1 /*
2 * linux/kernel/time/timekeeping.c
3 *
4 * Kernel timekeeping code and accessor functions
5 *
6 * This code was moved from linux/kernel/timer.c.
7 * Please see that file for copyright and history logs.
8 *
9 */
10
11 #include <linux/module.h>
12 #include <linux/interrupt.h>
13 #include <linux/percpu.h>
14 #include <linux/init.h>
15 #include <linux/mm.h>
16 #include <linux/sysdev.h>
17 #include <linux/clocksource.h>
18 #include <linux/jiffies.h>
19 #include <linux/time.h>
20 #include <linux/tick.h>
21
22
23 /*
24 * This read-write spinlock protects us from races in SMP while
25 * playing with xtime and avenrun.
26 */
27 __cacheline_aligned_in_smp DEFINE_SEQLOCK(xtime_lock);
28
29
30 /*
31 * The current time
32 * wall_to_monotonic is what we need to add to xtime (or xtime corrected
33 * for sub jiffie times) to get to monotonic time. Monotonic is pegged
34 * at zero at system boot time, so wall_to_monotonic will be negative,
35 * however, we will ALWAYS keep the tv_nsec part positive so we can use
36 * the usual normalization.
37 *
38 * wall_to_monotonic is moved after resume from suspend for the monotonic
39 * time not to jump. We need to add total_sleep_time to wall_to_monotonic
40 * to get the real boot based time offset.
41 *
42 * - wall_to_monotonic is no longer the boot time, getboottime must be
43 * used instead.
44 */
45 struct timespec xtime __attribute__ ((aligned (16)));
46 struct timespec wall_to_monotonic __attribute__ ((aligned (16)));
47 static unsigned long total_sleep_time; /* seconds */
48
49 /* flag for if timekeeping is suspended */
50 int __read_mostly timekeeping_suspended;
51
52 static struct timespec xtime_cache __attribute__ ((aligned (16)));
update_xtime_cache(u64 nsec)53 void update_xtime_cache(u64 nsec)
54 {
55 xtime_cache = xtime;
56 timespec_add_ns(&xtime_cache, nsec);
57 }
58
59 struct clocksource *clock;
60
61
62 #ifdef CONFIG_GENERIC_TIME
63 /**
64 * clocksource_forward_now - update clock to the current time
65 *
66 * Forward the current clock to update its state since the last call to
67 * update_wall_time(). This is useful before significant clock changes,
68 * as it avoids having to deal with this time offset explicitly.
69 */
clocksource_forward_now(void)70 static void clocksource_forward_now(void)
71 {
72 cycle_t cycle_now, cycle_delta;
73 s64 nsec;
74
75 cycle_now = clocksource_read(clock);
76 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
77 clock->cycle_last = cycle_now;
78
79 nsec = cyc2ns(clock, cycle_delta);
80 timespec_add_ns(&xtime, nsec);
81
82 nsec = ((s64)cycle_delta * clock->mult_orig) >> clock->shift;
83 clock->raw_time.tv_nsec += nsec;
84 }
85
86 /**
87 * getnstimeofday - Returns the time of day in a timespec
88 * @ts: pointer to the timespec to be set
89 *
90 * Returns the time of day in a timespec.
