Stefan Sonnenberg-Carstens wrote:
> Hi everyone,
>
> on our department someone came up with a "strange" theory,
> which I could not decline completely and wanted to get some insight
> from the network pros under you.
>
> We habe a SUN Cluster which has several IP adresses bound,
> for example 192.168.10.20 and 192.168.10.30.
Cluster? Is this OpenSolaris? It doesn't sound like it. I would
recommend directing questions about Solaris 10 and other non-OpenSolaris
releases to Oracle's support group.
> Now there is a TCP server running on 192.168.10.30 port 13000,
> and the client explicitly bound 192.168.10.20 port 38790 tcp to
> communicate with the server.
>
> The "theory" now is, that the packets, instead of getting copied
> only inside the kernel, somehow enter the device driver of
> the corresponding devices, hme0 and hme1.
How did you arrive at this theory? And if they theory is correct, why
exactly is it a problem? Is there some larger problem you're trying to
solve?
One way to check the theory would be to snoop the interfaces. Snoop
attaches to the device driver itself and asks for copies of the packets
passing through. That's why (at least prior to some recent OpenSolaris
enhancements) you won't see looped-back packets with snoop.
> Why should the kernel do that ?
> Shouldn't it be more likely, that the kernel sees both
> endpoints local-only, as if we used 127.0.0.1 and do no routing
> or device driver processing at all ?
Yes, it will do that -- it does not send traffic destined for local
addresses down to the drivers.
In general, the stack delivers packets based on the IREs. You can view
the IREs with "netstat -nrva". If this is some non-OpenSolaris release,
then you're looking for "UHA" and "UHL" entries. If it's the latest
OpenSolaris, then you're looking for "UC" and "UHL".
Either way, these entries are what determine how packets are handled.
If the destination specified is local ("UHL"), then the stack loops it
back internally. If it's non-local ("UHA" or "UC"), then it sends it
out through the driver.
That's all in regular Solaris. I don't know how (or if) the Sun Cluster
product may modify this behavior.
--
James Carlson 42.703N 71.076W <[email protected]>
_______________________________________________
networking-discuss mailing list
[email protected]
