blob: eb4af016ac65ff4fe6cf07b793ff8a0f77f6c074 [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>
20#include <linux/profile.h>
21#include <linux/sched.h>
venkatesh.pallipadi@intel.com8083e4a2008-08-04 11:59:11 -070022#include <linux/module.h>
Frederic Weisbecker00b42952012-11-07 21:03:07 +010023#include <linux/irq_work.h>
Frederic Weisbecker9014c452013-04-20 15:43:57 +020024#include <linux/posix-timers.h>
25#include <linux/perf_event.h>
Frederic Weisbecker2e709332013-07-10 00:55:25 +020026#include <linux/context_tracking.h>
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -080027
David S. Miller9e203bc2007-02-24 22:10:13 -080028#include <asm/irq_regs.h>
29
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -080030#include "tick-internal.h"
31
Frederic Weisbeckercb41a292013-04-20 17:35:50 +020032#include <trace/events/timer.h>
33
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -080034/*
35 * Per cpu nohz control structure
36 */
Frederic Weisbecker33a5f622012-10-11 17:52:56 +020037DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -080038
39/*
John Stultzd6ad4182012-02-28 16:50:11 -080040 * The time, when the last jiffy update happened. Protected by jiffies_lock.
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -080041 */
42static ktime_t last_jiffies_update;
43
Ingo Molnar289f4802007-02-16 01:28:15 -080044struct tick_sched *tick_get_tick_sched(int cpu)
45{
46 return &per_cpu(tick_cpu_sched, cpu);
47}
48
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -080049/*
50 * Must be called with interrupts disabled !
51 */
52static void tick_do_update_jiffies64(ktime_t now)
53{
54 unsigned long ticks = 0;
55 ktime_t delta;
56
Ingo Molnar7a14ce12008-05-12 15:43:53 +020057 /*
John Stultzd6ad4182012-02-28 16:50:11 -080058 * Do a quick check without holding jiffies_lock:
Ingo Molnar7a14ce12008-05-12 15:43:53 +020059 */
60 delta = ktime_sub(now, last_jiffies_update);
61 if (delta.tv64 < tick_period.tv64)
62 return;
63
John Stultzd6ad4182012-02-28 16:50:11 -080064 /* Reevalute with jiffies_lock held */
65 write_seqlock(&jiffies_lock);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -080066
67 delta = ktime_sub(now, last_jiffies_update);
68 if (delta.tv64 >= tick_period.tv64) {
69
70 delta = ktime_sub(delta, tick_period);
71 last_jiffies_update = ktime_add(last_jiffies_update,
72 tick_period);
73
74 /* Slow path for long timeouts */
75 if (unlikely(delta.tv64 >= tick_period.tv64)) {
76 s64 incr = ktime_to_ns(tick_period);
77
78 ticks = ktime_divns(delta, incr);
79
80 last_jiffies_update = ktime_add_ns(last_jiffies_update,
81 incr * ticks);
82 }
83 do_timer(++ticks);
Thomas Gleixner49d670f2008-09-22 18:56:01 +020084
85 /* Keep the tick_next_period variable up to date */
86 tick_next_period = ktime_add(last_jiffies_update, tick_period);
Viresh Kumar03e6bdc2014-04-15 10:54:40 +053087 } else {
88 write_sequnlock(&jiffies_lock);
89 return;
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -080090 }
John Stultzd6ad4182012-02-28 16:50:11 -080091 write_sequnlock(&jiffies_lock);
John Stultz47a1b7962013-12-12 13:10:55 -080092 update_wall_time();
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -080093}
94
95/*
96 * Initialize and return retrieve the jiffies update.
97 */
98static ktime_t tick_init_jiffy_update(void)
99{
100 ktime_t period;
101
John Stultzd6ad4182012-02-28 16:50:11 -0800102 write_seqlock(&jiffies_lock);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800103 /* Did we start the jiffies update yet ? */
104 if (last_jiffies_update.tv64 == 0)
105 last_jiffies_update = tick_next_period;
106 period = last_jiffies_update;
John Stultzd6ad4182012-02-28 16:50:11 -0800107 write_sequnlock(&jiffies_lock);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800108 return period;
109}
110
Frederic Weisbecker5bb96222012-10-15 02:03:27 +0200111
112static void tick_sched_do_timer(ktime_t now)
113{
114 int cpu = smp_processor_id();
115
Frederic Weisbecker3451d022011-08-10 23:21:01 +0200116#ifdef CONFIG_NO_HZ_COMMON
Frederic Weisbecker5bb96222012-10-15 02:03:27 +0200117 /*
118 * Check if the do_timer duty was dropped. We don't care about
119 * concurrency: This happens only when the cpu in charge went
120 * into a long sleep. If two cpus happen to assign themself to
121 * this duty, then the jiffies update is still serialized by
Thomas Gleixner9c3f9e22012-11-21 20:31:52 +0100122 * jiffies_lock.
Frederic Weisbecker5bb96222012-10-15 02:03:27 +0200123 */
Frederic Weisbeckera382bf92012-12-18 18:24:35 +0100124 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)
Frederic Weisbeckerc5bfece2013-04-12 16:45:34 +0200125 && !tick_nohz_full_cpu(cpu))
Frederic Weisbecker5bb96222012-10-15 02:03:27 +0200126 tick_do_timer_cpu = cpu;
127#endif
128
129 /* Check, if the jiffies need an update */
130 if (tick_do_timer_cpu == cpu)
131 tick_do_update_jiffies64(now);
132}
133
Frederic Weisbecker9e8f5592012-10-15 02:43:03 +0200134static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
135{
Frederic Weisbecker3451d022011-08-10 23:21:01 +0200136#ifdef CONFIG_NO_HZ_COMMON
Frederic Weisbecker9e8f5592012-10-15 02:43:03 +0200137 /*
138 * When we are idle and the tick is stopped, we have to touch
139 * the watchdog as we might not schedule for a really long
140 * time. This happens on complete idle SMP systems while
141 * waiting on the login prompt. We also increment the "start of
142 * idle" jiffy stamp so the idle accounting adjustment we do
143 * when we go busy again does not account too much ticks.
144 */
145 if (ts->tick_stopped) {
146 touch_softlockup_watchdog();
147 if (is_idle_task(current))
148 ts->idle_jiffies++;
149 }
Frederic Weisbecker94a57142012-10-15 16:17:16 +0200150#endif
Frederic Weisbecker9e8f5592012-10-15 02:43:03 +0200151 update_process_times(user_mode(regs));
152 profile_tick(CPU_PROFILING);
153}
154
Frederic Weisbeckerc5bfece2013-04-12 16:45:34 +0200155#ifdef CONFIG_NO_HZ_FULL
Frederic Weisbecker460775df2013-07-24 23:52:27 +0200156cpumask_var_t tick_nohz_full_mask;
Paul E. McKenneyc0f489d2014-06-04 13:46:03 -0700157cpumask_var_t housekeeping_mask;
Frederic Weisbecker73867dc2013-07-24 23:31:00 +0200158bool tick_nohz_full_running;
Frederic Weisbeckera8318812012-12-18 17:32:19 +0100159
Frederic Weisbecker9014c452013-04-20 15:43:57 +0200160static bool can_stop_full_tick(void)
161{
162 WARN_ON_ONCE(!irqs_disabled());
163
Frederic Weisbeckercb41a292013-04-20 17:35:50 +0200164 if (!sched_can_stop_tick()) {
165 trace_tick_stop(0, "more than 1 task in runqueue\n");
Frederic Weisbecker9014c452013-04-20 15:43:57 +0200166 return false;
Frederic Weisbeckercb41a292013-04-20 17:35:50 +0200167 }
Frederic Weisbecker9014c452013-04-20 15:43:57 +0200168
Frederic Weisbeckercb41a292013-04-20 17:35:50 +0200169 if (!posix_cpu_timers_can_stop_tick(current)) {
170 trace_tick_stop(0, "posix timers running\n");
Frederic Weisbecker9014c452013-04-20 15:43:57 +0200171 return false;
Frederic Weisbeckercb41a292013-04-20 17:35:50 +0200172 }
Frederic Weisbecker9014c452013-04-20 15:43:57 +0200173
Frederic Weisbeckercb41a292013-04-20 17:35:50 +0200174 if (!perf_event_can_stop_tick()) {
175 trace_tick_stop(0, "perf events running\n");
Frederic Weisbecker9014c452013-04-20 15:43:57 +0200176 return false;
Frederic Weisbeckercb41a292013-04-20 17:35:50 +0200177 }
Frederic Weisbecker9014c452013-04-20 15:43:57 +0200178
179 /* sched_clock_tick() needs us? */
180#ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
181 /*
182 * TODO: kick full dynticks CPUs when
183 * sched_clock_stable is set.
