net: add NUMA awareness to skb_attempt_defer_free()

Instead of sharing sd->defer_list & sd->defer_count with
many cpus, add one pair for each NUMA node.

Signed-off-by: Eric Dumazet <[email protected]>
Reviewed-by: Jason Xing <[email protected]>
Reviewed-by: Kuniyuki Iwashima <[email protected]>
Link: https://patch.msgid.link/[email protected]
Signed-off-by: Paolo Abeni <[email protected]>

Author: edumazet

include/linux/netdevice.h

@@ -3536,10 +3536,6 @@ struct softnet_data {
 
 	struct numa_drop_counters drop_counters;
 
-	/* Another possibly contended cache line */
-	struct llist_head	defer_list ____cacheline_aligned_in_smp;
-	atomic_long_t		defer_count;
-
 	int			defer_ipi_scheduled ____cacheline_aligned_in_smp;
 	call_single_data_t	defer_csd;
 };

include/net/hotdata.h

@@ -2,10 +2,16 @@
 #ifndef _NET_HOTDATA_H
 #define _NET_HOTDATA_H
 
+#include <linux/llist.h>
 #include <linux/types.h>
 #include <linux/netdevice.h>
 #include <net/protocol.h>
 
+struct skb_defer_node {
+	struct llist_head	defer_list;
+	atomic_long_t		defer_count;
+} ____cacheline_aligned_in_smp;
+
 /* Read mostly data used in network fast paths. */
 struct net_hotdata {
 #if IS_ENABLED(CONFIG_INET)
@@ -30,6 +36,7 @@ struct net_hotdata {
 	struct rps_sock_flow_table __rcu *rps_sock_flow_table;
 	u32			rps_cpu_mask;
 #endif
+	struct skb_defer_node __percpu *skb_defer_nodes;
 	int			gro_normal_batch;
 	int			netdev_budget;
 	int			netdev_budget_usecs;

net/core/dev.c

@@ -5180,8 +5180,9 @@ static void napi_schedule_rps(struct softnet_data *sd)
 	__napi_schedule_irqoff(&mysd->backlog);
 }
 
-void kick_defer_list_purge(struct softnet_data *sd, unsigned int cpu)
+void kick_defer_list_purge(unsigned int cpu)
 {
+	struct softnet_data *sd = &per_cpu(softnet_data, cpu);
 	unsigned long flags;
 
 	if (use_backlog_threads()) {
@@ -6715,18 +6716,24 @@ bool napi_complete_done(struct napi_struct *n, int work_done)
 }
 EXPORT_SYMBOL(napi_complete_done);
 
-static void skb_defer_free_flush(struct softnet_data *sd)
+static void skb_defer_free_flush(void)
 {
 	struct llist_node *free_list;
 	struct sk_buff *skb, *next;
+	struct skb_defer_node *sdn;
+	int node;
 
-	if (llist_empty(&sd->defer_list))
-		return;
-	atomic_long_set(&sd->defer_count, 0);
-	free_list = llist_del_all(&sd->defer_list);
+	for_each_node(node) {
+		sdn = this_cpu_ptr(net_hotdata.skb_defer_nodes) + node;
+
+		if (llist_empty(&sdn->defer_list))
+			continue;
+		atomic_long_set(&sdn->defer_count, 0);
+		free_list = llist_del_all(&sdn->defer_list);
 
-	llist_for_each_entry_safe(skb, next, free_list, ll_node) {
-		napi_consume_skb(skb, 1);
+		llist_for_each_entry_safe(skb, next, free_list, ll_node) {
+			napi_consume_skb(skb, 1);
+		}
 	}
 }
 
@@ -6854,7 +6861,7 @@ static void __napi_busy_loop(unsigned int napi_id,
 		if (work > 0)
 			__NET_ADD_STATS(dev_net(napi->dev),
 					LINUX_MIB_BUSYPOLLRXPACKETS, work);
-		skb_defer_free_flush(this_cpu_ptr(&softnet_data));
+		skb_defer_free_flush();
 		bpf_net_ctx_clear(bpf_net_ctx);
 		local_bh_enable();
 
