[PATCH net] netlink: don't call ->netlink_bind with table lock held

[PATCH net] netlink: don't call ->netlink_bind with table lock held

[Florian Westphal]

When I added support to allow generic netlink multicast groups to be
restricted to subscribers with CAP_NET_ADMIN I was unaware that a
genl_bind implementation already existed in the past.

It was reverted due to ABBA deadlock:

1. ->netlink_bind gets called with the table lock held.
2. genetlink bind callback is invoked, it grabs the genl lock.

But when a new genl subsystem is (un)registered, these two locks are
taken in reverse order.

One solution would be to revert again and add a comment in genl
referring 1e82a62fec613, "genetlink: remove genl_bind").

This would need a second change in mptcp to not expose the raw token
value anymore, e.g.  by hashing the token with a secret key so userspace
can still associate subflow events with the correct mptcp connection.

However, Paolo Abeni reminded me to double-check why the netlink table is
locked in the first place.

I can't find one.  netlink_bind() is already called without this lock
when userspace joins a group via NETLINK_ADD_MEMBERSHIP setsockopt.
Same holds for the netlink_unbind operation.

Digging through the history, commit f773608026ee1
("netlink: access nlk groups safely in netlink bind and getname")
expanded the lock scope.

commit 3a20773beeeeade ("net: netlink: cap max groups which will be considered in netlink_bind()")
... removed the nlk->ngroups access that the lock scope
extension was all about.

Reduce the lock scope again and always call ->netlink_bind without
the table lock.

The Fixes tag should be vs. the patch mentioned in the link below,
but that one got squash-merged into the patch that came earlier in the
series.

Fixes: 4d54cc32112d8d ("mptcp: avoid lock_fast usage in accept path")
Link: https://lore.kernel.org/mptcp/20210213000001.379332-8-mathew.j.martineau@linux.intel.com/T/#u
Cc: Cong Wang <xiyou.wangcong@gmail.com>
Cc: Xin Long <lucien.xin@gmail.com>
Cc: Johannes Berg <johannes.berg@intel.com>
Cc: Sean Tranchetti <stranche@codeaurora.org>
Cc: Paolo Abeni <pabeni@redhat.com>
Cc: Pablo Neira Ayuso <pablo@netfilter.org>
Signed-off-by: Florian Westphal <fw@strlen.de>
---
 net/netlink/af_netlink.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/net/netlink/af_netlink.c b/net/netlink/af_netlink.c
index dd488938447f..3a62f97acf39 100644
--- a/net/netlink/af_netlink.c
+++ b/net/netlink/af_netlink.c
@@ -1019,7 +1019,6 @@ static int netlink_bind(struct socket *sock, struct sockaddr *addr,
 			return -EINVAL;
 	}
 
-	netlink_lock_table();
 	if (nlk->netlink_bind && groups) {
 		int group;
 
@@ -1031,13 +1030,14 @@ static int netlink_bind(struct socket *sock, struct sockaddr *addr,
 			if (!err)
 				continue;
 			netlink_undo_bind(group, groups, sk);
-			goto unlock;
+			return err;
 		}
 	}
 
 	/* No need for barriers here as we return to user-space without
 	 * using any of the bound attributes.
 	 */
+	netlink_lock_table();
 	if (!bound) {
 		err = nladdr->nl_pid ?
 			netlink_insert(sk, nladdr->nl_pid) :
-- 
2.26.3

Re: [PATCH net] netlink: don't call ->netlink_bind with table lock held

[patchwork-bot+netdevbpf]

Hello:

This patch was applied to netdev/net.git (refs/heads/master):

On Fri, 16 Apr 2021 21:29:13 +0200 you wrote:
> When I added support to allow generic netlink multicast groups to be
> restricted to subscribers with CAP_NET_ADMIN I was unaware that a
> genl_bind implementation already existed in the past.
> 
> It was reverted due to ABBA deadlock:
> 
> 1. ->netlink_bind gets called with the table lock held.
> 2. genetlink bind callback is invoked, it grabs the genl lock.
> 
> [...]

Here is the summary with links:
  - [net] netlink: don't call ->netlink_bind with table lock held
    https://git.kernel.org/netdev/net/c/f2764bd4f6a8

You are awesome, thank you!
--
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html