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 *_arg
)
60 unsigned ib_mpidr
, ib_cpu
, ib_cluster
;
61 long volatile handshake
, **handshake_ptr
= _arg
;
63 pr_debug("%s\n", __func__
);
65 ib_mpidr
= cpu_logical_map(smp_processor_id());
66 ib_cpu
= MPIDR_AFFINITY_LEVEL(ib_mpidr
, 0);
67 ib_cluster
= MPIDR_AFFINITY_LEVEL(ib_mpidr
, 1);
69 /* Advertise our handshake location */
72 *handshake_ptr
= &handshake
;
77 * Our state has been saved at this point. Let's release our
80 mcpm_set_entry_vector(ib_cpu
, ib_cluster
, cpu_resume
);
84 * From this point, we must assume that our counterpart CPU might
85 * have taken over in its parallel world already, as if execution
86 * just returned from cpu_suspend(). It is therefore important to
87 * be very careful not to make any change the other guy is not
88 * expecting. This is why we need stack isolation.
90 * Fancy under cover tasks could be performed here. For now
95 * Let's wait until our inbound is alive.
102 /* Let's put ourself down. */
103 mcpm_cpu_power_down();
105 /* should never get here */
110 * Stack isolation. To ensure 'current' remains valid, we just use another
111 * piece of our thread's stack space which should be fairly lightly used.
112 * The selected area starts just above the thread_info structure located
113 * at the very bottom of the stack, aligned to a cache line, and indexed
114 * with the cluster number.
116 #define STACK_SIZE 512
117 extern void call_with_stack(void (*fn
)(void *), void *arg
, void *sp
);
118 static int bL_switchpoint(unsigned long _arg
)
120 unsigned int mpidr
= read_mpidr();
121 unsigned int clusterid
= MPIDR_AFFINITY_LEVEL(mpidr
, 1);
122 void *stack
= current_thread_info() + 1;
123 stack
= PTR_ALIGN(stack
, L1_CACHE_BYTES
);
124 stack
+= clusterid
* STACK_SIZE
+ STACK_SIZE
;
125 call_with_stack(bL_do_switch
, (void *)_arg
, stack
);
130 * Generic switcher interface
133 static unsigned int bL_gic_id
[MAX_CPUS_PER_CLUSTER
][MAX_NR_CLUSTERS
];
134 static int bL_switcher_cpu_pairing
[NR_CPUS
];
137 * bL_switch_to - Switch to a specific cluster for the current CPU
138 * @new_cluster_id: the ID of the cluster to switch to.
140 * This function must be called on the CPU to be switched.
141 * Returns 0 on success, else a negative status code.
143 static int bL_switch_to(unsigned int new_cluster_id
)
145 unsigned int mpidr
, this_cpu
, that_cpu
;
146 unsigned int ob_mpidr
, ob_cpu
, ob_cluster
, ib_mpidr
, ib_cpu
, ib_cluster
;
147 struct completion inbound_alive
;
148 struct tick_device
*tdev
;
149 enum clock_event_mode tdev_mode
;
150 long volatile *handshake_ptr
;
153 this_cpu
= smp_processor_id();
154 ob_mpidr
= read_mpidr();
155 ob_cpu
= MPIDR_AFFINITY_LEVEL(ob_mpidr
, 0);
156 ob_cluster
= MPIDR_AFFINITY_LEVEL(ob_mpidr
, 1);
157 BUG_ON(cpu_logical_map(this_cpu
) != ob_mpidr
);
159 if (new_cluster_id
== ob_cluster
)
162 that_cpu
= bL_switcher_cpu_pairing
[this_cpu
];
163 ib_mpidr
= cpu_logical_map(that_cpu
);
164 ib_cpu
= MPIDR_AFFINITY_LEVEL(ib_mpidr
, 0);
165 ib_cluster
= MPIDR_AFFINITY_LEVEL(ib_mpidr
, 1);
167 pr_debug("before switch: CPU %d MPIDR %#x -> %#x\n",
168 this_cpu
, ob_mpidr
, ib_mpidr
);
170 this_cpu
= smp_processor_id();
172 /* Close the gate for our entry vectors */
173 mcpm_set_entry_vector(ob_cpu
, ob_cluster
, NULL
);
174 mcpm_set_entry_vector(ib_cpu
, ib_cluster
, NULL
);
176 /* Install our "inbound alive" notifier. */
177 init_completion(&inbound_alive
);
178 ipi_nr
= register_ipi_completion(&inbound_alive
, this_cpu
);
179 ipi_nr
|= ((1 << 16) << bL_gic_id
[ob_cpu
][ob_cluster
]);
180 mcpm_set_early_poke(ib_cpu
, ib_cluster
, gic_get_sgir_physaddr(), ipi_nr
);
183 * Let's wake up the inbound CPU now in case it requires some delay
184 * to come online, but leave it gated in our entry vector code.
