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 * Get a global nanosecond time stamp for tracing.
63 static s64 get_ns(void)
67 return timespec_to_ns(&ts);
71 * bL switcher core code.
74 static void bL_do_switch(void *_arg)
76 unsigned ib_mpidr, ib_cpu, ib_cluster;
77 long volatile handshake, **handshake_ptr = _arg;
79 pr_debug("%s\n", __func__);
81 ib_mpidr = cpu_logical_map(smp_processor_id());
82 ib_cpu = MPIDR_AFFINITY_LEVEL(ib_mpidr, 0);
83 ib_cluster = MPIDR_AFFINITY_LEVEL(ib_mpidr, 1);
85 /* Advertise our handshake location */
88 *handshake_ptr = &handshake;
93 * Our state has been saved at this point. Let's release our
96 mcpm_set_entry_vector(ib_cpu, ib_cluster, cpu_resume);
100 * From this point, we must assume that our counterpart CPU might
101 * have taken over in its parallel world already, as if execution
102 * just returned from cpu_suspend(). It is therefore important to
103 * be very careful not to make any change the other guy is not
104 * expecting. This is why we need stack isolation.
106 * Fancy under cover tasks could be performed here. For now
111 * Let's wait until our inbound is alive.
118 /* Let's put ourself down. */
119 mcpm_cpu_power_down();
121 /* should never get here */
126 * Stack isolation. To ensure 'current' remains valid, we just use another
127 * piece of our thread's stack space which should be fairly lightly used.
128 * The selected area starts just above the thread_info structure located
129 * at the very bottom of the stack, aligned to a cache line, and indexed
130 * with the cluster number.
132 #define STACK_SIZE 512
133 extern void call_with_stack(void (*fn)(void *), void *arg, void *sp);
134 static int bL_switchpoint(unsigned long _arg)
136 unsigned int mpidr = read_mpidr();
137 unsigned int clusterid = MPIDR_AFFINITY_LEVEL(mpidr, 1);
138 void *stack = current_thread_info() + 1;
139 stack = PTR_ALIGN(stack, L1_CACHE_BYTES);
140 stack += clusterid * STACK_SIZE + STACK_SIZE;
141 call_with_stack(bL_do_switch, (void *)_arg, stack);
146 * Generic switcher interface
149 static unsigned int bL_gic_id[MAX_CPUS_PER_CLUSTER][MAX_NR_CLUSTERS];
150 static int bL_switcher_cpu_pairing[NR_CPUS];
153 * bL_switch_to - Switch to a specific cluster for the current CPU
154 * @new_cluster_id: the ID of the cluster to switch to.
156 * This function must be called on the CPU to be switched.
157 * Returns 0 on success, else a negative status code.
159 static int bL_switch_to(unsigned int new_cluster_id)
161 unsigned int mpidr, this_cpu, that_cpu;
162 unsigned int ob_mpidr, ob_cpu, ob_cluster, ib_mpidr, ib_cpu, ib_cluster;
163 struct completion inbound_alive;
164 struct tick_device *tdev;
165 enum clock_event_mode tdev_mode;
166 long volatile *handshake_ptr;
169 this_cpu = smp_processor_id();
170 ob_mpidr = read_mpidr();
171 ob_cpu = MPIDR_AFFINITY_LEVEL(ob_mpidr, 0);
172 ob_cluster = MPIDR_AFFINITY_LEVEL(ob_mpidr, 1);
173 BUG_ON(cpu_logical_map(this_cpu) != ob_mpidr);
175 if (new_cluster_id == ob_cluster)
178 that_cpu = bL_switcher_cpu_pairing[this_cpu];
179 ib_mpidr = cpu_logical_map(that_cpu);
180 ib_cpu = MPIDR_AFFINITY_LEVEL(ib_mpidr, 0);
181 ib_cluster = MPIDR_AFFINITY_LEVEL(ib_mpidr, 1);
183 pr_debug("before switch: CPU %d MPIDR %#x -> %#x\n",
184 this_cpu, ob_mpidr, ib_mpidr);
186 this_cpu = smp_processor_id();
188 /* Close the gate for our entry vectors */
189 mcpm_set_entry_vector(ob_cpu, ob_cluster, NULL);
190 mcpm_set_entry_vector(ib_cpu, ib_cluster, NULL);
192 /* Install our "inbound alive" notifier. */
193 init_completion(&inbound_alive);
194 ipi_nr = register_ipi_completion(&inbound_alive, this_cpu);
195 ipi_nr |= ((1 << 16) << bL_gic_id[ob_cpu][ob_cluster]);
196 mcpm_set_early_poke(ib_cpu, ib_cluster, gic_get_sgir_physaddr(), ipi_nr);
199 * Let's wake up the inbound CPU now in case it requires some delay
200 * to come online, but leave it gated in our entry vector code.
