Re: [PATCH bpf-next v2] bpf: arm64: Fix panic due to missing BTI at indirect jump targets

[Yonghong Song <yonghong.song@linux.dev>]

On 12/25/25 3:59 AM, Xu Kuohai wrote:
> On 12/24/2025 2:32 AM, Yonghong Song wrote:
>
> [...]
>
>>> +
>>> +/*
>>> + * This function collects possible indirect jump targets in a BPF 
>>> program. Since indirect jump
>>> + * targets can only be read from indirect arrays used as jump 
>>> table, it traverses all jump
>>> + * tables used by @prog. For each instruction found in the jump 
>>> tables, it sets the corresponding
>>> + * bit in @bitmap.
>>> + */
>>> +void bpf_prog_collect_indirect_targets(const struct bpf_prog *prog, 
>>> unsigned long *bitmap)
>>> +{
>>> +    struct bpf_insn_array *insn_array;
>>> +    struct bpf_map *map;
>>> +    u32 xlated_off;
>>> +    int i, j;
>>> +
>>> +    for (i = 0; i < prog->aux->used_map_cnt; i++) {
>>> +        map = prog->aux->used_maps[i];
>>> +        if (!is_jump_table(map))
>>> +            continue;
>>> +
>>> +        insn_array = cast_insn_array(map);
>>> +        for (j = 0; j < map->max_entries; j++) {
>>> +            xlated_off = insn_array->values[j].xlated_off;
>>> +            if (xlated_off == INSN_DELETED)
>>> +                continue;
>>> +            if (xlated_off < prog->aux->subprog_start)
>>> +                continue;
>>> +            xlated_off -= prog->aux->subprog_start;
>>> +            if (xlated_off >= prog->len)
>>> +                continue;
>>
>> The above codes are duplicated with bpf_prog_update_insn_ptrs().
>> Maybe we can have a helper for the above?
>>
>
> I tried using function callbacks to factor out the duplicated code,
> but the result felt a bit over-engineered. For these two functions,
> simple duplication seems clearer and simpler.

I am okay with this then.

>
>>> +            __set_bit(xlated_off, bitmap);
>>> +        }
>>> +    }
>>> +}
>>> +
>>> +void bpf_prog_set_insn_array_type(struct bpf_map *map, int type)
>>> +{
>>> +    struct bpf_insn_array *insn_array = cast_insn_array(map);
>>> +
>>> +    insn_array->type = type;
>>> +}
>>> diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
>>> index d6b8a77fbe3b..ee6f4ddfbb79 100644
>>> --- a/kernel/bpf/verifier.c
>>> +++ b/kernel/bpf/verifier.c
>>> @@ -20288,6 +20288,12 @@ static int check_indirect_jump(struct 
>>> bpf_verifier_env *env, struct bpf_insn *in
>>>           return -EINVAL;
>>>       }
>>> +    /*
>>> +     * Explicitly mark this map as a jump table such that it can be
>>> +     * distinguished later from other instruction arrays
>>> +     */
>>> +    bpf_prog_set_insn_array_type(map, BPF_INSN_ARRAY_JUMP_TABLE);
>>
>> I think we do not need this for now. If a new indirect_jump type is 
>> introduced,
>> verifier/jit can be adjusted that time if necessary.
>>
>
> As Anton noted, even though jump tables are currently the only type
> of instruction array, users may still create insn_arrays that are not
> used as jump tables. In such cases, there is no need to emit BTIs.

Okay. Thanks for explanation.

>
>>> +
>>>       for (i = 0; i < n - 1; i++) {
>>>           other_branch = push_stack(env, env->gotox_tmp_buf->items[i],
>>>                         env->insn_idx, env->cur_state->speculative);
>>
>>
>>
>