Re: [PATCH 2/2 net-next] net: move qdisc ingress filtering code where it belongs

[Alexei Starovoitov <ast@plumgrid.com>]

On 5/11/15 4:30 PM, Eric Dumazet wrote:
>
> For example , commit 7866a621043fbaca3d7389e9b9f69dd1a2e5a855
> helped a given workload, but probably made things slower for most common
> cases.

yes, indeed. reverting it improves netif_receive + drop in ip_rcv
from 41.1 to 42.6. I've been trying to come up with a simple way
to roll global ptype_all into skb->dev->ptype_all, but registering
device notifier seems overkill to remove one loop from netif_receive.
Also tried to partially remove pt_prev logic from the first half
of netif_receive and keep it only for deliver_ptype_list_skb() part,
but that didn't help.
Then tried this:
-static int __netif_receive_skb(struct sk_buff *skb)
+static inline int __netif_receive_skb(struct sk_buff *skb)
...
-static int netif_receive_skb_internal(struct sk_buff *skb)
+static inline int netif_receive_skb_internal(struct sk_buff *skb)

it helped to go from 41.1 to 43.1, but size increase not negligible:
    text	   data	    bss	    dec	    hex	filename
   55990	   1667	   2856	  60513	   ec61	dev.o.base
   56403	   1907	   2856	  61166	   eeee	dev.o.inline

inlining only one of them (either __netif_receive_skb or
netif_receive_skb_internal) gives minimal gain.
Still exploring...