]>
Commit | Line | Data |
---|---|---|
35728b82 | 1 | // SPDX-License-Identifier: GPL-2.0 |
79bf2bb3 | 2 | /* |
79bf2bb3 TG |
3 | * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de> |
4 | * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar | |
5 | * Copyright(C) 2006-2007 Timesys Corp., Thomas Gleixner | |
6 | * | |
7 | * No idle tick implementation for low and high resolution timers | |
8 | * | |
9 | * Started by: Thomas Gleixner and Ingo Molnar | |
79bf2bb3 TG |
10 | */ |
11 | #include <linux/cpu.h> | |
12 | #include <linux/err.h> | |
13 | #include <linux/hrtimer.h> | |
14 | #include <linux/interrupt.h> | |
15 | #include <linux/kernel_stat.h> | |
16 | #include <linux/percpu.h> | |
38b8d208 | 17 | #include <linux/nmi.h> |
79bf2bb3 | 18 | #include <linux/profile.h> |
3f07c014 | 19 | #include <linux/sched/signal.h> |
e6017571 | 20 | #include <linux/sched/clock.h> |
03441a34 | 21 | #include <linux/sched/stat.h> |
370c9135 | 22 | #include <linux/sched/nohz.h> |
8083e4ad | 23 | #include <linux/module.h> |
00b42959 | 24 | #include <linux/irq_work.h> |
9014c45d | 25 | #include <linux/posix-timers.h> |
2e709338 | 26 | #include <linux/context_tracking.h> |
62cb1188 | 27 | #include <linux/mm.h> |
79bf2bb3 | 28 | |
9e203bcc DM |
29 | #include <asm/irq_regs.h> |
30 | ||
79bf2bb3 TG |
31 | #include "tick-internal.h" |
32 | ||
cb41a290 FW |
33 | #include <trace/events/timer.h> |
34 | ||
79bf2bb3 | 35 | /* |
0de7611a | 36 | * Per-CPU nohz control structure |
79bf2bb3 | 37 | */ |
c1797baf | 38 | static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched); |
79bf2bb3 | 39 | |
289f480a IM |
40 | struct tick_sched *tick_get_tick_sched(int cpu) |
41 | { | |
42 | return &per_cpu(tick_cpu_sched, cpu); | |
43 | } | |
44 | ||
7809998a AB |
45 | #if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS) |
46 | /* | |
47 | * The time, when the last jiffy update happened. Protected by jiffies_lock. | |
48 | */ | |
49 | static ktime_t last_jiffies_update; | |
50 | ||
79bf2bb3 TG |
51 | /* |
52 | * Must be called with interrupts disabled ! | |
53 | */ | |
54 | static void tick_do_update_jiffies64(ktime_t now) | |
55 | { | |
56 | unsigned long ticks = 0; | |
57 | ktime_t delta; | |
58 | ||
7a14ce1d | 59 | /* |
d6ad4187 | 60 | * Do a quick check without holding jiffies_lock: |
de95a991 | 61 | * The READ_ONCE() pairs with two updates done later in this function. |
7a14ce1d | 62 | */ |
de95a991 | 63 | delta = ktime_sub(now, READ_ONCE(last_jiffies_update)); |
2456e855 | 64 | if (delta < tick_period) |
7a14ce1d IM |
65 | return; |
66 | ||
6168f8ed | 67 | /* Reevaluate with jiffies_lock held */ |
e5d4d175 TG |
68 | raw_spin_lock(&jiffies_lock); |
69 | write_seqcount_begin(&jiffies_seq); | |
79bf2bb3 TG |
70 | |
71 | delta = ktime_sub(now, last_jiffies_update); | |
2456e855 | 72 | if (delta >= tick_period) { |
79bf2bb3 TG |
73 | |
74 | delta = ktime_sub(delta, tick_period); | |
de95a991 ED |
75 | /* Pairs with the lockless read in this function. */ |
76 | WRITE_ONCE(last_jiffies_update, | |
77 | ktime_add(last_jiffies_update, tick_period)); | |
79bf2bb3 TG |
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 | ||
de95a991 ED |
85 | /* Pairs with the lockless read in this function. */ |
86 | WRITE_ONCE(last_jiffies_update, | |
87 | ktime_add_ns(last_jiffies_update, | |
88 | incr * ticks)); | |
79bf2bb3 TG |
89 | } |
90 | do_timer(++ticks); | |
49d670fb TG |
91 | |
92 | /* Keep the tick_next_period variable up to date */ | |
93 | tick_next_period = ktime_add(last_jiffies_update, tick_period); | |
03e6bdc5 | 94 | } else { |
e5d4d175 TG |
95 | write_seqcount_end(&jiffies_seq); |
96 | raw_spin_unlock(&jiffies_lock); | |
03e6bdc5 | 97 | return; |
79bf2bb3 | 98 | } |
e5d4d175 TG |
99 | write_seqcount_end(&jiffies_seq); |
100 | raw_spin_unlock(&jiffies_lock); | |
47a1b796 | 101 | update_wall_time(); |
79bf2bb3 TG |
102 | } |
103 | ||
104 | /* | |
105 | * Initialize and return retrieve the jiffies update. | |
106 | */ | |
107 | static ktime_t tick_init_jiffy_update(void) | |
108 | { | |
109 | ktime_t period; | |
110 | ||
e5d4d175 TG |
111 | raw_spin_lock(&jiffies_lock); |
112 | write_seqcount_begin(&jiffies_seq); | |
79bf2bb3 | 113 | /* Did we start the jiffies update yet ? */ |
2456e855 | 114 | if (last_jiffies_update == 0) |
79bf2bb3 TG |
115 | last_jiffies_update = tick_next_period; |
116 | period = last_jiffies_update; | |
e5d4d175 TG |
117 | write_seqcount_end(&jiffies_seq); |
118 | raw_spin_unlock(&jiffies_lock); | |
79bf2bb3 TG |
119 | return period; |
120 | } | |
121 | ||
ff7de620 | 122 | static void tick_sched_do_timer(struct tick_sched *ts, ktime_t now) |
5bb96226 FW |
123 | { |
124 | int cpu = smp_processor_id(); | |
125 | ||
3451d024 | 126 | #ifdef CONFIG_NO_HZ_COMMON |
5bb96226 FW |
127 | /* |
128 | * Check if the do_timer duty was dropped. We don't care about | |
0de7611a IM |
129 | * concurrency: This happens only when the CPU in charge went |
130 | * into a long sleep. If two CPUs happen to assign themselves to | |
5bb96226 | 131 | * this duty, then the jiffies update is still serialized by |
9c3f9e28 | 132 | * jiffies_lock. |
08ae95f4 NP |
133 | * |
134 | * If nohz_full is enabled, this should not happen because the | |
135 | * tick_do_timer_cpu never relinquishes. | |
5bb96226 | 136 | */ |
08ae95f4 NP |
137 | if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) { |
138 | #ifdef CONFIG_NO_HZ_FULL | |
139 | WARN_ON(tick_nohz_full_running); | |
140 | #endif | |
5bb96226 | 141 | tick_do_timer_cpu = cpu; |
08ae95f4 | 142 | } |
5bb96226 FW |
143 | #endif |
144 | ||
145 | /* Check, if the jiffies need an update */ | |
146 | if (tick_do_timer_cpu == cpu) | |
147 | tick_do_update_jiffies64(now); | |
ff7de620 RW |
148 | |
149 | if (ts->inidle) | |
150 | ts->got_idle_tick = 1; | |
5bb96226 FW |
151 | } |
152 | ||
9e8f559b FW |
153 | static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs) |
154 | { | |
3451d024 | 155 | #ifdef CONFIG_NO_HZ_COMMON |
9e8f559b FW |
156 | /* |
157 | * When we are idle and the tick is stopped, we have to touch | |
158 | * the watchdog as we might not schedule for a really long | |
159 | * time. This happens on complete idle SMP systems while | |
160 | * waiting on the login prompt. We also increment the "start of | |
161 | * idle" jiffy stamp so the idle accounting adjustment we do | |
162 | * when we go busy again does not account too much ticks. | |
163 | */ | |
164 | if (ts->tick_stopped) { | |
03e0d461 | 165 | touch_softlockup_watchdog_sched(); |
9e8f559b FW |
166 | if (is_idle_task(current)) |
167 | ts->idle_jiffies++; | |
411fe24e FW |
168 | /* |
169 | * In case the current tick fired too early past its expected | |
170 | * expiration, make sure we don't bypass the next clock reprogramming | |
171 | * to the same deadline. | |
172 | */ | |
173 | ts->next_tick = 0; | |
9e8f559b | 174 | } |
94a57140 | 175 | #endif |
9e8f559b FW |
