2 * coupled.c - helper functions to enter the same idle state on multiple cpus
4 * Copyright (c) 2011 Google, Inc.
6 * Author: Colin Cross <ccross@android.com>
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License as published by
10 * the Free Software Foundation; either version 2 of the License, or
11 * (at your option) any later version.
13 * This program is distributed in the hope that it will be useful, but WITHOUT
14 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
19 #include <linux/kernel.h>
20 #include <linux/cpu.h>
21 #include <linux/cpuidle.h>
22 #include <linux/mutex.h>
23 #include <linux/sched.h>
24 #include <linux/slab.h>
25 #include <linux/spinlock.h>
30 * DOC: Coupled cpuidle states
32 * On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
33 * cpus cannot be independently powered down, either due to
34 * sequencing restrictions (on Tegra 2, cpu 0 must be the last to
35 * power down), or due to HW bugs (on OMAP4460, a cpu powering up
36 * will corrupt the gic state unless the other cpu runs a work
37 * around). Each cpu has a power state that it can enter without
38 * coordinating with the other cpu (usually Wait For Interrupt, or
39 * WFI), and one or more "coupled" power states that affect blocks
40 * shared between the cpus (L2 cache, interrupt controller, and
41 * sometimes the whole SoC). Entering a coupled power state must
42 * be tightly controlled on both cpus.
44 * This file implements a solution, where each cpu will wait in the
45 * WFI state until all cpus are ready to enter a coupled state, at
46 * which point the coupled state function will be called on all
47 * cpus at approximately the same time.
49 * Once all cpus are ready to enter idle, they are woken by an smp
50 * cross call. At this point, there is a chance that one of the
51 * cpus will find work to do, and choose not to enter idle. A
52 * final pass is needed to guarantee that all cpus will call the
53 * power state enter function at the same time. During this pass,
54 * each cpu will increment the ready counter, and continue once the
55 * ready counter matches the number of online coupled cpus. If any
56 * cpu exits idle, the other cpus will decrement their counter and
59 * requested_state stores the deepest coupled idle state each cpu
60 * is ready for. It is assumed that the states are indexed from
61 * shallowest (highest power, lowest exit latency) to deepest
62 * (lowest power, highest exit latency). The requested_state
63 * variable is not locked. It is only written from the cpu that
64 * it stores (or by the on/offlining cpu if that cpu is offline),
65 * and only read after all the cpus are ready for the coupled idle
66 * state are are no longer updating it.
68 * Three atomic counters are used. alive_count tracks the number
69 * of cpus in the coupled set that are currently or soon will be
70 * online. waiting_count tracks the number of cpus that are in
71 * the waiting loop, in the ready loop, or in the coupled idle state.
72 * ready_count tracks the number of cpus that are in the ready loop
73 * or in the coupled idle state.
75 * To use coupled cpuidle states, a cpuidle driver must:
77 * Set struct cpuidle_device.coupled_cpus to the mask of all
78 * coupled cpus, usually the same as cpu_possible_mask if all cpus
79 * are part of the same cluster. The coupled_cpus mask must be
80 * set in the struct cpuidle_device for each cpu.
82 * Set struct cpuidle_device.safe_state to a state that is not a
83 * coupled state. This is usually WFI.
85 * Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
86 * state that affects multiple cpus.
88 * Provide a struct cpuidle_state.enter function for each state
89 * that affects multiple cpus. This function is guaranteed to be
90 * called on all cpus at approximately the same time. The driver
91 * should ensure that the cpus all abort together if any cpu tries
92 * to abort once the function is called. The function should return
93 * with interrupts still disabled.
