Netlink-based D-Bus

[Rodrigo Moya]

On Tue, 2012-03-20 at 18:13 +0200, Rémi Denis-Courmont wrote:
> Le mardi 20 mars 2012 17:49:56 Rodrigo Moya, vous avez écrit :
> > after some feedback and investigation, here's a plan for a new transport
> > for D-Bus, based on Netlink:
> > 
> > http://blogs.gnome.org/rodrigo/2012/03/20/netlink-based-d-bus/
> > 
> > any feedback would be appreciated, as we are starting the implementation
> > as we speak and would like to have all the details covered if possible.
> 
> If the goal is to support direct client-to-client I/O you're going to have to 
> tune the D-Bus specification to:
>
all traffic goes through the kernel dbus module, so no direct
client-to-client I/O

> - tolerate packet loss
>
the kernel module will be sending the messages to each peer
individually, so we can still detect when the receiver's message is
full, and try again after a little bit. So we think we can avoid this.

> - tolerate malformatted messages
> (i.e. drop them individually instead of giving up),
>
dbus-daemon does the validation of the messages, so it's true it won't
be doing that anymore, so there are 2 options, which are to do the
validation either in the kernel module or in client libs

> - drop support for file descriptors passing (oops!).
> 
we are still thinking how to solve this, any ideas are welcome. We can
have the netlink protocol have a way for peers to negotiate a side-band
channel.

> I think.
> 
> 
> As for large messages, I think D-Bus should not be used for that purpose in 
> the first place...
> 
totally agree, but there are users of D-Bus that use it for this, so we
might want to have a good performance on this.