blob: dd4b7b492c9b2c743c3431ff077160f7df5459a4 [file] [log] [blame]
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001/*
2 * linux/kernel/time/tick-sched.c
3 *
4 * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
5 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
6 * Copyright(C) 2006-2007 Timesys Corp., Thomas Gleixner
7 *
8 * No idle tick implementation for low and high resolution timers
9 *
10 * Started by: Thomas Gleixner and Ingo Molnar
11 *
Pavel Machekb10db7f2008-01-30 13:30:00 +010012 * Distribute under GPLv2.
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -080013 */
14#include <linux/cpu.h>
15#include <linux/err.h>
16#include <linux/hrtimer.h>
17#include <linux/interrupt.h>
18#include <linux/kernel_stat.h>
19#include <linux/percpu.h>
Ingo Molnar38b8d202017-02-08 18:51:31 +010020#include <linux/nmi.h>
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -080021#include <linux/profile.h>
Ingo Molnar3f07c012017-02-08 18:51:30 +010022#include <linux/sched/signal.h>
Ingo Molnare6017572017-02-01 16:36:40 +010023#include <linux/sched/clock.h>
Ingo Molnar03441a32017-02-08 18:51:35 +010024#include <linux/sched/stat.h>
Ingo Molnar370c9132017-02-08 18:51:35 +010025#include <linux/sched/nohz.h>
venkatesh.pallipadi@intel.com8083e4a2008-08-04 11:59:11 -070026#include <linux/module.h>
Frederic Weisbecker00b42952012-11-07 21:03:07 +010027#include <linux/irq_work.h>
Frederic Weisbecker9014c452013-04-20 15:43:57 +020028#include <linux/posix-timers.h>
Frederic Weisbecker2e709332013-07-10 00:55:25 +020029#include <linux/context_tracking.h>
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -080030
David S. Miller9e203bc2007-02-24 22:10:13 -080031#include <asm/irq_regs.h>
32
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -080033#include "tick-internal.h"
34
Frederic Weisbeckercb41a292013-04-20 17:35:50 +020035#include <trace/events/timer.h>
36
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -080037/*
Ingo Molnar0de76112016-07-01 12:42:35 +020038 * Per-CPU nohz control structure
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -080039 */
Thomas Gleixnerc1797ba2015-03-25 13:07:37 +010040static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -080041
Ingo Molnar289f4802007-02-16 01:28:15 -080042struct tick_sched *tick_get_tick_sched(int cpu)
43{
44 return &per_cpu(tick_cpu_sched, cpu);
45}
46
Arnd Bergmann7809998a2016-01-25 16:41:49 +010047#if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS)
48/*
49 * The time, when the last jiffy update happened. Protected by jiffies_lock.
50 */
51static ktime_t last_jiffies_update;
52
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -080053/*
54 * Must be called with interrupts disabled !
55 */
56static void tick_do_update_jiffies64(ktime_t now)
57{
58 unsigned long ticks = 0;
59 ktime_t delta;
60
Ingo Molnar7a14ce12008-05-12 15:43:53 +020061 /*
John Stultzd6ad4182012-02-28 16:50:11 -080062 * Do a quick check without holding jiffies_lock:
Ingo Molnar7a14ce12008-05-12 15:43:53 +020063 */
64 delta = ktime_sub(now, last_jiffies_update);
Thomas Gleixner2456e852016-12-25 11:38:40 +010065 if (delta < tick_period)
Ingo Molnar7a14ce12008-05-12 15:43:53 +020066 return;
67
Wei Jiangang6168f8e2016-06-29 12:51:50 +080068 /* Reevaluate with jiffies_lock held */
John Stultzd6ad4182012-02-28 16:50:11 -080069 write_seqlock(&jiffies_lock);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -080070
71 delta = ktime_sub(now, last_jiffies_update);
Thomas Gleixner2456e852016-12-25 11:38:40 +010072 if (delta >= tick_period) {
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -080073
74 delta = ktime_sub(delta, tick_period);
75 last_jiffies_update = ktime_add(last_jiffies_update,
76 tick_period);
77
78 /* Slow path for long timeouts */
Thomas Gleixner2456e852016-12-25 11:38:40 +010079 if (unlikely(delta >= tick_period)) {
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -080080 s64 incr = ktime_to_ns(tick_period);
81
82 ticks = ktime_divns(delta, incr);
83
84 last_jiffies_update = ktime_add_ns(last_jiffies_update,
85 incr * ticks);
86 }
87 do_timer(++ticks);
Thomas Gleixner49d670f2008-09-22 18:56:01 +020088
89 /* Keep the tick_next_period variable up to date */
90 tick_next_period = ktime_add(last_jiffies_update, tick_period);
Viresh Kumar03e6bdc2014-04-15 10:54:40 +053091 } else {
92 write_sequnlock(&jiffies_lock);
93 return;
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -080094 }
John Stultzd6ad4182012-02-28 16:50:11 -080095 write_sequnlock(&jiffies_lock);
John Stultz47a1b7962013-12-12 13:10:55 -080096 update_wall_time();
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -080097}
98
99/*
100 * Initialize and return retrieve the jiffies update.
101 */
102static ktime_t tick_init_jiffy_update(void)
103{
104 ktime_t period;
105
John Stultzd6ad4182012-02-28 16:50:11 -0800106 write_seqlock(&jiffies_lock);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800107 /* Did we start the jiffies update yet ? */
Thomas Gleixner2456e852016-12-25 11:38:40 +0100108 if (last_jiffies_update == 0)
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800109 last_jiffies_update = tick_next_period;
110 period = last_jiffies_update;
John Stultzd6ad4182012-02-28 16:50:11 -0800111 write_sequnlock(&jiffies_lock);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800112 return period;
113}
114
Frederic Weisbecker5bb96222012-10-15 02:03:27 +0200115
116static void tick_sched_do_timer(ktime_t now)
117{
118 int cpu = smp_processor_id();
119
Frederic Weisbecker3451d022011-08-10 23:21:01 +0200120#ifdef CONFIG_NO_HZ_COMMON
Frederic Weisbecker5bb96222012-10-15 02:03:27 +0200121 /*
122 * Check if the do_timer duty was dropped. We don't care about
Ingo Molnar0de76112016-07-01 12:42:35 +0200123 * concurrency: This happens only when the CPU in charge went
124 * into a long sleep. If two CPUs happen to assign themselves to
Frederic Weisbecker5bb96222012-10-15 02:03:27 +0200125 * this duty, then the jiffies update is still serialized by
Thomas Gleixner9c3f9e22012-11-21 20:31:52 +0100126 * jiffies_lock.
Frederic Weisbecker5bb96222012-10-15 02:03:27 +0200127 */
Frederic Weisbeckera382bf92012-12-18 18:24:35 +0100128 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)
Frederic Weisbeckerc5bfece2013-04-12 16:45:34 +0200129 && !tick_nohz_full_cpu(cpu))
Frederic Weisbecker5bb96222012-10-15 02:03:27 +0200130 tick_do_timer_cpu = cpu;
131#endif
132
133 /* Check, if the jiffies need an update */
134 if (tick_do_timer_cpu == cpu)
135 tick_do_update_jiffies64(now);
136}
137
Frederic Weisbecker9e8f5592012-10-15 02:43:03 +0200138static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
139{
Frederic Weisbecker3451d022011-08-10 23:21:01 +0200140#ifdef CONFIG_NO_HZ_COMMON
Frederic Weisbecker9e8f5592012-10-15 02:43:03 +0200141 /*
142 * When we are idle and the tick is stopped, we have to touch
143 * the watchdog as we might not schedule for a really long
144 * time. This happens on complete idle SMP systems while
145 * waiting on the login prompt. We also increment the "start of
146 * idle" jiffy stamp so the idle accounting adjustment we do
147 * when we go busy again does not account too much ticks.
148 */
149 if (ts->tick_stopped) {
Tejun Heo03e0d462015-12-08 11:28:04 -0500150 touch_softlockup_watchdog_sched();
Frederic Weisbecker9e8f5592012-10-15 02:43:03 +0200151 if (is_idle_task(current))
152 ts->idle_jiffies++;
Frederic Weisbecker411fe242017-04-21 16:00:54 +0200153 /*
154 * In case the current tick fired too early past its expected
155 * expiration, make sure we don't bypass the next clock reprogramming
156 * to the same deadline.
