blob: 344058f8501a2c2ee888189950b79f615e815a02 [file] [log] [blame]
Colin Cross4126c012012-05-07 17:57:41 -07001/*
2 * coupled.c - helper functions to enter the same idle state on multiple cpus
3 *
4 * Copyright (c) 2011 Google, Inc.
5 *
6 * Author: Colin Cross <ccross@android.com>
7 *
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.
12 *
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
16 * more details.
17 */
18
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>
26
27#include "cpuidle.h"
28
29/**
30 * DOC: Coupled cpuidle states
31 *
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.
43 *
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.
48 *
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
57 * retry.
58 *
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.
67 *
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.
74 *
75 * To use coupled cpuidle states, a cpuidle driver must:
76 *
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.
81 *
82 * Set struct cpuidle_device.safe_state to a state that is not a
83 * coupled state. This is usually WFI.
84 *
85 * Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
86 * state that affects multiple cpus.
87 *
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.
94 */
95
96/**
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
104 */
105struct cpuidle_coupled {
106 cpumask_t coupled_cpus;
107 int requested_state[NR_CPUS];
108 atomic_t ready_waiting_counts;
Colin Crossf9838272013-08-28 18:41:47 -0700109 atomic_t abort_barrier;
Colin Cross4126c012012-05-07 17:57:41 -0700110 int online_count;
111 int refcnt;
112 int prevent;
113};
114
115#define WAITING_BITS 16
116#define MAX_WAITING_CPUS (1 << WAITING_BITS)
117#define WAITING_MASK (MAX_WAITING_CPUS - 1)
118#define READY_MASK (~WAITING_MASK)
119
120#define CPUIDLE_COUPLED_NOT_IDLE (-1)
121
122static DEFINE_MUTEX(cpuidle_coupled_lock);
123static DEFINE_PER_CPU(struct call_single_data, cpuidle_coupled_poke_cb);
124
125/*
Colin Crossf9838272013-08-28 18:41:47 -0700126 * The cpuidle_coupled_poke_pending mask is used to avoid calling
Colin Cross4126c012012-05-07 17:57:41 -0700127 * __smp_call_function_single with the per cpu call_single_data struct already
128 * in use. This prevents a deadlock where two cpus are waiting for each others
129 * call_single_data struct to be available
130 */
Colin Crossf9838272013-08-28 18:41:47 -0700131static cpumask_t cpuidle_coupled_poke_pending;
132
133/*
134 * The cpuidle_coupled_poked mask is used to ensure that each cpu has been poked
135 * once to minimize entering the ready loop with a poke pending, which would
136 * require aborting and retrying.
137 */
138static cpumask_t cpuidle_coupled_poked;
Colin Cross4126c012012-05-07 17:57:41 -0700139
140/**
Colin Cross20ff51a2012-05-07 17:57:42 -0700141 * cpuidle_coupled_parallel_barrier - synchronize all online coupled cpus
142 * @dev: cpuidle_device of the calling cpu
143 * @a: atomic variable to hold the barrier
144 *
145 * No caller to this function will return from this function until all online
146 * cpus in the same coupled group have called this function. Once any caller
147 * has returned from this function, the barrier is immediately available for
148 * reuse.
149 *
Viresh Kumarcaf4a362013-10-03 21:26:41 +0530150 * The atomic variable must be initialized to 0 before any cpu calls
Colin Cross20ff51a2012-05-07 17:57:42 -0700151 * this function, will be reset to 0 before any cpu returns from this function.
152 *
153 * Must only be called from within a coupled idle state handler
154 * (state.enter when state.flags has CPUIDLE_FLAG_COUPLED set).
155 *
156 * Provides full smp barrier semantics before and after calling.
