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 monotomic)
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 #include "sched.h"
57 #include <linux/sched_clock.h>
58
59 /*
60 * Scheduler clock - returns current time in nanosec units.
61 * This is default implementation.
62 * Architectures and sub-architectures can override this.
63 */
64 unsigned long long __weak sched_clock(void)
65 {
66 return (unsigned long long)(jiffies - INITIAL_JIFFIES)
67 * (NSEC_PER_SEC / HZ);
68 }
69 EXPORT_SYMBOL_GPL(sched_clock);
70
71 static DEFINE_STATIC_KEY_FALSE(sched_clock_running);
72
73 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
74 /*
75 * We must start with !__sched_clock_stable because the unstable -> stable
76 * transition is accurate, while the stable -> unstable transition is not.
77 *
78 * Similarly we start with __sched_clock_stable_early, thereby assuming we
79 * will become stable, such that there's only a single 1 -> 0 transition.
80 */
81 static DEFINE_STATIC_KEY_FALSE(__sched_clock_stable);
82 static int __sched_clock_stable_early = 1;
83
84 /*
85 * We want: ktime_get_ns() + __gtod_offset == sched_clock() + __sched_clock_offset
86 */
87 __read_mostly u64 __sched_clock_offset;
88 static __read_mostly u64 __gtod_offset;
89
90 struct sched_clock_data {
91 u64 tick_raw;
92 u64 tick_gtod;
93 u64 clock;
94 };
95
96 static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
97
98 static inline struct sched_clock_data *this_scd(void)
99 {
100 return this_cpu_ptr(&sched_clock_data);
101 }
102
103 static inline struct sched_clock_data *cpu_sdc(int cpu)
104 {
105 return &per_cpu(sched_clock_data, cpu);
106 }
107
108 int sched_clock_stable(void)
109 {
110 return static_branch_likely(&__sched_clock_stable);
111 }
112
113 static void __scd_stamp(struct sched_clock_data *scd)
114 {
115 scd->tick_gtod = ktime_get_ns();
116 scd->tick_raw = sched_clock();
117 }
118
119 static void __set_sched_clock_stable(void)
120 {
121 struct sched_clock_data *scd;
122
123 /*
124 * Since we're still unstable and the tick is already running, we have
125 * to disable IRQs in order to get a consistent scd->tick* reading.
126 */
127 local_irq_disable();
128 scd = this_scd();
129 /*
130 * Attempt to make the (initial) unstable->stable transition continuous.
131 */
132 __sched_clock_offset = (scd->tick_gtod + __gtod_offset) - (scd->tick_raw);
133 local_irq_enable();
134
135 printk(KERN_INFO "sched_clock: Marking stable (%lld, %lld)->(%lld, %lld)\n",
136 scd->tick_gtod, __gtod_offset,
137 scd->tick_raw, __sched_clock_offset);
138
139 static_branch_enable(&__sched_clock_stable);
140 tick_dep_clear(TICK_DEP_BIT_CLOCK_UNSTABLE);
141 }
142
143 /*
144 * If we ever get here, we're screwed, because we found out -- typically after
145 * the fact -- that TSC wasn't good. This means all our clocksources (including
146 * ktime) could have reported wrong values.
147 *
148 * What we do here is an attempt to fix up and continue sort of where we left
149 * off in a coherent manner.
150 *
151 * The only way to fully avoid random clock jumps is to boot with:
152 * "tsc=unstable".
153 */
154 static void __sched_clock_work(struct work_struct *work)
155 {
156 struct sched_clock_data *scd;
157 int cpu;
158
159 /* take a current timestamp and set 'now' */
160 preempt_disable();
161 scd = this_scd();
162 __scd_stamp(scd);
163 scd->clock = scd->tick_gtod + __gtod_offset;
164 preempt_enable();
165
166 /* clone to all CPUs */
167 for_each_possible_cpu(cpu)
168 per_cpu(sched_clock_data, cpu) = *scd;
169
170 printk(KERN_WARNING "TSC found unstable after boot, most likely due to broken BIOS. Use 'tsc=unstable'.\n");
171 printk(KERN_INFO "sched_clock: Marking unstable (%lld, %lld)<-(%lld, %lld)\n",
172 scd->tick_gtod, __gtod_offset,
173 scd->tick_raw, __sched_clock_offset);
174
175 static_branch_disable(&__sched_clock_stable);
176 }
177
178 static DECLARE_WORK(sched_clock_work, __sched_clock_work);
179
180 static void __clear_sched_clock_stable(void)
181 {
182 if (!sched_clock_stable())
183 return;
184
185 tick_dep_set(TICK_DEP_BIT_CLOCK_UNSTABLE);
186 schedule_work(&sched_clock_work);
187 }
188
189 void clear_sched_clock_stable(void)
190 {
191 __sched_clock_stable_early = 0;
192
193 smp_mb(); /* matches sched_clock_init_late() */
194
195 if (static_key_count(&sched_clock_running.key) == 2)
196 __clear_sched_clock_stable();
197 }
198
199 static void __sched_clock_gtod_offset(void)
200 {
201 struct sched_clock_data *scd = this_scd();
202
203 __scd_stamp(scd);
204 __gtod_offset = (scd->tick_raw + __sched_clock_offset) - scd->tick_gtod;
205 }
206
207 void __init sched_clock_init(void)
208 {
209 /*
210 * Set __gtod_offset such that once we mark sched_clock_running,
211 * sched_clock_tick() continues where sched_clock() left off.
