On Tue, 2005-31-05 at 13:42 +0200, Thomas Graf wrote:
> > All "stuffs" (your bases is mine) accessible via in_devxx abstraction
> > should be devconfig configurable. I do concur that some of it may not
> > make sense - but that should be the exception rules...
> > Ok, lets say these tables are one such exception, but note:
> > In the future there could be multiple ARP tables for example (very
> > likely given all the virtualization approaches happening). In other
> > words different devices may point to different tables...
>
> Let's assume for a moment that there are no device specific parameters,
> just the default parameter set, e.g everything in the "arp_cache" block.
> According do your idea, everything would be part of a faked in_dev entry
> with ifindex==0 (or alike). We'd have the exact same issues as with
> RTM_NEIGHTBL, the requirement of unique table ids stays the same. There
> is one big difference though, what if we ever have neighbour tables which
> do not care about inet devices at all? Maybe some kind of new virtual
> devices for in-kernel communication? ;-> Sounds scary and not realistic,
An extra virtual device just to store the defaults will be overkill.
OTOH, a global allocated structure for defaults is ok. I think thats the
way it is done today.
> but my point is basically that neighbour tables itself have a global scope,
> the assignment to the device and device specific parameters is on the
> level of the neighbours itself. The connection to in_dev/... is one
> level higher with a scope of the relevant neighbour table. A neighbour
> table implementation is not required to support device specific
> parameters, the code neighbour code will always assign the default
> parameter set and leaves it up to the constructor of the implementation
> level to reassign a device specific parameter set. Which means that what
> the procfs and now also neightbl code does it kind of a hack anyway
> since the core neighbour code is not aware of device specific stuff
> anyway, all it does is provide an easy method to manage them in a
> efficient matter.
The neighbor code ought not be aware of devices; but whoever is
configuring must be. By configuring i mean "stiching together" the
different blocks (as in some netlink based program in user space). the
idevxxx is a "stiching" construct - and therefore it may be aware of
many other things that a standalone block doesnt. One could argue that
the ifindex in a netdevice is also under the same class. The ARP, NDISC,
netdevice, filters etc are the blocks being sown together.
Most of the times these blocks are built together in a topology that a
packet flows through.
> I would like to stay consistent to this architecture
> if possible to be able to follow every new additions.
>
> I hope my point is now clear, I wasn't so sure until now ;-> However,
> I think you have a good point, parameter sets have a strong relation
> to inet devices at the moment. If you still think that we should move
> _everything_ into a devconfig layer then I'll do it but I still support
> my initial prospoal.
>
I agree that we should leave the neighbor table-specific knobs out of
devconfig.
It will help, though, for devconfig to have a single reference to the
"instance" of the table(perhaps a name or a even a 32 bit pointer
address) that a device uses (thefore maintaining the abstraction); but
since there is only one table for v4 and v6 anyways i suppose it is
implicitly assumed to be that one table.
cheers,
jamal
|