Re: [PATCH bpf-next v4 2/2] selftests/bpf: Add overflow test for BTF dump zero-scan

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Medium] selftests/bpf: test_btf_dump_struct_data() fails on ENODATA check
--

commit 1be6c5280eb4868315725b469eaa863a95165c0b
Author: Gu Bin <gubin@cmss.chinamobile.com>

selftests/bpf: Add overflow test for BTF dump zero-scan

This patch adds a TEST_BTF_DUMP_DATA_OVER() selftest case in
test_btf_dump_struct_data() that dumps an all-zero union bpf_iter_link_info
with a data_sz one byte short of the type size, expecting
btf_dump__dump_type_data() to return -E2BIG.

> diff --git a/tools/testing/selftests/bpf/prog_tests/btf_dump.c b/tools/testing/selftests/bpf/prog_tests/btf_dump.c
> index 35f214e254da8..68f33f8a54ab8 100644
> --- a/tools/testing/selftests/bpf/prog_tests/btf_dump.c
> +++ b/tools/testing/selftests/bpf/prog_tests/btf_dump.c
> @@ -1156,6 +1156,15 @@ static void test_btf_dump_struct_data(struct btf *btf, struct btf_dump *d,
>  			   "(union bpf_iter_link_info){.map = (struct){.map_fd = (__u32)1,},.cgroup = (struct){.order = (enum bpf_cgroup_iter_order)BPF_CGROUP_ITER_SELF_ONLY,.cgroup_fd = (__u32)1,},.task = (struct){.tid = (__u32)1,.pid = (__u32)1,},}",
>  			   { .cgroup = { .order = 1, .cgroup_fd = 1, }});
>  
> +	/*
> +	 * overflow test for all-zero union: with a data_sz one byte short
> +	 * of the type size, the zero-scan must not read past data_end and
> +	 * must return -E2BIG.
> +	 */
> +	TEST_BTF_DUMP_DATA_OVER(btf, d, "union", str, union bpf_iter_link_info,
> +				sizeof(union bpf_iter_link_info) - 1,
> +				"(union bpf_iter_link_info){\n", {});

[Severity: Medium]
As reported by Alexei Starovoitov on v3 [1], does this new test pass when
running test_progs?

When btf_dump_struct_data() iterates over the members of union
bpf_iter_link_info at depth 1, it invokes btf_dump_type_data_check_zero() on
the nested structs (.map, .cgroup, and .task). In
tools/lib/bpf/btf_dump.c:btf_dump_type_data_check_zero():

	case BTF_KIND_STRUCT:
	case BTF_KIND_UNION: {
		...
		for (i = 0; i < n; i++, m++) {
			...
			err = btf_dump_type_data_check_zero(d, mtype, m->type, data + moffset / 8,
							    moffset % 8, bit_sz);
			if (err != ENODATA)
				return err;
		}
		return -ENODATA;
	}

When checking the first member of .cgroup (.cgroup.order, a 4-byte enum at
offset 0), the recursive call sees a zero value within data_end and returns
-ENODATA. Because the loop compares err against positive ENODATA instead of
-ENODATA, -ENODATA != ENODATA evaluates to true and the loop returns
-ENODATA immediately after i == 0 without ever reaching .cgroup.cgroup_id at
offset 8 (the only member extending past sizeof(union bpf_iter_link_info) - 1).

Consequently, all union members are skipped as zeroed,
btf_dump__dump_type_data() returns 16 instead of -E2BIG, and str contains
"(union bpf_iter_link_info){\n}", failing the assertions in btf_dump_data().

[1] https://lore.kernel.org/r/DMLAKLO9AB4T.X5JCIISSONLF@gmail.com

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261104025806.1693-1-gubin@cmss.chinamobile.com?part=2