@@ -7713,7 +7720,7 @@ static void napi_threaded_poll_loop(struct napi_struct *napi)
 			local_irq_disable();
 			net_rps_action_and_irq_enable(sd);
 		}
-		skb_defer_free_flush(sd);
+		skb_defer_free_flush();
 		bpf_net_ctx_clear(bpf_net_ctx);
 		local_bh_enable();
 
@@ -7755,7 +7762,7 @@ static __latent_entropy void net_rx_action(void)
 	for (;;) {
 		struct napi_struct *n;
 
-		skb_defer_free_flush(sd);
+		skb_defer_free_flush();
 
 		if (list_empty(&list)) {
 			if (list_empty(&repoll)) {
@@ -12989,7 +12996,6 @@ static int __init net_dev_init(void)
 		sd->cpu = i;
 #endif
 		INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd);
-		init_llist_head(&sd->defer_list);
 
 		gro_init(&sd->backlog.gro);
 		sd->backlog.poll = process_backlog;
@@ -12999,6 +13005,11 @@ static int __init net_dev_init(void)
 		if (net_page_pool_create(i))
 			goto out;
 	}
+	net_hotdata.skb_defer_nodes =
+		 __alloc_percpu(sizeof(struct skb_defer_node) * nr_node_ids,
+				__alignof__(struct skb_defer_node));
+	if (!net_hotdata.skb_defer_nodes)
+		goto out;
 	if (use_backlog_threads())
 		smpboot_register_percpu_thread(&backlog_threads);
 

net/core/dev.h

@@ -357,7 +357,7 @@ static inline void napi_assert_will_not_race(const struct napi_struct *napi)
 	WARN_ON(READ_ONCE(napi->list_owner) != -1);
 }
 
-void kick_defer_list_purge(struct softnet_data *sd, unsigned int cpu);
+void kick_defer_list_purge(unsigned int cpu);
 
 #define XMIT_RECURSION_LIMIT	8
 

net/core/skbuff.c

@@ -7185,9 +7185,9 @@ static void kfree_skb_napi_cache(struct sk_buff *skb)
  */
 void skb_attempt_defer_free(struct sk_buff *skb)
 {
+	struct skb_defer_node *sdn;
 	unsigned long defer_count;
 	int cpu = skb->alloc_cpu;
-	struct softnet_data *sd;
 	unsigned int defer_max;
 	bool kick;
 
@@ -7201,14 +7201,15 @@ nodefer:	kfree_skb_napi_cache(skb);
 	DEBUG_NET_WARN_ON_ONCE(skb_dst(skb));
 	DEBUG_NET_WARN_ON_ONCE(skb->destructor);
 
-	sd = &per_cpu(softnet_data, cpu);
+	sdn = per_cpu_ptr(net_hotdata.skb_defer_nodes, cpu) + numa_node_id();
+
 	defer_max = READ_ONCE(net_hotdata.sysctl_skb_defer_max);
-	defer_count = atomic_long_inc_return(&sd->defer_count);
+	defer_count = atomic_long_inc_return(&sdn->defer_count);
 
 	if (defer_count >= defer_max)
 		goto nodefer;
 
-	llist_add(&skb->ll_node, &sd->defer_list);
+	llist_add(&skb->ll_node, &sdn->defer_list);
 
 	/* Send an IPI every time queue reaches half capacity. */
 	kick = (defer_count - 1) == (defer_max >> 1);
@@ -7217,7 +7218,7 @@ nodefer:	kfree_skb_napi_cache(skb);
 	 * if we are unlucky enough (this seems very unlikely).
 	 */
 	if (unlikely(kick))
-		kick_defer_list_purge(sd, cpu);
+		kick_defer_list_purge(cpu);
 }
 
 static void skb_splice_csum_page(struct sk_buff *skb, struct page *page,