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/time.h>
24 #include <linux/clockchips.h>
25 #include <linux/hrtimer.h>
26 #include <linux/tick.h>
27 #include <linux/notifier.h>
29 #include <linux/mutex.h>
30 #include <linux/smp.h>
31 #include <linux/spinlock.h>
32 #include <linux/string.h>
33 #include <linux/sysfs.h>
34 #include <linux/irqchip/arm-gic.h>
35 #include <linux/moduleparam.h>
37 #include <asm/smp_plat.h>
38 #include <asm/cputype.h>
39 #include <asm/suspend.h>
41 #include <asm/bL_switcher.h>
43 #define CREATE_TRACE_POINTS
44 #include <trace/events/power_cpu_migrate.h>
48 * Use our own MPIDR accessors as the generic ones in asm/cputype.h have
49 * __attribute_const__ and we don't want the compiler to assume any
50 * constness here as the value _does_ change along some code paths.
53 static int read_mpidr(void)
56 asm volatile ("mrc p15, 0, %0, c0, c0, 5" : "=r" (id));
57 return id & MPIDR_HWID_BITMASK;
61 * bL switcher core code.
64 static void bL_do_switch(void *_arg)
66 unsigned ib_mpidr, ib_cpu, ib_cluster;
67 long volatile handshake, **handshake_ptr = _arg;
69 pr_debug("%s\n", __func__);
71 ib_mpidr = cpu_logical_map(smp_processor_id());
72 ib_cpu = MPIDR_AFFINITY_LEVEL(ib_mpidr, 0);
73 ib_cluster = MPIDR_AFFINITY_LEVEL(ib_mpidr, 1);
75 /* Advertise our handshake location */
78 *handshake_ptr = &handshake;
83 * Our state has been saved at this point. Let's release our
86 mcpm_set_entry_vector(ib_cpu, ib_cluster, cpu_resume);
90 * From this point, we must assume that our counterpart CPU might
91 * have taken over in its parallel world already, as if execution
92 * just returned from cpu_suspend(). It is therefore important to
93 * be very careful not to make any change the other guy is not
94 * expecting. This is why we need stack isolation.
96 * Fancy under cover tasks could be performed here. For now
101 * Let's wait until our inbound is alive.
108 /* Let's put ourself down. */
109 mcpm_cpu_power_down();
111 /* should never get here */
116 * Stack isolation. To ensure 'current' remains valid, we just use another
117 * piece of our thread's stack space which should be fairly lightly used.
118 * The selected area starts just above the thread_info structure located
119 * at the very bottom of the stack, aligned to a cache line, and indexed
120 * with the cluster number.
122 #define STACK_SIZE 512
123 extern void call_with_stack(void (*fn)(void *), void *arg, void *sp);
124 static int bL_switchpoint(unsigned long _arg)
126 unsigned int mpidr = read_mpidr();
127 unsigned int clusterid = MPIDR_AFFINITY_LEVEL(mpidr, 1);
128 void *stack = current_thread_info() + 1;
129 stack = PTR_ALIGN(stack, L1_CACHE_BYTES);
130 stack += clusterid * STACK_SIZE + STACK_SIZE;
131 call_with_stack(bL_do_switch, (void *)_arg, stack);
136 * Generic switcher interface
139 static unsigned int bL_gic_id[MAX_CPUS_PER_CLUSTER][MAX_NR_CLUSTERS];
140 static int bL_switcher_cpu_pairing[NR_CPUS];
143 * bL_switch_to - Switch to a specific cluster for the current CPU
144 * @new_cluster_id: the ID of the cluster to switch to.
146 * This function must be called on the CPU to be switched.
147 * Returns 0 on success, else a negative status code.