202 ret = mcpm_cpu_power_up(ib_cpu, ib_cluster);
204 pr_err("%s: mcpm_cpu_power_up() returned %d\n", __func__, ret);
209 * Raise a SGI on the inbound CPU to make sure it doesn't stall
210 * in a possible WFI, such as in bL_power_down().
212 gic_send_sgi(bL_gic_id[ib_cpu][ib_cluster], 0);
215 * Wait for the inbound to come up. This allows for other
216 * tasks to be scheduled in the mean time.
218 wait_for_completion(&inbound_alive);
219 mcpm_set_early_poke(ib_cpu, ib_cluster, 0, 0);
222 * From this point we are entering the switch critical zone
223 * and can't take any interrupts anymore.
227 trace_cpu_migrate_begin(get_ns(), ob_mpidr);
229 /* redirect GIC's SGIs to our counterpart */
230 gic_migrate_target(bL_gic_id[ib_cpu][ib_cluster]);
232 tdev = tick_get_device(this_cpu);
233 if (tdev && !cpumask_equal(tdev->evtdev->cpumask, cpumask_of(this_cpu)))
236 tdev_mode = tdev->evtdev->mode;
237 clockevents_set_mode(tdev->evtdev, CLOCK_EVT_MODE_SHUTDOWN);
240 ret = cpu_pm_enter();
242 /* we can not tolerate errors at this point */
244 panic("%s: cpu_pm_enter() returned %d\n", __func__, ret);
246 /* Swap the physical CPUs in the logical map for this logical CPU. */
247 cpu_logical_map(this_cpu) = ib_mpidr;
248 cpu_logical_map(that_cpu) = ob_mpidr;
250 /* Let's do the actual CPU switch. */
251 ret = cpu_suspend((unsigned long)&handshake_ptr, bL_switchpoint);
253 panic("%s: cpu_suspend() returned %d\n", __func__, ret);
255 /* We are executing on the inbound CPU at this point */
256 mpidr = read_mpidr();
257 pr_debug("after switch: CPU %d MPIDR %#x\n", this_cpu, mpidr);
258 BUG_ON(mpidr != ib_mpidr);
260 mcpm_cpu_powered_up();
265 clockevents_set_mode(tdev->evtdev, tdev_mode);
266 clockevents_program_event(tdev->evtdev,
267 tdev->evtdev->next_event, 1);
270 trace_cpu_migrate_finish(get_ns(), ib_mpidr);
278 pr_err("%s exiting with error %d\n", __func__, ret);
284 struct task_struct *task;
285 wait_queue_head_t wq;
287 struct completion started;
288 bL_switch_completion_handler completer;
289 void *completer_cookie;
292 static struct bL_thread bL_threads[NR_CPUS];
294 static int bL_switcher_thread(void *arg)
296 struct bL_thread *t = arg;
297 struct sched_param param = { .sched_priority = 1 };
299 bL_switch_completion_handler completer;
300 void *completer_cookie;
302 sched_setscheduler_nocheck(current, SCHED_FIFO, ¶m);
303 complete(&t->started);
306 if (signal_pending(current))
307 flush_signals(current);
308 wait_event_interruptible(t->wq,
309 t->wanted_cluster != -1 ||
310 kthread_should_stop());
313 cluster = t->wanted_cluster;
314 completer = t->completer;
315 completer_cookie = t->completer_cookie;
316 t->wanted_cluster = -1;
318 spin_unlock(&t->lock);
321 bL_switch_to(cluster);
324 completer(completer_cookie);
326 } while (!kthread_should_stop());
331 static struct task_struct *bL_switcher_thread_create(int cpu, void *arg)
333 struct task_struct *task;
335 task = kthread_create_on_node(bL_switcher_thread, arg,
336 cpu_to_node(cpu), "kswitcher_%d", cpu);
338 kthread_bind(task, cpu);
339 wake_up_process(task);
341 pr_err("%s failed for CPU %d\n", __func__, cpu);
346 * bL_switch_request_cb - Switch to a specific cluster for the given CPU,
347 * with completion notification via a callback
349 * @cpu: the CPU to switch
350 * @new_cluster_id: the ID of the cluster to switch to.
