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 long volatile *handshake_ptr
;
157 this_cpu
= smp_processor_id();
158 ob_mpidr
= read_mpidr();
159 ob_cpu
= MPIDR_AFFINITY_LEVEL(ob_mpidr
, 0);
160 ob_cluster
= MPIDR_AFFINITY_LEVEL(ob_mpidr
, 1);
161 BUG_ON(cpu_logical_map(this_cpu
) != ob_mpidr
);
163 if (new_cluster_id
== ob_cluster
)
166 that_cpu
= bL_switcher_cpu_pairing
[this_cpu
];
167 ib_mpidr
= cpu_logical_map(that_cpu
);
168 ib_cpu
= MPIDR_AFFINITY_LEVEL(ib_mpidr
, 0);
169 ib_cluster
= MPIDR_AFFINITY_LEVEL(ib_mpidr
, 1);
171 pr_debug("before switch: CPU %d MPIDR %#x -> %#x\n",
172 this_cpu
, ob_mpidr
, ib_mpidr
);
174 this_cpu
= smp_processor_id();
176 /* Close the gate for our entry vectors */
177 mcpm_set_entry_vector(ob_cpu
, ob_cluster
, NULL
);
178 mcpm_set_entry_vector(ib_cpu
, ib_cluster
, NULL
);
180 /* Install our "inbound alive" notifier. */
181 init_completion(&inbound_alive
);
182 ipi_nr
= register_ipi_completion(&inbound_alive
, this_cpu
);
183 ipi_nr
|= ((1 << 16) << bL_gic_id
[ob_cpu
][ob_cluster
]);
184 mcpm_set_early_poke(ib_cpu
, ib_cluster
, gic_get_sgir_physaddr(), ipi_nr
);
187 * Let's wake up the inbound CPU now in case it requires some delay
188 * to come online, but leave it gated in our entry vector code.
190 ret
= mcpm_cpu_power_up(ib_cpu
, ib_cluster
);
192 pr_err("%s: mcpm_cpu_power_up() returned %d\n", __func__
, ret
);
197 * Raise a SGI on the inbound CPU to make sure it doesn't stall
198 * in a possible WFI, such as in bL_power_down().
200 gic_send_sgi(bL_gic_id
[ib_cpu
][ib_cluster
], 0);
203 * Wait for the inbound to come up. This allows for other
204 * tasks to be scheduled in the mean time.
206 wait_for_completion(&inbound_alive
);
207 mcpm_set_early_poke(ib_cpu
, ib_cluster
, 0, 0);
210 * From this point we are entering the switch critical zone
211 * and can't take any interrupts anymore.
215 trace_cpu_migrate_begin(ktime_get_real_ns(), ob_mpidr
);
217 /* redirect GIC's SGIs to our counterpart */
218 gic_migrate_target(bL_gic_id
[ib_cpu
][ib_cluster
]);
220 tick_suspend_local();
222 ret
= cpu_pm_enter();
224 /* we can not tolerate errors at this point */
226 panic("%s: cpu_pm_enter() returned %d\n", __func__
, ret
);
228 /* Swap the physical CPUs in the logical map for this logical CPU. */
229 cpu_logical_map(this_cpu
) = ib_mpidr
;
230 cpu_logical_map(that_cpu
) = ob_mpidr
;
232 /* Let's do the actual CPU switch. */
233 ret
= cpu_suspend((unsigned long)&handshake_ptr
, bL_switchpoint
);
235 panic("%s: cpu_suspend() returned %d\n", __func__
, ret
);
237 /* We are executing on the inbound CPU at this point */
238 mpidr
= read_mpidr();
239 pr_debug("after switch: CPU %d MPIDR %#x\n", this_cpu
, mpidr
);
240 BUG_ON(mpidr
!= ib_mpidr
);
242 mcpm_cpu_powered_up();
248 trace_cpu_migrate_finish(ktime_get_real_ns(), ib_mpidr
);
256 pr_err("%s exiting with error %d\n", __func__
, ret
