2 * linux/kernel/time/timekeeping.c
4 * Kernel timekeeping code and accessor functions
6 * This code was moved from linux/kernel/timer.c.
7 * Please see that file for copyright and history logs.
11 #include <linux/timekeeper_internal.h>
12 #include <linux/module.h>
13 #include <linux/interrupt.h>
14 #include <linux/percpu.h>
15 #include <linux/init.h>
17 #include <linux/sched.h>
18 #include <linux/syscore_ops.h>
19 #include <linux/clocksource.h>
20 #include <linux/jiffies.h>
21 #include <linux/time.h>
22 #include <linux/tick.h>
23 #include <linux/stop_machine.h>
24 #include <linux/pvclock_gtod.h>
25 #include <linux/compiler.h>
27 #include "tick-internal.h"
28 #include "ntp_internal.h"
29 #include "timekeeping_internal.h"
31 #define TK_CLEAR_NTP (1 << 0)
32 #define TK_MIRROR (1 << 1)
33 #define TK_CLOCK_WAS_SET (1 << 2)
36 * The most important data for readout fits into a single 64 byte
41 struct timekeeper timekeeper;
42 } tk_core ____cacheline_aligned;
44 static DEFINE_RAW_SPINLOCK(timekeeper_lock);
45 static struct timekeeper shadow_timekeeper;
48 * struct tk_fast - NMI safe timekeeper
49 * @seq: Sequence counter for protecting updates. The lowest bit
50 * is the index for the tk_read_base array
51 * @base: tk_read_base array. Access is indexed by the lowest bit of
54 * See @update_fast_timekeeper() below.
58 struct tk_read_base base[2];
61 static struct tk_fast tk_fast_mono ____cacheline_aligned;
63 /* flag for if timekeeping is suspended */
64 int __read_mostly timekeeping_suspended;
66 /* Flag for if there is a persistent clock on this platform */
67 bool __read_mostly persistent_clock_exist = false;
69 static inline void tk_normalize_xtime(struct timekeeper *tk)
71 while (tk->tkr.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr.shift)) {
72 tk->tkr.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr.shift;
77 static inline struct timespec64 tk_xtime(struct timekeeper *tk)
81 ts.tv_sec = tk->xtime_sec;
82 ts.tv_nsec = (long)(tk->tkr.xtime_nsec >> tk->tkr.shift);
86 static void tk_set_xtime(struct timekeeper *tk, const struct timespec64 *ts)
88 tk->xtime_sec = ts->tv_sec;
89 tk->tkr.xtime_nsec = (u64)ts->tv_nsec << tk->tkr.shift;
92 static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts)
94 tk->xtime_sec += ts->tv_sec;
95 tk->tkr.xtime_nsec += (u64)ts->tv_nsec << tk->tkr.shift;
96 tk_normalize_xtime(tk);
99 static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm)
101 struct timespec64 tmp;
104 * Verify consistency of: offset_real = -wall_to_monotonic
105 * before modifying anything
107 set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec,
108 -tk->wall_to_monotonic.tv_nsec);
109 WARN_ON_ONCE(tk->offs_real.tv64 != timespec64_to_ktime(tmp).tv64);
110 tk->wall_to_monotonic = wtm;
111 set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
112 tk->offs_real = timespec64_to_ktime(tmp);
113 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0));
116 static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta)
118 tk->offs_boot = ktime_add(tk->offs_boot, delta);
122 * tk_setup_internals - Set up internals to use clocksource clock.
124 * @tk: The target timekeeper to setup.
125 * @clock: Pointer to clocksource.
127 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
128 * pair and interval request.
130 * Unless you're the timekeeping code, you should not be using this!
132 static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
135 u64 tmp, ntpinterval;
136 struct clocksource *old_clock;
138 old_clock = tk->tkr.clock;
139 tk->tkr.clock = clock;
140 tk->tkr.read = clock->read;
141 tk->tkr.mask = clock->mask;
142 tk->tkr.cycle_last = tk->tkr.read(clock);
144 /* Do the ns -> cycle conversion first, using original mult */
145 tmp = NTP_INTERVAL_LENGTH;
146 tmp <<= clock->shift;
148 tmp += clock->mult/2;
149 do_div(tmp, clock->mult);
153 interval = (cycle_t) tmp;
154 tk->cycle_interval = interval;
156 /* Go back from cycles -> shifted ns */
157 tk->xtime_interval = (u64) interval * clock->mult;
158 tk->xtime_remainder = ntpinterval - tk->xtime_interval;
160 ((u64) interval * clock->mult) >> clock->shift;
162 /* if changing clocks, convert xtime_nsec shift units */
164 int shift_change = clock->shift - old_clock->shift;
165 if (shift_change < 0)
166 tk->tkr.xtime_nsec >>= -shift_change;
168 tk->tkr.xtime_nsec <<= shift_change;
170 tk->tkr.shift = clock->shift;
173 tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
174 tk->ntp_tick = ntpinterval << tk->ntp_error_shift;
177 * The timekeeper keeps its own mult values for the currently
178 * active clocksource. These value will be adjusted via NTP
179 * to counteract clock drifting.
