[PATCH 00/10] cpuset: implement sane hierarchy behaviors

[PATCH 00/10] cpuset: implement sane hierarchy behaviors

[Li Zefan]

Currently some cpuset behaviors are not friendly when cpuset is co-mounted
with other cgroup controllers.

Now with this patchset if cpuset is mounted with sane_behavior option, it
behaves differently:

- Tasks will be kept in empty cpusets when hotplug happens and take masks
of ancestors with non-empty cpus/mems, instead of being moved to an ancestor.

- A task can be moved into an empty cpuset, and again it takes masks of
ancestors, so we drop a task into a newly created cgroup without the user
having to do anything for it.

As tasks can reside in empy cpusets, here're some rules:

- They can be moved to another cpuset, regardless it's empty or not.

- Though it takes masks from ancestors, it takes other configs from the
empty cpuset.

- If the ancestors' masks are changed, those tasks will also be updated
to take new masks.

The first 6 patches are actually cleanups which can be applied no matter
if the last 4 patches are acceptable or not.

Li Zefan (10):
      cpuset: remove redundant check in cpuset_cpus_allowed_fallback()
      cpuset: cleanup guarantee_online_{cpus|mems}()
      cpuset: remove unnecessary variable in cpuset_attach()
      cpuset: remove cpuset_test_cpumask()
      cpuset: don't update tasks' cpumask and nodemask in an empty cpuset
      cpuset: re-structure update_cpumask() a bit
      cpuset: record old_mems_allowed in struct cpuset
      cpuset: introduce effective_{cpumask|nodemask}_cpuset()
      cpuset: allow to keep tasks in empty cpusets
      cpuset: allow to move tasks to empty cpusets

--
 kernel/cpuset.c | 363 +++++++++++++++++++++++++++++++++++++++++++++++------------------
 1 file changed, 262 insertions(+), 101 deletions(-)

[PATCH 00/10] cpuset: implement sane hierarchy behaviors

[Li Zefan]

Currently some cpuset behaviors are not friendly when cpuset is co-mounted
with other cgroup controllers.

Now with this patchset if cpuset is mounted with sane_behavior option, it
behaves differently:

- Tasks will be kept in empty cpusets when hotplug happens and take masks
of ancestors with non-empty cpus/mems, instead of being moved to an ancestor.

- A task can be moved into an empty cpuset, and again it takes masks of
ancestors, so we drop a task into a newly created cgroup without the user
having to do anything for it.

As tasks can reside in empy cpusets, here're some rules:

- They can be moved to another cpuset, regardless it's empty or not.

- Though it takes masks from ancestors, it takes other configs from the
empty cpuset.

- If the ancestors' masks are changed, those tasks will also be updated
to take new masks.

The first 6 patches are actually cleanups which can be applied no matter
if the last 4 patches are acceptable or not.

Li Zefan (10):
      cpuset: remove redundant check in cpuset_cpus_allowed_fallback()
      cpuset: cleanup guarantee_online_{cpus|mems}()
      cpuset: remove unnecessary variable in cpuset_attach()
      cpuset: remove cpuset_test_cpumask()
      cpuset: don't update tasks' cpumask and nodemask in an empty cpuset
      cpuset: re-structure update_cpumask() a bit
      cpuset: record old_mems_allowed in struct cpuset
      cpuset: introduce effective_{cpumask|nodemask}_cpuset()
      cpuset: allow to keep tasks in empty cpusets
      cpuset: allow to move tasks to empty cpusets

--
 kernel/cpuset.c | 363 +++++++++++++++++++++++++++++++++++++++++++++++------------------
 1 file changed, 262 insertions(+), 101 deletions(-)

[PATCH 05/10] cpuset: don't update tasks' cpumask and nodemask in an empty cpuset

[Li Zefan]

I think this was introduced unintentionally when cpuset hotplug was
made asynchronous. Fortunately it does no harm, as updating tasks'
cpumask will just return failure and there's a guarantee_online_mems()
when updating nodemask, and then the tasks will be moved to an ancestor
cpuset.

Signed-off-by: Li Zefan <lizefan@huawei.com>
---
 kernel/cpuset.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index 8b0e433..2e4c5af 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -2013,7 +2013,9 @@ static void cpuset_propagate_hotplug_workfn(struct work_struct *work)
 		mutex_lock(&callback_mutex);
 		cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
 		mutex_unlock(&callback_mutex);
-		update_tasks_cpumask(cs, NULL);
+
+		if (!cpumask_empty(cs->cpus_allowed))
+			update_tasks_cpumask(cs, NULL);
 	}
 
 	/* remove offline mems from @cs */
@@ -2022,7 +2024,9 @@ static void cpuset_propagate_hotplug_workfn(struct work_struct *work)
 		mutex_lock(&callback_mutex);
 		nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
 		mutex_unlock(&callback_mutex);
-		update_tasks_nodemask(cs, &tmp_mems, NULL);
+
+		if (!nodes_empty(cs->mems_allowed))
+			update_tasks_nodemask(cs, &tmp_mems, NULL);
 	}
 
 	is_empty = cpumask_empty(cs->cpus_allowed) ||
-- 
1.8.0.2

Re: [PATCH 05/10] cpuset: don't update tasks' cpumask and nodemask in an empty cpuset

[Tejun Heo]

Hello, Li.

On Fri, Apr 19, 2013 at 08:27:05PM +0800, Li Zefan wrote:
> I think this was introduced unintentionally when cpuset hotplug was
> made asynchronous. Fortunately it does no harm, as updating tasks'
> cpumask will just return failure and there's a guarantee_online_mems()
> when updating nodemask, and then the tasks will be moved to an ancestor
> cpuset.

Yeah, which will update the masks to the proper values anyway, so it
was intentionally written that way as it didn't really matter either
way.  I suppose this helps future changes?  Maybe update the
description a bit?

Thanks.

-- 
tejun

Re: [PATCH 05/10] cpuset: don't update tasks' cpumask and nodemask in an empty cpuset

[Tejun Heo]

Hello, Li.

On Fri, Apr 19, 2013 at 08:27:05PM +0800, Li Zefan wrote:
> I think this was introduced unintentionally when cpuset hotplug was
> made asynchronous. Fortunately it does no harm, as updating tasks'
> cpumask will just return failure and there's a guarantee_online_mems()
> when updating nodemask, and then the tasks will be moved to an ancestor
> cpuset.

Yeah, which will update the masks to the proper values anyway, so it
was intentionally written that way as it didn't really matter either
way.  I suppose this helps future changes?  Maybe update the
description a bit?

Thanks.

-- 
tejun

Re: [PATCH 05/10] cpuset: don't update tasks' cpumask and nodemask in an empty cpuset

[Li Zefan]

On 2013/4/20 2:36, Tejun Heo wrote:
> Hello, Li.
> 
> On Fri, Apr 19, 2013 at 08:27:05PM +0800, Li Zefan wrote:
>> I think this was introduced unintentionally when cpuset hotplug was
>> made asynchronous. Fortunately it does no harm, as updating tasks'
>> cpumask will just return failure and there's a guarantee_online_mems()
>> when updating nodemask, and then the tasks will be moved to an ancestor
>> cpuset.
> 
> Yeah, which will update the masks to the proper values anyway, so it
> was intentionally written that way as it didn't really matter either
> way.  I suppose this helps future changes?  Maybe update the
> description a bit?
> 

I was trying to fix a bug: before moving tasks out of empty cpusets,
update_tasks_nodemask() is called, which calls do_migrate_pages(xx, from, to)
finally, from == node_to_be_offlined, to == empty_nodeamsk, so I guess
no pages will be migrated.

Then when those tasks are moving to an ancestor, do_migrate_pages(xx, from, to)
is called again, from == empty_nodemask, to == ancestor's nodemask,
so I guess again no pages will be migrated.

This bug exists even before your changes.

Re: [PATCH 05/10] cpuset: don't update tasks' cpumask and nodemask in an empty cpuset

[Li Zefan]

On 2013/4/20 2:36, Tejun Heo wrote:
> Hello, Li.
> 
> On Fri, Apr 19, 2013 at 08:27:05PM +0800, Li Zefan wrote:
>> I think this was introduced unintentionally when cpuset hotplug was
>> made asynchronous. Fortunately it does no harm, as updating tasks'
>> cpumask will just return failure and there's a guarantee_online_mems()
>> when updating nodemask, and then the tasks will be moved to an ancestor
>> cpuset.
> 
> Yeah, which will update the masks to the proper values anyway, so it
> was intentionally written that way as it didn't really matter either
> way.  I suppose this helps future changes?  Maybe update the
> description a bit?
> 

I was trying to fix a bug: before moving tasks out of empty cpusets,
update_tasks_nodemask() is called, which calls do_migrate_pages(xx, from, to)
finally, from == node_to_be_offlined, to == empty_nodeamsk, so I guess
no pages will be migrated.

Then when those tasks are moving to an ancestor, do_migrate_pages(xx, from, to)
is called again, from == empty_nodemask, to == ancestor's nodemask,
so I guess again no pages will be migrated.

This bug exists even before your changes.

Re: [PATCH 05/10] cpuset: don't update tasks' cpumask and nodemask in an empty cpuset

[Li Zefan]

On 2013/4/20 2:36, Tejun Heo wrote:
> Hello, Li.
> 
> On Fri, Apr 19, 2013 at 08:27:05PM +0800, Li Zefan wrote:
>> I think this was introduced unintentionally when cpuset hotplug was
>> made asynchronous. Fortunately it does no harm, as updating tasks'
>> cpumask will just return failure and there's a guarantee_online_mems()
>> when updating nodemask, and then the tasks will be moved to an ancestor
>> cpuset.
> 
> Yeah, which will update the masks to the proper values anyway, so it
> was intentionally written that way as it didn't really matter either
> way.  I suppose this helps future changes?  Maybe update the
> description a bit?
> 

I was trying to fix a bug: before moving tasks out of empty cpusets,
update_tasks_nodemask() is called, which calls do_migrate_pages(xx, from, to)
finally, from == node_to_be_offlined, to == empty_nodeamsk, so I guess
no pages will be migrated.

Then when those tasks are moving to an ancestor, do_migrate_pages(xx, from, to)
is called again, from == empty_nodemask, to == ancestor's nodemask,
so I guess again no pages will be migrated.

This bug exists even before your changes.

Re: [PATCH 05/10] cpuset: don't update tasks' cpumask and nodemask in an empty cpuset

[Tejun Heo]

Hello,

On Sat, Apr 20, 2013 at 09:06:44AM +0800, Li Zefan wrote:
> I was trying to fix a bug: before moving tasks out of empty cpusets,
> update_tasks_nodemask() is called, which calls do_migrate_pages(xx, from, to)
> finally, from == node_to_be_offlined, to == empty_nodeamsk, so I guess
> no pages will be migrated.
> 
> Then when those tasks are moving to an ancestor, do_migrate_pages(xx, from, to)
> is called again, from == empty_nodemask, to == ancestor's nodemask,
> so I guess again no pages will be migrated.

Oh, I see.  I was a bit confused by the description.  Maybe include
the above?

Thanks.

-- 
tejun

Re: [PATCH 05/10] cpuset: don't update tasks' cpumask and nodemask in an empty cpuset

[Tejun Heo]

Hello,

On Sat, Apr 20, 2013 at 09:06:44AM +0800, Li Zefan wrote:
> I was trying to fix a bug: before moving tasks out of empty cpusets,
> update_tasks_nodemask() is called, which calls do_migrate_pages(xx, from, to)
> finally, from == node_to_be_offlined, to == empty_nodeamsk, so I guess
> no pages will be migrated.
> 
> Then when those tasks are moving to an ancestor, do_migrate_pages(xx, from, to)
> is called again, from == empty_nodemask, to == ancestor's nodemask,
> so I guess again no pages will be migrated.

Oh, I see.  I was a bit confused by the description.  Maybe include
the above?

Thanks.

-- 
tejun

Re: [PATCH 05/10] cpuset: don't update tasks' cpumask and nodemask in an empty cpuset

[Tejun Heo]

Hello,

On Sat, Apr 20, 2013 at 09:06:44AM +0800, Li Zefan wrote:
> I was trying to fix a bug: before moving tasks out of empty cpusets,
> update_tasks_nodemask() is called, which calls do_migrate_pages(xx, from, to)
> finally, from == node_to_be_offlined, to == empty_nodeamsk, so I guess
> no pages will be migrated.
> 
> Then when those tasks are moving to an ancestor, do_migrate_pages(xx, from, to)
> is called again, from == empty_nodemask, to == ancestor's nodemask,
> so I guess again no pages will be migrated.

Oh, I see.  I was a bit confused by the description.  Maybe include
the above?

Thanks.

-- 
tejun

Re: [PATCH 05/10] cpuset: don't update tasks' cpumask and nodemask in an empty cpuset

[Tejun Heo]

Hello, Li.

On Fri, Apr 19, 2013 at 08:27:05PM +0800, Li Zefan wrote:
> I think this was introduced unintentionally when cpuset hotplug was
> made asynchronous. Fortunately it does no harm, as updating tasks'
> cpumask will just return failure and there's a guarantee_online_mems()
> when updating nodemask, and then the tasks will be moved to an ancestor
> cpuset.

Yeah, which will update the masks to the proper values anyway, so it
was intentionally written that way as it didn't really matter either
way.  I suppose this helps future changes?  Maybe update the
description a bit?

Thanks.

-- 
tejun

[PATCH 01/10] cpuset: remove redundant check in cpuset_cpus_allowed_fallback()

[Li Zefan]

task_cs() will never return NULL.

Signed-off-by: Li Zefan <lizefan-hv44wF8Li93QT0dZR+AlfA@public.gmane.org>
---
 kernel/cpuset.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index 866d78e..4a62dcc 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -2259,8 +2259,7 @@ void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
 
 	rcu_read_lock();
 	cs = task_cs(tsk);
-	if (cs)
-		do_set_cpus_allowed(tsk, cs->cpus_allowed);
+	do_set_cpus_allowed(tsk, cs->cpus_allowed);
 	rcu_read_unlock();
 
 	/*
-- 
1.8.0.2

[PATCH 01/10] cpuset: remove redundant check in cpuset_cpus_allowed_fallback()

[Li Zefan]

task_cs() will never return NULL.

Signed-off-by: Li Zefan <lizefan-hv44wF8Li93QT0dZR+AlfA@public.gmane.org>
---
 kernel/cpuset.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index 866d78e..4a62dcc 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -2259,8 +2259,7 @@ void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
 
 	rcu_read_lock();
 	cs = task_cs(tsk);
-	if (cs)
-		do_set_cpus_allowed(tsk, cs->cpus_allowed);
+	do_set_cpus_allowed(tsk, cs->cpus_allowed);
 	rcu_read_unlock();
 
 	/*
-- 
1.8.0.2

[PATCH 01/10] cpuset: remove redundant check in cpuset_cpus_allowed_fallback()

[Li Zefan]

task_cs() will never return NULL.

Signed-off-by: Li Zefan <lizefan@huawei.com>
---
 kernel/cpuset.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index 866d78e..4a62dcc 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -2259,8 +2259,7 @@ void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
 
 	rcu_read_lock();
 	cs = task_cs(tsk);
-	if (cs)
-		do_set_cpus_allowed(tsk, cs->cpus_allowed);
+	do_set_cpus_allowed(tsk, cs->cpus_allowed);
 	rcu_read_unlock();
 
 	/*
-- 
1.8.0.2

[PATCH 02/10] cpuset: cleanup guarantee_online_{cpus|mems}()

[Li Zefan]

- We never pass a NULL @cs to these functions.
- The top cpuset always has some online cpus/mems.