157 */
158 ts->next_tick = 0;
Frederic Weisbecker9e8f5592012-10-15 02:43:03 +0200159 }
Frederic Weisbecker94a57142012-10-15 16:17:16 +0200160#endif
Frederic Weisbecker9e8f5592012-10-15 02:43:03 +0200161 update_process_times(user_mode(regs));
162 profile_tick(CPU_PROFILING);
163}
Arnd Bergmann7809998a2016-01-25 16:41:49 +0100164#endif
Frederic Weisbecker9e8f5592012-10-15 02:43:03 +0200165
Frederic Weisbeckerc5bfece2013-04-12 16:45:34 +0200166#ifdef CONFIG_NO_HZ_FULL
Frederic Weisbecker460775df2013-07-24 23:52:27 +0200167cpumask_var_t tick_nohz_full_mask;
Paul E. McKenneyc0f489d2014-06-04 13:46:03 -0700168cpumask_var_t housekeeping_mask;
Frederic Weisbecker73867dc2013-07-24 23:31:00 +0200169bool tick_nohz_full_running;
Frederic Weisbeckerf009a7a2016-03-24 15:38:00 +0100170static atomic_t tick_dep_mask;
Frederic Weisbeckera8318812012-12-18 17:32:19 +0100171
Frederic Weisbeckerf009a7a2016-03-24 15:38:00 +0100172static bool check_tick_dependency(atomic_t *dep)
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200173{
Frederic Weisbeckerf009a7a2016-03-24 15:38:00 +0100174 int val = atomic_read(dep);
175
176 if (val & TICK_DEP_MASK_POSIX_TIMER) {
Frederic Weisbeckere6e6cc22015-12-11 03:27:25 +0100177 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
Frederic Weisbeckerf009a7a2016-03-24 15:38:00 +0100178 return true;
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200179 }
180
Frederic Weisbeckerf009a7a2016-03-24 15:38:00 +0100181 if (val & TICK_DEP_MASK_PERF_EVENTS) {
Frederic Weisbeckere6e6cc22015-12-11 03:27:25 +0100182 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
Frederic Weisbeckerf009a7a2016-03-24 15:38:00 +0100183 return true;
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200184 }
185
Frederic Weisbeckerf009a7a2016-03-24 15:38:00 +0100186 if (val & TICK_DEP_MASK_SCHED) {
Frederic Weisbeckere6e6cc22015-12-11 03:27:25 +0100187 trace_tick_stop(0, TICK_DEP_MASK_SCHED);
Frederic Weisbeckerf009a7a2016-03-24 15:38:00 +0100188 return true;
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200189 }
190
Frederic Weisbeckerf009a7a2016-03-24 15:38:00 +0100191 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
Frederic Weisbeckere6e6cc22015-12-11 03:27:25 +0100192 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
Frederic Weisbeckerf009a7a2016-03-24 15:38:00 +0100193 return true;
194 }
195
196 return false;
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200197}
198
Wanpeng Li57ccdf42016-09-07 18:51:13 +0800199static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
Frederic Weisbecker9014c452013-04-20 15:43:57 +0200200{
Frederic Weisbeckerebf3adb2017-11-06 16:01:20 +0100201 lockdep_assert_irqs_disabled();
Frederic Weisbecker9014c452013-04-20 15:43:57 +0200202
Wanpeng Li57ccdf42016-09-07 18:51:13 +0800203 if (unlikely(!cpu_online(cpu)))
204 return false;
205
Frederic Weisbeckerf009a7a2016-03-24 15:38:00 +0100206 if (check_tick_dependency(&tick_dep_mask))
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200207 return false;
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200208
Frederic Weisbeckerf009a7a2016-03-24 15:38:00 +0100209 if (check_tick_dependency(&ts->tick_dep_mask))
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200210 return false;
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200211
Frederic Weisbeckerf009a7a2016-03-24 15:38:00 +0100212 if (check_tick_dependency(&current->tick_dep_mask))
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200213 return false;
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200214
Frederic Weisbeckerf009a7a2016-03-24 15:38:00 +0100215 if (check_tick_dependency(&current->signal->tick_dep_mask))
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200216 return false;
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200217
Frederic Weisbecker9014c452013-04-20 15:43:57 +0200218 return true;
219}
220
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200221static void nohz_full_kick_func(struct irq_work *work)
Frederic Weisbecker76c24fb2013-04-18 00:15:40 +0200222{
Frederic Weisbecker73738a92015-05-27 19:22:08 +0200223 /* Empty, the tick restart happens on tick_nohz_irq_exit() */
Frederic Weisbecker76c24fb2013-04-18 00:15:40 +0200224}
225
226static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200227 .func = nohz_full_kick_func,
Frederic Weisbecker76c24fb2013-04-18 00:15:40 +0200228};
229
230/*
Frederic Weisbecker40bea032014-08-13 18:50:16 +0200231 * Kick this CPU if it's full dynticks in order to force it to
232 * re-evaluate its dependency on the tick and restart it if necessary.
233 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
234 * is NMI safe.
235 */
Frederic Weisbecker555e0c12015-07-16 17:42:29 +0200236static void tick_nohz_full_kick(void)
Frederic Weisbecker40bea032014-08-13 18:50:16 +0200237{
238 if (!tick_nohz_full_cpu(smp_processor_id()))
239 return;
240
Christoph Lameter56e4dea2014-10-27 10:49:45 -0500241 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
Frederic Weisbecker40bea032014-08-13 18:50:16 +0200242}
243
244/*
Frederic Weisbecker3d36aeb2014-06-04 16:17:33 +0200245 * Kick the CPU if it's full dynticks in order to force it to
Frederic Weisbecker76c24fb2013-04-18 00:15:40 +0200246 * re-evaluate its dependency on the tick and restart it if necessary.
247 */
Frederic Weisbecker3d36aeb2014-06-04 16:17:33 +0200248void tick_nohz_full_kick_cpu(int cpu)
Frederic Weisbecker76c24fb2013-04-18 00:15:40 +0200249{
Frederic Weisbecker3d36aeb2014-06-04 16:17:33 +0200250 if (!tick_nohz_full_cpu(cpu))
251 return;
252
253 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
Frederic Weisbecker76c24fb2013-04-18 00:15:40 +0200254}
255
Frederic Weisbecker76c24fb2013-04-18 00:15:40 +0200256/*
257 * Kick all full dynticks CPUs in order to force these to re-evaluate
258 * their dependency on the tick and restart it if necessary.
259 */
Frederic Weisbeckerb7878302015-07-17 22:25:49 +0200260static void tick_nohz_full_kick_all(void)
Frederic Weisbecker76c24fb2013-04-18 00:15:40 +0200261{
Frederic Weisbecker8537bb92015-12-07 16:55:23 +0100262 int cpu;
263
Frederic Weisbecker73867dc2013-07-24 23:31:00 +0200264 if (!tick_nohz_full_running)
Frederic Weisbecker76c24fb2013-04-18 00:15:40 +0200265 return;
266
267 preempt_disable();
Frederic Weisbecker8537bb92015-12-07 16:55:23 +0100268 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
269 tick_nohz_full_kick_cpu(cpu);
Frederic Weisbecker76c24fb2013-04-18 00:15:40 +0200270 preempt_enable();
271}
272
Frederic Weisbeckerf009a7a2016-03-24 15:38:00 +0100273static void tick_nohz_dep_set_all(atomic_t *dep,
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200274 enum tick_dep_bits bit)
275{
Frederic Weisbeckerf009a7a2016-03-24 15:38:00 +0100276 int prev;
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200277
Peter Zijlstraa1cc5bc2016-04-21 20:35:25 +0200278 prev = atomic_fetch_or(BIT(bit), dep);
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200279 if (!prev)
280 tick_nohz_full_kick_all();
281}
282
283/*
284 * Set a global tick dependency. Used by perf events that rely on freq and
285 * by unstable clock.
286 */
287void tick_nohz_dep_set(enum tick_dep_bits bit)
288{
289 tick_nohz_dep_set_all(&tick_dep_mask, bit);
290}
291
292void tick_nohz_dep_clear(enum tick_dep_bits bit)
293{
Frederic Weisbeckerf009a7a2016-03-24 15:38:00 +0100294 atomic_andnot(BIT(bit), &tick_dep_mask);
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200295}
296
297/*
298 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
299 * manage events throttling.
300 */
301void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
302{
Frederic Weisbeckerf009a7a2016-03-24 15:38:00 +0100303 int prev;
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200304 struct tick_sched *ts;
305
306 ts = per_cpu_ptr(&tick_cpu_sched, cpu);
307
Peter Zijlstraa1cc5bc2016-04-21 20:35:25 +0200308 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200309 if (!prev) {
310 preempt_disable();
311 /* Perf needs local kick that is NMI safe */
312 if (cpu == smp_processor_id()) {
313 tick_nohz_full_kick();
314 } else {
315 /* Remote irq work not NMI-safe */
316 if (!WARN_ON_ONCE(in_nmi()))
317 tick_nohz_full_kick_cpu(cpu);
318 }
319 preempt_enable();
320 }
321}
322
323void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
324{
325 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
326
Frederic Weisbeckerf009a7a2016-03-24 15:38:00 +0100327 atomic_andnot(BIT(bit), &ts->tick_dep_mask);
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200328}
329
330/*
331 * Set a per-task tick dependency. Posix CPU timers need this in order to elapse
332 * per task timers.
333 */
334void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
335{
336 /*
337 * We could optimize this with just kicking the target running the task
338 * if that noise matters for nohz full users.
