]>
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> |
79bf2bb3 | 30 | |
9e203bcc DM |
31 | #include <asm/irq_regs.h> |
32 | ||
79bf2bb3 TG |
33 | #include "tick-internal.h" |
34 | ||
cb41a290 FW |
35 | #include <trace/events/timer.h> |
36 | ||
79bf2bb3 | 37 | /* |
0de7611a | 38 | * Per-CPU nohz control structure |
79bf2bb3 | 39 | */ |
c1797baf | 40 | static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched); |
79bf2bb3 | 41 | |
289f480a IM |
42 | struct tick_sched *tick_get_tick_sched(int cpu) |
43 | { | |
44 | return &per_cpu(tick_cpu_sched, cpu); | |
45 | } | |
46 | ||
7809998a AB |
47 | #if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS) |
48 | /* | |
49 | * The time, when the last jiffy update happened. Protected by jiffies_lock. | |
50 | */ | |
51 | static ktime_t last_jiffies_update; | |
52 | ||
79bf2bb3 TG |
53 | /* |
54 | * Must be called with interrupts disabled ! | |
55 | */ | |
56 | static void tick_do_update_jiffies64(ktime_t now) | |
57 | { | |
58 | unsigned long ticks = 0; | |
59 | ktime_t delta; | |
60 | ||
7a14ce1d | 61 | /* |
d6ad4187 | 62 | * Do a quick check without holding jiffies_lock: |
7a14ce1d IM |
63 | */ |
64 | delta = ktime_sub(now, last_jiffies_update); | |
2456e855 | 65 | if (delta < tick_period) |
7a14ce1d IM |
66 | return; |
67 | ||
6168f8ed | 68 | /* Reevaluate with jiffies_lock held */ |
d6ad4187 | 69 | write_seqlock(&jiffies_lock); |
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); | |
75 | last_jiffies_update = ktime_add(last_jiffies_update, | |
76 | tick_period); | |
77 | ||
78 | /* Slow path for long timeouts */ | |
2456e855 | 79 | if (unlikely(delta >= tick_period)) { |
79bf2bb3 TG |
80 | s64 incr = ktime_to_ns(tick_period); |
81 | ||
82 | ticks = ktime_divns(delta, incr); | |
83 | ||
84 | last_jiffies_update = ktime_add_ns(last_jiffies_update, | |
85 | incr * ticks); | |
86 | } | |
87 | do_timer(++ticks); | |
49d670fb TG |
88 | |
89 | /* Keep the tick_next_period variable up to date */ | |
90 | tick_next_period = ktime_add(last_jiffies_update, tick_period); | |
03e6bdc5 VK |
91 | } else { |
92 | write_sequnlock(&jiffies_lock); | |
93 | return; | |
79bf2bb3 | 94 | } |
d6ad4187 | 95 | write_sequnlock(&jiffies_lock); |
47a1b796 | 96 | update_wall_time(); |
79bf2bb3 TG |
97 | } |
98 | ||
99 | /* | |
100 | * Initialize and return retrieve the jiffies update. | |
101 | */ | |
102 | static ktime_t tick_init_jiffy_update(void) | |
103 | { | |
104 | ktime_t period; | |
105 | ||
d6ad4187 | 106 | write_seqlock(&jiffies_lock); |
79bf2bb3 | 107 | /* Did we start the jiffies update yet ? */ |
2456e855 | 108 | if (last_jiffies_update == 0) |
79bf2bb3 TG |
109 | last_jiffies_update = tick_next_period; |
110 | period = last_jiffies_update; | |
d6ad4187 | 111 | write_sequnlock(&jiffies_lock); |
79bf2bb3 TG |
112 | return period; |
113 | } | |
114 | ||
5bb96226 FW |
115 | |
116 | static void tick_sched_do_timer(ktime_t now) | |
117 | { | |
118 | int cpu = smp_processor_id(); | |
119 | ||
3451d024 | 120 | #ifdef CONFIG_NO_HZ_COMMON |
5bb96226 FW |
121 | /* |
122 | * Check if the do_timer duty was dropped. We don't care about | |
0de7611a IM |
123 | * concurrency: This happens only when the CPU in charge went |
124 | * into a long sleep. If two CPUs happen to assign themselves to | |
5bb96226 | 125 | * this duty, then the jiffies update is still serialized by |
9c3f9e28 | 126 | * jiffies_lock. |
5bb96226 | 127 | */ |
a382bf93 | 128 | if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE) |
c5bfece2 | 129 | && !tick_nohz_full_cpu(cpu)) |
5bb96226 FW |
130 | tick_do_timer_cpu = cpu; |
131 | #endif | |
132 | ||
133 | /* Check, if the jiffies need an update */ | |
134 | if (tick_do_timer_cpu == cpu) | |
135 | tick_do_update_jiffies64(now); | |
136 | } | |
137 | ||
9e8f559b FW |
138 | static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs) |
139 | { | |
3451d024 | 140 | #ifdef CONFIG_NO_HZ_COMMON |
9e8f559b FW |
141 | /* |
142 | * When we are idle and the tick is stopped, we have to touch | |
143 | * the watchdog as we might not schedule for a really long | |
144 | * time. This happens on complete idle SMP systems while | |
145 | * waiting on the login prompt. We also increment the "start of | |
146 | * idle" jiffy stamp so the idle accounting adjustment we do | |
147 | * when we go busy again does not account too much ticks. | |
148 | */ | |
149 | if (ts->tick_stopped) { | |
03e0d461 | 150 | touch_softlockup_watchdog_sched(); |
9e8f559b FW |
151 | if (is_idle_task(current)) |
152 | ts->idle_jiffies++; | |
411fe24e FW |
153 | /* |
154 | * In case the current tick fired too early past its expected | |
155 | * expiration, make sure we don't bypass the next clock reprogramming | |
156 | * to the same deadline. | |
157 | */ | |
158 | ts->next_tick = 0; | |
9e8f559b | 159 | } |
94a57140 | 160 | #endif |
9e8f559b FW |
161 | update_process_times(user_mode(regs)); |
162 | profile_tick(CPU_PROFILING); | |
163 | } | |
7809998a | 164 | #endif |
9e8f559b | 165 | |
c5bfece2 | 166 | #ifdef CONFIG_NO_HZ_FULL |
460775df | 167 | cpumask_var_t tick_nohz_full_mask; |