91 */
getnstimeofday(struct timespec * ts)92 void getnstimeofday(struct timespec *ts)
93 {
94 cycle_t cycle_now, cycle_delta;
95 unsigned long seq;
96 s64 nsecs;
97
98 WARN_ON(timekeeping_suspended);
99
100 do {
101 seq = read_seqbegin(&xtime_lock);
102
103 *ts = xtime;
104
105 /* read clocksource: */
106 cycle_now = clocksource_read(clock);
107
108 /* calculate the delta since the last update_wall_time: */
109 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
110
111 /* convert to nanoseconds: */
112 nsecs = cyc2ns(clock, cycle_delta);
113
114 } while (read_seqretry(&xtime_lock, seq));
115
116 timespec_add_ns(ts, nsecs);
117 }
118
119 EXPORT_SYMBOL(getnstimeofday);
120
121 /**
122 * do_gettimeofday - Returns the time of day in a timeval
123 * @tv: pointer to the timeval to be set
124 *
125 * NOTE: Users should be converted to using getnstimeofday()
126 */
do_gettimeofday(struct timeval * tv)127 void do_gettimeofday(struct timeval *tv)
128 {
129 struct timespec now;
130
131 getnstimeofday(&now);
132 tv->tv_sec = now.tv_sec;
133 tv->tv_usec = now.tv_nsec/1000;
134 }
135
136 EXPORT_SYMBOL(do_gettimeofday);
137 /**
138 * do_settimeofday - Sets the time of day
139 * @tv: pointer to the timespec variable containing the new time
140 *
141 * Sets the time of day to the new time and update NTP and notify hrtimers
142 */
do_settimeofday(struct timespec * tv)143 int do_settimeofday(struct timespec *tv)
144 {
145 struct timespec ts_delta;
146 unsigned long flags;
147
148 if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
149 return -EINVAL;
150
151 write_seqlock_irqsave(&xtime_lock, flags);
152
153 clocksource_forward_now();
154
155 ts_delta.tv_sec = tv->tv_sec - xtime.tv_sec;
156 ts_delta.tv_nsec = tv->tv_nsec - xtime.tv_nsec;
157 wall_to_monotonic = timespec_sub(wall_to_monotonic, ts_delta);
158
159 xtime = *tv;
160
161 update_xtime_cache(0);
162
163 clock->error = 0;
164 ntp_clear();
165
166 update_vsyscall(&xtime, clock);
167
168 write_sequnlock_irqrestore(&xtime_lock, flags);
169
170 /* signal hrtimers about time change */
171 clock_was_set();
172
173 return 0;
174 }
175
176 EXPORT_SYMBOL(do_settimeofday);
177
178 /**
179 * change_clocksource - Swaps clocksources if a new one is available
180 *
181 * Accumulates current time interval and initializes new clocksource
182 */
change_clocksource(void)183 static void change_clocksource(void)
184 {
185 struct clocksource *new;
186
187 new = clocksource_get_next();
188
189 if (clock == new)
190 return;
191
192 clocksource_forward_now();
193
194 new->raw_time = clock->raw_time;
195
196 clock = new;
197 clock->cycle_last = 0;
198 clock->cycle_last = clocksource_read(new);
199 clock->error = 0;
200 clock->xtime_nsec = 0;
201 clocksource_calculate_interval(clock, NTP_INTERVAL_LENGTH);
202
203 tick_clock_notify();
204
205 /*
206 * We're holding xtime lock and waking up klogd would deadlock
207 * us on enqueue. So no printing!
208 printk(KERN_INFO "Time: %s clocksource has been installed.\n",
209 clock->name);
210 */
211 }
212 #else
clocksource_forward_now(void)213 static inline void clocksource_forward_now(void) { }
change_clocksource(void)214 static inline void change_clocksource(void) { }
215 #endif
216
217 /**
218 * getrawmonotonic - Returns the raw monotonic time in a timespec
219 * @ts: pointer to the timespec to be set
220 *
221 * Returns the raw monotonic time (completely un-modified by ntp)
222 */
getrawmonotonic(struct timespec * ts)223 void getrawmonotonic(struct timespec *ts)
224 {
225 unsigned long seq;
226 s64 nsecs;
227 cycle_t cycle_now, cycle_delta;
228
229 do {
230 seq = read_seqbegin(&xtime_lock);
231
232 /* read clocksource: */
233 cycle_now = clocksource_read(clock);
234
235 /* calculate the delta since the last update_wall_time: */
236 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
237
238 /* convert to nanoseconds: */
239 nsecs = ((s64)cycle_delta * clock->mult_orig) >> clock->shift;
240
241 *ts = clock->raw_time;
242
243 } while (read_seqretry(&xtime_lock, seq));
244
245 timespec_add_ns(ts, nsecs);
246 }
247 EXPORT_SYMBOL(getrawmonotonic);
248
249
250 /**
251 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
252 */
timekeeping_valid_for_hres(void)253 int timekeeping_valid_for_hres(void)
254 {
255 unsigned long seq;
256 int ret;
257
258 do {
259 seq = read_seqbegin(&xtime_lock);
260
261 ret = clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
262
263 } while (read_seqretry(&xtime_lock, seq));
264
265 return ret;
266 }
267
268 /**
269 * read_persistent_clock - Return time in seconds from the persistent clock.
