Re: [PATCH net-next] bpf/verifier: track liveness for pruning

[Daniel Borkmann <daniel@iogearbox.net>]

On 08/14/2017 07:55 PM, Edward Cree wrote:
> State of a register doesn't matter if it wasn't read in reaching an exit;
>   a write screens off all reads downstream of it from all explored_states
>   upstream of it.
> This allows us to prune many more branches; here are some processed insn
>   counts for some Cilium programs:
> Program                  before  after
> bpf_lb_opt_-DLB_L3.o       6515   3361
> bpf_lb_opt_-DLB_L4.o       8976   5176
> bpf_lb_opt_-DUNKNOWN.o     2960   1137
> bpf_lxc_opt_-DDROP_ALL.o  95412  48537
> bpf_lxc_opt_-DUNKNOWN.o  141706  79048
> bpf_netdev.o              24251  17995
> bpf_overlay.o             10999   9385
>
> The runtime is also improved; here are 'time' results in ms:
> Program                  before  after
> bpf_lb_opt_-DLB_L3.o         24      6
> bpf_lb_opt_-DLB_L4.o         26     11
> bpf_lb_opt_-DUNKNOWN.o       11      2
> bpf_lxc_opt_-DDROP_ALL.o   1288    152
> bpf_lxc_opt_-DUNKNOWN.o    1768    257
> bpf_netdev.o                 62     31
> bpf_overlay.o                15     13
>
> Signed-off-by: Edward Cree <ecree@solarflare.com>

Awesome work!

[...]
>   	if (arg_type == ARG_ANYTHING) {
>   		if (is_pointer_value(env, regno)) {
> @@ -1639,10 +1675,13 @@ static int check_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
>   	}
>
>   	/* reset caller saved regs */
> -	for (i = 0; i < CALLER_SAVED_REGS; i++)
> +	for (i = 0; i < CALLER_SAVED_REGS; i++) {
>   		mark_reg_not_init(regs, caller_saved[i]);
> +		check_reg_arg(env, i, DST_OP_NO_MARK);

Don't we need the same in check_ld_abs() since we treat it similar
to a function call?

> +	}
>
>   	/* update return register */
> +	check_reg_arg(env, BPF_REG_0, DST_OP_NO_MARK);
[...]