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/signal.h>
17 #include <uapi/linux/sched/types.h>
18 #include <linux/interrupt.h>
19 #include <linux/cpu_pm.h>
20 #include <linux/cpu.h>
21 #include <linux/cpumask.h>
22 #include <linux/kthread.h>
23 #include <linux/wait.h>
24 #include <linux/time.h>
25 #include <linux/clockchips.h>
26 #include <linux/hrtimer.h>
27 #include <linux/tick.h>
28 #include <linux/notifier.h>
30 #include <linux/mutex.h>
31 #include <linux/smp.h>
32 #include <linux/spinlock.h>
33 #include <linux/string.h>
34 #include <linux/sysfs.h>
35 #include <linux/irqchip/arm-gic.h>
36 #include <linux/moduleparam.h>
38 #include <asm/smp_plat.h>
39 #include <asm/cputype.h>
40 #include <asm/suspend.h>
42 #include <asm/bL_switcher.h>
44 #define CREATE_TRACE_POINTS
45 #include <trace/events/power_cpu_migrate.h>
49 * Use our own MPIDR accessors as the generic ones in asm/cputype.h have
50 * __attribute_const__ and we don't want the compiler to assume any
51 * constness here as the value _does_ change along some code paths.
54 static int read_mpidr(void)
57 asm volatile ("mrc p15, 0, %0, c0, c0, 5" : "=r" (id
));
58 return id
& MPIDR_HWID_BITMASK
;
62 * bL switcher core code.
65 static void bL_do_switch(void *_arg
)
67 unsigned ib_mpidr
, ib_cpu
, ib_cluster
;
68 long volatile handshake
, **handshake_ptr
= _arg
;
70 pr_debug("%s\n", __func__
);
72 ib_mpidr
= cpu_logical_map(smp_processor_id());
73 ib_cpu
= MPIDR_AFFINITY_LEVEL(ib_mpidr
, 0);
74 ib_cluster
= MPIDR_AFFINITY_LEVEL(ib_mpidr
, 1);
76 /* Advertise our handshake location */
79 *handshake_ptr
= &handshake
;
84 * Our state has been saved at this point. Let's release our
87 mcpm_set_entry_vector(ib_cpu
, ib_cluster
, cpu_resume
);
91 * From this point, we must assume that our counterpart CPU might
92 * have taken over in its parallel world already, as if execution
93 * just returned from cpu_suspend(). It is therefore important to
94 * be very careful not to make any change the other guy is not
95 * expecting. This is why we need stack isolation.
97 * Fancy under cover tasks could be performed here. For now
102 * Let's wait until our inbound is alive.
109 /* Let's put ourself down. */
110 mcpm_cpu_power_down();
112 /* should never get here */
117 * Stack isolation. To ensure 'current' remains valid, we just use another
118 * piece of our thread's stack space which should be fairly lightly used.
119 * The selected area starts just above the thread_info structure located
120 * at the very bottom of the stack, aligned to a cache line, and indexed
121 * with the cluster number.
123 #define STACK_SIZE 512
124 extern void call_with_stack(void (*fn
)(void *), void *arg
, void *sp
);
125 static int bL_switchpoint(unsigned long _arg
)
127 unsigned int mpidr
= read_mpidr();
128 unsigned int clusterid
= MPIDR_AFFINITY_LEVEL(mpidr
, 1);
129 void *stack
= current_thread_info() + 1;
130 stack
= PTR_ALIGN(stack
, L1_CACHE_BYTES
);
131 stack
+= clusterid
* STACK_SIZE
+ STACK_SIZE
;
132 call_with_stack(bL_do_switch
, (void *)_arg
, stack
);
137 * Generic switcher interface
140 static unsigned int bL_gic_id
[MAX_CPUS_PER_CLUSTER
][MAX_NR_CLUSTERS
];
141 static int bL_switcher_cpu_pairing
[NR_CPUS
];
144 * bL_switch_to - Switch to a specific cluster for the current CPU
145 * @new_cluster_id: the ID of the cluster to switch to.
147 * This function must be called on the CPU to be switched.
148 * Returns 0 on success, else a negative status code.
