]>
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> | |
38b8d208 | 20 | #include <linux/nmi.h> |
79bf2bb3 | 21 | #include <linux/profile.h> |
3f07c014 | 22 | #include <linux/sched/signal.h> |
e6017571 | 23 | #include <linux/sched/clock.h> |
03441a34 | 24 | #include <linux/sched/stat.h> |
370c9135 | 25 | #include <linux/sched/nohz.h> |
8083e4ad | 26 | #include <linux/module.h> |
00b42959 | 27 | #include <linux/irq_work.h> |
9014c45d | 28 | #include <linux/posix-timers.h> |
2e709338 | 29 | #include <linux/context_tracking.h> |
62cb1188 | 30 | #include <linux/mm.h> |
79bf2bb3 | 31 | |
9e203bcc DM |
32 | #include <asm/irq_regs.h> |
33 | ||
79bf2bb3 TG |
34 | #include "tick-internal.h" |
35 | ||
cb41a290 FW |
36 | #include <trace/events/timer.h> |
37 | ||
79bf2bb3 | 38 | /* |
0de7611a | 39 | * Per-CPU nohz control structure |
79bf2bb3 | 40 | */ |
c1797baf | 41 | static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched); |
79bf2bb3 | 42 | |
289f480a IM |
43 | struct tick_sched *tick_get_tick_sched(int cpu) |
44 | { | |
45 | return &per_cpu(tick_cpu_sched, cpu); | |
46 | } | |
47 | ||
7809998a AB |
48 | #if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS) |
49 | /* | |
50 | * The time, when the last jiffy update happened. Protected by jiffies_lock. | |
51 | */ | |
52 | static ktime_t last_jiffies_update; | |
53 | ||
79bf2bb3 TG |
54 | /* |
55 | * Must be called with interrupts disabled ! | |
56 | */ | |
57 | static void tick_do_update_jiffies64(ktime_t now) | |
58 | { | |
59 | unsigned long ticks = 0; | |
60 | ktime_t delta; | |
61 | ||
7a14ce1d | 62 | /* |
d6ad4187 | 63 | * Do a quick check without holding jiffies_lock: |
7a14ce1d IM |
64 | */ |
65 | delta = ktime_sub(now, last_jiffies_update); | |
2456e855 | 66 | if (delta < tick_period) |
7a14ce1d IM |
67 | return; |
68 | ||
6168f8ed | 69 | /* Reevaluate with jiffies_lock held */ |
d6ad4187 | 70 | write_seqlock(&jiffies_lock); |
79bf2bb3 TG |
71 | |
72 | delta = ktime_sub(now, last_jiffies_update); | |
2456e855 | 73 | if (delta >= tick_period) { |
79bf2bb3 TG |
74 | |
75 | delta = ktime_sub(delta, tick_period); | |
76 | last_jiffies_update = ktime_add(last_jiffies_update, | |
77 | tick_period); | |
78 | ||
79 | /* Slow path for long timeouts */ | |
2456e855 | 80 | if (unlikely(delta >= tick_period)) { |
79bf2bb3 TG |
81 | s64 incr = ktime_to_ns(tick_period); |
82 | ||
83 | ticks = ktime_divns(delta, incr); | |
84 | ||
85 | last_jiffies_update = ktime_add_ns(last_jiffies_update, | |
86 | incr * ticks); | |
87 | } | |
88 | do_timer(++ticks); | |
49d670fb TG |
89 | |
90 | /* Keep the tick_next_period variable up to date */ | |
91 | tick_next_period = ktime_add(last_jiffies_update, tick_period); | |
03e6bdc5 VK |
92 | } else { |
93 | write_sequnlock(&jiffies_lock); | |
94 | return; | |
79bf2bb3 | 95 | } |
d6ad4187 | 96 | write_sequnlock(&jiffies_lock); |
47a1b796 | 97 | update_wall_time(); |
79bf2bb3 TG |
98 | } |
99 | ||
100 | /* | |
101 | * Initialize and return retrieve the jiffies update. | |
102 | */ | |
103 | static ktime_t tick_init_jiffy_update(void) | |
104 | { | |
105 | ktime_t period; | |
106 | ||
d6ad4187 | 107 | write_seqlock(&jiffies_lock); |
79bf2bb3 | 108 | /* Did we start the jiffies update yet ? */ |
2456e855 | 109 | if (last_jiffies_update == 0) |
79bf2bb3 TG |
110 | last_jiffies_update = tick_next_period; |
111 | period = last_jiffies_update; | |
d6ad4187 | 112 | write_sequnlock(&jiffies_lock); |
79bf2bb3 TG |
113 | return period; |
114 | } | |
115 | ||
5bb96226 FW |
116 | |
117 | static void tick_sched_do_timer(ktime_t now) | |
118 | { | |
119 | int cpu = smp_processor_id(); | |
120 | ||
3451d024 | 121 | #ifdef CONFIG_NO_HZ_COMMON |
5bb96226 FW |
122 | /* |
123 | * Check if the do_timer duty was dropped. We don't care about | |
0de7611a IM |
124 | * concurrency: This happens only when the CPU in charge went |
125 | * into a long sleep. If two CPUs happen to assign themselves to | |
5bb96226 | 126 | * this duty, then the jiffies update is still serialized by |
9c3f9e28 | 127 | * jiffies_lock. |
5bb96226 | 128 | */ |
a382bf93 | 129 | if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE) |
c5bfece2 | 130 | && !tick_nohz_full_cpu(cpu)) |
5bb96226 FW |
131 | tick_do_timer_cpu = cpu; |
132 | #endif | |
133 | ||
134 | /* Check, if the jiffies need an update */ | |
135 | if (tick_do_timer_cpu == cpu) | |
136 | tick_do_update_jiffies64(now); | |
137 | } | |
138 | ||
9e8f559b FW |
139 | static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs) |
140 | { | |
3451d024 | 141 | #ifdef CONFIG_NO_HZ_COMMON |
9e8f559b FW |
142 | /* |
143 | * When we are idle and the tick is stopped, we have to touch | |
144 | * the watchdog as we might not schedule for a really long | |
145 | * time. This happens on complete idle SMP systems while | |
146 | * waiting on the login prompt. We also increment the "start of | |
147 | * idle" jiffy stamp so the idle accounting adjustment we do | |
148 | * when we go busy again does not account too much ticks. | |
149 | */ | |
150 | if (ts->tick_stopped) { | |
03e0d461 | 151 | touch_softlockup_watchdog_sched(); |
9e8f559b FW |
152 | if (is_idle_task(current)) |
153 | ts->idle_jiffies++; | |
411fe24e FW |
154 | /* |
155 | * In case the current tick fired too early past its expected | |
156 | * expiration, make sure we don't bypass the next clock reprogramming | |
157 | * to the same deadline. | |
158 | */ | |
159 | ts->next_tick = 0; | |
9e8f559b | 160 | } |
94a57140 | 161 | #endif |
9e8f559b FW |
162 | update_process_times(user_mode(regs)); |
163 | profile_tick(CPU_PROFILING); | |
164 | } | |
7809998a | 165 | #endif |
9e8f559b | 166 | |
c5bfece2 | 167 | #ifdef CONFIG_NO_HZ_FULL |
460775df | 168 | cpumask_var_t tick_nohz_full_mask; |
73867dcd | 169 | bool tick_nohz_full_running; |
f009a7a7 | 170 | static atomic_t tick_dep_mask; |
a831881b | 171 | |
f009a7a7 | 172 | static bool check_tick_dependency(atomic_t *dep) |
d027d45d | 173 | { |
f009a7a7 FW |
174 | int val = atomic_read(dep); |
175 | ||
176 | if (val & TICK_DEP_MASK_POSIX_TIMER) { | |
e6e6cc22 | 177 | trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER); |
f009a7a7 | 178 | return true; |
d027d45d FW |
179 | } |