157 */
158void cpuidle_coupled_parallel_barrier(struct cpuidle_device *dev, atomic_t *a)
159{
160 int n = dev->coupled->online_count;
161
Peter Zijlstra4e857c52014-03-17 18:06:10 +0100162 smp_mb__before_atomic();
Colin Cross20ff51a2012-05-07 17:57:42 -0700163 atomic_inc(a);
164
165 while (atomic_read(a) < n)
166 cpu_relax();
167
168 if (atomic_inc_return(a) == n * 2) {
169 atomic_set(a, 0);
170 return;
171 }
172
173 while (atomic_read(a) > n)
174 cpu_relax();
175}
176
177/**
Colin Cross4126c012012-05-07 17:57:41 -0700178 * cpuidle_state_is_coupled - check if a state is part of a coupled set
Colin Cross4126c012012-05-07 17:57:41 -0700179 * @drv: struct cpuidle_driver for the platform
180 * @state: index of the target state in drv->states
181 *
182 * Returns true if the target state is coupled with cpus besides this one
183 */
Xunlei Pang4c1ed5a2015-08-04 13:48:56 +0800184bool cpuidle_state_is_coupled(struct cpuidle_driver *drv, int state)
Colin Cross4126c012012-05-07 17:57:41 -0700185{
186 return drv->states[state].flags & CPUIDLE_FLAG_COUPLED;
187}
188
189/**
Xunlei Pangabceaa92015-08-31 11:34:05 +0800190 * cpuidle_coupled_state_verify - check if the coupled states are correctly set.
191 * @drv: struct cpuidle_driver for the platform
192 *
193 * Returns 0 for valid state values, a negative error code otherwise:
194 * * -EINVAL if any coupled state(safe_state_index) is wrongly set.
195 */
196int cpuidle_coupled_state_verify(struct cpuidle_driver *drv)
197{
198 int i;
199
200 for (i = drv->state_count - 1; i >= 0; i--) {
201 if (cpuidle_state_is_coupled(drv, i) &&
202 (drv->safe_state_index == i ||
203 drv->safe_state_index < 0 ||
204 drv->safe_state_index >= drv->state_count))
205 return -EINVAL;
206 }
207
208 return 0;
209}
210
211/**
Colin Cross4126c012012-05-07 17:57:41 -0700212 * cpuidle_coupled_set_ready - mark a cpu as ready
213 * @coupled: the struct coupled that contains the current cpu
214 */
215static inline void cpuidle_coupled_set_ready(struct cpuidle_coupled *coupled)
216{
217 atomic_add(MAX_WAITING_CPUS, &coupled->ready_waiting_counts);
218}
219
220/**
221 * cpuidle_coupled_set_not_ready - mark a cpu as not ready
222 * @coupled: the struct coupled that contains the current cpu
223 *
224 * Decrements the ready counter, unless the ready (and thus the waiting) counter
225 * is equal to the number of online cpus. Prevents a race where one cpu
226 * decrements the waiting counter and then re-increments it just before another
227 * cpu has decremented its ready counter, leading to the ready counter going
228 * down from the number of online cpus without going through the coupled idle
229 * state.
230 *
231 * Returns 0 if the counter was decremented successfully, -EINVAL if the ready
232 * counter was equal to the number of online cpus.
233 */
234static
235inline int cpuidle_coupled_set_not_ready(struct cpuidle_coupled *coupled)
236{
237 int all;
238 int ret;
239
Sivaram Nair92638e22012-12-18 13:52:54 +0100240 all = coupled->online_count | (coupled->online_count << WAITING_BITS);
Colin Cross4126c012012-05-07 17:57:41 -0700241 ret = atomic_add_unless(&coupled->ready_waiting_counts,
242 -MAX_WAITING_CPUS, all);
243
244 return ret ? 0 : -EINVAL;
245}
246
247/**
248 * cpuidle_coupled_no_cpus_ready - check if no cpus in a coupled set are ready
249 * @coupled: the struct coupled that contains the current cpu
250 *
251 * Returns true if all of the cpus in a coupled set are out of the ready loop.
