Merge branch 'net-rps-misc'

Eric Dumazet says:

====================
net: rps: misc changes

Make RPS/RFS a bit more efficient with better cache locality
and heuristics.

Aso shrink include/linux/netdevice.h a bit.

v2: fixed a build issue in patch 6/8 with CONFIG_RPS=n
(Jakub and kernel build bots)
====================

Signed-off-by: David S. Miller <[e-mail]>

include/linux/netdevice.h

···
 	struct softnet_data	*rps_ipi_list;
 #endif
 
+	unsigned int		received_rps;
 	bool			in_net_rx_action;
 	bool			in_napi_threaded_poll;
 
···
 	unsigned int		cpu;
 	unsigned int		input_queue_tail;
 #endif
-	unsigned int		received_rps;
-	unsigned int		dropped;
 	struct sk_buff_head	input_pkt_queue;
 	struct napi_struct	backlog;
+
+	atomic_t		dropped ____cacheline_aligned_in_smp;
 
 	/* Another possibly contended cache line */
 	spinlock_t		defer_lock ____cacheline_aligned_in_smp;
···
 	call_single_data_t	defer_csd;
 };
 
-static inline void input_queue_head_incr(struct softnet_data *sd)
-{
-#ifdef CONFIG_RPS
-	sd->input_queue_head++;
-#endif
-}
-
-static inline void input_queue_tail_incr_save(struct softnet_data *sd,
-					      unsigned int *qtail)
-{
-#ifdef CONFIG_RPS
-	*qtail = ++sd->input_queue_tail;
-#endif
-}
-
 DECLARE_PER_CPU_ALIGNED(struct softnet_data, softnet_data);
 
 static inline int dev_recursion_level(void)
···
 	return this_cpu_read(softnet_data.xmit.recursion);
 }
 
-#define XMIT_RECURSION_LIMIT	8
-static inline bool dev_xmit_recursion(void)
-{
-	return unlikely(__this_cpu_read(softnet_data.xmit.recursion) >
-			XMIT_RECURSION_LIMIT);
-}
-
-static inline void dev_xmit_recursion_inc(void)
-{
-	__this_cpu_inc(softnet_data.xmit.recursion);
-}
-
-static inline void dev_xmit_recursion_dec(void)
-{
-	__this_cpu_dec(softnet_data.xmit.recursion);
-}
-
-void kick_defer_list_purge(struct softnet_data *sd, unsigned int cpu);
 void __netif_schedule(struct Qdisc *q);
 void netif_schedule_queue(struct netdev_queue *txq);
 

include/net/rps.h

···
 #endif
 }
 
+static inline u32 rps_input_queue_tail_incr(struct softnet_data *sd)
+{
+#ifdef CONFIG_RPS
+	return ++sd->input_queue_tail;
+#else
+	return 0;
+#endif
+}
+
+static inline void rps_input_queue_tail_save(u32 *dest, u32 tail)
+{
+#ifdef CONFIG_RPS
+	WRITE_ONCE(*dest, tail);
+#endif
+}
+
+static inline void rps_input_queue_head_add(struct softnet_data *sd, int val)
+{
+#ifdef CONFIG_RPS
+	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 */

net/core/dev.c

···
 	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;
···
 		 */
 		if (unlikely(tcpu != next_cpu) &&
 		    (tcpu >= nr_cpu_ids || !cpu_online(tcpu) ||
-		     ((int)(per_cpu(softnet_data, tcpu).input_queue_head -
-		      rflow->last_qtail)) >= 0)) {
+		     ((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);
 		}
···
 		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 -
-			   rflow->last_qtail) <
+		    ((int)(READ_ONCE(per_cpu(softnet_data, cpu).input_queue_head) -
+			   READ_ONCE(rflow->last_qtail)) <
 		     (int)(10 * flow_table->mask)))
 			expire = false;
 	}
···
 	struct softnet_data *sd;
 	unsigned long flags;
 	unsigned int qlen;
+	int max_backlog;
+	u32 tail;
 
-	reason = SKB_DROP_REASON_NOT_SPECIFIED;
+	reason = SKB_DROP_REASON_DEV_READY;
+	if (!netif_running(skb->dev))
+		goto bad_dev;
+
+	reason = SKB_DROP_REASON_CPU_BACKLOG;
 	sd = &per_cpu(softnet_data, cpu);
 
