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
48#include <linux/uaccess.h>
49#include <asm/unistd.h>
50#include <asm/div64.h>
51#include <asm/timex.h>
52#include <asm/io.h>
53
54#include "tick-internal.h"
55
56#define CREATE_TRACE_POINTS
57#include <trace/events/timer.h>
58
59__visible u64 jiffies_64 __cacheline_aligned_in_smp = INITIAL_JIFFIES;
60
61EXPORT_SYMBOL(jiffies_64);
62
63/*
64 * The timer wheel has LVL_DEPTH array levels. Each level provides an array of
65 * LVL_SIZE buckets. Each level is driven by its own clock and therefor each
66 * level has a different granularity.
67 *
68 * The level granularity is: LVL_CLK_DIV ^ lvl
69 * The level clock frequency is: HZ / (LVL_CLK_DIV ^ level)
70 *
71 * The array level of a newly armed timer depends on the relative expiry
72 * time. The farther the expiry time is away the higher the array level and
73 * therefor the granularity becomes.
74 *
75 * Contrary to the original timer wheel implementation, which aims for 'exact'
76 * expiry of the timers, this implementation removes the need for recascading
77 * the timers into the lower array levels. The previous 'classic' timer wheel
78 * implementation of the kernel already violated the 'exact' expiry by adding
79 * slack to the expiry time to provide batched expiration. The granularity
80 * levels provide implicit batching.
81 *
82 * This is an optimization of the original timer wheel implementation for the
83 * majority of the timer wheel use cases: timeouts. The vast majority of
84 * timeout timers (networking, disk I/O ...) are canceled before expiry. If
85 * the timeout expires it indicates that normal operation is disturbed, so it
86 * does not matter much whether the timeout comes with a slight delay.
87 *
88 * The only exception to this are networking timers with a small expiry
89 * time. They rely on the granularity. Those fit into the first wheel level,
90 * which has HZ granularity.
91 *
92 * We don't have cascading anymore. timers with a expiry time above the
93 * capacity of the last wheel level are force expired at the maximum timeout
94 * value of the last wheel level. From data sampling we know that the maximum
95 * value observed is 5 days (network connection tracking), so this should not
96 * be an issue.
97 *
98 * The currently chosen array constants values are a good compromise between
99 * array size and granularity.
100 *
101 * This results in the following granularity and range levels:
102 *
103 * HZ 1000 steps
104 * Level Offset Granularity Range
105 * 0 0 1 ms 0 ms - 63 ms
106 * 1 64 8 ms 64 ms - 511 ms
107 * 2 128 64 ms 512 ms - 4095 ms (512ms - ~4s)
108 * 3 192 512 ms 4096 ms - 32767 ms (~4s - ~32s)
109 * 4 256 4096 ms (~4s) 32768 ms - 262143 ms (~32s - ~4m)
110 * 5 320 32768 ms (~32s) 262144 ms - 2097151 ms (~4m - ~34m)
111 * 6 384 262144 ms (~4m) 2097152 ms - 16777215 ms (~34m - ~4h)
112 * 7 448 2097152 ms (~34m) 16777216 ms - 134217727 ms (~4h - ~1d)
113 * 8 512 16777216 ms (~4h) 134217728 ms - 1073741822 ms (~1d - ~12d)
114 *
115 * HZ 300
116 * Level Offset Granularity Range
117 * 0 0 3 ms 0 ms - 210 ms
118 * 1 64 26 ms 213 ms - 1703 ms (213ms - ~1s)
119 * 2 128 213 ms 1706 ms - 13650 ms (~1s - ~13s)
120 * 3 192 1706 ms (~1s) 13653 ms - 109223 ms (~13s - ~1m)
121 * 4 256 13653 ms (~13s) 109226 ms - 873810 ms (~1m - ~14m)
122 * 5 320 109226 ms (~1m) 873813 ms - 6990503 ms (~14m - ~1h)
123 * 6 384 873813 ms (~14m) 6990506 ms - 55924050 ms (~1h - ~15h)
124 * 7 448 6990506 ms (~1h) 55924053 ms - 447392423 ms (~15h - ~5d)
125 * 8 512 55924053 ms (~15h) 447392426 ms - 3579139406 ms (~5d - ~41d)
126 *
127 * HZ 250
128 * Level Offset Granularity Range
129 * 0 0 4 ms 0 ms - 255 ms
130 * 1 64 32 ms 256 ms - 2047 ms (256ms - ~2s)
131 * 2 128 256 ms 2048 ms - 16383 ms (~2s - ~16s)
132 * 3 192 2048 ms (~2s) 16384 ms - 131071 ms (~16s - ~2m)
133 * 4 256 16384 ms (~16s) 131072 ms - 1048575 ms (~2m - ~17m)
134 * 5 320 131072 ms (~2m) 1048576 ms - 8388607 ms (~17m - ~2h)
135 * 6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
136 * 7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
137 * 8 512 67108864 ms (~18h) 536870912 ms - 4294967288 ms (~6d - ~49d)
138 *
139 * HZ 100
140 * Level Offset Granularity Range
141 * 0 0 10 ms 0 ms - 630 ms
142 * 1 64 80 ms 640 ms - 5110 ms (640ms - ~5s)
143 * 2 128 640 ms 5120 ms - 40950 ms (~5s - ~40s)
144 * 3 192 5120 ms (~5s) 40960 ms - 327670 ms (~40s - ~5m)
145 * 4 256 40960 ms (~40s) 327680 ms - 2621430 ms (~5m - ~43m)
146 * 5 320 327680 ms (~5m) 2621440 ms - 20971510 ms (~43m - ~5h)
147 * 6 384 2621440 ms (~43m) 20971520 ms - 167772150 ms (~5h - ~1d)
148 * 7 448 20971520 ms (~5h) 167772160 ms - 1342177270 ms (~1d - ~15d)
149 */
150
151/* Clock divisor for the next level */
152#define LVL_CLK_SHIFT 3
153#define LVL_CLK_DIV (1UL << LVL_CLK_SHIFT)
154#define LVL_CLK_MASK (LVL_CLK_DIV - 1)
155#define LVL_SHIFT(n) ((n) * LVL_CLK_SHIFT)
156#define LVL_GRAN(n) (1UL << LVL_SHIFT(n))
157
158/*
159 * The time start value for each level to select the bucket at enqueue
160 * time. We start from the last possible delta of the previous level
161 * so that we can later add an extra LVL_GRAN(n) to n (see calc_index()).
162 */
163#define LVL_START(n) ((LVL_SIZE - 1) << (((n) - 1) * LVL_CLK_SHIFT))
164
165/* Size of each clock level */
166#define LVL_BITS 6
167#define LVL_SIZE (1UL << LVL_BITS)
168#define LVL_MASK (LVL_SIZE - 1)
169#define LVL_OFFS(n) ((n) * LVL_SIZE)
170
171/* Level depth */
172#if HZ > 100
173# define LVL_DEPTH 9
174# else
175# define LVL_DEPTH 8
176#endif
177
178/* The cutoff (max. capacity of the wheel) */
179#define WHEEL_TIMEOUT_CUTOFF (LVL_START(LVL_DEPTH))
180#define WHEEL_TIMEOUT_MAX (WHEEL_TIMEOUT_CUTOFF - LVL_GRAN(LVL_DEPTH - 1))
181
182/*
183 * The resulting wheel size. If NOHZ is configured we allocate two
184 * wheels so we have a separate storage for the deferrable timers.
185 */
186#define WHEEL_SIZE (LVL_SIZE * LVL_DEPTH)
187
188#ifdef CONFIG_NO_HZ_COMMON
189# define NR_BASES 2
190# define BASE_STD 0
191# define BASE_DEF 1
192#else
193# define NR_BASES 1
194# define BASE_STD 0
195# define BASE_DEF 0
196#endif
197
198struct timer_base {
199 raw_spinlock_t lock;
200 struct timer_list *running_timer;
201#ifdef CONFIG_PREEMPT_RT
202 spinlock_t expiry_lock;
203 atomic_t timer_waiters;
204#endif
205 unsigned long clk;
206 unsigned long next_expiry;
207 unsigned int cpu;
208 bool next_expiry_recalc;
209 bool is_idle;
210 DECLARE_BITMAP(pending_map, WHEEL_SIZE);
211 struct hlist_head vectors[WHEEL_SIZE];
212} ____cacheline_aligned;
213
214static DEFINE_PER_CPU(struct timer_base, timer_bases[NR_BASES]);
215
216#ifdef CONFIG_NO_HZ_COMMON
217
218static DEFINE_STATIC_KEY_FALSE(timers_nohz_active);
219static DEFINE_MUTEX(timer_keys_mutex);
220
221static void timer_update_keys(struct work_struct *work);
222static DECLARE_WORK(timer_update_work, timer_update_keys);
223
224#ifdef CONFIG_SMP
225unsigned int sysctl_timer_migration = 1;
226
227DEFINE_STATIC_KEY_FALSE(timers_migration_enabled);
228
229static void timers_update_migration(void)
230{
231 if (sysctl_timer_migration && tick_nohz_active)
232 static_branch_enable(&timers_migration_enabled);
233 else
234 static_branch_disable(&timers_migration_enabled);
235}
236#else
237static inline void timers_update_migration(void) { }
238#endif /* !CONFIG_SMP */
239
240static void timer_update_keys(struct work_struct *work)
241{
242 mutex_lock(&timer_keys_mutex);
243 timers_update_migration();
244 static_branch_enable(&timers_nohz_active);
245 mutex_unlock(&timer_keys_mutex);
246}
247
248void timers_update_nohz(void)
249{
250 schedule_work(&timer_update_work);
251}
252
253int timer_migration_handler(struct ctl_table *table, int write,
254 void *buffer, size_t *lenp, loff_t *ppos)
255{
256 int ret;
257
258 mutex_lock(&timer_keys_mutex);
259 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
260 if (!ret && write)
261 timers_update_migration();
262 mutex_unlock(&timer_keys_mutex);
263 return ret;
264}
265
266static inline bool is_timers_nohz_active(void)
267{
268 return static_branch_unlikely(&timers_nohz_active);
269}
270#else
271static inline bool is_timers_nohz_active(void) { return false; }
272#endif /* NO_HZ_COMMON */
273
274static unsigned long round_jiffies_common(unsigned long j, int cpu,
275 bool force_up)
276{
277 int rem;
278 unsigned long original = j;
279
280 /*
281 * We don't want all cpus firing their timers at once hitting the
282 * same lock or cachelines, so we skew each extra cpu with an extra
283 * 3 jiffies. This 3 jiffies came originally from the mm/ code which
284 * already did this.