252 */
253static inline int cpuidle_coupled_no_cpus_ready(struct cpuidle_coupled *coupled)
254{
255 int r = atomic_read(&coupled->ready_waiting_counts) >> WAITING_BITS;
256 return r == 0;
257}
258
259/**
260 * cpuidle_coupled_cpus_ready - check if all cpus in a coupled set are ready
261 * @coupled: the struct coupled that contains the current cpu
262 *
263 * Returns true if all cpus coupled to this target state are in the ready loop
264 */
265static inline bool cpuidle_coupled_cpus_ready(struct cpuidle_coupled *coupled)
266{
267 int r = atomic_read(&coupled->ready_waiting_counts) >> WAITING_BITS;
268 return r == coupled->online_count;
269}
270
271/**
272 * cpuidle_coupled_cpus_waiting - check if all cpus in a coupled set are waiting
273 * @coupled: the struct coupled that contains the current cpu
274 *
275 * Returns true if all cpus coupled to this target state are in the wait loop
276 */
277static inline bool cpuidle_coupled_cpus_waiting(struct cpuidle_coupled *coupled)
278{
279 int w = atomic_read(&coupled->ready_waiting_counts) & WAITING_MASK;
280 return w == coupled->online_count;
281}
282
283/**
284 * cpuidle_coupled_no_cpus_waiting - check if no cpus in coupled set are waiting
285 * @coupled: the struct coupled that contains the current cpu
286 *
287 * Returns true if all of the cpus in a coupled set are out of the waiting loop.
288 */
289static inline int cpuidle_coupled_no_cpus_waiting(struct cpuidle_coupled *coupled)
290{
291 int w = atomic_read(&coupled->ready_waiting_counts) & WAITING_MASK;
292 return w == 0;
293}
294
295/**
296 * cpuidle_coupled_get_state - determine the deepest idle state
297 * @dev: struct cpuidle_device for this cpu
298 * @coupled: the struct coupled that contains the current cpu
299 *
300 * Returns the deepest idle state that all coupled cpus can enter
301 */
302static inline int cpuidle_coupled_get_state(struct cpuidle_device *dev,
303 struct cpuidle_coupled *coupled)
304{
305 int i;
306 int state = INT_MAX;
307
308 /*
309 * Read barrier ensures that read of requested_state is ordered after
310 * reads of ready_count. Matches the write barriers
311 * cpuidle_set_state_waiting.
312 */
313 smp_rmb();
314
Rusty Russellf9b531f2015-03-05 10:49:16 +1030315 for_each_cpu(i, &coupled->coupled_cpus)
Colin Cross4126c012012-05-07 17:57:41 -0700316 if (cpu_online(i) && coupled->requested_state[i] < state)
317 state = coupled->requested_state[i];
318
319 return state;
320}
321
Colin Crossf9838272013-08-28 18:41:47 -0700322static void cpuidle_coupled_handle_poke(void *info)
Colin Cross4126c012012-05-07 17:57:41 -0700323{
324 int cpu = (unsigned long)info;
Colin Crossf9838272013-08-28 18:41:47 -0700325 cpumask_set_cpu(cpu, &cpuidle_coupled_poked);
326 cpumask_clear_cpu(cpu, &cpuidle_coupled_poke_pending);
Colin Cross4126c012012-05-07 17:57:41 -0700327}
328
329/**
330 * cpuidle_coupled_poke - wake up a cpu that may be waiting
331 * @cpu: target cpu
332 *
333 * Ensures that the target cpu exits it's waiting idle state (if it is in it)
334 * and will see updates to waiting_count before it re-enters it's waiting idle
335 * state.
336 *
337 * If cpuidle_coupled_poked_mask is already set for the target cpu, that cpu
338 * either has or will soon have a pending IPI that will wake it out of idle,
339 * or it is currently processing the IPI and is not in idle.
340 */
341static void cpuidle_coupled_poke(int cpu)
342{
343 struct call_single_data *csd = &per_cpu(cpuidle_coupled_poke_cb, cpu);
344
Colin Crossf9838272013-08-28 18:41:47 -0700345 if (!cpumask_test_and_set_cpu(cpu, &cpuidle_coupled_poke_pending))
Frederic Weisbeckerc46fff22014-02-24 16:40:02 +0100346 smp_call_function_single_async(cpu, csd);
Colin Cross4126c012012-05-07 17:57:41 -0700347}
348
349/**
350 * cpuidle_coupled_poke_others - wake up all other cpus that may be waiting
351 * @dev: struct cpuidle_device for this cpu
352 * @coupled: the struct coupled that contains the current cpu
353 *
354 * Calls cpuidle_coupled_poke on all other online cpus.
