Re: [PATCH v5] net: batch skb dequeueing from softnet input_pkt_queue

[Changli Gao <xiaosuo@gmail.com>]

On Thu, Apr 22, 2010 at 7:37 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
>
> Please reorder things better.
>
> Most likely this function is called for one packet.
>
> In your version you take twice the rps_lock()/rps_unlock() path, so
> it'll be slower.
>
> Once to 'transfert' one list to process list
>
> Once to be able to do the 'label out:' post processing.
>

It doesn't make any difference. We have to hold rps_lock to update
input_pkt_queue_len, if we don't use another variable to record the
length of the process queue, or atomic variable.

I think it is better that using another variable to record the length
of the process queue, and updating it before process_backlog()
returns. For one packet, there is only one locking/unlocking. There is
only one issue you concerned: cache miss due to sum the two queues'
length. I'll change softnet_data to:

struct softnet_data {
        struct Qdisc            *output_queue;
        struct list_head        poll_list;
        struct sk_buff          *completion_queue;
        struct sk_buff_head     process_queue;

#ifdef CONFIG_RPS
        struct softnet_data     *rps_ipi_list;

        /* Elements below can be accessed between CPUs for RPS */
        struct call_single_data csd ____cacheline_aligned_in_smp;
        struct softnet_data     *rps_ipi_next;
        unsigned int            cpu;
        unsigned int            input_queue_head;
#endif
        unsigned int            process_queue_len;
        struct sk_buff_head     input_pkt_queue;
        struct napi_struct      backlog;
};

For one packets, we have to update process_queue_len in any way. For
more packets, we only change process_queue_len just before
process_backlog() returns. It means that process_queue_len change is
batched.

-- 
Regards，
Changli Gao(xiaosuo@gmail.com)