Message-ID: <CA+mtBx_J=fm3+k6_NzMH9KCdrYpZPwT2XBQe4r424eNDCKzazg@mail.gmail.com>
Date:	Thu, 3 May 2012 20:22:22 -0700
From:	Tom Herbert <therbert@...gle.com>
To:	Deng-Cheng Zhu <dczhu@...s.com>
Cc:	davem@...emloft.net, netdev@...r.kernel.org, eric.dumazet@...il.com
Subject: Re: [PATCH v2] RPS: Sparse connection optimizations - v2

> +struct cpu_flow {
> +       struct net_device *dev;
> +       u32 rxhash;
> +       unsigned long ts;
> +};

This seems like overkill, we already have the rps_flow_table and this
used in accelerated RFS so the device can also take advantage of
steering.  Maybe somehow program that table for your sparse flows?

Tom

> +#endif
> +
>  /*
>  * This structure holds an RPS map which can be of variable length.  The
>  * map is an array of CPUs.
> diff --git a/net/Kconfig b/net/Kconfig
> index e07272d..d5aa682 100644
> --- a/net/Kconfig
> +++ b/net/Kconfig
> @@ -222,6 +222,28 @@ config RPS
>        depends on SMP && SYSFS && USE_GENERIC_SMP_HELPERS
>        default y
>
> +config RPS_SPARSE_FLOW_OPTIMIZATION
> +       bool "RPS optimizations for sparse flows"
> +       depends on RPS
> +       default n
> +       ---help---
> +         This feature will try to map some network flows to consecutive
> +         CPUs in the RPS map. It will bring in some per packet overhead
> +         but should be able to do good to network throughput in the case
> +         of low number of connections while not much affecting other
> +         cases. (e.g. relatively consistent and high bandwidth in single
> +         connection tests).
> +
> +config NR_RPS_MAP_LOOPS
> +       int "Number of loops walking RPS map before hash indexing (1-5)"
> +       range 1 5
> +       depends on RPS_SPARSE_FLOW_OPTIMIZATION
> +       default "4"
> +       ---help---
> +         It defines how many loops to go through the RPS map while
> +         determing target CPU to process the incoming packet. After that,
> +         the decision will fall back on hash indexing the RPS map.
> +
>  config RFS_ACCEL
>        boolean
>        depends on RPS && GENERIC_HARDIRQS
> diff --git a/net/core/dev.c b/net/core/dev.c
> index c25d453..92e292b 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -2698,6 +2698,61 @@ set_rps_cpu(struct net_device *dev, struct sk_buff *skb,
>        return rflow;
>  }
>
> +#ifdef CONFIG_RPS_SPARSE_FLOW_OPTIMIZATION
> +static DEFINE_PER_CPU(struct cpu_flow [CONFIG_NR_RPS_MAP_LOOPS], cpu_flows);
> +static unsigned long hash_active;
> +
> +#define FLOW_INACTIVE(now, base) (time_after((now), (base) + HZ) || \
> +                        unlikely(time_before((now), (base))))
> +
> +static u16 find_cpu(const struct rps_map *map, const struct sk_buff *skb)
> +{
> +       struct cpu_flow *flow;
> +       u16 cpu;
> +       int i, l, do_alloc = 0;
> +       unsigned long now = jiffies;
> +
> +retry:
> +       for (l = 0; l < CONFIG_NR_RPS_MAP_LOOPS; l++) {
> +               for (i = map->len - 1; i >= 0; i--) {
> +                       cpu = map->cpus[i];
> +                       flow = &per_cpu(cpu_flows, cpu)[l];
> +
> +                       if (do_alloc) {
> +                               if (flow->dev == NULL ||
> +                                   FLOW_INACTIVE(now, flow->ts)) {
> +                                       flow->dev = skb->dev;
> +                                       flow->rxhash = skb->rxhash;
> +                                       flow->ts = now;
> +                                       return cpu;
> +                               }
> +                       } else {
> +                               /*
> +                                * Unlike hash indexing, this avoids packet
> +                                * processing imbalance across CPUs.
> +                                */
> +                               if (flow->rxhash == skb->rxhash &&
> +                                   flow->dev == skb->dev &&
> +                                   !FLOW_INACTIVE(now, flow->ts)) {
> +                                       flow->ts = now;
> +                                       return cpu;
> +                               }
> +                       }
> +               }
> +       }
> +
> +       if (FLOW_INACTIVE(now, hash_active) && do_alloc == 0) {
> +               do_alloc = 1;
> +               goto retry;
> +       }
> +
> +       /* For all other flows */
> +       hash_active = now;
> +
> +       return map->cpus[((u64) skb->rxhash * map->len) >> 32];
> +}
> +#endif
> +
>  /*
>  * get_rps_cpu is called from netif_receive_skb and returns the target
>  * CPU from the RPS map of the receiving queue for a given skb.
> @@ -2780,7 +2835,11 @@ static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb,
>        }
>
>        if (map) {
> +#ifdef CONFIG_RPS_SPARSE_FLOW_OPTIMIZATION
> +               tcpu = find_cpu(map, skb);
> +#else
>                tcpu = map->cpus[((u64) skb->rxhash * map->len) >> 32];
> +#endif
>
>                if (cpu_online(tcpu)) {
>                        cpu = tcpu;
> --
> 1.7.1
>
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists