2 * arch/arm/common/bL_switcher.c -- big.LITTLE cluster switcher core driver
4 * Created by: Nicolas Pitre, March 2012
5 * Copyright: (C) 2012-2013 Linaro Limited
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License version 2 as
9 * published by the Free Software Foundation.
12 #include <linux/atomic.h>
13 #include <linux/init.h>
14 #include <linux/kernel.h>
15 #include <linux/module.h>
16 #include <linux/sched.h>
17 #include <linux/interrupt.h>
18 #include <linux/cpu_pm.h>
19 #include <linux/cpu.h>
20 #include <linux/cpumask.h>
21 #include <linux/kthread.h>
22 #include <linux/wait.h>
23 #include <linux/clockchips.h>
24 #include <linux/hrtimer.h>
25 #include <linux/tick.h>
26 #include <linux/notifier.h>
28 #include <linux/mutex.h>
29 #include <linux/spinlock.h>
30 #include <linux/string.h>
31 #include <linux/sysfs.h>
32 #include <linux/irqchip/arm-gic.h>
33 #include <linux/moduleparam.h>
35 #include <asm/smp_plat.h>
36 #include <asm/suspend.h>
38 #include <asm/bL_switcher.h>
42 * Use our own MPIDR accessors as the generic ones in asm/cputype.h have
43 * __attribute_const__ and we don't want the compiler to assume any
44 * constness here as the value _does_ change along some code paths.
47 static int read_mpidr(void)
50 asm volatile ("mrc p15, 0, %0, c0, c0, 5" : "=r" (id));
51 return id & MPIDR_HWID_BITMASK;
55 * bL switcher core code.
58 static void bL_do_switch(void *_arg)
60 unsigned ib_mpidr, ib_cpu, ib_cluster;
61 long volatile handshake, **handshake_ptr = _arg;
63 pr_debug("%s\n", __func__);
65 ib_mpidr = cpu_logical_map(smp_processor_id());
66 ib_cpu = MPIDR_AFFINITY_LEVEL(ib_mpidr, 0);
67 ib_cluster = MPIDR_AFFINITY_LEVEL(ib_mpidr, 1);
69 /* Advertise our handshake location */
72 *handshake_ptr = &handshake;
77 * Our state has been saved at this point. Let's release our
80 mcpm_set_entry_vector(ib_cpu, ib_cluster, cpu_resume);
84 * From this point, we must assume that our counterpart CPU might
85 * have taken over in its parallel world already, as if execution
86 * just returned from cpu_suspend(). It is therefore important to
87 * be very careful not to make any change the other guy is not
88 * expecting. This is why we need stack isolation.
90 * Fancy under cover tasks could be performed here. For now
95 * Let's wait until our inbound is alive.
102 /* Let's put ourself down. */
103 mcpm_cpu_power_down();
105 /* should never get here */
110 * Stack isolation. To ensure 'current' remains valid, we just use another
111 * piece of our thread's stack space which should be fairly lightly used.
112 * The selected area starts just above the thread_info structure located
113 * at the very bottom of the stack, aligned to a cache line, and indexed
114 * with the cluster number.
116 #define STACK_SIZE 512
117 extern void call_with_stack(void (*fn)(void *), void *arg, void *sp);
118 static int bL_switchpoint(unsigned long _arg)
120 unsigned int mpidr = read_mpidr();
121 unsigned int clusterid = MPIDR_AFFINITY_LEVEL(mpidr, 1);
122 void *stack = current_thread_info() + 1;
123 stack = PTR_ALIGN(stack, L1_CACHE_BYTES);
124 stack += clusterid * STACK_SIZE + STACK_SIZE;
125 call_with_stack(bL_do_switch, (void *)_arg, stack);
130 * Generic switcher interface
133 static unsigned int bL_gic_id[MAX_CPUS_PER_CLUSTER][MAX_NR_CLUSTERS];
134 static int bL_switcher_cpu_pairing[NR_CPUS];
137 * bL_switch_to - Switch to a specific cluster for the current CPU
138 * @new_cluster_id: the ID of the cluster to switch to.
140 * This function must be called on the CPU to be switched.
141 * Returns 0 on success, else a negative status code.
