Re: [PATCH bpf-next 01/11] bpf: Move insn if/else into do_check_insn()

[Luis Gerhorst]

Eduard Zingerman <eddyz87@xxxxxxxxx> writes:
> On Thu, 2025-03-13 at 18:21 +0100, Luis Gerhorst wrote:
>> +		err = do_check_insn(env, insn, pop_log, &do_print_state, regs, state,
>> +				    &prev_insn_idx);
>
> - `regs` remains declared in do_check(), while nothing prevents
>   pushing its declaration to do_check_insn().
> - `state` is `env->cur_state`, so I'd avoid passing it as a parameter
>   (just to reduce count);
> - `prev_insn_idx` is unused by `do_check_insn`;
> - `pop_log` is not used by `do_check_insn`;

Changed for v2, thank you very much.

> - given that `insn` is presumed to correspond to `env->insn_idx` in
>   many places down the stack not sure about this parameter.

I don't have a strong opinion on this either. Unless someone objects I
will keep it as it matches the other check_*() functions like this.

>> +		if (err < 0) {
>> +			return err;
>> +		} else if (err == INSN_IDX_MODIFIED) {
>
> Also, I'd get rid of `INSN_IDX_MODIFIED` and move `env->insn_idx++`
> into `do_check_insn()`. This would save a few mental cycles when
> looking at the code with full patch-set applied:
>
> 		} else if (err == INSN_IDX_MODIFIED) {
> 			continue;
> 		} else if (err == PROCESS_BPF_EXIT) {
> 			goto process_bpf_exit;
> 		}
> 		WARN_ON_ONCE(err);
>
> 		if (state->speculative && cur_aux(env)->nospec_result) {
> 			... bunch of actions ...
> 		}
>
> 		env->insn_idx++;
>
> One needs to stop for a moment and think why "bunch of actions" is
> performed for regular index increment, but not for INSN_IDX_MODIFIED.

That certainly makes it more readable. I changed it for v2.

If we have an instruction that does not simply do `insn_idx++` but
jumps, the `nospec_result` check should never trigger. Otherwise, the
patched nospec might be skipped. Currently, this is satisfied because
`nospec_result` is only used for store-instructions. I will add a
comment and WARN_ON_ONCE to document that for v2.