]>
Commit | Line | Data |
---|---|---|
3e51f33f | 1 | /* |
97fb7a0a | 2 | * sched_clock() for unstable CPU clocks |
3e51f33f | 3 | * |
90eec103 | 4 | * Copyright (C) 2008 Red Hat, Inc., Peter Zijlstra |
3e51f33f | 5 | * |
c300ba25 SR |
6 | * Updates and enhancements: |
7 | * Copyright (C) 2008 Red Hat, Inc. Steven Rostedt <srostedt@redhat.com> | |
8 | * | |
3e51f33f PZ |
9 | * Based on code by: |
10 | * Ingo Molnar <mingo@redhat.com> | |
11 | * Guillaume Chazarain <guichaz@gmail.com> | |
12 | * | |
c676329a | 13 | * |
97fb7a0a | 14 | * What this file implements: |
c676329a PZ |
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. | |
97fb7a0a | 29 | * local_clock() -- is cpu_clock() on the current CPU. |
c676329a | 30 | * |
ef08f0ff PZ |
31 | * sched_clock_cpu(i) |
32 | * | |
97fb7a0a | 33 | * How it is implemented: |
c676329a PZ |
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) | |
3e51f33f PZ |
44 | * - sched_clock() |
45 | * - explicit idle events | |
46 | * | |
c676329a PZ |
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. | |
3e51f33f PZ |
50 | * |
51 | * Furthermore, explicit sleep and wakeup hooks allow us to account for time | |
52 | * that is otherwise invisible (TSC gets stopped). | |
53 | * | |
3e51f33f | 54 | */ |
325ea10c | 55 | #include "sched.h" |
5d2a4e91 | 56 | #include <linux/sched_clock.h> |
3e51f33f | 57 | |
2c3d103b HD |
58 | /* |
59 | * Scheduler clock - returns current time in nanosec units. | |
60 | * This is default implementation. | |
61 | * Architectures and sub-architectures can override this. | |
62 | */ | |
52f5684c | 63 | unsigned long long __weak sched_clock(void) |
2c3d103b | 64 | { |
92d23f70 R |
65 | return (unsigned long long)(jiffies - INITIAL_JIFFIES) |
66 | * (NSEC_PER_SEC / HZ); | |
2c3d103b | 67 | } |
b6ac23af | 68 | EXPORT_SYMBOL_GPL(sched_clock); |
3e51f33f | 69 | |
46457ea4 | 70 | static DEFINE_STATIC_KEY_FALSE(sched_clock_running); |
c1955a3d | 71 | |
3e51f33f | 72 | #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK |
acb04058 PZ |
73 | /* |
74 | * We must start with !__sched_clock_stable because the unstable -> stable | |
75 | * transition is accurate, while the stable -> unstable transition is not. | |
76 | * | |
77 | * Similarly we start with __sched_clock_stable_early, thereby assuming we | |
78 | * will become stable, such that there's only a single 1 -> 0 transition. | |
79 | */ | |
555570d7 | 80 | static DEFINE_STATIC_KEY_FALSE(__sched_clock_stable); |
acb04058 | 81 | static int __sched_clock_stable_early = 1; |
35af99e6 | 82 | |
5680d809 | 83 | /* |
698eff63 | 84 | * We want: ktime_get_ns() + __gtod_offset == sched_clock() + __sched_clock_offset |
5680d809 | 85 | */ |
698eff63 PZ |
86 | __read_mostly u64 __sched_clock_offset; |
87 | static __read_mostly u64 __gtod_offset; | |
5680d809 PZ |
88 | |
89 | struct sched_clock_data { | |
90 | u64 tick_raw; | |
91 | u64 tick_gtod; | |
92 | u64 clock; | |
93 | }; | |
94 | ||
95 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data); | |
96 | ||
97 | static inline struct sched_clock_data *this_scd(void) | |
98 | { | |
99 | return this_cpu_ptr(&sched_clock_data); | |
100 | } | |
101 | ||
102 | static inline struct sched_clock_data *cpu_sdc(int cpu) | |
103 | { | |
104 | return &per_cpu(sched_clock_data, cpu); | |
105 | } | |
106 | ||
35af99e6 PZ |
107 | int sched_clock_stable(void) |
108 | { | |
555570d7 | 109 | return static_branch_likely(&__sched_clock_stable); |
35af99e6 PZ |
110 | } |
