[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>

diff --git a/mm/memory.c b/mm/memory.c
index 1db40e9..222c13e 100644 (file)
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -410,7 +410,7 @@ 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);
@@ -418,17 +418,19 @@ again:
                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;