Merge branch 'linus' into x86/paravirt-spinlocks
[linux-2.6] / kernel / time / timekeeping.c
CommitLineData
8524070b 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 */
ba2a631b 27__cacheline_aligned_in_smp DEFINE_SEQLOCK(xtime_lock);
8524070b 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.
7c3f1a57
TJ
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.
8524070b 44 */
45struct timespec xtime __attribute__ ((aligned (16)));
46struct timespec wall_to_monotonic __attribute__ ((aligned (16)));
7c3f1a57 47static unsigned long total_sleep_time; /* seconds */
8524070b 48
17c38b74 49static struct timespec xtime_cache __attribute__ ((aligned (16)));
1001d0a9 50void update_xtime_cache(u64 nsec)
17c38b74 51{
52 xtime_cache = xtime;
53 timespec_add_ns(&xtime_cache, nsec);
54}
17c38b74 55
7dffa3c6 56struct clocksource *clock;
8524070b 57
58
59#ifdef CONFIG_GENERIC_TIME
60/**
61 * __get_nsec_offset - Returns nanoseconds since last call to periodic_hook
62 *
63 * private function, must hold xtime_lock lock when being
64 * called. Returns the number of nanoseconds since the
65 * last call to update_wall_time() (adjusted by NTP scaling)
66 */
67static inline s64 __get_nsec_offset(void)
68{
69 cycle_t cycle_now, cycle_delta;
70 s64 ns_offset;
71
72 /* read clocksource: */
73 cycle_now = clocksource_read(clock);
74
75 /* calculate the delta since the last update_wall_time: */
76 cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
77
78 /* convert to nanoseconds: */
79 ns_offset = cyc2ns(clock, cycle_delta);
80
81 return ns_offset;
82}
83
84/**
efd9ac86 85 * getnstimeofday - Returns the time of day in a timespec
8524070b 86 * @ts: pointer to the timespec to be set
87 *
efd9ac86 88 * Returns the time of day in a timespec.
8524070b 89 */
efd9ac86 90void getnstimeofday(struct timespec *ts)
8524070b 91{
92 unsigned long seq;
93 s64 nsecs;
94
95 do {
96 seq = read_seqbegin(&xtime_lock);
97
98 *ts = xtime;
99 nsecs = __get_nsec_offset();
100
101 } while (read_seqretry(&xtime_lock, seq));
102
103 timespec_add_ns(ts, nsecs);
104}
105
8524070b 106EXPORT_SYMBOL(getnstimeofday);
107
108/**
109 * do_gettimeofday - Returns the time of day in a timeval
110 * @tv: pointer to the timeval to be set
111 *
efd9ac86 112 * NOTE: Users should be converted to using getnstimeofday()
8524070b 113 */
114void do_gettimeofday(struct timeval *tv)
115{
116 struct timespec now;
117
efd9ac86 118 getnstimeofday(&now);
8524070b 119 tv->tv_sec = now.tv_sec;
120 tv->tv_usec = now.tv_nsec/1000;
121}
122
123EXPORT_SYMBOL(do_gettimeofday);
124/**
125 * do_settimeofday - Sets the time of day
126 * @tv: pointer to the timespec variable containing the new time
127 *
128 * Sets the time of day to the new time and update NTP and notify hrtimers
129 */
130int do_settimeofday(struct timespec *tv)
131{
132 unsigned long flags;
133 time_t wtm_sec, sec = tv->tv_sec;
134 long wtm_nsec, nsec = tv->tv_nsec;
135
136 if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
137 return -EINVAL;
138
139 write_seqlock_irqsave(&xtime_lock, flags);
140
141 nsec -= __get_nsec_offset();
142
143 wtm_sec = wall_to_monotonic.tv_sec + (xtime.tv_sec - sec);
144 wtm_nsec = wall_to_monotonic.tv_nsec + (xtime.tv_nsec - nsec);
145
146 set_normalized_timespec(&xtime, sec, nsec);
147 set_normalized_timespec(&wall_to_monotonic, wtm_sec, wtm_nsec);
1001d0a9 148 update_xtime_cache(0);
8524070b 149
150 clock->error = 0;
151 ntp_clear();
152
153 update_vsyscall(&xtime, clock);
154
155 write_sequnlock_irqrestore(&xtime_lock, flags);
156
157 /* signal hrtimers about time change */
158 clock_was_set();
159
160 return 0;
161}
162
163EXPORT_SYMBOL(do_settimeofday);
164
165/**
166 * change_clocksource - Swaps clocksources if a new one is available
167 *
168 * Accumulates current time interval and initializes new clocksource
169 */
170static void change_clocksource(void)
171{
172 struct clocksource *new;
173 cycle_t now;
174 u64 nsec;
175
176 new = clocksource_get_next();
177
178 if (clock == new)
179 return;
180
d8bb6f4c 181 new->cycle_last = 0;
8524070b 182 now = clocksource_read(new);
183 nsec = __get_nsec_offset();
184 timespec_add_ns(&xtime, nsec);
185
186 clock = new;
187 clock->cycle_last = now;
188
189 clock->error = 0;
190 clock->xtime_nsec = 0;
10a398d0 191 clocksource_calculate_interval(clock, NTP_INTERVAL_LENGTH);
8524070b 192
193 tick_clock_notify();
194
92896bd9
LT
195 /*
196 * We're holding xtime lock and waking up klogd would deadlock
197 * us on enqueue. So no printing!