181 tk->tkr.mult = clock->mult;
182 tk->ntp_err_mult = 0;
185 /* Timekeeper helper functions. */
187 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
188 static u32 default_arch_gettimeoffset(void) { return 0; }
189 u32 (*arch_gettimeoffset)(void) = default_arch_gettimeoffset;
191 static inline u32 arch_gettimeoffset(void) { return 0; }
194 static inline s64 timekeeping_get_ns(struct tk_read_base *tkr)
196 cycle_t cycle_now, delta;
199 /* read clocksource: */
200 cycle_now = tkr->read(tkr->clock);
202 /* calculate the delta since the last update_wall_time: */
203 delta = clocksource_delta(cycle_now, tkr->cycle_last, tkr->mask);
205 nsec = delta * tkr->mult + tkr->xtime_nsec;
208 /* If arch requires, add in get_arch_timeoffset() */
209 return nsec + arch_gettimeoffset();
212 static inline s64 timekeeping_get_ns_raw(struct timekeeper *tk)
214 struct clocksource *clock = tk->tkr.clock;
215 cycle_t cycle_now, delta;
218 /* read clocksource: */
219 cycle_now = tk->tkr.read(clock);
221 /* calculate the delta since the last update_wall_time: */
222 delta = clocksource_delta(cycle_now, tk->tkr.cycle_last, tk->tkr.mask);
224 /* convert delta to nanoseconds. */
225 nsec = clocksource_cyc2ns(delta, clock->mult, clock->shift);
227 /* If arch requires, add in get_arch_timeoffset() */
228 return nsec + arch_gettimeoffset();
232 * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
233 * @tkr: Timekeeping readout base from which we take the update
235 * We want to use this from any context including NMI and tracing /
236 * instrumenting the timekeeping code itself.
238 * So we handle this differently than the other timekeeping accessor
239 * functions which retry when the sequence count has changed. The
242 * smp_wmb(); <- Ensure that the last base[1] update is visible
244 * smp_wmb(); <- Ensure that the seqcount update is visible
245 * update(tkf->base[0], tkr);
246 * smp_wmb(); <- Ensure that the base[0] update is visible
248 * smp_wmb(); <- Ensure that the seqcount update is visible
249 * update(tkf->base[1], tkr);
251 * The reader side does:
257 * now = now(tkf->base[idx]);
259 * } while (seq != tkf->seq)
261 * As long as we update base[0] readers are forced off to
262 * base[1]. Once base[0] is updated readers are redirected to base[0]
263 * and the base[1] update takes place.
265 * So if a NMI hits the update of base[0] then it will use base[1]
266 * which is still consistent. In the worst case this can result is a
267 * slightly wrong timestamp (a few nanoseconds). See
268 * @ktime_get_mono_fast_ns.
270 static void update_fast_timekeeper(struct tk_read_base *tkr)
272 struct tk_read_base *base = tk_fast_mono.base;
274 /* Force readers off to base[1] */
275 raw_write_seqcount_latch(&tk_fast_mono.seq);
278 memcpy(base, tkr, sizeof(*base));
280 /* Force readers back to base[0] */
281 raw_write_seqcount_latch(&tk_fast_mono.seq);
284 memcpy(base + 1, base, sizeof(*base));
288 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
290 * This timestamp is not guaranteed to be monotonic across an update.
291 * The timestamp is calculated by:
293 * now = base_mono + clock_delta * slope
295 * So if the update lowers the slope, readers who are forced to the
296 * not yet updated second array are still using the old steeper slope.
305 * |12345678---> reader order
311 * So reader 6 will observe time going backwards versus reader 5.