339 */
340 tick_nohz_dep_set_all(&tsk->tick_dep_mask, bit);
341}
342
343void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
344{
Frederic Weisbeckerf009a7a2016-03-24 15:38:00 +0100345 atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200346}
347
348/*
349 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
350 * per process timers.
351 */
352void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit)
353{
354 tick_nohz_dep_set_all(&sig->tick_dep_mask, bit);
355}
356
357void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
358{
Frederic Weisbeckerf009a7a2016-03-24 15:38:00 +0100359 atomic_andnot(BIT(bit), &sig->tick_dep_mask);
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200360}
361
Frederic Weisbecker99e5ada2013-04-20 17:11:50 +0200362/*
363 * Re-evaluate the need for the tick as we switch the current task.
364 * It might need the tick due to per task/process properties:
Ingo Molnar0de76112016-07-01 12:42:35 +0200365 * perf events, posix CPU timers, ...
Frederic Weisbecker99e5ada2013-04-20 17:11:50 +0200366 */
Frederic Weisbeckerde734f82015-06-11 18:07:12 +0200367void __tick_nohz_task_switch(void)
Frederic Weisbecker99e5ada2013-04-20 17:11:50 +0200368{
369 unsigned long flags;
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200370 struct tick_sched *ts;
Frederic Weisbecker99e5ada2013-04-20 17:11:50 +0200371
Frederic Weisbecker99e5ada2013-04-20 17:11:50 +0200372 local_irq_save(flags);
373
Li Zhong6296ace2013-04-28 11:25:58 +0800374 if (!tick_nohz_full_cpu(smp_processor_id()))
375 goto out;
376
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200377 ts = this_cpu_ptr(&tick_cpu_sched);
Frederic Weisbecker99e5ada2013-04-20 17:11:50 +0200378
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200379 if (ts->tick_stopped) {
Frederic Weisbeckerf009a7a2016-03-24 15:38:00 +0100380 if (atomic_read(&current->tick_dep_mask) ||
381 atomic_read(&current->signal->tick_dep_mask))
Frederic Weisbeckerd027d452015-06-07 15:54:30 +0200382 tick_nohz_full_kick();
383 }
Li Zhong6296ace2013-04-28 11:25:58 +0800384out:
Frederic Weisbecker99e5ada2013-04-20 17:11:50 +0200385 local_irq_restore(flags);
386}
387
Frederic Weisbeckera8318812012-12-18 17:32:19 +0100388/* Parse the boot-time nohz CPU list from the kernel parameters. */
Frederic Weisbeckerc5bfece2013-04-12 16:45:34 +0200389static int __init tick_nohz_full_setup(char *str)
Frederic Weisbeckera8318812012-12-18 17:32:19 +0100390{
Frederic Weisbecker73867dc2013-07-24 23:31:00 +0200391 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
392 if (cpulist_parse(str, tick_nohz_full_mask) < 0) {
Joe Perchesa395d6a2016-03-22 14:28:09 -0700393 pr_warn("NO_HZ: Incorrect nohz_full cpumask\n");
Frederic Weisbecker4327b152014-08-17 22:02:55 +0200394 free_bootmem_cpumask_var(tick_nohz_full_mask);
Frederic Weisbecker0453b432013-03-27 02:18:34 +0100395 return 1;
396 }
Frederic Weisbecker73867dc2013-07-24 23:31:00 +0200397 tick_nohz_full_running = true;
Frederic Weisbecker0453b432013-03-27 02:18:34 +0100398
Frederic Weisbeckera8318812012-12-18 17:32:19 +0100399 return 1;
400}
Frederic Weisbeckerc5bfece2013-04-12 16:45:34 +0200401__setup("nohz_full=", tick_nohz_full_setup);
Frederic Weisbeckera8318812012-12-18 17:32:19 +0100402
Sebastian Andrzej Siewior31eff2432016-11-17 19:35:34 +0100403static int tick_nohz_cpu_down(unsigned int cpu)
Frederic Weisbeckera382bf92012-12-18 18:24:35 +0100404{
Sebastian Andrzej Siewior31eff2432016-11-17 19:35:34 +0100405 /*
406 * The boot CPU handles housekeeping duty (unbound timers,
407 * workqueues, timekeeping, ...) on behalf of full dynticks
408 * CPUs. It must remain online when nohz full is enabled.
409 */
410 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
411 return -EBUSY;
412 return 0;
Frederic Weisbeckera382bf92012-12-18 18:24:35 +0100413}
414
Frederic Weisbeckerf98823a2013-03-27 01:17:22 +0100415static int tick_nohz_init_all(void)
416{
417 int err = -1;
418
419#ifdef CONFIG_NO_HZ_FULL_ALL
Frederic Weisbecker73867dc2013-07-24 23:31:00 +0200420 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) {
Frederic Weisbecker4327b152014-08-17 22:02:55 +0200421 WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n");
Paul E. McKenneyc0f489d2014-06-04 13:46:03 -0700422 return err;
423 }
Frederic Weisbeckerf98823a2013-03-27 01:17:22 +0100424 err = 0;
Frederic Weisbecker73867dc2013-07-24 23:31:00 +0200425 cpumask_setall(tick_nohz_full_mask);
Frederic Weisbecker73867dc2013-07-24 23:31:00 +0200426 tick_nohz_full_running = true;
Frederic Weisbeckerf98823a2013-03-27 01:17:22 +0100427#endif
428 return err;
429}
430
Frederic Weisbeckerd1e43fa2013-03-26 23:47:24 +0100431void __init tick_nohz_init(void)
Frederic Weisbeckera8318812012-12-18 17:32:19 +0100432{
Sebastian Andrzej Siewior31eff2432016-11-17 19:35:34 +0100433 int cpu, ret;
Frederic Weisbeckerd1e43fa2013-03-26 23:47:24 +0100434
Frederic Weisbecker73867dc2013-07-24 23:31:00 +0200435 if (!tick_nohz_full_running) {
Frederic Weisbeckerf98823a2013-03-27 01:17:22 +0100436 if (tick_nohz_init_all() < 0)
437 return;
438 }
Frederic Weisbeckerd1e43fa2013-03-26 23:47:24 +0100439
Frederic Weisbecker4327b152014-08-17 22:02:55 +0200440 if (!alloc_cpumask_var(&housekeeping_mask, GFP_KERNEL)) {
441 WARN(1, "NO_HZ: Can't allocate not-full dynticks cpumask\n");
442 cpumask_clear(tick_nohz_full_mask);
443 tick_nohz_full_running = false;
444 return;
445 }
446
Frederic Weisbecker9b01f5b2014-08-18 01:36:07 +0200447 /*
448 * Full dynticks uses irq work to drive the tick rescheduling on safe
449 * locking contexts. But then we need irq work to raise its own
450 * interrupts to avoid circular dependency on the tick
451 */
452 if (!arch_irq_work_has_interrupt()) {
Joe Perchesa395d6a2016-03-22 14:28:09 -0700453 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n");
Frederic Weisbecker9b01f5b2014-08-18 01:36:07 +0200454 cpumask_clear(tick_nohz_full_mask);
455 cpumask_copy(housekeeping_mask, cpu_possible_mask);
456 tick_nohz_full_running = false;
457 return;
458 }
459
Frederic Weisbecker4327b152014-08-17 22:02:55 +0200460 cpu = smp_processor_id();
461
462 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
Joe Perchesa395d6a2016-03-22 14:28:09 -0700463 pr_warn("NO_HZ: Clearing %d from nohz_full range for timekeeping\n",
464 cpu);
Frederic Weisbecker4327b152014-08-17 22:02:55 +0200465 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
466 }
467
468 cpumask_andnot(housekeeping_mask,
469 cpu_possible_mask, tick_nohz_full_mask);
470
Frederic Weisbecker73867dc2013-07-24 23:31:00 +0200471 for_each_cpu(cpu, tick_nohz_full_mask)
Frederic Weisbecker2e709332013-07-10 00:55:25 +0200472 context_tracking_cpu_set(cpu);
473
Sebastian Andrzej Siewior31eff2432016-11-17 19:35:34 +0100474 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
475 "kernel/nohz:predown", NULL,
476 tick_nohz_cpu_down);
477 WARN_ON(ret < 0);
Tejun Heoffda22c2015-02-13 14:37:31 -0800478 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
479 cpumask_pr_args(tick_nohz_full_mask));
Frederic Weisbecker7c8bb6c2015-09-01 16:51:00 +0200480
481 /*
482 * We need at least one CPU to handle housekeeping work such
483 * as timekeeping, unbound timers, workqueues, ...
484 */
485 WARN_ON_ONCE(cpumask_empty(housekeeping_mask));
Frederic Weisbeckera8318812012-12-18 17:32:19 +0100486}
Frederic Weisbeckera8318812012-12-18 17:32:19 +0100487#endif
488
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800489/*
490 * NOHZ - aka dynamic tick functionality
491 */
Frederic Weisbecker3451d022011-08-10 23:21:01 +0200492#ifdef CONFIG_NO_HZ_COMMON
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800493/*
494 * NO HZ enabled ?