351 * @completer: switch completion callback. if non-NULL,
352 * @completer(@completer_cookie) will be called on completion of
353 * the switch, in non-atomic context.
354 * @completer_cookie: opaque context argument for @completer.
356 * This function causes a cluster switch on the given CPU by waking up
357 * the appropriate switcher thread. This function may or may not return
358 * before the switch has occurred.
360 * If a @completer callback function is supplied, it will be called when
361 * the switch is complete. This can be used to determine asynchronously
362 * when the switch is complete, regardless of when bL_switch_request()
363 * returns. When @completer is supplied, no new switch request is permitted
364 * for the affected CPU until after the switch is complete, and @completer
367 int bL_switch_request_cb(unsigned int cpu, unsigned int new_cluster_id,
368 bL_switch_completion_handler completer,
369 void *completer_cookie)
373 if (cpu >= ARRAY_SIZE(bL_threads)) {
374 pr_err("%s: cpu %d out of bounds\n", __func__, cpu);
378 t = &bL_threads[cpu];
381 return PTR_ERR(t->task);
387 spin_unlock(&t->lock);
390 t->completer = completer;
391 t->completer_cookie = completer_cookie;
392 t->wanted_cluster = new_cluster_id;
393 spin_unlock(&t->lock);
397 EXPORT_SYMBOL_GPL(bL_switch_request_cb);
400 * Activation and configuration code.
403 static DEFINE_MUTEX(bL_switcher_activation_lock);
404 static BLOCKING_NOTIFIER_HEAD(bL_activation_notifier);
405 static unsigned int bL_switcher_active;
406 static unsigned int bL_switcher_cpu_original_cluster[NR_CPUS];
407 static cpumask_t bL_switcher_removed_logical_cpus;
409 int bL_switcher_register_notifier(struct notifier_block *nb)
411 return blocking_notifier_chain_register(&bL_activation_notifier, nb);
413 EXPORT_SYMBOL_GPL(bL_switcher_register_notifier);
415 int bL_switcher_unregister_notifier(struct notifier_block *nb)
417 return blocking_notifier_chain_unregister(&bL_activation_notifier, nb);
419 EXPORT_SYMBOL_GPL(bL_switcher_unregister_notifier);
421 static int bL_activation_notify(unsigned long val)
425 ret = blocking_notifier_call_chain(&bL_activation_notifier, val, NULL);
426 if (ret & NOTIFY_STOP_MASK)
427 pr_err("%s: notifier chain failed with status 0x%x\n",
429 return notifier_to_errno(ret);
432 static void bL_switcher_restore_cpus(void)
436 for_each_cpu(i, &bL_switcher_removed_logical_cpus) {
437 struct device *cpu_dev = get_cpu_device(i);
438 int ret = device_online(cpu_dev);
440 dev_err(cpu_dev, "switcher: unable to restore CPU\n");
444 static int bL_switcher_halve_cpus(void)
446 int i, j, cluster_0, gic_id, ret;
447 unsigned int cpu, cluster, mask;
448 cpumask_t available_cpus;
450 /* First pass to validate what we have */
452 for_each_online_cpu(i) {
453 cpu = MPIDR_AFFINITY_LEVEL(cpu_logical_map(i), 0);
454 cluster = MPIDR_AFFINITY_LEVEL(cpu_logical_map(i), 1);
456 pr_err("%s: only dual cluster systems are supported\n", __func__);
459 if (WARN_ON(cpu >= MAX_CPUS_PER_CLUSTER))
461 mask |= (1 << cluster);
464 pr_err("%s: no CPU pairing possible\n", __func__);
469 * Now let's do the pairing. We match each CPU with another CPU
470 * from a different cluster. To get a uniform scheduling behavior
471 * without fiddling with CPU topology and compute capacity data,
472 * we'll use logical CPUs initially belonging to the same cluster.
474 memset(bL_switcher_cpu_pairing, -1, sizeof(bL_switcher_cpu_pairing));
475 cpumask_copy(&available_cpus, cpu_online_mask);
477 for_each_cpu(i, &available_cpus) {
479 cluster = MPIDR_AFFINITY_LEVEL(cpu_logical_map(i), 1);
482 if (cluster != cluster_0)
484 cpumask_clear_cpu(i, &available_cpus);
485 for_each_cpu(j, &available_cpus) {
486 cluster = MPIDR_AFFINITY_LEVEL(cpu_logical_map(j), 1);
488 * Let's remember the last match to create "odd"
489 * pairings on purpose in order for other code not
490 * to assume any relation between physical and
491 * logical CPU numbers.