Signed-off-by: Li Zefan <lizefan-hv44wF8Li93QT0dZR+AlfA@public.gmane.org>
---
 kernel/cpuset.c | 29 +++++++----------------------
 1 file changed, 7 insertions(+), 22 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index 4a62dcc..6f9d5e0 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -304,53 +304,38 @@ static struct file_system_type cpuset_fs_type = {
 /*
  * Return in pmask the portion of a cpusets's cpus_allowed that
  * are online.  If none are online, walk up the cpuset hierarchy
- * until we find one that does have some online cpus.  If we get
- * all the way to the top and still haven't found any online cpus,
- * return cpu_online_mask.  Or if passed a NULL cs from an exit'ing
- * task, return cpu_online_mask.
+ * until we find one that does have some online cpus.  The top
+ * cpuset always has some cpus online.
  *
  * One way or another, we guarantee to return some non-empty subset
  * of cpu_online_mask.
  *
  * Call with callback_mutex held.
  */
-
 static void guarantee_online_cpus(const struct cpuset *cs,
 				  struct cpumask *pmask)
 {
-	while (cs && !cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
+	while (!cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
 		cs = parent_cs(cs);
-	if (cs)
-		cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
-	else
-		cpumask_copy(pmask, cpu_online_mask);
-	BUG_ON(!cpumask_intersects(pmask, cpu_online_mask));
+	cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
 }
 
 /*
  * Return in *pmask the portion of a cpusets's mems_allowed that
  * are online, with memory.  If none are online with memory, walk
  * up the cpuset hierarchy until we find one that does have some
- * online mems.  If we get all the way to the top and still haven't
- * found any online mems, return node_states[N_MEMORY].
+ * online mems.  The top cpuset always has some mems online.
  *
  * One way or another, we guarantee to return some non-empty subset
  * of node_states[N_MEMORY].
  *
  * Call with callback_mutex held.
  */
-
 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
 {
-	while (cs && !nodes_intersects(cs->mems_allowed,
-					node_states[N_MEMORY]))
+	while (!nodes_intersects(cs->mems_allowed, node_states[N_MEMORY]))
 		cs = parent_cs(cs);
-	if (cs)
-		nodes_and(*pmask, cs->mems_allowed,
-					node_states[N_MEMORY]);
-	else
-		*pmask = node_states[N_MEMORY];
-	BUG_ON(!nodes_intersects(*pmask, node_states[N_MEMORY]));
+	nodes_and(*pmask, cs->mems_allowed, node_states[N_MEMORY]);
 }
 
 /*
-- 
1.8.0.2

[PATCH 02/10] cpuset: cleanup guarantee_online_{cpus|mems}()

[Li Zefan]

- We never pass a NULL @cs to these functions.
- The top cpuset always has some online cpus/mems.

Signed-off-by: Li Zefan <lizefan@huawei.com>
---
 kernel/cpuset.c | 29 +++++++----------------------
 1 file changed, 7 insertions(+), 22 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index 4a62dcc..6f9d5e0 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -304,53 +304,38 @@ static struct file_system_type cpuset_fs_type = {
 /*
  * Return in pmask the portion of a cpusets's cpus_allowed that
  * are online.  If none are online, walk up the cpuset hierarchy
- * until we find one that does have some online cpus.  If we get
- * all the way to the top and still haven't found any online cpus,
- * return cpu_online_mask.  Or if passed a NULL cs from an exit'ing
- * task, return cpu_online_mask.
+ * until we find one that does have some online cpus.  The top
+ * cpuset always has some cpus online.
  *
  * One way or another, we guarantee to return some non-empty subset
  * of cpu_online_mask.
  *
  * Call with callback_mutex held.
  */
-
 static void guarantee_online_cpus(const struct cpuset *cs,
 				  struct cpumask *pmask)
 {
-	while (cs && !cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
+	while (!cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
 		cs = parent_cs(cs);
-	if (cs)
-		cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
-	else
-		cpumask_copy(pmask, cpu_online_mask);
-	BUG_ON(!cpumask_intersects(pmask, cpu_online_mask));
+	cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
 }
 
 /*
  * Return in *pmask the portion of a cpusets's mems_allowed that
  * are online, with memory.  If none are online with memory, walk
  * up the cpuset hierarchy until we find one that does have some
- * online mems.  If we get all the way to the top and still haven't
- * found any online mems, return node_states[N_MEMORY].
+ * online mems.  The top cpuset always has some mems online.
  *
  * One way or another, we guarantee to return some non-empty subset
  * of node_states[N_MEMORY].
  *
  * Call with callback_mutex held.
  */
-
 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
 {
-	while (cs && !nodes_intersects(cs->mems_allowed,
-					node_states[N_MEMORY]))
+	while (!nodes_intersects(cs->mems_allowed, node_states[N_MEMORY]))
 		cs = parent_cs(cs);
-	if (cs)
-		nodes_and(*pmask, cs->mems_allowed,
-					node_states[N_MEMORY]);
-	else
-		*pmask = node_states[N_MEMORY];
-	BUG_ON(!nodes_intersects(*pmask, node_states[N_MEMORY]));
+	nodes_and(*pmask, cs->mems_allowed, node_states[N_MEMORY]);
 }
 
 /*
-- 
1.8.0.2

[PATCH 03/10] cpuset: remove unnecessary variable in cpuset_attach()

[Li Zefan]

We can just use oldcs->mems_allowed.

Signed-off-by: Li Zefan <lizefan-hv44wF8Li93QT0dZR+AlfA@public.gmane.org>
---
 kernel/cpuset.c | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index 6f9d5e0..66302a9 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -1406,8 +1406,7 @@ static cpumask_var_t cpus_attach;
 
 static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 {
-	/* static bufs protected by cpuset_mutex */
-	static nodemask_t cpuset_attach_nodemask_from;
+	/* static buf protected by cpuset_mutex */
 	static nodemask_t cpuset_attach_nodemask_to;
 	struct mm_struct *mm;
 	struct task_struct *task;
@@ -1441,13 +1440,12 @@ static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 	 * Change mm, possibly for multiple threads in a threadgroup. This is
 	 * expensive and may sleep.
 	 */
-	cpuset_attach_nodemask_from = oldcs->mems_allowed;
 	cpuset_attach_nodemask_to = cs->mems_allowed;
 	mm = get_task_mm(leader);
 	if (mm) {
 		mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
 		if (is_memory_migrate(cs))
-			cpuset_migrate_mm(mm, &cpuset_attach_nodemask_from,
+			cpuset_migrate_mm(mm, &oldcs->mems_allowed,
 					  &cpuset_attach_nodemask_to);
 		mmput(mm);
 	}
-- 
1.8.0.2

[PATCH 03/10] cpuset: remove unnecessary variable in cpuset_attach()

[Li Zefan]

We can just use oldcs->mems_allowed.

Signed-off-by: Li Zefan <lizefan-hv44wF8Li93QT0dZR+AlfA@public.gmane.org>
---
 kernel/cpuset.c | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index 6f9d5e0..66302a9 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -1406,8 +1406,7 @@ static cpumask_var_t cpus_attach;
 
 static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 {
-	/* static bufs protected by cpuset_mutex */
-	static nodemask_t cpuset_attach_nodemask_from;
+	/* static buf protected by cpuset_mutex */
 	static nodemask_t cpuset_attach_nodemask_to;
 	struct mm_struct *mm;
 	struct task_struct *task;
@@ -1441,13 +1440,12 @@ static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 	 * Change mm, possibly for multiple threads in a threadgroup. This is
 	 * expensive and may sleep.
 	 */
-	cpuset_attach_nodemask_from = oldcs->mems_allowed;
 	cpuset_attach_nodemask_to = cs->mems_allowed;
 	mm = get_task_mm(leader);
 	if (mm) {
 		mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
 		if (is_memory_migrate(cs))
-			cpuset_migrate_mm(mm, &cpuset_attach_nodemask_from,
+			cpuset_migrate_mm(mm, &oldcs->mems_allowed,
 					  &cpuset_attach_nodemask_to);
 		mmput(mm);
 	}
-- 
1.8.0.2

[PATCH 03/10] cpuset: remove unnecessary variable in cpuset_attach()

[Li Zefan]

We can just use oldcs->mems_allowed.

Signed-off-by: Li Zefan <lizefan@huawei.com>
---
 kernel/cpuset.c | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index 6f9d5e0..66302a9 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -1406,8 +1406,7 @@ static cpumask_var_t cpus_attach;
 
 static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 {
-	/* static bufs protected by cpuset_mutex */
-	static nodemask_t cpuset_attach_nodemask_from;
+	/* static buf protected by cpuset_mutex */
 	static nodemask_t cpuset_attach_nodemask_to;
 	struct mm_struct *mm;
 	struct task_struct *task;
@@ -1441,13 +1440,12 @@ static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 	 * Change mm, possibly for multiple threads in a threadgroup. This is
 	 * expensive and may sleep.
 	 */
-	cpuset_attach_nodemask_from = oldcs->mems_allowed;
 	cpuset_attach_nodemask_to = cs->mems_allowed;
 	mm = get_task_mm(leader);
 	if (mm) {
 		mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
 		if (is_memory_migrate(cs))
-			cpuset_migrate_mm(mm, &cpuset_attach_nodemask_from,
+			cpuset_migrate_mm(mm, &oldcs->mems_allowed,
 					  &cpuset_attach_nodemask_to);
 		mmput(mm);
 	}
-- 
1.8.0.2

[PATCH 04/10] cpuset: remove cpuset_test_cpumask()

[Li Zefan]

The test is done in set_cpus_allowed_ptr(), so it's redundant.

Signed-off-by: Li Zefan <lizefan-hv44wF8Li93QT0dZR+AlfA@public.gmane.org>
---
 kernel/cpuset.c | 19 +------------------
 1 file changed, 1 insertion(+), 18 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index 66302a9..8b0e433 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -783,23 +783,6 @@ void rebuild_sched_domains(void)
 }
 
 /**
- * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
- * @tsk: task to test
- * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
- *
- * Call with cpuset_mutex held.  May take callback_mutex during call.
- * Called for each task in a cgroup by cgroup_scan_tasks().
- * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
- * words, if its mask is not equal to its cpuset's mask).
- */
-static int cpuset_test_cpumask(struct task_struct *tsk,
-			       struct cgroup_scanner *scan)
-{
-	return !cpumask_equal(&tsk->cpus_allowed,
-			(cgroup_cs(scan->cg))->cpus_allowed);
-}
-
-/**
  * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
  * @tsk: task to test
  * @scan: struct cgroup_scanner containing the cgroup of the task
@@ -834,7 +817,7 @@ static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
 	struct cgroup_scanner scan;
 
 	scan.cg = cs->css.cgroup;
-	scan.test_task = cpuset_test_cpumask;
+	scan.test_task = NULL;
 	scan.process_task = cpuset_change_cpumask;
 	scan.heap = heap;
 	cgroup_scan_tasks(&scan);
-- 
1.8.0.2

[PATCH 04/10] cpuset: remove cpuset_test_cpumask()

[Li Zefan]

The test is done in set_cpus_allowed_ptr(), so it's redundant.

Signed-off-by: Li Zefan <lizefan@huawei.com>
---
 kernel/cpuset.c | 19 +------------------
 1 file changed, 1 insertion(+), 18 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index 66302a9..8b0e433 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -783,23 +783,6 @@ void rebuild_sched_domains(void)
 }
 
 /**
- * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
- * @tsk: task to test
- * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
- *
- * Call with cpuset_mutex held.  May take callback_mutex during call.
- * Called for each task in a cgroup by cgroup_scan_tasks().
- * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
- * words, if its mask is not equal to its cpuset's mask).
- */
-static int cpuset_test_cpumask(struct task_struct *tsk,
-			       struct cgroup_scanner *scan)
-{
-	return !cpumask_equal(&tsk->cpus_allowed,
-			(cgroup_cs(scan->cg))->cpus_allowed);
-}
-
-/**
  * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
  * @tsk: task to test
  * @scan: struct cgroup_scanner containing the cgroup of the task
@@ -834,7 +817,7 @@ static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
 	struct cgroup_scanner scan;
 
 	scan.cg = cs->css.cgroup;
-	scan.test_task = cpuset_test_cpumask;
+	scan.test_task = NULL;
 	scan.process_task = cpuset_change_cpumask;
 	scan.heap = heap;
 	cgroup_scan_tasks(&scan);
-- 
1.8.0.2

Re: [PATCH 04/10] cpuset: remove cpuset_test_cpumask()

[Tejun Heo]

On Fri, Apr 19, 2013 at 08:26:15PM +0800, Li Zefan wrote:
> The test is done in set_cpus_allowed_ptr(), so it's redundant.
> 
> Signed-off-by: Li Zefan <lizefan-hv44wF8Li93QT0dZR+AlfA@public.gmane.org>

For 0001-0004,

 Acked-by: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>

I'll apply these into for-3.11 once v3.10-rc1 drops.

Thanks.

-- 
tejun

Re: [PATCH 04/10] cpuset: remove cpuset_test_cpumask()

[Tejun Heo]

On Fri, Apr 19, 2013 at 08:26:15PM +0800, Li Zefan wrote:
> The test is done in set_cpus_allowed_ptr(), so it's redundant.
> 
> Signed-off-by: Li Zefan <lizefan@huawei.com>

For 0001-0004,

 Acked-by: Tejun Heo <tj@kernel.org>

I'll apply these into for-3.11 once v3.10-rc1 drops.

Thanks.

-- 
tejun

Re: [PATCH 04/10] cpuset: remove cpuset_test_cpumask()

[Tejun Heo]

On Fri, Apr 19, 2013 at 08:26:15PM +0800, Li Zefan wrote:
> The test is done in set_cpus_allowed_ptr(), so it's redundant.
> 
> Signed-off-by: Li Zefan <lizefan-hv44wF8Li93QT0dZR+AlfA@public.gmane.org>

For 0001-0004,

 Acked-by: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>

I'll apply these into for-3.11 once v3.10-rc1 drops.

Thanks.

-- 
tejun

[PATCH 05/10] cpuset: don't update tasks' cpumask and nodemask in an empty cpuset

[Li Zefan]

I think this was introduced unintentionally when cpuset hotplug was
made asynchronous. Fortunately it does no harm, as updating tasks'
cpumask will just return failure and there's a guarantee_online_mems()
when updating nodemask, and then the tasks will be moved to an ancestor
cpuset.

Signed-off-by: Li Zefan <lizefan-hv44wF8Li93QT0dZR+AlfA@public.gmane.org>
---
 kernel/cpuset.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index 8b0e433..2e4c5af 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -2013,7 +2013,9 @@ static void cpuset_propagate_hotplug_workfn(struct work_struct *work)
 		mutex_lock(&callback_mutex);
 		cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
 		mutex_unlock(&callback_mutex);
-		update_tasks_cpumask(cs, NULL);
+
+		if (!cpumask_empty(cs->cpus_allowed))
+			update_tasks_cpumask(cs, NULL);
 	}
 
 	/* remove offline mems from @cs */
@@ -2022,7 +2024,9 @@ static void cpuset_propagate_hotplug_workfn(struct work_struct *work)
 		mutex_lock(&callback_mutex);
 		nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
 		mutex_unlock(&callback_mutex);
-		update_tasks_nodemask(cs, &tmp_mems, NULL);
+
+		if (!nodes_empty(cs->mems_allowed))
+			update_tasks_nodemask(cs, &tmp_mems, NULL);
 	}
 
 	is_empty = cpumask_empty(cs->cpus_allowed) ||
-- 
1.8.0.2

[PATCH 06/10] cpuset: re-structure update_cpumask() a bit

[Li Zefan]

Check if cpus_allowed is to be changed before calling validate_change().