176 | update_process_times(user_mode(regs)); |
177 | profile_tick(CPU_PROFILING); | |
178 | } | |
7809998a | 179 | #endif |
9e8f559b | 180 | |
c5bfece2 | 181 | #ifdef CONFIG_NO_HZ_FULL |
460775df | 182 | cpumask_var_t tick_nohz_full_mask; |
73867dcd | 183 | bool tick_nohz_full_running; |
ae9e557b | 184 | EXPORT_SYMBOL_GPL(tick_nohz_full_running); |
f009a7a7 | 185 | static atomic_t tick_dep_mask; |
a831881b | 186 | |
f009a7a7 | 187 | static bool check_tick_dependency(atomic_t *dep) |
d027d45d | 188 | { |
f009a7a7 FW |
189 | int val = atomic_read(dep); |
190 | ||
191 | if (val & TICK_DEP_MASK_POSIX_TIMER) { | |
e6e6cc22 | 192 | trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER); |
f009a7a7 | 193 | return true; |
d027d45d FW |
194 | } |
195 | ||
f009a7a7 | 196 | if (val & TICK_DEP_MASK_PERF_EVENTS) { |
e6e6cc22 | 197 | trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS); |
f009a7a7 | 198 | return true; |
d027d45d FW |
199 | } |
200 | ||
f009a7a7 | 201 | if (val & TICK_DEP_MASK_SCHED) { |
e6e6cc22 | 202 | trace_tick_stop(0, TICK_DEP_MASK_SCHED); |
f009a7a7 | 203 | return true; |
d027d45d FW |
204 | } |
205 | ||
f009a7a7 | 206 | if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) { |
e6e6cc22 | 207 | trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE); |
f009a7a7 FW |
208 | return true; |
209 | } | |
210 | ||
01b4c399 FW |
211 | if (val & TICK_DEP_MASK_RCU) { |
212 | trace_tick_stop(0, TICK_DEP_MASK_RCU); | |
213 | return true; | |
214 | } | |
215 | ||
f009a7a7 | 216 | return false; |
d027d45d FW |
217 | } |
218 | ||
57ccdf44 | 219 | static bool can_stop_full_tick(int cpu, struct tick_sched *ts) |
9014c45d | 220 | { |
ebf3adba | 221 | lockdep_assert_irqs_disabled(); |
9014c45d | 222 | |
57ccdf44 WL |
223 | if (unlikely(!cpu_online(cpu))) |
224 | return false; | |
225 | ||
f009a7a7 | 226 | if (check_tick_dependency(&tick_dep_mask)) |
d027d45d | 227 | return false; |
d027d45d | 228 | |
f009a7a7 | 229 | if (check_tick_dependency(&ts->tick_dep_mask)) |
d027d45d | 230 | return false; |
d027d45d | 231 | |
f009a7a7 | 232 | if (check_tick_dependency(¤t->tick_dep_mask)) |
d027d45d | 233 | return false; |
d027d45d | 234 | |
f009a7a7 | 235 | if (check_tick_dependency(¤t->signal->tick_dep_mask)) |
d027d45d | 236 | return false; |
d027d45d | 237 | |
9014c45d FW |
238 | return true; |
239 | } | |
240 | ||
d027d45d | 241 | static void nohz_full_kick_func(struct irq_work *work) |
76c24fb0 | 242 | { |
73738a95 | 243 | /* Empty, the tick restart happens on tick_nohz_irq_exit() */ |
76c24fb0 FW |
244 | } |
245 | ||
246 | static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = { | |
d027d45d | 247 | .func = nohz_full_kick_func, |
49915ac3 | 248 | .flags = ATOMIC_INIT(IRQ_WORK_HARD_IRQ), |
76c24fb0 FW |
249 | }; |
250 | ||
40bea039 FW |
251 | /* |
252 | * Kick this CPU if it's full dynticks in order to force it to | |
253 | * re-evaluate its dependency on the tick and restart it if necessary. | |
254 | * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(), | |
255 | * is NMI safe. | |
256 | */ | |
555e0c1e | 257 | static void tick_nohz_full_kick(void) |
40bea039 FW |
258 | { |
259 | if (!tick_nohz_full_cpu(smp_processor_id())) | |
260 | return; | |
261 | ||
56e4dea8 | 262 | irq_work_queue(this_cpu_ptr(&nohz_full_kick_work)); |
40bea039 FW |
263 | } |
264 | ||
76c24fb0 | 265 | /* |
3d36aebc | 266 | * Kick the CPU if it's full dynticks in order to force it to |
76c24fb0 FW |
267 | * re-evaluate its dependency on the tick and restart it if necessary. |
268 | */ | |
3d36aebc | 269 | void tick_nohz_full_kick_cpu(int cpu) |
76c24fb0 | 270 | { |
3d36aebc FW |
271 | if (!tick_nohz_full_cpu(cpu)) |
272 | return; | |
273 | ||
274 | irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu); | |
76c24fb0 FW |
275 | } |
276 | ||
76c24fb0 FW |
277 | /* |
278 | * Kick all full dynticks CPUs in order to force these to re-evaluate | |
279 | * their dependency on the tick and restart it if necessary. | |
280 | */ | |
b7878300 | 281 | static void tick_nohz_full_kick_all(void) |
76c24fb0 | 282 | { |
8537bb95 FW |
283 | int cpu; |
284 | ||
73867dcd | 285 | if (!tick_nohz_full_running) |
76c24fb0 FW |
286 | return; |
287 | ||
288 | preempt_disable(); | |
8537bb95 FW |
289 | for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask) |
290 | tick_nohz_full_kick_cpu(cpu); | |
76c24fb0 FW |
291 | preempt_enable(); |
292 | } | |
293 | ||
f009a7a7 | 294 | static void tick_nohz_dep_set_all(atomic_t *dep, |
d027d45d FW |
295 | enum tick_dep_bits bit) |
296 | { | |
f009a7a7 | 297 | int prev; |
d027d45d | 298 | |
a1cc5bcf | 299 | prev = atomic_fetch_or(BIT(bit), dep); |
d027d45d FW |
300 | if (!prev) |
301 | tick_nohz_full_kick_all(); | |
302 | } | |
303 | ||
304 | /* | |
305 | * Set a global tick dependency. Used by perf events that rely on freq and | |
306 | * by unstable clock. | |
307 | */ | |
308 | void tick_nohz_dep_set(enum tick_dep_bits bit) | |
309 | { | |
310 | tick_nohz_dep_set_all(&tick_dep_mask, bit); | |
311 | } | |
312 | ||
313 | void tick_nohz_dep_clear(enum tick_dep_bits bit) | |
314 | { | |
f009a7a7 | 315 | atomic_andnot(BIT(bit), &tick_dep_mask); |
d027d45d FW |
316 | } |
317 | ||
318 | /* | |
319 | * Set per-CPU tick dependency. Used by scheduler and perf events in order to | |
320 | * manage events throttling. | |
321 | */ | |
322 | void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit) | |
323 | { | |
f009a7a7 | 324 | int prev; |
d027d45d FW |
325 | struct tick_sched *ts; |
326 | ||
327 | ts = per_cpu_ptr(&tick_cpu_sched, cpu); | |
328 | ||
a1cc5bcf | 329 | prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask); |
d027d45d FW |
330 | if (!prev) { |
331 | preempt_disable(); | |
332 | /* Perf needs local kick that is NMI safe */ | |
333 | if (cpu == smp_processor_id()) { | |
334 | tick_nohz_full_kick(); | |
335 | } else { | |
336 | /* Remote irq work not NMI-safe */ | |
337 | if (!WARN_ON_ONCE(in_nmi())) | |
338 | tick_nohz_full_kick_cpu(cpu); | |
339 | } | |
340 | preempt_enable(); | |
341 | } | |
342 | } | |
01b4c399 | 343 | EXPORT_SYMBOL_GPL(tick_nohz_dep_set_cpu); |
d027d45d FW |
344 | |
345 | void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit) | |
346 | { | |
347 | struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu); | |
348 | ||
f009a7a7 | 349 | atomic_andnot(BIT(bit), &ts->tick_dep_mask); |
d027d45d | 350 | } |
01b4c399 | 351 | EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_cpu); |
d027d45d FW |
352 | |
353 | /* | |
3c8920e2 FW |
354 | * Set a per-task tick dependency. RCU need this. Also posix CPU timers |
355 | * in order to elapse per task timers. | |
d027d45d FW |
356 | */ |
357 | void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit) | |
358 | { | |
3c8920e2 FW |
359 | if (!atomic_fetch_or(BIT(bit), &tsk->tick_dep_mask)) { |
360 | if (tsk == current) { | |
361 | preempt_disable(); | |
362 | tick_nohz_full_kick(); | |
363 | preempt_enable(); | |