97 * struct cpuidle_coupled - data for set of cpus that share a coupled idle state
98 * @coupled_cpus: mask of cpus that are part of the coupled set
99 * @requested_state: array of requested states for cpus in the coupled set
100 * @ready_waiting_counts: combined count of cpus in ready or waiting loops
101 * @online_count: count of cpus that are online
102 * @refcnt: reference count of cpuidle devices that are using this struct
103 * @prevent: flag to prevent coupled idle while a cpu is hotplugging
105 struct cpuidle_coupled
{
106 cpumask_t coupled_cpus
;
107 int requested_state
[NR_CPUS
];
108 atomic_t ready_waiting_counts
;
114 #define WAITING_BITS 16
115 #define MAX_WAITING_CPUS (1 << WAITING_BITS)
116 #define WAITING_MASK (MAX_WAITING_CPUS - 1)
117 #define READY_MASK (~WAITING_MASK)
119 #define CPUIDLE_COUPLED_NOT_IDLE (-1)
121 static DEFINE_MUTEX(cpuidle_coupled_lock
);
122 static DEFINE_PER_CPU(struct call_single_data
, cpuidle_coupled_poke_cb
);
125 * The cpuidle_coupled_poked_mask mask is used to avoid calling
126 * __smp_call_function_single with the per cpu call_single_data struct already
127 * in use. This prevents a deadlock where two cpus are waiting for each others
128 * call_single_data struct to be available
130 static cpumask_t cpuidle_coupled_poked_mask
;
133 * cpuidle_coupled_parallel_barrier - synchronize all online coupled cpus
134 * @dev: cpuidle_device of the calling cpu
135 * @a: atomic variable to hold the barrier
137 * No caller to this function will return from this function until all online
138 * cpus in the same coupled group have called this function. Once any caller
139 * has returned from this function, the barrier is immediately available for
142 * The atomic variable a must be initialized to 0 before any cpu calls
143 * this function, will be reset to 0 before any cpu returns from this function.
145 * Must only be called from within a coupled idle state handler
146 * (state.enter when state.flags has CPUIDLE_FLAG_COUPLED set).
148 * Provides full smp barrier semantics before and after calling.
150 void cpuidle_coupled_parallel_barrier(struct cpuidle_device
*dev
, atomic_t
*a
)
152 int n
= dev
->coupled
->online_count
;
154 smp_mb__before_atomic_inc();
157 while (atomic_read(a
) < n
)
160 if (atomic_inc_return(a
) == n
* 2) {
165 while (atomic_read(a
) > n
)
170 * cpuidle_state_is_coupled - check if a state is part of a coupled set
171 * @dev: struct cpuidle_device for the current cpu
172 * @drv: struct cpuidle_driver for the platform
173 * @state: index of the target state in drv->states
175 * Returns true if the target state is coupled with cpus besides this one
177 bool cpuidle_state_is_coupled(struct cpuidle_device
*dev
,
178 struct cpuidle_driver
*drv
, int state
)
180 return drv
->states
[state
].flags
& CPUIDLE_FLAG_COUPLED
;
184 * cpuidle_coupled_set_ready - mark a cpu as ready
185 * @coupled: the struct coupled that contains the current cpu
187 static inline void cpuidle_coupled_set_ready(struct cpuidle_coupled
*coupled
)
189 atomic_add(MAX_WAITING_CPUS
, &coupled
->ready_waiting_counts
);
193 * cpuidle_coupled_set_not_ready - mark a cpu as not ready
194 * @coupled: the struct coupled that contains the current cpu
196 * Decrements the ready counter, unless the ready (and thus the waiting) counter
197 * is equal to the number of online cpus. Prevents a race where one cpu
198 * decrements the waiting counter and then re-increments it just before another
199 * cpu has decremented its ready counter, leading to the ready counter going
200 * down from the number of online cpus without going through the coupled idle
203 * Returns 0 if the counter was decremented successfully, -EINVAL if the ready
204 * counter was equal to the number of online cpus.
207 inline int cpuidle_coupled_set_not_ready(struct cpuidle_coupled
*coupled
)
212 all
= coupled
->online_count
|| (coupled
->online_count
<< WAITING_BITS
);
213 ret
= atomic_add_unless(&coupled
->ready_waiting_counts
,
214 -MAX_WAITING_CPUS
, all
);
216 return ret
? 0 : -EINVAL
;
220 * cpuidle_coupled_no_cpus_ready - check if no cpus in a coupled set are ready
221 * @coupled: the struct coupled that contains the current cpu
223 * Returns true if all of the cpus in a coupled set are out of the ready loop.
