]>
Commit | Line | Data |
---|---|---|
79bf2bb3 TG |
1 | /* |
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 | * | |
b10db7f0 | 12 | * Distribute under GPLv2. |
79bf2bb3 TG |
13 | */ |
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> | |
8083e4ad | 22 | #include <linux/module.h> |
00b42959 | 23 | #include <linux/irq_work.h> |
9014c45d FW |
24 | #include <linux/posix-timers.h> |
25 | #include <linux/perf_event.h> | |
79bf2bb3 | 26 | |
9e203bcc DM |
27 | #include <asm/irq_regs.h> |
28 | ||
79bf2bb3 TG |
29 | #include "tick-internal.h" |
30 | ||
cb41a290 FW |
31 | #include <trace/events/timer.h> |
32 | ||
79bf2bb3 TG |
33 | /* |
34 | * Per cpu nohz control structure | |
35 | */ | |
33a5f626 | 36 | DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched); |
79bf2bb3 TG |
37 | |
38 | /* | |
d6ad4187 | 39 | * The time, when the last jiffy update happened. Protected by jiffies_lock. |
79bf2bb3 TG |
40 | */ |
41 | static ktime_t last_jiffies_update; | |
42 | ||
289f480a IM |
43 | struct tick_sched *tick_get_tick_sched(int cpu) |
44 | { | |
45 | return &per_cpu(tick_cpu_sched, cpu); | |
46 | } | |
47 | ||
79bf2bb3 TG |
48 | /* |
49 | * Must be called with interrupts disabled ! | |
50 | */ | |
51 | static void tick_do_update_jiffies64(ktime_t now) | |
52 | { | |
53 | unsigned long ticks = 0; | |
54 | ktime_t delta; | |
55 | ||
7a14ce1d | 56 | /* |
d6ad4187 | 57 | * Do a quick check without holding jiffies_lock: |
7a14ce1d IM |
58 | */ |
59 | delta = ktime_sub(now, last_jiffies_update); | |
60 | if (delta.tv64 < tick_period.tv64) | |
61 | return; | |
62 | ||
d6ad4187 JS |
63 | /* Reevalute with jiffies_lock held */ |
64 | write_seqlock(&jiffies_lock); | |
79bf2bb3 TG |
65 | |
66 | delta = ktime_sub(now, last_jiffies_update); | |
67 | if (delta.tv64 >= tick_period.tv64) { | |
68 | ||
69 | delta = ktime_sub(delta, tick_period); | |
70 | last_jiffies_update = ktime_add(last_jiffies_update, | |
71 | tick_period); | |
72 | ||
73 | /* Slow path for long timeouts */ | |
74 | if (unlikely(delta.tv64 >= tick_period.tv64)) { | |
75 | s64 incr = ktime_to_ns(tick_period); | |
76 | ||
77 | ticks = ktime_divns(delta, incr); | |
78 | ||
79 | last_jiffies_update = ktime_add_ns(last_jiffies_update, | |
80 | incr * ticks); | |
81 | } | |
82 | do_timer(++ticks); | |
49d670fb TG |
83 | |
84 | /* Keep the tick_next_period variable up to date */ | |
85 | tick_next_period = ktime_add(last_jiffies_update, tick_period); | |
79bf2bb3 | 86 | } |
d6ad4187 | 87 | write_sequnlock(&jiffies_lock); |
79bf2bb3 TG |
88 | } |
89 | ||
90 | /* | |
91 | * Initialize and return retrieve the jiffies update. | |
92 | */ | |
93 | static ktime_t tick_init_jiffy_update(void) | |
94 | { | |
95 | ktime_t period; | |
96 | ||
d6ad4187 | 97 | write_seqlock(&jiffies_lock); |
79bf2bb3 TG |
98 | /* Did we start the jiffies update yet ? */ |
99 | if (last_jiffies_update.tv64 == 0) | |
100 | last_jiffies_update = tick_next_period; | |
101 | period = last_jiffies_update; | |
d6ad4187 | 102 | write_sequnlock(&jiffies_lock); |
79bf2bb3 TG |
103 | return period; |
104 | } | |
105 | ||
5bb96226 FW |
106 | |
107 | static void tick_sched_do_timer(ktime_t now) | |
108 | { | |
109 | int cpu = smp_processor_id(); | |
110 | ||
3451d024 | 111 | #ifdef CONFIG_NO_HZ_COMMON |
5bb96226 FW |
112 | /* |
113 | * Check if the do_timer duty was dropped. We don't care about | |
114 | * concurrency: This happens only when the cpu in charge went | |
115 | * into a long sleep. If two cpus happen to assign themself to | |
116 | * this duty, then the jiffies update is still serialized by | |
9c3f9e28 | 117 | * jiffies_lock. |
5bb96226 | 118 | */ |
a382bf93 | 119 | if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE) |
c5bfece2 | 120 | && !tick_nohz_full_cpu(cpu)) |
5bb96226 FW |
121 | tick_do_timer_cpu = cpu; |
122 | #endif | |
123 | ||
124 | /* Check, if the jiffies need an update */ | |
125 | if (tick_do_timer_cpu == cpu) | |
126 | tick_do_update_jiffies64(now); | |
127 | } | |
128 | ||
9e8f559b FW |
129 | static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs) |
130 | { | |
3451d024 | 131 | #ifdef CONFIG_NO_HZ_COMMON |
9e8f559b FW |
132 | /* |
133 | * When we are idle and the tick is stopped, we have to touch | |
134 | * the watchdog as we might not schedule for a really long | |
135 | * time. This happens on complete idle SMP systems while | |
136 | * waiting on the login prompt. We also increment the "start of | |
137 | * idle" jiffy stamp so the idle accounting adjustment we do | |
138 | * when we go busy again does not account too much ticks. | |
139 | */ | |
140 | if (ts->tick_stopped) { | |
141 | touch_softlockup_watchdog(); | |
142 | if (is_idle_task(current)) | |
143 | ts->idle_jiffies++; | |
144 | } | |
94a57140 | 145 | #endif |
9e8f559b FW |
146 | update_process_times(user_mode(regs)); |
147 | profile_tick(CPU_PROFILING); | |
148 | } | |
149 | ||
c5bfece2 FW |
150 | #ifdef CONFIG_NO_HZ_FULL |
151 | static cpumask_var_t nohz_full_mask; | |
152 | bool have_nohz_full_mask; | |
a831881b | 153 | |
9014c45d FW |
154 | static bool can_stop_full_tick(void) |
155 | { | |
156 | WARN_ON_ONCE(!irqs_disabled()); | |
157 | ||
cb41a290 FW |
158 | if (!sched_can_stop_tick()) { |
159 | trace_tick_stop(0, "more than 1 task in runqueue\n"); | |
9014c45d | 160 | return false; |
cb41a290 | 161 | } |
9014c45d | 162 | |
cb41a290 FW |
163 | if (!posix_cpu_timers_can_stop_tick(current)) { |
164 | trace_tick_stop(0, "posix timers running\n"); | |
9014c45d | 165 | return false; |
