Re: [PATCH net-next 4/4] net: skbuff: fix l4_hash comment

[Antoine Tenart <atenart@kernel.org>]

Quoting Ilya Maximets (2023-05-18 01:00:40)
> On 5/17/23 14:05, Antoine Tenart wrote:
> > 
> > Even l4_hash w/o taking the rnd case into account does not guarantee a
> > stable hash for the lifetime of a flow; what happens if packets from the
> > same flow are received on two NICs using different keys and/or algs?
> 
> Following the same logic we can't really say that it "provides a uniform
> distribution over L4 flows" either.  The fact that L4 fields were used
> to calculate the hash, doesn't mean the hash function is any good.

Well drivers need to either trust the h/w in some ways or not use what
is provided if it's broken; or we can't be sure of anything. It's not
the same as an example where a valid setup can't guarantee a property by
design.

> > Now, I'll let some time to give a chance for others to chime in.
> 
> Sure.

I don't think we'll get more guidance and we failed to come to an
agreement so let's keep this as-is for now; I'll send a v2 w/o this
documentation change.

As for a way forward and the stability need, IMHO the hash needs to be
computed where it is used (with potential cache, not reusing skb->hash)
but if you feel this should be addressed at this level a patch doing so
might at least get others to comment (both ways).

Thanks,
Antoine