On Thu, Jul 07, 2022 at 09:02:33AM -0700, Stanislav Fomichev wrote:
> Syzkaller reports the following crash:
> RIP: 0010:check_return_code kernel/bpf/verifier.c:10575 [inline]
> RIP: 0010:do_check kernel/bpf/verifier.c:12346 [inline]
> RIP: 0010:do_check_common+0xb3d2/0xd250 kernel/bpf/verifier.c:14610
>
> With the following reproducer:
> bpf$PROG_LOAD_XDP(0x5, &(0x7f00000004c0)={0xd, 0x3, &(0x7f0000000000)=ANY=[@ANYBLOB="1800000000000019000000000000000095"], &(0x7f0000000300)='GPL\x00', 0x0, 0x0, 0x0, 0x0, 0x0, '\x00', 0x0, 0x2b, 0xffffffffffffffff, 0x8, 0x0, 0x0, 0x10, 0x0}, 0x80)
>
> Because we don't enforce expected_attach_type for XDP programs,
> we end up in hitting 'if (prog->expected_attach_type == BPF_LSM_CGROUP'
> part in check_return_code and follow up with testing
> `prog->aux->attach_func_proto->type`, but `prog->aux->attach_func_proto`
> is NULL.
>
> Add explicit prog_type check for the "Note, BPF_LSM_CGROUP that
> attach ..." condition. Also, don't skip return code check for
> LSM/STRUCT_OPS.
>
> The above actually brings an issue with existing selftest which
> tries to return EPERM from void inet_csk_clone. Fix the
> test (and move called_socket_clone to make sure it's not
> incremented in case of an error) and add a new one to explicitly
> verify this condition.
>
> v2:
> - Martin: don't add new helper, check prog_type instead
> - Martin: check expected_attach_type as well at the function entry
> - Update selftest to verify this condition
>
> Fixes: 69fd337a975c ("bpf: per-cgroup lsm flavor")
> Reported-by: syzbot+5cc0730bd4b4d2c5f152@xxxxxxxxxxxxxxxxxxxxxxxxx
> Signed-off-by: Stanislav Fomichev <sdf@xxxxxxxxxx>
> ---
> kernel/bpf/verifier.c | 2 ++
> .../testing/selftests/bpf/prog_tests/lsm_cgroup.c | 12 ++++++++++++
> tools/testing/selftests/bpf/progs/lsm_cgroup.c | 12 ++++++------
> .../selftests/bpf/progs/lsm_cgroup_nonvoid.c | 14 ++++++++++++++
> 4 files changed, 34 insertions(+), 6 deletions(-)
> create mode 100644 tools/testing/selftests/bpf/progs/lsm_cgroup_nonvoid.c
>
> diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> index df3ec6b05f05..2bc1e7252778 100644
> --- a/kernel/bpf/verifier.c
> +++ b/kernel/bpf/verifier.c
> @@ -10445,6 +10445,7 @@ static int check_return_code(struct bpf_verifier_env *env)
>
> /* LSM and struct_ops func-ptr's return type could be "void" */
> if (!is_subprog &&
> + prog->expected_attach_type != BPF_LSM_CGROUP &&
BPF_PROG_TYPE_STRUCT_OPS also uses the expected_attach_type,
so the expected_attach_type check should only be done for LSM prog alone.
Others lgtm.
> (prog_type == BPF_PROG_TYPE_STRUCT_OPS ||
> prog_type == BPF_PROG_TYPE_LSM) &&
> !prog->aux->attach_func_proto->type)
> @@ -10572,6 +10573,7 @@ static int check_return_code(struct bpf_verifier_env *env)
> if (!tnum_in(range, reg->var_off)) {
> verbose_invalid_scalar(env, reg, &range, "program exit", "R0");
> if (prog->expected_attach_type == BPF_LSM_CGROUP &&
> + prog_type == BPF_PROG_TYPE_LSM &&
> !prog->aux->attach_func_proto->type)
> verbose(env, "Note, BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
> return -EINVAL;
