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 *_unused)
60 unsigned ib_mpidr, ib_cpu, ib_cluster;
62 pr_debug("%s\n", __func__);
64 ib_mpidr = cpu_logical_map(smp_processor_id());
65 ib_cpu = MPIDR_AFFINITY_LEVEL(ib_mpidr, 0);
66 ib_cluster = MPIDR_AFFINITY_LEVEL(ib_mpidr, 1);
69 * Our state has been saved at this point. Let's release our
72 mcpm_set_entry_vector(ib_cpu, ib_cluster, cpu_resume);
76 * From this point, we must assume that our counterpart CPU might
77 * have taken over in its parallel world already, as if execution
78 * just returned from cpu_suspend(). It is therefore important to
79 * be very careful not to make any change the other guy is not
80 * expecting. This is why we need stack isolation.
82 * Fancy under cover tasks could be performed here. For now
86 /* Let's put ourself down. */
87 mcpm_cpu_power_down();
89 /* should never get here */
94 * Stack isolation. To ensure 'current' remains valid, we just use another
95 * piece of our thread's stack space which should be fairly lightly used.
96 * The selected area starts just above the thread_info structure located
97 * at the very bottom of the stack, aligned to a cache line, and indexed
98 * with the cluster number.
100 #define STACK_SIZE 512
101 extern void call_with_stack(void (*fn)(void *), void *arg, void *sp);
102 static int bL_switchpoint(unsigned long _arg)
104 unsigned int mpidr = read_mpidr();
105 unsigned int clusterid = MPIDR_AFFINITY_LEVEL(mpidr, 1);
106 void *stack = current_thread_info() + 1;
107 stack = PTR_ALIGN(stack, L1_CACHE_BYTES);
108 stack += clusterid * STACK_SIZE + STACK_SIZE;
109 call_with_stack(bL_do_switch, (void *)_arg, stack);
114 * Generic switcher interface
117 static unsigned int bL_gic_id[MAX_CPUS_PER_CLUSTER][MAX_NR_CLUSTERS];
118 static int bL_switcher_cpu_pairing[NR_CPUS];
121 * bL_switch_to - Switch to a specific cluster for the current CPU
122 * @new_cluster_id: the ID of the cluster to switch to.
124 * This function must be called on the CPU to be switched.
125 * Returns 0 on success, else a negative status code.
127 static int bL_switch_to(unsigned int new_cluster_id)
129 unsigned int mpidr, this_cpu, that_cpu;
130 unsigned int ob_mpidr, ob_cpu, ob_cluster, ib_mpidr, ib_cpu, ib_cluster;
131 struct tick_device *tdev;
132 enum clock_event_mode tdev_mode;
135 this_cpu = smp_processor_id();
136 ob_mpidr = read_mpidr();
137 ob_cpu = MPIDR_AFFINITY_LEVEL(ob_mpidr, 0);
138 ob_cluster = MPIDR_AFFINITY_LEVEL(ob_mpidr, 1);
139 BUG_ON(cpu_logical_map(this_cpu) != ob_mpidr);
141 if (new_cluster_id == ob_cluster)
144 that_cpu = bL_switcher_cpu_pairing[this_cpu];
145 ib_mpidr = cpu_logical_map(that_cpu);
146 ib_cpu = MPIDR_AFFINITY_LEVEL(ib_mpidr, 0);
147 ib_cluster = MPIDR_AFFINITY_LEVEL(ib_mpidr, 1);
149 pr_debug("before switch: CPU %d MPIDR %#x -> %#x\n",
150 this_cpu, ob_mpidr, ib_mpidr);
152 /* Close the gate for our entry vectors */
153 mcpm_set_entry_vector(ob_cpu, ob_cluster, NULL);
154 mcpm_set_entry_vector(ib_cpu, ib_cluster, NULL);
157 * Let's wake up the inbound CPU now in case it requires some delay
158 * to come online, but leave it gated in our entry vector code.
