Merge branch 'net-more-factorization-in-cleanup_net-paths'

Eric Dumazet says:

====================
net: more factorization in cleanup_net() paths

This series is inspired by recent syzbot reports hinting to RTNL and
workqueue abuses.

rtnl_lock() is unfair to (single threaded) cleanup_net(), because
many threads can cause contention on it.

This series adds a new (struct pernet_operations) method,
so that cleanup_net() can hold RTNL longer once it finally
acquires it.

It also factorizes unregister_netdevice_many(), to further
reduce stalls in cleanup_net().

Link: https://lore.kernel.org/netdev/CANn89iLJrrJs+6Vc==Un4rVKcpV0Eof4F_4w1_wQGxUCE2FWAg@mail.gmail.com/T/#u
https://lore.kernel.org/netdev/170688415193.5216.10499830272732622816@kwain/
====================

Link: https://lore.kernel.org/r/20240206144313.2050392-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

drivers/net/bareudp.c

@@ -760,23 +760,18 @@ static void bareudp_destroy_tunnels(struct net *net, struct list_head *head)
 		unregister_netdevice_queue(bareudp->dev, head);
 }
 
-static void __net_exit bareudp_exit_batch_net(struct list_head *net_list)
+static void __net_exit bareudp_exit_batch_rtnl(struct list_head *net_list,
+					       struct list_head *dev_kill_list)
 {
 	struct net *net;
-	LIST_HEAD(list);
 
-	rtnl_lock();
 	list_for_each_entry(net, net_list, exit_list)
-		bareudp_destroy_tunnels(net, &list);
-
-	/* unregister the devices gathered above */
-	unregister_netdevice_many(&list);
-	rtnl_unlock();
+		bareudp_destroy_tunnels(net, dev_kill_list);
 }
 
 static struct pernet_operations bareudp_net_ops = {
 	.init = bareudp_init_net,
-	.exit_batch = bareudp_exit_batch_net,
+	.exit_batch_rtnl = bareudp_exit_batch_rtnl,
 	.id   = &bareudp_net_id,
 	.size = sizeof(struct bareudp_net),
 };

drivers/net/bonding/bond_main.c

@@ -6416,28 +6416,41 @@ static int __net_init bond_net_init(struct net *net)
 	return 0;
 }
 
-static void __net_exit bond_net_exit_batch(struct list_head *net_list)
+/* According to commit 69b0216ac255 ("bonding: fix bonding_masters
+ * race condition in bond unloading") we need to remove sysfs files
+ * before we remove our devices (done later in bond_net_exit_batch_rtnl())
+ */
+static void __net_exit bond_net_pre_exit(struct net *net)
+{
+	struct bond_net *bn = net_generic(net, bond_net_id);
+
+	bond_destroy_sysfs(bn);
+}
+
+static void __net_exit bond_net_exit_batch_rtnl(struct list_head *net_list,
+						struct list_head *dev_kill_list)
 {
 	struct bond_net *bn;
 	struct net *net;
-	LIST_HEAD(list);
-
-	list_for_each_entry(net, net_list, exit_list) {
-		bn = net_generic(net, bond_net_id);
-		bond_destroy_sysfs(bn);
-	}
 
 	/* Kill off any bonds created after unregistering bond rtnl ops */
-	rtnl_lock();
 	list_for_each_entry(net, net_list, exit_list) {
 		struct bonding *bond, *tmp_bond;
 
 		bn = net_generic(net, bond_net_id);
 		list_for_each_entry_safe(bond, tmp_bond, &bn->dev_list, bond_list)
-			unregister_netdevice_queue(bond->dev, &list);
+			unregister_netdevice_queue(bond->dev, dev_kill_list);
 	}
-	unregister_netdevice_many(&list);
-	rtnl_unlock();
+}
+
+/* According to commit 23fa5c2caae0 ("bonding: destroy proc directory
+ * only after all bonds are gone") bond_destroy_proc_dir() is called
+ * after bond_net_exit_batch_rtnl() has completed.
+ */
+static void __net_exit bond_net_exit_batch(struct list_head *net_list)
+{
+	struct bond_net *bn;
+	struct net *net;
 