);
262 struct task_struct
*task
;
263 wait_queue_head_t wq
;
265 struct completion started
;
266 bL_switch_completion_handler completer
;
267 void *completer_cookie
;
270 static struct bL_thread bL_threads
[NR_CPUS
];
272 static int bL_switcher_thread(void *arg
)
274 struct bL_thread
*t
= arg
;
275 struct sched_param param
= { .sched_priority
= 1 };
277 bL_switch_completion_handler completer
;
278 void *completer_cookie
;
280 sched_setscheduler_nocheck(current
, SCHED_FIFO
, ¶m
);
281 complete(&t
->started
);
284 if (signal_pending(current
))
285 flush_signals(current
);
286 wait_event_interruptible(t
->wq
,
287 t
->wanted_cluster
!= -1 ||
288 kthread_should_stop());
291 cluster
= t
->wanted_cluster
;
292 completer
= t
->completer
;
293 completer_cookie
= t
->completer_cookie
;
294 t
->wanted_cluster
= -1;
296 spin_unlock(&t
->lock
);
299 bL_switch_to(cluster
);
302 completer(completer_cookie
);
304 } while (!kthread_should_stop());
309 static struct task_struct
*bL_switcher_thread_create(int cpu
, void *arg
)
311 struct task_struct
*task
;
313 task
= kthread_create_on_node(bL_switcher_thread
, arg
,
314 cpu_to_node(cpu
), "kswitcher_%d", cpu
);
316 kthread_bind(task
, cpu
);
317 wake_up_process(task
);
319 pr_err("%s failed for CPU %d\n", __func__
, cpu
);
324 * bL_switch_request_cb - Switch to a specific cluster for the given CPU,
325 * with completion notification via a callback
327 * @cpu: the CPU to switch
328 * @new_cluster_id: the ID of the cluster to switch to.
329 * @completer: switch completion callback. if non-NULL,
330 * @completer(@completer_cookie) will be called on completion of
331 * the switch, in non-atomic context.
332 * @completer_cookie: opaque context argument for @completer.
334 * This function causes a cluster switch on the given CPU by waking up
335 * the appropriate switcher thread. This function may or may not return
336 * before the switch has occurred.
338 * If a @completer callback function is supplied, it will be called when
339 * the switch is complete. This can be used to determine asynchronously
340 * when the switch is complete, regardless of when bL_switch_request()
341 * returns. When @completer is supplied, no new switch request is permitted
342 * for the affected CPU until after the switch is complete, and @completer
345 int bL_switch_request_cb(unsigned int cpu
, unsigned int new_cluster_id
,
346 bL_switch_completion_handler completer
,
347 void *completer_cookie
)
351 if (cpu
>= ARRAY_SIZE(bL_threads
)) {
352 pr_err("%s: cpu %d out of bounds\n", __func__
, cpu
);
356 t
= &bL_threads
[cpu
];
359 return PTR_ERR(t
->task
);
365 spin_unlock(&t
->lock
);
368 t
->completer
= completer
;
369 t
->completer_cookie
= completer_cookie
;
370 t
->wanted_cluster
= new_cluster_id
;
371 spin_unlock(&t
->lock
);
375 EXPORT_SYMBOL_GPL(bL_switch_request_cb
);
378 * Activation and configuration code.