143 static int bL_switch_to(unsigned int new_cluster_id)
145 unsigned int mpidr, this_cpu, that_cpu;
146 unsigned int ob_mpidr, ob_cpu, ob_cluster, ib_mpidr, ib_cpu, ib_cluster;
147 struct tick_device *tdev;
148 enum clock_event_mode tdev_mode;
149 long volatile *handshake_ptr;
152 this_cpu = smp_processor_id();
153 ob_mpidr = read_mpidr();
154 ob_cpu = MPIDR_AFFINITY_LEVEL(ob_mpidr, 0);
155 ob_cluster = MPIDR_AFFINITY_LEVEL(ob_mpidr, 1);
156 BUG_ON(cpu_logical_map(this_cpu) != ob_mpidr);
158 if (new_cluster_id == ob_cluster)
161 that_cpu = bL_switcher_cpu_pairing[this_cpu];
162 ib_mpidr = cpu_logical_map(that_cpu);
163 ib_cpu = MPIDR_AFFINITY_LEVEL(ib_mpidr, 0);
164 ib_cluster = MPIDR_AFFINITY_LEVEL(ib_mpidr, 1);
166 pr_debug("before switch: CPU %d MPIDR %#x -> %#x\n",
167 this_cpu, ob_mpidr, ib_mpidr);
169 /* Close the gate for our entry vectors */
170 mcpm_set_entry_vector(ob_cpu, ob_cluster, NULL);
171 mcpm_set_entry_vector(ib_cpu, ib_cluster, NULL);
174 * Let's wake up the inbound CPU now in case it requires some delay
175 * to come online, but leave it gated in our entry vector code.
177 ret = mcpm_cpu_power_up(ib_cpu, ib_cluster);
179 pr_err("%s: mcpm_cpu_power_up() returned %d\n", __func__, ret);
184 * From this point we are entering the switch critical zone
185 * and can't take any interrupts anymore.
190 /* redirect GIC's SGIs to our counterpart */
191 gic_migrate_target(bL_gic_id[ib_cpu][ib_cluster]);
194 * Raise a SGI on the inbound CPU to make sure it doesn't stall
195 * in a possible WFI, such as in mcpm_power_down().
197 arch_send_wakeup_ipi_mask(cpumask_of(this_cpu));
199 tdev = tick_get_device(this_cpu);
200 if (tdev && !cpumask_equal(tdev->evtdev->cpumask, cpumask_of(this_cpu)))
203 tdev_mode = tdev->evtdev->mode;
204 clockevents_set_mode(tdev->evtdev, CLOCK_EVT_MODE_SHUTDOWN);
207 ret = cpu_pm_enter();
209 /* we can not tolerate errors at this point */
211 panic("%s: cpu_pm_enter() returned %d\n", __func__, ret);
213 /* Swap the physical CPUs in the logical map for this logical CPU. */
214 cpu_logical_map(this_cpu) = ib_mpidr;
215 cpu_logical_map(that_cpu) = ob_mpidr;
217 /* Let's do the actual CPU switch. */
218 ret = cpu_suspend((unsigned long)&handshake_ptr, bL_switchpoint);
220 panic("%s: cpu_suspend() returned %d\n", __func__, ret);
222 /* We are executing on the inbound CPU at this point */
223 mpidr = read_mpidr();
224 pr_debug("after switch: CPU %d MPIDR %#x\n", this_cpu, mpidr);
225 BUG_ON(mpidr != ib_mpidr);
227 mcpm_cpu_powered_up();
232 clockevents_set_mode(tdev->evtdev, tdev_mode);
233 clockevents_program_event(tdev->evtdev,
234 tdev->evtdev->next_event, 1);
244 pr_err("%s exiting with error %d\n", __func__, ret);
250 struct task_struct *task;
251 wait_queue_head_t wq;
253 struct completion started;
254 bL_switch_completion_handler completer;
255 void *completer_cookie;
258 static struct bL_thread bL_threads[NR_CPUS];
260 static int bL_switcher_thread(void *arg)
262 struct bL_thread *t = arg;
263 struct sched_param param = { .sched_priority = 1 };
265 bL_switch_completion_handler completer;
266 void *completer_cookie;
268 sched_setscheduler_nocheck(current, SCHED_FIFO, ¶m);
269 complete(&t->started);
272 if (signal_pending(current))
273 flush_signals(current);
274 wait_event_interruptible(t->wq,
275 t->wanted_cluster != -1 ||
276 kthread_should_stop());
279 cluster = t->wanted_cluster;
280 completer = t->completer;
281 completer_cookie = t->completer_cookie;
282 t->wanted_cluster = -1;
284 spin_unlock(&t->lock);
287 bL_switch_to(cluster);
290 completer(completer_cookie);
292 } while (!kthread_should_stop());
297 static struct task_struct *bL_switcher_thread_create(int cpu, void *arg)
299 struct task_struct *task;
301 task = kthread_create_on_node(bL_switcher_thread, arg,
302 cpu_to_node(cpu), "kswitcher_%d", cpu);
304 kthread_bind(task, cpu);
305 wake_up_process(task);
307 pr_err("%s failed for CPU %d\n", __func__, cpu);
312 * bL_switch_request_cb - Switch to a specific cluster for the given CPU,
313 * with completion notification via a callback
315 * @cpu: the CPU to switch
316 * @new_cluster_id: the ID of the cluster to switch to.
