Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * sched_clock() for unstable CPU clocks
4 *
5 * Copyright (C) 2008 Red Hat, Inc., Peter Zijlstra
6 *
7 * Updates and enhancements:
8 * Copyright (C) 2008 Red Hat, Inc. Steven Rostedt <srostedt@redhat.com>
9 *
10 * Based on code by:
11 * Ingo Molnar <mingo@redhat.com>
12 * Guillaume Chazarain <guichaz@gmail.com>
13 *
14 *
15 * What this file implements:
16 *
17 * cpu_clock(i) provides a fast (execution time) high resolution
18 * clock with bounded drift between CPUs. The value of cpu_clock(i)
19 * is monotonic for constant i. The timestamp returned is in nanoseconds.
20 *
21 * ######################### BIG FAT WARNING ##########################
22 * # when comparing cpu_clock(i) to cpu_clock(j) for i != j, time can #
23 * # go backwards !! #
24 * ####################################################################
25 *
26 * There is no strict promise about the base, although it tends to start
27 * at 0 on boot (but people really shouldn't rely on that).
28 *
29 * cpu_clock(i) -- can be used from any context, including NMI.
30 * local_clock() -- is cpu_clock() on the current CPU.
31 *
32 * sched_clock_cpu(i)
33 *
34 * How it is implemented:
35 *
36 * The implementation either uses sched_clock() when
37 * !CONFIG_HAVE_UNSTABLE_SCHED_CLOCK, which means in that case the
38 * sched_clock() is assumed to provide these properties (mostly it means
39 * the architecture provides a globally synchronized highres time source).
40 *
41 * Otherwise it tries to create a semi stable clock from a mixture of other
42 * clocks, including:
43 *
44 * - GTOD (clock monotonic)
45 * - sched_clock()
46 * - explicit idle events
47 *
48 * We use GTOD as base and use sched_clock() deltas to improve resolution. The
49 * deltas are filtered to provide monotonicity and keeping it within an
50 * expected window.
51 *
52 * Furthermore, explicit sleep and wakeup hooks allow us to account for time
53 * that is otherwise invisible (TSC gets stopped).
54 *
55 */
56
57/*
58 * Scheduler clock - returns current time in nanosec units.
59 * This is default implementation.
60 * Architectures and sub-architectures can override this.
61 */
62notrace unsigned long long __weak sched_clock(void)
63{
64 return (unsigned long long)(jiffies - INITIAL_JIFFIES)
65 * (NSEC_PER_SEC / HZ);
66}
67EXPORT_SYMBOL_GPL(sched_clock);
68
69static DEFINE_STATIC_KEY_FALSE(sched_clock_running);
70
71#ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
72/*
73 * We must start with !__sched_clock_stable because the unstable -> stable
74 * transition is accurate, while the stable -> unstable transition is not.
75 *
76 * Similarly we start with __sched_clock_stable_early, thereby assuming we
77 * will become stable, such that there's only a single 1 -> 0 transition.
78 */
79static DEFINE_STATIC_KEY_FALSE(__sched_clock_stable);
80static int __sched_clock_stable_early = 1;
81
82/*
83 * We want: ktime_get_ns() + __gtod_offset == sched_clock() + __sched_clock_offset
84 */
85__read_mostly u64 __sched_clock_offset;
86static __read_mostly u64 __gtod_offset;
87
88struct sched_clock_data {
89 u64 tick_raw;
90 u64 tick_gtod;
91 u64 clock;
92};
93
94static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
95
96notrace static inline struct sched_clock_data *this_scd(void)
97{
98 return this_cpu_ptr(&sched_clock_data);
99}
100
101notrace static inline struct sched_clock_data *cpu_sdc(int cpu)
102{
103 return &per_cpu(sched_clock_data, cpu);
104}
105
106notrace int sched_clock_stable(void)
107{
108 return static_branch_likely(&__sched_clock_stable);
109}
110
111notrace static void __scd_stamp(struct sched_clock_data *scd)
112{
113 scd->tick_gtod = ktime_get_ns();
114 scd->tick_raw = sched_clock();
115}
116
117notrace static void __set_sched_clock_stable(void)
118{
119 struct sched_clock_data *scd;
120
121 /*
122 * Since we're still unstable and the tick is already running, we have
123 * to disable IRQs in order to get a consistent scd->tick* reading.
