On 26/06/09 11:03PM, Nuoqi Gui wrote:
> CFG construction records the modeled gotox target set in
> insn_aux_data->jt. It includes INSN_ARRAY maps based on whether the map
> target is in the current subprog. check_indirect_jump() later validates and
> follows the current PTR_TO_INSN register's actual INSN_ARRAY map. The
> verifier does not check that targets copied from that map match the targets
> that CFG construction modeled for this gotox instruction.
>
> This lets one gotox instruction observe two different INSN_ARRAY maps. CFG
> can select a map whose target is in the current subprog. Another path to
> the same gotox can carry a PTR_TO_INSN value from a map whose target points
> at a different subprog. The verifier then accepts an edge absent from the
> CFG.
>
> On x86, gotox becomes a raw indirect jump in the JIT image. Accepting a
> target not modeled by CFG can enter another subprog without a matching BPF
> call frame and crash when executed. Validation observed a GPF in
> bpf_test_run().
>
> Fix this by requiring every target copied from the actual PTR_TO_INSN map
> to be present in the CFG jump table built for the current gotox
> instruction.
> Reject the program before pushing verifier states for any unmodeled target.
>
> Fixes: 493d9e0d6083 ("bpf, x86: add support for indirect jumps")
> Signed-off-by: Nuoqi Gui <[email protected]>
> ---
> kernel/bpf/verifier.c | 26 ++++++++++++++++++++++++++
> 1 file changed, 26 insertions(+)
>
> diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> index ed7ba0e6a9ce..25fa90e731e3 100644
> --- a/kernel/bpf/verifier.c
> +++ b/kernel/bpf/verifier.c
> @@ -17124,6 +17124,23 @@ static int indirect_jump_min_max_index(struct
> bpf_verifier_env *env,
> return 0;
> }
>
> +static bool is_cfg_indirect_jump_target(struct bpf_verifier_env *env,
> + u32 target)
> +{
> + struct bpf_iarray *jt = env->insn_aux_data[env->insn_idx].jt;
> + int i;
> +
> + if (!jt)
> + return false;
> +
> + for (i = 0; i < jt->cnt; i++) {
> + if (jt->items[i] == target)
> + return true;
> + }
> +
> + return false;
> +}
> +
> /* gotox *dst_reg */
> static int check_indirect_jump(struct bpf_verifier_env *env, struct bpf_insn
> *insn)
> {
> @@ -17171,6 +17188,15 @@ static int check_indirect_jump(struct
> bpf_verifier_env *env, struct bpf_insn *in
> return -EINVAL;
> }
>
> + for (i = 0; i < n; i++) {
> + if (!is_cfg_indirect_jump_target(env,
> env->gotox_tmp_buf->items[i])) {
> + verbose(env,
> + "gotox target %u from map id=%d is not in the
> CFG jump table\n",
> + env->gotox_tmp_buf->items[i], map->id);
> + return -EINVAL;
> + }
> + }
Thanks for reporting the bug.
As for the fix, would it make more sense to either record maps in
check_cfg or to re-check subfunc boundaries here?
> for (i = 0; i < n - 1; i++) {
> mark_indirect_target(env, env->gotox_tmp_buf->items[i]);
> other_branch = push_stack(env, env->gotox_tmp_buf->items[i],
>
> --
> 2.34.1
>