317 * @completer: switch completion callback. if non-NULL,
318 * @completer(@completer_cookie) will be called on completion of
319 * the switch, in non-atomic context.
320 * @completer_cookie: opaque context argument for @completer.
322 * This function causes a cluster switch on the given CPU by waking up
323 * the appropriate switcher thread. This function may or may not return
324 * before the switch has occurred.
326 * If a @completer callback function is supplied, it will be called when
327 * the switch is complete. This can be used to determine asynchronously
328 * when the switch is complete, regardless of when bL_switch_request()
329 * returns. When @completer is supplied, no new switch request is permitted
330 * for the affected CPU until after the switch is complete, and @completer
333 int bL_switch_request_cb(unsigned int cpu, unsigned int new_cluster_id,
334 bL_switch_completion_handler completer,
335 void *completer_cookie)
339 if (cpu >= ARRAY_SIZE(bL_threads)) {
340 pr_err("%s: cpu %d out of bounds\n", __func__, cpu);
344 t = &bL_threads[cpu];
347 return PTR_ERR(t->task);
353 spin_unlock(&t->lock);
356 t->completer = completer;
357 t->completer_cookie = completer_cookie;
358 t->wanted_cluster = new_cluster_id;
359 spin_unlock(&t->lock);
363 EXPORT_SYMBOL_GPL(bL_switch_request_cb);
366 * Activation and configuration code.
369 static DEFINE_MUTEX(bL_switcher_activation_lock);
370 static BLOCKING_NOTIFIER_HEAD(bL_activation_notifier);
371 static unsigned int bL_switcher_active;
372 static unsigned int bL_switcher_cpu_original_cluster[NR_CPUS];
373 static cpumask_t bL_switcher_removed_logical_cpus;
375 int bL_switcher_register_notifier(struct notifier_block *nb)
377 return blocking_notifier_chain_register(&bL_activation_notifier, nb);
379 EXPORT_SYMBOL_GPL(bL_switcher_register_notifier);
381 int bL_switcher_unregister_notifier(struct notifier_block *nb)
383 return blocking_notifier_chain_unregister(&bL_activation_notifier, nb);
385 EXPORT_SYMBOL_GPL(bL_switcher_unregister_notifier);
387 static int bL_activation_notify(unsigned long val)
391 ret = blocking_notifier_call_chain(&bL_activation_notifier, val, NULL);
392 if (ret & NOTIFY_STOP_MASK)
393 pr_err("%s: notifier chain failed with status 0x%x\n",
395 return notifier_to_errno(ret);
398 static void bL_switcher_restore_cpus(void)
402 for_each_cpu(i, &bL_switcher_removed_logical_cpus)
406 static int bL_switcher_halve_cpus(void)
408 int i, j, cluster_0, gic_id, ret;
409 unsigned int cpu, cluster, mask;
410 cpumask_t available_cpus;
412 /* First pass to validate what we have */
414 for_each_online_cpu(i) {
415 cpu = MPIDR_AFFINITY_LEVEL(cpu_logical_map(i), 0);
416 cluster = MPIDR_AFFINITY_LEVEL(cpu_logical_map(i), 1);
418 pr_err("%s: only dual cluster systems are supported\n", __func__);
421 if (WARN_ON(cpu >= MAX_CPUS_PER_CLUSTER))
423 mask |= (1 << cluster);
426 pr_err("%s: no CPU pairing possible\n", __func__);
431 * Now let's do the pairing. We match each CPU with another CPU
432 * from a different cluster. To get a uniform scheduling behavior
433 * without fiddling with CPU topology and compute capacity data,
434 * we'll use logical CPUs initially belonging to the same cluster.