186 ret
= mcpm_cpu_power_up(ib_cpu
, ib_cluster
);
188 pr_err("%s: mcpm_cpu_power_up() returned %d\n", __func__
, ret
);
193 * Raise a SGI on the inbound CPU to make sure it doesn't stall
194 * in a possible WFI, such as in bL_power_down().
196 gic_send_sgi(bL_gic_id
[ib_cpu
][ib_cluster
], 0);
199 * Wait for the inbound to come up. This allows for other
200 * tasks to be scheduled in the mean time.
202 wait_for_completion(&inbound_alive
);
203 mcpm_set_early_poke(ib_cpu
, ib_cluster
, 0, 0);
206 * From this point we are entering the switch critical zone
207 * and can't take any interrupts anymore.
212 /* redirect GIC's SGIs to our counterpart */
213 gic_migrate_target(bL_gic_id
[ib_cpu
][ib_cluster
]);
215 tdev
= tick_get_device(this_cpu
);
216 if (tdev
&& !cpumask_equal(tdev
->evtdev
->cpumask
, cpumask_of(this_cpu
)))
219 tdev_mode
= tdev
->evtdev
->mode
;
220 clockevents_set_mode(tdev
->evtdev
, CLOCK_EVT_MODE_SHUTDOWN
);
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();
248 clockevents_set_mode(tdev
->evtdev
, tdev_mode
);
249 clockevents_program_event(tdev
->evtdev
,
250 tdev
->evtdev
->next_event
, 1);
260 pr_err("%s exiting with error %d\n", __func__
, ret
);
266 struct task_struct
*task
;
267 wait_queue_head_t wq
;
269 struct completion started
;
270 bL_switch_completion_handler completer
;
271 void *completer_cookie
;
274 static struct bL_thread bL_threads
[NR_CPUS
];
276 static int bL_switcher_thread(void *arg
)
278 struct bL_thread
*t
= arg
;
279 struct sched_param param
= { .sched_priority
= 1 };
281 bL_switch_completion_handler completer
;
282 void *completer_cookie
;
284 sched_setscheduler_nocheck(current
, SCHED_FIFO
, ¶m
);
285 complete(&t
->started
);
288 if (signal_pending(current
))
289 flush_signals(current
);
290 wait_event_interruptible(t
->wq
,
291 t
->wanted_cluster
!= -1 ||
292 kthread_should_stop());
295 cluster
= t
->wanted_cluster
;
296 completer
= t
->completer
;
297 completer_cookie
= t
->completer_cookie
;
298 t
->wanted_cluster
= -1;
300 spin_unlock(&t
->lock
);
303 bL_switch_to(cluster
);
306 completer(completer_cookie
);
308 } while (!kthread_should_stop());
313 static struct task_struct
*bL_switcher_thread_create(int cpu
, void *arg
)
315 struct task_struct
*task
;
317 task
= kthread_create_on_node(bL_switcher_thread
, arg
,
318 cpu_to_node(cpu
), "kswitcher_%d", cpu
);
320 kthread_bind(task
, cpu
);
321 wake_up_process(task
);
323 pr_err("%s failed for CPU %d\n", __func__
, cpu
);
328 * bL_switch_request_cb - Switch to a specific cluster for the given CPU,
329 * with completion notification via a callback
331 * @cpu: the CPU to switch
332 * @new_cluster_id: the ID of the cluster to switch to.
333 * @completer: switch completion callback. if non-NULL,
334 * @completer(@completer_cookie) will be called on completion of
335 * the switch, in non-atomic context.
336 * @completer_cookie: opaque context argument for @completer.
338 * This function causes a cluster switch on the given CPU by waking up
339 * the appropriate switcher thread. This function may or may not return
340 * before the switch has occurred.