c0f489d2 | 168 | cpumask_var_t housekeeping_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 FW |
200 | { |
201 | WARN_ON_ONCE(!irqs_disabled()); | |
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 | ||
a831881b | 388 | /* Parse the boot-time nohz CPU list from the kernel parameters. */ |
c5bfece2 | 389 | static int __init tick_nohz_full_setup(char *str) |
a831881b | 390 | { |
73867dcd FW |
391 | alloc_bootmem_cpumask_var(&tick_nohz_full_mask); |
392 | if (cpulist_parse(str, tick_nohz_full_mask) < 0) { | |
a395d6a7 | 393 | pr_warn("NO_HZ: Incorrect nohz_full cpumask\n"); |
4327b15f | 394 | free_bootmem_cpumask_var(tick_nohz_full_mask); |
0453b435 FW |
395 | return 1; |
396 | } | |
73867dcd | 397 | tick_nohz_full_running = true; |
0453b435 | 398 | |
a831881b FW |
399 | return 1; |
400 | } | |
c5bfece2 | 401 | __setup("nohz_full=", tick_nohz_full_setup); |
a831881b | 402 | |
31eff243 | 403 | static int tick_nohz_cpu_down(unsigned int cpu) |
a382bf93 | 404 | { |
31eff243 SAS |
405 | /* |
406 | * The boot CPU handles housekeeping duty (unbound timers, | |
407 | * workqueues, timekeeping, ...) on behalf of full dynticks | |
408 | * CPUs. It must remain online when nohz full is enabled. | |
409 | */ | |
410 | if (tick_nohz_full_running && tick_do_timer_cpu == cpu) | |
411 | return -EBUSY; | |
412 | return 0; | |
a382bf93 FW |
413 | } |
414 | ||
f98823ac FW |
415 | static int tick_nohz_init_all(void) |
416 | { | |
417 | int err = -1; | |
418 | ||
419 | #ifdef CONFIG_NO_HZ_FULL_ALL | |
73867dcd | 420 | if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) { |
4327b15f | 421 | WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n"); |
c0f489d2 PM |
422 | return err; |
423 | } | |
f98823ac | 424 | err = 0; |
73867dcd | 425 | cpumask_setall(tick_nohz_full_mask); |
73867dcd | 426 | tick_nohz_full_running = true; |
f98823ac FW |
427 | #endif |
428 | return err; | |
429 | } | |
430 | ||
d1e43fa5 | 431 | void __init tick_nohz_init(void) |
a831881b | 432 | { |
31eff243 | 433 | int cpu, ret; |
d1e43fa5 | 434 | |
73867dcd | 435 | if (!tick_nohz_full_running) { |
f98823ac FW |
436 | if (tick_nohz_init_all() < 0) |
437 | return; | |
438 | } | |
d1e43fa5 | 439 | |
4327b15f FW |
440 | if (!alloc_cpumask_var(&housekeeping_mask, GFP_KERNEL)) { |
441 | WARN(1, "NO_HZ: Can't allocate not-full dynticks cpumask\n"); | |
442 | cpumask_clear(tick_nohz_full_mask); | |
443 | tick_nohz_full_running = false; | |
444 | return; | |
445 | } | |
446 | ||
9b01f5bf FW |
447 | /* |
448 | * Full dynticks uses irq work to drive the tick rescheduling on safe | |
449 | * locking contexts. But then we need irq work to raise its own | |
450 | * interrupts to avoid circular dependency on the tick | |
451 | */ | |
452 | if (!arch_irq_work_has_interrupt()) { | |
a395d6a7 | 453 | pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n"); |
9b01f5bf FW |
454 | cpumask_clear(tick_nohz_full_mask); |
455 | cpumask_copy(housekeeping_mask, cpu_possible_mask); | |
456 | tick_nohz_full_running = false; | |
457 | return; | |
458 | } | |
459 | ||
4327b15f FW |
460 | cpu = smp_processor_id(); |
461 | ||
462 | if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) { | |
a395d6a7 JP |
463 | pr_warn("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", |
464 | cpu); | |
4327b15f FW |
465 | cpumask_clear_cpu(cpu, tick_nohz_full_mask); |
466 | } | |
467 | ||
468 | cpumask_andnot(housekeeping_mask, | |
469 | cpu_possible_mask, tick_nohz_full_mask); | |
470 | ||
73867dcd | 471 | for_each_cpu(cpu, tick_nohz_full_mask) |
2e709338 FW |
472 | context_tracking_cpu_set(cpu); |
473 | ||
31eff243 SAS |
474 | ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, |
475 | "kernel/nohz:predown", NULL, | |
476 | tick_nohz_cpu_down); | |
477 | WARN_ON(ret < 0); | |
ffda22c1 TH |
478 | pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n", |
479 | cpumask_pr_args(tick_nohz_full_mask)); | |
7c8bb6cb FW |
480 | |
481 | /* | |
482 | * We need at least one CPU to handle housekeeping work such | |
483 | * as timekeeping, unbound timers, workqueues, ... | |
484 | */ | |
485 | WARN_ON_ONCE(cpumask_empty(housekeeping_mask)); | |
a831881b | 486 | } |
a831881b FW |
487 | #endif |
488 | ||
79bf2bb3 TG |
489 | /* |
490 | * NOHZ - aka dynamic tick functionality | |
491 | */ | |
3451d024 | 492 | #ifdef CONFIG_NO_HZ_COMMON |
79bf2bb3 TG |
493 | /* |
494 | * NO HZ enabled ? | |
495 | */ | |
4cc7ecb7 | 496 | bool tick_nohz_enabled __read_mostly = true; |
bc7a34b8 | 497 | unsigned long tick_nohz_active __read_mostly; |
79bf2bb3 TG |
498 | /* |
499 | * Enable / Disable tickless mode | |
500 | */ | |
501 | static int __init setup_tick_nohz(char *str) | |
502 | { | |
4cc7ecb7 | 503 | return (kstrtobool(str, &tick_nohz_enabled) == 0); |
79bf2bb3 TG |
504 | } |
505 | ||
506 | __setup("nohz=", setup_tick_nohz); | |
507 | ||
c1797baf TG |
508 | int tick_nohz_tick_stopped(void) |
509 | { | |
510 | return __this_cpu_read(tick_cpu_sched.tick_stopped); | |
511 | } | |
512 | ||
79bf2bb3 TG |
513 | /** |
514 | * tick_nohz_update_jiffies - update jiffies when idle was interrupted | |
515 | * | |
516 | * Called from interrupt entry when the CPU was idle | |