150 static int bL_switch_to(unsigned int new_cluster_id
)
152 unsigned int mpidr
, this_cpu
, that_cpu
;
153 unsigned int ob_mpidr
, ob_cpu
, ob_cluster
, ib_mpidr
, ib_cpu
, ib_cluster
;
154 struct completion inbound_alive
;
155 long volatile *handshake_ptr
;
158 this_cpu
= smp_processor_id();
159 ob_mpidr
= read_mpidr();
160 ob_cpu
= MPIDR_AFFINITY_LEVEL(ob_mpidr
, 0);
161 ob_cluster
= MPIDR_AFFINITY_LEVEL(ob_mpidr
, 1);
162 BUG_ON(cpu_logical_map(this_cpu
) != ob_mpidr
);
164 if (new_cluster_id
== ob_cluster
)
167 that_cpu
= bL_switcher_cpu_pairing
[this_cpu
];
168 ib_mpidr
= cpu_logical_map(that_cpu
);
169 ib_cpu
= MPIDR_AFFINITY_LEVEL(ib_mpidr
, 0);
170 ib_cluster
= MPIDR_AFFINITY_LEVEL(ib_mpidr
, 1);
172 pr_debug("before switch: CPU %d MPIDR %#x -> %#x\n",
173 this_cpu
, ob_mpidr
, ib_mpidr
);
175 this_cpu
= smp_processor_id();
177 /* Close the gate for our entry vectors */
178 mcpm_set_entry_vector(ob_cpu
, ob_cluster
, NULL
);
179 mcpm_set_entry_vector(ib_cpu
, ib_cluster
, NULL
);
181 /* Install our "inbound alive" notifier. */
182 init_completion(&inbound_alive
);
183 ipi_nr
= register_ipi_completion(&inbound_alive
, this_cpu
);
184 ipi_nr
|= ((1 << 16) << bL_gic_id
[ob_cpu
][ob_cluster
]);
185 mcpm_set_early_poke(ib_cpu
, ib_cluster
, gic_get_sgir_physaddr(), ipi_nr
);
188 * Let's wake up the inbound CPU now in case it requires some delay
189 * to come online, but leave it gated in our entry vector code.
191 ret
= mcpm_cpu_power_up(ib_cpu
, ib_cluster
);
193 pr_err("%s: mcpm_cpu_power_up() returned %d\n", __func__
, ret
);
198 * Raise a SGI on the inbound CPU to make sure it doesn't stall
199 * in a possible WFI, such as in bL_power_down().
201 gic_send_sgi(bL_gic_id
[ib_cpu
][ib_cluster
], 0);
204 * Wait for the inbound to come up. This allows for other
205 * tasks to be scheduled in the mean time.
207 wait_for_completion(&inbound_alive
);
208 mcpm_set_early_poke(ib_cpu
, ib_cluster
, 0, 0);
211 * From this point we are entering the switch critical zone
212 * and can't take any interrupts anymore.
