Comment # 4 on bug 997639 from Neil Brown

That patch is imperfect.
+        schedule_timeout(HZ/10);
should be
+        schedule_timeout_uninterruptible(HZ/10);

There is a similar bug in fscache_object_sleep_till_congested()

A quick look suggests the problem goes back quite a way, probably at least to 
Commit: 97e7449a7ad8 ("autofs4: fix indirect mount pending expire race")
in 2.6.27.

So I probably need to examine each kernel and see if there is a problem to fix,
rather than trying to backport.