111 | ||
cf15ca8d PZ |
112 | static void __scd_stamp(struct sched_clock_data *scd) |
113 | { | |
114 | scd->tick_gtod = ktime_get_ns(); | |
115 | scd->tick_raw = sched_clock(); | |
116 | } | |
117 | ||
d375b4e0 | 118 | static void __set_sched_clock_stable(void) |
35af99e6 | 119 | { |
45aea321 | 120 | struct sched_clock_data *scd; |
5680d809 | 121 | |
45aea321 PZ |
122 | /* |
123 | * Since we're still unstable and the tick is already running, we have | |
124 | * to disable IRQs in order to get a consistent scd->tick* reading. | |
125 | */ | |
126 | local_irq_disable(); | |
127 | scd = this_scd(); | |
5680d809 PZ |
128 | /* |
129 | * Attempt to make the (initial) unstable->stable transition continuous. | |
130 | */ | |
698eff63 | 131 | __sched_clock_offset = (scd->tick_gtod + __gtod_offset) - (scd->tick_raw); |
45aea321 | 132 | local_irq_enable(); |
5680d809 PZ |
133 | |
134 | printk(KERN_INFO "sched_clock: Marking stable (%lld, %lld)->(%lld, %lld)\n", | |
698eff63 PZ |
135 | scd->tick_gtod, __gtod_offset, |
136 | scd->tick_raw, __sched_clock_offset); | |
5680d809 | 137 | |
555570d7 | 138 | static_branch_enable(&__sched_clock_stable); |
4f49b90a | 139 | tick_dep_clear(TICK_DEP_BIT_CLOCK_UNSTABLE); |
d375b4e0 PZ |
140 | } |
141 | ||
cf15ca8d PZ |
142 | /* |
143 | * If we ever get here, we're screwed, because we found out -- typically after | |
144 | * the fact -- that TSC wasn't good. This means all our clocksources (including | |
145 | * ktime) could have reported wrong values. | |
146 | * | |
147 | * What we do here is an attempt to fix up and continue sort of where we left | |
148 | * off in a coherent manner. | |
149 | * | |
150 | * The only way to fully avoid random clock jumps is to boot with: | |
151 | * "tsc=unstable". | |
152 | */ | |
71fdb70e PZ |
153 | static void __sched_clock_work(struct work_struct *work) |
154 | { | |
cf15ca8d PZ |
155 | struct sched_clock_data *scd; |
156 | int cpu; | |
157 | ||
158 | /* take a current timestamp and set 'now' */ | |
159 | preempt_disable(); | |
160 | scd = this_scd(); | |
161 | __scd_stamp(scd); | |
162 | scd->clock = scd->tick_gtod + __gtod_offset; | |
163 | preempt_enable(); | |
164 | ||
165 | /* clone to all CPUs */ | |
166 | for_each_possible_cpu(cpu) | |
167 | per_cpu(sched_clock_data, cpu) = *scd; | |
168 | ||
7708d5f0 | 169 | printk(KERN_WARNING "TSC found unstable after boot, most likely due to broken BIOS. Use 'tsc=unstable'.\n"); |
cf15ca8d PZ |
170 | printk(KERN_INFO "sched_clock: Marking unstable (%lld, %lld)<-(%lld, %lld)\n", |
171 | scd->tick_gtod, __gtod_offset, | |
172 | scd->tick_raw, __sched_clock_offset); | |
173 | ||
71fdb70e PZ |
174 | static_branch_disable(&__sched_clock_stable); |
175 | } | |
176 | ||
177 | static DECLARE_WORK(sched_clock_work, __sched_clock_work); | |
178 | ||
179 | static void __clear_sched_clock_stable(void) | |
35af99e6 | 180 | { |
cf15ca8d PZ |
181 | if (!sched_clock_stable()) |
182 | return; | |
5680d809 | 183 | |
4f49b90a | 184 | tick_dep_set(TICK_DEP_BIT_CLOCK_UNSTABLE); |
cf15ca8d | 185 | schedule_work(&sched_clock_work); |
71fdb70e | 186 | } |
6577e42a PZ |
187 | |
188 | void clear_sched_clock_stable(void) | |
189 | { | |
d375b4e0 PZ |
190 | __sched_clock_stable_early = 0; |
191 | ||
9881b024 | 192 | smp_mb(); /* matches sched_clock_init_late() */ |
d375b4e0 | 193 | |
46457ea4 | 194 | if (static_key_count(&sched_clock_running.key) == 2) |
71fdb70e | 195 | __clear_sched_clock_stable(); |
6577e42a PZ |
196 | } |