184 */
Peter Zijlstra35af99e2013-11-28 19:38:42 +0100185 if (!sched_clock_stable()) {
Frederic Weisbeckercb41a292013-04-20 17:35:50 +0200186 trace_tick_stop(0, "unstable sched clock\n");
Steven Rostedte12d0272013-05-10 17:12:28 -0400187 /*
188 * Don't allow the user to think they can get
189 * full NO_HZ with this machine.
190 */
Frederic Weisbecker73867dc2013-07-24 23:31:00 +0200191 WARN_ONCE(tick_nohz_full_running,
Steven Rostedt543487c2013-07-16 10:22:12 -0400192 "NO_HZ FULL will not work with unstable sched clock");
Frederic Weisbecker9014c452013-04-20 15:43:57 +0200193 return false;
Frederic Weisbeckercb41a292013-04-20 17:35:50 +0200194 }
Frederic Weisbecker9014c452013-04-20 15:43:57 +0200195#endif
196
197 return true;
198}
199
200static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now);
201
Frederic Weisbecker76c24fb2013-04-18 00:15:40 +0200202/*
203 * Re-evaluate the need for the tick on the current CPU
204 * and restart it if necessary.
205 */
Frederic Weisbeckerd13508f2013-07-24 23:52:27 +0200206void __tick_nohz_full_check(void)
Frederic Weisbecker76c24fb2013-04-18 00:15:40 +0200207{
Frederic Weisbecker9014c452013-04-20 15:43:57 +0200208 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
209
210 if (tick_nohz_full_cpu(smp_processor_id())) {
211 if (ts->tick_stopped && !is_idle_task(current)) {
212 if (!can_stop_full_tick())
213 tick_nohz_restart_sched_tick(ts, ktime_get());
214 }
215 }
Frederic Weisbecker76c24fb2013-04-18 00:15:40 +0200216}
217
218static void nohz_full_kick_work_func(struct irq_work *work)
219{
Frederic Weisbeckerd13508f2013-07-24 23:52:27 +0200220 __tick_nohz_full_check();
Frederic Weisbecker76c24fb2013-04-18 00:15:40 +0200221}
222
223static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = {
224 .func = nohz_full_kick_work_func,
225};
226
227/*
Frederic Weisbecker40bea032014-08-13 18:50:16 +0200228 * Kick this CPU if it's full dynticks in order to force it to
229 * re-evaluate its dependency on the tick and restart it if necessary.
230 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
231 * is NMI safe.
232 */
233void tick_nohz_full_kick(void)
234{
235 if (!tick_nohz_full_cpu(smp_processor_id()))
236 return;
237
238 irq_work_queue(&__get_cpu_var(nohz_full_kick_work));
239}
240
241/*
Frederic Weisbecker3d36aeb2014-06-04 16:17:33 +0200242 * Kick the CPU if it's full dynticks in order to force it to
Frederic Weisbecker76c24fb2013-04-18 00:15:40 +0200243 * re-evaluate its dependency on the tick and restart it if necessary.
244 */
Frederic Weisbecker3d36aeb2014-06-04 16:17:33 +0200245void tick_nohz_full_kick_cpu(int cpu)
Frederic Weisbecker76c24fb2013-04-18 00:15:40 +0200246{
Frederic Weisbecker3d36aeb2014-06-04 16:17:33 +0200247 if (!tick_nohz_full_cpu(cpu))
248 return;
249
250 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
Frederic Weisbecker76c24fb2013-04-18 00:15:40 +0200251}
252
253static void nohz_full_kick_ipi(void *info)
254{
Frederic Weisbeckerd13508f2013-07-24 23:52:27 +0200255 __tick_nohz_full_check();
Frederic Weisbecker76c24fb2013-04-18 00:15:40 +0200256}
257
258/*
259 * Kick all full dynticks CPUs in order to force these to re-evaluate
260 * their dependency on the tick and restart it if necessary.
261 */
262void tick_nohz_full_kick_all(void)
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 Weisbecker73867dc2013-07-24 23:31:00 +0200268 smp_call_function_many(tick_nohz_full_mask,
Frederic Weisbecker76c24fb2013-04-18 00:15:40 +0200269 nohz_full_kick_ipi, NULL, false);
Frederic Weisbeckerc2e7fcf2013-08-02 18:29:56 +0200270 tick_nohz_full_kick();
Frederic Weisbecker76c24fb2013-04-18 00:15:40 +0200271 preempt_enable();
272}
273
Frederic Weisbecker99e5ada2013-04-20 17:11:50 +0200274/*
275 * Re-evaluate the need for the tick as we switch the current task.
276 * It might need the tick due to per task/process properties:
277 * perf events, posix cpu timers, ...
278 */
Frederic Weisbeckerd13508f2013-07-24 23:52:27 +0200279void __tick_nohz_task_switch(struct task_struct *tsk)
Frederic Weisbecker99e5ada2013-04-20 17:11:50 +0200280{
281 unsigned long flags;
282
Frederic Weisbecker99e5ada2013-04-20 17:11:50 +0200283 local_irq_save(flags);
284
Li Zhong6296ace2013-04-28 11:25:58 +0800285 if (!tick_nohz_full_cpu(smp_processor_id()))
286 goto out;
287
Frederic Weisbecker99e5ada2013-04-20 17:11:50 +0200288 if (tick_nohz_tick_stopped() && !can_stop_full_tick())
289 tick_nohz_full_kick();
290
Li Zhong6296ace2013-04-28 11:25:58 +0800291out:
Frederic Weisbecker99e5ada2013-04-20 17:11:50 +0200292 local_irq_restore(flags);
293}
294
Frederic Weisbeckera8318812012-12-18 17:32:19 +0100295/* Parse the boot-time nohz CPU list from the kernel parameters. */
Frederic Weisbeckerc5bfece2013-04-12 16:45:34 +0200296static int __init tick_nohz_full_setup(char *str)
Frederic Weisbeckera8318812012-12-18 17:32:19 +0100297{
Frederic Weisbecker73867dc2013-07-24 23:31:00 +0200298 alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
299 if (cpulist_parse(str, tick_nohz_full_mask) < 0) {
Frederic Weisbeckerc5bfece2013-04-12 16:45:34 +0200300 pr_warning("NOHZ: Incorrect nohz_full cpumask\n");
Frederic Weisbecker4327b152014-08-17 22:02:55 +0200301 free_bootmem_cpumask_var(tick_nohz_full_mask);
Frederic Weisbecker0453b432013-03-27 02:18:34 +0100302 return 1;
303 }
Frederic Weisbecker73867dc2013-07-24 23:31:00 +0200304 tick_nohz_full_running = true;
Frederic Weisbecker0453b432013-03-27 02:18:34 +0100305
Frederic Weisbeckera8318812012-12-18 17:32:19 +0100306 return 1;
307}
Frederic Weisbeckerc5bfece2013-04-12 16:45:34 +0200308__setup("nohz_full=", tick_nohz_full_setup);
Frederic Weisbeckera8318812012-12-18 17:32:19 +0100309
Paul Gortmaker0db06282013-06-19 14:53:51 -0400310static int tick_nohz_cpu_down_callback(struct notifier_block *nfb,
Frederic Weisbeckera382bf92012-12-18 18:24:35 +0100311 unsigned long action,
312 void *hcpu)
313{
314 unsigned int cpu = (unsigned long)hcpu;
315
316 switch (action & ~CPU_TASKS_FROZEN) {
317 case CPU_DOWN_PREPARE:
318 /*
319 * If we handle the timekeeping duty for full dynticks CPUs,
320 * we can't safely shutdown that CPU.
321 */
Frederic Weisbecker73867dc2013-07-24 23:31:00 +0200322 if (tick_nohz_full_running && tick_do_timer_cpu == cpu)
Li Zhong1a7f8292013-05-17 16:44:04 +0800323 return NOTIFY_BAD;
Frederic Weisbeckera382bf92012-12-18 18:24:35 +0100324 break;
325 }
326 return NOTIFY_OK;
327}
328
Frederic Weisbecker1034fc22013-03-26 15:04:50 +0100329/*
330 * Worst case string length in chunks of CPU range seems 2 steps
331 * separations: 0,2,4,6,...
332 * This is NR_CPUS + sizeof('\0')
333 */
Frederic Weisbeckerc5bfece2013-04-12 16:45:34 +0200334static char __initdata nohz_full_buf[NR_CPUS + 1];
Frederic Weisbecker1034fc22013-03-26 15:04:50 +0100335
Frederic Weisbeckerf98823a2013-03-27 01:17:22 +0100336static int tick_nohz_init_all(void)
337{
338 int err = -1;
339
340#ifdef CONFIG_NO_HZ_FULL_ALL
Frederic Weisbecker73867dc2013-07-24 23:31:00 +0200341 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) {
Frederic Weisbecker4327b152014-08-17 22:02:55 +0200342 WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n");
Paul E. McKenneyc0f489d2014-06-04 13:46:03 -0700343 return err;
344 }
Frederic Weisbeckerf98823a2013-03-27 01:17:22 +0100345 err = 0;
Frederic Weisbecker73867dc2013-07-24 23:31:00 +0200346 cpumask_setall(tick_nohz_full_mask);
Frederic Weisbecker73867dc2013-07-24 23:31:00 +0200347 tick_nohz_full_running = true;
Frederic Weisbeckerf98823a2013-03-27 01:17:22 +0100348#endif
349 return err;
350}
351
Frederic Weisbeckerd1e43fa2013-03-26 23:47:24 +0100352void __init tick_nohz_init(void)
Frederic Weisbeckera8318812012-12-18 17:32:19 +0100353{
Frederic Weisbeckerd1e43fa2013-03-26 23:47:24 +0100354 int cpu;
355
Frederic Weisbecker73867dc2013-07-24 23:31:00 +0200356 if (!tick_nohz_full_running) {
Frederic Weisbeckerf98823a2013-03-27 01:17:22 +0100357 if (tick_nohz_init_all() < 0)
358 return;
359 }
Frederic Weisbeckerd1e43fa2013-03-26 23:47:24 +0100360
Frederic Weisbecker4327b152014-08-17 22:02:55 +0200361 if (!alloc_cpumask_var(&housekeeping_mask, GFP_KERNEL)) {
362 WARN(1, "NO_HZ: Can't allocate not-full dynticks cpumask\n");
363 cpumask_clear(tick_nohz_full_mask);
364 tick_nohz_full_running = false;
365 return;
366 }
367
368 cpu = smp_processor_id();
369
370 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
371 pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu);
372 cpumask_clear_cpu(cpu, tick_nohz_full_mask);
373 }
374
375 cpumask_andnot(housekeeping_mask,
376 cpu_possible_mask, tick_nohz_full_mask);
377
Frederic Weisbecker73867dc2013-07-24 23:31:00 +0200378 for_each_cpu(cpu, tick_nohz_full_mask)
Frederic Weisbecker2e709332013-07-10 00:55:25 +0200379 context_tracking_cpu_set(cpu);
380
Frederic Weisbeckerd1e43fa2013-03-26 23:47:24 +0100381 cpu_notifier(tick_nohz_cpu_down_callback, 0);
Frederic Weisbecker73867dc2013-07-24 23:31:00 +0200382 cpulist_scnprintf(nohz_full_buf, sizeof(nohz_full_buf), tick_nohz_full_mask);
Frederic Weisbeckerc5bfece2013-04-12 16:45:34 +0200383 pr_info("NO_HZ: Full dynticks CPUs: %s.\n", nohz_full_buf);
Frederic Weisbeckera8318812012-12-18 17:32:19 +0100384}
Frederic Weisbeckera8318812012-12-18 17:32:19 +0100385#endif
386
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800387/*
388 * NOHZ - aka dynamic tick functionality
389 */
Frederic Weisbecker3451d022011-08-10 23:21:01 +0200390#ifdef CONFIG_NO_HZ_COMMON
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800391/*
392 * NO HZ enabled ?
393 */
Thomas Gleixnerd689fe22013-11-13 21:01:57 +0100394static int tick_nohz_enabled __read_mostly = 1;
395int tick_nohz_active __read_mostly;
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800396/*
397 * Enable / Disable tickless mode
398 */
399static int __init setup_tick_nohz(char *str)
400{
401 if (!strcmp(str, "off"))
402 tick_nohz_enabled = 0;
403 else if (!strcmp(str, "on"))
404 tick_nohz_enabled = 1;
405 else
406 return 0;
407 return 1;
408}
409
410__setup("nohz=", setup_tick_nohz);
411
412/**
413 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
414 *
415 * Called from interrupt entry when the CPU was idle
416 *
417 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
418 * must be updated. Otherwise an interrupt handler could use a stale jiffy
419 * value. We do this unconditionally on any cpu, as we don't know whether the
420 * cpu, which has the update task assigned is in a long sleep.
421 */
Martin Schwidefskyeed3b9c2009-09-29 14:25:15 +0200422static void tick_nohz_update_jiffies(ktime_t now)
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800423{
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800424 unsigned long flags;
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800425
Frederic Weisbeckere8fcaa52013-08-07 22:28:01 +0200426 __this_cpu_write(tick_cpu_sched.idle_waketime, now);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800427
428 local_irq_save(flags);
429 tick_do_update_jiffies64(now);
430 local_irq_restore(flags);
Ingo Molnar02ff3752008-05-12 15:43:53 +0200431
432 touch_softlockup_watchdog();
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800433}
434
Arjan van de Ven595aac42010-05-09 08:22:45 -0700435/*
436 * Updates the per cpu time idle statistics counters
437 */
Arjan van de Ven8d63bf92010-05-09 08:24:03 -0700438static void
Peter Zijlstra8c215bd2010-07-01 09:07:17 +0200439update_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 -0700440{
441 ktime_t delta;
442
Arjan van de Ven595aac42010-05-09 08:22:45 -0700443 if (ts->idle_active) {
444 delta = ktime_sub(now, ts->idle_entrytime);
Peter Zijlstra8c215bd2010-07-01 09:07:17 +0200445 if (nr_iowait_cpu(cpu) > 0)
Arjan van de Ven0224cf42010-05-09 08:25:23 -0700446 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
Michal Hocko6beea0c2011-08-24 09:37:48 +0200447 else
448 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
Arjan van de Ven8c7b09f2010-05-09 08:23:23 -0700449 ts->idle_entrytime = now;
Arjan van de Ven595aac42010-05-09 08:22:45 -0700450 }
Arjan van de Ven8d63bf92010-05-09 08:24:03 -0700451
Arjan van de Vene0e37c22010-05-09 08:24:39 -0700452 if (last_update_time)
Arjan van de Ven8d63bf92010-05-09 08:24:03 -0700453 *last_update_time = ktime_to_us(now);
454
Arjan van de Ven595aac42010-05-09 08:22:45 -0700455}
456
Frederic Weisbeckere8fcaa52013-08-07 22:28:01 +0200457static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
Venki Pallipadi6378ddb2008-01-30 13:30:04 +0100458{
Frederic Weisbeckere8fcaa52013-08-07 22:28:01 +0200459 update_ts_time_stats(smp_processor_id(), ts, now, NULL);
Martin Schwidefskyeed3b9c2009-09-29 14:25:15 +0200460 ts->idle_active = 0;
Peter Zijlstra56c74262008-09-01 16:44:23 +0200461
Martin Schwidefskyeed3b9c2009-09-29 14:25:15 +0200462 sched_clock_idle_wakeup_event(0);
Venki Pallipadi6378ddb2008-01-30 13:30:04 +0100463}
464
Frederic Weisbeckere8fcaa52013-08-07 22:28:01 +0200465static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
Venki Pallipadi6378ddb2008-01-30 13:30:04 +0100466{
Michal Hocko430ee882011-12-01 17:00:22 +0100467 ktime_t now = ktime_get();
Arjan van de Ven595aac42010-05-09 08:22:45 -0700468
Venki Pallipadi6378ddb2008-01-30 13:30:04 +0100469 ts->idle_entrytime = now;
470 ts->idle_active = 1;
Peter Zijlstra56c74262008-09-01 16:44:23 +0200471 sched_clock_idle_sleep_event();
Venki Pallipadi6378ddb2008-01-30 13:30:04 +0100472 return now;
473}
474
Arjan van de Venb1f724c2010-05-09 08:22:08 -0700475/**
476 * get_cpu_idle_time_us - get the total idle time of a cpu
477 * @cpu: CPU number to query
Michal Hocko09a1d342011-08-24 09:39:30 +0200478 * @last_update_time: variable to store update time in. Do not update
479 * counters if NULL.
Arjan van de Venb1f724c2010-05-09 08:22:08 -0700480 *
481 * Return the cummulative idle time (since boot) for a given
Michal Hocko6beea0c2011-08-24 09:37:48 +0200482 * CPU, in microseconds.
Arjan van de Venb1f724c2010-05-09 08:22:08 -0700483 *
484 * This time is measured via accounting rather than sampling,
485 * and is as accurate as ktime_get() is.
486 *
487 * This function returns -1 if NOHZ is not enabled.
488 */
Venki Pallipadi6378ddb2008-01-30 13:30:04 +0100489u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
490{
491 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
Michal Hocko09a1d342011-08-24 09:39:30 +0200492 ktime_t now, idle;
Venki Pallipadi6378ddb2008-01-30 13:30:04 +0100493
Thomas Gleixnerd689fe22013-11-13 21:01:57 +0100494 if (!tick_nohz_active)
venkatesh.pallipadi@intel.com8083e4a2008-08-04 11:59:11 -0700495 return -1;
496
Michal Hocko09a1d342011-08-24 09:39:30 +0200497 now = ktime_get();
498 if (last_update_time) {
499 update_ts_time_stats(cpu, ts, now, last_update_time);
500 idle = ts->idle_sleeptime;
501 } else {
502 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
503 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
venkatesh.pallipadi@intel.com8083e4a2008-08-04 11:59:11 -0700504
Michal Hocko09a1d342011-08-24 09:39:30 +0200505 idle = ktime_add(ts->idle_sleeptime, delta);
506 } else {
507 idle = ts->idle_sleeptime;
508 }
509 }
510
511 return ktime_to_us(idle);
512
Venki Pallipadi6378ddb2008-01-30 13:30:04 +0100513}
venkatesh.pallipadi@intel.com8083e4a2008-08-04 11:59:11 -0700514EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
Venki Pallipadi6378ddb2008-01-30 13:30:04 +0100515
Michal Hocko6beea0c2011-08-24 09:37:48 +0200516/**
Arjan van de Ven0224cf42010-05-09 08:25:23 -0700517 * get_cpu_iowait_time_us - get the total iowait time of a cpu
518 * @cpu: CPU number to query
Michal Hocko09a1d342011-08-24 09:39:30 +0200519 * @last_update_time: variable to store update time in. Do not update
520 * counters if NULL.
Arjan van de Ven0224cf42010-05-09 08:25:23 -0700521 *
522 * Return the cummulative iowait time (since boot) for a given
523 * CPU, in microseconds.
524 *
525 * This time is measured via accounting rather than sampling,
526 * and is as accurate as ktime_get() is.
527 *
528 * This function returns -1 if NOHZ is not enabled.
529 */
530u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
531{
532 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
Michal Hocko09a1d342011-08-24 09:39:30 +0200533 ktime_t now, iowait;
Arjan van de Ven0224cf42010-05-09 08:25:23 -0700534
Thomas Gleixnerd689fe22013-11-13 21:01:57 +0100535 if (!tick_nohz_active)
Arjan van de Ven0224cf42010-05-09 08:25:23 -0700536 return -1;
537
Michal Hocko09a1d342011-08-24 09:39:30 +0200538 now = ktime_get();
539 if (last_update_time) {
540 update_ts_time_stats(cpu, ts, now, last_update_time);
541 iowait = ts->iowait_sleeptime;
542 } else {
543 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
544 ktime_t delta = ktime_sub(now, ts->idle_entrytime);
Arjan van de Ven0224cf42010-05-09 08:25:23 -0700545
Michal Hocko09a1d342011-08-24 09:39:30 +0200546 iowait = ktime_add(ts->iowait_sleeptime, delta);
547 } else {
548 iowait = ts->iowait_sleeptime;
549 }
550 }
551
552 return ktime_to_us(iowait);
Arjan van de Ven0224cf42010-05-09 08:25:23 -0700553}
554EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
555
Frederic Weisbecker84bf1bc2011-08-01 01:25:38 +0200556static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
557 ktime_t now, int cpu)
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800558{
Frederic Weisbecker280f0672011-10-07 18:22:06 +0200559 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies;
Frederic Weisbecker84bf1bc2011-08-01 01:25:38 +0200560 ktime_t last_update, expires, ret = { .tv64 = 0 };
Paul E. McKenneyaa9b16302012-05-10 16:41:44 -0700561 unsigned long rcu_delta_jiffies;
Len Brown4f86d3a2007-10-03 18:58:00 -0400562 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
Jon Hunter98962462009-08-18 12:45:10 -0500563 u64 time_delta;
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800564
Frederic Weisbecker855a0fc2013-12-17 00:16:37 +0100565 time_delta = timekeeping_max_deferment();
566
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800567 /* Read jiffies and the time when jiffies were updated last */
568 do {
John Stultzd6ad4182012-02-28 16:50:11 -0800569 seq = read_seqbegin(&jiffies_lock);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800570 last_update = last_jiffies_update;
571 last_jiffies = jiffies;
John Stultzd6ad4182012-02-28 16:50:11 -0800572 } while (read_seqretry(&jiffies_lock, seq));
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800573
Frederic Weisbecker74876a92012-10-12 18:00:23 +0200574 if (rcu_needs_cpu(cpu, &rcu_delta_jiffies) ||
Frederic Weisbecker00b42952012-11-07 21:03:07 +0100575 arch_needs_cpu(cpu) || irq_work_needs_cpu()) {
Martin Schwidefsky3c5d92a2009-09-29 14:25:16 +0200576 next_jiffies = last_jiffies + 1;
Ingo Molnar6ba9b342007-02-19 18:11:56 +0000577 delta_jiffies = 1;
Martin Schwidefsky3c5d92a2009-09-29 14:25:16 +0200578 } else {
579 /* Get the next timer wheel timer */
580 next_jiffies = get_next_timer_interrupt(last_jiffies);
581 delta_jiffies = next_jiffies - last_jiffies;
Paul E. McKenneyaa9b16302012-05-10 16:41:44 -0700582 if (rcu_delta_jiffies < delta_jiffies) {
583 next_jiffies = last_jiffies + rcu_delta_jiffies;
584 delta_jiffies = rcu_delta_jiffies;
585 }
Martin Schwidefsky3c5d92a2009-09-29 14:25:16 +0200586 }
Ingo Molnar47aa8b62013-04-26 10:05:59 +0200587
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800588 /*
Ingo Molnar47aa8b62013-04-26 10:05:59 +0200589 * Do not stop the tick, if we are only one off (or less)
590 * or if the cpu is required for RCU:
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800591 */
Ingo Molnar47aa8b62013-04-26 10:05:59 +0200592 if (!ts->tick_stopped && delta_jiffies <= 1)
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800593 goto out;
594
595 /* Schedule the tick, if we are at least one jiffie off */
596 if ((long)delta_jiffies >= 1) {
597
Woodruff, Richard00147442008-12-01 14:18:11 -0800598 /*
Woodruff, Richard00147442008-12-01 14:18:11 -0800599 * If this cpu is the one which updates jiffies, then
600 * give up the assignment and let it be taken by the
601 * cpu which runs the tick timer next, which might be
602 * this cpu as well. If we don't drop this here the
603 * jiffies might be stale and do_timer() never
Thomas Gleixner27185012009-11-12 22:12:06 +0100604 * invoked. Keep track of the fact that it was the one
605 * which had the do_timer() duty last. If this cpu is
606 * the one which had the do_timer() duty last, we
607 * limit the sleep time to the timekeeping
608 * max_deferement value which we retrieved
609 * above. Otherwise we can sleep as long as we want.