517 | * | |
518 | * In case the sched_tick was stopped on this CPU, we have to check if jiffies | |
519 | * must be updated. Otherwise an interrupt handler could use a stale jiffy | |
0de7611a IM |
520 | * value. We do this unconditionally on any CPU, as we don't know whether the |
521 | * CPU, which has the update task assigned is in a long sleep. | |
79bf2bb3 | 522 | */ |
eed3b9cf | 523 | static void tick_nohz_update_jiffies(ktime_t now) |
79bf2bb3 | 524 | { |
79bf2bb3 | 525 | unsigned long flags; |
79bf2bb3 | 526 | |
e8fcaa5c | 527 | __this_cpu_write(tick_cpu_sched.idle_waketime, now); |
79bf2bb3 TG |
528 | |
529 | local_irq_save(flags); | |
530 | tick_do_update_jiffies64(now); | |
531 | local_irq_restore(flags); | |
02ff3755 | 532 | |
03e0d461 | 533 | touch_softlockup_watchdog_sched(); |
79bf2bb3 TG |
534 | } |
535 | ||
595aac48 | 536 | /* |
0de7611a | 537 | * Updates the per-CPU time idle statistics counters |
595aac48 | 538 | */ |
8d63bf94 | 539 | static void |
8c215bd3 | 540 | update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time) |
6378ddb5 | 541 | { |
eed3b9cf | 542 | ktime_t delta; |
6378ddb5 | 543 | |
595aac48 AV |
544 | if (ts->idle_active) { |
545 | delta = ktime_sub(now, ts->idle_entrytime); | |
8c215bd3 | 546 | if (nr_iowait_cpu(cpu) > 0) |
0224cf4c | 547 | ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); |
6beea0cd MH |
548 | else |
549 | ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); | |
8c7b09f4 | 550 | ts->idle_entrytime = now; |
595aac48 | 551 | } |
8d63bf94 | 552 | |
e0e37c20 | 553 | if (last_update_time) |
8d63bf94 AV |
554 | *last_update_time = ktime_to_us(now); |
555 | ||
595aac48 AV |
556 | } |
557 | ||
e8fcaa5c | 558 | static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now) |
595aac48 | 559 | { |
e8fcaa5c | 560 | update_ts_time_stats(smp_processor_id(), ts, now, NULL); |
eed3b9cf | 561 | ts->idle_active = 0; |
56c7426b | 562 | |
ac1e843f | 563 | sched_clock_idle_wakeup_event(); |
6378ddb5 VP |
564 | } |
565 | ||
e8fcaa5c | 566 | static ktime_t tick_nohz_start_idle(struct tick_sched *ts) |
6378ddb5 | 567 | { |
430ee881 | 568 | ktime_t now = ktime_get(); |
595aac48 | 569 | |
6378ddb5 VP |
570 | ts->idle_entrytime = now; |
571 | ts->idle_active = 1; | |
56c7426b | 572 | sched_clock_idle_sleep_event(); |
6378ddb5 VP |
573 | return now; |
574 | } | |
575 | ||
b1f724c3 | 576 | /** |
0de7611a | 577 | * get_cpu_idle_time_us - get the total idle time of a CPU |
b1f724c3 | 578 | * @cpu: CPU number to query |
09a1d34f MH |
579 | * @last_update_time: variable to store update time in. Do not update |
580 | * counters if NULL. | |
b1f724c3 | 581 | * |
6168f8ed | 582 | * Return the cumulative idle time (since boot) for a given |
6beea0cd | 583 | * CPU, in microseconds. |
b1f724c3 AV |
584 | * |
585 | * This time is measured via accounting rather than sampling, | |
586 | * and is as accurate as ktime_get() is. | |
587 | * | |
588 | * This function returns -1 if NOHZ is not enabled. | |
589 | */ | |
6378ddb5 VP |
590 | u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) |
591 | { | |
592 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
09a1d34f | 593 | ktime_t now, idle; |
6378ddb5 | 594 | |
d689fe22 | 595 | if (!tick_nohz_active) |
8083e4ad | 596 | return -1; |
597 | ||
09a1d34f MH |
598 | now = ktime_get(); |
599 | if (last_update_time) { | |
600 | update_ts_time_stats(cpu, ts, now, last_update_time); | |
601 | idle = ts->idle_sleeptime; | |
602 | } else { | |
603 | if (ts->idle_active && !nr_iowait_cpu(cpu)) { | |
604 | ktime_t delta = ktime_sub(now, ts->idle_entrytime); | |
605 | ||
606 | idle = ktime_add(ts->idle_sleeptime, delta); | |
607 | } else { | |
608 | idle = ts->idle_sleeptime; | |
609 | } | |
610 | } | |
611 | ||
612 | return ktime_to_us(idle); | |
8083e4ad | 613 | |
6378ddb5 | 614 | } |
8083e4ad | 615 | EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); |
6378ddb5 | 616 | |
6beea0cd | 617 | /** |
0de7611a | 618 | * get_cpu_iowait_time_us - get the total iowait time of a CPU |
0224cf4c | 619 | * @cpu: CPU number to query |
09a1d34f MH |
620 | * @last_update_time: variable to store update time in. Do not update |
621 | * counters if NULL. | |
0224cf4c | 622 | * |
6168f8ed | 623 | * Return the cumulative iowait time (since boot) for a given |
0224cf4c AV |
624 | * CPU, in microseconds. |
625 | * | |
626 | * This time is measured via accounting rather than sampling, | |
627 | * and is as accurate as ktime_get() is. | |
628 | * | |
629 | * This function returns -1 if NOHZ is not enabled. | |
630 | */ | |
631 | u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) | |
632 | { | |
633 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
09a1d34f | 634 | ktime_t now, iowait; |
0224cf4c | 635 | |
d689fe22 | 636 | if (!tick_nohz_active) |
0224cf4c AV |
637 | return -1; |
638 | ||
09a1d34f MH |
639 | now = ktime_get(); |
640 | if (last_update_time) { | |
641 | update_ts_time_stats(cpu, ts, now, last_update_time); | |
642 | iowait = ts->iowait_sleeptime; | |
643 | } else { | |
644 | if (ts->idle_active && nr_iowait_cpu(cpu) > 0) { | |