8524070b 198 printk(KERN_INFO "Time: %s clocksource has been installed.\n",
199 clock->name);
92896bd9 200 */
8524070b 201}
202#else
203static inline void change_clocksource(void) { }
6a669ee8 204static inline s64 __get_nsec_offset(void) { return 0; }
8524070b 205#endif
206
207/**
cf4fc6cb 208 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
8524070b 209 */
cf4fc6cb 210int timekeeping_valid_for_hres(void)
8524070b 211{
212 unsigned long seq;
213 int ret;
214
215 do {
216 seq = read_seqbegin(&xtime_lock);
217
218 ret = clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
219
220 } while (read_seqretry(&xtime_lock, seq));
221
222 return ret;
223}
224
225/**
226 * read_persistent_clock - Return time in seconds from the persistent clock.
227 *
228 * Weak dummy function for arches that do not yet support it.
229 * Returns seconds from epoch using the battery backed persistent clock.
230 * Returns zero if unsupported.
231 *
232 * XXX - Do be sure to remove it once all arches implement it.
233 */
234unsigned long __attribute__((weak)) read_persistent_clock(void)
235{
236 return 0;
237}
238
239/*
240 * timekeeping_init - Initializes the clocksource and common timekeeping values
241 */
242void __init timekeeping_init(void)
243{
244 unsigned long flags;
245 unsigned long sec = read_persistent_clock();
246
247 write_seqlock_irqsave(&xtime_lock, flags);
248
7dffa3c6 249 ntp_init();
8524070b 250
251 clock = clocksource_get_next();
10a398d0 252 clocksource_calculate_interval(clock, NTP_INTERVAL_LENGTH);
8524070b 253 clock->cycle_last = clocksource_read(clock);
254
255 xtime.tv_sec = sec;
256 xtime.tv_nsec = 0;
257 set_normalized_timespec(&wall_to_monotonic,
258 -xtime.tv_sec, -xtime.tv_nsec);
1001d0a9 259 update_xtime_cache(0);
7c3f1a57 260 total_sleep_time = 0;
8524070b 261 write_sequnlock_irqrestore(&xtime_lock, flags);
262}
263
264/* flag for if timekeeping is suspended */
265static int timekeeping_suspended;
266/* time in seconds when suspend began */
267static unsigned long timekeeping_suspend_time;
6a669ee8
TG
268/* xtime offset when we went into suspend */
269static s64 timekeeping_suspend_nsecs;
8524070b 270
271/**
272 * timekeeping_resume - Resumes the generic timekeeping subsystem.
273 * @dev: unused
274 *
275 * This is for the generic clocksource timekeeping.
276 * xtime/wall_to_monotonic/jiffies/etc are
277 * still managed by arch specific suspend/resume code.