This won't change any behavior, but later it will allow us to do this:

 # mkdir /cpuset/child
 # echo $$ > /cpuset/child/tasks	/* empty cpuset */
 # echo > /cpuset/child/cpuset.cpus	/* do nothing, won't fail */

Without this patch, the last operation will fail.

Signed-off-by: Li Zefan <lizefan-hv44wF8Li93QT0dZR+AlfA@public.gmane.org>
---
 kernel/cpuset.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index 2e4c5af..63a5343 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -855,14 +855,15 @@ static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
 		if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
 			return -EINVAL;
 	}
-	retval = validate_change(cs, trialcs);
-	if (retval < 0)
-		return retval;
 
 	/* Nothing to do if the cpus didn't change */
 	if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
 		return 0;
 
+	retval = validate_change(cs, trialcs);
+	if (retval < 0)
+		return retval;
+
 	retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
 	if (retval)
 		return retval;
-- 
1.8.0.2

[PATCH 06/10] cpuset: re-structure update_cpumask() a bit

[Li Zefan]

Check if cpus_allowed is to be changed before calling validate_change().

This won't change any behavior, but later it will allow us to do this:

 # mkdir /cpuset/child
 # echo $$ > /cpuset/child/tasks	/* empty cpuset */
 # echo > /cpuset/child/cpuset.cpus	/* do nothing, won't fail */

Without this patch, the last operation will fail.

Signed-off-by: Li Zefan <lizefan@huawei.com>
---
 kernel/cpuset.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index 2e4c5af..63a5343 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -855,14 +855,15 @@ static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
 		if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
 			return -EINVAL;
 	}
-	retval = validate_change(cs, trialcs);
-	if (retval < 0)
-		return retval;
 
 	/* Nothing to do if the cpus didn't change */
 	if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
 		return 0;
 
+	retval = validate_change(cs, trialcs);
+	if (retval < 0)
+		return retval;
+
 	retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
 	if (retval)
 		return retval;
-- 
1.8.0.2

[PATCH 06/10] cpuset: re-structure update_cpumask() a bit

[Li Zefan]

Check if cpus_allowed is to be changed before calling validate_change().

This won't change any behavior, but later it will allow us to do this:

 # mkdir /cpuset/child
 # echo $$ > /cpuset/child/tasks	/* empty cpuset */
 # echo > /cpuset/child/cpuset.cpus	/* do nothing, won't fail */

Without this patch, the last operation will fail.

Signed-off-by: Li Zefan <lizefan-hv44wF8Li93QT0dZR+AlfA@public.gmane.org>
---
 kernel/cpuset.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index 2e4c5af..63a5343 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -855,14 +855,15 @@ static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
 		if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
 			return -EINVAL;
 	}
-	retval = validate_change(cs, trialcs);
-	if (retval < 0)
-		return retval;
 
 	/* Nothing to do if the cpus didn't change */
 	if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
 		return 0;
 
+	retval = validate_change(cs, trialcs);
+	if (retval < 0)
+		return retval;
+
 	retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
 	if (retval)
 		return retval;
-- 
1.8.0.2

[PATCH 07/10] cpuset: record old_mems_allowed in struct cpuset

[Li Zefan]

When we update a cpuset's mems_allowed and thus update tasks'
mems_allowed, it's required to pass the old mems_allowed and new
mems_allowed to cpuset_migrate_mm().

Currently we save old mems_allowed in a temp local variable before
changing cpuset->mems_allowed. This patch changes it by saving
old mems_allowed in cpuset->old_mems_allowed.

This currently won't change any behavior, but it will later allow
us to keep tasks in empty cpusets.

Signed-off-by: Li Zefan <lizefan-hv44wF8Li93QT0dZR+AlfA@public.gmane.org>
---
 kernel/cpuset.c | 62 +++++++++++++++++++++++++++++++++------------------------
 1 file changed, 36 insertions(+), 26 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index 63a5343..c4f9ebd 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -87,6 +87,18 @@ struct cpuset {
 	cpumask_var_t cpus_allowed;	/* CPUs allowed to tasks in cpuset */
 	nodemask_t mems_allowed;	/* Memory Nodes allowed to tasks */
 
+	/*
+	 * This is old Memory Nodes tasks took on.
+	 *
+	 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
+	 * - A new cpuset's old_mems_allowed is initialized when some
+	 *   task is moved into it.
+	 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
+	 *   cpuset.mems_allowed and have tasks' nodemask updated, and
+	 *   then old_mems_allowed is updated to mems_allowed.
+	 */
+	nodemask_t old_mems_allowed;
+
 	struct fmeter fmeter;		/* memory_pressure filter */
 
 	/*
@@ -975,16 +987,12 @@ static void cpuset_change_task_nodemask(struct task_struct *tsk,
 static void cpuset_change_nodemask(struct task_struct *p,
 				   struct cgroup_scanner *scan)
 {
+	struct cpuset *cs = cgroup_cs(scan->cg);
 	struct mm_struct *mm;
-	struct cpuset *cs;
 	int migrate;
-	const nodemask_t *oldmem = scan->data;
-	static nodemask_t newmems;	/* protected by cpuset_mutex */
-
-	cs = cgroup_cs(scan->cg);
-	guarantee_online_mems(cs, &newmems);
+	nodemask_t *newmems = scan->data;
 
-	cpuset_change_task_nodemask(p, &newmems);
+	cpuset_change_task_nodemask(p, newmems);
 
 	mm = get_task_mm(p);
 	if (!mm)
@@ -994,7 +1002,7 @@ static void cpuset_change_nodemask(struct task_struct *p,
 
 	mpol_rebind_mm(mm, &cs->mems_allowed);
 	if (migrate)
-		cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
+		cpuset_migrate_mm(mm, &cs->old_mems_allowed, newmems);
 	mmput(mm);
 }
 
@@ -1003,25 +1011,26 @@ static void *cpuset_being_rebound;
 /**
  * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
  * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
- * @oldmem: old mems_allowed of cpuset cs
  * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
  *
  * Called with cpuset_mutex held
  * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
  * if @heap != NULL.
  */
-static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
-				 struct ptr_heap *heap)
+static void update_tasks_nodemask(struct cpuset *cs, struct ptr_heap *heap)
 {
+	static nodemask_t newmems;	/* protected by cpuset_mutex */
 	struct cgroup_scanner scan;
 
 	cpuset_being_rebound = cs;		/* causes mpol_dup() rebind */
 
+	guarantee_online_mems(cs, &newmems);
+
 	scan.cg = cs->css.cgroup;
 	scan.test_task = NULL;
 	scan.process_task = cpuset_change_nodemask;
 	scan.heap = heap;
-	scan.data = (nodemask_t *)oldmem;
+	scan.data = &newmems;
 
 	/*
 	 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
@@ -1035,6 +1044,12 @@ static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
 	 */
 	cgroup_scan_tasks(&scan);
 
+	/*
+	 * All the tasks' nodemasks have been updated, update
+	 * cs->old_mems_allowed.
+	 */
+	cs->old_mems_allowed = newmems;
+
 	/* We're done rebinding vmas to this cpuset's new mems_allowed. */
 	cpuset_being_rebound = NULL;
 }
@@ -1055,13 +1070,9 @@ static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
 			   const char *buf)
 {
-	NODEMASK_ALLOC(nodemask_t, oldmem, GFP_KERNEL);
 	int retval;
 	struct ptr_heap heap;
 
-	if (!oldmem)
-		return -ENOMEM;
-
 	/*
 	 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
 	 * it's read-only
@@ -1090,8 +1101,8 @@ static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
 			goto done;
 		}
 	}
-	*oldmem = cs->mems_allowed;
-	if (nodes_equal(*oldmem, trialcs->mems_allowed)) {
+
+	if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
 		retval = 0;		/* Too easy - nothing to do */
 		goto done;
 	}
@@ -1107,11 +1118,10 @@ static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
 	cs->mems_allowed = trialcs->mems_allowed;
 	mutex_unlock(&callback_mutex);
 
-	update_tasks_nodemask(cs, oldmem, &heap);
+	update_tasks_nodemask(cs, &heap);
 
 	heap_free(&heap);
 done:
-	NODEMASK_FREE(oldmem);
 	return retval;
 }
 
@@ -1424,7 +1434,6 @@ static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 	 * Change mm, possibly for multiple threads in a threadgroup. This is
 	 * expensive and may sleep.
 	 */
-	cpuset_attach_nodemask_to = cs->mems_allowed;
 	mm = get_task_mm(leader);
 	if (mm) {
 		mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
@@ -1434,6 +1443,8 @@ static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 		mmput(mm);
 	}
 
+	cs->old_mems_allowed = cpuset_attach_nodemask_to;
+
 	cs->attach_in_progress--;
 
 	/*
@@ -2000,7 +2011,7 @@ static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
 static void cpuset_propagate_hotplug_workfn(struct work_struct *work)
 {
 	static cpumask_t off_cpus;
-	static nodemask_t off_mems, tmp_mems;
+	static nodemask_t off_mems;
 	struct cpuset *cs = container_of(work, struct cpuset, hotplug_work);
 	bool is_empty;
 
@@ -2021,13 +2032,12 @@ static void cpuset_propagate_hotplug_workfn(struct work_struct *work)
 
 	/* remove offline mems from @cs */
 	if (!nodes_empty(off_mems)) {
-		tmp_mems = cs->mems_allowed;
 		mutex_lock(&callback_mutex);
 		nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
 		mutex_unlock(&callback_mutex);
 
 		if (!nodes_empty(cs->mems_allowed))
-			update_tasks_nodemask(cs, &tmp_mems, NULL);
+			update_tasks_nodemask(cs, NULL);
 	}
 
 	is_empty = cpumask_empty(cs->cpus_allowed) ||
@@ -2119,11 +2129,10 @@ static void cpuset_hotplug_workfn(struct work_struct *work)
 
 	/* synchronize mems_allowed to N_MEMORY */
 	if (mems_updated) {
-		tmp_mems = top_cpuset.mems_allowed;
 		mutex_lock(&callback_mutex);
 		top_cpuset.mems_allowed = new_mems;
 		mutex_unlock(&callback_mutex);
-		update_tasks_nodemask(&top_cpuset, &tmp_mems, NULL);
+		update_tasks_nodemask(&top_cpuset, NULL);
 	}
 
 	/* if cpus or mems went down, we need to propagate to descendants */
@@ -2196,6 +2205,7 @@ void __init cpuset_init_smp(void)
 {
 	cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
 	top_cpuset.mems_allowed = node_states[N_MEMORY];
+	top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
 
 	hotplug_memory_notifier(cpuset_track_online_nodes, 10);
 
-- 
1.8.0.2

[PATCH 07/10] cpuset: record old_mems_allowed in struct cpuset

[Li Zefan]

When we update a cpuset's mems_allowed and thus update tasks'
mems_allowed, it's required to pass the old mems_allowed and new
mems_allowed to cpuset_migrate_mm().

Currently we save old mems_allowed in a temp local variable before
changing cpuset->mems_allowed. This patch changes it by saving
old mems_allowed in cpuset->old_mems_allowed.

This currently won't change any behavior, but it will later allow
us to keep tasks in empty cpusets.

Signed-off-by: Li Zefan <lizefan@huawei.com>
---
 kernel/cpuset.c | 62 +++++++++++++++++++++++++++++++++------------------------
 1 file changed, 36 insertions(+), 26 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index 63a5343..c4f9ebd 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -87,6 +87,18 @@ struct cpuset {
 	cpumask_var_t cpus_allowed;	/* CPUs allowed to tasks in cpuset */
 	nodemask_t mems_allowed;	/* Memory Nodes allowed to tasks */
 
+	/*
+	 * This is old Memory Nodes tasks took on.
+	 *
+	 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
+	 * - A new cpuset's old_mems_allowed is initialized when some
+	 *   task is moved into it.
+	 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
+	 *   cpuset.mems_allowed and have tasks' nodemask updated, and
+	 *   then old_mems_allowed is updated to mems_allowed.
+	 */
+	nodemask_t old_mems_allowed;
+
 	struct fmeter fmeter;		/* memory_pressure filter */
 
 	/*
@@ -975,16 +987,12 @@ static void cpuset_change_task_nodemask(struct task_struct *tsk,
 static void cpuset_change_nodemask(struct task_struct *p,
 				   struct cgroup_scanner *scan)
 {
+	struct cpuset *cs = cgroup_cs(scan->cg);
 	struct mm_struct *mm;
-	struct cpuset *cs;
 	int migrate;
-	const nodemask_t *oldmem = scan->data;
-	static nodemask_t newmems;	/* protected by cpuset_mutex */
-
-	cs = cgroup_cs(scan->cg);
-	guarantee_online_mems(cs, &newmems);
+	nodemask_t *newmems = scan->data;
 
-	cpuset_change_task_nodemask(p, &newmems);
+	cpuset_change_task_nodemask(p, newmems);
 
 	mm = get_task_mm(p);
 	if (!mm)
@@ -994,7 +1002,7 @@ static void cpuset_change_nodemask(struct task_struct *p,
 
 	mpol_rebind_mm(mm, &cs->mems_allowed);
 	if (migrate)
-		cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
+		cpuset_migrate_mm(mm, &cs->old_mems_allowed, newmems);
 	mmput(mm);
 }
 
@@ -1003,25 +1011,26 @@ static void *cpuset_being_rebound;
 /**
  * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
  * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
- * @oldmem: old mems_allowed of cpuset cs
  * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
  *
  * Called with cpuset_mutex held
  * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
  * if @heap != NULL.
  */
-static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
-				 struct ptr_heap *heap)
+static void update_tasks_nodemask(struct cpuset *cs, struct ptr_heap *heap)
 {
+	static nodemask_t newmems;	/* protected by cpuset_mutex */
 	struct cgroup_scanner scan;
 
 	cpuset_being_rebound = cs;		/* causes mpol_dup() rebind */
 
+	guarantee_online_mems(cs, &newmems);
+
 	scan.cg = cs->css.cgroup;
 	scan.test_task = NULL;
 	scan.process_task = cpuset_change_nodemask;
 	scan.heap = heap;
-	scan.data = (nodemask_t *)oldmem;
+	scan.data = &newmems;
 
 	/*
 	 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
@@ -1035,6 +1044,12 @@ static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
 	 */
 	cgroup_scan_tasks(&scan);
 
+	/*
+	 * All the tasks' nodemasks have been updated, update
+	 * cs->old_mems_allowed.
+	 */
+	cs->old_mems_allowed = newmems;
+
 	/* We're done rebinding vmas to this cpuset's new mems_allowed. */
 	cpuset_being_rebound = NULL;
 }
@@ -1055,13 +1070,9 @@ static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
 			   const char *buf)
 {
-	NODEMASK_ALLOC(nodemask_t, oldmem, GFP_KERNEL);
 	int retval;
 	struct ptr_heap heap;
 
-	if (!oldmem)
-		return -ENOMEM;
-
 	/*
 	 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
 	 * it's read-only
@@ -1090,8 +1101,8 @@ static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
 			goto done;
 		}
 	}
-	*oldmem = cs->mems_allowed;
-	if (nodes_equal(*oldmem, trialcs->mems_allowed)) {
+
+	if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
 		retval = 0;		/* Too easy - nothing to do */
 		goto done;
 	}
@@ -1107,11 +1118,10 @@ static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
 	cs->mems_allowed = trialcs->mems_allowed;
 	mutex_unlock(&callback_mutex);
 
