cifs: avoid deadlocks while updating iface

We use cifs_tcp_ses_lock to protect a lot of things.
Not only does it protect the lists of connections, sessions,
tree connects, open file lists, etc., we also use it to
protect some fields in each of it's entries.

In this case, cifs_mark_ses_for_reconnect takes the
cifs_tcp_ses_lock to traverse the lists, and then calls
cifs_update_iface. However, that can end up calling
cifs_put_tcp_session, which picks up the same lock again.

Avoid this by taking a ref for the session, drop the lock,
and then call update iface.

Also, in cifs_update_iface, avoid nested locking of iface_lock
and chan_lock, as much as possible. When unavoidable, we need
to pick iface_lock first.

Signed-off-by: Shyam Prasad N <sprasad@microsoft.com>
Signed-off-by: Steve French <stfrench@microsoft.com>

1 files changed, 12 insertions, 3 deletions

diff --git a/fs/cifs/connect.c b/fs/cifs/connect.c
index e666d2643ede..8d56325915d0 100644
--- a/fs/cifs/connect.c
+++ b/fs/cifs/connect.c
@@ -236,7 +236,7 @@ cifs_mark_tcp_ses_conns_for_reconnect(struct TCP_Server_Info *server,
 				      bool mark_smb_session)
 {
 	struct TCP_Server_Info *pserver;
-	struct cifs_ses *ses;
+	struct cifs_ses *ses, *nses;
 	struct cifs_tcon *tcon;
 
 	/*
@@ -250,10 +250,19 @@ cifs_mark_tcp_ses_conns_for_reconnect(struct TCP_Server_Info *server,
 
 
 	spin_lock(&cifs_tcp_ses_lock);
-	list_for_each_entry(ses, &pserver->smb_ses_list, smb_ses_list) {
+	list_for_each_entry_safe(ses, nses, &pserver->smb_ses_list, smb_ses_list) {
 		/* check if iface is still active */
-		if (!cifs_chan_is_iface_active(ses, server))
+		if (!cifs_chan_is_iface_active(ses, server)) {
+			/*
+			 * HACK: drop the lock before calling
+			 * cifs_chan_update_iface to avoid deadlock
+			 */
+			ses->ses_count++;
+			spin_unlock(&cifs_tcp_ses_lock);
 			cifs_chan_update_iface(ses, server);
+			spin_lock(&cifs_tcp_ses_lock);
+			ses->ses_count--;
+		}
 
 		spin_lock(&ses->chan_lock);
 		if (!mark_smb_session && cifs_chan_needs_reconnect(ses, server))