Because of the batching logic in shrink_zone() it turns out that the
high-level `priority' variable is not a reliable indicator of the amount of
scanning stress.
This is because with small inactive lists, `priority' can pass through quite
a number of decrements before we do _any_ scanning at all: instead, we've
simply been increasing the value of zone->nr_inactive_scan, waiting for it to
exceed SWAP_CLUSTER_MAX, at which point we do some scanning for real.
So the patch stops using `priority' as a measure of scanning distress and uses
(total pages scanned) / (total pages reclaimed)
instead. The code at present fairly arbitrarily decides that if this exceeds
1.5, reclaim is struggling and we need to throttle.
---
25-akpm/mm/vmscan.c | 17 +++++++++--------
1 files changed, 9 insertions(+), 8 deletions(-)
diff -puN mm/vmscan.c~vmscan-less-sleepiness mm/vmscan.c
--- 25/mm/vmscan.c~vmscan-less-sleepiness 2004-03-25 22:25:48.367046584 -0800
+++ 25-akpm/mm/vmscan.c 2004-03-25 22:27:34.007986720 -0800
@@ -876,11 +876,9 @@ int try_to_free_pages(struct zone **zone
if (total_scanned > SWAP_CLUSTER_MAX + SWAP_CLUSTER_MAX/2) {
wakeup_bdflush(laptop_mode ? 0 : total_scanned);
do_writepage = 1;
+ blk_congestion_wait(WRITE, HZ/10);
}
- /* Take a nap, wait for some writeback to complete */
- if (scanned && priority < DEF_PRIORITY - 2)
- blk_congestion_wait(WRITE, HZ/10);
}
if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY))
out_of_memory();
@@ -1009,12 +1007,15 @@ scan:
continue; /* swsusp: need to do more work */
if (all_zones_ok)
break; /* kswapd: all done */
- /*
- * OK, kswapd is getting into trouble. Take a nap, then take
- * another pass across the zones.
- */
- if (total_scanned && priority < DEF_PRIORITY - 2)
+ if (do_writepage) {
+ /*
+ * OK, kswapd is getting into trouble. Take a nap,
+ * then take another pass across the zones. We reuse
+ * do_writepage here, just because it happens to be
+ * exactly what we want...
+ */
blk_congestion_wait(WRITE, HZ/10);
+ }
}
out:
for (i = 0; i < pgdat->nr_zones; i++) {
_