-	update_tasks_nodemask(cs, oldmem, &heap);
+	update_tasks_nodemask(cs, &heap);
 
 	heap_free(&heap);
 done:
-	NODEMASK_FREE(oldmem);
 	return retval;
 }
 
@@ -1424,7 +1434,6 @@ static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 	 * Change mm, possibly for multiple threads in a threadgroup. This is
 	 * expensive and may sleep.
 	 */
-	cpuset_attach_nodemask_to = cs->mems_allowed;
 	mm = get_task_mm(leader);
 	if (mm) {
 		mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
@@ -1434,6 +1443,8 @@ static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 		mmput(mm);
 	}
 
+	cs->old_mems_allowed = cpuset_attach_nodemask_to;
+
 	cs->attach_in_progress--;
 
 	/*
@@ -2000,7 +2011,7 @@ static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
 static void cpuset_propagate_hotplug_workfn(struct work_struct *work)
 {
 	static cpumask_t off_cpus;
-	static nodemask_t off_mems, tmp_mems;
+	static nodemask_t off_mems;
 	struct cpuset *cs = container_of(work, struct cpuset, hotplug_work);
 	bool is_empty;
 
@@ -2021,13 +2032,12 @@ static void cpuset_propagate_hotplug_workfn(struct work_struct *work)
 
 	/* remove offline mems from @cs */
 	if (!nodes_empty(off_mems)) {
-		tmp_mems = cs->mems_allowed;
 		mutex_lock(&callback_mutex);
 		nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
 		mutex_unlock(&callback_mutex);
 
 		if (!nodes_empty(cs->mems_allowed))
-			update_tasks_nodemask(cs, &tmp_mems, NULL);
+			update_tasks_nodemask(cs, NULL);
 	}
 
 	is_empty = cpumask_empty(cs->cpus_allowed) ||
@@ -2119,11 +2129,10 @@ static void cpuset_hotplug_workfn(struct work_struct *work)
 
 	/* synchronize mems_allowed to N_MEMORY */
 	if (mems_updated) {
-		tmp_mems = top_cpuset.mems_allowed;
 		mutex_lock(&callback_mutex);
 		top_cpuset.mems_allowed = new_mems;
 		mutex_unlock(&callback_mutex);
-		update_tasks_nodemask(&top_cpuset, &tmp_mems, NULL);
+		update_tasks_nodemask(&top_cpuset, NULL);
 	}
 
 	/* if cpus or mems went down, we need to propagate to descendants */
@@ -2196,6 +2205,7 @@ void __init cpuset_init_smp(void)
 {
 	cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
 	top_cpuset.mems_allowed = node_states[N_MEMORY];
+	top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
 
 	hotplug_memory_notifier(cpuset_track_online_nodes, 10);
 
-- 
1.8.0.2

[PATCH 07/10] cpuset: record old_mems_allowed in struct cpuset

[Li Zefan]

When we update a cpuset's mems_allowed and thus update tasks'
mems_allowed, it's required to pass the old mems_allowed and new
mems_allowed to cpuset_migrate_mm().

Currently we save old mems_allowed in a temp local variable before
changing cpuset->mems_allowed. This patch changes it by saving
old mems_allowed in cpuset->old_mems_allowed.

This currently won't change any behavior, but it will later allow
us to keep tasks in empty cpusets.

Signed-off-by: Li Zefan <lizefan-hv44wF8Li93QT0dZR+AlfA@public.gmane.org>
---
 kernel/cpuset.c | 62 +++++++++++++++++++++++++++++++++------------------------
 1 file changed, 36 insertions(+), 26 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index 63a5343..c4f9ebd 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -87,6 +87,18 @@ struct cpuset {
 	cpumask_var_t cpus_allowed;	/* CPUs allowed to tasks in cpuset */
 	nodemask_t mems_allowed;	/* Memory Nodes allowed to tasks */
 
+	/*
+	 * This is old Memory Nodes tasks took on.
+	 *
+	 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
+	 * - A new cpuset's old_mems_allowed is initialized when some
+	 *   task is moved into it.
+	 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
+	 *   cpuset.mems_allowed and have tasks' nodemask updated, and
+	 *   then old_mems_allowed is updated to mems_allowed.
+	 */
+	nodemask_t old_mems_allowed;
+
 	struct fmeter fmeter;		/* memory_pressure filter */
 
 	/*
@@ -975,16 +987,12 @@ static void cpuset_change_task_nodemask(struct task_struct *tsk,
 static void cpuset_change_nodemask(struct task_struct *p,
 				   struct cgroup_scanner *scan)
 {
+	struct cpuset *cs = cgroup_cs(scan->cg);
 	struct mm_struct *mm;
-	struct cpuset *cs;
 	int migrate;
-	const nodemask_t *oldmem = scan->data;
-	static nodemask_t newmems;	/* protected by cpuset_mutex */
-
-	cs = cgroup_cs(scan->cg);
-	guarantee_online_mems(cs, &newmems);
+	nodemask_t *newmems = scan->data;
 
-	cpuset_change_task_nodemask(p, &newmems);
+	cpuset_change_task_nodemask(p, newmems);
 
 	mm = get_task_mm(p);
 	if (!mm)
@@ -994,7 +1002,7 @@ static void cpuset_change_nodemask(struct task_struct *p,
 
 	mpol_rebind_mm(mm, &cs->mems_allowed);
 	if (migrate)
-		cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
+		cpuset_migrate_mm(mm, &cs->old_mems_allowed, newmems);
 	mmput(mm);
 }
 
@@ -1003,25 +1011,26 @@ static void *cpuset_being_rebound;
 /**
  * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
  * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
- * @oldmem: old mems_allowed of cpuset cs
  * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
  *
  * Called with cpuset_mutex held
  * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
  * if @heap != NULL.
  */
-static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
-				 struct ptr_heap *heap)
+static void update_tasks_nodemask(struct cpuset *cs, struct ptr_heap *heap)
 {
+	static nodemask_t newmems;	/* protected by cpuset_mutex */
 	struct cgroup_scanner scan;
 
 	cpuset_being_rebound = cs;		/* causes mpol_dup() rebind */
 
+	guarantee_online_mems(cs, &newmems);
+
 	scan.cg = cs->css.cgroup;
 	scan.test_task = NULL;
 	scan.process_task = cpuset_change_nodemask;
 	scan.heap = heap;
-	scan.data = (nodemask_t *)oldmem;
+	scan.data = &newmems;
 
 	/*
 	 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
@@ -1035,6 +1044,12 @@ static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
 	 */
 	cgroup_scan_tasks(&scan);
 
+	/*
+	 * All the tasks' nodemasks have been updated, update
+	 * cs->old_mems_allowed.
+	 */
+	cs->old_mems_allowed = newmems;
+
 	/* We're done rebinding vmas to this cpuset's new mems_allowed. */
 	cpuset_being_rebound = NULL;
 }
@@ -1055,13 +1070,9 @@ static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
 			   const char *buf)
 {
-	NODEMASK_ALLOC(nodemask_t, oldmem, GFP_KERNEL);
 	int retval;
 	struct ptr_heap heap;
 
-	if (!oldmem)
-		return -ENOMEM;
-
 	/*
 	 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
 	 * it's read-only
@@ -1090,8 +1101,8 @@ static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
 			goto done;
 		}
 	}
-	*oldmem = cs->mems_allowed;
-	if (nodes_equal(*oldmem, trialcs->mems_allowed)) {
+
+	if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
 		retval = 0;		/* Too easy - nothing to do */
 		goto done;
 	}
@@ -1107,11 +1118,10 @@ static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
 	cs->mems_allowed = trialcs->mems_allowed;
 	mutex_unlock(&callback_mutex);
 
-	update_tasks_nodemask(cs, oldmem, &heap);
+	update_tasks_nodemask(cs, &heap);
 
 	heap_free(&heap);
 done:
-	NODEMASK_FREE(oldmem);
 	return retval;
 }
 
@@ -1424,7 +1434,6 @@ static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 	 * Change mm, possibly for multiple threads in a threadgroup. This is
 	 * expensive and may sleep.
 	 */
-	cpuset_attach_nodemask_to = cs->mems_allowed;
 	mm = get_task_mm(leader);
 	if (mm) {
 		mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
@@ -1434,6 +1443,8 @@ static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 		mmput(mm);
 	}
 
+	cs->old_mems_allowed = cpuset_attach_nodemask_to;
+
 	cs->attach_in_progress--;
 
 	/*
@@ -2000,7 +2011,7 @@ static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
 static void cpuset_propagate_hotplug_workfn(struct work_struct *work)
 {
 	static cpumask_t off_cpus;
-	static nodemask_t off_mems, tmp_mems;
+	static nodemask_t off_mems;
 	struct cpuset *cs = container_of(work, struct cpuset, hotplug_work);
 	bool is_empty;
 
@@ -2021,13 +2032,12 @@ static void cpuset_propagate_hotplug_workfn(struct work_struct *work)
 
 	/* remove offline mems from @cs */
 	if (!nodes_empty(off_mems)) {
-		tmp_mems = cs->mems_allowed;
 		mutex_lock(&callback_mutex);
 		nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
 		mutex_unlock(&callback_mutex);
 
 		if (!nodes_empty(cs->mems_allowed))
-			update_tasks_nodemask(cs, &tmp_mems, NULL);
+			update_tasks_nodemask(cs, NULL);
 	}
 
 	is_empty = cpumask_empty(cs->cpus_allowed) ||
@@ -2119,11 +2129,10 @@ static void cpuset_hotplug_workfn(struct work_struct *work)
 
 	/* synchronize mems_allowed to N_MEMORY */
 	if (mems_updated) {
-		tmp_mems = top_cpuset.mems_allowed;
 		mutex_lock(&callback_mutex);
 		top_cpuset.mems_allowed = new_mems;
 		mutex_unlock(&callback_mutex);
-		update_tasks_nodemask(&top_cpuset, &tmp_mems, NULL);
+		update_tasks_nodemask(&top_cpuset, NULL);
 	}
 
 	/* if cpus or mems went down, we need to propagate to descendants */
@@ -2196,6 +2205,7 @@ void __init cpuset_init_smp(void)
 {
 	cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
 	top_cpuset.mems_allowed = node_states[N_MEMORY];
+	top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
 
 	hotplug_memory_notifier(cpuset_track_online_nodes, 10);
 
-- 
1.8.0.2

[PATCH 08/10] cpuset: introduce effective_{cpumask|nodemask}_cpuset()

[Li Zefan]

effective_cpumask_cpuset() returns an ancestor cpuset which has
non-empty cpumask.

If a cpuset is empty and the tasks in it need to update their
cpus_allowed, they take on the ancestor cpuset's cpumask.

This currently won't change any behavior, but it will later allow us
to keep tasks in empty cpusets.

Signed-off-by: Li Zefan <lizefan-hv44wF8Li93QT0dZR+AlfA@public.gmane.org>
---
 kernel/cpuset.c | 76 ++++++++++++++++++++++++++++++++++++++++++++++++---------
 1 file changed, 65 insertions(+), 11 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index c4f9ebd..741e652 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -794,6 +794,45 @@ void rebuild_sched_domains(void)
 	mutex_unlock(&cpuset_mutex);
 }
 
+/*
+ * effective_cpumask_cpuset - return nearest ancestor with non-empty cpus
+ * @cs: the cpuset in interest
+ *
+ * A cpuset's effective cpumask is the cpumask of the nearest ancestor
+ * with non-empty cpus. We use effective cpumask whenever:
+ * - we update tasks' cpus_allowed. (they take on the ancestor's cpumask
+ *   if the cpuset they reside in has no cpus)
+ * - we want to retrieve task_cs(tsk)'s cpus_allowed.
+ *
+ * Called with cpuset_mutex held. cpuset_cpus_allowed_fallback() is an
+ * exception. See comments there.
+ */
+static struct cpuset *effective_cpumask_cpuset(struct cpuset *cs)
+{
+	while (cpumask_empty(cs->cpus_allowed))
+		cs = parent_cs(cs);
+	return cs;
+}
+
+/*
+ * effective_nodemask_cpuset - return nearest ancestor with non-empty mems
+ * @cs: the cpuset in interest
+ *
+ * A cpuset's effective nodemask is the nodemask of the nearest ancestor
+ * with non-empty memss. We use effective nodemask whenever:
+ * - we update tasks' mems_allowed. (they take on the ancestor's nodemask
+ *   if the cpuset they reside in has no mems)
+ * - we want to retrieve task_cs(tsk)'s mems_allowed.
+ *
+ * Called with cpuset_mutex held.
+ */
+static struct cpuset *effective_nodemask_cpuset(struct cpuset *cs)
+{
+	while (nodes_empty(cs->mems_allowed))
+		cs = parent_cs(cs);
+	return cs;
+}
+
 /**
  * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
  * @tsk: task to test
@@ -808,7 +847,10 @@ void rebuild_sched_domains(void)
 static void cpuset_change_cpumask(struct task_struct *tsk,
 				  struct cgroup_scanner *scan)
 {
-	set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
+	struct cpuset *cpus_cs;
+	
+	cpus_cs = effective_cpumask_cpuset(cgroup_cs(scan->cg));
+	set_cpus_allowed_ptr(tsk, cpus_cs->cpus_allowed);
 }
 
 /**
@@ -923,12 +965,14 @@ static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
 							const nodemask_t *to)
 {
 	struct task_struct *tsk = current;
+	struct cpuset *mems_cs;
 
 	tsk->mems_allowed = *to;
 
 	do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
 
-	guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
+	mems_cs = effective_nodemask_cpuset(task_cs(tsk));
+	guarantee_online_mems(mems_cs, &tsk->mems_allowed);
 }
 
 /*
@@ -1021,10 +1065,11 @@ static void update_tasks_nodemask(struct cpuset *cs, struct ptr_heap *heap)
 {
 	static nodemask_t newmems;	/* protected by cpuset_mutex */
 	struct cgroup_scanner scan;
+	struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
 
 	cpuset_being_rebound = cs;		/* causes mpol_dup() rebind */
 
-	guarantee_online_mems(cs, &newmems);
+	guarantee_online_mems(mems_cs, &newmems);
 
 	scan.cg = cs->css.cgroup;
 	scan.test_task = NULL;
@@ -1408,6 +1453,8 @@ static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 	struct cgroup *oldcgrp = cgroup_taskset_cur_cgroup(tset);
 	struct cpuset *cs = cgroup_cs(cgrp);
 	struct cpuset *oldcs = cgroup_cs(oldcgrp);
+	struct cpuset *cpus_cs = effective_cpumask_cpuset(cs);
+	struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
 
 	mutex_lock(&cpuset_mutex);
 
@@ -1415,9 +1462,9 @@ static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 	if (cs == &top_cpuset)
 		cpumask_copy(cpus_attach, cpu_possible_mask);
 	else
-		guarantee_online_cpus(cs, cpus_attach);
+		guarantee_online_cpus(cpus_cs, cpus_attach);
 
-	guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
+	guarantee_online_mems(mems_cs, &cpuset_attach_nodemask_to);
 
 	cgroup_taskset_for_each(task, cgrp, tset) {
 		/*
@@ -1436,9 +1483,11 @@ static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 	 */
 	mm = get_task_mm(leader);
 	if (mm) {
+		struct cpuset *mems_oldcs = effective_nodemask_cpuset(oldcs);
+
 		mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
 		if (is_memory_migrate(cs))
-			cpuset_migrate_mm(mm, &oldcs->mems_allowed,
+			cpuset_migrate_mm(mm, &mems_oldcs->mems_allowed,
 					  &cpuset_attach_nodemask_to);
 		mmput(mm);
 	}
@@ -2227,20 +2276,23 @@ void __init cpuset_init_smp(void)
 
 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
 {
+	struct cpuset *cpus_cs;
+
 	mutex_lock(&callback_mutex);
 	task_lock(tsk);
-	guarantee_online_cpus(task_cs(tsk), pmask);
+	cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
+	guarantee_online_cpus(cpus_cs, pmask);
 	task_unlock(tsk);
 	mutex_unlock(&callback_mutex);
 }
 
 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
 {
-	const struct cpuset *cs;
+	const struct cpuset *cpus_cs;
 
 	rcu_read_lock();
-	cs = task_cs(tsk);
-	do_set_cpus_allowed(tsk, cs->cpus_allowed);
+	cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
+	do_set_cpus_allowed(tsk, cpus_cs->cpus_allowed);
 	rcu_read_unlock();
 
 	/*
@@ -2279,11 +2331,13 @@ void cpuset_init_current_mems_allowed(void)
 
 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
 {
+	struct cpuset *mems_cs;
 	nodemask_t mask;
 
 	mutex_lock(&callback_mutex);
 	task_lock(tsk);
-	guarantee_online_mems(task_cs(tsk), &mask);
+	mems_cs = effective_nodemask_cpuset(task_cs(tsk));
+	guarantee_online_mems(mems_cs, &mask);
 	task_unlock(tsk);
 	mutex_unlock(&callback_mutex);
 
-- 
1.8.0.2

[PATCH 08/10] cpuset: introduce effective_{cpumask|nodemask}_cpuset()

[Li Zefan]

effective_cpumask_cpuset() returns an ancestor cpuset which has
non-empty cpumask.