225 static inline int cpuidle_coupled_no_cpus_ready(struct cpuidle_coupled
*coupled
)
227 int r
= atomic_read(&coupled
->ready_waiting_counts
) >> WAITING_BITS
;
232 * cpuidle_coupled_cpus_ready - check if all cpus in a coupled set are ready
233 * @coupled: the struct coupled that contains the current cpu
235 * Returns true if all cpus coupled to this target state are in the ready loop
237 static inline bool cpuidle_coupled_cpus_ready(struct cpuidle_coupled
*coupled
)
239 int r
= atomic_read(&coupled
->ready_waiting_counts
) >> WAITING_BITS
;
240 return r
== coupled
->online_count
;
244 * cpuidle_coupled_cpus_waiting - check if all cpus in a coupled set are waiting
245 * @coupled: the struct coupled that contains the current cpu
247 * Returns true if all cpus coupled to this target state are in the wait loop
249 static inline bool cpuidle_coupled_cpus_waiting(struct cpuidle_coupled
*coupled
)
251 int w
= atomic_read(&coupled
->ready_waiting_counts
) & WAITING_MASK
;
252 return w
== coupled
->online_count
;
256 * cpuidle_coupled_no_cpus_waiting - check if no cpus in coupled set are waiting
257 * @coupled: the struct coupled that contains the current cpu
259 * Returns true if all of the cpus in a coupled set are out of the waiting loop.
261 static inline int cpuidle_coupled_no_cpus_waiting(struct cpuidle_coupled
*coupled
)
263 int w
= atomic_read(&coupled
->ready_waiting_counts
) & WAITING_MASK
;
268 * cpuidle_coupled_get_state - determine the deepest idle state
269 * @dev: struct cpuidle_device for this cpu
270 * @coupled: the struct coupled that contains the current cpu
272 * Returns the deepest idle state that all coupled cpus can enter
274 static inline int cpuidle_coupled_get_state(struct cpuidle_device
*dev
,
275 struct cpuidle_coupled
*coupled
)
281 * Read barrier ensures that read of requested_state is ordered after
282 * reads of ready_count. Matches the write barriers
283 * cpuidle_set_state_waiting.
287 for_each_cpu_mask(i
, coupled
->coupled_cpus
)
288 if (cpu_online(i
) && coupled
->requested_state
[i
] < state
)
289 state
= coupled
->requested_state
[i
];
294 static void cpuidle_coupled_poked(void *info
)
296 int cpu
= (unsigned long)info
;
297 cpumask_clear_cpu(cpu
, &cpuidle_coupled_poked_mask
);
301 * cpuidle_coupled_poke - wake up a cpu that may be waiting
304 * Ensures that the target cpu exits it's waiting idle state (if it is in it)
305 * and will see updates to waiting_count before it re-enters it's waiting idle
308 * If cpuidle_coupled_poked_mask is already set for the target cpu, that cpu
309 * either has or will soon have a pending IPI that will wake it out of idle,
310 * or it is currently processing the IPI and is not in idle.
312 static void cpuidle_coupled_poke(int cpu
)
314 struct call_single_data
*csd
= &per_cpu(cpuidle_coupled_poke_cb
, cpu
);
316 if (!cpumask_test_and_set_cpu(cpu
, &cpuidle_coupled_poked_mask
))
317 __smp_call_function_single(cpu
, csd
, 0);
321 * cpuidle_coupled_poke_others - wake up all other cpus that may be waiting
322 * @dev: struct cpuidle_device for this cpu
323 * @coupled: the struct coupled that contains the current cpu
325 * Calls cpuidle_coupled_poke on all other online cpus.
327 static void cpuidle_coupled_poke_others(int this_cpu
,
328 struct cpuidle_coupled
*coupled
)
332 for_each_cpu_mask(cpu
, coupled
->coupled_cpus
)
333 if (cpu
!= this_cpu
&& cpu_online(cpu
))
334 cpuidle_coupled_poke(cpu
);
338 * cpuidle_coupled_set_waiting - mark this cpu as in the wait loop
339 * @dev: struct cpuidle_device for this cpu
340 * @coupled: the struct coupled that contains the current cpu
341 * @next_state: the index in drv->states of the requested state for this cpu
343 * Updates the requested idle state for the specified cpuidle device,
344 * poking all coupled cpus out of idle if necessary to let them see the new
347 static void cpuidle_coupled_set_waiting(int cpu
,
348 struct cpuidle_coupled
*coupled
, int next_state
)
352 coupled
->requested_state
[cpu
] = next_state
;
355 * If this is the last cpu to enter the waiting state, poke
356 * all the other cpus out of their waiting state so they can
357 * enter a deeper state. This can race with one of the cpus
358 * exiting the waiting state due to an interrupt and
359 * decrementing waiting_count, see comment below.