313 * While other CPUs are likely to be able observe that, the only way
314 * for a CPU local observation is when an NMI hits in the middle of
315 * the update. Timestamps taken from that NMI context might be ahead
316 * of the following timestamps. Callers need to be aware of that and
319 u64 notrace ktime_get_mono_fast_ns(void)
321 struct tk_read_base *tkr;
326 seq = raw_read_seqcount(&tk_fast_mono.seq);
327 tkr = tk_fast_mono.base + (seq & 0x01);
328 now = ktime_to_ns(tkr->base_mono) + timekeeping_get_ns(tkr);
330 } while (read_seqcount_retry(&tk_fast_mono.seq, seq));
333 EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
335 /* Suspend-time cycles value for halted fast timekeeper. */
336 static cycle_t cycles_at_suspend;
338 static cycle_t dummy_clock_read(struct clocksource *cs)
340 return cycles_at_suspend;
344 * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
345 * @tk: Timekeeper to snapshot.
347 * It generally is unsafe to access the clocksource after timekeeping has been
348 * suspended, so take a snapshot of the readout base of @tk and use it as the
349 * fast timekeeper's readout base while suspended. It will return the same
350 * number of cycles every time until timekeeping is resumed at which time the
351 * proper readout base for the fast timekeeper will be restored automatically.
353 static void halt_fast_timekeeper(struct timekeeper *tk)
355 static struct tk_read_base tkr_dummy;
356 struct tk_read_base *tkr = &tk->tkr;
358 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
359 cycles_at_suspend = tkr->read(tkr->clock);
360 tkr_dummy.read = dummy_clock_read;
361 update_fast_timekeeper(&tkr_dummy);
364 #ifdef CONFIG_GENERIC_TIME_VSYSCALL_OLD
366 static inline void update_vsyscall(struct timekeeper *tk)
368 struct timespec xt, wm;
370 xt = timespec64_to_timespec(tk_xtime(tk));
371 wm = timespec64_to_timespec(tk->wall_to_monotonic);
372 update_vsyscall_old(&xt, &wm, tk->tkr.clock, tk->tkr.mult,
376 static inline void old_vsyscall_fixup(struct timekeeper *tk)
381 * Store only full nanoseconds into xtime_nsec after rounding
382 * it up and add the remainder to the error difference.
383 * XXX - This is necessary to avoid small 1ns inconsistnecies caused
384 * by truncating the remainder in vsyscalls. However, it causes
385 * additional work to be done in timekeeping_adjust(). Once
386 * the vsyscall implementations are converted to use xtime_nsec
387 * (shifted nanoseconds), and CONFIG_GENERIC_TIME_VSYSCALL_OLD
388 * users are removed, this can be killed.
390 remainder = tk->tkr.xtime_nsec & ((1ULL << tk->tkr.shift) - 1);
391 tk->tkr.xtime_nsec -= remainder;
392 tk->tkr.xtime_nsec += 1ULL << tk->tkr.shift;
393 tk->ntp_error += remainder << tk->ntp_error_shift;
394 tk->ntp_error -= (1ULL << tk->tkr.shift) << tk->ntp_error_shift;
397 #define old_vsyscall_fixup(tk)
400 static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
402 static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
404 raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
408 * pvclock_gtod_register_notifier - register a pvclock timedata update listener
410 int pvclock_gtod_register_notifier(struct notifier_block *nb)
412 struct timekeeper *tk = &tk_core.timekeeper;
416 raw_spin_lock_irqsave(&timekeeper_lock, flags);
417 ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
418 update_pvclock_gtod(tk, true);
419 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
423 EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
426 * pvclock_gtod_unregister_notifier - unregister a pvclock
427 * timedata update listener
429 int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
434 raw_spin_lock_irqsave(&timekeeper_lock, flags);
435 ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
436 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
440 EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
443 * Update the ktime_t based scalar nsec members of the timekeeper
445 static inline void tk_update_ktime_data(struct timekeeper *tk)
451 * The xtime based monotonic readout is:
452 * nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
453 * The ktime based monotonic readout is:
454 * nsec = base_mono + now();
455 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
457 seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
458 nsec = (u32) tk->wall_to_monotonic.tv_nsec;
459 tk->tkr.base_mono = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
461 /* Update the monotonic raw base */
462 tk->base_raw = timespec64_to_ktime(tk->raw_time);
465 * The sum of the nanoseconds portions of xtime and
466 * wall_to_monotonic can be greater/equal one second. Take
467 * this into account before updating tk->ktime_sec.