180 | ||
f009a7a7 | 181 | if (val & TICK_DEP_MASK_PERF_EVENTS) { |
e6e6cc22 | 182 | trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS); |
f009a7a7 | 183 | return true; |
d027d45d FW |
184 | } |
185 | ||
f009a7a7 | 186 | if (val & TICK_DEP_MASK_SCHED) { |
e6e6cc22 | 187 | trace_tick_stop(0, TICK_DEP_MASK_SCHED); |
f009a7a7 | 188 | return true; |
d027d45d FW |
189 | } |
190 | ||
f009a7a7 | 191 | if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) { |
e6e6cc22 | 192 | trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE); |
f009a7a7 FW |
193 | return true; |
194 | } | |
195 | ||
196 | return false; | |
d027d45d FW |
197 | } |
198 | ||
57ccdf44 | 199 | static bool can_stop_full_tick(int cpu, struct tick_sched *ts) |
9014c45d | 200 | { |
ebf3adba | 201 | lockdep_assert_irqs_disabled(); |
9014c45d | 202 | |
57ccdf44 WL |
203 | if (unlikely(!cpu_online(cpu))) |
204 | return false; | |
205 | ||
f009a7a7 | 206 | if (check_tick_dependency(&tick_dep_mask)) |
d027d45d | 207 | return false; |
d027d45d | 208 | |
f009a7a7 | 209 | if (check_tick_dependency(&ts->tick_dep_mask)) |
d027d45d | 210 | return false; |
d027d45d | 211 | |
f009a7a7 | 212 | if (check_tick_dependency(¤t->tick_dep_mask)) |
d027d45d | 213 | return false; |
d027d45d | 214 | |
f009a7a7 | 215 | if (check_tick_dependency(¤t->signal->tick_dep_mask)) |
d027d45d | 216 | return false; |
d027d45d | 217 | |
9014c45d FW |
218 | return true; |
219 | } | |
220 | ||
d027d45d | 221 | static void nohz_full_kick_func(struct irq_work *work) |
76c24fb0 | 222 | { |
73738a95 | 223 | /* Empty, the tick restart happens on tick_nohz_irq_exit() */ |
76c24fb0 FW |
224 | } |
225 | ||
226 | static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = { | |
d027d45d | 227 | .func = nohz_full_kick_func, |
76c24fb0 FW |
228 | }; |
229 | ||
40bea039 FW |
230 | /* |
231 | * Kick this CPU if it's full dynticks in order to force it to | |
232 | * re-evaluate its dependency on the tick and restart it if necessary. | |
233 | * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(), | |
234 | * is NMI safe. | |
235 | */ | |
555e0c1e | 236 | static void tick_nohz_full_kick(void) |
40bea039 FW |
237 | { |
238 | if (!tick_nohz_full_cpu(smp_processor_id())) | |
239 | return; | |
240 | ||
56e4dea8 | 241 | irq_work_queue(this_cpu_ptr(&nohz_full_kick_work)); |
40bea039 FW |
242 | } |
243 | ||
76c24fb0 | 244 | /* |
3d36aebc | 245 | * Kick the CPU if it's full dynticks in order to force it to |
76c24fb0 FW |
246 | * re-evaluate its dependency on the tick and restart it if necessary. |
247 | */ | |
3d36aebc | 248 | void tick_nohz_full_kick_cpu(int cpu) |
76c24fb0 | 249 | { |
3d36aebc FW |
250 | if (!tick_nohz_full_cpu(cpu)) |
251 | return; | |
252 | ||
253 | irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu); | |
76c24fb0 FW |
254 | } |
255 | ||
76c24fb0 FW |
256 | /* |
257 | * Kick all full dynticks CPUs in order to force these to re-evaluate | |
258 | * their dependency on the tick and restart it if necessary. | |
259 | */ | |
b7878300 | 260 | static void tick_nohz_full_kick_all(void) |
76c24fb0 | 261 | { |
8537bb95 FW |
262 | int cpu; |
263 | ||
73867dcd | 264 | if (!tick_nohz_full_running) |
76c24fb0 FW |
265 | return; |
266 | ||
267 | preempt_disable(); | |
8537bb95 FW |
268 | for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask) |
269 | tick_nohz_full_kick_cpu(cpu); | |
76c24fb0 FW |
270 | preempt_enable(); |
271 | } | |
272 | ||
f009a7a7 | 273 | static void tick_nohz_dep_set_all(atomic_t *dep, |
d027d45d FW |
274 | enum tick_dep_bits bit) |
275 | { | |
f009a7a7 | 276 | int prev; |
d027d45d | 277 | |
a1cc5bcf | 278 | prev = atomic_fetch_or(BIT(bit), dep); |
d027d45d FW |
279 | if (!prev) |
280 | tick_nohz_full_kick_all(); | |
281 | } | |
282 | ||
283 | /* | |
284 | * Set a global tick dependency. Used by perf events that rely on freq and | |
285 | * by unstable clock. | |
286 | */ | |
287 | void tick_nohz_dep_set(enum tick_dep_bits bit) | |
288 | { | |
289 | tick_nohz_dep_set_all(&tick_dep_mask, bit); | |
290 | } | |
291 | ||
292 | void tick_nohz_dep_clear(enum tick_dep_bits bit) | |
293 | { | |
f009a7a7 | 294 | atomic_andnot(BIT(bit), &tick_dep_mask); |
d027d45d FW |
295 | } |
296 | ||
297 | /* | |
298 | * Set per-CPU tick dependency. Used by scheduler and perf events in order to | |
299 | * manage events throttling. | |
300 | */ | |
301 | void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit) | |
302 | { | |
f009a7a7 | 303 | int prev; |
d027d45d FW |
304 | struct tick_sched *ts; |
305 | ||
306 | ts = per_cpu_ptr(&tick_cpu_sched, cpu); | |
307 | ||
a1cc5bcf | 308 | prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask); |
d027d45d FW |
309 | if (!prev) { |
310 | preempt_disable(); | |
311 | /* Perf needs local kick that is NMI safe */ | |
312 | if (cpu == smp_processor_id()) { | |
313 | tick_nohz_full_kick(); | |
314 | } else { | |
315 | /* Remote irq work not NMI-safe */ | |
316 | if (!WARN_ON_ONCE(in_nmi())) | |
317 | tick_nohz_full_kick_cpu(cpu); | |
318 | } | |
319 | preempt_enable(); | |
320 | } | |
321 | } | |
322 | ||
323 | void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit) | |
324 | { | |
325 | struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu); | |
326 | ||
f009a7a7 | 327 | atomic_andnot(BIT(bit), &ts->tick_dep_mask); |
d027d45d FW |
328 | } |
329 | ||
330 | /* | |
331 | * Set a per-task tick dependency. Posix CPU timers need this in order to elapse | |
332 | * per task timers. | |
333 | */ | |
334 | void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit) | |
335 | { | |
336 | /* | |
337 | * We could optimize this with just kicking the target running the task | |
338 | * if that noise matters for nohz full users. | |
339 | */ | |
340 | tick_nohz_dep_set_all(&tsk->tick_dep_mask, bit); | |
341 | } | |
342 | ||
343 | void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit) | |
344 | { | |
f009a7a7 | 345 | atomic_andnot(BIT(bit), &tsk->tick_dep_mask); |
d027d45d FW |
346 | } |
347 | ||
348 | /* | |
349 | * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse | |
350 | * per process timers. | |
351 | */ | |
352 | void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit) | |
