[PATCH] unmap_vmas latency improvement

unmap_vmas() will cause scheduling latency when tearing down really big vmas
on !CONFIG_PREEMPT.  That's a bit unkind to the non-preempt case, so let's do
a cond_resched() after zapping 1024 pages.

mm/memory.c

@@ -491,9 +491,9 @@ void unmap_page_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
 #define ZAP_BLOCK_SIZE	(256 * PAGE_SIZE)
 #endif
 
-/* No preempt: go for the best straight-line efficiency */
+/* No preempt: go for improved straight-line efficiency */
 #if !defined(CONFIG_PREEMPT)
-#define ZAP_BLOCK_SIZE	(~(0UL))
+#define ZAP_BLOCK_SIZE	(1024 * PAGE_SIZE)
 #endif
 
 /**