160 ret = mcpm_cpu_power_up(ib_cpu, ib_cluster);
162 pr_err("%s: mcpm_cpu_power_up() returned %d\n", __func__, ret);
167 * From this point we are entering the switch critical zone
168 * and can't take any interrupts anymore.
173 /* redirect GIC's SGIs to our counterpart */
174 gic_migrate_target(bL_gic_id[ib_cpu][ib_cluster]);
177 * Raise a SGI on the inbound CPU to make sure it doesn't stall
178 * in a possible WFI, such as in mcpm_power_down().
180 arch_send_wakeup_ipi_mask(cpumask_of(this_cpu));
182 tdev = tick_get_device(this_cpu);
183 if (tdev && !cpumask_equal(tdev->evtdev->cpumask, cpumask_of(this_cpu)))
186 tdev_mode = tdev->evtdev->mode;
187 clockevents_set_mode(tdev->evtdev, CLOCK_EVT_MODE_SHUTDOWN);
190 ret = cpu_pm_enter();
192 /* we can not tolerate errors at this point */
194 panic("%s: cpu_pm_enter() returned %d\n", __func__, ret);
196 /* Swap the physical CPUs in the logical map for this logical CPU. */
197 cpu_logical_map(this_cpu) = ib_mpidr;
198 cpu_logical_map(that_cpu) = ob_mpidr;
200 /* Let's do the actual CPU switch. */
201 ret = cpu_suspend(0, bL_switchpoint);
203 panic("%s: cpu_suspend() returned %d\n", __func__, ret);
205 /* We are executing on the inbound CPU at this point */
206 mpidr = read_mpidr();
207 pr_debug("after switch: CPU %d MPIDR %#x\n", this_cpu, mpidr);
208 BUG_ON(mpidr != ib_mpidr);
210 mcpm_cpu_powered_up();
215 clockevents_set_mode(tdev->evtdev, tdev_mode);
216 clockevents_program_event(tdev->evtdev,
217 tdev->evtdev->next_event, 1);
224 pr_err("%s exiting with error %d\n", __func__, ret);
230 struct task_struct *task;
231 wait_queue_head_t wq;
233 struct completion started;
234 bL_switch_completion_handler completer;
235 void *completer_cookie;
238 static struct bL_thread bL_threads[NR_CPUS];
240 static int bL_switcher_thread(void *arg)
242 struct bL_thread *t = arg;
243 struct sched_param param = { .sched_priority = 1 };
245 bL_switch_completion_handler completer;
246 void *completer_cookie;
248 sched_setscheduler_nocheck(current, SCHED_FIFO, ¶m);
249 complete(&t->started);
252 if (signal_pending(current))
253 flush_signals(current);
254 wait_event_interruptible(t->wq,
255 t->wanted_cluster != -1 ||
256 kthread_should_stop());
259 cluster = t->wanted_cluster;
260 completer = t->completer;
261 completer_cookie = t->completer_cookie;
262 t->wanted_cluster = -1;
264 spin_unlock(&t->lock);
267 bL_switch_to(cluster);
270 completer(completer_cookie);
272 } while (!kthread_should_stop());
277 static struct task_struct *bL_switcher_thread_create(int cpu, void *arg)
279 struct task_struct *task;
281 task = kthread_create_on_node(bL_switcher_thread, arg,
282 cpu_to_node(cpu), "kswitcher_%d", cpu);
284 kthread_bind(task, cpu);
285 wake_up_process(task);
287 pr_err("%s failed for CPU %d\n", __func__, cpu);
292 * bL_switch_request_cb - Switch to a specific cluster for the given CPU,
293 * with completion notification via a callback
295 * @cpu: the CPU to switch
296 * @new_cluster_id: the ID of the cluster to switch to.
297 * @completer: switch completion callback. if non-NULL,
298 * @completer(@completer_cookie) will be called on completion of
299 * the switch, in non-atomic context.
300 * @completer_cookie: opaque context argument for @completer.