355 */
356static void cpuidle_coupled_poke_others(int this_cpu,
357 struct cpuidle_coupled *coupled)
358{
359 int cpu;
360
Rusty Russellf9b531f2015-03-05 10:49:16 +1030361 for_each_cpu(cpu, &coupled->coupled_cpus)
Colin Cross4126c012012-05-07 17:57:41 -0700362 if (cpu != this_cpu && cpu_online(cpu))
363 cpuidle_coupled_poke(cpu);
364}
365
366/**
367 * cpuidle_coupled_set_waiting - mark this cpu as in the wait loop
368 * @dev: struct cpuidle_device for this cpu
369 * @coupled: the struct coupled that contains the current cpu
370 * @next_state: the index in drv->states of the requested state for this cpu
371 *
Colin Crossf9838272013-08-28 18:41:47 -0700372 * Updates the requested idle state for the specified cpuidle device.
373 * Returns the number of waiting cpus.
Colin Cross4126c012012-05-07 17:57:41 -0700374 */
Colin Crossf9838272013-08-28 18:41:47 -0700375static int cpuidle_coupled_set_waiting(int cpu,
Colin Cross4126c012012-05-07 17:57:41 -0700376 struct cpuidle_coupled *coupled, int next_state)
377{
Colin Cross4126c012012-05-07 17:57:41 -0700378 coupled->requested_state[cpu] = next_state;
379
380 /*
Colin Cross4126c012012-05-07 17:57:41 -0700381 * The atomic_inc_return provides a write barrier to order the write
382 * to requested_state with the later write that increments ready_count.
383 */
Colin Crossf9838272013-08-28 18:41:47 -0700384 return atomic_inc_return(&coupled->ready_waiting_counts) & WAITING_MASK;
Colin Cross4126c012012-05-07 17:57:41 -0700385}
386
387/**
388 * cpuidle_coupled_set_not_waiting - mark this cpu as leaving the wait loop
389 * @dev: struct cpuidle_device for this cpu
390 * @coupled: the struct coupled that contains the current cpu
391 *
392 * Removes the requested idle state for the specified cpuidle device.
393 */
394static void cpuidle_coupled_set_not_waiting(int cpu,
395 struct cpuidle_coupled *coupled)
396{
397 /*
398 * Decrementing waiting count can race with incrementing it in
399 * cpuidle_coupled_set_waiting, but that's OK. Worst case, some
400 * cpus will increment ready_count and then spin until they
401 * notice that this cpu has cleared it's requested_state.
402 */
403 atomic_dec(&coupled->ready_waiting_counts);
404
405 coupled->requested_state[cpu] = CPUIDLE_COUPLED_NOT_IDLE;
406}
407
408/**
409 * cpuidle_coupled_set_done - mark this cpu as leaving the ready loop
410 * @cpu: the current cpu
411 * @coupled: the struct coupled that contains the current cpu
412 *
413 * Marks this cpu as no longer in the ready and waiting loops. Decrements
414 * the waiting count first to prevent another cpu looping back in and seeing
415 * this cpu as waiting just before it exits idle.
416 */
417static void cpuidle_coupled_set_done(int cpu, struct cpuidle_coupled *coupled)
418{
419 cpuidle_coupled_set_not_waiting(cpu, coupled);
420 atomic_sub(MAX_WAITING_CPUS, &coupled->ready_waiting_counts);
421}
422
423/**
424 * cpuidle_coupled_clear_pokes - spin until the poke interrupt is processed
425 * @cpu - this cpu
426 *
427 * Turns on interrupts and spins until any outstanding poke interrupts have
428 * been processed and the poke bit has been cleared.
429 *
430 * Other interrupts may also be processed while interrupts are enabled, so
Colin Cross9e19b732013-08-23 12:45:12 -0700431 * need_resched() must be tested after this function returns to make sure
Colin Cross4126c012012-05-07 17:57:41 -0700432 * the interrupt didn't schedule work that should take the cpu out of idle.
433 *
Colin Cross9e19b732013-08-23 12:45:12 -0700434 * Returns 0 if no poke was pending, 1 if a poke was cleared.
