Message ID | 1321535606.2751.35.camel@edumazet-HP-Compaq-6005-Pro-SFF-PC |
---|---|
State | Accepted, archived |
Delegated to: | David Miller |
Headers | show |
Le jeudi 17 novembre 2011 à 14:13 +0100, Eric Dumazet a écrit : > Most machines dont use RPS/RFS, and pay a fair amount of instructions in > netif_receive_skb() / netif_rx() / get_rps_cpu() just to discover > RPS/RFS is not setup. > > Add a jump_label named rps_needed. > > If no device rps_map or global rps_sock_flow_table is setup, > netif_receive_skb() / netif_rx() do a single instruction instead of many > ones, including conditional jumps. > > jmp +0 (if CONFIG_JUMP_LABEL=y) > > Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> > CC: Tom Herbert <therbert@google.com> I wonder if we could use this jump_label infrastructure to shortcut expensive netfilter nf_hook_slow() / nf_iterate on empty tables/chains... -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
From: Eric Dumazet <eric.dumazet@gmail.com> Date: Thu, 17 Nov 2011 22:34:49 +0100 > I wonder if we could use this jump_label infrastructure to shortcut > expensive netfilter nf_hook_slow() / nf_iterate on empty > tables/chains... That would be nice, I can't see why it wouldn't be possible. -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
From: Eric Dumazet <eric.dumazet@gmail.com> Date: Thu, 17 Nov 2011 14:13:26 +0100 > Most machines dont use RPS/RFS, and pay a fair amount of instructions in > netif_receive_skb() / netif_rx() / get_rps_cpu() just to discover > RPS/RFS is not setup. > > Add a jump_label named rps_needed. > > If no device rps_map or global rps_sock_flow_table is setup, > netif_receive_skb() / netif_rx() do a single instruction instead of many > ones, including conditional jumps. > > jmp +0 (if CONFIG_JUMP_LABEL=y) > > Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> > CC: Tom Herbert <therbert@google.com> > --- > V2: add netif_rx() optim as well Applied, thanks. -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Le jeudi 17 novembre 2011 à 17:04 -0500, David Miller a écrit : > From: Eric Dumazet <eric.dumazet@gmail.com> > Date: Thu, 17 Nov 2011 22:34:49 +0100 > > > I wonder if we could use this jump_label infrastructure to shortcut > > expensive netfilter nf_hook_slow() / nf_iterate on empty > > tables/chains... > > That would be nice, I can't see why it wouldn't be possible. Almost done :) Only problem is arch_static_branch() has a "i" (key), so some calls wont be optimized (hooknum in a variable, not a constant) for example in nf_ct_frag6_output() Not a big deal, most calls can be optimized away. -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h index 4d5698a..0bbe030 100644 --- a/include/linux/netdevice.h +++ b/include/linux/netdevice.h @@ -214,6 +214,11 @@ enum { #include <linux/cache.h> #include <linux/skbuff.h> +#ifdef CONFIG_RPS +#include <linux/jump_label.h> +extern struct jump_label_key rps_needed; +#endif + struct neighbour; struct neigh_parms; struct sk_buff; diff --git a/net/core/dev.c b/net/core/dev.c index 26c49d5..f789599 100644 --- a/net/core/dev.c +++ b/net/core/dev.c @@ -2711,6 +2711,8 @@ EXPORT_SYMBOL(__skb_get_rxhash); struct rps_sock_flow_table __rcu *rps_sock_flow_table __read_mostly; EXPORT_SYMBOL(rps_sock_flow_table); +struct jump_label_key rps_needed __read_mostly; + static struct rps_dev_flow * set_rps_cpu(struct net_device *dev, struct sk_buff *skb, struct rps_dev_flow *rflow, u16 next_cpu) @@ -2994,7 +2996,7 @@ int netif_rx(struct sk_buff *skb) trace_netif_rx(skb); #ifdef CONFIG_RPS - { + if (static_branch(&rps_needed)) { struct rps_dev_flow voidflow, *rflow = &voidflow; int cpu; @@ -3009,14 +3011,13 @@ int netif_rx(struct sk_buff *skb) rcu_read_unlock(); preempt_enable(); - } -#else + } else +#endif { unsigned int qtail; ret = enqueue_to_backlog(skb, get_cpu(), &qtail); put_cpu(); } -#endif return ret; } EXPORT_SYMBOL(netif_rx); @@ -3359,7 +3360,7 @@ int netif_receive_skb(struct sk_buff *skb) return NET_RX_SUCCESS; #ifdef CONFIG_RPS - { + if (static_branch(&rps_needed)) { struct rps_dev_flow voidflow, *rflow = &voidflow; int cpu, ret; @@ -3370,16 +3371,12 @@ int netif_receive_skb(struct sk_buff *skb) if (cpu >= 0) { ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); rcu_read_unlock(); - } else { - rcu_read_unlock(); - ret = __netif_receive_skb(skb); + return ret; } - - return ret; + rcu_read_unlock(); } -#else - return __netif_receive_skb(skb); #endif + return __netif_receive_skb(skb); } EXPORT_SYMBOL(netif_receive_skb); diff --git a/net/core/net-sysfs.c b/net/core/net-sysfs.c index 602b141..db6c2f8 100644 --- a/net/core/net-sysfs.c +++ b/net/core/net-sysfs.c @@ -606,9 +606,12 @@ static ssize_t store_rps_map(struct netdev_rx_queue *queue, rcu_assign_pointer(queue->rps_map, map); spin_unlock(&rps_map_lock); - if (old_map) + if (map) + jump_label_inc(&rps_needed); + if (old_map) { kfree_rcu(old_map, rcu); - + jump_label_dec(&rps_needed); + } free_cpumask_var(mask); return len; } diff --git a/net/core/sysctl_net_core.c b/net/core/sysctl_net_core.c index 77a65f0..d05559d 100644 --- a/net/core/sysctl_net_core.c +++ b/net/core/sysctl_net_core.c @@ -68,8 +68,13 @@ static int rps_sock_flow_sysctl(ctl_table *table, int write, if (sock_table != orig_sock_table) { rcu_assign_pointer(rps_sock_flow_table, sock_table); - synchronize_rcu(); - vfree(orig_sock_table); + if (sock_table) + jump_label_inc(&rps_needed); + if (orig_sock_table) { + jump_label_dec(&rps_needed); + synchronize_rcu(); + vfree(orig_sock_table); + } } }
Most machines dont use RPS/RFS, and pay a fair amount of instructions in netif_receive_skb() / netif_rx() / get_rps_cpu() just to discover RPS/RFS is not setup. Add a jump_label named rps_needed. If no device rps_map or global rps_sock_flow_table is setup, netif_receive_skb() / netif_rx() do a single instruction instead of many ones, including conditional jumps. jmp +0 (if CONFIG_JUMP_LABEL=y) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> CC: Tom Herbert <therbert@google.com> --- V2: add netif_rx() optim as well include/linux/netdevice.h | 5 +++++ net/core/dev.c | 21 +++++++++------------ net/core/net-sysfs.c | 7 +++++-- net/core/sysctl_net_core.c | 9 +++++++-- 4 files changed, 26 insertions(+), 16 deletions(-) -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html