Commit | Line | Data |
---|---|---|
3e51f33f PZ |
1 | /* |
2 | * sched_clock for unstable cpu clocks | |
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 PZ |
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. | |
c676329a PZ |
29 | * local_clock() -- is cpu_clock() on the current cpu. |
30 | * | |
ef08f0ff PZ |
31 | * sched_clock_cpu(i) |
32 | * | |
c676329a PZ |
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) | |
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 | */ |
3e51f33f | 55 | #include <linux/spinlock.h> |
6409c4da | 56 | #include <linux/hardirq.h> |
9984de1a | 57 | #include <linux/export.h> |
b342501c IM |
58 | #include <linux/percpu.h> |
59 | #include <linux/ktime.h> | |
60 | #include <linux/sched.h> | |
35af99e6 | 61 | #include <linux/static_key.h> |
6577e42a | 62 | #include <linux/workqueue.h> |
52f5684c | 63 | #include <linux/compiler.h> |
4f49b90a | 64 | #include <linux/tick.h> |
3e51f33f | 65 | |
2c3d103b HD |
66 | /* |
67 | * Scheduler clock - returns current time in nanosec units. | |
68 | * This is default implementation. | |
69 | * Architectures and sub-architectures can override this. | |
70 | */ | |
52f5684c | 71 | unsigned long long __weak sched_clock(void) |
2c3d103b | 72 | { |
92d23f70 R |
73 | return (unsigned long long)(jiffies - INITIAL_JIFFIES) |
74 | * (NSEC_PER_SEC / HZ); | |
2c3d103b | 75 | } |
b6ac23af | 76 | EXPORT_SYMBOL_GPL(sched_clock); |
3e51f33f | 77 | |
5bb6b1ea | 78 | __read_mostly int sched_clock_running; |
c1955a3d | 79 | |
3e51f33f | 80 | #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK |
35af99e6 | 81 | static struct static_key __sched_clock_stable = STATIC_KEY_INIT; |
d375b4e0 | 82 | static int __sched_clock_stable_early; |
35af99e6 PZ |
83 | |
84 | int sched_clock_stable(void) | |
85 | { | |
d375b4e0 | 86 | return static_key_false(&__sched_clock_stable); |
35af99e6 PZ |
87 | } |
88 | ||
d375b4e0 | 89 | static void __set_sched_clock_stable(void) |
35af99e6 PZ |
90 | { |
91 | if (!sched_clock_stable()) | |
d375b4e0 | 92 | static_key_slow_inc(&__sched_clock_stable); |
4f49b90a FW |
93 | |
94 | tick_dep_clear(TICK_DEP_BIT_CLOCK_UNSTABLE); | |
d375b4e0 PZ |
95 | } |
96 | ||
97 | void set_sched_clock_stable(void) | |
98 | { | |
99 | __sched_clock_stable_early = 1; | |
100 | ||
101 | smp_mb(); /* matches sched_clock_init() */ | |
102 | ||
103 | if (!sched_clock_running) | |
104 | return; | |
105 | ||
106 | __set_sched_clock_stable(); | |
35af99e6 PZ |
107 | } |
108 | ||
6577e42a | 109 | static void __clear_sched_clock_stable(struct work_struct *work) |
35af99e6 PZ |
110 | { |
111 | /* XXX worry about clock continuity */ | |
112 | if (sched_clock_stable()) | |
d375b4e0 | 113 | static_key_slow_dec(&__sched_clock_stable); |
4f49b90a FW |
114 | |
115 | tick_dep_set(TICK_DEP_BIT_CLOCK_UNSTABLE); | |
35af99e6 | 116 | } |
3e51f33f | 117 | |
6577e42a PZ |
118 | static DECLARE_WORK(sched_clock_work, __clear_sched_clock_stable); |
119 | ||
120 | void clear_sched_clock_stable(void) | |
121 | { | |
d375b4e0 PZ |
122 | __sched_clock_stable_early = 0; |
123 | ||
124 | smp_mb(); /* matches sched_clock_init() */ | |
125 | ||
126 | if (!sched_clock_running) | |
127 | return; | |
128 | ||
129 | schedule_work(&sched_clock_work); | |
6577e42a PZ |
130 | } |
131 | ||
3e51f33f | 132 | struct sched_clock_data { |
3e51f33f PZ |
133 | u64 tick_raw; |
134 | u64 tick_gtod; | |
135 | u64 clock; | |
136 | }; | |
137 | ||
138 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data); | |
