Re: [PATCH net-next] igb: use build_skb()

[Jeff Kirsher <jeffrey.t.kirsher@intel.com>]

[-- Attachment #1: Type: text/plain, Size: 1383 bytes --]

On Thu, 2012-08-02 at 17:51 +0200, Eric Dumazet wrote:
> From: Eric Dumazet <edumazet@google.com>
> 
> By using netdev_alloc_frag() & build_skb() instead of legacy
> netdev_alloc_skb_ip_align() calls, we reduce number of cache misses in
> RX path and size of working set.
> 
> For a given rx workload, number of 'inuse' sk_buff can be reduced to a
> very minimum, especially when packets are dropped by our stack.
> 
> (Before this patch, default sk_buff allocation was 2048 sk_buffs in rx
> ring buffer)
> 
> They are initialized right before being delivered to stack, so can
> stay
> hot in cpu caches.
> 
> Ethernet header prefetching is more effective (old prefetch of
> skb->data
> paid a stall to access skb->data pointer)
> 
> I have 15% performance increase in a RX stress test, removing SLUB
> slow
> path in the profiles.
> 
> Signed-off-by: Eric Dumazet <edumazet@google.com>
> Cc: Alexander Duyck <alexander.h.duyck@intel.com>
> ---
>  drivers/net/ethernet/intel/igb/igb.h         |    8 ++
>  drivers/net/ethernet/intel/igb/igb_ethtool.c |   14 ++--
>  drivers/net/ethernet/intel/igb/igb_main.c    |   56 ++++++++++-------
>  3 files changed, 49 insertions(+), 29 deletions(-) 

Thanks Eric, I have added this as well to my queue.

NOTE- you had some trailing whitespace errors in igb.h which I cleaned
up before adding to my queue.

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 836 bytes --]