Colin Cross4126c012012-05-07 17:57:41 -0700435 */
436static int cpuidle_coupled_clear_pokes(int cpu)
437{
Colin Cross9e19b732013-08-23 12:45:12 -0700438 if (!cpumask_test_cpu(cpu, &cpuidle_coupled_poke_pending))
439 return 0;
440
Colin Cross4126c012012-05-07 17:57:41 -0700441 local_irq_enable();
Colin Crossf9838272013-08-28 18:41:47 -0700442 while (cpumask_test_cpu(cpu, &cpuidle_coupled_poke_pending))
Colin Cross4126c012012-05-07 17:57:41 -0700443 cpu_relax();
444 local_irq_disable();
445
Colin Cross9e19b732013-08-23 12:45:12 -0700446 return 1;
Colin Cross4126c012012-05-07 17:57:41 -0700447}
448
Colin Crossf9838272013-08-28 18:41:47 -0700449static bool cpuidle_coupled_any_pokes_pending(struct cpuidle_coupled *coupled)
450{
451 cpumask_t cpus;
452 int ret;
453
454 cpumask_and(&cpus, cpu_online_mask, &coupled->coupled_cpus);
455 ret = cpumask_and(&cpus, &cpuidle_coupled_poke_pending, &cpus);
456
457 return ret;
458}
459
Colin Cross4126c012012-05-07 17:57:41 -0700460/**
461 * cpuidle_enter_state_coupled - attempt to enter a state with coupled cpus
462 * @dev: struct cpuidle_device for the current cpu
463 * @drv: struct cpuidle_driver for the platform
464 * @next_state: index of the requested state in drv->states
465 *
466 * Coordinate with coupled cpus to enter the target state. This is a two
467 * stage process. In the first stage, the cpus are operating independently,
468 * and may call into cpuidle_enter_state_coupled at completely different times.
469 * To save as much power as possible, the first cpus to call this function will
470 * go to an intermediate state (the cpuidle_device's safe state), and wait for
471 * all the other cpus to call this function. Once all coupled cpus are idle,
472 * the second stage will start. Each coupled cpu will spin until all cpus have
473 * guaranteed that they will call the target_state.
474 *
475 * This function must be called with interrupts disabled. It may enable
476 * interrupts while preparing for idle, and it will always return with
477 * interrupts enabled.
478 */
479int cpuidle_enter_state_coupled(struct cpuidle_device *dev,
480 struct cpuidle_driver *drv, int next_state)
481{
482 int entered_state = -1;
483 struct cpuidle_coupled *coupled = dev->coupled;
Colin Crossf9838272013-08-28 18:41:47 -0700484 int w;
Colin Cross4126c012012-05-07 17:57:41 -0700485
486 if (!coupled)
487 return -EINVAL;
488
489 while (coupled->prevent) {
Colin Cross9e19b732013-08-23 12:45:12 -0700490 cpuidle_coupled_clear_pokes(dev->cpu);
491 if (need_resched()) {
Colin Cross4126c012012-05-07 17:57:41 -0700492 local_irq_enable();
493 return entered_state;
494 }
495 entered_state = cpuidle_enter_state(dev, drv,
Xunlei Pangba6a8602015-08-04 13:48:55 +0800496 drv->safe_state_index);
Colin Cross59e99852013-08-23 12:45:10 -0700497 local_irq_disable();
Colin Cross4126c012012-05-07 17:57:41 -0700498 }
499
500 /* Read barrier ensures online_count is read after prevent is cleared */
501 smp_rmb();
502
Colin Crossf9838272013-08-28 18:41:47 -0700503reset:
504 cpumask_clear_cpu(dev->cpu, &cpuidle_coupled_poked);
505
506 w = cpuidle_coupled_set_waiting(dev->cpu, coupled, next_state);
507 /*
508 * If this is the last cpu to enter the waiting state, poke
509 * all the other cpus out of their waiting state so they can
510 * enter a deeper state. This can race with one of the cpus
511 * exiting the waiting state due to an interrupt and
512 * decrementing waiting_count, see comment below.