285 * The skew is done by adding 3*cpunr, then round, then subtract this
286 * extra offset again.
287 */
288 j += cpu * 3;
289
290 rem = j % HZ;
291
292 /*
293 * If the target jiffie is just after a whole second (which can happen
294 * due to delays of the timer irq, long irq off times etc etc) then
295 * we should round down to the whole second, not up. Use 1/4th second
296 * as cutoff for this rounding as an extreme upper bound for this.
297 * But never round down if @force_up is set.
298 */
299 if (rem < HZ/4 && !force_up) /* round down */
300 j = j - rem;
301 else /* round up */
302 j = j - rem + HZ;
303
304 /* now that we have rounded, subtract the extra skew again */
305 j -= cpu * 3;
306
307 /*
308 * Make sure j is still in the future. Otherwise return the
309 * unmodified value.
310 */
311 return time_is_after_jiffies(j) ? j : original;
312}
313
314/**
315 * __round_jiffies - function to round jiffies to a full second
316 * @j: the time in (absolute) jiffies that should be rounded
317 * @cpu: the processor number on which the timeout will happen
318 *
319 * __round_jiffies() rounds an absolute time in the future (in jiffies)
320 * up or down to (approximately) full seconds. This is useful for timers
321 * for which the exact time they fire does not matter too much, as long as
322 * they fire approximately every X seconds.
323 *
324 * By rounding these timers to whole seconds, all such timers will fire
325 * at the same time, rather than at various times spread out. The goal
326 * of this is to have the CPU wake up less, which saves power.
327 *
328 * The exact rounding is skewed for each processor to avoid all
329 * processors firing at the exact same time, which could lead
330 * to lock contention or spurious cache line bouncing.
331 *
332 * The return value is the rounded version of the @j parameter.
333 */
334unsigned long __round_jiffies(unsigned long j, int cpu)
335{
336 return round_jiffies_common(j, cpu, false);
337}
338EXPORT_SYMBOL_GPL(__round_jiffies);
339
340/**
341 * __round_jiffies_relative - function to round jiffies to a full second
342 * @j: the time in (relative) jiffies that should be rounded
343 * @cpu: the processor number on which the timeout will happen
344 *
345 * __round_jiffies_relative() rounds a time delta in the future (in jiffies)
346 * up or down to (approximately) full seconds. This is useful for timers
347 * for which the exact time they fire does not matter too much, as long as
348 * they fire approximately every X seconds.
349 *
350 * By rounding these timers to whole seconds, all such timers will fire
351 * at the same time, rather than at various times spread out. The goal
352 * of this is to have the CPU wake up less, which saves power.
353 *
354 * The exact rounding is skewed for each processor to avoid all
355 * processors firing at the exact same time, which could lead
356 * to lock contention or spurious cache line bouncing.
357 *
358 * The return value is the rounded version of the @j parameter.
359 */
360unsigned long __round_jiffies_relative(unsigned long j, int cpu)
361{
362 unsigned long j0 = jiffies;
363
364 /* Use j0 because jiffies might change while we run */
365 return round_jiffies_common(j + j0, cpu, false) - j0;
366}
367EXPORT_SYMBOL_GPL(__round_jiffies_relative);
368
369/**
370 * round_jiffies - function to round jiffies to a full second
371 * @j: the time in (absolute) jiffies that should be rounded
372 *
373 * round_jiffies() rounds an absolute time in the future (in jiffies)
374 * up or down to (approximately) full seconds. This is useful for timers
375 * for which the exact time they fire does not matter too much, as long as
376 * they fire approximately every X seconds.
377 *
378 * By rounding these timers to whole seconds, all such timers will fire
379 * at the same time, rather than at various times spread out. The goal
380 * of this is to have the CPU wake up less, which saves power.
381 *
382 * The return value is the rounded version of the @j parameter.
383 */
384unsigned long round_jiffies(unsigned long j)
385{
386 return round_jiffies_common(j, raw_smp_processor_id(), false);
387}
388EXPORT_SYMBOL_GPL(round_jiffies);
389
390/**
391 * round_jiffies_relative - function to round jiffies to a full second
392 * @j: the time in (relative) jiffies that should be rounded
393 *
394 * round_jiffies_relative() rounds a time delta in the future (in jiffies)
395 * up or down to (approximately) full seconds. This is useful for timers
396 * for which the exact time they fire does not matter too much, as long as
397 * they fire approximately every X seconds.
398 *
399 * By rounding these timers to whole seconds, all such timers will fire
400 * at the same time, rather than at various times spread out. The goal
401 * of this is to have the CPU wake up less, which saves power.
402 *
403 * The return value is the rounded version of the @j parameter.
404 */
405unsigned long round_jiffies_relative(unsigned long j)
406{
407 return __round_jiffies_relative(j, raw_smp_processor_id());
408}
409EXPORT_SYMBOL_GPL(round_jiffies_relative);
410
411/**
412 * __round_jiffies_up - function to round jiffies up to a full second
413 * @j: the time in (absolute) jiffies that should be rounded
414 * @cpu: the processor number on which the timeout will happen
415 *
416 * This is the same as __round_jiffies() except that it will never
417 * round down. This is useful for timeouts for which the exact time
418 * of firing does not matter too much, as long as they don't fire too
419 * early.
420 */
421unsigned long __round_jiffies_up(unsigned long j, int cpu)
422{
423 return round_jiffies_common(j, cpu, true);
424}
425EXPORT_SYMBOL_GPL(__round_jiffies_up);
426
427/**
428 * __round_jiffies_up_relative - function to round jiffies up to a full second
429 * @j: the time in (relative) jiffies that should be rounded
430 * @cpu: the processor number on which the timeout will happen
431 *
432 * This is the same as __round_jiffies_relative() except that it will never
433 * round down. This is useful for timeouts for which the exact time
434 * of firing does not matter too much, as long as they don't fire too
435 * early.
436 */
437unsigned long __round_jiffies_up_relative(unsigned long j, int cpu)
438{
439 unsigned long j0 = jiffies;
440
441 /* Use j0 because jiffies might change while we run */
442 return round_jiffies_common(j + j0, cpu, true) - j0;
443}
444EXPORT_SYMBOL_GPL(__round_jiffies_up_relative);
445
446/**
447 * round_jiffies_up - function to round jiffies up to a full second
448 * @j: the time in (absolute) jiffies that should be rounded
449 *
450 * This is the same as round_jiffies() except that it will never
451 * round down. This is useful for timeouts for which the exact time
452 * of firing does not matter too much, as long as they don't fire too
453 * early.
454 */
455unsigned long round_jiffies_up(unsigned long j)
456{
457 return round_jiffies_common(j, raw_smp_processor_id(), true);
458}
459EXPORT_SYMBOL_GPL(round_jiffies_up);
460
461/**
462 * round_jiffies_up_relative - function to round jiffies up to a full second
463 * @j: the time in (relative) jiffies that should be rounded
464 *
465 * This is the same as round_jiffies_relative() except that it will never
466 * round down. This is useful for timeouts for which the exact time
467 * of firing does not matter too much, as long as they don't fire too
468 * early.
469 */
470unsigned long round_jiffies_up_relative(unsigned long j)
471{
472 return __round_jiffies_up_relative(j, raw_smp_processor_id());
473}
474EXPORT_SYMBOL_GPL(round_jiffies_up_relative);
475
476
477static inline unsigned int timer_get_idx(struct timer_list *timer)
478{
479 return (timer->flags & TIMER_ARRAYMASK) >> TIMER_ARRAYSHIFT;
480}
481
482static inline void timer_set_idx(struct timer_list *timer, unsigned int idx)
483{
484 timer->flags = (timer->flags & ~TIMER_ARRAYMASK) |
485 idx << TIMER_ARRAYSHIFT;
486}
487
488/*
489 * Helper function to calculate the array index for a given expiry
490 * time.
