[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
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -410,7 +410,7 @@
{
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 @@
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;