+	qlen = skb_queue_len_lockless(&sd->input_pkt_queue);
+	max_backlog = READ_ONCE(net_hotdata.max_backlog);
+	if (unlikely(qlen > max_backlog))
+		goto cpu_backlog_drop;
 	backlog_lock_irq_save(sd, &flags);
-	if (!netif_running(skb->dev))
-		goto drop;
 	qlen = skb_queue_len(&sd->input_pkt_queue);
-	if (qlen <= READ_ONCE(net_hotdata.max_backlog) &&
-	    !skb_flow_limit(skb, qlen)) {
-		if (qlen) {
-enqueue:
-			__skb_queue_tail(&sd->input_pkt_queue, skb);
-			input_queue_tail_incr_save(sd, qtail);
-			backlog_unlock_irq_restore(sd, &flags);
-			return NET_RX_SUCCESS;
+	if (qlen <= max_backlog && !skb_flow_limit(skb, qlen)) {
+		if (!qlen) {
+			/* Schedule NAPI for backlog device. We can use
+			 * non atomic operation as we own the queue lock.
+			 */
+			if (!__test_and_set_bit(NAPI_STATE_SCHED,
+						&sd->backlog.state))
+				napi_schedule_rps(sd);
 		}
+		__skb_queue_tail(&sd->input_pkt_queue, skb);
+		tail = rps_input_queue_tail_incr(sd);
+		backlog_unlock_irq_restore(sd, &flags);
 
-		/* Schedule NAPI for backlog device
-		 * We can use non atomic operation since we own the queue lock
-		 */
-		if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state))
-			napi_schedule_rps(sd);
-		goto enqueue;
+		/* save the tail outside of the critical section */
+		rps_input_queue_tail_save(qtail, tail);
+		return NET_RX_SUCCESS;
 	}
-	reason = SKB_DROP_REASON_CPU_BACKLOG;
 
-drop:
-	sd->dropped++;
 	backlog_unlock_irq_restore(sd, &flags);
 
+cpu_backlog_drop:
+	atomic_inc(&sd->dropped);
+bad_dev:
 	dev_core_stats_rx_dropped_inc(skb->dev);
 	kfree_skb_reason(skb, reason);
 	return NET_RX_DROP;
···
 		if (skb->dev->reg_state == NETREG_UNREGISTERING) {
 			__skb_unlink(skb, &sd->input_pkt_queue);
 			dev_kfree_skb_irq(skb);
-			input_queue_head_incr(sd);
+			rps_input_queue_head_incr(sd);
 		}
 	}
 	backlog_unlock_irq_enable(sd);
···
 		if (skb->dev->reg_state == NETREG_UNREGISTERING) {
 			__skb_unlink(skb, &sd->process_queue);
 			kfree_skb(skb);
-			input_queue_head_incr(sd);
+			rps_input_queue_head_incr(sd);
 		}
 	}
 	local_bh_enable();
···
 			rcu_read_lock();
 			__netif_receive_skb(skb);
 			rcu_read_unlock();
-			input_queue_head_incr(sd);
-			if (++work >= quota)
+			if (++work >= quota) {
+				rps_input_queue_head_add(sd, work);
 				return work;
+			}
 
 		}
 
···
 		backlog_unlock_irq_enable(sd);
 	}
 
+	if (work)
+		rps_input_queue_head_add(sd, work);
 	return work;
 }
 
···
 	/* Process offline CPU's input_pkt_queue */
 	while ((skb = __skb_dequeue(&oldsd->process_queue))) {
 		netif_rx(skb);
-		input_queue_head_incr(oldsd);
+		rps_input_queue_head_incr(oldsd);
 	}
 	while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) {
 		netif_rx(skb);
-		input_queue_head_incr(oldsd);
+		rps_input_queue_head_incr(oldsd);
 	}
 
 	return 0;

net/core/dev.h

···
 
 #include <linux/types.h>
 #include <linux/rwsem.h>
+#include <linux/netdevice.h>
 
 struct net;
-struct net_device;
-struct netdev_bpf;
-struct netdev_phys_item_id;
 struct netlink_ext_ack;
 struct cpumask;
 
···
 #endif
 
 struct napi_struct *napi_by_id(unsigned int napi_id);
+void kick_defer_list_purge(struct softnet_data *sd, unsigned int cpu);
+
+#define XMIT_RECURSION_LIMIT	8
+static inline bool dev_xmit_recursion(void)
+{
+	return unlikely(__this_cpu_read(softnet_data.xmit.recursion) >
+			XMIT_RECURSION_LIMIT);
+}
+
+static inline void dev_xmit_recursion_inc(void)
+{
+	__this_cpu_inc(softnet_data.xmit.recursion);
+}
+
+static inline void dev_xmit_recursion_dec(void)
+{
+	__this_cpu_dec(softnet_data.xmit.recursion);
+}
+
 #endif

net/core/net-procfs.c

···
 	seq_printf(seq,
 		   "%08x %08x %08x %08x %08x %08x %08x %08x %08x %08x %08x %08x %08x "
 		   "%08x %08x\n",
-		   sd->processed, sd->dropped, sd->time_squeeze, 0,
+		   sd->processed, atomic_read(&sd->dropped),
+		   sd->time_squeeze, 0,
 		   0, 0, 0, 0, /* was fastroute */
 		   0,	/* was cpu_collision */
 		   sd->received_rps, flow_limit_count,