If a cpuset is empty and the tasks in it need to update their
cpus_allowed, they take on the ancestor cpuset's cpumask.

This currently won't change any behavior, but it will later allow us
to keep tasks in empty cpusets.

Signed-off-by: Li Zefan <lizefan@huawei.com>
---
 kernel/cpuset.c | 76 ++++++++++++++++++++++++++++++++++++++++++++++++---------
 1 file changed, 65 insertions(+), 11 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index c4f9ebd..741e652 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -794,6 +794,45 @@ void rebuild_sched_domains(void)
 	mutex_unlock(&cpuset_mutex);
 }
 
+/*
+ * effective_cpumask_cpuset - return nearest ancestor with non-empty cpus
+ * @cs: the cpuset in interest
+ *
+ * A cpuset's effective cpumask is the cpumask of the nearest ancestor
+ * with non-empty cpus. We use effective cpumask whenever:
+ * - we update tasks' cpus_allowed. (they take on the ancestor's cpumask
+ *   if the cpuset they reside in has no cpus)
+ * - we want to retrieve task_cs(tsk)'s cpus_allowed.
+ *
+ * Called with cpuset_mutex held. cpuset_cpus_allowed_fallback() is an
+ * exception. See comments there.
+ */
+static struct cpuset *effective_cpumask_cpuset(struct cpuset *cs)
+{
+	while (cpumask_empty(cs->cpus_allowed))
+		cs = parent_cs(cs);
+	return cs;
+}
+
+/*
+ * effective_nodemask_cpuset - return nearest ancestor with non-empty mems
+ * @cs: the cpuset in interest
+ *
+ * A cpuset's effective nodemask is the nodemask of the nearest ancestor
+ * with non-empty memss. We use effective nodemask whenever:
+ * - we update tasks' mems_allowed. (they take on the ancestor's nodemask
+ *   if the cpuset they reside in has no mems)
+ * - we want to retrieve task_cs(tsk)'s mems_allowed.
+ *
+ * Called with cpuset_mutex held.
+ */
+static struct cpuset *effective_nodemask_cpuset(struct cpuset *cs)
+{
+	while (nodes_empty(cs->mems_allowed))
+		cs = parent_cs(cs);
+	return cs;
+}
+
 /**
  * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
  * @tsk: task to test
@@ -808,7 +847,10 @@ void rebuild_sched_domains(void)
 static void cpuset_change_cpumask(struct task_struct *tsk,
 				  struct cgroup_scanner *scan)
 {
-	set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
+	struct cpuset *cpus_cs;
+	
+	cpus_cs = effective_cpumask_cpuset(cgroup_cs(scan->cg));
+	set_cpus_allowed_ptr(tsk, cpus_cs->cpus_allowed);
 }
 
 /**
@@ -923,12 +965,14 @@ static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
 							const nodemask_t *to)
 {
 	struct task_struct *tsk = current;
+	struct cpuset *mems_cs;
 
 	tsk->mems_allowed = *to;
 
 	do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
 
-	guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
+	mems_cs = effective_nodemask_cpuset(task_cs(tsk));
+	guarantee_online_mems(mems_cs, &tsk->mems_allowed);
 }
 
 /*
@@ -1021,10 +1065,11 @@ static void update_tasks_nodemask(struct cpuset *cs, struct ptr_heap *heap)
 {
 	static nodemask_t newmems;	/* protected by cpuset_mutex */
 	struct cgroup_scanner scan;
+	struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
 
 	cpuset_being_rebound = cs;		/* causes mpol_dup() rebind */
 
-	guarantee_online_mems(cs, &newmems);
+	guarantee_online_mems(mems_cs, &newmems);
 
 	scan.cg = cs->css.cgroup;
 	scan.test_task = NULL;
@@ -1408,6 +1453,8 @@ static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 	struct cgroup *oldcgrp = cgroup_taskset_cur_cgroup(tset);
 	struct cpuset *cs = cgroup_cs(cgrp);
 	struct cpuset *oldcs = cgroup_cs(oldcgrp);
+	struct cpuset *cpus_cs = effective_cpumask_cpuset(cs);
+	struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
 
 	mutex_lock(&cpuset_mutex);
 
@@ -1415,9 +1462,9 @@ static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 	if (cs == &top_cpuset)
 		cpumask_copy(cpus_attach, cpu_possible_mask);
 	else
-		guarantee_online_cpus(cs, cpus_attach);
+		guarantee_online_cpus(cpus_cs, cpus_attach);
 
-	guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
+	guarantee_online_mems(mems_cs, &cpuset_attach_nodemask_to);
 
 	cgroup_taskset_for_each(task, cgrp, tset) {
 		/*
@@ -1436,9 +1483,11 @@ static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 	 */
 	mm = get_task_mm(leader);
 	if (mm) {
+		struct cpuset *mems_oldcs = effective_nodemask_cpuset(oldcs);
+
 		mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
 		if (is_memory_migrate(cs))
-			cpuset_migrate_mm(mm, &oldcs->mems_allowed,
+			cpuset_migrate_mm(mm, &mems_oldcs->mems_allowed,
 					  &cpuset_attach_nodemask_to);
 		mmput(mm);
 	}
@@ -2227,20 +2276,23 @@ void __init cpuset_init_smp(void)
 
 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
 {
+	struct cpuset *cpus_cs;
+
 	mutex_lock(&callback_mutex);
 	task_lock(tsk);
-	guarantee_online_cpus(task_cs(tsk), pmask);
+	cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
+	guarantee_online_cpus(cpus_cs, pmask);
 	task_unlock(tsk);
 	mutex_unlock(&callback_mutex);
 }
 
 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
 {
-	const struct cpuset *cs;
+	const struct cpuset *cpus_cs;
 
 	rcu_read_lock();
-	cs = task_cs(tsk);
-	do_set_cpus_allowed(tsk, cs->cpus_allowed);
+	cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
+	do_set_cpus_allowed(tsk, cpus_cs->cpus_allowed);
 	rcu_read_unlock();
 
 	/*
@@ -2279,11 +2331,13 @@ void cpuset_init_current_mems_allowed(void)
 
 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
 {
+	struct cpuset *mems_cs;
 	nodemask_t mask;
 
 	mutex_lock(&callback_mutex);
 	task_lock(tsk);
-	guarantee_online_mems(task_cs(tsk), &mask);
+	mems_cs = effective_nodemask_cpuset(task_cs(tsk));
+	guarantee_online_mems(mems_cs, &mask);
 	task_unlock(tsk);
 	mutex_unlock(&callback_mutex);
 
-- 
1.8.0.2

[PATCH 08/10] cpuset: introduce effective_{cpumask|nodemask}_cpuset()

[Li Zefan]

effective_cpumask_cpuset() returns an ancestor cpuset which has
non-empty cpumask.

If a cpuset is empty and the tasks in it need to update their
cpus_allowed, they take on the ancestor cpuset's cpumask.

This currently won't change any behavior, but it will later allow us
to keep tasks in empty cpusets.

Signed-off-by: Li Zefan <lizefan-hv44wF8Li93QT0dZR+AlfA@public.gmane.org>
---
 kernel/cpuset.c | 76 ++++++++++++++++++++++++++++++++++++++++++++++++---------
 1 file changed, 65 insertions(+), 11 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index c4f9ebd..741e652 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -794,6 +794,45 @@ void rebuild_sched_domains(void)
 	mutex_unlock(&cpuset_mutex);
 }
 
+/*
+ * effective_cpumask_cpuset - return nearest ancestor with non-empty cpus
+ * @cs: the cpuset in interest
+ *
+ * A cpuset's effective cpumask is the cpumask of the nearest ancestor
+ * with non-empty cpus. We use effective cpumask whenever:
+ * - we update tasks' cpus_allowed. (they take on the ancestor's cpumask
+ *   if the cpuset they reside in has no cpus)
+ * - we want to retrieve task_cs(tsk)'s cpus_allowed.
+ *
+ * Called with cpuset_mutex held. cpuset_cpus_allowed_fallback() is an
+ * exception. See comments there.
+ */
+static struct cpuset *effective_cpumask_cpuset(struct cpuset *cs)
+{
+	while (cpumask_empty(cs->cpus_allowed))
+		cs = parent_cs(cs);
+	return cs;
+}
+
+/*
+ * effective_nodemask_cpuset - return nearest ancestor with non-empty mems
+ * @cs: the cpuset in interest
+ *
+ * A cpuset's effective nodemask is the nodemask of the nearest ancestor
+ * with non-empty memss. We use effective nodemask whenever:
+ * - we update tasks' mems_allowed. (they take on the ancestor's nodemask
+ *   if the cpuset they reside in has no mems)
+ * - we want to retrieve task_cs(tsk)'s mems_allowed.
+ *
+ * Called with cpuset_mutex held.
+ */
+static struct cpuset *effective_nodemask_cpuset(struct cpuset *cs)
+{
+	while (nodes_empty(cs->mems_allowed))
+		cs = parent_cs(cs);
+	return cs;
+}
+
 /**
  * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
  * @tsk: task to test
@@ -808,7 +847,10 @@ void rebuild_sched_domains(void)
 static void cpuset_change_cpumask(struct task_struct *tsk,
 				  struct cgroup_scanner *scan)
 {
-	set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
+	struct cpuset *cpus_cs;
+	
+	cpus_cs = effective_cpumask_cpuset(cgroup_cs(scan->cg));
+	set_cpus_allowed_ptr(tsk, cpus_cs->cpus_allowed);
 }
 
 /**
@@ -923,12 +965,14 @@ static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
 							const nodemask_t *to)
 {
 	struct task_struct *tsk = current;
+	struct cpuset *mems_cs;
 
 	tsk->mems_allowed = *to;
 
 	do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
 
-	guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
+	mems_cs = effective_nodemask_cpuset(task_cs(tsk));
+	guarantee_online_mems(mems_cs, &tsk->mems_allowed);
 }
 
 /*
@@ -1021,10 +1065,11 @@ static void update_tasks_nodemask(struct cpuset *cs, struct ptr_heap *heap)
 {
 	static nodemask_t newmems;	/* protected by cpuset_mutex */
 	struct cgroup_scanner scan;
+	struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
 
 	cpuset_being_rebound = cs;		/* causes mpol_dup() rebind */
 
-	guarantee_online_mems(cs, &newmems);
+	guarantee_online_mems(mems_cs, &newmems);
 
 	scan.cg = cs->css.cgroup;
 	scan.test_task = NULL;
@@ -1408,6 +1453,8 @@ static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 	struct cgroup *oldcgrp = cgroup_taskset_cur_cgroup(tset);
 	struct cpuset *cs = cgroup_cs(cgrp);
 	struct cpuset *oldcs = cgroup_cs(oldcgrp);
+	struct cpuset *cpus_cs = effective_cpumask_cpuset(cs);
+	struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
 
 	mutex_lock(&cpuset_mutex);
 
@@ -1415,9 +1462,9 @@ static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 	if (cs == &top_cpuset)
 		cpumask_copy(cpus_attach, cpu_possible_mask);
 	else
-		guarantee_online_cpus(cs, cpus_attach);
+		guarantee_online_cpus(cpus_cs, cpus_attach);
 
-	guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
+	guarantee_online_mems(mems_cs, &cpuset_attach_nodemask_to);
 
 	cgroup_taskset_for_each(task, cgrp, tset) {
 		/*
@@ -1436,9 +1483,11 @@ static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 	 */
 	mm = get_task_mm(leader);
 	if (mm) {
+		struct cpuset *mems_oldcs = effective_nodemask_cpuset(oldcs);
+
 		mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
 		if (is_memory_migrate(cs))
-			cpuset_migrate_mm(mm, &oldcs->mems_allowed,
+			cpuset_migrate_mm(mm, &mems_oldcs->mems_allowed,
 					  &cpuset_attach_nodemask_to);
 		mmput(mm);
 	}
@@ -2227,20 +2276,23 @@ void __init cpuset_init_smp(void)
 
 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
 {
+	struct cpuset *cpus_cs;
+
 	mutex_lock(&callback_mutex);
 	task_lock(tsk);
-	guarantee_online_cpus(task_cs(tsk), pmask);
+	cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
+	guarantee_online_cpus(cpus_cs, pmask);
 	task_unlock(tsk);
 	mutex_unlock(&callback_mutex);
 }
 
 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
 {
-	const struct cpuset *cs;
+	const struct cpuset *cpus_cs;
 
 	rcu_read_lock();
-	cs = task_cs(tsk);
-	do_set_cpus_allowed(tsk, cs->cpus_allowed);
+	cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
+	do_set_cpus_allowed(tsk, cpus_cs->cpus_allowed);
 	rcu_read_unlock();
 
 	/*
@@ -2279,11 +2331,13 @@ void cpuset_init_current_mems_allowed(void)
 
 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
 {
+	struct cpuset *mems_cs;
 	nodemask_t mask;
 
 	mutex_lock(&callback_mutex);
 	task_lock(tsk);
-	guarantee_online_mems(task_cs(tsk), &mask);
+	mems_cs = effective_nodemask_cpuset(task_cs(tsk));
+	guarantee_online_mems(mems_cs, &mask);
 	task_unlock(tsk);
 	mutex_unlock(&callback_mutex);
 
-- 
1.8.0.2

[PATCH 09/10] cpuset: allow to keep tasks in empty cpusets

[Li Zefan]

To achieve this:

- We call update_tasks_cpumask/nodemask() for empty cpusets when
hotplug happens, instead of moving tasks out of them.

- When a cpuset's masks are changed by writing cpuset.cpus/mems,
we also update tasks in child cpusets which are empty.