361 * The atomic_inc_return provides a write barrier to order the write
362 * to requested_state with the later write that increments ready_count.
364 w
= atomic_inc_return(&coupled
->ready_waiting_counts
) & WAITING_MASK
;
365 if (w
== coupled
->online_count
)
366 cpuidle_coupled_poke_others(cpu
, coupled
);
370 * cpuidle_coupled_set_not_waiting - mark this cpu as leaving the wait loop
371 * @dev: struct cpuidle_device for this cpu
372 * @coupled: the struct coupled that contains the current cpu
374 * Removes the requested idle state for the specified cpuidle device.
376 static void cpuidle_coupled_set_not_waiting(int cpu
,
377 struct cpuidle_coupled
*coupled
)
380 * Decrementing waiting count can race with incrementing it in
381 * cpuidle_coupled_set_waiting, but that's OK. Worst case, some
382 * cpus will increment ready_count and then spin until they
383 * notice that this cpu has cleared it's requested_state.
385 atomic_dec(&coupled
->ready_waiting_counts
);
387 coupled
->requested_state
[cpu
] = CPUIDLE_COUPLED_NOT_IDLE
;
391 * cpuidle_coupled_set_done - mark this cpu as leaving the ready loop
392 * @cpu: the current cpu
393 * @coupled: the struct coupled that contains the current cpu
395 * Marks this cpu as no longer in the ready and waiting loops. Decrements
396 * the waiting count first to prevent another cpu looping back in and seeing
397 * this cpu as waiting just before it exits idle.
399 static void cpuidle_coupled_set_done(int cpu
, struct cpuidle_coupled
*coupled
)
401 cpuidle_coupled_set_not_waiting(cpu
, coupled
);
402 atomic_sub(MAX_WAITING_CPUS
, &coupled
->ready_waiting_counts
);
406 * cpuidle_coupled_clear_pokes - spin until the poke interrupt is processed
409 * Turns on interrupts and spins until any outstanding poke interrupts have
410 * been processed and the poke bit has been cleared.
412 * Other interrupts may also be processed while interrupts are enabled, so
413 * need_resched() must be tested after turning interrupts off again to make sure
414 * the interrupt didn't schedule work that should take the cpu out of idle.
416 * Returns 0 if need_resched was false, -EINTR if need_resched was true.
418 static int cpuidle_coupled_clear_pokes(int cpu
)
421 while (cpumask_test_cpu(cpu
, &cpuidle_coupled_poked_mask
))
425 return need_resched() ? -EINTR
: 0;
429 * cpuidle_enter_state_coupled - attempt to enter a state with coupled cpus
430 * @dev: struct cpuidle_device for the current cpu
431 * @drv: struct cpuidle_driver for the platform
432 * @next_state: index of the requested state in drv->states
434 * Coordinate with coupled cpus to enter the target state. This is a two
435 * stage process. In the first stage, the cpus are operating independently,
436 * and may call into cpuidle_enter_state_coupled at completely different times.
437 * To save as much power as possible, the first cpus to call this function will
438 * go to an intermediate state (the cpuidle_device's safe state), and wait for
439 * all the other cpus to call this function. Once all coupled cpus are idle,
440 * the second stage will start. Each coupled cpu will spin until all cpus have
441 * guaranteed that they will call the target_state.
443 * This function must be called with interrupts disabled. It may enable
444 * interrupts while preparing for idle, and it will always return with
445 * interrupts enabled.