139 | ||
140 | static inline struct sched_clock_data *this_scd(void) | |
141 | { | |
22127e93 | 142 | return this_cpu_ptr(&sched_clock_data); |
3e51f33f PZ |
143 | } |
144 | ||
145 | static inline struct sched_clock_data *cpu_sdc(int cpu) | |
146 | { | |
147 | return &per_cpu(sched_clock_data, cpu); | |
148 | } | |
149 | ||
150 | void sched_clock_init(void) | |
151 | { | |
152 | u64 ktime_now = ktime_to_ns(ktime_get()); | |
3e51f33f PZ |
153 | int cpu; |
154 | ||
155 | for_each_possible_cpu(cpu) { | |
156 | struct sched_clock_data *scd = cpu_sdc(cpu); | |
157 | ||
a381759d | 158 | scd->tick_raw = 0; |
3e51f33f PZ |
159 | scd->tick_gtod = ktime_now; |
160 | scd->clock = ktime_now; | |
161 | } | |
a381759d PZ |
162 | |
163 | sched_clock_running = 1; | |
d375b4e0 PZ |
164 | |
165 | /* | |
166 | * Ensure that it is impossible to not do a static_key update. | |
167 | * | |
168 | * Either {set,clear}_sched_clock_stable() must see sched_clock_running | |
169 | * and do the update, or we must see their __sched_clock_stable_early | |
170 | * and do the update, or both. | |
171 | */ | |
172 | smp_mb(); /* matches {set,clear}_sched_clock_stable() */ | |
173 | ||
174 | if (__sched_clock_stable_early) | |
175 | __set_sched_clock_stable(); | |
176 | else | |
177 | __clear_sched_clock_stable(NULL); | |
3e51f33f PZ |
178 | } |
179 | ||
354879bb | 180 | /* |
b342501c | 181 | * min, max except they take wrapping into account |
354879bb PZ |
182 | */ |
183 | ||
184 | static inline u64 wrap_min(u64 x, u64 y) | |
185 | { | |
186 | return (s64)(x - y) < 0 ? x : y; | |
187 | } | |
188 | ||
189 | static inline u64 wrap_max(u64 x, u64 y) | |
190 | { | |
191 | return (s64)(x - y) > 0 ? x : y; | |
192 | } | |
193 | ||
3e51f33f PZ |
194 | /* |
195 | * update the percpu scd from the raw @now value | |
196 | * | |
197 | * - filter out backward motion | |
354879bb | 198 | * - use the GTOD tick value to create a window to filter crazy TSC values |
3e51f33f | 199 | */ |
def0a9b2 | 200 | static u64 sched_clock_local(struct sched_clock_data *scd) |
3e51f33f | 201 | { |
def0a9b2 PZ |
202 | u64 now, clock, old_clock, min_clock, max_clock; |
203 | s64 delta; | |
3e51f33f | 204 | |
def0a9b2 PZ |
205 | again: |
206 | now = sched_clock(); | |
207 | delta = now - scd->tick_raw; | |
354879bb PZ |
208 | if (unlikely(delta < 0)) |
209 | delta = 0; | |
3e51f33f | 210 | |
def0a9b2 PZ |
211 | old_clock = scd->clock; |
212 | ||
354879bb PZ |
213 | /* |
214 | * scd->clock = clamp(scd->tick_gtod + delta, | |
b342501c IM |
215 | * max(scd->tick_gtod, scd->clock), |
216 | * scd->tick_gtod + TICK_NSEC); | |
354879bb | 217 | */ |
3e51f33f | 218 | |
354879bb | 219 | clock = scd->tick_gtod + delta; |
def0a9b2 PZ |
220 | min_clock = wrap_max(scd->tick_gtod, old_clock); |
221 | max_clock = wrap_max(old_clock, scd->tick_gtod + TICK_NSEC); | |
3e51f33f | 222 | |
354879bb PZ |
223 | clock = wrap_max(clock, min_clock); |
224 | clock = wrap_min(clock, max_clock); | |
3e51f33f | 225 | |
152f9d07 | 226 | if (cmpxchg64(&scd->clock, old_clock, clock) != old_clock) |
def0a9b2 | 227 | goto again; |
56b90612 | 228 | |
def0a9b2 | 229 | return clock; |
3e51f33f PZ |
230 | } |
231 | ||
def0a9b2 | 232 | static u64 sched_clock_remote(struct sched_clock_data *scd) |
3e51f33f | 233 | { |
def0a9b2 PZ |
234 | struct sched_clock_data *my_scd = this_scd(); |
235 | u64 this_clock, remote_clock; | |
236 | u64 *ptr, old_val, val; | |
237 | ||
a1cbcaa9 TG |
238 | #if BITS_PER_LONG != 64 |
