PERFORCE change 161793 for review

[Brooks Davis]

On Mon, May 11, 2009 at 04:05:32AM +0200, Marko Zec wrote:
> On Saturday 09 May 2009 21:15:41 Brooks Davis wrote:
> > On Fri, May 08, 2009 at 09:44:20PM +0000, Marko Zec wrote:
> > > http://perforce.freebsd.org/chv.cgi?CH=161793
> > >
> > > Change 161793 by zec at zec_tpx32 on 2009/05/08 21:43:51
> > >
> > > 	A controversial hack: when allocating if_unit numbers for
> > > 	cloning ifnets, do NOT resort to bitmap mappings, but
> > > 	iterate through all existing ifnets in a particular vnet
> > > 	as long as a free unit number is found.
> > >
> > > 	The problem here is that we are switchng from a O(n) to
> > > 	a O(n**2) search method with more work being done in each
> > > 	iteration.  So this should be revisited RSN.
> >
> > IMO the best solution for now is to retain the global namespace and
> > thus not worry about virtualizing this any more than we do for em0.
> 
> It's not that simple - don't forget that we need to have lo0 in each vnet due 
> to expectations from various userland applications.  Moreover, users of 
> cloning ifnets would expect to get (as an example) vlan0, vlan1, vlan2 when 
> issuing ifconfig vlan create, not vlan539, vlan1, vlan123 etc.  The same goes 
> for tunneling interfaces, ng_eiface, ng_iface etc.  I think the general 
> principle behind vimage / vnets was not to allow for any information leakages 
> between the isolated contexts, and retaining global namespace for cloning 
> ifnets would significantly violate that principle.

lo0 will need to be an exception, but I'd be inclined to only violate
name uniqueness in that case.

I agree that there is a longterm argument that vimage instances should
look like stand alone machines in the areas that are virtualized, but at
this stage of the 8.0 I think we're much better off not attempting to
find all the possible edge cases.

> > This must not make it in to current since if it did, users with tunnel
> > brokers would be very unhappy.
> 
> Precisely for the sake of tunnel brokers, I think each vnet should have its 
> own namespace for cloning devices - and this has nothing to do with "em0" or 
> any other physical ifnet which does not use the if_clone infrastructure.  I 
> agree that the O(n**2) code for searching for free ifunits should not get 
> commited into svn, but Julian and I already discussed how we could virtualize 
> the existing O(n) bitmapped algorithm in if_clone.c, and will try to submit 

Tunnel brokers mostly likely won't care AT ALL what name the interfaces
have.  They will case that performance will be terrible due to the n^2
algorithm.

-- Brooks

> 
> Marko
> 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 187 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/p4-projects/attachments/20090512/daf292d1/attachment.pgp