 	list_for_each_entry(net, net_list, exit_list) {
 		bn = net_generic(net, bond_net_id);
@@ -6447,6 +6460,8 @@ static void __net_exit bond_net_exit_batch(struct list_head *net_list)
 
 static struct pernet_operations bond_net_ops = {
 	.init = bond_net_init,
+	.pre_exit = bond_net_pre_exit,
+	.exit_batch_rtnl = bond_net_exit_batch_rtnl,
 	.exit_batch = bond_net_exit_batch,
 	.id   = &bond_net_id,
 	.size = sizeof(struct bond_net),

drivers/net/geneve.c

@@ -1900,29 +1900,26 @@ static void geneve_destroy_tunnels(struct net *net, struct list_head *head)
 	}
 }
 
-static void __net_exit geneve_exit_batch_net(struct list_head *net_list)
+static void __net_exit geneve_exit_batch_rtnl(struct list_head *net_list,
+					      struct list_head *dev_to_kill)
 {
 	struct net *net;
-	LIST_HEAD(list);
 
-	rtnl_lock();
 	list_for_each_entry(net, net_list, exit_list)
-		geneve_destroy_tunnels(net, &list);
-
-	/* unregister the devices gathered above */
-	unregister_netdevice_many(&list);
-	rtnl_unlock();
+		geneve_destroy_tunnels(net, dev_to_kill);
+}
 
-	list_for_each_entry(net, net_list, exit_list) {
-		const struct geneve_net *gn = net_generic(net, geneve_net_id);
+static void __net_exit geneve_exit_net(struct net *net)
+{
+	const struct geneve_net *gn = net_generic(net, geneve_net_id);
 
-		WARN_ON_ONCE(!list_empty(&gn->sock_list));
-	}
+	WARN_ON_ONCE(!list_empty(&gn->sock_list));
 }
 
 static struct pernet_operations geneve_net_ops = {
 	.init = geneve_init_net,
-	.exit_batch = geneve_exit_batch_net,
+	.exit_batch_rtnl = geneve_exit_batch_rtnl,
+	.exit = geneve_exit_net,
 	.id   = &geneve_net_id,
 	.size = sizeof(struct geneve_net),
 };

drivers/net/gtp.c

@@ -1876,23 +1876,23 @@ static int __net_init gtp_net_init(struct net *net)
 	return 0;
 }
 
-static void __net_exit gtp_net_exit(struct net *net)
+static void __net_exit gtp_net_exit_batch_rtnl(struct list_head *net_list,
+					       struct list_head *dev_to_kill)
 {
-	struct gtp_net *gn = net_generic(net, gtp_net_id);
-	struct gtp_dev *gtp;
-	LIST_HEAD(list);
+	struct net *net;
 
-	rtnl_lock();
-	list_for_each_entry(gtp, &gn->gtp_dev_list, list)
-		gtp_dellink(gtp->dev, &list);
+	list_for_each_entry(net, net_list, exit_list) {
+		struct gtp_net *gn = net_generic(net, gtp_net_id);
+		struct gtp_dev *gtp;
 
-	unregister_netdevice_many(&list);
-	rtnl_unlock();
+		list_for_each_entry(gtp, &gn->gtp_dev_list, list)
+			gtp_dellink(gtp->dev, dev_to_kill);
+	}
 }
 
 static struct pernet_operations gtp_net_ops = {
 	.init	= gtp_net_init,
-	.exit	= gtp_net_exit,
+	.exit_batch_rtnl = gtp_net_exit_batch_rtnl,
 	.id	= &gtp_net_id,
 	.size	= sizeof(struct gtp_net),
 };

drivers/net/vxlan/vxlan_core.c

@@ -4826,55 +4826,43 @@ static __net_init int vxlan_init_net(struct net *net)
 					 NULL);
 }
 
-static void vxlan_destroy_tunnels(struct net *net, struct list_head *head)
+static void __net_exit vxlan_destroy_tunnels(struct vxlan_net *vn,
+					     struct list_head *dev_to_kill)
 {
-	struct vxlan_net *vn = net_generic(net, vxlan_net_id);
 	struct vxlan_dev *vxlan, *next;
-	struct net_device *dev, *aux;
-
-	for_each_netdev_safe(net, dev, aux)
-		if (dev->rtnl_link_ops == &vxlan_link_ops)
-			unregister_netdevice_queue(dev, head);
-
-	list_for_each_entry_safe(vxlan, next, &vn->vxlan_list, next) {
-		/* If vxlan->dev is in the same netns, it has already been added
-		 * to the list by the previous loop.
-		 */
-		if (!net_eq(dev_net(vxlan->dev), net))
-			unregister_netdevice_queue(vxlan->dev, head);
-	}
 