491 */
492static inline unsigned calc_index(unsigned long expires, unsigned lvl,
493 unsigned long *bucket_expiry)
494{
495
496 /*
497 * The timer wheel has to guarantee that a timer does not fire
498 * early. Early expiry can happen due to:
499 * - Timer is armed at the edge of a tick
500 * - Truncation of the expiry time in the outer wheel levels
501 *
502 * Round up with level granularity to prevent this.
503 */
504 expires = (expires + LVL_GRAN(lvl)) >> LVL_SHIFT(lvl);
505 *bucket_expiry = expires << LVL_SHIFT(lvl);
506 return LVL_OFFS(lvl) + (expires & LVL_MASK);
507}
508
509static int calc_wheel_index(unsigned long expires, unsigned long clk,
510 unsigned long *bucket_expiry)
511{
512 unsigned long delta = expires - clk;
513 unsigned int idx;
514
515 if (delta < LVL_START(1)) {
516 idx = calc_index(expires, 0, bucket_expiry);
517 } else if (delta < LVL_START(2)) {
518 idx = calc_index(expires, 1, bucket_expiry);
519 } else if (delta < LVL_START(3)) {
520 idx = calc_index(expires, 2, bucket_expiry);
521 } else if (delta < LVL_START(4)) {
522 idx = calc_index(expires, 3, bucket_expiry);
523 } else if (delta < LVL_START(5)) {
524 idx = calc_index(expires, 4, bucket_expiry);
525 } else if (delta < LVL_START(6)) {
526 idx = calc_index(expires, 5, bucket_expiry);
527 } else if (delta < LVL_START(7)) {
528 idx = calc_index(expires, 6, bucket_expiry);
529 } else if (LVL_DEPTH > 8 && delta < LVL_START(8)) {
530 idx = calc_index(expires, 7, bucket_expiry);
531 } else if ((long) delta < 0) {
532 idx = clk & LVL_MASK;
533 *bucket_expiry = clk;
534 } else {
535 /*
536 * Force expire obscene large timeouts to expire at the
537 * capacity limit of the wheel.
538 */
539 if (delta >= WHEEL_TIMEOUT_CUTOFF)
540 expires = clk + WHEEL_TIMEOUT_MAX;
541
542 idx = calc_index(expires, LVL_DEPTH - 1, bucket_expiry);
543 }
544 return idx;
545}
546
547static void
548trigger_dyntick_cpu(struct timer_base *base, struct timer_list *timer)
549{
550 if (!is_timers_nohz_active())
551 return;
552
553 /*
554 * TODO: This wants some optimizing similar to the code below, but we
555 * will do that when we switch from push to pull for deferrable timers.
556 */
557 if (timer->flags & TIMER_DEFERRABLE) {
558 if (tick_nohz_full_cpu(base->cpu))
559 wake_up_nohz_cpu(base->cpu);
560 return;
561 }
562
563 /*
564 * We might have to IPI the remote CPU if the base is idle and the
565 * timer is not deferrable. If the other CPU is on the way to idle
566 * then it can't set base->is_idle as we hold the base lock:
567 */
568 if (base->is_idle)
569 wake_up_nohz_cpu(base->cpu);
570}
571
572/*
573 * Enqueue the timer into the hash bucket, mark it pending in
574 * the bitmap, store the index in the timer flags then wake up
575 * the target CPU if needed.
576 */
577static void enqueue_timer(struct timer_base *base, struct timer_list *timer,
578 unsigned int idx, unsigned long bucket_expiry)
579{
580
581 hlist_add_head(&timer->entry, base->vectors + idx);
582 __set_bit(idx, base->pending_map);
583 timer_set_idx(timer, idx);
584
585 trace_timer_start(timer, timer->expires, timer->flags);
586
587 /*
588 * Check whether this is the new first expiring timer. The
589 * effective expiry time of the timer is required here
590 * (bucket_expiry) instead of timer->expires.
591 */
592 if (time_before(bucket_expiry, base->next_expiry)) {
593 /*
594 * Set the next expiry time and kick the CPU so it
595 * can reevaluate the wheel:
596 */
597 base->next_expiry = bucket_expiry;
598 base->next_expiry_recalc = false;
599 trigger_dyntick_cpu(base, timer);
600 }
601}
602
603static void internal_add_timer(struct timer_base *base, struct timer_list *timer)
604{
605 unsigned long bucket_expiry;
606 unsigned int idx;
607
608 idx = calc_wheel_index(timer->expires, base->clk, &bucket_expiry);
609 enqueue_timer(base, timer, idx, bucket_expiry);
610}
611
612#ifdef CONFIG_DEBUG_OBJECTS_TIMERS
613
614static struct debug_obj_descr timer_debug_descr;
615
616static void *timer_debug_hint(void *addr)
617{
618 return ((struct timer_list *) addr)->function;
619}
620
621static bool timer_is_static_object(void *addr)
622{
623 struct timer_list *timer = addr;
624
625 return (timer->entry.pprev == NULL &&
626 timer->entry.next == TIMER_ENTRY_STATIC);
627}
628
629/*
630 * fixup_init is called when:
631 * - an active object is initialized
632 */
633static bool timer_fixup_init(void *addr, enum debug_obj_state state)
634{
635 struct timer_list *timer = addr;
636
637 switch (state) {
638 case ODEBUG_STATE_ACTIVE:
639 del_timer_sync(timer);
640 debug_object_init(timer, &timer_debug_descr);
641 return true;
642 default:
643 return false;
644 }
645}
646
647/* Stub timer callback for improperly used timers. */
648static void stub_timer(struct timer_list *unused)
649{
650 WARN_ON(1);
651}
652
653/*
654 * fixup_activate is called when:
655 * - an active object is activated
656 * - an unknown non-static object is activated
657 */
658static bool timer_fixup_activate(void *addr, enum debug_obj_state state)
659{
660 struct timer_list *timer = addr;
661
662 switch (state) {
663 case ODEBUG_STATE_NOTAVAILABLE:
664 timer_setup(timer, stub_timer, 0);
665 return true;
666
667 case ODEBUG_STATE_ACTIVE:
668 WARN_ON(1);
669 fallthrough;
670 default:
671 return false;
672 }
673}
674
675/*
676 * fixup_free is called when:
677 * - an active object is freed
678 */
679static bool timer_fixup_free(void *addr, enum debug_obj_state state)
680{
681 struct timer_list *timer = addr;
682
683 switch (state) {
684 case ODEBUG_STATE_ACTIVE:
685 del_timer_sync(timer);
686 debug_object_free(timer, &timer_debug_descr);
687 return true;
688 default:
689 return false;
690 }
691}
692
693/*
694 * fixup_assert_init is called when:
695 * - an untracked/uninit-ed object is found
696 */
697static bool timer_fixup_assert_init(void *addr, enum debug_obj_state state)
698{
699 struct timer_list *timer = addr;
700
701 switch (state) {
702 case ODEBUG_STATE_NOTAVAILABLE:
703 timer_setup(timer, stub_timer, 0);
704 return true;
705 default:
706 return false;
707 }
708}
709
710static struct debug_obj_descr timer_debug_descr = {
711 .name = "timer_list",
712 .debug_hint = timer_debug_hint,
713 .is_static_object = timer_is_static_object,
714 .fixup_init = timer_fixup_init,
715 .fixup_activate = timer_fixup_activate,
716 .fixup_free = timer_fixup_free,
717 .fixup_assert_init = timer_fixup_assert_init,
718};
719
720static inline void debug_timer_init(struct timer_list *timer)
721{
722 debug_object_init(timer, &timer_debug_descr);
723}
724
725static inline void debug_timer_activate(struct timer_list *timer)
726{
727 debug_object_activate(timer, &timer_debug_descr);
728}
729
730static inline void debug_timer_deactivate(struct timer_list *timer)
731{
732 debug_object_deactivate(timer, &timer_debug_descr);
733}
734
735static inline void debug_timer_free(struct timer_list *timer)
736{
737 debug_object_free(timer, &timer_debug_descr);
738}
739
740static inline void debug_timer_assert_init(struct timer_list *timer)
741{
742 debug_object_assert_init(timer, &timer_debug_descr);
743}
744
745static void do_init_timer(struct timer_list *timer,
746 void (*func)(struct timer_list *),
747 unsigned int flags,
748 const char *name, struct lock_class_key *key);
749
750void init_timer_on_stack_key(struct timer_list *timer,
751 void (*func)(struct timer_list *),
752 unsigned int flags,
753 const char *name, struct lock_class_key *key)
754{
755 debug_object_init_on_stack(timer, &timer_debug_descr);
756 do_init_timer(timer, func, flags, name, key);
757}
758EXPORT_SYMBOL_GPL(init_timer_on_stack_key);
759
760void destroy_timer_on_stack(struct timer_list *timer)
761{
762 debug_object_free(timer, &timer_debug_descr);
763}
764EXPORT_SYMBOL_GPL(destroy_timer_on_stack);
765
766#else
767static inline void debug_timer_init(struct timer_list *timer) { }
768static inline void debug_timer_activate(struct timer_list *timer) { }
769static inline void debug_timer_deactivate(struct timer_list *timer) { }
770static inline void debug_timer_assert_init(struct timer_list *timer) { }
771#endif
772
773static inline void debug_init(struct timer_list *timer)
774{
775 debug_timer_init(timer);
776 trace_timer_init(timer);
777}
778
779static inline void debug_deactivate(struct timer_list *timer)
780{
781 debug_timer_deactivate(timer);
782 trace_timer_cancel(timer);
783}
784
785static inline void debug_assert_init(struct timer_list *timer)
786{
787 debug_timer_assert_init(timer);
788}
789
790static void do_init_timer(struct timer_list *timer,
791 void (*func)(struct timer_list *),
792 unsigned int flags,
793 const char *name, struct lock_class_key *key)
794{
795 timer->entry.pprev = NULL;
796 timer->function = func;
797 timer->flags = flags | raw_smp_processor_id();
798 lockdep_init_map(&timer->lockdep_map, name, key, 0);
799}
800
801/**
802 * init_timer_key - initialize a timer
803 * @timer: the timer to be initialized
804 * @func: timer callback function
805 * @flags: timer flags
806 * @name: name of the timer
807 * @key: lockdep class key of the fake lock used for tracking timer
808 * sync lock dependencies
809 *
810 * init_timer_key() must be done to a timer prior calling *any* of the
811 * other timer functions.
