cpuset: Add an error state to cpuset.sched.partition

When external events like CPU offlining or user events like changing
the cpu list of an ancestor cpuset happen, update_cpumasks_hier()
will be called to update the effective cpus of each of the affected
cpusets. That will then call update_parent_subparts_cpumask() if
partitions are impacted.

Currently, these events may cause update_parent_subparts_cpumask()
to return error if none of the requested cpus are available or it will
consume all the cpus in the parent partition root. Handling these errors
is problematic as the states may become inconsistent.

Instead of letting update_parent_subparts_cpumask() return error, a new
error state (-1) is added to the partition_root_state flag to designate
the fact that the partition is no longer valid. IOW, it is no longer a
real partition root, but the CS_CPU_EXCLUSIVE flag will still be set
as it can be changed back to a real one if favorable change happens
later on.

This new error state is set internally and user cannot write this new
value to "cpuset.sched.partition".

Signed-off-by: Waiman Long <longman@redhat.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Tejun Heo <tj@kernel.org>

kernel/cgroup/cpuset.c

@@ -153,10 +153,19 @@ struct cpuset {
  * Partition root states:
  *
  *   0 - not a partition root
+ *
  *   1 - partition root
+ *
+ *  -1 - invalid partition root
+ *       None of the cpus in cpus_allowed can be put into the parent's
+ *       subparts_cpus. In this case, the cpuset is not a real partition
+ *       root anymore.  However, the CPU_EXCLUSIVE bit will still be set
+ *       and the cpuset can be restored back to a partition root if the
+ *       parent cpuset can give more CPUs back to this child cpuset.
  */
 #define PRS_DISABLED		0
 #define PRS_ENABLED		1
+#define PRS_ERROR		-1
 
 /*
  * Temporary cpumasks for working with partitions that are passed among
@@ -251,7 +260,7 @@ static inline int is_spread_slab(const struct cpuset *cs)
 
 static inline int is_partition_root(const struct cpuset *cs)
 {
-	return cs->partition_root_state;
+	return cs->partition_root_state > 0;
 }
 
 static struct cpuset top_cpuset = {
@@ -1021,9 +1030,12 @@ enum subparts_cmd {
  *
  * For partcmd_update, if the optional newmask is specified, the cpu
  * list is to be changed from cpus_allowed to newmask. Otherwise,
- * cpus_allowed is assumed to remain the same.  The function will return
- * 1 if changes to parent's subparts_cpus and effective_cpus happen or 0
- * otherwise. In case of error, an error code will be returned.
+ * cpus_allowed is assumed to remain the same. The cpuset should either
+ * be a partition root or an invalid partition root. The partition root
+ * state may change if newmask is NULL and none of the requested CPUs can
+ * be granted by the parent. The function will return 1 if changes to
+ * parent's subparts_cpus and effective_cpus happen or 0 otherwise.
+ * Error code should only be returned when newmask is non-NULL.
  *
  * The partcmd_enable and partcmd_disable commands are used by
  * update_prstate(). The partcmd_update command is used by
@@ -1046,6 +1058,7 @@ static int update_parent_subparts_cpumask(struct cpuset *cpuset, int cmd,
 	struct cpuset *parent = parent_cs(cpuset);
 	int adding;	/* Moving cpus from effective_cpus to subparts_cpus */
 	int deleting;	/* Moving cpus from subparts_cpus to effective_cpus */
+	bool part_error = false;	/* Partition error? */
 
 	lockdep_assert_held(&cpuset_mutex);
 
@@ -1114,13 +1127,48 @@ static int update_parent_subparts_cpumask(struct cpuset *cpuset, int cmd,
 		 * addmask = cpus_allowed & parent->effectiveb_cpus
 		 *
 		 * Note that parent's subparts_cpus may have been
-		 * pre-shrunk in case the CPUs granted to the parent
-		 * by the grandparent changes. So no deletion is needed.
+		 * pre-shrunk in case there is a change in the cpu list.
+		 * So no deletion is needed.
 		 */
 		adding = cpumask_and(tmp->addmask, cpuset->cpus_allowed,
 				     parent->effective_cpus);
-		if (cpumask_equal(tmp->addmask, parent->effective_cpus))
-			return -EINVAL;
+		part_error = cpumask_equal(tmp->addmask,
+					   parent->effective_cpus);
+	}
+
+	if (cmd == partcmd_update) {
+		int prev_prs = cpuset->partition_root_state;
+
+		/*
+		 * Check for possible transition between PRS_ENABLED
+		 * and PRS_ERROR.
+		 */
+		switch (cpuset->partition_root_state) {
+		case PRS_ENABLED:
+			if (part_error)
+				cpuset->partition_root_state = PRS_ERROR;
+			break;
+		case PRS_ERROR:
+			if (!part_error)
+				cpuset->partition_root_state = PRS_ENABLED;
+			break;
+		}
+		/*
+		 * Set part_error if previously in invalid state.
+		 */
+		part_error = (prev_prs == PRS_ERROR);
+	}
+
+	if (!part_error && (cpuset->partition_root_state == PRS_ERROR))
+		return 0;	/* Nothing need to be done */
+
+	if (cpuset->partition_root_state == PRS_ERROR) {
+		/*
+		 * Remove all its cpus from parent's subparts_cpus.
+		 */
+		adding = false;
+		deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
+				       parent->subparts_cpus);
 	}
 
 	if (!adding && !deleting)
@@ -1172,28 +1220,21 @@ static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp)
 	rcu_read_lock();
 	cpuset_for_each_descendant_pre(cp, pos_css, cs) {
 		struct cpuset *parent = parent_cs(cp);
-		bool cs_empty;
 
 		compute_effective_cpumask(tmp->new_cpus, cp, parent);
-		cs_empty = cpumask_empty(tmp->new_cpus);
-
-		/*
-		 * A partition root cannot have empty effective_cpus
-		 */
-		WARN_ON_ONCE(cs_empty && is_partition_root(cp));
 
