[PATCH] cgroup/cpuset: Optimize update_tasks_nodemask()

[PATCH] cgroup/cpuset: Optimize update_tasks_nodemask()

[haifeng.xu]

When change the 'cpuset.mems' under some cgroup, system will hung
for a long time. From the dmesg, many processes or theads are
stuck in fork/exit. The reason is show as follows.

thread A:
cpuset_write_resmask /* takes cpuset_rwsem */
  ...
    update_tasks_nodemask
      mpol_rebind_mm /* waits mmap_lock */

thread B:
worker_thread
  ...
    cpuset_migrate_mm_workfn
      do_migrate_pages /* takes mmap_lock */

thread C:
cgroup_procs_write /* takes cgroup_mutex and cgroup_threadgroup_rwsem */
  ...
    cpuset_can_attach
      percpu_down_write /* waits cpuset_rwsem */

Once update the nodemasks of cpuset, thread A wakes up thread B to
migrate mm. But when thread A iterates through all tasks, including
child threads and group leader, it has to wait the mmap_lock which
has been take by thread B. Unfortunately, thread C wants to migrate
tasks into cgroup at this moment, it must wait thread A to release
cpuset_rwsem. If thread B spends much time to migrate mm, the
fork/exit which acquire cgroup_threadgroup_rwsem also need to
wait for a long time.

There is no need to migrate the mm of child threads which is
shared with group leader. Just iterate through the group
leader only.

Signed-off-by: haifeng.xu <haifeng.xu@shopee.com>
---
 kernel/cgroup/cpuset.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
index 589827ccda8b..43cbd09546d0 100644
--- a/kernel/cgroup/cpuset.c
+++ b/kernel/cgroup/cpuset.c
@@ -1968,6 +1968,9 @@ static void update_tasks_nodemask(struct cpuset *cs)
 
 		cpuset_change_task_nodemask(task, &newmems);
 
+		if (!thread_group_leader(task))
+			continue;
+
 		mm = get_task_mm(task);
 		if (!mm)
 			continue;
-- 
2.25.1

Re: [PATCH] cgroup/cpuset: Optimize update_tasks_nodemask()

[Tejun Heo]

On Wed, Nov 23, 2022 at 08:21:57AM +0000, haifeng.xu wrote:
> When change the 'cpuset.mems' under some cgroup, system will hung
> for a long time. From the dmesg, many processes or theads are
> stuck in fork/exit. The reason is show as follows.
> 
> thread A:
> cpuset_write_resmask /* takes cpuset_rwsem */
>   ...
>     update_tasks_nodemask
>       mpol_rebind_mm /* waits mmap_lock */
> 
> thread B:
> worker_thread
>   ...
>     cpuset_migrate_mm_workfn
>       do_migrate_pages /* takes mmap_lock */
> 
> thread C:
> cgroup_procs_write /* takes cgroup_mutex and cgroup_threadgroup_rwsem */
>   ...
>     cpuset_can_attach
>       percpu_down_write /* waits cpuset_rwsem */
> 
> Once update the nodemasks of cpuset, thread A wakes up thread B to
> migrate mm. But when thread A iterates through all tasks, including
> child threads and group leader, it has to wait the mmap_lock which
> has been take by thread B. Unfortunately, thread C wants to migrate
> tasks into cgroup at this moment, it must wait thread A to release
> cpuset_rwsem. If thread B spends much time to migrate mm, the
> fork/exit which acquire cgroup_threadgroup_rwsem also need to
> wait for a long time.
> 
> There is no need to migrate the mm of child threads which is
> shared with group leader. 

This is only a problem in cgroup1 and cgroup1 doesn't require the threads of
a given task to be in the same cgroup. I don't think you can optimize it
this way.

Thanks.

-- 
tejun

Re: [PATCH] cgroup/cpuset: Optimize update_tasks_nodemask()

[Waiman Long]

On 11/23/22 12:05, Tejun Heo wrote:
> On Wed, Nov 23, 2022 at 08:21:57AM +0000, haifeng.xu wrote:
>> When change the 'cpuset.mems' under some cgroup, system will hung
>> for a long time. From the dmesg, many processes or theads are
>> stuck in fork/exit. The reason is show as follows.
>>
>> thread A:
>> cpuset_write_resmask /* takes cpuset_rwsem */
>>    ...
>>      update_tasks_nodemask
>>        mpol_rebind_mm /* waits mmap_lock */
>>
>> thread B:
>> worker_thread
>>    ...
>>      cpuset_migrate_mm_workfn
>>        do_migrate_pages /* takes mmap_lock */
>>
>> thread C:
>> cgroup_procs_write /* takes cgroup_mutex and cgroup_threadgroup_rwsem */
>>    ...
>>      cpuset_can_attach
>>        percpu_down_write /* waits cpuset_rwsem */
>>
>> Once update the nodemasks of cpuset, thread A wakes up thread B to
>> migrate mm. But when thread A iterates through all tasks, including
>> child threads and group leader, it has to wait the mmap_lock which
>> has been take by thread B. Unfortunately, thread C wants to migrate
>> tasks into cgroup at this moment, it must wait thread A to release
>> cpuset_rwsem. If thread B spends much time to migrate mm, the
>> fork/exit which acquire cgroup_threadgroup_rwsem also need to
>> wait for a long time.
>>
>> There is no need to migrate the mm of child threads which is
>> shared with group leader.
> This is only a problem in cgroup1 and cgroup1 doesn't require the threads of
> a given task to be in the same cgroup. I don't think you can optimize it
> this way.

I think it is an issue anyway if different threads of a process are in 
different cpusets with different node mask. It is not a configuration 
that should be used at all.

