clear vm_reclaimable when we free pages.
Subsystems (3)
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.
mm/filemap.c renamed from mm/filemap.c~ +7 -0
@@ -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);
mm/page_alloc.c renamed from mm/page_alloc.c~ +5 -0
@@ -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 ();
Signature · 1 line
http://www.codemonkey.org.uk
Quoted 19 lines · 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_unreclaimableflag 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]>mm/filemap.c renamed from mm/filemap.c~ +1
-0@@ -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 ();
mm/filemap.c EXPORT_SYMBOL(unlock_page); · line 487
Quoted 10 lines
mm/filemap.c renamed from mm/filemap.c~ +6
-0}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 ;)
Quoted 5 lines
mm/filemap.c renamed from mm/filemap.c~ +0
-0EXPORT_SYMBOL (end_page_writeback);mm/page_alloc.c renamed from mm/page_alloc.c~ +0
-0@@ -722,6 +722,11 @@ static void fastcall free_hot_cold_page(if (pcp->count >= pcp->high) {
mm/page_alloc.c static void fastcall free_hot_cold_page( · line 732
Quoted 10 lines
mm/page_alloc.c renamed from mm/page_alloc.c~ +5
-0free_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.