212 *
213 * Even if TSC is buggered, we're still UP at this point so it
214 * can't really be out of sync.
215 */
216 local_irq_disable();
217 __sched_clock_gtod_offset();
218 local_irq_enable();
219
220 static_branch_inc(&sched_clock_running);
221 }
222 /*
223 * We run this as late_initcall() such that it runs after all built-in drivers,
224 * notably: acpi_processor and intel_idle, which can mark the TSC as unstable.
225 */
226 static int __init sched_clock_init_late(void)
227 {
228 static_branch_inc(&sched_clock_running);
229 /*
230 * Ensure that it is impossible to not do a static_key update.
231 *
232 * Either {set,clear}_sched_clock_stable() must see sched_clock_running
233 * and do the update, or we must see their __sched_clock_stable_early
234 * and do the update, or both.
235 */
236 smp_mb(); /* matches {set,clear}_sched_clock_stable() */
237
238 if (__sched_clock_stable_early)
239 __set_sched_clock_stable();
240
241 return 0;
242 }
243 late_initcall(sched_clock_init_late);
244
245 /*
246 * min, max except they take wrapping into account
247 */
248
249 static inline u64 wrap_min(u64 x, u64 y)
250 {
251 return (s64)(x - y) < 0 ? x : y;
252 }
253
254 static inline u64 wrap_max(u64 x, u64 y)
255 {
256 return (s64)(x - y) > 0 ? x : y;
257 }
258
259 /*
260 * update the percpu scd from the raw @now value
261 *
262 * - filter out backward motion
263 * - use the GTOD tick value to create a window to filter crazy TSC values
264 */
265 static u64 sched_clock_local(struct sched_clock_data *scd)
266 {
267 u64 now, clock, old_clock, min_clock, max_clock, gtod;
268 s64 delta;
269
270 again:
271 now = sched_clock();
272 delta = now - scd->tick_raw;
273 if (unlikely(delta < 0))
274 delta = 0;
275
276 old_clock = scd->clock;
277
278 /*
279 * scd->clock = clamp(scd->tick_gtod + delta,
280 * max(scd->tick_gtod, scd->clock),
281 * scd->tick_gtod + TICK_NSEC);
282 */
283
284 gtod = scd->tick_gtod + __gtod_offset;
285 clock = gtod + delta;
286 min_clock = wrap_max(gtod, old_clock);
287 max_clock = wrap_max(old_clock, gtod + TICK_NSEC);
288
289 clock = wrap_max(clock, min_clock);
290 clock = wrap_min(clock, max_clock);
291
292 if (cmpxchg64(&scd->clock, old_clock, clock) != old_clock)
293 goto again;
294
295 return clock;
296 }
297
298 static u64 sched_clock_remote(struct sched_clock_data *scd)
299 {
300 struct sched_clock_data *my_scd = this_scd();
301 u64 this_clock, remote_clock;
302 u64 *ptr, old_val, val;
303
304 #if BITS_PER_LONG != 64
305 again:
306 /*
307 * Careful here: The local and the remote clock values need to
308 * be read out atomic as we need to compare the values and
309 * then update either the local or the remote side. So the
310 * cmpxchg64 below only protects one readout.
311 *
312 * We must reread via sched_clock_local() in the retry case on
313 * 32-bit kernels as an NMI could use sched_clock_local() via the
314 * tracer and hit between the readout of
315 * the low 32-bit and the high 32-bit portion.