This patch makes update_tasks_nodemask() somewhat similar to 
cpuset_attach() where all tasks are iterated to update the node mask but 
only the task leaders are required to update the mm. For a non-group 
leader task, maybe we can check if the group leader is in the same 
cpuset. If so, we can skip the mm update. Do we need similar change in 
cpuset_attach()?

I do think the "migrate = is_memory_migrate(cs);" line can be moved 
outside of the loop, though. Of course, that won't help much in this case.

Cheers,
Longman

Re: [PATCH] cgroup/cpuset: Optimize update_tasks_nodemask()

[Tejun Heo]

On Wed, Nov 23, 2022 at 01:48:46PM -0500, Waiman Long wrote:
> I think it is an issue anyway if different threads of a process are in
> different cpusets with different node mask. It is not a configuration that
> should be used at all.

Anything memory related is in the same boat and people still use them
reaching whatever end results they reach. Given the whole thing is pretty
ill-defined, I don't wanna change the behavior now.

> This patch makes update_tasks_nodemask() somewhat similar to cpuset_attach()
> where all tasks are iterated to update the node mask but only the task
> leaders are required to update the mm. For a non-group leader task, maybe we
> can check if the group leader is in the same cpuset. If so, we can skip the
> mm update. Do we need similar change in cpuset_attach()?

The leader isn't special tho. We just wanna avoid visiting the same mm more
than once, right?

Thanks.

-- 
tejun

Re: [PATCH] cgroup/cpuset: Optimize update_tasks_nodemask()

[Waiman Long]

On 11/23/22 13:54, Tejun Heo wrote:
> On Wed, Nov 23, 2022 at 01:48:46PM -0500, Waiman Long wrote:
>> I think it is an issue anyway if different threads of a process are in
>> different cpusets with different node mask. It is not a configuration that
>> should be used at all.
> Anything memory related is in the same boat and people still use them
> reaching whatever end results they reach. Given the whole thing is pretty
> ill-defined, I don't wanna change the behavior now.
I am just saying that this is not a good config. I don't have any 
intention to change the existing behavior at all.
>
>> This patch makes update_tasks_nodemask() somewhat similar to cpuset_attach()
>> where all tasks are iterated to update the node mask but only the task
>> leaders are required to update the mm. For a non-group leader task, maybe we
>> can check if the group leader is in the same cpuset. If so, we can skip the
>> mm update. Do we need similar change in cpuset_attach()?
> The leader isn't special tho. We just wanna avoid visiting the same mm more
> than once, right?

Right, the group leader is just a marker to make it easier to avoid 
duplicating the work for the same mm. If the group leader happens to be 
in another cpuset, it will suffer some performance consequence.

Cheers,
Longman

Re: [PATCH] cgroup/cpuset: Optimize update_tasks_nodemask()

[Tejun Heo]

On Wed, Nov 23, 2022 at 02:05:59PM -0500, Waiman Long wrote:
> Right, the group leader is just a marker to make it easier to avoid
> duplicating the work for the same mm. If the group leader happens to be in
> another cpuset, it will suffer some performance consequence.

Yeah, I guess that's gonna be good enough for most cases.

Thanks.

-- 
tejun

Re: [PATCH] cgroup/cpuset: Optimize update_tasks_nodemask()

[Waiman Long]

[-- Attachment #1: Type: text/plain, Size: 2016 bytes --]

On 11/23/22 03:21, haifeng.xu wrote:
> When change the 'cpuset.mems' under some cgroup, system will hung
> for a long time. From the dmesg, many processes or theads are
> stuck in fork/exit. The reason is show as follows.
>
> thread A:
> cpuset_write_resmask /* takes cpuset_rwsem */
>    ...
>      update_tasks_nodemask
>        mpol_rebind_mm /* waits mmap_lock */
>
> thread B:
> worker_thread
>    ...
>      cpuset_migrate_mm_workfn
>        do_migrate_pages /* takes mmap_lock */
>
> thread C:
> cgroup_procs_write /* takes cgroup_mutex and cgroup_threadgroup_rwsem */
>    ...
>      cpuset_can_attach
>        percpu_down_write /* waits cpuset_rwsem */
>
> Once update the nodemasks of cpuset, thread A wakes up thread B to
> migrate mm. But when thread A iterates through all tasks, including
> child threads and group leader, it has to wait the mmap_lock which
> has been take by thread B. Unfortunately, thread C wants to migrate
> tasks into cgroup at this moment, it must wait thread A to release
> cpuset_rwsem. If thread B spends much time to migrate mm, the
> fork/exit which acquire cgroup_threadgroup_rwsem also need to
> wait for a long time.
>
> There is no need to migrate the mm of child threads which is
> shared with group leader. Just iterate through the group
> leader only.
>
> Signed-off-by: haifeng.xu <haifeng.xu@shopee.com>
> ---
>   kernel/cgroup/cpuset.c | 3 +++
>   1 file changed, 3 insertions(+)
>
> diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
> index 589827ccda8b..43cbd09546d0 100644
> --- a/kernel/cgroup/cpuset.c
> +++ b/kernel/cgroup/cpuset.c
> @@ -1968,6 +1968,9 @@ static void update_tasks_nodemask(struct cpuset *cs)
>   
>   		cpuset_change_task_nodemask(task, &newmems);
>   
> +		if (!thread_group_leader(task))
> +			continue;
> +
>   		mm = get_task_mm(task);
>   		if (!mm)
>   			continue;

Could you try the attached test patch to see if it can fix your problem? 
Something along the line of this patch will be more acceptable.