645 | ktime_t delta = ktime_sub(now, ts->idle_entrytime); | |
0224cf4c | 646 | |
09a1d34f MH |
647 | iowait = ktime_add(ts->iowait_sleeptime, delta); |
648 | } else { | |
649 | iowait = ts->iowait_sleeptime; | |
650 | } | |
651 | } | |
0224cf4c | 652 | |
09a1d34f | 653 | return ktime_to_us(iowait); |
0224cf4c AV |
654 | } |
655 | EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); | |
656 | ||
0ff53d09 TG |
657 | static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) |
658 | { | |
659 | hrtimer_cancel(&ts->sched_timer); | |
660 | hrtimer_set_expires(&ts->sched_timer, ts->last_tick); | |
661 | ||
662 | /* Forward the time to expire in the future */ | |
663 | hrtimer_forward(&ts->sched_timer, now, tick_period); | |
664 | ||
665 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) | |
666 | hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); | |
667 | else | |
668 | tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); | |
411fe24e FW |
669 | |
670 | /* | |
671 | * Reset to make sure next tick stop doesn't get fooled by past | |
672 | * cached clock deadline. | |
673 | */ | |
674 | ts->next_tick = 0; | |
0ff53d09 TG |
675 | } |
676 | ||
84bf1bcc FW |
677 | static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts, |
678 | ktime_t now, int cpu) | |
79bf2bb3 | 679 | { |
22127e93 | 680 | struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev); |
c1ad348b TG |
681 | u64 basemono, next_tick, next_tmr, next_rcu, delta, expires; |
682 | unsigned long seq, basejiff; | |
683 | ktime_t tick; | |
855a0fc3 | 684 | |
79bf2bb3 TG |
685 | /* Read jiffies and the time when jiffies were updated last */ |
686 | do { | |
d6ad4187 | 687 | seq = read_seqbegin(&jiffies_lock); |
2456e855 | 688 | basemono = last_jiffies_update; |
c1ad348b | 689 | basejiff = jiffies; |
d6ad4187 | 690 | } while (read_seqretry(&jiffies_lock, seq)); |
c1ad348b | 691 | ts->last_jiffies = basejiff; |
79bf2bb3 | 692 | |
c1ad348b | 693 | if (rcu_needs_cpu(basemono, &next_rcu) || |
fe0f4976 | 694 | arch_needs_cpu() || irq_work_needs_cpu()) { |
c1ad348b | 695 | next_tick = basemono + TICK_NSEC; |
3c5d92a0 | 696 | } else { |
c1ad348b TG |
697 | /* |
698 | * Get the next pending timer. If high resolution | |
699 | * timers are enabled this only takes the timer wheel | |
700 | * timers into account. If high resolution timers are | |
701 | * disabled this also looks at the next expiring | |
702 | * hrtimer. | |
703 | */ | |
704 | next_tmr = get_next_timer_interrupt(basejiff, basemono); | |
705 | ts->next_timer = next_tmr; | |
706 | /* Take the next rcu event into account */ | |
707 | next_tick = next_rcu < next_tmr ? next_rcu : next_tmr; | |
3c5d92a0 | 708 | } |
47aa8b6c | 709 | |
c1ad348b TG |
710 | /* |
711 | * If the tick is due in the next period, keep it ticking or | |
82bbe34b | 712 | * force prod the timer. |
c1ad348b TG |
713 | */ |
714 | delta = next_tick - basemono; | |
715 | if (delta <= (u64)TICK_NSEC) { | |
a683f390 TG |
716 | /* |
717 | * Tell the timer code that the base is not idle, i.e. undo | |
718 | * the effect of get_next_timer_interrupt(): | |
719 | */ | |
720 | timer_clear_idle(); | |
82bbe34b PZ |
721 | /* |
722 | * We've not stopped the tick yet, and there's a timer in the | |
723 | * next period, so no point in stopping it either, bail. | |
724 | */ | |
f99973e1 FW |
725 | if (!ts->tick_stopped) { |
726 | tick = 0; | |
157d29e1 TG |
727 | goto out; |
728 | } | |
729 | } | |
730 | ||
79bf2bb3 | 731 | /* |
0de7611a IM |
732 | * If this CPU is the one which updates jiffies, then give up |
733 | * the assignment and let it be taken by the CPU which runs | |
734 | * the tick timer next, which might be this CPU as well. If we | |
157d29e1 TG |
735 | * don't drop this here the jiffies might be stale and |
736 | * do_timer() never invoked. Keep track of the fact that it | |
0de7611a | 737 | * was the one which had the do_timer() duty last. If this CPU |
157d29e1 | 738 | * is the one which had the do_timer() duty last, we limit the |
6168f8ed | 739 | * sleep time to the timekeeping max_deferment value. |
c1ad348b | 740 | * Otherwise we can sleep as long as we want. |
79bf2bb3 | 741 | */ |
c1ad348b | 742 | delta = timekeeping_max_deferment(); |
157d29e1 TG |
743 | if (cpu == tick_do_timer_cpu) { |
744 | tick_do_timer_cpu = TICK_DO_TIMER_NONE; | |
745 | ts->do_timer_last = 1; | |
746 | } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { | |
c1ad348b | 747 | delta = KTIME_MAX; |
157d29e1 TG |
748 | ts->do_timer_last = 0; |
749 | } else if (!ts->do_timer_last) { | |
c1ad348b | 750 | delta = KTIME_MAX; |
157d29e1 | 751 | } |
27185016 | 752 | |
265f22a9 | 753 | #ifdef CONFIG_NO_HZ_FULL |
c1ad348b | 754 | /* Limit the tick delta to the maximum scheduler deferment */ |
157d29e1 | 755 | if (!ts->inidle) |
c1ad348b | 756 | delta = min(delta, scheduler_tick_max_deferment()); |
265f22a9 FW |
757 | #endif |
758 | ||
c1ad348b TG |
759 | /* Calculate the next expiry time */ |
760 | if (delta < (KTIME_MAX - basemono)) | |
761 | expires = basemono + delta; | |
157d29e1 | 762 | else |
c1ad348b TG |
763 | expires = KTIME_MAX; |
764 | ||
765 | expires = min_t(u64, expires, next_tick); | |
2456e855 | 766 | tick = expires; |
00147449 | 767 | |
157d29e1 | 768 | /* Skip reprogram of event if its not changed */ |
411fe24e FW |
769 | if (ts->tick_stopped && (expires == ts->next_tick)) { |
770 | /* Sanity check: make sure clockevent is actually programmed */ | |
d4af6d93 | 771 | if (tick == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer)) |
ce6cf9a1 | 772 | goto out; |
411fe24e FW |
773 | |
774 | WARN_ON_ONCE(1); | |
775 | printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n", | |
776 | basemono, ts->next_tick, dev->next_event, | |
777 | hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer)); | |
ce6cf9a1 | 778 | } |
84bf1bcc | 779 | |
157d29e1 TG |
780 | /* |
781 | * nohz_stop_sched_tick can be called several times before | |
782 | * the nohz_restart_sched_tick is called. This happens when | |
783 | * interrupts arrive which do not cause a reschedule. In the | |
784 | * first call we save the current tick time, so we can restart | |
785 | * the scheduler tick in nohz_restart_sched_tick. | |
786 | */ | |
787 | if (!ts->tick_stopped) { | |
3c85d6db | 788 | calc_load_nohz_start(); |
1f41906a | 789 | cpu_load_update_nohz_start(); |
d3ed7824 | 790 | |
157d29e1 TG |
791 | ts->last_tick = hrtimer_get_expires(&ts->sched_timer); |
792 | ts->tick_stopped = 1; | |
e6e6cc22 | 793 | trace_tick_stop(1, TICK_DEP_MASK_NONE); |
157d29e1 | 794 | } |
eaad084b | 795 | |
411fe24e FW |
796 | ts->next_tick = tick; |
797 | ||
157d29e1 | 798 | /* |
c1ad348b TG |
799 | * If the expiration time == KTIME_MAX, then we simply stop |
800 | * the tick timer. | |
157d29e1 | 801 | */ |
c1ad348b | 802 | if (unlikely(expires == KTIME_MAX)) { |
157d29e1 TG |
803 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) |
804 | hrtimer_cancel(&ts->sched_timer); | |
805 | goto out; | |
79bf2bb3 | 806 | } |
0ff53d09 | 807 | |
d4af6d93 FW |
808 | hrtimer_set_expires(&ts->sched_timer, tick); |
809 | ||
157d29e1 | 810 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) |
d4af6d93 | 811 | hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); |
157d29e1 | 812 | else |
c1ad348b | 813 | tick_program_event(tick, 1); |
79bf2bb3 | 814 | out: |
411fe24e FW |
815 | /* |
816 | * Update the estimated sleep length until the next timer | |
817 | * (not only the tick). | |
818 | */ | |
4f86d3a8 | 819 | ts->sleep_length = ktime_sub(dev->next_event, now); |
c1ad348b | 820 | return tick; |
280f0677 FW |
821 | } |
822 | ||
1f41906a | 823 | static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now) |
59d2c7ca FW |
824 | { |
825 | /* Update jiffies first */ | |
826 | tick_do_update_jiffies64(now); | |
1f41906a | 827 | cpu_load_update_nohz_stop(); |
59d2c7ca | 828 | |
a683f390 TG |
829 | /* |
830 | * Clear the timer idle flag, so we avoid IPIs on remote queueing and | |
831 | * the clock forward checks in the enqueue path: | |
832 | */ | |
833 | timer_clear_idle(); | |
834 | ||
3c85d6db | 835 | calc_load_nohz_stop(); |
03e0d461 | 836 | touch_softlockup_watchdog_sched(); |
59d2c7ca FW |
837 | /* |
838 | * Cancel the scheduled timer and restore the tick | |
839 | */ | |
840 | ts->tick_stopped = 0; | |
841 | ts->idle_exittime = now; | |
842 | ||
843 | tick_nohz_restart(ts, now); | |
844 | } | |
73738a95 FW |
845 | |
846 | static void tick_nohz_full_update_tick(struct tick_sched *ts) | |
5811d996 FW |
847 | { |
848 | #ifdef CONFIG_NO_HZ_FULL | |
e9a2eb40 | 849 | int cpu = smp_processor_id(); |
5811d996 | 850 | |
59449359 | 851 | if (!tick_nohz_full_cpu(cpu)) |
e9a2eb40 | 852 | return; |
5811d996 | 853 | |
e9a2eb40 AS |
854 | if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE) |
855 | return; | |
5811d996 | 856 | |
57ccdf44 | 857 | if (can_stop_full_tick(cpu, ts)) |
73738a95 FW |
858 | tick_nohz_stop_sched_tick(ts, ktime_get(), cpu); |
859 | else if (ts->tick_stopped) | |
1f41906a | 860 | tick_nohz_restart_sched_tick(ts, ktime_get()); |
5811d996 FW |
861 | #endif |
862 | } | |
863 | ||
5b39939a FW |
864 | static bool can_stop_idle_tick(int cpu, struct tick_sched *ts) |
865 | { | |
866 | /* | |
0de7611a | 867 | * If this CPU is offline and it is the one which updates |
5b39939a | 868 | * jiffies, then give up the assignment and let it be taken by |
0de7611a | 869 | * the CPU which runs the tick timer next. If we don't drop |
5b39939a FW |
870 | * this here the jiffies might be stale and do_timer() never |
871 | * invoked. | |
872 | */ | |
873 | if (unlikely(!cpu_online(cpu))) { | |
874 | if (cpu == tick_do_timer_cpu) | |
875 | tick_do_timer_cpu = TICK_DO_TIMER_NONE; | |
411fe24e FW |
876 | /* |
877 | * Make sure the CPU doesn't get fooled by obsolete tick | |
878 | * deadline if it comes back online later. | |
879 | */ | |
880 | ts->next_tick = 0; | |
f7ea0fd6 | 881 | return false; |
5b39939a FW |
882 | } |
883 | ||
0e576acb | 884 | if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) { |
2456e855 | 885 | ts->sleep_length = NSEC_PER_SEC / HZ; |
5b39939a | 886 | return false; |
0e576acb | 887 | } |
5b39939a FW |
888 | |
889 | if (need_resched()) | |
890 | return false; | |
891 | ||
892 | if (unlikely(local_softirq_pending() && cpu_online(cpu))) { | |