812 */
813void init_timer_key(struct timer_list *timer,
814 void (*func)(struct timer_list *), unsigned int flags,
815 const char *name, struct lock_class_key *key)
816{
817 debug_init(timer);
818 do_init_timer(timer, func, flags, name, key);
819}
820EXPORT_SYMBOL(init_timer_key);
821
822static inline void detach_timer(struct timer_list *timer, bool clear_pending)
823{
824 struct hlist_node *entry = &timer->entry;
825
826 debug_deactivate(timer);
827
828 __hlist_del(entry);
829 if (clear_pending)
830 entry->pprev = NULL;
831 entry->next = LIST_POISON2;
832}
833
834static int detach_if_pending(struct timer_list *timer, struct timer_base *base,
835 bool clear_pending)
836{
837 unsigned idx = timer_get_idx(timer);
838
839 if (!timer_pending(timer))
840 return 0;
841
842 if (hlist_is_singular_node(&timer->entry, base->vectors + idx)) {
843 __clear_bit(idx, base->pending_map);
844 base->next_expiry_recalc = true;
845 }
846
847 detach_timer(timer, clear_pending);
848 return 1;
849}
850
851static inline struct timer_base *get_timer_cpu_base(u32 tflags, u32 cpu)
852{
853 struct timer_base *base = per_cpu_ptr(&timer_bases[BASE_STD], cpu);
854
855 /*
856 * If the timer is deferrable and NO_HZ_COMMON is set then we need
857 * to use the deferrable base.
858 */
859 if (IS_ENABLED(CONFIG_NO_HZ_COMMON) && (tflags & TIMER_DEFERRABLE))
860 base = per_cpu_ptr(&timer_bases[BASE_DEF], cpu);
861 return base;
862}
863
864static inline struct timer_base *get_timer_this_cpu_base(u32 tflags)
865{
866 struct timer_base *base = this_cpu_ptr(&timer_bases[BASE_STD]);
867
868 /*
869 * If the timer is deferrable and NO_HZ_COMMON is set then we need
870 * to use the deferrable base.
871 */
872 if (IS_ENABLED(CONFIG_NO_HZ_COMMON) && (tflags & TIMER_DEFERRABLE))
873 base = this_cpu_ptr(&timer_bases[BASE_DEF]);
874 return base;
875}
876
877static inline struct timer_base *get_timer_base(u32 tflags)
878{
879 return get_timer_cpu_base(tflags, tflags & TIMER_CPUMASK);
880}
881
882static inline struct timer_base *
883get_target_base(struct timer_base *base, unsigned tflags)
884{
885#if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
886 if (static_branch_likely(&timers_migration_enabled) &&
887 !(tflags & TIMER_PINNED))
888 return get_timer_cpu_base(tflags, get_nohz_timer_target());
889#endif
890 return get_timer_this_cpu_base(tflags);
891}
892
893static inline void forward_timer_base(struct timer_base *base)
894{
895 unsigned long jnow = READ_ONCE(jiffies);
896
897 /*
898 * No need to forward if we are close enough below jiffies.
899 * Also while executing timers, base->clk is 1 offset ahead
900 * of jiffies to avoid endless requeuing to current jffies.
901 */
902 if ((long)(jnow - base->clk) < 1)
903 return;
904
905 /*
906 * If the next expiry value is > jiffies, then we fast forward to
907 * jiffies otherwise we forward to the next expiry value.
908 */
909 if (time_after(base->next_expiry, jnow)) {
910 base->clk = jnow;
911 } else {
912 if (WARN_ON_ONCE(time_before(base->next_expiry, base->clk)))
913 return;
914 base->clk = base->next_expiry;
915 }
916}
917
918
919/*
920 * We are using hashed locking: Holding per_cpu(timer_bases[x]).lock means
921 * that all timers which are tied to this base are locked, and the base itself
922 * is locked too.
923 *
924 * So __run_timers/migrate_timers can safely modify all timers which could
925 * be found in the base->vectors array.
926 *
927 * When a timer is migrating then the TIMER_MIGRATING flag is set and we need
928 * to wait until the migration is done.
929 */
930static struct timer_base *lock_timer_base(struct timer_list *timer,
931 unsigned long *flags)
932 __acquires(timer->base->lock)
933{
934 for (;;) {
935 struct timer_base *base;
936 u32 tf;
937
938 /*
939 * We need to use READ_ONCE() here, otherwise the compiler
940 * might re-read @tf between the check for TIMER_MIGRATING
941 * and spin_lock().
942 */
943 tf = READ_ONCE(timer->flags);
944
945 if (!(tf & TIMER_MIGRATING)) {
946 base = get_timer_base(tf);
947 raw_spin_lock_irqsave(&base->lock, *flags);
948 if (timer->flags == tf)
949 return base;
950 raw_spin_unlock_irqrestore(&base->lock, *flags);
951 }
952 cpu_relax();
953 }
954}
955
956#define MOD_TIMER_PENDING_ONLY 0x01
957#define MOD_TIMER_REDUCE 0x02
958#define MOD_TIMER_NOTPENDING 0x04
959
960static inline int
961__mod_timer(struct timer_list *timer, unsigned long expires, unsigned int options)
962{
963 unsigned long clk = 0, flags, bucket_expiry;
964 struct timer_base *base, *new_base;
965 unsigned int idx = UINT_MAX;
966 int ret = 0;
967
968 BUG_ON(!timer->function);
969
970 /*
971 * This is a common optimization triggered by the networking code - if
972 * the timer is re-modified to have the same timeout or ends up in the
973 * same array bucket then just return:
974 */
975 if (!(options & MOD_TIMER_NOTPENDING) && timer_pending(timer)) {
976 /*
977 * The downside of this optimization is that it can result in
978 * larger granularity than you would get from adding a new
979 * timer with this expiry.
980 */
981 long diff = timer->expires - expires;
982
983 if (!diff)
984 return 1;
985 if (options & MOD_TIMER_REDUCE && diff <= 0)
986 return 1;
987
988 /*
989 * We lock timer base and calculate the bucket index right
990 * here. If the timer ends up in the same bucket, then we
991 * just update the expiry time and avoid the whole
992 * dequeue/enqueue dance.
993 */
994 base = lock_timer_base(timer, &flags);
995 forward_timer_base(base);
996
997 if (timer_pending(timer) && (options & MOD_TIMER_REDUCE) &&
998 time_before_eq(timer->expires, expires)) {
999 ret = 1;
1000 goto out_unlock;
1001 }
1002
1003 clk = base->clk;
1004 idx = calc_wheel_index(expires, clk, &bucket_expiry);
1005
1006 /*
1007 * Retrieve and compare the array index of the pending
1008 * timer. If it matches set the expiry to the new value so a
1009 * subsequent call will exit in the expires check above.
1010 */
1011 if (idx == timer_get_idx(timer)) {
1012 if (!(options & MOD_TIMER_REDUCE))
1013 timer->expires = expires;
1014 else if (time_after(timer->expires, expires))
1015 timer->expires = expires;
1016 ret = 1;
1017 goto out_unlock;
1018 }
1019 } else {
1020 base = lock_timer_base(timer, &flags);
1021 forward_timer_base(base);
1022 }
1023
1024 ret = detach_if_pending(timer, base, false);
1025 if (!ret && (options & MOD_TIMER_PENDING_ONLY))
1026 goto out_unlock;
1027
1028 new_base = get_target_base(base, timer->flags);
1029
1030 if (base != new_base) {
1031 /*
1032 * We are trying to schedule the timer on the new base.
1033 * However we can't change timer's base while it is running,
1034 * otherwise del_timer_sync() can't detect that the timer's
1035 * handler yet has not finished. This also guarantees that the
1036 * timer is serialized wrt itself.
1037 */
1038 if (likely(base->running_timer != timer)) {
1039 /* See the comment in lock_timer_base() */
1040 timer->flags |= TIMER_MIGRATING;
1041
1042 raw_spin_unlock(&base->lock);
1043 base = new_base;
1044 raw_spin_lock(&base->lock);
1045 WRITE_ONCE(timer->flags,
1046 (timer->flags & ~TIMER_BASEMASK) | base->cpu);
1047 forward_timer_base(base);
1048 }
1049 }
1050
1051 debug_timer_activate(timer);
1052
1053 timer->expires = expires;
1054 /*
1055 * If 'idx' was calculated above and the base time did not advance
1056 * between calculating 'idx' and possibly switching the base, only
1057 * enqueue_timer() is required. Otherwise we need to (re)calculate
1058 * the wheel index via internal_add_timer().