239 | again: | |
240 | /* | |
241 | * Careful here: The local and the remote clock values need to | |
242 | * be read out atomic as we need to compare the values and | |
243 | * then update either the local or the remote side. So the | |
244 | * cmpxchg64 below only protects one readout. | |
245 | * | |
246 | * We must reread via sched_clock_local() in the retry case on | |
247 | * 32bit as an NMI could use sched_clock_local() via the | |
248 | * tracer and hit between the readout of | |
249 | * the low32bit and the high 32bit portion. | |
250 | */ | |
251 | this_clock = sched_clock_local(my_scd); | |
252 | /* | |
253 | * We must enforce atomic readout on 32bit, otherwise the | |
254 | * update on the remote cpu can hit inbetween the readout of | |
255 | * the low32bit and the high 32bit portion. | |
256 | */ | |
257 | remote_clock = cmpxchg64(&scd->clock, 0, 0); | |
258 | #else | |
259 | /* | |
260 | * On 64bit the read of [my]scd->clock is atomic versus the | |
261 | * update, so we can avoid the above 32bit dance. | |
262 | */ | |
def0a9b2 PZ |
263 | sched_clock_local(my_scd); |
264 | again: | |
265 | this_clock = my_scd->clock; | |
266 | remote_clock = scd->clock; | |
a1cbcaa9 | 267 | #endif |
def0a9b2 PZ |
268 | |
269 | /* | |
270 | * Use the opportunity that we have both locks | |
271 | * taken to couple the two clocks: we take the | |
272 | * larger time as the latest time for both | |
273 | * runqueues. (this creates monotonic movement) | |
274 | */ | |
275 | if (likely((s64)(remote_clock - this_clock) < 0)) { | |
276 | ptr = &scd->clock; | |
277 | old_val = remote_clock; | |
278 | val = this_clock; | |
3e51f33f | 279 | } else { |
def0a9b2 PZ |
280 | /* |
281 | * Should be rare, but possible: | |
282 | */ | |
283 | ptr = &my_scd->clock; | |
284 | old_val = this_clock; | |
285 | val = remote_clock; | |
3e51f33f | 286 | } |
def0a9b2 | 287 | |
152f9d07 | 288 | if (cmpxchg64(ptr, old_val, val) != old_val) |
def0a9b2 PZ |
289 | goto again; |
290 | ||
291 | return val; | |
3e51f33f PZ |
292 | } |
293 | ||
c676329a PZ |
294 | /* |
295 | * Similar to cpu_clock(), but requires local IRQs to be disabled. | |
296 | * | |
297 | * See cpu_clock(). | |
298 | */ | |
3e51f33f PZ |
299 | u64 sched_clock_cpu(int cpu) |
300 | { | |
b342501c | 301 | struct sched_clock_data *scd; |
def0a9b2 PZ |
302 | u64 clock; |
303 | ||
35af99e6 | 304 | if (sched_clock_stable()) |
b342501c | 305 | return sched_clock(); |
a381759d | 306 | |
a381759d PZ |
307 | if (unlikely(!sched_clock_running)) |
308 | return 0ull; | |
309 | ||
96b3d28b | 310 | preempt_disable_notrace(); |
def0a9b2 | 311 | scd = cpu_sdc(cpu); |
3e51f33f | 312 | |
def0a9b2 PZ |
313 | if (cpu != smp_processor_id()) |
314 | clock = sched_clock_remote(scd); | |
315 | else | |
316 | clock = sched_clock_local(scd); | |
96b3d28b | 317 | preempt_enable_notrace(); |
e4e4e534 | 318 | |
3e51f33f PZ |
319 | return clock; |
320 | } | |
321 | ||
322 | void sched_clock_tick(void) | |
323 | { | |
8325d9c0 | 324 | struct sched_clock_data *scd; |
3e51f33f PZ |
325 | u64 now, now_gtod; |
326 | ||
35af99e6 | 327 | if (sched_clock_stable()) |
8325d9c0 PZ |
328 | return; |
329 | ||
a381759d PZ |
330 | if (unlikely(!sched_clock_running)) |
331 | return; | |
332 | ||
3e51f33f PZ |
333 | WARN_ON_ONCE(!irqs_disabled()); |
334 | ||
8325d9c0 | 335 | scd = this_scd(); |
3e51f33f | 336 | now_gtod = ktime_to_ns(ktime_get()); |
a83bc47c | 337 | now = sched_clock(); |
3e51f33f | 338 | |
3e51f33f PZ |
339 | scd->tick_raw = now; |