124 */
125 local_irq_disable();
126 scd = this_scd();
127 /*
128 * Attempt to make the (initial) unstable->stable transition continuous.
129 */
130 __sched_clock_offset = (scd->tick_gtod + __gtod_offset) - (scd->tick_raw);
131 local_irq_enable();
132
133 printk(KERN_INFO "sched_clock: Marking stable (%lld, %lld)->(%lld, %lld)\n",
134 scd->tick_gtod, __gtod_offset,
135 scd->tick_raw, __sched_clock_offset);
136
137 static_branch_enable(&__sched_clock_stable);
138 tick_dep_clear(TICK_DEP_BIT_CLOCK_UNSTABLE);
139}
140
141/*
142 * If we ever get here, we're screwed, because we found out -- typically after
143 * the fact -- that TSC wasn't good. This means all our clocksources (including
144 * ktime) could have reported wrong values.
145 *
146 * What we do here is an attempt to fix up and continue sort of where we left
147 * off in a coherent manner.
148 *
149 * The only way to fully avoid random clock jumps is to boot with:
150 * "tsc=unstable".
151 */
152notrace static void __sched_clock_work(struct work_struct *work)
153{
154 struct sched_clock_data *scd;
155 int cpu;
156
157 /* take a current timestamp and set 'now' */
158 preempt_disable();
159 scd = this_scd();
160 __scd_stamp(scd);
161 scd->clock = scd->tick_gtod + __gtod_offset;
162 preempt_enable();
163
164 /* clone to all CPUs */
165 for_each_possible_cpu(cpu)
166 per_cpu(sched_clock_data, cpu) = *scd;
167
168 printk(KERN_WARNING "TSC found unstable after boot, most likely due to broken BIOS. Use 'tsc=unstable'.\n");
169 printk(KERN_INFO "sched_clock: Marking unstable (%lld, %lld)<-(%lld, %lld)\n",
170 scd->tick_gtod, __gtod_offset,
171 scd->tick_raw, __sched_clock_offset);
172
173 static_branch_disable(&__sched_clock_stable);
174}
175
176static DECLARE_WORK(sched_clock_work, __sched_clock_work);
177
178notrace static void __clear_sched_clock_stable(void)
179{
180 if (!sched_clock_stable())
181 return;
182
183 tick_dep_set(TICK_DEP_BIT_CLOCK_UNSTABLE);
184 schedule_work(&sched_clock_work);
185}
186
187notrace void clear_sched_clock_stable(void)
188{
189 __sched_clock_stable_early = 0;
190
191 smp_mb(); /* matches sched_clock_init_late() */
192
193 if (static_key_count(&sched_clock_running.key) == 2)
194 __clear_sched_clock_stable();
195}
196
197notrace static void __sched_clock_gtod_offset(void)
198{
199 struct sched_clock_data *scd = this_scd();
200
201 __scd_stamp(scd);
202 __gtod_offset = (scd->tick_raw + __sched_clock_offset) - scd->tick_gtod;
203}
204
205void __init sched_clock_init(void)
206{
207 /*
208 * Set __gtod_offset such that once we mark sched_clock_running,
209 * sched_clock_tick() continues where sched_clock() left off.
210 *
211 * Even if TSC is buggered, we're still UP at this point so it
212 * can't really be out of sync.
213 */
214 local_irq_disable();
215 __sched_clock_gtod_offset();
216 local_irq_enable();
217
218 static_branch_inc(&sched_clock_running);
219}
220/*
221 * We run this as late_initcall() such that it runs after all built-in drivers,
222 * notably: acpi_processor and intel_idle, which can mark the TSC as unstable.