Woodruff, Richard00147442008-12-01 14:18:11 -0800610 */
Thomas Gleixner27185012009-11-12 22:12:06 +0100611 if (cpu == tick_do_timer_cpu) {
Woodruff, Richard00147442008-12-01 14:18:11 -0800612 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
Thomas Gleixner27185012009-11-12 22:12:06 +0100613 ts->do_timer_last = 1;
614 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
615 time_delta = KTIME_MAX;
616 ts->do_timer_last = 0;
617 } else if (!ts->do_timer_last) {
618 time_delta = KTIME_MAX;
619 }
620
Frederic Weisbecker265f22a2013-05-03 03:39:05 +0200621#ifdef CONFIG_NO_HZ_FULL
622 if (!ts->inidle) {
623 time_delta = min(time_delta,
624 scheduler_tick_max_deferment());
625 }
626#endif
627
Thomas Gleixner27185012009-11-12 22:12:06 +0100628 /*
Jon Hunter98962462009-08-18 12:45:10 -0500629 * calculate the expiry time for the next timer wheel
630 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
631 * that there is no timer pending or at least extremely
632 * far into the future (12 days for HZ=1000). In this
633 * case we set the expiry to the end of time.
634 */
635 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) {
636 /*
637 * Calculate the time delta for the next timer event.
638 * If the time delta exceeds the maximum time delta
639 * permitted by the current clocksource then adjust
640 * the time delta accordingly to ensure the
641 * clocksource does not wrap.
642 */
643 time_delta = min_t(u64, time_delta,
644 tick_period.tv64 * delta_jiffies);
Jon Hunter98962462009-08-18 12:45:10 -0500645 }
Woodruff, Richard00147442008-12-01 14:18:11 -0800646
Thomas Gleixner27185012009-11-12 22:12:06 +0100647 if (time_delta < KTIME_MAX)
648 expires = ktime_add_ns(last_update, time_delta);
649 else
650 expires.tv64 = KTIME_MAX;
Woodruff, Richard00147442008-12-01 14:18:11 -0800651
Woodruff, Richard00147442008-12-01 14:18:11 -0800652 /* Skip reprogram of event if its not changed */
653 if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
654 goto out;
655
Frederic Weisbecker84bf1bc2011-08-01 01:25:38 +0200656 ret = expires;
657
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800658 /*
659 * nohz_stop_sched_tick can be called several times before
660 * the nohz_restart_sched_tick is called. This happens when
661 * interrupts arrive which do not cause a reschedule. In the
662 * first call we save the current tick time, so we can restart
663 * the scheduler tick in nohz_restart_sched_tick.
664 */
665 if (!ts->tick_stopped) {
Alex Shic1cc0172012-09-10 15:10:58 +0800666 nohz_balance_enter_idle(cpu);
Peter Zijlstra5167e8d2012-06-22 15:52:09 +0200667 calc_load_enter_idle();
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -0700668
Frederic Weisbeckerf5d411c2011-07-31 17:44:12 +0200669 ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800670 ts->tick_stopped = 1;
Frederic Weisbeckercb41a292013-04-20 17:35:50 +0200671 trace_tick_stop(1, " ");
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800672 }
Thomas Gleixnerd3ed7822007-05-08 00:30:03 -0700673
Thomas Gleixnereaad0842007-05-29 23:47:39 +0200674 /*
Jon Hunter98962462009-08-18 12:45:10 -0500675 * If the expiration time == KTIME_MAX, then
676 * in this case we simply stop the tick timer.
Thomas Gleixnereaad0842007-05-29 23:47:39 +0200677 */
Jon Hunter98962462009-08-18 12:45:10 -0500678 if (unlikely(expires.tv64 == KTIME_MAX)) {
Thomas Gleixnereaad0842007-05-29 23:47:39 +0200679 if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
680 hrtimer_cancel(&ts->sched_timer);
681 goto out;
682 }
683
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800684 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
685 hrtimer_start(&ts->sched_timer, expires,
Arun R Bharadwaj5c333862009-04-16 12:14:37 +0530686 HRTIMER_MODE_ABS_PINNED);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800687 /* Check, if the timer was already in the past */
688 if (hrtimer_active(&ts->sched_timer))
689 goto out;
Pavel Machek4c9dc642008-01-30 13:30:00 +0100690 } else if (!tick_program_event(expires, 0))
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800691 goto out;
692 /*
693 * We are past the event already. So we crossed a
694 * jiffie boundary. Update jiffies and raise the
695 * softirq.
696 */
697 tick_do_update_jiffies64(ktime_get());
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800698 }
699 raise_softirq_irqoff(TIMER_SOFTIRQ);
700out:
701 ts->next_jiffies = next_jiffies;
702 ts->last_jiffies = last_jiffies;
Len Brown4f86d3a2007-10-03 18:58:00 -0400703 ts->sleep_length = ktime_sub(dev->next_event, now);
Frederic Weisbecker84bf1bc2011-08-01 01:25:38 +0200704
705 return ret;
Frederic Weisbecker280f0672011-10-07 18:22:06 +0200706}
707
Frederic Weisbecker5811d992013-04-20 16:40:31 +0200708static void tick_nohz_full_stop_tick(struct tick_sched *ts)
709{
710#ifdef CONFIG_NO_HZ_FULL
Alex Shie9a2eb42013-11-28 14:27:11 +0800711 int cpu = smp_processor_id();
Frederic Weisbecker5811d992013-04-20 16:40:31 +0200712
Alex Shie9a2eb42013-11-28 14:27:11 +0800713 if (!tick_nohz_full_cpu(cpu) || is_idle_task(current))
714 return;
Frederic Weisbecker5811d992013-04-20 16:40:31 +0200715
Alex Shie9a2eb42013-11-28 14:27:11 +0800716 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE)
717 return;
Frederic Weisbecker5811d992013-04-20 16:40:31 +0200718
Alex Shie9a2eb42013-11-28 14:27:11 +0800719 if (!can_stop_full_tick())
720 return;
Frederic Weisbecker5811d992013-04-20 16:40:31 +0200721
Alex Shie9a2eb42013-11-28 14:27:11 +0800722 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu);
Frederic Weisbecker5811d992013-04-20 16:40:31 +0200723#endif
724}
725
Frederic Weisbecker5b399392011-08-01 00:06:10 +0200726static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
727{
728 /*
729 * If this cpu is offline and it is the one which updates
730 * jiffies, then give up the assignment and let it be taken by
731 * the cpu which runs the tick timer next. If we don't drop
732 * this here the jiffies might be stale and do_timer() never
733 * invoked.
