Commit 9a3284fa authored by Michal Koutný's avatar Michal Koutný Committed by Tejun Heo

cgroup: Optimize single thread migration

There are reports of users who use thread migrations between cgroups and
they report performance drop after d59cfc09 ("sched, cgroup: replace
signal_struct->group_rwsem with a global percpu_rwsem"). The effect is
pronounced on machines with more CPUs.

The migration is affected by forking noise happening in the background,
after the mentioned commit a migrating thread must wait for all
(forking) processes on the system, not only of its threadgroup.

There are several places that need to synchronize with migration:
	a) do_exit,
	b) de_thread,
	c) copy_process,
	d) cgroup_update_dfl_csses,
	e) parallel migration (cgroup_{proc,thread}s_write).

In the case of self-migrating thread, we relax the synchronization on
cgroup_threadgroup_rwsem to avoid the cost of waiting. d) and e) are
excluded with cgroup_mutex, c) does not matter in case of single thread
migration and the executing thread cannot exec(2) or exit(2) while it is
writing into cgroup.threads. In case of do_exit because of signal
delivery, we either exit before the migration or finish the migration
(of not yet PF_EXITING thread) and die afterwards.

This patch handles only the case of self-migration by writing "0" into
cgroup.threads. For simplicity, we always take cgroup_threadgroup_rwsem
with numeric PIDs.

This change improves migration dependent workload performance similar
to per-signal_struct state.
Signed-off-by: default avatarMichal Koutný <mkoutny@suse.com>
Signed-off-by: default avatarTejun Heo <tj@kernel.org>
parent e7c7b1d8
......@@ -231,9 +231,10 @@ int cgroup_migrate(struct task_struct *leader, bool threadgroup,
int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
bool threadgroup);
struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup)
struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup,
bool *locked)
__acquires(&cgroup_threadgroup_rwsem);
void cgroup_procs_write_finish(struct task_struct *task)
void cgroup_procs_write_finish(struct task_struct *task, bool locked)
__releases(&cgroup_threadgroup_rwsem);
void cgroup_lock_and_drain_offline(struct cgroup *cgrp);
......
......@@ -495,12 +495,13 @@ static ssize_t __cgroup1_procs_write(struct kernfs_open_file *of,
struct task_struct *task;
const struct cred *cred, *tcred;
ssize_t ret;
bool locked;
cgrp = cgroup_kn_lock_live(of->kn, false);
if (!cgrp)
return -ENODEV;
task = cgroup_procs_write_start(buf, threadgroup);
task = cgroup_procs_write_start(buf, threadgroup, &locked);
ret = PTR_ERR_OR_ZERO(task);
if (ret)
goto out_unlock;
......@@ -522,7 +523,7 @@ static ssize_t __cgroup1_procs_write(struct kernfs_open_file *of,
ret = cgroup_attach_task(cgrp, task, threadgroup);
out_finish:
cgroup_procs_write_finish(task);
cgroup_procs_write_finish(task, locked);
out_unlock:
cgroup_kn_unlock(of->kn);
......
......@@ -2824,7 +2824,8 @@ int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
return ret;
}
struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup)
struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup,
bool *locked)
__acquires(&cgroup_threadgroup_rwsem)
{
struct task_struct *tsk;
......@@ -2833,7 +2834,21 @@ struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup)
if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
return ERR_PTR(-EINVAL);
/*
* If we migrate a single thread, we don't care about threadgroup
* stability. If the thread is `current`, it won't exit(2) under our
* hands or change PID through exec(2). We exclude
* cgroup_update_dfl_csses and other cgroup_{proc,thread}s_write
* callers by cgroup_mutex.
* Therefore, we can skip the global lock.
*/
lockdep_assert_held(&cgroup_mutex);
if (pid || threadgroup) {
percpu_down_write(&cgroup_threadgroup_rwsem);
*locked = true;
} else {
*locked = false;
}
rcu_read_lock();
if (pid) {
......@@ -2864,13 +2879,16 @@ struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup)
goto out_unlock_rcu;
out_unlock_threadgroup:
if (*locked) {
percpu_up_write(&cgroup_threadgroup_rwsem);
*locked = false;
}
out_unlock_rcu:
rcu_read_unlock();
return tsk;
}
void cgroup_procs_write_finish(struct task_struct *task)
void cgroup_procs_write_finish(struct task_struct *task, bool locked)
__releases(&cgroup_threadgroup_rwsem)
{
struct cgroup_subsys *ss;
......@@ -2879,6 +2897,7 @@ void cgroup_procs_write_finish(struct task_struct *task)
/* release reference from cgroup_procs_write_start() */
put_task_struct(task);
if (locked)
percpu_up_write(&cgroup_threadgroup_rwsem);
for_each_subsys(ss, ssid)
if (ss->post_attach)
......@@ -4754,12 +4773,13 @@ static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
struct cgroup *src_cgrp, *dst_cgrp;
struct task_struct *task;
ssize_t ret;
bool locked;
dst_cgrp = cgroup_kn_lock_live(of->kn, false);
if (!dst_cgrp)
return -ENODEV;
task = cgroup_procs_write_start(buf, true);
task = cgroup_procs_write_start(buf, true, &locked);
ret = PTR_ERR_OR_ZERO(task);
if (ret)
goto out_unlock;
......@@ -4777,7 +4797,7 @@ static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
ret = cgroup_attach_task(dst_cgrp, task, true);
out_finish:
cgroup_procs_write_finish(task);
cgroup_procs_write_finish(task, locked);
out_unlock:
cgroup_kn_unlock(of->kn);
......@@ -4795,6 +4815,7 @@ static ssize_t cgroup_threads_write(struct kernfs_open_file *of,
struct cgroup *src_cgrp, *dst_cgrp;
struct task_struct *task;
ssize_t ret;
bool locked;
buf = strstrip(buf);
......@@ -4802,7 +4823,7 @@ static ssize_t cgroup_threads_write(struct kernfs_open_file *of,
if (!dst_cgrp)
return -ENODEV;
task = cgroup_procs_write_start(buf, false);
task = cgroup_procs_write_start(buf, false, &locked);
ret = PTR_ERR_OR_ZERO(task);
if (ret)
goto out_unlock;
......@@ -4826,7 +4847,7 @@ static ssize_t cgroup_threads_write(struct kernfs_open_file *of,
ret = cgroup_attach_task(dst_cgrp, task, false);
out_finish:
cgroup_procs_write_finish(task);
cgroup_procs_write_finish(task, locked);
out_unlock:
cgroup_kn_unlock(of->kn);
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment