Skip to contentSkip to conversation

Query regarding deadlock involving cgroup_threadgroup_rwsem and cpu_hotplug_lock

PATCH Applied Imran Khan 5 participants 15 messages Jul 20, 2022 – Aug 17, 2022 1 message not in the archive

Subsystems (2)
lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

Imran Khan

author
#1 lore raw

Hello everyone,

I am seeing a deadlock between cgroup_threadgroup_rwsem and cpu_hotplug_lock in 5.4 kernel.

Due to some missing drivers I don't have this test setup for latest upstream kernel but looking at the code the issue seems to be present in the latest kernel as well. If needed I can provide stack traces and other relevant info from the vmcore that I have got from 5.4 setup.

The description of the problem is as follows (I am using 5.19-rc7 as reference below):

__cgroup_procs_write acquires cgroup_threadgroup_rwsem via cgroup_procs_write_start and then invokes cgroup_attach_task. Now cgroup_attach_task can invoke following call chain:

cgroup_attach_task --> cgroup_migrate --> cgroup_migrate_execute --> cpuset_attach

Here cpuset_attach tries to take cpu_hotplug_lock.

But by this time if some other context

1. is already in the middle of cpu hotplug and has acquired cpu_hotplug_lock in _cpu_up but
2. has not yet reached CPUHP_ONLINE state and
3. one of the intermediate hotplug states (in my case CPUHP_AP_ONLINE_DYN ) has a callback which involves creation of a thread (or invocation of copy_process via some other path) the invoked copy_process will get blocked on cgroup_threadgroup_rwsem in following call chain:

copy_process --> cgroup_can_fork --> cgroup_css_set_fork -->
cgroup_threadgroup_change_begin

I am looking for suggestions to fix this deadlock.

Or if I am missing something in the above analysis and the above mention scenario can't happen in latest upstream kernel, then please let me know as that would help me in back porting relevant changes to 5.4 kernel because the issue definitely exists in 5.4 kernel.

Thanks,
-- Imran

author
#2 lore raw

Hello everyone,

I am seeing a deadlock between cgroup_threadgroup_rwsem and cpu_hotplug_lock in 5.4 kernel.

Due to some missing drivers I don't have this test setup for latest upstream kernel but looking at the code the issue seems to be present in the latest kernel as well. If needed I can provide stack traces and other relevant info from the vmcore that I have got from 5.4 setup.

The description of the problem is as follows (I am using 5.19-rc7 as reference below):

__cgroup_procs_write acquires cgroup_threadgroup_rwsem via cgroup_procs_write_start and then invokes cgroup_attach_task. Now cgroup_attach_task can invoke following call chain:

cgroup_attach_task --> cgroup_migrate --> cgroup_migrate_execute --> cpuset_attach

Here cpuset_attach tries to take cpu_hotplug_lock.

But by this time if some other context

1. is already in the middle of cpu hotplug and has acquired cpu_hotplug_lock in _cpu_up but
2. has not yet reached CPUHP_ONLINE state and
3. one of the intermediate hotplug states (in my case CPUHP_AP_ONLINE_DYN ) has a callback which involves creation of a thread (or invocation of copy_process via some other path) the invoked copy_process will get blocked on cgroup_threadgroup_rwsem in following call chain:

copy_process --> cgroup_can_fork --> cgroup_css_set_fork -->
cgroup_threadgroup_change_begin

I am looking for suggestions to fix this deadlock.

Or if I am missing something in the above analysis and the above mention scenario can't happen in latest upstream kernel, then please let me know as that would help me in back porting relevant changes to 5.4 kernel because the issue definitely exists in 5.4 kernel.

Thanks,
-- Imran

#3 lore raw

Hi,

Quoted 31 lines · Imran Khan

Hello everyone,

I am seeing a deadlock between cgroup_threadgroup_rwsem and cpu_hotplug_lock in 5.4 kernel.

Due to some missing drivers I don't have this test setup for latest upstream kernel but looking at the code the issue seems to be present in the latest kernel as well. If needed I can provide stack traces and other relevant info from the vmcore that I have got from 5.4 setup.

The description of the problem is as follows (I am using 5.19-rc7 as reference below):

__cgroup_procs_write acquires cgroup_threadgroup_rwsem via cgroup_procs_write_start and then invokes cgroup_attach_task. Now cgroup_attach_task can invoke following call chain:

cgroup_attach_task --> cgroup_migrate --> cgroup_migrate_execute --> cpuset_attach

Here cpuset_attach tries to take cpu_hotplug_lock.

But by this time if some other context

1. is already in the middle of cpu hotplug and has acquired cpu_hotplug_lock in _cpu_up but
2. has not yet reached CPUHP_ONLINE state and
3. one of the intermediate hotplug states (in my case CPUHP_AP_ONLINE_DYN ) has a callback which involves creation of a thread (or invocation of copy_process via some other path) the invoked copy_process will get blocked on cgroup_threadgroup_rwsem in following call chain:

Quoted 2 lines
    copy_process --> cgroup_can_fork --> cgroup_css_set_fork -->

cgroup_threadgroup_change_begin

Similar discussion is at [1], not sure on the conclusion.

[1]
https://lore.kernel.org/lkml/20220705123705.764-1-xuewen.yan-1tVvrHeaX6nQT0dZR+AlfA@public.gmane.org/

-Mukesh

Quoted 11 lines

I am looking for suggestions to fix this deadlock.

Or if I am missing something in the above analysis and the above mention scenario can't happen in latest upstream kernel, then please let me know as that would help me in back porting relevant changes to 5.4 kernel because the issue definitely exists in 5.4 kernel.

Thanks,
-- Imran

#4 lore raw

Looks like these patches are the fixes.

https://lore.kernel.org/all/YtDvN0wJ6CKaEPN8-NiLfg/[email protected]/#r

Would let Tejun confirm this .

-Mukesh

Quoted 55 lines · Mukesh Ojha

Hi,

Quoted 31 lines · Imran Khan

Hello everyone,

I am seeing a deadlock between cgroup_threadgroup_rwsem and cpu_hotplug_lock in 5.4 kernel.

Due to some missing drivers I don't have this test setup for latest upstream kernel but looking at the code the issue seems to be present in the latest kernel as well. If needed I can provide stack traces and other relevant info from the vmcore that I have got from 5.4 setup.

The description of the problem is as follows (I am using 5.19-rc7 as reference below):

__cgroup_procs_write acquires cgroup_threadgroup_rwsem via cgroup_procs_write_start and then invokes cgroup_attach_task. Now cgroup_attach_task can invoke following call chain:

cgroup_attach_task --> cgroup_migrate --> cgroup_migrate_execute --> cpuset_attach

Here cpuset_attach tries to take cpu_hotplug_lock.

But by this time if some other context

1. is already in the middle of cpu hotplug and has acquired cpu_hotplug_lock in _cpu_up but
2. has not yet reached CPUHP_ONLINE state and
3. one of the intermediate hotplug states (in my case CPUHP_AP_ONLINE_DYN ) has a callback which involves creation of a thread (or invocation of copy_process via some other path) the invoked copy_process will get blocked on cgroup_threadgroup_rwsem in following call chain:

Quoted 2 lines

    copy_process --> cgroup_can_fork --> cgroup_css_set_fork --> cgroup_threadgroup_change_begin

Similar discussion is at [1], not sure on the conclusion.

[1]
https://lore.kernel.org/lkml/20220705123705.764-1-xuewen.yan-1tVvrHeaX6nQT0dZR+AlfA@public.gmane.org/

-Mukesh

Quoted 11 lines

I am looking for suggestions to fix this deadlock.

Or if I am missing something in the above analysis and the above mention scenario can't happen in latest upstream kernel, then please let me know as that would help me in back porting relevant changes to 5.4 kernel because the issue definitely exists in 5.4 kernel.

Thanks,
-- Imran

maintainer
#6 lore raw
Quoted 9 lines · Tejun Heo
Quoted 5 lines · Mukesh Ojha

Looks like these patches are the fixes.

https://lore.kernel.org/all/YtDvN0wJ6CKaEPN8-NiLfg/[email protected]/#r

Would let Tejun confirm this .

Yeah, looks like the same issue. I'll write up a patch later this week / early next unless someone beats me to it.

https://lore.kernel.org/lkml/20220705123705.764-1-xuewen.yan-1tVvrHeaX6nQT0dZR+AlfA@public.gmane.org/ is the thread with the same issue. Let's follow up there.

Thanks.

Signature · 1 line
tejun
#7 lore raw

Hi Tejun,

Quoted 6 lines · Tejun Heo
Quoted 5 lines · Tejun Heo
Quoted 4 lines · Mukesh Ojha

Looks like these patches are the fixes.

https://lore.kernel.org/all/YtDvN0wJ6CKaEPN8-NiLfg/[email protected]/#r

Quoted 7 lines
Quoted 4 lines
Quoted 1 line

Would let Tejun confirm this .

Yeah, looks like the same issue. I'll write up a patch later this week / early next unless someone beats me to it.

https://lore.kernel.org/lkml/20220705123705.764-1-xuewen.yan-1tVvrHeaX6nQT0dZR+AlfA@public.gmane.org/ is the thread with the same issue. Let's follow up there.

Since, i am not part of the above thread, is the reason i am commenting here.

The original patch of yours [1] and the revert of [2] is fixing the issue and it is also confirmed here [3]. Can we get proper fix merge on your tree?

[1] https://lore.kernel.org/lkml/YuGbYCfAG81mZBnN-NiLfg/[email protected]/

[2] https://lore.kernel.org/all/20220121101210.84926-1-zhangqiao22-hv44wF8Li93QT0dZR+AlfA@public.gmane.org/

[3] https://lore.kernel.org/lkml/CAB8ipk-72V-bYRfL-VcSRSyXTeQqkBVj+1d5MHSVV5CTar9a0Q-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org/

-Mukesh

Quoted 3 lines

Thanks.

maintainer
#8 lore raw

+Cc: Zhao Gongyi <[email protected]>, Zhang Qiao <[email protected]>

Quoted 8 lines · Mukesh Ojha

The original patch of yours [1] and the revert of [2] is fixing the issue and it is also confirmed here [3].
Can we get proper fix merge on your tree?

[1] https://lore.kernel.org/lkml/[email protected]/

[2]
https://lore.kernel.org/all/[email protected]/

The revert + Tejun's patch looks fine wrt the problem of the reverted patch (just moves cpus_read_lock to upper callers).

I'd just suggest a comment that'd explicitly document also the lock order that we stick to, IIUC, it should be:

cpu_hotplug_lock // cpus_read_lock
cgroup_threadgroup_rwsem
cpuset_rwsem

Michal

Quoted 4 lines

[3] https://lore.kernel.org/lkml/CAB8ipk-72V-bYRfL-VcSRSyXTeQqkBVj+1d5MHSVV5CTar9a0Q@mail.gmail.com/

-Mukesh

#9 lore raw

Hi Michal

Quoted 8 lines · Michal Koutný

+Cc: Zhao Gongyi <[email protected]>, Zhang Qiao <[email protected]>

Quoted 4 lines · Mukesh Ojha

The original patch of yours [1] and the revert of [2] is fixing the issue and it is also confirmed here [3].
Can we get proper fix merge on your tree?

Quoted 7 lines
Quoted 4 lines

[1] https://lore.kernel.org/lkml/YuGbYCfAG81mZBnN-NiLfg/[email protected]/

[2]
https://lore.kernel.org/all/[email protected]/

The revert + Tejun's patch looks fine wrt the problem of the reverted patch (just moves cpus_read_lock to upper callers).

Your means is that the problem should be fixed by [1]+[2]'s revert ? I just tested the case which reverted the [2]. Need I test with [1] and [2]?

Thanks!

Quoted 14 lines

I'd just suggest a comment that'd explicitly document also the lock order that we stick to, IIUC, it should be:

cpu_hotplug_lock // cpus_read_lock
cgroup_threadgroup_rwsem
cpuset_rwsem

Michal

Quoted 4 lines

[3] https://lore.kernel.org/lkml/CAB8ipk-72V-bYRfL-VcSRSyXTeQqkBVj+1d5MHSVV5CTar9a0Q-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org/

-Mukesh

maintainer
#10 lore raw
Quoted 1 line · Xuewen Yan

Your means is that the problem should be fixed by [1]+[2]'s revert ?

I understood that was already the combination you had tested. You write in [T] that [1] alone causes (another) deadlock and therefore the revert of [2] was suggested.

Quoted 1 line

I just tested the case which reverted the [2]. Need I test with [1] and [2]?

It'd be better (unless you haven't already :-), my reasoning is for the [1]+[2] combo.

Thanks,
Michal

[T] https://lore.kernel.org/r/CAB8ipk_gCLtvEahsp2DvPJf4NxRsM8WCYmmH=yTd7zQE+81_Yg-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org/

#11 lore raw
Quoted 11 lines · Michal Koutný
Quoted 1 line · Xuewen Yan

Your means is that the problem should be fixed by [1]+[2]'s revert ?

I understood that was already the combination you had tested. You write in [T] that [1] alone causes (another) deadlock and therefore the revert of [2] was suggested.

Quoted 1 line

I just tested the case which reverted the [2]. Need I test with [1] and [2]?

It'd be better (unless you haven't already :-), my reasoning is for the [1]+[2] combo.

Feel free to add my

Reported-and-tested-by Mukesh Ojha <[email protected]>

-Mukesh

Quoted 5 lines

Thanks,
Michal

[T] https://lore.kernel.org/r/CAB8ipk_gCLtvEahsp2DvPJf4NxRsM8WCYmmH=yTd7zQE+81_Yg-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org/

maintainer
#12 lore raw

PATCH 1/1 cgroup: Fix threadgroup_rwsem <-> cpus_read_lock() deadlock

Bringing up a CPU may involve creating new tasks which requires read-locking threadgroup_rwsem, so threadgroup_rwsem nests inside cpus_read_lock(). However, cpuset's ->attach(), which may be called with thredagroup_rwsem write-locked, also wants to disable CPU hotplug and acquires cpus_read_lock(), leading to a deadlock.

Fix it by guaranteeing that ->attach() is always called with CPU hotplug disabled and removing cpus_read_lock() call from cpuset_attach().

---
Hello, sorry about the delay.

So, the previous patch + the revert isn't quite correct because we sometimes elide both cpus_read_lock() and threadgroup_rwsem together and cpuset_attach() woudl end up running without CPU hotplug enabled. Can you please test whether this patch fixes the problem?

Thanks.

 kernel/cgroup/cgroup.c |   77 ++++++++++++++++++++++++++++++++++---------------
 kernel/cgroup/cpuset.c |    3 -
 2 files changed, 55 insertions(+), 25 deletions(-)
kernel/cgroup/cgroup.c +54 -23
index ffaccd6373f1e..52502f34fae8c 100644
@@ -2369,6 +2369,47 @@ int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
 }
 EXPORT_SYMBOL_GPL(task_cgroup_path);
 
+/**
+ * cgroup_attach_lock - Lock for ->attach()
+ * @lock_threadgroup: whether to down_write cgroup_threadgroup_rwsem
+ *
+ * cgroup migration sometimes needs to stabilize threadgroups against forks and
+ * exits by write-locking cgroup_threadgroup_rwsem. However, some ->attach()
+ * implementations (e.g. cpuset), also need to disable CPU hotplug.
+ * Unfortunately, letting ->attach() operations acquire cpus_read_lock() can
+ * lead to deadlocks.
+ *
+ * Bringing up a CPU may involve creating new tasks which requires read-locking
+ * threadgroup_rwsem, so threadgroup_rwsem nests inside cpus_read_lock(). If we
+ * call an ->attach() which acquires the cpus lock while write-locking
+ * threadgroup_rwsem, the locking order is reversed and we end up waiting for an
+ * on-going CPU hotplug operation which in turn is waiting for the
+ * threadgroup_rwsem to be released to create new tasks. For more details:
+ *
+ *   http://lkml.kernel.org/r/20220711174629.uehfmqegcwn2lqzu@wubuntu
+ *
+ * Resolve the situation by always acquiring cpus_read_lock() before optionally
+ * write-locking cgroup_threadgroup_rwsem. This allows ->attach() to assume that
+ * CPU hotplug is disabled on entry.
+ */
+static void cgroup_attach_lock(bool lock_threadgroup)
+{
+	cpus_read_lock();
+	if (lock_threadgroup)
+		percpu_down_write(&cgroup_threadgroup_rwsem);
+}
+
+/**
+ * cgroup_attach_unlock - Undo cgroup_attach_lock()
+ * @lock_threadgroup: whether to up_write cgroup_threadgroup_rwsem
+ */
+static void cgroup_attach_unlock(bool lock_threadgroup)
+{
+	if (lock_threadgroup)
+		percpu_up_write(&cgroup_threadgroup_rwsem);
+	cpus_read_unlock();
+}
+
 /**
  * cgroup_migrate_add_task - add a migration target task to a migration context
  * @task: target task
@@ -2841,8 +2882,7 @@ int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
 }
 
 struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup,
-					     bool *locked)
-	__acquires(&cgroup_threadgroup_rwsem)
+					     bool *threadgroup_locked)
 {
 	struct task_struct *tsk;
 	pid_t pid;
@@ -2859,12 +2899,8 @@ struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup,
 	 * 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;
-	}
+	*threadgroup_locked = pid || threadgroup;
+	cgroup_attach_lock(*threadgroup_locked);
 
 	rcu_read_lock();
 	if (pid) {
@@ -2895,17 +2931,14 @@ 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;
-	}
+	cgroup_attach_unlock(*threadgroup_locked);
+	*threadgroup_locked = false;
 out_unlock_rcu:
 	rcu_read_unlock();
 	return tsk;
 }
 
-void cgroup_procs_write_finish(struct task_struct *task, bool locked)
-	__releases(&cgroup_threadgroup_rwsem)
+void cgroup_procs_write_finish(struct task_struct *task, bool threadgroup_locked)
 {
 	struct cgroup_subsys *ss;
 	int ssid;
@@ -2913,8 +2946,8 @@ void cgroup_procs_write_finish(struct task_struct *task, bool locked)
 	/* release reference from cgroup_procs_write_start() */
 	put_task_struct(task);
 
-	if (locked)
-		percpu_up_write(&cgroup_threadgroup_rwsem);
+	cgroup_attach_unlock(threadgroup_locked);
+
 	for_each_subsys(ss, ssid)
 		if (ss->post_attach)
 			ss->post_attach();
@@ -3000,8 +3033,7 @@ static int cgroup_update_dfl_csses(struct cgroup *cgrp)
 	 * write-locking can be skipped safely.
 	 */
 	has_tasks = !list_empty(&mgctx.preloaded_src_csets);
-	if (has_tasks)
-		percpu_down_write(&cgroup_threadgroup_rwsem);
+	cgroup_attach_lock(has_tasks);
 
 	/* NULL dst indicates self on default hierarchy */
 	ret = cgroup_migrate_prepare_dst(&mgctx);
@@ -3022,8 +3054,7 @@ static int cgroup_update_dfl_csses(struct cgroup *cgrp)
 	ret = cgroup_migrate_execute(&mgctx);
 out_finish:
 	cgroup_migrate_finish(&mgctx);
-	if (has_tasks)
-		percpu_up_write(&cgroup_threadgroup_rwsem);
+	cgroup_attach_unlock(has_tasks);
 	return ret;
 }
 
@@ -4971,13 +5002,13 @@ static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
 	struct task_struct *task;
 	const struct cred *saved_cred;
 	ssize_t ret;
-	bool locked;
+	bool threadgroup_locked;
 
 	dst_cgrp = cgroup_kn_lock_live(of->kn, false);
 	if (!dst_cgrp)
 		return -ENODEV;
 
-	task = cgroup_procs_write_start(buf, threadgroup, &locked);
+	task = cgroup_procs_write_start(buf, threadgroup, &threadgroup_locked);
 	ret = PTR_ERR_OR_ZERO(task);
 	if (ret)
 		goto out_unlock;
@@ -5003,7 +5034,7 @@ static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
 	ret = cgroup_attach_task(dst_cgrp, task, threadgroup);
 
 out_finish:
-	cgroup_procs_write_finish(task, locked);
+	cgroup_procs_write_finish(task, threadgroup_locked);
 out_unlock:
 	cgroup_kn_unlock(of->kn);
 
kernel/cgroup/cpuset.c +1 -2
index 58aadfda9b8b3..1f3a55297f39d 100644
@@ -2289,7 +2289,7 @@ static void cpuset_attach(struct cgroup_taskset *tset)
 	cgroup_taskset_first(tset, &css);
 	cs = css_cs(css);
 
-	cpus_read_lock();
+	lockdep_assert_cpus_held();	/* see cgroup_attach_lock() */
 	percpu_down_write(&cpuset_rwsem);
 
 	guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
@@ -2343,7 +2343,6 @@ static void cpuset_attach(struct cgroup_taskset *tset)
 		wake_up(&cpuset_attach_wq);
 
 	percpu_up_write(&cpuset_rwsem);
-	cpus_read_unlock();
 }
 
 /* The various types of files and directories in a cpuset file system */
author
#13 lore raw

Re: [PATCH cgroup/for-6.0-fixes] cgroup: Fix threadgroup_rwsem <-> cpus_read_lock() deadlock

Hello Tejun,

Quoted 9 lines · Tejun Heo

Bringing up a CPU may involve creating new tasks which requires read-locking threadgroup_rwsem, so threadgroup_rwsem nests inside cpus_read_lock(). However, cpuset's ->attach(), which may be called with thredagroup_rwsem write-locked, also wants to disable CPU hotplug and acquires cpus_read_lock(), leading to a deadlock.

Fix it by guaranteeing that ->attach() is always called with CPU hotplug disabled and removing cpus_read_lock() call from cpuset_attach().

Quoted 9 lines

---
Hello, sorry about the delay.

So, the previous patch + the revert isn't quite correct because we sometimes elide both cpus_read_lock() and threadgroup_rwsem together and cpuset_attach() woudl end up running without CPU hotplug enabled. Can you please test whether this patch fixes the problem?

This fixes the issue seen in my setup. As my setup is 5.4 based I used cgroup_attach_lock/unlock(true) in the backport version of your patch.

Feel free to add my

Reviewed-and-tested-by Imran Khan <[email protected]>

Thanks,
-- Imran

#14 lore raw

Re: [PATCH cgroup/for-6.0-fixes] cgroup: Fix threadgroup_rwsem <-> cpus_read_lock() deadlock

In reply to Tejun Heo

Hi Tejun

Quoted 3 lines · Tejun Heo

Bringing up a CPU may involve creating new tasks which requires read-locking threadgroup_rwsem, so threadgroup_rwsem nests inside cpus_read_lock().

Indeed, it is creating new kthreads. And not only creating new kthread, but also destroying kthread. the backtrace is:

__switch_to
__schedule
schedule
percpu_rwsem_wait <<< wait for cgroup_threadgroup_rwsem
__percpu_down_read
exit_signals
do_exit
kthread

Quoted 31 lines

However, cpuset's ->attach(), which may be called with thredagroup_rwsem write-locked, also wants to disable CPU hotplug and acquires cpus_read_lock(), leading to a deadlock.

Fix it by guaranteeing that ->attach() is always called with CPU hotplug disabled and removing cpus_read_lock() call from cpuset_attach().

---
Hello, sorry about the delay.

So, the previous patch + the revert isn't quite correct because we sometimes elide both cpus_read_lock() and threadgroup_rwsem together and cpuset_attach() woudl end up running without CPU hotplug enabled. Can you please test whether this patch fixes the problem?