495 */
Kees Cook4cc7ecb72016-03-17 14:23:00 -0700496bool tick_nohz_enabled __read_mostly = true;
Thomas Gleixnerbc7a34b2015-05-26 22:50:33 +0000497unsigned long tick_nohz_active __read_mostly;
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800498/*
499 * Enable / Disable tickless mode
500 */
501static int __init setup_tick_nohz(char *str)
502{
Kees Cook4cc7ecb72016-03-17 14:23:00 -0700503 return (kstrtobool(str, &tick_nohz_enabled) == 0);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800504}
505
506__setup("nohz=", setup_tick_nohz);
507
Thomas Gleixnerc1797ba2015-03-25 13:07:37 +0100508int tick_nohz_tick_stopped(void)
509{
510 return __this_cpu_read(tick_cpu_sched.tick_stopped);
511}
512
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800513/**
514 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
515 *
516 * Called from interrupt entry when the CPU was idle
517 *
518 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
519 * must be updated. Otherwise an interrupt handler could use a stale jiffy
Ingo Molnar0de76112016-07-01 12:42:35 +0200520 * value. We do this unconditionally on any CPU, as we don't know whether the
521 * CPU, which has the update task assigned is in a long sleep.
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800522 */
Martin Schwidefskyeed3b9c2009-09-29 14:25:15 +0200523static void tick_nohz_update_jiffies(ktime_t now)
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800524{
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800525 unsigned long flags;
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800526
Frederic Weisbeckere8fcaa52013-08-07 22:28:01 +0200527 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800528
529 local_irq_save(flags);
530 tick_do_update_jiffies64(now);
531 local_irq_restore(flags);
Ingo Molnar02ff3752008-05-12 15:43:53 +0200532
Tejun Heo03e0d462015-12-08 11:28:04 -0500533 touch_softlockup_watchdog_sched();
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800534}
535
Arjan van de Ven595aac42010-05-09 08:22:45 -0700536/*
Ingo Molnar0de76112016-07-01 12:42:35 +0200537 * Updates the per-CPU time idle statistics counters
Arjan van de Ven595aac42010-05-09 08:22:45 -0700538 */
Arjan van de Ven8d63bf92010-05-09 08:24:03 -0700539static void
Peter Zijlstra8c215bd2010-07-01 09:07:17 +0200540update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
Arjan van de Ven595aac42010-05-09 08:22:45 -0700541{
542 ktime_t delta;
543
Arjan van de Ven595aac42010-05-09 08:22:45 -0700544 if (ts->idle_active) {
545 delta = ktime_sub(now, ts->idle_entrytime);
Peter Zijlstra8c215bd2010-07-01 09:07:17 +0200546 if (nr_iowait_cpu(cpu) > 0)
Arjan van de Ven0224cf42010-05-09 08:25:23 -0700547 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
Michal Hocko6beea0c2011-08-24 09:37:48 +0200548 else
549 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
Arjan van de Ven8c7b09f2010-05-09 08:23:23 -0700550 ts->idle_entrytime = now;
Arjan van de Ven595aac42010-05-09 08:22:45 -0700551 }
Arjan van de Ven8d63bf92010-05-09 08:24:03 -0700552
Arjan van de Vene0e37c22010-05-09 08:24:39 -0700553 if (last_update_time)
Arjan van de Ven8d63bf92010-05-09 08:24:03 -0700554 *last_update_time = ktime_to_us(now);
555
Arjan van de Ven595aac42010-05-09 08:22:45 -0700556}
557
Frederic Weisbeckere8fcaa52013-08-07 22:28:01 +0200558static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
Venki Pallipadi6378ddb2008-01-30 13:30:04 +0100559{
Frederic Weisbeckere8fcaa52013-08-07 22:28:01 +0200560 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
Martin Schwidefskyeed3b9c2009-09-29 14:25:15 +0200561 ts->idle_active = 0;
Peter Zijlstra56c74262008-09-01 16:44:23 +0200562
Peter Zijlstraac1e8432017-04-21 12:26:23 +0200563 sched_clock_idle_wakeup_event();
Venki Pallipadi6378ddb2008-01-30 13:30:04 +0100564}
565
Frederic Weisbeckere8fcaa52013-08-07 22:28:01 +0200566static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
Venki Pallipadi6378ddb2008-01-30 13:30:04 +0100567{
Michal Hocko430ee882011-12-01 17:00:22 +0100568 ktime_t now = ktime_get();
Arjan van de Ven595aac42010-05-09 08:22:45 -0700569
Venki Pallipadi6378ddb2008-01-30 13:30:04 +0100570 ts->idle_entrytime = now;
571 ts->idle_active = 1;
Peter Zijlstra56c74262008-09-01 16:44:23 +0200572 sched_clock_idle_sleep_event();
Venki Pallipadi6378ddb2008-01-30 13:30:04 +0100573 return now;
574}
575
Arjan van de Venb1f724c2010-05-09 08:22:08 -0700576/**
Ingo Molnar0de76112016-07-01 12:42:35 +0200577 * get_cpu_idle_time_us - get the total idle time of a CPU
Arjan van de Venb1f724c2010-05-09 08:22:08 -0700578 * @cpu: CPU number to query
Michal Hocko09a1d342011-08-24 09:39:30 +0200579 * @last_update_time: variable to store update time in. Do not update
580 * counters if NULL.
Arjan van de Venb1f724c2010-05-09 08:22:08 -0700581 *
Wei Jiangang6168f8e2016-06-29 12:51:50 +0800582 * Return the cumulative idle time (since boot) for a given
Michal Hocko6beea0c2011-08-24 09:37:48 +0200583 * CPU, in microseconds.
Arjan van de Venb1f724c2010-05-09 08:22:08 -0700584 *
585 * This time is measured via accounting rather than sampling,
586 * and is as accurate as ktime_get() is.
587 *
588 * This function returns -1 if NOHZ is not enabled.
589 */
Venki Pallipadi6378ddb2008-01-30 13:30:04 +0100590u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
591{
592 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
Michal Hocko09a1d342011-08-24 09:39:30 +0200593 ktime_t now, idle;
Venki Pallipadi6378ddb2008-01-30 13:30:04 +0100594
Thomas Gleixnerd689fe22013-11-13 21:01:57 +0100595 if (!tick_nohz_active)
venkatesh.pallipadi@intel.com8083e4a2008-08-04 11:59:11 -0700596 return -1;
597
Michal Hocko09a1d342011-08-24 09:39:30 +0200598 now = ktime_get();
599 if (last_update_time) {
600 update_ts_time_stats(cpu, ts, now, last_update_time);
601 idle = ts->idle_sleeptime;
602 } else {
603 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
604 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
venkatesh.pallipadi@intel.com8083e4a2008-08-04 11:59:11 -0700605
Michal Hocko09a1d342011-08-24 09:39:30 +0200606 idle = ktime_add(ts->idle_sleeptime, delta);
607 } else {
608 idle = ts->idle_sleeptime;
609 }
610 }
611
612 return ktime_to_us(idle);
613
Venki Pallipadi6378ddb2008-01-30 13:30:04 +0100614}
venkatesh.pallipadi@intel.com8083e4a2008-08-04 11:59:11 -0700615EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
Venki Pallipadi6378ddb2008-01-30 13:30:04 +0100616
Michal Hocko6beea0c2011-08-24 09:37:48 +0200617/**
Ingo Molnar0de76112016-07-01 12:42:35 +0200618 * get_cpu_iowait_time_us - get the total iowait time of a CPU
Arjan van de Ven0224cf42010-05-09 08:25:23 -0700619 * @cpu: CPU number to query
Michal Hocko09a1d342011-08-24 09:39:30 +0200620 * @last_update_time: variable to store update time in. Do not update
621 * counters if NULL.
Arjan van de Ven0224cf42010-05-09 08:25:23 -0700622 *
Wei Jiangang6168f8e2016-06-29 12:51:50 +0800623 * Return the cumulative iowait time (since boot) for a given
Arjan van de Ven0224cf42010-05-09 08:25:23 -0700624 * CPU, in microseconds.
625 *
626 * This time is measured via accounting rather than sampling,
627 * and is as accurate as ktime_get() is.
628 *
629 * This function returns -1 if NOHZ is not enabled.
