On 09/21/2016 01:09 AM, Thomas Graf wrote:
On 09/20/16 at 03:49pm, Tom Herbert wrote:
On Tue, Sep 20, 2016 at 3:44 PM, Thomas Graf <tg...@suug.ch> wrote:
On 09/20/16 at 03:00pm, Tom Herbert wrote:
+static inline int __xdp_hook_run(struct list_head *list_head,
+                              struct xdp_buff *xdp)
+     struct xdp_hook_ops *elem;
+     int ret = XDP_PASS;
+     list_for_each_entry(elem, list_head, list) {
+             ret = elem->hook(elem->priv, xdp);
+             if (ret != XDP_PASS)
+                     break;
+     }

Walking over a linear list? Really? :-) I thought this was supposed
to be fast, no compromises made.

Can you suggest an alternative?

Single BPF program that encodes whatever logic is required. This is
what BPF is for. If it absolutely has to run two programs in sequence
then it can still do that even though I really don't see much of a
point of doing that in a high performance environment.

Agreed, if there's one thing I would change in cls_bpf, then it's getting
rid of the (uapi unfortunately) list of classifiers and just make it a
single one, because that's all that is needed, and chaining/pipelining
can be done via tail calls for example. This whole list + callback likely
also makes things slower. Why not let more drivers come first to support
the current xdp model we have, and then we can move common parts to the
driver-independent core code?

I'm not even sure yet I understand full purpose of this yet ;-)

Reply via email to