NULL
branch.

A follow-on patch will add a selftest that verifies this behavior.

Signed-off-by: David Vernet <void@manifault.com>
Link: https://lore.kernel.org/r/20230602150112.1494194-1-void@manifault.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

kernel/bpf/verifier.c

@@ -197,6 +197,7 @@ static int ref_set_non_owning(struct bpf_verifier_env *env,
 			      struct bpf_reg_state *reg);
 static void specialize_kfunc(struct bpf_verifier_env *env,
 			     u32 func_id, u16 offset, unsigned long *addr);
+static bool is_trusted_reg(const struct bpf_reg_state *reg);
 
 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
 {
@@ -442,8 +443,11 @@ static bool type_may_be_null(u32 type)
 	return type & PTR_MAYBE_NULL;
 }
 
-static bool reg_type_not_null(enum bpf_reg_type type)
+static bool reg_not_null(const struct bpf_reg_state *reg)
 {
+	enum bpf_reg_type type;
+
+	type = reg->type;
 	if (type_may_be_null(type))
 		return false;
 
@@ -453,6 +457,7 @@ static bool reg_type_not_null(enum bpf_reg_type type)
 		type == PTR_TO_MAP_VALUE ||
 		type == PTR_TO_MAP_KEY ||
 		type == PTR_TO_SOCK_COMMON ||
+		(type == PTR_TO_BTF_ID && is_trusted_reg(reg)) ||
 		type == PTR_TO_MEM;
 }
 
@@ -13170,7 +13175,7 @@ static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
 			   bool is_jmp32)
 {
 	if (__is_pointer_value(false, reg)) {
-		if (!reg_type_not_null(reg->type))
+		if (!reg_not_null(reg))
 			return -1;
 
 		/* If pointer is valid tests against zero will fail so we can