381 static DEFINE_MUTEX(bL_switcher_activation_lock
);
382 static BLOCKING_NOTIFIER_HEAD(bL_activation_notifier
);
383 static unsigned int bL_switcher_active
;
384 static unsigned int bL_switcher_cpu_original_cluster
[NR_CPUS
];
385 static cpumask_t bL_switcher_removed_logical_cpus
;
387 int bL_switcher_register_notifier(struct notifier_block
*nb
)
389 return blocking_notifier_chain_register(&bL_activation_notifier
, nb
);
391 EXPORT_SYMBOL_GPL(bL_switcher_register_notifier
);
393 int bL_switcher_unregister_notifier(struct notifier_block
*nb
)
395 return blocking_notifier_chain_unregister(&bL_activation_notifier
, nb
);
397 EXPORT_SYMBOL_GPL(bL_switcher_unregister_notifier
);
399 static int bL_activation_notify(unsigned long val
)
403 ret
= blocking_notifier_call_chain(&bL_activation_notifier
, val
, NULL
);
404 if (ret
& NOTIFY_STOP_MASK
)
405 pr_err("%s: notifier chain failed with status 0x%x\n",
407 return notifier_to_errno(ret
);
410 static void bL_switcher_restore_cpus(void)
414 for_each_cpu(i
, &bL_switcher_removed_logical_cpus
) {
415 struct device
*cpu_dev
= get_cpu_device(i
);
416 int ret
= device_online(cpu_dev
);
418 dev_err(cpu_dev
, "switcher: unable to restore CPU\n");
422 static int bL_switcher_halve_cpus(void)
424 int i
, j
, cluster_0
, gic_id
, ret
;
425 unsigned int cpu
, cluster
, mask
;
426 cpumask_t available_cpus
;
428 /* First pass to validate what we have */
430 for_each_online_cpu(i
) {
431 cpu
= MPIDR_AFFINITY_LEVEL(cpu_logical_map(i
), 0);
432 cluster
= MPIDR_AFFINITY_LEVEL(cpu_logical_map(i
), 1);
434 pr_err("%s: only dual cluster systems are supported\n", __func__
);
437 if (WARN_ON(cpu
>= MAX_CPUS_PER_CLUSTER
))
439 mask
|= (1 << cluster
);
442 pr_err("%s: no CPU pairing possible\n", __func__
);
447 * Now let's do the pairing. We match each CPU with another CPU
448 * from a different cluster. To get a uniform scheduling behavior
449 * without fiddling with CPU topology and compute capacity data,
450 * we'll use logical CPUs initially belonging to the same cluster.
452 memset(bL_switcher_cpu_pairing
, -1, sizeof(bL_switcher_cpu_pairing
));
453 cpumask_copy(&available_cpus
, cpu_online_mask
);
455 for_each_cpu(i
, &available_cpus
) {
457 cluster
= MPIDR_AFFINITY_LEVEL(cpu_logical_map(i
), 1);
460 if (cluster
!= cluster_0
)
462 cpumask_clear_cpu(i
, &available_cpus
);
463 for_each_cpu(j
, &available_cpus
) {
464 cluster
= MPIDR_AFFINITY_LEVEL(cpu_logical_map(j
), 1);
466 * Let's remember the last match to create "odd"
467 * pairings on purpose in order for other code not
468 * to assume any relation between physical and
469 * logical CPU numbers.
471 if (cluster
!= cluster_0
)
475 bL_switcher_cpu_pairing
[i
] = match
;
476 cpumask_clear_cpu(match
, &available_cpus
);
477 pr_info("CPU%d paired with CPU%d\n", i
, match
);
482 * Now we disable the unwanted CPUs i.e. everything that has no
483 * pairing information (that includes the pairing counterparts).