cb41a290 | 166 | } |
9014c45d | 167 | |
cb41a290 FW |
168 | if (!perf_event_can_stop_tick()) { |
169 | trace_tick_stop(0, "perf events running\n"); | |
9014c45d | 170 | return false; |
cb41a290 | 171 | } |
9014c45d FW |
172 | |
173 | /* sched_clock_tick() needs us? */ | |
174 | #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK | |
175 | /* | |
176 | * TODO: kick full dynticks CPUs when | |
177 | * sched_clock_stable is set. | |
178 | */ | |
cb41a290 FW |
179 | if (!sched_clock_stable) { |
180 | trace_tick_stop(0, "unstable sched clock\n"); | |
9014c45d | 181 | return false; |
cb41a290 | 182 | } |
9014c45d FW |
183 | #endif |
184 | ||
185 | return true; | |
186 | } | |
187 | ||
188 | static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now); | |
189 | ||
76c24fb0 FW |
190 | /* |
191 | * Re-evaluate the need for the tick on the current CPU | |
192 | * and restart it if necessary. | |
193 | */ | |
ff442c51 | 194 | void tick_nohz_full_check(void) |
76c24fb0 | 195 | { |
9014c45d FW |
196 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); |
197 | ||
198 | if (tick_nohz_full_cpu(smp_processor_id())) { | |
199 | if (ts->tick_stopped && !is_idle_task(current)) { | |
200 | if (!can_stop_full_tick()) | |
201 | tick_nohz_restart_sched_tick(ts, ktime_get()); | |
202 | } | |
203 | } | |
76c24fb0 FW |
204 | } |
205 | ||
206 | static void nohz_full_kick_work_func(struct irq_work *work) | |
207 | { | |
208 | tick_nohz_full_check(); | |
209 | } | |
210 | ||
211 | static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = { | |
212 | .func = nohz_full_kick_work_func, | |
213 | }; | |
214 | ||
215 | /* | |
216 | * Kick the current CPU if it's full dynticks in order to force it to | |
217 | * re-evaluate its dependency on the tick and restart it if necessary. | |
218 | */ | |
219 | void tick_nohz_full_kick(void) | |
220 | { | |
221 | if (tick_nohz_full_cpu(smp_processor_id())) | |
222 | irq_work_queue(&__get_cpu_var(nohz_full_kick_work)); | |
223 | } | |
224 | ||
225 | static void nohz_full_kick_ipi(void *info) | |
226 | { | |
227 | tick_nohz_full_check(); | |
228 | } | |
229 | ||
230 | /* | |
231 | * Kick all full dynticks CPUs in order to force these to re-evaluate | |
232 | * their dependency on the tick and restart it if necessary. | |
233 | */ | |
234 | void tick_nohz_full_kick_all(void) | |
235 | { | |
236 | if (!have_nohz_full_mask) | |
237 | return; | |
238 | ||
239 | preempt_disable(); | |
240 | smp_call_function_many(nohz_full_mask, | |
241 | nohz_full_kick_ipi, NULL, false); | |
242 | preempt_enable(); | |
243 | } | |
244 | ||
99e5ada9 FW |
245 | /* |
246 | * Re-evaluate the need for the tick as we switch the current task. | |
247 | * It might need the tick due to per task/process properties: | |
248 | * perf events, posix cpu timers, ... | |
249 | */ | |
250 | void tick_nohz_task_switch(struct task_struct *tsk) | |
251 | { | |
252 | unsigned long flags; | |
253 | ||
254 | if (!tick_nohz_full_cpu(smp_processor_id())) | |
255 | return; | |
256 | ||
257 | local_irq_save(flags); | |
258 | ||
259 | if (tick_nohz_tick_stopped() && !can_stop_full_tick()) | |
260 | tick_nohz_full_kick(); | |
261 | ||
262 | local_irq_restore(flags); | |
263 | } | |
264 | ||
c5bfece2 | 265 | int tick_nohz_full_cpu(int cpu) |
a831881b | 266 | { |
c5bfece2 | 267 | if (!have_nohz_full_mask) |
a831881b FW |
268 | return 0; |
269 | ||
c5bfece2 | 270 | return cpumask_test_cpu(cpu, nohz_full_mask); |
a831881b FW |
271 | } |
272 | ||
273 | /* Parse the boot-time nohz CPU list from the kernel parameters. */ | |
c5bfece2 | 274 | static int __init tick_nohz_full_setup(char *str) |
a831881b | 275 | { |
0453b435 FW |
276 | int cpu; |
277 | ||
c5bfece2 | 278 | alloc_bootmem_cpumask_var(&nohz_full_mask); |
0453b435 | 279 | if (cpulist_parse(str, nohz_full_mask) < 0) { |
c5bfece2 | 280 | pr_warning("NOHZ: Incorrect nohz_full cpumask\n"); |
0453b435 FW |
281 | return 1; |
282 | } | |
283 | ||
284 | cpu = smp_processor_id(); | |
285 | if (cpumask_test_cpu(cpu, nohz_full_mask)) { | |
286 | pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu); | |
287 | cpumask_clear_cpu(cpu, nohz_full_mask); | |
288 | } | |
289 | have_nohz_full_mask = true; | |
290 | ||
a831881b FW |
291 | return 1; |
292 | } | |
c5bfece2 | 293 | __setup("nohz_full=", tick_nohz_full_setup); |
a831881b | 294 | |
a382bf93 FW |
295 | static int __cpuinit tick_nohz_cpu_down_callback(struct notifier_block *nfb, |
296 | unsigned long action, | |
297 | void *hcpu) | |
298 | { | |
299 | unsigned int cpu = (unsigned long)hcpu; | |
300 | ||
301 | switch (action & ~CPU_TASKS_FROZEN) { | |
302 | case CPU_DOWN_PREPARE: | |
303 | /* | |
304 | * If we handle the timekeeping duty for full dynticks CPUs, | |
305 | * we can't safely shutdown that CPU. | |
306 | */ | |
c5bfece2 | 307 | if (have_nohz_full_mask && tick_do_timer_cpu == cpu) |
a382bf93 FW |
308 | return -EINVAL; |
309 | break; | |
310 | } | |
311 | return NOTIFY_OK; | |
312 | } | |
313 | ||
1034fc2f FW |
314 | /* |
315 | * Worst case string length in chunks of CPU range seems 2 steps | |
316 | * separations: 0,2,4,6,... | |
317 | * This is NR_CPUS + sizeof('\0') | |
318 | */ | |
c5bfece2 | 319 | static char __initdata nohz_full_buf[NR_CPUS + 1]; |
1034fc2f | 320 | |
f98823ac FW |
321 | static int tick_nohz_init_all(void) |
322 | { | |
323 | int err = -1; | |
324 | ||
325 | #ifdef CONFIG_NO_HZ_FULL_ALL | |
326 | if (!alloc_cpumask_var(&nohz_full_mask, GFP_KERNEL)) { | |
327 | pr_err("NO_HZ: Can't allocate full dynticks cpumask\n"); | |
328 | return err; | |
329 | } | |
330 | err = 0; | |
331 | cpumask_setall(nohz_full_mask); | |
332 | cpumask_clear_cpu(smp_processor_id(), nohz_full_mask); | |
333 | have_nohz_full_mask = true; | |
334 | #endif | |
335 | return err; | |
336 | } | |
337 | ||
d1e43fa5 | 338 | void __init tick_nohz_init(void) |
a831881b | 339 | { |
d1e43fa5 FW |
340 | int cpu; |
341 | ||
f98823ac FW |
342 | if (!have_nohz_full_mask) { |
343 | if (tick_nohz_init_all() < 0) | |
344 | return; | |
345 | } | |
d1e43fa5 FW |
346 | |
347 | cpu_notifier(tick_nohz_cpu_down_callback, 0); | |
348 | ||
349 | /* Make sure full dynticks CPU are also RCU nocbs */ | |
350 | for_each_cpu(cpu, nohz_full_mask) { | |
351 | if (!rcu_is_nocb_cpu(cpu)) { | |
352 | pr_warning("NO_HZ: CPU %d is not RCU nocb: " | |
353 | "cleared from nohz_full range", cpu); | |
354 | cpumask_clear_cpu(cpu, nohz_full_mask); | |
355 | } | |
356 | } | |
a831881b | 357 | |
c5bfece2 FW |
358 | cpulist_scnprintf(nohz_full_buf, sizeof(nohz_full_buf), nohz_full_mask); |
359 | pr_info("NO_HZ: Full dynticks CPUs: %s.\n", nohz_full_buf); | |
a831881b | 360 | } |
a831881b | 361 | #else |
c5bfece2 | 362 | #define have_nohz_full_mask (0) |
a831881b FW |
363 | #endif |
364 | ||
79bf2bb3 TG |
365 | /* |
366 | * NOHZ - aka dynamic tick functionality | |
367 | */ | |
3451d024 | 368 | #ifdef CONFIG_NO_HZ_COMMON |
79bf2bb3 TG |
369 | /* |
370 | * NO HZ enabled ? | |
371 | */ | |
9d2ad243 | 372 | int tick_nohz_enabled __read_mostly = 1; |
79bf2bb3 TG |
373 | |
374 | /* | |
375 | * Enable / Disable tickless mode | |
376 | */ | |
377 | static int __init setup_tick_nohz(char *str) | |
378 | { | |
379 | if (!strcmp(str, "off")) | |
380 | tick_nohz_enabled = 0; | |
381 | else if (!strcmp(str, "on")) | |
382 | tick_nohz_enabled = 1; | |
383 | else | |
384 | return 0; | |
385 | return 1; | |
386 | } | |
387 | ||
388 | __setup("nohz=", setup_tick_nohz); | |
389 | ||
390 | /** | |
391 | * tick_nohz_update_jiffies - update jiffies when idle was interrupted | |
392 | * | |
393 | * Called from interrupt entry when the CPU was idle | |
394 | * | |
395 | * In case the sched_tick was stopped on this CPU, we have to check if jiffies | |
396 | * must be updated. Otherwise an interrupt handler could use a stale jiffy | |
397 | * value. We do this unconditionally on any cpu, as we don't know whether the | |
398 | * cpu, which has the update task assigned is in a long sleep. | |
399 | */ | |
eed3b9cf | 400 | static void tick_nohz_update_jiffies(ktime_t now) |
79bf2bb3 TG |
401 | { |
402 | int cpu = smp_processor_id(); | |
403 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
404 | unsigned long flags; | |
79bf2bb3 | 405 | |
5df7fa1c | 406 | ts->idle_waketime = now; |
79bf2bb3 TG |
407 | |
408 | local_irq_save(flags); | |
409 | tick_do_update_jiffies64(now); | |
410 | local_irq_restore(flags); | |
02ff3755 IM |
411 | |
412 | touch_softlockup_watchdog(); | |
79bf2bb3 TG |
413 | } |
414 | ||
595aac48 AV |
415 | /* |
416 | * Updates the per cpu time idle statistics counters | |
417 | */ | |
8d63bf94 | 418 | static void |
8c215bd3 | 419 | update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time) |
6378ddb5 | 420 | { |
eed3b9cf | 421 | ktime_t delta; |
6378ddb5 | 422 | |
595aac48 AV |
423 | if (ts->idle_active) { |
424 | delta = ktime_sub(now, ts->idle_entrytime); | |
8c215bd3 | 425 | if (nr_iowait_cpu(cpu) > 0) |
0224cf4c | 426 | ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); |
6beea0cd MH |
427 | else |
428 | ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); | |
8c7b09f4 | 429 | ts->idle_entrytime = now; |
595aac48 | 430 | } |
8d63bf94 | 431 | |
e0e37c20 | 432 | if (last_update_time) |
8d63bf94 AV |
433 | *last_update_time = ktime_to_us(now); |
434 | ||
595aac48 AV |
435 | } |
436 | ||
437 | static void tick_nohz_stop_idle(int cpu, ktime_t now) | |
438 | { | |
439 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
440 | ||
8c215bd3 | 441 | update_ts_time_stats(cpu, ts, now, NULL); |
eed3b9cf | 442 | ts->idle_active = 0; |
56c7426b | 443 | |
eed3b9cf | 444 | sched_clock_idle_wakeup_event(0); |
6378ddb5 VP |
445 | } |
446 | ||
8c215bd3 | 447 | static ktime_t tick_nohz_start_idle(int cpu, struct tick_sched *ts) |
6378ddb5 | 448 | { |
430ee881 | 449 | ktime_t now = ktime_get(); |
595aac48 | 450 | |
6378ddb5 VP |
451 | ts->idle_entrytime = now; |
452 | ts->idle_active = 1; | |
56c7426b | 453 | sched_clock_idle_sleep_event(); |
6378ddb5 VP |
454 | return now; |
455 | } | |
456 | ||
b1f724c3 AV |
457 | /** |
458 | * get_cpu_idle_time_us - get the total idle time of a cpu | |
459 | * @cpu: CPU number to query | |
09a1d34f MH |
460 | * @last_update_time: variable to store update time in. Do not update |
461 | * counters if NULL. | |
b1f724c3 AV |
462 | * |
463 | * Return the cummulative idle time (since boot) for a given | |
6beea0cd | 464 | * CPU, in microseconds. |
b1f724c3 AV |
465 | * |
466 | * This time is measured via accounting rather than sampling, | |
467 | * and is as accurate as ktime_get() is. | |
468 | * | |
469 | * This function returns -1 if NOHZ is not enabled. | |
470 | */ | |
6378ddb5 VP |
471 | u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) |
472 | { | |
473 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
09a1d34f | 474 | ktime_t now, idle; |
6378ddb5 | 475 | |
8083e4ad | 476 | if (!tick_nohz_enabled) |
477 | return -1; | |
478 | ||
09a1d34f MH |
479 | now = ktime_get(); |
480 | if (last_update_time) { | |