734 */
735 if (unlikely(!cpu_online(cpu))) {
736 if (cpu == tick_do_timer_cpu)
737 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
Thomas Gleixnerf7ea0fd2013-05-13 21:40:27 +0200738 return false;
Frederic Weisbecker5b399392011-08-01 00:06:10 +0200739 }
740
Thomas Gleixner0e576acb2013-11-29 12:18:13 +0100741 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) {
742 ts->sleep_length = (ktime_t) { .tv64 = NSEC_PER_SEC/HZ };
Frederic Weisbecker5b399392011-08-01 00:06:10 +0200743 return false;
Thomas Gleixner0e576acb2013-11-29 12:18:13 +0100744 }
Frederic Weisbecker5b399392011-08-01 00:06:10 +0200745
746 if (need_resched())
747 return false;
748
749 if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
750 static int ratelimit;
751
Paul E. McKenney803b0eb2012-08-23 08:34:07 -0700752 if (ratelimit < 10 &&
753 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
Rado Vrbovskycfea7d72013-02-08 12:37:30 -0500754 pr_warn("NOHZ: local_softirq_pending %02x\n",
755 (unsigned int) local_softirq_pending());
Frederic Weisbecker5b399392011-08-01 00:06:10 +0200756 ratelimit++;
757 }
758 return false;
759 }
760
Frederic Weisbecker460775df2013-07-24 23:52:27 +0200761 if (tick_nohz_full_enabled()) {
Frederic Weisbeckera382bf92012-12-18 18:24:35 +0100762 /*
763 * Keep the tick alive to guarantee timekeeping progression
764 * if there are full dynticks CPUs around
765 */
766 if (tick_do_timer_cpu == cpu)
767 return false;
768 /*
769 * Boot safety: make sure the timekeeping duty has been
770 * assigned before entering dyntick-idle mode,
771 */
772 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
773 return false;
774 }
775
Frederic Weisbecker5b399392011-08-01 00:06:10 +0200776 return true;
777}
778
Frederic Weisbecker19f5f732011-07-27 17:29:28 +0200779static void __tick_nohz_idle_enter(struct tick_sched *ts)
780{
Frederic Weisbecker84bf1bc2011-08-01 01:25:38 +0200781 ktime_t now, expires;
Frederic Weisbecker5b399392011-08-01 00:06:10 +0200782 int cpu = smp_processor_id();
Frederic Weisbecker19f5f732011-07-27 17:29:28 +0200783
Frederic Weisbeckere8fcaa52013-08-07 22:28:01 +0200784 now = tick_nohz_start_idle(ts);
Frederic Weisbecker2ac0d982011-07-28 04:00:47 +0200785
Frederic Weisbecker5b399392011-08-01 00:06:10 +0200786 if (can_stop_idle_tick(cpu, ts)) {
787 int was_stopped = ts->tick_stopped;
788
789 ts->idle_calls++;
Frederic Weisbecker84bf1bc2011-08-01 01:25:38 +0200790
791 expires = tick_nohz_stop_sched_tick(ts, now, cpu);
792 if (expires.tv64 > 0LL) {
793 ts->idle_sleeps++;
794 ts->idle_expires = expires;
795 }
Frederic Weisbecker5b399392011-08-01 00:06:10 +0200796
797 if (!was_stopped && ts->tick_stopped)
798 ts->idle_jiffies = ts->last_jiffies;
799 }
Frederic Weisbecker280f0672011-10-07 18:22:06 +0200800}
801
802/**
803 * tick_nohz_idle_enter - stop the idle tick from the idle task
804 *
805 * When the next event is more than a tick into the future, stop the idle tick
806 * Called when we start the idle loop.
Frederic Weisbecker2bbb6812011-10-08 16:01:00 +0200807 *
Frederic Weisbecker1268fbc2011-11-17 18:48:14 +0100808 * The arch is responsible of calling:
Frederic Weisbecker2bbb6812011-10-08 16:01:00 +0200809 *
810 * - rcu_idle_enter() after its last use of RCU before the CPU is put
811 * to sleep.
812 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
Frederic Weisbecker280f0672011-10-07 18:22:06 +0200813 */
Frederic Weisbecker1268fbc2011-11-17 18:48:14 +0100814void tick_nohz_idle_enter(void)
Frederic Weisbecker280f0672011-10-07 18:22:06 +0200815{
816 struct tick_sched *ts;
817
Frederic Weisbecker1268fbc2011-11-17 18:48:14 +0100818 WARN_ON_ONCE(irqs_disabled());
819
Linus Torvalds0db49b72012-01-06 08:33:28 -0800820 /*
821 * Update the idle state in the scheduler domain hierarchy
822 * when tick_nohz_stop_sched_tick() is called from the idle loop.
823 * State will be updated to busy during the first busy tick after
824 * exiting idle.
825 */
826 set_cpu_sd_state_idle();
827
Frederic Weisbecker1268fbc2011-11-17 18:48:14 +0100828 local_irq_disable();
829
Frederic Weisbecker280f0672011-10-07 18:22:06 +0200830 ts = &__get_cpu_var(tick_cpu_sched);
Frederic Weisbecker280f0672011-10-07 18:22:06 +0200831 ts->inidle = 1;
Frederic Weisbecker19f5f732011-07-27 17:29:28 +0200832 __tick_nohz_idle_enter(ts);
Frederic Weisbecker1268fbc2011-11-17 18:48:14 +0100833
834 local_irq_enable();
Frederic Weisbecker280f0672011-10-07 18:22:06 +0200835}
Jacob Pan4dbd2772013-01-04 11:12:43 +0000836EXPORT_SYMBOL_GPL(tick_nohz_idle_enter);
Frederic Weisbecker280f0672011-10-07 18:22:06 +0200837
838/**
839 * tick_nohz_irq_exit - update next tick event from interrupt exit
840 *
841 * When an interrupt fires while we are idle and it doesn't cause
842 * a reschedule, it may still add, modify or delete a timer, enqueue
843 * an RCU callback, etc...
844 * So we need to re-calculate and reprogram the next tick event.
