]> xenbits.xensource.com Git - legacy/linux-2.6.18-xen.git/commitdiff
xenbus: do not hold transaction_mutex when returning to userspace xen-3.4.2
authorKeir Fraser <keir.fraser@citrix.com>
Wed, 4 Nov 2009 18:13:32 +0000 (18:13 +0000)
committerKeir Fraser <keir.fraser@citrix.com>
Wed, 4 Nov 2009 18:13:32 +0000 (18:13 +0000)
  ================================================
  [ BUG: lock held when returning to user space! ]
  ------------------------------------------------
  xenstore-list/3522 is leaving the kernel with locks still held!
  1 lock held by xenstore-list/3522:
   #0:  (&xs_state.transaction_mutex){......}, at: [<c026dc6f>]
   xenbus_dev_request_and_reply+0x8f/0xa0

The canonical fix for this type of issue appears to be to maintain a
count manually rather than using an rwsem so do that here.

Signed-off-by: Ian Campbell <ian.campbell@citrix.com>
drivers/xen/xenbus/xenbus_xs.c

index c51bbd03ea577fccc2e476040994ccc57ed85d0f..99a8614c7aad9c018e1fd687508c1e1e5f3ba8d0 100644 (file)
@@ -84,6 +84,14 @@ struct xs_handle {
        /*
         * Mutex ordering: transaction_mutex -> watch_mutex -> request_mutex.
         * response_mutex is never taken simultaneously with the other three.
+        *
+        * transaction_mutex must be held before incrementing
+        * transaction_count. The mutex is held when a suspend is in
+        * progress to prevent new transactions starting.
+        *
+        * When decrementing transaction_count to zero the wait queue
+        * should be woken up, the suspend code waits for count to
+        * reach zero.
         */
 
        /* One request at a time. */
@@ -93,7 +101,9 @@ struct xs_handle {
        struct mutex response_mutex;
 
        /* Protect transactions against save/restore. */
-       struct rw_semaphore transaction_mutex;
+       struct mutex transaction_mutex;
+       atomic_t transaction_count;
+       wait_queue_head_t transaction_wq;
 
        /* Protect watch (de)register against save/restore. */
        struct rw_semaphore watch_mutex;
@@ -165,6 +175,31 @@ static void *read_reply(enum xsd_sockmsg_type *type, unsigned int *len)
        return body;
 }
 
+static void transaction_start(void)
+{
+       mutex_lock(&xs_state.transaction_mutex);
+       atomic_inc(&xs_state.transaction_count);
+       mutex_unlock(&xs_state.transaction_mutex);
+}
+
+static void transaction_end(void)
+{
+       if (atomic_dec_and_test(&xs_state.transaction_count))
+               wake_up(&xs_state.transaction_wq);
+}
+
+static void transaction_suspend(void)
+{
+       mutex_lock(&xs_state.transaction_mutex);
+       wait_event(xs_state.transaction_wq,
+                  atomic_read(&xs_state.transaction_count) == 0);
+}
+
+static void transaction_resume(void)
+{
+       mutex_unlock(&xs_state.transaction_mutex);
+}
+
 void *xenbus_dev_request_and_reply(struct xsd_sockmsg *msg)
 {
        void *ret;
@@ -172,7 +207,7 @@ void *xenbus_dev_request_and_reply(struct xsd_sockmsg *msg)
        int err;
 
        if (req_msg.type == XS_TRANSACTION_START)
-               down_read(&xs_state.transaction_mutex);
+               transaction_start();
 
        mutex_lock(&xs_state.request_mutex);
 
@@ -188,7 +223,7 @@ void *xenbus_dev_request_and_reply(struct xsd_sockmsg *msg)
        if ((req_msg.type == XS_TRANSACTION_END) ||
            ((req_msg.type == XS_TRANSACTION_START) &&
             (msg->type == XS_ERROR)))
-               up_read(&xs_state.transaction_mutex);
+               transaction_end();
 
        return ret;
 }
@@ -440,11 +475,11 @@ int xenbus_transaction_start(struct xenbus_transaction *t)
 {
        char *id_str;
 
-       down_read(&xs_state.transaction_mutex);
+       transaction_start();
 
        id_str = xs_single(XBT_NIL, XS_TRANSACTION_START, "", NULL);
        if (IS_ERR(id_str)) {
-               up_read(&xs_state.transaction_mutex);
+               transaction_end();
                return PTR_ERR(id_str);
        }
 
@@ -469,7 +504,7 @@ int xenbus_transaction_end(struct xenbus_transaction t, int abort)
 
        err = xs_error(xs_single(t, XS_TRANSACTION_END, abortstr, NULL));
 
-       up_read(&xs_state.transaction_mutex);
+       transaction_end();
 
        return err;
 }
@@ -670,7 +705,7 @@ EXPORT_SYMBOL_GPL(unregister_xenbus_watch);
 
 void xs_suspend(void)
 {
-       down_write(&xs_state.transaction_mutex);
+       transaction_suspend();
        down_write(&xs_state.watch_mutex);
        mutex_lock(&xs_state.request_mutex);
        mutex_lock(&xs_state.response_mutex);
@@ -683,7 +718,7 @@ void xs_resume(void)
 
        mutex_unlock(&xs_state.response_mutex);
        mutex_unlock(&xs_state.request_mutex);
-       up_write(&xs_state.transaction_mutex);
+       transaction_resume();
 
        /* No need for watches_lock: the watch_mutex is sufficient. */
        list_for_each_entry(watch, &watches, list) {
@@ -699,7 +734,7 @@ void xs_suspend_cancel(void)
        mutex_unlock(&xs_state.response_mutex);
        mutex_unlock(&xs_state.request_mutex);
        up_write(&xs_state.watch_mutex);
-       up_write(&xs_state.transaction_mutex);
+       mutex_unlock(&xs_state.transaction_mutex);
 }
 
 static int xenwatch_handle_callback(void *data)
@@ -880,8 +915,10 @@ int xs_init(void)
 
        mutex_init(&xs_state.request_mutex);
        mutex_init(&xs_state.response_mutex);
-       init_rwsem(&xs_state.transaction_mutex);
+       mutex_init(&xs_state.transaction_mutex);
        init_rwsem(&xs_state.watch_mutex);
+       atomic_set(&xs_state.transaction_count, 0);
+       init_waitqueue_head(&xs_state.transaction_wq);
 
        task = kthread_run(xenwatch_thread, NULL, "xenwatch");
        if (IS_ERR(task))