302 * This function causes a cluster switch on the given CPU by waking up
303 * the appropriate switcher thread. This function may or may not return
304 * before the switch has occurred.
306 * If a @completer callback function is supplied, it will be called when
307 * the switch is complete. This can be used to determine asynchronously
308 * when the switch is complete, regardless of when bL_switch_request()
309 * returns. When @completer is supplied, no new switch request is permitted
310 * for the affected CPU until after the switch is complete, and @completer
313 int bL_switch_request_cb(unsigned int cpu, unsigned int new_cluster_id,
314 bL_switch_completion_handler completer,
315 void *completer_cookie)
319 if (cpu >= ARRAY_SIZE(bL_threads)) {
320 pr_err("%s: cpu %d out of bounds\n", __func__, cpu);
324 t = &bL_threads[cpu];
327 return PTR_ERR(t->task);
333 spin_unlock(&t->lock);
336 t->completer = completer;
337 t->completer_cookie = completer_cookie;
338 t->wanted_cluster = new_cluster_id;
339 spin_unlock(&t->lock);
343 EXPORT_SYMBOL_GPL(bL_switch_request_cb);
346 * Activation and configuration code.
349 static DEFINE_MUTEX(bL_switcher_activation_lock);
350 static BLOCKING_NOTIFIER_HEAD(bL_activation_notifier);
351 static unsigned int bL_switcher_active;
352 static unsigned int bL_switcher_cpu_original_cluster[NR_CPUS];
353 static cpumask_t bL_switcher_removed_logical_cpus;
355 int bL_switcher_register_notifier(struct notifier_block *nb)
357 return blocking_notifier_chain_register(&bL_activation_notifier, nb);
359 EXPORT_SYMBOL_GPL(bL_switcher_register_notifier);
361 int bL_switcher_unregister_notifier(struct notifier_block *nb)
363 return blocking_notifier_chain_unregister(&bL_activation_notifier, nb);
365 EXPORT_SYMBOL_GPL(bL_switcher_unregister_notifier);
367 static int bL_activation_notify(unsigned long val)
371 ret = blocking_notifier_call_chain(&bL_activation_notifier, val, NULL);
372 if (ret & NOTIFY_STOP_MASK)
373 pr_err("%s: notifier chain failed with status 0x%x\n",
375 return notifier_to_errno(ret);
378 static void bL_switcher_restore_cpus(void)
382 for_each_cpu(i, &bL_switcher_removed_logical_cpus)
386 static int bL_switcher_halve_cpus(void)
388 int i, j, cluster_0, gic_id, ret;
389 unsigned int cpu, cluster, mask;
390 cpumask_t available_cpus;
392 /* First pass to validate what we have */
394 for_each_online_cpu(i) {
395 cpu = MPIDR_AFFINITY_LEVEL(cpu_logical_map(i), 0);
396 cluster = MPIDR_AFFINITY_LEVEL(cpu_logical_map(i), 1);
398 pr_err("%s: only dual cluster systems are supported\n", __func__);
401 if (WARN_ON(cpu >= MAX_CPUS_PER_CLUSTER))
403 mask |= (1 << cluster);
406 pr_err("%s: no CPU pairing possible\n", __func__);
411 * Now let's do the pairing. We match each CPU with another CPU
412 * from a different cluster. To get a uniform scheduling behavior
413 * without fiddling with CPU topology and compute capacity data,
414 * we'll use logical CPUs initially belonging to the same cluster.
416 memset(bL_switcher_cpu_pairing, -1, sizeof(bL_switcher_cpu_pairing));
417 cpumask_copy(&available_cpus, cpu_online_mask);
419 for_each_cpu(i, &available_cpus) {
421 cluster = MPIDR_AFFINITY_LEVEL(cpu_logical_map(i), 1);
424 if (cluster != cluster_0)
426 cpumask_clear_cpu(i, &available_cpus);
427 for_each_cpu(j, &available_cpus) {
428 cluster = MPIDR_AFFINITY_LEVEL(cpu_logical_map(j), 1);
430 * Let's remember the last match to create "odd"
431 * pairings on purpose in order for other code not
432 * to assume any relation between physical and
433 * logical CPU numbers.
