]>
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 | |
9881b024 PZ |
80 | void sched_clock_init(void) |
81 | { | |
82 | sched_clock_running = 1; | |
83 | } | |
84 | ||
3e51f33f | 85 | #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK |
acb04058 PZ |
86 | /* |
87 | * We must start with !__sched_clock_stable because the unstable -> stable | |
88 | * transition is accurate, while the stable -> unstable transition is not. | |
89 | * | |
90 | * Similarly we start with __sched_clock_stable_early, thereby assuming we | |
91 | * will become stable, such that there's only a single 1 -> 0 transition. | |
92 | */ | |
555570d7 | 93 | static DEFINE_STATIC_KEY_FALSE(__sched_clock_stable); |
acb04058 | 94 | static int __sched_clock_stable_early = 1; |
35af99e6 | 95 | |
5680d809 PZ |
96 | /* |
97 | * We want: ktime_get_ns() + gtod_offset == sched_clock() + raw_offset | |
98 | */ | |
99 | static __read_mostly u64 raw_offset; | |
100 | static __read_mostly u64 gtod_offset; | |
101 | ||
102 | struct sched_clock_data { | |
103 | u64 tick_raw; | |
104 | u64 tick_gtod; | |
105 | u64 clock; | |
106 | }; | |
107 | ||
108 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data); | |
109 | ||
110 | static inline struct sched_clock_data *this_scd(void) | |
111 | { | |
112 | return this_cpu_ptr(&sched_clock_data); | |
113 | } | |
114 | ||
115 | static inline struct sched_clock_data *cpu_sdc(int cpu) | |
116 | { | |
117 | return &per_cpu(sched_clock_data, cpu); | |
118 | } | |
119 | ||
35af99e6 PZ |
120 | int sched_clock_stable(void) |
121 | { | |
555570d7 | 122 | return static_branch_likely(&__sched_clock_stable); |
35af99e6 PZ |
123 | } |
124 | ||
d375b4e0 | 125 | static void __set_sched_clock_stable(void) |
35af99e6 | 126 | { |
5680d809 PZ |
127 | struct sched_clock_data *scd = this_scd(); |
128 | ||
129 | /* | |
130 | * Attempt to make the (initial) unstable->stable transition continuous. | |
131 | */ | |
132 | raw_offset = (scd->tick_gtod + gtod_offset) - (scd->tick_raw); | |
133 | ||
134 | printk(KERN_INFO "sched_clock: Marking stable (%lld, %lld)->(%lld, %lld)\n", | |
135 | scd->tick_gtod, gtod_offset, | |
136 | scd->tick_raw, raw_offset); | |
137 | ||
555570d7 | 138 | static_branch_enable(&__sched_clock_stable); |
4f49b90a | 139 | tick_dep_clear(TICK_DEP_BIT_CLOCK_UNSTABLE); |
d375b4e0 PZ |
140 | } |
141 | ||
6577e42a | 142 | static void __clear_sched_clock_stable(struct work_struct *work) |
35af99e6 | 143 | { |
5680d809 PZ |
144 | struct sched_clock_data *scd = this_scd(); |
145 | ||
146 | /* | |
147 | * Attempt to make the stable->unstable transition continuous. | |
148 | * | |
149 | * Trouble is, this is typically called from the TSC watchdog | |
150 | * timer, which is late per definition. This means the tick | |
151 | * values can already be screwy. | |
152 | * | |
153 | * Still do what we can. | |
154 | */ | |
155 | gtod_offset = (scd->tick_raw + raw_offset) - (scd->tick_gtod); | |
156 | ||
157 | printk(KERN_INFO "sched_clock: Marking unstable (%lld, %lld)<-(%lld, %lld)\n", | |
158 | scd->tick_gtod, gtod_offset, | |
159 | scd->tick_raw, raw_offset); | |
160 | ||
555570d7 | 161 | static_branch_disable(&__sched_clock_stable); |
4f49b90a | 162 | tick_dep_set(TICK_DEP_BIT_CLOCK_UNSTABLE); |
35af99e6 | 163 | } |
3e51f33f | 164 | |
6577e42a PZ |
165 | static DECLARE_WORK(sched_clock_work, __clear_sched_clock_stable); |
166 | ||
167 | void clear_sched_clock_stable(void) | |
168 | { | |
d375b4e0 PZ |
169 | __sched_clock_stable_early = 0; |
170 | ||
9881b024 | 171 | smp_mb(); /* matches sched_clock_init_late() */ |
d375b4e0 | 172 | |
9881b024 PZ |
173 | if (sched_clock_running == 2) |
174 | schedule_work(&sched_clock_work); | |
6577e42a PZ |
175 | } |
176 | ||
9881b024 | 177 | void sched_clock_init_late(void) |
3e51f33f | 178 | { |
9881b024 | 179 | sched_clock_running = 2; |
d375b4e0 PZ |
180 | /* |
181 | * Ensure that it is impossible to not do a static_key update. | |
182 | * | |
