Merge branch 'for-3.11-cpuset' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/cgroup

Pull cpuset changes from Tejun Heo:
 "cpuset has always been rather odd about its configurations - a cgroup
  right after creation didn't allow any task executions before
  configuration, changing configuration in the parent modifies the
  descendants irreversibly and so on.  These behaviors are inherently
  nasty and almost hostile against sharing the hierarchy with other
  controllers making it very difficult to use in unified hierarchy.

  Li is currently in the process of updating the behaviors for
  __DEVEL__sane_behavior which is the bulk of changes in this pull
  request.  It isn't complete yet and the behaviors will change further
  but all changes are gated behind sane_behavior.  In the process, the
  rather hairy work-item punting which was used to work around the
  limitations of cgroup descendant iterator was simplified."

* 'for-3.11-cpuset' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/cgroup:
  cpuset: rename @cont to @cgrp
  cpuset: fix to migrate mm correctly in a corner case
  cpuset: allow to move tasks to empty cpusets
  cpuset: allow to keep tasks in empty cpusets
  cpuset: introduce effective_{cpumask|nodemask}_cpuset()
  cpuset: record old_mems_allowed in struct cpuset
  cpuset: remove async hotplug propagation work
  cpuset: let hotplug propagation work wait for task attaching
  cpuset: re-structure update_cpumask() a bit
  cpuset: remove cpuset_test_cpumask()
  cpuset: remove unnecessary variable in cpuset_attach()
  cpuset: cleanup guarantee_online_{cpus|mems}()
  cpuset: remove redundant check in cpuset_cpus_allowed_fallback()

2 files changed, 295 insertions, 192 deletions

diff --git a/include/linux/cgroup.h b/include/linux/cgroup.h
index 8db53974f7b..fd097ecfcd9 100644
--- a/include/linux/cgroup.h
+++ b/include/linux/cgroup.h
@@ -261,13 +261,20 @@ enum {
 	 *
 	 * - Remount is disallowed.
 	 *
+	 * - rename(2) is disallowed.
+	 *
 	 * - "tasks" is removed.  Everything should be at process
 	 *   granularity.  Use "cgroup.procs" instead.
 	 *
 	 * - "release_agent" and "notify_on_release" are removed.
 	 *   Replacement notification mechanism will be implemented.
 	 *
-	 * - rename(2) is disallowed.
+	 * - cpuset: 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.
+	 *
+	 * - cpuset: a task can be moved into an empty cpuset, and again it
+	 *   takes masks of ancestors.
 	 *
 	 * - memcg: use_hierarchy is on by default and the cgroup file for
 	 *   the flag is not created.
diff --git a/kernel/cpuset.c b/kernel/cpuset.c
index 902d13fc2b1..e5657788fed 100644
--- a/kernel/cpuset.c
+++ b/kernel/cpuset.c
@@ -59,6 +59,7 @@
 #include <linux/mutex.h>
 #include <linux/workqueue.h>
 #include <linux/cgroup.h>
+#include <linux/wait.h>
 
 /*
  * Tracks how many cpusets are currently defined in system.
@@ -87,6 +88,18 @@ 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 */
 
 	/*
@@ -100,14 +113,12 @@ struct cpuset {
 
 	/* for custom sched domain */
 	int relax_domain_level;
-
-	struct work_struct hotplug_work;
 };
 
 /* Retrieve the cpuset for a cgroup */
-static inline struct cpuset *cgroup_cs(struct cgroup *cont)
+static inline struct cpuset *cgroup_cs(struct cgroup *cgrp)
 {
-	return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
+	return container_of(cgroup_subsys_state(cgrp, cpuset_subsys_id),
 			    struct cpuset, css);
 }
 
@@ -267,14 +278,11 @@ static DEFINE_MUTEX(callback_mutex);
 /*
  * CPU / memory hotplug is handled asynchronously.
  */
-static struct workqueue_struct *cpuset_propagate_hotplug_wq;
-
 static void cpuset_hotplug_workfn(struct work_struct *work);
-static void cpuset_propagate_hotplug_workfn(struct work_struct *work);
-static void schedule_cpuset_propagate_hotplug(struct cpuset *cs);
-
 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
 