340 | scd->tick_gtod = now_gtod; | |
def0a9b2 | 341 | sched_clock_local(scd); |
3e51f33f PZ |
342 | } |
343 | ||
344 | /* | |
345 | * We are going deep-idle (irqs are disabled): | |
346 | */ | |
347 | void sched_clock_idle_sleep_event(void) | |
348 | { | |
349 | sched_clock_cpu(smp_processor_id()); | |
350 | } | |
351 | EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event); | |
352 | ||
353 | /* | |
354 | * We just idled delta nanoseconds (called with irqs disabled): | |
355 | */ | |
356 | void sched_clock_idle_wakeup_event(u64 delta_ns) | |
357 | { | |
1c5745aa TG |
358 | if (timekeeping_suspended) |
359 | return; | |
360 | ||
354879bb | 361 | sched_clock_tick(); |
03e0d461 | 362 | touch_softlockup_watchdog_sched(); |
3e51f33f PZ |
363 | } |
364 | EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event); | |
365 | ||
c676329a PZ |
366 | /* |
367 | * As outlined at the top, provides a fast, high resolution, nanosecond | |
368 | * time source that is monotonic per cpu argument and has bounded drift | |
369 | * between cpus. | |
370 | * | |
371 | * ######################### BIG FAT WARNING ########################## | |
372 | * # when comparing cpu_clock(i) to cpu_clock(j) for i != j, time can # | |
373 | * # go backwards !! # | |
374 | * #################################################################### | |
375 | */ | |
376 | u64 cpu_clock(int cpu) | |
b9f8fcd5 | 377 | { |
d375b4e0 | 378 | if (!sched_clock_stable()) |
35af99e6 PZ |
379 | return sched_clock_cpu(cpu); |
380 | ||
381 | return sched_clock(); | |
b9f8fcd5 DM |
382 | } |
383 | ||
c676329a PZ |
384 | /* |
385 | * Similar to cpu_clock() for the current cpu. Time will only be observed | |
386 | * to be monotonic if care is taken to only compare timestampt taken on the | |
387 | * same CPU. | |
388 | * | |
389 | * See cpu_clock(). | |
390 | */ | |
391 | u64 local_clock(void) | |
392 | { | |
d375b4e0 | 393 | if (!sched_clock_stable()) |
35af99e6 PZ |
394 | return sched_clock_cpu(raw_smp_processor_id()); |
395 | ||
396 | return sched_clock(); | |
c676329a PZ |
397 | } |
398 | ||
8325d9c0 PZ |
399 | #else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */ |
400 | ||
401 | void sched_clock_init(void) | |
402 | { | |
403 | sched_clock_running = 1; | |
404 | } | |
405 | ||
406 | u64 sched_clock_cpu(int cpu) | |
407 | { | |
408 | if (unlikely(!sched_clock_running)) | |
409 | return 0; | |
410 | ||
411 | return sched_clock(); | |
412 | } | |
413 | ||
c676329a | 414 | u64 cpu_clock(int cpu) |
76a2a6ee | 415 | { |
35af99e6 | 416 | return sched_clock(); |
b9f8fcd5 | 417 | } |
76a2a6ee | 418 | |
c676329a PZ |
419 | u64 local_clock(void) |
420 | { | |
35af99e6 | 421 | return sched_clock(); |
c676329a PZ |
422 | } |
423 | ||
b9f8fcd5 | 424 | #endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */ |
76a2a6ee | 425 | |
4c9fe8ad | 426 | EXPORT_SYMBOL_GPL(cpu_clock); |
c676329a | 427 | EXPORT_SYMBOL_GPL(local_clock); |
545a2bf7 CB |
428 | |
429 | /* | |
430 | * Running clock - returns the time that has elapsed while a guest has been | |
431 | * running. | |
432 | * On a guest this value should be local_clock minus the time the guest was | |
433 | * suspended by the hypervisor (for any reason). | |
434 | * On bare metal this function should return the same as local_clock. | |
435 | * Architectures and sub-architectures can override this. | |
436 | */ | |
437 | u64 __weak running_clock(void) | |
438 | { | |
439 | return local_clock(); | |
440 | } |