Thanks,
Longman


[-- Attachment #2: test.patch --]
[-- Type: text/x-patch, Size: 1275 bytes --]

diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
index b474289c15b8..9c17b6d4877c 100644
--- a/kernel/cgroup/cpuset.c
+++ b/kernel/cgroup/cpuset.c
@@ -1942,6 +1942,7 @@ static void update_tasks_nodemask(struct cpuset *cs)
 	static nodemask_t newmems;	/* protected by cpuset_rwsem */
 	struct css_task_iter it;
 	struct task_struct *task;
+	bool migrate;
 
 	cpuset_being_rebound = cs;		/* causes mpol_dup() rebind */
 
@@ -1957,19 +1958,25 @@ static void update_tasks_nodemask(struct cpuset *cs)
 	 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
 	 * is idempotent.  Also migrate pages in each mm to new nodes.
 	 */
+	migrate = is_memory_migrate(cs);
 	css_task_iter_start(&cs->css, 0, &it);
 	while ((task = css_task_iter_next(&it))) {
 		struct mm_struct *mm;
-		bool migrate;
 
 		cpuset_change_task_nodemask(task, &newmems);
 
+		/*
+		 * Skip mm update if a non group leader task and its group
+		 * leader are in the same cpuset.
+		 */
+		if (!thread_group_leader(task) &&
+		   (task_cs(task->group_leader) == cs))
+			continue;
+
 		mm = get_task_mm(task);
 		if (!mm)
 			continue;
 
-		migrate = is_memory_migrate(cs);
-
 		mpol_rebind_mm(mm, &cs->mems_allowed);
 		if (migrate)
 			cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);

Re: [PATCH] cgroup/cpuset: Optimize update_tasks_nodemask()

[Haifeng Xu]

On 2022/11/24 04:23, Waiman Long wrote:
> On 11/23/22 03:21, haifeng.xu wrote:
>> When change the 'cpuset.mems' under some cgroup, system will hung
>> for a long time. From the dmesg, many processes or theads are
>> stuck in fork/exit. The reason is show as follows.
>>
>> thread A:
>> cpuset_write_resmask /* takes cpuset_rwsem */
>>    ...
>>      update_tasks_nodemask
>>        mpol_rebind_mm /* waits mmap_lock */
>>
>> thread B:
>> worker_thread
>>    ...
>>      cpuset_migrate_mm_workfn
>>        do_migrate_pages /* takes mmap_lock */
>>
>> thread C:
>> cgroup_procs_write /* takes cgroup_mutex and cgroup_threadgroup_rwsem */
>>    ...
>>      cpuset_can_attach
>>        percpu_down_write /* waits cpuset_rwsem */
>>
>> Once update the nodemasks of cpuset, thread A wakes up thread B to
>> migrate mm. But when thread A iterates through all tasks, including
>> child threads and group leader, it has to wait the mmap_lock which
>> has been take by thread B. Unfortunately, thread C wants to migrate
>> tasks into cgroup at this moment, it must wait thread A to release
>> cpuset_rwsem. If thread B spends much time to migrate mm, the
>> fork/exit which acquire cgroup_threadgroup_rwsem also need to
>> wait for a long time.
>>
>> There is no need to migrate the mm of child threads which is
>> shared with group leader. Just iterate through the group
>> leader only.
>>
>> Signed-off-by: haifeng.xu <haifeng.xu@shopee.com>
>> ---
>>   kernel/cgroup/cpuset.c | 3 +++
>>   1 file changed, 3 insertions(+)
>>
>> diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
>> index 589827ccda8b..43cbd09546d0 100644
>> --- a/kernel/cgroup/cpuset.c
>> +++ b/kernel/cgroup/cpuset.c
>> @@ -1968,6 +1968,9 @@ static void update_tasks_nodemask(struct cpuset
>> *cs)
>>             cpuset_change_task_nodemask(task, &newmems);
>>   +        if (!thread_group_leader(task))
>> +            continue;
>> +
>>           mm = get_task_mm(task);
>>           if (!mm)
>>               continue;
> 
> Could you try the attached test patch to see if it can fix your problem?
> Something along the line of this patch will be more acceptable.
> 
> Thanks,
> Longman
> 

Hi, Longman.
Thanks for your patch, but there are still some problems.

1）
  (group leader, node: 0,1)
         cgroup0
         /     \
        /       \
    cgroup1   cgroup2
   (threads)  (threads)

If set node 0 in cgroup1 and node 1 in cgroup2, both of them will update
the mm. And the nodemask of mm depends on who set the node last.

2）
   (process, node: 0,1)
         cgroup0
         /     \
        /       \
    cgroup1   cgroup2
   (node: 0)  (node: 1)

If migrate thread from cgroup0 to cgroup1 or cgroup2, cpuset_attach
won't update the mm. So the nodemask of thread, including mems_allowed
and mempolicy（updated in cpuset_change_task_nodemask）, is different from
the vm_policy in vma(updated in mpol_rebind_mm).


In a word, if threads have different cpusets with different nodemask, it
will cause inconsistent memory behavior.

Thanks,
Haifeng.

Re: [PATCH] cgroup/cpuset: Optimize update_tasks_nodemask()

[Waiman Long]