493 if (cluster != cluster_0)
497 bL_switcher_cpu_pairing[i] = match;
498 cpumask_clear_cpu(match, &available_cpus);
499 pr_info("CPU%d paired with CPU%d\n", i, match);
504 * Now we disable the unwanted CPUs i.e. everything that has no
505 * pairing information (that includes the pairing counterparts).
507 cpumask_clear(&bL_switcher_removed_logical_cpus);
508 for_each_online_cpu(i) {
509 cpu = MPIDR_AFFINITY_LEVEL(cpu_logical_map(i), 0);
510 cluster = MPIDR_AFFINITY_LEVEL(cpu_logical_map(i), 1);
512 /* Let's take note of the GIC ID for this CPU */
513 gic_id = gic_get_cpu_id(i);
515 pr_err("%s: bad GIC ID for CPU %d\n", __func__, i);
516 bL_switcher_restore_cpus();
519 bL_gic_id[cpu][cluster] = gic_id;
520 pr_info("GIC ID for CPU %u cluster %u is %u\n",
521 cpu, cluster, gic_id);
523 if (bL_switcher_cpu_pairing[i] != -1) {
524 bL_switcher_cpu_original_cluster[i] = cluster;
528 ret = device_offline(get_cpu_device(i));
530 bL_switcher_restore_cpus();
533 cpumask_set_cpu(i, &bL_switcher_removed_logical_cpus);
539 /* Determine the logical CPU a given physical CPU is grouped on. */
540 int bL_switcher_get_logical_index(u32 mpidr)
544 if (!bL_switcher_active)
547 mpidr &= MPIDR_HWID_BITMASK;
548 for_each_online_cpu(cpu) {
549 int pairing = bL_switcher_cpu_pairing[cpu];
552 if ((mpidr == cpu_logical_map(cpu)) ||
553 (mpidr == cpu_logical_map(pairing)))
559 static void bL_switcher_trace_trigger_cpu(void *__always_unused info)
561 trace_cpu_migrate_current(get_ns(), read_mpidr());
564 int bL_switcher_trace_trigger(void)
570 bL_switcher_trace_trigger_cpu(NULL);
571 ret = smp_call_function(bL_switcher_trace_trigger_cpu, NULL, true);
577 EXPORT_SYMBOL_GPL(bL_switcher_trace_trigger);
579 static int bL_switcher_enable(void)
583 mutex_lock(&bL_switcher_activation_lock);
584 lock_device_hotplug();
585 if (bL_switcher_active) {
586 unlock_device_hotplug();
587 mutex_unlock(&bL_switcher_activation_lock);
591 pr_info("big.LITTLE switcher initializing\n");
593 ret = bL_activation_notify(BL_NOTIFY_PRE_ENABLE);
597 ret = bL_switcher_halve_cpus();
601 bL_switcher_trace_trigger();
603 for_each_online_cpu(cpu) {
604 struct bL_thread *t = &bL_threads[cpu];
605 spin_lock_init(&t->lock);
606 init_waitqueue_head(&t->wq);
607 init_completion(&t->started);
608 t->wanted_cluster = -1;
609 t->task = bL_switcher_thread_create(cpu, t);
612 bL_switcher_active = 1;
613 bL_activation_notify(BL_NOTIFY_POST_ENABLE);
614 pr_info("big.LITTLE switcher initialized\n");
618 pr_warn("big.LITTLE switcher initialization failed\n");
619 bL_activation_notify(BL_NOTIFY_POST_DISABLE);
622 unlock_device_hotplug();
623 mutex_unlock(&bL_switcher_activation_lock);
629 static void bL_switcher_disable(void)
631 unsigned int cpu, cluster;
633 struct task_struct *task;
635 mutex_lock(&bL_switcher_activation_lock);
636 lock_device_hotplug();
638 if (!bL_switcher_active)
641 if (bL_activation_notify(BL_NOTIFY_PRE_DISABLE) != 0) {
642 bL_activation_notify(BL_NOTIFY_POST_ENABLE);
646 bL_switcher_active = 0;
649 * To deactivate the switcher, we must shut down the switcher
650 * threads to prevent any other requests from being accepted.
651 * Then, if the final cluster for given logical CPU is not the
652 * same as the original one, we'll recreate a switcher thread
653 * just for the purpose of switching the CPU back without any
654 * possibility for interference from external requests.