149 static int bL_switch_to(unsigned int new_cluster_id)
151 unsigned int mpidr, this_cpu, that_cpu;
152 unsigned int ob_mpidr, ob_cpu, ob_cluster, ib_mpidr, ib_cpu, ib_cluster;
153 struct completion inbound_alive;
154 struct tick_device *tdev;
155 enum clock_event_mode tdev_mode;
156 long volatile *handshake_ptr;
159 this_cpu = smp_processor_id();
160 ob_mpidr = read_mpidr();
161 ob_cpu = MPIDR_AFFINITY_LEVEL(ob_mpidr, 0);
162 ob_cluster = MPIDR_AFFINITY_LEVEL(ob_mpidr, 1);
163 BUG_ON(cpu_logical_map(this_cpu) != ob_mpidr);
165 if (new_cluster_id == ob_cluster)
168 that_cpu = bL_switcher_cpu_pairing[this_cpu];
169 ib_mpidr = cpu_logical_map(that_cpu);
170 ib_cpu = MPIDR_AFFINITY_LEVEL(ib_mpidr, 0);
171 ib_cluster = MPIDR_AFFINITY_LEVEL(ib_mpidr, 1);
173 pr_debug("before switch: CPU %d MPIDR %#x -> %#x\n",
174 this_cpu, ob_mpidr, ib_mpidr);
176 this_cpu = smp_processor_id();
178 /* Close the gate for our entry vectors */
179 mcpm_set_entry_vector(ob_cpu, ob_cluster, NULL);
180 mcpm_set_entry_vector(ib_cpu, ib_cluster, NULL);
182 /* Install our "inbound alive" notifier. */
183 init_completion(&inbound_alive);
184 ipi_nr = register_ipi_completion(&inbound_alive, this_cpu);
185 ipi_nr |= ((1 << 16) << bL_gic_id[ob_cpu][ob_cluster]);
186 mcpm_set_early_poke(ib_cpu, ib_cluster, gic_get_sgir_physaddr(), ipi_nr);
189 * Let's wake up the inbound CPU now in case it requires some delay
190 * to come online, but leave it gated in our entry vector code.
192 ret = mcpm_cpu_power_up(ib_cpu, ib_cluster);
194 pr_err("%s: mcpm_cpu_power_up() returned %d\n", __func__, ret);
199 * Raise a SGI on the inbound CPU to make sure it doesn't stall
200 * in a possible WFI, such as in bL_power_down().
202 gic_send_sgi(bL_gic_id[ib_cpu][ib_cluster], 0);
205 * Wait for the inbound to come up. This allows for other
206 * tasks to be scheduled in the mean time.
208 wait_for_completion(&inbound_alive);
209 mcpm_set_early_poke(ib_cpu, ib_cluster, 0, 0);
212 * From this point we are entering the switch critical zone
213 * and can't take any interrupts anymore.
217 trace_cpu_migrate_begin(ktime_get_real_ns(), ob_mpidr);
219 /* redirect GIC's SGIs to our counterpart */
220 gic_migrate_target(bL_gic_id[ib_cpu][ib_cluster]);
222 tdev = tick_get_device(this_cpu);
223 if (tdev && !cpumask_equal(tdev->evtdev->cpumask, cpumask_of(this_cpu)))
226 tdev_mode = tdev->evtdev->mode;
227 clockevents_set_mode(tdev->evtdev, CLOCK_EVT_MODE_SHUTDOWN);
230 ret = cpu_pm_enter();
232 /* we can not tolerate errors at this point */
234 panic("%s: cpu_pm_enter() returned %d\n", __func__, ret);
236 /* Swap the physical CPUs in the logical map for this logical CPU. */
237 cpu_logical_map(this_cpu) = ib_mpidr;
238 cpu_logical_map(that_cpu) = ob_mpidr;
240 /* Let's do the actual CPU switch. */
241 ret = cpu_suspend((unsigned long)&handshake_ptr, bL_switchpoint);
243 panic("%s: cpu_suspend() returned %d\n", __func__, ret);
245 /* We are executing on the inbound CPU at this point */
246 mpidr = read_mpidr();
247 pr_debug("after switch: CPU %d MPIDR %#x\n", this_cpu, mpidr);
248 BUG_ON(mpidr != ib_mpidr);