216 trace_cpu_migrate_begin(ktime_get_real_ns(), ob_mpidr
);
218 /* redirect GIC's SGIs to our counterpart */
219 gic_migrate_target(bL_gic_id
[ib_cpu
][ib_cluster
]);
221 tick_suspend_local();
223 ret
= cpu_pm_enter();
225 /* we can not tolerate errors at this point */
227 panic("%s: cpu_pm_enter() returned %d\n", __func__
, ret
);
229 /* Swap the physical CPUs in the logical map for this logical CPU. */
230 cpu_logical_map(this_cpu
) = ib_mpidr
;
231 cpu_logical_map(that_cpu
) = ob_mpidr
;
233 /* Let's do the actual CPU switch. */
234 ret
= cpu_suspend((unsigned long)&handshake_ptr
, bL_switchpoint
);
236 panic("%s: cpu_suspend() returned %d\n", __func__
, ret
);
238 /* We are executing on the inbound CPU at this point */
239 mpidr
= read_mpidr();
240 pr_debug("after switch: CPU %d MPIDR %#x\n", this_cpu
, mpidr
);
241 BUG_ON(mpidr
!= ib_mpidr
);
243 mcpm_cpu_powered_up();
249 trace_cpu_migrate_finish(ktime_get_real_ns(), ib_mpidr
);
257 pr_err("%s exiting with error %d\n", __func__
, ret
);
263 struct task_struct
*task
;
264 wait_queue_head_t wq
;
266 struct completion started
;
267 bL_switch_completion_handler completer
;
268 void *completer_cookie
;
271 static struct bL_thread bL_threads
[NR_CPUS
];
273 static int bL_switcher_thread(void *arg
)
275 struct bL_thread
*t
= arg
;
276 struct sched_param param
= { .sched_priority
= 1 };
278 bL_switch_completion_handler completer
;
279 void *completer_cookie
;
281 sched_setscheduler_nocheck(current
, SCHED_FIFO
, ¶m
);
282 complete(&t
->started
);
285 if (signal_pending(current
))
286 flush_signals(current
);
287 wait_event_interruptible(t
->wq
,
288 t
->wanted_cluster
!= -1 ||
289 kthread_should_stop());
292 cluster
= t
->wanted_cluster
;
293 completer
= t
->completer
;
294 completer_cookie
= t
->completer_cookie
;
295 t
->wanted_cluster
= -1;
297 spin_unlock(&t
->lock
);
300 bL_switch_to(cluster
);
303 completer(completer_cookie
);
305 } while (!kthread_should_stop());
310 static struct task_struct
*bL_switcher_thread_create(int cpu
, void *arg
)
312 struct task_struct
*task
;
314 task
= kthread_create_on_node(bL_switcher_thread
, arg
,
315 cpu_to_node(cpu
), "kswitcher_%d", cpu
);
317 kthread_bind(task
, cpu
);
318 wake_up_process(task
);
320 pr_err("%s failed for CPU %d\n", __func__
, cpu
);
325 * bL_switch_request_cb - Switch to a specific cluster for the given CPU,
326 * with completion notification via a callback
328 * @cpu: the CPU to switch
329 * @new_cluster_id: the ID of the cluster to switch to.
330 * @completer: switch completion callback. if non-NULL,
331 * @completer(@completer_cookie) will be called on completion of
332 * the switch, in non-atomic context.
333 * @completer_cookie: opaque context argument for @completer.
335 * This function causes a cluster switch on the given CPU by waking up
336 * the appropriate switcher thread. This function may or may not return
337 * before the switch has occurred.
339 * If a @completer callback function is supplied, it will be called when
340 * the switch is complete. This can be used to determine asynchronously
341 * when the switch is complete, regardless of when bL_switch_request()
342 * returns. When @completer is supplied, no new switch request is permitted
343 * for the affected CPU until after the switch is complete, and @completer
346 int bL_switch_request_cb(unsigned int cpu
, unsigned int new_cluster_id
,
347 bL_switch_completion_handler completer
,
348 void *completer_cookie
)
352 if (cpu
>= ARRAY_SIZE(bL_threads
)) {
353 pr_err("%s: cpu %d out of bounds\n", __func__
, cpu
);
357 t
= &bL_threads
[cpu
];
360 return PTR_ERR(t
->task
);
366 spin_unlock(&t
->lock
);
369 t
->completer
= completer
;
370 t
->completer_cookie
= completer_cookie
;
371 t
->wanted_cluster
= new_cluster_id
;
372 spin_unlock(&t
->lock
);
376 EXPORT_SYMBOL_GPL(bL_switch_request_cb
);
379 * Activation and configuration code.