845 */
846void tick_nohz_irq_exit(void)
847{
848 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
849
Rafael J. Wysocki14851912013-07-27 01:41:34 +0200850 if (ts->inidle)
Frederic Weisbecker5811d992013-04-20 16:40:31 +0200851 __tick_nohz_idle_enter(ts);
Rafael J. Wysocki14851912013-07-27 01:41:34 +0200852 else
Frederic Weisbecker5811d992013-04-20 16:40:31 +0200853 tick_nohz_full_stop_tick(ts);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800854}
855
856/**
Len Brown4f86d3a2007-10-03 18:58:00 -0400857 * tick_nohz_get_sleep_length - return the length of the current sleep
858 *
859 * Called from power state control code with interrupts disabled
860 */
861ktime_t tick_nohz_get_sleep_length(void)
862{
863 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
864
865 return ts->sleep_length;
866}
867
Thomas Gleixnerc34bec52008-10-17 10:04:34 +0200868static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
869{
870 hrtimer_cancel(&ts->sched_timer);
Frederic Weisbeckerf5d411c2011-07-31 17:44:12 +0200871 hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
Thomas Gleixnerc34bec52008-10-17 10:04:34 +0200872
873 while (1) {
874 /* Forward the time to expire in the future */
875 hrtimer_forward(&ts->sched_timer, now, tick_period);
876
877 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
Thomas Gleixner268a3dc2008-10-22 09:48:06 +0200878 hrtimer_start_expires(&ts->sched_timer,
Arun R Bharadwaj5c333862009-04-16 12:14:37 +0530879 HRTIMER_MODE_ABS_PINNED);
Thomas Gleixnerc34bec52008-10-17 10:04:34 +0200880 /* Check, if the timer was already in the past */
881 if (hrtimer_active(&ts->sched_timer))
882 break;
883 } else {
Thomas Gleixner268a3dc2008-10-22 09:48:06 +0200884 if (!tick_program_event(
885 hrtimer_get_expires(&ts->sched_timer), 0))
Thomas Gleixnerc34bec52008-10-17 10:04:34 +0200886 break;
887 }
Neal Cardwell6f103922012-03-27 15:09:37 -0400888 /* Reread time and update jiffies */
Thomas Gleixnerc34bec52008-10-17 10:04:34 +0200889 now = ktime_get();
Neal Cardwell6f103922012-03-27 15:09:37 -0400890 tick_do_update_jiffies64(now);
Thomas Gleixnerc34bec52008-10-17 10:04:34 +0200891 }
892}
893
Frederic Weisbecker19f5f732011-07-27 17:29:28 +0200894static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800895{
Venki Pallipadi6378ddb2008-01-30 13:30:04 +0100896 /* Update jiffies first */
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800897 tick_do_update_jiffies64(now);
Peter Zijlstra5aaa0b72012-05-17 17:15:29 +0200898 update_cpu_load_nohz();
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800899
Charles Wang749c8812012-08-20 16:02:33 +0800900 calc_load_exit_idle();
Frederic Weisbecker2ac0d982011-07-28 04:00:47 +0200901 touch_softlockup_watchdog();
902 /*
903 * Cancel the scheduled timer and restore the tick
904 */
905 ts->tick_stopped = 0;
906 ts->idle_exittime = now;
907
908 tick_nohz_restart(ts, now);
909}
910
911static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
912{
Frederic Weisbecker3f4724e2012-07-16 18:00:34 +0200913#ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
Frederic Weisbecker2ac0d982011-07-28 04:00:47 +0200914 unsigned long ticks;
Frederic Weisbecker3f4724e2012-07-16 18:00:34 +0200915
916 if (vtime_accounting_enabled())
917 return;
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800918 /*
919 * We stopped the tick in idle. Update process times would miss the
920 * time we slept as update_process_times does only a 1 tick
921 * accounting. Enforce that this is accounted to idle !
922 */
923 ticks = jiffies - ts->idle_jiffies;
924 /*
925 * We might be one off. Do not randomly account a huge number of ticks!
926 */
Martin Schwidefsky79741dd2008-12-31 15:11:38 +0100927 if (ticks && ticks < LONG_MAX)
928 account_idle_ticks(ticks);
929#endif
Frederic Weisbecker19f5f732011-07-27 17:29:28 +0200930}
931
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800932/**
933 * tick_nohz_idle_exit - restart the idle tick from the idle task
934 *
935 * Restart the idle tick when the CPU is woken up from idle
936 * This also exit the RCU extended quiescent state. The CPU
937 * can use RCU again after this function is called.
938 */
939void tick_nohz_idle_exit(void)
940{
Frederic Weisbeckere8fcaa52013-08-07 22:28:01 +0200941 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800942 ktime_t now;
943
944 local_irq_disable();
945
946 WARN_ON_ONCE(!ts->inidle);
947
948 ts->inidle = 0;
949
950 if (ts->idle_active || ts->tick_stopped)
951 now = ktime_get();
952
953 if (ts->idle_active)
Frederic Weisbeckere8fcaa52013-08-07 22:28:01 +0200954 tick_nohz_stop_idle(ts, now);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800955
Frederic Weisbecker2ac0d982011-07-28 04:00:47 +0200956 if (ts->tick_stopped) {
Frederic Weisbecker19f5f732011-07-27 17:29:28 +0200957 tick_nohz_restart_sched_tick(ts, now);
Frederic Weisbecker2ac0d982011-07-28 04:00:47 +0200958 tick_nohz_account_idle_ticks(ts);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800959 }
960
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800961 local_irq_enable();
962}
Jacob Pan4dbd2772013-01-04 11:12:43 +0000963EXPORT_SYMBOL_GPL(tick_nohz_idle_exit);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800964
965static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
966{
967 hrtimer_forward(&ts->sched_timer, now, tick_period);
Arjan van de Vencc584b22008-09-01 15:02:30 -0700968 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800969}
970
971/*
972 * The nohz low res interrupt handler
973 */
974static void tick_nohz_handler(struct clock_event_device *dev)
975{
976 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
977 struct pt_regs *regs = get_irq_regs();
978 ktime_t now = ktime_get();
979
980 dev->next_event.tv64 = KTIME_MAX;
981
Frederic Weisbecker5bb96222012-10-15 02:03:27 +0200982 tick_sched_do_timer(now);
Frederic Weisbecker9e8f5592012-10-15 02:43:03 +0200983 tick_sched_handle(ts, regs);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800984
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -0800985 while (tick_nohz_reprogram(ts, now)) {
986 now = ktime_get();
987 tick_do_update_jiffies64(now);
988 }
989}
990
991/**
992 * tick_nohz_switch_to_nohz - switch to nohz mode
993 */
994static void tick_nohz_switch_to_nohz(void)
995{
996 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
997 ktime_t next;
998
Viresh Kumar27630532014-04-15 10:54:41 +0530999 if (!tick_nohz_enabled)
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001000 return;
1001
1002 local_irq_disable();
1003 if (tick_switch_to_oneshot(tick_nohz_handler)) {
1004 local_irq_enable();
1005 return;
1006 }
Thomas Gleixnerd689fe22013-11-13 21:01:57 +01001007 tick_nohz_active = 1;
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001008 ts->nohz_mode = NOHZ_MODE_LOWRES;
1009
1010 /*
1011 * Recycle the hrtimer in ts, so we can share the
1012 * hrtimer_forward with the highres code.
1013 */
1014 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1015 /* Get the next period */
1016 next = tick_init_jiffy_update();
1017
1018 for (;;) {
Arjan van de Vencc584b22008-09-01 15:02:30 -07001019 hrtimer_set_expires(&ts->sched_timer, next);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001020 if (!tick_program_event(next, 0))
1021 break;
1022 next = ktime_add(next, tick_period);
1023 }
1024 local_irq_enable();
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001025}
1026
Thomas Gleixnerfb02fbc2008-10-17 10:01:23 +02001027/*
1028 * When NOHZ is enabled and the tick is stopped, we need to kick the
1029 * tick timer from irq_enter() so that the jiffies update is kept
1030 * alive during long running softirqs. That's ugly as hell, but
1031 * correctness is key even if we need to fix the offending softirq in
1032 * the first place.
1033 *
1034 * Note, this is different to tick_nohz_restart. We just kick the
1035 * timer and do not touch the other magic bits which need to be done
1036 * when idle is left.
1037 */
Frederic Weisbeckere8fcaa52013-08-07 22:28:01 +02001038static void tick_nohz_kick_tick(struct tick_sched *ts, ktime_t now)
Thomas Gleixnerfb02fbc2008-10-17 10:01:23 +02001039{
Thomas Gleixnerae992862008-11-10 13:20:23 +01001040#if 0
1041 /* Switch back to 2.6.27 behaviour */
Martin Schwidefskyeed3b9c2009-09-29 14:25:15 +02001042 ktime_t delta;
Thomas Gleixnerfb02fbc2008-10-17 10:01:23 +02001043
Thomas Gleixnerc4bd8222008-10-21 20:17:35 +02001044 /*
1045 * Do not touch the tick device, when the next expiry is either
1046 * already reached or less/equal than the tick period.