630 */
631u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
632{
633 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
Michal Hocko09a1d342011-08-24 09:39:30 +0200634 ktime_t now, iowait;
Arjan van de Ven0224cf42010-05-09 08:25:23 -0700635
Thomas Gleixnerd689fe22013-11-13 21:01:57 +0100636 if (!tick_nohz_active)
Arjan van de Ven0224cf42010-05-09 08:25:23 -0700637 return -1;
638
Michal Hocko09a1d342011-08-24 09:39:30 +0200639 now = ktime_get();
640 if (last_update_time) {
641 update_ts_time_stats(cpu, ts, now, last_update_time);
642 iowait = ts->iowait_sleeptime;
643 } else {
644 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
645 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
Arjan van de Ven0224cf42010-05-09 08:25:23 -0700646
Michal Hocko09a1d342011-08-24 09:39:30 +0200647 iowait = ktime_add(ts->iowait_sleeptime, delta);
648 } else {
649 iowait = ts->iowait_sleeptime;
650 }
651 }
652
653 return ktime_to_us(iowait);
Arjan van de Ven0224cf42010-05-09 08:25:23 -0700654}
655EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
656
Thomas Gleixner0ff53d02015-04-14 21:08:54 +0000657static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
658{
659 hrtimer_cancel(&ts->sched_timer);
660 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
661
662 /* Forward the time to expire in the future */
663 hrtimer_forward(&ts->sched_timer, now, tick_period);
664
665 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
666 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
667 else
668 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
Frederic Weisbecker411fe242017-04-21 16:00:54 +0200669
670 /*
671 * Reset to make sure next tick stop doesn't get fooled by past
672 * cached clock deadline.
673 */
674 ts->next_tick = 0;
Thomas Gleixner0ff53d02015-04-14 21:08:54 +0000675}
676
Frederic Weisbecker84bf1bc2011-08-01 01:25:38 +0200677static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
678 ktime_t now, int cpu)
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800679{
Christoph Lameter22127e92014-08-17 12:30:25 -0500680 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
Thomas Gleixnerc1ad3482015-04-14 21:08:58 +0000681 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires;
682 unsigned long seq, basejiff;
683 ktime_t tick;
Frederic Weisbecker855a0fc2013-12-17 00:16:37 +0100684
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800685 /* Read jiffies and the time when jiffies were updated last */
686 do {
John Stultzd6ad4182012-02-28 16:50:11 -0800687 seq = read_seqbegin(&jiffies_lock);
Thomas Gleixner2456e852016-12-25 11:38:40 +0100688 basemono = last_jiffies_update;
Thomas Gleixnerc1ad3482015-04-14 21:08:58 +0000689 basejiff = jiffies;
John Stultzd6ad4182012-02-28 16:50:11 -0800690 } while (read_seqretry(&jiffies_lock, seq));
Thomas Gleixnerc1ad3482015-04-14 21:08:58 +0000691 ts->last_jiffies = basejiff;
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800692
Thomas Gleixnerc1ad3482015-04-14 21:08:58 +0000693 if (rcu_needs_cpu(basemono, &next_rcu) ||
Martin Schwidefskyfe0f4972014-09-30 17:37:52 +0200694 arch_needs_cpu() || irq_work_needs_cpu()) {
Thomas Gleixnerc1ad3482015-04-14 21:08:58 +0000695 next_tick = basemono + TICK_NSEC;
Martin Schwidefsky3c5d92a2009-09-29 14:25:16 +0200696 } else {
Thomas Gleixnerc1ad3482015-04-14 21:08:58 +0000697 /*
698 * Get the next pending timer. If high resolution
699 * timers are enabled this only takes the timer wheel
700 * timers into account. If high resolution timers are
701 * disabled this also looks at the next expiring
702 * hrtimer.
703 */
704 next_tmr = get_next_timer_interrupt(basejiff, basemono);
705 ts->next_timer = next_tmr;
706 /* Take the next rcu event into account */
707 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr;
Martin Schwidefsky3c5d92a2009-09-29 14:25:16 +0200708 }
Ingo Molnar47aa8b62013-04-26 10:05:59 +0200709
Thomas Gleixnerc1ad3482015-04-14 21:08:58 +0000710 /*
711 * If the tick is due in the next period, keep it ticking or
Peter Zijlstra82bbe342015-11-19 17:21:06 +0100712 * force prod the timer.
Thomas Gleixnerc1ad3482015-04-14 21:08:58 +0000713 */
714 delta = next_tick - basemono;
715 if (delta <= (u64)TICK_NSEC) {
Thomas Gleixnera683f392016-07-04 09:50:36 +0000716 /*
717 * Tell the timer code that the base is not idle, i.e. undo
718 * the effect of get_next_timer_interrupt():
719 */
720 timer_clear_idle();
Peter Zijlstra82bbe342015-11-19 17:21:06 +0100721 /*
722 * We've not stopped the tick yet, and there's a timer in the
723 * next period, so no point in stopping it either, bail.
724 */
Frederic Weisbeckerf99973e2017-06-01 16:47:09 +0200725 if (!ts->tick_stopped) {
726 tick = 0;
Thomas Gleixnereaad0842007-05-29 23:47:39 +0200727 goto out;
Thomas Gleixner157d29e2015-04-14 21:08:56 +0000728 }
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800729 }
Thomas Gleixner0ff53d02015-04-14 21:08:54 +0000730
Thomas Gleixner157d29e2015-04-14 21:08:56 +0000731 /*
Ingo Molnar0de76112016-07-01 12:42:35 +0200732 * If this CPU is the one which updates jiffies, then give up
733 * the assignment and let it be taken by the CPU which runs
734 * the tick timer next, which might be this CPU as well. If we
Thomas Gleixner157d29e2015-04-14 21:08:56 +0000735 * don't drop this here the jiffies might be stale and
736 * do_timer() never invoked. Keep track of the fact that it
Ingo Molnar0de76112016-07-01 12:42:35 +0200737 * was the one which had the do_timer() duty last. If this CPU
Thomas Gleixner157d29e2015-04-14 21:08:56 +0000738 * is the one which had the do_timer() duty last, we limit the
Wei Jiangang6168f8e2016-06-29 12:51:50 +0800739 * sleep time to the timekeeping max_deferment value.
Thomas Gleixnerc1ad3482015-04-14 21:08:58 +0000740 * Otherwise we can sleep as long as we want.
Thomas Gleixner157d29e2015-04-14 21:08:56 +0000741 */
Thomas Gleixnerc1ad3482015-04-14 21:08:58 +0000742 delta = timekeeping_max_deferment();
Thomas Gleixner157d29e2015-04-14 21:08:56 +0000743 if (cpu == tick_do_timer_cpu) {
744 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
745 ts->do_timer_last = 1;
746 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
Thomas Gleixnerc1ad3482015-04-14 21:08:58 +0000747 delta = KTIME_MAX;
Thomas Gleixner157d29e2015-04-14 21:08:56 +0000748 ts->do_timer_last = 0;
749 } else if (!ts->do_timer_last) {
Thomas Gleixnerc1ad3482015-04-14 21:08:58 +0000750 delta = KTIME_MAX;
Thomas Gleixner157d29e2015-04-14 21:08:56 +0000751 }
752
753#ifdef CONFIG_NO_HZ_FULL
Thomas Gleixnerc1ad3482015-04-14 21:08:58 +0000754 /* Limit the tick delta to the maximum scheduler deferment */
Thomas Gleixner157d29e2015-04-14 21:08:56 +0000755 if (!ts->inidle)
Thomas Gleixnerc1ad3482015-04-14 21:08:58 +0000756 delta = min(delta, scheduler_tick_max_deferment());
Thomas Gleixner157d29e2015-04-14 21:08:56 +0000757#endif
758
Thomas Gleixnerc1ad3482015-04-14 21:08:58 +0000759 /* Calculate the next expiry time */
760 if (delta < (KTIME_MAX - basemono))
761 expires = basemono + delta;
Thomas Gleixner157d29e2015-04-14 21:08:56 +0000762 else
Thomas Gleixnerc1ad3482015-04-14 21:08:58 +0000763 expires = KTIME_MAX;
764
765 expires = min_t(u64, expires, next_tick);
Thomas Gleixner2456e852016-12-25 11:38:40 +0100766 tick = expires;
Thomas Gleixner157d29e2015-04-14 21:08:56 +0000767
768 /* Skip reprogram of event if its not changed */
Frederic Weisbecker411fe242017-04-21 16:00:54 +0200769 if (ts->tick_stopped && (expires == ts->next_tick)) {
770 /* Sanity check: make sure clockevent is actually programmed */
Frederic Weisbeckerd4af6d92017-06-13 06:04:14 +0200771 if (tick == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
Frederic Weisbeckerce6cf9a2017-05-11 16:36:19 +0200772 goto out;
Frederic Weisbecker411fe242017-04-21 16:00:54 +0200773
774 WARN_ON_ONCE(1);
775 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n",
776 basemono, ts->next_tick, dev->next_event,
777 hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer));
Frederic Weisbeckerce6cf9a2017-05-11 16:36:19 +0200778 }
Thomas Gleixner157d29e2015-04-14 21:08:56 +0000779
Thomas Gleixner157d29e2015-04-14 21:08:56 +0000780 /*
781 * nohz_stop_sched_tick can be called several times before
782 * the nohz_restart_sched_tick is called. This happens when
783 * interrupts arrive which do not cause a reschedule. In the
784 * first call we save the current tick time, so we can restart
785 * the scheduler tick in nohz_restart_sched_tick.
