From c62fdf5b11ef12b89ac2450c25c12bffc3d924c5 Mon Sep 17 00:00:00 2001 From: Eric Dumazet Date: Fri, 29 Mar 2024 15:42:24 +0000 Subject: [PATCH] net: rps: add rps_input_queue_head_add() helper process_backlog() can batch increments of sd->input_queue_head, saving some memory bandwidth. Also add READ_ONCE()/WRITE_ONCE() annotations around sd->input_queue_head accesses. Signed-off-by: Eric Dumazet Signed-off-by: David S. Miller --- include/net/rps.h | 9 +++++++-- net/core/dev.c | 13 ++++++++----- 2 files changed, 15 insertions(+), 7 deletions(-) diff --git a/include/net/rps.h b/include/net/rps.h index 10ca25731c1e..a93401d23d66 100644 --- a/include/net/rps.h +++ b/include/net/rps.h @@ -138,11 +138,16 @@ static inline void rps_input_queue_tail_save(u32 *dest, u32 tail) #endif } -static inline void rps_input_queue_head_incr(struct softnet_data *sd) +static inline void rps_input_queue_head_add(struct softnet_data *sd, int val) { #ifdef CONFIG_RPS - sd->input_queue_head++; + WRITE_ONCE(sd->input_queue_head, sd->input_queue_head + val); #endif } +static inline void rps_input_queue_head_incr(struct softnet_data *sd) +{ + rps_input_queue_head_add(sd, 1); +} + #endif /* _NET_RPS_H */ diff --git a/net/core/dev.c b/net/core/dev.c index 79073bbc9a64..818699dea9d7 100644 --- a/net/core/dev.c +++ b/net/core/dev.c @@ -4528,7 +4528,7 @@ set_rps_cpu(struct net_device *dev, struct sk_buff *skb, out: #endif rflow->last_qtail = - per_cpu(softnet_data, next_cpu).input_queue_head; + READ_ONCE(per_cpu(softnet_data, next_cpu).input_queue_head); } rflow->cpu = next_cpu; @@ -4610,7 +4610,7 @@ static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb, */ if (unlikely(tcpu != next_cpu) && (tcpu >= nr_cpu_ids || !cpu_online(tcpu) || - ((int)(per_cpu(softnet_data, tcpu).input_queue_head - + ((int)(READ_ONCE(per_cpu(softnet_data, tcpu).input_queue_head) - READ_ONCE(rflow->last_qtail))) >= 0)) { tcpu = next_cpu; rflow = set_rps_cpu(dev, skb, rflow, next_cpu); @@ -4665,7 +4665,7 @@ bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index, rflow = &flow_table->flows[flow_id]; cpu = READ_ONCE(rflow->cpu); if (rflow->filter == filter_id && cpu < nr_cpu_ids && - ((int)(per_cpu(softnet_data, cpu).input_queue_head - + ((int)(READ_ONCE(per_cpu(softnet_data, cpu).input_queue_head) - READ_ONCE(rflow->last_qtail)) < (int)(10 * flow_table->mask))) expire = false; @@ -6045,9 +6045,10 @@ static int process_backlog(struct napi_struct *napi, int quota) rcu_read_lock(); __netif_receive_skb(skb); rcu_read_unlock(); - rps_input_queue_head_incr(sd); - if (++work >= quota) + if (++work >= quota) { + rps_input_queue_head_add(sd, work); return work; + } } @@ -6070,6 +6071,8 @@ static int process_backlog(struct napi_struct *napi, int quota) backlog_unlock_irq_enable(sd); } + if (work) + rps_input_queue_head_add(sd, work); return work; }