On 10/06/2015 07:50 PM, Alexei Starovoitov wrote:
On 10/6/15 1:39 AM, Daniel Borkmann wrote:
[...] Also classic BPF would then need to test for it, since a socket
filter
doesn't really know whether native eBPF is loaded there or a
classic-to-eBPF
transformed one, and classic never makes use of this. Anyway, it
could be done
by adding a bit flag cb_access:1 to the bpf_prog, set it during eBPF
verification phase, and test it inside sk_filter() if I see it
correctly.
That could also be done in an unlikely() branch, to keep the cost to
the non-eBPF
case near zero.
Yes, agreed. For the time being, the majority of users are coming from the
classic BPF side anyway and the unlikely() could still be changed later on
if it should not be the case anymore. The flag and bpf_func would share the
same cacheline as well.
was also thinking that we can do it only in paths that actually
have multiple protocol layers, since today bpf is mainly used with
tcpdump(raw_socket) and new af_packet fanout both have cb cleared
on RX, because it just came out of alloc_skb and no layers were called,
and on TX we can clear 20 bytes in dev_queue_xmit_nit().
af_unix/netlink also have clean skb. Need to analyze tun and sctp...
but it feels overly fragile to save a branch in sk_filter,
so planning to go with
if(unlikely(prog->cb_access)) memset in sk_filter().
I was also thinking that for dev_queue_xmit_nit(), since we do the skb_clone()
there, to have a clone version (w/o affecting performance of the current one)
that instead of copying cb[] over, it would just do a memset(). But that would
just be limited to AF_PACKET, and doesn't catch all sk_filter() users.
Thanks,
Daniel
--
To unsubscribe from this list: send the line "unsubscribe linux-api" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