On 11/23/22 22:33, Haifeng Xu wrote:
>
> On 2022/11/24 04:23, Waiman Long wrote:
>> On 11/23/22 03:21, haifeng.xu wrote:
>>> When change the 'cpuset.mems' under some cgroup, system will hung
>>> for a long time. From the dmesg, many processes or theads are
>>> stuck in fork/exit. The reason is show as follows.
>>>
>>> thread A:
>>> cpuset_write_resmask /* takes cpuset_rwsem */
>>>     ...
>>>       update_tasks_nodemask
>>>         mpol_rebind_mm /* waits mmap_lock */
>>>
>>> thread B:
>>> worker_thread
>>>     ...
>>>       cpuset_migrate_mm_workfn
>>>         do_migrate_pages /* takes mmap_lock */
>>>
>>> thread C:
>>> cgroup_procs_write /* takes cgroup_mutex and cgroup_threadgroup_rwsem */
>>>     ...
>>>       cpuset_can_attach
>>>         percpu_down_write /* waits cpuset_rwsem */
>>>
>>> Once update the nodemasks of cpuset, thread A wakes up thread B to
>>> migrate mm. But when thread A iterates through all tasks, including
>>> child threads and group leader, it has to wait the mmap_lock which
>>> has been take by thread B. Unfortunately, thread C wants to migrate
>>> tasks into cgroup at this moment, it must wait thread A to release
>>> cpuset_rwsem. If thread B spends much time to migrate mm, the
>>> fork/exit which acquire cgroup_threadgroup_rwsem also need to
>>> wait for a long time.
>>>
>>> There is no need to migrate the mm of child threads which is
>>> shared with group leader. Just iterate through the group
>>> leader only.
>>>
>>> Signed-off-by: haifeng.xu <haifeng.xu@shopee.com>
>>> ---
>>>    kernel/cgroup/cpuset.c | 3 +++
>>>    1 file changed, 3 insertions(+)
>>>
>>> diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
>>> index 589827ccda8b..43cbd09546d0 100644
>>> --- a/kernel/cgroup/cpuset.c
>>> +++ b/kernel/cgroup/cpuset.c
>>> @@ -1968,6 +1968,9 @@ static void update_tasks_nodemask(struct cpuset
>>> *cs)
>>>              cpuset_change_task_nodemask(task, &newmems);
>>>    +        if (!thread_group_leader(task))
>>> +            continue;
>>> +
>>>            mm = get_task_mm(task);
>>>            if (!mm)
>>>                continue;
>> Could you try the attached test patch to see if it can fix your problem?
>> Something along the line of this patch will be more acceptable.
>>
>> Thanks,
>> Longman
>>
> Hi, Longman.
> Thanks for your patch, but there are still some problems.
>
> 1）
>    (group leader, node: 0,1)
>           cgroup0
>           /     \
>          /       \
>      cgroup1   cgroup2
>     (threads)  (threads)
>
> If set node 0 in cgroup1 and node 1 in cgroup2, both of them will update
> the mm. And the nodemask of mm depends on who set the node last.

Yes, that is the existing behavior. It was not that well defined in the 
past and so it is somewhat ambiguous as to what we need to do about it.

BTW, cgroup1 has a memory_migrate flag which will force page migration 
if set. I guess you may have it set in your case as it will introduce a 
lot more delay as page migration takes time. That is probably the reason 
why you are seeing a long delay. So one possible solution is to turn 
this flag off. Cgroup v2 doesn't have this flag.

>
> 2）
>     (process, node: 0,1)
>           cgroup0
>           /     \
>          /       \
>      cgroup1   cgroup2
>     (node: 0)  (node: 1)
>
> If migrate thread from cgroup0 to cgroup1 or cgroup2, cpuset_attach
> won't update the mm. So the nodemask of thread, including mems_allowed
> and mempolicy（updated in cpuset_change_task_nodemask）, is different from
> the vm_policy in vma(updated in mpol_rebind_mm).

Yes, that can be the case.

>
>
> In a word, if threads have different cpusets with different nodemask, it
> will cause inconsistent memory behavior.

So do you have suggestion of what we need to do going forward?

Cheers,
Longman

Re: [PATCH] cgroup/cpuset: Optimize update_tasks_nodemask()

[Haifeng Xu]

On 2022/11/24 12:24, Waiman Long wrote:
> On 11/23/22 22:33, Haifeng Xu wrote:
>>
>> On 2022/11/24 04:23, Waiman Long wrote:
>>> On 11/23/22 03:21, haifeng.xu wrote:
>>>> When change the 'cpuset.mems' under some cgroup, system will hung
>>>> for a long time. From the dmesg, many processes or theads are
>>>> stuck in fork/exit. The reason is show as follows.
>>>>
>>>> thread A:
>>>> cpuset_write_resmask /* takes cpuset_rwsem */
>>>>     ...
>>>>       update_tasks_nodemask
>>>>         mpol_rebind_mm /* waits mmap_lock */
>>>>
>>>> thread B:
>>>> worker_thread
>>>>     ...
>>>>       cpuset_migrate_mm_workfn
>>>>         do_migrate_pages /* takes mmap_lock */
>>>>
>>>> thread C:
>>>> cgroup_procs_write /* takes cgroup_mutex and
>>>> cgroup_threadgroup_rwsem */
>>>>     ...
>>>>       cpuset_can_attach
>>>>         percpu_down_write /* waits cpuset_rwsem */
>>>>
>>>> Once update the nodemasks of cpuset, thread A wakes up thread B to
>>>> migrate mm. But when thread A iterates through all tasks, including
>>>> child threads and group leader, it has to wait the mmap_lock which
>>>> has been take by thread B. Unfortunately, thread C wants to migrate
>>>> tasks into cgroup at this moment, it must wait thread A to release
>>>> cpuset_rwsem. If thread B spends much time to migrate mm, the
>>>> fork/exit which acquire cgroup_threadgroup_rwsem also need to
>>>> wait for a long time.
>>>>
>>>> There is no need to migrate the mm of child threads which is
>>>> shared with group leader. Just iterate through the group
>>>> leader only.
>>>>
>>>> Signed-off-by: haifeng.xu <haifeng.xu@shopee.com>
>>>> ---
>>>>    kernel/cgroup/cpuset.c | 3 +++
>>>>    1 file changed, 3 insertions(+)
>>>>
>>>> diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
>>>> index 589827ccda8b..43cbd09546d0 100644
>>>> --- a/kernel/cgroup/cpuset.c
>>>> +++ b/kernel/cgroup/cpuset.c
>>>> @@ -1968,6 +1968,9 @@ static void update_tasks_nodemask(struct cpuset
>>>> *cs)
>>>>              cpuset_change_task_nodemask(task, &newmems);
>>>>    +        if (!thread_group_leader(task))
>>>> +            continue;
>>>> +
>>>>            mm = get_task_mm(task);
>>>>            if (!mm)
>>>>                continue;
>>> Could you try the attached test patch to see if it can fix your problem?
>>> Something along the line of this patch will be more acceptable.
>>>
>>> Thanks,
>>> Longman
>>>
>> Hi, Longman.
>> Thanks for your patch, but there are still some problems.
>>
>> 1）
>>    (group leader, node: 0,1)
>>           cgroup0
>>           /     \
>>          /       \
>>      cgroup1   cgroup2
>>     (threads)  (threads)
>>
>> If set node 0 in cgroup1 and node 1 in cgroup2, both of them will update
>> the mm. And the nodemask of mm depends on who set the node last.
> 
> Yes, that is the existing behavior. It was not that well defined in the
> past and so it is somewhat ambiguous as to what we need to do about it.
> 

