Re: [PATCH RFC bpf-next v6 4/6] bpf/helpers: mark the callback of bpf_timer_set_sleepable_cb() as sleepable

[Eduard Zingerman]

On Mon, 2024-04-08 at 10:09 +0200, Benjamin Tissoires wrote:
> Now that we have bpf_timer_set_sleepable_cb() available and working, we
> can tag the attached callback as sleepable, and let the verifier check
> in the correct context the calls and kfuncs.
> 
> Signed-off-by: Benjamin Tissoires <bentiss@xxxxxxxxxx>
> 
> ---

I think this patch is fine with one nit regarding in_sleepable().
Acked-by: Eduard Zingerman <eddyz87@xxxxxxxxx>

> @@ -5279,7 +5281,8 @@ static int map_kptr_match_type(struct bpf_verifier_env *env,
>  
>  static bool in_sleepable(struct bpf_verifier_env *env)
>  {
> -	return env->prog->sleepable;
> +	return env->prog->sleepable ||
> +	       (env->cur_state && env->cur_state->in_sleepable);
>  }

Sorry, I already raised this before.
As far as I understand the 'env->cur_state' check is needed because
this function is used from do_misc_fixups():

		if (is_storage_get_function(insn->imm)) {
			if (!in_sleepable(env) ||
			    env->insn_aux_data[i + delta].storage_get_func_atomic)
				insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_ATOMIC);
			else
				insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_KERNEL);
			insn_buf[1] = *insn;
			cnt = 2;
			...
		}

For a timer callback function 'env->prog->sleepable' would be false.
Which means that inside sleepable callback function GFP_ATOMIC would
be used in cases where GFP_KERNEL would be sufficient.
An alternative would be to check (and set) sleepable flag not for a
full program but for a subprogram.

Whether or not this is something worth addressing I don't know.

[...]