1059 */
1060 if (idx != UINT_MAX && clk == base->clk)
1061 enqueue_timer(base, timer, idx, bucket_expiry);
1062 else
1063 internal_add_timer(base, timer);
1064
1065out_unlock:
1066 raw_spin_unlock_irqrestore(&base->lock, flags);
1067
1068 return ret;
1069}
1070
1071/**
1072 * mod_timer_pending - modify a pending timer's timeout
1073 * @timer: the pending timer to be modified
1074 * @expires: new timeout in jiffies
1075 *
1076 * mod_timer_pending() is the same for pending timers as mod_timer(),
1077 * but will not re-activate and modify already deleted timers.
1078 *
1079 * It is useful for unserialized use of timers.
1080 */
1081int mod_timer_pending(struct timer_list *timer, unsigned long expires)
1082{
1083 return __mod_timer(timer, expires, MOD_TIMER_PENDING_ONLY);
1084}
1085EXPORT_SYMBOL(mod_timer_pending);
1086
1087/**
1088 * mod_timer - modify a timer's timeout
1089 * @timer: the timer to be modified
1090 * @expires: new timeout in jiffies
1091 *
1092 * mod_timer() is a more efficient way to update the expire field of an
1093 * active timer (if the timer is inactive it will be activated)
1094 *
1095 * mod_timer(timer, expires) is equivalent to:
1096 *
1097 * del_timer(timer); timer->expires = expires; add_timer(timer);
1098 *
1099 * Note that if there are multiple unserialized concurrent users of the
1100 * same timer, then mod_timer() is the only safe way to modify the timeout,
1101 * since add_timer() cannot modify an already running timer.
1102 *
1103 * The function returns whether it has modified a pending timer or not.
1104 * (ie. mod_timer() of an inactive timer returns 0, mod_timer() of an
1105 * active timer returns 1.)
1106 */
1107int mod_timer(struct timer_list *timer, unsigned long expires)
1108{
1109 return __mod_timer(timer, expires, 0);
1110}
1111EXPORT_SYMBOL(mod_timer);
1112
1113/**
1114 * timer_reduce - Modify a timer's timeout if it would reduce the timeout
1115 * @timer: The timer to be modified
1116 * @expires: New timeout in jiffies
1117 *
1118 * timer_reduce() is very similar to mod_timer(), except that it will only
1119 * modify a running timer if that would reduce the expiration time (it will
1120 * start a timer that isn't running).
1121 */
1122int timer_reduce(struct timer_list *timer, unsigned long expires)
1123{
1124 return __mod_timer(timer, expires, MOD_TIMER_REDUCE);
1125}
1126EXPORT_SYMBOL(timer_reduce);
1127
1128/**
1129 * add_timer - start a timer
1130 * @timer: the timer to be added
1131 *
1132 * The kernel will do a ->function(@timer) callback from the
1133 * timer interrupt at the ->expires point in the future. The
1134 * current time is 'jiffies'.
1135 *
1136 * The timer's ->expires, ->function fields must be set prior calling this
1137 * function.
1138 *
1139 * Timers with an ->expires field in the past will be executed in the next
1140 * timer tick.
1141 */
1142void add_timer(struct timer_list *timer)
1143{
1144 BUG_ON(timer_pending(timer));
1145 __mod_timer(timer, timer->expires, MOD_TIMER_NOTPENDING);
1146}
1147EXPORT_SYMBOL(add_timer);
1148
1149/**
1150 * add_timer_on - start a timer on a particular CPU
1151 * @timer: the timer to be added
1152 * @cpu: the CPU to start it on
1153 *
1154 * This is not very scalable on SMP. Double adds are not possible.
1155 */
1156void add_timer_on(struct timer_list *timer, int cpu)
1157{
1158 struct timer_base *new_base, *base;
1159 unsigned long flags;
1160
1161 BUG_ON(timer_pending(timer) || !timer->function);
1162
1163 new_base = get_timer_cpu_base(timer->flags, cpu);
1164
1165 /*
1166 * If @timer was on a different CPU, it should be migrated with the
1167 * old base locked to prevent other operations proceeding with the
1168 * wrong base locked. See lock_timer_base().
1169 */
1170 base = lock_timer_base(timer, &flags);
1171 if (base != new_base) {
1172 timer->flags |= TIMER_MIGRATING;
1173
1174 raw_spin_unlock(&base->lock);
1175 base = new_base;
1176 raw_spin_lock(&base->lock);
1177 WRITE_ONCE(timer->flags,
1178 (timer->flags & ~TIMER_BASEMASK) | cpu);
1179 }
1180 forward_timer_base(base);
1181
1182 debug_timer_activate(timer);
1183 internal_add_timer(base, timer);
1184 raw_spin_unlock_irqrestore(&base->lock, flags);
1185}
1186EXPORT_SYMBOL_GPL(add_timer_on);
1187
1188/**
1189 * del_timer - deactivate a timer.
1190 * @timer: the timer to be deactivated
1191 *
1192 * del_timer() deactivates a timer - this works on both active and inactive
1193 * timers.
1194 *
1195 * The function returns whether it has deactivated a pending timer or not.
1196 * (ie. del_timer() of an inactive timer returns 0, del_timer() of an
1197 * active timer returns 1.)
1198 */
1199int del_timer(struct timer_list *timer)
1200{
1201 struct timer_base *base;
1202 unsigned long flags;
1203 int ret = 0;
1204
1205 debug_assert_init(timer);
1206
1207 if (timer_pending(timer)) {
1208 base = lock_timer_base(timer, &flags);
1209 ret = detach_if_pending(timer, base, true);
1210 raw_spin_unlock_irqrestore(&base->lock, flags);
1211 }
1212
1213 return ret;
1214}
1215EXPORT_SYMBOL(del_timer);
1216
1217/**
1218 * try_to_del_timer_sync - Try to deactivate a timer
1219 * @timer: timer to delete
1220 *
1221 * This function tries to deactivate a timer. Upon successful (ret >= 0)
1222 * exit the timer is not queued and the handler is not running on any CPU.
1223 */
1224int try_to_del_timer_sync(struct timer_list *timer)
1225{
1226 struct timer_base *base;
1227 unsigned long flags;
1228 int ret = -1;
1229
1230 debug_assert_init(timer);
1231
1232 base = lock_timer_base(timer, &flags);
1233
1234 if (base->running_timer != timer)
1235 ret = detach_if_pending(timer, base, true);
1236
1237 raw_spin_unlock_irqrestore(&base->lock, flags);
1238
1239 return ret;
1240}
1241EXPORT_SYMBOL(try_to_del_timer_sync);
1242
1243#ifdef CONFIG_PREEMPT_RT
1244static __init void timer_base_init_expiry_lock(struct timer_base *base)
1245{
1246 spin_lock_init(&base->expiry_lock);
1247}
1248
1249static inline void timer_base_lock_expiry(struct timer_base *base)
1250{
1251 spin_lock(&base->expiry_lock);
1252}
1253
1254static inline void timer_base_unlock_expiry(struct timer_base *base)
1255{
1256 spin_unlock(&base->expiry_lock);
1257}
1258
1259/*
1260 * The counterpart to del_timer_wait_running().
1261 *
1262 * If there is a waiter for base->expiry_lock, then it was waiting for the
1263 * timer callback to finish. Drop expiry_lock and reaquire it. That allows
1264 * the waiter to acquire the lock and make progress.
1265 */
1266static void timer_sync_wait_running(struct timer_base *base)
1267{
1268 if (atomic_read(&base->timer_waiters)) {
1269 spin_unlock(&base->expiry_lock);
1270 spin_lock(&base->expiry_lock);
1271 }
1272}
1273
1274/*
1275 * This function is called on PREEMPT_RT kernels when the fast path
1276 * deletion of a timer failed because the timer callback function was
1277 * running.
1278 *
1279 * This prevents priority inversion, if the softirq thread on a remote CPU
1280 * got preempted, and it prevents a life lock when the task which tries to
1281 * delete a timer preempted the softirq thread running the timer callback
1282 * function.
1283 */
1284static void del_timer_wait_running(struct timer_list *timer)
1285{
1286 u32 tf;
1287
1288 tf = READ_ONCE(timer->flags);
1289 if (!(tf & TIMER_MIGRATING)) {
1290 struct timer_base *base = get_timer_base(tf);
1291
1292 /*
1293 * Mark the base as contended and grab the expiry lock,
1294 * which is held by the softirq across the timer
1295 * callback. Drop the lock immediately so the softirq can
1296 * expire the next timer. In theory the timer could already
1297 * be running again, but that's more than unlikely and just
1298 * causes another wait loop.
1299 */
1300 atomic_inc(&base->timer_waiters);
1301 spin_lock_bh(&base->expiry_lock);
1302 atomic_dec(&base->timer_waiters);
1303 spin_unlock_bh(&base->expiry_lock);
1304 }
1305}
1306#else
1307static inline void timer_base_init_expiry_lock(struct timer_base *base) { }
1308static inline void timer_base_lock_expiry(struct timer_base *base) { }
1309static inline void timer_base_unlock_expiry(struct timer_base *base) { }
1310static inline void timer_sync_wait_running(struct timer_base *base) { }
1311static inline void del_timer_wait_running(struct timer_list *timer) { }
1312#endif
1313
1314#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
1315/**
1316 * del_timer_sync - deactivate a timer and wait for the handler to finish.
