idr: make idr_remove_all() do removal -before- free_layer()

Fix a problem in the IDR system, where an idr_remove_all() hands a data
element to call_rcu() (via free_layer()) before making that data element
inaccessible to new readers.  This is very bad, and results in readers
still having a reference to this data element at the end of the grace
period.

Tests on large machines that concurrently map and unmap user-space memory
within the same multithreaded process result in crashes within about five
minutes.  Applying this patch increases the kernel's longevity to the
three-to-eight-hour range.

There appear to be other similar problems in idr_get_empty_slot() and
sub_remove(), but I fixed the easy one in idr_remove_all() first.  It is
therefore no surprise that failures still occur.

Located-by: Milton Miller II <miltonm@austin.ibm.com>
Tested-by: Milton Miller II <miltonm@austin.ibm.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Manfred Spraul <manfred@colorfullife.com>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

lib/idr.c

@@ -449,6 +449,7 @@ void idr_remove_all(struct idr *idp)
 
 	n = idp->layers * IDR_BITS;
 	p = idp->top;
+	rcu_assign_pointer(idp->top, NULL);
 	max = 1 << n;
 
 	id = 0;
@@ -467,7 +468,6 @@ void idr_remove_all(struct idr *idp)
 			p = *--paa;
 		}
 	}
-	rcu_assign_pointer(idp->top, NULL);
 	idp->layers = 0;
 }
 EXPORT_SYMBOL(idr_remove_all);