513 */
514 if (w == coupled->online_count) {
515 cpumask_set_cpu(dev->cpu, &cpuidle_coupled_poked);
516 cpuidle_coupled_poke_others(dev->cpu, coupled);
517 }
Colin Cross4126c012012-05-07 17:57:41 -0700518
519retry:
520 /*
521 * Wait for all coupled cpus to be idle, using the deepest state
Colin Crossf9838272013-08-28 18:41:47 -0700522 * allowed for a single cpu. If this was not the poking cpu, wait
523 * for at least one poke before leaving to avoid a race where
524 * two cpus could arrive at the waiting loop at the same time,
525 * but the first of the two to arrive could skip the loop without
526 * processing the pokes from the last to arrive.
Colin Cross4126c012012-05-07 17:57:41 -0700527 */
Colin Crossf9838272013-08-28 18:41:47 -0700528 while (!cpuidle_coupled_cpus_waiting(coupled) ||
529 !cpumask_test_cpu(dev->cpu, &cpuidle_coupled_poked)) {
Colin Cross9e19b732013-08-23 12:45:12 -0700530 if (cpuidle_coupled_clear_pokes(dev->cpu))
531 continue;
532
533 if (need_resched()) {
Colin Cross4126c012012-05-07 17:57:41 -0700534 cpuidle_coupled_set_not_waiting(dev->cpu, coupled);
535 goto out;
536 }
537
538 if (coupled->prevent) {
539 cpuidle_coupled_set_not_waiting(dev->cpu, coupled);
540 goto out;
541 }
542
543 entered_state = cpuidle_enter_state(dev, drv,
Xunlei Pangba6a8602015-08-04 13:48:55 +0800544 drv->safe_state_index);
Colin Cross59e99852013-08-23 12:45:10 -0700545 local_irq_disable();
Colin Cross4126c012012-05-07 17:57:41 -0700546 }
547
Colin Cross9e19b732013-08-23 12:45:12 -0700548 cpuidle_coupled_clear_pokes(dev->cpu);
549 if (need_resched()) {
Colin Cross4126c012012-05-07 17:57:41 -0700550 cpuidle_coupled_set_not_waiting(dev->cpu, coupled);
551 goto out;
552 }
553
554 /*
Colin Crossf9838272013-08-28 18:41:47 -0700555 * Make sure final poke status for this cpu is visible before setting
556 * cpu as ready.
557 */
558 smp_wmb();
559
560 /*
Colin Cross4126c012012-05-07 17:57:41 -0700561 * All coupled cpus are probably idle. There is a small chance that
562 * one of the other cpus just became active. Increment the ready count,
563 * and spin until all coupled cpus have incremented the counter. Once a
564 * cpu has incremented the ready counter, it cannot abort idle and must
565 * spin until either all cpus have incremented the ready counter, or
566 * another cpu leaves idle and decrements the waiting counter.
567 */
568
569 cpuidle_coupled_set_ready(coupled);
570 while (!cpuidle_coupled_cpus_ready(coupled)) {
571 /* Check if any other cpus bailed out of idle. */
572 if (!cpuidle_coupled_cpus_waiting(coupled))
573 if (!cpuidle_coupled_set_not_ready(coupled))
574 goto retry;
575
576 cpu_relax();
577 }
578
Colin Crossf9838272013-08-28 18:41:47 -0700579 /*
580 * Make sure read of all cpus ready is done before reading pending pokes
581 */
582 smp_rmb();
583
584 /*
585 * There is a small chance that a cpu left and reentered idle after this
586 * cpu saw that all cpus were waiting. The cpu that reentered idle will
587 * have sent this cpu a poke, which will still be pending after the
588 * ready loop. The pending interrupt may be lost by the interrupt
589 * controller when entering the deep idle state. It's not possible to
590 * clear a pending interrupt without turning interrupts on and handling
591 * it, and it's too late to turn on interrupts here, so reset the
592 * coupled idle state of all cpus and retry.
593 */
594 if (cpuidle_coupled_any_pokes_pending(coupled)) {
595 cpuidle_coupled_set_done(dev->cpu, coupled);
596 /* Wait for all cpus to see the pending pokes */
597 cpuidle_coupled_parallel_barrier(dev, &coupled->abort_barrier);
598 goto reset;
599 }
600
Colin Cross4126c012012-05-07 17:57:41 -0700601 /* all cpus have acked the coupled state */
602 next_state = cpuidle_coupled_get_state(dev, coupled);
603
604 entered_state = cpuidle_enter_state(dev, drv, next_state);
605
606 cpuidle_coupled_set_done(dev->cpu, coupled);
607
608out:
609 /*
610 * Normal cpuidle states are expected to return with irqs enabled.
611 * That leads to an inefficiency where a cpu receiving an interrupt
612 * that brings it out of idle will process that interrupt before
613 * exiting the idle enter function and decrementing ready_count. All
614 * other cpus will need to spin waiting for the cpu that is processing
615 * the interrupt. If the driver returns with interrupts disabled,
616 * all other cpus will loop back into the safe idle state instead of
617 * spinning, saving power.
618 *
619 * Calling local_irq_enable here allows coupled states to return with
620 * interrupts disabled, but won't cause problems for drivers that
621 * exit with interrupts enabled.
622 */
623 local_irq_enable();
624
625 /*
626 * Wait until all coupled cpus have exited idle. There is no risk that
627 * a cpu exits and re-enters the ready state because this cpu has
628 * already decremented its waiting_count.
629 */
630 while (!cpuidle_coupled_no_cpus_ready(coupled))
631 cpu_relax();
632
633 return entered_state;
634}
635
636static void cpuidle_coupled_update_online_cpus(struct cpuidle_coupled *coupled)
637{
638 cpumask_t cpus;
639 cpumask_and(&cpus, cpu_online_mask, &coupled->coupled_cpus);
640 coupled->online_count = cpumask_weight(&cpus);
641}
642
643/**
644 * cpuidle_coupled_register_device - register a coupled cpuidle device
645 * @dev: struct cpuidle_device for the current cpu
646 *
647 * Called from cpuidle_register_device to handle coupled idle init. Finds the
648 * cpuidle_coupled struct for this set of coupled cpus, or creates one if none
649 * exists yet.
650 */
651int cpuidle_coupled_register_device(struct cpuidle_device *dev)
652{
653 int cpu;
654 struct cpuidle_device *other_dev;
655 struct call_single_data *csd;
656 struct cpuidle_coupled *coupled;
657
658 if (cpumask_empty(&dev->coupled_cpus))
659 return 0;
660
Rusty Russellf9b531f2015-03-05 10:49:16 +1030661 for_each_cpu(cpu, &dev->coupled_cpus) {
Colin Cross4126c012012-05-07 17:57:41 -0700662 other_dev = per_cpu(cpuidle_devices, cpu);
663 if (other_dev && other_dev->coupled) {
664 coupled = other_dev->coupled;
665 goto have_coupled;
666 }
667 }
668
669 /* No existing coupled info found, create a new one */
670 coupled = kzalloc(sizeof(struct cpuidle_coupled), GFP_KERNEL);
671 if (!coupled)
672 return -ENOMEM;
673
674 coupled->coupled_cpus = dev->coupled_cpus;
675
676have_coupled:
677 dev->coupled = coupled;
678 if (WARN_ON(!cpumask_equal(&dev->coupled_cpus, &coupled->coupled_cpus)))
679 coupled->prevent++;
680
681 cpuidle_coupled_update_online_cpus(coupled);
682
683 coupled->refcnt++;
684
685 csd = &per_cpu(cpuidle_coupled_poke_cb, dev->cpu);
Colin Crossf9838272013-08-28 18:41:47 -0700686 csd->func = cpuidle_coupled_handle_poke;
Colin Cross4126c012012-05-07 17:57:41 -0700687 csd->info = (void *)(unsigned long)dev->cpu;
688
689 return 0;
690}
691
692/**
693 * cpuidle_coupled_unregister_device - unregister a coupled cpuidle device
694 * @dev: struct cpuidle_device for the current cpu
695 *
696 * Called from cpuidle_unregister_device to tear down coupled idle. Removes the
697 * cpu from the coupled idle set, and frees the cpuidle_coupled_info struct if
698 * this was the last cpu in the set.