1317 * @timer: the timer to be deactivated
1318 *
1319 * This function only differs from del_timer() on SMP: besides deactivating
1320 * the timer it also makes sure the handler has finished executing on other
1321 * CPUs.
1322 *
1323 * Synchronization rules: Callers must prevent restarting of the timer,
1324 * otherwise this function is meaningless. It must not be called from
1325 * interrupt contexts unless the timer is an irqsafe one. The caller must
1326 * not hold locks which would prevent completion of the timer's
1327 * handler. The timer's handler must not call add_timer_on(). Upon exit the
1328 * timer is not queued and the handler is not running on any CPU.
1329 *
1330 * Note: For !irqsafe timers, you must not hold locks that are held in
1331 * interrupt context while calling this function. Even if the lock has
1332 * nothing to do with the timer in question. Here's why::
1333 *
1334 * CPU0 CPU1
1335 * ---- ----
1336 * <SOFTIRQ>
1337 * call_timer_fn();
1338 * base->running_timer = mytimer;
1339 * spin_lock_irq(somelock);
1340 * <IRQ>
1341 * spin_lock(somelock);
1342 * del_timer_sync(mytimer);
1343 * while (base->running_timer == mytimer);
1344 *
1345 * Now del_timer_sync() will never return and never release somelock.
1346 * The interrupt on the other CPU is waiting to grab somelock but
1347 * it has interrupted the softirq that CPU0 is waiting to finish.
1348 *
1349 * The function returns whether it has deactivated a pending timer or not.
1350 */
1351int del_timer_sync(struct timer_list *timer)
1352{
1353 int ret;
1354
1355#ifdef CONFIG_LOCKDEP
1356 unsigned long flags;
1357
1358 /*
1359 * If lockdep gives a backtrace here, please reference
1360 * the synchronization rules above.
1361 */
1362 local_irq_save(flags);
1363 lock_map_acquire(&timer->lockdep_map);
1364 lock_map_release(&timer->lockdep_map);
1365 local_irq_restore(flags);
1366#endif
1367 /*
1368 * don't use it in hardirq context, because it
1369 * could lead to deadlock.
1370 */
1371 WARN_ON(in_irq() && !(timer->flags & TIMER_IRQSAFE));
1372
1373 do {
1374 ret = try_to_del_timer_sync(timer);
1375
1376 if (unlikely(ret < 0)) {
1377 del_timer_wait_running(timer);
1378 cpu_relax();
1379 }
1380 } while (ret < 0);
1381
1382 return ret;
1383}
1384EXPORT_SYMBOL(del_timer_sync);
1385#endif
1386
1387static void call_timer_fn(struct timer_list *timer,
1388 void (*fn)(struct timer_list *),
1389 unsigned long baseclk)
1390{
1391 int count = preempt_count();
1392
1393#ifdef CONFIG_LOCKDEP
1394 /*
1395 * It is permissible to free the timer from inside the
1396 * function that is called from it, this we need to take into
1397 * account for lockdep too. To avoid bogus "held lock freed"
1398 * warnings as well as problems when looking into
1399 * timer->lockdep_map, make a copy and use that here.
1400 */
1401 struct lockdep_map lockdep_map;
1402
1403 lockdep_copy_map(&lockdep_map, &timer->lockdep_map);
1404#endif
1405 /*
1406 * Couple the lock chain with the lock chain at
1407 * del_timer_sync() by acquiring the lock_map around the fn()
1408 * call here and in del_timer_sync().
1409 */
1410 lock_map_acquire(&lockdep_map);
1411
1412 trace_timer_expire_entry(timer, baseclk);
1413 fn(timer);
1414 trace_timer_expire_exit(timer);
1415
1416 lock_map_release(&lockdep_map);
1417
1418 if (count != preempt_count()) {
1419 WARN_ONCE(1, "timer: %pS preempt leak: %08x -> %08x\n",
1420 fn, count, preempt_count());
1421 /*
1422 * Restore the preempt count. That gives us a decent
1423 * chance to survive and extract information. If the
1424 * callback kept a lock held, bad luck, but not worse
1425 * than the BUG() we had.
1426 */
1427 preempt_count_set(count);
1428 }
1429}
1430
1431static void expire_timers(struct timer_base *base, struct hlist_head *head)
1432{
1433 /*
1434 * This value is required only for tracing. base->clk was
1435 * incremented directly before expire_timers was called. But expiry
1436 * is related to the old base->clk value.
1437 */
1438 unsigned long baseclk = base->clk - 1;
1439
1440 while (!hlist_empty(head)) {
1441 struct timer_list *timer;
1442 void (*fn)(struct timer_list *);
1443
1444 timer = hlist_entry(head->first, struct timer_list, entry);
1445
1446 base->running_timer = timer;
1447 detach_timer(timer, true);
1448
1449 fn = timer->function;
1450
1451 if (timer->flags & TIMER_IRQSAFE) {
1452 raw_spin_unlock(&base->lock);
1453 call_timer_fn(timer, fn, baseclk);
1454 base->running_timer = NULL;
1455 raw_spin_lock(&base->lock);
1456 } else {
1457 raw_spin_unlock_irq(&base->lock);
1458 call_timer_fn(timer, fn, baseclk);
1459 base->running_timer = NULL;
1460 timer_sync_wait_running(base);
1461 raw_spin_lock_irq(&base->lock);
1462 }
1463 }
1464}
1465
1466static int collect_expired_timers(struct timer_base *base,
1467 struct hlist_head *heads)
1468{
1469 unsigned long clk = base->clk = base->next_expiry;
1470 struct hlist_head *vec;
1471 int i, levels = 0;
1472 unsigned int idx;
1473
1474 for (i = 0; i < LVL_DEPTH; i++) {
1475 idx = (clk & LVL_MASK) + i * LVL_SIZE;
1476
1477 if (__test_and_clear_bit(idx, base->pending_map)) {
1478 vec = base->vectors + idx;
1479 hlist_move_list(vec, heads++);
1480 levels++;
1481 }
1482 /* Is it time to look at the next level? */
1483 if (clk & LVL_CLK_MASK)
1484 break;
1485 /* Shift clock for the next level granularity */
1486 clk >>= LVL_CLK_SHIFT;
1487 }
1488 return levels;
1489}
1490
1491/*
1492 * Find the next pending bucket of a level. Search from level start (@offset)
1493 * + @clk upwards and if nothing there, search from start of the level
1494 * (@offset) up to @offset + clk.
1495 */
1496static int next_pending_bucket(struct timer_base *base, unsigned offset,
1497 unsigned clk)
1498{
1499 unsigned pos, start = offset + clk;
1500 unsigned end = offset + LVL_SIZE;
1501
1502 pos = find_next_bit(base->pending_map, end, start);
1503 if (pos < end)
1504 return pos - start;
1505
1506 pos = find_next_bit(base->pending_map, start, offset);
1507 return pos < start ? pos + LVL_SIZE - start : -1;
1508}
1509
1510/*
1511 * Search the first expiring timer in the various clock levels. Caller must
1512 * hold base->lock.
1513 */
1514static unsigned long __next_timer_interrupt(struct timer_base *base)
1515{
1516 unsigned long clk, next, adj;
1517 unsigned lvl, offset = 0;
1518
1519 next = base->clk + NEXT_TIMER_MAX_DELTA;
1520 clk = base->clk;
1521 for (lvl = 0; lvl < LVL_DEPTH; lvl++, offset += LVL_SIZE) {
1522 int pos = next_pending_bucket(base, offset, clk & LVL_MASK);
1523 unsigned long lvl_clk = clk & LVL_CLK_MASK;
1524
1525 if (pos >= 0) {
1526 unsigned long tmp = clk + (unsigned long) pos;
1527
1528 tmp <<= LVL_SHIFT(lvl);
1529 if (time_before(tmp, next))
1530 next = tmp;
1531
1532 /*
1533 * If the next expiration happens before we reach
1534 * the next level, no need to check further.
1535 */
1536 if (pos <= ((LVL_CLK_DIV - lvl_clk) & LVL_CLK_MASK))
1537 break;
1538 }
1539 /*
1540 * Clock for the next level. If the current level clock lower
1541 * bits are zero, we look at the next level as is. If not we
1542 * need to advance it by one because that's going to be the
1543 * next expiring bucket in that level. base->clk is the next
1544 * expiring jiffie. So in case of:
1545 *
1546 * LVL5 LVL4 LVL3 LVL2 LVL1 LVL0
1547 * 0 0 0 0 0 0
1548 *
1549 * we have to look at all levels @index 0. With
1550 *
1551 * LVL5 LVL4 LVL3 LVL2 LVL1 LVL0
1552 * 0 0 0 0 0 2
1553 *
1554 * LVL0 has the next expiring bucket @index 2. The upper
1555 * levels have the next expiring bucket @index 1.
1556 *
1557 * In case that the propagation wraps the next level the same
1558 * rules apply:
1559 *
1560 * LVL5 LVL4 LVL3 LVL2 LVL1 LVL0
1561 * 0 0 0 0 F 2
1562 *
1563 * So after looking at LVL0 we get:
1564 *
1565 * LVL5 LVL4 LVL3 LVL2 LVL1
1566 * 0 0 0 1 0
1567 *
1568 * So no propagation from LVL1 to LVL2 because that happened
1569 * with the add already, but then we need to propagate further
1570 * from LVL2 to LVL3.