270 *
271 * Weak dummy function for arches that do not yet support it.
272 * Returns seconds from epoch using the battery backed persistent clock.
273 * Returns zero if unsupported.
274 *
275 * XXX - Do be sure to remove it once all arches implement it.
276 */
read_persistent_clock(void)277 unsigned long __attribute__((weak)) read_persistent_clock(void)
278 {
279 return 0;
280 }
281
282 /*
283 * timekeeping_init - Initializes the clocksource and common timekeeping values
284 */
timekeeping_init(void)285 void __init timekeeping_init(void)
286 {
287 unsigned long flags;
288 unsigned long sec = read_persistent_clock();
289
290 write_seqlock_irqsave(&xtime_lock, flags);
291
292 ntp_init();
293
294 clock = clocksource_get_next();
295 clocksource_calculate_interval(clock, NTP_INTERVAL_LENGTH);
296 clock->cycle_last = clocksource_read(clock);
297
298 xtime.tv_sec = sec;
299 xtime.tv_nsec = 0;
300 set_normalized_timespec(&wall_to_monotonic,
301 -xtime.tv_sec, -xtime.tv_nsec);
302 update_xtime_cache(0);
303 total_sleep_time = 0;
304 write_sequnlock_irqrestore(&xtime_lock, flags);
305 }
306
307 /* time in seconds when suspend began */
308 static unsigned long timekeeping_suspend_time;
309
310 /**
311 * timekeeping_resume - Resumes the generic timekeeping subsystem.
312 * @dev: unused
313 *
314 * This is for the generic clocksource timekeeping.
315 * xtime/wall_to_monotonic/jiffies/etc are
316 * still managed by arch specific suspend/resume code.
317 */
timekeeping_resume(struct sys_device * dev)318 static int timekeeping_resume(struct sys_device *dev)
319 {
320 unsigned long flags;
321 unsigned long now = read_persistent_clock();
322
323 clocksource_resume();
324
325 write_seqlock_irqsave(&xtime_lock, flags);
326
327 if (now && (now > timekeeping_suspend_time)) {
328 unsigned long sleep_length = now - timekeeping_suspend_time;
329
330 xtime.tv_sec += sleep_length;
331 wall_to_monotonic.tv_sec -= sleep_length;
332 total_sleep_time += sleep_length;
333 }
334 update_xtime_cache(0);
335 /* re-base the last cycle value */
336 clock->cycle_last = 0;
337 clock->cycle_last = clocksource_read(clock);
338 clock->error = 0;
339 timekeeping_suspended = 0;
340 write_sequnlock_irqrestore(&xtime_lock, flags);
341
342 touch_softlockup_watchdog();
343
344 clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
345
346 /* Resume hrtimers */
347 hres_timers_resume();
348
349 return 0;
350 }
351
timekeeping_suspend(struct sys_device * dev,pm_message_t state)352 static int timekeeping_suspend(struct sys_device *dev, pm_message_t state)
353 {
354 unsigned long flags;
355
356 timekeeping_suspend_time = read_persistent_clock();
357
358 write_seqlock_irqsave(&xtime_lock, flags);
359 clocksource_forward_now();
360 timekeeping_suspended = 1;
361 write_sequnlock_irqrestore(&xtime_lock, flags);
362
363 clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
364
365 return 0;
366 }
367
368 /* sysfs resume/suspend bits for timekeeping */
369 static struct sysdev_class timekeeping_sysclass = {
370 .name = "timekeeping",
371 .resume = timekeeping_resume,
372 .suspend = timekeeping_suspend,
373 };
374
375 static struct sys_device device_timer = {
376 .id = 0,
377 .cls = &timekeeping_sysclass,
378 };
379
timekeeping_init_device(void)380 static int __init timekeeping_init_device(void)
381 {
382 int error = sysdev_class_register(&timekeeping_sysclass);
383 if (!error)
384 error = sysdev_register(&device_timer);
385 return error;
386 }
387
388 device_initcall(timekeeping_init_device);
389
390 /*
391 * If the error is already larger, we look ahead even further
392 * to compensate for late or lost adjustments.
