net: reduce RTNL hold duration in unregister_netdevice_many_notify() (part 2)

One synchronize_net() call is currently done while holding RTNL.

This is source of RTNL contention in workloads adding and deleting
many network namespaces per second, because synchronize_rcu()
and synchronize_rcu_expedited() can use 60+ ms in some cases.

For cleanup_net() use, temporarily release RTNL
while calling the last synchronize_net().

This should be safe, because devices are no longer visible
to other threads after unlist_netdevice() call
and setting dev->reg_state to NETREG_UNREGISTERING.

In any case, the new netdev_lock() / netdev_unlock()
infrastructure that we are adding should allow
to fix potential issues, with a combination
of a per-device mutex and dev->reg_state awareness.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Jesse Brandeburg <jbrandeburg@cloudflare.com>
Link: https://patch.msgid.link/20250114205531.967841-6-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

diff --git a/net/core/dev.c b/net/core/dev.c
index 3924a4af68b81ba81301635cd195f9626d00693b..3c87cb1cb87716c249473fbf02e5ce50a012160c 100644 (file)
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -11628,9 +11628,8 @@ void unregister_netdevice_many_notify(struct list_head *head,
 
        rtnl_drop_if_cleanup_net();
        flush_all_backlogs();
-       rtnl_acquire_if_cleanup_net();
-       /* TODO: move this before the prior rtnl_acquire_if_cleanup_net() */
        synchronize_net();
+       rtnl_acquire_if_cleanup_net();
 
        list_for_each_entry(dev, head, unreg_list) {
                struct sk_buff *skb = NULL;