1571 *
1572 * So the simple check whether the lower bits of the current
1573 * level are 0 or not is sufficient for all cases.
1574 */
1575 adj = lvl_clk ? 1 : 0;
1576 clk >>= LVL_CLK_SHIFT;
1577 clk += adj;
1578 }
1579
1580 base->next_expiry_recalc = false;
1581
1582 return next;
1583}
1584
1585#ifdef CONFIG_NO_HZ_COMMON
1586/*
1587 * Check, if the next hrtimer event is before the next timer wheel
1588 * event:
1589 */
1590static u64 cmp_next_hrtimer_event(u64 basem, u64 expires)
1591{
1592 u64 nextevt = hrtimer_get_next_event();
1593
1594 /*
1595 * If high resolution timers are enabled
1596 * hrtimer_get_next_event() returns KTIME_MAX.
1597 */
1598 if (expires <= nextevt)
1599 return expires;
1600
1601 /*
1602 * If the next timer is already expired, return the tick base
1603 * time so the tick is fired immediately.
1604 */
1605 if (nextevt <= basem)
1606 return basem;
1607
1608 /*
1609 * Round up to the next jiffie. High resolution timers are
1610 * off, so the hrtimers are expired in the tick and we need to
1611 * make sure that this tick really expires the timer to avoid
1612 * a ping pong of the nohz stop code.
1613 *
1614 * Use DIV_ROUND_UP_ULL to prevent gcc calling __divdi3
1615 */
1616 return DIV_ROUND_UP_ULL(nextevt, TICK_NSEC) * TICK_NSEC;
1617}
1618
1619/**
1620 * get_next_timer_interrupt - return the time (clock mono) of the next timer
1621 * @basej: base time jiffies
1622 * @basem: base time clock monotonic
1623 *
1624 * Returns the tick aligned clock monotonic time of the next pending
1625 * timer or KTIME_MAX if no timer is pending.
1626 */
1627u64 get_next_timer_interrupt(unsigned long basej, u64 basem)
1628{
1629 struct timer_base *base = this_cpu_ptr(&timer_bases[BASE_STD]);
1630 u64 expires = KTIME_MAX;
1631 unsigned long nextevt;
1632 bool is_max_delta;
1633
1634 /*
1635 * Pretend that there is no timer pending if the cpu is offline.
1636 * Possible pending timers will be migrated later to an active cpu.
1637 */
1638 if (cpu_is_offline(smp_processor_id()))
1639 return expires;
1640
1641 raw_spin_lock(&base->lock);
1642 if (base->next_expiry_recalc)
1643 base->next_expiry = __next_timer_interrupt(base);
1644 nextevt = base->next_expiry;
1645 is_max_delta = (nextevt == base->clk + NEXT_TIMER_MAX_DELTA);
1646
1647 /*
1648 * We have a fresh next event. Check whether we can forward the
1649 * base. We can only do that when @basej is past base->clk
1650 * otherwise we might rewind base->clk.
1651 */
1652 if (time_after(basej, base->clk)) {
1653 if (time_after(nextevt, basej))
1654 base->clk = basej;
1655 else if (time_after(nextevt, base->clk))
1656 base->clk = nextevt;
1657 }
1658
1659 if (time_before_eq(nextevt, basej)) {
1660 expires = basem;
1661 base->is_idle = false;
1662 } else {
1663 if (!is_max_delta)
1664 expires = basem + (u64)(nextevt - basej) * TICK_NSEC;
1665 /*
1666 * If we expect to sleep more than a tick, mark the base idle.
1667 * Also the tick is stopped so any added timer must forward
1668 * the base clk itself to keep granularity small. This idle
1669 * logic is only maintained for the BASE_STD base, deferrable
1670 * timers may still see large granularity skew (by design).
1671 */
1672 if ((expires - basem) > TICK_NSEC)
1673 base->is_idle = true;
1674 }
1675 raw_spin_unlock(&base->lock);
1676
1677 return cmp_next_hrtimer_event(basem, expires);
1678}
1679
1680/**
1681 * timer_clear_idle - Clear the idle state of the timer base
1682 *
1683 * Called with interrupts disabled
1684 */
1685void timer_clear_idle(void)
1686{
1687 struct timer_base *base = this_cpu_ptr(&timer_bases[BASE_STD]);
1688
1689 /*
1690 * We do this unlocked. The worst outcome is a remote enqueue sending
1691 * a pointless IPI, but taking the lock would just make the window for
1692 * sending the IPI a few instructions smaller for the cost of taking
1693 * the lock in the exit from idle path.
1694 */
1695 base->is_idle = false;
1696}
1697#endif
1698
1699/*
1700 * Called from the timer interrupt handler to charge one tick to the current
1701 * process. user_tick is 1 if the tick is user time, 0 for system.
1702 */
1703void update_process_times(int user_tick)
1704{
1705 struct task_struct *p = current;
1706
1707 /* Note: this timer irq context must be accounted for as well. */
1708 account_process_tick(p, user_tick);
1709 run_local_timers();
1710 rcu_sched_clock_irq(user_tick);
1711#ifdef CONFIG_IRQ_WORK
1712 if (in_irq())
1713 irq_work_tick();
1714#endif
1715 scheduler_tick();
1716 if (IS_ENABLED(CONFIG_POSIX_TIMERS))
1717 run_posix_cpu_timers();
1718
1719 /* The current CPU might make use of net randoms without receiving IRQs
1720 * to renew them often enough. Let's update the net_rand_state from a
1721 * non-constant value that's not affine to the number of calls to make
1722 * sure it's updated when there's some activity (we don't care in idle).
1723 */
1724 this_cpu_add(net_rand_state.s1, rol32(jiffies, 24) + user_tick);
1725}
1726
1727/**
1728 * __run_timers - run all expired timers (if any) on this CPU.
1729 * @base: the timer vector to be processed.
1730 */
1731static inline void __run_timers(struct timer_base *base)
1732{
1733 struct hlist_head heads[LVL_DEPTH];
1734 int levels;
1735
1736 if (time_before(jiffies, base->next_expiry))
1737 return;
1738
1739 timer_base_lock_expiry(base);
1740 raw_spin_lock_irq(&base->lock);
1741
1742 while (time_after_eq(jiffies, base->clk) &&
1743 time_after_eq(jiffies, base->next_expiry)) {
1744 levels = collect_expired_timers(base, heads);
1745 /*
1746 * The only possible reason for not finding any expired
1747 * timer at this clk is that all matching timers have been
1748 * dequeued.
1749 */
1750 WARN_ON_ONCE(!levels && !base->next_expiry_recalc);
1751 base->clk++;
1752 base->next_expiry = __next_timer_interrupt(base);
1753
1754 while (levels--)
1755 expire_timers(base, heads + levels);
1756 }
1757 raw_spin_unlock_irq(&base->lock);
1758 timer_base_unlock_expiry(base);
1759}
1760
1761/*
1762 * This function runs timers and the timer-tq in bottom half context.
1763 */
1764static __latent_entropy void run_timer_softirq(struct softirq_action *h)
1765{
1766 struct timer_base *base = this_cpu_ptr(&timer_bases[BASE_STD]);
1767
1768 __run_timers(base);
1769 if (IS_ENABLED(CONFIG_NO_HZ_COMMON))
1770 __run_timers(this_cpu_ptr(&timer_bases[BASE_DEF]));
1771}
1772
1773/*
1774 * Called by the local, per-CPU timer interrupt on SMP.
1775 */
1776void run_local_timers(void)
1777{
1778 struct timer_base *base = this_cpu_ptr(&timer_bases[BASE_STD]);
1779
1780 hrtimer_run_queues();
1781 /* Raise the softirq only if required. */
1782 if (time_before(jiffies, base->next_expiry)) {
1783 if (!IS_ENABLED(CONFIG_NO_HZ_COMMON))
1784 return;
1785 /* CPU is awake, so check the deferrable base. */
1786 base++;
1787 if (time_before(jiffies, base->next_expiry))
1788 return;
1789 }
1790 raise_softirq(TIMER_SOFTIRQ);
1791}
1792
1793/*
1794 * Since schedule_timeout()'s timer is defined on the stack, it must store
1795 * the target task on the stack as well.
1796 */
1797struct process_timer {
1798 struct timer_list timer;
1799 struct task_struct *task;
1800};
1801
1802static void process_timeout(struct timer_list *t)
1803{
1804 struct process_timer *timeout = from_timer(timeout, t, timer);
1805
1806 wake_up_process(timeout->task);
1807}
1808
1809/**
1810 * schedule_timeout - sleep until timeout
1811 * @timeout: timeout value in jiffies
1812 *
1813 * Make the current task sleep until @timeout jiffies have elapsed.
1814 * The function behavior depends on the current task state
1815 * (see also set_current_state() description):
1816 *
1817 * %TASK_RUNNING - the scheduler is called, but the task does not sleep
1818 * at all. That happens because sched_submit_work() does nothing for
1819 * tasks in %TASK_RUNNING state.
1820 *
1821 * %TASK_UNINTERRUPTIBLE - at least @timeout jiffies are guaranteed to
1822 * pass before the routine returns unless the current task is explicitly
1823 * woken up, (e.g. by wake_up_process()).