250 mcpm_cpu_powered_up();
255 clockevents_set_mode(tdev->evtdev, tdev_mode);
256 clockevents_program_event(tdev->evtdev,
257 tdev->evtdev->next_event, 1);
260 trace_cpu_migrate_finish(ktime_get_real_ns(), ib_mpidr);
268 pr_err("%s exiting with error %d\n", __func__, ret);
274 struct task_struct *task;
275 wait_queue_head_t wq;
277 struct completion started;
278 bL_switch_completion_handler completer;
279 void *completer_cookie;
282 static struct bL_thread bL_threads[NR_CPUS];
284 static int bL_switcher_thread(void *arg)
286 struct bL_thread *t = arg;
287 struct sched_param param = { .sched_priority = 1 };
289 bL_switch_completion_handler completer;
290 void *completer_cookie;
292 sched_setscheduler_nocheck(current, SCHED_FIFO, ¶m);
293 complete(&t->started);
296 if (signal_pending(current))
297 flush_signals(current);
298 wait_event_interruptible(t->wq,
299 t->wanted_cluster != -1 ||
300 kthread_should_stop());
303 cluster = t->wanted_cluster;
304 completer = t->completer;
305 completer_cookie = t->completer_cookie;
306 t->wanted_cluster = -1;
308 spin_unlock(&t->lock);
311 bL_switch_to(cluster);
314 completer(completer_cookie);
316 } while (!kthread_should_stop());
321 static struct task_struct *bL_switcher_thread_create(int cpu, void *arg)
323 struct task_struct *task;
325 task = kthread_create_on_node(bL_switcher_thread, arg,
326 cpu_to_node(cpu), "kswitcher_%d", cpu);
328 kthread_bind(task, cpu);
329 wake_up_process(task);
331 pr_err("%s failed for CPU %d\n", __func__, cpu);
336 * bL_switch_request_cb - Switch to a specific cluster for the given CPU,
337 * with completion notification via a callback
339 * @cpu: the CPU to switch
340 * @new_cluster_id: the ID of the cluster to switch to.
341 * @completer: switch completion callback. if non-NULL,
342 * @completer(@completer_cookie) will be called on completion of
343 * the switch, in non-atomic context.
344 * @completer_cookie: opaque context argument for @completer.
346 * This function causes a cluster switch on the given CPU by waking up
347 * the appropriate switcher thread. This function may or may not return
348 * before the switch has occurred.
350 * If a @completer callback function is supplied, it will be called when
351 * the switch is complete. This can be used to determine asynchronously
352 * when the switch is complete, regardless of when bL_switch_request()
353 * returns. When @completer is supplied, no new switch request is permitted
354 * for the affected CPU until after the switch is complete, and @completer
357 int bL_switch_request_cb(unsigned int cpu, unsigned int new_cluster_id,
358 bL_switch_completion_handler completer,
359 void *completer_cookie)
363 if (cpu >= ARRAY_SIZE(bL_threads)) {
364 pr_err("%s: cpu %d out of bounds\n", __func__, cpu);
368 t = &bL_threads[cpu];
371 return PTR_ERR(t->task);
377 spin_unlock(&t->lock);
380 t->completer = completer;
381 t->completer_cookie = completer_cookie;
382 t->wanted_cluster = new_cluster_id;
383 spin_unlock(&t->lock);
387 EXPORT_SYMBOL_GPL(bL_switch_request_cb);
390 * Activation and configuration code.