353 | { | |
354 | tick_nohz_dep_set_all(&sig->tick_dep_mask, bit); | |
355 | } | |
356 | ||
357 | void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit) | |
358 | { | |
f009a7a7 | 359 | atomic_andnot(BIT(bit), &sig->tick_dep_mask); |
d027d45d FW |
360 | } |
361 | ||
99e5ada9 FW |
362 | /* |
363 | * Re-evaluate the need for the tick as we switch the current task. | |
364 | * It might need the tick due to per task/process properties: | |
0de7611a | 365 | * perf events, posix CPU timers, ... |
99e5ada9 | 366 | */ |
de734f89 | 367 | void __tick_nohz_task_switch(void) |
99e5ada9 FW |
368 | { |
369 | unsigned long flags; | |
d027d45d | 370 | struct tick_sched *ts; |
99e5ada9 | 371 | |
99e5ada9 FW |
372 | local_irq_save(flags); |
373 | ||
6296ace4 LZ |
374 | if (!tick_nohz_full_cpu(smp_processor_id())) |
375 | goto out; | |
376 | ||
d027d45d | 377 | ts = this_cpu_ptr(&tick_cpu_sched); |
99e5ada9 | 378 | |
d027d45d | 379 | if (ts->tick_stopped) { |
f009a7a7 FW |
380 | if (atomic_read(¤t->tick_dep_mask) || |
381 | atomic_read(¤t->signal->tick_dep_mask)) | |
d027d45d FW |
382 | tick_nohz_full_kick(); |
383 | } | |
6296ace4 | 384 | out: |
99e5ada9 FW |
385 | local_irq_restore(flags); |
386 | } | |
387 | ||
6f1982fe FW |
388 | /* Get the boot-time nohz CPU list from the kernel parameters. */ |
389 | void __init tick_nohz_full_setup(cpumask_var_t cpumask) | |
a831881b | 390 | { |
73867dcd | 391 | alloc_bootmem_cpumask_var(&tick_nohz_full_mask); |
6f1982fe | 392 | cpumask_copy(tick_nohz_full_mask, cpumask); |
73867dcd | 393 | tick_nohz_full_running = true; |
a831881b | 394 | } |
a831881b | 395 | |
31eff243 | 396 | static int tick_nohz_cpu_down(unsigned int cpu) |
a382bf93 | 397 | { |
31eff243 SAS |
398 | /* |
399 | * The boot CPU handles housekeeping duty (unbound timers, | |
400 | * workqueues, timekeeping, ...) on behalf of full dynticks | |
401 | * CPUs. It must remain online when nohz full is enabled. | |
402 | */ | |
403 | if (tick_nohz_full_running && tick_do_timer_cpu == cpu) | |
404 | return -EBUSY; | |
405 | return 0; | |
a382bf93 FW |
406 | } |
407 | ||
f98823ac FW |
408 | static int tick_nohz_init_all(void) |
409 | { | |
410 | int err = -1; | |
411 | ||
412 | #ifdef CONFIG_NO_HZ_FULL_ALL | |
73867dcd | 413 | if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) { |
4327b15f | 414 | WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n"); |
c0f489d2 PM |
415 | return err; |
416 | } | |
f98823ac | 417 | err = 0; |
73867dcd | 418 | cpumask_setall(tick_nohz_full_mask); |
73867dcd | 419 | tick_nohz_full_running = true; |
f98823ac FW |
420 | #endif |
421 | return err; | |
422 | } | |
423 | ||
d1e43fa5 | 424 | void __init tick_nohz_init(void) |
a831881b | 425 | { |
31eff243 | 426 | int cpu, ret; |
d1e43fa5 | 427 | |
73867dcd | 428 | if (!tick_nohz_full_running) { |
f98823ac FW |
429 | if (tick_nohz_init_all() < 0) |
430 | return; | |
431 | } | |
d1e43fa5 | 432 | |
9b01f5bf FW |
433 | /* |
434 | * Full dynticks uses irq work to drive the tick rescheduling on safe | |
435 | * locking contexts. But then we need irq work to raise its own | |
436 | * interrupts to avoid circular dependency on the tick | |
437 | */ | |
438 | if (!arch_irq_work_has_interrupt()) { | |
a395d6a7 | 439 | pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n"); |
9b01f5bf | 440 | cpumask_clear(tick_nohz_full_mask); |
9b01f5bf FW |
441 | tick_nohz_full_running = false; |
442 | return; | |
443 | } | |
444 | ||
4327b15f FW |
445 | cpu = smp_processor_id(); |
446 | ||
447 | if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) { | |
a395d6a7 JP |
448 | pr_warn("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", |
449 | cpu); | |
4327b15f FW |
450 | cpumask_clear_cpu(cpu, tick_nohz_full_mask); |
451 | } | |
452 | ||
73867dcd | 453 | for_each_cpu(cpu, tick_nohz_full_mask) |
2e709338 FW |
454 | context_tracking_cpu_set(cpu); |
455 | ||
31eff243 SAS |
456 | ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, |
457 | "kernel/nohz:predown", NULL, | |
458 | tick_nohz_cpu_down); | |
459 | WARN_ON(ret < 0); | |
ffda22c1 TH |
460 | pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n", |
461 | cpumask_pr_args(tick_nohz_full_mask)); | |
a831881b | 462 | } |
a831881b FW |
463 | #endif |
464 | ||
79bf2bb3 TG |
465 | /* |
466 | * NOHZ - aka dynamic tick functionality | |
467 | */ | |
3451d024 | 468 | #ifdef CONFIG_NO_HZ_COMMON |
79bf2bb3 TG |
469 | /* |
470 | * NO HZ enabled ? | |
471 | */ | |
4cc7ecb7 | 472 | bool tick_nohz_enabled __read_mostly = true; |
bc7a34b8 | 473 | unsigned long tick_nohz_active __read_mostly; |
79bf2bb3 TG |
474 | /* |
475 | * Enable / Disable tickless mode | |
476 | */ | |
477 | static int __init setup_tick_nohz(char *str) | |
478 | { | |
4cc7ecb7 | 479 | return (kstrtobool(str, &tick_nohz_enabled) == 0); |
79bf2bb3 TG |
480 | } |
481 | ||
482 | __setup("nohz=", setup_tick_nohz); | |
483 | ||
a3642983 | 484 | bool tick_nohz_tick_stopped(void) |
c1797baf TG |
485 | { |
486 | return __this_cpu_read(tick_cpu_sched.tick_stopped); | |
487 | } | |
488 | ||
79bf2bb3 TG |
489 | /** |
490 | * tick_nohz_update_jiffies - update jiffies when idle was interrupted | |
491 | * | |
492 | * Called from interrupt entry when the CPU was idle | |
493 | * | |
494 | * In case the sched_tick was stopped on this CPU, we have to check if jiffies | |
495 | * must be updated. Otherwise an interrupt handler could use a stale jiffy | |
0de7611a IM |
496 | * value. We do this unconditionally on any CPU, as we don't know whether the |
497 | * CPU, which has the update task assigned is in a long sleep. | |
79bf2bb3 | 498 | */ |
eed3b9cf | 499 | static void tick_nohz_update_jiffies(ktime_t now) |
79bf2bb3 | 500 | { |
79bf2bb3 | 501 | unsigned long flags; |
79bf2bb3 | 502 | |
e8fcaa5c | 503 | __this_cpu_write(tick_cpu_sched.idle_waketime, now); |
79bf2bb3 TG |
504 | |
505 | local_irq_save(flags); | |
506 | tick_do_update_jiffies64(now); | |
507 | local_irq_restore(flags); | |
02ff3755 | 508 | |
03e0d461 | 509 | touch_softlockup_watchdog_sched(); |
79bf2bb3 TG |
510 | } |
511 | ||
595aac48 | 512 | /* |
0de7611a | 513 | * Updates the per-CPU time idle statistics counters |
595aac48 | 514 | */ |
8d63bf94 | 515 | static void |
8c215bd3 | 516 | update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time) |
6378ddb5 | 517 | { |
eed3b9cf | 518 | ktime_t delta; |
6378ddb5 | 519 | |
595aac48 AV |
520 | if (ts->idle_active) { |
521 | delta = ktime_sub(now, ts->idle_entrytime); | |
8c215bd3 | 522 | if (nr_iowait_cpu(cpu) > 0) |
0224cf4c | 523 | ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); |
6beea0cd MH |
524 | else |
525 | ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); | |
8c7b09f4 | 526 | ts->idle_entrytime = now; |
595aac48 | 527 | } |
8d63bf94 | 528 | |
e0e37c20 | 529 | if (last_update_time) |
8d63bf94 AV |
530 | *last_update_time = ktime_to_us(now); |
531 | ||
595aac48 AV |
532 | } |
533 | ||
e8fcaa5c | 534 | static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now) |
595aac48 | 535 | { |
e8fcaa5c | 536 | update_ts_time_stats(smp_processor_id(), ts, now, NULL); |
eed3b9cf | 537 | ts->idle_active = 0; |
56c7426b | 538 | |
ac1e843f | 539 | sched_clock_idle_wakeup_event(); |
6378ddb5 VP |
540 | } |
541 | ||
e8fcaa5c | 542 | static ktime_t tick_nohz_start_idle(struct tick_sched *ts) |
6378ddb5 | 543 | { |
430ee881 | 544 | ktime_t now = ktime_get(); |
595aac48 | 545 | |
6378ddb5 VP |
546 | ts->idle_entrytime = now; |
547 | ts->idle_active = 1; | |
56c7426b | 548 | sched_clock_idle_sleep_event(); |
6378ddb5 VP |
549 | return now; |
550 | } | |
551 | ||
b1f724c3 | 552 | /** |
0de7611a | 553 | * get_cpu_idle_time_us - get the total idle time of a CPU |
b1f724c3 | 554 | * @cpu: CPU number to query |
09a1d34f MH |
555 | * @last_update_time: variable to store update time in. Do not update |
556 | * counters if NULL. | |
b1f724c3 | 557 | * |
6168f8ed | 558 | * Return the cumulative idle time (since boot) for a given |
6beea0cd | 559 | * CPU, in microseconds. |
b1f724c3 AV |
560 | * |
561 | * This time is measured via accounting rather than sampling, | |
562 | * and is as accurate as ktime_get() is. | |
563 | * | |
564 | * This function returns -1 if NOHZ is not enabled. | |
565 | */ | |
6378ddb5 VP |
566 | u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) |
567 | { | |
568 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
09a1d34f | 569 | ktime_t now, idle; |
6378ddb5 | 570 | |
d689fe22 | 571 | if (!tick_nohz_active) |
8083e4ad | 572 | return -1; |
573 | ||
09a1d34f MH |
574 | now = ktime_get(); |
575 | if (last_update_time) { | |
576 | update_ts_time_stats(cpu, ts, now, last_update_time); | |
577 | idle = ts->idle_sleeptime; | |
578 | } else { | |
579 | if (ts->idle_active && !nr_iowait_cpu(cpu)) { | |
580 | ktime_t delta = ktime_sub(now, ts->idle_entrytime); | |
581 | ||
582 | idle = ktime_add(ts->idle_sleeptime, delta); | |
583 | } else { | |
584 | idle = ts->idle_sleeptime; | |
585 | } | |
586 | } | |
587 | ||
588 | return ktime_to_us(idle); | |
8083e4ad | 589 | |
6378ddb5 | 590 | } |
8083e4ad | 591 | EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); |
6378ddb5 | 592 | |
6beea0cd | 593 | /** |
0de7611a | 594 | * get_cpu_iowait_time_us - get the total iowait time of a CPU |
0224cf4c | 595 | * @cpu: CPU number to query |
09a1d34f MH |
596 | * @last_update_time: variable to store update time in. Do not update |
597 | * counters if NULL. | |
0224cf4c | 598 | * |
6168f8ed | 599 | * Return the cumulative iowait time (since boot) for a given |
0224cf4c AV |
600 | * CPU, in microseconds. |
601 | * | |
602 | * This time is measured via accounting rather than sampling, | |
603 | * and is as accurate as ktime_get() is. | |
604 | * | |
605 | * This function returns -1 if NOHZ is not enabled. | |
606 | */ | |
607 | u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) | |
608 | { | |
609 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
09a1d34f | 610 | ktime_t now, iowait; |
0224cf4c | 611 | |
d689fe22 | 612 | if (!tick_nohz_active) |
0224cf4c AV |
613 | return -1; |
614 | ||
09a1d34f MH |
615 | now = ktime_get(); |
616 | if (last_update_time) { | |
617 | update_ts_time_stats(cpu, ts, now, last_update_time); | |
618 | iowait = ts->iowait_sleeptime; | |
619 | } else { | |
620 | if (ts->idle_active && nr_iowait_cpu(cpu) > 0) { | |
621 | ktime_t delta = ktime_sub(now, ts->idle_entrytime); | |
0224cf4c | 622 | |
09a1d34f MH |
623 | iowait = ktime_add(ts->iowait_sleeptime, delta); |
624 | } else { | |
625 | iowait = ts->iowait_sleeptime; | |
626 | } | |
627 | } | |
0224cf4c | 628 | |
09a1d34f | 629 | return ktime_to_us(iowait); |
0224cf4c AV |
630 | } |
631 | EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); | |
632 | ||
0ff53d09 TG |
633 | static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) |
634 | { | |
635 | hrtimer_cancel(&ts->sched_timer); | |
636 | hrtimer_set_expires(&ts->sched_timer, ts->last_tick); | |
637 | ||
638 | /* Forward the time to expire in the future */ | |
639 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
640 | ||
641 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) | |
642 | hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); | |
643 | else | |
644 | tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); | |
411fe24e FW |
645 | |
646 | /* | |
647 | * Reset to make sure next tick stop doesn't get fooled by past | |
648 | * cached clock deadline. | |
649 | */ | |
650 | ts->next_tick = 0; | |
0ff53d09 TG |
651 | } |
652 | ||
5d62c183 TG |
653 | static inline bool local_timer_softirq_pending(void) |
654 | { | |
655 | return local_softirq_pending() & TIMER_SOFTIRQ; | |
656 | } | |
657 | ||
84bf1bcc FW |
658 | static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts, |
659 | ktime_t now, int cpu) | |
79bf2bb3 | 660 | { |
22127e93 | 661 | struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev); |
c1ad348b TG |
662 | u64 basemono, next_tick, next_tmr, next_rcu, delta, expires; |
663 | unsigned long seq, basejiff; | |
664 | ktime_t tick; | |
855a0fc3 | 665 | |
79bf2bb3 TG |
666 | /* Read jiffies and the time when jiffies were updated last */ |
667 | do { | |
d6ad4187 | 668 | seq = read_seqbegin(&jiffies_lock); |
2456e855 | 669 | basemono = last_jiffies_update; |
c1ad348b | 670 | basejiff = jiffies; |
d6ad4187 | 671 | } while (read_seqretry(&jiffies_lock, seq)); |
c1ad348b | 672 | ts->last_jiffies = basejiff; |
