[PATCH] mm: copy_pte_range progress fix

My latency breaking in copy_pte_range didn't work as intended: instead of
checking at regularish intervals, after the first interval it checked every
time around the loop, too impatient to be preempted.  Fix that.

Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

mm/memory.c

@@ -410,25 +410,27 @@ static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 {
 	pte_t *src_pte, *dst_pte;
 	unsigned long vm_flags = vma->vm_flags;
-	int progress;
+	int progress = 0;
 
 again:
 	dst_pte = pte_alloc_map(dst_mm, dst_pmd, addr);
 	if (!dst_pte)
 		return -ENOMEM;
 	src_pte = pte_offset_map_nested(src_pmd, addr);
 
-	progress = 0;
 	spin_lock(&src_mm->page_table_lock);
 	do {
 		/*
 		 * We are holding two locks at this point - either of them
 		 * could generate latencies in another task on another CPU.
 		 */
-		if (progress >= 32 && (need_resched() ||
-		    need_lockbreak(&src_mm->page_table_lock) ||
-		    need_lockbreak(&dst_mm->page_table_lock)))
-			break;
+		if (progress >= 32) {
+			progress = 0;
+			if (need_resched() ||
+			    need_lockbreak(&src_mm->page_table_lock) ||
+			    need_lockbreak(&dst_mm->page_table_lock))
+				break;
+		}
 		if (pte_none(*src_pte)) {
 			progress++;
 			continue;