+static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
+
 /*
  * This is ugly, but preserves the userspace API for existing cpuset
  * users. If someone tries to mount the "cpuset" filesystem, we
@@ -304,53 +312,38 @@ 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]);
 }
 
 /*
@@ -440,7 +433,7 @@ static void free_trial_cpuset(struct cpuset *trial)
 
 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
 {
-	struct cgroup *cont;
+	struct cgroup *cgrp;
 	struct cpuset *c, *par;
 	int ret;
 
@@ -448,7 +441,7 @@ static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
 
 	/* Each of our child cpusets must be a subset of us */
 	ret = -EBUSY;
-	cpuset_for_each_child(c, cont, cur)
+	cpuset_for_each_child(c, cgrp, cur)
 		if (!is_cpuset_subset(c, trial))
 			goto out;
 
@@ -469,7 +462,7 @@ static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
 	 * overlap
 	 */
 	ret = -EINVAL;
-	cpuset_for_each_child(c, cont, par) {
+	cpuset_for_each_child(c, cgrp, par) {
 		if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
 		    c != cur &&
 		    cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
@@ -486,7 +479,7 @@ 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;
 
@@ -798,21 +791,43 @@ void rebuild_sched_domains(void)
 	mutex_unlock(&cpuset_mutex);
 }
 
-/**
- * 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
+/*
+ * effective_cpumask_cpuset - return nearest ancestor with non-empty cpus
+ * @cs: the cpuset in interest
  *
- * 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).
+ * 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 int cpuset_test_cpumask(struct task_struct *tsk,
-			       struct cgroup_scanner *scan)
+static struct cpuset *effective_cpumask_cpuset(struct cpuset *cs)
 {
-	return !cpumask_equal(&tsk->cpus_allowed,
-			(cgroup_cs(scan->cg))->cpus_allowed);
+	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;
 }
 
 /**
@@ -829,7 +844,10 @@ static int cpuset_test_cpumask(struct task_struct *tsk,
 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);
 }
 
 /**
@@ -850,12 +868,51 @@ 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);
 }
 
+/*
+ * 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;
+		}
+		if (!css_tryget(&cp->css))
+			continue;
+		rcu_read_unlock();
+
+		update_tasks_cpumask(cp, heap);
+
+		rcu_read_lock();
+		css_put(&cp->css);
+	}
+	rcu_read_unlock();
+}
+
 /**
  * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
  * @cs: the cpuset to consider
@@ -888,14 +945,15 @@ 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;
@@ -906,11 +964,7 @@ 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);
 
@@ -943,12 +997,14 @@ 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);
 }
 
 /*
@@ -1007,16 +1063,12 @@ 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)
@@ -1026,7 +1078,7 @@ 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);
 }
 
@@ -1035,25 +1087,27 @@ 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;
+	struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
 
 	cpuset_being_rebound = cs;		/* causes mpol_dup() rebind */
 
+	guarantee_online_mems(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
@@ -1067,11 +1121,56 @@ 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;
 }
 
 /*
+ * 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;
+		}
+		if (!css_tryget(&cp->css))
+			continue;
+		rcu_read_unlock();
+
+		update_tasks_nodemask(cp, heap);
+
+		rcu_read_lock();
+		css_put(&cp->css);
+	}
+	rcu_read_unlock();
+}
+
+/*
  * Handle user request to change the 'mems' memory placement
  * of a cpuset.  Needs to validate the request, update the
  * cpusets mems_allowed, and for each task in the cpuset,
@@ -1087,13 +1186,9 @@ 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
@@ -1122,8 +1217,8 @@ 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;
 	}
@@ -1139,11 +1234,10 @@ 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_hier(cs, true, &heap);
 
 	heap_free(&heap);
 done:
-	NODEMASK_FREE(oldmem);
 	return retval;
 }
 
@@ -1372,8 +1466,13 @@ 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) {
@@ -1422,8 +1521,7 @@ 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;
@@ -1431,6 +1529,8 @@ 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);
 