79bf2bb3 | 673 | |
5d62c183 TG |
674 | /* |
675 | * Keep the periodic tick, when RCU, architecture or irq_work | |
676 | * requests it. | |
677 | * Aside of that check whether the local timer softirq is | |
678 | * pending. If so its a bad idea to call get_next_timer_interrupt() | |
679 | * because there is an already expired timer, so it will request | |
680 | * immeditate expiry, which rearms the hardware timer with a | |
681 | * minimal delta which brings us back to this place | |
682 | * immediately. Lather, rinse and repeat... | |
683 | */ | |
684 | if (rcu_needs_cpu(basemono, &next_rcu) || arch_needs_cpu() || | |
685 | irq_work_needs_cpu() || local_timer_softirq_pending()) { | |
c1ad348b | 686 | next_tick = basemono + TICK_NSEC; |
3c5d92a0 | 687 | } else { |
c1ad348b TG |
688 | /* |
689 | * Get the next pending timer. If high resolution | |
690 | * timers are enabled this only takes the timer wheel | |
691 | * timers into account. If high resolution timers are | |
692 | * disabled this also looks at the next expiring | |
693 | * hrtimer. | |
694 | */ | |
695 | next_tmr = get_next_timer_interrupt(basejiff, basemono); | |
696 | ts->next_timer = next_tmr; | |
697 | /* Take the next rcu event into account */ | |
698 | next_tick = next_rcu < next_tmr ? next_rcu : next_tmr; | |
3c5d92a0 | 699 | } |
47aa8b6c | 700 | |
c1ad348b TG |
701 | /* |
702 | * If the tick is due in the next period, keep it ticking or | |
82bbe34b | 703 | * force prod the timer. |
c1ad348b TG |
704 | */ |
705 | delta = next_tick - basemono; | |
706 | if (delta <= (u64)TICK_NSEC) { | |
a683f390 TG |
707 | /* |
708 | * Tell the timer code that the base is not idle, i.e. undo | |
709 | * the effect of get_next_timer_interrupt(): | |
710 | */ | |
711 | timer_clear_idle(); | |
82bbe34b PZ |
712 | /* |
713 | * We've not stopped the tick yet, and there's a timer in the | |
714 | * next period, so no point in stopping it either, bail. | |
715 | */ | |
f99973e1 FW |
716 | if (!ts->tick_stopped) { |
717 | tick = 0; | |
157d29e1 TG |
718 | goto out; |
719 | } | |
720 | } | |
721 | ||
79bf2bb3 | 722 | /* |
0de7611a IM |
723 | * If this CPU is the one which updates jiffies, then give up |
724 | * the assignment and let it be taken by the CPU which runs | |
725 | * the tick timer next, which might be this CPU as well. If we | |
157d29e1 TG |
726 | * don't drop this here the jiffies might be stale and |
727 | * do_timer() never invoked. Keep track of the fact that it | |
0de7611a | 728 | * was the one which had the do_timer() duty last. If this CPU |
157d29e1 | 729 | * is the one which had the do_timer() duty last, we limit the |
6168f8ed | 730 | * sleep time to the timekeeping max_deferment value. |
c1ad348b | 731 | * Otherwise we can sleep as long as we want. |
79bf2bb3 | 732 | */ |
c1ad348b | 733 | delta = timekeeping_max_deferment(); |
157d29e1 TG |
734 | if (cpu == tick_do_timer_cpu) { |
735 | tick_do_timer_cpu = TICK_DO_TIMER_NONE; | |
736 | ts->do_timer_last = 1; | |
737 | } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { | |
c1ad348b | 738 | delta = KTIME_MAX; |
157d29e1 TG |
739 | ts->do_timer_last = 0; |
740 | } else if (!ts->do_timer_last) { | |
c1ad348b | 741 | delta = KTIME_MAX; |
157d29e1 | 742 | } |
27185016 | 743 | |
265f22a9 | 744 | #ifdef CONFIG_NO_HZ_FULL |
c1ad348b | 745 | /* Limit the tick delta to the maximum scheduler deferment */ |
157d29e1 | 746 | if (!ts->inidle) |
c1ad348b | 747 | delta = min(delta, scheduler_tick_max_deferment()); |
265f22a9 FW |
748 | #endif |
749 | ||
c1ad348b TG |
750 | /* Calculate the next expiry time */ |
751 | if (delta < (KTIME_MAX - basemono)) | |
752 | expires = basemono + delta; | |
157d29e1 | 753 | else |
c1ad348b TG |
754 | expires = KTIME_MAX; |
755 | ||
756 | expires = min_t(u64, expires, next_tick); | |
2456e855 | 757 | tick = expires; |
00147449 | 758 | |
157d29e1 | 759 | /* Skip reprogram of event if its not changed */ |
411fe24e FW |
760 | if (ts->tick_stopped && (expires == ts->next_tick)) { |
761 | /* Sanity check: make sure clockevent is actually programmed */ | |
d4af6d93 | 762 | if (tick == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer)) |
ce6cf9a1 | 763 | goto out; |
411fe24e FW |
764 | |
765 | WARN_ON_ONCE(1); | |
766 | printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n", | |
767 | basemono, ts->next_tick, dev->next_event, | |
768 | hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer)); | |
ce6cf9a1 | 769 | } |
84bf1bcc | 770 | |
157d29e1 TG |
771 | /* |
772 | * nohz_stop_sched_tick can be called several times before | |
773 | * the nohz_restart_sched_tick is called. This happens when | |
774 | * interrupts arrive which do not cause a reschedule. In the | |
775 | * first call we save the current tick time, so we can restart | |
776 | * the scheduler tick in nohz_restart_sched_tick. | |
777 | */ | |
778 | if (!ts->tick_stopped) { | |
3c85d6db | 779 | calc_load_nohz_start(); |
1f41906a | 780 | cpu_load_update_nohz_start(); |
62cb1188 | 781 | quiet_vmstat(); |
d3ed7824 | 782 | |
157d29e1 TG |
783 | ts->last_tick = hrtimer_get_expires(&ts->sched_timer); |
784 | ts->tick_stopped = 1; | |
e6e6cc22 | 785 | trace_tick_stop(1, TICK_DEP_MASK_NONE); |
157d29e1 | 786 | } |
eaad084b | 787 | |
411fe24e FW |
788 | ts->next_tick = tick; |
789 | ||
157d29e1 | 790 | /* |
c1ad348b TG |
791 | * If the expiration time == KTIME_MAX, then we simply stop |
792 | * the tick timer. | |
157d29e1 | 793 | */ |
c1ad348b | 794 | if (unlikely(expires == KTIME_MAX)) { |
157d29e1 TG |
795 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) |
796 | hrtimer_cancel(&ts->sched_timer); | |
797 | goto out; | |
79bf2bb3 | 798 | } |
0ff53d09 | 799 | |
d4af6d93 FW |
800 | hrtimer_set_expires(&ts->sched_timer, tick); |
801 | ||
157d29e1 | 802 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) |
d4af6d93 | 803 | hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); |
157d29e1 | 804 | else |
c1ad348b | 805 | tick_program_event(tick, 1); |
79bf2bb3 | 806 | out: |
411fe24e FW |
807 | /* |
808 | * Update the estimated sleep length until the next timer | |
809 | * (not only the tick). | |
810 | */ | |
4f86d3a8 | 811 | ts->sleep_length = ktime_sub(dev->next_event, now); |