197 | ||
5d2a4e91 PT |
198 | static void __sched_clock_gtod_offset(void) |
199 | { | |
9407f5a7 PZ |
200 | struct sched_clock_data *scd = this_scd(); |
201 | ||
202 | __scd_stamp(scd); | |
203 | __gtod_offset = (scd->tick_raw + __sched_clock_offset) - scd->tick_gtod; | |
5d2a4e91 PT |
204 | } |
205 | ||
206 | void __init sched_clock_init(void) | |
207 | { | |
857baa87 PT |
208 | /* |
209 | * Set __gtod_offset such that once we mark sched_clock_running, | |
210 | * sched_clock_tick() continues where sched_clock() left off. | |
211 | * | |
212 | * Even if TSC is buggered, we're still UP at this point so it | |
213 | * can't really be out of sync. | |
214 | */ | |
9407f5a7 | 215 | local_irq_disable(); |
857baa87 | 216 | __sched_clock_gtod_offset(); |
9407f5a7 | 217 | local_irq_enable(); |
857baa87 | 218 | |
46457ea4 | 219 | static_branch_inc(&sched_clock_running); |
5d2a4e91 | 220 | } |
2e44b7dd PZ |
221 | /* |
222 | * We run this as late_initcall() such that it runs after all built-in drivers, | |
223 | * notably: acpi_processor and intel_idle, which can mark the TSC as unstable. | |
224 | */ | |
225 | static int __init sched_clock_init_late(void) | |
3e51f33f | 226 | { |
46457ea4 | 227 | static_branch_inc(&sched_clock_running); |
d375b4e0 PZ |
228 | /* |
229 | * Ensure that it is impossible to not do a static_key update. | |
230 | * | |
231 | * Either {set,clear}_sched_clock_stable() must see sched_clock_running | |
232 | * and do the update, or we must see their __sched_clock_stable_early | |
233 | * and do the update, or both. | |
234 | */ | |
235 | smp_mb(); /* matches {set,clear}_sched_clock_stable() */ | |
236 | ||
237 | if (__sched_clock_stable_early) | |
238 | __set_sched_clock_stable(); | |
2e44b7dd PZ |
239 | |
240 | return 0; | |
3e51f33f | 241 | } |
2e44b7dd | 242 | late_initcall(sched_clock_init_late); |
3e51f33f | 243 | |
354879bb | 244 | /* |
b342501c | 245 | * min, max except they take wrapping into account |
354879bb PZ |
246 | */ |
247 | ||
248 | static inline u64 wrap_min(u64 x, u64 y) | |
249 | { | |
250 | return (s64)(x - y) < 0 ? x : y; | |
251 | } | |
252 | ||
253 | static inline u64 wrap_max(u64 x, u64 y) | |
254 | { | |
255 | return (s64)(x - y) > 0 ? x : y; | |
256 | } | |
257 | ||
3e51f33f PZ |
258 | /* |
259 | * update the percpu scd from the raw @now value | |
260 | * | |
261 | * - filter out backward motion | |
354879bb | 262 | * - use the GTOD tick value to create a window to filter crazy TSC values |
3e51f33f | 263 | */ |
def0a9b2 | 264 | static u64 sched_clock_local(struct sched_clock_data *scd) |
3e51f33f | 265 | { |
7b09cc5a | 266 | u64 now, clock, old_clock, min_clock, max_clock, gtod; |
def0a9b2 | 267 | s64 delta; |
3e51f33f | 268 | |
def0a9b2 PZ |
269 | again: |
270 | now = sched_clock(); | |
271 | delta = now - scd->tick_raw; | |
354879bb PZ |
272 | if (unlikely(delta < 0)) |
273 | delta = 0; | |
3e51f33f | 274 | |
def0a9b2 PZ |
275 | old_clock = scd->clock; |
276 | ||
354879bb PZ |
277 | /* |
278 | * scd->clock = clamp(scd->tick_gtod + delta, | |
b342501c IM |
279 | * max(scd->tick_gtod, scd->clock), |
280 | * scd->tick_gtod + TICK_NSEC); | |
354879bb | 281 | */ |
3e51f33f | 282 | |
7b09cc5a PT |
283 | gtod = scd->tick_gtod + __gtod_offset; |
284 | clock = gtod + delta; | |
285 | min_clock = wrap_max(gtod, old_clock); | |
286 | max_clock = wrap_max(old_clock, gtod + TICK_NSEC); | |
3e51f33f | 287 | |
354879bb PZ |
288 | clock = wrap_max(clock, min_clock); |