481 | update_ts_time_stats(cpu, ts, now, last_update_time); | |
482 | idle = ts->idle_sleeptime; | |
483 | } else { | |
484 | if (ts->idle_active && !nr_iowait_cpu(cpu)) { | |
485 | ktime_t delta = ktime_sub(now, ts->idle_entrytime); | |
486 | ||
487 | idle = ktime_add(ts->idle_sleeptime, delta); | |
488 | } else { | |
489 | idle = ts->idle_sleeptime; | |
490 | } | |
491 | } | |
492 | ||
493 | return ktime_to_us(idle); | |
8083e4ad | 494 | |
6378ddb5 | 495 | } |
8083e4ad | 496 | EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); |
6378ddb5 | 497 | |
6beea0cd | 498 | /** |
0224cf4c AV |
499 | * get_cpu_iowait_time_us - get the total iowait time of a cpu |
500 | * @cpu: CPU number to query | |
09a1d34f MH |
501 | * @last_update_time: variable to store update time in. Do not update |
502 | * counters if NULL. | |
0224cf4c AV |
503 | * |
504 | * Return the cummulative iowait time (since boot) for a given | |
505 | * CPU, in microseconds. | |
506 | * | |
507 | * This time is measured via accounting rather than sampling, | |
508 | * and is as accurate as ktime_get() is. | |
509 | * | |
510 | * This function returns -1 if NOHZ is not enabled. | |
511 | */ | |
512 | u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) | |
513 | { | |
514 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
09a1d34f | 515 | ktime_t now, iowait; |
0224cf4c AV |
516 | |
517 | if (!tick_nohz_enabled) | |
518 | return -1; | |
519 | ||
09a1d34f MH |
520 | now = ktime_get(); |
521 | if (last_update_time) { | |
522 | update_ts_time_stats(cpu, ts, now, last_update_time); | |
523 | iowait = ts->iowait_sleeptime; | |
524 | } else { | |
525 | if (ts->idle_active && nr_iowait_cpu(cpu) > 0) { | |
526 | ktime_t delta = ktime_sub(now, ts->idle_entrytime); | |
0224cf4c | 527 | |
09a1d34f MH |
528 | iowait = ktime_add(ts->iowait_sleeptime, delta); |
529 | } else { | |
530 | iowait = ts->iowait_sleeptime; | |
531 | } | |
532 | } | |
0224cf4c | 533 | |
09a1d34f | 534 | return ktime_to_us(iowait); |
0224cf4c AV |
535 | } |
536 | EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); | |
537 | ||
84bf1bcc FW |
538 | static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts, |
539 | ktime_t now, int cpu) | |
79bf2bb3 | 540 | { |
280f0677 | 541 | unsigned long seq, last_jiffies, next_jiffies, delta_jiffies; |
84bf1bcc | 542 | ktime_t last_update, expires, ret = { .tv64 = 0 }; |
aa9b1630 | 543 | unsigned long rcu_delta_jiffies; |
4f86d3a8 | 544 | struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev; |
98962465 | 545 | u64 time_delta; |
79bf2bb3 | 546 | |
79bf2bb3 TG |
547 | /* Read jiffies and the time when jiffies were updated last */ |
548 | do { | |
d6ad4187 | 549 | seq = read_seqbegin(&jiffies_lock); |
79bf2bb3 TG |
550 | last_update = last_jiffies_update; |
551 | last_jiffies = jiffies; | |
27185016 | 552 | time_delta = timekeeping_max_deferment(); |
d6ad4187 | 553 | } while (read_seqretry(&jiffies_lock, seq)); |
79bf2bb3 | 554 | |
74876a98 | 555 | if (rcu_needs_cpu(cpu, &rcu_delta_jiffies) || |
00b42959 | 556 | arch_needs_cpu(cpu) || irq_work_needs_cpu()) { |
3c5d92a0 | 557 | next_jiffies = last_jiffies + 1; |
6ba9b346 | 558 | delta_jiffies = 1; |
3c5d92a0 MS |
559 | } else { |
560 | /* Get the next timer wheel timer */ | |
561 | next_jiffies = get_next_timer_interrupt(last_jiffies); | |
562 | delta_jiffies = next_jiffies - last_jiffies; | |
aa9b1630 PM |
563 | if (rcu_delta_jiffies < delta_jiffies) { |
564 | next_jiffies = last_jiffies + rcu_delta_jiffies; | |
565 | delta_jiffies = rcu_delta_jiffies; | |
566 | } | |
3c5d92a0 | 567 | } |
47aa8b6c | 568 | |
79bf2bb3 | 569 | /* |
47aa8b6c IM |
570 | * Do not stop the tick, if we are only one off (or less) |
571 | * or if the cpu is required for RCU: | |
79bf2bb3 | 572 | */ |
47aa8b6c | 573 | if (!ts->tick_stopped && delta_jiffies <= 1) |
79bf2bb3 TG |
574 | goto out; |
575 | ||
576 | /* Schedule the tick, if we are at least one jiffie off */ | |
577 | if ((long)delta_jiffies >= 1) { | |
578 | ||
00147449 WR |
579 | /* |
580 | * If this cpu is the one which updates jiffies, then | |
581 | * give up the assignment and let it be taken by the | |
582 | * cpu which runs the tick timer next, which might be | |
583 | * this cpu as well. If we don't drop this here the | |
584 | * jiffies might be stale and do_timer() never | |
27185016 TG |
585 | * invoked. Keep track of the fact that it was the one |
586 | * which had the do_timer() duty last. If this cpu is | |
587 | * the one which had the do_timer() duty last, we | |
588 | * limit the sleep time to the timekeeping | |
589 | * max_deferement value which we retrieved | |
590 | * above. Otherwise we can sleep as long as we want. | |
00147449 | 591 | */ |
27185016 | 592 | if (cpu == tick_do_timer_cpu) { |
00147449 | 593 | tick_do_timer_cpu = TICK_DO_TIMER_NONE; |
27185016 TG |
594 | ts->do_timer_last = 1; |
595 | } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { | |
596 | time_delta = KTIME_MAX; | |
597 | ts->do_timer_last = 0; | |
598 | } else if (!ts->do_timer_last) { | |
599 | time_delta = KTIME_MAX; | |
600 | } | |
601 | ||
00147449 | 602 | /* |
98962465 JH |
603 | * calculate the expiry time for the next timer wheel |
604 | * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals | |
605 | * that there is no timer pending or at least extremely | |
606 | * far into the future (12 days for HZ=1000). In this | |
607 | * case we set the expiry to the end of time. | |
608 | */ | |
609 | if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) { | |
