Merge branch 'net-lockless-skb_attempt_defer_free'

Eric Dumazet says:

====================
net: lockless skb_attempt_defer_free()

Platforms with many cpus and relatively slow inter connect show
a significant spinlock contention in skb_attempt_defer_free().

This series refactors this infrastructure to be NUMA aware,
and lockless.

Tested on various platforms, including AMD Zen 2/3/4
and Intel Granite Rapids, showing significant cost reductions
under network stress (more than 20 Mpps).
====================

Link: https://patch.msgid.link/20250928084934.3266948-1-edumazet@google.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>

3 files changed, 35 insertions, 32 deletions

diff --git a/net/core/dev.c b/net/core/dev.c
index 8b54fdf0289a..a64cef2c537e 100644
--- a/net/core/dev.c
+++ b/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()) {
@@ -6715,24 +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;
 
-	/* Paired with WRITE_ONCE() in skb_attempt_defer_free() */
-	if (!READ_ONCE(sd->defer_list))
-		return;
+	for_each_node(node) {
+		sdn = this_cpu_ptr(net_hotdata.skb_defer_nodes) + node;
 
-	spin_lock(&sd->defer_lock);
-	skb = sd->defer_list;
-	sd->defer_list = NULL;
-	sd->defer_count = 0;
-	spin_unlock(&sd->defer_lock);
+		if (llist_empty(&sdn->defer_list))
+			continue;
+		atomic_long_set(&sdn->defer_count, 0);
+		free_list = llist_del_all(&sdn->defer_list);
 
-	while (skb != NULL) {
-		next = skb->next;
-		napi_consume_skb(skb, 1);
-		skb = next;
+		llist_for_each_entry_safe(skb, next, free_list, ll_node) {
+			napi_consume_skb(skb, 1);
+		}
 	}
 }
 
@@ -6860,7 +6861,7 @@ count:
 		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();
 
@@ -7719,7 +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();
 
@@ -7761,7 +7762,7 @@ start:
 	for (;;) {
 		struct napi_struct *n;
 
-		skb_defer_free_flush(sd);
+		skb_defer_free_flush();
 
 		if (list_empty(&list)) {
 			if (list_empty(&repoll)) {
@@ -12995,7 +12996,6 @@ static int __init net_dev_init(void)
 		sd->cpu = i;
 #endif
 		INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd);
-		spin_lock_init(&sd->defer_lock);
 
 		gro_init(&sd->backlog.gro);
 		sd->backlog.poll = process_backlog;
@@ -13005,6 +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);
 
diff --git a/net/core/dev.h b/net/core/dev.h
index d6b08d435479..900880e8b5b4 100644
--- a/net/core/dev.h
+++ b/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
 
diff --git a/net/core/skbuff.c b/net/core/skbuff.c
index 618afd59afff..bc12790017b0 100644
--- a/net/core/skbuff.c
+++ b/net/core/skbuff.c
@@ -7185,8 +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;
 
@@ -7200,27 +7201,24 @@ 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);
-	if (READ_ONCE(sd->defer_count) >= defer_max)
+	defer_count = atomic_long_inc_return(&sdn->defer_count);
+
+	if (defer_count >= defer_max)
 		goto nodefer;
 
-	spin_lock_bh(&sd->defer_lock);
-	/* Send an IPI every time queue reaches half capacity. */
-	kick = sd->defer_count == (defer_max >> 1);
-	/* Paired with the READ_ONCE() few lines above */
-	WRITE_ONCE(sd->defer_count, sd->defer_count + 1);
+	llist_add(&skb->ll_node, &sdn->defer_list);
 
-	skb->next = sd->defer_list;
-	/* Paired with READ_ONCE() in skb_defer_free_flush() */
-	WRITE_ONCE(sd->defer_list, skb);
-	spin_unlock_bh(&sd->defer_lock);
+	/* Send an IPI every time queue reaches half capacity. */
+	kick = (defer_count - 1) == (defer_max >> 1);
 
 	/* Make sure to trigger NET_RX_SOFTIRQ on the remote CPU
 	 * 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,