The test patch works if the child threads are in same cpuset with group
leader which has same logic with my patch. But if they are in different
cpusets, the test patch will fail because the contention of mmap_lock
still exsits and seems similar to the original logic.

> BTW, cgroup1 has a memory_migrate flag which will force page migration
> if set. I guess you may have it set in your case as it will introduce a
> lot more delay as page migration takes time. That is probably the reason
> why you are seeing a long delay. So one possible solution is to turn
> this flag off. Cgroup v2 doesn't have this flag.
> 

Dou you mean 'CS_MEMORY_MIGRATE'? This flag can be turn off in Cgroup
v1, but it has been set in Cgroup v2 (cpuset_css_alloc) in default and
couldn't be changed.

>>
>> 2）
>>     (process, node: 0,1)
>>           cgroup0
>>           /     \
>>          /       \
>>      cgroup1   cgroup2
>>     (node: 0)  (node: 1)
>>
>> If migrate thread from cgroup0 to cgroup1 or cgroup2, cpuset_attach
>> won't update the mm. So the nodemask of thread, including mems_allowed
>> and mempolicy（updated in cpuset_change_task_nodemask）, is different
>> from
>> the vm_policy in vma(updated in mpol_rebind_mm).
> 
> Yes, that can be the case.
> 
>>
>>
>> In a word, if threads have different cpusets with different nodemask, it
>> will cause inconsistent memory behavior.
> 
> So do you have suggestion of what we need to do going forward?

Should we prevent thread from migrating to those cgroups which have
different nodemask with the cgroup that contains the group leader?

In addition, the group leader and child threads should be in same cgroup
tree, also the level of cgroup containes group leader must be higher
than these cgroups contain child threads, so update_nodemask will work.

Or just disable thread migration in cpuset？It's easy to achieve but will
affect cpu bind.
> 
> Cheers,
> Longman
> 
> 

Re: [PATCH] cgroup/cpuset: Optimize update_tasks_nodemask()

[Waiman Long]

On 11/24/22 02:49, Haifeng Xu wrote:
>
> On 2022/11/24 12:24, Waiman Long wrote:
>> On 11/23/22 22:33, Haifeng Xu wrote:
>>> On 2022/11/24 04:23, Waiman Long wrote:
>>>> On 11/23/22 03:21, haifeng.xu wrote:
>>>>> When change the 'cpuset.mems' under some cgroup, system will hung
>>>>> for a long time. From the dmesg, many processes or theads are
>>>>> stuck in fork/exit. The reason is show as follows.
>>>>>
>>>>> thread A:
>>>>> cpuset_write_resmask /* takes cpuset_rwsem */
>>>>>      ...
>>>>>        update_tasks_nodemask
>>>>>          mpol_rebind_mm /* waits mmap_lock */
>>>>>
>>>>> thread B:
>>>>> worker_thread
>>>>>      ...
>>>>>        cpuset_migrate_mm_workfn
>>>>>          do_migrate_pages /* takes mmap_lock */
>>>>>
>>>>> thread C:
>>>>> cgroup_procs_write /* takes cgroup_mutex and
>>>>> cgroup_threadgroup_rwsem */
>>>>>      ...
>>>>>        cpuset_can_attach
>>>>>          percpu_down_write /* waits cpuset_rwsem */
>>>>>
>>>>> Once update the nodemasks of cpuset, thread A wakes up thread B to
>>>>> migrate mm. But when thread A iterates through all tasks, including
>>>>> child threads and group leader, it has to wait the mmap_lock which
>>>>> has been take by thread B. Unfortunately, thread C wants to migrate
>>>>> tasks into cgroup at this moment, it must wait thread A to release
>>>>> cpuset_rwsem. If thread B spends much time to migrate mm, the
>>>>> fork/exit which acquire cgroup_threadgroup_rwsem also need to
>>>>> wait for a long time.
>>>>>
>>>>> There is no need to migrate the mm of child threads which is
>>>>> shared with group leader. Just iterate through the group
>>>>> leader only.
>>>>>
>>>>> Signed-off-by: haifeng.xu <haifeng.xu@shopee.com>
>>>>> ---
>>>>>     kernel/cgroup/cpuset.c | 3 +++
>>>>>     1 file changed, 3 insertions(+)
>>>>>
>>>>> diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
>>>>> index 589827ccda8b..43cbd09546d0 100644
>>>>> --- a/kernel/cgroup/cpuset.c
>>>>> +++ b/kernel/cgroup/cpuset.c
>>>>> @@ -1968,6 +1968,9 @@ static void update_tasks_nodemask(struct cpuset
>>>>> *cs)
>>>>>               cpuset_change_task_nodemask(task, &newmems);
>>>>>     +        if (!thread_group_leader(task))
>>>>> +            continue;
>>>>> +
>>>>>             mm = get_task_mm(task);
>>>>>             if (!mm)
>>>>>                 continue;
>>>> Could you try the attached test patch to see if it can fix your problem?
>>>> Something along the line of this patch will be more acceptable.
>>>>
>>>> Thanks,
>>>> Longman
>>>>
>>> Hi, Longman.
>>> Thanks for your patch, but there are still some problems.
>>>
>>> 1）
>>>     (group leader, node: 0,1)
>>>            cgroup0
>>>            /     \
>>>           /       \
>>>       cgroup1   cgroup2
>>>      (threads)  (threads)
>>>
>>> If set node 0 in cgroup1 and node 1 in cgroup2, both of them will update
>>> the mm. And the nodemask of mm depends on who set the node last.
>> Yes, that is the existing behavior. It was not that well defined in the
>> past and so it is somewhat ambiguous as to what we need to do about it.
>>
> The test patch works if the child threads are in same cpuset with group
> leader which has same logic with my patch. But if they are in different
> cpusets, the test patch will fail because the contention of mmap_lock
> still exsits and seems similar to the original logic.

