bonding: propagation of offload settings

bonding: propagation of offload settings

[Simon Horman]

Hi,

I am wondering what the desired behaviour for the propagating
of offload settings between master and slaves when the settings
are modified using ethtool.

I have observed the following (using Linus' latest tree, 2.6.36+)

#1 Disabling gro on a slave device propagates to the master
   but not other slaves

   bond1: generic-receive-offload: on
   eth1: generic-receive-offload: on
   eth4: generic-receive-offload: on

   # ethtool -K eth1 gro off

   bond1: generic-receive-offload: off
   eth1: generic-receive-offload: on
   eth4: generic-receive-offload: off

   This seems to occur regardless of if the slave is the
   active slave or not.

#2 No other propagation of settings occurs


It seems to me that from a user point of view it may make more sense to:

a) propagate settings from the master to the slaves and;
b) possibly disallow setting the slaves directly

Re: bonding: propagation of offload settings

[David Miller]

From: Simon Horman <horms@verge.net.au>
Date: Sat, 30 Oct 2010 11:54:35 +0900

> It seems to me that from a user point of view it may make more sense to:
> 
> a) propagate settings from the master to the slaves and;
> b) possibly disallow setting the slaves directly

Yeah, good question.

Propagation from master to slave(s) would have difficult semantics.
If any of the slave changes fail (f.e. unsupported feature or memory
allocation failure) we'd have to unwind all of the slaves which did
accept the change without error.

What if the unwind operation fails, due to lack of resources?  A lot
of state would thus need to be tracked to support this reasonably.

However we pretty much have to respect whatever changes get made
directly to the slaves, since the master must at all times claim
support for only the lowest common denominator, feature wise, amongst
that master's slaves.