On Mon, 02 May 2016 07:34:28 -0700 Eric Dumazet <eric.duma...@gmail.com> wrote:
> On Mon, 2016-05-02 at 09:49 +0200, Jesper Dangaard Brouer wrote: > > > What about using bulk free of SKBs here? > > > > There is a very high probability that we are hitting SLUB slowpath, > > which involves an expensive locked cmpxchg_double per packet. Instead > > we can amortize this cost via kmem_cache_free_bulk(). > > > > Maybe extend kfree_skb_list() to hide the slab/kmem_cache call? > > Sounds tricky, because of skb destructors. skb are complex objects. > > For each skb, need to free the frags, skb->head, and skb. It is not that complicated, inside kfree_skb_list(), we just call skb_release_all(skb) on each SKB first, and then bulk free the SKB's themselves in the end. Example see, _kfree_skb_defer(). The question is where to store the SKB array needed by kmem_cache_free_bulk. The easy option is just to use the stack of kfree_skb_list(), but we have to be careful about the stack size, it might not be so good because skb_release_all() can be deep and via skb_release_data() invoke kfree_skb_list() a second time. -- Best regards, Jesper Dangaard Brouer MSc.CS, Principal Kernel Engineer at Red Hat Author of http://www.iptv-analyzer.org LinkedIn: http://www.linkedin.com/in/brouer
