There are two places where we reclaim free pages, but we never
update the all_unreclaimable flag for the relevant zone.
This patch helped reduce the frequency of oom-kills under high load
for us a while back, and we've been carrying it since.
I posted this a few months back and from what I recall it didn't
get a great deal of interest.
Signed-off-by: Dave Jones <[email protected]>
--- linux-2.6/mm/filemap.c~ 2005-12-10 01:47:15.000000000 -0500
+++ linux-2.6/mm/filemap.c 2005-12-10 01:47:46.000000000 -0500
@@ -471,11 +471,18 @@ EXPORT_SYMBOL(unlock_page);
*/
void end_page_writeback(struct page *page)
{
+ struct zone *zone = page_zone(page);
if (!TestClearPageReclaim(page) || rotate_reclaimable_page(page)) {
if (!test_clear_page_writeback(page))
BUG();
}
smp_mb__after_clear_bit();
+ if (zone->all_unreclaimable) {
+ spin_lock(&zone->lock);
+ zone->all_unreclaimable = 0;
+ zone->pages_scanned = 0;
+ spin_unlock(&zone->lock);
+ }
wake_up_page(page, PG_writeback);
}
EXPORT_SYMBOL(end_page_writeback);
--- linux-2.6/mm/page_alloc.c~ 2006-01-09 13:40:03.000000000 -0500
+++ linux-2.6/mm/page_alloc.c 2006-01-09 13:40:50.000000000 -0500
@@ -722,6 +722,11 @@ static void fastcall free_hot_cold_page(
if (pcp->count >= pcp->high) {
free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
pcp->count -= pcp->batch;
+ } else if (zone->all_unreclaimable) {
+ spin_lock(&zone->lock);
+ zone->all_unreclaimable = 0;
+ zone->pages_scanned = 0;
+ spin_unlock(&zone->lock);
}
local_irq_restore(flags);
put_cpu();
--
http://www.codemonkey.org.uk
On Tue, 1 Aug 2006 13:57:16 -0400
Dave Jones <[email protected]> wrote:
> There are two places where we reclaim free pages, but we never
> update the all_unreclaimable flag for the relevant zone.
> This patch helped reduce the frequency of oom-kills under high load
> for us a while back, and we've been carrying it since.
> I posted this a few months back and from what I recall it didn't
> get a great deal of interest.
>
> Signed-off-by: Dave Jones <[email protected]>
>
> --- linux-2.6/mm/filemap.c~ 2005-12-10 01:47:15.000000000 -0500
> +++ linux-2.6/mm/filemap.c 2005-12-10 01:47:46.000000000 -0500
> @@ -471,11 +471,18 @@ EXPORT_SYMBOL(unlock_page);
> */
> void end_page_writeback(struct page *page)
> {
> + struct zone *zone = page_zone(page);
> if (!TestClearPageReclaim(page) || rotate_reclaimable_page(page)) {
> if (!test_clear_page_writeback(page))
> BUG();
> }
> smp_mb__after_clear_bit();
> + if (zone->all_unreclaimable) {
> + spin_lock(&zone->lock);
> + zone->all_unreclaimable = 0;
> + zone->pages_scanned = 0;
> + spin_unlock(&zone->lock);
> + }
> wake_up_page(page, PG_writeback);
> }
We're not actually freeing the page here though. We _might_ be making it
reclaimable, but we don't know that. If page reclaim later _does_ reclaim
the page, then ->all_unreclaimable will get cleared then. A little later,
but if that's enough to make a difference, I suspect we're already rather
doomed.
Also, if rotate_reclaimable_page() returned non-zero we know this page
isn't immediately reclaimable, so the patch shouldn't be clearing
->all_unreclaimable in that case.
And zone->lock is supposed to be irq-safe. Yes, we're in an interrupt, but
many different interrupt sources will vector into this code - if we take an
interrupt from /dev/sda while serving a /dev/hda interrupt we'll deadlock
here. File a bug against lockdep ;)
> EXPORT_SYMBOL(end_page_writeback);
> --- linux-2.6/mm/page_alloc.c~ 2006-01-09 13:40:03.000000000 -0500
> +++ linux-2.6/mm/page_alloc.c 2006-01-09 13:40:50.000000000 -0500
> @@ -722,6 +722,11 @@ static void fastcall free_hot_cold_page(
> if (pcp->count >= pcp->high) {
> free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
> pcp->count -= pcp->batch;
> + } else if (zone->all_unreclaimable) {
> + spin_lock(&zone->lock);
> + zone->all_unreclaimable = 0;
> + zone->pages_scanned = 0;
> + spin_unlock(&zone->lock);
> }
> local_irq_restore(flags);
> put_cpu();
free_pages_bulk() will clear ->all_unreclaimable. If this really makes a
difference then we're already skating along the raggedy edge.
There is a string of oom-killer patches in -mm from Nick. I suspect we'll
be in much better shape after those are merged.