SUNRPC: fix race to sk_err after xs_error_report

Since commit 4f8943f8 ("SUNRPC: Replace direct task wakeups from
softirq context") there has been a race to the value of the sk_err if both
XPRT_SOCK_WAKE_ERROR and XPRT_SOCK_WAKE_DISCONNECT are set.  In that case,
we may end up losing the sk_err value that existed when xs_error_report was
called.

Fix this by reverting to the previous behavior: instead of using SO_ERROR
to retrieve the value at a later time (which might also return sk_err_soft),
copy the sk_err value onto struct sock_xprt, and use that value to wake
pending tasks.
Signed-off-by: NBenjamin Coddington <bcodding@redhat.com>
Fixes: 4f8943f8 ("SUNRPC: Replace direct task wakeups from softirq context")
Signed-off-by: NAnna Schumaker <Anna.Schumaker@Netapp.com>

include/linux/sunrpc/xprtsock.h

@@ -61,6 +61,7 @@ struct sock_xprt {
 	struct mutex		recv_mutex;
 	struct sockaddr_storage	srcaddr;
 	unsigned short		srcport;
+	int			xprt_err;
 
 	/*
 	 * UDP socket buffer size parameters

net/sunrpc/xprtsock.c

@@ -1249,19 +1249,21 @@ static void xs_error_report(struct sock *sk)
 {
 	struct sock_xprt *transport;
 	struct rpc_xprt *xprt;
-	int err;
 
 	read_lock_bh(&sk->sk_callback_lock);
 	if (!(xprt = xprt_from_sock(sk)))
 		goto out;
 
 	transport = container_of(xprt, struct sock_xprt, xprt);
-	err = -sk->sk_err;
-	if (err == 0)
+	transport->xprt_err = -sk->sk_err;
+	if (transport->xprt_err == 0)
 		goto out;
 	dprintk("RPC:       xs_error_report client %p, error=%d...\n",
-			xprt, -err);
-	trace_rpc_socket_error(xprt, sk->sk_socket, err);
+			xprt, -transport->xprt_err);
+	trace_rpc_socket_error(xprt, sk->sk_socket, transport->xprt_err);
+
+	/* barrier ensures xprt_err is set before XPRT_SOCK_WAKE_ERROR */
+	smp_mb__before_atomic();
 	xs_run_error_worker(transport, XPRT_SOCK_WAKE_ERROR);
  out:
 	read_unlock_bh(&sk->sk_callback_lock);
@@ -2476,7 +2478,6 @@ static void xs_wake_write(struct sock_xprt *transport)
 static void xs_wake_error(struct sock_xprt *transport)
 {
 	int sockerr;
-	int sockerr_len = sizeof(sockerr);
 
 	if (!test_bit(XPRT_SOCK_WAKE_ERROR, &transport->sock_state))
 		return;
@@ -2485,9 +2486,7 @@ static void xs_wake_error(struct sock_xprt *transport)
 		goto out;
 	if (!test_and_clear_bit(XPRT_SOCK_WAKE_ERROR, &transport->sock_state))
 		goto out;
-	if (kernel_getsockopt(transport->sock, SOL_SOCKET, SO_ERROR,
-				(char *)&sockerr, &sockerr_len) != 0)
-		goto out;
+	sockerr = xchg(&transport->xprt_err, 0);
 	if (sockerr < 0)
 		xprt_wake_pending_tasks(&transport->xprt, sockerr);
 out: