Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Kernel internal timers
4 *
5 * Copyright (C) 1991, 1992 Linus Torvalds
6 *
7 * 1997-01-28 Modified by Finn Arne Gangstad to make timers scale better.
8 *
9 * 1997-09-10 Updated NTP code according to technical memorandum Jan '96
10 * "A Kernel Model for Precision Timekeeping" by Dave Mills
11 * 1998-12-24 Fixed a xtime SMP race (we need the xtime_lock rw spinlock to
12 * serialize accesses to xtime/lost_ticks).
13 * Copyright (C) 1998 Andrea Arcangeli
14 * 1999-03-10 Improved NTP compatibility by Ulrich Windl
15 * 2002-05-31 Move sys_sysinfo here and make its locking sane, Robert Love
16 * 2000-10-05 Implemented scalable SMP per-CPU timer handling.
17 * Copyright (C) 2000, 2001, 2002 Ingo Molnar
18 * Designed by David S. Miller, Alexey Kuznetsov and Ingo Molnar
19 */
20
21#include <linux/kernel_stat.h>
22#include <linux/export.h>
23#include <linux/interrupt.h>
24#include <linux/percpu.h>
25#include <linux/init.h>
26#include <linux/mm.h>
27#include <linux/swap.h>
28#include <linux/pid_namespace.h>
29#include <linux/notifier.h>
30#include <linux/thread_info.h>
31#include <linux/time.h>
32#include <linux/jiffies.h>
33#include <linux/posix-timers.h>
34#include <linux/cpu.h>
35#include <linux/syscalls.h>
36#include <linux/delay.h>
37#include <linux/tick.h>
38#include <linux/kallsyms.h>
39#include <linux/irq_work.h>
40#include <linux/sched/signal.h>
41#include <linux/sched/sysctl.h>
42#include <linux/sched/nohz.h>
43#include <linux/sched/debug.h>
44#include <linux/slab.h>
45#include <linux/compat.h>
46#include <linux/random.h>
47#include <linux/sysctl.h>
48
49#include <linux/uaccess.h>
50#include <asm/unistd.h>
51#include <asm/div64.h>
52#include <asm/timex.h>
53#include <asm/io.h>
54
55#include "tick-internal.h"
56
57#define CREATE_TRACE_POINTS
58#include <trace/events/timer.h>
59
60__visible u64 jiffies_64 __cacheline_aligned_in_smp = INITIAL_JIFFIES;
61
62EXPORT_SYMBOL(jiffies_64);
63
64/*
65 * The timer wheel has LVL_DEPTH array levels. Each level provides an array of
66 * LVL_SIZE buckets. Each level is driven by its own clock and therefor each
67 * level has a different granularity.
68 *
69 * The level granularity is: LVL_CLK_DIV ^ lvl
70 * The level clock frequency is: HZ / (LVL_CLK_DIV ^ level)
71 *
72 * The array level of a newly armed timer depends on the relative expiry
73 * time. The farther the expiry time is away the higher the array level and
74 * therefor the granularity becomes.
75 *
76 * Contrary to the original timer wheel implementation, which aims for 'exact'
77 * expiry of the timers, this implementation removes the need for recascading
78 * the timers into the lower array levels. The previous 'classic' timer wheel
79 * implementation of the kernel already violated the 'exact' expiry by adding
80 * slack to the expiry time to provide batched expiration. The granularity
81 * levels provide implicit batching.
82 *
83 * This is an optimization of the original timer wheel implementation for the
84 * majority of the timer wheel use cases: timeouts. The vast majority of
85 * timeout timers (networking, disk I/O ...) are canceled before expiry. If
86 * the timeout expires it indicates that normal operation is disturbed, so it
87 * does not matter much whether the timeout comes with a slight delay.
88 *
89 * The only exception to this are networking timers with a small expiry
90 * time. They rely on the granularity. Those fit into the first wheel level,
91 * which has HZ granularity.
92 *
93 * We don't have cascading anymore. timers with a expiry time above the
94 * capacity of the last wheel level are force expired at the maximum timeout
95 * value of the last wheel level. From data sampling we know that the maximum
96 * value observed is 5 days (network connection tracking), so this should not
97 * be an issue.
98 *
99 * The currently chosen array constants values are a good compromise between
100 * array size and granularity.
101 *
102 * This results in the following granularity and range levels:
103 *
104 * HZ 1000 steps
105 * Level Offset Granularity Range
106 * 0 0 1 ms 0 ms - 63 ms
107 * 1 64 8 ms 64 ms - 511 ms
108 * 2 128 64 ms 512 ms - 4095 ms (512ms - ~4s)
109 * 3 192 512 ms 4096 ms - 32767 ms (~4s - ~32s)
110 * 4 256 4096 ms (~4s) 32768 ms - 262143 ms (~32s - ~4m)
111 * 5 320 32768 ms (~32s) 262144 ms - 2097151 ms (~4m - ~34m)
112 * 6 384 262144 ms (~4m) 2097152 ms - 16777215 ms (~34m - ~4h)
113 * 7 448 2097152 ms (~34m) 16777216 ms - 134217727 ms (~4h - ~1d)
114 * 8 512 16777216 ms (~4h) 134217728 ms - 1073741822 ms (~1d - ~12d)
115 *
116 * HZ 300
117 * Level Offset Granularity Range
118 * 0 0 3 ms 0 ms - 210 ms
119 * 1 64 26 ms 213 ms - 1703 ms (213ms - ~1s)
120 * 2 128 213 ms 1706 ms - 13650 ms (~1s - ~13s)
121 * 3 192 1706 ms (~1s) 13653 ms - 109223 ms (~13s - ~1m)
122 * 4 256 13653 ms (~13s) 109226 ms - 873810 ms (~1m - ~14m)
123 * 5 320 109226 ms (~1m) 873813 ms - 6990503 ms (~14m - ~1h)
124 * 6 384 873813 ms (~14m) 6990506 ms - 55924050 ms (~1h - ~15h)
125 * 7 448 6990506 ms (~1h) 55924053 ms - 447392423 ms (~15h - ~5d)
126 * 8 512 55924053 ms (~15h) 447392426 ms - 3579139406 ms (~5d - ~41d)
127 *
128 * HZ 250
129 * Level Offset Granularity Range
130 * 0 0 4 ms 0 ms - 255 ms
131 * 1 64 32 ms 256 ms - 2047 ms (256ms - ~2s)
132 * 2 128 256 ms 2048 ms - 16383 ms (~2s - ~16s)
133 * 3 192 2048 ms (~2s) 16384 ms - 131071 ms (~16s - ~2m)
134 * 4 256 16384 ms (~16s) 131072 ms - 1048575 ms (~2m - ~17m)
135 * 5 320 131072 ms (~2m) 1048576 ms - 8388607 ms (~17m - ~2h)
136 * 6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
137 * 7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
138 * 8 512 67108864 ms (~18h) 536870912 ms - 4294967288 ms (~6d - ~49d)
139 *
140 * HZ 100
141 * Level Offset Granularity Range
142 * 0 0 10 ms 0 ms - 630 ms
143 * 1 64 80 ms 640 ms - 5110 ms (640ms - ~5s)
144 * 2 128 640 ms 5120 ms - 40950 ms (~5s - ~40s)
145 * 3 192 5120 ms (~5s) 40960 ms - 327670 ms (~40s - ~5m)
146 * 4 256 40960 ms (~40s) 327680 ms - 2621430 ms (~5m - ~43m)
147 * 5 320 327680 ms (~5m) 2621440 ms - 20971510 ms (~43m - ~5h)
148 * 6 384 2621440 ms (~43m) 20971520 ms - 167772150 ms (~5h - ~1d)
149 * 7 448 20971520 ms (~5h) 167772160 ms - 1342177270 ms (~1d - ~15d)
150 */
151
152/* Clock divisor for the next level */
153#define LVL_CLK_SHIFT 3
154#define LVL_CLK_DIV (1UL << LVL_CLK_SHIFT)
155#define LVL_CLK_MASK (LVL_CLK_DIV - 1)
156#define LVL_SHIFT(n) ((n) * LVL_CLK_SHIFT)
157#define LVL_GRAN(n) (1UL << LVL_SHIFT(n))
158
159/*
160 * The time start value for each level to select the bucket at enqueue
161 * time. We start from the last possible delta of the previous level
162 * so that we can later add an extra LVL_GRAN(n) to n (see calc_index()).
163 */
164#define LVL_START(n) ((LVL_SIZE - 1) << (((n) - 1) * LVL_CLK_SHIFT))
165
166/* Size of each clock level */
167#define LVL_BITS 6
168#define LVL_SIZE (1UL << LVL_BITS)
169#define LVL_MASK (LVL_SIZE - 1)
170#define LVL_OFFS(n) ((n) * LVL_SIZE)
171
172/* Level depth */
173#if HZ > 100
174# define LVL_DEPTH 9
175# else
176# define LVL_DEPTH 8
177#endif
178
179/* The cutoff (max. capacity of the wheel) */
180#define WHEEL_TIMEOUT_CUTOFF (LVL_START(LVL_DEPTH))
181#define WHEEL_TIMEOUT_MAX (WHEEL_TIMEOUT_CUTOFF - LVL_GRAN(LVL_DEPTH - 1))
182
183/*
184 * The resulting wheel size. If NOHZ is configured we allocate two
185 * wheels so we have a separate storage for the deferrable timers.
186 */
187#define WHEEL_SIZE (LVL_SIZE * LVL_DEPTH)
188
189#ifdef CONFIG_NO_HZ_COMMON
190# define NR_BASES 2
191# define BASE_STD 0
192# define BASE_DEF 1
193#else
194# define NR_BASES 1
195# define BASE_STD 0
196# define BASE_DEF 0
197#endif
198
199struct timer_base {
200 raw_spinlock_t lock;
201 struct timer_list *running_timer;
202#ifdef CONFIG_PREEMPT_RT
203 spinlock_t expiry_lock;
204 atomic_t timer_waiters;
205#endif
206 unsigned long clk;
207 unsigned long next_expiry;
208 unsigned int cpu;
209 bool next_expiry_recalc;
210 bool is_idle;
211 bool timers_pending;
212 DECLARE_BITMAP(pending_map, WHEEL_SIZE);
213 struct hlist_head vectors[WHEEL_SIZE];
214} ____cacheline_aligned;
215
216static DEFINE_PER_CPU(struct timer_base, timer_bases[NR_BASES]);
217
218#ifdef CONFIG_NO_HZ_COMMON
219
220static DEFINE_STATIC_KEY_FALSE(timers_nohz_active);
221static DEFINE_MUTEX(timer_keys_mutex);
222
223static void timer_update_keys(struct work_struct *work);
224static DECLARE_WORK(timer_update_work, timer_update_keys);
225
226#ifdef CONFIG_SMP
227static unsigned int sysctl_timer_migration = 1;
228
229DEFINE_STATIC_KEY_FALSE(timers_migration_enabled);
230
231static void timers_update_migration(void)
232{
233 if (sysctl_timer_migration && tick_nohz_active)
234 static_branch_enable(&timers_migration_enabled);
235 else
236 static_branch_disable(&timers_migration_enabled);
237}
238
239#ifdef CONFIG_SYSCTL
240static int timer_migration_handler(struct ctl_table *table, int write,
241 void *buffer, size_t *lenp, loff_t *ppos)
242{
243 int ret;
244
245 mutex_lock(&timer_keys_mutex);
246 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
247 if (!ret && write)
248 timers_update_migration();
249 mutex_unlock(&timer_keys_mutex);
250 return ret;
251}
252
253static struct ctl_table timer_sysctl[] = {
254 {
255 .procname = "timer_migration",
256 .data = &sysctl_timer_migration,
257 .maxlen = sizeof(unsigned int),
258 .mode = 0644,
259 .proc_handler = timer_migration_handler,
260 .extra1 = SYSCTL_ZERO,
261 .extra2 = SYSCTL_ONE,
262 },
263 {}
264};
265
266static int __init timer_sysctl_init(void)
267{
268 register_sysctl("kernel", timer_sysctl);
269 return 0;
270}
271device_initcall(timer_sysctl_init);
272#endif /* CONFIG_SYSCTL */
273#else /* CONFIG_SMP */
274static inline void timers_update_migration(void) { }
275#endif /* !CONFIG_SMP */
276
277static void timer_update_keys(struct work_struct *work)
278{
279 mutex_lock(&timer_keys_mutex);
280 timers_update_migration();
281 static_branch_enable(&timers_nohz_active);
282 mutex_unlock(&timer_keys_mutex);
283}
284
285void timers_update_nohz(void)
286{
287 schedule_work(&timer_update_work);
288}
289
290static inline bool is_timers_nohz_active(void)
291{
292 return static_branch_unlikely(&timers_nohz_active);
293}
294#else
295static inline bool is_timers_nohz_active(void) { return false; }
296#endif /* NO_HZ_COMMON */
297
298static unsigned long round_jiffies_common(unsigned long j, int cpu,
299 bool force_up)
300{
301 int rem;
302 unsigned long original = j;
303
304 /*
305 * We don't want all cpus firing their timers at once hitting the
306 * same lock or cachelines, so we skew each extra cpu with an extra
307 * 3 jiffies. This 3 jiffies came originally from the mm/ code which
308 * already did this.
309 * The skew is done by adding 3*cpunr, then round, then subtract this
310 * extra offset again.
311 */
312 j += cpu * 3;
313
314 rem = j % HZ;
315
316 /*
317 * If the target jiffie is just after a whole second (which can happen
318 * due to delays of the timer irq, long irq off times etc etc) then
319 * we should round down to the whole second, not up. Use 1/4th second
320 * as cutoff for this rounding as an extreme upper bound for this.
321 * But never round down if @force_up is set.
322 */
323 if (rem < HZ/4 && !force_up) /* round down */
324 j = j - rem;
325 else /* round up */
326 j = j - rem + HZ;
327
328 /* now that we have rounded, subtract the extra skew again */
329 j -= cpu * 3;
330
331 /*
332 * Make sure j is still in the future. Otherwise return the
333 * unmodified value.
334 */
335 return time_is_after_jiffies(j) ? j : original;
336}
337
338/**
339 * __round_jiffies - function to round jiffies to a full second
340 * @j: the time in (absolute) jiffies that should be rounded
341 * @cpu: the processor number on which the timeout will happen
342 *
343 * __round_jiffies() rounds an absolute time in the future (in jiffies)
344 * up or down to (approximately) full seconds. This is useful for timers
345 * for which the exact time they fire does not matter too much, as long as
346 * they fire approximately every X seconds.
347 *
348 * By rounding these timers to whole seconds, all such timers will fire
349 * at the same time, rather than at various times spread out. The goal
350 * of this is to have the CPU wake up less, which saves power.
351 *
352 * The exact rounding is skewed for each processor to avoid all
353 * processors firing at the exact same time, which could lead
354 * to lock contention or spurious cache line bouncing.
355 *
356 * The return value is the rounded version of the @j parameter.
357 */
358unsigned long __round_jiffies(unsigned long j, int cpu)
359{
360 return round_jiffies_common(j, cpu, false);
361}
362EXPORT_SYMBOL_GPL(__round_jiffies);
363
364/**
365 * __round_jiffies_relative - function to round jiffies to a full second
366 * @j: the time in (relative) jiffies that should be rounded
367 * @cpu: the processor number on which the timeout will happen
368 *
369 * __round_jiffies_relative() rounds a time delta in the future (in jiffies)
370 * up or down to (approximately) full seconds. This is useful for timers
371 * for which the exact time they fire does not matter too much, as long as
372 * they fire approximately every X seconds.
373 *
374 * By rounding these timers to whole seconds, all such timers will fire
375 * at the same time, rather than at various times spread out. The goal
376 * of this is to have the CPU wake up less, which saves power.
377 *
378 * The exact rounding is skewed for each processor to avoid all
379 * processors firing at the exact same time, which could lead
380 * to lock contention or spurious cache line bouncing.
381 *
382 * The return value is the rounded version of the @j parameter.
383 */
384unsigned long __round_jiffies_relative(unsigned long j, int cpu)
385{
386 unsigned long j0 = jiffies;
387
388 /* Use j0 because jiffies might change while we run */
389 return round_jiffies_common(j + j0, cpu, false) - j0;
390}
391EXPORT_SYMBOL_GPL(__round_jiffies_relative);
392
393/**
394 * round_jiffies - function to round jiffies to a full second
395 * @j: the time in (absolute) jiffies that should be rounded
396 *
397 * round_jiffies() rounds an absolute time in the future (in jiffies)
398 * up or down to (approximately) full seconds. This is useful for timers
399 * for which the exact time they fire does not matter too much, as long as
400 * they fire approximately every X seconds.
401 *
402 * By rounding these timers to whole seconds, all such timers will fire
403 * at the same time, rather than at various times spread out. The goal
404 * of this is to have the CPU wake up less, which saves power.
405 *
406 * The return value is the rounded version of the @j parameter.
407 */
408unsigned long round_jiffies(unsigned long j)
409{
410 return round_jiffies_common(j, raw_smp_processor_id(), false);
411}
412EXPORT_SYMBOL_GPL(round_jiffies);
413
414/**
415 * round_jiffies_relative - function to round jiffies to a full second
416 * @j: the time in (relative) jiffies that should be rounded
417 *
418 * round_jiffies_relative() rounds a time delta in the future (in jiffies)
419 * up or down to (approximately) full seconds. This is useful for timers
420 * for which the exact time they fire does not matter too much, as long as
421 * they fire approximately every X seconds.
422 *
423 * By rounding these timers to whole seconds, all such timers will fire
424 * at the same time, rather than at various times spread out. The goal
425 * of this is to have the CPU wake up less, which saves power.
426 *
427 * The return value is the rounded version of the @j parameter.
428 */
429unsigned long round_jiffies_relative(unsigned long j)
430{
431 return __round_jiffies_relative(j, raw_smp_processor_id());
432}
433EXPORT_SYMBOL_GPL(round_jiffies_relative);
434
435/**
436 * __round_jiffies_up - function to round jiffies up to a full second
437 * @j: the time in (absolute) jiffies that should be rounded
438 * @cpu: the processor number on which the timeout will happen
439 *
440 * This is the same as __round_jiffies() except that it will never
441 * round down. This is useful for timeouts for which the exact time
442 * of firing does not matter too much, as long as they don't fire too
443 * early.
444 */
445unsigned long __round_jiffies_up(unsigned long j, int cpu)
446{
447 return round_jiffies_common(j, cpu, true);
448}
449EXPORT_SYMBOL_GPL(__round_jiffies_up);
450
451/**
452 * __round_jiffies_up_relative - function to round jiffies up to a full second
453 * @j: the time in (relative) jiffies that should be rounded
454 * @cpu: the processor number on which the timeout will happen
455 *
456 * This is the same as __round_jiffies_relative() except that it will never
457 * round down. This is useful for timeouts for which the exact time
458 * of firing does not matter too much, as long as they don't fire too
459 * early.
460 */
461unsigned long __round_jiffies_up_relative(unsigned long j, int cpu)
462{
463 unsigned long j0 = jiffies;
464
465 /* Use j0 because jiffies might change while we run */
466 return round_jiffies_common(j + j0, cpu, true) - j0;
467}
468EXPORT_SYMBOL_GPL(__round_jiffies_up_relative);
469
470/**
471 * round_jiffies_up - function to round jiffies up to a full second
472 * @j: the time in (absolute) jiffies that should be rounded
473 *
474 * This is the same as round_jiffies() except that it will never
475 * round down. This is useful for timeouts for which the exact time
476 * of firing does not matter too much, as long as they don't fire too
477 * early.
478 */
479unsigned long round_jiffies_up(unsigned long j)
480{
481 return round_jiffies_common(j, raw_smp_processor_id(), true);
482}
483EXPORT_SYMBOL_GPL(round_jiffies_up);
484
485/**
486 * round_jiffies_up_relative - function to round jiffies up to a full second
487 * @j: the time in (relative) jiffies that should be rounded
488 *
489 * This is the same as round_jiffies_relative() except that it will never
490 * round down. This is useful for timeouts for which the exact time
491 * of firing does not matter too much, as long as they don't fire too
492 * early.
493 */
494unsigned long round_jiffies_up_relative(unsigned long j)
495{
496 return __round_jiffies_up_relative(j, raw_smp_processor_id());
497}
498EXPORT_SYMBOL_GPL(round_jiffies_up_relative);
499
500
501static inline unsigned int timer_get_idx(struct timer_list *timer)
502{
503 return (timer->flags & TIMER_ARRAYMASK) >> TIMER_ARRAYSHIFT;
504}
505
506static inline void timer_set_idx(struct timer_list *timer, unsigned int idx)
507{
508 timer->flags = (timer->flags & ~TIMER_ARRAYMASK) |
509 idx << TIMER_ARRAYSHIFT;
510}
511
512/*
513 * Helper function to calculate the array index for a given expiry
514 * time.
515 */
516static inline unsigned calc_index(unsigned long expires, unsigned lvl,
517 unsigned long *bucket_expiry)
518{
519
520 /*
521 * The timer wheel has to guarantee that a timer does not fire
522 * early. Early expiry can happen due to:
523 * - Timer is armed at the edge of a tick
524 * - Truncation of the expiry time in the outer wheel levels
525 *
526 * Round up with level granularity to prevent this.
527 */
528 expires = (expires >> LVL_SHIFT(lvl)) + 1;
529 *bucket_expiry = expires << LVL_SHIFT(lvl);
530 return LVL_OFFS(lvl) + (expires & LVL_MASK);
531}
532
533static int calc_wheel_index(unsigned long expires, unsigned long clk,
534 unsigned long *bucket_expiry)
535{
536 unsigned long delta = expires - clk;
537 unsigned int idx;
538
539 if (delta < LVL_START(1)) {
540 idx = calc_index(expires, 0, bucket_expiry);
541 } else if (delta < LVL_START(2)) {
542 idx = calc_index(expires, 1, bucket_expiry);
543 } else if (delta < LVL_START(3)) {
544 idx = calc_index(expires, 2, bucket_expiry);
545 } else if (delta < LVL_START(4)) {
546 idx = calc_index(expires, 3, bucket_expiry);
547 } else if (delta < LVL_START(5)) {
548 idx = calc_index(expires, 4, bucket_expiry);
549 } else if (delta < LVL_START(6)) {
550 idx = calc_index(expires, 5, bucket_expiry);
551 } else if (delta < LVL_START(7)) {
552 idx = calc_index(expires, 6, bucket_expiry);
553 } else if (LVL_DEPTH > 8 && delta < LVL_START(8)) {
554 idx = calc_index(expires, 7, bucket_expiry);
555 } else if ((long) delta < 0) {
556 idx = clk & LVL_MASK;
557 *bucket_expiry = clk;
558 } else {
559 /*
560 * Force expire obscene large timeouts to expire at the
561 * capacity limit of the wheel.
562 */
563 if (delta >= WHEEL_TIMEOUT_CUTOFF)
564 expires = clk + WHEEL_TIMEOUT_MAX;
565
566 idx = calc_index(expires, LVL_DEPTH - 1, bucket_expiry);
567 }
568 return idx;
569}
570
571static void
572trigger_dyntick_cpu(struct timer_base *base, struct timer_list *timer)
573{
574 if (!is_timers_nohz_active())
575 return;
576
577 /*
578 * TODO: This wants some optimizing similar to the code below, but we
579 * will do that when we switch from push to pull for deferrable timers.
