Re: [PATCH net-next 0/6] net: Call skb_get_hash in qdiscs

[David Miller <davem@davemloft.net>]

From: Eric Dumazet <eric.dumazet@gmail.com>
Date: Wed, 04 Mar 2015 12:03:03 -0800

> Yes, but then your patch is all about reducing flow compares to a single
> u32 comparison in qdiscs (and elsewhere)
> 
> choke for example explicitly wants to make sure we drop a companion
> if incoming packet belongs to the same flow.
> 
> Relying on a 'strong hash' whatever it can be was not considered in
> Choke paper. There is no mention of a stochastic match.
> 
> If we no longer can store the keys in skb->cb[], fine (although I claim
> skb->cb[] size should be irrelevant, see our discussion on this topic
> with Florian)
>  -> Just recompute the keys, using a local variable, from packet
> content. Yes, it will be more expensive, but hey, we get what we want.
> 
> Same for sfq : your skb_get_hash_perturb() doesn't address the point I
> made earlier.
> 
> It is only giving a false sense of security.
> I would rather not spread it.
> (Note that there is no documentation or changelog to explain the
> pro/cons)
> 
> I doubt OVS would condense their flow keys in a single u32...

I'm largely siding with Eric on this.  And the Choke argument is
a strong one.

Therefore I'm deferring this series for now, more thought and work
is definitely needed.