223 */
224static int __init sched_clock_init_late(void)
225{
226 static_branch_inc(&sched_clock_running);
227 /*
228 * Ensure that it is impossible to not do a static_key update.
229 *
230 * Either {set,clear}_sched_clock_stable() must see sched_clock_running
231 * and do the update, or we must see their __sched_clock_stable_early
232 * and do the update, or both.
233 */
234 smp_mb(); /* matches {set,clear}_sched_clock_stable() */
235
236 if (__sched_clock_stable_early)
237 __set_sched_clock_stable();
238
239 return 0;
240}
241late_initcall(sched_clock_init_late);
242
243/*
244 * min, max except they take wrapping into account
245 */
246
247notrace static inline u64 wrap_min(u64 x, u64 y)
248{
249 return (s64)(x - y) < 0 ? x : y;
250}
251
252notrace static inline u64 wrap_max(u64 x, u64 y)
253{
254 return (s64)(x - y) > 0 ? x : y;
255}
256
257/*
258 * update the percpu scd from the raw @now value
259 *
260 * - filter out backward motion
261 * - use the GTOD tick value to create a window to filter crazy TSC values
262 */
263notrace static u64 sched_clock_local(struct sched_clock_data *scd)
264{
265 u64 now, clock, old_clock, min_clock, max_clock, gtod;
266 s64 delta;
267
268again:
269 now = sched_clock();
270 delta = now - scd->tick_raw;
271 if (unlikely(delta < 0))
272 delta = 0;
273
274 old_clock = scd->clock;
275
276 /*
277 * scd->clock = clamp(scd->tick_gtod + delta,
278 * max(scd->tick_gtod, scd->clock),
279 * scd->tick_gtod + TICK_NSEC);
280 */
281
282 gtod = scd->tick_gtod + __gtod_offset;
283 clock = gtod + delta;
284 min_clock = wrap_max(gtod, old_clock);
285 max_clock = wrap_max(old_clock, gtod + TICK_NSEC);
286
287 clock = wrap_max(clock, min_clock);
288 clock = wrap_min(clock, max_clock);
289
290 if (!try_cmpxchg64(&scd->clock, &old_clock, clock))
291 goto again;
292
293 return clock;
294}
295
296notrace static u64 sched_clock_remote(struct sched_clock_data *scd)
297{
298 struct sched_clock_data *my_scd = this_scd();
299 u64 this_clock, remote_clock;
300 u64 *ptr, old_val, val;
301
302#if BITS_PER_LONG != 64
303again:
304 /*
305 * Careful here: The local and the remote clock values need to
306 * be read out atomic as we need to compare the values and
307 * then update either the local or the remote side. So the
308 * cmpxchg64 below only protects one readout.
309 *
310 * We must reread via sched_clock_local() in the retry case on
311 * 32-bit kernels as an NMI could use sched_clock_local() via the
312 * tracer and hit between the readout of
313 * the low 32-bit and the high 32-bit portion.
314 */
315 this_clock = sched_clock_local(my_scd);
316 /*
317 * We must enforce atomic readout on 32-bit, otherwise the
318 * update on the remote CPU can hit inbetween the readout of
319 * the low 32-bit and the high 32-bit portion.
320 */
321 remote_clock = cmpxchg64(&scd->clock, 0, 0);
322#else
323 /*
324 * On 64-bit kernels the read of [my]scd->clock is atomic versus the
325 * update, so we can avoid the above 32-bit dance.
326 */
327 sched_clock_local(my_scd);
328again:
329 this_clock = my_scd->clock;
330 remote_clock = scd->clock;
331#endif
332
333 /*
334 * Use the opportunity that we have both locks
335 * taken to couple the two clocks: we take the
336 * larger time as the latest time for both
337 * runqueues. (this creates monotonic movement)
338 */
339 if (likely((s64)(remote_clock - this_clock) < 0)) {
340 ptr = &scd->clock;
341 old_val = remote_clock;
342 val = this_clock;
343 } else {
344 /*
345 * Should be rare, but possible:
346 */
347 ptr = &my_scd->clock;
348 old_val = this_clock;
349 val = remote_clock;
350 }
351
352 if (!try_cmpxchg64(ptr, &old_val, val))
353 goto again;
354
355 return val;
356}
357
358/*
359 * Similar to cpu_clock(), but requires local IRQs to be disabled.
360 *
361 * See cpu_clock().
362 */
363notrace u64 sched_clock_cpu(int cpu)
364{
365 struct sched_clock_data *scd;
366 u64 clock;
367
368 if (sched_clock_stable())
369 return sched_clock() + __sched_clock_offset;
370
371 if (!static_branch_likely(&sched_clock_running))
372 return sched_clock();
373
374 preempt_disable_notrace();
375 scd = cpu_sdc(cpu);
376
377 if (cpu != smp_processor_id())
378 clock = sched_clock_remote(scd);
379 else
380 clock = sched_clock_local(scd);
381 preempt_enable_notrace();
382
383 return clock;
384}
385EXPORT_SYMBOL_GPL(sched_clock_cpu);
386
387notrace void sched_clock_tick(void)
388{
389 struct sched_clock_data *scd;
390
391 if (sched_clock_stable())
392 return;
393
394 if (!static_branch_likely(&sched_clock_running))
395 return;
396
397 lockdep_assert_irqs_disabled();
398
399 scd = this_scd();
400 __scd_stamp(scd);
401 sched_clock_local(scd);
402}
403
404notrace void sched_clock_tick_stable(void)
405{
406 if (!sched_clock_stable())
407 return;
408
409 /*
410 * Called under watchdog_lock.
411 *
412 * The watchdog just found this TSC to (still) be stable, so now is a
413 * good moment to update our __gtod_offset. Because once we find the
414 * TSC to be unstable, any computation will be computing crap.
415 */
416 local_irq_disable();
417 __sched_clock_gtod_offset();
418 local_irq_enable();
419}
420
421/*
422 * We are going deep-idle (irqs are disabled):
423 */
424notrace void sched_clock_idle_sleep_event(void)
425{
426 sched_clock_cpu(smp_processor_id());
427}
428EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
429
430/*
431 * We just idled; resync with ktime.
432 */
433notrace void sched_clock_idle_wakeup_event(void)
434{
435 unsigned long flags;
436
437 if (sched_clock_stable())
438 return;
439
440 if (unlikely(timekeeping_suspended))
441 return;
442
443 local_irq_save(flags);
444 sched_clock_tick();
445 local_irq_restore(flags);
446}
447EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
448
449#else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
450
451void __init sched_clock_init(void)
452{
453 static_branch_inc(&sched_clock_running);
454 local_irq_disable();
455 generic_sched_clock_init();
456 local_irq_enable();
457}
458
459notrace u64 sched_clock_cpu(int cpu)
460{
461 if (!static_branch_likely(&sched_clock_running))
462 return 0;
463
464 return sched_clock();
465}
466
467#endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
468
469/*
470 * Running clock - returns the time that has elapsed while a guest has been
471 * running.
472 * On a guest this value should be local_clock minus the time the guest was
473 * suspended by the hypervisor (for any reason).
474 * On bare metal this function should return the same as local_clock.
475 * Architectures and sub-architectures can override this.
476 */
477notrace u64 __weak running_clock(void)
478{
479 return local_clock();
480}
1/*
2 * sched_clock for unstable cpu clocks
3 *
4 * Copyright (C) 2008 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
5 *
6 * Updates and enhancements:
7 * Copyright (C) 2008 Red Hat, Inc. Steven Rostedt <srostedt@redhat.com>
8 *
9 * Based on code by:
10 * Ingo Molnar <mingo@redhat.com>
11 * Guillaume Chazarain <guichaz@gmail.com>
12 *
13 *
14 * What:
15 *
16 * cpu_clock(i) provides a fast (execution time) high resolution
17 * clock with bounded drift between CPUs. The value of cpu_clock(i)
18 * is monotonic for constant i. The timestamp returned is in nanoseconds.