364 | } else { | |
365 | /* | |
366 | * Some future tick_nohz_full_kick_task() | |
367 | * should optimize this. | |
368 | */ | |
369 | tick_nohz_full_kick_all(); | |
370 | } | |
371 | } | |
d027d45d | 372 | } |
ae9e557b | 373 | EXPORT_SYMBOL_GPL(tick_nohz_dep_set_task); |
d027d45d FW |
374 | |
375 | void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit) | |
376 | { | |
f009a7a7 | 377 | atomic_andnot(BIT(bit), &tsk->tick_dep_mask); |
d027d45d | 378 | } |
ae9e557b | 379 | EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_task); |
d027d45d FW |
380 | |
381 | /* | |
382 | * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse | |
383 | * per process timers. | |
384 | */ | |
385 | void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit) | |
386 | { | |
387 | tick_nohz_dep_set_all(&sig->tick_dep_mask, bit); | |
388 | } | |
389 | ||
390 | void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit) | |
391 | { | |
f009a7a7 | 392 | atomic_andnot(BIT(bit), &sig->tick_dep_mask); |
d027d45d FW |
393 | } |
394 | ||
99e5ada9 FW |
395 | /* |
396 | * Re-evaluate the need for the tick as we switch the current task. | |
397 | * It might need the tick due to per task/process properties: | |
0de7611a | 398 | * perf events, posix CPU timers, ... |
99e5ada9 | 399 | */ |
de734f89 | 400 | void __tick_nohz_task_switch(void) |
99e5ada9 FW |
401 | { |
402 | unsigned long flags; | |
d027d45d | 403 | struct tick_sched *ts; |
99e5ada9 | 404 | |
99e5ada9 FW |
405 | local_irq_save(flags); |
406 | ||
6296ace4 LZ |
407 | if (!tick_nohz_full_cpu(smp_processor_id())) |
408 | goto out; | |
409 | ||
d027d45d | 410 | ts = this_cpu_ptr(&tick_cpu_sched); |
99e5ada9 | 411 | |
d027d45d | 412 | if (ts->tick_stopped) { |
f009a7a7 FW |
413 | if (atomic_read(¤t->tick_dep_mask) || |
414 | atomic_read(¤t->signal->tick_dep_mask)) | |
d027d45d FW |
415 | tick_nohz_full_kick(); |
416 | } | |
6296ace4 | 417 | out: |
99e5ada9 FW |
418 | local_irq_restore(flags); |
419 | } | |
420 | ||
6f1982fe FW |
421 | /* Get the boot-time nohz CPU list from the kernel parameters. */ |
422 | void __init tick_nohz_full_setup(cpumask_var_t cpumask) | |
a831881b | 423 | { |
73867dcd | 424 | alloc_bootmem_cpumask_var(&tick_nohz_full_mask); |
6f1982fe | 425 | cpumask_copy(tick_nohz_full_mask, cpumask); |
73867dcd | 426 | tick_nohz_full_running = true; |
a831881b | 427 | } |
ae9e557b | 428 | EXPORT_SYMBOL_GPL(tick_nohz_full_setup); |
a831881b | 429 | |
31eff243 | 430 | static int tick_nohz_cpu_down(unsigned int cpu) |
a382bf93 | 431 | { |
31eff243 | 432 | /* |
08ae95f4 NP |
433 | * The tick_do_timer_cpu CPU handles housekeeping duty (unbound |
434 | * timers, workqueues, timekeeping, ...) on behalf of full dynticks | |
31eff243 SAS |
435 | * CPUs. It must remain online when nohz full is enabled. |
436 | */ | |
437 | if (tick_nohz_full_running && tick_do_timer_cpu == cpu) | |
438 | return -EBUSY; | |
439 | return 0; | |
a382bf93 FW |
440 | } |
441 | ||
d1e43fa5 | 442 | void __init tick_nohz_init(void) |
a831881b | 443 | { |
31eff243 | 444 | int cpu, ret; |
d1e43fa5 | 445 | |
a7c8655b PM |
446 | if (!tick_nohz_full_running) |
447 | return; | |
d1e43fa5 | 448 | |
9b01f5bf FW |
449 | /* |
450 | * Full dynticks uses irq work to drive the tick rescheduling on safe | |
451 | * locking contexts. But then we need irq work to raise its own | |
452 | * interrupts to avoid circular dependency on the tick | |
453 | */ | |
454 | if (!arch_irq_work_has_interrupt()) { | |
a395d6a7 | 455 | pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n"); |
9b01f5bf | 456 | cpumask_clear(tick_nohz_full_mask); |
9b01f5bf FW |
457 | tick_nohz_full_running = false; |
458 | return; | |
459 | } | |
460 | ||
08ae95f4 NP |
461 | if (IS_ENABLED(CONFIG_PM_SLEEP_SMP) && |
462 | !IS_ENABLED(CONFIG_PM_SLEEP_SMP_NONZERO_CPU)) { | |
463 | cpu = smp_processor_id(); | |
4327b15f | 464 | |
08ae95f4 NP |
465 | if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) { |
466 | pr_warn("NO_HZ: Clearing %d from nohz_full range " | |
467 | "for timekeeping\n", cpu); | |
468 | cpumask_clear_cpu(cpu, tick_nohz_full_mask); | |
469 | } | |
4327b15f FW |
470 | } |
471 | ||
73867dcd | 472 | for_each_cpu(cpu, tick_nohz_full_mask) |
2e709338 FW |
473 | context_tracking_cpu_set(cpu); |
474 | ||
31eff243 SAS |
475 | ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, |
476 | "kernel/nohz:predown", NULL, | |
477 | tick_nohz_cpu_down); | |
478 | WARN_ON(ret < 0); | |
ffda22c1 TH |
479 | pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n", |
480 | cpumask_pr_args(tick_nohz_full_mask)); | |
a831881b | 481 | } |
a831881b FW |
482 | #endif |
483 | ||
79bf2bb3 TG |
484 | /* |
485 | * NOHZ - aka dynamic tick functionality | |
486 | */ | |
3451d024 | 487 | #ifdef CONFIG_NO_HZ_COMMON |
79bf2bb3 TG |
488 | /* |
489 | * NO HZ enabled ? | |
490 | */ | |
4cc7ecb7 | 491 | bool tick_nohz_enabled __read_mostly = true; |
bc7a34b8 | 492 | unsigned long tick_nohz_active __read_mostly; |
79bf2bb3 TG |
493 | /* |
494 | * Enable / Disable tickless mode | |
495 | */ | |
496 | static int __init setup_tick_nohz(char *str) | |
497 | { | |
4cc7ecb7 | 498 | return (kstrtobool(str, &tick_nohz_enabled) == 0); |
79bf2bb3 TG |
499 | } |
500 | ||
501 | __setup("nohz=", setup_tick_nohz); | |
502 | ||
a3642983 | 503 | bool tick_nohz_tick_stopped(void) |
c1797baf | 504 | { |
2bc629a6 FW |
505 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
506 | ||
507 | return ts->tick_stopped; | |
c1797baf TG |
508 | } |
509 | ||
22ab8bc0 FW |
510 | bool tick_nohz_tick_stopped_cpu(int cpu) |
511 | { | |
512 | struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu); | |
513 | ||
514 | return ts->tick_stopped; | |
515 | } | |
516 | ||
79bf2bb3 TG |
517 | /** |
518 | * tick_nohz_update_jiffies - update jiffies when idle was interrupted | |
519 | * | |
520 | * Called from interrupt entry when the CPU was idle | |
521 | * | |
522 | * In case the sched_tick was stopped on this CPU, we have to check if jiffies | |
523 | * must be updated. Otherwise an interrupt handler could use a stale jiffy | |
0de7611a IM |
524 | * value. We do this unconditionally on any CPU, as we don't know whether the |
525 | * CPU, which has the update task assigned is in a long sleep. | |
79bf2bb3 | 526 | */ |
eed3b9cf | 527 | static void tick_nohz_update_jiffies(ktime_t now) |
79bf2bb3 | 528 | { |
79bf2bb3 | 529 | unsigned long flags; |
79bf2bb3 | 530 | |
e8fcaa5c | 531 | __this_cpu_write(tick_cpu_sched.idle_waketime, now); |
79bf2bb3 TG |
532 | |
533 | local_irq_save(flags); | |
534 | tick_do_update_jiffies64(now); | |
535 | local_irq_restore(flags); | |
02ff3755 | 536 | |
03e0d461 | 537 | touch_softlockup_watchdog_sched(); |
79bf2bb3 TG |
538 | } |
539 | ||
595aac48 | 540 | /* |
0de7611a | 541 | * Updates the per-CPU time idle statistics counters |
595aac48 | 542 | */ |
8d63bf94 | 543 | static void |
8c215bd3 | 544 | update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time) |
6378ddb5 | 545 | { |
eed3b9cf | 546 | ktime_t delta; |
6378ddb5 | 547 | |
595aac48 AV |
548 | if (ts->idle_active) { |
549 | delta = ktime_sub(now, ts->idle_entrytime); | |
8c215bd3 | 550 | if (nr_iowait_cpu(cpu) > 0) |
0224cf4c | 551 | ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); |
6beea0cd MH |
552 | else |
553 | ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); | |
8c7b09f4 | 554 | ts->idle_entrytime = now; |
595aac48 | 555 | } |
8d63bf94 | 556 | |
e0e37c20 | 557 | if (last_update_time) |
8d63bf94 AV |
558 | *last_update_time = ktime_to_us(now); |
559 | ||
595aac48 AV |
560 | } |
561 | ||
e8fcaa5c | 562 | static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now) |
595aac48 | 563 | { |
e8fcaa5c | 564 | update_ts_time_stats(smp_processor_id(), ts, now, NULL); |
eed3b9cf | 565 | ts->idle_active = 0; |
56c7426b | 566 | |
ac1e843f | 567 | sched_clock_idle_wakeup_event(); |
6378ddb5 VP |
568 | } |
569 | ||
0e776768 | 570 | static void tick_nohz_start_idle(struct tick_sched *ts) |
6378ddb5 | 571 | { |
0e776768 | 572 | ts->idle_entrytime = ktime_get(); |
6378ddb5 | 573 | ts->idle_active = 1; |
56c7426b | 574 | sched_clock_idle_sleep_event(); |
6378ddb5 VP |
575 | } |
576 | ||
b1f724c3 | 577 | /** |
0de7611a | 578 | * get_cpu_idle_time_us - get the total idle time of a CPU |
b1f724c3 | 579 | * @cpu: CPU number to query |
09a1d34f MH |
580 | * @last_update_time: variable to store update time in. Do not update |
581 | * counters if NULL. | |
b1f724c3 | 582 | * |
6168f8ed | 583 | * Return the cumulative idle time (since boot) for a given |
6beea0cd | 584 | * CPU, in microseconds. |
b1f724c3 AV |
585 | * |
586 | * This time is measured via accounting rather than sampling, | |
587 | * and is as accurate as ktime_get() is. | |
588 | * | |
589 | * This function returns -1 if NOHZ is not enabled. | |
590 | */ | |
6378ddb5 VP |
591 | u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) |
592 | { | |
593 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
09a1d34f | 594 | ktime_t now, idle; |
6378ddb5 | 595 | |
d689fe22 | 596 | if (!tick_nohz_active) |
8083e4ad | 597 | return -1; |
598 | ||
09a1d34f MH |
599 | now = ktime_get(); |
600 | if (last_update_time) { | |
601 | update_ts_time_stats(cpu, ts, now, last_update_time); | |
602 | idle = ts->idle_sleeptime; | |
603 | } else { | |
604 | if (ts->idle_active && !nr_iowait_cpu(cpu)) { | |
605 | ktime_t delta = ktime_sub(now, ts->idle_entrytime); | |
606 | ||
607 | idle = ktime_add(ts->idle_sleeptime, delta); | |
608 | } else { | |
609 | idle = ts->idle_sleeptime; | |
610 | } | |
611 | } | |
612 | ||
613 | return ktime_to_us(idle); | |
8083e4ad | 614 | |
6378ddb5 | 615 | } |
8083e4ad | 616 | EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); |
6378ddb5 | 617 | |
6beea0cd | 618 | /** |
0de7611a | 619 | * get_cpu_iowait_time_us - get the total iowait time of a CPU |
0224cf4c | 620 | * @cpu: CPU number to query |
09a1d34f MH |
621 | * @last_update_time: variable to store update time in. Do not update |
622 | * counters if NULL. | |
0224cf4c | 623 | * |
6168f8ed | 624 | * Return the cumulative iowait time (since boot) for a given |
0224cf4c AV |
625 | * CPU, in microseconds. |
626 | * | |
627 | * This time is measured via accounting rather than sampling, | |
628 | * and is as accurate as ktime_get() is. | |
629 | * | |
630 | * This function returns -1 if NOHZ is not enabled. | |
631 | */ | |
632 | u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) | |
633 | { | |
634 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
09a1d34f | 635 | ktime_t now, iowait; |
0224cf4c | 636 | |
d689fe22 | 637 | if (!tick_nohz_active) |
0224cf4c AV |
638 | return -1; |
639 | ||
09a1d34f MH |
640 | now = ktime_get(); |
641 | if (last_update_time) { | |
642 | update_ts_time_stats(cpu, ts, now, last_update_time); | |
643 | iowait = ts->iowait_sleeptime; | |
644 | } else { | |
645 | if (ts->idle_active && nr_iowait_cpu(cpu) > 0) { | |
646 | ktime_t delta = ktime_sub(now, ts->idle_entrytime); | |
0224cf4c | 647 | |
09a1d34f MH |
648 | iowait = ktime_add(ts->iowait_sleeptime, delta); |
649 | } else { | |
650 | iowait = ts->iowait_sleeptime; | |
651 | } | |
652 | } | |
0224cf4c | 653 | |
09a1d34f | 654 | return ktime_to_us(iowait); |
0224cf4c AV |
655 | } |
656 | EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); | |
657 | ||
0ff53d09 TG |
658 | static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) |
659 | { | |
660 | hrtimer_cancel(&ts->sched_timer); | |
661 | hrtimer_set_expires(&ts->sched_timer, ts->last_tick); | |
662 | ||
663 | /* Forward the time to expire in the future */ | |
664 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
665 | ||
902a9f9c SAS |
666 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { |
667 | hrtimer_start_expires(&ts->sched_timer, | |
668 | HRTIMER_MODE_ABS_PINNED_HARD); | |
669 | } else { | |
0ff53d09 | 670 | tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); |
902a9f9c | 671 | } |
411fe24e FW |
672 | |
673 | /* | |
674 | * Reset to make sure next tick stop doesn't get fooled by past | |
675 | * cached clock deadline. | |
676 | */ | |
677 | ts->next_tick = 0; | |
0ff53d09 TG |
678 | } |
679 | ||
5d62c183 TG |
680 | static inline bool local_timer_softirq_pending(void) |
681 | { | |
80d20d35 | 682 | return local_softirq_pending() & BIT(TIMER_SOFTIRQ); |
5d62c183 TG |
683 | } |
684 | ||
23a8d888 | 685 | static ktime_t tick_nohz_next_event(struct tick_sched *ts, int cpu) |
79bf2bb3 | 686 | { |
c1ad348b | 687 | u64 basemono, next_tick, next_tmr, next_rcu, delta, expires; |
e1e41b6c RV |
688 | unsigned long basejiff; |
689 | unsigned int seq; | |
855a0fc3 | 690 | |
79bf2bb3 TG |
691 | /* Read jiffies and the time when jiffies were updated last */ |
692 | do { | |
e5d4d175 | 693 | seq = read_seqcount_begin(&jiffies_seq); |
2456e855 | 694 | basemono = last_jiffies_update; |
c1ad348b | 695 | basejiff = jiffies; |
e5d4d175 | 696 | } while (read_seqcount_retry(&jiffies_seq, seq)); |
c1ad348b | 697 | ts->last_jiffies = basejiff; |
23a8d888 | 698 | ts->timer_expires_base = basemono; |
79bf2bb3 | 699 | |
5d62c183 TG |
700 | /* |
701 | * Keep the periodic tick, when RCU, architecture or irq_work | |
702 | * requests it. | |
703 | * Aside of that check whether the local timer softirq is | |
704 | * pending. If so its a bad idea to call get_next_timer_interrupt() | |
705 | * because there is an already expired timer, so it will request | |
706 | * immeditate expiry, which rearms the hardware timer with a | |
707 | * minimal delta which brings us back to this place | |
708 | * immediately. Lather, rinse and repeat... | |
709 | */ | |
710 | if (rcu_needs_cpu(basemono, &next_rcu) || arch_needs_cpu() || | |
711 | irq_work_needs_cpu() || local_timer_softirq_pending()) { | |
c1ad348b | 712 | next_tick = basemono + TICK_NSEC; |
3c5d92a0 | 713 | } else { |
c1ad348b TG |