786 */
787 if (!ts->tick_stopped) {
Frederic Weisbecker3c85d6d2017-06-19 04:12:00 +0200788 calc_load_nohz_start();
Frederic Weisbecker1f419062016-04-13 15:56:51 +0200789 cpu_load_update_nohz_start();
Thomas Gleixner157d29e2015-04-14 21:08:56 +0000790
791 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
792 ts->tick_stopped = 1;
Frederic Weisbeckere6e6cc22015-12-11 03:27:25 +0100793 trace_tick_stop(1, TICK_DEP_MASK_NONE);
Thomas Gleixner157d29e2015-04-14 21:08:56 +0000794 }
795
Frederic Weisbecker411fe242017-04-21 16:00:54 +0200796 ts->next_tick = tick;
797
Thomas Gleixner157d29e2015-04-14 21:08:56 +0000798 /*
Thomas Gleixnerc1ad3482015-04-14 21:08:58 +0000799 * If the expiration time == KTIME_MAX, then we simply stop
800 * the tick timer.
Thomas Gleixner157d29e2015-04-14 21:08:56 +0000801 */
Thomas Gleixnerc1ad3482015-04-14 21:08:58 +0000802 if (unlikely(expires == KTIME_MAX)) {
Thomas Gleixner157d29e2015-04-14 21:08:56 +0000803 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
804 hrtimer_cancel(&ts->sched_timer);
805 goto out;
806 }
807
Frederic Weisbeckerd4af6d92017-06-13 06:04:14 +0200808 hrtimer_set_expires(&ts->sched_timer, tick);
809
Thomas Gleixner157d29e2015-04-14 21:08:56 +0000810 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
Frederic Weisbeckerd4af6d92017-06-13 06:04:14 +0200811 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
Thomas Gleixner157d29e2015-04-14 21:08:56 +0000812 else
Thomas Gleixnerc1ad3482015-04-14 21:08:58 +0000813 tick_program_event(tick, 1);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800814out:
Frederic Weisbecker411fe242017-04-21 16:00:54 +0200815 /*
816 * Update the estimated sleep length until the next timer
817 * (not only the tick).
818 */
Len Brown4f86d3a2007-10-03 18:58:00 -0400819 ts->sleep_length = ktime_sub(dev->next_event, now);
Thomas Gleixnerc1ad3482015-04-14 21:08:58 +0000820 return tick;
Frederic Weisbecker280f0672011-10-07 18:22:06 +0200821}
822
Frederic Weisbecker1f419062016-04-13 15:56:51 +0200823static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
Frederic Weisbecker59d2c7c2015-05-29 14:42:15 +0200824{
825 /* Update jiffies first */
826 tick_do_update_jiffies64(now);
Frederic Weisbecker1f419062016-04-13 15:56:51 +0200827 cpu_load_update_nohz_stop();
Frederic Weisbecker59d2c7c2015-05-29 14:42:15 +0200828
Thomas Gleixnera683f392016-07-04 09:50:36 +0000829 /*
830 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
831 * the clock forward checks in the enqueue path:
832 */
833 timer_clear_idle();
834
Frederic Weisbecker3c85d6d2017-06-19 04:12:00 +0200835 calc_load_nohz_stop();
Tejun Heo03e0d462015-12-08 11:28:04 -0500836 touch_softlockup_watchdog_sched();
Frederic Weisbecker59d2c7c2015-05-29 14:42:15 +0200837 /*
838 * Cancel the scheduled timer and restore the tick
839 */
840 ts->tick_stopped = 0;
841 ts->idle_exittime = now;
842
843 tick_nohz_restart(ts, now);
844}
Frederic Weisbecker73738a92015-05-27 19:22:08 +0200845
846static void tick_nohz_full_update_tick(struct tick_sched *ts)
Frederic Weisbecker5811d992013-04-20 16:40:31 +0200847{
848#ifdef CONFIG_NO_HZ_FULL
Alex Shie9a2eb42013-11-28 14:27:11 +0800849 int cpu = smp_processor_id();
Frederic Weisbecker5811d992013-04-20 16:40:31 +0200850
Frederic Weisbecker59449352015-05-27 15:42:42 +0200851 if (!tick_nohz_full_cpu(cpu))
Alex Shie9a2eb42013-11-28 14:27:11 +0800852 return;
Frederic Weisbecker5811d992013-04-20 16:40:31 +0200853
Alex Shie9a2eb42013-11-28 14:27:11 +0800854 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
855 return;
Frederic Weisbecker5811d992013-04-20 16:40:31 +0200856
Wanpeng Li57ccdf42016-09-07 18:51:13 +0800857 if (can_stop_full_tick(cpu, ts))
Frederic Weisbecker73738a92015-05-27 19:22:08 +0200858 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu);
859 else if (ts->tick_stopped)
Frederic Weisbecker1f419062016-04-13 15:56:51 +0200860 tick_nohz_restart_sched_tick(ts, ktime_get());
Frederic Weisbecker5811d992013-04-20 16:40:31 +0200861#endif
862}
863
Frederic Weisbecker5b399392011-08-01 00:06:10 +0200864static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
865{
866 /*
Ingo Molnar0de76112016-07-01 12:42:35 +0200867 * If this CPU is offline and it is the one which updates
Frederic Weisbecker5b399392011-08-01 00:06:10 +0200868 * jiffies, then give up the assignment and let it be taken by
Ingo Molnar0de76112016-07-01 12:42:35 +0200869 * the CPU which runs the tick timer next. If we don't drop
Frederic Weisbecker5b399392011-08-01 00:06:10 +0200870 * this here the jiffies might be stale and do_timer() never
871 * invoked.
872 */
873 if (unlikely(!cpu_online(cpu))) {
874 if (cpu == tick_do_timer_cpu)
875 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
Frederic Weisbecker411fe242017-04-21 16:00:54 +0200876 /*
877 * Make sure the CPU doesn't get fooled by obsolete tick
878 * deadline if it comes back online later.
879 */
880 ts->next_tick = 0;
Thomas Gleixnerf7ea0fd2013-05-13 21:40:27 +0200881 return false;
Frederic Weisbecker5b399392011-08-01 00:06:10 +0200882 }
883
Thomas Gleixner0e576acb2013-11-29 12:18:13 +0100884 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) {
Thomas Gleixner2456e852016-12-25 11:38:40 +0100885 ts->sleep_length = NSEC_PER_SEC / HZ;
Frederic Weisbecker5b399392011-08-01 00:06:10 +0200886 return false;
Thomas Gleixner0e576acb2013-11-29 12:18:13 +0100887 }
Frederic Weisbecker5b399392011-08-01 00:06:10 +0200888
889 if (need_resched())
890 return false;
891
892 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
893 static int ratelimit;
894
Paul E. McKenney803b0eb2012-08-23 08:34:07 -0700895 if (ratelimit < 10 &&
896 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
Rado Vrbovskycfea7d72013-02-08 12:37:30 -0500897 pr_warn("NOHZ: local_softirq_pending %02x\n",
898 (unsigned int) local_softirq_pending());
Frederic Weisbecker5b399392011-08-01 00:06:10 +0200899 ratelimit++;
900 }
901 return false;
902 }
903
Frederic Weisbecker460775df2013-07-24 23:52:27 +0200904 if (tick_nohz_full_enabled()) {
Frederic Weisbeckera382bf92012-12-18 18:24:35 +0100905 /*
906 * Keep the tick alive to guarantee timekeeping progression
907 * if there are full dynticks CPUs around
908 */
909 if (tick_do_timer_cpu == cpu)
910 return false;
911 /*
912 * Boot safety: make sure the timekeeping duty has been
913 * assigned before entering dyntick-idle mode,
914 */
915 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
916 return false;
917 }
918
Frederic Weisbecker5b399392011-08-01 00:06:10 +0200919 return true;
920}
921
Frederic Weisbecker19f5f732011-07-27 17:29:28 +0200922static void __tick_nohz_idle_enter(struct tick_sched *ts)
923{
Frederic Weisbecker84bf1bc2011-08-01 01:25:38 +0200924 ktime_t now, expires;
Frederic Weisbecker5b399392011-08-01 00:06:10 +0200925 int cpu = smp_processor_id();
Frederic Weisbecker19f5f732011-07-27 17:29:28 +0200926
Wanpeng Li08d072592016-09-02 14:38:23 +0800927 now = tick_nohz_start_idle(ts);
928
Frederic Weisbecker5b399392011-08-01 00:06:10 +0200929 if (can_stop_idle_tick(cpu, ts)) {
930 int was_stopped = ts->tick_stopped;
931
932 ts->idle_calls++;
Frederic Weisbecker84bf1bc2011-08-01 01:25:38 +0200933
934 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
Thomas Gleixner2456e852016-12-25 11:38:40 +0100935 if (expires > 0LL) {
Frederic Weisbecker84bf1bc2011-08-01 01:25:38 +0200936 ts->idle_sleeps++;
937 ts->idle_expires = expires;
938 }
Frederic Weisbecker5b399392011-08-01 00:06:10 +0200939
Frederic Weisbeckera0db9712017-06-19 04:12:01 +0200940 if (!was_stopped && ts->tick_stopped) {
Frederic Weisbecker5b399392011-08-01 00:06:10 +0200941 ts->idle_jiffies = ts->last_jiffies;
Frederic Weisbeckera0db9712017-06-19 04:12:01 +0200942 nohz_balance_enter_idle(cpu);
943 }
Frederic Weisbecker5b399392011-08-01 00:06:10 +0200944 }
Frederic Weisbecker280f0672011-10-07 18:22:06 +0200945}
946
947/**
948 * tick_nohz_idle_enter - stop the idle tick from the idle task
949 *
950 * When the next event is more than a tick into the future, stop the idle tick
951 * Called when we start the idle loop.
