memcg: replace ss->id_lock with a rwlock
authorAndrew Bresticker <abrestic@google.com>
Wed, 2 Nov 2011 20:40:29 +0000 (13:40 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 2 Nov 2011 23:07:03 +0000 (16:07 -0700)
While back-porting Johannes Weiner's patch "mm: memcg-aware global
reclaim" for an internal effort, we noticed a significant performance
regression during page-reclaim heavy workloads due to high contention of
the ss->id_lock.  This lock protects idr map, and serializes calls to
idr_get_next() in css_get_next() (which is used during the memcg hierarchy
walk).

Since idr_get_next() is just doing a look up, we need only serialize it
with respect to idr_remove()/idr_get_new().  By making the ss->id_lock a
rwlock, contention is greatly reduced and performance improves.

Tested: cat a 256m file from a ramdisk in a 128m container 50 times on
each core (one file + container per core) in parallel on a NUMA machine.
Result is the time for the test to complete in 1 of the containers.
Both kernels included Johannes' memcg-aware global reclaim patches.

Before rwlock patch: 1710.778s
After rwlock patch: 152.227s

Signed-off-by: Andrew Bresticker <abrestic@google.com>
Cc: Paul Menage <menage@gmail.com>
Cc: Li Zefan <lizf@cn.fujitsu.com>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Ying Han <yinghan@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
include/linux/cgroup.h
kernel/cgroup.c

index da7e4bc..1b7f9d5 100644 (file)
@@ -516,7 +516,7 @@ struct cgroup_subsys {
        struct list_head sibling;
        /* used when use_id == true */
        struct idr idr;
-       spinlock_t id_lock;
+       rwlock_t id_lock;
 
        /* should be defined only by modular subsystems */
        struct module *module;
index 8386b21..d9d5648 100644 (file)
@@ -4883,9 +4883,9 @@ void free_css_id(struct cgroup_subsys *ss, struct cgroup_subsys_state *css)
 
        rcu_assign_pointer(id->css, NULL);
        rcu_assign_pointer(css->id, NULL);
-       spin_lock(&ss->id_lock);
+       write_lock(&ss->id_lock);
        idr_remove(&ss->idr, id->id);
-       spin_unlock(&ss->id_lock);
+       write_unlock(&ss->id_lock);
        kfree_rcu(id, rcu_head);
 }
 EXPORT_SYMBOL_GPL(free_css_id);
@@ -4911,10 +4911,10 @@ static struct css_id *get_new_cssid(struct cgroup_subsys *ss, int depth)
                error = -ENOMEM;
                goto err_out;
        }
-       spin_lock(&ss->id_lock);
+       write_lock(&ss->id_lock);
        /* Don't use 0. allocates an ID of 1-65535 */
        error = idr_get_new_above(&ss->idr, newid, 1, &myid);
-       spin_unlock(&ss->id_lock);
+       write_unlock(&ss->id_lock);
 
        /* Returns error when there are no free spaces for new ID.*/
        if (error) {
@@ -4929,9 +4929,9 @@ static struct css_id *get_new_cssid(struct cgroup_subsys *ss, int depth)
        return newid;
 remove_idr:
        error = -ENOSPC;
-       spin_lock(&ss->id_lock);
+       write_lock(&ss->id_lock);
        idr_remove(&ss->idr, myid);
-       spin_unlock(&ss->id_lock);
+       write_unlock(&ss->id_lock);
 err_out:
        kfree(newid);
        return ERR_PTR(error);
@@ -4943,7 +4943,7 @@ static int __init_or_module cgroup_init_idr(struct cgroup_subsys *ss,
 {
        struct css_id *newid;
 
-       spin_lock_init(&ss->id_lock);
+       rwlock_init(&ss->id_lock);
        idr_init(&ss->idr);
 
        newid = get_new_cssid(ss, 0);
@@ -5038,9 +5038,9 @@ css_get_next(struct cgroup_subsys *ss, int id,
                 * scan next entry from bitmap(tree), tmpid is updated after
                 * idr_get_next().
                 */
-               spin_lock(&ss->id_lock);
+               read_lock(&ss->id_lock);
                tmp = idr_get_next(&ss->idr, &tmpid);
-               spin_unlock(&ss->id_lock);
+               read_unlock(&ss->id_lock);
 
                if (!tmp)
                        break;