That is true. I am thinking about adding a nodemask to mm_struct so that 
we can figure out if we need to propagate the changes down to all the 
VMAs and do the migration. That will enable us to avoid doing wasteful 
work.

Current node mask handling isn't that efficient especially for distros 
that have a relatively large NODES_SHIFT value. Some work may also be 
need in this area.

>> BTW, cgroup1 has a memory_migrate flag which will force page migration
>> if set. I guess you may have it set in your case as it will introduce a
>> lot more delay as page migration takes time. That is probably the reason
>> why you are seeing a long delay. So one possible solution is to turn
>> this flag off. Cgroup v2 doesn't have this flag.
>>
> Dou you mean 'CS_MEMORY_MIGRATE'? This flag can be turn off in Cgroup
> v1, but it has been set in Cgroup v2 (cpuset_css_alloc) in default and
> couldn't be changed.
You are right. Cgroup v2 has CS_MEMORY_MIGRATE enabled by default and 
can't be turned off.
>
>>> 2）
>>>      (process, node: 0,1)
>>>            cgroup0
>>>            /     \
>>>           /       \
>>>       cgroup1   cgroup2
>>>      (node: 0)  (node: 1)
>>>
>>> If migrate thread from cgroup0 to cgroup1 or cgroup2, cpuset_attach
>>> won't update the mm. So the nodemask of thread, including mems_allowed
>>> and mempolicy（updated in cpuset_change_task_nodemask）, is different
>>> from
>>> the vm_policy in vma(updated in mpol_rebind_mm).
>> Yes, that can be the case.
>>
>>>
>>> In a word, if threads have different cpusets with different nodemask, it
>>> will cause inconsistent memory behavior.
>> So do you have suggestion of what we need to do going forward?
> Should we prevent thread from migrating to those cgroups which have
> different nodemask with the cgroup that contains the group leader?
>
> In addition, the group leader and child threads should be in same cgroup
> tree, also the level of cgroup containes group leader must be higher
> than these cgroups contain child threads, so update_nodemask will work.
>
> Or just disable thread migration in cpuset？It's easy to achieve but will
> affect cpu bind.

As said above, my current inclination is to add a nodemask to mm_struct 
and revise the way nodemask is being handled. That will take some time.

Cheers,
Longman

Re: [PATCH] cgroup/cpuset: Optimize update_tasks_nodemask()

[Haifeng Xu]

