[PATCH] pagevec freeze

[Hugh Dickins]

refill_inactive_zone stuck circling the one-page active_list with
lru_lock and interrupts disabled, if concurrent release_pages has
done its put_page_testzero and is now spinning for that lru_lock.

(I haven't done a patch for include/linux/pagevec.h, but when working
on this freeze, I noticed that pagevec_add becomes rather unfriendly
once you get into NMIs and BUGs - overshoots PAGEVEC_SIZE and corrupts
beyond.  Perhaps you'd say all bets are off then anyway, or perhaps
pagevec_space should be signed, and tested before rather than after?)

--- 2.5.42-mm3/mm/vmscan.c	Tue Oct 15 06:43:41 2002
+++ linux/mm/vmscan.c	Tue Oct 15 13:14:35 2002
@@ -542,10 +542,13 @@
 	long mapped_ratio;
 	long distress;
 	long swap_tendency;
+	int max_scan;
 
 	lru_add_drain();
 	spin_lock_irq(&zone->lru_lock);
-	while (nr_pages && !list_empty(&zone->active_list)) {
+	max_scan = zone->nr_active;
+	while (max_scan && nr_pages && !list_empty(&zone->active_list)) {
+		max_scan--;
 		page = list_entry(zone->active_list.prev, struct page, lru);
 		prefetchw_prev_lru_page(page, &zone->active_list, flags);
 		if (!TestClearPageLRU(page))