[PATCH 1/6] netfilter: conntrack: resched in nf_ct_iterate_cleanup

@@ -21,6 +21,10 @@
 #include <net/ipv6.h>
 #include <net/netfilter/ipv6/nf_nat_masquerade.h>
 
+#define MAX_WORK_COUNT	16
+
+static atomic_t v6_worker_count;
+
 unsigned int
 nf_nat_masquerade_ipv6(struct sk_buff *skb, const struct nf_nat_range *range,
 		       const struct net_device *out)

@@ -78,14 +82,78 @@ static struct notifier_block masq_dev_notifier = {
 	.notifier_call	= masq_device_event,
 };
 
+struct masq_dev_work {
+	struct work_struct work;
+	struct net *net;
+	int ifindex;
+};
+
+static void iterate_cleanup_work(struct work_struct *work)
+{
+	struct masq_dev_work *w;
+	long index;
+
+	w = container_of(work, struct masq_dev_work, work);
+
+	index = w->ifindex;
+	nf_ct_iterate_cleanup(w->net, device_cmp, (void *)index, 0, 0);
+
+	put_net(w->net);
+	kfree(w);
+	atomic_dec(&v6_worker_count);
+	module_put(THIS_MODULE);
+}
+
+/* ipv6 inet notifier is an atomic notifier, i.e. we cannot
+ * schedule.
+ *
+ * Unfortunately, nf_ct_iterate_cleanup can run for a long
+ * time if there are lots of conntracks and the system
+ * handles high softirq load, so it frequently calls cond_resched
+ * while iterating the conntrack table.
+ *
+ * So we defer nf_ct_iterate_cleanup walk to the system workqueue.
+ *
+ * As we can have 'a lot' of inet_events (depending on amount
+ * of ipv6 addresses being deleted), we also need to add an upper
+ * limit to the number of queued work items.
+ */
 static int masq_inet_event(struct notifier_block *this,
 			   unsigned long event, void *ptr)
 {
 	struct inet6_ifaddr *ifa = ptr;
-	struct netdev_notifier_info info;
+	const struct net_device *dev;
+	struct masq_dev_work *w;
+	struct net *net;
+
+	if (event != NETDEV_DOWN ||
+	    atomic_read(&v6_worker_count) >= MAX_WORK_COUNT)
+		return NOTIFY_DONE;
+
+	dev = ifa->idev->dev;
+	net = maybe_get_net(dev_net(dev));
+	if (!net)
+		return NOTIFY_DONE;
 
-	netdev_notifier_info_init(&info, ifa->idev->dev);
-	return masq_device_event(this, event, &info);
+	if (!try_module_get(THIS_MODULE))
+		goto err_module;
+
+	w = kmalloc(sizeof(*w), GFP_ATOMIC);
+	if (w) {
+		atomic_inc(&v6_worker_count);
+
+		INIT_WORK(&w->work, iterate_cleanup_work);
+		w->ifindex = dev->ifindex;
+		w->net = net;
+		schedule_work(&w->work);
+
+		return NOTIFY_DONE;
+	}
+
+	module_put(THIS_MODULE);
+ err_module:
+	put_net(net);
+	return NOTIFY_DONE;
 }
 
 static struct notifier_block masq_inet_notifier = {

@@ -1412,6 +1412,7 @@ get_next_corpse(struct net *net, int (*iter)(struct nf_conn *i, void *data),
 		}
 		spin_unlock(lockp);
 		local_bh_enable();
+		cond_resched();
 	}
 
 	for_each_possible_cpu(cpu) {

@@ -1424,6 +1425,7 @@ get_next_corpse(struct net *net, int (*iter)(struct nf_conn *i, void *data),
 				set_bit(IPS_DYING_BIT, &ct->status);
 		}
 		spin_unlock_bh(&pcpu->lock);
+		cond_resched();
 	}
 	return NULL;
 found:

@@ -1440,6 +1442,8 @@ void nf_ct_iterate_cleanup(struct net *net,
 	struct nf_conn *ct;
 	unsigned int bucket = 0;
 
+	might_sleep();
+
 	while ((ct = get_next_corpse(net, iter, data, &bucket)) != NULL) {
 		/* Time to push up daises... */
 		if (del_timer(&ct->timeout))

@@ -1448,6 +1452,7 @@ void nf_ct_iterate_cleanup(struct net *net,
 		/* ... else the timer will get him soon. */
 
 		nf_ct_put(ct);
+		cond_resched();
 	}
 }
 EXPORT_SYMBOL_GPL(nf_ct_iterate_cleanup);