447 int cpuidle_enter_state_coupled(struct cpuidle_device
*dev
,
448 struct cpuidle_driver
*drv
, int next_state
)
450 int entered_state
= -1;
451 struct cpuidle_coupled
*coupled
= dev
->coupled
;
456 while (coupled
->prevent
) {
457 if (cpuidle_coupled_clear_pokes(dev
->cpu
)) {
459 return entered_state
;
461 entered_state
= cpuidle_enter_state(dev
, drv
,
462 dev
->safe_state_index
);
465 /* Read barrier ensures online_count is read after prevent is cleared */
468 cpuidle_coupled_set_waiting(dev
->cpu
, coupled
, next_state
);
472 * Wait for all coupled cpus to be idle, using the deepest state
473 * allowed for a single cpu.
475 while (!cpuidle_coupled_cpus_waiting(coupled
)) {
476 if (cpuidle_coupled_clear_pokes(dev
->cpu
)) {
477 cpuidle_coupled_set_not_waiting(dev
->cpu
, coupled
);
481 if (coupled
->prevent
) {
482 cpuidle_coupled_set_not_waiting(dev
->cpu
, coupled
);
486 entered_state
= cpuidle_enter_state(dev
, drv
,
487 dev
->safe_state_index
);
490 if (cpuidle_coupled_clear_pokes(dev
->cpu
)) {
491 cpuidle_coupled_set_not_waiting(dev
->cpu
, coupled
);
496 * All coupled cpus are probably idle. There is a small chance that
497 * one of the other cpus just became active. Increment the ready count,
498 * and spin until all coupled cpus have incremented the counter. Once a
499 * cpu has incremented the ready counter, it cannot abort idle and must
500 * spin until either all cpus have incremented the ready counter, or
501 * another cpu leaves idle and decrements the waiting counter.
504 cpuidle_coupled_set_ready(coupled
);
505 while (!cpuidle_coupled_cpus_ready(coupled
)) {
506 /* Check if any other cpus bailed out of idle. */
507 if (!cpuidle_coupled_cpus_waiting(coupled
))
508 if (!cpuidle_coupled_set_not_ready(coupled
))
514 /* all cpus have acked the coupled state */
515 next_state
= cpuidle_coupled_get_state(dev
, coupled
);
517 entered_state
= cpuidle_enter_state(dev
, drv
, next_state
);
519 cpuidle_coupled_set_done(dev
->cpu
, coupled
);
523 * Normal cpuidle states are expected to return with irqs enabled.
524 * That leads to an inefficiency where a cpu receiving an interrupt
525 * that brings it out of idle will process that interrupt before
526 * exiting the idle enter function and decrementing ready_count. All
527 * other cpus will need to spin waiting for the cpu that is processing
528 * the interrupt. If the driver returns with interrupts disabled,
529 * all other cpus will loop back into the safe idle state instead of
530 * spinning, saving power.
532 * Calling local_irq_enable here allows coupled states to return with
533 * interrupts disabled, but won't cause problems for drivers that
534 * exit with interrupts enabled.
539 * Wait until all coupled cpus have exited idle. There is no risk that
540 * a cpu exits and re-enters the ready state because this cpu has
541 * already decremented its waiting_count.