393 static DEFINE_MUTEX(bL_switcher_activation_lock);
394 static BLOCKING_NOTIFIER_HEAD(bL_activation_notifier);
395 static unsigned int bL_switcher_active;
396 static unsigned int bL_switcher_cpu_original_cluster[NR_CPUS];
397 static cpumask_t bL_switcher_removed_logical_cpus;
399 int bL_switcher_register_notifier(struct notifier_block *nb)
401 return blocking_notifier_chain_register(&bL_activation_notifier, nb);
403 EXPORT_SYMBOL_GPL(bL_switcher_register_notifier);
405 int bL_switcher_unregister_notifier(struct notifier_block *nb)
407 return blocking_notifier_chain_unregister(&bL_activation_notifier, nb);
409 EXPORT_SYMBOL_GPL(bL_switcher_unregister_notifier);
411 static int bL_activation_notify(unsigned long val)
415 ret = blocking_notifier_call_chain(&bL_activation_notifier, val, NULL);
416 if (ret & NOTIFY_STOP_MASK)
417 pr_err("%s: notifier chain failed with status 0x%x\n",
419 return notifier_to_errno(ret);
422 static void bL_switcher_restore_cpus(void)
426 for_each_cpu(i, &bL_switcher_removed_logical_cpus) {
427 struct device *cpu_dev = get_cpu_device(i);
428 int ret = device_online(cpu_dev);
430 dev_err(cpu_dev, "switcher: unable to restore CPU\n");
434 static int bL_switcher_halve_cpus(void)
436 int i, j, cluster_0, gic_id, ret;
437 unsigned int cpu, cluster, mask;
438 cpumask_t available_cpus;
440 /* First pass to validate what we have */
442 for_each_online_cpu(i) {
443 cpu = MPIDR_AFFINITY_LEVEL(cpu_logical_map(i), 0);
444 cluster = MPIDR_AFFINITY_LEVEL(cpu_logical_map(i), 1);
446 pr_err("%s: only dual cluster systems are supported\n", __func__);
449 if (WARN_ON(cpu >= MAX_CPUS_PER_CLUSTER))
451 mask |= (1 << cluster);
454 pr_err("%s: no CPU pairing possible\n", __func__);
459 * Now let's do the pairing. We match each CPU with another CPU
460 * from a different cluster. To get a uniform scheduling behavior
461 * without fiddling with CPU topology and compute capacity data,
462 * we'll use logical CPUs initially belonging to the same cluster.
464 memset(bL_switcher_cpu_pairing, -1, sizeof(bL_switcher_cpu_pairing));
465 cpumask_copy(&available_cpus, cpu_online_mask);
467 for_each_cpu(i, &available_cpus) {
469 cluster = MPIDR_AFFINITY_LEVEL(cpu_logical_map(i), 1);
472 if (cluster != cluster_0)
474 cpumask_clear_cpu(i, &available_cpus);
475 for_each_cpu(j, &available_cpus) {
476 cluster = MPIDR_AFFINITY_LEVEL(cpu_logical_map(j), 1);
478 * Let's remember the last match to create "odd"
479 * pairings on purpose in order for other code not
480 * to assume any relation between physical and
481 * logical CPU numbers.
483 if (cluster != cluster_0)
487 bL_switcher_cpu_pairing[i] = match;
488 cpumask_clear_cpu(match, &available_cpus);
489 pr_info("CPU%d paired with CPU%d\n", i, match);
494 * Now we disable the unwanted CPUs i.e. everything that has no
495 * pairing information (that includes the pairing counterparts).