382 static DEFINE_MUTEX(bL_switcher_activation_lock
);
383 static BLOCKING_NOTIFIER_HEAD(bL_activation_notifier
);
384 static unsigned int bL_switcher_active
;
385 static unsigned int bL_switcher_cpu_original_cluster
[NR_CPUS
];
386 static cpumask_t bL_switcher_removed_logical_cpus
;
388 int bL_switcher_register_notifier(struct notifier_block
*nb
)
390 return blocking_notifier_chain_register(&bL_activation_notifier
, nb
);
392 EXPORT_SYMBOL_GPL(bL_switcher_register_notifier
);
394 int bL_switcher_unregister_notifier(struct notifier_block
*nb
)
396 return blocking_notifier_chain_unregister(&bL_activation_notifier
, nb
);
398 EXPORT_SYMBOL_GPL(bL_switcher_unregister_notifier
);
400 static int bL_activation_notify(unsigned long val
)
404 ret
= blocking_notifier_call_chain(&bL_activation_notifier
, val
, NULL
);
405 if (ret
& NOTIFY_STOP_MASK
)
406 pr_err("%s: notifier chain failed with status 0x%x\n",
408 return notifier_to_errno(ret
);
411 static void bL_switcher_restore_cpus(void)
415 for_each_cpu(i
, &bL_switcher_removed_logical_cpus
) {
416 struct device
*cpu_dev
= get_cpu_device(i
);
417 int ret
= device_online(cpu_dev
);
419 dev_err(cpu_dev
, "switcher: unable to restore CPU\n");
423 static int bL_switcher_halve_cpus(void)
425 int i
, j
, cluster_0
, gic_id
, ret
;
426 unsigned int cpu
, cluster
, mask
;
427 cpumask_t available_cpus
;
429 /* First pass to validate what we have */
431 for_each_online_cpu(i
) {
432 cpu
= MPIDR_AFFINITY_LEVEL(cpu_logical_map(i
), 0);
433 cluster
= MPIDR_AFFINITY_LEVEL(cpu_logical_map(i
), 1);
435 pr_err("%s: only dual cluster systems are supported\n", __func__
);
438 if (WARN_ON(cpu
>= MAX_CPUS_PER_CLUSTER
))
440 mask
|= (1 << cluster
);
443 pr_err("%s: no CPU pairing possible\n", __func__
);
448 * Now let's do the pairing. We match each CPU with another CPU
449 * from a different cluster. To get a uniform scheduling behavior
450 * without fiddling with CPU topology and compute capacity data,
451 * we'll use logical CPUs initially belonging to the same cluster.
453 memset(bL_switcher_cpu_pairing
, -1, sizeof(bL_switcher_cpu_pairing
));
454 cpumask_copy(&available_cpus
, cpu_online_mask
);
456 for_each_cpu(i
, &available_cpus
) {
458 cluster
= MPIDR_AFFINITY_LEVEL(cpu_logical_map(i
), 1);
461 if (cluster
!= cluster_0
)
463 cpumask_clear_cpu(i
, &available_cpus
);
464 for_each_cpu(j
, &available_cpus
) {
465 cluster
= MPIDR_AFFINITY_LEVEL(cpu_logical_map(j
), 1);
467 * Let's remember the last match to create "odd"
468 * pairings on purpose in order for other code not
469 * to assume any relation between physical and
470 * logical CPU numbers.
472 if (cluster
!= cluster_0
)
476 bL_switcher_cpu_pairing
[i
] = match
;
477 cpumask_clear_cpu(match
, &available_cpus
);
478 pr_info("CPU%d paired with CPU%d\n", i
, match
);
483 * Now we disable the unwanted CPUs i.e. everything that has no
484 * pairing information (that includes the pairing counterparts).