610 | /* | |
611 | * Calculate the time delta for the next timer event. | |
612 | * If the time delta exceeds the maximum time delta | |
613 | * permitted by the current clocksource then adjust | |
614 | * the time delta accordingly to ensure the | |
615 | * clocksource does not wrap. | |
616 | */ | |
617 | time_delta = min_t(u64, time_delta, | |
618 | tick_period.tv64 * delta_jiffies); | |
98962465 | 619 | } |
00147449 | 620 | |
27185016 TG |
621 | if (time_delta < KTIME_MAX) |
622 | expires = ktime_add_ns(last_update, time_delta); | |
623 | else | |
624 | expires.tv64 = KTIME_MAX; | |
00147449 | 625 | |
00147449 WR |
626 | /* Skip reprogram of event if its not changed */ |
627 | if (ts->tick_stopped && ktime_equal(expires, dev->next_event)) | |
628 | goto out; | |
629 | ||
84bf1bcc FW |
630 | ret = expires; |
631 | ||
79bf2bb3 TG |
632 | /* |
633 | * nohz_stop_sched_tick can be called several times before | |
634 | * the nohz_restart_sched_tick is called. This happens when | |
635 | * interrupts arrive which do not cause a reschedule. In the | |
636 | * first call we save the current tick time, so we can restart | |
637 | * the scheduler tick in nohz_restart_sched_tick. | |
638 | */ | |
639 | if (!ts->tick_stopped) { | |
c1cc017c | 640 | nohz_balance_enter_idle(cpu); |
5167e8d5 | 641 | calc_load_enter_idle(); |
46cb4b7c | 642 | |
f5d411c9 | 643 | ts->last_tick = hrtimer_get_expires(&ts->sched_timer); |
79bf2bb3 | 644 | ts->tick_stopped = 1; |
cb41a290 | 645 | trace_tick_stop(1, " "); |
79bf2bb3 | 646 | } |
d3ed7824 | 647 | |
eaad084b | 648 | /* |
98962465 JH |
649 | * If the expiration time == KTIME_MAX, then |
650 | * in this case we simply stop the tick timer. | |
eaad084b | 651 | */ |
98962465 | 652 | if (unlikely(expires.tv64 == KTIME_MAX)) { |
eaad084b TG |
653 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) |
654 | hrtimer_cancel(&ts->sched_timer); | |
655 | goto out; | |
656 | } | |
657 | ||
79bf2bb3 TG |
658 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { |
659 | hrtimer_start(&ts->sched_timer, expires, | |
5c333864 | 660 | HRTIMER_MODE_ABS_PINNED); |
79bf2bb3 TG |
661 | /* Check, if the timer was already in the past */ |
662 | if (hrtimer_active(&ts->sched_timer)) | |
663 | goto out; | |
4c9dc641 | 664 | } else if (!tick_program_event(expires, 0)) |
79bf2bb3 TG |
665 | goto out; |
666 | /* | |
667 | * We are past the event already. So we crossed a | |
668 | * jiffie boundary. Update jiffies and raise the | |
669 | * softirq. | |
670 | */ | |
671 | tick_do_update_jiffies64(ktime_get()); | |
79bf2bb3 TG |
672 | } |
673 | raise_softirq_irqoff(TIMER_SOFTIRQ); | |
674 | out: | |
675 | ts->next_jiffies = next_jiffies; | |
676 | ts->last_jiffies = last_jiffies; | |
4f86d3a8 | 677 | ts->sleep_length = ktime_sub(dev->next_event, now); |
84bf1bcc FW |
678 | |
679 | return ret; | |
280f0677 FW |
680 | } |
681 | ||
5811d996 FW |
682 | static void tick_nohz_full_stop_tick(struct tick_sched *ts) |
683 | { | |
684 | #ifdef CONFIG_NO_HZ_FULL | |
685 | int cpu = smp_processor_id(); | |
686 | ||
687 | if (!tick_nohz_full_cpu(cpu) || is_idle_task(current)) | |
688 | return; | |
689 | ||
690 | if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE) | |
691 | return; | |
692 | ||
693 | if (!can_stop_full_tick()) | |
694 | return; | |
695 | ||
696 | tick_nohz_stop_sched_tick(ts, ktime_get(), cpu); | |
697 | #endif | |
698 | } | |
699 | ||
5b39939a FW |
700 | static bool can_stop_idle_tick(int cpu, struct tick_sched *ts) |
701 | { | |
702 | /* | |
703 | * If this cpu is offline and it is the one which updates | |
704 | * jiffies, then give up the assignment and let it be taken by | |
705 | * the cpu which runs the tick timer next. If we don't drop | |
706 | * this here the jiffies might be stale and do_timer() never | |
707 | * invoked. | |
708 | */ | |
709 | if (unlikely(!cpu_online(cpu))) { | |
710 | if (cpu == tick_do_timer_cpu) | |
711 | tick_do_timer_cpu = TICK_DO_TIMER_NONE; | |
712 | } | |
713 | ||
714 | if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) | |
715 | return false; | |
716 | ||
717 | if (need_resched()) | |
718 | return false; | |
719 | ||
720 | if (unlikely(local_softirq_pending() && cpu_online(cpu))) { | |
721 | static int ratelimit; | |
722 | ||
803b0eba PM |
723 | if (ratelimit < 10 && |
724 | (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) { | |
5b39939a FW |
725 | printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n", |
726 | (unsigned int) local_softirq_pending()); | |
727 | ratelimit++; | |
728 | } | |
729 | return false; | |
730 | } | |
731 | ||
c5bfece2 | 732 | if (have_nohz_full_mask) { |
a382bf93 FW |
733 | /* |
734 | * Keep the tick alive to guarantee timekeeping progression | |
735 | * if there are full dynticks CPUs around | |
736 | */ | |
737 | if (tick_do_timer_cpu == cpu) | |
738 | return false; | |
739 | /* | |
740 | * Boot safety: make sure the timekeeping duty has been | |
741 | * assigned before entering dyntick-idle mode, | |
742 | */ | |
743 | if (tick_do_timer_cpu == TICK_DO_TIMER_NONE) | |
744 | return false; | |
745 | } | |
746 | ||
5b39939a FW |
747 | return true; |
748 | } | |
749 | ||
19f5f736 FW |
750 | static void __tick_nohz_idle_enter(struct tick_sched *ts) |
751 | { | |
84bf1bcc | 752 | ktime_t now, expires; |
5b39939a | 753 | int cpu = smp_processor_id(); |
19f5f736 | 754 | |
5b39939a | 755 | now = tick_nohz_start_idle(cpu, ts); |
2ac0d98f | 756 | |
5b39939a FW |
757 | if (can_stop_idle_tick(cpu, ts)) { |