435 if (cluster != cluster_0)
439 bL_switcher_cpu_pairing[i] = match;
440 cpumask_clear_cpu(match, &available_cpus);
441 pr_info("CPU%d paired with CPU%d\n", i, match);
446 * Now we disable the unwanted CPUs i.e. everything that has no
447 * pairing information (that includes the pairing counterparts).
449 cpumask_clear(&bL_switcher_removed_logical_cpus);
450 for_each_online_cpu(i) {
451 cpu = MPIDR_AFFINITY_LEVEL(cpu_logical_map(i), 0);
452 cluster = MPIDR_AFFINITY_LEVEL(cpu_logical_map(i), 1);
454 /* Let's take note of the GIC ID for this CPU */
455 gic_id = gic_get_cpu_id(i);
457 pr_err("%s: bad GIC ID for CPU %d\n", __func__, i);
458 bL_switcher_restore_cpus();
461 bL_gic_id[cpu][cluster] = gic_id;
462 pr_info("GIC ID for CPU %u cluster %u is %u\n",
463 cpu, cluster, gic_id);
465 if (bL_switcher_cpu_pairing[i] != -1) {
466 bL_switcher_cpu_original_cluster[i] = cluster;
472 bL_switcher_restore_cpus();
475 cpumask_set_cpu(i, &bL_switcher_removed_logical_cpus);
481 static int bL_switcher_enable(void)
485 mutex_lock(&bL_switcher_activation_lock);
486 cpu_hotplug_driver_lock();
487 if (bL_switcher_active) {
488 cpu_hotplug_driver_unlock();
489 mutex_unlock(&bL_switcher_activation_lock);
493 pr_info("big.LITTLE switcher initializing\n");
495 ret = bL_activation_notify(BL_NOTIFY_PRE_ENABLE);
499 ret = bL_switcher_halve_cpus();
503 for_each_online_cpu(cpu) {
504 struct bL_thread *t = &bL_threads[cpu];
505 spin_lock_init(&t->lock);
506 init_waitqueue_head(&t->wq);
507 init_completion(&t->started);
508 t->wanted_cluster = -1;
509 t->task = bL_switcher_thread_create(cpu, t);
512 bL_switcher_active = 1;
513 bL_activation_notify(BL_NOTIFY_POST_ENABLE);
514 pr_info("big.LITTLE switcher initialized\n");
518 pr_warn("big.LITTLE switcher initialization failed\n");
519 bL_activation_notify(BL_NOTIFY_POST_DISABLE);
522 cpu_hotplug_driver_unlock();
523 mutex_unlock(&bL_switcher_activation_lock);
529 static void bL_switcher_disable(void)
531 unsigned int cpu, cluster;
533 struct task_struct *task;
535 mutex_lock(&bL_switcher_activation_lock);
536 cpu_hotplug_driver_lock();
538 if (!bL_switcher_active)
541 if (bL_activation_notify(BL_NOTIFY_PRE_DISABLE) != 0) {
542 bL_activation_notify(BL_NOTIFY_POST_ENABLE);
546 bL_switcher_active = 0;
549 * To deactivate the switcher, we must shut down the switcher
550 * threads to prevent any other requests from being accepted.
551 * Then, if the final cluster for given logical CPU is not the
552 * same as the original one, we'll recreate a switcher thread
553 * just for the purpose of switching the CPU back without any
554 * possibility for interference from external requests.