183 | * Either {set,clear}_sched_clock_stable() must see sched_clock_running | |
184 | * and do the update, or we must see their __sched_clock_stable_early | |
185 | * and do the update, or both. | |
186 | */ | |
187 | smp_mb(); /* matches {set,clear}_sched_clock_stable() */ | |
188 | ||
189 | if (__sched_clock_stable_early) | |
190 | __set_sched_clock_stable(); | |
3e51f33f PZ |
191 | } |
192 | ||
354879bb | 193 | /* |
b342501c | 194 | * min, max except they take wrapping into account |
354879bb PZ |
195 | */ |
196 | ||
197 | static inline u64 wrap_min(u64 x, u64 y) | |
198 | { | |
199 | return (s64)(x - y) < 0 ? x : y; | |
200 | } | |
201 | ||
202 | static inline u64 wrap_max(u64 x, u64 y) | |
203 | { | |
204 | return (s64)(x - y) > 0 ? x : y; | |
205 | } | |
206 | ||
3e51f33f PZ |
207 | /* |
208 | * update the percpu scd from the raw @now value | |
209 | * | |
210 | * - filter out backward motion | |
354879bb | 211 | * - use the GTOD tick value to create a window to filter crazy TSC values |
3e51f33f | 212 | */ |
def0a9b2 | 213 | static u64 sched_clock_local(struct sched_clock_data *scd) |
3e51f33f | 214 | { |
def0a9b2 PZ |
215 | u64 now, clock, old_clock, min_clock, max_clock; |
216 | s64 delta; | |
3e51f33f | 217 | |
def0a9b2 PZ |
218 | again: |
219 | now = sched_clock(); | |
220 | delta = now - scd->tick_raw; | |
354879bb PZ |
221 | if (unlikely(delta < 0)) |
222 | delta = 0; | |
3e51f33f | 223 | |
def0a9b2 PZ |
224 | old_clock = scd->clock; |
225 | ||
354879bb PZ |
226 | /* |
227 | * scd->clock = clamp(scd->tick_gtod + delta, | |
b342501c IM |
228 | * max(scd->tick_gtod, scd->clock), |
229 | * scd->tick_gtod + TICK_NSEC); | |
354879bb | 230 | */ |
3e51f33f | 231 | |
5680d809 | 232 | clock = scd->tick_gtod + gtod_offset + delta; |
def0a9b2 PZ |
233 | min_clock = wrap_max(scd->tick_gtod, old_clock); |
234 | max_clock = wrap_max(old_clock, scd->tick_gtod + TICK_NSEC); | |
3e51f33f | 235 | |
354879bb PZ |
236 | clock = wrap_max(clock, min_clock); |
237 | clock = wrap_min(clock, max_clock); | |
3e51f33f | 238 | |
152f9d07 | 239 | if (cmpxchg64(&scd->clock, old_clock, clock) != old_clock) |
def0a9b2 | 240 | goto again; |
56b90612 | 241 | |
def0a9b2 | 242 | return clock; |
3e51f33f PZ |
243 | } |
244 | ||
def0a9b2 | 245 | static u64 sched_clock_remote(struct sched_clock_data *scd) |
3e51f33f | 246 | { |
def0a9b2 PZ |
247 | struct sched_clock_data *my_scd = this_scd(); |
248 | u64 this_clock, remote_clock; | |
249 | u64 *ptr, old_val, val; | |
250 | ||
a1cbcaa9 TG |
251 | #if BITS_PER_LONG != 64 |
252 | again: | |
253 | /* | |
254 | * Careful here: The local and the remote clock values need to | |
255 | * be read out atomic as we need to compare the values and | |
256 | * then update either the local or the remote side. So the | |
257 | * cmpxchg64 below only protects one readout. | |
258 | * | |
259 | * We must reread via sched_clock_local() in the retry case on | |
260 | * 32bit as an NMI could use sched_clock_local() via the | |
261 | * tracer and hit between the readout of | |
262 | * the low32bit and the high 32bit portion. | |
263 | */ | |
264 | this_clock = sched_clock_local(my_scd); | |
265 | /* | |
266 | * We must enforce atomic readout on 32bit, otherwise the | |
267 | * update on the remote cpu can hit inbetween the readout of | |
268 | * the low32bit and the high 32bit portion. | |
269 | */ | |
270 | remote_clock = cmpxchg64(&scd->clock, 0, 0); | |
271 | #else | |
272 | /* | |
273 | * On 64bit the read of [my]scd->clock is atomic versus the | |
274 | * update, so we can avoid the above 32bit dance. | |
275 | */ | |
def0a9b2 PZ |
276 | sched_clock_local(my_scd); |
277 | again: | |
278 | this_clock = my_scd->clock; | |
279 | remote_clock = scd->clock; | |
a1cbcaa9 | 280 | #endif |
def0a9b2 PZ |
281 | |
282 | /* | |
283 | * Use the opportunity that we have both locks | |
284 | * taken to couple the two clocks: we take the | |
285 | * larger time as the latest time for both | |
286 | * runqueues. (this creates monotonic movement) | |
287 | */ | |
288 | if (likely((s64)(remote_clock - this_clock) < 0)) { | |
289 | ptr = &scd->clock; | |
290 | old_val = remote_clock; | |
291 | val = this_clock; | |
3e51f33f | 292 | } else { |
def0a9b2 PZ |
293 | /* |
294 | * Should be rare, but possible: | |
295 | */ | |
296 | ptr = &my_scd->clock; | |
297 | old_val = this_clock; | |
298 | val = remote_clock; | |
3e51f33f | 299 | } |
def0a9b2 | 300 | |
152f9d07 | 301 | if (cmpxchg64(ptr, old_val, val) != old_val) |
def0a9b2 PZ |
302 | goto again; |
303 | ||
304 | return val; | |
3e51f33f PZ |
305 | } |
306 | ||
c676329a PZ |
307 | /* |
308 | * Similar to cpu_clock(), but requires local IRQs to be disabled. | |
309 | * | |
310 | * See cpu_clock(). | |
311 | */ | |
3e51f33f PZ |
312 | u64 sched_clock_cpu(int cpu) |
313 | { | |
b342501c | 314 | struct sched_clock_data *scd; |
def0a9b2 PZ |
315 | u64 clock; |
316 | ||
35af99e6 | 317 | if (sched_clock_stable()) |
5680d809 | 318 | return sched_clock() + raw_offset; |
a381759d | 319 | |
a381759d PZ |
320 | if (unlikely(!sched_clock_running)) |
321 | return 0ull; | |
322 | ||
96b3d28b | 323 | preempt_disable_notrace(); |
def0a9b2 | 324 | scd = cpu_sdc(cpu); |
3e51f33f | 325 | |
def0a9b2 PZ |
326 | if (cpu != smp_processor_id()) |
327 | clock = sched_clock_remote(scd); | |
328 | else | |
329 | clock = sched_clock_local(scd); | |
96b3d28b | 330 | preempt_enable_notrace(); |
e4e4e534 | 331 | |
3e51f33f PZ |
332 | return clock; |
333 | } | |
2c923e94 | 334 | EXPORT_SYMBOL_GPL(sched_clock_cpu); |
3e51f33f PZ |
335 | |
336 | void sched_clock_tick(void) | |
337 | { | |
8325d9c0 | 338 | struct sched_clock_data *scd; |
a381759d | 339 | |
3e51f33f PZ |
340 | WARN_ON_ONCE(!irqs_disabled()); |
341 | ||
5680d809 PZ |
342 | /* |
343 | * Update these values even if sched_clock_stable(), because it can | |
344 | * become unstable at any point in time at which point we need some | |
345 | * values to fall back on. | |
346 | * | |
347 | * XXX arguably we can skip this if we expose tsc_clocksource_reliable | |
348 | */ | |
8325d9c0 | 349 | scd = this_scd(); |
5680d809 PZ |
350 | scd->tick_raw = sched_clock(); |
351 | scd->tick_gtod = ktime_get_ns(); | |
3e51f33f | 352 | |
5680d809 PZ |
353 | if (!sched_clock_stable() && likely(sched_clock_running)) |
354 | sched_clock_local(scd); | |
3e51f33f PZ |
355 | } |
356 | ||
357 | /* | |
358 | * We are going deep-idle (irqs are disabled): | |
359 | */ | |
360 | void sched_clock_idle_sleep_event(void) | |
361 | { | |
362 | sched_clock_cpu(smp_processor_id()); | |
363 | } | |
364 | EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event); | |
365 | ||
366 | /* | |
367 | * We just idled delta nanoseconds (called with irqs disabled): | |
368 | */ | |
369 | void sched_clock_idle_wakeup_event(u64 delta_ns) | |
370 | { | |
1c5745aa TG |
371 | if (timekeeping_suspended) |
372 | return; | |
373 | ||
354879bb | 374 | sched_clock_tick(); |
03e0d461 | 375 | touch_softlockup_watchdog_sched(); |
3e51f33f PZ |
376 | } |
377 | EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event); | |
378 | ||
8325d9c0 PZ |
379 | #else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */ |
380 | ||
8325d9c0 PZ |
381 | u64 sched_clock_cpu(int cpu) |
382 | { | |
383 | if (unlikely(!sched_clock_running)) | |
384 | return 0; | |
385 | ||
386 | return sched_clock(); | |
387 | } | |
9881b024 | 388 | |
b9f8fcd5 | 389 | #endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */ |
76a2a6ee | 390 | |
545a2bf7 CB |
391 | /* |
392 | * Running clock - returns the time that has elapsed while a guest has been | |
393 | * running. | |
394 | * On a guest this value should be local_clock minus the time the guest was | |
395 | * suspended by the hypervisor (for any reason). | |
396 | * On bare metal this function should return the same as local_clock. | |
397 | * Architectures and sub-architectures can override this. | |
398 | */ | |
399 | u64 __weak running_clock(void) | |
400 | { | |
401 | return local_clock(); | |
402 | } |