c1ad348b | 812 | return tick; |
280f0677 FW |
813 | } |
814 | ||
1f41906a | 815 | static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now) |
59d2c7ca FW |
816 | { |
817 | /* Update jiffies first */ | |
818 | tick_do_update_jiffies64(now); | |
1f41906a | 819 | cpu_load_update_nohz_stop(); |
59d2c7ca | 820 | |
a683f390 TG |
821 | /* |
822 | * Clear the timer idle flag, so we avoid IPIs on remote queueing and | |
823 | * the clock forward checks in the enqueue path: | |
824 | */ | |
825 | timer_clear_idle(); | |
826 | ||
3c85d6db | 827 | calc_load_nohz_stop(); |
03e0d461 | 828 | touch_softlockup_watchdog_sched(); |
59d2c7ca FW |
829 | /* |
830 | * Cancel the scheduled timer and restore the tick | |
831 | */ | |
832 | ts->tick_stopped = 0; | |
833 | ts->idle_exittime = now; | |
834 | ||
835 | tick_nohz_restart(ts, now); | |
836 | } | |
73738a95 FW |
837 | |
838 | static void tick_nohz_full_update_tick(struct tick_sched *ts) | |
5811d996 FW |
839 | { |
840 | #ifdef CONFIG_NO_HZ_FULL | |
e9a2eb40 | 841 | int cpu = smp_processor_id(); |
5811d996 | 842 | |
59449359 | 843 | if (!tick_nohz_full_cpu(cpu)) |
e9a2eb40 | 844 | return; |
5811d996 | 845 | |
e9a2eb40 AS |
846 | if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE) |
847 | return; | |
5811d996 | 848 | |
57ccdf44 | 849 | if (can_stop_full_tick(cpu, ts)) |
73738a95 FW |
850 | tick_nohz_stop_sched_tick(ts, ktime_get(), cpu); |
851 | else if (ts->tick_stopped) | |
1f41906a | 852 | tick_nohz_restart_sched_tick(ts, ktime_get()); |
5811d996 FW |
853 | #endif |
854 | } | |
855 | ||
5b39939a FW |
856 | static bool can_stop_idle_tick(int cpu, struct tick_sched *ts) |
857 | { | |
858 | /* | |
0de7611a | 859 | * If this CPU is offline and it is the one which updates |
5b39939a | 860 | * jiffies, then give up the assignment and let it be taken by |
0de7611a | 861 | * the CPU which runs the tick timer next. If we don't drop |
5b39939a FW |
862 | * this here the jiffies might be stale and do_timer() never |
863 | * invoked. | |
864 | */ | |
865 | if (unlikely(!cpu_online(cpu))) { | |
866 | if (cpu == tick_do_timer_cpu) | |
867 | tick_do_timer_cpu = TICK_DO_TIMER_NONE; | |
411fe24e FW |
868 | /* |
869 | * Make sure the CPU doesn't get fooled by obsolete tick | |
870 | * deadline if it comes back online later. | |
871 | */ | |
872 | ts->next_tick = 0; | |
f7ea0fd6 | 873 | return false; |
5b39939a FW |
874 | } |
875 | ||
0e576acb | 876 | if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) { |
2456e855 | 877 | ts->sleep_length = NSEC_PER_SEC / HZ; |
5b39939a | 878 | return false; |
0e576acb | 879 | } |
5b39939a FW |
880 | |
881 | if (need_resched()) | |
882 | return false; | |
883 | ||
884 | if (unlikely(local_softirq_pending() && cpu_online(cpu))) { | |
885 | static int ratelimit; | |
886 | ||
803b0eba PM |
887 | if (ratelimit < 10 && |
888 | (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) { | |
cfea7d7e RV |
889 | pr_warn("NOHZ: local_softirq_pending %02x\n", |
890 | (unsigned int) local_softirq_pending()); | |
5b39939a FW |
891 | ratelimit++; |
892 | } | |
893 | return false; | |
894 | } | |
895 | ||
460775df | 896 | if (tick_nohz_full_enabled()) { |
a382bf93 FW |
897 | /* |
898 | * Keep the tick alive to guarantee timekeeping progression | |
899 | * if there are full dynticks CPUs around | |
900 | */ | |
901 | if (tick_do_timer_cpu == cpu) | |
902 | return false; | |
903 | /* | |
904 | * Boot safety: make sure the timekeeping duty has been | |
905 | * assigned before entering dyntick-idle mode, | |
906 | */ | |
907 | if (tick_do_timer_cpu == TICK_DO_TIMER_NONE) | |
908 | return false; | |
909 | } | |
910 | ||
5b39939a FW |
911 | return true; |
912 | } | |
913 | ||
19f5f736 FW |
914 | static void __tick_nohz_idle_enter(struct tick_sched *ts) |
915 | { | |
84bf1bcc | 916 | ktime_t now, expires; |
5b39939a | 917 | int cpu = smp_processor_id(); |
19f5f736 | 918 | |
08d07259 WL |
919 | now = tick_nohz_start_idle(ts); |
920 | ||
5b39939a FW |
921 | if (can_stop_idle_tick(cpu, ts)) { |
922 | int was_stopped = ts->tick_stopped; | |
923 | ||
924 | ts->idle_calls++; | |
84bf1bcc FW |
925 | |
926 | expires = tick_nohz_stop_sched_tick(ts, now, cpu); | |
2456e855 | 927 | if (expires > 0LL) { |
84bf1bcc FW |
928 | ts->idle_sleeps++; |
929 | ts->idle_expires = expires; | |
930 | } | |
5b39939a | 931 | |
a0db971e | 932 | if (!was_stopped && ts->tick_stopped) { |
5b39939a | 933 | ts->idle_jiffies = ts->last_jiffies; |
a0db971e FW |
934 | nohz_balance_enter_idle(cpu); |
935 | } | |
5b39939a | 936 | } |
280f0677 FW |
937 | } |
938 | ||
939 | /** | |
940 | * tick_nohz_idle_enter - stop the idle tick from the idle task | |
941 | * | |
942 | * When the next event is more than a tick into the future, stop the idle tick | |
943 | * Called when we start the idle loop. | |
2bbb6817 | 944 | * |
1268fbc7 | 945 | * The arch is responsible of calling: |
2bbb6817 FW |
946 | * |
947 | * - rcu_idle_enter() after its last use of RCU before the CPU is put | |
948 | * to sleep. | |
949 | * - rcu_idle_exit() before the first use of RCU after the CPU is woken up. | |
280f0677 | 950 | */ |
1268fbc7 | 951 | void tick_nohz_idle_enter(void) |
280f0677 FW |
952 | { |
953 | struct tick_sched *ts; | |
954 | ||
ebf3adba | 955 | lockdep_assert_irqs_enabled(); |
0db49b72 | 956 | /* |
0de7611a IM |
957 | * Update the idle state in the scheduler domain hierarchy |
958 | * when tick_nohz_stop_sched_tick() is called from the idle loop. | |
959 | * State will be updated to busy during the first busy tick after | |
960 | * exiting idle. | |
961 | */ | |
0db49b72 LT |
962 | set_cpu_sd_state_idle(); |
963 | ||
1268fbc7 FW |
964 | local_irq_disable(); |
965 | ||
22127e93 | 966 | ts = this_cpu_ptr(&tick_cpu_sched); |
280f0677 | 967 | ts->inidle = 1; |
19f5f736 | 968 | __tick_nohz_idle_enter(ts); |
1268fbc7 FW |
969 | |
970 | local_irq_enable(); | |
280f0677 FW |
971 | } |
972 | ||
973 | /** | |
974 | * tick_nohz_irq_exit - update next tick event from interrupt exit | |
975 | * | |
976 | * When an interrupt fires while we are idle and it doesn't cause | |
977 | * a reschedule, it may still add, modify or delete a timer, enqueue | |