758 | int was_stopped = ts->tick_stopped; | |
759 | ||
760 | ts->idle_calls++; | |
84bf1bcc FW |
761 | |
762 | expires = tick_nohz_stop_sched_tick(ts, now, cpu); | |
763 | if (expires.tv64 > 0LL) { | |
764 | ts->idle_sleeps++; | |
765 | ts->idle_expires = expires; | |
766 | } | |
5b39939a FW |
767 | |
768 | if (!was_stopped && ts->tick_stopped) | |
769 | ts->idle_jiffies = ts->last_jiffies; | |
770 | } | |
280f0677 FW |
771 | } |
772 | ||
773 | /** | |
774 | * tick_nohz_idle_enter - stop the idle tick from the idle task | |
775 | * | |
776 | * When the next event is more than a tick into the future, stop the idle tick | |
777 | * Called when we start the idle loop. | |
2bbb6817 | 778 | * |
1268fbc7 | 779 | * The arch is responsible of calling: |
2bbb6817 FW |
780 | * |
781 | * - rcu_idle_enter() after its last use of RCU before the CPU is put | |
782 | * to sleep. | |
783 | * - rcu_idle_exit() before the first use of RCU after the CPU is woken up. | |
280f0677 | 784 | */ |
1268fbc7 | 785 | void tick_nohz_idle_enter(void) |
280f0677 FW |
786 | { |
787 | struct tick_sched *ts; | |
788 | ||
1268fbc7 FW |
789 | WARN_ON_ONCE(irqs_disabled()); |
790 | ||
0db49b72 LT |
791 | /* |
792 | * Update the idle state in the scheduler domain hierarchy | |
793 | * when tick_nohz_stop_sched_tick() is called from the idle loop. | |
794 | * State will be updated to busy during the first busy tick after | |
795 | * exiting idle. | |
796 | */ | |
797 | set_cpu_sd_state_idle(); | |
798 | ||
1268fbc7 FW |
799 | local_irq_disable(); |
800 | ||
280f0677 FW |
801 | ts = &__get_cpu_var(tick_cpu_sched); |
802 | /* | |
803 | * set ts->inidle unconditionally. even if the system did not | |
804 | * switch to nohz mode the cpu frequency governers rely on the | |
805 | * update of the idle time accounting in tick_nohz_start_idle(). | |
806 | */ | |
807 | ts->inidle = 1; | |
19f5f736 | 808 | __tick_nohz_idle_enter(ts); |
1268fbc7 FW |
809 | |
810 | local_irq_enable(); | |
280f0677 | 811 | } |
4dbd2771 | 812 | EXPORT_SYMBOL_GPL(tick_nohz_idle_enter); |
280f0677 FW |
813 | |
814 | /** | |
815 | * tick_nohz_irq_exit - update next tick event from interrupt exit | |
816 | * | |
817 | * When an interrupt fires while we are idle and it doesn't cause | |
818 | * a reschedule, it may still add, modify or delete a timer, enqueue | |
819 | * an RCU callback, etc... | |
820 | * So we need to re-calculate and reprogram the next tick event. | |
821 | */ | |
822 | void tick_nohz_irq_exit(void) | |
823 | { | |
824 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
825 | ||
5811d996 FW |
826 | if (ts->inidle) { |
827 | /* Cancel the timer because CPU already waken up from the C-states*/ | |
828 | menu_hrtimer_cancel(); | |
829 | __tick_nohz_idle_enter(ts); | |
830 | } else { | |
831 | tick_nohz_full_stop_tick(ts); | |
832 | } | |
79bf2bb3 TG |
833 | } |
834 | ||
4f86d3a8 LB |
835 | /** |
836 | * tick_nohz_get_sleep_length - return the length of the current sleep | |
837 | * | |
838 | * Called from power state control code with interrupts disabled | |
839 | */ | |
840 | ktime_t tick_nohz_get_sleep_length(void) | |
841 | { | |
842 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
843 | ||
844 | return ts->sleep_length; | |
845 | } | |
846 | ||
c34bec5a TG |
847 | static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) |
848 | { | |
849 | hrtimer_cancel(&ts->sched_timer); | |
f5d411c9 | 850 | hrtimer_set_expires(&ts->sched_timer, ts->last_tick); |
c34bec5a TG |
851 | |
852 | while (1) { | |
853 | /* Forward the time to expire in the future */ | |
854 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
855 | ||
856 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { | |
268a3dcf | 857 | hrtimer_start_expires(&ts->sched_timer, |
5c333864 | 858 | HRTIMER_MODE_ABS_PINNED); |
c34bec5a TG |
859 | /* Check, if the timer was already in the past */ |
860 | if (hrtimer_active(&ts->sched_timer)) | |
861 | break; | |
862 | } else { | |
268a3dcf TG |
863 | if (!tick_program_event( |
864 | hrtimer_get_expires(&ts->sched_timer), 0)) | |
c34bec5a TG |
865 | break; |
866 | } | |
6f103929 | 867 | /* Reread time and update jiffies */ |
c34bec5a | 868 | now = ktime_get(); |
6f103929 | 869 | tick_do_update_jiffies64(now); |
c34bec5a TG |
870 | } |
871 | } | |
872 | ||
19f5f736 | 873 | static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now) |
79bf2bb3 | 874 | { |
79bf2bb3 | 875 | /* Update jiffies first */ |
79bf2bb3 | 876 | tick_do_update_jiffies64(now); |
5aaa0b7a | 877 | update_cpu_load_nohz(); |
79bf2bb3 | 878 | |
749c8814 | 879 | calc_load_exit_idle(); |
2ac0d98f FW |
880 | touch_softlockup_watchdog(); |
881 | /* | |
882 | * Cancel the scheduled timer and restore the tick | |
883 | */ | |
884 | ts->tick_stopped = 0; | |
885 | ts->idle_exittime = now; | |
886 | ||
887 | tick_nohz_restart(ts, now); | |
888 | } | |
889 | ||
890 | static void tick_nohz_account_idle_ticks(struct tick_sched *ts) | |
891 | { | |
3f4724ea | 892 | #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE |
2ac0d98f | 893 | unsigned long ticks; |
3f4724ea FW |
894 | |
895 | if (vtime_accounting_enabled()) | |
896 | return; | |
79bf2bb3 TG |
897 | /* |
898 | * We stopped the tick in idle. Update process times would miss the | |
899 | * time we slept as update_process_times does only a 1 tick | |
900 | * accounting. Enforce that this is accounted to idle ! | |
901 | */ | |
902 | ticks = jiffies - ts->idle_jiffies; | |
903 | /* | |
904 | * We might be one off. Do not randomly account a huge number of ticks! | |