278 */
279static int timekeeping_resume(struct sys_device *dev)
280{
281 unsigned long flags;
282 unsigned long now = read_persistent_clock();
283
d10ff3fb
TG
284 clocksource_resume();
285
8524070b 286 write_seqlock_irqsave(&xtime_lock, flags);
287
288 if (now && (now > timekeeping_suspend_time)) {
289 unsigned long sleep_length = now - timekeeping_suspend_time;
290
291 xtime.tv_sec += sleep_length;
292 wall_to_monotonic.tv_sec -= sleep_length;
7c3f1a57 293 total_sleep_time += sleep_length;
8524070b 294 }
6a669ee8
TG
295 /* Make sure that we have the correct xtime reference */
296 timespec_add_ns(&xtime, timekeeping_suspend_nsecs);
1001d0a9 297 update_xtime_cache(0);
8524070b 298 /* re-base the last cycle value */
d8bb6f4c 299 clock->cycle_last = 0;
8524070b 300 clock->cycle_last = clocksource_read(clock);
301 clock->error = 0;
302 timekeeping_suspended = 0;
303 write_sequnlock_irqrestore(&xtime_lock, flags);
304
305 touch_softlockup_watchdog();
306
307 clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
308
309 /* Resume hrtimers */
310 hres_timers_resume();
311
312 return 0;
313}
314
315static int timekeeping_suspend(struct sys_device *dev, pm_message_t state)
316{
317 unsigned long flags;
318
3be90950
TG
319 timekeeping_suspend_time = read_persistent_clock();
320
8524070b 321 write_seqlock_irqsave(&xtime_lock, flags);
6a669ee8
TG
322 /* Get the current xtime offset */
323 timekeeping_suspend_nsecs = __get_nsec_offset();
8524070b 324 timekeeping_suspended = 1;
8524070b 325 write_sequnlock_irqrestore(&xtime_lock, flags);
326
327 clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
328
329 return 0;
330}
331
332/* sysfs resume/suspend bits for timekeeping */
333static struct sysdev_class timekeeping_sysclass = {
af5ca3f4 334 .name = "timekeeping",
8524070b 335 .resume = timekeeping_resume,
336 .suspend = timekeeping_suspend,
8524070b 337};
338
339static struct sys_device device_timer = {
340 .id = 0,
341 .cls = &timekeeping_sysclass,
342};
343
344static int __init timekeeping_init_device(void)
345{
346 int error = sysdev_class_register(&timekeeping_sysclass);
347 if (!error)
348 error = sysdev_register(&device_timer);
349 return error;
350}
351
352device_initcall(timekeeping_init_device);
353
354/*
355 * If the error is already larger, we look ahead even further
356 * to compensate for late or lost adjustments.
357 */
358static __always_inline int clocksource_bigadjust(s64 error, s64 *interval,
359 s64 *offset)
360{
361 s64 tick_error, i;
362 u32 look_ahead, adj;
363 s32 error2, mult;
364
365 /*
366 * Use the current error value to determine how much to look ahead.
367 * The larger the error the slower we adjust for it to avoid problems
368 * with losing too many ticks, otherwise we would overadjust and
369 * produce an even larger error. The smaller the adjustment the
370 * faster we try to adjust for it, as lost ticks can do less harm
3eb05676 371 * here. This is tuned so that an error of about 1 msec is adjusted
8524070b 372 * within about 1 sec (or 2^20 nsec in 2^SHIFT_HZ ticks).
373 */
7fc5c784 374 error2 = clock->error >> (NTP_SCALE_SHIFT + 22 - 2 * SHIFT_HZ);
8524070b 375 error2 = abs(error2);
376 for (look_ahead = 0; error2 > 0; look_ahead++)
377 error2 >>= 2;
378
379 /*
380 * Now calculate the error in (1 << look_ahead) ticks, but first
381 * remove the single look ahead already included in the error.
382 */
8383c423 383 tick_error = tick_length >> (NTP_SCALE_SHIFT - clock->shift + 1);
8524070b 384 tick_error -= clock->xtime_interval >> 1;
385 error = ((error - tick_error) >> look_ahead) + tick_error;
386
387 /* Finally calculate the adjustment shift value. */
388 i = *interval;
389 mult = 1;
390 if (error < 0) {
391 error = -error;
392 *interval = -*interval;
393 *offset = -*offset;
394 mult = -1;
395 }
396 for (adj = 0; error > i; adj++)
397 error >>= 1;
398
399 *interval <<= adj;
400 *offset <<= adj;
401 return mult << adj;
402}
403
404/*
405 * Adjust the multiplier to reduce the error value,
406 * this is optimized for the most common adjustments of -1,0,1,
407 * for other values we can do a bit more work.
