Re: keepalive per socket settings patch

Christos Zoulas wrote:
> On Jun 19, 10:01pm, darrenrnetbsd.org (Darren Reed)
wrote:
> -- Subject: Re: keepalive per socket settings patch
>
> | Christos Zoulas wrote:
> | > Hi,
> | >
> | > This is pretty straight forward... The question
is do we let setsockopt
> | > specify arbitrary values, or do we cap them to
the global settings like
> | > I do now? Comments? If there are no
disagreements I will commit this after
> | > I write the documentation. If there are, speak
now and I won't bother
> | > with the docs.
> | >   
> | 
> | The most common criticism I've discussed with people
about the
> | keep alive packets is the problem that you're
addressing: different
> | applications want different keep alive values.
> | 
> | I don't think there should be an upper bound, as this
reduces the
> | amount of extra traffic/work generated in a given
period.
> | 
> | With IPFilter, the largest TCP idle timeout default
is 5 days,
> | aimed at the ssh/telnet/etc window that you look at
"every now
> | and then".  I don't know if that is a good
suggestion for where
> | to start with an upper limit or not.
> | 
> | The bigger problem is the lower bound....
> | Is it too much to allow an application to specify it
as being 500ms?
> | And what if apache does this for all n000
connections?
> | Should that just be a configuration error or ...?
>
> All good questions. The idea for the upper limit being
the sysctl
> value seemed reasonable at the time. After I
implemented it, I
> looked at linux and it does almost the same (for some
of them it
> defines max = tcp_keepfoo / PR_SLOWHZ). Can large
values be used
> as a DOS attack? What impact do small values have on
the system?
> The other thing that we can do, and I have not done
yet, is to
> define 0 as infinity (i.e. no timeout). Linux disallows
0; we accept
> it and timeout immediately.
>   

Can we make it behave differently if a NULL pointer is
passed vs a value 
of 0
and mimic the behaviour of select?
However, I don't know if getsockopt() could return the
difference, could it?

Given that you can simply turn off keep alives for a
connection, I can't 
see how
large values could somehow introduce a new DoS vector that
doesn't already
exist today.

Experimenting with small values and seeing what impact it
has on both the
network and CPU when there are 1000s of connections could be
of some
interest.

Darren