893 | static int ratelimit; | |
894 | ||
803b0eba PM |
895 | if (ratelimit < 10 && |
896 | (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) { | |
cfea7d7e RV |
897 | pr_warn("NOHZ: local_softirq_pending %02x\n", |
898 | (unsigned int) local_softirq_pending()); | |
5b39939a FW |
899 | ratelimit++; |
900 | } | |
901 | return false; | |
902 | } | |
903 | ||
460775df | 904 | if (tick_nohz_full_enabled()) { |
a382bf93 FW |
905 | /* |
906 | * Keep the tick alive to guarantee timekeeping progression | |
907 | * if there are full dynticks CPUs around | |
908 | */ | |
909 | if (tick_do_timer_cpu == cpu) | |
910 | return false; | |
911 | /* | |
912 | * Boot safety: make sure the timekeeping duty has been | |
913 | * assigned before entering dyntick-idle mode, | |
914 | */ | |
915 | if (tick_do_timer_cpu == TICK_DO_TIMER_NONE) | |
916 | return false; | |
917 | } | |
918 | ||
5b39939a FW |
919 | return true; |
920 | } | |
921 | ||
19f5f736 FW |
922 | static void __tick_nohz_idle_enter(struct tick_sched *ts) |
923 | { | |
84bf1bcc | 924 | ktime_t now, expires; |
5b39939a | 925 | int cpu = smp_processor_id(); |
19f5f736 | 926 | |
08d07259 WL |
927 | now = tick_nohz_start_idle(ts); |
928 | ||
5b39939a FW |
929 | if (can_stop_idle_tick(cpu, ts)) { |
930 | int was_stopped = ts->tick_stopped; | |
931 | ||
932 | ts->idle_calls++; | |
84bf1bcc FW |
933 | |
934 | expires = tick_nohz_stop_sched_tick(ts, now, cpu); | |
2456e855 | 935 | if (expires > 0LL) { |
84bf1bcc FW |
936 | ts->idle_sleeps++; |
937 | ts->idle_expires = expires; | |
938 | } | |
5b39939a | 939 | |
a0db971e | 940 | if (!was_stopped && ts->tick_stopped) { |
5b39939a | 941 | ts->idle_jiffies = ts->last_jiffies; |
a0db971e FW |
942 | nohz_balance_enter_idle(cpu); |
943 | } | |
5b39939a | 944 | } |
280f0677 FW |
945 | } |
946 | ||
947 | /** | |
948 | * tick_nohz_idle_enter - stop the idle tick from the idle task | |
949 | * | |
950 | * When the next event is more than a tick into the future, stop the idle tick | |
951 | * Called when we start the idle loop. | |
2bbb6817 | 952 | * |
1268fbc7 | 953 | * The arch is responsible of calling: |
2bbb6817 FW |
954 | * |
955 | * - rcu_idle_enter() after its last use of RCU before the CPU is put | |
956 | * to sleep. | |
957 | * - rcu_idle_exit() before the first use of RCU after the CPU is woken up. | |
280f0677 | 958 | */ |
1268fbc7 | 959 | void tick_nohz_idle_enter(void) |
280f0677 FW |
960 | { |
961 | struct tick_sched *ts; | |
962 | ||
1268fbc7 FW |
963 | WARN_ON_ONCE(irqs_disabled()); |
964 | ||
0db49b72 | 965 | /* |
0de7611a IM |
966 | * Update the idle state in the scheduler domain hierarchy |
967 | * when tick_nohz_stop_sched_tick() is called from the idle loop. | |
968 | * State will be updated to busy during the first busy tick after | |
969 | * exiting idle. | |
970 | */ | |
0db49b72 LT |
971 | set_cpu_sd_state_idle(); |
972 | ||
1268fbc7 FW |
973 | local_irq_disable(); |
974 | ||
22127e93 | 975 | ts = this_cpu_ptr(&tick_cpu_sched); |
280f0677 | 976 | ts->inidle = 1; |
19f5f736 | 977 | __tick_nohz_idle_enter(ts); |
1268fbc7 FW |
978 | |
979 | local_irq_enable(); | |
280f0677 FW |
980 | } |
981 | ||
982 | /** | |
983 | * tick_nohz_irq_exit - update next tick event from interrupt exit | |
984 | * | |
985 | * When an interrupt fires while we are idle and it doesn't cause | |
986 | * a reschedule, it may still add, modify or delete a timer, enqueue | |
987 | * an RCU callback, etc... | |
988 | * So we need to re-calculate and reprogram the next tick event. | |
989 | */ | |
990 | void tick_nohz_irq_exit(void) | |
991 | { | |
22127e93 | 992 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
280f0677 | 993 | |
14851912 | 994 | if (ts->inidle) |
5811d996 | 995 | __tick_nohz_idle_enter(ts); |
14851912 | 996 | else |
73738a95 | 997 | tick_nohz_full_update_tick(ts); |
79bf2bb3 TG |
998 | } |
999 | ||
4f86d3a8 LB |
1000 | /** |
1001 | * tick_nohz_get_sleep_length - return the length of the current sleep | |
1002 | * | |
1003 | * Called from power state control code with interrupts disabled | |
1004 | */ | |
1005 | ktime_t tick_nohz_get_sleep_length(void) | |
1006 | { | |
22127e93 | 1007 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
4f86d3a8 LB |
1008 | |
1009 | return ts->sleep_length; | |
1010 | } | |
1011 | ||
b7eaf1aa RW |
1012 | /** |
1013 | * tick_nohz_get_idle_calls - return the current idle calls counter value | |
1014 | * | |
1015 | * Called from the schedutil frequency scaling governor in scheduler context. | |
1016 | */ | |
1017 | unsigned long tick_nohz_get_idle_calls(void) | |
1018 | { | |
1019 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); | |
1020 | ||
1021 | return ts->idle_calls; | |
1022 | } | |
1023 | ||
2ac0d98f FW |
1024 | static void tick_nohz_account_idle_ticks(struct tick_sched *ts) |
1025 | { | |
3f4724ea | 1026 | #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE |
2ac0d98f | 1027 | unsigned long ticks; |
3f4724ea | 1028 | |
55dbdcfa | 1029 | if (vtime_accounting_cpu_enabled()) |
3f4724ea | 1030 | return; |
79bf2bb3 TG |
1031 | /* |
1032 | * We stopped the tick in idle. Update process times would miss the | |
