On Wed, 27 Mar 2019 12:02:13 +0000 Edward Cree <ec...@solarflare.com> wrote:
> On 26/03/2019 14:43, Jesper Dangaard Brouer wrote: > > On Mon, 25 Mar 2019 13:42:39 +0000 > > Ioana Ciornei <ioana.cior...@nxp.com> wrote: > > > >> Take advantage of the software Rx batching by using > >> netif_receive_skb_list instead of napi_gro_receive. > >> > >> Signed-off-by: Ioana Ciornei <ioana.cior...@nxp.com> > >> --- > > Nice to see more people/drivers using: netif_receive_skb_list() > > > > We should likely add a similar napi_gro_receive_list() function. > > I had a patch series that did that; last posting was v3 back in > November: https://marc.info/?l=linux-netdev&m=154221888012410&w=2 > However, Eric raised some issues, also some Mellanox folks privately > reported that using it in their driver regressed performance, and > I've been too busy since to make progress with it. Since you seem > to be much better than me at perf investigations, Jesper, maybe you > could take over the series? I'm hoping Florian Westphal might also have some cycles for this? (We talked about doing this during NetDevConf-0x13, because if we can make more driver use these SKB-lists, then it makes sense to let iptables/nftables build a SKB-list of packets to drop, instead of doing it individually, and then we leverage Felix'es work on bulk free in kfree_skb_list). I'm currently coding up use of netif_receive_skb_list() in CPUMAP redirect. As this makes is easier for e.g. Florian (and others) to play with this API, as we no-longer depend on a device driver having this (although we do depend on XDP_REDIRECT in a driver). And the trick to get this as faster than GRO (that basically recycle the same SKB) is to use the slub/kmem_cache bulk API for SKBs. -- Best regards, Jesper Dangaard Brouer MSc.CS, Principal Kernel Engineer at Red Hat LinkedIn: http://www.linkedin.com/in/brouer