Frederic Weisbecker2bbb6812011-10-08 16:01:00 +0200952 *
Frederic Weisbecker1268fbc2011-11-17 18:48:14 +0100953 * The arch is responsible of calling:
Frederic Weisbecker2bbb6812011-10-08 16:01:00 +0200954 *
955 * - rcu_idle_enter() after its last use of RCU before the CPU is put
956 * to sleep.
957 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
Frederic Weisbecker280f0672011-10-07 18:22:06 +0200958 */
Frederic Weisbecker1268fbc2011-11-17 18:48:14 +0100959void tick_nohz_idle_enter(void)
Frederic Weisbecker280f0672011-10-07 18:22:06 +0200960{
961 struct tick_sched *ts;
962
Frederic Weisbeckerebf3adb2017-11-06 16:01:20 +0100963 lockdep_assert_irqs_enabled();
Linus Torvalds0db49b72012-01-06 08:33:28 -0800964 /*
Ingo Molnar0de76112016-07-01 12:42:35 +0200965 * Update the idle state in the scheduler domain hierarchy
966 * when tick_nohz_stop_sched_tick() is called from the idle loop.
967 * State will be updated to busy during the first busy tick after
968 * exiting idle.
969 */
Linus Torvalds0db49b72012-01-06 08:33:28 -0800970 set_cpu_sd_state_idle();
971
Frederic Weisbecker1268fbc2011-11-17 18:48:14 +0100972 local_irq_disable();
973
Christoph Lameter22127e92014-08-17 12:30:25 -0500974 ts = this_cpu_ptr(&tick_cpu_sched);
Frederic Weisbecker280f0672011-10-07 18:22:06 +0200975 ts->inidle = 1;
Frederic Weisbecker19f5f732011-07-27 17:29:28 +0200976 __tick_nohz_idle_enter(ts);
Frederic Weisbecker1268fbc2011-11-17 18:48:14 +0100977
978 local_irq_enable();
Frederic Weisbecker280f0672011-10-07 18:22:06 +0200979}
980
981/**
982 * tick_nohz_irq_exit - update next tick event from interrupt exit
983 *
984 * When an interrupt fires while we are idle and it doesn't cause
985 * a reschedule, it may still add, modify or delete a timer, enqueue
986 * an RCU callback, etc...
987 * So we need to re-calculate and reprogram the next tick event.
988 */
989void tick_nohz_irq_exit(void)
990{
Christoph Lameter22127e92014-08-17 12:30:25 -0500991 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
Frederic Weisbecker280f0672011-10-07 18:22:06 +0200992
Rafael J. Wysocki14851912013-07-27 01:41:34 +0200993 if (ts->inidle)
Frederic Weisbecker5811d992013-04-20 16:40:31 +0200994 __tick_nohz_idle_enter(ts);
Rafael J. Wysocki14851912013-07-27 01:41:34 +0200995 else
Frederic Weisbecker73738a92015-05-27 19:22:08 +0200996 tick_nohz_full_update_tick(ts);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800997}
998
999/**
Len Brown4f86d3a2007-10-03 18:58:00 -04001000 * tick_nohz_get_sleep_length - return the length of the current sleep
1001 *
1002 * Called from power state control code with interrupts disabled
1003 */
1004ktime_t tick_nohz_get_sleep_length(void)
1005{
Christoph Lameter22127e92014-08-17 12:30:25 -05001006 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
Len Brown4f86d3a2007-10-03 18:58:00 -04001007
1008 return ts->sleep_length;
1009}
1010
Rafael J. Wysockib7eaf1a2017-03-22 00:08:50 +01001011/**
1012 * tick_nohz_get_idle_calls - return the current idle calls counter value
1013 *
1014 * Called from the schedutil frequency scaling governor in scheduler context.
1015 */
1016unsigned long tick_nohz_get_idle_calls(void)
1017{
1018 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1019
1020 return ts->idle_calls;
1021}
1022
Frederic Weisbecker2ac0d982011-07-28 04:00:47 +02001023static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
1024{
Frederic Weisbecker3f4724e2012-07-16 18:00:34 +02001025#ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
Frederic Weisbecker2ac0d982011-07-28 04:00:47 +02001026 unsigned long ticks;
Frederic Weisbecker3f4724e2012-07-16 18:00:34 +02001027
Frederic Weisbecker55dbdcf2015-11-19 16:47:32 +01001028 if (vtime_accounting_cpu_enabled())
Frederic Weisbecker3f4724e2012-07-16 18:00:34 +02001029 return;
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001030 /*
1031 * We stopped the tick in idle. Update process times would miss the
1032 * time we slept as update_process_times does only a 1 tick
1033 * accounting. Enforce that this is accounted to idle !
1034 */
1035 ticks = jiffies - ts->idle_jiffies;
1036 /*
1037 * We might be one off. Do not randomly account a huge number of ticks!
1038 */
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01001039 if (ticks && ticks < LONG_MAX)
1040 account_idle_ticks(ticks);
1041#endif
Frederic Weisbecker19f5f732011-07-27 17:29:28 +02001042}
1043
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001044/**
1045 * tick_nohz_idle_exit - restart the idle tick from the idle task
1046 *
1047 * Restart the idle tick when the CPU is woken up from idle
1048 * This also exit the RCU extended quiescent state. The CPU
1049 * can use RCU again after this function is called.
1050 */
1051void tick_nohz_idle_exit(void)
1052{
Christoph Lameter4a32fea2014-08-17 12:30:27 -05001053 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001054 ktime_t now;
1055
1056 local_irq_disable();
1057
1058 WARN_ON_ONCE(!ts->inidle);
1059
1060 ts->inidle = 0;
1061
1062 if (ts->idle_active || ts->tick_stopped)
1063 now = ktime_get();
1064
1065 if (ts->idle_active)
Frederic Weisbeckere8fcaa52013-08-07 22:28:01 +02001066 tick_nohz_stop_idle(ts, now);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001067
Frederic Weisbecker2ac0d982011-07-28 04:00:47 +02001068 if (ts->tick_stopped) {
Frederic Weisbecker1f419062016-04-13 15:56:51 +02001069 tick_nohz_restart_sched_tick(ts, now);
Frederic Weisbecker2ac0d982011-07-28 04:00:47 +02001070 tick_nohz_account_idle_ticks(ts);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001071 }
1072
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001073 local_irq_enable();
1074}
1075
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001076/*
1077 * The nohz low res interrupt handler
1078 */
1079static void tick_nohz_handler(struct clock_event_device *dev)
1080{
Christoph Lameter22127e92014-08-17 12:30:25 -05001081 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001082 struct pt_regs *regs = get_irq_regs();
1083 ktime_t now = ktime_get();
1084
Thomas Gleixner2456e852016-12-25 11:38:40 +01001085 dev->next_event = KTIME_MAX;
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001086
Frederic Weisbecker5bb96222012-10-15 02:03:27 +02001087 tick_sched_do_timer(now);
Frederic Weisbecker9e8f5592012-10-15 02:43:03 +02001088 tick_sched_handle(ts, regs);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001089
Viresh Kumarb5e995e2014-06-12 16:24:41 +05301090 /* No need to reprogram if we are running tickless */
1091 if (unlikely(ts->tick_stopped))
1092 return;
1093
Thomas Gleixner0ff53d02015-04-14 21:08:54 +00001094 hrtimer_forward(&ts->sched_timer, now, tick_period);
1095 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001096}
1097
Thomas Gleixnerbc7a34b2015-05-26 22:50:33 +00001098static inline void tick_nohz_activate(struct tick_sched *ts, int mode)
1099{
1100 if (!tick_nohz_enabled)
1101 return;
1102 ts->nohz_mode = mode;
1103 /* One update is enough */
1104 if (!test_and_set_bit(0, &tick_nohz_active))
Thomas Gleixner683be132015-05-26 22:50:35 +00001105 timers_update_migration(true);
Thomas Gleixnerbc7a34b2015-05-26 22:50:33 +00001106}
1107
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001108/**
1109 * tick_nohz_switch_to_nohz - switch to nohz mode
1110 */
1111static void tick_nohz_switch_to_nohz(void)
1112{
Christoph Lameter22127e92014-08-17 12:30:25 -05001113 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001114 ktime_t next;
1115
Viresh Kumar27630532014-04-15 10:54:41 +05301116 if (!tick_nohz_enabled)
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001117 return;
1118
Thomas Gleixner6b442bc2015-05-07 14:35:59 +02001119 if (tick_switch_to_oneshot(tick_nohz_handler))
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001120 return;
Thomas Gleixner6b442bc2015-05-07 14:35:59 +02001121
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001122 /*
1123 * Recycle the hrtimer in ts, so we can share the
1124 * hrtimer_forward with the highres code.