19 *
20 * ######################### BIG FAT WARNING ##########################
21 * # when comparing cpu_clock(i) to cpu_clock(j) for i != j, time can #
22 * # go backwards !! #
23 * ####################################################################
24 *
25 * There is no strict promise about the base, although it tends to start
26 * at 0 on boot (but people really shouldn't rely on that).
27 *
28 * cpu_clock(i) -- can be used from any context, including NMI.
29 * local_clock() -- is cpu_clock() on the current cpu.
30 *
31 * sched_clock_cpu(i)
32 *
33 * How:
34 *
35 * The implementation either uses sched_clock() when
36 * !CONFIG_HAVE_UNSTABLE_SCHED_CLOCK, which means in that case the
37 * sched_clock() is assumed to provide these properties (mostly it means
38 * the architecture provides a globally synchronized highres time source).
39 *
40 * Otherwise it tries to create a semi stable clock from a mixture of other
41 * clocks, including:
42 *
43 * - GTOD (clock monotomic)
44 * - sched_clock()
45 * - explicit idle events
46 *
47 * We use GTOD as base and use sched_clock() deltas to improve resolution. The
48 * deltas are filtered to provide monotonicity and keeping it within an
49 * expected window.
50 *
51 * Furthermore, explicit sleep and wakeup hooks allow us to account for time
52 * that is otherwise invisible (TSC gets stopped).
53 *
54 */
55#include <linux/spinlock.h>
56#include <linux/hardirq.h>
57#include <linux/export.h>
58#include <linux/percpu.h>
59#include <linux/ktime.h>
60#include <linux/sched.h>
61#include <linux/static_key.h>
62#include <linux/workqueue.h>
63#include <linux/compiler.h>
64
65/*
66 * Scheduler clock - returns current time in nanosec units.
67 * This is default implementation.
68 * Architectures and sub-architectures can override this.
69 */
70unsigned long long __weak sched_clock(void)
71{
72 return (unsigned long long)(jiffies - INITIAL_JIFFIES)
73 * (NSEC_PER_SEC / HZ);
74}
75EXPORT_SYMBOL_GPL(sched_clock);
76
77__read_mostly int sched_clock_running;
78
79#ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
80static struct static_key __sched_clock_stable = STATIC_KEY_INIT;
81static int __sched_clock_stable_early;
82
83int sched_clock_stable(void)
84{
85 return static_key_false(&__sched_clock_stable);
86}
87
88static void __set_sched_clock_stable(void)
89{
90 if (!sched_clock_stable())
91 static_key_slow_inc(&__sched_clock_stable);
92}
93
94void set_sched_clock_stable(void)
95{
96 __sched_clock_stable_early = 1;
97
98 smp_mb(); /* matches sched_clock_init() */
99
100 if (!sched_clock_running)
101 return;
102
103 __set_sched_clock_stable();
104}
105
106static void __clear_sched_clock_stable(struct work_struct *work)
107{
108 /* XXX worry about clock continuity */
109 if (sched_clock_stable())
110 static_key_slow_dec(&__sched_clock_stable);
111}
112
113static DECLARE_WORK(sched_clock_work, __clear_sched_clock_stable);
114
115void clear_sched_clock_stable(void)
116{
117 __sched_clock_stable_early = 0;
118
119 smp_mb(); /* matches sched_clock_init() */
120
121 if (!sched_clock_running)
122 return;
123
124 schedule_work(&sched_clock_work);
125}
126
127struct sched_clock_data {
128 u64 tick_raw;
129 u64 tick_gtod;
130 u64 clock;
131};
132
133static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
134
135static inline struct sched_clock_data *this_scd(void)
136{
137 return &__get_cpu_var(sched_clock_data);
138}
139
140static inline struct sched_clock_data *cpu_sdc(int cpu)
141{
142 return &per_cpu(sched_clock_data, cpu);
143}
144
145void sched_clock_init(void)
146{
147 u64 ktime_now = ktime_to_ns(ktime_get());
148 int cpu;
149
150 for_each_possible_cpu(cpu) {
151 struct sched_clock_data *scd = cpu_sdc(cpu);
152
153 scd->tick_raw = 0;
154 scd->tick_gtod = ktime_now;
155 scd->clock = ktime_now;
156 }
157
158 sched_clock_running = 1;
159
160 /*
161 * Ensure that it is impossible to not do a static_key update.
