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>

1 files changed, 30 insertions, 1 deletions

diff --git a/net/core/net_namespace.c b/net/core/net_namespace.c
index 72799533426b..233ec0cdd011 100644
--- a/net/core/net_namespace.c
+++ b/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");
@@ -358,6 +359,15 @@ out_undo:
 	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);
 }