1125 */
1126 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1127 /* Get the next period */
1128 next = tick_init_jiffy_update();
1129
Thomas Gleixner0ff53d02015-04-14 21:08:54 +00001130 hrtimer_set_expires(&ts->sched_timer, next);
Wanpeng Li1ca8ec52016-01-27 19:26:07 +08001131 hrtimer_forward_now(&ts->sched_timer, tick_period);
1132 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
Thomas Gleixnerbc7a34b2015-05-26 22:50:33 +00001133 tick_nohz_activate(ts, NOHZ_MODE_LOWRES);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001134}
1135
Frederic Weisbecker5acac1b2013-12-04 18:28:20 +01001136static inline void tick_nohz_irq_enter(void)
Martin Schwidefskyeed3b9c2009-09-29 14:25:15 +02001137{
Christoph Lameter4a32fea2014-08-17 12:30:27 -05001138 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
Martin Schwidefskyeed3b9c2009-09-29 14:25:15 +02001139 ktime_t now;
1140
1141 if (!ts->idle_active && !ts->tick_stopped)
1142 return;
1143 now = ktime_get();
1144 if (ts->idle_active)
Frederic Weisbeckere8fcaa52013-08-07 22:28:01 +02001145 tick_nohz_stop_idle(ts, now);
Thomas Gleixnerff006732016-07-04 09:50:35 +00001146 if (ts->tick_stopped)
Martin Schwidefskyeed3b9c2009-09-29 14:25:15 +02001147 tick_nohz_update_jiffies(now);
Martin Schwidefskyeed3b9c2009-09-29 14:25:15 +02001148}
1149
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001150#else
1151
1152static inline void tick_nohz_switch_to_nohz(void) { }
Frederic Weisbecker5acac1b2013-12-04 18:28:20 +01001153static inline void tick_nohz_irq_enter(void) { }
Thomas Gleixnerbc7a34b2015-05-26 22:50:33 +00001154static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { }
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001155
Frederic Weisbecker3451d022011-08-10 23:21:01 +02001156#endif /* CONFIG_NO_HZ_COMMON */
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001157
1158/*
Thomas Gleixner719254fa2008-10-17 09:59:47 +02001159 * Called from irq_enter to notify about the possible interruption of idle()
1160 */
Frederic Weisbecker5acac1b2013-12-04 18:28:20 +01001161void tick_irq_enter(void)
Thomas Gleixner719254fa2008-10-17 09:59:47 +02001162{
Frederic Weisbeckere8fcaa52013-08-07 22:28:01 +02001163 tick_check_oneshot_broadcast_this_cpu();
Frederic Weisbecker5acac1b2013-12-04 18:28:20 +01001164 tick_nohz_irq_enter();
Thomas Gleixner719254fa2008-10-17 09:59:47 +02001165}
1166
1167/*
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001168 * High resolution timer specific code
1169 */
1170#ifdef CONFIG_HIGH_RES_TIMERS
1171/*
Pavel Machek4c9dc642008-01-30 13:30:00 +01001172 * We rearm the timer until we get disabled by the idle code.
Chuansheng Liu351f1812012-10-25 01:07:35 +08001173 * Called with interrupts disabled.
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001174 */
1175static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1176{
1177 struct tick_sched *ts =
1178 container_of(timer, struct tick_sched, sched_timer);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001179 struct pt_regs *regs = get_irq_regs();
1180 ktime_t now = ktime_get();
Thomas Gleixnerd3ed7822007-05-08 00:30:03 -07001181
Frederic Weisbecker5bb96222012-10-15 02:03:27 +02001182 tick_sched_do_timer(now);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001183
1184 /*
1185 * Do not call, when we are not in irq context and have
1186 * no valid regs pointer
1187 */
Frederic Weisbecker9e8f5592012-10-15 02:43:03 +02001188 if (regs)
1189 tick_sched_handle(ts, regs);
Frederic Weisbecker7c259042017-05-15 14:56:50 +02001190 else
1191 ts->next_tick = 0;
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001192
Viresh Kumar2a16fc92014-06-12 16:24:41 +05301193 /* No need to reprogram if we are in idle or full dynticks mode */
1194 if (unlikely(ts->tick_stopped))
1195 return HRTIMER_NORESTART;
1196
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001197 hrtimer_forward(timer, now, tick_period);
1198
1199 return HRTIMER_RESTART;
1200}
1201
Mike Galbraith5307c952012-05-08 12:20:58 +02001202static int sched_skew_tick;
1203
Thomas Gleixner62cf20b2012-05-25 14:08:57 +02001204static int __init skew_tick(char *str)
1205{
1206 get_option(&str, &sched_skew_tick);
1207
1208 return 0;
1209}
1210early_param("skew_tick", skew_tick);
1211
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001212/**
1213 * tick_setup_sched_timer - setup the tick emulation timer
1214 */
1215void tick_setup_sched_timer(void)
1216{
Christoph Lameter22127e92014-08-17 12:30:25 -05001217 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001218 ktime_t now = ktime_get();
1219
1220 /*
1221 * Emulate tick processing via per-CPU hrtimers:
1222 */
1223 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1224 ts->sched_timer.function = tick_sched_timer;
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001225
Ingo Molnar0de76112016-07-01 12:42:35 +02001226 /* Get the next period (per-CPU) */
Arjan van de Vencc584b22008-09-01 15:02:30 -07001227 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001228
Thomas Gleixner9c3f9e22012-11-21 20:31:52 +01001229 /* Offset the tick to avert jiffies_lock contention. */
Mike Galbraith5307c952012-05-08 12:20:58 +02001230 if (sched_skew_tick) {
1231 u64 offset = ktime_to_ns(tick_period) >> 1;
1232 do_div(offset, num_possible_cpus());
1233 offset *= smp_processor_id();
1234 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1235 }
1236
Thomas Gleixnerafc08b12015-04-14 21:08:52 +00001237 hrtimer_forward(&ts->sched_timer, now, tick_period);
1238 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
Thomas Gleixnerbc7a34b2015-05-26 22:50:33 +00001239 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001240}
Miao Xie3c4fbe52008-08-20 16:37:38 -07001241#endif /* HIGH_RES_TIMERS */
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001242
Frederic Weisbecker3451d022011-08-10 23:21:01 +02001243#if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001244void tick_cancel_sched_timer(int cpu)
1245{
1246 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1247
Miao Xie3c4fbe52008-08-20 16:37:38 -07001248# ifdef CONFIG_HIGH_RES_TIMERS
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001249 if (ts->sched_timer.base)
1250 hrtimer_cancel(&ts->sched_timer);
Miao Xie3c4fbe52008-08-20 16:37:38 -07001251# endif
Karsten Wiesea7901762008-03-04 14:59:55 -08001252
Thomas Gleixner4b0c0f22013-05-03 15:02:50 +02001253 memset(ts, 0, sizeof(*ts));
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001254}
Miao Xie3c4fbe52008-08-20 16:37:38 -07001255#endif
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001256
1257/**
1258 * Async notification about clocksource changes
1259 */
1260void tick_clock_notify(void)
1261{
1262 int cpu;
1263
1264 for_each_possible_cpu(cpu)
1265 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1266}
1267
1268/*
1269 * Async notification about clock event changes
1270 */
1271void tick_oneshot_notify(void)
1272{
Christoph Lameter22127e92014-08-17 12:30:25 -05001273 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001274
1275 set_bit(0, &ts->check_clocks);
1276}
1277
1278/**
1279 * Check, if a change happened, which makes oneshot possible.
1280 *
1281 * Called cyclic from the hrtimer softirq (driven by the timer
1282 * softirq) allow_nohz signals, that we can switch into low-res nohz
1283 * mode, because high resolution timers are disabled (either compile
Thomas Gleixner6b442bc2015-05-07 14:35:59 +02001284 * or runtime). Called with interrupts disabled.
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001285 */
1286int tick_check_oneshot_change(int allow_nohz)
1287{
Christoph Lameter22127e92014-08-17 12:30:25 -05001288 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001289
1290 if (!test_and_clear_bit(0, &ts->check_clocks))
1291 return 0;
1292
1293 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1294 return 0;
1295
Li Zefancf4fc6c2008-02-08 04:19:24 -08001296 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001297 return 0;
1298
1299 if (!allow_nohz)
1300 return 1;
1301
1302 tick_nohz_switch_to_nohz();
1303 return 0;
1304}