@@ -1438,9 +1538,9 @@ 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) {
 		/*
@@ -1457,26 +1557,32 @@ 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) {
+		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, &cpuset_attach_nodemask_from,
+
+		/*
+		 * old_mems_allowed is the same with mems_allowed here, except
+		 * if this task is being moved automatically due to hotplug.
+		 * In that case @mems_allowed has been updated and is empty,
+		 * so @old_mems_allowed is the right nodesets that we migrate
+		 * mm from.
+		 */
+		if (is_memory_migrate(cs)) {
+			cpuset_migrate_mm(mm, &mems_oldcs->old_mems_allowed,
 					  &cpuset_attach_nodemask_to);
+		}
 		mmput(mm);
 	}
 
-	cs->attach_in_progress--;
+	cs->old_mems_allowed = cpuset_attach_nodemask_to;
 
-	/*
-	 * We may have raced with CPU/memory hotunplug.  Trigger hotplug
-	 * 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))
-		schedule_cpuset_propagate_hotplug(cs);
+	cs->attach_in_progress--;
+	if (!cs->attach_in_progress)
+		wake_up(&cpuset_attach_wq);
 
 	mutex_unlock(&cpuset_mutex);
 }
@@ -1588,13 +1694,8 @@ static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
 	 * resources, wait for the previously scheduled operations before
 	 * proceeding, so that we don't end up keep removing tasks added
 	 * after execution capability is restored.
-	 *
-	 * Flushing cpuset_hotplug_work is enough to synchronize against
-	 * hotplug hanlding; however, cpuset_attach() may schedule
-	 * propagation work directly.  Flush the workqueue too.
 	 */
 	flush_work(&cpuset_hotplug_work);
-	flush_workqueue(cpuset_propagate_hotplug_wq);
 
 	mutex_lock(&cpuset_mutex);
 	if (!is_cpuset_online(cs))
@@ -1658,13 +1759,13 @@ static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs)
 	return count;
 }
 
-static ssize_t cpuset_common_file_read(struct cgroup *cont,
+static ssize_t cpuset_common_file_read(struct cgroup *cgrp,
 				       struct cftype *cft,
 				       struct file *file,
 				       char __user *buf,
 				       size_t nbytes, loff_t *ppos)
 {
-	struct cpuset *cs = cgroup_cs(cont);
+	struct cpuset *cs = cgroup_cs(cgrp);
 	cpuset_filetype_t type = cft->private;
 	char *page;
 	ssize_t retval = 0;
@@ -1694,9 +1795,9 @@ out:
 	return retval;
 }
 
-static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
+static u64 cpuset_read_u64(struct cgroup *cgrp, struct cftype *cft)
 {
-	struct cpuset *cs = cgroup_cs(cont);
+	struct cpuset *cs = cgroup_cs(cgrp);
 	cpuset_filetype_t type = cft->private;
 	switch (type) {
 	case FILE_CPU_EXCLUSIVE:
@@ -1725,9 +1826,9 @@ static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
 	return 0;
 }
 
-static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
+static s64 cpuset_read_s64(struct cgroup *cgrp, struct cftype *cft)
 {
-	struct cpuset *cs = cgroup_cs(cont);
+	struct cpuset *cs = cgroup_cs(cgrp);
 	cpuset_filetype_t type = cft->private;
 	switch (type) {
 	case FILE_SCHED_RELAX_DOMAIN_LEVEL:
@@ -1839,14 +1940,14 @@ static struct cftype files[] = {
 
 /*
  *	cpuset_css_alloc - allocate a cpuset css
- *	cont:	control group that the new cpuset will be part of
+ *	cgrp:	control group that the new cpuset will be part of
  */
 