905 | */ | |
79741dd3 MS |
906 | if (ticks && ticks < LONG_MAX) |
907 | account_idle_ticks(ticks); | |
908 | #endif | |
19f5f736 FW |
909 | } |
910 | ||
79bf2bb3 | 911 | /** |
280f0677 | 912 | * tick_nohz_idle_exit - restart the idle tick from the idle task |
79bf2bb3 TG |
913 | * |
914 | * Restart the idle tick when the CPU is woken up from idle | |
280f0677 FW |
915 | * This also exit the RCU extended quiescent state. The CPU |
916 | * can use RCU again after this function is called. | |
79bf2bb3 | 917 | */ |
280f0677 | 918 | void tick_nohz_idle_exit(void) |
79bf2bb3 TG |
919 | { |
920 | int cpu = smp_processor_id(); | |
921 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
6378ddb5 | 922 | ktime_t now; |
79bf2bb3 | 923 | |
6378ddb5 | 924 | local_irq_disable(); |
2bbb6817 | 925 | |
15f827be FW |
926 | WARN_ON_ONCE(!ts->inidle); |
927 | ||
928 | ts->inidle = 0; | |
929 | ||
69a37bea YS |
930 | /* Cancel the timer because CPU already waken up from the C-states*/ |
931 | menu_hrtimer_cancel(); | |
15f827be | 932 | if (ts->idle_active || ts->tick_stopped) |
eed3b9cf MS |
933 | now = ktime_get(); |
934 | ||
935 | if (ts->idle_active) | |
936 | tick_nohz_stop_idle(cpu, now); | |
6378ddb5 | 937 | |
2ac0d98f | 938 | if (ts->tick_stopped) { |
19f5f736 | 939 | tick_nohz_restart_sched_tick(ts, now); |
2ac0d98f | 940 | tick_nohz_account_idle_ticks(ts); |
6378ddb5 | 941 | } |
79bf2bb3 | 942 | |
79bf2bb3 TG |
943 | local_irq_enable(); |
944 | } | |
4dbd2771 | 945 | EXPORT_SYMBOL_GPL(tick_nohz_idle_exit); |
79bf2bb3 TG |
946 | |
947 | static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now) | |
948 | { | |
949 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
cc584b21 | 950 | return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0); |
79bf2bb3 TG |
951 | } |
952 | ||
953 | /* | |
954 | * The nohz low res interrupt handler | |
955 | */ | |
956 | static void tick_nohz_handler(struct clock_event_device *dev) | |
957 | { | |
958 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
959 | struct pt_regs *regs = get_irq_regs(); | |
960 | ktime_t now = ktime_get(); | |
961 | ||
962 | dev->next_event.tv64 = KTIME_MAX; | |
963 | ||
5bb96226 | 964 | tick_sched_do_timer(now); |
9e8f559b | 965 | tick_sched_handle(ts, regs); |
79bf2bb3 | 966 | |
79bf2bb3 TG |
967 | while (tick_nohz_reprogram(ts, now)) { |
968 | now = ktime_get(); | |
969 | tick_do_update_jiffies64(now); | |
970 | } | |
971 | } | |
972 | ||
973 | /** | |
974 | * tick_nohz_switch_to_nohz - switch to nohz mode | |
975 | */ | |
976 | static void tick_nohz_switch_to_nohz(void) | |
977 | { | |
978 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
979 | ktime_t next; | |
980 | ||
981 | if (!tick_nohz_enabled) | |
982 | return; | |
983 | ||
984 | local_irq_disable(); | |
985 | if (tick_switch_to_oneshot(tick_nohz_handler)) { | |
986 | local_irq_enable(); | |
987 | return; | |
988 | } | |
989 | ||
990 | ts->nohz_mode = NOHZ_MODE_LOWRES; | |
991 | ||
992 | /* | |
993 | * Recycle the hrtimer in ts, so we can share the | |
994 | * hrtimer_forward with the highres code. | |
995 | */ | |
996 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); | |
997 | /* Get the next period */ | |
998 | next = tick_init_jiffy_update(); | |
999 | ||
1000 | for (;;) { | |
cc584b21 | 1001 | hrtimer_set_expires(&ts->sched_timer, next); |
79bf2bb3 TG |
1002 | if (!tick_program_event(next, 0)) |
1003 | break; | |
1004 | next = ktime_add(next, tick_period); | |
1005 | } | |
1006 | local_irq_enable(); | |
79bf2bb3 TG |
1007 | } |
1008 | ||
fb02fbc1 TG |
1009 | /* |
1010 | * When NOHZ is enabled and the tick is stopped, we need to kick the | |
1011 | * tick timer from irq_enter() so that the jiffies update is kept | |
1012 | * alive during long running softirqs. That's ugly as hell, but | |
1013 | * correctness is key even if we need to fix the offending softirq in | |
1014 | * the first place. | |
1015 | * | |
1016 | * Note, this is different to tick_nohz_restart. We just kick the | |
1017 | * timer and do not touch the other magic bits which need to be done | |
1018 | * when idle is left. | |
1019 | */ | |
eed3b9cf | 1020 | static void tick_nohz_kick_tick(int cpu, ktime_t now) |
fb02fbc1 | 1021 | { |
ae99286b TG |
1022 | #if 0 |
1023 | /* Switch back to 2.6.27 behaviour */ | |
1024 | ||
fb02fbc1 | 1025 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); |
eed3b9cf | 1026 | ktime_t delta; |
fb02fbc1 | 1027 | |
c4bd822e TG |
1028 | /* |
1029 | * Do not touch the tick device, when the next expiry is either | |
1030 | * already reached or less/equal than the tick period. | |
1031 | */ | |
268a3dcf | 1032 | delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now); |
c4bd822e TG |
1033 | if (delta.tv64 <= tick_period.tv64) |
1034 | return; | |
1035 | ||
1036 | tick_nohz_restart(ts, now); | |
ae99286b | 1037 | #endif |
fb02fbc1 TG |
1038 | } |
1039 | ||
eed3b9cf MS |
1040 | static inline void tick_check_nohz(int cpu) |
1041 | { | |
1042 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
1043 | ktime_t now; | |
1044 | ||
1045 | if (!ts->idle_active && !ts->tick_stopped) | |
1046 | return; | |
1047 | now = ktime_get(); | |
1048 | if (ts->idle_active) | |
1049 | tick_nohz_stop_idle(cpu, now); | |
1050 | if (ts->tick_stopped) { | |
1051 | tick_nohz_update_jiffies(now); | |
1052 | tick_nohz_kick_tick(cpu, now); | |
1053 | } | |
1054 | } | |
1055 | ||
79bf2bb3 TG |
1056 | #else |
1057 | ||
1058 | static inline void tick_nohz_switch_to_nohz(void) { } | |