342 * If a @completer callback function is supplied, it will be called when
343 * the switch is complete. This can be used to determine asynchronously
344 * when the switch is complete, regardless of when bL_switch_request()
345 * returns. When @completer is supplied, no new switch request is permitted
346 * for the affected CPU until after the switch is complete, and @completer
349 int bL_switch_request_cb(unsigned int cpu
, unsigned int new_cluster_id
,
350 bL_switch_completion_handler completer
,
351 void *completer_cookie
)
355 if (cpu
>= ARRAY_SIZE(bL_threads
)) {
356 pr_err("%s: cpu %d out of bounds\n", __func__
, cpu
);
360 t
= &bL_threads
[cpu
];
363 return PTR_ERR(t
->task
);
369 spin_unlock(&t
->lock
);
372 t
->completer
= completer
;
373 t
->completer_cookie
= completer_cookie
;
374 t
->wanted_cluster
= new_cluster_id
;
375 spin_unlock(&t
->lock
);
379 EXPORT_SYMBOL_GPL(bL_switch_request_cb
);
382 * Activation and configuration code.
385 static DEFINE_MUTEX(bL_switcher_activation_lock
);
386 static BLOCKING_NOTIFIER_HEAD(bL_activation_notifier
);
387 static unsigned int bL_switcher_active
;
388 static unsigned int bL_switcher_cpu_original_cluster
[NR_CPUS
];
389 static cpumask_t bL_switcher_removed_logical_cpus
;
391 int bL_switcher_register_notifier(struct notifier_block
*nb
)
393 return blocking_notifier_chain_register(&bL_activation_notifier
, nb
);
395 EXPORT_SYMBOL_GPL(bL_switcher_register_notifier
);
397 int bL_switcher_unregister_notifier(struct notifier_block
*nb
)
399 return blocking_notifier_chain_unregister(&bL_activation_notifier
, nb
);
401 EXPORT_SYMBOL_GPL(bL_switcher_unregister_notifier
);
403 static int bL_activation_notify(unsigned long val
)
407 ret
= blocking_notifier_call_chain(&bL_activation_notifier
, val
, NULL
);
408 if (ret
& NOTIFY_STOP_MASK
)
409 pr_err("%s: notifier chain failed with status 0x%x\n",
411 return notifier_to_errno(ret
);
414 static void bL_switcher_restore_cpus(void)
418 for_each_cpu(i
, &bL_switcher_removed_logical_cpus
)
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
;
508 bL_switcher_restore_cpus();
511 cpumask_set_cpu(i
, &bL_switcher_removed_logical_cpus
);
517 static int bL_switcher_enable(void)
521 mutex_lock(&bL_switcher_activation_lock
);
522 cpu_hotplug_driver_lock();
523 if (bL_switcher_active
) {
524 cpu_hotplug_driver_unlock();
525 mutex_unlock(&bL_switcher_activation_lock
);
529 pr_info("big.LITTLE switcher initializing\n");
531 ret
= bL_activation_notify(BL_NOTIFY_PRE_ENABLE
);
535 ret
= bL_switcher_halve_cpus();
539 for_each_online_cpu(cpu
) {
540 struct bL_thread
*t
= &bL_threads
[cpu
];
541 spin_lock_init(&t
->lock
);
542 init_waitqueue_head(&t
->wq
);
543 init_completion(&t
->started
);
544 t
->wanted_cluster
= -1;
545 t
->task
= bL_switcher_thread_create(cpu
, t
);
548 bL_switcher_active
= 1;
549 bL_activation_notify(BL_NOTIFY_POST_ENABLE
);
550 pr_info("big.LITTLE switcher initialized\n");
554 pr_warn("big.LITTLE switcher initialization failed\n");
555 bL_activation_notify(BL_NOTIFY_POST_DISABLE
);
558 cpu_hotplug_driver_unlock();
559 mutex_unlock(&bL_switcher_activation_lock
);
565 static void bL_switcher_disable(void)
567 unsigned int cpu
, cluster
;
569 struct task_struct
*task
;
571 mutex_lock(&bL_switcher_activation_lock
);
572 cpu_hotplug_driver_lock();
574 if (!bL_switcher_active
)
577 if (bL_activation_notify(BL_NOTIFY_PRE_DISABLE
) != 0) {
578 bL_activation_notify(BL_NOTIFY_POST_ENABLE
);
582 bL_switcher_active
= 0;
585 * To deactivate the switcher, we must shut down the switcher
586 * threads to prevent any other requests from being accepted.
587 * Then, if the final cluster for given logical CPU is not the
588 * same as the original one, we'll recreate a switcher thread
589 * just for the purpose of switching the CPU back without any
590 * possibility for interference from external requests.
