nfsd: Fix slot wake up race in the nfsv4.1 callback code
authorTrond Myklebust <trond.myklebust@primarydata.com>
Wed, 19 Nov 2014 17:47:50 +0000 (12:47 -0500)
committerBen Hutchings <ben@decadent.org.uk>
Sun, 14 Dec 2014 16:23:58 +0000 (16:23 +0000)
commit c6c15e1ed303ffc47e696ea1c9a9df1761c1f603 upstream.

The currect code for nfsd41_cb_get_slot() and nfsd4_cb_done() has no
locking in order to guarantee atomicity, and so allows for races of
the form.

Task 1                                  Task 2
======                                  ======
if (test_and_set_bit(0) != 0) {
                                        clear_bit(0)
                                        rpc_wake_up_next(queue)
        rpc_sleep_on(queue)
        return false;
}

This patch breaks the race condition by adding a retest of the bit
after the call to rpc_sleep_on().

Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Signed-off-by: Ben Hutchings <ben@decadent.org.uk>
fs/nfsd/nfs4callback.c

index 809a38a..51b4f43 100644 (file)
@@ -785,8 +785,12 @@ static bool nfsd41_cb_get_slot(struct nfs4_client *clp, struct rpc_task *task)
 {
        if (test_and_set_bit(0, &clp->cl_cb_slot_busy) != 0) {
                rpc_sleep_on(&clp->cl_cb_waitq, task, NULL);
-               dprintk("%s slot is busy\n", __func__);
-               return false;
+               /* Race breaker */
+               if (test_and_set_bit(0, &clp->cl_cb_slot_busy) != 0) {
+                       dprintk("%s slot is busy\n", __func__);
+                       return false;
+               }
+               rpc_wake_up_queued_task(&clp->cl_cb_waitq, task);
        }
        return true;
 }