436 memset(bL_switcher_cpu_pairing, -1, sizeof(bL_switcher_cpu_pairing));
437 cpumask_copy(&available_cpus, cpu_online_mask);
439 for_each_cpu(i, &available_cpus) {
441 cluster = MPIDR_AFFINITY_LEVEL(cpu_logical_map(i), 1);
444 if (cluster != cluster_0)
446 cpumask_clear_cpu(i, &available_cpus);
447 for_each_cpu(j, &available_cpus) {
448 cluster = MPIDR_AFFINITY_LEVEL(cpu_logical_map(j), 1);
450 * Let's remember the last match to create "odd"
451 * pairings on purpose in order for other code not
452 * to assume any relation between physical and
453 * logical CPU numbers.
455 if (cluster != cluster_0)
459 bL_switcher_cpu_pairing[i] = match;
460 cpumask_clear_cpu(match, &available_cpus);
461 pr_info("CPU%d paired with CPU%d\n", i, match);
466 * Now we disable the unwanted CPUs i.e. everything that has no
467 * pairing information (that includes the pairing counterparts).
469 cpumask_clear(&bL_switcher_removed_logical_cpus);
470 for_each_online_cpu(i) {
471 cpu = MPIDR_AFFINITY_LEVEL(cpu_logical_map(i), 0);
472 cluster = MPIDR_AFFINITY_LEVEL(cpu_logical_map(i), 1);
474 /* Let's take note of the GIC ID for this CPU */
475 gic_id = gic_get_cpu_id(i);
477 pr_err("%s: bad GIC ID for CPU %d\n", __func__, i);
478 bL_switcher_restore_cpus();
481 bL_gic_id[cpu][cluster] = gic_id;
482 pr_info("GIC ID for CPU %u cluster %u is %u\n",
483 cpu, cluster, gic_id);
485 if (bL_switcher_cpu_pairing[i] != -1) {
486 bL_switcher_cpu_original_cluster[i] = cluster;
492 bL_switcher_restore_cpus();
495 cpumask_set_cpu(i, &bL_switcher_removed_logical_cpus);
501 static int bL_switcher_enable(void)
505 mutex_lock(&bL_switcher_activation_lock);
506 cpu_hotplug_driver_lock();
507 if (bL_switcher_active) {
508 cpu_hotplug_driver_unlock();
509 mutex_unlock(&bL_switcher_activation_lock);
513 pr_info("big.LITTLE switcher initializing\n");
515 ret = bL_activation_notify(BL_NOTIFY_PRE_ENABLE);
519 ret = bL_switcher_halve_cpus();
523 for_each_online_cpu(cpu) {
524 struct bL_thread *t = &bL_threads[cpu];
525 spin_lock_init(&t->lock);
526 init_waitqueue_head(&t->wq);
527 init_completion(&t->started);
528 t->wanted_cluster = -1;
529 t->task = bL_switcher_thread_create(cpu, t);
532 bL_switcher_active = 1;
533 bL_activation_notify(BL_NOTIFY_POST_ENABLE);
534 pr_info("big.LITTLE switcher initialized\n");
538 pr_warn("big.LITTLE switcher initialization failed\n");
539 bL_activation_notify(BL_NOTIFY_POST_DISABLE);
542 cpu_hotplug_driver_unlock();
543 mutex_unlock(&bL_switcher_activation_lock);
549 static void bL_switcher_disable(void)
551 unsigned int cpu, cluster;
553 struct task_struct *task;
555 mutex_lock(&bL_switcher_activation_lock);
556 cpu_hotplug_driver_lock();
558 if (!bL_switcher_active)
561 if (bL_activation_notify(BL_NOTIFY_PRE_DISABLE) != 0) {
562 bL_activation_notify(BL_NOTIFY_POST_ENABLE);
566 bL_switcher_active = 0;
569 * To deactivate the switcher, we must shut down the switcher
570 * threads to prevent any other requests from being accepted.