162 *
163 * Either {set,clear}_sched_clock_stable() must see sched_clock_running
164 * and do the update, or we must see their __sched_clock_stable_early
165 * and do the update, or both.
166 */
167 smp_mb(); /* matches {set,clear}_sched_clock_stable() */
168
169 if (__sched_clock_stable_early)
170 __set_sched_clock_stable();
171 else
172 __clear_sched_clock_stable(NULL);
173}
174
175/*
176 * min, max except they take wrapping into account
177 */
178
179static inline u64 wrap_min(u64 x, u64 y)
180{
181 return (s64)(x - y) < 0 ? x : y;
182}
183
184static inline u64 wrap_max(u64 x, u64 y)
185{
186 return (s64)(x - y) > 0 ? x : y;
187}
188
189/*
190 * update the percpu scd from the raw @now value
191 *
192 * - filter out backward motion
193 * - use the GTOD tick value to create a window to filter crazy TSC values
194 */
195static u64 sched_clock_local(struct sched_clock_data *scd)
196{
197 u64 now, clock, old_clock, min_clock, max_clock;
198 s64 delta;
199
200again:
201 now = sched_clock();
202 delta = now - scd->tick_raw;
203 if (unlikely(delta < 0))
204 delta = 0;
205
206 old_clock = scd->clock;
207
208 /*
209 * scd->clock = clamp(scd->tick_gtod + delta,
210 * max(scd->tick_gtod, scd->clock),
211 * scd->tick_gtod + TICK_NSEC);
212 */
213
214 clock = scd->tick_gtod + delta;
215 min_clock = wrap_max(scd->tick_gtod, old_clock);
216 max_clock = wrap_max(old_clock, scd->tick_gtod + TICK_NSEC);
217
218 clock = wrap_max(clock, min_clock);
219 clock = wrap_min(clock, max_clock);
220
221 if (cmpxchg64(&scd->clock, old_clock, clock) != old_clock)
222 goto again;
223
224 return clock;
225}
226
227static u64 sched_clock_remote(struct sched_clock_data *scd)
228{
229 struct sched_clock_data *my_scd = this_scd();
230 u64 this_clock, remote_clock;
231 u64 *ptr, old_val, val;
232
233#if BITS_PER_LONG != 64
234again:
235 /*
236 * Careful here: The local and the remote clock values need to
237 * be read out atomic as we need to compare the values and
238 * then update either the local or the remote side. So the
239 * cmpxchg64 below only protects one readout.
240 *
241 * We must reread via sched_clock_local() in the retry case on
242 * 32bit as an NMI could use sched_clock_local() via the
243 * tracer and hit between the readout of
244 * the low32bit and the high 32bit portion.
245 */
246 this_clock = sched_clock_local(my_scd);
247 /*
248 * We must enforce atomic readout on 32bit, otherwise the
249 * update on the remote cpu can hit inbetween the readout of
250 * the low32bit and the high 32bit portion.
251 */
252 remote_clock = cmpxchg64(&scd->clock, 0, 0);
253#else
254 /*
255 * On 64bit the read of [my]scd->clock is atomic versus the
256 * update, so we can avoid the above 32bit dance.
