Re: Get rxhash fixes and RFS support in tun

[Stephen Hemminger <stephen@networkplumber.org>]

On Wed, 20 Nov 2013 17:09:01 -0800
Tom Herbert <therbert@google.com> wrote:

> On Wed, Nov 20, 2013 at 4:02 PM, David Miller <davem@davemloft.net> wrote:
> > From: Tom Herbert <therbert@google.com>
> > Date: Wed, 20 Nov 2013 12:25:48 -0800 (PST)
> >  
> >> This patch series fixes some subtle bugs in tun use of skb->rxhash, all
> >> rxhash hash not be cleared appropraitely, and adds support for tun flows
> >> to work with RFS.
> >>
> >> Testing, in particular with tun, hasn't been completed yet.  
> >
> > I think this needs to be reworked slightly.
> >
> > We really only have two boolean states:
> >
> > 1) Is the rxhash value in this SKB valid?
> >
> > 2) Is it a full L4 tuple hash?
> >
> > You are adding a "this is a SW computed hash" boolean state but I do
> > not think you should distinguish sw vs. hw especially.  If the
> > hardware computed the rxhash on a tunneled packet in the
> > pre-decapsulated state, we very much want to recompute it, in
> > software, upon tunnel decapsulation in ip_tunnel_core.c
> >  
> In either case it would be recomputed in SW if L4 hash was not set
> (i.e. no flow_dissector done finding L4).  If L4 hash is set, that
> should refer to the hash of the inner 4-tuple, so I don't think you'd
> need to recompute it.  I suppose there could be a case like encap in
> UDP where there are potentially two 4-tuples to deal with, but then
> the rxhash could just be cleared in decap to force recomputing hash
> over the inner packet-- even better I would hope that the trick of
> using outer source port to hold the hash of the inner packet (like in
> nvgre) is used so that the hash on the outer header is good for L4.
> 
> The primary reason for sw_rhash is to know whether it is a comparable
> value to match against that of a flow whose hash was computed in SW
> (tun case).

You can't guarantee that two hardware hashes from different devices
are the same.