408 */
71120f18 409static void clocksource_adjust(s64 offset)
8524070b 410{
411 s64 error, interval = clock->cycle_interval;
412 int adj;
413
7fc5c784 414 error = clock->error >> (NTP_SCALE_SHIFT - clock->shift - 1);
8524070b 415 if (error > interval) {
416 error >>= 2;
417 if (likely(error <= interval))
418 adj = 1;
419 else
420 adj = clocksource_bigadjust(error, &interval, &offset);
421 } else if (error < -interval) {
422 error >>= 2;
423 if (likely(error >= -interval)) {
424 adj = -1;
425 interval = -interval;
426 offset = -offset;
427 } else
428 adj = clocksource_bigadjust(error, &interval, &offset);
429 } else
430 return;
431
432 clock->mult += adj;
433 clock->xtime_interval += interval;
434 clock->xtime_nsec -= offset;
435 clock->error -= (interval - offset) <<
7fc5c784 436 (NTP_SCALE_SHIFT - clock->shift);
8524070b 437}
438
439/**
440 * update_wall_time - Uses the current clocksource to increment the wall time
441 *
442 * Called from the timer interrupt, must hold a write on xtime_lock.
443 */
444void update_wall_time(void)
445{
446 cycle_t offset;
447
448 /* Make sure we're fully resumed: */
449 if (unlikely(timekeeping_suspended))
450 return;
451
452#ifdef CONFIG_GENERIC_TIME
453 offset = (clocksource_read(clock) - clock->cycle_last) & clock->mask;
454#else
455 offset = clock->cycle_interval;
456#endif
457 clock->xtime_nsec += (s64)xtime.tv_nsec << clock->shift;
458
459 /* normally this loop will run just once, however in the
460 * case of lost or late ticks, it will accumulate correctly.
461 */
462 while (offset >= clock->cycle_interval) {
463 /* accumulate one interval */
464 clock->xtime_nsec += clock->xtime_interval;
465 clock->cycle_last += clock->cycle_interval;
466 offset -= clock->cycle_interval;
467
468 if (clock->xtime_nsec >= (u64)NSEC_PER_SEC << clock->shift) {
469 clock->xtime_nsec -= (u64)NSEC_PER_SEC << clock->shift;
470 xtime.tv_sec++;
471 second_overflow();
472 }
473
8524070b 474 /* accumulate error between NTP and clock interval */
8383c423 475 clock->error += tick_length;
7fc5c784 476 clock->error -= clock->xtime_interval << (NTP_SCALE_SHIFT - clock->shift);
8524070b 477 }
478
479 /* correct the clock when NTP error is too big */
71120f18 480 clocksource_adjust(offset);
8524070b 481
482 /* store full nanoseconds into xtime */
483 xtime.tv_nsec = (s64)clock->xtime_nsec >> clock->shift;
484 clock->xtime_nsec -= (s64)xtime.tv_nsec << clock->shift;
485
17c38b74 486 update_xtime_cache(cyc2ns(clock, offset));
487
8524070b 488 /* check to see if there is a new clocksource to use */
489 change_clocksource();
490 update_vsyscall(&xtime, clock);
491}
7c3f1a57
TJ
492
493/**
494 * getboottime - Return the real time of system boot.
495 * @ts: pointer to the timespec to be set
496 *
497 * Returns the time of day in a timespec.
498 *
499 * This is based on the wall_to_monotonic offset and the total suspend
500 * time. Calls to settimeofday will affect the value returned (which
501 * basically means that however wrong your real time clock is at boot time,
502 * you get the right time here).
503 */
504void getboottime(struct timespec *ts)
505{
506 set_normalized_timespec(ts,
507 - (wall_to_monotonic.tv_sec + total_sleep_time),
508 - wall_to_monotonic.tv_nsec);
509}
510
511/**
512 * monotonic_to_bootbased - Convert the monotonic time to boot based.
513 * @ts: pointer to the timespec to be converted
514 */
515void monotonic_to_bootbased(struct timespec *ts)
516{
517 ts->tv_sec += total_sleep_time;
518}
2c6b47de 519
17c38b74 520unsigned long get_seconds(void)
521{
522 return xtime_cache.tv_sec;
523}
524EXPORT_SYMBOL(get_seconds);
525
526
2c6b47de 527struct timespec current_kernel_time(void)
528{
529 struct timespec now;
530 unsigned long seq;
531
532 do {
533 seq = read_seqbegin(&xtime_lock);
534
17c38b74 535 now = xtime_cache;
2c6b47de 536 } while (read_seqretry(&xtime_lock, seq));
537
538 return now;
539}
2c6b47de 540EXPORT_SYMBOL(current_kernel_time);