On 2022/11/25 07:00, Waiman Long wrote:
> On 11/24/22 02:49, Haifeng Xu wrote:
>>
>> On 2022/11/24 12:24, Waiman Long wrote:
>>> On 11/23/22 22:33, Haifeng Xu wrote:
>>>> On 2022/11/24 04:23, Waiman Long wrote:
>>>>> On 11/23/22 03:21, haifeng.xu wrote:
>>>>>> When change the 'cpuset.mems' under some cgroup, system will hung
>>>>>> for a long time. From the dmesg, many processes or theads are
>>>>>> stuck in fork/exit. The reason is show as follows.
>>>>>>
>>>>>> thread A:
>>>>>> cpuset_write_resmask /* takes cpuset_rwsem */
>>>>>>      ...
>>>>>>        update_tasks_nodemask
>>>>>>          mpol_rebind_mm /* waits mmap_lock */
>>>>>>
>>>>>> thread B:
>>>>>> worker_thread
>>>>>>      ...
>>>>>>        cpuset_migrate_mm_workfn
>>>>>>          do_migrate_pages /* takes mmap_lock */
>>>>>>
>>>>>> thread C:
>>>>>> cgroup_procs_write /* takes cgroup_mutex and
>>>>>> cgroup_threadgroup_rwsem */
>>>>>>      ...
>>>>>>        cpuset_can_attach
>>>>>>          percpu_down_write /* waits cpuset_rwsem */
>>>>>>
>>>>>> Once update the nodemasks of cpuset, thread A wakes up thread B to
>>>>>> migrate mm. But when thread A iterates through all tasks, including
>>>>>> child threads and group leader, it has to wait the mmap_lock which
>>>>>> has been take by thread B. Unfortunately, thread C wants to migrate
>>>>>> tasks into cgroup at this moment, it must wait thread A to release
>>>>>> cpuset_rwsem. If thread B spends much time to migrate mm, the
>>>>>> fork/exit which acquire cgroup_threadgroup_rwsem also need to
>>>>>> wait for a long time.
>>>>>>
>>>>>> There is no need to migrate the mm of child threads which is
>>>>>> shared with group leader. Just iterate through the group
>>>>>> leader only.
>>>>>>
>>>>>> Signed-off-by: haifeng.xu <haifeng.xu@shopee.com>
>>>>>> ---
>>>>>>     kernel/cgroup/cpuset.c | 3 +++
>>>>>>     1 file changed, 3 insertions(+)
>>>>>>
>>>>>> diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
>>>>>> index 589827ccda8b..43cbd09546d0 100644
>>>>>> --- a/kernel/cgroup/cpuset.c
>>>>>> +++ b/kernel/cgroup/cpuset.c
>>>>>> @@ -1968,6 +1968,9 @@ static void update_tasks_nodemask(struct cpuset
>>>>>> *cs)
>>>>>>               cpuset_change_task_nodemask(task, &newmems);
>>>>>>     +        if (!thread_group_leader(task))
>>>>>> +            continue;
>>>>>> +
>>>>>>             mm = get_task_mm(task);
>>>>>>             if (!mm)
>>>>>>                 continue;
>>>>> Could you try the attached test patch to see if it can fix your
>>>>> problem?
>>>>> Something along the line of this patch will be more acceptable.
>>>>>
>>>>> Thanks,
>>>>> Longman
>>>>>
>>>> Hi, Longman.
>>>> Thanks for your patch, but there are still some problems.
>>>>
>>>> 1）
>>>>     (group leader, node: 0,1)
>>>>            cgroup0
>>>>            /     \
>>>>           /       \
>>>>       cgroup1   cgroup2
>>>>      (threads)  (threads)
>>>>
>>>> If set node 0 in cgroup1 and node 1 in cgroup2, both of them will
>>>> update
>>>> the mm. And the nodemask of mm depends on who set the node last.
>>> Yes, that is the existing behavior. It was not that well defined in the
>>> past and so it is somewhat ambiguous as to what we need to do about it.
>>>
>> The test patch works if the child threads are in same cpuset with group
>> leader which has same logic with my patch. But if they are in different
>> cpusets, the test patch will fail because the contention of mmap_lock
>> still exsits and seems similar to the original logic.
> 
> That is true. I am thinking about adding a nodemask to mm_struct so that
> we can figure out if we need to propagate the changes down to all the
> VMAs and do the migration. That will enable us to avoid doing wasteful
> work.
> 
> Current node mask handling isn't that efficient especially for distros
> that have a relatively large NODES_SHIFT value. Some work may also be
> need in this area.
> 
>>> BTW, cgroup1 has a memory_migrate flag which will force page migration
>>> if set. I guess you may have it set in your case as it will introduce a
>>> lot more delay as page migration takes time. That is probably the reason
>>> why you are seeing a long delay. So one possible solution is to turn
>>> this flag off. Cgroup v2 doesn't have this flag.
>>>
>> Dou you mean 'CS_MEMORY_MIGRATE'? This flag can be turn off in Cgroup
>> v1, but it has been set in Cgroup v2 (cpuset_css_alloc) in default and
>> couldn't be changed.
> You are right. Cgroup v2 has CS_MEMORY_MIGRATE enabled by default and
> can't be turned off.
>>
>>>> 2）
>>>>      (process, node: 0,1)
>>>>            cgroup0
>>>>            /     \
>>>>           /       \
>>>>       cgroup1   cgroup2
>>>>      (node: 0)  (node: 1)
>>>>
>>>> If migrate thread from cgroup0 to cgroup1 or cgroup2, cpuset_attach
>>>> won't update the mm. So the nodemask of thread, including mems_allowed
>>>> and mempolicy（updated in cpuset_change_task_nodemask）, is different
>>>> from
>>>> the vm_policy in vma(updated in mpol_rebind_mm).
>>> Yes, that can be the case.
>>>
>>>>
>>>> In a word, if threads have different cpusets with different
>>>> nodemask, it
>>>> will cause inconsistent memory behavior.
>>> So do you have suggestion of what we need to do going forward?
>> Should we prevent thread from migrating to those cgroups which have
>> different nodemask with the cgroup that contains the group leader?
>>
>> In addition, the group leader and child threads should be in same cgroup
>> tree, also the level of cgroup containes group leader must be higher
>> than these cgroups contain child threads, so update_nodemask will work.
>>
>> Or just disable thread migration in cpuset？It's easy to achieve but will
>> affect cpu bind.
> 
> As said above, my current inclination is to add a nodemask to mm_struct
> and revise the way nodemask is being handled. That will take some time>
> Cheers,
> Longman
> 

OK, thanks.

Re: [PATCH] cgroup/cpuset: Optimize update_tasks_nodemask()

[Haifeng Xu]

