Re: [PATCH net-next] r8152: reduce memory copy for rx

[Eric Dumazet <eric.dumazet@gmail.com>]

On Wed, 2014-12-03 at 07:05 +0000, Hayes Wang wrote:
> Eric Dumazet [mailto:eric.dumazet@gmail.com] 
> > Sent: Wednesday, December 03, 2014 2:08 PM
> [...]
> > Better performance for what workload exactly ?
> 
> I test it by using Chariot with 4 Tx and 4 Rx.
> It has about 4% improvement.
> 

Have you tried using more concurrent RX flows, in a possibly lossy
environment (so that TCP is forced to queue packets in out of order
queue) ?

> > cloning in rx path has many drawbacks, with skb->truesize 
> > being usually wrong.
> 
> Excuse me. I find the skb_clone() would copy the
> truesize from original skb. Do you mean the value
> may not be suitable for the cloned skb?

With cloning, (skb->len / skb->truesize) will eventually be very very
small, forcing TCP stack to perform collapses (copies !!!) under
pressure.

> 
> Could other method do the same thing? Or, do you
> think keeping the original one is better?


skb cloning prevents GRO and TCP coalescing from working.

netfilter might also be forced to copy whole frame in case a mangle is
needed (eg with NAT ...)

I would rather try to implement GRO, and/or using fragments instead of
pure linear skbs.

(skb->head would be around 128 or 256 bytes, and you attach to skb the
frame as a page fragment)