486 cpumask_clear(&bL_switcher_removed_logical_cpus
);
487 for_each_online_cpu(i
) {
488 cpu
= MPIDR_AFFINITY_LEVEL(cpu_logical_map(i
), 0);
489 cluster
= MPIDR_AFFINITY_LEVEL(cpu_logical_map(i
), 1);
491 /* Let's take note of the GIC ID for this CPU */
492 gic_id
= gic_get_cpu_id(i
);
494 pr_err("%s: bad GIC ID for CPU %d\n", __func__
, i
);
495 bL_switcher_restore_cpus();
498 bL_gic_id
[cpu
][cluster
] = gic_id
;
499 pr_info("GIC ID for CPU %u cluster %u is %u\n",
500 cpu
, cluster
, gic_id
);
502 if (bL_switcher_cpu_pairing
[i
] != -1) {
503 bL_switcher_cpu_original_cluster
[i
] = cluster
;
507 ret
= device_offline(get_cpu_device(i
));
509 bL_switcher_restore_cpus();
512 cpumask_set_cpu(i
, &bL_switcher_removed_logical_cpus
);
518 /* Determine the logical CPU a given physical CPU is grouped on. */
519 int bL_switcher_get_logical_index(u32 mpidr
)
523 if (!bL_switcher_active
)
526 mpidr
&= MPIDR_HWID_BITMASK
;
527 for_each_online_cpu(cpu
) {
528 int pairing
= bL_switcher_cpu_pairing
[cpu
];
531 if ((mpidr
== cpu_logical_map(cpu
)) ||
532 (mpidr
== cpu_logical_map(pairing
)))
538 static void bL_switcher_trace_trigger_cpu(void *__always_unused info
)
540 trace_cpu_migrate_current(ktime_get_real_ns(), read_mpidr());
543 int bL_switcher_trace_trigger(void)
549 bL_switcher_trace_trigger_cpu(NULL
);
550 ret
= smp_call_function(bL_switcher_trace_trigger_cpu
, NULL
, true);
556 EXPORT_SYMBOL_GPL(bL_switcher_trace_trigger
);
558 static int bL_switcher_enable(void)
562 mutex_lock(&bL_switcher_activation_lock
);
563 lock_device_hotplug();
564 if (bL_switcher_active
) {
565 unlock_device_hotplug();
566 mutex_unlock(&bL_switcher_activation_lock
);
570 pr_info("big.LITTLE switcher initializing\n");
572 ret
= bL_activation_notify(BL_NOTIFY_PRE_ENABLE
);
576 ret
= bL_switcher_halve_cpus();
580 bL_switcher_trace_trigger();
582 for_each_online_cpu(cpu
) {
583 struct bL_thread
*t
= &bL_threads
[cpu
];
584 spin_lock_init(&t
->lock
);
585 init_waitqueue_head(&t
->wq
);
586 init_completion(&t
->started
);
587 t
->wanted_cluster
= -1;
588 t
->task
= bL_switcher_thread_create(cpu
, t
);
591 bL_switcher_active
= 1;
592 bL_activation_notify(BL_NOTIFY_POST_ENABLE
);
593 pr_info("big.LITTLE switcher initialized\n");
597 pr_warn("big.LITTLE switcher initialization failed\n");
598 bL_activation_notify(BL_NOTIFY_POST_DISABLE
);
601 unlock_device_hotplug();
602 mutex_unlock(&bL_switcher_activation_lock
);
608 static void bL_switcher_disable(void)
610 unsigned int cpu
, cluster
;
612 struct task_struct
*task
;
614 mutex_lock(&bL_switcher_activation_lock
);
615 lock_device_hotplug();
617 if (!bL_switcher_active
)
620 if (bL_activation_notify(BL_NOTIFY_PRE_DISABLE
) != 0) {
621 bL_activation_notify(BL_NOTIFY_POST_ENABLE
);
625 bL_switcher_active
= 0;
628 * To deactivate the switcher, we must shut down the switcher
629 * threads to prevent any other requests from being accepted.
630 * Then, if the final cluster for given logical CPU is not the
631 * same as the original one, we'll recreate a switcher thread
632 * just for the purpose of switching the CPU back without any
633 * possibility for interference from external requests.