699 */
700void cpuidle_coupled_unregister_device(struct cpuidle_device *dev)
701{
702 struct cpuidle_coupled *coupled = dev->coupled;
703
704 if (cpumask_empty(&dev->coupled_cpus))
705 return;
706
707 if (--coupled->refcnt)
708 kfree(coupled);
709 dev->coupled = NULL;
710}
711
712/**
713 * cpuidle_coupled_prevent_idle - prevent cpus from entering a coupled state
714 * @coupled: the struct coupled that contains the cpu that is changing state
715 *
716 * Disables coupled cpuidle on a coupled set of cpus. Used to ensure that
717 * cpu_online_mask doesn't change while cpus are coordinating coupled idle.
718 */
719static void cpuidle_coupled_prevent_idle(struct cpuidle_coupled *coupled)
720{
721 int cpu = get_cpu();
722
723 /* Force all cpus out of the waiting loop. */
724 coupled->prevent++;
725 cpuidle_coupled_poke_others(cpu, coupled);
726 put_cpu();
727 while (!cpuidle_coupled_no_cpus_waiting(coupled))
728 cpu_relax();
729}
730
731/**
732 * cpuidle_coupled_allow_idle - allows cpus to enter a coupled state
733 * @coupled: the struct coupled that contains the cpu that is changing state
734 *
735 * Enables coupled cpuidle on a coupled set of cpus. Used to ensure that
736 * cpu_online_mask doesn't change while cpus are coordinating coupled idle.
737 */
738static void cpuidle_coupled_allow_idle(struct cpuidle_coupled *coupled)
739{
740 int cpu = get_cpu();
741
742 /*
743 * Write barrier ensures readers see the new online_count when they
744 * see prevent == 0.
745 */
746 smp_wmb();
747 coupled->prevent--;
748 /* Force cpus out of the prevent loop. */
749 cpuidle_coupled_poke_others(cpu, coupled);
750 put_cpu();
751}
752
753/**
754 * cpuidle_coupled_cpu_notify - notifier called during hotplug transitions
755 * @nb: notifier block
756 * @action: hotplug transition
757 * @hcpu: target cpu number
758 *
759 * Called when a cpu is brought on or offline using hotplug. Updates the
760 * coupled cpu set appropriately
761 */
762static int cpuidle_coupled_cpu_notify(struct notifier_block *nb,
763 unsigned long action, void *hcpu)
764{
765 int cpu = (unsigned long)hcpu;
766 struct cpuidle_device *dev;
767
Colin Cross63c6ba42012-08-15 22:10:50 +0200768 switch (action & ~CPU_TASKS_FROZEN) {
769 case CPU_UP_PREPARE:
770 case CPU_DOWN_PREPARE:
771 case CPU_ONLINE:
772 case CPU_DEAD:
773 case CPU_UP_CANCELED:
774 case CPU_DOWN_FAILED:
775 break;
776 default:
777 return NOTIFY_OK;
778 }
779
Colin Cross4126c012012-05-07 17:57:41 -0700780 mutex_lock(&cpuidle_lock);
781
782 dev = per_cpu(cpuidle_devices, cpu);
Jon Medhurst (Tixy)5fbbb902012-08-15 22:11:00 +0200783 if (!dev || !dev->coupled)
Colin Cross4126c012012-05-07 17:57:41 -0700784 goto out;
785
786 switch (action & ~CPU_TASKS_FROZEN) {
787 case CPU_UP_PREPARE:
788 case CPU_DOWN_PREPARE:
789 cpuidle_coupled_prevent_idle(dev->coupled);
790 break;
791 case CPU_ONLINE:
792 case CPU_DEAD:
793 cpuidle_coupled_update_online_cpus(dev->coupled);
794 /* Fall through */
795 case CPU_UP_CANCELED:
796 case CPU_DOWN_FAILED:
797 cpuidle_coupled_allow_idle(dev->coupled);
798 break;
799 }
800
801out:
802 mutex_unlock(&cpuidle_lock);
803 return NOTIFY_OK;
804}
805
806static struct notifier_block cpuidle_coupled_cpu_notifier = {
807 .notifier_call = cpuidle_coupled_cpu_notify,
808};
809
810static int __init cpuidle_coupled_init(void)
811{
812 return register_cpu_notifier(&cpuidle_coupled_cpu_notifier);
813}
814core_initcall(cpuidle_coupled_init);