Thanks.

 kernel/cgroup/cgroup.c |   77 ++++++++++++++++++++++++++++++++++---------------
 kernel/cgroup/cpuset.c |    3 -
 2 files changed, 55 insertions(+), 25 deletions(-)
kernel/cgroup/cgroup.c +1 -0
index ffaccd6373f1e..52502f34fae8c 100644
@@ -2369,6 +2369,47 @@ int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
 }
 EXPORT_SYMBOL_GPL(task_cgroup_path);
 
+/**

kernel/cgroup/cgroup.c int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen) · line 2382

Quoted 10 lines
kernel/cgroup/cgroup.c +10 -0
+ * cgroup_attach_lock - Lock for ->attach()
+ * @lock_threadgroup: whether to down_write cgroup_threadgroup_rwsem
+ *
+ * cgroup migration sometimes needs to stabilize threadgroups against forks and
+ * exits by write-locking cgroup_threadgroup_rwsem. However, some ->attach()
+ * implementations (e.g. cpuset), also need to disable CPU hotplug.
+ * Unfortunately, letting ->attach() operations acquire cpus_read_lock() can
+ * lead to deadlocks.
+ *
+ * Bringing up a CPU may involve creating new tasks which requires read-locking

Is it better to change to creating new kthreads and destroying kthreads?

Quoted 101 lines
kernel/cgroup/cgroup.c +39 -18
+ * threadgroup_rwsem, so threadgroup_rwsem nests inside cpus_read_lock(). If we
+ * call an ->attach() which acquires the cpus lock while write-locking
+ * threadgroup_rwsem, the locking order is reversed and we end up waiting for an
+ * on-going CPU hotplug operation which in turn is waiting for the
+ * threadgroup_rwsem to be released to create new tasks. For more details:
+ *
+ *   http://lkml.kernel.org/r/20220711174629.uehfmqegcwn2lqzu@wubuntu
+ *
+ * Resolve the situation by always acquiring cpus_read_lock() before optionally
+ * write-locking cgroup_threadgroup_rwsem. This allows ->attach() to assume that
+ * CPU hotplug is disabled on entry.
+ */
+static void cgroup_attach_lock(bool lock_threadgroup)
+{
+       cpus_read_lock();
+       if (lock_threadgroup)
+               percpu_down_write(&cgroup_threadgroup_rwsem);
+}
+
+/**
+ * cgroup_attach_unlock - Undo cgroup_attach_lock()
+ * @lock_threadgroup: whether to up_write cgroup_threadgroup_rwsem
+ */
+static void cgroup_attach_unlock(bool lock_threadgroup)
+{
+       if (lock_threadgroup)
+               percpu_up_write(&cgroup_threadgroup_rwsem);
+       cpus_read_unlock();
+}
+
 /**
  * cgroup_migrate_add_task - add a migration target task to a migration context
  * @task: target task
@@ -2841,8 +2882,7 @@ int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
 }
 
 struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup,
-                                            bool *locked)
-       __acquires(&cgroup_threadgroup_rwsem)
+                                            bool *threadgroup_locked)
 {
        struct task_struct *tsk;
        pid_t pid;
@@ -2859,12 +2899,8 @@ struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup,
         * 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;
-       }
+       *threadgroup_locked = pid || threadgroup;
+       cgroup_attach_lock(*threadgroup_locked);
 
        rcu_read_lock();
        if (pid) {
@@ -2895,17 +2931,14 @@ 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;
-       }
+       cgroup_attach_unlock(*threadgroup_locked);
+       *threadgroup_locked = false;
 out_unlock_rcu:
        rcu_read_unlock();
        return tsk;
 }
 
-void cgroup_procs_write_finish(struct task_struct *task, bool locked)
-       __releases(&cgroup_threadgroup_rwsem)
+void cgroup_procs_write_finish(struct task_struct *task, bool threadgroup_locked)
 {
        struct cgroup_subsys *ss;
        int ssid;
@@ -2913,8 +2946,8 @@ void cgroup_procs_write_finish(struct task_struct *task, bool locked)
        /* release reference from cgroup_procs_write_start() */
        put_task_struct(task);
 
-       if (locked)
-               percpu_up_write(&cgroup_threadgroup_rwsem);
+       cgroup_attach_unlock(threadgroup_locked);
+
        for_each_subsys(ss, ssid)
                if (ss->post_attach)
                        ss->post_attach();
@@ -3000,8 +3033,7 @@ static int cgroup_update_dfl_csses(struct cgroup *cgrp)
         * write-locking can be skipped safely.
         */
        has_tasks = !list_empty(&mgctx.preloaded_src_csets);
-       if (has_tasks)
-               percpu_down_write(&cgroup_threadgroup_rwsem);
+       cgroup_attach_lock(has_tasks);
 