1047 */
Thomas Gleixner268a3dc2008-10-22 09:48:06 +02001048 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
Thomas Gleixnerc4bd8222008-10-21 20:17:35 +02001049 if (delta.tv64 <= tick_period.tv64)
1050 return;
1051
1052 tick_nohz_restart(ts, now);
Thomas Gleixnerae992862008-11-10 13:20:23 +01001053#endif
Thomas Gleixnerfb02fbc2008-10-17 10:01:23 +02001054}
1055
Frederic Weisbecker5acac1b2013-12-04 18:28:20 +01001056static inline void tick_nohz_irq_enter(void)
Martin Schwidefskyeed3b9c2009-09-29 14:25:15 +02001057{
Frederic Weisbeckere8fcaa52013-08-07 22:28:01 +02001058 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
Martin Schwidefskyeed3b9c2009-09-29 14:25:15 +02001059 ktime_t now;
1060
1061 if (!ts->idle_active && !ts->tick_stopped)
1062 return;
1063 now = ktime_get();
1064 if (ts->idle_active)
Frederic Weisbeckere8fcaa52013-08-07 22:28:01 +02001065 tick_nohz_stop_idle(ts, now);
Martin Schwidefskyeed3b9c2009-09-29 14:25:15 +02001066 if (ts->tick_stopped) {
1067 tick_nohz_update_jiffies(now);
Frederic Weisbeckere8fcaa52013-08-07 22:28:01 +02001068 tick_nohz_kick_tick(ts, now);
Martin Schwidefskyeed3b9c2009-09-29 14:25:15 +02001069 }
1070}
1071
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001072#else
1073
1074static inline void tick_nohz_switch_to_nohz(void) { }
Frederic Weisbecker5acac1b2013-12-04 18:28:20 +01001075static inline void tick_nohz_irq_enter(void) { }
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001076
Frederic Weisbecker3451d022011-08-10 23:21:01 +02001077#endif /* CONFIG_NO_HZ_COMMON */
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001078
1079/*
Thomas Gleixner719254fa2008-10-17 09:59:47 +02001080 * Called from irq_enter to notify about the possible interruption of idle()
1081 */
Frederic Weisbecker5acac1b2013-12-04 18:28:20 +01001082void tick_irq_enter(void)
Thomas Gleixner719254fa2008-10-17 09:59:47 +02001083{
Frederic Weisbeckere8fcaa52013-08-07 22:28:01 +02001084 tick_check_oneshot_broadcast_this_cpu();
Frederic Weisbecker5acac1b2013-12-04 18:28:20 +01001085 tick_nohz_irq_enter();
Thomas Gleixner719254fa2008-10-17 09:59:47 +02001086}
1087
1088/*
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001089 * High resolution timer specific code
1090 */
1091#ifdef CONFIG_HIGH_RES_TIMERS
1092/*
Pavel Machek4c9dc642008-01-30 13:30:00 +01001093 * We rearm the timer until we get disabled by the idle code.
Chuansheng Liu351f1812012-10-25 01:07:35 +08001094 * Called with interrupts disabled.
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001095 */
1096static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
1097{
1098 struct tick_sched *ts =
1099 container_of(timer, struct tick_sched, sched_timer);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001100 struct pt_regs *regs = get_irq_regs();
1101 ktime_t now = ktime_get();
Thomas Gleixnerd3ed7822007-05-08 00:30:03 -07001102
Frederic Weisbecker5bb96222012-10-15 02:03:27 +02001103 tick_sched_do_timer(now);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001104
1105 /*
1106 * Do not call, when we are not in irq context and have
1107 * no valid regs pointer
1108 */
Frederic Weisbecker9e8f5592012-10-15 02:43:03 +02001109 if (regs)
1110 tick_sched_handle(ts, regs);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001111
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001112 hrtimer_forward(timer, now, tick_period);
1113
1114 return HRTIMER_RESTART;
1115}
1116
Mike Galbraith5307c952012-05-08 12:20:58 +02001117static int sched_skew_tick;
1118
Thomas Gleixner62cf20b2012-05-25 14:08:57 +02001119static int __init skew_tick(char *str)
1120{
1121 get_option(&str, &sched_skew_tick);
1122
1123 return 0;
1124}
1125early_param("skew_tick", skew_tick);
1126
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001127/**
1128 * tick_setup_sched_timer - setup the tick emulation timer
1129 */
1130void tick_setup_sched_timer(void)
1131{
1132 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1133 ktime_t now = ktime_get();
1134
1135 /*
1136 * Emulate tick processing via per-CPU hrtimers:
1137 */
1138 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1139 ts->sched_timer.function = tick_sched_timer;
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001140
john stultz37045402007-07-21 04:37:35 -07001141 /* Get the next period (per cpu) */
Arjan van de Vencc584b22008-09-01 15:02:30 -07001142 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001143
Thomas Gleixner9c3f9e22012-11-21 20:31:52 +01001144 /* Offset the tick to avert jiffies_lock contention. */
Mike Galbraith5307c952012-05-08 12:20:58 +02001145 if (sched_skew_tick) {
1146 u64 offset = ktime_to_ns(tick_period) >> 1;
1147 do_div(offset, num_possible_cpus());
1148 offset *= smp_processor_id();
1149 hrtimer_add_expires_ns(&ts->sched_timer, offset);
1150 }
1151
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001152 for (;;) {
1153 hrtimer_forward(&ts->sched_timer, now, tick_period);
Arun R Bharadwaj5c333862009-04-16 12:14:37 +05301154 hrtimer_start_expires(&ts->sched_timer,
1155 HRTIMER_MODE_ABS_PINNED);
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001156 /* Check, if the timer was already in the past */
1157 if (hrtimer_active(&ts->sched_timer))
1158 break;
1159 now = ktime_get();
1160 }
1161
Frederic Weisbecker3451d022011-08-10 23:21:01 +02001162#ifdef CONFIG_NO_HZ_COMMON
Thomas Gleixnerd689fe22013-11-13 21:01:57 +01001163 if (tick_nohz_enabled) {
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001164 ts->nohz_mode = NOHZ_MODE_HIGHRES;
Thomas Gleixnerd689fe22013-11-13 21:01:57 +01001165 tick_nohz_active = 1;
1166 }
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001167#endif
1168}
Miao Xie3c4fbe52008-08-20 16:37:38 -07001169#endif /* HIGH_RES_TIMERS */
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001170
Frederic Weisbecker3451d022011-08-10 23:21:01 +02001171#if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001172void tick_cancel_sched_timer(int cpu)
1173{
1174 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1175
Miao Xie3c4fbe52008-08-20 16:37:38 -07001176# ifdef CONFIG_HIGH_RES_TIMERS
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001177 if (ts->sched_timer.base)
1178 hrtimer_cancel(&ts->sched_timer);
Miao Xie3c4fbe52008-08-20 16:37:38 -07001179# endif
Karsten Wiesea7901762008-03-04 14:59:55 -08001180
Thomas Gleixner4b0c0f22013-05-03 15:02:50 +02001181 memset(ts, 0, sizeof(*ts));
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001182}
Miao Xie3c4fbe52008-08-20 16:37:38 -07001183#endif
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001184
1185/**
1186 * Async notification about clocksource changes
1187 */
1188void tick_clock_notify(void)
1189{
1190 int cpu;
1191
1192 for_each_possible_cpu(cpu)
1193 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1194}
1195
1196/*
1197 * Async notification about clock event changes
1198 */
1199void tick_oneshot_notify(void)
1200{
1201 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1202
1203 set_bit(0, &ts->check_clocks);
1204}
1205
1206/**
1207 * Check, if a change happened, which makes oneshot possible.
1208 *
1209 * Called cyclic from the hrtimer softirq (driven by the timer
1210 * softirq) allow_nohz signals, that we can switch into low-res nohz
1211 * mode, because high resolution timers are disabled (either compile
1212 * or runtime).
1213 */
1214int tick_check_oneshot_change(int allow_nohz)
1215{
1216 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
1217
1218 if (!test_and_clear_bit(0, &ts->check_clocks))
1219 return 0;
1220
1221 if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
1222 return 0;
1223
Li Zefancf4fc6c2008-02-08 04:19:24 -08001224 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
Thomas Gleixner79bf2bb2007-02-16 01:28:03 -08001225 return 0;
1226
1227 if (!allow_nohz)
1228 return 1;
1229
1230 tick_nohz_switch_to_nohz();
1231 return 0;
1232}