497 cpumask_clear(&bL_switcher_removed_logical_cpus);
498 for_each_online_cpu(i) {
499 cpu = MPIDR_AFFINITY_LEVEL(cpu_logical_map(i), 0);
500 cluster = MPIDR_AFFINITY_LEVEL(cpu_logical_map(i), 1);
502 /* Let's take note of the GIC ID for this CPU */
503 gic_id = gic_get_cpu_id(i);
505 pr_err("%s: bad GIC ID for CPU %d\n", __func__, i);
506 bL_switcher_restore_cpus();
509 bL_gic_id[cpu][cluster] = gic_id;
510 pr_info("GIC ID for CPU %u cluster %u is %u\n",
511 cpu, cluster, gic_id);
513 if (bL_switcher_cpu_pairing[i] != -1) {
514 bL_switcher_cpu_original_cluster[i] = cluster;
518 ret = device_offline(get_cpu_device(i));
520 bL_switcher_restore_cpus();
523 cpumask_set_cpu(i, &bL_switcher_removed_logical_cpus);
529 /* Determine the logical CPU a given physical CPU is grouped on. */
530 int bL_switcher_get_logical_index(u32 mpidr)
534 if (!bL_switcher_active)
537 mpidr &= MPIDR_HWID_BITMASK;
538 for_each_online_cpu(cpu) {
539 int pairing = bL_switcher_cpu_pairing[cpu];
542 if ((mpidr == cpu_logical_map(cpu)) ||
543 (mpidr == cpu_logical_map(pairing)))
549 static void bL_switcher_trace_trigger_cpu(void *__always_unused info)
551 trace_cpu_migrate_current(ktime_get_real_ns(), read_mpidr());
554 int bL_switcher_trace_trigger(void)
560 bL_switcher_trace_trigger_cpu(NULL);
561 ret = smp_call_function(bL_switcher_trace_trigger_cpu, NULL, true);
567 EXPORT_SYMBOL_GPL(bL_switcher_trace_trigger);
569 static int bL_switcher_enable(void)
573 mutex_lock(&bL_switcher_activation_lock);
574 lock_device_hotplug();
575 if (bL_switcher_active) {
576 unlock_device_hotplug();
577 mutex_unlock(&bL_switcher_activation_lock);
581 pr_info("big.LITTLE switcher initializing\n");
583 ret = bL_activation_notify(BL_NOTIFY_PRE_ENABLE);
587 ret = bL_switcher_halve_cpus();
591 bL_switcher_trace_trigger();
593 for_each_online_cpu(cpu) {
594 struct bL_thread *t = &bL_threads[cpu];
595 spin_lock_init(&t->lock);
596 init_waitqueue_head(&t->wq);
597 init_completion(&t->started);
598 t->wanted_cluster = -1;
599 t->task = bL_switcher_thread_create(cpu, t);
602 bL_switcher_active = 1;
603 bL_activation_notify(BL_NOTIFY_POST_ENABLE);
604 pr_info("big.LITTLE switcher initialized\n");
608 pr_warn("big.LITTLE switcher initialization failed\n");
609 bL_activation_notify(BL_NOTIFY_POST_DISABLE);
612 unlock_device_hotplug();
613 mutex_unlock(&bL_switcher_activation_lock);
619 static void bL_switcher_disable(void)
621 unsigned int cpu, cluster;
623 struct task_struct *task;
625 mutex_lock(&bL_switcher_activation_lock);
626 lock_device_hotplug();
628 if (!bL_switcher_active)
631 if (bL_activation_notify(BL_NOTIFY_PRE_DISABLE) != 0) {
632 bL_activation_notify(BL_NOTIFY_POST_ENABLE);
636 bL_switcher_active = 0;
639 * To deactivate the switcher, we must shut down the switcher
640 * threads to prevent any other requests from being accepted.
641 * Then, if the final cluster for given logical CPU is not the
642 * same as the original one, we'll recreate a switcher thread
643 * just for the purpose of switching the CPU back without any
644 * possibility for interference from external requests.
