bugfix for memory cgroup controller: migration under memory controller fix
While using memory control cgroup, page-migration under it works as following.
==
1. uncharge all refs at try to unmap.
2. charge regs again remove_migration_ptes()
==
This is simple but has following problems.
==
The page is uncharged and charged back again if *mapped*.
- This means that cgroup before migration can be different from one after
migration
- If page is not mapped but charged as page cache, charge is just ignored
(because not mapped, it will not be uncharged before migration)
This is memory leak.
==
This patch tries to keep memory cgroup at page migration by increasing
one refcnt during it. 3 functions are added.
mem_cgroup_prepare_migration() --- increase refcnt of page->page_cgroup
mem_cgroup_end_migration() --- decrease refcnt of page->page_cgroup
mem_cgroup_page_migration() --- copy page->page_cgroup from old page to
new page.
During migration
- old page is under PG_locked.
- new page is under PG_locked, too.
- both old page and new page is not on LRU.
These 3 facts guarantee that page_cgroup() migration has no race.
Tested and worked well in x86_64/fake-NUMA box.
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Balbir Singh <balbir@linux.vnet.ibm.com>
Cc: Pavel Emelianov <xemul@openvz.org>
Cc: Paul Menage <menage@google.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Kirill Korotaev <dev@sw.ru>
Cc: Herbert Poetzl <herbert@13thfloor.at>
Cc: David Rientjes <rientjes@google.com>
Cc: Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 3270ce7..128f45c 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -492,6 +492,49 @@
}
}
}
+/*
+ * Returns non-zero if a page (under migration) has valid page_cgroup member.
+ * Refcnt of page_cgroup is incremented.
+ */
+
+int mem_cgroup_prepare_migration(struct page *page)
+{
+ struct page_cgroup *pc;
+ int ret = 0;
+ lock_page_cgroup(page);
+ pc = page_get_page_cgroup(page);
+ if (pc && atomic_inc_not_zero(&pc->ref_cnt))
+ ret = 1;
+ unlock_page_cgroup(page);
+ return ret;
+}
+
+void mem_cgroup_end_migration(struct page *page)
+{
+ struct page_cgroup *pc = page_get_page_cgroup(page);
+ mem_cgroup_uncharge(pc);
+}
+/*
+ * We know both *page* and *newpage* are now not-on-LRU and Pg_locked.
+ * And no race with uncharge() routines because page_cgroup for *page*
+ * has extra one reference by mem_cgroup_prepare_migration.
+ */
+
+void mem_cgroup_page_migration(struct page *page, struct page *newpage)
+{
+ struct page_cgroup *pc;
+retry:
+ pc = page_get_page_cgroup(page);
+ if (!pc)
+ return;
+ if (clear_page_cgroup(page, pc) != pc)
+ goto retry;
+ pc->page = newpage;
+ lock_page_cgroup(newpage);
+ page_assign_page_cgroup(newpage, pc);
+ unlock_page_cgroup(newpage);
+ return;
+}
int mem_cgroup_write_strategy(char *buf, unsigned long long *tmp)
{
diff --git a/mm/migrate.c b/mm/migrate.c
index 7637941..a73504f 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -593,9 +593,10 @@
else
rc = fallback_migrate_page(mapping, newpage, page);
- if (!rc)
+ if (!rc) {
+ mem_cgroup_page_migration(page, newpage);
remove_migration_ptes(page, newpage);
- else
+ } else
newpage->mapping = NULL;
unlock_page(newpage);
@@ -614,6 +615,7 @@
int *result = NULL;
struct page *newpage = get_new_page(page, private, &result);
int rcu_locked = 0;
+ int charge = 0;
if (!newpage)
return -ENOMEM;
@@ -673,14 +675,19 @@
goto rcu_unlock;
}
+ charge = mem_cgroup_prepare_migration(page);
/* Establish migration ptes or remove ptes */
try_to_unmap(page, 1);
if (!page_mapped(page))
rc = move_to_new_page(newpage, page);
- if (rc)
+ if (rc) {
remove_migration_ptes(page, page);
+ if (charge)
+ mem_cgroup_end_migration(page);
+ } else if (charge)
+ mem_cgroup_end_migration(newpage);
rcu_unlock:
if (rcu_locked)
rcu_read_unlock();