sched: Fix Core-wide rq->lock for uninitialized CPUs

Eugene tripped over the case where rq_lock(), as called in a
for_each_possible_cpu() loop came apart because rq->core hadn't been
setup yet.

This is a somewhat unusual, but valid case.

Rework things such that rq->core is initialized to point at itself. IOW
initialize each CPU as a single threaded Core. CPU online will then join
the new CPU (thread) to an existing Core where needed.

For completeness sake, have CPU offline fully undo the state so as to
not presume the topology will match the next time it comes online.

Fixes: 9edeaea1bc45 ("sched: Core-wide rq->lock")
Reported-by: Eugene Syromiatnikov <[e-mail]>
Signed-off-by: Peter Zijlstra (Intel) <[e-mail]>
Reviewed-by: Josh Don <[e-mail]>
Tested-by: Eugene Syromiatnikov <[e-mail]>
Link: https://lkml.kernel.org/r/[e-mail]

kernel/sched/core.c

···
 static atomic_t sched_core_count;
 static struct cpumask sched_core_mask;
 
+static void sched_core_lock(int cpu, unsigned long *flags)
+{
+	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
+	int t, i = 0;
+
+	local_irq_save(*flags);
+	for_each_cpu(t, smt_mask)
+		raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
+}
+
+static void sched_core_unlock(int cpu, unsigned long *flags)
+{
+	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
+	int t;
+
+	for_each_cpu(t, smt_mask)
+		raw_spin_unlock(&cpu_rq(t)->__lock);
+	local_irq_restore(*flags);
+}
+
 static void __sched_core_flip(bool enabled)
 {
-	int cpu, t, i;
+	unsigned long flags;
+	int cpu, t;
 
 	cpus_read_lock();
 
···
 	for_each_cpu(cpu, &sched_core_mask) {
 		const struct cpumask *smt_mask = cpu_smt_mask(cpu);
 
-		i = 0;
-		local_irq_disable();
-		for_each_cpu(t, smt_mask) {
-			/* supports up to SMT8 */
-			raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
-		}
+		sched_core_lock(cpu, &flags);
 
 		for_each_cpu(t, smt_mask)
 			cpu_rq(t)->core_enabled = enabled;
 
-		for_each_cpu(t, smt_mask)
-			raw_spin_unlock(&cpu_rq(t)->__lock);
-		local_irq_enable();
+		sched_core_unlock(cpu, &flags);
 
 		cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
 	}
···
 	queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
 }
 
-static inline void sched_core_cpu_starting(unsigned int cpu)
+static void sched_core_cpu_starting(unsigned int cpu)
 {
 	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
-	struct rq *rq, *core_rq = NULL;
-	int i;
+	struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
+	unsigned long flags;
+	int t;
 
-	core_rq = cpu_rq(cpu)->core;
+	sched_core_lock(cpu, &flags);
 
-	if (!core_rq) {
-		for_each_cpu(i, smt_mask) {
-			rq = cpu_rq(i);
-			if (rq->core && rq->core == rq)
-				core_rq = rq;
-		}
+	WARN_ON_ONCE(rq->core != rq);
 
-		if (!core_rq)
-			core_rq = cpu_rq(cpu);
+	/* if we're the first, we'll be our own leader */
+	if (cpumask_weight(smt_mask) == 1)
+		goto unlock;
 
-		for_each_cpu(i, smt_mask) {
-			rq = cpu_rq(i);
-
-			WARN_ON_ONCE(rq->core && rq->core != core_rq);
-			rq->core = core_rq;
+	/* find the leader */
+	for_each_cpu(t, smt_mask) {
+		if (t == cpu)
+			continue;
+		rq = cpu_rq(t);
+		if (rq->core == rq) {
+			core_rq = rq;
+			break;
 		}
 	}
+
+	if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
+		goto unlock;
+
+	/* install and validate core_rq */
+	for_each_cpu(t, smt_mask) {
+		rq = cpu_rq(t);
+
+		if (t == cpu)
+			rq->core = core_rq;
+
+		WARN_ON_ONCE(rq->core != core_rq);
+	}
+
+unlock:
+	sched_core_unlock(cpu, &flags);
 }
+
+static void sched_core_cpu_deactivate(unsigned int cpu)
+{
+	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
+	struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
+	unsigned long flags;
+	int t;
+
+	sched_core_lock(cpu, &flags);
+
+	/* if we're the last man standing, nothing to do */
+	if (cpumask_weight(smt_mask) == 1) {
+		WARN_ON_ONCE(rq->core != rq);
+		goto unlock;
+	}
+
+	/* if we're not the leader, nothing to do */
+	if (rq->core != rq)
+		goto unlock;
+
+	/* find a new leader */
+	for_each_cpu(t, smt_mask) {
+		if (t == cpu)
+			continue;
+		core_rq = cpu_rq(t);
+		break;
+	}
+
+	if (WARN_ON_ONCE(!core_rq)) /* impossible */
+		goto unlock;
+
+	/* copy the shared state to the new leader */
+	core_rq->core_task_seq      = rq->core_task_seq;
+	core_rq->core_pick_seq      = rq->core_pick_seq;
+	core_rq->core_cookie        = rq->core_cookie;
+	core_rq->core_forceidle     = rq->core_forceidle;
+	core_rq->core_forceidle_seq = rq->core_forceidle_seq;
+
+	/* install new leader */
+	for_each_cpu(t, smt_mask) {
+		rq = cpu_rq(t);
+		rq->core = core_rq;
+	}
+
+unlock:
+	sched_core_unlock(cpu, &flags);
+}
+
+static inline void sched_core_cpu_dying(unsigned int cpu)
+{
+	struct rq *rq = cpu_rq(cpu);
+
+	if (rq->core != rq)
+		rq->core = rq;
+}
+
 #else /* !CONFIG_SCHED_CORE */
 
 static inline void sched_core_cpu_starting(unsigned int cpu) {}
+static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
+static inline void sched_core_cpu_dying(unsigned int cpu) {}
 
 static struct task_struct *
 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
···
 	 */
 	if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
 		static_branch_dec_cpuslocked(&sched_smt_present);
+
+	sched_core_cpu_deactivate(cpu);
 #endif
 
 	if (!sched_smp_initialized)
···
 	calc_load_migrate(rq);
 	update_max_interval();
 	hrtick_clear(rq);
+	sched_core_cpu_dying(cpu);
 	return 0;
 }
 #endif
···
 		atomic_set(&rq->nr_iowait, 0);
 
 #ifdef CONFIG_SCHED_CORE
-		rq->core = NULL;
+		rq->core = rq;
 		rq->core_pick = NULL;
 		rq->core_enabled = 0;
 		rq->core_tree = RB_ROOT;

kernel/sched/sched.h

···
 	unsigned int		core_sched_seq;
 	struct rb_root		core_tree;
 
-	/* shared state */
+	/* shared state -- careful with sched_core_cpu_deactivate() */
 	unsigned int		core_task_seq;
 	unsigned int		core_pick_seq;
 	unsigned long		core_cookie;