714 | /* |
715 | * Get the next pending timer. If high resolution | |
716 | * timers are enabled this only takes the timer wheel | |
717 | * timers into account. If high resolution timers are | |
718 | * disabled this also looks at the next expiring | |
719 | * hrtimer. | |
720 | */ | |
721 | next_tmr = get_next_timer_interrupt(basejiff, basemono); | |
722 | ts->next_timer = next_tmr; | |
723 | /* Take the next rcu event into account */ | |
724 | next_tick = next_rcu < next_tmr ? next_rcu : next_tmr; | |
3c5d92a0 | 725 | } |
47aa8b6c | 726 | |
c1ad348b TG |
727 | /* |
728 | * If the tick is due in the next period, keep it ticking or | |
82bbe34b | 729 | * force prod the timer. |
c1ad348b TG |
730 | */ |
731 | delta = next_tick - basemono; | |
732 | if (delta <= (u64)TICK_NSEC) { | |
a683f390 TG |
733 | /* |
734 | * Tell the timer code that the base is not idle, i.e. undo | |
735 | * the effect of get_next_timer_interrupt(): | |
736 | */ | |
737 | timer_clear_idle(); | |
82bbe34b PZ |
738 | /* |
739 | * We've not stopped the tick yet, and there's a timer in the | |
740 | * next period, so no point in stopping it either, bail. | |
741 | */ | |
f99973e1 | 742 | if (!ts->tick_stopped) { |
23a8d888 | 743 | ts->timer_expires = 0; |
157d29e1 TG |
744 | goto out; |
745 | } | |
746 | } | |
747 | ||
23a8d888 RW |
748 | /* |
749 | * If this CPU is the one which had the do_timer() duty last, we limit | |
750 | * the sleep time to the timekeeping max_deferment value. | |
751 | * Otherwise we can sleep as long as we want. | |
752 | */ | |
753 | delta = timekeeping_max_deferment(); | |
754 | if (cpu != tick_do_timer_cpu && | |
755 | (tick_do_timer_cpu != TICK_DO_TIMER_NONE || !ts->do_timer_last)) | |
756 | delta = KTIME_MAX; | |
757 | ||
758 | /* Calculate the next expiry time */ | |
759 | if (delta < (KTIME_MAX - basemono)) | |
760 | expires = basemono + delta; | |
761 | else | |
762 | expires = KTIME_MAX; | |
763 | ||
764 | ts->timer_expires = min_t(u64, expires, next_tick); | |
765 | ||
766 | out: | |
767 | return ts->timer_expires; | |
768 | } | |
769 | ||
770 | static void tick_nohz_stop_tick(struct tick_sched *ts, int cpu) | |
771 | { | |
772 | struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev); | |
773 | u64 basemono = ts->timer_expires_base; | |
774 | u64 expires = ts->timer_expires; | |
775 | ktime_t tick = expires; | |
776 | ||
777 | /* Make sure we won't be trying to stop it twice in a row. */ | |
778 | ts->timer_expires_base = 0; | |
779 | ||
79bf2bb3 | 780 | /* |
0de7611a IM |
781 | * If this CPU is the one which updates jiffies, then give up |
782 | * the assignment and let it be taken by the CPU which runs | |
783 | * the tick timer next, which might be this CPU as well. If we | |
157d29e1 TG |
784 | * don't drop this here the jiffies might be stale and |
785 | * do_timer() never invoked. Keep track of the fact that it | |
23a8d888 | 786 | * was the one which had the do_timer() duty last. |
79bf2bb3 | 787 | */ |
157d29e1 TG |
788 | if (cpu == tick_do_timer_cpu) { |
789 | tick_do_timer_cpu = TICK_DO_TIMER_NONE; | |
790 | ts->do_timer_last = 1; | |
791 | } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { | |
157d29e1 | 792 | ts->do_timer_last = 0; |
157d29e1 | 793 | } |
27185016 | 794 | |
157d29e1 | 795 | /* Skip reprogram of event if its not changed */ |
411fe24e FW |
796 | if (ts->tick_stopped && (expires == ts->next_tick)) { |
797 | /* Sanity check: make sure clockevent is actually programmed */ | |
d4af6d93 | 798 | if (tick == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer)) |
23a8d888 | 799 | return; |
411fe24e FW |
800 | |
801 | WARN_ON_ONCE(1); | |
802 | printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n", | |
803 | basemono, ts->next_tick, dev->next_event, | |
804 | hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer)); | |
ce6cf9a1 | 805 | } |
84bf1bcc | 806 | |
157d29e1 TG |
807 | /* |
808 | * nohz_stop_sched_tick can be called several times before | |
809 | * the nohz_restart_sched_tick is called. This happens when | |
810 | * interrupts arrive which do not cause a reschedule. In the | |
811 | * first call we save the current tick time, so we can restart | |
812 | * the scheduler tick in nohz_restart_sched_tick. | |
813 | */ | |
814 | if (!ts->tick_stopped) { | |
3c85d6db | 815 | calc_load_nohz_start(); |
62cb1188 | 816 | quiet_vmstat(); |
d3ed7824 | 817 | |
157d29e1 TG |
818 | ts->last_tick = hrtimer_get_expires(&ts->sched_timer); |
819 | ts->tick_stopped = 1; | |
e6e6cc22 | 820 | trace_tick_stop(1, TICK_DEP_MASK_NONE); |
157d29e1 | 821 | } |
eaad084b | 822 | |
411fe24e FW |
823 | ts->next_tick = tick; |
824 | ||
157d29e1 | 825 | /* |
c1ad348b TG |
826 | * If the expiration time == KTIME_MAX, then we simply stop |
827 | * the tick timer. | |
157d29e1 | 828 | */ |
c1ad348b | 829 | if (unlikely(expires == KTIME_MAX)) { |
157d29e1 TG |
830 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) |
831 | hrtimer_cancel(&ts->sched_timer); | |
23a8d888 | 832 | return; |
79bf2bb3 | 833 | } |
0ff53d09 | 834 | |
1f71addd | 835 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { |
902a9f9c SAS |
836 | hrtimer_start(&ts->sched_timer, tick, |
837 | HRTIMER_MODE_ABS_PINNED_HARD); | |
1f71addd TG |
838 | } else { |
839 | hrtimer_set_expires(&ts->sched_timer, tick); | |
c1ad348b | 840 | tick_program_event(tick, 1); |
1f71addd | 841 | } |
280f0677 FW |
842 | } |
843 | ||
23a8d888 RW |
844 | static void tick_nohz_retain_tick(struct tick_sched *ts) |
845 | { | |
846 | ts->timer_expires_base = 0; | |
847 | } | |
848 | ||
849 | #ifdef CONFIG_NO_HZ_FULL | |
850 | static void tick_nohz_stop_sched_tick(struct tick_sched *ts, int cpu) | |
851 | { | |
852 | if (tick_nohz_next_event(ts, cpu)) | |
853 | tick_nohz_stop_tick(ts, cpu); | |
854 | else | |
855 | tick_nohz_retain_tick(ts); | |
856 | } | |
857 | #endif /* CONFIG_NO_HZ_FULL */ | |
858 | ||
1f41906a | 859 | static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now) |
59d2c7ca FW |
860 | { |
861 | /* Update jiffies first */ | |
862 | tick_do_update_jiffies64(now); | |
59d2c7ca | 863 | |
a683f390 TG |
864 | /* |
865 | * Clear the timer idle flag, so we avoid IPIs on remote queueing and | |
866 | * the clock forward checks in the enqueue path: | |
867 | */ | |
868 | timer_clear_idle(); | |
869 | ||
3c85d6db | 870 | calc_load_nohz_stop(); |
03e0d461 | 871 | touch_softlockup_watchdog_sched(); |
59d2c7ca FW |
872 | /* |
873 | * Cancel the scheduled timer and restore the tick | |
874 | */ | |
875 | ts->tick_stopped = 0; | |
876 | ts->idle_exittime = now; | |
877 | ||
878 | tick_nohz_restart(ts, now); | |
879 | } | |
73738a95 FW |
880 | |
881 | static void tick_nohz_full_update_tick(struct tick_sched *ts) | |
5811d996 FW |
882 | { |
883 | #ifdef CONFIG_NO_HZ_FULL | |
e9a2eb40 | 884 | int cpu = smp_processor_id(); |
5811d996 | 885 | |
59449359 | 886 | if (!tick_nohz_full_cpu(cpu)) |
e9a2eb40 | 887 | return; |
