Re: [PATCH 0/5] Make nicer CONFIG_NET_NS=n case code

[David Miller <davem@davemloft.net>]

From: Eric Dumazet <dada1@cosmosbay.com>
Date: Wed, 31 Oct 2007 23:40:59 +0100

> Eric W. Biederman a écrit :
> > Eric Dumazet <dada1@cosmosbay.com> writes:
> > 
> > 
> >> Definitly wanted here. Thank you.
> >> One more refcounting on each socket creation/deletion was expensive.
> > 
> > Really?  Have you actually measured that?  If the overhead is
> > measurable and expensive we may want to look at per cpu counters or
> > something like that.  So far I don't have any numbers that say any
> > of the network namespace work inherently has any overhead.
> 
> It seems that on some old opterons (two 246 for example),
> "if (atomic_dec_and_test(&net->count))" is rather expensive yes :(

P4 chips are generally very poor at mispredicted branches and
atomics.  So every atomic you remove from the socket paths
gives a noticable improvement on them.

Network device reference counting is such a stupid problem.  There has
to be a way to get rid of it on the packet side.

I think we could get rid of all of the device refcounting from packets
if we:

1) Formalize "SKB roots".  This is every place a packet
   could sit in the transmit path.

2) On device unregister:

	a) wait for RCU quiesce period
	b) stop_machine_run(skb_walk_roots, netdev, NR_CPUS);

   skb_walk_roots is a function that walks all the places in
   #1, rewriting the packet to point to loopback or whatever
   instead of 'netdev' which we are trying to unregister.

This gives us two things.

First, we no longer would need to rectount net devices
for packet references.

Second, we have a debugging framework for all those dreaded SKB leaks
that keep devices from being unloadable.  As we walk the roots
we'll see where all packets referencing a device actually are.