580 */
581 if (timer->flags & TIMER_DEFERRABLE) {
582 if (tick_nohz_full_cpu(base->cpu))
583 wake_up_nohz_cpu(base->cpu);
584 return;
585 }
586
587 /*
588 * We might have to IPI the remote CPU if the base is idle and the
589 * timer is not deferrable. If the other CPU is on the way to idle
590 * then it can't set base->is_idle as we hold the base lock:
591 */
592 if (base->is_idle)
593 wake_up_nohz_cpu(base->cpu);
594}
595
596/*
597 * Enqueue the timer into the hash bucket, mark it pending in
598 * the bitmap, store the index in the timer flags then wake up
599 * the target CPU if needed.
600 */
601static void enqueue_timer(struct timer_base *base, struct timer_list *timer,
602 unsigned int idx, unsigned long bucket_expiry)
603{
604
605 hlist_add_head(&timer->entry, base->vectors + idx);
606 __set_bit(idx, base->pending_map);
607 timer_set_idx(timer, idx);
608
609 trace_timer_start(timer, timer->expires, timer->flags);
610
611 /*
612 * Check whether this is the new first expiring timer. The
613 * effective expiry time of the timer is required here
614 * (bucket_expiry) instead of timer->expires.
615 */
616 if (time_before(bucket_expiry, base->next_expiry)) {
617 /*
618 * Set the next expiry time and kick the CPU so it
619 * can reevaluate the wheel:
620 */
621 base->next_expiry = bucket_expiry;
622 base->timers_pending = true;
623 base->next_expiry_recalc = false;
624 trigger_dyntick_cpu(base, timer);
625 }
626}
627
628static void internal_add_timer(struct timer_base *base, struct timer_list *timer)
629{
630 unsigned long bucket_expiry;
631 unsigned int idx;
632
633 idx = calc_wheel_index(timer->expires, base->clk, &bucket_expiry);
634 enqueue_timer(base, timer, idx, bucket_expiry);
635}
636
637#ifdef CONFIG_DEBUG_OBJECTS_TIMERS
638
639static const struct debug_obj_descr timer_debug_descr;
640
641struct timer_hint {
642 void (*function)(struct timer_list *t);
643 long offset;
644};
645
646#define TIMER_HINT(fn, container, timr, hintfn) \
647 { \
648 .function = fn, \
649 .offset = offsetof(container, hintfn) - \
650 offsetof(container, timr) \
651 }
652
653static const struct timer_hint timer_hints[] = {
654 TIMER_HINT(delayed_work_timer_fn,
655 struct delayed_work, timer, work.func),
656 TIMER_HINT(kthread_delayed_work_timer_fn,
657 struct kthread_delayed_work, timer, work.func),
658};
659
660static void *timer_debug_hint(void *addr)
661{
662 struct timer_list *timer = addr;
663 int i;
664
665 for (i = 0; i < ARRAY_SIZE(timer_hints); i++) {
666 if (timer_hints[i].function == timer->function) {
667 void (**fn)(void) = addr + timer_hints[i].offset;
668
669 return *fn;
670 }
671 }
672
673 return timer->function;
674}
675
676static bool timer_is_static_object(void *addr)
677{
678 struct timer_list *timer = addr;
679
680 return (timer->entry.pprev == NULL &&
681 timer->entry.next == TIMER_ENTRY_STATIC);
682}
683
684/*
685 * fixup_init is called when:
686 * - an active object is initialized
687 */
688static bool timer_fixup_init(void *addr, enum debug_obj_state state)
689{
690 struct timer_list *timer = addr;
691
692 switch (state) {
693 case ODEBUG_STATE_ACTIVE:
694 del_timer_sync(timer);
695 debug_object_init(timer, &timer_debug_descr);
696 return true;
697 default:
698 return false;
699 }
700}
701
702/* Stub timer callback for improperly used timers. */
703static void stub_timer(struct timer_list *unused)
704{
705 WARN_ON(1);
706}
707
708/*
709 * fixup_activate is called when:
710 * - an active object is activated
711 * - an unknown non-static object is activated
712 */
713static bool timer_fixup_activate(void *addr, enum debug_obj_state state)
714{
715 struct timer_list *timer = addr;
716
717 switch (state) {
718 case ODEBUG_STATE_NOTAVAILABLE:
719 timer_setup(timer, stub_timer, 0);
720 return true;
721
722 case ODEBUG_STATE_ACTIVE:
723 WARN_ON(1);
724 fallthrough;
725 default:
726 return false;
727 }
728}
729
730/*
731 * fixup_free is called when:
732 * - an active object is freed
733 */
734static bool timer_fixup_free(void *addr, enum debug_obj_state state)
735{
736 struct timer_list *timer = addr;
737
738 switch (state) {
739 case ODEBUG_STATE_ACTIVE:
740 del_timer_sync(timer);
741 debug_object_free(timer, &timer_debug_descr);
742 return true;
743 default:
744 return false;
745 }
746}
747
748/*
749 * fixup_assert_init is called when:
750 * - an untracked/uninit-ed object is found
751 */
752static bool timer_fixup_assert_init(void *addr, enum debug_obj_state state)
753{
754 struct timer_list *timer = addr;
755
756 switch (state) {
757 case ODEBUG_STATE_NOTAVAILABLE:
758 timer_setup(timer, stub_timer, 0);
759 return true;
760 default:
761 return false;
762 }
763}
764
765static const struct debug_obj_descr timer_debug_descr = {
766 .name = "timer_list",
767 .debug_hint = timer_debug_hint,
768 .is_static_object = timer_is_static_object,
769 .fixup_init = timer_fixup_init,
770 .fixup_activate = timer_fixup_activate,
771 .fixup_free = timer_fixup_free,
772 .fixup_assert_init = timer_fixup_assert_init,
773};
774
775static inline void debug_timer_init(struct timer_list *timer)
776{
777 debug_object_init(timer, &timer_debug_descr);
778}
779
780static inline void debug_timer_activate(struct timer_list *timer)
781{
782 debug_object_activate(timer, &timer_debug_descr);
783}
784
785static inline void debug_timer_deactivate(struct timer_list *timer)
786{
787 debug_object_deactivate(timer, &timer_debug_descr);
788}
789
790static inline void debug_timer_assert_init(struct timer_list *timer)
791{
792 debug_object_assert_init(timer, &timer_debug_descr);
793}
794
795static void do_init_timer(struct timer_list *timer,
796 void (*func)(struct timer_list *),
797 unsigned int flags,
798 const char *name, struct lock_class_key *key);
799
800void init_timer_on_stack_key(struct timer_list *timer,
801 void (*func)(struct timer_list *),
802 unsigned int flags,
803 const char *name, struct lock_class_key *key)
804{
805 debug_object_init_on_stack(timer, &timer_debug_descr);
806 do_init_timer(timer, func, flags, name, key);
807}
808EXPORT_SYMBOL_GPL(init_timer_on_stack_key);
809
810void destroy_timer_on_stack(struct timer_list *timer)
811{
812 debug_object_free(timer, &timer_debug_descr);
813}
814EXPORT_SYMBOL_GPL(destroy_timer_on_stack);
815
816#else
817static inline void debug_timer_init(struct timer_list *timer) { }
818static inline void debug_timer_activate(struct timer_list *timer) { }
819static inline void debug_timer_deactivate(struct timer_list *timer) { }
820static inline void debug_timer_assert_init(struct timer_list *timer) { }
821#endif
822
823static inline void debug_init(struct timer_list *timer)
824{
825 debug_timer_init(timer);
826 trace_timer_init(timer);
827}
828
829static inline void debug_deactivate(struct timer_list *timer)
830{
831 debug_timer_deactivate(timer);
832 trace_timer_cancel(timer);
833}
834
835static inline void debug_assert_init(struct timer_list *timer)
836{
837 debug_timer_assert_init(timer);
838}
839
840static void do_init_timer(struct timer_list *timer,
841 void (*func)(struct timer_list *),
842 unsigned int flags,
843 const char *name, struct lock_class_key *key)
844{
845 timer->entry.pprev = NULL;
846 timer->function = func;
847 if (WARN_ON_ONCE(flags & ~TIMER_INIT_FLAGS))
848 flags &= TIMER_INIT_FLAGS;
849 timer->flags = flags | raw_smp_processor_id();
850 lockdep_init_map(&timer->lockdep_map, name, key, 0);
851}
852
853/**
854 * init_timer_key - initialize a timer
855 * @timer: the timer to be initialized
856 * @func: timer callback function
857 * @flags: timer flags
858 * @name: name of the timer
859 * @key: lockdep class key of the fake lock used for tracking timer
860 * sync lock dependencies
861 *
862 * init_timer_key() must be done to a timer prior calling *any* of the
863 * other timer functions.
864 */
865void init_timer_key(struct timer_list *timer,
866 void (*func)(struct timer_list *), unsigned int flags,
867 const char *name, struct lock_class_key *key)
868{
869 debug_init(timer);
870 do_init_timer(timer, func, flags, name, key);
871}
872EXPORT_SYMBOL(init_timer_key);
873
874static inline void detach_timer(struct timer_list *timer, bool clear_pending)
875{
876 struct hlist_node *entry = &timer->entry;
877
878 debug_deactivate(timer);
879
880 __hlist_del(entry);
881 if (clear_pending)
882 entry->pprev = NULL;
883 entry->next = LIST_POISON2;
884}
885
886static int detach_if_pending(struct timer_list *timer, struct timer_base *base,
887 bool clear_pending)
888{
889 unsigned idx = timer_get_idx(timer);
890
891 if (!timer_pending(timer))
892 return 0;
893
894 if (hlist_is_singular_node(&timer->entry, base->vectors + idx)) {
895 __clear_bit(idx, base->pending_map);
896 base->next_expiry_recalc = true;
897 }
898
899 detach_timer(timer, clear_pending);
900 return 1;
901}
902
903static inline struct timer_base *get_timer_cpu_base(u32 tflags, u32 cpu)
904{
905 struct timer_base *base = per_cpu_ptr(&timer_bases[BASE_STD], cpu);
906
907 /*
908 * If the timer is deferrable and NO_HZ_COMMON is set then we need
909 * to use the deferrable base.
910 */
911 if (IS_ENABLED(CONFIG_NO_HZ_COMMON) && (tflags & TIMER_DEFERRABLE))
912 base = per_cpu_ptr(&timer_bases[BASE_DEF], cpu);
913 return base;
914}
915
916static inline struct timer_base *get_timer_this_cpu_base(u32 tflags)
917{
918 struct timer_base *base = this_cpu_ptr(&timer_bases[BASE_STD]);
919
920 /*
921 * If the timer is deferrable and NO_HZ_COMMON is set then we need
922 * to use the deferrable base.
923 */
924 if (IS_ENABLED(CONFIG_NO_HZ_COMMON) && (tflags & TIMER_DEFERRABLE))
925 base = this_cpu_ptr(&timer_bases[BASE_DEF]);
926 return base;
927}
928
929static inline struct timer_base *get_timer_base(u32 tflags)
930{
931 return get_timer_cpu_base(tflags, tflags & TIMER_CPUMASK);
932}
933
934static inline struct timer_base *
935get_target_base(struct timer_base *base, unsigned tflags)
936{
937#if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
938 if (static_branch_likely(&timers_migration_enabled) &&
939 !(tflags & TIMER_PINNED))
940 return get_timer_cpu_base(tflags, get_nohz_timer_target());
941#endif
942 return get_timer_this_cpu_base(tflags);
943}
944
945static inline void forward_timer_base(struct timer_base *base)
946{
947 unsigned long jnow = READ_ONCE(jiffies);
948
949 /*
950 * No need to forward if we are close enough below jiffies.
951 * Also while executing timers, base->clk is 1 offset ahead
952 * of jiffies to avoid endless requeuing to current jiffies.
953 */
954 if ((long)(jnow - base->clk) < 1)
955 return;
956
957 /*
958 * If the next expiry value is > jiffies, then we fast forward to
959 * jiffies otherwise we forward to the next expiry value.
960 */
961 if (time_after(base->next_expiry, jnow)) {
962 base->clk = jnow;
963 } else {
964 if (WARN_ON_ONCE(time_before(base->next_expiry, base->clk)))
965 return;
966 base->clk = base->next_expiry;
967 }
968}
969
970
971/*
972 * We are using hashed locking: Holding per_cpu(timer_bases[x]).lock means
973 * that all timers which are tied to this base are locked, and the base itself
974 * is locked too.
975 *
976 * So __run_timers/migrate_timers can safely modify all timers which could
977 * be found in the base->vectors array.
978 *
979 * When a timer is migrating then the TIMER_MIGRATING flag is set and we need
980 * to wait until the migration is done.
981 */
982static struct timer_base *lock_timer_base(struct timer_list *timer,
983 unsigned long *flags)
984 __acquires(timer->base->lock)
985{
986 for (;;) {
987 struct timer_base *base;
988 u32 tf;
989
990 /*
991 * We need to use READ_ONCE() here, otherwise the compiler
992 * might re-read @tf between the check for TIMER_MIGRATING
993 * and spin_lock().
994 */
995 tf = READ_ONCE(timer->flags);
996
997 if (!(tf & TIMER_MIGRATING)) {
998 base = get_timer_base(tf);
999 raw_spin_lock_irqsave(&base->lock, *flags);
1000 if (timer->flags == tf)
1001 return base;
1002 raw_spin_unlock_irqrestore(&base->lock, *flags);
1003 }
1004 cpu_relax();
1005 }
1006}
1007
1008#define MOD_TIMER_PENDING_ONLY 0x01
1009#define MOD_TIMER_REDUCE 0x02
1010#define MOD_TIMER_NOTPENDING 0x04
1011
1012static inline int
1013__mod_timer(struct timer_list *timer, unsigned long expires, unsigned int options)
1014{
1015 unsigned long clk = 0, flags, bucket_expiry;
1016 struct timer_base *base, *new_base;
1017 unsigned int idx = UINT_MAX;
1018 int ret = 0;
1019
1020 debug_assert_init(timer);
1021
1022 /*
1023 * This is a common optimization triggered by the networking code - if
1024 * the timer is re-modified to have the same timeout or ends up in the
1025 * same array bucket then just return:
1026 */
1027 if (!(options & MOD_TIMER_NOTPENDING) && timer_pending(timer)) {
1028 /*
1029 * The downside of this optimization is that it can result in
1030 * larger granularity than you would get from adding a new
1031 * timer with this expiry.
1032 */
1033 long diff = timer->expires - expires;
1034
1035 if (!diff)
1036 return 1;
1037 if (options & MOD_TIMER_REDUCE && diff <= 0)
1038 return 1;
1039
1040 /*
1041 * We lock timer base and calculate the bucket index right
1042 * here. If the timer ends up in the same bucket, then we
1043 * just update the expiry time and avoid the whole
1044 * dequeue/enqueue dance.
1045 */
1046 base = lock_timer_base(timer, &flags);
1047 /*
1048 * Has @timer been shutdown? This needs to be evaluated
1049 * while holding base lock to prevent a race against the
1050 * shutdown code.
1051 */
1052 if (!timer->function)
1053 goto out_unlock;
1054
1055 forward_timer_base(base);
1056
1057 if (timer_pending(timer) && (options & MOD_TIMER_REDUCE) &&
1058 time_before_eq(timer->expires, expires)) {
1059 ret = 1;
1060 goto out_unlock;
1061 }
1062
1063 clk = base->clk;
1064 idx = calc_wheel_index(expires, clk, &bucket_expiry);
1065
1066 /*
1067 * Retrieve and compare the array index of the pending
1068 * timer. If it matches set the expiry to the new value so a
1069 * subsequent call will exit in the expires check above.
1070 */
1071 if (idx == timer_get_idx(timer)) {
1072 if (!(options & MOD_TIMER_REDUCE))
1073 timer->expires = expires;
1074 else if (time_after(timer->expires, expires))
1075 timer->expires = expires;
1076 ret = 1;
1077 goto out_unlock;
1078 }
1079 } else {
1080 base = lock_timer_base(timer, &flags);
1081 /*
1082 * Has @timer been shutdown? This needs to be evaluated
1083 * while holding base lock to prevent a race against the
1084 * shutdown code.
1085 */
1086 if (!timer->function)
1087 goto out_unlock;
1088
1089 forward_timer_base(base);
1090 }
1091
1092 ret = detach_if_pending(timer, base, false);
1093 if (!ret && (options & MOD_TIMER_PENDING_ONLY))
1094 goto out_unlock;
1095
1096 new_base = get_target_base(base, timer->flags);
1097
1098 if (base != new_base) {
1099 /*
1100 * We are trying to schedule the timer on the new base.
1101 * However we can't change timer's base while it is running,
1102 * otherwise timer_delete_sync() can't detect that the timer's
1103 * handler yet has not finished. This also guarantees that the
1104 * timer is serialized wrt itself.
1105 */
1106 if (likely(base->running_timer != timer)) {
1107 /* See the comment in lock_timer_base() */
1108 timer->flags |= TIMER_MIGRATING;
1109
1110 raw_spin_unlock(&base->lock);
1111 base = new_base;
1112 raw_spin_lock(&base->lock);
1113 WRITE_ONCE(timer->flags,
1114 (timer->flags & ~TIMER_BASEMASK) | base->cpu);
1115 forward_timer_base(base);
1116 }
1117 }
1118
1119 debug_timer_activate(timer);
1120
1121 timer->expires = expires;
1122 /*
1123 * If 'idx' was calculated above and the base time did not advance
1124 * between calculating 'idx' and possibly switching the base, only
1125 * enqueue_timer() is required. Otherwise we need to (re)calculate
1126 * the wheel index via internal_add_timer().
1127 */
1128 if (idx != UINT_MAX && clk == base->clk)
1129 enqueue_timer(base, timer, idx, bucket_expiry);
1130 else
1131 internal_add_timer(base, timer);
1132
1133out_unlock:
1134 raw_spin_unlock_irqrestore(&base->lock, flags);
1135
1136 return ret;
1137}
1138
1139/**
1140 * mod_timer_pending - Modify a pending timer's timeout
1141 * @timer: The pending timer to be modified
1142 * @expires: New absolute timeout in jiffies
1143 *
1144 * mod_timer_pending() is the same for pending timers as mod_timer(), but
1145 * will not activate inactive timers.
1146 *
1147 * If @timer->function == NULL then the start operation is silently
1148 * discarded.
1149 *
1150 * Return:
1151 * * %0 - The timer was inactive and not modified or was in
1152 * shutdown state and the operation was discarded
1153 * * %1 - The timer was active and requeued to expire at @expires
1154 */
1155int mod_timer_pending(struct timer_list *timer, unsigned long expires)
1156{
1157 return __mod_timer(timer, expires, MOD_TIMER_PENDING_ONLY);
1158}
1159EXPORT_SYMBOL(mod_timer_pending);
1160
1161/**
1162 * mod_timer - Modify a timer's timeout
1163 * @timer: The timer to be modified
1164 * @expires: New absolute timeout in jiffies
1165 *
1166 * mod_timer(timer, expires) is equivalent to:
1167 *
1168 * del_timer(timer); timer->expires = expires; add_timer(timer);
1169 *
1170 * mod_timer() is more efficient than the above open coded sequence. In
1171 * case that the timer is inactive, the del_timer() part is a NOP. The
1172 * timer is in any case activated with the new expiry time @expires.
1173 *
1174 * Note that if there are multiple unserialized concurrent users of the
1175 * same timer, then mod_timer() is the only safe way to modify the timeout,
1176 * since add_timer() cannot modify an already running timer.
1177 *
1178 * If @timer->function == NULL then the start operation is silently
1179 * discarded. In this case the return value is 0 and meaningless.
1180 *
1181 * Return:
1182 * * %0 - The timer was inactive and started or was in shutdown
1183 * state and the operation was discarded
1184 * * %1 - The timer was active and requeued to expire at @expires or
1185 * the timer was active and not modified because @expires did
1186 * not change the effective expiry time
1187 */
1188int mod_timer(struct timer_list *timer, unsigned long expires)
1189{
1190 return __mod_timer(timer, expires, 0);
1191}
1192EXPORT_SYMBOL(mod_timer);
1193
1194/**
1195 * timer_reduce - Modify a timer's timeout if it would reduce the timeout
1196 * @timer: The timer to be modified
1197 * @expires: New absolute timeout in jiffies
1198 *
1199 * timer_reduce() is very similar to mod_timer(), except that it will only
1200 * modify an enqueued timer if that would reduce the expiration time. If
1201 * @timer is not enqueued it starts the timer.
1202 *
1203 * If @timer->function == NULL then the start operation is silently
1204 * discarded.
1205 *
1206 * Return:
1207 * * %0 - The timer was inactive and started or was in shutdown
1208 * state and the operation was discarded
1209 * * %1 - The timer was active and requeued to expire at @expires or
1210 * the timer was active and not modified because @expires
1211 * did not change the effective expiry time such that the
1212 * timer would expire earlier than already scheduled
1213 */
1214int timer_reduce(struct timer_list *timer, unsigned long expires)
1215{
1216 return __mod_timer(timer, expires, MOD_TIMER_REDUCE);
1217}
1218EXPORT_SYMBOL(timer_reduce);
1219
1220/**
1221 * add_timer - Start a timer
1222 * @timer: The timer to be started
1223 *
1224 * Start @timer to expire at @timer->expires in the future. @timer->expires
1225 * is the absolute expiry time measured in 'jiffies'. When the timer expires
1226 * timer->function(timer) will be invoked from soft interrupt context.
1227 *
1228 * The @timer->expires and @timer->function fields must be set prior
1229 * to calling this function.
1230 *
1231 * If @timer->function == NULL then the start operation is silently
1232 * discarded.
1233 *
1234 * If @timer->expires is already in the past @timer will be queued to
1235 * expire at the next timer tick.
1236 *
1237 * This can only operate on an inactive timer. Attempts to invoke this on
1238 * an active timer are rejected with a warning.
1239 */
1240void add_timer(struct timer_list *timer)
1241{
1242 if (WARN_ON_ONCE(timer_pending(timer)))
1243 return;
1244 __mod_timer(timer, timer->expires, MOD_TIMER_NOTPENDING);
1245}
1246EXPORT_SYMBOL(add_timer);
1247
1248/**
1249 * add_timer_on - Start a timer on a particular CPU
1250 * @timer: The timer to be started
1251 * @cpu: The CPU to start it on
1252 *
1253 * Same as add_timer() except that it starts the timer on the given CPU.
1254 *
1255 * See add_timer() for further details.
1256 */
1257void add_timer_on(struct timer_list *timer, int cpu)
1258{
1259 struct timer_base *new_base, *base;
1260 unsigned long flags;
1261
1262 debug_assert_init(timer);
1263
1264 if (WARN_ON_ONCE(timer_pending(timer)))
1265 return;
1266
1267 new_base = get_timer_cpu_base(timer->flags, cpu);
1268
1269 /*
1270 * If @timer was on a different CPU, it should be migrated with the
1271 * old base locked to prevent other operations proceeding with the
1272 * wrong base locked. See lock_timer_base().
1273 */
1274 base = lock_timer_base(timer, &flags);
1275 /*
1276 * Has @timer been shutdown? This needs to be evaluated while
1277 * holding base lock to prevent a race against the shutdown code.
1278 */
1279 if (!timer->function)
1280 goto out_unlock;
1281
1282 if (base != new_base) {
1283 timer->flags |= TIMER_MIGRATING;
1284
1285 raw_spin_unlock(&base->lock);
1286 base = new_base;
1287 raw_spin_lock(&base->lock);
1288 WRITE_ONCE(timer->flags,
1289 (timer->flags & ~TIMER_BASEMASK) | cpu);
1290 }
1291 forward_timer_base(base);
1292
1293 debug_timer_activate(timer);
1294 internal_add_timer(base, timer);
1295out_unlock:
1296 raw_spin_unlock_irqrestore(&base->lock, flags);
1297}
1298EXPORT_SYMBOL_GPL(add_timer_on);
1299
1300/**
1301 * __timer_delete - Internal function: Deactivate a timer
1302 * @timer: The timer to be deactivated
1303 * @shutdown: If true, this indicates that the timer is about to be
1304 * shutdown permanently.