656 for_each_online_cpu(cpu) {
657 t = &bL_threads[cpu];
660 if (!task || IS_ERR(task))
663 /* no more switch may happen on this CPU at this point */
664 cluster = MPIDR_AFFINITY_LEVEL(cpu_logical_map(cpu), 1);
665 if (cluster == bL_switcher_cpu_original_cluster[cpu])
667 init_completion(&t->started);
668 t->wanted_cluster = bL_switcher_cpu_original_cluster[cpu];
669 task = bL_switcher_thread_create(cpu, t);
671 wait_for_completion(&t->started);
673 cluster = MPIDR_AFFINITY_LEVEL(cpu_logical_map(cpu), 1);
674 if (cluster == bL_switcher_cpu_original_cluster[cpu])
677 /* If execution gets here, we're in trouble. */
678 pr_crit("%s: unable to restore original cluster for CPU %d\n",
680 pr_crit("%s: CPU %d can't be restored\n",
681 __func__, bL_switcher_cpu_pairing[cpu]);
682 cpumask_clear_cpu(bL_switcher_cpu_pairing[cpu],
683 &bL_switcher_removed_logical_cpus);
686 bL_switcher_restore_cpus();
687 bL_switcher_trace_trigger();
689 bL_activation_notify(BL_NOTIFY_POST_DISABLE);
692 unlock_device_hotplug();
693 mutex_unlock(&bL_switcher_activation_lock);
696 static ssize_t bL_switcher_active_show(struct kobject *kobj,
697 struct kobj_attribute *attr, char *buf)
699 return sprintf(buf, "%u\n", bL_switcher_active);
702 static ssize_t bL_switcher_active_store(struct kobject *kobj,
703 struct kobj_attribute *attr, const char *buf, size_t count)
709 bL_switcher_disable();
713 ret = bL_switcher_enable();
719 return (ret >= 0) ? count : ret;
722 static ssize_t bL_switcher_trace_trigger_store(struct kobject *kobj,
723 struct kobj_attribute *attr, const char *buf, size_t count)
725 int ret = bL_switcher_trace_trigger();
727 return ret ? ret : count;
730 static struct kobj_attribute bL_switcher_active_attr =
731 __ATTR(active, 0644, bL_switcher_active_show, bL_switcher_active_store);
733 static struct kobj_attribute bL_switcher_trace_trigger_attr =
734 __ATTR(trace_trigger, 0200, NULL, bL_switcher_trace_trigger_store);
736 static struct attribute *bL_switcher_attrs[] = {
737 &bL_switcher_active_attr.attr,
738 &bL_switcher_trace_trigger_attr.attr,
742 static struct attribute_group bL_switcher_attr_group = {
743 .attrs = bL_switcher_attrs,
746 static struct kobject *bL_switcher_kobj;
748 static int __init bL_switcher_sysfs_init(void)
752 bL_switcher_kobj = kobject_create_and_add("bL_switcher", kernel_kobj);
753 if (!bL_switcher_kobj)
755 ret = sysfs_create_group(bL_switcher_kobj, &bL_switcher_attr_group);
757 kobject_put(bL_switcher_kobj);
761 #endif /* CONFIG_SYSFS */
763 bool bL_switcher_get_enabled(void)
765 mutex_lock(&bL_switcher_activation_lock);
767 return bL_switcher_active;
769 EXPORT_SYMBOL_GPL(bL_switcher_get_enabled);
771 void bL_switcher_put_enabled(void)
773 mutex_unlock(&bL_switcher_activation_lock);
775 EXPORT_SYMBOL_GPL(bL_switcher_put_enabled);
778 * Veto any CPU hotplug operation on those CPUs we've removed
779 * while the switcher is active.
780 * We're just not ready to deal with that given the trickery involved.
782 static int bL_switcher_hotplug_callback(struct notifier_block *nfb,
783 unsigned long action, void *hcpu)
785 if (bL_switcher_active) {
786 int pairing = bL_switcher_cpu_pairing[(unsigned long)hcpu];
787 switch (action & 0xf) {
789 case CPU_DOWN_PREPARE:
797 static bool no_bL_switcher;
798 core_param(no_bL_switcher, no_bL_switcher, bool, 0644);
800 static int __init bL_switcher_init(void)
804 if (!mcpm_is_available())
807 cpu_notifier(bL_switcher_hotplug_callback, 0);
809 if (!no_bL_switcher) {
810 ret = bL_switcher_enable();
816 ret = bL_switcher_sysfs_init();
818 pr_err("%s: unable to create sysfs entry\n", __func__);
824 late_initcall(bL_switcher_init);