289 | clock = wrap_min(clock, max_clock); | |
3e51f33f | 290 | |
152f9d07 | 291 | if (cmpxchg64(&scd->clock, old_clock, clock) != old_clock) |
def0a9b2 | 292 | goto again; |
56b90612 | 293 | |
def0a9b2 | 294 | return clock; |
3e51f33f PZ |
295 | } |
296 | ||
def0a9b2 | 297 | static u64 sched_clock_remote(struct sched_clock_data *scd) |
3e51f33f | 298 | { |
def0a9b2 PZ |
299 | struct sched_clock_data *my_scd = this_scd(); |
300 | u64 this_clock, remote_clock; | |
301 | u64 *ptr, old_val, val; | |
302 | ||
a1cbcaa9 TG |
303 | #if BITS_PER_LONG != 64 |
304 | again: | |
305 | /* | |
306 | * Careful here: The local and the remote clock values need to | |
307 | * be read out atomic as we need to compare the values and | |
308 | * then update either the local or the remote side. So the | |
309 | * cmpxchg64 below only protects one readout. | |
310 | * | |
311 | * We must reread via sched_clock_local() in the retry case on | |
97fb7a0a | 312 | * 32-bit kernels as an NMI could use sched_clock_local() via the |
a1cbcaa9 | 313 | * tracer and hit between the readout of |
97fb7a0a | 314 | * the low 32-bit and the high 32-bit portion. |
a1cbcaa9 TG |
315 | */ |
316 | this_clock = sched_clock_local(my_scd); | |
317 | /* | |
97fb7a0a IM |
318 | * We must enforce atomic readout on 32-bit, otherwise the |
319 | * update on the remote CPU can hit inbetween the readout of | |
320 | * the low 32-bit and the high 32-bit portion. | |
a1cbcaa9 TG |
321 | */ |
322 | remote_clock = cmpxchg64(&scd->clock, 0, 0); | |
323 | #else | |
324 | /* | |
97fb7a0a IM |
325 | * On 64-bit kernels the read of [my]scd->clock is atomic versus the |
326 | * update, so we can avoid the above 32-bit dance. | |
a1cbcaa9 | 327 | */ |
def0a9b2 PZ |
328 | sched_clock_local(my_scd); |
329 | again: | |
330 | this_clock = my_scd->clock; | |
331 | remote_clock = scd->clock; | |
a1cbcaa9 | 332 | #endif |
def0a9b2 PZ |
333 | |
334 | /* | |
335 | * Use the opportunity that we have both locks | |
336 | * taken to couple the two clocks: we take the | |
337 | * larger time as the latest time for both | |
338 | * runqueues. (this creates monotonic movement) | |
339 | */ | |
340 | if (likely((s64)(remote_clock - this_clock) < 0)) { | |
341 | ptr = &scd->clock; | |
342 | old_val = remote_clock; | |
343 | val = this_clock; | |
3e51f33f | 344 | } else { |
def0a9b2 PZ |
345 | /* |
346 | * Should be rare, but possible: | |
347 | */ | |
348 | ptr = &my_scd->clock; | |
349 | old_val = this_clock; | |
350 | val = remote_clock; | |
3e51f33f | 351 | } |
def0a9b2 | 352 | |
152f9d07 | 353 | if (cmpxchg64(ptr, old_val, val) != old_val) |
def0a9b2 PZ |
354 | goto again; |
355 | ||
356 | return val; | |
3e51f33f PZ |
357 | } |
358 | ||
c676329a PZ |
359 | /* |
360 | * Similar to cpu_clock(), but requires local IRQs to be disabled. | |
361 | * | |
362 | * See cpu_clock(). | |
363 | */ | |
3e51f33f PZ |
364 | u64 sched_clock_cpu(int cpu) |
365 | { | |
b342501c | 366 | struct sched_clock_data *scd; |
def0a9b2 PZ |
367 | u64 clock; |
368 | ||
35af99e6 | 369 | if (sched_clock_stable()) |
698eff63 | 370 | return sched_clock() + __sched_clock_offset; |
a381759d | 371 | |
46457ea4 | 372 | if (!static_branch_unlikely(&sched_clock_running)) |
857baa87 | 373 | return sched_clock(); |
a381759d | 374 | |
96b3d28b | 375 | preempt_disable_notrace(); |
def0a9b2 | 376 | scd = cpu_sdc(cpu); |
3e51f33f | 377 | |
def0a9b2 PZ |
378 | if (cpu != smp_processor_id()) |
379 | clock = sched_clock_remote(scd); | |
380 | else | |
