Re: Minimum MTU Mess

[Jarod Wilson <jarod@redhat.com>]

On Mon, Sep 12, 2016 at 04:41:40AM +0200, Andrew Lunn wrote:
> > Actually breaking this up into easily digestable/mergeable chunks is going
> > to be kind of entertaining... Suggestions welcomed on that. First up is
> > obviously the core change, which touches just net/ethernet/eth.c,
> > net/core/dev.c, include/linux/netdevice.h and
> > include/uapi/linux/if_ether.h, and should let existing code continue to
> > Just Work(tm), though devices using ether_setup() that had no MTU range
> > checking (or one or the other missing) will wind up with new bounds.
> 
> Hi Jarod
> 
> Did you find any drivers which support jumbo packets, but don't have
> checks? These drivers, if there are any, need handling first, before
> this core change is made. Otherwise you introduce regressions.

Surprisingly, very few. There was dvb_net using eth_change_mtu and it's
max of 1500 while setting it's initial mtu to 4096, and I swear there was
at least one or two drivers that had no upper bounds checking at all that
I set max_mtu to IP_MAX_MTU (65535), but it's possible I missed something.

At the moment, things are roughly chopped up into:

1) core change
2) deprecate eth_change_mtu and remove in-kernel users
3) set m{in,ax}_mtu in ethernet drivers
4) set m{in,ax}_mtu in wireless drivers
5) set m{in,ax}_mtu in wan drivers
6) set m{in,ax}_mtu in usb ethernet drivers
7) set m{in,ax}_mtu in net infra
8) set m{in,ax}_mtu in virt drivers
9) set m{in,ax}_mtu in misc drivers

The ethernet drivers one is by far the largest, and was thinking I'd start
breaking that up next.

-- 
Jarod Wilson
jarod@redhat.com