On Wed, Jul 06, 2022 at 10:48:57AM -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.
>
> Let's add a new btf_func_returns_void() wrapper which is more defensive
> and use it in the places where we currently do '!->type' check.
>
> Fixes: 69fd337a975c ("bpf: per-cgroup lsm flavor")
> Reported-by: syzbot+5cc0730bd4b4d2c5f152@xxxxxxxxxxxxxxxxxxxxxxxxx
> Signed-off-by: Stanislav Fomichev <sdf@xxxxxxxxxx>
> ---
> include/linux/btf.h | 5 +++++
> kernel/bpf/trampoline.c | 2 +-
> kernel/bpf/verifier.c | 8 ++++----
> 3 files changed, 10 insertions(+), 5 deletions(-)
>
> diff --git a/include/linux/btf.h b/include/linux/btf.h
> index 1bfed7fa0428..17ba7d27a8ad 100644
> --- a/include/linux/btf.h
> +++ b/include/linux/btf.h
> @@ -302,6 +302,11 @@ static inline u16 btf_func_linkage(const struct btf_type *t)
> return BTF_INFO_VLEN(t->info);
> }
>
> +static inline bool btf_func_returns_void(const struct btf_type *t)
> +{
> + return t && !t->type;
> +}
> +
> static inline bool btf_type_kflag(const struct btf_type *t)
> {
> return BTF_INFO_KFLAG(t->info);
> diff --git a/kernel/bpf/trampoline.c b/kernel/bpf/trampoline.c
> index 6cd226584c33..9c4cb4c8a5fa 100644
> --- a/kernel/bpf/trampoline.c
> +++ b/kernel/bpf/trampoline.c
> @@ -400,7 +400,7 @@ static enum bpf_tramp_prog_type bpf_attach_type_to_tramp(struct bpf_prog *prog)
> case BPF_TRACE_FEXIT:
> return BPF_TRAMP_FEXIT;
> case BPF_LSM_MAC:
> - if (!prog->aux->attach_func_proto->type)
> + if (btf_func_returns_void(prog->aux->attach_func_proto))
> /* The function returns void, we cannot modify its
> * return value.
> */
> diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> index df3ec6b05f05..e3ee6f70939b 100644
> --- a/kernel/bpf/verifier.c
> +++ b/kernel/bpf/verifier.c
> @@ -7325,7 +7325,7 @@ static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn
> break;
> case BPF_FUNC_set_retval:
> if (env->prog->expected_attach_type == BPF_LSM_CGROUP) {
> - if (!env->prog->aux->attach_func_proto->type) {
> + if (btf_func_returns_void(env->prog->aux->attach_func_proto)) {
> /* Make sure programs that attach to void
> * hooks don't try to modify return value.
> */
> @@ -10447,7 +10447,7 @@ static int check_return_code(struct bpf_verifier_env *env)
> if (!is_subprog &&
> (prog_type == BPF_PROG_TYPE_STRUCT_OPS ||
> prog_type == BPF_PROG_TYPE_LSM) &&
> - !prog->aux->attach_func_proto->type)
> + btf_func_returns_void(prog->aux->attach_func_proto))
> return 0;
It seems there is another problem here.
It returns early here for prog_type == BPF_PROG_TYPE_LSM.
It should only do that for expected_attach_type != BPF_LSM_CGROUP.
Otherwise, the later verbose(env, "Note, BPF_LSM_CGROUP...") will
not get a chance.
>
> /* eBPF calling convention is such that R0 is used
> @@ -10547,7 +10547,7 @@ static int check_return_code(struct bpf_verifier_env *env)
> */
> return 0;
> }
> - if (!env->prog->aux->attach_func_proto->type) {
> + if (btf_func_returns_void(env->prog->aux->attach_func_proto)) {
> /* Make sure programs that attach to void
> * hooks don't try to modify return value.
> */
> @@ -10572,7 +10572,7 @@ 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 &&
I think the problem is more like missing
prog_type == BPF_PROG_TYPE_LSM [&& expected_attach_type == BPF_LSM_CGROUP] here
instead of testing !attach_func_proto in all places.
> - !prog->aux->attach_func_proto->type)
> + btf_func_returns_void(prog->aux->attach_func_proto))
> verbose(env, "Note, BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
> return -EINVAL;
> }
> --
> 2.37.0.rc0.161.g10f37bed90-goog
>
