Re: [PATCH] bpf: optimize constant blinding

[Edward Cree <ecree@solarflare.com>]

On 17/06/2019 21:40, Jiong Wang wrote:
> Now if we don't split patch when patch an insn inside patch, instead, if we
> replace the patched insn using what you suggested, then the logic looks to
> me becomes even more complex, something like
>
>    for (idx = 0; idx < insn_cnt; idx++) {
>      if (insns[idx] is not BPF_LIST_INSN) {
>        do_insn(...)
>      }
>      else if (insns[idx] is BPF_LIST_INSN) {
>        list = pool_base + insn.imm;
>        while (list) {
>          insn = list_head->insn;
>          if (insn is BF_LIST_INSN) {
>            sub_list = ...
>            while ()
>              do_insn()
>            continue;
>          }
>          do_insn(...)
>          list = pool_base + list->next;
>        }
>      }
>    }
Why can't do_insn() just go like:
    if (insn is BPF_LIST_INSN)
        for (idx = 0; idx < LIST_COUNT(insn); idx++)
            do_insn(pool_base + LIST_START(insn) + idx);
    else
        rest of processing
?

Alternatively, iterate with something more sophisticated than 'idx++'
 (standard recursion-to-loop transformation).
You shouldn't ever need a for() tower statically in the code...

> So, I am thinking what Alexei and Andrii suggested make sense, just use
> single data structure (singly linked list) to represent everything, so the
> insn traversal etc could be simple
But then you have to also store orig_insn_idx with each insn, so you can
 calculate the new jump offsets when you linearise.  Having an array of
 patched_orig_insns gives you that for free.