Signed-off-by: Li Zefan <lizefan-hv44wF8Li93QT0dZR+AlfA@public.gmane.org>
---
 kernel/cpuset.c | 171 +++++++++++++++++++++++++++++++++++++++++++++++---------
 1 file changed, 146 insertions(+), 25 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index 741e652..95e9394 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -99,6 +99,9 @@ struct cpuset {
 	 */
 	nodemask_t old_mems_allowed;
 
+	/* used in cpuset_update_nodemask_workfn() */
+	struct ptr_heap *heap;
+
 	struct fmeter fmeter;		/* memory_pressure filter */
 
 	/*
@@ -114,6 +117,7 @@ struct cpuset {
 	int relax_domain_level;
 
 	struct work_struct hotplug_work;
+	struct work_struct update_nodemask_work;
 };
 
 /* Retrieve the cpuset for a cgroup */
@@ -276,6 +280,8 @@ static struct cpuset top_cpuset = {
 static DEFINE_MUTEX(cpuset_mutex);
 static DEFINE_MUTEX(callback_mutex);
 
+static struct workqueue_struct *cpuset_update_nodemask_wq;
+
 /*
  * CPU / memory hotplug is handled asynchronously.
  */
@@ -877,6 +883,39 @@ static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
 	cgroup_scan_tasks(&scan);
 }
 
+/*
+ * update_tasks_cpumask_hier - Update the cpumasks of tasks in the hierarchy.
+ * @root_cs: the root cpuset of the hierarchy
+ * @update_root: update root cpuset or not?
+ * @heap: the heap used by cgroup_scan_tasks()
+ *
+ * This will update cpumasks of tasks in @root_cs and all other empty cpusets
+ * which take on cpumask of @root_cs.
+ *
+ * Called with cpuset_mutex held
+ */
+static void update_tasks_cpumask_hier(struct cpuset *root_cs,
+				      bool update_root, struct ptr_heap *heap)
+{
+	struct cpuset *cp;
+	struct cgroup *pos_cgrp;
+
+	if (update_root)
+		update_tasks_cpumask(root_cs, heap);
+
+	rcu_read_lock();
+	cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) {
+		/* skip the whole subtree if @cp have some CPU */
+		if (!cpumask_empty(cp->cpus_allowed)) {
+			pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
+			continue;
+		}
+
+		update_tasks_cpumask(cp, heap);
+	}
+	rcu_read_unlock();
+}
+
 /**
  * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
  * @cs: the cpuset to consider
@@ -928,11 +967,7 @@ static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
 	cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
 	mutex_unlock(&callback_mutex);
 
-	/*
-	 * Scan tasks in the cpuset, and update the cpumasks of any
-	 * that need an update.
-	 */
-	update_tasks_cpumask(cs, &heap);
+	update_tasks_cpumask_hier(cs, true, &heap);
 
 	heap_free(&heap);
 
@@ -1099,6 +1134,78 @@ static void update_tasks_nodemask(struct cpuset *cs, struct ptr_heap *heap)
 	cpuset_being_rebound = NULL;
 }
 
+static void cpuset_update_nodemask_workfn(struct work_struct *work)
+{
+	struct cpuset *cs = container_of(work, struct cpuset,
+					update_nodemask_work);
+
+	update_tasks_nodemask(cs, cs->heap);
+	css_put(&cs->css);
+}
+
+static void schedule_update_tasks_nodemask(struct cpuset *cs,
+					   struct ptr_heap *heap)
+{
+	bool queued;
+
+	/* Will be released when the work item finishes executing. */
+	if (!css_tryget(&cs->css))
+		return;
+
+	/*
+	 * The caller will flush the workqueue with cpuset_mutex held,
+	 * so it's not possible a work item was already queued, and
+	 * we're sure cs->heap is valid.
+	 */
+	cs->heap = heap;
+	queued = queue_work(cpuset_update_nodemask_wq,
+			    &cs->update_nodemask_work);
+	if (queued) {
+		WARN_ON(1);
+		css_put(&cs->css);
+	}
+}
+
+/*
+ * update_tasks_nodemask_hier - Update the nodemasks of tasks in the hierarchy.
+ * @cs: the root cpuset of the hierarchy
+ * @update_root: update the root cpuset or not?
+ * @heap: the heap used by cgroup_scan_tasks()
+ *
+ * This will update nodemasks of tasks in @root_cs and all other empty cpusets
+ * which take on nodemask of @root_cs.
+ *
+ * Called with cpuset_mutex held
+ */
+static void update_tasks_nodemask_hier(struct cpuset *root_cs,
+				       bool update_root, struct ptr_heap *heap)
+{
+	struct cpuset *cp;
+	struct cgroup *pos_cgrp;
+
+	if (update_root)
+		update_tasks_nodemask(root_cs, heap);
+
+	rcu_read_lock();
+	cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) {
+		/* skip the whole subtree if @cp have some CPU */
+		if (!nodes_empty(cp->mems_allowed)) {
+			pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
+			continue;
+		}
+
+		schedule_update_tasks_nodemask(cp, heap);
+	}
+	rcu_read_unlock();
+
+	/*
+	 * The only reason we use workqueue is update_tasks_nodemask() can't
+	 * be called in rcu_read_lock(). Flush the workqueue to make sure
+	 * all the updates are done before we return.
+	 */
+	flush_workqueue(cpuset_update_nodemask_wq);
+}
+
 /*
  * Handle user request to change the 'mems' memory placement
  * of a cpuset.  Needs to validate the request, update the
@@ -1163,7 +1270,7 @@ static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
 	cs->mems_allowed = trialcs->mems_allowed;
 	mutex_unlock(&callback_mutex);
 
-	update_tasks_nodemask(cs, &heap);
+	update_tasks_nodemask_hier(cs, true, &heap);
 
 	heap_free(&heap);
 done:
@@ -1888,6 +1995,7 @@ static struct cgroup_subsys_state *cpuset_css_alloc(struct cgroup *cont)
 	nodes_clear(cs->mems_allowed);
 	fmeter_init(&cs->fmeter);
 	INIT_WORK(&cs->hotplug_work, cpuset_propagate_hotplug_workfn);
+	INIT_WORK(&cs->update_nodemask_work, cpuset_update_nodemask_workfn);
 	cs->relax_domain_level = -1;
 
 	return &cs->css;
@@ -2063,31 +2171,36 @@ static void cpuset_propagate_hotplug_workfn(struct work_struct *work)
 	static nodemask_t off_mems;
 	struct cpuset *cs = container_of(work, struct cpuset, hotplug_work);
 	bool is_empty;
+	bool sane = cgroup_sane_behavior(cs->css.cgroup);
 
 	mutex_lock(&cpuset_mutex);
 
 	cpumask_andnot(&off_cpus, cs->cpus_allowed, top_cpuset.cpus_allowed);
 	nodes_andnot(off_mems, cs->mems_allowed, top_cpuset.mems_allowed);
 
-	/* remove offline cpus from @cs */
-	if (!cpumask_empty(&off_cpus)) {
-		mutex_lock(&callback_mutex);
-		cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
-		mutex_unlock(&callback_mutex);
+	mutex_lock(&callback_mutex);
+	cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
+	mutex_unlock(&callback_mutex);
 
-		if (!cpumask_empty(cs->cpus_allowed))
-			update_tasks_cpumask(cs, NULL);
-	}
+	/*
+	 * If sane_behavior flag is set, we need to update tasks' cpumask
+	 * for empy cpuset to take on ancestor's cpumask
+	 */
+	if ((sane && cpumask_empty(cs->cpus_allowed)) ||
+	    (!cpumask_empty(&off_cpus) && !cpumask_empty(cs->cpus_allowed)))
+		update_tasks_cpumask(cs, NULL);
 
-	/* remove offline mems from @cs */
-	if (!nodes_empty(off_mems)) {
-		mutex_lock(&callback_mutex);
-		nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
-		mutex_unlock(&callback_mutex);
+	mutex_lock(&callback_mutex);
+	nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
+	mutex_unlock(&callback_mutex);
 
-		if (!nodes_empty(cs->mems_allowed))
-			update_tasks_nodemask(cs, NULL);
-	}
+	/*
+	 * If sane_behavior flag is set, we need to update tasks' nodemask
+	 * for empy cpuset to take on ancestor's nodemask
+	 */
+	if ((sane && nodes_empty(cs->mems_allowed)) ||
+	    (!nodes_empty(off_mems) && !nodes_empty(cs->mems_allowed)))
+		update_tasks_nodemask(cs, NULL);
 
 	is_empty = cpumask_empty(cs->cpus_allowed) ||
 		nodes_empty(cs->mems_allowed);
@@ -2095,11 +2208,13 @@ static void cpuset_propagate_hotplug_workfn(struct work_struct *work)
 	mutex_unlock(&cpuset_mutex);
 
 	/*
-	 * If @cs became empty, move tasks to the nearest ancestor with
-	 * execution resources.  This is full cgroup operation which will
+	 * If sane_behavior flag is set, we'll keep tasks in empty cpusets.
+	 *
+	 * Otherwise move tasks to the nearest ancestor with execution
+	 *  resources.  This is full cgroup operation which will
 	 * also call back into cpuset.  Should be done outside any lock.
 	 */
-	if (is_empty)
+	if (!sane && is_empty)
 		remove_tasks_in_empty_cpuset(cs);
 
 	/* the following may free @cs, should be the last operation */
@@ -2174,6 +2289,7 @@ static void cpuset_hotplug_workfn(struct work_struct *work)
 		cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
 		mutex_unlock(&callback_mutex);
 		/* we don't mess with cpumasks of tasks in top_cpuset */
+		update_tasks_cpumask_hier(&top_cpuset, false, NULL);
 	}
 
 	/* synchronize mems_allowed to N_MEMORY */
@@ -2182,6 +2298,7 @@ static void cpuset_hotplug_workfn(struct work_struct *work)
 		top_cpuset.mems_allowed = new_mems;
 		mutex_unlock(&callback_mutex);
 		update_tasks_nodemask(&top_cpuset, NULL);
+		update_tasks_nodemask_hier(&top_cpuset, false, NULL);
 	}
 
 	/* if cpus or mems went down, we need to propagate to descendants */
@@ -2261,6 +2378,10 @@ void __init cpuset_init_smp(void)
 	cpuset_propagate_hotplug_wq =
 		alloc_ordered_workqueue("cpuset_hotplug", 0);
 	BUG_ON(!cpuset_propagate_hotplug_wq);
+
+	cpuset_update_nodemask_wq =
+		create_workqueue("cpuset_update_nodemask");
+	BUG_ON(!cpuset_update_nodemask_wq);
 }
 
 /**
-- 
1.8.0.2

[PATCH 09/10] cpuset: allow to keep tasks in empty cpusets

[Li Zefan]

To achieve this:

- We call update_tasks_cpumask/nodemask() for empty cpusets when
hotplug happens, instead of moving tasks out of them.

- When a cpuset's masks are changed by writing cpuset.cpus/mems,
we also update tasks in child cpusets which are empty.

Signed-off-by: Li Zefan <lizefan@huawei.com>
---
 kernel/cpuset.c | 171 +++++++++++++++++++++++++++++++++++++++++++++++---------
 1 file changed, 146 insertions(+), 25 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index 741e652..95e9394 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -99,6 +99,9 @@ struct cpuset {
 	 */
 	nodemask_t old_mems_allowed;
 
+	/* used in cpuset_update_nodemask_workfn() */
+	struct ptr_heap *heap;
+
 	struct fmeter fmeter;		/* memory_pressure filter */
 
 	/*
@@ -114,6 +117,7 @@ struct cpuset {
 	int relax_domain_level;
 
 	struct work_struct hotplug_work;
+	struct work_struct update_nodemask_work;
 };
 
 /* Retrieve the cpuset for a cgroup */
@@ -276,6 +280,8 @@ static struct cpuset top_cpuset = {
 static DEFINE_MUTEX(cpuset_mutex);
 static DEFINE_MUTEX(callback_mutex);
 
+static struct workqueue_struct *cpuset_update_nodemask_wq;
+
 /*
  * CPU / memory hotplug is handled asynchronously.
  */
@@ -877,6 +883,39 @@ static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
 	cgroup_scan_tasks(&scan);
 }
 
+/*
+ * update_tasks_cpumask_hier - Update the cpumasks of tasks in the hierarchy.
+ * @root_cs: the root cpuset of the hierarchy
+ * @update_root: update root cpuset or not?
+ * @heap: the heap used by cgroup_scan_tasks()
+ *
+ * This will update cpumasks of tasks in @root_cs and all other empty cpusets
+ * which take on cpumask of @root_cs.
+ *
+ * Called with cpuset_mutex held
+ */
+static void update_tasks_cpumask_hier(struct cpuset *root_cs,
+				      bool update_root, struct ptr_heap *heap)
+{
+	struct cpuset *cp;
+	struct cgroup *pos_cgrp;
+
+	if (update_root)
+		update_tasks_cpumask(root_cs, heap);
+
+	rcu_read_lock();
+	cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) {
+		/* skip the whole subtree if @cp have some CPU */
+		if (!cpumask_empty(cp->cpus_allowed)) {
+			pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
+			continue;
+		}
+
+		update_tasks_cpumask(cp, heap);
+	}
+	rcu_read_unlock();
+}
+
 /**
  * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
  * @cs: the cpuset to consider
@@ -928,11 +967,7 @@ static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
 	cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
 	mutex_unlock(&callback_mutex);
 
-	/*
-	 * Scan tasks in the cpuset, and update the cpumasks of any
-	 * that need an update.
-	 */
-	update_tasks_cpumask(cs, &heap);
+	update_tasks_cpumask_hier(cs, true, &heap);
 
 	heap_free(&heap);
 
@@ -1099,6 +1134,78 @@ static void update_tasks_nodemask(struct cpuset *cs, struct ptr_heap *heap)
 	cpuset_being_rebound = NULL;
 }
 
+static void cpuset_update_nodemask_workfn(struct work_struct *work)
+{
+	struct cpuset *cs = container_of(work, struct cpuset,
+					update_nodemask_work);
+
+	update_tasks_nodemask(cs, cs->heap);
+	css_put(&cs->css);
+}
+
+static void schedule_update_tasks_nodemask(struct cpuset *cs,
+					   struct ptr_heap *heap)
+{
+	bool queued;
+
+	/* Will be released when the work item finishes executing. */
+	if (!css_tryget(&cs->css))
+		return;
+
+	/*
+	 * The caller will flush the workqueue with cpuset_mutex held,
+	 * so it's not possible a work item was already queued, and
+	 * we're sure cs->heap is valid.
+	 */
+	cs->heap = heap;
+	queued = queue_work(cpuset_update_nodemask_wq,
+			    &cs->update_nodemask_work);
+	if (queued) {
+		WARN_ON(1);
+		css_put(&cs->css);
+	}
+}
+
+/*
+ * update_tasks_nodemask_hier - Update the nodemasks of tasks in the hierarchy.
+ * @cs: the root cpuset of the hierarchy
+ * @update_root: update the root cpuset or not?
+ * @heap: the heap used by cgroup_scan_tasks()
+ *
+ * This will update nodemasks of tasks in @root_cs and all other empty cpusets
+ * which take on nodemask of @root_cs.
+ *
+ * Called with cpuset_mutex held
+ */
+static void update_tasks_nodemask_hier(struct cpuset *root_cs,
+				       bool update_root, struct ptr_heap *heap)
+{
+	struct cpuset *cp;
+	struct cgroup *pos_cgrp;
+
+	if (update_root)
+		update_tasks_nodemask(root_cs, heap);
+
+	rcu_read_lock();
+	cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) {
+		/* skip the whole subtree if @cp have some CPU */
+		if (!nodes_empty(cp->mems_allowed)) {
+			pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
+			continue;
+		}
+
+		schedule_update_tasks_nodemask(cp, heap);
+	}
+	rcu_read_unlock();
+
+	/*
+	 * The only reason we use workqueue is update_tasks_nodemask() can't
+	 * be called in rcu_read_lock(). Flush the workqueue to make sure
+	 * all the updates are done before we return.
+	 */
+	flush_workqueue(cpuset_update_nodemask_wq);
+}
+
 /*
  * Handle user request to change the 'mems' memory placement
  * of a cpuset.  Needs to validate the request, update the
@@ -1163,7 +1270,7 @@ static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
 	cs->mems_allowed = trialcs->mems_allowed;
 	mutex_unlock(&callback_mutex);
 
