Re: [patch net-next-2.6] net: convert bonding to use rx_handler

Fri, Feb 18, 2011 at 05:14:30PM CET, eric.dumazet@gmail.com wrote:

Le vendredi 18 février 2011 à 16:50 +0100, Patrick McHardy a écrit :

quoted

On 18.02.2011 15:58, Jiri Pirko wrote:

quoted

Fri, Feb 18, 2011 at 03:46:45PM CET, kaber@trash.net wrote:

quoted

Am 18.02.2011 15:27, schrieb Eric Dumazet:

quoted

Le vendredi 18 février 2011 à 15:14 +0100, Jiri Pirko a écrit :

quoted

Do not know how to do it better. As for percpu variable, not only
origdev would have to be remembered but also probably skb pointer to
know if it's the first run on the skb or not. Can't really figure out a
better solution. Can you?

I'll try and let you know.

Why not simply do a lookup on skb->iif?

Well I was trying to avoid iterating over list of devices for each
incoming frame.

Well, there are a couple of holes on 64 bit, perhaps you can rearrange
things and eliminate either iif or input_dev without increasing size
since they appear to be redundant.

Jiri

I dont understand why netif_rx() is needed in your patch.

I used netif_rx() because bridge and macvlan does that too. I did not see
a reason to not to do the same.

Can we stack 10 bond devices or so ???

If we avoid this stage and call the real thing (netif_receive_skb()),
then we dont need adding a field in each skb, since it can be carried by
a global variable (per cpu of course)

I'm probably missing something. How do netif_receive_skb() and
netif_rx() differ in this point of view, since both are calling:
"ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);"
?

Still I see a problem with the percpu global variable. We would have to
store skb pointer there as well and in each __netif_receive_skb() call it
would have to be checked if it's different from the current one.
In that case store new skb and orig_Dev.

Leaving aside that global variables are evil in general, I still think
this is not nicer solution then to add skb->input_dev (although I
understand your arguments).

bond_handle_frame() being called from __netif_receive_skb() I believe it
can use netif_receive_skb() instead of netif_rx().

Same remark for vlan_on_bond_hook()