1824 *
1825 * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
1826 * delivered to the current task or the current task is explicitly woken
1827 * up.
1828 *
1829 * The current task state is guaranteed to be %TASK_RUNNING when this
1830 * routine returns.
1831 *
1832 * Specifying a @timeout value of %MAX_SCHEDULE_TIMEOUT will schedule
1833 * the CPU away without a bound on the timeout. In this case the return
1834 * value will be %MAX_SCHEDULE_TIMEOUT.
1835 *
1836 * Returns 0 when the timer has expired otherwise the remaining time in
1837 * jiffies will be returned. In all cases the return value is guaranteed
1838 * to be non-negative.
1839 */
1840signed long __sched schedule_timeout(signed long timeout)
1841{
1842 struct process_timer timer;
1843 unsigned long expire;
1844
1845 switch (timeout)
1846 {
1847 case MAX_SCHEDULE_TIMEOUT:
1848 /*
1849 * These two special cases are useful to be comfortable
1850 * in the caller. Nothing more. We could take
1851 * MAX_SCHEDULE_TIMEOUT from one of the negative value
1852 * but I' d like to return a valid offset (>=0) to allow
1853 * the caller to do everything it want with the retval.
1854 */
1855 schedule();
1856 goto out;
1857 default:
1858 /*
1859 * Another bit of PARANOID. Note that the retval will be
1860 * 0 since no piece of kernel is supposed to do a check
1861 * for a negative retval of schedule_timeout() (since it
1862 * should never happens anyway). You just have the printk()
1863 * that will tell you if something is gone wrong and where.
1864 */
1865 if (timeout < 0) {
1866 printk(KERN_ERR "schedule_timeout: wrong timeout "
1867 "value %lx\n", timeout);
1868 dump_stack();
1869 current->state = TASK_RUNNING;
1870 goto out;
1871 }
1872 }
1873
1874 expire = timeout + jiffies;
1875
1876 timer.task = current;
1877 timer_setup_on_stack(&timer.timer, process_timeout, 0);
1878 __mod_timer(&timer.timer, expire, MOD_TIMER_NOTPENDING);
1879 schedule();
1880 del_singleshot_timer_sync(&timer.timer);
1881
1882 /* Remove the timer from the object tracker */
1883 destroy_timer_on_stack(&timer.timer);
1884
1885 timeout = expire - jiffies;
1886
1887 out:
1888 return timeout < 0 ? 0 : timeout;
1889}
1890EXPORT_SYMBOL(schedule_timeout);
1891
1892/*
1893 * We can use __set_current_state() here because schedule_timeout() calls
1894 * schedule() unconditionally.
1895 */
1896signed long __sched schedule_timeout_interruptible(signed long timeout)
1897{
1898 __set_current_state(TASK_INTERRUPTIBLE);
1899 return schedule_timeout(timeout);
1900}
1901EXPORT_SYMBOL(schedule_timeout_interruptible);
1902
1903signed long __sched schedule_timeout_killable(signed long timeout)
1904{
1905 __set_current_state(TASK_KILLABLE);
1906 return schedule_timeout(timeout);
1907}
1908EXPORT_SYMBOL(schedule_timeout_killable);
1909
1910signed long __sched schedule_timeout_uninterruptible(signed long timeout)
1911{
1912 __set_current_state(TASK_UNINTERRUPTIBLE);
1913 return schedule_timeout(timeout);
1914}
1915EXPORT_SYMBOL(schedule_timeout_uninterruptible);
1916
1917/*
1918 * Like schedule_timeout_uninterruptible(), except this task will not contribute
1919 * to load average.
1920 */
1921signed long __sched schedule_timeout_idle(signed long timeout)
1922{
1923 __set_current_state(TASK_IDLE);
1924 return schedule_timeout(timeout);
1925}
1926EXPORT_SYMBOL(schedule_timeout_idle);
1927
1928#ifdef CONFIG_HOTPLUG_CPU
1929static void migrate_timer_list(struct timer_base *new_base, struct hlist_head *head)
1930{
1931 struct timer_list *timer;
1932 int cpu = new_base->cpu;
1933
1934 while (!hlist_empty(head)) {
1935 timer = hlist_entry(head->first, struct timer_list, entry);
1936 detach_timer(timer, false);
1937 timer->flags = (timer->flags & ~TIMER_BASEMASK) | cpu;
1938 internal_add_timer(new_base, timer);
1939 }
1940}
1941
1942int timers_prepare_cpu(unsigned int cpu)
1943{
1944 struct timer_base *base;
1945 int b;
1946
1947 for (b = 0; b < NR_BASES; b++) {
1948 base = per_cpu_ptr(&timer_bases[b], cpu);
1949 base->clk = jiffies;
1950 base->next_expiry = base->clk + NEXT_TIMER_MAX_DELTA;
1951 base->is_idle = false;
1952 }
1953 return 0;
1954}
1955
1956int timers_dead_cpu(unsigned int cpu)
1957{
1958 struct timer_base *old_base;
1959 struct timer_base *new_base;
1960 int b, i;
1961
1962 BUG_ON(cpu_online(cpu));
1963
1964 for (b = 0; b < NR_BASES; b++) {
1965 old_base = per_cpu_ptr(&timer_bases[b], cpu);
1966 new_base = get_cpu_ptr(&timer_bases[b]);
1967 /*
1968 * The caller is globally serialized and nobody else
1969 * takes two locks at once, deadlock is not possible.
1970 */
1971 raw_spin_lock_irq(&new_base->lock);
1972 raw_spin_lock_nested(&old_base->lock, SINGLE_DEPTH_NESTING);
1973
1974 /*
1975 * The current CPUs base clock might be stale. Update it
1976 * before moving the timers over.
1977 */
1978 forward_timer_base(new_base);
1979
1980 BUG_ON(old_base->running_timer);
1981
1982 for (i = 0; i < WHEEL_SIZE; i++)
1983 migrate_timer_list(new_base, old_base->vectors + i);
1984
1985 raw_spin_unlock(&old_base->lock);
1986 raw_spin_unlock_irq(&new_base->lock);
1987 put_cpu_ptr(&timer_bases);
1988 }
1989 return 0;
1990}
1991
1992#endif /* CONFIG_HOTPLUG_CPU */
1993
1994static void __init init_timer_cpu(int cpu)
1995{
1996 struct timer_base *base;
1997 int i;
1998
1999 for (i = 0; i < NR_BASES; i++) {
2000 base = per_cpu_ptr(&timer_bases[i], cpu);
2001 base->cpu = cpu;
2002 raw_spin_lock_init(&base->lock);
2003 base->clk = jiffies;
2004 base->next_expiry = base->clk + NEXT_TIMER_MAX_DELTA;
2005 timer_base_init_expiry_lock(base);
2006 }
2007}
2008
2009static void __init init_timer_cpus(void)
2010{
2011 int cpu;
2012
2013 for_each_possible_cpu(cpu)
2014 init_timer_cpu(cpu);
2015}
2016
2017void __init init_timers(void)
2018{
2019 init_timer_cpus();
2020 posix_cputimers_init_work();
2021 open_softirq(TIMER_SOFTIRQ, run_timer_softirq);
2022}
2023
2024/**
2025 * msleep - sleep safely even with waitqueue interruptions
2026 * @msecs: Time in milliseconds to sleep for
2027 */
2028void msleep(unsigned int msecs)
2029{
2030 unsigned long timeout = msecs_to_jiffies(msecs) + 1;
2031
2032 while (timeout)
2033 timeout = schedule_timeout_uninterruptible(timeout);
2034}
2035
2036EXPORT_SYMBOL(msleep);
2037
2038/**
2039 * msleep_interruptible - sleep waiting for signals
2040 * @msecs: Time in milliseconds to sleep for
2041 */
2042unsigned long msleep_interruptible(unsigned int msecs)
2043{
2044 unsigned long timeout = msecs_to_jiffies(msecs) + 1;
2045
2046 while (timeout && !signal_pending(current))
2047 timeout = schedule_timeout_interruptible(timeout);
2048 return jiffies_to_msecs(timeout);
2049}
2050
2051EXPORT_SYMBOL(msleep_interruptible);
2052
2053/**
2054 * usleep_range - Sleep for an approximate time
2055 * @min: Minimum time in usecs to sleep
2056 * @max: Maximum time in usecs to sleep
2057 *
2058 * In non-atomic context where the exact wakeup time is flexible, use
2059 * usleep_range() instead of udelay(). The sleep improves responsiveness
2060 * by avoiding the CPU-hogging busy-wait of udelay(), and the range reduces
2061 * power usage by allowing hrtimers to take advantage of an already-
2062 * scheduled interrupt instead of scheduling a new one just for this sleep.
2063 */
2064void __sched usleep_range(unsigned long min, unsigned long max)
2065{
2066 ktime_t exp = ktime_add_us(ktime_get(), min);
2067 u64 delta = (u64)(max - min) * NSEC_PER_USEC;
2068
2069 for (;;) {
2070 __set_current_state(TASK_UNINTERRUPTIBLE);
2071 /* Do not return before the requested sleep time has elapsed */
2072 if (!schedule_hrtimeout_range(&exp, delta, HRTIMER_MODE_ABS))
2073 break;
2074 }
2075}
2076EXPORT_SYMBOL(usleep_range);