+	list_for_each_entry_safe(vxlan, next, &vn->vxlan_list, next)
+		vxlan_dellink(vxlan->dev, dev_to_kill);
 }
 
-static void __net_exit vxlan_exit_batch_net(struct list_head *net_list)
+static void __net_exit vxlan_exit_batch_rtnl(struct list_head *net_list,
+					     struct list_head *dev_to_kill)
 {
 	struct net *net;
-	LIST_HEAD(list);
-	unsigned int h;
 
+	ASSERT_RTNL();
 	list_for_each_entry(net, net_list, exit_list) {
 		struct vxlan_net *vn = net_generic(net, vxlan_net_id);
 
-		unregister_nexthop_notifier(net, &vn->nexthop_notifier_block);
-	}
-	rtnl_lock();
-	list_for_each_entry(net, net_list, exit_list)
-		vxlan_destroy_tunnels(net, &list);
+		__unregister_nexthop_notifier(net, &vn->nexthop_notifier_block);
 
-	unregister_netdevice_many(&list);
-	rtnl_unlock();
+		vxlan_destroy_tunnels(vn, dev_to_kill);
+	}
+}
 
-	list_for_each_entry(net, net_list, exit_list) {
-		struct vxlan_net *vn = net_generic(net, vxlan_net_id);
+static void __net_exit vxlan_exit_net(struct net *net)
+{
+	struct vxlan_net *vn = net_generic(net, vxlan_net_id);
+	unsigned int h;
 
-		for (h = 0; h < PORT_HASH_SIZE; ++h)
-			WARN_ON_ONCE(!hlist_empty(&vn->sock_list[h]));
-	}
+	for (h = 0; h < PORT_HASH_SIZE; ++h)
+		WARN_ON_ONCE(!hlist_empty(&vn->sock_list[h]));
 }
 
 static struct pernet_operations vxlan_net_ops = {
 	.init = vxlan_init_net,
-	.exit_batch = vxlan_exit_batch_net,
+	.exit_batch_rtnl = vxlan_exit_batch_rtnl,
+	.exit = vxlan_exit_net,
 	.id   = &vxlan_net_id,
 	.size = sizeof(struct vxlan_net),
 };

include/net/ip_tunnels.h

@@ -284,7 +284,8 @@ int ip_tunnel_init_net(struct net *net, unsigned int ip_tnl_net_id,
 		       struct rtnl_link_ops *ops, char *devname);
 
 void ip_tunnel_delete_nets(struct list_head *list_net, unsigned int id,
-			   struct rtnl_link_ops *ops);
+			   struct rtnl_link_ops *ops,
+			   struct list_head *dev_to_kill);
 
 void ip_tunnel_xmit(struct sk_buff *skb, struct net_device *dev,
 		    const struct iphdr *tnl_params, const u8 protocol);

include/net/net_namespace.h

@@ -448,6 +448,9 @@ struct pernet_operations {
 	void (*pre_exit)(struct net *net);
 	void (*exit)(struct net *net);
 	void (*exit_batch)(struct list_head *net_exit_list);
+	/* Following method is called with RTNL held. */
+	void (*exit_batch_rtnl)(struct list_head *net_exit_list,
+				struct list_head *dev_kill_list);
 	unsigned int *id;
 	size_t size;
 };

include/net/nexthop.h

@@ -218,6 +218,7 @@ struct nh_notifier_info {
 
 int register_nexthop_notifier(struct net *net, struct notifier_block *nb,
 			      struct netlink_ext_ack *extack);
+int __unregister_nexthop_notifier(struct net *net, struct notifier_block *nb);
 int unregister_nexthop_notifier(struct net *net, struct notifier_block *nb);
 void nexthop_set_hw_flags(struct net *net, u32 id, bool offload, bool trap);
 void nexthop_bucket_set_hw_flags(struct net *net, u32 id, u16 bucket_index,

net/bridge/br.c

@@ -356,26 +356,21 @@ void br_opt_toggle(struct net_bridge *br, enum net_bridge_opts opt, bool on)
 		clear_bit(opt, &br->options);
 }
 
