According to the comments, this was how it's been done years ago, but
apparently took an xbt pointer from elsewhere back then. The code was
removed because of consistency issues: cancellation wont't roll back
the saved xbdev->state.

Still, unsolicited writes to the state field remain an issue,
especially if device shutdown takes thread synchronization, and subtle
races cause accidental recreation of the device node.

Fixed by reintroducing the transaction. An internal one is sufficient,
so the xbdev->state value remains consistent.

Also fixes the original hack to prevent infinite recursion. Instead of
bailing out on the first attempt to switch to Closing, checks call
depth now.

Signed-off-by: Daniel Stodden <daniel.stodden@xxxxxxxxxx>

diff -r be2785bfda86 -r b863073c3633 drivers/xen/xenbus/xenbus_client.c
--- a/drivers/xen/xenbus/xenbus_client.c	Fri Apr 30 14:58:59 2010 -0700
+++ b/drivers/xen/xenbus/xenbus_client.c	Fri Apr 30 14:58:59 2010 -0700
@@ -134,6 +134,64 @@
 }
 EXPORT_SYMBOL_GPL(xenbus_watch_pathfmt);
 
+static void xenbus_switch_fatal(struct xenbus_device *, int, int,
+				const char *, ...);
+
+static int
+__xenbus_switch_state(struct xenbus_device *dev,
+		      enum xenbus_state state, int depth)
+{
+	/* We check whether the state is currently set to the given value, and
+	   if not, then the state is set.  We don't want to unconditionally
+	   write the given state, because we don't want to fire watches
+	   unnecessarily.  Furthermore, if the node has gone, we don't write
+	   to it, as the device will be tearing down, and we don't want to
+	   resurrect that directory.
+
+	   Note that, because of this cached value of our state, this
+	   function will not take a caller's Xenstore transaction
+	   (something it was trying to in the past) because dev->state
+	   would not get reset if the transaction was aborted.
+	 */
+
+	struct xenbus_transaction xbt;
+	int current_state;
+	int err, abort;
+
+	if (state == dev->state)
+		return 0;
+
+again:
+	abort = 1;
+
+	err = xenbus_transaction_start(&xbt);
+	if (err) {
+		xenbus_switch_fatal(dev, depth, err, "starting transaction");
+		return 0;
+	}
+
+	err = xenbus_scanf(xbt, dev->nodename, "state", "%d", &current_state);
+	if (err != 1)
+		goto abort;
+
+	err = xenbus_printf(xbt, dev->nodename, "state", "%d", state);
+	if (err) {
+		xenbus_switch_fatal(dev, depth, err, "writing new state");
+		goto abort;
+	}
+
+	abort = 0;
+abort:
+	err = xenbus_transaction_end(xbt, abort);
+	if (err) {
+		if (err == -EAGAIN && !abort)
+			goto again;
+		xenbus_switch_fatal(dev, depth, err, "ending transaction");
+	} else
+		dev->state = state;
+
+	return 0;
+}
 
 /**
  * xenbus_switch_state
@@ -146,42 +204,9 @@
  */
 int xenbus_switch_state(struct xenbus_device *dev, enum xenbus_state state)
 {
-	/* We check whether the state is currently set to the given value, and
-	   if not, then the state is set.  We don't want to unconditionally
-	   write the given state, because we don't want to fire watches
-	   unnecessarily.  Furthermore, if the node has gone, we don't write
-	   to it, as the device will be tearing down, and we don't want to
-	   resurrect that directory.
+	return __xenbus_switch_state(dev, state, 0);
+}
 
-	   Note that, because of this cached value of our state, this function
-	   will not work inside a Xenstore transaction (something it was
-	   trying to in the past) because dev->state would not get reset if
-	   the transaction was aborted.
-
-	 */
-
-	int current_state;
-	int err;
-
-	if (state == dev->state)
-		return 0;
-
-	err = xenbus_scanf(XBT_NIL, dev->nodename, "state", "%d",
-			   &current_state);
-	if (err != 1)
-		return 0;
-
-	err = xenbus_printf(XBT_NIL, dev->nodename, "state", "%d", state);
-	if (err) {
-		if (state != XenbusStateClosing) /* Avoid looping */
-			xenbus_dev_fatal(dev, err, "writing new state");
-		return err;
-	}
-
-	dev->state = state;
-
-	return 0;
-}
 EXPORT_SYMBOL_GPL(xenbus_switch_state);
 
 int xenbus_frontend_closed(struct xenbus_device *dev)
@@ -285,6 +310,23 @@
 EXPORT_SYMBOL_GPL(xenbus_dev_fatal);
 
 /**
+ * Equivalent to xenbus_dev_fatal(dev, err, fmt, args), but helps
+ * avoiding recursion within xenbus_switch_state.
+ */
+static void xenbus_switch_fatal(struct xenbus_device *dev, int depth, int err,
+				const char *fmt, ...)
+{
+	va_list ap;
+
+	va_start(ap, fmt);
+	xenbus_va_dev_error(dev, err, fmt, ap);
+	va_end(ap);
+
+	if (!depth)
+		__xenbus_switch_state(dev, XenbusStateClosing, 1);
+}
+
+/**
  * xenbus_grant_ring
  * @dev: xenbus device
  * @ring_mfn: mfn of ring to grant