Can we rely on ethernet header padding?

Can we rely on ethernet header padding?

[Michal Kubecek]

Hello,

a customer of ours ran into

  http://bugzilla.netfilter.org/show_bug.cgi?id=765

They checked that commit a504b86e prevents the crash but I'm not sure it
is sufficient.

The crash happens when br_nf_pre_routing_finish_bridge() calls
neigh_hh_bridge() which copies not only destination MAC address but also
the padding with it. IIUC this is for performance reasons (so that
aligned 8 bytes are copied rather than 6).

But I wonder whether we can rely on the fact that every skb on an
ethernet-like device has ethernet header padded at least to the 16 bytes
expected by neigh_hh_bridge() and neigh_hh_output() or whether the
bridge code should make sure. I tried to look for such test but couldn't
find any, even if commit a504b86e description mentions reallocating the
skb rather than a crash.

Thanks in advance,
                                                          Michal Kubecek

Re: Can we rely on ethernet header padding?

[Eric Dumazet]

On Tue, 2013-03-19 at 16:05 +0100, Michal Kubecek wrote:
> Hello,
> 
> a customer of ours ran into
> 
>   http://bugzilla.netfilter.org/show_bug.cgi?id=765
> 
> They checked that commit a504b86e prevents the crash but I'm not sure it
> is sufficient.
> 
> The crash happens when br_nf_pre_routing_finish_bridge() calls
> neigh_hh_bridge() which copies not only destination MAC address but also
> the padding with it. IIUC this is for performance reasons (so that
> aligned 8 bytes are copied rather than 6).
> 
> But I wonder whether we can rely on the fact that every skb on an
> ethernet-like device has ethernet header padded at least to the 16 bytes
> expected by neigh_hh_bridge() and neigh_hh_output() or whether the
> bridge code should make sure. I tried to look for such test but couldn't
> find any, even if commit a504b86e description mentions reallocating the
> skb rather than a crash.

Thats a side effect.

Before calling netif_rx() the driver usually calls eth_type_trans()
to pull the ethernet header, so there is the room for 14 bytes.

Normally a driver has NET_SKB_PAD bytes of headroom before the ethernet
header, so the bridge code is safe only if all drivers use this
NET_SKB_PAD padding on receive side. And they really should for
performance reasons.

Better not touch bridge code to catch offending drivers

Re: Can we rely on ethernet header padding?

[Michal Kubecek]

On Tue, Mar 19, 2013 at 08:21:35AM -0700, Eric Dumazet wrote:
> 
> Normally a driver has NET_SKB_PAD bytes of headroom before the ethernet
> header, so the bridge code is safe only if all drivers use this
> NET_SKB_PAD padding on receive side. And they really should for
> performance reasons.
> 
> Better not touch bridge code to catch offending drivers

That makes sense. Thank you for your reply.

                                                   Michal Kubecek

Re: Can we rely on ethernet header padding?

[Stephen Hemminger]

On Tue, 19 Mar 2013 16:48:06 +0100
Michal Kubecek <mkubecek@suse.cz> wrote:

> On Tue, Mar 19, 2013 at 08:21:35AM -0700, Eric Dumazet wrote:
> > 
> > Normally a driver has NET_SKB_PAD bytes of headroom before the ethernet
> > header, so the bridge code is safe only if all drivers use this
> > NET_SKB_PAD padding on receive side. And they really should for
> > performance reasons.
> > 
> > Better not touch bridge code to catch offending drivers
> 
> That makes sense. Thank you for your reply.
> 

My view is that the bridge code must check before assuming headroom.
But because of that, it means a packet copy would be necessary for cases
where packets arrive without enough headroom.

Re: Can we rely on ethernet header padding?

[Michal Kubecek]

On Tue, Mar 19, 2013 at 02:25:53PM -0700, Stephen Hemminger wrote:
> 
> My view is that the bridge code must check before assuming headroom.
> But because of that, it means a packet copy would be necessary for cases
> where packets arrive without enough headroom.

Yes, but for most users this is better than a crash. How about issuing a
WARN_ONCE() in such case so that we avoid writing out of the buffer but
still allow developers (and users who check logs) to catch the problem?

                                                          Michal Kubecek