On 2022/11/25 07:00, Waiman Long wrote:
> On 11/24/22 02:49, Haifeng Xu wrote:
>>
>> On 2022/11/24 12:24, Waiman Long wrote:
>>> On 11/23/22 22:33, Haifeng Xu wrote:
>>>> On 2022/11/24 04:23, Waiman Long wrote:
>>>>> On 11/23/22 03:21, haifeng.xu wrote:
>>>>>> When change the 'cpuset.mems' under some cgroup, system will hung
>>>>>> for a long time. From the dmesg, many processes or theads are
>>>>>> stuck in fork/exit. The reason is show as follows.
>>>>>>
>>>>>> thread A:
>>>>>> cpuset_write_resmask /* takes cpuset_rwsem */
>>>>>>      ...
>>>>>>        update_tasks_nodemask
>>>>>>          mpol_rebind_mm /* waits mmap_lock */
>>>>>>
>>>>>> thread B:
>>>>>> worker_thread
>>>>>>      ...
>>>>>>        cpuset_migrate_mm_workfn
>>>>>>          do_migrate_pages /* takes mmap_lock */
>>>>>>
>>>>>> thread C:
>>>>>> cgroup_procs_write /* takes cgroup_mutex and
>>>>>> cgroup_threadgroup_rwsem */
>>>>>>      ...
>>>>>>        cpuset_can_attach
>>>>>>          percpu_down_write /* waits cpuset_rwsem */
>>>>>>
>>>>>> Once update the nodemasks of cpuset, thread A wakes up thread B to
>>>>>> migrate mm. But when thread A iterates through all tasks, including
>>>>>> child threads and group leader, it has to wait the mmap_lock which
>>>>>> has been take by thread B. Unfortunately, thread C wants to migrate
>>>>>> tasks into cgroup at this moment, it must wait thread A to release
>>>>>> cpuset_rwsem. If thread B spends much time to migrate mm, the
>>>>>> fork/exit which acquire cgroup_threadgroup_rwsem also need to
>>>>>> wait for a long time.
>>>>>>
>>>>>> There is no need to migrate the mm of child threads which is
>>>>>> shared with group leader. Just iterate through the group
>>>>>> leader only.
>>>>>>
>>>>>> Signed-off-by: haifeng.xu <haifeng.xu@shopee.com>
>>>>>> ---
>>>>>>     kernel/cgroup/cpuset.c | 3 +++
>>>>>>     1 file changed, 3 insertions(+)
>>>>>>
>>>>>> diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
>>>>>> index 589827ccda8b..43cbd09546d0 100644
>>>>>> --- a/kernel/cgroup/cpuset.c
>>>>>> +++ b/kernel/cgroup/cpuset.c
>>>>>> @@ -1968,6 +1968,9 @@ static void update_tasks_nodemask(struct cpuset
>>>>>> *cs)
>>>>>>               cpuset_change_task_nodemask(task, &newmems);
>>>>>>     +        if (!thread_group_leader(task))
>>>>>> +            continue;
>>>>>> +
>>>>>>             mm = get_task_mm(task);
>>>>>>             if (!mm)
>>>>>>                 continue;
>>>>> Could you try the attached test patch to see if it can fix your problem?
>>>>> Something along the line of this patch will be more acceptable.
>>>>>
>>>>> Thanks,
>>>>> Longman
>>>>>
>>>> Hi, Longman.
>>>> Thanks for your patch, but there are still some problems.
>>>>
>>>> 1）
>>>>     (group leader, node: 0,1)
>>>>            cgroup0
>>>>            /     \
>>>>           /       \
>>>>       cgroup1   cgroup2
>>>>      (threads)  (threads)
>>>>
>>>> If set node 0 in cgroup1 and node 1 in cgroup2, both of them will update
>>>> the mm. And the nodemask of mm depends on who set the node last.
>>> Yes, that is the existing behavior. It was not that well defined in the
>>> past and so it is somewhat ambiguous as to what we need to do about it.
>>>
>> The test patch works if the child threads are in same cpuset with group
>> leader which has same logic with my patch. But if they are in different
>> cpusets, the test patch will fail because the contention of mmap_lock
>> still exsits and seems similar to the original logic.
> 
> That is true. I am thinking about adding a nodemask to mm_struct so that we can figure out if we need to propagate the changes down to all the VMAs and do the migration. That will enable us to avoid doing wasteful work.
> 
> Current node mask handling isn't that efficient especially for distros that have a relatively large NODES_SHIFT value. Some work may also be need in this area.
> 
>>> BTW, cgroup1 has a memory_migrate flag which will force page migration
>>> if set. I guess you may have it set in your case as it will introduce a
>>> lot more delay as page migration takes time. That is probably the reason
>>> why you are seeing a long delay. So one possible solution is to turn
>>> this flag off. Cgroup v2 doesn't have this flag.
>>>
>> Dou you mean 'CS_MEMORY_MIGRATE'? This flag can be turn off in Cgroup
>> v1, but it has been set in Cgroup v2 (cpuset_css_alloc) in default and
>> couldn't be changed.
> You are right. Cgroup v2 has CS_MEMORY_MIGRATE enabled by default and can't be turned off.

Hi, Longman.
'dfl_files' is just a minimal set. Shall we enable memory_migrate feature in Cgroup V2?
So it can be turned off and help to solve the problem.
>>
>>>> 2）
>>>>      (process, node: 0,1)
>>>>            cgroup0
>>>>            /     \
>>>>           /       \
>>>>       cgroup1   cgroup2
>>>>      (node: 0)  (node: 1)
>>>>
>>>> If migrate thread from cgroup0 to cgroup1 or cgroup2, cpuset_attach
>>>> won't update the mm. So the nodemask of thread, including mems_allowed
>>>> and mempolicy（updated in cpuset_change_task_nodemask）, is different
>>>> from
>>>> the vm_policy in vma(updated in mpol_rebind_mm).
>>> Yes, that can be the case.
>>>
>>>>
>>>> In a word, if threads have different cpusets with different nodemask, it
>>>> will cause inconsistent memory behavior.
>>> So do you have suggestion of what we need to do going forward?
>> Should we prevent thread from migrating to those cgroups which have
>> different nodemask with the cgroup that contains the group leader?
>>
>> In addition, the group leader and child threads should be in same cgroup
>> tree, also the level of cgroup containes group leader must be higher
>> than these cgroups contain child threads, so update_nodemask will work.
>>
>> Or just disable thread migration in cpuset？It's easy to achieve but will
>> affect cpu bind.
> 
> As said above, my current inclination is to add a nodemask to mm_struct and revise the way nodemask is being handled. That will take some time.
> 
> Cheers,
> Longman
>