556 for_each_online_cpu(cpu) {
557 t = &bL_threads[cpu];
560 if (!task || IS_ERR(task))
563 /* no more switch may happen on this CPU at this point */
564 cluster = MPIDR_AFFINITY_LEVEL(cpu_logical_map(cpu), 1);
565 if (cluster == bL_switcher_cpu_original_cluster[cpu])
567 init_completion(&t->started);
568 t->wanted_cluster = bL_switcher_cpu_original_cluster[cpu];
569 task = bL_switcher_thread_create(cpu, t);
571 wait_for_completion(&t->started);
573 cluster = MPIDR_AFFINITY_LEVEL(cpu_logical_map(cpu), 1);
574 if (cluster == bL_switcher_cpu_original_cluster[cpu])
577 /* If execution gets here, we're in trouble. */
578 pr_crit("%s: unable to restore original cluster for CPU %d\n",
580 pr_crit("%s: CPU %d can't be restored\n",
581 __func__, bL_switcher_cpu_pairing[cpu]);
582 cpumask_clear_cpu(bL_switcher_cpu_pairing[cpu],
583 &bL_switcher_removed_logical_cpus);
586 bL_switcher_restore_cpus();
587 bL_activation_notify(BL_NOTIFY_POST_DISABLE);
590 cpu_hotplug_driver_unlock();
591 mutex_unlock(&bL_switcher_activation_lock);
594 static ssize_t bL_switcher_active_show(struct kobject *kobj,
595 struct kobj_attribute *attr, char *buf)
597 return sprintf(buf, "%u\n", bL_switcher_active);
600 static ssize_t bL_switcher_active_store(struct kobject *kobj,
601 struct kobj_attribute *attr, const char *buf, size_t count)
607 bL_switcher_disable();
611 ret = bL_switcher_enable();
617 return (ret >= 0) ? count : ret;
620 static struct kobj_attribute bL_switcher_active_attr =
621 __ATTR(active, 0644, bL_switcher_active_show, bL_switcher_active_store);
623 static struct attribute *bL_switcher_attrs[] = {
624 &bL_switcher_active_attr.attr,
628 static struct attribute_group bL_switcher_attr_group = {
629 .attrs = bL_switcher_attrs,
632 static struct kobject *bL_switcher_kobj;
634 static int __init bL_switcher_sysfs_init(void)
638 bL_switcher_kobj = kobject_create_and_add("bL_switcher", kernel_kobj);
639 if (!bL_switcher_kobj)
641 ret = sysfs_create_group(bL_switcher_kobj, &bL_switcher_attr_group);
643 kobject_put(bL_switcher_kobj);
647 #endif /* CONFIG_SYSFS */
649 bool bL_switcher_get_enabled(void)
651 mutex_lock(&bL_switcher_activation_lock);
653 return bL_switcher_active;
655 EXPORT_SYMBOL_GPL(bL_switcher_get_enabled);
657 void bL_switcher_put_enabled(void)
659 mutex_unlock(&bL_switcher_activation_lock);
661 EXPORT_SYMBOL_GPL(bL_switcher_put_enabled);
664 * Veto any CPU hotplug operation on those CPUs we've removed
665 * while the switcher is active.
666 * We're just not ready to deal with that given the trickery involved.
668 static int bL_switcher_hotplug_callback(struct notifier_block *nfb,
669 unsigned long action, void *hcpu)
671 if (bL_switcher_active) {
672 int pairing = bL_switcher_cpu_pairing[(unsigned long)hcpu];
673 switch (action & 0xf) {
675 case CPU_DOWN_PREPARE:
683 static bool no_bL_switcher;
684 core_param(no_bL_switcher, no_bL_switcher, bool, 0644);
686 static int __init bL_switcher_init(void)
690 if (MAX_NR_CLUSTERS != 2) {
691 pr_err("%s: only dual cluster systems are supported\n", __func__);
695 cpu_notifier(bL_switcher_hotplug_callback, 0);
697 if (!no_bL_switcher) {
698 ret = bL_switcher_enable();
704 ret = bL_switcher_sysfs_init();
706 pr_err("%s: unable to create sysfs entry\n", __func__);
712 late_initcall(bL_switcher_init);