485 cpumask_clear(&bL_switcher_removed_logical_cpus
);
486 for_each_online_cpu(i
) {
487 cpu
= MPIDR_AFFINITY_LEVEL(cpu_logical_map(i
), 0);
488 cluster
= MPIDR_AFFINITY_LEVEL(cpu_logical_map(i
), 1);
490 /* Let's take note of the GIC ID for this CPU */
491 gic_id
= gic_get_cpu_id(i
);
493 pr_err("%s: bad GIC ID for CPU %d\n", __func__
, i
);
494 bL_switcher_restore_cpus();
497 bL_gic_id
[cpu
][cluster
] = gic_id
;
498 pr_info("GIC ID for CPU %u cluster %u is %u\n",
499 cpu
, cluster
, gic_id
);
501 if (bL_switcher_cpu_pairing
[i
] != -1) {
502 bL_switcher_cpu_original_cluster
[i
] = cluster
;
506 ret
= device_offline(get_cpu_device(i
));
508 bL_switcher_restore_cpus();
511 cpumask_set_cpu(i
, &bL_switcher_removed_logical_cpus
);
517 /* Determine the logical CPU a given physical CPU is grouped on. */
518 int bL_switcher_get_logical_index(u32 mpidr
)
522 if (!bL_switcher_active
)
525 mpidr
&= MPIDR_HWID_BITMASK
;
526 for_each_online_cpu(cpu
) {
527 int pairing
= bL_switcher_cpu_pairing
[cpu
];
530 if ((mpidr
== cpu_logical_map(cpu
)) ||
531 (mpidr
== cpu_logical_map(pairing
)))
537 static void bL_switcher_trace_trigger_cpu(void *__always_unused info
)
539 trace_cpu_migrate_current(ktime_get_real_ns(), read_mpidr());
542 int bL_switcher_trace_trigger(void)
548 bL_switcher_trace_trigger_cpu(NULL
);
549 ret
= smp_call_function(bL_switcher_trace_trigger_cpu
, NULL
, true);
555 EXPORT_SYMBOL_GPL(bL_switcher_trace_trigger
);
557 static int bL_switcher_enable(void)
561 mutex_lock(&bL_switcher_activation_lock
);
562 lock_device_hotplug();
563 if (bL_switcher_active
) {
564 unlock_device_hotplug();
565 mutex_unlock(&bL_switcher_activation_lock
);
569 pr_info("big.LITTLE switcher initializing\n");
571 ret
= bL_activation_notify(BL_NOTIFY_PRE_ENABLE
);
575 ret
= bL_switcher_halve_cpus();
579 bL_switcher_trace_trigger();
581 for_each_online_cpu(cpu
) {
582 struct bL_thread
*t
= &bL_threads
[cpu
];
583 spin_lock_init(&t
->lock
);
584 init_waitqueue_head(&t
->wq
);
585 init_completion(&t
->started
);
586 t
->wanted_cluster
= -1;
587 t
->task
= bL_switcher_thread_create(cpu
, t
);
590 bL_switcher_active
= 1;
591 bL_activation_notify(BL_NOTIFY_POST_ENABLE
);
592 pr_info("big.LITTLE switcher initialized\n");
596 pr_warn("big.LITTLE switcher initialization failed\n");
597 bL_activation_notify(BL_NOTIFY_POST_DISABLE
);
600 unlock_device_hotplug();
601 mutex_unlock(&bL_switcher_activation_lock
);
607 static void bL_switcher_disable(void)
609 unsigned int cpu
, cluster
;
611 struct task_struct
*task
;
613 mutex_lock(&bL_switcher_activation_lock
);
614 lock_device_hotplug();
616 if (!bL_switcher_active
)
619 if (bL_activation_notify(BL_NOTIFY_PRE_DISABLE
) != 0) {
620 bL_activation_notify(BL_NOTIFY_POST_ENABLE
);
624 bL_switcher_active
= 0;
627 * To deactivate the switcher, we must shut down the switcher
628 * threads to prevent any other requests from being accepted.
629 * Then, if the final cluster for given logical CPU is not the
630 * same as the original one, we'll recreate a switcher thread
631 * just for the purpose of switching the CPU back without any
632 * possibility for interference from external requests.