-	update_tasks_nodemask(cs, &heap);
+	update_tasks_nodemask_hier(cs, true, &heap);
 
 	heap_free(&heap);
 done:
@@ -1888,6 +1995,7 @@ static struct cgroup_subsys_state *cpuset_css_alloc(struct cgroup *cont)
 	nodes_clear(cs->mems_allowed);
 	fmeter_init(&cs->fmeter);
 	INIT_WORK(&cs->hotplug_work, cpuset_propagate_hotplug_workfn);
+	INIT_WORK(&cs->update_nodemask_work, cpuset_update_nodemask_workfn);
 	cs->relax_domain_level = -1;
 
 	return &cs->css;
@@ -2063,31 +2171,36 @@ static void cpuset_propagate_hotplug_workfn(struct work_struct *work)
 	static nodemask_t off_mems;
 	struct cpuset *cs = container_of(work, struct cpuset, hotplug_work);
 	bool is_empty;
+	bool sane = cgroup_sane_behavior(cs->css.cgroup);
 
 	mutex_lock(&cpuset_mutex);
 
 	cpumask_andnot(&off_cpus, cs->cpus_allowed, top_cpuset.cpus_allowed);
 	nodes_andnot(off_mems, cs->mems_allowed, top_cpuset.mems_allowed);
 
-	/* remove offline cpus from @cs */
-	if (!cpumask_empty(&off_cpus)) {
-		mutex_lock(&callback_mutex);
-		cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
-		mutex_unlock(&callback_mutex);
+	mutex_lock(&callback_mutex);
+	cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
+	mutex_unlock(&callback_mutex);
 
-		if (!cpumask_empty(cs->cpus_allowed))
-			update_tasks_cpumask(cs, NULL);
-	}
+	/*
+	 * If sane_behavior flag is set, we need to update tasks' cpumask
+	 * for empy cpuset to take on ancestor's cpumask
+	 */
+	if ((sane && cpumask_empty(cs->cpus_allowed)) ||
+	    (!cpumask_empty(&off_cpus) && !cpumask_empty(cs->cpus_allowed)))
+		update_tasks_cpumask(cs, NULL);
 
-	/* remove offline mems from @cs */
-	if (!nodes_empty(off_mems)) {
-		mutex_lock(&callback_mutex);
-		nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
-		mutex_unlock(&callback_mutex);
+	mutex_lock(&callback_mutex);
+	nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
+	mutex_unlock(&callback_mutex);
 
-		if (!nodes_empty(cs->mems_allowed))
-			update_tasks_nodemask(cs, NULL);
-	}
+	/*
+	 * If sane_behavior flag is set, we need to update tasks' nodemask
+	 * for empy cpuset to take on ancestor's nodemask
+	 */
+	if ((sane && nodes_empty(cs->mems_allowed)) ||
+	    (!nodes_empty(off_mems) && !nodes_empty(cs->mems_allowed)))
+		update_tasks_nodemask(cs, NULL);
 
 	is_empty = cpumask_empty(cs->cpus_allowed) ||
 		nodes_empty(cs->mems_allowed);
@@ -2095,11 +2208,13 @@ static void cpuset_propagate_hotplug_workfn(struct work_struct *work)
 	mutex_unlock(&cpuset_mutex);
 
 	/*
-	 * If @cs became empty, move tasks to the nearest ancestor with
-	 * execution resources.  This is full cgroup operation which will
+	 * If sane_behavior flag is set, we'll keep tasks in empty cpusets.
+	 *
+	 * Otherwise move tasks to the nearest ancestor with execution
+	 *  resources.  This is full cgroup operation which will
 	 * also call back into cpuset.  Should be done outside any lock.
 	 */
-	if (is_empty)
+	if (!sane && is_empty)
 		remove_tasks_in_empty_cpuset(cs);
 
 	/* the following may free @cs, should be the last operation */
@@ -2174,6 +2289,7 @@ static void cpuset_hotplug_workfn(struct work_struct *work)
 		cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
 		mutex_unlock(&callback_mutex);
 		/* we don't mess with cpumasks of tasks in top_cpuset */
+		update_tasks_cpumask_hier(&top_cpuset, false, NULL);
 	}
 
 	/* synchronize mems_allowed to N_MEMORY */
@@ -2182,6 +2298,7 @@ static void cpuset_hotplug_workfn(struct work_struct *work)
 		top_cpuset.mems_allowed = new_mems;
 		mutex_unlock(&callback_mutex);
 		update_tasks_nodemask(&top_cpuset, NULL);
+		update_tasks_nodemask_hier(&top_cpuset, false, NULL);
 	}
 
 	/* if cpus or mems went down, we need to propagate to descendants */
@@ -2261,6 +2378,10 @@ void __init cpuset_init_smp(void)
 	cpuset_propagate_hotplug_wq =
 		alloc_ordered_workqueue("cpuset_hotplug", 0);
 	BUG_ON(!cpuset_propagate_hotplug_wq);
+
+	cpuset_update_nodemask_wq =
+		create_workqueue("cpuset_update_nodemask");
+	BUG_ON(!cpuset_update_nodemask_wq);
 }
 
 /**
-- 
1.8.0.2

Re: [PATCH 09/10] cpuset: allow to keep tasks in empty cpusets

[Tejun Heo]

Hello,

On Fri, Apr 19, 2013 at 08:29:24PM +0800, Li Zefan wrote:
> +static void update_tasks_cpumask_hier(struct cpuset *root_cs,
> +				      bool update_root, struct ptr_heap *heap)
> +{
> +	struct cpuset *cp;
> +	struct cgroup *pos_cgrp;
> +
> +	if (update_root)
> +		update_tasks_cpumask(root_cs, heap);
> +
> +	rcu_read_lock();
> +	cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) {
> +		/* skip the whole subtree if @cp have some CPU */
> +		if (!cpumask_empty(cp->cpus_allowed)) {
> +			pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
> +			continue;
> +		}
> +
> +		update_tasks_cpumask(cp, heap);
> +	}
> +	rcu_read_unlock();

I don't think we can call update_tasks_cpumask() under
rcu_read_lock().  It calls into set_cpus_allowed_ptr() which may
block, so you'll probably have to punt it to a work item like how
migration is being done.  Another approach would be converting cgroup
to use SRCU instead, which would lessen pain on other places too.  The
only problem there would be that srcu_read_lock() is a bit more
expensive than rcu_read_lock().  I'm not sure whether that'd show up
in some hot path or not.  Ideas?

Thanks.

-- 
tejun

Re: [PATCH 09/10] cpuset: allow to keep tasks in empty cpusets

[Tejun Heo]

Hello,

On Fri, Apr 19, 2013 at 08:29:24PM +0800, Li Zefan wrote:
> +static void update_tasks_cpumask_hier(struct cpuset *root_cs,
> +				      bool update_root, struct ptr_heap *heap)
> +{
> +	struct cpuset *cp;
> +	struct cgroup *pos_cgrp;
> +
> +	if (update_root)
> +		update_tasks_cpumask(root_cs, heap);
> +
> +	rcu_read_lock();
> +	cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) {
> +		/* skip the whole subtree if @cp have some CPU */
> +		if (!cpumask_empty(cp->cpus_allowed)) {
> +			pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
> +			continue;
> +		}
> +
> +		update_tasks_cpumask(cp, heap);
> +	}
> +	rcu_read_unlock();

I don't think we can call update_tasks_cpumask() under
rcu_read_lock().  It calls into set_cpus_allowed_ptr() which may
block, so you'll probably have to punt it to a work item like how
migration is being done.  Another approach would be converting cgroup
to use SRCU instead, which would lessen pain on other places too.  The
only problem there would be that srcu_read_lock() is a bit more
expensive than rcu_read_lock().  I'm not sure whether that'd show up
in some hot path or not.  Ideas?

Thanks.

-- 
tejun

Re: [PATCH 09/10] cpuset: allow to keep tasks in empty cpusets

[Li Zefan]

On 2013/4/20 4:58, Tejun Heo wrote:
> Hello,
> 
> On Fri, Apr 19, 2013 at 08:29:24PM +0800, Li Zefan wrote:
>> +static void update_tasks_cpumask_hier(struct cpuset *root_cs,
>> +				      bool update_root, struct ptr_heap *heap)
>> +{
>> +	struct cpuset *cp;
>> +	struct cgroup *pos_cgrp;
>> +
>> +	if (update_root)
>> +		update_tasks_cpumask(root_cs, heap);
>> +
>> +	rcu_read_lock();
>> +	cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) {
>> +		/* skip the whole subtree if @cp have some CPU */
>> +		if (!cpumask_empty(cp->cpus_allowed)) {
>> +			pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
>> +			continue;
>> +		}
>> +
>> +		update_tasks_cpumask(cp, heap);
>> +	}
>> +	rcu_read_unlock();
> 
> I don't think we can call update_tasks_cpumask() under
> rcu_read_lock().  It calls into set_cpus_allowed_ptr() which may
> block, so you'll probably have to punt it to a work item like how

Oh, will fix.

> migration is being done.  Another approach would be converting cgroup
> to use SRCU instead, which would lessen pain on other places too.  The
> only problem there would be that srcu_read_lock() is a bit more
> expensive than rcu_read_lock().  I'm not sure whether that'd show up
> in some hot path or not.  Ideas?
> 

I guess we can live with rcu_read_lock() for now, and see if we can
change it to srcu when other significant changes are done in all
cgroup controllers. (hierarchy support in blkcg, etc..)

Re: [PATCH 09/10] cpuset: allow to keep tasks in empty cpusets

[Li Zefan]

On 2013/4/20 4:58, Tejun Heo wrote:
> Hello,
> 
> On Fri, Apr 19, 2013 at 08:29:24PM +0800, Li Zefan wrote:
>> +static void update_tasks_cpumask_hier(struct cpuset *root_cs,
>> +				      bool update_root, struct ptr_heap *heap)
>> +{
>> +	struct cpuset *cp;
>> +	struct cgroup *pos_cgrp;
>> +
>> +	if (update_root)
>> +		update_tasks_cpumask(root_cs, heap);
>> +
>> +	rcu_read_lock();
>> +	cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) {
>> +		/* skip the whole subtree if @cp have some CPU */
>> +		if (!cpumask_empty(cp->cpus_allowed)) {
>> +			pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
>> +			continue;
>> +		}
>> +
>> +		update_tasks_cpumask(cp, heap);
>> +	}
>> +	rcu_read_unlock();
> 
> I don't think we can call update_tasks_cpumask() under
> rcu_read_lock().  It calls into set_cpus_allowed_ptr() which may
> block, so you'll probably have to punt it to a work item like how

Oh, will fix.

> migration is being done.  Another approach would be converting cgroup
> to use SRCU instead, which would lessen pain on other places too.  The
> only problem there would be that srcu_read_lock() is a bit more
> expensive than rcu_read_lock().  I'm not sure whether that'd show up
> in some hot path or not.  Ideas?
> 

I guess we can live with rcu_read_lock() for now, and see if we can
change it to srcu when other significant changes are done in all
cgroup controllers. (hierarchy support in blkcg, etc..)

Re: [PATCH 09/10] cpuset: allow to keep tasks in empty cpusets

[Li Zefan]

On 2013/4/20 4:58, Tejun Heo wrote:
> Hello,
> 
> On Fri, Apr 19, 2013 at 08:29:24PM +0800, Li Zefan wrote:
>> +static void update_tasks_cpumask_hier(struct cpuset *root_cs,
>> +				      bool update_root, struct ptr_heap *heap)
>> +{
>> +	struct cpuset *cp;
>> +	struct cgroup *pos_cgrp;
>> +
>> +	if (update_root)
>> +		update_tasks_cpumask(root_cs, heap);
>> +
>> +	rcu_read_lock();
>> +	cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) {
>> +		/* skip the whole subtree if @cp have some CPU */
>> +		if (!cpumask_empty(cp->cpus_allowed)) {
>> +			pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
>> +			continue;
>> +		}
>> +
>> +		update_tasks_cpumask(cp, heap);
>> +	}
>> +	rcu_read_unlock();
> 
> I don't think we can call update_tasks_cpumask() under
> rcu_read_lock().  It calls into set_cpus_allowed_ptr() which may
> block, so you'll probably have to punt it to a work item like how

Oh, will fix.

> migration is being done.  Another approach would be converting cgroup
> to use SRCU instead, which would lessen pain on other places too.  The
> only problem there would be that srcu_read_lock() is a bit more
> expensive than rcu_read_lock().  I'm not sure whether that'd show up
> in some hot path or not.  Ideas?
> 

I guess we can live with rcu_read_lock() for now, and see if we can
change it to srcu when other significant changes are done in all
cgroup controllers. (hierarchy support in blkcg, etc..)

Re: [PATCH 09/10] cpuset: allow to keep tasks in empty cpusets

[Tejun Heo]

Hello, Li.

On Sat, Apr 20, 2013 at 09:22:09AM +0800, Li Zefan wrote:
> > migration is being done.  Another approach would be converting cgroup
> > to use SRCU instead, which would lessen pain on other places too.  The
> > only problem there would be that srcu_read_lock() is a bit more
> > expensive than rcu_read_lock().  I'm not sure whether that'd show up
> > in some hot path or not.  Ideas?
> 
> I guess we can live with rcu_read_lock() for now, and see if we can
> change it to srcu when other significant changes are done in all
> cgroup controllers. (hierarchy support in blkcg, etc..)

Ooh, I'm almost done with converting blk-throttle to support hierarchy
as cfq is already ready, blkcg will be ready very soon.  The only
thing left to do in terms of hierarchy support would be internal
weight for cpu.  :)

Thanks.

-- 
tejun

Re: [PATCH 09/10] cpuset: allow to keep tasks in empty cpusets

[Tejun Heo]

Hello, Li.

On Sat, Apr 20, 2013 at 09:22:09AM +0800, Li Zefan wrote:
> > migration is being done.  Another approach would be converting cgroup
> > to use SRCU instead, which would lessen pain on other places too.  The
> > only problem there would be that srcu_read_lock() is a bit more
> > expensive than rcu_read_lock().  I'm not sure whether that'd show up
> > in some hot path or not.  Ideas?
> 
> I guess we can live with rcu_read_lock() for now, and see if we can
> change it to srcu when other significant changes are done in all
> cgroup controllers. (hierarchy support in blkcg, etc..)

Ooh, I'm almost done with converting blk-throttle to support hierarchy
as cfq is already ready, blkcg will be ready very soon.  The only
thing left to do in terms of hierarchy support would be internal
weight for cpu.  :)

Thanks.

-- 
tejun

[PATCH 09/10] cpuset: allow to keep tasks in empty cpusets

[Li Zefan]

To achieve this:

- We call update_tasks_cpumask/nodemask() for empty cpusets when
hotplug happens, instead of moving tasks out of them.