eed3b9cf | 1059 | static inline void tick_check_nohz(int cpu) { } |
79bf2bb3 | 1060 | |
3451d024 | 1061 | #endif /* CONFIG_NO_HZ_COMMON */ |
79bf2bb3 | 1062 | |
719254fa TG |
1063 | /* |
1064 | * Called from irq_enter to notify about the possible interruption of idle() | |
1065 | */ | |
1066 | void tick_check_idle(int cpu) | |
1067 | { | |
fb02fbc1 | 1068 | tick_check_oneshot_broadcast(cpu); |
eed3b9cf | 1069 | tick_check_nohz(cpu); |
719254fa TG |
1070 | } |
1071 | ||
79bf2bb3 TG |
1072 | /* |
1073 | * High resolution timer specific code | |
1074 | */ | |
1075 | #ifdef CONFIG_HIGH_RES_TIMERS | |
1076 | /* | |
4c9dc641 | 1077 | * We rearm the timer until we get disabled by the idle code. |
351f181f | 1078 | * Called with interrupts disabled. |
79bf2bb3 TG |
1079 | */ |
1080 | static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) | |
1081 | { | |
1082 | struct tick_sched *ts = | |
1083 | container_of(timer, struct tick_sched, sched_timer); | |
79bf2bb3 TG |
1084 | struct pt_regs *regs = get_irq_regs(); |
1085 | ktime_t now = ktime_get(); | |
d3ed7824 | 1086 | |
5bb96226 | 1087 | tick_sched_do_timer(now); |
79bf2bb3 TG |
1088 | |
1089 | /* | |
1090 | * Do not call, when we are not in irq context and have | |
1091 | * no valid regs pointer | |
1092 | */ | |
9e8f559b FW |
1093 | if (regs) |
1094 | tick_sched_handle(ts, regs); | |
79bf2bb3 | 1095 | |
79bf2bb3 TG |
1096 | hrtimer_forward(timer, now, tick_period); |
1097 | ||
1098 | return HRTIMER_RESTART; | |
1099 | } | |
1100 | ||
5307c955 MG |
1101 | static int sched_skew_tick; |
1102 | ||
62cf20b3 TG |
1103 | static int __init skew_tick(char *str) |
1104 | { | |
1105 | get_option(&str, &sched_skew_tick); | |
1106 | ||
1107 | return 0; | |
1108 | } | |
1109 | early_param("skew_tick", skew_tick); | |
1110 | ||
79bf2bb3 TG |
1111 | /** |
1112 | * tick_setup_sched_timer - setup the tick emulation timer | |
1113 | */ | |
1114 | void tick_setup_sched_timer(void) | |
1115 | { | |
1116 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
1117 | ktime_t now = ktime_get(); | |
1118 | ||
1119 | /* | |
1120 | * Emulate tick processing via per-CPU hrtimers: | |
1121 | */ | |
1122 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); | |
1123 | ts->sched_timer.function = tick_sched_timer; | |
79bf2bb3 | 1124 | |
3704540b | 1125 | /* Get the next period (per cpu) */ |
cc584b21 | 1126 | hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); |
79bf2bb3 | 1127 | |
9c3f9e28 | 1128 | /* Offset the tick to avert jiffies_lock contention. */ |
5307c955 MG |
1129 | if (sched_skew_tick) { |
1130 | u64 offset = ktime_to_ns(tick_period) >> 1; | |
1131 | do_div(offset, num_possible_cpus()); | |
1132 | offset *= smp_processor_id(); | |
1133 | hrtimer_add_expires_ns(&ts->sched_timer, offset); | |
1134 | } | |
1135 | ||
79bf2bb3 TG |
1136 | for (;;) { |
1137 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
5c333864 AB |
1138 | hrtimer_start_expires(&ts->sched_timer, |
1139 | HRTIMER_MODE_ABS_PINNED); | |
79bf2bb3 TG |
1140 | /* Check, if the timer was already in the past */ |
1141 | if (hrtimer_active(&ts->sched_timer)) | |
1142 | break; | |
1143 | now = ktime_get(); | |
1144 | } | |
1145 | ||
3451d024 | 1146 | #ifdef CONFIG_NO_HZ_COMMON |
29c158e8 | 1147 | if (tick_nohz_enabled) |
79bf2bb3 TG |
1148 | ts->nohz_mode = NOHZ_MODE_HIGHRES; |
1149 | #endif | |
1150 | } | |
3c4fbe5e | 1151 | #endif /* HIGH_RES_TIMERS */ |
79bf2bb3 | 1152 | |
3451d024 | 1153 | #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS |
79bf2bb3 TG |
1154 | void tick_cancel_sched_timer(int cpu) |
1155 | { | |
1156 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
1157 | ||
3c4fbe5e | 1158 | # ifdef CONFIG_HIGH_RES_TIMERS |
79bf2bb3 TG |
1159 | if (ts->sched_timer.base) |
1160 | hrtimer_cancel(&ts->sched_timer); | |
3c4fbe5e | 1161 | # endif |
a7901766 | 1162 | |
79bf2bb3 TG |
1163 | ts->nohz_mode = NOHZ_MODE_INACTIVE; |
1164 | } | |
3c4fbe5e | 1165 | #endif |
79bf2bb3 TG |
1166 | |
1167 | /** | |
1168 | * Async notification about clocksource changes | |
1169 | */ | |
1170 | void tick_clock_notify(void) | |
1171 | { | |
1172 | int cpu; | |
1173 | ||
1174 | for_each_possible_cpu(cpu) | |
1175 | set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); | |
1176 | } | |
1177 | ||
1178 | /* | |
1179 | * Async notification about clock event changes | |
1180 | */ | |
1181 | void tick_oneshot_notify(void) | |
1182 | { | |
1183 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
1184 | ||
1185 | set_bit(0, &ts->check_clocks); | |
1186 | } | |
1187 | ||
1188 | /** | |
1189 | * Check, if a change happened, which makes oneshot possible. | |
1190 | * | |
1191 | * Called cyclic from the hrtimer softirq (driven by the timer | |
1192 | * softirq) allow_nohz signals, that we can switch into low-res nohz | |
1193 | * mode, because high resolution timers are disabled (either compile | |
1194 | * or runtime). | |
1195 | */ | |
1196 | int tick_check_oneshot_change(int allow_nohz) | |
1197 | { | |
1198 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); | |
1199 | ||
1200 | if (!test_and_clear_bit(0, &ts->check_clocks)) | |
1201 | return 0; | |
1202 | ||
1203 | if (ts->nohz_mode != NOHZ_MODE_INACTIVE) | |
1204 | return 0; | |
1205 | ||
cf4fc6cb | 1206 | if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) |
79bf2bb3 TG |
1207 | return 0; |
1208 | ||
1209 | if (!allow_nohz) | |
1210 | return 1; | |
1211 | ||
1212 | tick_nohz_switch_to_nohz(); | |
1213 | return 0; | |
1214 | } |