login
Search
PG_swapbacked flag of newpage should be set(if needed) before
mem_cgroup_prepare_migration, because mem_cgroup_charge_common
checks the flag and determines whether it sets PAGE_CGROUP_FLAG_FILE or not.
Before this patch, if migrating shmem/tmpfs pages, newpage would be
charged with PAGE_CGROUP_FLAG_FILE set, while oldpage has been charged
without the flag.
The problem here is mem_cgroup_move_lists doesn't clear(or set)
the PAGE_CGROUP_FLAG_FILE flag, so pc->flags of the newpage
remains PAGE_CGROUP_FLAG_FILE set even when the pc is moved to
another lru(anon) by mem_cgroup_move_lists. And this leads to
incorrect MEM_CGROUP_ZSTAT.
(In my test, I see an underflow of MEM_CGROUP_ZSTAT(active_file).
As a result, mem_cgroup_calc_reclaim returns very huge number and
causes soft lockup on page reclaim.)
I'm not sure if mem_cgroup_move_lists should handle PAGE_CGROUP_FLAG_FILE
or not(I suppose it should be used to move between active <-> inactive,
not anon <-> file), I moved SetPageSwapBacked(newpage) before
mem_cgroup_prepare_migration.
Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
---
mm/migrate.c | 4 ++--
1 files changed, 2 insertions(+), 2 deletions(-)
diff --git a/mm/migrate.c b/mm/migrate.c
index 577d481..7343463 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -586,8 +586,6 @@ static int move_to_new_page(struct page *newpage, struct page *page)
/* Prepare mapping for the new page.*/
newpage->index = page->index;
newpage->mapping = page->mapping;
- if (PageSwapBacked(page))
- SetPageSwapBacked(newpage);
mapping = page_mapping(page);
if (!mapping)
@@ -636,6 +634,8 @@ static int unmap_and_move(new_page_t get_new_page, unsigned long private,
goto move_newpage;
}
+ if (PageSwapBacked(page))
+ SetPageSwapBacked(newpage);
charge = mem_cgroup_prepare_migration(page, newpage);
if (charge == -ENOMEM) {
rc = -ENOMEM;
--
On Wed, 17 Sep 2008 13:31:49 +0900 Nice catch ! Thank you. Hmm, should I add MEM_CGROUP_CHARGE_TYPE_SHMEM rather than setting flag to newpage ? --
On Wed, 17 Sep 2008 14:46:59 +0900 I acked but.. can't this change moved into memcontrol.c ? Thanks, -Kame --
Hmm, something like this?
---
@@ -734,6 +734,9 @@ int mem_cgroup_prepare_migration(struct page *page, struct page *newpa
if (mem_cgroup_subsys.disabled)
return 0;
+ if (PageSwapBacked(page))
+ SetPageSwapBacked(newpage);
+
lock_page_cgroup(page);
pc = page_get_page_cgroup(page);
if (pc) {
---
Or, adding MEM_CGROUP_CHARGE_TYPE_SHMEM and
---
@@ -740,7 +740,10 @@ int mem_cgroup_prepare_migration(struct page *page, struct page *newp
mem = pc->mem_cgroup;
css_get(&mem->css);
if (pc->flags & PAGE_CGROUP_FLAG_CACHE)
- ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
+ if (page_is_file_cache(page))
+ ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
+ else
+ ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
}
unlock_page_cgroup(page);
if (mem) {
---
(Of course, mem_cgroup_charge_common should be modified too.)
Thanks,
Daisuke Nishimura.
--
On Wed, 17 Sep 2008 15:19:51 +0900 like this :) I don't want to change logic in migration.c (and this is special case handling for memcg.) Thanks, -Kame --
OK. I'll rewrite and resend it later. Thanks, Daisuke Nishimura. --
Before this patch, if migrating shmem/tmpfs pages, newpage would be
charged with PAGE_CGROUP_FLAG_FILE set, while oldpage has been charged
without the flag.
The problem here is mem_cgroup_move_lists doesn't clear(or set)
the PAGE_CGROUP_FLAG_FILE flag, so pc->flags of the newpage
remains PAGE_CGROUP_FLAG_FILE set even when the pc is moved to
another lru(anon) by mem_cgroup_move_lists. And this leads to
incorrect MEM_CGROUP_ZSTAT.
(In my test, I see an underflow of MEM_CGROUP_ZSTAT(active_file).
As a result, mem_cgroup_calc_reclaim returns very huge number and
causes soft lockup on page reclaim.)
I'm not sure if mem_cgroup_move_lists should handle PAGE_CGROUP_FLAG_FILE
or not(I suppose it should be used to move between active <-> inactive,
not anon <-> file), I added MEM_CGROUP_CHARGE_TYPE_SHMEM for precharge
at shmem's page migration.
ChangeLog: v1->v2
- instead of modifying migrate.c, modify memcontrol.c only.
- add MEM_CGROUP_CHARGE_TYPE_SHMEM.
Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
---
mm/memcontrol.c | 13 ++++++++++---
1 files changed, 10 insertions(+), 3 deletions(-)
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 2979d22..ef8812d 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -179,6 +179,7 @@ enum charge_type {
MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
MEM_CGROUP_CHARGE_TYPE_MAPPED,
MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
+ MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
};
/*
@@ -579,8 +580,10 @@ static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
pc->flags |= PAGE_CGROUP_FLAG_FILE;
else
pc->flags |= PAGE_CGROUP_FLAG_ACTIVE;
- } else
+ } else if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
pc->flags = PAGE_CGROUP_FLAG_ACTIVE;
+ else /* MEM_CGROUP_CHARGE_TYPE_SHMEM */
+ pc->flags = PAGE_CGROUP_FLAG_CACHE | PAGE_CGROUP_FLAG_ACTIVE;
lock_page_cgroup(page);
if (unlikely(page_get_page_cgroup(page))) {
@@ -739,8 +742,12 @@ int ...On Wed, 17 Sep 2008 16:55:44 +0900 I'll fix mem_cgroup_charge_cache_page() to use TYPE_SHMEM later. Thank you. Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> --
On Wed, 17 Sep 2008 16:55:44 +0900 I queued this as a fix against vmscan-split-lru-lists-into-anon-file-sets.patch. Was that appropriate? If the bug you're fixing here is also present in mainline then I'll need to ask for a tested patch against mainline, please. --
I don't think this bug exist in mainline, where memcg have only two ZSTAT(active/inactive) and mem_cgroup_move_lists can handle them properly. Thanks, Daisuke Nishimura. --
On Wed, 17 Sep 2008 15:51:12 -0700 I think this bug depends on split-lru patch set. Mayne not in mainline yet...? Thanks, -Kame --