1305 *
1306 * If @shutdown is true then @timer->function is set to NULL under the
1307 * timer base lock which prevents further rearming of the time. In that
1308 * case any attempt to rearm @timer after this function returns will be
1309 * silently ignored.
1310 *
1311 * Return:
1312 * * %0 - The timer was not pending
1313 * * %1 - The timer was pending and deactivated
1314 */
1315static int __timer_delete(struct timer_list *timer, bool shutdown)
1316{
1317 struct timer_base *base;
1318 unsigned long flags;
1319 int ret = 0;
1320
1321 debug_assert_init(timer);
1322
1323 /*
1324 * If @shutdown is set then the lock has to be taken whether the
1325 * timer is pending or not to protect against a concurrent rearm
1326 * which might hit between the lockless pending check and the lock
1327 * aquisition. By taking the lock it is ensured that such a newly
1328 * enqueued timer is dequeued and cannot end up with
1329 * timer->function == NULL in the expiry code.
1330 *
1331 * If timer->function is currently executed, then this makes sure
1332 * that the callback cannot requeue the timer.
1333 */
1334 if (timer_pending(timer) || shutdown) {
1335 base = lock_timer_base(timer, &flags);
1336 ret = detach_if_pending(timer, base, true);
1337 if (shutdown)
1338 timer->function = NULL;
1339 raw_spin_unlock_irqrestore(&base->lock, flags);
1340 }
1341
1342 return ret;
1343}
1344
1345/**
1346 * timer_delete - Deactivate a timer
1347 * @timer: The timer to be deactivated
1348 *
1349 * The function only deactivates a pending timer, but contrary to
1350 * timer_delete_sync() it does not take into account whether the timer's
1351 * callback function is concurrently executed on a different CPU or not.
1352 * It neither prevents rearming of the timer. If @timer can be rearmed
1353 * concurrently then the return value of this function is meaningless.
1354 *
1355 * Return:
1356 * * %0 - The timer was not pending
1357 * * %1 - The timer was pending and deactivated
1358 */
1359int timer_delete(struct timer_list *timer)
1360{
1361 return __timer_delete(timer, false);
1362}
1363EXPORT_SYMBOL(timer_delete);
1364
1365/**
1366 * timer_shutdown - Deactivate a timer and prevent rearming
1367 * @timer: The timer to be deactivated
1368 *
1369 * The function does not wait for an eventually running timer callback on a
1370 * different CPU but it prevents rearming of the timer. Any attempt to arm
1371 * @timer after this function returns will be silently ignored.
1372 *
1373 * This function is useful for teardown code and should only be used when
1374 * timer_shutdown_sync() cannot be invoked due to locking or context constraints.
1375 *
1376 * Return:
1377 * * %0 - The timer was not pending
1378 * * %1 - The timer was pending
1379 */
1380int timer_shutdown(struct timer_list *timer)
1381{
1382 return __timer_delete(timer, true);
1383}
1384EXPORT_SYMBOL_GPL(timer_shutdown);
1385
1386/**
1387 * __try_to_del_timer_sync - Internal function: Try to deactivate a timer
1388 * @timer: Timer to deactivate
1389 * @shutdown: If true, this indicates that the timer is about to be
1390 * shutdown permanently.
1391 *
1392 * If @shutdown is true then @timer->function is set to NULL under the
1393 * timer base lock which prevents further rearming of the timer. Any
1394 * attempt to rearm @timer after this function returns will be silently
1395 * ignored.
1396 *
1397 * This function cannot guarantee that the timer cannot be rearmed
1398 * right after dropping the base lock if @shutdown is false. That
1399 * needs to be prevented by the calling code if necessary.
1400 *
1401 * Return:
1402 * * %0 - The timer was not pending
1403 * * %1 - The timer was pending and deactivated
1404 * * %-1 - The timer callback function is running on a different CPU
1405 */
1406static int __try_to_del_timer_sync(struct timer_list *timer, bool shutdown)
1407{
1408 struct timer_base *base;
1409 unsigned long flags;
1410 int ret = -1;
1411
1412 debug_assert_init(timer);
1413
1414 base = lock_timer_base(timer, &flags);
1415
1416 if (base->running_timer != timer)
1417 ret = detach_if_pending(timer, base, true);
1418 if (shutdown)
1419 timer->function = NULL;
1420
1421 raw_spin_unlock_irqrestore(&base->lock, flags);
1422
1423 return ret;
1424}
1425
1426/**
1427 * try_to_del_timer_sync - Try to deactivate a timer
1428 * @timer: Timer to deactivate
1429 *
1430 * This function tries to deactivate a timer. On success the timer is not
1431 * queued and the timer callback function is not running on any CPU.
1432 *
1433 * This function does not guarantee that the timer cannot be rearmed right
1434 * after dropping the base lock. That needs to be prevented by the calling
1435 * code if necessary.
1436 *
1437 * Return:
1438 * * %0 - The timer was not pending
1439 * * %1 - The timer was pending and deactivated
1440 * * %-1 - The timer callback function is running on a different CPU
1441 */
1442int try_to_del_timer_sync(struct timer_list *timer)
1443{
1444 return __try_to_del_timer_sync(timer, false);
1445}
1446EXPORT_SYMBOL(try_to_del_timer_sync);
1447
1448#ifdef CONFIG_PREEMPT_RT
1449static __init void timer_base_init_expiry_lock(struct timer_base *base)
1450{
1451 spin_lock_init(&base->expiry_lock);
1452}
1453
1454static inline void timer_base_lock_expiry(struct timer_base *base)
1455{
1456 spin_lock(&base->expiry_lock);
1457}
1458
1459static inline void timer_base_unlock_expiry(struct timer_base *base)
1460{
1461 spin_unlock(&base->expiry_lock);
1462}
1463
1464/*
1465 * The counterpart to del_timer_wait_running().
1466 *
1467 * If there is a waiter for base->expiry_lock, then it was waiting for the
1468 * timer callback to finish. Drop expiry_lock and reacquire it. That allows
1469 * the waiter to acquire the lock and make progress.
1470 */
1471static void timer_sync_wait_running(struct timer_base *base)
1472{
1473 if (atomic_read(&base->timer_waiters)) {
1474 raw_spin_unlock_irq(&base->lock);
1475 spin_unlock(&base->expiry_lock);
1476 spin_lock(&base->expiry_lock);
1477 raw_spin_lock_irq(&base->lock);
1478 }
1479}
1480
1481/*
1482 * This function is called on PREEMPT_RT kernels when the fast path
1483 * deletion of a timer failed because the timer callback function was
1484 * running.
1485 *
1486 * This prevents priority inversion, if the softirq thread on a remote CPU
1487 * got preempted, and it prevents a life lock when the task which tries to
1488 * delete a timer preempted the softirq thread running the timer callback
1489 * function.
1490 */
1491static void del_timer_wait_running(struct timer_list *timer)
1492{
1493 u32 tf;
1494
1495 tf = READ_ONCE(timer->flags);
1496 if (!(tf & (TIMER_MIGRATING | TIMER_IRQSAFE))) {
1497 struct timer_base *base = get_timer_base(tf);
1498
1499 /*
1500 * Mark the base as contended and grab the expiry lock,
1501 * which is held by the softirq across the timer
1502 * callback. Drop the lock immediately so the softirq can
1503 * expire the next timer. In theory the timer could already
1504 * be running again, but that's more than unlikely and just
1505 * causes another wait loop.
1506 */
1507 atomic_inc(&base->timer_waiters);
1508 spin_lock_bh(&base->expiry_lock);
1509 atomic_dec(&base->timer_waiters);
1510 spin_unlock_bh(&base->expiry_lock);
1511 }
1512}
1513#else
1514static inline void timer_base_init_expiry_lock(struct timer_base *base) { }
1515static inline void timer_base_lock_expiry(struct timer_base *base) { }
1516static inline void timer_base_unlock_expiry(struct timer_base *base) { }
1517static inline void timer_sync_wait_running(struct timer_base *base) { }
1518static inline void del_timer_wait_running(struct timer_list *timer) { }
1519#endif
1520
1521/**
1522 * __timer_delete_sync - Internal function: Deactivate a timer and wait
1523 * for the handler to finish.
1524 * @timer: The timer to be deactivated
1525 * @shutdown: If true, @timer->function will be set to NULL under the
1526 * timer base lock which prevents rearming of @timer
1527 *
1528 * If @shutdown is not set the timer can be rearmed later. If the timer can
1529 * be rearmed concurrently, i.e. after dropping the base lock then the
1530 * return value is meaningless.
1531 *
1532 * If @shutdown is set then @timer->function is set to NULL under timer
1533 * base lock which prevents rearming of the timer. Any attempt to rearm
1534 * a shutdown timer is silently ignored.
1535 *
1536 * If the timer should be reused after shutdown it has to be initialized
1537 * again.
1538 *
1539 * Return:
1540 * * %0 - The timer was not pending
1541 * * %1 - The timer was pending and deactivated
1542 */
1543static int __timer_delete_sync(struct timer_list *timer, bool shutdown)
1544{
1545 int ret;
1546
1547#ifdef CONFIG_LOCKDEP
1548 unsigned long flags;
1549
1550 /*
1551 * If lockdep gives a backtrace here, please reference
1552 * the synchronization rules above.
1553 */
1554 local_irq_save(flags);
1555 lock_map_acquire(&timer->lockdep_map);
1556 lock_map_release(&timer->lockdep_map);
1557 local_irq_restore(flags);
1558#endif
1559 /*
1560 * don't use it in hardirq context, because it
1561 * could lead to deadlock.
1562 */
1563 WARN_ON(in_hardirq() && !(timer->flags & TIMER_IRQSAFE));
1564
1565 /*
1566 * Must be able to sleep on PREEMPT_RT because of the slowpath in
1567 * del_timer_wait_running().
1568 */
1569 if (IS_ENABLED(CONFIG_PREEMPT_RT) && !(timer->flags & TIMER_IRQSAFE))
1570 lockdep_assert_preemption_enabled();
1571
1572 do {
1573 ret = __try_to_del_timer_sync(timer, shutdown);
1574
1575 if (unlikely(ret < 0)) {
1576 del_timer_wait_running(timer);
1577 cpu_relax();
1578 }
1579 } while (ret < 0);
1580
1581 return ret;
1582}
1583
1584/**
1585 * timer_delete_sync - Deactivate a timer and wait for the handler to finish.
1586 * @timer: The timer to be deactivated
1587 *
1588 * Synchronization rules: Callers must prevent restarting of the timer,
1589 * otherwise this function is meaningless. It must not be called from
1590 * interrupt contexts unless the timer is an irqsafe one. The caller must
1591 * not hold locks which would prevent completion of the timer's callback
1592 * function. The timer's handler must not call add_timer_on(). Upon exit
1593 * the timer is not queued and the handler is not running on any CPU.
1594 *
1595 * For !irqsafe timers, the caller must not hold locks that are held in
1596 * interrupt context. Even if the lock has nothing to do with the timer in
1597 * question. Here's why::
1598 *
1599 * CPU0 CPU1
1600 * ---- ----
1601 * <SOFTIRQ>
1602 * call_timer_fn();
1603 * base->running_timer = mytimer;
1604 * spin_lock_irq(somelock);
1605 * <IRQ>
1606 * spin_lock(somelock);
1607 * timer_delete_sync(mytimer);
1608 * while (base->running_timer == mytimer);
1609 *
1610 * Now timer_delete_sync() will never return and never release somelock.
1611 * The interrupt on the other CPU is waiting to grab somelock but it has
1612 * interrupted the softirq that CPU0 is waiting to finish.
1613 *
1614 * This function cannot guarantee that the timer is not rearmed again by
1615 * some concurrent or preempting code, right after it dropped the base
1616 * lock. If there is the possibility of a concurrent rearm then the return
1617 * value of the function is meaningless.
1618 *
1619 * If such a guarantee is needed, e.g. for teardown situations then use
1620 * timer_shutdown_sync() instead.
1621 *
1622 * Return:
1623 * * %0 - The timer was not pending
1624 * * %1 - The timer was pending and deactivated
1625 */
1626int timer_delete_sync(struct timer_list *timer)
1627{
1628 return __timer_delete_sync(timer, false);
1629}
1630EXPORT_SYMBOL(timer_delete_sync);
1631
1632/**
1633 * timer_shutdown_sync - Shutdown a timer and prevent rearming
1634 * @timer: The timer to be shutdown
1635 *
1636 * When the function returns it is guaranteed that:
1637 * - @timer is not queued
1638 * - The callback function of @timer is not running
1639 * - @timer cannot be enqueued again. Any attempt to rearm
1640 * @timer is silently ignored.
1641 *
1642 * See timer_delete_sync() for synchronization rules.
1643 *
1644 * This function is useful for final teardown of an infrastructure where
1645 * the timer is subject to a circular dependency problem.
1646 *
1647 * A common pattern for this is a timer and a workqueue where the timer can
1648 * schedule work and work can arm the timer. On shutdown the workqueue must
1649 * be destroyed and the timer must be prevented from rearming. Unless the
1650 * code has conditionals like 'if (mything->in_shutdown)' to prevent that
1651 * there is no way to get this correct with timer_delete_sync().
1652 *
1653 * timer_shutdown_sync() is solving the problem. The correct ordering of
1654 * calls in this case is:
1655 *
1656 * timer_shutdown_sync(&mything->timer);
1657 * workqueue_destroy(&mything->workqueue);
1658 *
1659 * After this 'mything' can be safely freed.
1660 *
1661 * This obviously implies that the timer is not required to be functional
1662 * for the rest of the shutdown operation.
1663 *
1664 * Return:
1665 * * %0 - The timer was not pending
1666 * * %1 - The timer was pending
1667 */
1668int timer_shutdown_sync(struct timer_list *timer)
1669{
1670 return __timer_delete_sync(timer, true);
1671}
1672EXPORT_SYMBOL_GPL(timer_shutdown_sync);
1673
1674static void call_timer_fn(struct timer_list *timer,
1675 void (*fn)(struct timer_list *),
1676 unsigned long baseclk)
1677{
1678 int count = preempt_count();
1679
1680#ifdef CONFIG_LOCKDEP
1681 /*
1682 * It is permissible to free the timer from inside the
1683 * function that is called from it, this we need to take into
1684 * account for lockdep too. To avoid bogus "held lock freed"
1685 * warnings as well as problems when looking into
1686 * timer->lockdep_map, make a copy and use that here.
1687 */
1688 struct lockdep_map lockdep_map;
1689
1690 lockdep_copy_map(&lockdep_map, &timer->lockdep_map);
1691#endif
1692 /*
1693 * Couple the lock chain with the lock chain at
1694 * timer_delete_sync() by acquiring the lock_map around the fn()
1695 * call here and in timer_delete_sync().
1696 */
1697 lock_map_acquire(&lockdep_map);
1698
1699 trace_timer_expire_entry(timer, baseclk);
1700 fn(timer);
1701 trace_timer_expire_exit(timer);
1702
1703 lock_map_release(&lockdep_map);
1704
1705 if (count != preempt_count()) {
1706 WARN_ONCE(1, "timer: %pS preempt leak: %08x -> %08x\n",
1707 fn, count, preempt_count());
1708 /*
1709 * Restore the preempt count. That gives us a decent
1710 * chance to survive and extract information. If the
1711 * callback kept a lock held, bad luck, but not worse
1712 * than the BUG() we had.
1713 */
1714 preempt_count_set(count);
1715 }
1716}
1717
1718static void expire_timers(struct timer_base *base, struct hlist_head *head)
1719{
1720 /*
1721 * This value is required only for tracing. base->clk was
1722 * incremented directly before expire_timers was called. But expiry
1723 * is related to the old base->clk value.
1724 */
1725 unsigned long baseclk = base->clk - 1;
1726
1727 while (!hlist_empty(head)) {
1728 struct timer_list *timer;
1729 void (*fn)(struct timer_list *);
1730
1731 timer = hlist_entry(head->first, struct timer_list, entry);
1732
1733 base->running_timer = timer;
1734 detach_timer(timer, true);
1735
1736 fn = timer->function;
1737
1738 if (WARN_ON_ONCE(!fn)) {
1739 /* Should never happen. Emphasis on should! */
1740 base->running_timer = NULL;
1741 continue;
1742 }
1743
1744 if (timer->flags & TIMER_IRQSAFE) {
1745 raw_spin_unlock(&base->lock);
1746 call_timer_fn(timer, fn, baseclk);
1747 raw_spin_lock(&base->lock);
1748 base->running_timer = NULL;
1749 } else {
1750 raw_spin_unlock_irq(&base->lock);
1751 call_timer_fn(timer, fn, baseclk);
1752 raw_spin_lock_irq(&base->lock);
1753 base->running_timer = NULL;
1754 timer_sync_wait_running(base);
1755 }
1756 }
1757}
1758
1759static int collect_expired_timers(struct timer_base *base,
1760 struct hlist_head *heads)
1761{
1762 unsigned long clk = base->clk = base->next_expiry;
1763 struct hlist_head *vec;
1764 int i, levels = 0;
1765 unsigned int idx;
1766
1767 for (i = 0; i < LVL_DEPTH; i++) {
1768 idx = (clk & LVL_MASK) + i * LVL_SIZE;
1769
1770 if (__test_and_clear_bit(idx, base->pending_map)) {
1771 vec = base->vectors + idx;
1772 hlist_move_list(vec, heads++);
1773 levels++;
1774 }
1775 /* Is it time to look at the next level? */
1776 if (clk & LVL_CLK_MASK)
1777 break;
1778 /* Shift clock for the next level granularity */
1779 clk >>= LVL_CLK_SHIFT;
1780 }
1781 return levels;
1782}
1783
1784/*
1785 * Find the next pending bucket of a level. Search from level start (@offset)
1786 * + @clk upwards and if nothing there, search from start of the level
1787 * (@offset) up to @offset + clk.
1788 */
1789static int next_pending_bucket(struct timer_base *base, unsigned offset,
1790 unsigned clk)
1791{
1792 unsigned pos, start = offset + clk;
1793 unsigned end = offset + LVL_SIZE;
1794
1795 pos = find_next_bit(base->pending_map, end, start);
1796 if (pos < end)
1797 return pos - start;
1798
1799 pos = find_next_bit(base->pending_map, start, offset);
1800 return pos < start ? pos + LVL_SIZE - start : -1;
1801}
1802
1803/*
1804 * Search the first expiring timer in the various clock levels. Caller must
1805 * hold base->lock.
1806 */
1807static unsigned long __next_timer_interrupt(struct timer_base *base)
1808{
1809 unsigned long clk, next, adj;
1810 unsigned lvl, offset = 0;
1811
1812 next = base->clk + NEXT_TIMER_MAX_DELTA;
1813 clk = base->clk;
1814 for (lvl = 0; lvl < LVL_DEPTH; lvl++, offset += LVL_SIZE) {
1815 int pos = next_pending_bucket(base, offset, clk & LVL_MASK);
1816 unsigned long lvl_clk = clk & LVL_CLK_MASK;
1817
1818 if (pos >= 0) {
1819 unsigned long tmp = clk + (unsigned long) pos;
1820
1821 tmp <<= LVL_SHIFT(lvl);
1822 if (time_before(tmp, next))
1823 next = tmp;
1824
1825 /*
1826 * If the next expiration happens before we reach
1827 * the next level, no need to check further.
1828 */
1829 if (pos <= ((LVL_CLK_DIV - lvl_clk) & LVL_CLK_MASK))
1830 break;
1831 }
1832 /*
1833 * Clock for the next level. If the current level clock lower
1834 * bits are zero, we look at the next level as is. If not we
1835 * need to advance it by one because that's going to be the
1836 * next expiring bucket in that level. base->clk is the next
1837 * expiring jiffie. So in case of:
1838 *
1839 * LVL5 LVL4 LVL3 LVL2 LVL1 LVL0
1840 * 0 0 0 0 0 0
1841 *
1842 * we have to look at all levels @index 0. With
1843 *
1844 * LVL5 LVL4 LVL3 LVL2 LVL1 LVL0
1845 * 0 0 0 0 0 2
1846 *
1847 * LVL0 has the next expiring bucket @index 2. The upper
1848 * levels have the next expiring bucket @index 1.
1849 *
1850 * In case that the propagation wraps the next level the same
1851 * rules apply:
1852 *
1853 * LVL5 LVL4 LVL3 LVL2 LVL1 LVL0
1854 * 0 0 0 0 F 2
1855 *
1856 * So after looking at LVL0 we get:
1857 *
1858 * LVL5 LVL4 LVL3 LVL2 LVL1
1859 * 0 0 0 1 0
1860 *
1861 * So no propagation from LVL1 to LVL2 because that happened
1862 * with the add already, but then we need to propagate further
1863 * from LVL2 to LVL3.
1864 *
1865 * So the simple check whether the lower bits of the current
1866 * level are 0 or not is sufficient for all cases.
1867 */
1868 adj = lvl_clk ? 1 : 0;
1869 clk >>= LVL_CLK_SHIFT;
1870 clk += adj;
1871 }
1872
1873 base->next_expiry_recalc = false;
1874 base->timers_pending = !(next == base->clk + NEXT_TIMER_MAX_DELTA);
1875
1876 return next;
1877}
1878
1879#ifdef CONFIG_NO_HZ_COMMON
1880/*
1881 * Check, if the next hrtimer event is before the next timer wheel
1882 * event:
1883 */
1884static u64 cmp_next_hrtimer_event(u64 basem, u64 expires)
1885{
1886 u64 nextevt = hrtimer_get_next_event();
1887
1888 /*
1889 * If high resolution timers are enabled
1890 * hrtimer_get_next_event() returns KTIME_MAX.
1891 */
1892 if (expires <= nextevt)
1893 return expires;
1894
1895 /*
1896 * If the next timer is already expired, return the tick base
1897 * time so the tick is fired immediately.
1898 */
1899 if (nextevt <= basem)
1900 return basem;
1901
1902 /*
1903 * Round up to the next jiffie. High resolution timers are
1904 * off, so the hrtimers are expired in the tick and we need to
1905 * make sure that this tick really expires the timer to avoid
1906 * a ping pong of the nohz stop code.
1907 *
1908 * Use DIV_ROUND_UP_ULL to prevent gcc calling __divdi3
1909 */
1910 return DIV_ROUND_UP_ULL(nextevt, TICK_NSEC) * TICK_NSEC;
1911}
1912
1913/**
1914 * get_next_timer_interrupt - return the time (clock mono) of the next timer
1915 * @basej: base time jiffies
1916 * @basem: base time clock monotonic
1917 *
1918 * Returns the tick aligned clock monotonic time of the next pending
1919 * timer or KTIME_MAX if no timer is pending.
1920 */
1921u64 get_next_timer_interrupt(unsigned long basej, u64 basem)
1922{
1923 struct timer_base *base = this_cpu_ptr(&timer_bases[BASE_STD]);
1924 u64 expires = KTIME_MAX;
1925 unsigned long nextevt;
1926
1927 /*
1928 * Pretend that there is no timer pending if the cpu is offline.
1929 * Possible pending timers will be migrated later to an active cpu.
1930 */
1931 if (cpu_is_offline(smp_processor_id()))
1932 return expires;
1933
1934 raw_spin_lock(&base->lock);
1935 if (base->next_expiry_recalc)
1936 base->next_expiry = __next_timer_interrupt(base);
1937 nextevt = base->next_expiry;
1938
1939 /*
1940 * We have a fresh next event. Check whether we can forward the
1941 * base. We can only do that when @basej is past base->clk
1942 * otherwise we might rewind base->clk.
1943 */
1944 if (time_after(basej, base->clk)) {
1945 if (time_after(nextevt, basej))
1946 base->clk = basej;
1947 else if (time_after(nextevt, base->clk))
1948 base->clk = nextevt;
1949 }
1950
1951 if (time_before_eq(nextevt, basej)) {
1952 expires = basem;
1953 base->is_idle = false;
1954 } else {
1955 if (base->timers_pending)
1956 expires = basem + (u64)(nextevt - basej) * TICK_NSEC;
1957 /*
1958 * If we expect to sleep more than a tick, mark the base idle.
1959 * Also the tick is stopped so any added timer must forward
1960 * the base clk itself to keep granularity small. This idle
1961 * logic is only maintained for the BASE_STD base, deferrable
1962 * timers may still see large granularity skew (by design).
1963 */
1964 if ((expires - basem) > TICK_NSEC)
1965 base->is_idle = true;
1966 }
1967 raw_spin_unlock(&base->lock);
1968
1969 return cmp_next_hrtimer_event(basem, expires);
1970}
1971
1972/**
1973 * timer_clear_idle - Clear the idle state of the timer base
1974 *
1975 * Called with interrupts disabled
1976 */
1977void timer_clear_idle(void)
1978{
1979 struct timer_base *base = this_cpu_ptr(&timer_bases[BASE_STD]);
1980
1981 /*
1982 * We do this unlocked. The worst outcome is a remote enqueue sending
1983 * a pointless IPI, but taking the lock would just make the window for
1984 * sending the IPI a few instructions smaller for the cost of taking
1985 * the lock in the exit from idle path.
1986 */
1987 base->is_idle = false;
1988}
1989#endif
1990
1991/**
1992 * __run_timers - run all expired timers (if any) on this CPU.
1993 * @base: the timer vector to be processed.
1994 */
1995static inline void __run_timers(struct timer_base *base)
1996{
1997 struct hlist_head heads[LVL_DEPTH];
1998 int levels;
1999
2000 if (time_before(jiffies, base->next_expiry))
2001 return;
2002
2003 timer_base_lock_expiry(base);
2004 raw_spin_lock_irq(&base->lock);
2005
2006 while (time_after_eq(jiffies, base->clk) &&
2007 time_after_eq(jiffies, base->next_expiry)) {
2008 levels = collect_expired_timers(base, heads);
2009 /*
2010 * The two possible reasons for not finding any expired
2011 * timer at this clk are that all matching timers have been
2012 * dequeued or no timer has been queued since
2013 * base::next_expiry was set to base::clk +
2014 * NEXT_TIMER_MAX_DELTA.
2015 */
2016 WARN_ON_ONCE(!levels && !base->next_expiry_recalc
2017 && base->timers_pending);
2018 base->clk++;
2019 base->next_expiry = __next_timer_interrupt(base);
2020
2021 while (levels--)
2022 expire_timers(base, heads + levels);
2023 }
2024 raw_spin_unlock_irq(&base->lock);
2025 timer_base_unlock_expiry(base);
2026}
2027
2028/*
2029 * This function runs timers and the timer-tq in bottom half context.
2030 */
2031static __latent_entropy void run_timer_softirq(struct softirq_action *h)
2032{
2033 struct timer_base *base = this_cpu_ptr(&timer_bases[BASE_STD]);
2034
2035 __run_timers(base);
2036 if (IS_ENABLED(CONFIG_NO_HZ_COMMON))
2037 __run_timers(this_cpu_ptr(&timer_bases[BASE_DEF]));
2038}
2039
2040/*
2041 * Called by the local, per-CPU timer interrupt on SMP.
2042 */
2043static void run_local_timers(void)
2044{
2045 struct timer_base *base = this_cpu_ptr(&timer_bases[BASE_STD]);
2046
2047 hrtimer_run_queues();
2048 /* Raise the softirq only if required. */
2049 if (time_before(jiffies, base->next_expiry)) {
2050 if (!IS_ENABLED(CONFIG_NO_HZ_COMMON))
2051 return;
2052 /* CPU is awake, so check the deferrable base. */
2053 base++;
2054 if (time_before(jiffies, base->next_expiry))
2055 return;
2056 }
2057 raise_softirq(TIMER_SOFTIRQ);
2058}
2059
2060/*
2061 * Called from the timer interrupt handler to charge one tick to the current
2062 * process. user_tick is 1 if the tick is user time, 0 for system.
2063 */
2064void update_process_times(int user_tick)
2065{
2066 struct task_struct *p = current;
2067
2068 /* Note: this timer irq context must be accounted for as well. */
2069 account_process_tick(p, user_tick);
2070 run_local_timers();
2071 rcu_sched_clock_irq(user_tick);
2072#ifdef CONFIG_IRQ_WORK
2073 if (in_irq())
2074 irq_work_tick();
2075#endif
2076 scheduler_tick();
2077 if (IS_ENABLED(CONFIG_POSIX_TIMERS))
2078 run_posix_cpu_timers();
2079}
2080
2081/*
2082 * Since schedule_timeout()'s timer is defined on the stack, it must store
2083 * the target task on the stack as well.
2084 */
2085struct process_timer {
2086 struct timer_list timer;
2087 struct task_struct *task;
2088};
2089
2090static void process_timeout(struct timer_list *t)
2091{
2092 struct process_timer *timeout = from_timer(timeout, t, timer);
2093
2094 wake_up_process(timeout->task);
2095}
2096
2097/**
2098 * schedule_timeout - sleep until timeout
2099 * @timeout: timeout value in jiffies
2100 *
2101 * Make the current task sleep until @timeout jiffies have elapsed.
2102 * The function behavior depends on the current task state
2103 * (see also set_current_state() description):
2104 *
2105 * %TASK_RUNNING - the scheduler is called, but the task does not sleep
2106 * at all. That happens because sched_submit_work() does nothing for
2107 * tasks in %TASK_RUNNING state.
2108 *
2109 * %TASK_UNINTERRUPTIBLE - at least @timeout jiffies are guaranteed to
2110 * pass before the routine returns unless the current task is explicitly
2111 * woken up, (e.g. by wake_up_process()).
2112 *
2113 * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
2114 * delivered to the current task or the current task is explicitly woken
2115 * up.
2116 *
2117 * The current task state is guaranteed to be %TASK_RUNNING when this
2118 * routine returns.
2119 *
2120 * Specifying a @timeout value of %MAX_SCHEDULE_TIMEOUT will schedule
2121 * the CPU away without a bound on the timeout. In this case the return
2122 * value will be %MAX_SCHEDULE_TIMEOUT.
2123 *
2124 * Returns 0 when the timer has expired otherwise the remaining time in
2125 * jiffies will be returned. In all cases the return value is guaranteed
2126 * to be non-negative.
2127 */
2128signed long __sched schedule_timeout(signed long timeout)
2129{
2130 struct process_timer timer;
2131 unsigned long expire;
2132
2133 switch (timeout)
2134 {
2135 case MAX_SCHEDULE_TIMEOUT:
2136 /*
2137 * These two special cases are useful to be comfortable
2138 * in the caller. Nothing more. We could take
2139 * MAX_SCHEDULE_TIMEOUT from one of the negative value
2140 * but I' d like to return a valid offset (>=0) to allow
2141 * the caller to do everything it want with the retval.
2142 */
2143 schedule();
2144 goto out;
2145 default:
2146 /*
2147 * Another bit of PARANOID. Note that the retval will be
2148 * 0 since no piece of kernel is supposed to do a check
2149 * for a negative retval of schedule_timeout() (since it
2150 * should never happens anyway). You just have the printk()
2151 * that will tell you if something is gone wrong and where.
2152 */
2153 if (timeout < 0) {
2154 printk(KERN_ERR "schedule_timeout: wrong timeout "
2155 "value %lx\n", timeout);
2156 dump_stack();
2157 __set_current_state(TASK_RUNNING);
2158 goto out;
2159 }
2160 }
2161
2162 expire = timeout + jiffies;
2163
2164 timer.task = current;
2165 timer_setup_on_stack(&timer.timer, process_timeout, 0);
2166 __mod_timer(&timer.timer, expire, MOD_TIMER_NOTPENDING);
2167 schedule();
2168 del_timer_sync(&timer.timer);
2169
2170 /* Remove the timer from the object tracker */
2171 destroy_timer_on_stack(&timer.timer);
2172
2173 timeout = expire - jiffies;
2174
2175 out:
2176 return timeout < 0 ? 0 : timeout;
2177}
2178EXPORT_SYMBOL(schedule_timeout);
2179
2180/*
2181 * We can use __set_current_state() here because schedule_timeout() calls
2182 * schedule() unconditionally.
2183 */
2184signed long __sched schedule_timeout_interruptible(signed long timeout)
2185{
2186 __set_current_state(TASK_INTERRUPTIBLE);
2187 return schedule_timeout(timeout);
2188}
2189EXPORT_SYMBOL(schedule_timeout_interruptible);
2190
2191signed long __sched schedule_timeout_killable(signed long timeout)
2192{
2193 __set_current_state(TASK_KILLABLE);
2194 return schedule_timeout(timeout);
2195}
2196EXPORT_SYMBOL(schedule_timeout_killable);
2197
2198signed long __sched schedule_timeout_uninterruptible(signed long timeout)
2199{
2200 __set_current_state(TASK_UNINTERRUPTIBLE);
2201 return schedule_timeout(timeout);
2202}
2203EXPORT_SYMBOL(schedule_timeout_uninterruptible);
2204
2205/*
2206 * Like schedule_timeout_uninterruptible(), except this task will not contribute
2207 * to load average.
2208 */
2209signed long __sched schedule_timeout_idle(signed long timeout)
2210{
2211 __set_current_state(TASK_IDLE);
2212 return schedule_timeout(timeout);
2213}
2214EXPORT_SYMBOL(schedule_timeout_idle);
2215
2216#ifdef CONFIG_HOTPLUG_CPU
2217static void migrate_timer_list(struct timer_base *new_base, struct hlist_head *head)
2218{
2219 struct timer_list *timer;
2220 int cpu = new_base->cpu;
2221
2222 while (!hlist_empty(head)) {
2223 timer = hlist_entry(head->first, struct timer_list, entry);
2224 detach_timer(timer, false);
2225 timer->flags = (timer->flags & ~TIMER_BASEMASK) | cpu;
2226 internal_add_timer(new_base, timer);
2227 }
2228}
2229
2230int timers_prepare_cpu(unsigned int cpu)
2231{
2232 struct timer_base *base;
2233 int b;
2234
2235 for (b = 0; b < NR_BASES; b++) {
2236 base = per_cpu_ptr(&timer_bases[b], cpu);
2237 base->clk = jiffies;
2238 base->next_expiry = base->clk + NEXT_TIMER_MAX_DELTA;
2239 base->next_expiry_recalc = false;
2240 base->timers_pending = false;
2241 base->is_idle = false;
2242 }
2243 return 0;
2244}
2245
2246int timers_dead_cpu(unsigned int cpu)
2247{
2248 struct timer_base *old_base;
2249 struct timer_base *new_base;
2250 int b, i;
2251
2252 for (b = 0; b < NR_BASES; b++) {
2253 old_base = per_cpu_ptr(&timer_bases[b], cpu);
2254 new_base = get_cpu_ptr(&timer_bases[b]);
2255 /*
2256 * The caller is globally serialized and nobody else
2257 * takes two locks at once, deadlock is not possible.
2258 */
2259 raw_spin_lock_irq(&new_base->lock);
2260 raw_spin_lock_nested(&old_base->lock, SINGLE_DEPTH_NESTING);
2261
2262 /*
2263 * The current CPUs base clock might be stale. Update it
2264 * before moving the timers over.
2265 */
2266 forward_timer_base(new_base);
2267
2268 WARN_ON_ONCE(old_base->running_timer);
2269 old_base->running_timer = NULL;
2270
2271 for (i = 0; i < WHEEL_SIZE; i++)
2272 migrate_timer_list(new_base, old_base->vectors + i);
2273
2274 raw_spin_unlock(&old_base->lock);
2275 raw_spin_unlock_irq(&new_base->lock);
2276 put_cpu_ptr(&timer_bases);
2277 }
2278 return 0;
2279}
2280
2281#endif /* CONFIG_HOTPLUG_CPU */
2282
2283static void __init init_timer_cpu(int cpu)
2284{
2285 struct timer_base *base;
2286 int i;
2287
2288 for (i = 0; i < NR_BASES; i++) {
2289 base = per_cpu_ptr(&timer_bases[i], cpu);
2290 base->cpu = cpu;
2291 raw_spin_lock_init(&base->lock);
2292 base->clk = jiffies;
2293 base->next_expiry = base->clk + NEXT_TIMER_MAX_DELTA;
2294 timer_base_init_expiry_lock(base);
2295 }
2296}
2297
2298static void __init init_timer_cpus(void)
2299{
2300 int cpu;
2301
2302 for_each_possible_cpu(cpu)
2303 init_timer_cpu(cpu);
2304}
2305
2306void __init init_timers(void)
2307{
2308 init_timer_cpus();
2309 posix_cputimers_init_work();
2310 open_softirq(TIMER_SOFTIRQ, run_timer_softirq);
2311}
2312
2313/**
2314 * msleep - sleep safely even with waitqueue interruptions
2315 * @msecs: Time in milliseconds to sleep for
2316 */
2317void msleep(unsigned int msecs)
2318{
2319 unsigned long timeout = msecs_to_jiffies(msecs) + 1;
2320
2321 while (timeout)
2322 timeout = schedule_timeout_uninterruptible(timeout);
2323}
2324
2325EXPORT_SYMBOL(msleep);
2326
2327/**
2328 * msleep_interruptible - sleep waiting for signals
2329 * @msecs: Time in milliseconds to sleep for
2330 */
2331unsigned long msleep_interruptible(unsigned int msecs)
2332{
2333 unsigned long timeout = msecs_to_jiffies(msecs) + 1;
2334
2335 while (timeout && !signal_pending(current))
2336 timeout = schedule_timeout_interruptible(timeout);
2337 return jiffies_to_msecs(timeout);
2338}
2339
2340EXPORT_SYMBOL(msleep_interruptible);
2341
2342/**
2343 * usleep_range_state - Sleep for an approximate time in a given state
2344 * @min: Minimum time in usecs to sleep
2345 * @max: Maximum time in usecs to sleep
2346 * @state: State of the current task that will be while sleeping
2347 *
2348 * In non-atomic context where the exact wakeup time is flexible, use
2349 * usleep_range_state() instead of udelay(). The sleep improves responsiveness
2350 * by avoiding the CPU-hogging busy-wait of udelay(), and the range reduces
2351 * power usage by allowing hrtimers to take advantage of an already-
2352 * scheduled interrupt instead of scheduling a new one just for this sleep.
2353 */
2354void __sched usleep_range_state(unsigned long min, unsigned long max,
2355 unsigned int state)
2356{
2357 ktime_t exp = ktime_add_us(ktime_get(), min);
2358 u64 delta = (u64)(max - min) * NSEC_PER_USEC;
2359
2360 for (;;) {
2361 __set_current_state(state);
2362 /* Do not return before the requested sleep time has elapsed */
2363 if (!schedule_hrtimeout_range(&exp, delta, HRTIMER_MODE_ABS))
2364 break;
2365 }
2366}
2367EXPORT_SYMBOL(usleep_range_state);
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Kernel internal timers
4 *
5 * Copyright (C) 1991, 1992 Linus Torvalds
6 *
7 * 1997-01-28 Modified by Finn Arne Gangstad to make timers scale better.
8 *
9 * 1997-09-10 Updated NTP code according to technical memorandum Jan '96
10 * "A Kernel Model for Precision Timekeeping" by Dave Mills
11 * 1998-12-24 Fixed a xtime SMP race (we need the xtime_lock rw spinlock to
12 * serialize accesses to xtime/lost_ticks).
13 * Copyright (C) 1998 Andrea Arcangeli
14 * 1999-03-10 Improved NTP compatibility by Ulrich Windl
15 * 2002-05-31 Move sys_sysinfo here and make its locking sane, Robert Love
16 * 2000-10-05 Implemented scalable SMP per-CPU timer handling.
17 * Copyright (C) 2000, 2001, 2002 Ingo Molnar
18 * Designed by David S. Miller, Alexey Kuznetsov and Ingo Molnar
19 */
20
21#include <linux/kernel_stat.h>
22#include <linux/export.h>
23#include <linux/interrupt.h>
24#include <linux/percpu.h>
25#include <linux/init.h>
26#include <linux/mm.h>
27#include <linux/swap.h>
28#include <linux/pid_namespace.h>
29#include <linux/notifier.h>
30#include <linux/thread_info.h>
31#include <linux/time.h>
32#include <linux/jiffies.h>
33#include <linux/posix-timers.h>
34#include <linux/cpu.h>
35#include <linux/syscalls.h>
36#include <linux/delay.h>
37#include <linux/tick.h>
38#include <linux/kallsyms.h>
39#include <linux/irq_work.h>
40#include <linux/sched/signal.h>
41#include <linux/sched/sysctl.h>
42#include <linux/sched/nohz.h>
43#include <linux/sched/debug.h>
44#include <linux/slab.h>
45#include <linux/compat.h>
46#include <linux/random.h>
47#include <linux/sysctl.h>
48
49#include <linux/uaccess.h>
50#include <asm/unistd.h>
51#include <asm/div64.h>
52#include <asm/timex.h>
53#include <asm/io.h>
54
55#include "tick-internal.h"
56#include "timer_migration.h"
57
58#define CREATE_TRACE_POINTS
59#include <trace/events/timer.h>
60
61__visible u64 jiffies_64 __cacheline_aligned_in_smp = INITIAL_JIFFIES;
62
63EXPORT_SYMBOL(jiffies_64);
64
65/*
66 * The timer wheel has LVL_DEPTH array levels. Each level provides an array of
67 * LVL_SIZE buckets. Each level is driven by its own clock and therefore each
68 * level has a different granularity.
69 *
70 * The level granularity is: LVL_CLK_DIV ^ level
71 * The level clock frequency is: HZ / (LVL_CLK_DIV ^ level)
72 *
73 * The array level of a newly armed timer depends on the relative expiry
74 * time. The farther the expiry time is away the higher the array level and
75 * therefore the granularity becomes.
76 *
77 * Contrary to the original timer wheel implementation, which aims for 'exact'
78 * expiry of the timers, this implementation removes the need for recascading
79 * the timers into the lower array levels. The previous 'classic' timer wheel
80 * implementation of the kernel already violated the 'exact' expiry by adding
81 * slack to the expiry time to provide batched expiration. The granularity
82 * levels provide implicit batching.
83 *
84 * This is an optimization of the original timer wheel implementation for the
85 * majority of the timer wheel use cases: timeouts. The vast majority of
86 * timeout timers (networking, disk I/O ...) are canceled before expiry. If
87 * the timeout expires it indicates that normal operation is disturbed, so it
88 * does not matter much whether the timeout comes with a slight delay.
89 *
90 * The only exception to this are networking timers with a small expiry
91 * time. They rely on the granularity. Those fit into the first wheel level,
92 * which has HZ granularity.
93 *
94 * We don't have cascading anymore. timers with a expiry time above the
95 * capacity of the last wheel level are force expired at the maximum timeout
96 * value of the last wheel level. From data sampling we know that the maximum
97 * value observed is 5 days (network connection tracking), so this should not
98 * be an issue.
99 *
100 * The currently chosen array constants values are a good compromise between
101 * array size and granularity.
102 *
103 * This results in the following granularity and range levels:
104 *
105 * HZ 1000 steps
106 * Level Offset Granularity Range
107 * 0 0 1 ms 0 ms - 63 ms
108 * 1 64 8 ms 64 ms - 511 ms
109 * 2 128 64 ms 512 ms - 4095 ms (512ms - ~4s)
110 * 3 192 512 ms 4096 ms - 32767 ms (~4s - ~32s)
111 * 4 256 4096 ms (~4s) 32768 ms - 262143 ms (~32s - ~4m)
112 * 5 320 32768 ms (~32s) 262144 ms - 2097151 ms (~4m - ~34m)
113 * 6 384 262144 ms (~4m) 2097152 ms - 16777215 ms (~34m - ~4h)
114 * 7 448 2097152 ms (~34m) 16777216 ms - 134217727 ms (~4h - ~1d)
115 * 8 512 16777216 ms (~4h) 134217728 ms - 1073741822 ms (~1d - ~12d)
116 *
117 * HZ 300
118 * Level Offset Granularity Range
119 * 0 0 3 ms 0 ms - 210 ms
120 * 1 64 26 ms 213 ms - 1703 ms (213ms - ~1s)
121 * 2 128 213 ms 1706 ms - 13650 ms (~1s - ~13s)
122 * 3 192 1706 ms (~1s) 13653 ms - 109223 ms (~13s - ~1m)
123 * 4 256 13653 ms (~13s) 109226 ms - 873810 ms (~1m - ~14m)
124 * 5 320 109226 ms (~1m) 873813 ms - 6990503 ms (~14m - ~1h)
125 * 6 384 873813 ms (~14m) 6990506 ms - 55924050 ms (~1h - ~15h)
126 * 7 448 6990506 ms (~1h) 55924053 ms - 447392423 ms (~15h - ~5d)
127 * 8 512 55924053 ms (~15h) 447392426 ms - 3579139406 ms (~5d - ~41d)
128 *
129 * HZ 250
130 * Level Offset Granularity Range
131 * 0 0 4 ms 0 ms - 255 ms
132 * 1 64 32 ms 256 ms - 2047 ms (256ms - ~2s)
133 * 2 128 256 ms 2048 ms - 16383 ms (~2s - ~16s)
134 * 3 192 2048 ms (~2s) 16384 ms - 131071 ms (~16s - ~2m)
135 * 4 256 16384 ms (~16s) 131072 ms - 1048575 ms (~2m - ~17m)
136 * 5 320 131072 ms (~2m) 1048576 ms - 8388607 ms (~17m - ~2h)
137 * 6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
138 * 7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
139 * 8 512 67108864 ms (~18h) 536870912 ms - 4294967288 ms (~6d - ~49d)
140 *
141 * HZ 100
142 * Level Offset Granularity Range
143 * 0 0 10 ms 0 ms - 630 ms
144 * 1 64 80 ms 640 ms - 5110 ms (640ms - ~5s)
145 * 2 128 640 ms 5120 ms - 40950 ms (~5s - ~40s)
146 * 3 192 5120 ms (~5s) 40960 ms - 327670 ms (~40s - ~5m)
147 * 4 256 40960 ms (~40s) 327680 ms - 2621430 ms (~5m - ~43m)
148 * 5 320 327680 ms (~5m) 2621440 ms - 20971510 ms (~43m - ~5h)
149 * 6 384 2621440 ms (~43m) 20971520 ms - 167772150 ms (~5h - ~1d)
150 * 7 448 20971520 ms (~5h) 167772160 ms - 1342177270 ms (~1d - ~15d)
151 */
152
153/* Clock divisor for the next level */
154#define LVL_CLK_SHIFT 3
155#define LVL_CLK_DIV (1UL << LVL_CLK_SHIFT)
156#define LVL_CLK_MASK (LVL_CLK_DIV - 1)
157#define LVL_SHIFT(n) ((n) * LVL_CLK_SHIFT)
158#define LVL_GRAN(n) (1UL << LVL_SHIFT(n))
159
160/*
161 * The time start value for each level to select the bucket at enqueue
162 * time. We start from the last possible delta of the previous level
163 * so that we can later add an extra LVL_GRAN(n) to n (see calc_index()).
