2009-09-02 00:58:01

by Hugh Dickins

Subject: [PATCH mmotm] vmscan move pgdeactivate modification to shrink_active_list fix2

A second fix to the ill-starred
vmscan-move-pgdeactivate-modification-to-shrink_active_list.patch
which, once corrected to update the right counters by the first fix,
builds up absurdly large Active counts in /proc/meminfo.

nr_rotated is not the number of pages added back to the active list
(maybe it once was, maybe it should be again: but if so that's not
any business for a code rearrangement patch). shrink_active_list()
needs to keep a separate nr_reactivated count of those.

Signed-off-by: Hugh Dickins <[email protected]>
---
Or... revert the offending patch and its first fix.

mm/vmscan.c | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)

--- mmotm/mm/vmscan.c 2009-08-28 18:30:33.000000000 +0100
+++ linux/mm/vmscan.c 2009-09-02 01:28:34.000000000 +0100
@@ -1306,6 +1306,7 @@ static void shrink_active_list(unsigned
struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
unsigned long nr_rotated = 0;
unsigned long nr_deactivated = 0;
+ unsigned long nr_reactivated = 0;

lru_add_drain();
spin_lock_irq(&zone->lru_lock);
@@ -1354,6 +1355,7 @@ static void shrink_active_list(unsigned
*/
if ((vm_flags & VM_EXEC) && !PageAnon(page)) {
list_add(&page->lru, &l_active);
+ nr_reactivated++;
continue;
}
}
@@ -1382,7 +1384,7 @@ static void shrink_active_list(unsigned
__count_vm_events(PGDEACTIVATE, nr_deactivated);
__mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
__mod_zone_page_state(zone, NR_ACTIVE_ANON + file * LRU_FILE,
- nr_rotated);
+ nr_reactivated);
__mod_zone_page_state(zone, NR_INACTIVE_ANON + file * LRU_FILE,
nr_deactivated);
spin_unlock_irq(&zone->lru_lock);

2009-09-02 01:30:28

by KOSAKI Motohiro

Subject: Re: [PATCH mmotm] vmscan move pgdeactivate modification to shrink_active_list fix2

> A second fix to the ill-starred
> vmscan-move-pgdeactivate-modification-to-shrink_active_list.patch
> which, once corrected to update the right counters by the first fix,
> builds up absurdly large Active counts in /proc/meminfo.
>
> nr_rotated is not the number of pages added back to the active list
> (maybe it once was, maybe it should be again: but if so that's not
> any business for a code rearrangement patch). shrink_active_list()
> needs to keep a separate nr_reactivated count of those.
>
> Signed-off-by: Hugh Dickins <[email protected]>
> ---
> Or... revert the offending patch and its first fix.

Yes, The original patch author should be layoff :(

Andrew, can you please drop the patch?
I need to clean my brain and the patch need proper additional test.

I plan to resubmit it at next -rc2 or -rc3.


I'm sorry.