- When a cpuset's masks are changed by writing cpuset.cpus/mems,
we also update tasks in child cpusets which are empty.

Signed-off-by: Li Zefan <lizefan-hv44wF8Li93QT0dZR+AlfA@public.gmane.org>
---
 kernel/cpuset.c | 171 +++++++++++++++++++++++++++++++++++++++++++++++---------
 1 file changed, 146 insertions(+), 25 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index 741e652..95e9394 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -99,6 +99,9 @@ struct cpuset {
 	 */
 	nodemask_t old_mems_allowed;
 
+	/* used in cpuset_update_nodemask_workfn() */
+	struct ptr_heap *heap;
+
 	struct fmeter fmeter;		/* memory_pressure filter */
 
 	/*
@@ -114,6 +117,7 @@ struct cpuset {
 	int relax_domain_level;
 
 	struct work_struct hotplug_work;
+	struct work_struct update_nodemask_work;
 };
 
 /* Retrieve the cpuset for a cgroup */
@@ -276,6 +280,8 @@ static struct cpuset top_cpuset = {
 static DEFINE_MUTEX(cpuset_mutex);
 static DEFINE_MUTEX(callback_mutex);
 
+static struct workqueue_struct *cpuset_update_nodemask_wq;
+
 /*
  * CPU / memory hotplug is handled asynchronously.
  */
@@ -877,6 +883,39 @@ static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
 	cgroup_scan_tasks(&scan);
 }
 
+/*
+ * update_tasks_cpumask_hier - Update the cpumasks of tasks in the hierarchy.
+ * @root_cs: the root cpuset of the hierarchy
+ * @update_root: update root cpuset or not?
+ * @heap: the heap used by cgroup_scan_tasks()
+ *
+ * This will update cpumasks of tasks in @root_cs and all other empty cpusets
+ * which take on cpumask of @root_cs.
+ *
+ * Called with cpuset_mutex held
+ */
+static void update_tasks_cpumask_hier(struct cpuset *root_cs,
+				      bool update_root, struct ptr_heap *heap)
+{
+	struct cpuset *cp;
+	struct cgroup *pos_cgrp;
+
+	if (update_root)
+		update_tasks_cpumask(root_cs, heap);
+
+	rcu_read_lock();
+	cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) {
+		/* skip the whole subtree if @cp have some CPU */
+		if (!cpumask_empty(cp->cpus_allowed)) {
+			pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
+			continue;
+		}
+
+		update_tasks_cpumask(cp, heap);
+	}
+	rcu_read_unlock();
+}
+
 /**
  * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
  * @cs: the cpuset to consider
@@ -928,11 +967,7 @@ static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
 	cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
 	mutex_unlock(&callback_mutex);
 
-	/*
-	 * Scan tasks in the cpuset, and update the cpumasks of any
-	 * that need an update.
-	 */
-	update_tasks_cpumask(cs, &heap);
+	update_tasks_cpumask_hier(cs, true, &heap);
 
 	heap_free(&heap);
 
@@ -1099,6 +1134,78 @@ static void update_tasks_nodemask(struct cpuset *cs, struct ptr_heap *heap)
 	cpuset_being_rebound = NULL;
 }
 
+static void cpuset_update_nodemask_workfn(struct work_struct *work)
+{
+	struct cpuset *cs = container_of(work, struct cpuset,
+					update_nodemask_work);
+
+	update_tasks_nodemask(cs, cs->heap);
+	css_put(&cs->css);
+}
+
+static void schedule_update_tasks_nodemask(struct cpuset *cs,
+					   struct ptr_heap *heap)
+{
+	bool queued;
+
+	/* Will be released when the work item finishes executing. */
+	if (!css_tryget(&cs->css))
+		return;
+
+	/*
+	 * The caller will flush the workqueue with cpuset_mutex held,
+	 * so it's not possible a work item was already queued, and
+	 * we're sure cs->heap is valid.
+	 */
+	cs->heap = heap;
+	queued = queue_work(cpuset_update_nodemask_wq,
+			    &cs->update_nodemask_work);
+	if (queued) {
+		WARN_ON(1);
+		css_put(&cs->css);
+	}
+}
+
+/*
+ * update_tasks_nodemask_hier - Update the nodemasks of tasks in the hierarchy.
+ * @cs: the root cpuset of the hierarchy
+ * @update_root: update the root cpuset or not?
+ * @heap: the heap used by cgroup_scan_tasks()
+ *
+ * This will update nodemasks of tasks in @root_cs and all other empty cpusets
+ * which take on nodemask of @root_cs.
+ *
+ * Called with cpuset_mutex held
+ */
+static void update_tasks_nodemask_hier(struct cpuset *root_cs,
+				       bool update_root, struct ptr_heap *heap)
+{
+	struct cpuset *cp;
+	struct cgroup *pos_cgrp;
+
+	if (update_root)
+		update_tasks_nodemask(root_cs, heap);
+
+	rcu_read_lock();
+	cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) {
+		/* skip the whole subtree if @cp have some CPU */
+		if (!nodes_empty(cp->mems_allowed)) {
+			pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
+			continue;
+		}
+
+		schedule_update_tasks_nodemask(cp, heap);
+	}
+	rcu_read_unlock();
+
+	/*
+	 * The only reason we use workqueue is update_tasks_nodemask() can't
+	 * be called in rcu_read_lock(). Flush the workqueue to make sure
+	 * all the updates are done before we return.
+	 */
+	flush_workqueue(cpuset_update_nodemask_wq);
+}
+
 /*
  * Handle user request to change the 'mems' memory placement
  * of a cpuset.  Needs to validate the request, update the
@@ -1163,7 +1270,7 @@ static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
 	cs->mems_allowed = trialcs->mems_allowed;
 	mutex_unlock(&callback_mutex);
 
-	update_tasks_nodemask(cs, &heap);
+	update_tasks_nodemask_hier(cs, true, &heap);
 
 	heap_free(&heap);
 done:
@@ -1888,6 +1995,7 @@ static struct cgroup_subsys_state *cpuset_css_alloc(struct cgroup *cont)
 	nodes_clear(cs->mems_allowed);
 	fmeter_init(&cs->fmeter);
 	INIT_WORK(&cs->hotplug_work, cpuset_propagate_hotplug_workfn);
+	INIT_WORK(&cs->update_nodemask_work, cpuset_update_nodemask_workfn);
 	cs->relax_domain_level = -1;
 
 	return &cs->css;
@@ -2063,31 +2171,36 @@ static void cpuset_propagate_hotplug_workfn(struct work_struct *work)
 	static nodemask_t off_mems;
 	struct cpuset *cs = container_of(work, struct cpuset, hotplug_work);
 	bool is_empty;
+	bool sane = cgroup_sane_behavior(cs->css.cgroup);
 
 	mutex_lock(&cpuset_mutex);
 
 	cpumask_andnot(&off_cpus, cs->cpus_allowed, top_cpuset.cpus_allowed);
 	nodes_andnot(off_mems, cs->mems_allowed, top_cpuset.mems_allowed);
 
-	/* remove offline cpus from @cs */
-	if (!cpumask_empty(&off_cpus)) {
-		mutex_lock(&callback_mutex);
-		cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
-		mutex_unlock(&callback_mutex);
+	mutex_lock(&callback_mutex);
+	cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
+	mutex_unlock(&callback_mutex);
 
-		if (!cpumask_empty(cs->cpus_allowed))
-			update_tasks_cpumask(cs, NULL);
-	}
+	/*
+	 * If sane_behavior flag is set, we need to update tasks' cpumask
+	 * for empy cpuset to take on ancestor's cpumask
+	 */
+	if ((sane && cpumask_empty(cs->cpus_allowed)) ||
+	    (!cpumask_empty(&off_cpus) && !cpumask_empty(cs->cpus_allowed)))
+		update_tasks_cpumask(cs, NULL);
 
-	/* remove offline mems from @cs */
-	if (!nodes_empty(off_mems)) {
-		mutex_lock(&callback_mutex);
-		nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
-		mutex_unlock(&callback_mutex);
+	mutex_lock(&callback_mutex);
+	nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
+	mutex_unlock(&callback_mutex);
 
-		if (!nodes_empty(cs->mems_allowed))
-			update_tasks_nodemask(cs, NULL);
-	}
+	/*
+	 * If sane_behavior flag is set, we need to update tasks' nodemask
+	 * for empy cpuset to take on ancestor's nodemask
+	 */
+	if ((sane && nodes_empty(cs->mems_allowed)) ||
+	    (!nodes_empty(off_mems) && !nodes_empty(cs->mems_allowed)))
+		update_tasks_nodemask(cs, NULL);
 
 	is_empty = cpumask_empty(cs->cpus_allowed) ||
 		nodes_empty(cs->mems_allowed);
@@ -2095,11 +2208,13 @@ static void cpuset_propagate_hotplug_workfn(struct work_struct *work)
 	mutex_unlock(&cpuset_mutex);
 
 	/*
-	 * If @cs became empty, move tasks to the nearest ancestor with
-	 * execution resources.  This is full cgroup operation which will
+	 * If sane_behavior flag is set, we'll keep tasks in empty cpusets.
+	 *
+	 * Otherwise move tasks to the nearest ancestor with execution
+	 *  resources.  This is full cgroup operation which will
 	 * also call back into cpuset.  Should be done outside any lock.
 	 */
-	if (is_empty)
+	if (!sane && is_empty)
 		remove_tasks_in_empty_cpuset(cs);
 
 	/* the following may free @cs, should be the last operation */
@@ -2174,6 +2289,7 @@ static void cpuset_hotplug_workfn(struct work_struct *work)
 		cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
 		mutex_unlock(&callback_mutex);
 		/* we don't mess with cpumasks of tasks in top_cpuset */
+		update_tasks_cpumask_hier(&top_cpuset, false, NULL);
 	}
 
 	/* synchronize mems_allowed to N_MEMORY */
@@ -2182,6 +2298,7 @@ static void cpuset_hotplug_workfn(struct work_struct *work)
 		top_cpuset.mems_allowed = new_mems;
 		mutex_unlock(&callback_mutex);
 		update_tasks_nodemask(&top_cpuset, NULL);
+		update_tasks_nodemask_hier(&top_cpuset, false, NULL);
 	}
 
 	/* if cpus or mems went down, we need to propagate to descendants */
@@ -2261,6 +2378,10 @@ void __init cpuset_init_smp(void)
 	cpuset_propagate_hotplug_wq =
 		alloc_ordered_workqueue("cpuset_hotplug", 0);
 	BUG_ON(!cpuset_propagate_hotplug_wq);
+
+	cpuset_update_nodemask_wq =
+		create_workqueue("cpuset_update_nodemask");
+	BUG_ON(!cpuset_update_nodemask_wq);
 }
 
 /**
-- 
1.8.0.2

[PATCH 10/10] cpuset: allow to move tasks to empty cpusets

[Li Zefan]

Currently some cpuset behaviors are not friendly when cpuset is co-mounted
with other cgroup controllers.

Now with this patchset if cpuset is mounted with sane_behavior option,
it behaves differently:

- Tasks will be kept in empty cpusets when hotplug happens and take
  masks of ancestors with non-empty cpus/mems, instead of being moved to
  an ancestor.

- A task can be moved into an empty cpuset, and again it takes masks of
  ancestors, so we drop a task into a newly created cgroup without the
  user having to do anything for it.

As tasks can reside in empy cpusets, here're some rules:

- They can be moved to another cpuset, regardless it's empty or not.

- Though it takes masks from ancestors, it takes other configs from the
  empty cpuset.

- If the ancestors' masks are changed, those tasks will also be updated
  to take new masks.

Signed-off-by: Li Zefan <lizefan-hv44wF8Li93QT0dZR+AlfA@public.gmane.org>
---
 kernel/cpuset.c | 12 +++++++++---
 1 file changed, 9 insertions(+), 3 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index 95e9394..646c794 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -489,7 +489,7 @@ static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
 	 */
 	ret = -ENOSPC;
 	if ((cgroup_task_count(cur->css.cgroup) || cur->attach_in_progress) &&
-	    (cpumask_empty(trial->cpus_allowed) ||
+	    (cpumask_empty(trial->cpus_allowed) &&
 	     nodes_empty(trial->mems_allowed)))
 		goto out;
 
@@ -1502,8 +1502,13 @@ static int cpuset_can_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 
 	mutex_lock(&cpuset_mutex);
 
+	/*
+	 * We allow to move tasks into an empty cpuset if sane_behavior
+	 * flag is set.
+	 */
 	ret = -ENOSPC;
-	if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
+	if (!cgroup_sane_behavior(cgrp) &&
+	    (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
 		goto out_unlock;
 
 	cgroup_taskset_for_each(task, cgrp, tset) {
@@ -1608,7 +1613,8 @@ static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 	 * propagation if @cs doesn't have any CPU or memory.  It will move
 	 * the newly added tasks to the nearest parent which can execute.
 	 */
-	if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
+	if (!cgroup_sane_behavior(cgrp) &&
+	    (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
 		schedule_cpuset_propagate_hotplug(cs);
 
 	mutex_unlock(&cpuset_mutex);
-- 
1.8.0.2

[PATCH 10/10] cpuset: allow to move tasks to empty cpusets

[Li Zefan]

Currently some cpuset behaviors are not friendly when cpuset is co-mounted
with other cgroup controllers.

Now with this patchset if cpuset is mounted with sane_behavior option,
it behaves differently:

- Tasks will be kept in empty cpusets when hotplug happens and take
  masks of ancestors with non-empty cpus/mems, instead of being moved to
  an ancestor.

- A task can be moved into an empty cpuset, and again it takes masks of
  ancestors, so we drop a task into a newly created cgroup without the
  user having to do anything for it.

As tasks can reside in empy cpusets, here're some rules:

- They can be moved to another cpuset, regardless it's empty or not.

- Though it takes masks from ancestors, it takes other configs from the
  empty cpuset.

- If the ancestors' masks are changed, those tasks will also be updated
  to take new masks.

Signed-off-by: Li Zefan <lizefan@huawei.com>
---
 kernel/cpuset.c | 12 +++++++++---
 1 file changed, 9 insertions(+), 3 deletions(-)

diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index 95e9394..646c794 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -489,7 +489,7 @@ static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
 	 */
 	ret = -ENOSPC;
 	if ((cgroup_task_count(cur->css.cgroup) || cur->attach_in_progress) &&
-	    (cpumask_empty(trial->cpus_allowed) ||
+	    (cpumask_empty(trial->cpus_allowed) &&
 	     nodes_empty(trial->mems_allowed)))
 		goto out;
 
@@ -1502,8 +1502,13 @@ static int cpuset_can_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 
 	mutex_lock(&cpuset_mutex);
 
+	/*
+	 * We allow to move tasks into an empty cpuset if sane_behavior
+	 * flag is set.
+	 */
 	ret = -ENOSPC;
-	if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
+	if (!cgroup_sane_behavior(cgrp) &&
+	    (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
 		goto out_unlock;
 
 	cgroup_taskset_for_each(task, cgrp, tset) {
@@ -1608,7 +1613,8 @@ static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 	 * propagation if @cs doesn't have any CPU or memory.  It will move
 	 * the newly added tasks to the nearest parent which can execute.
 	 */
-	if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
+	if (!cgroup_sane_behavior(cgrp) &&
+	    (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
 		schedule_cpuset_propagate_hotplug(cs);
 
 	mutex_unlock(&cpuset_mutex);
-- 
1.8.0.2