-static struct cgroup_subsys_state *cpuset_css_alloc(struct cgroup *cont)
+static struct cgroup_subsys_state *cpuset_css_alloc(struct cgroup *cgrp)
 {
 	struct cpuset *cs;
 
-	if (!cont->parent)
+	if (!cgrp->parent)
 		return &top_cpuset.css;
 
 	cs = kzalloc(sizeof(*cs), GFP_KERNEL);
@@ -1861,7 +1962,6 @@ static struct cgroup_subsys_state *cpuset_css_alloc(struct cgroup *cont)
 	cpumask_clear(cs->cpus_allowed);
 	nodes_clear(cs->mems_allowed);
 	fmeter_init(&cs->fmeter);
-	INIT_WORK(&cs->hotplug_work, cpuset_propagate_hotplug_workfn);
 	cs->relax_domain_level = -1;
 
 	return &cs->css;
@@ -1942,9 +2042,9 @@ static void cpuset_css_offline(struct cgroup *cgrp)
  * will call rebuild_sched_domains_locked().
  */
 
-static void cpuset_css_free(struct cgroup *cont)
+static void cpuset_css_free(struct cgroup *cgrp)
 {
-	struct cpuset *cs = cgroup_cs(cont);
+	struct cpuset *cs = cgroup_cs(cgrp);
 
 	free_cpumask_var(cs->cpus_allowed);
 	kfree(cs);
@@ -2024,41 +2124,64 @@ static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
 }
 
 /**
- * cpuset_propagate_hotplug_workfn - propagate CPU/memory hotplug to a cpuset
+ * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
  * @cs: cpuset in interest
  *
  * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
  * offline, update @cs accordingly.  If @cs ends up with no CPU or memory,
  * all its tasks are moved to the nearest ancestor with both resources.
  */
-static void cpuset_propagate_hotplug_workfn(struct work_struct *work)
+static void cpuset_hotplug_update_tasks(struct cpuset *cs)
 {
 	static cpumask_t off_cpus;
-	static nodemask_t off_mems, tmp_mems;
-	struct cpuset *cs = container_of(work, struct cpuset, hotplug_work);
+	static nodemask_t off_mems;
 	bool is_empty;
+	bool sane = cgroup_sane_behavior(cs->css.cgroup);
+
+retry:
+	wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
 
 	mutex_lock(&cpuset_mutex);
 
+	/*
+	 * We have raced with task attaching. We wait until attaching
+	 * is finished, so we won't attach a task to an empty cpuset.
+	 */
+	if (cs->attach_in_progress) {
+		mutex_unlock(&cpuset_mutex);
+		goto retry;
+	}
+
 	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 sane_behavior flag is set, we need to update tasks' cpumask
+	 * for empty cpuset to take on ancestor's cpumask. Otherwise, don't
+	 * call update_tasks_cpumask() if the cpuset becomes empty, as
+	 * the tasks in it will be migrated to an ancestor.
+	 */
+	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)) {
-		tmp_mems = cs->mems_allowed;
-		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);
-	}
+	mutex_lock(&callback_mutex);
+	nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
+	mutex_unlock(&callback_mutex);
+
+	/*
+	 * If sane_behavior flag is set, we need to update tasks' nodemask
+	 * for empty cpuset to take on ancestor's nodemask. Otherwise, don't
+	 * call update_tasks_nodemask() if the cpuset becomes empty, as
+	 * the tasks in it will be migratd to an ancestor.
+	 */
+	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);
@@ -2066,40 +2189,14 @@ 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 */
-	css_put(&cs->css);
-}
-
-/**
- * schedule_cpuset_propagate_hotplug - schedule hotplug propagation to a cpuset
- * @cs: cpuset of interest
- *
- * Schedule cpuset_propagate_hotplug_workfn() which will update CPU and
- * memory masks according to top_cpuset.
- */
-static void schedule_cpuset_propagate_hotplug(struct cpuset *cs)
-{
-	/*
-	 * Pin @cs.  The refcnt will be released when the work item
-	 * finishes executing.
-	 */
-	if (!css_tryget(&cs->css))
-		return;
-
-	/*
-	 * Queue @cs->hotplug_work.  If already pending, lose the css ref.
-	 * cpuset_propagate_hotplug_wq is ordered and propagation will
-	 * happen in the order this function is called.
-	 */
-	if (!queue_work(cpuset_propagate_hotplug_wq, &cs->hotplug_work))
-		css_put(&cs->css);
 }
 