393 */
clocksource_bigadjust(s64 error,s64 * interval,s64 * offset)394 static __always_inline int clocksource_bigadjust(s64 error, s64 *interval,
395 s64 *offset)
396 {
397 s64 tick_error, i;
398 u32 look_ahead, adj;
399 s32 error2, mult;
400
401 /*
402 * Use the current error value to determine how much to look ahead.
403 * The larger the error the slower we adjust for it to avoid problems
404 * with losing too many ticks, otherwise we would overadjust and
405 * produce an even larger error. The smaller the adjustment the
406 * faster we try to adjust for it, as lost ticks can do less harm
407 * here. This is tuned so that an error of about 1 msec is adjusted
408 * within about 1 sec (or 2^20 nsec in 2^SHIFT_HZ ticks).
409 */
410 error2 = clock->error >> (NTP_SCALE_SHIFT + 22 - 2 * SHIFT_HZ);
411 error2 = abs(error2);
412 for (look_ahead = 0; error2 > 0; look_ahead++)
413 error2 >>= 2;
414
415 /*
416 * Now calculate the error in (1 << look_ahead) ticks, but first
417 * remove the single look ahead already included in the error.
418 */
419 tick_error = tick_length >> (NTP_SCALE_SHIFT - clock->shift + 1);
420 tick_error -= clock->xtime_interval >> 1;
421 error = ((error - tick_error) >> look_ahead) + tick_error;
422
423 /* Finally calculate the adjustment shift value. */
424 i = *interval;
425 mult = 1;
426 if (error < 0) {
427 error = -error;
428 *interval = -*interval;
429 *offset = -*offset;
430 mult = -1;
431 }
432 for (adj = 0; error > i; adj++)
433 error >>= 1;
434
435 *interval <<= adj;
436 *offset <<= adj;
437 return mult << adj;
438 }
439
440 /*
441 * Adjust the multiplier to reduce the error value,
442 * this is optimized for the most common adjustments of -1,0,1,
443 * for other values we can do a bit more work.
444 */
clocksource_adjust(s64 offset)445 static void clocksource_adjust(s64 offset)
446 {
447 s64 error, interval = clock->cycle_interval;
448 int adj;
449
450 error = clock->error >> (NTP_SCALE_SHIFT - clock->shift - 1);
451 if (error > interval) {
452 error >>= 2;
453 if (likely(error <= interval))
454 adj = 1;
455 else
456 adj = clocksource_bigadjust(error, &interval, &offset);
457 } else if (error < -interval) {
458 error >>= 2;
459 if (likely(error >= -interval)) {
460 adj = -1;
461 interval = -interval;
462 offset = -offset;
463 } else
464 adj = clocksource_bigadjust(error, &interval, &offset);
465 } else
466 return;
467
468 clock->mult += adj;
469 clock->xtime_interval += interval;
470 clock->xtime_nsec -= offset;
471 clock->error -= (interval - offset) <<
472 (NTP_SCALE_SHIFT - clock->shift);
473 }
474
475 /**
476 * update_wall_time - Uses the current clocksource to increment the wall time
477 *
478 * Called from the timer interrupt, must hold a write on xtime_lock.
479 */
update_wall_time(void)480 void update_wall_time(void)
481 {
482 cycle_t offset;
483
484 /* Make sure we're fully resumed: */
485 if (unlikely(timekeeping_suspended))
486 return;
487
488 #ifdef CONFIG_GENERIC_TIME
489 offset = (clocksource_read(clock) - clock->cycle_last) & clock->mask;
490 #else
491 offset = clock->cycle_interval;
492 #endif
493 clock->xtime_nsec = (s64)xtime.tv_nsec << clock->shift;
494
495 /* normally this loop will run just once, however in the
496 * case of lost or late ticks, it will accumulate correctly.