978 | * an RCU callback, etc... | |
979 | * So we need to re-calculate and reprogram the next tick event. | |
980 | */ | |
981 | void tick_nohz_irq_exit(void) | |
982 | { | |
22127e93 | 983 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
280f0677 | 984 | |
14851912 | 985 | if (ts->inidle) |
5811d996 | 986 | __tick_nohz_idle_enter(ts); |
14851912 | 987 | else |
73738a95 | 988 | tick_nohz_full_update_tick(ts); |
79bf2bb3 TG |
989 | } |
990 | ||
4f86d3a8 LB |
991 | /** |
992 | * tick_nohz_get_sleep_length - return the length of the current sleep | |
993 | * | |
994 | * Called from power state control code with interrupts disabled | |
995 | */ | |
996 | ktime_t tick_nohz_get_sleep_length(void) | |
997 | { | |
22127e93 | 998 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
4f86d3a8 LB |
999 | |
1000 | return ts->sleep_length; | |
1001 | } | |
1002 | ||
466a2b42 JF |
1003 | /** |
1004 | * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value | |
1005 | * for a particular CPU. | |
1006 | * | |
1007 | * Called from the schedutil frequency scaling governor in scheduler context. | |
1008 | */ | |
1009 | unsigned long tick_nohz_get_idle_calls_cpu(int cpu) | |
1010 | { | |
1011 | struct tick_sched *ts = tick_get_tick_sched(cpu); | |
1012 | ||
1013 | return ts->idle_calls; | |
1014 | } | |
1015 | ||
b7eaf1aa RW |
1016 | /** |
1017 | * tick_nohz_get_idle_calls - return the current idle calls counter value | |
1018 | * | |
1019 | * Called from the schedutil frequency scaling governor in scheduler context. | |
1020 | */ | |
1021 | unsigned long tick_nohz_get_idle_calls(void) | |
1022 | { | |
1023 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); | |
1024 | ||
1025 | return ts->idle_calls; | |
1026 | } | |
1027 | ||
2ac0d98f FW |
1028 | static void tick_nohz_account_idle_ticks(struct tick_sched *ts) |
1029 | { | |
3f4724ea | 1030 | #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE |
2ac0d98f | 1031 | unsigned long ticks; |
3f4724ea | 1032 | |
55dbdcfa | 1033 | if (vtime_accounting_cpu_enabled()) |
3f4724ea | 1034 | return; |
79bf2bb3 TG |
1035 | /* |
1036 | * We stopped the tick in idle. Update process times would miss the | |
1037 | * time we slept as update_process_times does only a 1 tick | |
1038 | * accounting. Enforce that this is accounted to idle ! | |
1039 | */ | |
1040 | ticks = jiffies - ts->idle_jiffies; | |
1041 | /* | |
1042 | * We might be one off. Do not randomly account a huge number of ticks! | |
1043 | */ | |
79741dd3 MS |
1044 | if (ticks && ticks < LONG_MAX) |
1045 | account_idle_ticks(ticks); | |
1046 | #endif | |
19f5f736 FW |
1047 | } |
1048 | ||
79bf2bb3 | 1049 | /** |
280f0677 | 1050 | * tick_nohz_idle_exit - restart the idle tick from the idle task |
79bf2bb3 TG |
1051 | * |
1052 | * Restart the idle tick when the CPU is woken up from idle | |
280f0677 FW |
1053 | * This also exit the RCU extended quiescent state. The CPU |
1054 | * can use RCU again after this function is called. | |
79bf2bb3 | 1055 | */ |
280f0677 | 1056 | void tick_nohz_idle_exit(void) |
79bf2bb3 | 1057 | { |
4a32fea9 | 1058 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
6378ddb5 | 1059 | ktime_t now; |
79bf2bb3 | 1060 | |
6378ddb5 | 1061 | local_irq_disable(); |
2bbb6817 | 1062 | |
15f827be FW |
1063 | WARN_ON_ONCE(!ts->inidle); |
1064 | ||
1065 | ts->inidle = 0; | |
1066 | ||
1067 | if (ts->idle_active || ts->tick_stopped) | |
eed3b9cf MS |
1068 | now = ktime_get(); |
1069 | ||
1070 | if (ts->idle_active) | |
e8fcaa5c | 1071 | tick_nohz_stop_idle(ts, now); |
6378ddb5 | 1072 | |
2ac0d98f | 1073 | if (ts->tick_stopped) { |
1f41906a | 1074 | tick_nohz_restart_sched_tick(ts, now); |
2ac0d98f | 1075 | tick_nohz_account_idle_ticks(ts); |
6378ddb5 | 1076 | } |
79bf2bb3 | 1077 | |
79bf2bb3 TG |
1078 | local_irq_enable(); |
1079 | } | |
1080 | ||
79bf2bb3 TG |
1081 | /* |
1082 | * The nohz low res interrupt handler | |
1083 | */ | |
1084 | static void tick_nohz_handler(struct clock_event_device *dev) | |
1085 | { | |
22127e93 | 1086 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
79bf2bb3 TG |
1087 | struct pt_regs *regs = get_irq_regs(); |
1088 | ktime_t now = ktime_get(); | |
1089 | ||
2456e855 | 1090 | dev->next_event = KTIME_MAX; |
79bf2bb3 | 1091 | |
5bb96226 | 1092 | tick_sched_do_timer(now); |
9e8f559b | 1093 | tick_sched_handle(ts, regs); |
79bf2bb3 | 1094 | |
b5e995e6 VK |
1095 | /* No need to reprogram if we are running tickless */ |
1096 | if (unlikely(ts->tick_stopped)) | |
1097 | return; | |
1098 | ||
0ff53d09 TG |
1099 | hrtimer_forward(&ts->sched_timer, now, tick_period); |
1100 | tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); | |
79bf2bb3 TG |
1101 | } |
1102 | ||
bc7a34b8 TG |
1103 | static inline void tick_nohz_activate(struct tick_sched *ts, int mode) |
1104 | { | |
1105 | if (!tick_nohz_enabled) | |
1106 | return; | |
1107 | ts->nohz_mode = mode; | |
1108 | /* One update is enough */ | |
1109 | if (!test_and_set_bit(0, &tick_nohz_active)) | |
ae67bada | 1110 | timers_update_nohz(); |
bc7a34b8 TG |
1111 | } |
1112 | ||
79bf2bb3 TG |
1113 | /** |
1114 | * tick_nohz_switch_to_nohz - switch to nohz mode | |
1115 | */ | |
1116 | static void tick_nohz_switch_to_nohz(void) | |
1117 | { | |
22127e93 | 1118 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
79bf2bb3 TG |
1119 | ktime_t next; |
1120 | ||
27630532 | 1121 | if (!tick_nohz_enabled) |
79bf2bb3 TG |
1122 | return; |
1123 | ||
6b442bc8 | 1124 | if (tick_switch_to_oneshot(tick_nohz_handler)) |
79bf2bb3 | 1125 | return; |
6b442bc8 | 1126 | |
79bf2bb3 TG |
1127 | /* |
1128 | * Recycle the hrtimer in ts, so we can share the | |
1129 | * hrtimer_forward with the highres code. | |
1130 | */ | |
1131 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); | |
1132 | /* Get the next period */ | |
1133 | next = tick_init_jiffy_update(); | |
1134 | ||
0ff53d09 | 1135 | hrtimer_set_expires(&ts->sched_timer, next); |
1ca8ec53 WL |
1136 | hrtimer_forward_now(&ts->sched_timer, tick_period); |
1137 | tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); | |
bc7a34b8 | 1138 | tick_nohz_activate(ts, NOHZ_MODE_LOWRES); |
79bf2bb3 TG |
1139 | } |
1140 | ||
5acac1be | 1141 | static inline void tick_nohz_irq_enter(void) |