571 * Then, if the final cluster for given logical CPU is not the
572 * same as the original one, we'll recreate a switcher thread
573 * just for the purpose of switching the CPU back without any
574 * possibility for interference from external requests.
576 for_each_online_cpu(cpu) {
577 t = &bL_threads[cpu];
580 if (!task || IS_ERR(task))
583 /* no more switch may happen on this CPU at this point */
584 cluster = MPIDR_AFFINITY_LEVEL(cpu_logical_map(cpu), 1);
585 if (cluster == bL_switcher_cpu_original_cluster[cpu])
587 init_completion(&t->started);
588 t->wanted_cluster = bL_switcher_cpu_original_cluster[cpu];
589 task = bL_switcher_thread_create(cpu, t);
591 wait_for_completion(&t->started);
593 cluster = MPIDR_AFFINITY_LEVEL(cpu_logical_map(cpu), 1);
594 if (cluster == bL_switcher_cpu_original_cluster[cpu])
597 /* If execution gets here, we're in trouble. */
598 pr_crit("%s: unable to restore original cluster for CPU %d\n",
600 pr_crit("%s: CPU %d can't be restored\n",
601 __func__, bL_switcher_cpu_pairing[cpu]);
602 cpumask_clear_cpu(bL_switcher_cpu_pairing[cpu],
603 &bL_switcher_removed_logical_cpus);
606 bL_switcher_restore_cpus();
607 bL_activation_notify(BL_NOTIFY_POST_DISABLE);
610 cpu_hotplug_driver_unlock();
611 mutex_unlock(&bL_switcher_activation_lock);
614 static ssize_t bL_switcher_active_show(struct kobject *kobj,
615 struct kobj_attribute *attr, char *buf)
617 return sprintf(buf, "%u\n", bL_switcher_active);
620 static ssize_t bL_switcher_active_store(struct kobject *kobj,
621 struct kobj_attribute *attr, const char *buf, size_t count)
627 bL_switcher_disable();
631 ret = bL_switcher_enable();
637 return (ret >= 0) ? count : ret;
640 static struct kobj_attribute bL_switcher_active_attr =
641 __ATTR(active, 0644, bL_switcher_active_show, bL_switcher_active_store);
643 static struct attribute *bL_switcher_attrs[] = {
644 &bL_switcher_active_attr.attr,
648 static struct attribute_group bL_switcher_attr_group = {
649 .attrs = bL_switcher_attrs,
652 static struct kobject *bL_switcher_kobj;
654 static int __init bL_switcher_sysfs_init(void)
658 bL_switcher_kobj = kobject_create_and_add("bL_switcher", kernel_kobj);
659 if (!bL_switcher_kobj)
661 ret = sysfs_create_group(bL_switcher_kobj, &bL_switcher_attr_group);
663 kobject_put(bL_switcher_kobj);
667 #endif /* CONFIG_SYSFS */
669 bool bL_switcher_get_enabled(void)
671 mutex_lock(&bL_switcher_activation_lock);
673 return bL_switcher_active;
675 EXPORT_SYMBOL_GPL(bL_switcher_get_enabled);
677 void bL_switcher_put_enabled(void)
679 mutex_unlock(&bL_switcher_activation_lock);
681 EXPORT_SYMBOL_GPL(bL_switcher_put_enabled);
684 * Veto any CPU hotplug operation on those CPUs we've removed
685 * while the switcher is active.
686 * We're just not ready to deal with that given the trickery involved.
688 static int bL_switcher_hotplug_callback(struct notifier_block *nfb,
689 unsigned long action, void *hcpu)
691 if (bL_switcher_active) {
692 int pairing = bL_switcher_cpu_pairing[(unsigned long)hcpu];
693 switch (action & 0xf) {
695 case CPU_DOWN_PREPARE:
703 static bool no_bL_switcher;
704 core_param(no_bL_switcher, no_bL_switcher, bool, 0644);
706 static int __init bL_switcher_init(void)
710 if (MAX_NR_CLUSTERS != 2) {
711 pr_err("%s: only dual cluster systems are supported\n", __func__);
715 cpu_notifier(bL_switcher_hotplug_callback, 0);
717 if (!no_bL_switcher) {
718 ret = bL_switcher_enable();
724 ret = bL_switcher_sysfs_init();
726 pr_err("%s: unable to create sysfs entry\n", __func__);
732 late_initcall(bL_switcher_init);