543 while (!cpuidle_coupled_no_cpus_ready(coupled
))
546 return entered_state
;
549 static void cpuidle_coupled_update_online_cpus(struct cpuidle_coupled
*coupled
)
552 cpumask_and(&cpus
, cpu_online_mask
, &coupled
->coupled_cpus
);
553 coupled
->online_count
= cpumask_weight(&cpus
);
557 * cpuidle_coupled_register_device - register a coupled cpuidle device
558 * @dev: struct cpuidle_device for the current cpu
560 * Called from cpuidle_register_device to handle coupled idle init. Finds the
561 * cpuidle_coupled struct for this set of coupled cpus, or creates one if none
564 int cpuidle_coupled_register_device(struct cpuidle_device
*dev
)
567 struct cpuidle_device
*other_dev
;
568 struct call_single_data
*csd
;
569 struct cpuidle_coupled
*coupled
;
571 if (cpumask_empty(&dev
->coupled_cpus
))
574 for_each_cpu_mask(cpu
, dev
->coupled_cpus
) {
575 other_dev
= per_cpu(cpuidle_devices
, cpu
);
576 if (other_dev
&& other_dev
->coupled
) {
577 coupled
= other_dev
->coupled
;
582 /* No existing coupled info found, create a new one */
583 coupled
= kzalloc(sizeof(struct cpuidle_coupled
), GFP_KERNEL
);
587 coupled
->coupled_cpus
= dev
->coupled_cpus
;
590 dev
->coupled
= coupled
;
591 if (WARN_ON(!cpumask_equal(&dev
->coupled_cpus
, &coupled
->coupled_cpus
)))
594 cpuidle_coupled_update_online_cpus(coupled
);
598 csd
= &per_cpu(cpuidle_coupled_poke_cb
, dev
->cpu
);
599 csd
->func
= cpuidle_coupled_poked
;
600 csd
->info
= (void *)(unsigned long)dev
->cpu
;
606 * cpuidle_coupled_unregister_device - unregister a coupled cpuidle device
607 * @dev: struct cpuidle_device for the current cpu
609 * Called from cpuidle_unregister_device to tear down coupled idle. Removes the
610 * cpu from the coupled idle set, and frees the cpuidle_coupled_info struct if
611 * this was the last cpu in the set.
613 void cpuidle_coupled_unregister_device(struct cpuidle_device
*dev
)
615 struct cpuidle_coupled
*coupled
= dev
->coupled
;
617 if (cpumask_empty(&dev
->coupled_cpus
))
620 if (--coupled
->refcnt
)
626 * cpuidle_coupled_prevent_idle - prevent cpus from entering a coupled state
627 * @coupled: the struct coupled that contains the cpu that is changing state
629 * Disables coupled cpuidle on a coupled set of cpus. Used to ensure that
630 * cpu_online_mask doesn't change while cpus are coordinating coupled idle.
632 static void cpuidle_coupled_prevent_idle(struct cpuidle_coupled
*coupled
)
636 /* Force all cpus out of the waiting loop. */
638 cpuidle_coupled_poke_others(cpu
, coupled
);
640 while (!cpuidle_coupled_no_cpus_waiting(coupled
))
645 * cpuidle_coupled_allow_idle - allows cpus to enter a coupled state
646 * @coupled: the struct coupled that contains the cpu that is changing state
648 * Enables coupled cpuidle on a coupled set of cpus. Used to ensure that
649 * cpu_online_mask doesn't change while cpus are coordinating coupled idle.
651 static void cpuidle_coupled_allow_idle(struct cpuidle_coupled
*coupled
)
656 * Write barrier ensures readers see the new online_count when they
661 /* Force cpus out of the prevent loop. */
662 cpuidle_coupled_poke_others(cpu
, coupled
);
667 * cpuidle_coupled_cpu_notify - notifier called during hotplug transitions
668 * @nb: notifier block
669 * @action: hotplug transition
670 * @hcpu: target cpu number
672 * Called when a cpu is brought on or offline using hotplug. Updates the
673 * coupled cpu set appropriately
675 static int cpuidle_coupled_cpu_notify(struct notifier_block
*nb
,
676 unsigned long action
, void *hcpu
)
678 int cpu
= (unsigned long)hcpu
;
679 struct cpuidle_device
*dev
;
681 mutex_lock(&cpuidle_lock
);
683 dev
= per_cpu(cpuidle_devices
, cpu
);
687 switch (action
& ~CPU_TASKS_FROZEN
) {
689 case CPU_DOWN_PREPARE
:
690 cpuidle_coupled_prevent_idle(dev
->coupled
);
694 cpuidle_coupled_update_online_cpus(dev
->coupled
);
696 case CPU_UP_CANCELED
:
697 case CPU_DOWN_FAILED
:
698 cpuidle_coupled_allow_idle(dev
->coupled
);
703 mutex_unlock(&cpuidle_lock
);
707 static struct notifier_block cpuidle_coupled_cpu_notifier
= {
708 .notifier_call
= cpuidle_coupled_cpu_notify
,
711 static int __init
cpuidle_coupled_init(void)
713 return register_cpu_notifier(&cpuidle_coupled_cpu_notifier
);
715 core_initcall(cpuidle_coupled_init
);