5811d996 | 888 | |
e9a2eb40 AS |
889 | if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE) |
890 | return; | |
5811d996 | 891 | |
57ccdf44 | 892 | if (can_stop_full_tick(cpu, ts)) |
23a8d888 | 893 | tick_nohz_stop_sched_tick(ts, cpu); |
73738a95 | 894 | else if (ts->tick_stopped) |
1f41906a | 895 | tick_nohz_restart_sched_tick(ts, ktime_get()); |
5811d996 FW |
896 | #endif |
897 | } | |
898 | ||
5b39939a FW |
899 | static bool can_stop_idle_tick(int cpu, struct tick_sched *ts) |
900 | { | |
901 | /* | |
0de7611a | 902 | * If this CPU is offline and it is the one which updates |
5b39939a | 903 | * jiffies, then give up the assignment and let it be taken by |
0de7611a | 904 | * the CPU which runs the tick timer next. If we don't drop |
5b39939a FW |
905 | * this here the jiffies might be stale and do_timer() never |
906 | * invoked. | |
907 | */ | |
908 | if (unlikely(!cpu_online(cpu))) { | |
909 | if (cpu == tick_do_timer_cpu) | |
910 | tick_do_timer_cpu = TICK_DO_TIMER_NONE; | |
411fe24e FW |
911 | /* |
912 | * Make sure the CPU doesn't get fooled by obsolete tick | |
913 | * deadline if it comes back online later. | |
914 | */ | |
915 | ts->next_tick = 0; | |
f7ea0fd6 | 916 | return false; |
5b39939a FW |
917 | } |
918 | ||
23a8d888 | 919 | if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) |
5b39939a FW |
920 | return false; |
921 | ||
922 | if (need_resched()) | |
923 | return false; | |
924 | ||
d59e0ba1 | 925 | if (unlikely(local_softirq_pending())) { |
5b39939a FW |
926 | static int ratelimit; |
927 | ||
803b0eba PM |
928 | if (ratelimit < 10 && |
929 | (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) { | |
bca37119 | 930 | pr_warn("NOHZ tick-stop error: Non-RCU local softirq work is pending, handler #%02x!!!\n", |
cfea7d7e | 931 | (unsigned int) local_softirq_pending()); |
5b39939a FW |
932 | ratelimit++; |
933 | } | |
934 | return false; | |
935 | } | |
936 | ||
460775df | 937 | if (tick_nohz_full_enabled()) { |
a382bf93 FW |
938 | /* |
939 | * Keep the tick alive to guarantee timekeeping progression | |
940 | * if there are full dynticks CPUs around | |
941 | */ | |
942 | if (tick_do_timer_cpu == cpu) | |
943 | return false; | |
08ae95f4 NP |
944 | |
945 | /* Should not happen for nohz-full */ | |
946 | if (WARN_ON_ONCE(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) | |
a382bf93 FW |
947 | return false; |
948 | } | |
949 | ||
5b39939a FW |
950 | return true; |
951 | } | |
952 | ||
0e776768 | 953 | static void __tick_nohz_idle_stop_tick(struct tick_sched *ts) |
19f5f736 | 954 | { |
0e776768 | 955 | ktime_t expires; |
5b39939a | 956 | int cpu = smp_processor_id(); |
19f5f736 | 957 | |
554c8aa8 RW |
958 | /* |
959 | * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the | |
960 | * tick timer expiration time is known already. | |
961 | */ | |
962 | if (ts->timer_expires_base) | |
963 | expires = ts->timer_expires; | |
964 | else if (can_stop_idle_tick(cpu, ts)) | |
965 | expires = tick_nohz_next_event(ts, cpu); | |
966 | else | |
967 | return; | |
23a8d888 RW |
968 | |
969 | ts->idle_calls++; | |
08d07259 | 970 | |
23a8d888 | 971 | if (expires > 0LL) { |
5b39939a FW |
972 | int was_stopped = ts->tick_stopped; |
973 | ||
23a8d888 | 974 | tick_nohz_stop_tick(ts, cpu); |
84bf1bcc | 975 | |
23a8d888 RW |
976 | ts->idle_sleeps++; |
977 | ts->idle_expires = expires; | |
5b39939a | 978 | |
a0db971e | 979 | if (!was_stopped && ts->tick_stopped) { |
5b39939a | 980 | ts->idle_jiffies = ts->last_jiffies; |
a0db971e FW |
981 | nohz_balance_enter_idle(cpu); |
982 | } | |
23a8d888 RW |
983 | } else { |
984 | tick_nohz_retain_tick(ts); | |
5b39939a | 985 | } |
280f0677 FW |
986 | } |
987 | ||
988 | /** | |
0e776768 | 989 | * tick_nohz_idle_stop_tick - stop the idle tick from the idle task |
280f0677 FW |
990 | * |
991 | * When the next event is more than a tick into the future, stop the idle tick | |
0e776768 RW |
992 | */ |
993 | void tick_nohz_idle_stop_tick(void) | |
994 | { | |
995 | __tick_nohz_idle_stop_tick(this_cpu_ptr(&tick_cpu_sched)); | |
996 | } | |
997 | ||
554c8aa8 RW |
998 | void tick_nohz_idle_retain_tick(void) |
999 | { | |
1000 | tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched)); | |
1001 | /* | |
1002 | * Undo the effect of get_next_timer_interrupt() called from | |
1003 | * tick_nohz_next_event(). | |
1004 | */ | |
1005 | timer_clear_idle(); | |
1006 | } | |
1007 | ||
0e776768 RW |
1008 | /** |
1009 | * tick_nohz_idle_enter - prepare for entering idle on the current CPU | |
2bbb6817 | 1010 | * |
0e776768 | 1011 | * Called when we start the idle loop. |
280f0677 | 1012 | */ |
1268fbc7 | 1013 | void tick_nohz_idle_enter(void) |
280f0677 FW |
1014 | { |
1015 | struct tick_sched *ts; | |
1016 | ||
ebf3adba | 1017 | lockdep_assert_irqs_enabled(); |
0db49b72 | 1018 | |
1268fbc7 FW |
1019 | local_irq_disable(); |
1020 | ||
22127e93 | 1021 | ts = this_cpu_ptr(&tick_cpu_sched); |
23a8d888 RW |
1022 | |
1023 | WARN_ON_ONCE(ts->timer_expires_base); | |
1024 | ||
280f0677 | 1025 | ts->inidle = 1; |
0e776768 | 1026 | tick_nohz_start_idle(ts); |
1268fbc7 FW |
1027 | |
1028 | local_irq_enable(); | |
280f0677 FW |
1029 | } |
1030 | ||
1031 | /** | |
1032 | * tick_nohz_irq_exit - update next tick event from interrupt exit | |
1033 | * | |
1034 | * When an interrupt fires while we are idle and it doesn't cause | |
1035 | * a reschedule, it may still add, modify or delete a timer, enqueue | |
1036 | * an RCU callback, etc... | |
1037 | * So we need to re-calculate and reprogram the next tick event. | |
1038 | */ | |
1039 | void tick_nohz_irq_exit(void) | |
1040 | { | |
22127e93 | 1041 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
280f0677 | 1042 | |
14851912 | 1043 | if (ts->inidle) |
0e776768 | 1044 | tick_nohz_start_idle(ts); |
14851912 | 1045 | else |
73738a95 | 1046 | tick_nohz_full_update_tick(ts); |
79bf2bb3 TG |
1047 | } |
1048 | ||
4f86d3a8 | 1049 | /** |
45f1ff59 RW |
1050 | * tick_nohz_idle_got_tick - Check whether or not the tick handler has run |
1051 | */ | |
1052 | bool tick_nohz_idle_got_tick(void) | |
1053 | { | |
1054 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); | |
1055 | ||
2bc629a6 FW |
1056 | if (ts->got_idle_tick) { |
1057 | ts->got_idle_tick = 0; | |
45f1ff59 RW |
1058 | return true; |
1059 | } | |
1060 | return false; | |
1061 | } | |
1062 | ||
6f9b83ac UH |
1063 | /** |
1064 | * tick_nohz_get_next_hrtimer - return the next expiration time for the hrtimer | |
1065 | * or the tick, whatever that expires first. Note that, if the tick has been | |
1066 | * stopped, it returns the next hrtimer. | |
1067 | * | |
1068 | * Called from power state control code with interrupts disabled | |
1069 | */ | |
1070 | ktime_t tick_nohz_get_next_hrtimer(void) | |
1071 | { | |
1072 | return __this_cpu_read(tick_cpu_device.evtdev)->next_event; | |
1073 | } | |
1074 | ||
4f86d3a8 | 1075 | /** |