316 */
317 this_clock = sched_clock_local(my_scd);
318 /*
319 * We must enforce atomic readout on 32-bit, otherwise the
320 * update on the remote CPU can hit inbetween the readout of
321 * the low 32-bit and the high 32-bit portion.
322 */
323 remote_clock = cmpxchg64(&scd->clock, 0, 0);
324 #else
325 /*
326 * On 64-bit kernels the read of [my]scd->clock is atomic versus the
327 * update, so we can avoid the above 32-bit dance.
328 */
329 sched_clock_local(my_scd);
330 again:
331 this_clock = my_scd->clock;
332 remote_clock = scd->clock;
333 #endif
334
335 /*
336 * Use the opportunity that we have both locks
337 * taken to couple the two clocks: we take the
338 * larger time as the latest time for both
339 * runqueues. (this creates monotonic movement)
340 */
341 if (likely((s64)(remote_clock - this_clock) < 0)) {
342 ptr = &scd->clock;
343 old_val = remote_clock;
344 val = this_clock;
345 } else {
346 /*
347 * Should be rare, but possible:
348 */
349 ptr = &my_scd->clock;
350 old_val = this_clock;
351 val = remote_clock;
352 }
353
354 if (cmpxchg64(ptr, old_val, val) != old_val)
355 goto again;
356
357 return val;
358 }
359
360 /*
361 * Similar to cpu_clock(), but requires local IRQs to be disabled.
362 *
363 * See cpu_clock().
364 */
365 u64 sched_clock_cpu(int cpu)
366 {
367 struct sched_clock_data *scd;
368 u64 clock;
369
370 if (sched_clock_stable())
371 return sched_clock() + __sched_clock_offset;
372
373 if (!static_branch_unlikely(&sched_clock_running))
374 return sched_clock();
375
376 preempt_disable_notrace();
377 scd = cpu_sdc(cpu);
378
379 if (cpu != smp_processor_id())
380 clock = sched_clock_remote(scd);
381 else
382 clock = sched_clock_local(scd);
383 preempt_enable_notrace();
384
385 return clock;
386 }
387 EXPORT_SYMBOL_GPL(sched_clock_cpu);
388
389 void sched_clock_tick(void)
390 {
391 struct sched_clock_data *scd;
392
393 if (sched_clock_stable())
394 return;
395
396 if (!static_branch_unlikely(&sched_clock_running))
397 return;
398
399 lockdep_assert_irqs_disabled();
400
401 scd = this_scd();
402 __scd_stamp(scd);
403 sched_clock_local(scd);
404 }
405
406 void sched_clock_tick_stable(void)
407 {
408 if (!sched_clock_stable())
409 return;
410
411 /*
412 * Called under watchdog_lock.
413 *
414 * The watchdog just found this TSC to (still) be stable, so now is a
415 * good moment to update our __gtod_offset. Because once we find the
416 * TSC to be unstable, any computation will be computing crap.
417 */
418 local_irq_disable();
419 __sched_clock_gtod_offset();
420 local_irq_enable();
421 }
422
423 /*
424 * We are going deep-idle (irqs are disabled):
425 */
426 void sched_clock_idle_sleep_event(void)
427 {
428 sched_clock_cpu(smp_processor_id());
429 }
430 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
431
432 /*
433 * We just idled; resync with ktime.
434 */
435 void sched_clock_idle_wakeup_event(void)
436 {
437 unsigned long flags;
438
439 if (sched_clock_stable())
440 return;
441
442 if (unlikely(timekeeping_suspended))
443 return;
444
445 local_irq_save(flags);
446 sched_clock_tick();
447 local_irq_restore(flags);
448 }
449 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
450
451 #else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
452
453 void __init sched_clock_init(void)
454 {
455 static_branch_inc(&sched_clock_running);
456 local_irq_disable();
457 generic_sched_clock_init();
458 local_irq_enable();
459 }
460
461 u64 sched_clock_cpu(int cpu)
462 {
463 if (!static_branch_unlikely(&sched_clock_running))
464 return 0;
465
466 return sched_clock();
467 }
468
469 #endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
470
471 /*
472 * Running clock - returns the time that has elapsed while a guest has been
473 * running.
474 * On a guest this value should be local_clock minus the time the guest was
475 * suspended by the hypervisor (for any reason).
476 * On bare metal this function should return the same as local_clock.
477 * Architectures and sub-architectures can override this.
478 */
479 u64 __weak running_clock(void)
480 {
481 return local_clock();
482 }