cvs commit: src/sys/netinet tcp_subr.c

[gnn at FreeBSD.org]

At Tue, 12 Sep 2006 02:38:13 +0400,
Yar Tikhiy wrote:
> 
> On Mon, Sep 11, 2006 at 06:28:26PM +0400, Gleb Smirnoff wrote:
> > On Mon, Sep 11, 2006 at 12:36:57AM -0500, Mike Silbersack wrote:
> > M> >It stops creating any state (nor socket neither compressed tcptw) for the
> > M> >TCP connection, where both endpoints were local. This save a lot of
> > M> >resources on servers running HTTP accelerators, or database 
> > M> >servers+clients.
> > M> 
> > M> I think that you should change the description to say "Do not create TCP 
> > M> TIME_WAIT entries for local <-> local connections" (or something similar) 
> > M> to remove the ambiguity, and turn it on by default for 7.  This seems like 
> > M> a good idea, I'm jealous that I didn't think of it. :)
> > 
> > OOh, the sysctl name and its description were discussed with ru@, who
> > is our famous nitpicker. If you want to change name or description -
> > mail to him. /me hides
> 
> (It was me who suggested a different name for the sysctl in question.)
> 
> IMHO it's time to define a general style for new sysctl names.
> The first two points could be a) "foo" vs. "nofoo" for boolean
> knobs, and b) underscore usage.
> 
> Presently there seem to be few "nofoo" knobs in the system, e.g.:
> 
> 	# sysctl -aN | fgrep .no
> 	kern.nodump_coredump		-- should be coredump_nodump
> 	net.inet6.icmp6.nodeinfo	-- false match
> 	debug.nosleepwithlocks
> 	dev.fxp.0.noflow
> 
> I don't think new ones should be introduced; their logic would be
> unnecessarily vague.
> 
> OTOH, there are enough sysctl names containing an underscore to
> encourage using it in complex names.  An exception could be made
> for really short names, such as consisting of not more than 2 words,
> each being not longer than 4 letters.
> 
> What do you think about that, folks?

I am for it, in particular if it is documented.

Later,
George