Re: [PATCH v2 0/2] Tracepoint for tcp retransmission

[Eric Dumazet <eric.dumazet@gmail.com>]

Le lundi 06 février 2012 à 11:18 -0500, Steven Rostedt a écrit :
> On Mon, 2012-02-06 at 10:53 -0500, Neil Horman wrote:
> 
> > Yes, that was his point.  Mine was that, given Satoru's needs, a netfilter hook
> > provides 90% of what they need (as its been described here).  If you're
> > unfamiliar with them, netfilter hooks are those standard points in the network
> > input/output/forwarding paths at which we can watch and maniupulate network
> > traffic.  They're what iptables/ip6tables/ebtables/etc use to do everything they
> > do.  I'm suggesting that they use those existing hooks to monitor outgoing
> > traffic for whatever information they want (in this case retransmitted unacked
> > tcp sequence numbers).  Or whatever else their interested in.
> > 
> > As for the other 10% (recording failed retransmits, which don't generate
> > traffic), we have existing tracepoints that can be used.
> 
> I haven't looked at the details here, but I'm wondering if netfilter
> could benefit from static_branch() calls (aka jump-labels). I'm sure
> selinux could too. Basically, when netfilter is disabled, the fast path
> would just include a nop, with no if() branch testing at all. When you
> enable netfilters, it would then add a jmp to the code that actually
> does the tests.

Hmm... I thought I already did that.

Or maybe I intended so and forgot.

Ah no, its there in commit a2d7ec58ac09
(netfilter: use jump_label for nf_hooks)