PERFORCE change 167260 for review

[Marko Zec]

On Wednesday 12 August 2009 23:58:46 Julian Elischer wrote:
> Marko Zec wrote:
...
> > @@ -710,22 +715,36 @@
> >  	.pr_input =		div_input,
> >  	.pr_ctlinput =		div_ctlinput,
> >  	.pr_ctloutput =		ip_ctloutput,
> > -	.pr_init =		NULL,
> > +	.pr_init =		div_init,
> >  	.pr_usrreqs =		&div_usrreqs
>
> If you are going to make pr_init() called for every vnet then
> pr_destroy should be as well. But in fact that is not really safe.
> (either of them)
>
> The trouble is that we can not guarantee that other protocols can
> handle being called multiple times in their init and destroy methods.
> Especially 3rd party protocols.
>
> We need to ensure only protocols that have been converted to run
> with multiple vnets are ever called with multiple vnets.
>
> for this reason the only safe way to do this is via the VNET_SYSINIT
> and VNET_SYSUNINIT calls.

That would mean you would have to convert most if not all of the existing 
things that hang off of protosw-s in netinet, netinet6 etc. to use 
VNET_SYSINT / VNET_SYSUNIT instead of protosw->pr_init().  So the short 
answer is no.

I cannot recall that we ever discussed or planned to be able to mix 
virtualized with non-virtualized protocols in the same kernel.  That would be 
a horrible mess, and I cannot even imagine having say a multi-instance INET 
with a single-instance INET6 kernel, shared among all the vnets.  To start 
with, how would you decide that you're not allowed to process an IPv6 packet 
received on the wire in a non-default vnet in such an environment?  Do we 
have the infrastructure in place necessary for preventing doing say a 
ifconfig lo0 ::1 in a non-default vnet in such an hypotetical setup?  The 
answer is no.

VNET_SYSINIT is nice, but proper special-casing changes required to support 
single-instance protocols to work only with vnet0 and not with the other 
protocols are simply not there, and I hope will never be, because I fear they 
would be highly intrusive, difficult to verify and maintain, and probably 
also have an impact on performance.

A proper solution for the issue you are raising could be something that would 
prevent modules assuming our stack is compiled as single-instance to be 
kldloaded if the kernel was actually built with multi-instance stack support.  
I think Robert (cc-ed) had some ideas on how to accomplish this by having 
such modules depend on a magic global variable (say __no_vnet_support) to be 
available.

All the current "base" protocols are already using pr_init() in multi-instance 
mode in options VIMAGE case.  So I see no reason for ip_divert not being 
allowed to leverage on the same mechanism.

Re. pr_destroy(), you're right, patch already submitted to p4...

Marko