381 | clock = sched_clock_local(scd); | |
96b3d28b | 382 | preempt_enable_notrace(); |
e4e4e534 | 383 | |
3e51f33f PZ |
384 | return clock; |
385 | } | |
2c923e94 | 386 | EXPORT_SYMBOL_GPL(sched_clock_cpu); |
3e51f33f PZ |
387 | |
388 | void sched_clock_tick(void) | |
389 | { | |
8325d9c0 | 390 | struct sched_clock_data *scd; |
a381759d | 391 | |
b421b22b PZ |
392 | if (sched_clock_stable()) |
393 | return; | |
394 | ||
46457ea4 | 395 | if (!static_branch_unlikely(&sched_clock_running)) |
b421b22b PZ |
396 | return; |
397 | ||
2c11dba0 | 398 | lockdep_assert_irqs_disabled(); |
3e51f33f | 399 | |
8325d9c0 | 400 | scd = this_scd(); |
cf15ca8d | 401 | __scd_stamp(scd); |
b421b22b PZ |
402 | sched_clock_local(scd); |
403 | } | |
404 | ||
405 | void sched_clock_tick_stable(void) | |
406 | { | |
b421b22b PZ |
407 | if (!sched_clock_stable()) |
408 | return; | |
409 | ||
410 | /* | |
411 | * Called under watchdog_lock. | |
412 | * | |
413 | * The watchdog just found this TSC to (still) be stable, so now is a | |
414 | * good moment to update our __gtod_offset. Because once we find the | |
415 | * TSC to be unstable, any computation will be computing crap. | |
416 | */ | |
417 | local_irq_disable(); | |
5d2a4e91 | 418 | __sched_clock_gtod_offset(); |
b421b22b | 419 | local_irq_enable(); |
3e51f33f PZ |
420 | } |
421 | ||
422 | /* | |
423 | * We are going deep-idle (irqs are disabled): | |
424 | */ | |
425 | void sched_clock_idle_sleep_event(void) | |
426 | { | |
427 | sched_clock_cpu(smp_processor_id()); | |
428 | } | |
429 | EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event); | |
430 | ||
431 | /* | |
f9fccdb9 | 432 | * We just idled; resync with ktime. |
3e51f33f | 433 | */ |
ac1e843f | 434 | void sched_clock_idle_wakeup_event(void) |
3e51f33f | 435 | { |
f9fccdb9 PZ |
436 | unsigned long flags; |
437 | ||
438 | if (sched_clock_stable()) | |
439 | return; | |
440 | ||
441 | if (unlikely(timekeeping_suspended)) | |
1c5745aa TG |
442 | return; |
443 | ||
f9fccdb9 | 444 | local_irq_save(flags); |
354879bb | 445 | sched_clock_tick(); |
f9fccdb9 | 446 | local_irq_restore(flags); |
3e51f33f PZ |
447 | } |
448 | EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event); | |
449 | ||
8325d9c0 PZ |
450 | #else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */ |
451 | ||
5d2a4e91 PT |
452 | void __init sched_clock_init(void) |
453 | { | |
46457ea4 | 454 | static_branch_inc(&sched_clock_running); |
bd9f943e | 455 | local_irq_disable(); |
5d2a4e91 | 456 | generic_sched_clock_init(); |
bd9f943e | 457 | local_irq_enable(); |
5d2a4e91 PT |
458 | } |
459 | ||
8325d9c0 PZ |
460 | u64 sched_clock_cpu(int cpu) |
461 | { | |
46457ea4 | 462 | if (!static_branch_unlikely(&sched_clock_running)) |
8325d9c0 PZ |
463 | return 0; |
464 | ||
465 | return sched_clock(); | |
466 | } | |
9881b024 | 467 | |
b9f8fcd5 | 468 | #endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */ |
76a2a6ee | 469 | |
545a2bf7 CB |
470 | /* |
471 | * Running clock - returns the time that has elapsed while a guest has been | |
472 | * running. | |
473 | * On a guest this value should be local_clock minus the time the guest was | |
474 | * suspended by the hypervisor (for any reason). | |
475 | * On bare metal this function should return the same as local_clock. | |
476 | * Architectures and sub-architectures can override this. | |
477 | */ | |
478 | u64 __weak running_clock(void) | |
479 | { | |
480 | return local_clock(); | |
481 | } |