554c8aa8 | 1076 | * tick_nohz_get_sleep_length - return the expected length of the current sleep |
296bb1e5 | 1077 | * @delta_next: duration until the next event if the tick cannot be stopped |
4f86d3a8 LB |
1078 | * |
1079 | * Called from power state control code with interrupts disabled | |
1080 | */ | |
296bb1e5 | 1081 | ktime_t tick_nohz_get_sleep_length(ktime_t *delta_next) |
4f86d3a8 | 1082 | { |
554c8aa8 | 1083 | struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev); |
22127e93 | 1084 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
554c8aa8 RW |
1085 | int cpu = smp_processor_id(); |
1086 | /* | |
1087 | * The idle entry time is expected to be a sufficient approximation of | |
1088 | * the current time at this point. | |
1089 | */ | |
1090 | ktime_t now = ts->idle_entrytime; | |
1091 | ktime_t next_event; | |
1092 | ||
1093 | WARN_ON_ONCE(!ts->inidle); | |
1094 | ||
296bb1e5 RW |
1095 | *delta_next = ktime_sub(dev->next_event, now); |
1096 | ||
554c8aa8 | 1097 | if (!can_stop_idle_tick(cpu, ts)) |
296bb1e5 | 1098 | return *delta_next; |
554c8aa8 RW |
1099 | |
1100 | next_event = tick_nohz_next_event(ts, cpu); | |
1101 | if (!next_event) | |
296bb1e5 | 1102 | return *delta_next; |
554c8aa8 RW |
1103 | |
1104 | /* | |
1105 | * If the next highres timer to expire is earlier than next_event, the | |
1106 | * idle governor needs to know that. | |
1107 | */ | |
1108 | next_event = min_t(u64, next_event, | |
1109 | hrtimer_next_event_without(&ts->sched_timer)); | |
4f86d3a8 | 1110 | |
554c8aa8 | 1111 | return ktime_sub(next_event, now); |
4f86d3a8 LB |
1112 | } |
1113 | ||
466a2b42 JF |
1114 | /** |
1115 | * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value | |
1116 | * for a particular CPU. | |
1117 | * | |
1118 | * Called from the schedutil frequency scaling governor in scheduler context. | |
1119 | */ | |
1120 | unsigned long tick_nohz_get_idle_calls_cpu(int cpu) | |
1121 | { | |
1122 | struct tick_sched *ts = tick_get_tick_sched(cpu); | |
1123 | ||
1124 | return ts->idle_calls; | |
1125 | } | |
1126 | ||
b7eaf1aa RW |
1127 | /** |
1128 | * tick_nohz_get_idle_calls - return the current idle calls counter value | |
1129 | * | |
1130 | * Called from the schedutil frequency scaling governor in scheduler context. | |
1131 | */ | |
1132 | unsigned long tick_nohz_get_idle_calls(void) | |
1133 | { | |
1134 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); | |
1135 | ||
1136 | return ts->idle_calls; | |
1137 | } | |
1138 | ||
2ac0d98f FW |
1139 | static void tick_nohz_account_idle_ticks(struct tick_sched *ts) |
1140 | { | |
3f4724ea | 1141 | #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE |
2ac0d98f | 1142 | unsigned long ticks; |
3f4724ea | 1143 | |
e44fcb4b | 1144 | if (vtime_accounting_enabled_this_cpu()) |
3f4724ea | 1145 | return; |
79bf2bb3 TG |
1146 | /* |
1147 | * We stopped the tick in idle. Update process times would miss the | |
1148 | * time we slept as update_process_times does only a 1 tick | |
1149 | * accounting. Enforce that this is accounted to idle ! | |
1150 | */ | |
1151 | ticks = jiffies - ts->idle_jiffies; | |
1152 | /* | |
1153 | * We might be one off. Do not randomly account a huge number of ticks! | |
1154 | */ | |
79741dd3 MS |
1155 | if (ticks && ticks < LONG_MAX) |
1156 | account_idle_ticks(ticks); | |
1157 | #endif | |
19f5f736 FW |
1158 | } |
1159 | ||
2aaf709a RW |
1160 | static void __tick_nohz_idle_restart_tick(struct tick_sched *ts, ktime_t now) |
1161 | { | |
1162 | tick_nohz_restart_sched_tick(ts, now); | |
1163 | tick_nohz_account_idle_ticks(ts); | |
1164 | } | |
1165 | ||
1166 | void tick_nohz_idle_restart_tick(void) | |
1167 | { | |
1168 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); | |
1169 | ||
1170 | if (ts->tick_stopped) | |
1171 | __tick_nohz_idle_restart_tick(ts, ktime_get()); | |
1172 | } | |
1173 | ||
79bf2bb3 | 1174 | /** |
280f0677 | 1175 | * tick_nohz_idle_exit - restart the idle tick from the idle task |
79bf2bb3 TG |
1176 | * |
1177 | * Restart the idle tick when the CPU is woken up from idle | |
280f0677 FW |
1178 | * This also exit the RCU extended quiescent state. The CPU |
1179 | * can use RCU again after this function is called. | |
79bf2bb3 | 1180 | */ |
280f0677 | 1181 | void tick_nohz_idle_exit(void) |
79bf2bb3 | 1182 | { |
4a32fea9 | 1183 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
bbe9a70a | 1184 | bool idle_active, tick_stopped; |
6378ddb5 | 1185 | ktime_t now; |
79bf2bb3 | 1186 | |
6378ddb5 | 1187 | local_irq_disable(); |
2bbb6817 | 1188 | |
15f827be | 1189 | WARN_ON_ONCE(!ts->inidle); |
23a8d888 | 1190 | WARN_ON_ONCE(ts->timer_expires_base); |
15f827be FW |
1191 | |
1192 | ts->inidle = 0; | |
bbe9a70a AB |
1193 | idle_active = ts->idle_active; |
1194 | tick_stopped = ts->tick_stopped; | |
15f827be | 1195 | |
bbe9a70a | 1196 | if (idle_active || tick_stopped) |
eed3b9cf MS |
1197 | now = ktime_get(); |
1198 | ||
bbe9a70a | 1199 | if (idle_active) |
e8fcaa5c | 1200 | tick_nohz_stop_idle(ts, now); |
6378ddb5 | 1201 | |
bbe9a70a | 1202 | if (tick_stopped) |
2aaf709a | 1203 | __tick_nohz_idle_restart_tick(ts, now); |
79bf2bb3 | 1204 | |
79bf2bb3 TG |
1205 | local_irq_enable(); |
1206 | } | |
1207 | ||
79bf2bb3 TG |
1208 | /* |
1209 | * The nohz low res interrupt handler | |
1210 | */ | |
1211 | static void tick_nohz_handler(struct clock_event_device *dev) | |
1212 | { | |
22127e93 | 1213 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
79bf2bb3 TG |
1214 | struct pt_regs *regs = get_irq_regs(); |
1215 | ktime_t now = ktime_get(); | |
1216 | ||
2456e855 | 1217 | dev->next_event = KTIME_MAX; |
79bf2bb3 | 1218 | |
ff7de620 | 1219 | tick_sched_do_timer(ts, now); |
9e8f559b | 1220 | tick_sched_handle(ts, regs); |
79bf2bb3 | 1221 | |
b5e995e6 VK |
1222 | /* No need to reprogram if we are running tickless */ |
1223 | if (unlikely(ts->tick_stopped)) | |
1224 | return; | |
1225 | ||
0ff53d09 TG |
1226 | hrtimer_forward(&ts->sched_timer, now, tick_period); |
1227 | tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); | |
79bf2bb3 TG |
1228 | } |
1229 | ||
bc7a34b8 TG |
1230 | static inline void tick_nohz_activate(struct tick_sched *ts, int mode) |
1231 | { | |
1232 | if (!tick_nohz_enabled) | |
1233 | return; | |
1234 | ts->nohz_mode = mode; | |
1235 | /* One update is enough */ | |
1236 | if (!test_and_set_bit(0, &tick_nohz_active)) | |
ae67bada | 1237 | timers_update_nohz(); |
bc7a34b8 TG |
1238 | } |
1239 | ||
79bf2bb3 TG |
1240 | /** |
1241 | * tick_nohz_switch_to_nohz - switch to nohz mode | |
1242 | */ | |
1243 | static void tick_nohz_switch_to_nohz(void) | |
1244 | { | |
22127e93 | 1245 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
79bf2bb3 TG |
1246 | ktime_t next; |
1247 | ||
27630532 | 1248 | if (!tick_nohz_enabled) |
79bf2bb3 TG |
1249 | return; |
1250 | ||
6b442bc8 | 1251 | if (tick_switch_to_oneshot(tick_nohz_handler)) |
79bf2bb3 | 1252 | return; |