635 for_each_online_cpu(cpu
) {
636 t
= &bL_threads
[cpu
];
639 if (!task
|| IS_ERR(task
))
642 /* no more switch may happen on this CPU at this point */
643 cluster
= MPIDR_AFFINITY_LEVEL(cpu_logical_map(cpu
), 1);
644 if (cluster
== bL_switcher_cpu_original_cluster
[cpu
])
646 init_completion(&t
->started
);
647 t
->wanted_cluster
= bL_switcher_cpu_original_cluster
[cpu
];
648 task
= bL_switcher_thread_create(cpu
, t
);
650 wait_for_completion(&t
->started
);
652 cluster
= MPIDR_AFFINITY_LEVEL(cpu_logical_map(cpu
), 1);
653 if (cluster
== bL_switcher_cpu_original_cluster
[cpu
])
656 /* If execution gets here, we're in trouble. */
657 pr_crit("%s: unable to restore original cluster for CPU %d\n",
659 pr_crit("%s: CPU %d can't be restored\n",
660 __func__
, bL_switcher_cpu_pairing
[cpu
]);
661 cpumask_clear_cpu(bL_switcher_cpu_pairing
[cpu
],
662 &bL_switcher_removed_logical_cpus
);
665 bL_switcher_restore_cpus();
666 bL_switcher_trace_trigger();
668 bL_activation_notify(BL_NOTIFY_POST_DISABLE
);
671 unlock_device_hotplug();
672 mutex_unlock(&bL_switcher_activation_lock
);
675 static ssize_t
bL_switcher_active_show(struct kobject
*kobj
,
676 struct kobj_attribute
*attr
, char *buf
)
678 return sprintf(buf
, "%u\n", bL_switcher_active
);
681 static ssize_t
bL_switcher_active_store(struct kobject
*kobj
,
682 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
688 bL_switcher_disable();
692 ret
= bL_switcher_enable();
698 return (ret
>= 0) ? count
: ret
;
701 static ssize_t
bL_switcher_trace_trigger_store(struct kobject
*kobj
,
702 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
704 int ret
= bL_switcher_trace_trigger();
706 return ret
? ret
: count
;
709 static struct kobj_attribute bL_switcher_active_attr
=
710 __ATTR(active
, 0644, bL_switcher_active_show
, bL_switcher_active_store
);
712 static struct kobj_attribute bL_switcher_trace_trigger_attr
=
713 __ATTR(trace_trigger
, 0200, NULL
, bL_switcher_trace_trigger_store
);
715 static struct attribute
*bL_switcher_attrs
[] = {
716 &bL_switcher_active_attr
.attr
,
717 &bL_switcher_trace_trigger_attr
.attr
,
721 static struct attribute_group bL_switcher_attr_group
= {
722 .attrs
= bL_switcher_attrs
,
725 static struct kobject
*bL_switcher_kobj
;
727 static int __init
bL_switcher_sysfs_init(void)
731 bL_switcher_kobj
= kobject_create_and_add("bL_switcher", kernel_kobj
);
732 if (!bL_switcher_kobj
)
734 ret
= sysfs_create_group(bL_switcher_kobj
, &bL_switcher_attr_group
);
736 kobject_put(bL_switcher_kobj
);
740 #endif /* CONFIG_SYSFS */
742 bool bL_switcher_get_enabled(void)
744 mutex_lock(&bL_switcher_activation_lock
);
746 return bL_switcher_active
;
748 EXPORT_SYMBOL_GPL(bL_switcher_get_enabled
);
750 void bL_switcher_put_enabled(void)
752 mutex_unlock(&bL_switcher_activation_lock
);
754 EXPORT_SYMBOL_GPL(bL_switcher_put_enabled
);
757 * Veto any CPU hotplug operation on those CPUs we've removed
758 * while the switcher is active.
759 * We're just not ready to deal with that given the trickery involved.
761 static int bL_switcher_cpu_pre(unsigned int cpu
)
765 if (!bL_switcher_active
)
768 pairing
= bL_switcher_cpu_pairing
[cpu
];
775 static bool no_bL_switcher
;
776 core_param(no_bL_switcher
, no_bL_switcher
, bool, 0644);
778 static int __init
bL_switcher_init(void)
782 if (!mcpm_is_available())
785 cpuhp_setup_state_nocalls(CPUHP_ARM_BL_PREPARE
, "arm/bl:prepare",
786 bL_switcher_cpu_pre
, NULL
);
787 ret
= cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN
, "arm/bl:predown",
788 NULL
, bL_switcher_cpu_pre
);
790 cpuhp_remove_state_nocalls(CPUHP_ARM_BL_PREPARE
);
791 pr_err("bL_switcher: Failed to allocate a hotplug state\n");
794 if (!no_bL_switcher
) {
795 ret
= bL_switcher_enable();
801 ret
= bL_switcher_sysfs_init();
803 pr_err("%s: unable to create sysfs entry\n", __func__
);
809 late_initcall(bL_switcher_init
);