        /* NULL dst indicates self on default hierarchy */
        ret = cgroup_migrate_prepare_dst(&mgctx);

kernel/cgroup/cgroup.c static int cgroup_update_dfl_csses(struct cgroup *cgrp) · line 3057

Quoted 6 lines
kernel/cgroup/cgroup.c +1 -2
@@ -3022,8 +3054,7 @@ static int cgroup_update_dfl_csses(struct cgroup *cgrp)
        ret = cgroup_migrate_execute(&mgctx);
 out_finish:
        cgroup_migrate_finish(&mgctx);
-       if (has_tasks)
-               percpu_up_write(&cgroup_threadgroup_rwsem);
+       cgroup_attach_unlock(has_tasks);

In kernel5.15, I just set cgroup_attach_lock/unlock(true).

Quoted 42 lines
kernel/cgroup/cgroup.c +3 -3
        return ret;
 }
 
@@ -4971,13 +5002,13 @@ static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
        struct task_struct *task;
        const struct cred *saved_cred;
        ssize_t ret;
-       bool locked;
+       bool threadgroup_locked;
 
        dst_cgrp = cgroup_kn_lock_live(of->kn, false);
        if (!dst_cgrp)
                return -ENODEV;
 
-       task = cgroup_procs_write_start(buf, threadgroup, &locked);
+       task = cgroup_procs_write_start(buf, threadgroup, &threadgroup_locked);
        ret = PTR_ERR_OR_ZERO(task);
        if (ret)
                goto out_unlock;
@@ -5003,7 +5034,7 @@ static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
        ret = cgroup_attach_task(dst_cgrp, task, threadgroup);
 
 out_finish:
-       cgroup_procs_write_finish(task, locked);
+       cgroup_procs_write_finish(task, threadgroup_locked);
 out_unlock:
        cgroup_kn_unlock(of->kn);
 
kernel/cgroup/cpuset.c +1 -1
index 58aadfda9b8b3..1f3a55297f39d 100644
@@ -2289,7 +2289,7 @@ static void cpuset_attach(struct cgroup_taskset *tset)
        cgroup_taskset_first(tset, &css);
        cs = css_cs(css);
 
-       cpus_read_lock();
+       lockdep_assert_cpus_held();     /* see cgroup_attach_lock() */
        percpu_down_write(&cpuset_rwsem);
 
        guarantee_online_mems(cs, &cpuset_attach_nodemask_to);

kernel/cgroup/cpuset.c static void cpuset_attach(struct cgroup_taskset *tset) · line 2348

Quoted 7 lines
kernel/cgroup/cpuset.c +0 -1
@@ -2343,7 +2343,6 @@ static void cpuset_attach(struct cgroup_taskset *tset)
                wake_up(&cpuset_attach_wq);
 
        percpu_up_write(&cpuset_rwsem);
-       cpus_read_unlock();
 }
 
 /* The various types of files and directories in a cpuset file system */

I backported your patch. to kernel5.4 and kernel5.15, and just setting cgroup_attach_lock/unlock(true) when there are conflicts. And the deadlock has not occured.

Reported-and-tested-by Xuewen Yan <[email protected]>

Thanks!

maintainer Applied
#15 lore raw

Re: [PATCH cgroup/for-6.0-fixes] cgroup: Fix threadgroup_rwsem <-> cpus_read_lock() deadlock

In reply to Tejun Heo
Quoted 10 lines · Tejun Heo

Bringing up a CPU may involve creating new tasks which requires read-locking threadgroup_rwsem, so threadgroup_rwsem nests inside cpus_read_lock(). However, cpuset's ->attach(), which may be called with thredagroup_rwsem write-locked, also wants to disable CPU hotplug and acquires cpus_read_lock(), leading to a deadlock.

Fix it by guaranteeing that ->attach() is always called with CPU hotplug disabled and removing cpus_read_lock() call from cpuset_attach().

Applied to cgroup/for-6.0-fixes w/ commit message and comment update suggested by Xuewen and Fixes / stable tags added.

Thanks.

Signature · 1 line
tejun