1033 | * time we slept as update_process_times does only a 1 tick | |
1034 | * accounting. Enforce that this is accounted to idle ! | |
1035 | */ | |
1036 | ticks = jiffies - ts->idle_jiffies; | |
1037 | /* | |
1038 | * We might be one off. Do not randomly account a huge number of ticks! | |
1039 | */ | |
79741dd3 MS |
1040 | if (ticks && ticks < LONG_MAX) |
1041 | account_idle_ticks(ticks); | |
1042 | #endif | |
19f5f736 FW |
1043 | } |
1044 | ||
79bf2bb3 | 1045 | /** |
280f0677 | 1046 | * tick_nohz_idle_exit - restart the idle tick from the idle task |
79bf2bb3 TG |
1047 | * |
1048 | * Restart the idle tick when the CPU is woken up from idle | |
280f0677 FW |
1049 | * This also exit the RCU extended quiescent state. The CPU |
1050 | * can use RCU again after this function is called. | |
79bf2bb3 | 1051 | */ |
280f0677 | 1052 | void tick_nohz_idle_exit(void) |
79bf2bb3 | 1053 | { |
4a32fea9 | 1054 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
6378ddb5 | 1055 | ktime_t now; |
79bf2bb3 | 1056 | |
6378ddb5 | 1057 | local_irq_disable(); |
2bbb6817 | 1058 | |
15f827be FW |
1059 | WARN_ON_ONCE(!ts->inidle); |
1060 | ||
1061 | ts->inidle = 0; | |
1062 | ||
1063 | if (ts->idle_active || ts->tick_stopped) | |
eed3b9cf MS |
1064 | now = ktime_get(); |
1065 | ||
1066 | if (ts->idle_active) | |
e8fcaa5c | 1067 | tick_nohz_stop_idle(ts, now); |
6378ddb5 | 1068 | |
2ac0d98f | 1069 | if (ts->tick_stopped) { |
1f41906a | 1070 | tick_nohz_restart_sched_tick(ts, now); |
2ac0d98f | 1071 | tick_nohz_account_idle_ticks(ts); |
6378ddb5 | 1072 | } |
79bf2bb3 | 1073 | |
79bf2bb3 TG |
1074 | local_irq_enable(); |
1075 | } | |
1076 | ||
79bf2bb3 TG |
1077 | /* |
1078 | * The nohz low res interrupt handler | |
1079 | */ | |
1080 | static void tick_nohz_handler(struct clock_event_device *dev) | |
1081 | { | |
22127e93 | 1082 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
79bf2bb3 TG |
1083 | struct pt_regs *regs = get_irq_regs(); |
1084 | ktime_t now = ktime_get(); | |
1085 | ||
2456e855 | 1086 | dev->next_event = KTIME_MAX; |
79bf2bb3 | 1087 | |
5bb96226 | 1088 | tick_sched_do_timer(now); |
9e8f559b | 1089 | tick_sched_handle(ts, regs); |
79bf2bb3 | 1090 | |
b5e995e6 VK |
1091 | /* No need to reprogram if we are running tickless */ |
1092 | if (unlikely(ts->tick_stopped)) | |
1093 | return; | |
1094 | ||
0ff53d09 TG |
1095 | hrtimer_forward(&ts->sched_timer, now, tick_period); |
1096 | tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); | |
79bf2bb3 TG |
1097 | } |
1098 | ||
bc7a34b8 TG |
1099 | static inline void tick_nohz_activate(struct tick_sched *ts, int mode) |
1100 | { | |
1101 | if (!tick_nohz_enabled) | |
1102 | return; | |
1103 | ts->nohz_mode = mode; | |
1104 | /* One update is enough */ | |
1105 | if (!test_and_set_bit(0, &tick_nohz_active)) | |
683be13a | 1106 | timers_update_migration(true); |
bc7a34b8 TG |
1107 | } |
1108 | ||
79bf2bb3 TG |
1109 | /** |
1110 | * tick_nohz_switch_to_nohz - switch to nohz mode | |
1111 | */ | |
1112 | static void tick_nohz_switch_to_nohz(void) | |
1113 | { | |
22127e93 | 1114 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
79bf2bb3 TG |
1115 | ktime_t next; |
1116 | ||
27630532 | 1117 | if (!tick_nohz_enabled) |
79bf2bb3 TG |
1118 | return; |
1119 | ||
6b442bc8 | 1120 | if (tick_switch_to_oneshot(tick_nohz_handler)) |
79bf2bb3 | 1121 | return; |
6b442bc8 | 1122 | |
79bf2bb3 TG |
1123 | /* |
1124 | * Recycle the hrtimer in ts, so we can share the | |
1125 | * hrtimer_forward with the highres code. | |
1126 | */ | |
1127 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); | |
1128 | /* Get the next period */ | |
1129 | next = tick_init_jiffy_update(); | |
1130 | ||
0ff53d09 | 1131 | hrtimer_set_expires(&ts->sched_timer, next); |
1ca8ec53 WL |
1132 | hrtimer_forward_now(&ts->sched_timer, tick_period); |
1133 | tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); | |
bc7a34b8 | 1134 | tick_nohz_activate(ts, NOHZ_MODE_LOWRES); |
79bf2bb3 TG |
1135 | } |
1136 | ||
5acac1be | 1137 | static inline void tick_nohz_irq_enter(void) |
eed3b9cf | 1138 | { |
4a32fea9 | 1139 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
eed3b9cf MS |
1140 | ktime_t now; |
1141 | ||
1142 | if (!ts->idle_active && !ts->tick_stopped) | |
1143 | return; | |
1144 | now = ktime_get(); | |
1145 | if (ts->idle_active) | |
e8fcaa5c | 1146 | tick_nohz_stop_idle(ts, now); |
ff006732 | 1147 | if (ts->tick_stopped) |
eed3b9cf | 1148 | tick_nohz_update_jiffies(now); |
eed3b9cf MS |
1149 | } |
1150 | ||
79bf2bb3 TG |
1151 | #else |
1152 | ||
1153 | static inline void tick_nohz_switch_to_nohz(void) { } | |
5acac1be | 1154 | static inline void tick_nohz_irq_enter(void) { } |
bc7a34b8 | 1155 | static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { } |
79bf2bb3 | 1156 | |
3451d024 | 1157 | #endif /* CONFIG_NO_HZ_COMMON */ |
79bf2bb3 | 1158 | |
719254fa TG |
1159 | /* |
1160 | * Called from irq_enter to notify about the possible interruption of idle() | |
1161 | */ | |
5acac1be | 1162 | void tick_irq_enter(void) |
719254fa | 1163 | { |
e8fcaa5c | 1164 | tick_check_oneshot_broadcast_this_cpu(); |