634 for_each_online_cpu(cpu
) {
635 t
= &bL_threads
[cpu
];
638 if (!task
|| IS_ERR(task
))
641 /* no more switch may happen on this CPU at this point */
642 cluster
= MPIDR_AFFINITY_LEVEL(cpu_logical_map(cpu
), 1);
643 if (cluster
== bL_switcher_cpu_original_cluster
[cpu
])
645 init_completion(&t
->started
);
646 t
->wanted_cluster
= bL_switcher_cpu_original_cluster
[cpu
];
647 task
= bL_switcher_thread_create(cpu
, t
);
649 wait_for_completion(&t
->started
);
651 cluster
= MPIDR_AFFINITY_LEVEL(cpu_logical_map(cpu
), 1);
652 if (cluster
== bL_switcher_cpu_original_cluster
[cpu
])
655 /* If execution gets here, we're in trouble. */
656 pr_crit("%s: unable to restore original cluster for CPU %d\n",
658 pr_crit("%s: CPU %d can't be restored\n",
659 __func__
, bL_switcher_cpu_pairing
[cpu
]);
660 cpumask_clear_cpu(bL_switcher_cpu_pairing
[cpu
],
661 &bL_switcher_removed_logical_cpus
);
664 bL_switcher_restore_cpus();
665 bL_switcher_trace_trigger();
667 bL_activation_notify(BL_NOTIFY_POST_DISABLE
);
670 unlock_device_hotplug();
671 mutex_unlock(&bL_switcher_activation_lock
);
674 static ssize_t
bL_switcher_active_show(struct kobject
*kobj
,
675 struct kobj_attribute
*attr
, char *buf
)
677 return sprintf(buf
, "%u\n", bL_switcher_active
);
680 static ssize_t
bL_switcher_active_store(struct kobject
*kobj
,
681 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
687 bL_switcher_disable();
691 ret
= bL_switcher_enable();
697 return (ret
>= 0) ? count
: ret
;
700 static ssize_t
bL_switcher_trace_trigger_store(struct kobject
*kobj
,
701 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
703 int ret
= bL_switcher_trace_trigger();
705 return ret
? ret
: count
;
708 static struct kobj_attribute bL_switcher_active_attr
=
709 __ATTR(active
, 0644, bL_switcher_active_show
, bL_switcher_active_store
);
711 static struct kobj_attribute bL_switcher_trace_trigger_attr
=
712 __ATTR(trace_trigger
, 0200, NULL
, bL_switcher_trace_trigger_store
);
714 static struct attribute
*bL_switcher_attrs
[] = {
715 &bL_switcher_active_attr
.attr
,
716 &bL_switcher_trace_trigger_attr
.attr
,
720 static struct attribute_group bL_switcher_attr_group
= {
721 .attrs
= bL_switcher_attrs
,
724 static struct kobject
*bL_switcher_kobj
;
726 static int __init
bL_switcher_sysfs_init(void)
730 bL_switcher_kobj
= kobject_create_and_add("bL_switcher", kernel_kobj
);
731 if (!bL_switcher_kobj
)
733 ret
= sysfs_create_group(bL_switcher_kobj
, &bL_switcher_attr_group
);
735 kobject_put(bL_switcher_kobj
);
739 #endif /* CONFIG_SYSFS */
741 bool bL_switcher_get_enabled(void)
743 mutex_lock(&bL_switcher_activation_lock
);
745 return bL_switcher_active
;
747 EXPORT_SYMBOL_GPL(bL_switcher_get_enabled
);
749 void bL_switcher_put_enabled(void)
751 mutex_unlock(&bL_switcher_activation_lock
);
753 EXPORT_SYMBOL_GPL(bL_switcher_put_enabled
);
756 * Veto any CPU hotplug operation on those CPUs we've removed
757 * while the switcher is active.
758 * We're just not ready to deal with that given the trickery involved.
760 static int bL_switcher_hotplug_callback(struct notifier_block
*nfb
,
761 unsigned long action
, void *hcpu
)
763 if (bL_switcher_active
) {
764 int pairing
= bL_switcher_cpu_pairing
[(unsigned long)hcpu
];
765 switch (action
& 0xf) {
767 case CPU_DOWN_PREPARE
:
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 cpu_notifier(bL_switcher_hotplug_callback
, 0);
787 if (!no_bL_switcher
) {
788 ret
= bL_switcher_enable();
794 ret
= bL_switcher_sysfs_init();
796 pr_err("%s: unable to create sysfs entry\n", __func__
);
802 late_initcall(bL_switcher_init
);