497 */
498 while (offset >= clock->cycle_interval) {
499 /* accumulate one interval */
500 offset -= clock->cycle_interval;
501 clock->cycle_last += clock->cycle_interval;
502
503 clock->xtime_nsec += clock->xtime_interval;
504 if (clock->xtime_nsec >= (u64)NSEC_PER_SEC << clock->shift) {
505 clock->xtime_nsec -= (u64)NSEC_PER_SEC << clock->shift;
506 xtime.tv_sec++;
507 second_overflow();
508 }
509
510 clock->raw_time.tv_nsec += clock->raw_interval;
511 if (clock->raw_time.tv_nsec >= NSEC_PER_SEC) {
512 clock->raw_time.tv_nsec -= NSEC_PER_SEC;
513 clock->raw_time.tv_sec++;
514 }
515
516 /* accumulate error between NTP and clock interval */
517 clock->error += tick_length;
518 clock->error -= clock->xtime_interval << (NTP_SCALE_SHIFT - clock->shift);
519 }
520
521 /* correct the clock when NTP error is too big */
522 clocksource_adjust(offset);
523
524 /*
525 * Since in the loop above, we accumulate any amount of time
526 * in xtime_nsec over a second into xtime.tv_sec, its possible for
527 * xtime_nsec to be fairly small after the loop. Further, if we're
528 * slightly speeding the clocksource up in clocksource_adjust(),
529 * its possible the required corrective factor to xtime_nsec could
530 * cause it to underflow.
531 *
532 * Now, we cannot simply roll the accumulated second back, since
533 * the NTP subsystem has been notified via second_overflow. So
534 * instead we push xtime_nsec forward by the amount we underflowed,
535 * and add that amount into the error.
536 *
537 * We'll correct this error next time through this function, when
538 * xtime_nsec is not as small.
539 */
540 if (unlikely((s64)clock->xtime_nsec < 0)) {
541 s64 neg = -(s64)clock->xtime_nsec;
542 clock->xtime_nsec = 0;
543 clock->error += neg << (NTP_SCALE_SHIFT - clock->shift);
544 }
545
546 /* store full nanoseconds into xtime after rounding it up and
547 * add the remainder to the error difference.
548 */
549 xtime.tv_nsec = ((s64)clock->xtime_nsec >> clock->shift) + 1;
550 clock->xtime_nsec -= (s64)xtime.tv_nsec << clock->shift;
551 clock->error += clock->xtime_nsec << (NTP_SCALE_SHIFT - clock->shift);
552
553 update_xtime_cache(cyc2ns(clock, offset));
554
555 /* check to see if there is a new clocksource to use */
556 change_clocksource();
557 update_vsyscall(&xtime, clock);
558 }
559
560 /**
561 * getboottime - Return the real time of system boot.
562 * @ts: pointer to the timespec to be set
563 *
564 * Returns the time of day in a timespec.
565 *
566 * This is based on the wall_to_monotonic offset and the total suspend
567 * time. Calls to settimeofday will affect the value returned (which
568 * basically means that however wrong your real time clock is at boot time,
569 * you get the right time here).
570 */
getboottime(struct timespec * ts)571 void getboottime(struct timespec *ts)
572 {
573 set_normalized_timespec(ts,
574 - (wall_to_monotonic.tv_sec + total_sleep_time),
575 - wall_to_monotonic.tv_nsec);
576 }
577
578 /**
579 * monotonic_to_bootbased - Convert the monotonic time to boot based.
580 * @ts: pointer to the timespec to be converted
581 */
monotonic_to_bootbased(struct timespec * ts)582 void monotonic_to_bootbased(struct timespec *ts)
583 {
584 ts->tv_sec += total_sleep_time;
585 }
586
get_seconds(void)587 unsigned long get_seconds(void)
588 {
589 return xtime_cache.tv_sec;
590 }
591 EXPORT_SYMBOL(get_seconds);
592
593
current_kernel_time(void)594 struct timespec current_kernel_time(void)
595 {
596 struct timespec now;
597 unsigned long seq;
598
599 do {
600 seq = read_seqbegin(&xtime_lock);
601
602 now = xtime_cache;
603 } while (read_seqretry(&xtime_lock, seq));
604
605 return now;
606 }
607 EXPORT_SYMBOL(current_kernel_time);
608