164 */
165#define LVL_START(n) ((LVL_SIZE - 1) << (((n) - 1) * LVL_CLK_SHIFT))
166
167/* Size of each clock level */
168#define LVL_BITS 6
169#define LVL_SIZE (1UL << LVL_BITS)
170#define LVL_MASK (LVL_SIZE - 1)
171#define LVL_OFFS(n) ((n) * LVL_SIZE)
172
173/* Level depth */
174#if HZ > 100
175# define LVL_DEPTH 9
176# else
177# define LVL_DEPTH 8
178#endif
179
180/* The cutoff (max. capacity of the wheel) */
181#define WHEEL_TIMEOUT_CUTOFF (LVL_START(LVL_DEPTH))
182#define WHEEL_TIMEOUT_MAX (WHEEL_TIMEOUT_CUTOFF - LVL_GRAN(LVL_DEPTH - 1))
183
184/*
185 * The resulting wheel size. If NOHZ is configured we allocate two
186 * wheels so we have a separate storage for the deferrable timers.
187 */
188#define WHEEL_SIZE (LVL_SIZE * LVL_DEPTH)
189
190#ifdef CONFIG_NO_HZ_COMMON
191/*
192 * If multiple bases need to be locked, use the base ordering for lock
193 * nesting, i.e. lowest number first.
194 */
195# define NR_BASES 3
196# define BASE_LOCAL 0
197# define BASE_GLOBAL 1
198# define BASE_DEF 2
199#else
200# define NR_BASES 1
201# define BASE_LOCAL 0
202# define BASE_GLOBAL 0
203# define BASE_DEF 0
204#endif
205
206/**
207 * struct timer_base - Per CPU timer base (number of base depends on config)
208 * @lock: Lock protecting the timer_base
209 * @running_timer: When expiring timers, the lock is dropped. To make
210 * sure not to race against deleting/modifying a
211 * currently running timer, the pointer is set to the
212 * timer, which expires at the moment. If no timer is
213 * running, the pointer is NULL.
214 * @expiry_lock: PREEMPT_RT only: Lock is taken in softirq around
215 * timer expiry callback execution and when trying to
216 * delete a running timer and it wasn't successful in
217 * the first glance. It prevents priority inversion
218 * when callback was preempted on a remote CPU and a
219 * caller tries to delete the running timer. It also
220 * prevents a life lock, when the task which tries to
221 * delete a timer preempted the softirq thread which
222 * is running the timer callback function.
223 * @timer_waiters: PREEMPT_RT only: Tells, if there is a waiter
224 * waiting for the end of the timer callback function
225 * execution.
226 * @clk: clock of the timer base; is updated before enqueue
227 * of a timer; during expiry, it is 1 offset ahead of
228 * jiffies to avoid endless requeuing to current
229 * jiffies
230 * @next_expiry: expiry value of the first timer; it is updated when
231 * finding the next timer and during enqueue; the
232 * value is not valid, when next_expiry_recalc is set
233 * @cpu: Number of CPU the timer base belongs to
234 * @next_expiry_recalc: States, whether a recalculation of next_expiry is
235 * required. Value is set true, when a timer was
236 * deleted.
237 * @is_idle: Is set, when timer_base is idle. It is triggered by NOHZ
238 * code. This state is only used in standard
239 * base. Deferrable timers, which are enqueued remotely
240 * never wake up an idle CPU. So no matter of supporting it
241 * for this base.
242 * @timers_pending: Is set, when a timer is pending in the base. It is only
243 * reliable when next_expiry_recalc is not set.
244 * @pending_map: bitmap of the timer wheel; each bit reflects a
245 * bucket of the wheel. When a bit is set, at least a
246 * single timer is enqueued in the related bucket.
247 * @vectors: Array of lists; Each array member reflects a bucket
248 * of the timer wheel. The list contains all timers
249 * which are enqueued into a specific bucket.
250 */
251struct timer_base {
252 raw_spinlock_t lock;
253 struct timer_list *running_timer;
254#ifdef CONFIG_PREEMPT_RT
255 spinlock_t expiry_lock;
256 atomic_t timer_waiters;
257#endif
258 unsigned long clk;
259 unsigned long next_expiry;
260 unsigned int cpu;
261 bool next_expiry_recalc;
262 bool is_idle;
263 bool timers_pending;
264 DECLARE_BITMAP(pending_map, WHEEL_SIZE);
265 struct hlist_head vectors[WHEEL_SIZE];
266} ____cacheline_aligned;
267
268static DEFINE_PER_CPU(struct timer_base, timer_bases[NR_BASES]);
269
270#ifdef CONFIG_NO_HZ_COMMON
271
272static DEFINE_STATIC_KEY_FALSE(timers_nohz_active);
273static DEFINE_MUTEX(timer_keys_mutex);
274
275static void timer_update_keys(struct work_struct *work);
276static DECLARE_WORK(timer_update_work, timer_update_keys);
277
278#ifdef CONFIG_SMP
279static unsigned int sysctl_timer_migration = 1;
280
281DEFINE_STATIC_KEY_FALSE(timers_migration_enabled);
282
283static void timers_update_migration(void)
284{
285 if (sysctl_timer_migration && tick_nohz_active)
286 static_branch_enable(&timers_migration_enabled);
287 else
288 static_branch_disable(&timers_migration_enabled);
289}
290
291#ifdef CONFIG_SYSCTL
292static int timer_migration_handler(struct ctl_table *table, int write,
293 void *buffer, size_t *lenp, loff_t *ppos)
294{
295 int ret;
296
297 mutex_lock(&timer_keys_mutex);
298 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
299 if (!ret && write)
300 timers_update_migration();
301 mutex_unlock(&timer_keys_mutex);
302 return ret;
303}
304
305static struct ctl_table timer_sysctl[] = {
306 {
307 .procname = "timer_migration",
308 .data = &sysctl_timer_migration,
309 .maxlen = sizeof(unsigned int),
310 .mode = 0644,
311 .proc_handler = timer_migration_handler,
312 .extra1 = SYSCTL_ZERO,
313 .extra2 = SYSCTL_ONE,
314 },
315 {}
316};
317
318static int __init timer_sysctl_init(void)
319{
320 register_sysctl("kernel", timer_sysctl);
321 return 0;
322}
323device_initcall(timer_sysctl_init);
324#endif /* CONFIG_SYSCTL */
325#else /* CONFIG_SMP */
326static inline void timers_update_migration(void) { }
327#endif /* !CONFIG_SMP */
328
329static void timer_update_keys(struct work_struct *work)
330{
331 mutex_lock(&timer_keys_mutex);
332 timers_update_migration();
333 static_branch_enable(&timers_nohz_active);
334 mutex_unlock(&timer_keys_mutex);
335}
336
337void timers_update_nohz(void)
338{
339 schedule_work(&timer_update_work);
340}
341
342static inline bool is_timers_nohz_active(void)
343{
344 return static_branch_unlikely(&timers_nohz_active);
345}
346#else
347static inline bool is_timers_nohz_active(void) { return false; }
348#endif /* NO_HZ_COMMON */
349
350static unsigned long round_jiffies_common(unsigned long j, int cpu,
351 bool force_up)
352{
353 int rem;
354 unsigned long original = j;
355
356 /*
357 * We don't want all cpus firing their timers at once hitting the
358 * same lock or cachelines, so we skew each extra cpu with an extra
359 * 3 jiffies. This 3 jiffies came originally from the mm/ code which
360 * already did this.
361 * The skew is done by adding 3*cpunr, then round, then subtract this
362 * extra offset again.
363 */
364 j += cpu * 3;
365
366 rem = j % HZ;
367
368 /*
369 * If the target jiffie is just after a whole second (which can happen
370 * due to delays of the timer irq, long irq off times etc etc) then
371 * we should round down to the whole second, not up. Use 1/4th second
372 * as cutoff for this rounding as an extreme upper bound for this.
373 * But never round down if @force_up is set.
374 */
375 if (rem < HZ/4 && !force_up) /* round down */
376 j = j - rem;
377 else /* round up */
378 j = j - rem + HZ;
379
380 /* now that we have rounded, subtract the extra skew again */
381 j -= cpu * 3;
382
383 /*
384 * Make sure j is still in the future. Otherwise return the
385 * unmodified value.
386 */
387 return time_is_after_jiffies(j) ? j : original;
388}
389
390/**
391 * __round_jiffies - function to round jiffies to a full second
392 * @j: the time in (absolute) jiffies that should be rounded
393 * @cpu: the processor number on which the timeout will happen
394 *
395 * __round_jiffies() rounds an absolute time in the future (in jiffies)
396 * up or down to (approximately) full seconds. This is useful for timers
397 * for which the exact time they fire does not matter too much, as long as
398 * they fire approximately every X seconds.
399 *
400 * By rounding these timers to whole seconds, all such timers will fire
401 * at the same time, rather than at various times spread out. The goal
402 * of this is to have the CPU wake up less, which saves power.
403 *
404 * The exact rounding is skewed for each processor to avoid all
405 * processors firing at the exact same time, which could lead
406 * to lock contention or spurious cache line bouncing.
407 *
408 * The return value is the rounded version of the @j parameter.
409 */
410unsigned long __round_jiffies(unsigned long j, int cpu)
411{
412 return round_jiffies_common(j, cpu, false);
413}
414EXPORT_SYMBOL_GPL(__round_jiffies);
415
416/**
417 * __round_jiffies_relative - function to round jiffies to a full second
418 * @j: the time in (relative) jiffies that should be rounded
419 * @cpu: the processor number on which the timeout will happen
420 *
421 * __round_jiffies_relative() rounds a time delta in the future (in jiffies)
422 * up or down to (approximately) full seconds. This is useful for timers
423 * for which the exact time they fire does not matter too much, as long as
424 * they fire approximately every X seconds.
425 *
426 * By rounding these timers to whole seconds, all such timers will fire
427 * at the same time, rather than at various times spread out. The goal
428 * of this is to have the CPU wake up less, which saves power.
429 *
430 * The exact rounding is skewed for each processor to avoid all
431 * processors firing at the exact same time, which could lead
432 * to lock contention or spurious cache line bouncing.
433 *
434 * The return value is the rounded version of the @j parameter.
435 */
436unsigned long __round_jiffies_relative(unsigned long j, int cpu)
437{
438 unsigned long j0 = jiffies;
439
440 /* Use j0 because jiffies might change while we run */
441 return round_jiffies_common(j + j0, cpu, false) - j0;
442}
443EXPORT_SYMBOL_GPL(__round_jiffies_relative);
444
445/**
446 * round_jiffies - function to round jiffies to a full second
447 * @j: the time in (absolute) jiffies that should be rounded
448 *
449 * round_jiffies() rounds an absolute time in the future (in jiffies)
450 * up or down to (approximately) full seconds. This is useful for timers
451 * for which the exact time they fire does not matter too much, as long as
452 * they fire approximately every X seconds.
453 *
454 * By rounding these timers to whole seconds, all such timers will fire
455 * at the same time, rather than at various times spread out. The goal
456 * of this is to have the CPU wake up less, which saves power.
457 *
458 * The return value is the rounded version of the @j parameter.
459 */
460unsigned long round_jiffies(unsigned long j)
461{
462 return round_jiffies_common(j, raw_smp_processor_id(), false);
463}
464EXPORT_SYMBOL_GPL(round_jiffies);
465
466/**
467 * round_jiffies_relative - function to round jiffies to a full second
468 * @j: the time in (relative) jiffies that should be rounded
469 *
470 * round_jiffies_relative() rounds a time delta in the future (in jiffies)
471 * up or down to (approximately) full seconds. This is useful for timers
472 * for which the exact time they fire does not matter too much, as long as
473 * they fire approximately every X seconds.
474 *
475 * By rounding these timers to whole seconds, all such timers will fire
476 * at the same time, rather than at various times spread out. The goal
477 * of this is to have the CPU wake up less, which saves power.
478 *
479 * The return value is the rounded version of the @j parameter.
480 */
481unsigned long round_jiffies_relative(unsigned long j)
482{
483 return __round_jiffies_relative(j, raw_smp_processor_id());
484}
485EXPORT_SYMBOL_GPL(round_jiffies_relative);
486
487/**
488 * __round_jiffies_up - function to round jiffies up to a full second
489 * @j: the time in (absolute) jiffies that should be rounded
490 * @cpu: the processor number on which the timeout will happen
491 *
492 * This is the same as __round_jiffies() except that it will never
493 * round down. This is useful for timeouts for which the exact time
494 * of firing does not matter too much, as long as they don't fire too
495 * early.
496 */
497unsigned long __round_jiffies_up(unsigned long j, int cpu)
498{
499 return round_jiffies_common(j, cpu, true);
500}
501EXPORT_SYMBOL_GPL(__round_jiffies_up);
502
503/**
504 * __round_jiffies_up_relative - function to round jiffies up to a full second
505 * @j: the time in (relative) jiffies that should be rounded
506 * @cpu: the processor number on which the timeout will happen
507 *
508 * This is the same as __round_jiffies_relative() except that it will never
509 * round down. This is useful for timeouts for which the exact time
510 * of firing does not matter too much, as long as they don't fire too
511 * early.
512 */
513unsigned long __round_jiffies_up_relative(unsigned long j, int cpu)
514{
515 unsigned long j0 = jiffies;
516
517 /* Use j0 because jiffies might change while we run */
518 return round_jiffies_common(j + j0, cpu, true) - j0;
519}
520EXPORT_SYMBOL_GPL(__round_jiffies_up_relative);
521
522/**
523 * round_jiffies_up - function to round jiffies up to a full second
524 * @j: the time in (absolute) jiffies that should be rounded
525 *
526 * This is the same as round_jiffies() except that it will never
527 * round down. This is useful for timeouts for which the exact time
528 * of firing does not matter too much, as long as they don't fire too
529 * early.
530 */
531unsigned long round_jiffies_up(unsigned long j)
532{
533 return round_jiffies_common(j, raw_smp_processor_id(), true);
534}
535EXPORT_SYMBOL_GPL(round_jiffies_up);
536
537/**
538 * round_jiffies_up_relative - function to round jiffies up to a full second
539 * @j: the time in (relative) jiffies that should be rounded
540 *
541 * This is the same as round_jiffies_relative() except that it will never
542 * round down. This is useful for timeouts for which the exact time
543 * of firing does not matter too much, as long as they don't fire too
544 * early.
545 */
546unsigned long round_jiffies_up_relative(unsigned long j)
547{
548 return __round_jiffies_up_relative(j, raw_smp_processor_id());
549}
550EXPORT_SYMBOL_GPL(round_jiffies_up_relative);
551
552
553static inline unsigned int timer_get_idx(struct timer_list *timer)
554{
555 return (timer->flags & TIMER_ARRAYMASK) >> TIMER_ARRAYSHIFT;
556}
557
558static inline void timer_set_idx(struct timer_list *timer, unsigned int idx)
559{
560 timer->flags = (timer->flags & ~TIMER_ARRAYMASK) |
561 idx << TIMER_ARRAYSHIFT;
562}
563
564/*
565 * Helper function to calculate the array index for a given expiry
566 * time.
567 */
568static inline unsigned calc_index(unsigned long expires, unsigned lvl,
569 unsigned long *bucket_expiry)
570{
571
572 /*
573 * The timer wheel has to guarantee that a timer does not fire
574 * early. Early expiry can happen due to:
575 * - Timer is armed at the edge of a tick
576 * - Truncation of the expiry time in the outer wheel levels
577 *
578 * Round up with level granularity to prevent this.
579 */
580 expires = (expires >> LVL_SHIFT(lvl)) + 1;
581 *bucket_expiry = expires << LVL_SHIFT(lvl);
582 return LVL_OFFS(lvl) + (expires & LVL_MASK);
583}
584
585static int calc_wheel_index(unsigned long expires, unsigned long clk,
586 unsigned long *bucket_expiry)
587{
588 unsigned long delta = expires - clk;
589 unsigned int idx;
590
591 if (delta < LVL_START(1)) {
592 idx = calc_index(expires, 0, bucket_expiry);
593 } else if (delta < LVL_START(2)) {
594 idx = calc_index(expires, 1, bucket_expiry);
595 } else if (delta < LVL_START(3)) {
596 idx = calc_index(expires, 2, bucket_expiry);
597 } else if (delta < LVL_START(4)) {
598 idx = calc_index(expires, 3, bucket_expiry);
599 } else if (delta < LVL_START(5)) {
600 idx = calc_index(expires, 4, bucket_expiry);
601 } else if (delta < LVL_START(6)) {
602 idx = calc_index(expires, 5, bucket_expiry);
603 } else if (delta < LVL_START(7)) {
604 idx = calc_index(expires, 6, bucket_expiry);
605 } else if (LVL_DEPTH > 8 && delta < LVL_START(8)) {
606 idx = calc_index(expires, 7, bucket_expiry);
607 } else if ((long) delta < 0) {
608 idx = clk & LVL_MASK;
609 *bucket_expiry = clk;
610 } else {
611 /*
612 * Force expire obscene large timeouts to expire at the
613 * capacity limit of the wheel.
614 */
615 if (delta >= WHEEL_TIMEOUT_CUTOFF)
616 expires = clk + WHEEL_TIMEOUT_MAX;
617
618 idx = calc_index(expires, LVL_DEPTH - 1, bucket_expiry);
619 }
620 return idx;
621}
622
623static void
624trigger_dyntick_cpu(struct timer_base *base, struct timer_list *timer)
625{
626 /*
627 * Deferrable timers do not prevent the CPU from entering dynticks and
628 * are not taken into account on the idle/nohz_full path. An IPI when a
629 * new deferrable timer is enqueued will wake up the remote CPU but
630 * nothing will be done with the deferrable timer base. Therefore skip
631 * the remote IPI for deferrable timers completely.
632 */
633 if (!is_timers_nohz_active() || timer->flags & TIMER_DEFERRABLE)
634 return;
635
636 /*
637 * We might have to IPI the remote CPU if the base is idle and the
638 * timer is pinned. If it is a non pinned timer, it is only queued
639 * on the remote CPU, when timer was running during queueing. Then
640 * everything is handled by remote CPU anyway. If the other CPU is
641 * on the way to idle then it can't set base->is_idle as we hold
642 * the base lock:
643 */
644 if (base->is_idle) {
645 WARN_ON_ONCE(!(timer->flags & TIMER_PINNED ||
646 tick_nohz_full_cpu(base->cpu)));
647 wake_up_nohz_cpu(base->cpu);
648 }
649}
650
651/*
652 * Enqueue the timer into the hash bucket, mark it pending in
653 * the bitmap, store the index in the timer flags then wake up
654 * the target CPU if needed.
655 */
656static void enqueue_timer(struct timer_base *base, struct timer_list *timer,
657 unsigned int idx, unsigned long bucket_expiry)
658{
659
660 hlist_add_head(&timer->entry, base->vectors + idx);
661 __set_bit(idx, base->pending_map);
662 timer_set_idx(timer, idx);
663
664 trace_timer_start(timer, bucket_expiry);
665
666 /*
667 * Check whether this is the new first expiring timer. The
668 * effective expiry time of the timer is required here
669 * (bucket_expiry) instead of timer->expires.
670 */
671 if (time_before(bucket_expiry, base->next_expiry)) {
672 /*
673 * Set the next expiry time and kick the CPU so it
674 * can reevaluate the wheel:
675 */
676 base->next_expiry = bucket_expiry;
677 base->timers_pending = true;
678 base->next_expiry_recalc = false;
679 trigger_dyntick_cpu(base, timer);
680 }
681}
682
683static void internal_add_timer(struct timer_base *base, struct timer_list *timer)
684{
685 unsigned long bucket_expiry;
686 unsigned int idx;
687
688 idx = calc_wheel_index(timer->expires, base->clk, &bucket_expiry);
689 enqueue_timer(base, timer, idx, bucket_expiry);
690}
691
692#ifdef CONFIG_DEBUG_OBJECTS_TIMERS
693
694static const struct debug_obj_descr timer_debug_descr;
695
696struct timer_hint {
697 void (*function)(struct timer_list *t);
698 long offset;
699};
700
701#define TIMER_HINT(fn, container, timr, hintfn) \
702 { \
703 .function = fn, \
704 .offset = offsetof(container, hintfn) - \
705 offsetof(container, timr) \
706 }
707
708static const struct timer_hint timer_hints[] = {
709 TIMER_HINT(delayed_work_timer_fn,
710 struct delayed_work, timer, work.func),
711 TIMER_HINT(kthread_delayed_work_timer_fn,
712 struct kthread_delayed_work, timer, work.func),
713};
714
715static void *timer_debug_hint(void *addr)
716{
717 struct timer_list *timer = addr;
718 int i;
719
720 for (i = 0; i < ARRAY_SIZE(timer_hints); i++) {
721 if (timer_hints[i].function == timer->function) {
722 void (**fn)(void) = addr + timer_hints[i].offset;
723
724 return *fn;
725 }
726 }
727
728 return timer->function;
729}
730
731static bool timer_is_static_object(void *addr)
732{
733 struct timer_list *timer = addr;
734
735 return (timer->entry.pprev == NULL &&
736 timer->entry.next == TIMER_ENTRY_STATIC);
737}
738
739/*
740 * timer_fixup_init is called when:
741 * - an active object is initialized
742 */
743static bool timer_fixup_init(void *addr, enum debug_obj_state state)
744{
745 struct timer_list *timer = addr;
746
747 switch (state) {
748 case ODEBUG_STATE_ACTIVE:
749 del_timer_sync(timer);
750 debug_object_init(timer, &timer_debug_descr);
751 return true;
752 default:
753 return false;
754 }
755}
756
757/* Stub timer callback for improperly used timers. */
758static void stub_timer(struct timer_list *unused)
759{
760 WARN_ON(1);
761}
762
763/*
764 * timer_fixup_activate is called when:
765 * - an active object is activated
766 * - an unknown non-static object is activated
767 */
768static bool timer_fixup_activate(void *addr, enum debug_obj_state state)
769{
770 struct timer_list *timer = addr;
771
772 switch (state) {
773 case ODEBUG_STATE_NOTAVAILABLE:
774 timer_setup(timer, stub_timer, 0);
775 return true;
776
777 case ODEBUG_STATE_ACTIVE:
778 WARN_ON(1);
779 fallthrough;
780 default:
781 return false;
782 }
783}
784
785/*
786 * timer_fixup_free is called when:
787 * - an active object is freed
788 */
789static bool timer_fixup_free(void *addr, enum debug_obj_state state)
790{
791 struct timer_list *timer = addr;
792
793 switch (state) {
794 case ODEBUG_STATE_ACTIVE:
795 del_timer_sync(timer);
796 debug_object_free(timer, &timer_debug_descr);
797 return true;
798 default:
799 return false;
800 }
801}
802
803/*
804 * timer_fixup_assert_init is called when:
805 * - an untracked/uninit-ed object is found
806 */
807static bool timer_fixup_assert_init(void *addr, enum debug_obj_state state)
808{
809 struct timer_list *timer = addr;
810
811 switch (state) {
812 case ODEBUG_STATE_NOTAVAILABLE:
813 timer_setup(timer, stub_timer, 0);
814 return true;
815 default:
816 return false;
817 }
818}
819
820static const struct debug_obj_descr timer_debug_descr = {
821 .name = "timer_list",
822 .debug_hint = timer_debug_hint,
823 .is_static_object = timer_is_static_object,
824 .fixup_init = timer_fixup_init,
825 .fixup_activate = timer_fixup_activate,
826 .fixup_free = timer_fixup_free,
827 .fixup_assert_init = timer_fixup_assert_init,
828};
829
830static inline void debug_timer_init(struct timer_list *timer)
831{
832 debug_object_init(timer, &timer_debug_descr);
833}
834
835static inline void debug_timer_activate(struct timer_list *timer)
836{
837 debug_object_activate(timer, &timer_debug_descr);
838}
839
840static inline void debug_timer_deactivate(struct timer_list *timer)
841{
842 debug_object_deactivate(timer, &timer_debug_descr);
843}
844
845static inline void debug_timer_assert_init(struct timer_list *timer)
846{
847 debug_object_assert_init(timer, &timer_debug_descr);
848}
849
850static void do_init_timer(struct timer_list *timer,
851 void (*func)(struct timer_list *),
852 unsigned int flags,
853 const char *name, struct lock_class_key *key);
854
855void init_timer_on_stack_key(struct timer_list *timer,
856 void (*func)(struct timer_list *),
857 unsigned int flags,
858 const char *name, struct lock_class_key *key)
859{
860 debug_object_init_on_stack(timer, &timer_debug_descr);
861 do_init_timer(timer, func, flags, name, key);
862}
863EXPORT_SYMBOL_GPL(init_timer_on_stack_key);
864
865void destroy_timer_on_stack(struct timer_list *timer)
866{
867 debug_object_free(timer, &timer_debug_descr);
868}
869EXPORT_SYMBOL_GPL(destroy_timer_on_stack);
870
871#else
872static inline void debug_timer_init(struct timer_list *timer) { }
873static inline void debug_timer_activate(struct timer_list *timer) { }
874static inline void debug_timer_deactivate(struct timer_list *timer) { }
875static inline void debug_timer_assert_init(struct timer_list *timer) { }
876#endif
877
878static inline void debug_init(struct timer_list *timer)
879{
880 debug_timer_init(timer);
881 trace_timer_init(timer);
882}
883
884static inline void debug_deactivate(struct timer_list *timer)
885{
886 debug_timer_deactivate(timer);
887 trace_timer_cancel(timer);
888}
889
890static inline void debug_assert_init(struct timer_list *timer)
891{
892 debug_timer_assert_init(timer);
893}
894
895static void do_init_timer(struct timer_list *timer,
896 void (*func)(struct timer_list *),
897 unsigned int flags,
898 const char *name, struct lock_class_key *key)
899{
900 timer->entry.pprev = NULL;
901 timer->function = func;
902 if (WARN_ON_ONCE(flags & ~TIMER_INIT_FLAGS))
903 flags &= TIMER_INIT_FLAGS;
904 timer->flags = flags | raw_smp_processor_id();
905 lockdep_init_map(&timer->lockdep_map, name, key, 0);
906}
907
908/**
909 * init_timer_key - initialize a timer
910 * @timer: the timer to be initialized
911 * @func: timer callback function
912 * @flags: timer flags
913 * @name: name of the timer
914 * @key: lockdep class key of the fake lock used for tracking timer
915 * sync lock dependencies
916 *
917 * init_timer_key() must be done to a timer prior to calling *any* of the
918 * other timer functions.
919 */
920void init_timer_key(struct timer_list *timer,
921 void (*func)(struct timer_list *), unsigned int flags,
922 const char *name, struct lock_class_key *key)
923{
924 debug_init(timer);
925 do_init_timer(timer, func, flags, name, key);
926}
927EXPORT_SYMBOL(init_timer_key);
928
929static inline void detach_timer(struct timer_list *timer, bool clear_pending)
930{
931 struct hlist_node *entry = &timer->entry;
932
933 debug_deactivate(timer);
934
935 __hlist_del(entry);
936 if (clear_pending)
937 entry->pprev = NULL;
938 entry->next = LIST_POISON2;
939}
940
941static int detach_if_pending(struct timer_list *timer, struct timer_base *base,
942 bool clear_pending)
943{
944 unsigned idx = timer_get_idx(timer);
945
946 if (!timer_pending(timer))
947 return 0;
948
949 if (hlist_is_singular_node(&timer->entry, base->vectors + idx)) {
950 __clear_bit(idx, base->pending_map);
951 base->next_expiry_recalc = true;
952 }
953
954 detach_timer(timer, clear_pending);
955 return 1;
956}
957
958static inline struct timer_base *get_timer_cpu_base(u32 tflags, u32 cpu)
959{
960 int index = tflags & TIMER_PINNED ? BASE_LOCAL : BASE_GLOBAL;
961 struct timer_base *base;
962
963 base = per_cpu_ptr(&timer_bases[index], cpu);
964
965 /*
966 * If the timer is deferrable and NO_HZ_COMMON is set then we need
967 * to use the deferrable base.
968 */
969 if (IS_ENABLED(CONFIG_NO_HZ_COMMON) && (tflags & TIMER_DEFERRABLE))
970 base = per_cpu_ptr(&timer_bases[BASE_DEF], cpu);
971 return base;
972}
973
974static inline struct timer_base *get_timer_this_cpu_base(u32 tflags)
975{
976 int index = tflags & TIMER_PINNED ? BASE_LOCAL : BASE_GLOBAL;
977 struct timer_base *base;
978
979 base = this_cpu_ptr(&timer_bases[index]);
980
981 /*
982 * If the timer is deferrable and NO_HZ_COMMON is set then we need
983 * to use the deferrable base.
984 */
985 if (IS_ENABLED(CONFIG_NO_HZ_COMMON) && (tflags & TIMER_DEFERRABLE))
986 base = this_cpu_ptr(&timer_bases[BASE_DEF]);
987 return base;
988}
989
990static inline struct timer_base *get_timer_base(u32 tflags)
991{
992 return get_timer_cpu_base(tflags, tflags & TIMER_CPUMASK);
993}
994
995static inline void __forward_timer_base(struct timer_base *base,
996 unsigned long basej)
997{
998 /*
999 * Check whether we can forward the base. We can only do that when
1000 * @basej is past base->clk otherwise we might rewind base->clk.
1001 */
1002 if (time_before_eq(basej, base->clk))
1003 return;
1004
1005 /*
1006 * If the next expiry value is > jiffies, then we fast forward to
1007 * jiffies otherwise we forward to the next expiry value.
1008 */
1009 if (time_after(base->next_expiry, basej)) {
1010 base->clk = basej;
1011 } else {
1012 if (WARN_ON_ONCE(time_before(base->next_expiry, base->clk)))
1013 return;
1014 base->clk = base->next_expiry;
1015 }
1016
1017}
1018
1019static inline void forward_timer_base(struct timer_base *base)
1020{
1021 __forward_timer_base(base, READ_ONCE(jiffies));
1022}
1023
1024/*
1025 * We are using hashed locking: Holding per_cpu(timer_bases[x]).lock means
1026 * that all timers which are tied to this base are locked, and the base itself
1027 * is locked too.
1028 *
1029 * So __run_timers/migrate_timers can safely modify all timers which could
1030 * be found in the base->vectors array.
1031 *
1032 * When a timer is migrating then the TIMER_MIGRATING flag is set and we need
1033 * to wait until the migration is done.
1034 */
1035static struct timer_base *lock_timer_base(struct timer_list *timer,
1036 unsigned long *flags)
1037 __acquires(timer->base->lock)
1038{
1039 for (;;) {
1040 struct timer_base *base;
1041 u32 tf;
1042
1043 /*
1044 * We need to use READ_ONCE() here, otherwise the compiler
1045 * might re-read @tf between the check for TIMER_MIGRATING
1046 * and spin_lock().
1047 */
1048 tf = READ_ONCE(timer->flags);
1049
1050 if (!(tf & TIMER_MIGRATING)) {
1051 base = get_timer_base(tf);
1052 raw_spin_lock_irqsave(&base->lock, *flags);
1053 if (timer->flags == tf)
1054 return base;
1055 raw_spin_unlock_irqrestore(&base->lock, *flags);
1056 }
1057 cpu_relax();
1058 }
1059}
1060
1061#define MOD_TIMER_PENDING_ONLY 0x01
1062#define MOD_TIMER_REDUCE 0x02
1063#define MOD_TIMER_NOTPENDING 0x04
1064
1065static inline int
1066__mod_timer(struct timer_list *timer, unsigned long expires, unsigned int options)
1067{
1068 unsigned long clk = 0, flags, bucket_expiry;
1069 struct timer_base *base, *new_base;
1070 unsigned int idx = UINT_MAX;
1071 int ret = 0;
1072
1073 debug_assert_init(timer);
1074
1075 /*
1076 * This is a common optimization triggered by the networking code - if
1077 * the timer is re-modified to have the same timeout or ends up in the
1078 * same array bucket then just return:
1079 */
1080 if (!(options & MOD_TIMER_NOTPENDING) && timer_pending(timer)) {
1081 /*
1082 * The downside of this optimization is that it can result in
1083 * larger granularity than you would get from adding a new
1084 * timer with this expiry.
1085 */
1086 long diff = timer->expires - expires;
1087
1088 if (!diff)
1089 return 1;
1090 if (options & MOD_TIMER_REDUCE && diff <= 0)
1091 return 1;
1092
1093 /*
1094 * We lock timer base and calculate the bucket index right
1095 * here. If the timer ends up in the same bucket, then we
1096 * just update the expiry time and avoid the whole
1097 * dequeue/enqueue dance.
1098 */
1099 base = lock_timer_base(timer, &flags);
1100 /*
1101 * Has @timer been shutdown? This needs to be evaluated
1102 * while holding base lock to prevent a race against the
1103 * shutdown code.
1104 */
1105 if (!timer->function)
1106 goto out_unlock;
1107
1108 forward_timer_base(base);
1109
1110 if (timer_pending(timer) && (options & MOD_TIMER_REDUCE) &&
1111 time_before_eq(timer->expires, expires)) {
1112 ret = 1;
1113 goto out_unlock;
1114 }
1115
1116 clk = base->clk;
1117 idx = calc_wheel_index(expires, clk, &bucket_expiry);
1118
1119 /*
1120 * Retrieve and compare the array index of the pending
1121 * timer. If it matches set the expiry to the new value so a
1122 * subsequent call will exit in the expires check above.
1123 */
1124 if (idx == timer_get_idx(timer)) {
1125 if (!(options & MOD_TIMER_REDUCE))
1126 timer->expires = expires;
1127 else if (time_after(timer->expires, expires))
1128 timer->expires = expires;
1129 ret = 1;
1130 goto out_unlock;
1131 }
1132 } else {
1133 base = lock_timer_base(timer, &flags);
1134 /*
1135 * Has @timer been shutdown? This needs to be evaluated
1136 * while holding base lock to prevent a race against the
1137 * shutdown code.
1138 */
1139 if (!timer->function)
1140 goto out_unlock;
1141
1142 forward_timer_base(base);
1143 }
1144
1145 ret = detach_if_pending(timer, base, false);
1146 if (!ret && (options & MOD_TIMER_PENDING_ONLY))
1147 goto out_unlock;
1148
1149 new_base = get_timer_this_cpu_base(timer->flags);
1150
1151 if (base != new_base) {
1152 /*
1153 * We are trying to schedule the timer on the new base.
1154 * However we can't change timer's base while it is running,
1155 * otherwise timer_delete_sync() can't detect that the timer's
1156 * handler yet has not finished. This also guarantees that the
1157 * timer is serialized wrt itself.
1158 */
1159 if (likely(base->running_timer != timer)) {
1160 /* See the comment in lock_timer_base() */
1161 timer->flags |= TIMER_MIGRATING;
1162
1163 raw_spin_unlock(&base->lock);
1164 base = new_base;
1165 raw_spin_lock(&base->lock);
1166 WRITE_ONCE(timer->flags,
1167 (timer->flags & ~TIMER_BASEMASK) | base->cpu);
1168 forward_timer_base(base);
1169 }
1170 }
1171
1172 debug_timer_activate(timer);
1173
1174 timer->expires = expires;
1175 /*
1176 * If 'idx' was calculated above and the base time did not advance
1177 * between calculating 'idx' and possibly switching the base, only
1178 * enqueue_timer() is required. Otherwise we need to (re)calculate
1179 * the wheel index via internal_add_timer().
1180 */
1181 if (idx != UINT_MAX && clk == base->clk)
1182 enqueue_timer(base, timer, idx, bucket_expiry);
1183 else
1184 internal_add_timer(base, timer);
1185
1186out_unlock:
1187 raw_spin_unlock_irqrestore(&base->lock, flags);
1188
1189 return ret;
1190}
1191
1192/**
1193 * mod_timer_pending - Modify a pending timer's timeout
1194 * @timer: The pending timer to be modified
1195 * @expires: New absolute timeout in jiffies
1196 *
1197 * mod_timer_pending() is the same for pending timers as mod_timer(), but
1198 * will not activate inactive timers.
1199 *
1200 * If @timer->function == NULL then the start operation is silently
1201 * discarded.
1202 *
1203 * Return:
1204 * * %0 - The timer was inactive and not modified or was in
1205 * shutdown state and the operation was discarded
1206 * * %1 - The timer was active and requeued to expire at @expires
1207 */
1208int mod_timer_pending(struct timer_list *timer, unsigned long expires)
1209{
1210 return __mod_timer(timer, expires, MOD_TIMER_PENDING_ONLY);
1211}
1212EXPORT_SYMBOL(mod_timer_pending);
1213
1214/**
1215 * mod_timer - Modify a timer's timeout
1216 * @timer: The timer to be modified
1217 * @expires: New absolute timeout in jiffies
1218 *
1219 * mod_timer(timer, expires) is equivalent to:
1220 *
1221 * del_timer(timer); timer->expires = expires; add_timer(timer);
1222 *
1223 * mod_timer() is more efficient than the above open coded sequence. In
1224 * case that the timer is inactive, the del_timer() part is a NOP. The
1225 * timer is in any case activated with the new expiry time @expires.
1226 *
1227 * Note that if there are multiple unserialized concurrent users of the
1228 * same timer, then mod_timer() is the only safe way to modify the timeout,
1229 * since add_timer() cannot modify an already running timer.
1230 *
1231 * If @timer->function == NULL then the start operation is silently
1232 * discarded. In this case the return value is 0 and meaningless.
1233 *
1234 * Return:
1235 * * %0 - The timer was inactive and started or was in shutdown
1236 * state and the operation was discarded
1237 * * %1 - The timer was active and requeued to expire at @expires or
1238 * the timer was active and not modified because @expires did
1239 * not change the effective expiry time
1240 */
1241int mod_timer(struct timer_list *timer, unsigned long expires)
1242{
1243 return __mod_timer(timer, expires, 0);
1244}
1245EXPORT_SYMBOL(mod_timer);
1246
1247/**
1248 * timer_reduce - Modify a timer's timeout if it would reduce the timeout
1249 * @timer: The timer to be modified
1250 * @expires: New absolute timeout in jiffies
1251 *
1252 * timer_reduce() is very similar to mod_timer(), except that it will only
1253 * modify an enqueued timer if that would reduce the expiration time. If
1254 * @timer is not enqueued it starts the timer.
1255 *
1256 * If @timer->function == NULL then the start operation is silently
1257 * discarded.
1258 *
1259 * Return:
1260 * * %0 - The timer was inactive and started or was in shutdown
1261 * state and the operation was discarded
1262 * * %1 - The timer was active and requeued to expire at @expires or
1263 * the timer was active and not modified because @expires
1264 * did not change the effective expiry time such that the
1265 * timer would expire earlier than already scheduled
1266 */
1267int timer_reduce(struct timer_list *timer, unsigned long expires)
1268{
1269 return __mod_timer(timer, expires, MOD_TIMER_REDUCE);
1270}
1271EXPORT_SYMBOL(timer_reduce);
1272
1273/**
1274 * add_timer - Start a timer
1275 * @timer: The timer to be started
1276 *
1277 * Start @timer to expire at @timer->expires in the future. @timer->expires
1278 * is the absolute expiry time measured in 'jiffies'. When the timer expires
1279 * timer->function(timer) will be invoked from soft interrupt context.
1280 *
1281 * The @timer->expires and @timer->function fields must be set prior
1282 * to calling this function.
1283 *
1284 * If @timer->function == NULL then the start operation is silently
1285 * discarded.
1286 *
1287 * If @timer->expires is already in the past @timer will be queued to
1288 * expire at the next timer tick.
1289 *
1290 * This can only operate on an inactive timer. Attempts to invoke this on
1291 * an active timer are rejected with a warning.
1292 */
1293void add_timer(struct timer_list *timer)
1294{
1295 if (WARN_ON_ONCE(timer_pending(timer)))
1296 return;
1297 __mod_timer(timer, timer->expires, MOD_TIMER_NOTPENDING);
1298}
1299EXPORT_SYMBOL(add_timer);
1300
1301/**
1302 * add_timer_local() - Start a timer on the local CPU
1303 * @timer: The timer to be started
1304 *
1305 * Same as add_timer() except that the timer flag TIMER_PINNED is set.
1306 *
1307 * See add_timer() for further details.
1308 */
1309void add_timer_local(struct timer_list *timer)
1310{
1311 if (WARN_ON_ONCE(timer_pending(timer)))
1312 return;
1313 timer->flags |= TIMER_PINNED;
1314 __mod_timer(timer, timer->expires, MOD_TIMER_NOTPENDING);
1315}
1316EXPORT_SYMBOL(add_timer_local);
1317
1318/**
1319 * add_timer_global() - Start a timer without TIMER_PINNED flag set
1320 * @timer: The timer to be started
1321 *
1322 * Same as add_timer() except that the timer flag TIMER_PINNED is unset.
1323 *
1324 * See add_timer() for further details.
1325 */
1326void add_timer_global(struct timer_list *timer)
1327{
1328 if (WARN_ON_ONCE(timer_pending(timer)))
1329 return;
1330 timer->flags &= ~TIMER_PINNED;
1331 __mod_timer(timer, timer->expires, MOD_TIMER_NOTPENDING);
1332}
1333EXPORT_SYMBOL(add_timer_global);
1334
1335/**
1336 * add_timer_on - Start a timer on a particular CPU
1337 * @timer: The timer to be started
1338 * @cpu: The CPU to start it on
1339 *
1340 * Same as add_timer() except that it starts the timer on the given CPU and
1341 * the TIMER_PINNED flag is set. When timer shouldn't be a pinned timer in
1342 * the next round, add_timer_global() should be used instead as it unsets
1343 * the TIMER_PINNED flag.
1344 *
1345 * See add_timer() for further details.
1346 */
1347void add_timer_on(struct timer_list *timer, int cpu)
1348{
1349 struct timer_base *new_base, *base;
1350 unsigned long flags;
1351
1352 debug_assert_init(timer);
1353
1354 if (WARN_ON_ONCE(timer_pending(timer)))
1355 return;
1356
1357 /* Make sure timer flags have TIMER_PINNED flag set */
1358 timer->flags |= TIMER_PINNED;
1359
1360 new_base = get_timer_cpu_base(timer->flags, cpu);
1361
1362 /*
1363 * If @timer was on a different CPU, it should be migrated with the
1364 * old base locked to prevent other operations proceeding with the
1365 * wrong base locked. See lock_timer_base().
1366 */
1367 base = lock_timer_base(timer, &flags);
1368 /*
1369 * Has @timer been shutdown? This needs to be evaluated while
1370 * holding base lock to prevent a race against the shutdown code.
1371 */
1372 if (!timer->function)
1373 goto out_unlock;
1374
1375 if (base != new_base) {
1376 timer->flags |= TIMER_MIGRATING;
1377
1378 raw_spin_unlock(&base->lock);
1379 base = new_base;
1380 raw_spin_lock(&base->lock);
1381 WRITE_ONCE(timer->flags,
1382 (timer->flags & ~TIMER_BASEMASK) | cpu);
1383 }
1384 forward_timer_base(base);
1385
1386 debug_timer_activate(timer);
1387 internal_add_timer(base, timer);
1388out_unlock:
1389 raw_spin_unlock_irqrestore(&base->lock, flags);
1390}
1391EXPORT_SYMBOL_GPL(add_timer_on);
1392
1393/**
1394 * __timer_delete - Internal function: Deactivate a timer
1395 * @timer: The timer to be deactivated
1396 * @shutdown: If true, this indicates that the timer is about to be
1397 * shutdown permanently.
1398 *
1399 * If @shutdown is true then @timer->function is set to NULL under the
1400 * timer base lock which prevents further rearming of the time. In that
1401 * case any attempt to rearm @timer after this function returns will be
1402 * silently ignored.
1403 *
1404 * Return:
1405 * * %0 - The timer was not pending
1406 * * %1 - The timer was pending and deactivated
1407 */
1408static int __timer_delete(struct timer_list *timer, bool shutdown)
1409{
1410 struct timer_base *base;
1411 unsigned long flags;
1412 int ret = 0;
1413
1414 debug_assert_init(timer);
1415
1416 /*
1417 * If @shutdown is set then the lock has to be taken whether the
1418 * timer is pending or not to protect against a concurrent rearm
1419 * which might hit between the lockless pending check and the lock
1420 * acquisition. By taking the lock it is ensured that such a newly
1421 * enqueued timer is dequeued and cannot end up with
1422 * timer->function == NULL in the expiry code.
1423 *
1424 * If timer->function is currently executed, then this makes sure
1425 * that the callback cannot requeue the timer.
1426 */
1427 if (timer_pending(timer) || shutdown) {
1428 base = lock_timer_base(timer, &flags);
1429 ret = detach_if_pending(timer, base, true);
1430 if (shutdown)
1431 timer->function = NULL;
1432 raw_spin_unlock_irqrestore(&base->lock, flags);
1433 }
1434
1435 return ret;
1436}
1437
1438/**
1439 * timer_delete - Deactivate a timer
1440 * @timer: The timer to be deactivated
1441 *
1442 * The function only deactivates a pending timer, but contrary to
1443 * timer_delete_sync() it does not take into account whether the timer's
1444 * callback function is concurrently executed on a different CPU or not.
1445 * It neither prevents rearming of the timer. If @timer can be rearmed
1446 * concurrently then the return value of this function is meaningless.
1447 *
1448 * Return:
1449 * * %0 - The timer was not pending
1450 * * %1 - The timer was pending and deactivated
1451 */
1452int timer_delete(struct timer_list *timer)
1453{
1454 return __timer_delete(timer, false);
1455}
1456EXPORT_SYMBOL(timer_delete);
1457
1458/**
1459 * timer_shutdown - Deactivate a timer and prevent rearming
1460 * @timer: The timer to be deactivated
1461 *
1462 * The function does not wait for an eventually running timer callback on a
1463 * different CPU but it prevents rearming of the timer. Any attempt to arm
1464 * @timer after this function returns will be silently ignored.
1465 *
1466 * This function is useful for teardown code and should only be used when
1467 * timer_shutdown_sync() cannot be invoked due to locking or context constraints.
1468 *
1469 * Return:
1470 * * %0 - The timer was not pending
1471 * * %1 - The timer was pending
1472 */
1473int timer_shutdown(struct timer_list *timer)
1474{
1475 return __timer_delete(timer, true);
1476}
1477EXPORT_SYMBOL_GPL(timer_shutdown);
1478
1479/**
1480 * __try_to_del_timer_sync - Internal function: Try to deactivate a timer
1481 * @timer: Timer to deactivate
1482 * @shutdown: If true, this indicates that the timer is about to be
1483 * shutdown permanently.
1484 *
1485 * If @shutdown is true then @timer->function is set to NULL under the
1486 * timer base lock which prevents further rearming of the timer. Any
1487 * attempt to rearm @timer after this function returns will be silently
1488 * ignored.
1489 *
1490 * This function cannot guarantee that the timer cannot be rearmed
1491 * right after dropping the base lock if @shutdown is false. That
1492 * needs to be prevented by the calling code if necessary.
1493 *
1494 * Return:
1495 * * %0 - The timer was not pending
1496 * * %1 - The timer was pending and deactivated
1497 * * %-1 - The timer callback function is running on a different CPU
1498 */
1499static int __try_to_del_timer_sync(struct timer_list *timer, bool shutdown)
1500{
1501 struct timer_base *base;
1502 unsigned long flags;
1503 int ret = -1;
1504
1505 debug_assert_init(timer);
1506
1507 base = lock_timer_base(timer, &flags);
1508
1509 if (base->running_timer != timer)
1510 ret = detach_if_pending(timer, base, true);
1511 if (shutdown)
1512 timer->function = NULL;
1513
1514 raw_spin_unlock_irqrestore(&base->lock, flags);
1515
1516 return ret;
1517}
1518
1519/**
1520 * try_to_del_timer_sync - Try to deactivate a timer
1521 * @timer: Timer to deactivate
1522 *
1523 * This function tries to deactivate a timer. On success the timer is not
1524 * queued and the timer callback function is not running on any CPU.
1525 *
1526 * This function does not guarantee that the timer cannot be rearmed right
1527 * after dropping the base lock. That needs to be prevented by the calling
1528 * code if necessary.
1529 *
1530 * Return:
1531 * * %0 - The timer was not pending
1532 * * %1 - The timer was pending and deactivated
1533 * * %-1 - The timer callback function is running on a different CPU
1534 */
1535int try_to_del_timer_sync(struct timer_list *timer)
1536{
1537 return __try_to_del_timer_sync(timer, false);
1538}
1539EXPORT_SYMBOL(try_to_del_timer_sync);
1540
1541#ifdef CONFIG_PREEMPT_RT
1542static __init void timer_base_init_expiry_lock(struct timer_base *base)
1543{
1544 spin_lock_init(&base->expiry_lock);
1545}
1546
1547static inline void timer_base_lock_expiry(struct timer_base *base)
1548{
1549 spin_lock(&base->expiry_lock);
1550}
1551
1552static inline void timer_base_unlock_expiry(struct timer_base *base)
1553{
1554 spin_unlock(&base->expiry_lock);
1555}
1556
1557/*
1558 * The counterpart to del_timer_wait_running().
1559 *
1560 * If there is a waiter for base->expiry_lock, then it was waiting for the
1561 * timer callback to finish. Drop expiry_lock and reacquire it. That allows
1562 * the waiter to acquire the lock and make progress.
1563 */
1564static void timer_sync_wait_running(struct timer_base *base)
1565{
1566 if (atomic_read(&base->timer_waiters)) {
1567 raw_spin_unlock_irq(&base->lock);
1568 spin_unlock(&base->expiry_lock);
1569 spin_lock(&base->expiry_lock);
1570 raw_spin_lock_irq(&base->lock);
1571 }
1572}
1573
1574/*
1575 * This function is called on PREEMPT_RT kernels when the fast path
1576 * deletion of a timer failed because the timer callback function was
1577 * running.
1578 *
1579 * This prevents priority inversion, if the softirq thread on a remote CPU
1580 * got preempted, and it prevents a life lock when the task which tries to
1581 * delete a timer preempted the softirq thread running the timer callback
1582 * function.
1583 */
1584static void del_timer_wait_running(struct timer_list *timer)
1585{
1586 u32 tf;
1587
1588 tf = READ_ONCE(timer->flags);
1589 if (!(tf & (TIMER_MIGRATING | TIMER_IRQSAFE))) {
1590 struct timer_base *base = get_timer_base(tf);
1591
1592 /*
1593 * Mark the base as contended and grab the expiry lock,
1594 * which is held by the softirq across the timer
1595 * callback. Drop the lock immediately so the softirq can
1596 * expire the next timer. In theory the timer could already
1597 * be running again, but that's more than unlikely and just
1598 * causes another wait loop.
1599 */
1600 atomic_inc(&base->timer_waiters);
1601 spin_lock_bh(&base->expiry_lock);
1602 atomic_dec(&base->timer_waiters);
1603 spin_unlock_bh(&base->expiry_lock);
1604 }
1605}
1606#else
1607static inline void timer_base_init_expiry_lock(struct timer_base *base) { }
1608static inline void timer_base_lock_expiry(struct timer_base *base) { }
1609static inline void timer_base_unlock_expiry(struct timer_base *base) { }
1610static inline void timer_sync_wait_running(struct timer_base *base) { }
1611static inline void del_timer_wait_running(struct timer_list *timer) { }
1612#endif
1613
1614/**
1615 * __timer_delete_sync - Internal function: Deactivate a timer and wait
1616 * for the handler to finish.
1617 * @timer: The timer to be deactivated
1618 * @shutdown: If true, @timer->function will be set to NULL under the
1619 * timer base lock which prevents rearming of @timer
1620 *
1621 * If @shutdown is not set the timer can be rearmed later. If the timer can
1622 * be rearmed concurrently, i.e. after dropping the base lock then the
1623 * return value is meaningless.
1624 *
1625 * If @shutdown is set then @timer->function is set to NULL under timer
1626 * base lock which prevents rearming of the timer. Any attempt to rearm
1627 * a shutdown timer is silently ignored.
1628 *
1629 * If the timer should be reused after shutdown it has to be initialized
1630 * again.
1631 *
1632 * Return:
1633 * * %0 - The timer was not pending
1634 * * %1 - The timer was pending and deactivated
1635 */
1636static int __timer_delete_sync(struct timer_list *timer, bool shutdown)
1637{
1638 int ret;
1639
1640#ifdef CONFIG_LOCKDEP
1641 unsigned long flags;
1642
1643 /*
1644 * If lockdep gives a backtrace here, please reference
1645 * the synchronization rules above.
1646 */
1647 local_irq_save(flags);
1648 lock_map_acquire(&timer->lockdep_map);
1649 lock_map_release(&timer->lockdep_map);
1650 local_irq_restore(flags);
1651#endif
1652 /*
1653 * don't use it in hardirq context, because it
1654 * could lead to deadlock.
1655 */
1656 WARN_ON(in_hardirq() && !(timer->flags & TIMER_IRQSAFE));
1657
1658 /*
1659 * Must be able to sleep on PREEMPT_RT because of the slowpath in
1660 * del_timer_wait_running().
1661 */
1662 if (IS_ENABLED(CONFIG_PREEMPT_RT) && !(timer->flags & TIMER_IRQSAFE))
1663 lockdep_assert_preemption_enabled();
1664
1665 do {
1666 ret = __try_to_del_timer_sync(timer, shutdown);
1667
1668 if (unlikely(ret < 0)) {
1669 del_timer_wait_running(timer);
1670 cpu_relax();
1671 }
1672 } while (ret < 0);
1673
1674 return ret;
1675}
1676
1677/**
1678 * timer_delete_sync - Deactivate a timer and wait for the handler to finish.
1679 * @timer: The timer to be deactivated
1680 *
1681 * Synchronization rules: Callers must prevent restarting of the timer,
1682 * otherwise this function is meaningless. It must not be called from
1683 * interrupt contexts unless the timer is an irqsafe one. The caller must
1684 * not hold locks which would prevent completion of the timer's callback
1685 * function. The timer's handler must not call add_timer_on(). Upon exit
1686 * the timer is not queued and the handler is not running on any CPU.
1687 *
1688 * For !irqsafe timers, the caller must not hold locks that are held in
1689 * interrupt context. Even if the lock has nothing to do with the timer in
1690 * question. Here's why::
1691 *
1692 * CPU0 CPU1
1693 * ---- ----
1694 * <SOFTIRQ>
1695 * call_timer_fn();
1696 * base->running_timer = mytimer;
1697 * spin_lock_irq(somelock);
1698 * <IRQ>
1699 * spin_lock(somelock);
1700 * timer_delete_sync(mytimer);
1701 * while (base->running_timer == mytimer);
1702 *
1703 * Now timer_delete_sync() will never return and never release somelock.
1704 * The interrupt on the other CPU is waiting to grab somelock but it has
1705 * interrupted the softirq that CPU0 is waiting to finish.
1706 *
1707 * This function cannot guarantee that the timer is not rearmed again by
1708 * some concurrent or preempting code, right after it dropped the base
1709 * lock. If there is the possibility of a concurrent rearm then the return
1710 * value of the function is meaningless.
1711 *
1712 * If such a guarantee is needed, e.g. for teardown situations then use
1713 * timer_shutdown_sync() instead.
1714 *
1715 * Return:
1716 * * %0 - The timer was not pending
1717 * * %1 - The timer was pending and deactivated
1718 */
1719int timer_delete_sync(struct timer_list *timer)
1720{
1721 return __timer_delete_sync(timer, false);
1722}
1723EXPORT_SYMBOL(timer_delete_sync);
1724
1725/**
1726 * timer_shutdown_sync - Shutdown a timer and prevent rearming
1727 * @timer: The timer to be shutdown
1728 *
1729 * When the function returns it is guaranteed that:
1730 * - @timer is not queued
1731 * - The callback function of @timer is not running
1732 * - @timer cannot be enqueued again. Any attempt to rearm
1733 * @timer is silently ignored.
1734 *
1735 * See timer_delete_sync() for synchronization rules.
1736 *
1737 * This function is useful for final teardown of an infrastructure where
1738 * the timer is subject to a circular dependency problem.
1739 *
1740 * A common pattern for this is a timer and a workqueue where the timer can
1741 * schedule work and work can arm the timer. On shutdown the workqueue must
1742 * be destroyed and the timer must be prevented from rearming. Unless the
1743 * code has conditionals like 'if (mything->in_shutdown)' to prevent that
1744 * there is no way to get this correct with timer_delete_sync().
1745 *
1746 * timer_shutdown_sync() is solving the problem. The correct ordering of
1747 * calls in this case is:
1748 *
1749 * timer_shutdown_sync(&mything->timer);
1750 * workqueue_destroy(&mything->workqueue);
1751 *
1752 * After this 'mything' can be safely freed.
1753 *
1754 * This obviously implies that the timer is not required to be functional
1755 * for the rest of the shutdown operation.
1756 *
1757 * Return:
1758 * * %0 - The timer was not pending
1759 * * %1 - The timer was pending
1760 */
1761int timer_shutdown_sync(struct timer_list *timer)
1762{
1763 return __timer_delete_sync(timer, true);
1764}
1765EXPORT_SYMBOL_GPL(timer_shutdown_sync);
1766
1767static void call_timer_fn(struct timer_list *timer,
1768 void (*fn)(struct timer_list *),
1769 unsigned long baseclk)
1770{
1771 int count = preempt_count();
1772
1773#ifdef CONFIG_LOCKDEP
1774 /*
1775 * It is permissible to free the timer from inside the
1776 * function that is called from it, this we need to take into
1777 * account for lockdep too. To avoid bogus "held lock freed"
1778 * warnings as well as problems when looking into
1779 * timer->lockdep_map, make a copy and use that here.
1780 */
1781 struct lockdep_map lockdep_map;
1782
1783 lockdep_copy_map(&lockdep_map, &timer->lockdep_map);
1784#endif
1785 /*
1786 * Couple the lock chain with the lock chain at
1787 * timer_delete_sync() by acquiring the lock_map around the fn()
1788 * call here and in timer_delete_sync().
1789 */
1790 lock_map_acquire(&lockdep_map);
1791
1792 trace_timer_expire_entry(timer, baseclk);
1793 fn(timer);
1794 trace_timer_expire_exit(timer);
1795
1796 lock_map_release(&lockdep_map);
1797
1798 if (count != preempt_count()) {
1799 WARN_ONCE(1, "timer: %pS preempt leak: %08x -> %08x\n",
1800 fn, count, preempt_count());
1801 /*
1802 * Restore the preempt count. That gives us a decent
1803 * chance to survive and extract information. If the
1804 * callback kept a lock held, bad luck, but not worse
1805 * than the BUG() we had.
1806 */
1807 preempt_count_set(count);
1808 }
1809}
1810
1811static void expire_timers(struct timer_base *base, struct hlist_head *head)
1812{
1813 /*
1814 * This value is required only for tracing. base->clk was
1815 * incremented directly before expire_timers was called. But expiry
1816 * is related to the old base->clk value.
1817 */
1818 unsigned long baseclk = base->clk - 1;
1819
1820 while (!hlist_empty(head)) {
1821 struct timer_list *timer;
1822 void (*fn)(struct timer_list *);
1823
1824 timer = hlist_entry(head->first, struct timer_list, entry);
1825
1826 base->running_timer = timer;
1827 detach_timer(timer, true);
1828
1829 fn = timer->function;
1830
1831 if (WARN_ON_ONCE(!fn)) {
1832 /* Should never happen. Emphasis on should! */
1833 base->running_timer = NULL;
1834 continue;
1835 }
1836
1837 if (timer->flags & TIMER_IRQSAFE) {
1838 raw_spin_unlock(&base->lock);
1839 call_timer_fn(timer, fn, baseclk);
1840 raw_spin_lock(&base->lock);
1841 base->running_timer = NULL;
1842 } else {
1843 raw_spin_unlock_irq(&base->lock);
1844 call_timer_fn(timer, fn, baseclk);
1845 raw_spin_lock_irq(&base->lock);
1846 base->running_timer = NULL;
1847 timer_sync_wait_running(base);
1848 }
1849 }
1850}
1851
1852static int collect_expired_timers(struct timer_base *base,
1853 struct hlist_head *heads)
1854{
1855 unsigned long clk = base->clk = base->next_expiry;
1856 struct hlist_head *vec;
1857 int i, levels = 0;
1858 unsigned int idx;
1859
1860 for (i = 0; i < LVL_DEPTH; i++) {
1861 idx = (clk & LVL_MASK) + i * LVL_SIZE;
1862
1863 if (__test_and_clear_bit(idx, base->pending_map)) {
1864 vec = base->vectors + idx;
1865 hlist_move_list(vec, heads++);
1866 levels++;
1867 }
1868 /* Is it time to look at the next level? */
1869 if (clk & LVL_CLK_MASK)
1870 break;
1871 /* Shift clock for the next level granularity */
1872 clk >>= LVL_CLK_SHIFT;
1873 }
1874 return levels;
1875}
1876
1877/*
1878 * Find the next pending bucket of a level. Search from level start (@offset)
1879 * + @clk upwards and if nothing there, search from start of the level
1880 * (@offset) up to @offset + clk.
1881 */
1882static int next_pending_bucket(struct timer_base *base, unsigned offset,
1883 unsigned clk)
1884{
1885 unsigned pos, start = offset + clk;
1886 unsigned end = offset + LVL_SIZE;
1887
1888 pos = find_next_bit(base->pending_map, end, start);
1889 if (pos < end)
1890 return pos - start;
1891
1892 pos = find_next_bit(base->pending_map, start, offset);
1893 return pos < start ? pos + LVL_SIZE - start : -1;
1894}
1895
1896/*
1897 * Search the first expiring timer in the various clock levels. Caller must
1898 * hold base->lock.
1899 *
1900 * Store next expiry time in base->next_expiry.
1901 */
1902static void next_expiry_recalc(struct timer_base *base)
1903{
1904 unsigned long clk, next, adj;
1905 unsigned lvl, offset = 0;
1906
1907 next = base->clk + NEXT_TIMER_MAX_DELTA;
1908 clk = base->clk;
1909 for (lvl = 0; lvl < LVL_DEPTH; lvl++, offset += LVL_SIZE) {
1910 int pos = next_pending_bucket(base, offset, clk & LVL_MASK);
1911 unsigned long lvl_clk = clk & LVL_CLK_MASK;
1912
1913 if (pos >= 0) {
1914 unsigned long tmp = clk + (unsigned long) pos;
1915
1916 tmp <<= LVL_SHIFT(lvl);
1917 if (time_before(tmp, next))
1918 next = tmp;
1919
1920 /*
1921 * If the next expiration happens before we reach
1922 * the next level, no need to check further.
1923 */
1924 if (pos <= ((LVL_CLK_DIV - lvl_clk) & LVL_CLK_MASK))
1925 break;
1926 }
1927 /*
1928 * Clock for the next level. If the current level clock lower
1929 * bits are zero, we look at the next level as is. If not we
1930 * need to advance it by one because that's going to be the
1931 * next expiring bucket in that level. base->clk is the next
1932 * expiring jiffie. So in case of:
1933 *
1934 * LVL5 LVL4 LVL3 LVL2 LVL1 LVL0
1935 * 0 0 0 0 0 0
1936 *
1937 * we have to look at all levels @index 0. With
1938 *
1939 * LVL5 LVL4 LVL3 LVL2 LVL1 LVL0
1940 * 0 0 0 0 0 2
1941 *
1942 * LVL0 has the next expiring bucket @index 2. The upper
1943 * levels have the next expiring bucket @index 1.
1944 *
1945 * In case that the propagation wraps the next level the same
1946 * rules apply:
1947 *
1948 * LVL5 LVL4 LVL3 LVL2 LVL1 LVL0
1949 * 0 0 0 0 F 2
1950 *
1951 * So after looking at LVL0 we get:
1952 *
1953 * LVL5 LVL4 LVL3 LVL2 LVL1
1954 * 0 0 0 1 0
1955 *
1956 * So no propagation from LVL1 to LVL2 because that happened
1957 * with the add already, but then we need to propagate further
1958 * from LVL2 to LVL3.
1959 *
1960 * So the simple check whether the lower bits of the current
1961 * level are 0 or not is sufficient for all cases.
1962 */
1963 adj = lvl_clk ? 1 : 0;
1964 clk >>= LVL_CLK_SHIFT;
1965 clk += adj;
1966 }
1967
1968 base->next_expiry = next;
1969 base->next_expiry_recalc = false;
1970 base->timers_pending = !(next == base->clk + NEXT_TIMER_MAX_DELTA);
1971}
1972
1973#ifdef CONFIG_NO_HZ_COMMON
1974/*
1975 * Check, if the next hrtimer event is before the next timer wheel
1976 * event:
1977 */
1978static u64 cmp_next_hrtimer_event(u64 basem, u64 expires)
1979{
1980 u64 nextevt = hrtimer_get_next_event();
1981
1982 /*
1983 * If high resolution timers are enabled
1984 * hrtimer_get_next_event() returns KTIME_MAX.
1985 */
1986 if (expires <= nextevt)
1987 return expires;
1988
1989 /*
1990 * If the next timer is already expired, return the tick base
1991 * time so the tick is fired immediately.
1992 */
1993 if (nextevt <= basem)
1994 return basem;
1995
1996 /*
1997 * Round up to the next jiffie. High resolution timers are
1998 * off, so the hrtimers are expired in the tick and we need to
1999 * make sure that this tick really expires the timer to avoid
2000 * a ping pong of the nohz stop code.
2001 *
2002 * Use DIV_ROUND_UP_ULL to prevent gcc calling __divdi3
2003 */
2004 return DIV_ROUND_UP_ULL(nextevt, TICK_NSEC) * TICK_NSEC;
2005}
2006
2007static unsigned long next_timer_interrupt(struct timer_base *base,
2008 unsigned long basej)
2009{
2010 if (base->next_expiry_recalc)
2011 next_expiry_recalc(base);
2012
2013 /*
2014 * Move next_expiry for the empty base into the future to prevent an
2015 * unnecessary raise of the timer softirq when the next_expiry value
2016 * will be reached even if there is no timer pending.
2017 *
2018 * This update is also required to make timer_base::next_expiry values
2019 * easy comparable to find out which base holds the first pending timer.
2020 */
2021 if (!base->timers_pending)
2022 base->next_expiry = basej + NEXT_TIMER_MAX_DELTA;
2023
2024 return base->next_expiry;
2025}
2026
2027static unsigned long fetch_next_timer_interrupt(unsigned long basej, u64 basem,
2028 struct timer_base *base_local,
2029 struct timer_base *base_global,
2030 struct timer_events *tevt)
2031{
2032 unsigned long nextevt, nextevt_local, nextevt_global;
2033 bool local_first;
2034
2035 nextevt_local = next_timer_interrupt(base_local, basej);
2036 nextevt_global = next_timer_interrupt(base_global, basej);
2037
2038 local_first = time_before_eq(nextevt_local, nextevt_global);
2039
2040 nextevt = local_first ? nextevt_local : nextevt_global;
2041
2042 /*
2043 * If the @nextevt is at max. one tick away, use @nextevt and store
2044 * it in the local expiry value. The next global event is irrelevant in
2045 * this case and can be left as KTIME_MAX.
2046 */
2047 if (time_before_eq(nextevt, basej + 1)) {
2048 /* If we missed a tick already, force 0 delta */
2049 if (time_before(nextevt, basej))
2050 nextevt = basej;
2051 tevt->local = basem + (u64)(nextevt - basej) * TICK_NSEC;
2052
2053 /*
2054 * This is required for the remote check only but it doesn't
2055 * hurt, when it is done for both call sites:
2056 *
2057 * * The remote callers will only take care of the global timers
2058 * as local timers will be handled by CPU itself. When not
2059 * updating tevt->global with the already missed first global
2060 * timer, it is possible that it will be missed completely.
2061 *
2062 * * The local callers will ignore the tevt->global anyway, when
2063 * nextevt is max. one tick away.
2064 */
2065 if (!local_first)
2066 tevt->global = tevt->local;
2067 return nextevt;
2068 }
2069
2070 /*
2071 * Update tevt.* values:
2072 *
2073 * If the local queue expires first, then the global event can be
2074 * ignored. If the global queue is empty, nothing to do either.
2075 */
2076 if (!local_first && base_global->timers_pending)
2077 tevt->global = basem + (u64)(nextevt_global - basej) * TICK_NSEC;
2078
2079 if (base_local->timers_pending)
2080 tevt->local = basem + (u64)(nextevt_local - basej) * TICK_NSEC;
2081
2082 return nextevt;
2083}
2084
2085# ifdef CONFIG_SMP
2086/**
2087 * fetch_next_timer_interrupt_remote() - Store next timers into @tevt
2088 * @basej: base time jiffies
2089 * @basem: base time clock monotonic
2090 * @tevt: Pointer to the storage for the expiry values
2091 * @cpu: Remote CPU
2092 *
2093 * Stores the next pending local and global timer expiry values in the
2094 * struct pointed to by @tevt. If a queue is empty the corresponding
2095 * field is set to KTIME_MAX. If local event expires before global
2096 * event, global event is set to KTIME_MAX as well.
2097 *
2098 * Caller needs to make sure timer base locks are held (use
2099 * timer_lock_remote_bases() for this purpose).
2100 */
2101void fetch_next_timer_interrupt_remote(unsigned long basej, u64 basem,
2102 struct timer_events *tevt,
2103 unsigned int cpu)
2104{
2105 struct timer_base *base_local, *base_global;
2106
2107 /* Preset local / global events */
2108 tevt->local = tevt->global = KTIME_MAX;
2109
2110 base_local = per_cpu_ptr(&timer_bases[BASE_LOCAL], cpu);
2111 base_global = per_cpu_ptr(&timer_bases[BASE_GLOBAL], cpu);
2112
2113 lockdep_assert_held(&base_local->lock);
2114 lockdep_assert_held(&base_global->lock);
2115
2116 fetch_next_timer_interrupt(basej, basem, base_local, base_global, tevt);
2117}
2118
2119/**
2120 * timer_unlock_remote_bases - unlock timer bases of cpu
2121 * @cpu: Remote CPU
2122 *
2123 * Unlocks the remote timer bases.
2124 */
2125void timer_unlock_remote_bases(unsigned int cpu)
2126 __releases(timer_bases[BASE_LOCAL]->lock)
2127 __releases(timer_bases[BASE_GLOBAL]->lock)
2128{
2129 struct timer_base *base_local, *base_global;
2130
2131 base_local = per_cpu_ptr(&timer_bases[BASE_LOCAL], cpu);
2132 base_global = per_cpu_ptr(&timer_bases[BASE_GLOBAL], cpu);
2133
2134 raw_spin_unlock(&base_global->lock);
2135 raw_spin_unlock(&base_local->lock);
2136}
2137
2138/**
2139 * timer_lock_remote_bases - lock timer bases of cpu
2140 * @cpu: Remote CPU
2141 *
2142 * Locks the remote timer bases.
2143 */
2144void timer_lock_remote_bases(unsigned int cpu)
2145 __acquires(timer_bases[BASE_LOCAL]->lock)
2146 __acquires(timer_bases[BASE_GLOBAL]->lock)
2147{
2148 struct timer_base *base_local, *base_global;
2149
2150 base_local = per_cpu_ptr(&timer_bases[BASE_LOCAL], cpu);
2151 base_global = per_cpu_ptr(&timer_bases[BASE_GLOBAL], cpu);
2152
2153 lockdep_assert_irqs_disabled();
2154
2155 raw_spin_lock(&base_local->lock);
2156 raw_spin_lock_nested(&base_global->lock, SINGLE_DEPTH_NESTING);
2157}
2158
2159/**
2160 * timer_base_is_idle() - Return whether timer base is set idle
2161 *
2162 * Returns value of local timer base is_idle value.
2163 */
2164bool timer_base_is_idle(void)
2165{
2166 return __this_cpu_read(timer_bases[BASE_LOCAL].is_idle);
2167}
2168
2169static void __run_timer_base(struct timer_base *base);
2170
2171/**
2172 * timer_expire_remote() - expire global timers of cpu
2173 * @cpu: Remote CPU
2174 *
2175 * Expire timers of global base of remote CPU.
2176 */
2177void timer_expire_remote(unsigned int cpu)
2178{
2179 struct timer_base *base = per_cpu_ptr(&timer_bases[BASE_GLOBAL], cpu);
2180
2181 __run_timer_base(base);
2182}
2183
2184static void timer_use_tmigr(unsigned long basej, u64 basem,
2185 unsigned long *nextevt, bool *tick_stop_path,
2186 bool timer_base_idle, struct timer_events *tevt)
2187{
2188 u64 next_tmigr;
2189
2190 if (timer_base_idle)
2191 next_tmigr = tmigr_cpu_new_timer(tevt->global);
2192 else if (tick_stop_path)
2193 next_tmigr = tmigr_cpu_deactivate(tevt->global);
2194 else
2195 next_tmigr = tmigr_quick_check(tevt->global);
2196
2197 /*
2198 * If the CPU is the last going idle in timer migration hierarchy, make
2199 * sure the CPU will wake up in time to handle remote timers.
2200 * next_tmigr == KTIME_MAX if other CPUs are still active.
2201 */
2202 if (next_tmigr < tevt->local) {
2203 u64 tmp;
2204
2205 /* If we missed a tick already, force 0 delta */
2206 if (next_tmigr < basem)
2207 next_tmigr = basem;
2208
2209 tmp = div_u64(next_tmigr - basem, TICK_NSEC);
2210
2211 *nextevt = basej + (unsigned long)tmp;
2212 tevt->local = next_tmigr;
2213 }
2214}
2215# else
2216static void timer_use_tmigr(unsigned long basej, u64 basem,
2217 unsigned long *nextevt, bool *tick_stop_path,
2218 bool timer_base_idle, struct timer_events *tevt)
2219{
2220 /*
2221 * Make sure first event is written into tevt->local to not miss a
2222 * timer on !SMP systems.
2223 */
2224 tevt->local = min_t(u64, tevt->local, tevt->global);
2225}
2226# endif /* CONFIG_SMP */
2227
2228static inline u64 __get_next_timer_interrupt(unsigned long basej, u64 basem,
2229 bool *idle)
2230{
2231 struct timer_events tevt = { .local = KTIME_MAX, .global = KTIME_MAX };
2232 struct timer_base *base_local, *base_global;
2233 unsigned long nextevt;
2234 bool idle_is_possible;
2235
2236 /*
2237 * When the CPU is offline, the tick is cancelled and nothing is supposed
2238 * to try to stop it.
2239 */
2240 if (WARN_ON_ONCE(cpu_is_offline(smp_processor_id()))) {
2241 if (idle)
2242 *idle = true;
2243 return tevt.local;
2244 }
2245
2246 base_local = this_cpu_ptr(&timer_bases[BASE_LOCAL]);
2247 base_global = this_cpu_ptr(&timer_bases[BASE_GLOBAL]);
2248
2249 raw_spin_lock(&base_local->lock);
2250 raw_spin_lock_nested(&base_global->lock, SINGLE_DEPTH_NESTING);
2251
2252 nextevt = fetch_next_timer_interrupt(basej, basem, base_local,
2253 base_global, &tevt);
2254
2255 /*
2256 * If the next event is only one jiffie ahead there is no need to call
2257 * timer migration hierarchy related functions. The value for the next
2258 * global timer in @tevt struct equals then KTIME_MAX. This is also
2259 * true, when the timer base is idle.
2260 *
2261 * The proper timer migration hierarchy function depends on the callsite
2262 * and whether timer base is idle or not. @nextevt will be updated when
2263 * this CPU needs to handle the first timer migration hierarchy
2264 * event. See timer_use_tmigr() for detailed information.
2265 */
2266 idle_is_possible = time_after(nextevt, basej + 1);
2267 if (idle_is_possible)
2268 timer_use_tmigr(basej, basem, &nextevt, idle,
2269 base_local->is_idle, &tevt);
2270
2271 /*
2272 * We have a fresh next event. Check whether we can forward the
2273 * base.
2274 */
2275 __forward_timer_base(base_local, basej);
2276 __forward_timer_base(base_global, basej);
2277
2278 /*
2279 * Set base->is_idle only when caller is timer_base_try_to_set_idle()
2280 */
2281 if (idle) {
2282 /*
2283 * Bases are idle if the next event is more than a tick
2284 * away. Caution: @nextevt could have changed by enqueueing a
2285 * global timer into timer migration hierarchy. Therefore a new
2286 * check is required here.
2287 *
2288 * If the base is marked idle then any timer add operation must
2289 * forward the base clk itself to keep granularity small. This
2290 * idle logic is only maintained for the BASE_LOCAL and
2291 * BASE_GLOBAL base, deferrable timers may still see large
2292 * granularity skew (by design).
2293 */
2294 if (!base_local->is_idle && time_after(nextevt, basej + 1)) {
2295 base_local->is_idle = true;
2296 /*
2297 * Global timers queued locally while running in a task
2298 * in nohz_full mode need a self-IPI to kick reprogramming
2299 * in IRQ tail.
2300 */
2301 if (tick_nohz_full_cpu(base_local->cpu))
2302 base_global->is_idle = true;
2303 trace_timer_base_idle(true, base_local->cpu);
2304 }
2305 *idle = base_local->is_idle;
2306
2307 /*
2308 * When timer base is not set idle, undo the effect of
2309 * tmigr_cpu_deactivate() to prevent inconsistent states - active
2310 * timer base but inactive timer migration hierarchy.
2311 *
2312 * When timer base was already marked idle, nothing will be
2313 * changed here.
2314 */
2315 if (!base_local->is_idle && idle_is_possible)
2316 tmigr_cpu_activate();
2317 }
2318
2319 raw_spin_unlock(&base_global->lock);
2320 raw_spin_unlock(&base_local->lock);
2321
2322 return cmp_next_hrtimer_event(basem, tevt.local);
2323}
2324
2325/**
2326 * get_next_timer_interrupt() - return the time (clock mono) of the next timer
2327 * @basej: base time jiffies
2328 * @basem: base time clock monotonic
2329 *
2330 * Returns the tick aligned clock monotonic time of the next pending timer or
2331 * KTIME_MAX if no timer is pending. If timer of global base was queued into
2332 * timer migration hierarchy, first global timer is not taken into account. If
2333 * it was the last CPU of timer migration hierarchy going idle, first global
2334 * event is taken into account.
2335 */
2336u64 get_next_timer_interrupt(unsigned long basej, u64 basem)
2337{
2338 return __get_next_timer_interrupt(basej, basem, NULL);
2339}
2340
2341/**
2342 * timer_base_try_to_set_idle() - Try to set the idle state of the timer bases
2343 * @basej: base time jiffies
2344 * @basem: base time clock monotonic
2345 * @idle: pointer to store the value of timer_base->is_idle on return;
2346 * *idle contains the information whether tick was already stopped
2347 *
2348 * Returns the tick aligned clock monotonic time of the next pending timer or
2349 * KTIME_MAX if no timer is pending. When tick was already stopped KTIME_MAX is
2350 * returned as well.
2351 */
2352u64 timer_base_try_to_set_idle(unsigned long basej, u64 basem, bool *idle)
2353{
2354 if (*idle)
2355 return KTIME_MAX;
2356
2357 return __get_next_timer_interrupt(basej, basem, idle);
2358}
2359
2360/**
2361 * timer_clear_idle - Clear the idle state of the timer base
2362 *
2363 * Called with interrupts disabled
2364 */
2365void timer_clear_idle(void)
2366{
2367 /*
2368 * We do this unlocked. The worst outcome is a remote pinned timer
2369 * enqueue sending a pointless IPI, but taking the lock would just
2370 * make the window for sending the IPI a few instructions smaller
2371 * for the cost of taking the lock in the exit from idle
2372 * path. Required for BASE_LOCAL only.
2373 */
2374 __this_cpu_write(timer_bases[BASE_LOCAL].is_idle, false);
2375 if (tick_nohz_full_cpu(smp_processor_id()))
2376 __this_cpu_write(timer_bases[BASE_GLOBAL].is_idle, false);
2377 trace_timer_base_idle(false, smp_processor_id());
2378
2379 /* Activate without holding the timer_base->lock */
2380 tmigr_cpu_activate();
2381}
2382#endif
2383
2384/**
2385 * __run_timers - run all expired timers (if any) on this CPU.
2386 * @base: the timer vector to be processed.
2387 */
2388static inline void __run_timers(struct timer_base *base)
2389{
2390 struct hlist_head heads[LVL_DEPTH];
2391 int levels;
2392
2393 lockdep_assert_held(&base->lock);
2394
2395 if (base->running_timer)
2396 return;
2397
2398 while (time_after_eq(jiffies, base->clk) &&
2399 time_after_eq(jiffies, base->next_expiry)) {
2400 levels = collect_expired_timers(base, heads);
2401 /*
2402 * The two possible reasons for not finding any expired
2403 * timer at this clk are that all matching timers have been
2404 * dequeued or no timer has been queued since
2405 * base::next_expiry was set to base::clk +
2406 * NEXT_TIMER_MAX_DELTA.
2407 */
2408 WARN_ON_ONCE(!levels && !base->next_expiry_recalc
2409 && base->timers_pending);
2410 /*
2411 * While executing timers, base->clk is set 1 offset ahead of
2412 * jiffies to avoid endless requeuing to current jiffies.
2413 */
2414 base->clk++;
2415 next_expiry_recalc(base);
2416
2417 while (levels--)
2418 expire_timers(base, heads + levels);
2419 }
2420}
2421
2422static void __run_timer_base(struct timer_base *base)
2423{
2424 if (time_before(jiffies, base->next_expiry))
2425 return;
2426
2427 timer_base_lock_expiry(base);
2428 raw_spin_lock_irq(&base->lock);
2429 __run_timers(base);
2430 raw_spin_unlock_irq(&base->lock);
2431 timer_base_unlock_expiry(base);
2432}
2433
2434static void run_timer_base(int index)
2435{
2436 struct timer_base *base = this_cpu_ptr(&timer_bases[index]);
2437
2438 __run_timer_base(base);
2439}
2440
2441/*
2442 * This function runs timers and the timer-tq in bottom half context.
2443 */
2444static __latent_entropy void run_timer_softirq(struct softirq_action *h)
2445{
2446 run_timer_base(BASE_LOCAL);
2447 if (IS_ENABLED(CONFIG_NO_HZ_COMMON)) {
2448 run_timer_base(BASE_GLOBAL);
2449 run_timer_base(BASE_DEF);
2450
2451 if (is_timers_nohz_active())
2452 tmigr_handle_remote();
2453 }
2454}
2455
2456/*
2457 * Called by the local, per-CPU timer interrupt on SMP.
2458 */
2459static void run_local_timers(void)
2460{
2461 struct timer_base *base = this_cpu_ptr(&timer_bases[BASE_LOCAL]);
2462
2463 hrtimer_run_queues();
2464
2465 for (int i = 0; i < NR_BASES; i++, base++) {
2466 /* Raise the softirq only if required. */
2467 if (time_after_eq(jiffies, base->next_expiry) ||
2468 (i == BASE_DEF && tmigr_requires_handle_remote())) {
2469 raise_softirq(TIMER_SOFTIRQ);
2470 return;
2471 }
2472 }
2473}
2474
2475/*
2476 * Called from the timer interrupt handler to charge one tick to the current
2477 * process. user_tick is 1 if the tick is user time, 0 for system.
2478 */
2479void update_process_times(int user_tick)
2480{
2481 struct task_struct *p = current;
2482
2483 /* Note: this timer irq context must be accounted for as well. */
2484 account_process_tick(p, user_tick);
2485 run_local_timers();
2486 rcu_sched_clock_irq(user_tick);
2487#ifdef CONFIG_IRQ_WORK
2488 if (in_irq())
2489 irq_work_tick();
2490#endif
2491 scheduler_tick();
2492 if (IS_ENABLED(CONFIG_POSIX_TIMERS))
2493 run_posix_cpu_timers();
2494}
2495
2496/*
2497 * Since schedule_timeout()'s timer is defined on the stack, it must store
2498 * the target task on the stack as well.
2499 */
2500struct process_timer {
2501 struct timer_list timer;
2502 struct task_struct *task;
2503};
2504
2505static void process_timeout(struct timer_list *t)
2506{
2507 struct process_timer *timeout = from_timer(timeout, t, timer);
2508
2509 wake_up_process(timeout->task);
2510}
2511
2512/**
2513 * schedule_timeout - sleep until timeout
2514 * @timeout: timeout value in jiffies
2515 *
2516 * Make the current task sleep until @timeout jiffies have elapsed.
2517 * The function behavior depends on the current task state
2518 * (see also set_current_state() description):
2519 *
2520 * %TASK_RUNNING - the scheduler is called, but the task does not sleep
2521 * at all. That happens because sched_submit_work() does nothing for
2522 * tasks in %TASK_RUNNING state.
2523 *
2524 * %TASK_UNINTERRUPTIBLE - at least @timeout jiffies are guaranteed to
2525 * pass before the routine returns unless the current task is explicitly
2526 * woken up, (e.g. by wake_up_process()).
2527 *
2528 * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
2529 * delivered to the current task or the current task is explicitly woken
2530 * up.
2531 *
2532 * The current task state is guaranteed to be %TASK_RUNNING when this
2533 * routine returns.
2534 *
2535 * Specifying a @timeout value of %MAX_SCHEDULE_TIMEOUT will schedule
2536 * the CPU away without a bound on the timeout. In this case the return
2537 * value will be %MAX_SCHEDULE_TIMEOUT.
2538 *
2539 * Returns 0 when the timer has expired otherwise the remaining time in
2540 * jiffies will be returned. In all cases the return value is guaranteed
2541 * to be non-negative.
2542 */
2543signed long __sched schedule_timeout(signed long timeout)
2544{
2545 struct process_timer timer;
2546 unsigned long expire;
2547
2548 switch (timeout)
2549 {
2550 case MAX_SCHEDULE_TIMEOUT:
2551 /*
2552 * These two special cases are useful to be comfortable
2553 * in the caller. Nothing more. We could take
2554 * MAX_SCHEDULE_TIMEOUT from one of the negative value
2555 * but I' d like to return a valid offset (>=0) to allow
2556 * the caller to do everything it want with the retval.
2557 */
2558 schedule();
2559 goto out;
2560 default:
2561 /*
2562 * Another bit of PARANOID. Note that the retval will be
2563 * 0 since no piece of kernel is supposed to do a check
2564 * for a negative retval of schedule_timeout() (since it
2565 * should never happens anyway). You just have the printk()
2566 * that will tell you if something is gone wrong and where.
2567 */
2568 if (timeout < 0) {
2569 printk(KERN_ERR "schedule_timeout: wrong timeout "
2570 "value %lx\n", timeout);
2571 dump_stack();
2572 __set_current_state(TASK_RUNNING);
2573 goto out;
2574 }
2575 }
2576
2577 expire = timeout + jiffies;
2578
2579 timer.task = current;
2580 timer_setup_on_stack(&timer.timer, process_timeout, 0);
2581 __mod_timer(&timer.timer, expire, MOD_TIMER_NOTPENDING);
2582 schedule();
2583 del_timer_sync(&timer.timer);
2584
2585 /* Remove the timer from the object tracker */
2586 destroy_timer_on_stack(&timer.timer);
2587
2588 timeout = expire - jiffies;
2589
2590 out:
2591 return timeout < 0 ? 0 : timeout;
2592}
2593EXPORT_SYMBOL(schedule_timeout);
2594
2595/*
2596 * We can use __set_current_state() here because schedule_timeout() calls
2597 * schedule() unconditionally.
2598 */
2599signed long __sched schedule_timeout_interruptible(signed long timeout)
2600{
2601 __set_current_state(TASK_INTERRUPTIBLE);
2602 return schedule_timeout(timeout);
2603}
2604EXPORT_SYMBOL(schedule_timeout_interruptible);
2605
2606signed long __sched schedule_timeout_killable(signed long timeout)
2607{
2608 __set_current_state(TASK_KILLABLE);
2609 return schedule_timeout(timeout);
2610}
2611EXPORT_SYMBOL(schedule_timeout_killable);
2612
2613signed long __sched schedule_timeout_uninterruptible(signed long timeout)
2614{
2615 __set_current_state(TASK_UNINTERRUPTIBLE);
2616 return schedule_timeout(timeout);
2617}
2618EXPORT_SYMBOL(schedule_timeout_uninterruptible);
2619
2620/*
2621 * Like schedule_timeout_uninterruptible(), except this task will not contribute
2622 * to load average.
2623 */
2624signed long __sched schedule_timeout_idle(signed long timeout)
2625{
2626 __set_current_state(TASK_IDLE);
2627 return schedule_timeout(timeout);
2628}
2629EXPORT_SYMBOL(schedule_timeout_idle);
2630
2631#ifdef CONFIG_HOTPLUG_CPU
2632static void migrate_timer_list(struct timer_base *new_base, struct hlist_head *head)
2633{
2634 struct timer_list *timer;
2635 int cpu = new_base->cpu;
2636
2637 while (!hlist_empty(head)) {
2638 timer = hlist_entry(head->first, struct timer_list, entry);
2639 detach_timer(timer, false);
2640 timer->flags = (timer->flags & ~TIMER_BASEMASK) | cpu;
2641 internal_add_timer(new_base, timer);
2642 }
2643}
2644
2645int timers_prepare_cpu(unsigned int cpu)
2646{
2647 struct timer_base *base;
2648 int b;
2649
2650 for (b = 0; b < NR_BASES; b++) {
2651 base = per_cpu_ptr(&timer_bases[b], cpu);
2652 base->clk = jiffies;
2653 base->next_expiry = base->clk + NEXT_TIMER_MAX_DELTA;
2654 base->next_expiry_recalc = false;
2655 base->timers_pending = false;
2656 base->is_idle = false;
2657 }
2658 return 0;
2659}
2660
2661int timers_dead_cpu(unsigned int cpu)
2662{
2663 struct timer_base *old_base;
2664 struct timer_base *new_base;
2665 int b, i;
2666
2667 for (b = 0; b < NR_BASES; b++) {
2668 old_base = per_cpu_ptr(&timer_bases[b], cpu);
2669 new_base = get_cpu_ptr(&timer_bases[b]);
2670 /*
2671 * The caller is globally serialized and nobody else
2672 * takes two locks at once, deadlock is not possible.
2673 */
2674 raw_spin_lock_irq(&new_base->lock);
2675 raw_spin_lock_nested(&old_base->lock, SINGLE_DEPTH_NESTING);
2676
2677 /*
2678 * The current CPUs base clock might be stale. Update it
2679 * before moving the timers over.
2680 */
2681 forward_timer_base(new_base);
2682
2683 WARN_ON_ONCE(old_base->running_timer);
2684 old_base->running_timer = NULL;
2685
2686 for (i = 0; i < WHEEL_SIZE; i++)
2687 migrate_timer_list(new_base, old_base->vectors + i);
2688
2689 raw_spin_unlock(&old_base->lock);
2690 raw_spin_unlock_irq(&new_base->lock);
2691 put_cpu_ptr(&timer_bases);
2692 }
2693 return 0;
2694}
2695
2696#endif /* CONFIG_HOTPLUG_CPU */
2697
2698static void __init init_timer_cpu(int cpu)
2699{
2700 struct timer_base *base;
2701 int i;
2702
2703 for (i = 0; i < NR_BASES; i++) {
2704 base = per_cpu_ptr(&timer_bases[i], cpu);
2705 base->cpu = cpu;
2706 raw_spin_lock_init(&base->lock);
2707 base->clk = jiffies;
2708 base->next_expiry = base->clk + NEXT_TIMER_MAX_DELTA;
2709 timer_base_init_expiry_lock(base);
2710 }
2711}
2712
2713static void __init init_timer_cpus(void)
2714{
2715 int cpu;
2716
2717 for_each_possible_cpu(cpu)
2718 init_timer_cpu(cpu);
2719}
2720
2721void __init init_timers(void)
2722{
2723 init_timer_cpus();
2724 posix_cputimers_init_work();
2725 open_softirq(TIMER_SOFTIRQ, run_timer_softirq);
2726}
2727
2728/**
2729 * msleep - sleep safely even with waitqueue interruptions
2730 * @msecs: Time in milliseconds to sleep for
2731 */
2732void msleep(unsigned int msecs)
2733{
2734 unsigned long timeout = msecs_to_jiffies(msecs) + 1;
2735
2736 while (timeout)
2737 timeout = schedule_timeout_uninterruptible(timeout);
2738}
2739
2740EXPORT_SYMBOL(msleep);
2741
2742/**
2743 * msleep_interruptible - sleep waiting for signals
2744 * @msecs: Time in milliseconds to sleep for
2745 */
2746unsigned long msleep_interruptible(unsigned int msecs)
2747{
2748 unsigned long timeout = msecs_to_jiffies(msecs) + 1;
2749
2750 while (timeout && !signal_pending(current))
2751 timeout = schedule_timeout_interruptible(timeout);
2752 return jiffies_to_msecs(timeout);
2753}
2754
2755EXPORT_SYMBOL(msleep_interruptible);
2756
2757/**
2758 * usleep_range_state - Sleep for an approximate time in a given state
2759 * @min: Minimum time in usecs to sleep
2760 * @max: Maximum time in usecs to sleep
2761 * @state: State of the current task that will be while sleeping
2762 *
2763 * In non-atomic context where the exact wakeup time is flexible, use
2764 * usleep_range_state() instead of udelay(). The sleep improves responsiveness
2765 * by avoiding the CPU-hogging busy-wait of udelay(), and the range reduces
2766 * power usage by allowing hrtimers to take advantage of an already-
2767 * scheduled interrupt instead of scheduling a new one just for this sleep.
2768 */
2769void __sched usleep_range_state(unsigned long min, unsigned long max,
2770 unsigned int state)
2771{
2772 ktime_t exp = ktime_add_us(ktime_get(), min);
2773 u64 delta = (u64)(max - min) * NSEC_PER_USEC;
2774
2775 for (;;) {
2776 __set_current_state(state);
2777 /* Do not return before the requested sleep time has elapsed */
2778 if (!schedule_hrtimeout_range(&exp, delta, HRTIMER_MODE_ABS))
2779 break;
2780 }
2781}
2782EXPORT_SYMBOL(usleep_range_state);