6b442bc8 | 1253 | |
79bf2bb3 TG |
1254 | /* |
1255 | * Recycle the hrtimer in ts, so we can share the | |
1256 | * hrtimer_forward with the highres code. | |
1257 | */ | |
71fed982 | 1258 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD); |
79bf2bb3 TG |
1259 | /* Get the next period */ |
1260 | next = tick_init_jiffy_update(); | |
1261 | ||
0ff53d09 | 1262 | hrtimer_set_expires(&ts->sched_timer, next); |
1ca8ec53 WL |
1263 | hrtimer_forward_now(&ts->sched_timer, tick_period); |
1264 | tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); | |
bc7a34b8 | 1265 | tick_nohz_activate(ts, NOHZ_MODE_LOWRES); |
79bf2bb3 TG |
1266 | } |
1267 | ||
5acac1be | 1268 | static inline void tick_nohz_irq_enter(void) |
eed3b9cf | 1269 | { |
4a32fea9 | 1270 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
eed3b9cf MS |
1271 | ktime_t now; |
1272 | ||
1273 | if (!ts->idle_active && !ts->tick_stopped) | |
1274 | return; | |
1275 | now = ktime_get(); | |
1276 | if (ts->idle_active) | |
e8fcaa5c | 1277 | tick_nohz_stop_idle(ts, now); |
ff006732 | 1278 | if (ts->tick_stopped) |
eed3b9cf | 1279 | tick_nohz_update_jiffies(now); |
eed3b9cf MS |
1280 | } |
1281 | ||
79bf2bb3 TG |
1282 | #else |
1283 | ||
1284 | static inline void tick_nohz_switch_to_nohz(void) { } | |
5acac1be | 1285 | static inline void tick_nohz_irq_enter(void) { } |
bc7a34b8 | 1286 | static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { } |
79bf2bb3 | 1287 | |
3451d024 | 1288 | #endif /* CONFIG_NO_HZ_COMMON */ |
79bf2bb3 | 1289 | |
719254fa TG |
1290 | /* |
1291 | * Called from irq_enter to notify about the possible interruption of idle() | |
1292 | */ | |
5acac1be | 1293 | void tick_irq_enter(void) |
719254fa | 1294 | { |
e8fcaa5c | 1295 | tick_check_oneshot_broadcast_this_cpu(); |
5acac1be | 1296 | tick_nohz_irq_enter(); |
719254fa TG |
1297 | } |
1298 | ||
79bf2bb3 TG |
1299 | /* |
1300 | * High resolution timer specific code | |
1301 | */ | |
1302 | #ifdef CONFIG_HIGH_RES_TIMERS | |
1303 | /* | |
4c9dc641 | 1304 | * We rearm the timer until we get disabled by the idle code. |
351f181f | 1305 | * Called with interrupts disabled. |
79bf2bb3 TG |
1306 | */ |
1307 | static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) | |
1308 | { | |
1309 | struct tick_sched *ts = | |
1310 | container_of(timer, struct tick_sched, sched_timer); | |
79bf2bb3 TG |
1311 | struct pt_regs *regs = get_irq_regs(); |
1312 | ktime_t now = ktime_get(); | |
d3ed7824 | 1313 | |
ff7de620 | 1314 | tick_sched_do_timer(ts, now); |
79bf2bb3 TG |
1315 | |
1316 | /* | |
1317 | * Do not call, when we are not in irq context and have | |
1318 | * no valid regs pointer | |
1319 | */ | |
9e8f559b FW |
1320 | if (regs) |
1321 | tick_sched_handle(ts, regs); | |
7c259045 FW |
1322 | else |
1323 | ts->next_tick = 0; | |
79bf2bb3 | 1324 | |
2a16fc93 VK |
1325 | /* No need to reprogram if we are in idle or full dynticks mode */ |
1326 | if (unlikely(ts->tick_stopped)) | |
1327 | return HRTIMER_NORESTART; | |
1328 | ||
79bf2bb3 TG |
1329 | hrtimer_forward(timer, now, tick_period); |
1330 | ||
1331 | return HRTIMER_RESTART; | |
1332 | } | |
1333 | ||
5307c955 MG |
1334 | static int sched_skew_tick; |
1335 | ||
62cf20b3 TG |
1336 | static int __init skew_tick(char *str) |
1337 | { | |
1338 | get_option(&str, &sched_skew_tick); | |
1339 | ||
1340 | return 0; | |
1341 | } | |
1342 | early_param("skew_tick", skew_tick); | |
1343 | ||
79bf2bb3 TG |
1344 | /** |
1345 | * tick_setup_sched_timer - setup the tick emulation timer | |
1346 | */ | |
1347 | void tick_setup_sched_timer(void) | |
1348 | { | |
22127e93 | 1349 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
79bf2bb3 TG |
1350 | ktime_t now = ktime_get(); |
1351 | ||
1352 | /* | |
1353 | * Emulate tick processing via per-CPU hrtimers: | |
1354 | */ | |
902a9f9c | 1355 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD); |
79bf2bb3 | 1356 | ts->sched_timer.function = tick_sched_timer; |
79bf2bb3 | 1357 | |
0de7611a | 1358 | /* Get the next period (per-CPU) */ |
cc584b21 | 1359 | hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); |
79bf2bb3 | 1360 | |
9c3f9e28 | 1361 | /* Offset the tick to avert jiffies_lock contention. */ |
5307c955 MG |
1362 | if (sched_skew_tick) { |
1363 | u64 offset = ktime_to_ns(tick_period) >> 1; | |
1364 | do_div(offset, num_possible_cpus()); | |
1365 | offset *= smp_processor_id(); | |
1366 | hrtimer_add_expires_ns(&ts->sched_timer, offset); | |
1367 | } | |
1368 | ||
afc08b15 | 1369 | hrtimer_forward(&ts->sched_timer, now, tick_period); |
902a9f9c | 1370 | hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED_HARD); |
bc7a34b8 | 1371 | tick_nohz_activate(ts, NOHZ_MODE_HIGHRES); |
79bf2bb3 | 1372 | } |
3c4fbe5e | 1373 | #endif /* HIGH_RES_TIMERS */ |
79bf2bb3 | 1374 | |
3451d024 | 1375 | #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS |
79bf2bb3 TG |
1376 | void tick_cancel_sched_timer(int cpu) |
1377 | { | |
1378 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
1379 | ||
3c4fbe5e | 1380 | # ifdef CONFIG_HIGH_RES_TIMERS |
79bf2bb3 TG |
1381 | if (ts->sched_timer.base) |
1382 | hrtimer_cancel(&ts->sched_timer); | |
3c4fbe5e | 1383 | # endif |
a7901766 | 1384 | |
4b0c0f29 | 1385 | memset(ts, 0, sizeof(*ts)); |
79bf2bb3 | 1386 | } |
3c4fbe5e | 1387 | #endif |
79bf2bb3 TG |
1388 | |
1389 | /** | |
1390 | * Async notification about clocksource changes | |
1391 | */ | |
1392 | void tick_clock_notify(void) | |
1393 | { | |
1394 | int cpu; | |
1395 | ||
1396 | for_each_possible_cpu(cpu) | |
1397 | set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); | |
1398 | } | |
1399 | ||
1400 | /* | |
1401 | * Async notification about clock event changes | |
1402 | */ | |
1403 | void tick_oneshot_notify(void) | |
1404 | { | |
22127e93 | 1405 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
79bf2bb3 TG |
1406 | |
1407 | set_bit(0, &ts->check_clocks); | |
1408 | } | |
1409 | ||
1410 | /** | |
1411 | * Check, if a change happened, which makes oneshot possible. | |
1412 | * | |
1413 | * Called cyclic from the hrtimer softirq (driven by the timer | |
1414 | * softirq) allow_nohz signals, that we can switch into low-res nohz | |
1415 | * mode, because high resolution timers are disabled (either compile | |
6b442bc8 | 1416 | * or runtime). Called with interrupts disabled. |
79bf2bb3 TG |
1417 | */ |
1418 | int tick_check_oneshot_change(int allow_nohz) | |
1419 | { | |
22127e93 | 1420 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
79bf2bb3 TG |
1421 | |
1422 | if (!test_and_clear_bit(0, &ts->check_clocks)) | |
1423 | return 0; | |
1424 | ||
1425 | if (ts->nohz_mode != NOHZ_MODE_INACTIVE) | |
1426 | return 0; | |
1427 | ||
cf4fc6cb | 1428 | if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) |
79bf2bb3 TG |
1429 | return 0; |
1430 | ||
1431 | if (!allow_nohz) | |
1432 | return 1; | |
1433 | ||
1434 | tick_nohz_switch_to_nohz(); | |
1435 | return 0; | |
1436 | } |