257 */
258 sched_clock_local(my_scd);
259again:
260 this_clock = my_scd->clock;
261 remote_clock = scd->clock;
262#endif
263
264 /*
265 * Use the opportunity that we have both locks
266 * taken to couple the two clocks: we take the
267 * larger time as the latest time for both
268 * runqueues. (this creates monotonic movement)
269 */
270 if (likely((s64)(remote_clock - this_clock) < 0)) {
271 ptr = &scd->clock;
272 old_val = remote_clock;
273 val = this_clock;
274 } else {
275 /*
276 * Should be rare, but possible:
277 */
278 ptr = &my_scd->clock;
279 old_val = this_clock;
280 val = remote_clock;
281 }
282
283 if (cmpxchg64(ptr, old_val, val) != old_val)
284 goto again;
285
286 return val;
287}
288
289/*
290 * Similar to cpu_clock(), but requires local IRQs to be disabled.
291 *
292 * See cpu_clock().
293 */
294u64 sched_clock_cpu(int cpu)
295{
296 struct sched_clock_data *scd;
297 u64 clock;
298
299 if (sched_clock_stable())
300 return sched_clock();
301
302 if (unlikely(!sched_clock_running))
303 return 0ull;
304
305 preempt_disable_notrace();
306 scd = cpu_sdc(cpu);
307
308 if (cpu != smp_processor_id())
309 clock = sched_clock_remote(scd);
310 else
311 clock = sched_clock_local(scd);
312 preempt_enable_notrace();
313
314 return clock;
315}
316
317void sched_clock_tick(void)
318{
319 struct sched_clock_data *scd;
320 u64 now, now_gtod;
321
322 if (sched_clock_stable())
323 return;
324
325 if (unlikely(!sched_clock_running))
326 return;
327
328 WARN_ON_ONCE(!irqs_disabled());
329
330 scd = this_scd();
331 now_gtod = ktime_to_ns(ktime_get());
332 now = sched_clock();
333
334 scd->tick_raw = now;
335 scd->tick_gtod = now_gtod;
336 sched_clock_local(scd);
337}
338
339/*
340 * We are going deep-idle (irqs are disabled):
341 */
342void sched_clock_idle_sleep_event(void)
343{
344 sched_clock_cpu(smp_processor_id());
345}
346EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
347
348/*
349 * We just idled delta nanoseconds (called with irqs disabled):
350 */
351void sched_clock_idle_wakeup_event(u64 delta_ns)
352{
353 if (timekeeping_suspended)
354 return;
355
356 sched_clock_tick();
357 touch_softlockup_watchdog();
358}
359EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
360
361/*
362 * As outlined at the top, provides a fast, high resolution, nanosecond
363 * time source that is monotonic per cpu argument and has bounded drift
364 * between cpus.
365 *
366 * ######################### BIG FAT WARNING ##########################
367 * # when comparing cpu_clock(i) to cpu_clock(j) for i != j, time can #
368 * # go backwards !! #
369 * ####################################################################
370 */
371u64 cpu_clock(int cpu)
372{
373 if (!sched_clock_stable())
374 return sched_clock_cpu(cpu);
375
376 return sched_clock();
377}
378
379/*
380 * Similar to cpu_clock() for the current cpu. Time will only be observed
381 * to be monotonic if care is taken to only compare timestampt taken on the
382 * same CPU.
383 *
384 * See cpu_clock().
385 */
386u64 local_clock(void)
387{
388 if (!sched_clock_stable())
389 return sched_clock_cpu(raw_smp_processor_id());
390
391 return sched_clock();
392}
393
394#else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
395
396void sched_clock_init(void)
397{
398 sched_clock_running = 1;
399}
400
401u64 sched_clock_cpu(int cpu)
402{
403 if (unlikely(!sched_clock_running))
404 return 0;
405
406 return sched_clock();
407}
408
409u64 cpu_clock(int cpu)
410{
411 return sched_clock();
412}
413
414u64 local_clock(void)
415{
416 return sched_clock();
417}
418
419#endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
420
421EXPORT_SYMBOL_GPL(cpu_clock);
422EXPORT_SYMBOL_GPL(local_clock);