592 for_each_online_cpu(cpu
) {
593 t
= &bL_threads
[cpu
];
596 if (!task
|| IS_ERR(task
))
599 /* no more switch may happen on this CPU at this point */
600 cluster
= MPIDR_AFFINITY_LEVEL(cpu_logical_map(cpu
), 1);
601 if (cluster
== bL_switcher_cpu_original_cluster
[cpu
])
603 init_completion(&t
->started
);
604 t
->wanted_cluster
= bL_switcher_cpu_original_cluster
[cpu
];
605 task
= bL_switcher_thread_create(cpu
, t
);
607 wait_for_completion(&t
->started
);
609 cluster
= MPIDR_AFFINITY_LEVEL(cpu_logical_map(cpu
), 1);
610 if (cluster
== bL_switcher_cpu_original_cluster
[cpu
])
613 /* If execution gets here, we're in trouble. */
614 pr_crit("%s: unable to restore original cluster for CPU %d\n",
616 pr_crit("%s: CPU %d can't be restored\n",
617 __func__
, bL_switcher_cpu_pairing
[cpu
]);
618 cpumask_clear_cpu(bL_switcher_cpu_pairing
[cpu
],
619 &bL_switcher_removed_logical_cpus
);
622 bL_switcher_restore_cpus();
623 bL_activation_notify(BL_NOTIFY_POST_DISABLE
);
626 cpu_hotplug_driver_unlock();
627 mutex_unlock(&bL_switcher_activation_lock
);
630 static ssize_t
bL_switcher_active_show(struct kobject
*kobj
,
631 struct kobj_attribute
*attr
, char *buf
)
633 return sprintf(buf
, "%u\n", bL_switcher_active
);
636 static ssize_t
bL_switcher_active_store(struct kobject
*kobj
,
637 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
643 bL_switcher_disable();
647 ret
= bL_switcher_enable();
653 return (ret
>= 0) ? count
: ret
;
656 static struct kobj_attribute bL_switcher_active_attr
=
657 __ATTR(active
, 0644, bL_switcher_active_show
, bL_switcher_active_store
);
659 static struct attribute
*bL_switcher_attrs
[] = {
660 &bL_switcher_active_attr
.attr
,
664 static struct attribute_group bL_switcher_attr_group
= {
665 .attrs
= bL_switcher_attrs
,
668 static struct kobject
*bL_switcher_kobj
;
670 static int __init
bL_switcher_sysfs_init(void)
674 bL_switcher_kobj
= kobject_create_and_add("bL_switcher", kernel_kobj
);
675 if (!bL_switcher_kobj
)
677 ret
= sysfs_create_group(bL_switcher_kobj
, &bL_switcher_attr_group
);
679 kobject_put(bL_switcher_kobj
);
683 #endif /* CONFIG_SYSFS */
685 bool bL_switcher_get_enabled(void)
687 mutex_lock(&bL_switcher_activation_lock
);
689 return bL_switcher_active
;
691 EXPORT_SYMBOL_GPL(bL_switcher_get_enabled
);
693 void bL_switcher_put_enabled(void)
695 mutex_unlock(&bL_switcher_activation_lock
);
697 EXPORT_SYMBOL_GPL(bL_switcher_put_enabled
);
700 * Veto any CPU hotplug operation on those CPUs we've removed
701 * while the switcher is active.
702 * We're just not ready to deal with that given the trickery involved.
704 static int bL_switcher_hotplug_callback(struct notifier_block
*nfb
,
705 unsigned long action
, void *hcpu
)
707 if (bL_switcher_active
) {
708 int pairing
= bL_switcher_cpu_pairing
[(unsigned long)hcpu
];
709 switch (action
& 0xf) {
711 case CPU_DOWN_PREPARE
:
719 static bool no_bL_switcher
;
720 core_param(no_bL_switcher
, no_bL_switcher
, bool, 0644);
722 static int __init
bL_switcher_init(void)
726 if (MAX_NR_CLUSTERS
!= 2) {
727 pr_err("%s: only dual cluster systems are supported\n", __func__
);
731 cpu_notifier(bL_switcher_hotplug_callback
, 0);
733 if (!no_bL_switcher
) {
734 ret
= bL_switcher_enable();
740 ret
= bL_switcher_sysfs_init();
742 pr_err("%s: unable to create sysfs entry\n", __func__
);
748 late_initcall(bL_switcher_init
);