Re: sunrpc

[Andi Kleen <ak@muc.de>]

On Sun, May 16, 1999 at 01:21:50PM +0200, Philip Blundell wrote:
> >Alexey Kuznetsov's TI-RPC port does IPv6 (and also is fully thread safe etc.)
> 
> Thanks, I'll take a look.  I'm more interested in the application interface 
> that Sun chose than the implementation details.
> 
> Sometime last week I got really fired up and made about half the changes 
> necessary for the kernel implementation of RPC to handle IPv6 too.  Now that 
> 2.3 is starting perhaps I'll drag that out again and try to finish it off. :-)

I believe Alexey has some code for that too, perhaps it is worth to ask him first.

When you do that it would be nice if you could take a look at the sunrpc/UDP ARP bug 
too: the arp/ndisc code expects the upper layers to signal "forward process" regularly
via dst_confirm() on the route. When they don't it'll resend ARP requests in regular 
intervals, that happens in 2.2 currently for NFS. 

The fix is to cache the destination entry somewhere in the RPC request structure, and
call dst_confirm in the input path for any successfull reply. Doing a reverse route
lookup is too slow for the fast path, the route needs to be cached (and passed via
sk->dst_cache to the UDP output functions, this only works if no sunrpc sockets
are shared between threads)

So while doing sunrpc socket cleanup for IPv6 it would be nice if you could take 
that problem in mind.


-Andi

-- 
This is like TV. I don't like TV.