On Mon, Dec 04, 2023 at 05:18:31PM -0800, Alexei Starovoitov wrote:
> How about
> +int get_cfi_offset(void)
> +{
> + switch (cfi_mode) {
> + case CFI_FINEIBT:
> + return 16;
> + case CFI_KCFI:
> +#ifdef CONFIG_CALL_PADDING
> + return 16;
> +#else
> + return 5;
> +#endif
> + default:
> + return 0;
> + }
> +}
Yeah, that works. I'll go make it happen.
> Separately we need to deal with bpf_for_each_array_elem()
> which doesn't look easy.
> And fix tcp_set_ca_state() as well (which is even harder).
>
> Just to see where places like these are I did:
> +__nocfi
> BPF_CALL_4(bpf_loop, u32, nr_loops, void *, callback_fn, void *, callback_ctx,
> +__nocfi
> static long bpf_for_each_hash_elem(struct bpf_map *map,
> bpf_callback_t callback_fn,
> +__nocfi
> static enum hrtimer_restart bpf_timer_cb(struct hrtimer *hrtimer)
> +__nocfi
> static int __bpf_rbtree_add(struct bpf_rb_root *root,
> +__nocfi
> BPF_CALL_4(bpf_user_ringbuf_drain, struct bpf_map *, map,
> +__nocfi
> void tcp_set_ca_state(struct sock *sk, const u8 ca_state)
> +__nocfi
> void tcp_init_congestion_control(struct sock *sk)
> +__nocfi
> void tcp_enter_loss(struct sock *sk)
> +__nocfi
> static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
> +__nocfi
> static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
>
> and more... Which is clearly not a direction to go.
>
> Instead of annotating callers is there a way to say that
> all bpf_callback_t calls are nocfi?
Well, ideally they would all actually use CFI, I'll go figure out how
all this works and think about it. Thanks!
> I feel the patches scratched the iceberg.
Yeah, clearly :/ I'll go stare at it all.