646 for_each_online_cpu(cpu) {
647 t = &bL_threads[cpu];
650 if (!task || IS_ERR(task))
653 /* no more switch may happen on this CPU at this point */
654 cluster = MPIDR_AFFINITY_LEVEL(cpu_logical_map(cpu), 1);
655 if (cluster == bL_switcher_cpu_original_cluster[cpu])
657 init_completion(&t->started);
658 t->wanted_cluster = bL_switcher_cpu_original_cluster[cpu];
659 task = bL_switcher_thread_create(cpu, t);
661 wait_for_completion(&t->started);
663 cluster = MPIDR_AFFINITY_LEVEL(cpu_logical_map(cpu), 1);
664 if (cluster == bL_switcher_cpu_original_cluster[cpu])
667 /* If execution gets here, we're in trouble. */
668 pr_crit("%s: unable to restore original cluster for CPU %d\n",
670 pr_crit("%s: CPU %d can't be restored\n",
671 __func__, bL_switcher_cpu_pairing[cpu]);
672 cpumask_clear_cpu(bL_switcher_cpu_pairing[cpu],
673 &bL_switcher_removed_logical_cpus);
676 bL_switcher_restore_cpus();
677 bL_switcher_trace_trigger();
679 bL_activation_notify(BL_NOTIFY_POST_DISABLE);
682 unlock_device_hotplug();
683 mutex_unlock(&bL_switcher_activation_lock);
686 static ssize_t bL_switcher_active_show(struct kobject *kobj,
687 struct kobj_attribute *attr, char *buf)
689 return sprintf(buf, "%u\n", bL_switcher_active);
692 static ssize_t bL_switcher_active_store(struct kobject *kobj,
693 struct kobj_attribute *attr, const char *buf, size_t count)
699 bL_switcher_disable();
703 ret = bL_switcher_enable();
709 return (ret >= 0) ? count : ret;
712 static ssize_t bL_switcher_trace_trigger_store(struct kobject *kobj,
713 struct kobj_attribute *attr, const char *buf, size_t count)
715 int ret = bL_switcher_trace_trigger();
717 return ret ? ret : count;
720 static struct kobj_attribute bL_switcher_active_attr =
721 __ATTR(active, 0644, bL_switcher_active_show, bL_switcher_active_store);
723 static struct kobj_attribute bL_switcher_trace_trigger_attr =
724 __ATTR(trace_trigger, 0200, NULL, bL_switcher_trace_trigger_store);
726 static struct attribute *bL_switcher_attrs[] = {
727 &bL_switcher_active_attr.attr,
728 &bL_switcher_trace_trigger_attr.attr,
732 static struct attribute_group bL_switcher_attr_group = {
733 .attrs = bL_switcher_attrs,
736 static struct kobject *bL_switcher_kobj;
738 static int __init bL_switcher_sysfs_init(void)
742 bL_switcher_kobj = kobject_create_and_add("bL_switcher", kernel_kobj);
743 if (!bL_switcher_kobj)
745 ret = sysfs_create_group(bL_switcher_kobj, &bL_switcher_attr_group);
747 kobject_put(bL_switcher_kobj);
751 #endif /* CONFIG_SYSFS */
753 bool bL_switcher_get_enabled(void)
755 mutex_lock(&bL_switcher_activation_lock);
757 return bL_switcher_active;
759 EXPORT_SYMBOL_GPL(bL_switcher_get_enabled);
761 void bL_switcher_put_enabled(void)
763 mutex_unlock(&bL_switcher_activation_lock);
765 EXPORT_SYMBOL_GPL(bL_switcher_put_enabled);
768 * Veto any CPU hotplug operation on those CPUs we've removed
769 * while the switcher is active.
770 * We're just not ready to deal with that given the trickery involved.
772 static int bL_switcher_hotplug_callback(struct notifier_block *nfb,
773 unsigned long action, void *hcpu)
775 if (bL_switcher_active) {
776 int pairing = bL_switcher_cpu_pairing[(unsigned long)hcpu];
777 switch (action & 0xf) {
779 case CPU_DOWN_PREPARE:
787 static bool no_bL_switcher;
788 core_param(no_bL_switcher, no_bL_switcher, bool, 0644);
790 static int __init bL_switcher_init(void)
794 if (!mcpm_is_available())
797 cpu_notifier(bL_switcher_hotplug_callback, 0);
799 if (!no_bL_switcher) {
800 ret = bL_switcher_enable();
806 ret = bL_switcher_sysfs_init();
808 pr_err("%s: unable to create sysfs entry\n", __func__);
814 late_initcall(bL_switcher_init);