-static void __net_exit br_net_exit_batch(struct list_head *net_list)
+static void __net_exit br_net_exit_batch_rtnl(struct list_head *net_list,
+					      struct list_head *dev_to_kill)
 {
 	struct net_device *dev;
 	struct net *net;
-	LIST_HEAD(list);
-
-	rtnl_lock();
 
+	ASSERT_RTNL();
 	list_for_each_entry(net, net_list, exit_list)
 		for_each_netdev(net, dev)
 			if (netif_is_bridge_master(dev))
-				br_dev_delete(dev, &list);
-
-	unregister_netdevice_many(&list);
-
-	rtnl_unlock();
+				br_dev_delete(dev, dev_to_kill);
 }
 
 static struct pernet_operations br_net_ops = {
-	.exit_batch	= br_net_exit_batch,
+	.exit_batch_rtnl = br_net_exit_batch_rtnl,
 };
 
 static const struct stp_proto br_stp_proto = {

net/core/net_namespace.c

@@ -318,8 +318,9 @@ static __net_init int setup_net(struct net *net, struct user_namespace *user_ns)
 {
 	/* Must be called with pernet_ops_rwsem held */
 	const struct pernet_operations *ops, *saved_ops;
-	int error = 0;
 	LIST_HEAD(net_exit_list);
+	LIST_HEAD(dev_kill_list);
+	int error = 0;
 
 	refcount_set(&net->ns.count, 1);
 	ref_tracker_dir_init(&net->refcnt_tracker, 128, "net refcnt");
@@ -357,6 +358,15 @@ static __net_init int setup_net(struct net *net, struct user_namespace *user_ns)
 
 	synchronize_rcu();
 
+	ops = saved_ops;
+	rtnl_lock();
+	list_for_each_entry_continue_reverse(ops, &pernet_list, list) {
+		if (ops->exit_batch_rtnl)
+			ops->exit_batch_rtnl(&net_exit_list, &dev_kill_list);
+	}
+	unregister_netdevice_many(&dev_kill_list);
+	rtnl_unlock();
+
 	ops = saved_ops;
 	list_for_each_entry_continue_reverse(ops, &pernet_list, list)
 		ops_exit_list(ops, &net_exit_list);
@@ -573,6 +583,7 @@ static void cleanup_net(struct work_struct *work)
 	struct net *net, *tmp, *last;
 	struct llist_node *net_kill_list;
 	LIST_HEAD(net_exit_list);
+	LIST_HEAD(dev_kill_list);
 
 	/* Atomically snapshot the list of namespaces to cleanup */
 	net_kill_list = llist_del_all(&cleanup_list);
@@ -613,6 +624,14 @@ static void cleanup_net(struct work_struct *work)
 	 */
 	synchronize_rcu();
 
+	rtnl_lock();
+	list_for_each_entry_reverse(ops, &pernet_list, list) {
+		if (ops->exit_batch_rtnl)
+			ops->exit_batch_rtnl(&net_exit_list, &dev_kill_list);
+	}
+	unregister_netdevice_many(&dev_kill_list);
+	rtnl_unlock();
+
 	/* Run all of the network namespace exit methods */
 	list_for_each_entry_reverse(ops, &pernet_list, list)
 		ops_exit_list(ops, &net_exit_list);
@@ -1193,7 +1212,17 @@ static void free_exit_list(struct pernet_operations *ops, struct list_head *net_
 {
 	ops_pre_exit_list(ops, net_exit_list);
 	synchronize_rcu();
+
+	if (ops->exit_batch_rtnl) {
+		LIST_HEAD(dev_kill_list);
+
+		rtnl_lock();
+		ops->exit_batch_rtnl(net_exit_list, &dev_kill_list);
+		unregister_netdevice_many(&dev_kill_list);
+		rtnl_unlock();
+	}
 	ops_exit_list(ops, net_exit_list);
+
 	ops_free_list(ops, net_exit_list);
 }