eed3b9cf | 1142 | { |
4a32fea9 | 1143 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
eed3b9cf MS |
1144 | ktime_t now; |
1145 | ||
1146 | if (!ts->idle_active && !ts->tick_stopped) | |
1147 | return; | |
1148 | now = ktime_get(); | |
1149 | if (ts->idle_active) | |
e8fcaa5c | 1150 | tick_nohz_stop_idle(ts, now); |
ff006732 | 1151 | if (ts->tick_stopped) |
eed3b9cf | 1152 | tick_nohz_update_jiffies(now); |
eed3b9cf MS |
1153 | } |
1154 | ||
79bf2bb3 TG |
1155 | #else |
1156 | ||
1157 | static inline void tick_nohz_switch_to_nohz(void) { } | |
5acac1be | 1158 | static inline void tick_nohz_irq_enter(void) { } |
bc7a34b8 | 1159 | static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { } |
79bf2bb3 | 1160 | |
3451d024 | 1161 | #endif /* CONFIG_NO_HZ_COMMON */ |
79bf2bb3 | 1162 | |
719254fa TG |
1163 | /* |
1164 | * Called from irq_enter to notify about the possible interruption of idle() | |
1165 | */ | |
5acac1be | 1166 | void tick_irq_enter(void) |
719254fa | 1167 | { |
e8fcaa5c | 1168 | tick_check_oneshot_broadcast_this_cpu(); |
5acac1be | 1169 | tick_nohz_irq_enter(); |
719254fa TG |
1170 | } |
1171 | ||
79bf2bb3 TG |
1172 | /* |
1173 | * High resolution timer specific code | |
1174 | */ | |
1175 | #ifdef CONFIG_HIGH_RES_TIMERS | |
1176 | /* | |
4c9dc641 | 1177 | * We rearm the timer until we get disabled by the idle code. |
351f181f | 1178 | * Called with interrupts disabled. |
79bf2bb3 TG |
1179 | */ |
1180 | static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) | |
1181 | { | |
1182 | struct tick_sched *ts = | |
1183 | container_of(timer, struct tick_sched, sched_timer); | |
79bf2bb3 TG |
1184 | struct pt_regs *regs = get_irq_regs(); |
1185 | ktime_t now = ktime_get(); | |
d3ed7824 | 1186 | |
5bb96226 | 1187 | tick_sched_do_timer(now); |
79bf2bb3 TG |
1188 | |
1189 | /* | |
1190 | * Do not call, when we are not in irq context and have | |
1191 | * no valid regs pointer | |
1192 | */ | |
9e8f559b FW |
1193 | if (regs) |
1194 | tick_sched_handle(ts, regs); | |
7c259045 FW |
1195 | else |
1196 | ts->next_tick = 0; | |
79bf2bb3 | 1197 | |
2a16fc93 VK |
1198 | /* No need to reprogram if we are in idle or full dynticks mode */ |
1199 | if (unlikely(ts->tick_stopped)) | |
1200 | return HRTIMER_NORESTART; | |
1201 | ||
79bf2bb3 TG |
1202 | hrtimer_forward(timer, now, tick_period); |
1203 | ||
1204 | return HRTIMER_RESTART; | |
1205 | } | |
1206 | ||
5307c955 MG |
1207 | static int sched_skew_tick; |
1208 | ||
62cf20b3 TG |
1209 | static int __init skew_tick(char *str) |
1210 | { | |
1211 | get_option(&str, &sched_skew_tick); | |
1212 | ||
1213 | return 0; | |
1214 | } | |
1215 | early_param("skew_tick", skew_tick); | |
1216 | ||
79bf2bb3 TG |
1217 | /** |
1218 | * tick_setup_sched_timer - setup the tick emulation timer | |
1219 | */ | |
1220 | void tick_setup_sched_timer(void) | |
1221 | { | |
22127e93 | 1222 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
79bf2bb3 TG |
1223 | ktime_t now = ktime_get(); |
1224 | ||
1225 | /* | |
1226 | * Emulate tick processing via per-CPU hrtimers: | |
1227 | */ | |
1228 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); | |
1229 | ts->sched_timer.function = tick_sched_timer; | |
79bf2bb3 | 1230 | |
0de7611a | 1231 | /* Get the next period (per-CPU) */ |
cc584b21 | 1232 | hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); |
79bf2bb3 | 1233 | |
9c3f9e28 | 1234 | /* Offset the tick to avert jiffies_lock contention. */ |
5307c955 MG |
1235 | if (sched_skew_tick) { |
1236 | u64 offset = ktime_to_ns(tick_period) >> 1; | |
1237 | do_div(offset, num_possible_cpus()); | |
1238 | offset *= smp_processor_id(); | |
1239 | hrtimer_add_expires_ns(&ts->sched_timer, offset); | |
1240 | } | |
1241 | ||
afc08b15 TG |
1242 | hrtimer_forward(&ts->sched_timer, now, tick_period); |
1243 | hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); | |
bc7a34b8 | 1244 | tick_nohz_activate(ts, NOHZ_MODE_HIGHRES); |
79bf2bb3 | 1245 | } |
3c4fbe5e | 1246 | #endif /* HIGH_RES_TIMERS */ |
79bf2bb3 | 1247 | |
3451d024 | 1248 | #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS |
79bf2bb3 TG |
1249 | void tick_cancel_sched_timer(int cpu) |
1250 | { | |
1251 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
1252 | ||
3c4fbe5e | 1253 | # ifdef CONFIG_HIGH_RES_TIMERS |
79bf2bb3 TG |
1254 | if (ts->sched_timer.base) |
1255 | hrtimer_cancel(&ts->sched_timer); | |
3c4fbe5e | 1256 | # endif |
a7901766 | 1257 | |
4b0c0f29 | 1258 | memset(ts, 0, sizeof(*ts)); |
79bf2bb3 | 1259 | } |
3c4fbe5e | 1260 | #endif |
79bf2bb3 TG |
1261 | |
1262 | /** | |
1263 | * Async notification about clocksource changes | |
1264 | */ | |
1265 | void tick_clock_notify(void) | |
1266 | { | |
1267 | int cpu; | |
1268 | ||
1269 | for_each_possible_cpu(cpu) | |
1270 | set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); | |
1271 | } | |
1272 | ||
1273 | /* | |
1274 | * Async notification about clock event changes | |
1275 | */ | |
1276 | void tick_oneshot_notify(void) | |
1277 | { | |
22127e93 | 1278 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
79bf2bb3 TG |
1279 | |
1280 | set_bit(0, &ts->check_clocks); | |
1281 | } | |
1282 | ||
1283 | /** | |
1284 | * Check, if a change happened, which makes oneshot possible. | |
1285 | * | |
1286 | * Called cyclic from the hrtimer softirq (driven by the timer | |
1287 | * softirq) allow_nohz signals, that we can switch into low-res nohz | |
1288 | * mode, because high resolution timers are disabled (either compile | |
6b442bc8 | 1289 | * or runtime). Called with interrupts disabled. |
79bf2bb3 TG |
1290 | */ |
1291 | int tick_check_oneshot_change(int allow_nohz) | |
1292 | { | |
22127e93 | 1293 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
79bf2bb3 TG |
1294 | |
1295 | if (!test_and_clear_bit(0, &ts->check_clocks)) | |
1296 | return 0; | |
1297 | ||
1298 | if (ts->nohz_mode != NOHZ_MODE_INACTIVE) | |
1299 | return 0; | |
1300 | ||
cf4fc6cb | 1301 | if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) |
79bf2bb3 TG |
1302 | return 0; |
1303 | ||
1304 | if (!allow_nohz) | |
1305 | return 1; | |
1306 | ||
1307 | tick_nohz_switch_to_nohz(); | |
1308 | return 0; | |
1309 | } |