 /**
@@ -2112,18 +2209,17 @@ static void schedule_cpuset_propagate_hotplug(struct cpuset *cs)
  * actively using CPU hotplug but making no active use of cpusets.
  *
  * Non-root cpusets are only affected by offlining.  If any CPUs or memory
- * nodes have been taken down, cpuset_propagate_hotplug() is invoked on all
- * descendants.
+ * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
+ * all descendants.
  *
  * Note that CPU offlining during suspend is ignored.  We don't modify
  * cpusets across suspend/resume cycles at all.
  */
 static void cpuset_hotplug_workfn(struct work_struct *work)
 {
-	static cpumask_t new_cpus, tmp_cpus;
-	static nodemask_t new_mems, tmp_mems;
+	static cpumask_t new_cpus;
+	static nodemask_t new_mems;
 	bool cpus_updated, mems_updated;
-	bool cpus_offlined, mems_offlined;
 
 	mutex_lock(&cpuset_mutex);
 
@@ -2132,12 +2228,7 @@ static void cpuset_hotplug_workfn(struct work_struct *work)
 	new_mems = node_states[N_MEMORY];
 
 	cpus_updated = !cpumask_equal(top_cpuset.cpus_allowed, &new_cpus);
-	cpus_offlined = cpumask_andnot(&tmp_cpus, top_cpuset.cpus_allowed,
-				       &new_cpus);
-
 	mems_updated = !nodes_equal(top_cpuset.mems_allowed, new_mems);
-	nodes_andnot(tmp_mems, top_cpuset.mems_allowed, new_mems);
-	mems_offlined = !nodes_empty(tmp_mems);
 
 	/* synchronize cpus_allowed to cpu_active_mask */
 	if (cpus_updated) {
@@ -2149,28 +2240,32 @@ 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 */
-	if (cpus_offlined || mems_offlined) {
+	mutex_unlock(&cpuset_mutex);
+
+	/* if cpus or mems changed, we need to propagate to descendants */
+	if (cpus_updated || mems_updated) {
 		struct cpuset *cs;
 		struct cgroup *pos_cgrp;
 
 		rcu_read_lock();
-		cpuset_for_each_descendant_pre(cs, pos_cgrp, &top_cpuset)
-			schedule_cpuset_propagate_hotplug(cs);
-		rcu_read_unlock();
-	}
+		cpuset_for_each_descendant_pre(cs, pos_cgrp, &top_cpuset) {
+			if (!css_tryget(&cs->css))
+				continue;
+			rcu_read_unlock();
 
-	mutex_unlock(&cpuset_mutex);
+			cpuset_hotplug_update_tasks(cs);
 
-	/* wait for propagations to finish */
-	flush_workqueue(cpuset_propagate_hotplug_wq);
+			rcu_read_lock();
+			css_put(&cs->css);
+		}
+		rcu_read_unlock();
+	}
 
 	/* rebuild sched domains if cpus_allowed has changed */
 	if (cpus_updated)
@@ -2219,12 +2314,9 @@ 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;
 
 	register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
-
-	cpuset_propagate_hotplug_wq =
-		alloc_ordered_workqueue("cpuset_hotplug", 0);
-	BUG_ON(!cpuset_propagate_hotplug_wq);
 }
 
 /**
@@ -2240,21 +2332,23 @@ 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);
-	if (cs)
-		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();
 
 	/*
@@ -2293,11 +2387,13 @@ 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);