5acac1be | 1165 | tick_nohz_irq_enter(); |
719254fa TG |
1166 | } |
1167 | ||
79bf2bb3 TG |
1168 | /* |
1169 | * High resolution timer specific code | |
1170 | */ | |
1171 | #ifdef CONFIG_HIGH_RES_TIMERS | |
1172 | /* | |
4c9dc641 | 1173 | * We rearm the timer until we get disabled by the idle code. |
351f181f | 1174 | * Called with interrupts disabled. |
79bf2bb3 TG |
1175 | */ |
1176 | static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) | |
1177 | { | |
1178 | struct tick_sched *ts = | |
1179 | container_of(timer, struct tick_sched, sched_timer); | |
79bf2bb3 TG |
1180 | struct pt_regs *regs = get_irq_regs(); |
1181 | ktime_t now = ktime_get(); | |
d3ed7824 | 1182 | |
5bb96226 | 1183 | tick_sched_do_timer(now); |
79bf2bb3 TG |
1184 | |
1185 | /* | |
1186 | * Do not call, when we are not in irq context and have | |
1187 | * no valid regs pointer | |
1188 | */ | |
9e8f559b FW |
1189 | if (regs) |
1190 | tick_sched_handle(ts, regs); | |
7c259045 FW |
1191 | else |
1192 | ts->next_tick = 0; | |
79bf2bb3 | 1193 | |
2a16fc93 VK |
1194 | /* No need to reprogram if we are in idle or full dynticks mode */ |
1195 | if (unlikely(ts->tick_stopped)) | |
1196 | return HRTIMER_NORESTART; | |
1197 | ||
79bf2bb3 TG |
1198 | hrtimer_forward(timer, now, tick_period); |
1199 | ||
1200 | return HRTIMER_RESTART; | |
1201 | } | |
1202 | ||
5307c955 MG |
1203 | static int sched_skew_tick; |
1204 | ||
62cf20b3 TG |
1205 | static int __init skew_tick(char *str) |
1206 | { | |
1207 | get_option(&str, &sched_skew_tick); | |
1208 | ||
1209 | return 0; | |
1210 | } | |
1211 | early_param("skew_tick", skew_tick); | |
1212 | ||
79bf2bb3 TG |
1213 | /** |
1214 | * tick_setup_sched_timer - setup the tick emulation timer | |
1215 | */ | |
1216 | void tick_setup_sched_timer(void) | |
1217 | { | |
22127e93 | 1218 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
79bf2bb3 TG |
1219 | ktime_t now = ktime_get(); |
1220 | ||
1221 | /* | |
1222 | * Emulate tick processing via per-CPU hrtimers: | |
1223 | */ | |
1224 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); | |
1225 | ts->sched_timer.function = tick_sched_timer; | |
79bf2bb3 | 1226 | |
0de7611a | 1227 | /* Get the next period (per-CPU) */ |
cc584b21 | 1228 | hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); |
79bf2bb3 | 1229 | |
9c3f9e28 | 1230 | /* Offset the tick to avert jiffies_lock contention. */ |
5307c955 MG |
1231 | if (sched_skew_tick) { |
1232 | u64 offset = ktime_to_ns(tick_period) >> 1; | |
1233 | do_div(offset, num_possible_cpus()); | |
1234 | offset *= smp_processor_id(); | |
1235 | hrtimer_add_expires_ns(&ts->sched_timer, offset); | |
1236 | } | |
1237 | ||
afc08b15 TG |
1238 | hrtimer_forward(&ts->sched_timer, now, tick_period); |
1239 | hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); | |
bc7a34b8 | 1240 | tick_nohz_activate(ts, NOHZ_MODE_HIGHRES); |
79bf2bb3 | 1241 | } |
3c4fbe5e | 1242 | #endif /* HIGH_RES_TIMERS */ |
79bf2bb3 | 1243 | |
3451d024 | 1244 | #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS |
79bf2bb3 TG |
1245 | void tick_cancel_sched_timer(int cpu) |
1246 | { | |
1247 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); | |
1248 | ||
3c4fbe5e | 1249 | # ifdef CONFIG_HIGH_RES_TIMERS |
79bf2bb3 TG |
1250 | if (ts->sched_timer.base) |
1251 | hrtimer_cancel(&ts->sched_timer); | |
3c4fbe5e | 1252 | # endif |
a7901766 | 1253 | |
4b0c0f29 | 1254 | memset(ts, 0, sizeof(*ts)); |
79bf2bb3 | 1255 | } |
3c4fbe5e | 1256 | #endif |
79bf2bb3 TG |
1257 | |
1258 | /** | |
1259 | * Async notification about clocksource changes | |
1260 | */ | |
1261 | void tick_clock_notify(void) | |
1262 | { | |
1263 | int cpu; | |
1264 | ||
1265 | for_each_possible_cpu(cpu) | |
1266 | set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); | |
1267 | } | |
1268 | ||
1269 | /* | |
1270 | * Async notification about clock event changes | |
1271 | */ | |
1272 | void tick_oneshot_notify(void) | |
1273 | { | |
22127e93 | 1274 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
79bf2bb3 TG |
1275 | |
1276 | set_bit(0, &ts->check_clocks); | |
1277 | } | |
1278 | ||
1279 | /** | |
1280 | * Check, if a change happened, which makes oneshot possible. | |
1281 | * | |
1282 | * Called cyclic from the hrtimer softirq (driven by the timer | |
1283 | * softirq) allow_nohz signals, that we can switch into low-res nohz | |
1284 | * mode, because high resolution timers are disabled (either compile | |
6b442bc8 | 1285 | * or runtime). Called with interrupts disabled. |
79bf2bb3 TG |
1286 | */ |
1287 | int tick_check_oneshot_change(int allow_nohz) | |
1288 | { | |
22127e93 | 1289 | struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); |
79bf2bb3 TG |
1290 | |
1291 | if (!test_and_clear_bit(0, &ts->check_clocks)) | |
1292 | return 0; | |
1293 | ||
1294 | if (ts->nohz_mode != NOHZ_MODE_INACTIVE) | |
1295 | return 0; | |
1296 | ||
cf4fc6cb | 1297 | if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) |
79bf2bb3 TG |
1298 | return 0; |
1299 | ||
1300 | if (!allow_nohz) | |
1301 | return 1; | |
1302 | ||
1303 | tick_nohz_switch_to_nohz(); | |
1304 | return 0; | |
1305 | } |