 		/*
 		 * If it becomes empty, inherit the effective mask of the
 		 * parent, which is guaranteed to have some CPUs.
 		 */
-		if (is_in_v2_mode() && cs_empty)
+		if (is_in_v2_mode() && cpumask_empty(tmp->new_cpus))
 			cpumask_copy(tmp->new_cpus, parent->effective_cpus);
 
 		/*
 		 * Skip the whole subtree if the cpumask remains the same
 		 * and has no partition root state.
 		 */
-		if (!is_partition_root(cp) &&
+		if (!cp->partition_root_state &&
 		    cpumask_equal(tmp->new_cpus, cp->effective_cpus)) {
 			pos_css = css_rightmost_descendant(pos_css);
 			continue;
@@ -1205,11 +1246,44 @@ static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp)
 		 * update_tasks_cpumask() again for tasks in the parent
 		 * cpuset if the parent's subparts_cpus changes.
 		 */
-		if ((cp != cs) && cp->partition_root_state &&
-		    update_parent_subparts_cpumask(cp, partcmd_update,
-						   NULL, tmp)) {
-			if (parent != &top_cpuset)
-				update_tasks_cpumask(parent);
+		if ((cp != cs) && cp->partition_root_state) {
+			switch (parent->partition_root_state) {
+			case PRS_DISABLED:
+				/*
+				 * If parent is not a partition root or an
+				 * invalid partition root, clear the state
+				 * state and the CS_CPU_EXCLUSIVE flag.
+				 */
+				WARN_ON_ONCE(cp->partition_root_state
+					     != PRS_ERROR);
+				cp->partition_root_state = 0;
+
+				/*
+				 * clear_bit() is an atomic operation and
+				 * readers aren't interested in the state
+				 * of CS_CPU_EXCLUSIVE anyway. So we can
+				 * just update the flag without holding
+				 * the callback_lock.
+				 */
+				clear_bit(CS_CPU_EXCLUSIVE, &cp->flags);
+				break;
+
+			case PRS_ENABLED:
+				if (update_parent_subparts_cpumask(cp, partcmd_update, NULL, tmp))
+					update_tasks_cpumask(parent);
+				break;
+
+			case PRS_ERROR:
+				/*
+				 * When parent is invalid, it has to be too.
+				 */
+				cp->partition_root_state = PRS_ERROR;
+				if (cp->nr_subparts_cpus) {
+					cp->nr_subparts_cpus = 0;
+					cpumask_clear(cp->subparts_cpus);
+				}
+				break;
+			}
 		}
 
 		if (!css_tryget_online(&cp->css))
@@ -1219,13 +1293,33 @@ static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp)
 		spin_lock_irq(&callback_lock);
 
 		cpumask_copy(cp->effective_cpus, tmp->new_cpus);
-		if (cp->nr_subparts_cpus) {
+		if (cp->nr_subparts_cpus &&
+		   (cp->partition_root_state != PRS_ENABLED)) {
+			cp->nr_subparts_cpus = 0;
+			cpumask_clear(cp->subparts_cpus);
+		} else if (cp->nr_subparts_cpus) {
 			/*
 			 * Make sure that effective_cpus & subparts_cpus
 			 * are mutually exclusive.
+			 *
+			 * In the unlikely event that effective_cpus
+			 * becomes empty. we clear cp->nr_subparts_cpus and
+			 * let its child partition roots to compete for
+			 * CPUs again.
 			 */
 			cpumask_andnot(cp->effective_cpus, cp->effective_cpus,
 				       cp->subparts_cpus);
+			if (cpumask_empty(cp->effective_cpus)) {
+				cpumask_copy(cp->effective_cpus, tmp->new_cpus);
+				cpumask_clear(cp->subparts_cpus);
+				cp->nr_subparts_cpus = 0;
+			} else if (!cpumask_subset(cp->subparts_cpus,
+						   tmp->new_cpus)) {
+				cpumask_andnot(cp->subparts_cpus,
+					cp->subparts_cpus, tmp->new_cpus);
+				cp->nr_subparts_cpus
+					= cpumask_weight(cp->subparts_cpus);
+			}
 		}
 		spin_unlock_irq(&callback_lock);
 
@@ -1702,7 +1796,7 @@ static int update_prstate(struct cpuset *cs, int val)
 		return 0;
 
 	/*
-	 * Cannot force a partial or erroneous partition root to a full
+	 * Cannot force a partial or invalid partition root to a full
 	 * partition root.
 	 */
 	if (val && cs->partition_root_state)
@@ -1733,6 +1827,17 @@ static int update_prstate(struct cpuset *cs, int val)
 		}
 		cs->partition_root_state = PRS_ENABLED;
 	} else {
+		/*
+		 * Turning off partition root will clear the
+		 * CS_CPU_EXCLUSIVE bit.
+		 */
+		if (cs->partition_root_state == PRS_ERROR) {
+			cs->partition_root_state = 0;
+			update_flag(CS_CPU_EXCLUSIVE, cs, 0);
+			err = 0;
+			goto out;
+		}
+
 		err = update_parent_subparts_cpumask(cs, partcmd_disable,
 						     NULL, &tmp);
 		if (err)