cifs: fix race between call_async() and reconnect()
cifs_call_async() queues the MID to the pending list and calls
smb_send_rqst(). If smb_send_rqst() performs a partial send, it sets
the tcpStatus to CifsNeedReconnect and returns an error code to
cifs_call_async(). In this case, cifs_call_async() removes the MID
from the list and returns to the caller.
However, cifs_call_async() releases the server mutex _before_ removing
the MID. This means that a cifs_reconnect() can race with this function
and manage to remove the MID from the list and delete the entry before
cifs_call_async() calls cifs_delete_mid(). This leads to various
crashes due to the use after free in cifs_delete_mid().
Task1 Task2
cifs_call_async():
- rc = -EAGAIN
- mutex_unlock(srv_mutex)
cifs_reconnect():
- mutex_lock(srv_mutex)
- mutex_unlock(srv_mutex)
- list_delete(mid)
- mid->callback()
cifs_writev_callback():
- mutex_lock(srv_mutex)
- delete(mid)
- mutex_unlock(srv_mutex)
- cifs_delete_mid(mid) <---- use after free
Fix this by removing the MID in cifs_call_async() before releasing the
srv_mutex. Also hold the srv_mutex in cifs_reconnect() until the MIDs
are moved out of the pending list.
Signed-off-by: Rabin Vincent <rabin.vincent@axis.com>
Acked-by: Shirish Pargaonkar <shirishpargaonkar@gmail.com>
CC: Stable <stable@vger.kernel.org>
Signed-off-by: Steve French <sfrench@localhost.localdomain>
diff --git a/fs/cifs/transport.c b/fs/cifs/transport.c
index 2a24c52..87abe8e 100644
--- a/fs/cifs/transport.c
+++ b/fs/cifs/transport.c
@@ -576,14 +576,16 @@
cifs_in_send_dec(server);
cifs_save_when_sent(mid);
- if (rc < 0)
+ if (rc < 0) {
server->sequence_number -= 2;
+ cifs_delete_mid(mid);
+ }
+
mutex_unlock(&server->srv_mutex);
if (rc == 0)
return 0;
- cifs_delete_mid(mid);
add_credits_and_wake_if(server, credits, optype);
return rc;
}