smb: client: Fix netns refcount imbalance causing leaks and use-after-free

Commit ef7134c ("smb: client: Fix use-after-free of network
namespace.") attempted to fix a netns use-after-free issue by manually
adjusting reference counts via sk->sk_net_refcnt and sock_inuse_add().

However, a later commit e9f2517 ("smb: client: fix TCP timers deadlock
after rmmod") pointed out that the approach of manually setting
sk->sk_net_refcnt in the first commit was technically incorrect, as
sk->sk_net_refcnt should only be set for user sockets. It led to issues
like TCP timers not being cleared properly on close. The second commit
moved to a model of just holding an extra netns reference for
server->ssocket using get_net(), and dropping it when the server is torn
down.

But there remain some gaps in the get_net()/put_net() balancing added by
these commits. The incomplete reference handling in these fixes results
in two issues:

1. Netns refcount leaks[1]

The problem process is as follows:

```
mount.cifs                        cifsd

cifs_do_mount
  cifs_mount
    cifs_mount_get_session
      cifs_get_tcp_session
        get_net()  /* First get net. */
        ip_connect
          generic_ip_connect /* Try port 445 */
            get_net()
            ->connect() /* Failed */
            put_net()
          generic_ip_connect /* Try port 139 */
            get_net() /* Missing matching put_net() for this get_net().*/
      cifs_get_smb_ses
        cifs_negotiate_protocol
          smb2_negotiate
            SMB2_negotiate
              cifs_send_recv
                wait_for_response
                                 cifs_demultiplex_thread
                                   cifs_read_from_socket
                                     cifs_readv_from_socket
                                       cifs_reconnect
                                         cifs_abort_connection
                                           sock_release();
                                           server->ssocket = NULL;
                                           /* Missing put_net() here. */
                                           generic_ip_connect
                                             get_net()
                                             ->connect() /* Failed */
                                             put_net()
                                             sock_release();
                                             server->ssocket = NULL;
          free_rsp_buf
    ...
                                   clean_demultiplex_info
                                     /* It's only called once here. */
                                     put_net()
```

When cifs_reconnect() is triggered, the server->ssocket is released
without a corresponding put_net() for the reference acquired in
generic_ip_connect() before. it ends up calling generic_ip_connect()
again to retry get_net(). After that, server->ssocket is set to NULL
in the error path of generic_ip_connect(), and the net count cannot be
released in the final clean_demultiplex_info() function.

2. Potential use-after-free

The current refcounting scheme can lead to a potential use-after-free issue
in the following scenario:

```
 cifs_do_mount
   cifs_mount
     cifs_mount_get_session
       cifs_get_tcp_session
         get_net()  /* First get net */
           ip_connect
             generic_ip_connect
               get_net()
               bind_socket
	         kernel_bind /* failed */
               put_net()
         /* after out_err_crypto_release label */
         put_net()
         /* after out_err label */
         put_net()
```

In the exception handling process where binding the socket fails, the
get_net() and put_net() calls are unbalanced, which may cause the
server->net reference count to drop to zero and be prematurely released.

To address both issues, this patch ties the netns reference counting to
the server->ssocket and server lifecycles. The extra reference is now
acquired when the server or socket is created, and released when the
socket is destroyed or the server is torn down.

[1]: https://bugzilla.kernel.org/show_bug.cgi?id=219792

Fixes: ef7134c ("smb: client: Fix use-after-free of network namespace.")
Fixes: e9f2517 ("smb: client: fix TCP timers deadlock after rmmod")
Signed-off-by: Wang Zhaolong <wangzhaolong1@huawei.com>
Signed-off-by: Steve French <stfrench@microsoft.com>

Author: Wang Zhaolong

fs/smb/client/connect.c

@@ -300,6 +300,7 @@ cifs_abort_connection(struct TCP_Server_Info *server)
 			 server->ssocket->flags);
 		sock_release(server->ssocket);
 		server->ssocket = NULL;
+		put_net(cifs_net_ns(server));
 	}
 	server->sequence_number = 0;
 	server->session_estab = false;
@@ -3265,8 +3266,12 @@ generic_ip_connect(struct TCP_Server_Info *server)
 		/*
 		 * Grab netns reference for the socket.
 		 *
-		 * It'll be released here, on error, or in clean_demultiplex_info() upon server
-		 * teardown.
+		 * This reference will be released in several situations:
+		 * - In the failure path before the cifsd thread is started.
+		 * - In the all place where server->socket is released, it is
+		 *   also set to NULL.
+		 * - Ultimately in clean_demultiplex_info(), during the final
+		 *   teardown.
 		 */
 		get_net(net);
 
@@ -3282,10 +3287,8 @@ generic_ip_connect(struct TCP_Server_Info *server)
 	}
 
 	rc = bind_socket(server);
-	if (rc < 0) {
-		put_net(cifs_net_ns(server));
+	if (rc < 0)
 		return rc;
-	}
 
 	/*
 	 * Eventually check for other socket options to change from
@@ -3331,9 +3334,6 @@ generic_ip_connect(struct TCP_Server_Info *server)
 	if (sport == htons(RFC1001_PORT))
 		rc = ip_rfc1001_connect(server);
 
-	if (rc < 0)
-		put_net(cifs_net_ns(server));
-
 	return rc;
 }