469 nsec += (u32)(tk->tkr.xtime_nsec >> tk->tkr.shift);
470 if (nsec >= NSEC_PER_SEC)
472 tk->ktime_sec = seconds;
475 /* must hold timekeeper_lock */
476 static void timekeeping_update(struct timekeeper *tk, unsigned int action)
478 if (action & TK_CLEAR_NTP) {
483 tk_update_ktime_data(tk);
486 update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
488 if (action & TK_MIRROR)
489 memcpy(&shadow_timekeeper, &tk_core.timekeeper,
490 sizeof(tk_core.timekeeper));
492 update_fast_timekeeper(&tk->tkr);
496 * timekeeping_forward_now - update clock to the current time
498 * Forward the current clock to update its state since the last call to
499 * update_wall_time(). This is useful before significant clock changes,
500 * as it avoids having to deal with this time offset explicitly.
502 static void timekeeping_forward_now(struct timekeeper *tk)
504 struct clocksource *clock = tk->tkr.clock;
505 cycle_t cycle_now, delta;
508 cycle_now = tk->tkr.read(clock);
509 delta = clocksource_delta(cycle_now, tk->tkr.cycle_last, tk->tkr.mask);
510 tk->tkr.cycle_last = cycle_now;
512 tk->tkr.xtime_nsec += delta * tk->tkr.mult;
514 /* If arch requires, add in get_arch_timeoffset() */
515 tk->tkr.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr.shift;
517 tk_normalize_xtime(tk);
519 nsec = clocksource_cyc2ns(delta, clock->mult, clock->shift);
520 timespec64_add_ns(&tk->raw_time, nsec);
524 * __getnstimeofday64 - Returns the time of day in a timespec64.
525 * @ts: pointer to the timespec to be set
527 * Updates the time of day in the timespec.
528 * Returns 0 on success, or -ve when suspended (timespec will be undefined).
530 int __getnstimeofday64(struct timespec64 *ts)
532 struct timekeeper *tk = &tk_core.timekeeper;
537 seq = read_seqcount_begin(&tk_core.seq);
539 ts->tv_sec = tk->xtime_sec;
540 nsecs = timekeeping_get_ns(&tk->tkr);
542 } while (read_seqcount_retry(&tk_core.seq, seq));
545 timespec64_add_ns(ts, nsecs);
548 * Do not bail out early, in case there were callers still using
549 * the value, even in the face of the WARN_ON.
551 if (unlikely(timekeeping_suspended))
555 EXPORT_SYMBOL(__getnstimeofday64);
558 * getnstimeofday64 - Returns the time of day in a timespec64.
559 * @ts: pointer to the timespec64 to be set
561 * Returns the time of day in a timespec64 (WARN if suspended).
563 void getnstimeofday64(struct timespec64 *ts)
565 WARN_ON(__getnstimeofday64(ts));
567 EXPORT_SYMBOL(getnstimeofday64);
569 ktime_t ktime_get(void)
571 struct timekeeper *tk = &tk_core.timekeeper;
576 WARN_ON(timekeeping_suspended);
579 seq = read_seqcount_begin(&tk_core.seq);
580 base = tk->tkr.base_mono;
581 nsecs = timekeeping_get_ns(&tk->tkr);
583 } while (read_seqcount_retry(&tk_core.seq, seq));
585 return ktime_add_ns(base, nsecs);
587 EXPORT_SYMBOL_GPL(ktime_get);
589 static ktime_t *offsets[TK_OFFS_MAX] = {
590 [TK_OFFS_REAL] = &tk_core.timekeeper.offs_real,
591 [TK_OFFS_BOOT] = &tk_core.timekeeper.offs_boot,
592 [TK_OFFS_TAI] = &tk_core.timekeeper.offs_tai,
595 ktime_t ktime_get_with_offset(enum tk_offsets offs)
597 struct timekeeper *tk = &tk_core.timekeeper;
599 ktime_t base, *offset = offsets[offs];
602 WARN_ON(timekeeping_suspended);
605 seq = read_seqcount_begin(&tk_core.seq);
606 base = ktime_add(tk->tkr.base_mono, *offset);
607 nsecs = timekeeping_get_ns(&tk->tkr);
609 } while (read_seqcount_retry(&tk_core.seq, seq));
611 return ktime_add_ns(base, nsecs);
614 EXPORT_SYMBOL_GPL(ktime_get_with_offset);
617 * ktime_mono_to_any() - convert mononotic time to any other time
618 * @tmono: time to convert.
619 * @offs: which offset to use
621 ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
623 ktime_t *offset = offsets[offs];
628 seq = read_seqcount_begin(&tk_core.seq);
629 tconv = ktime_add(tmono, *offset);
630 } while (read_seqcount_retry(&tk_core.seq, seq));
634 EXPORT_SYMBOL_GPL(ktime_mono_to_any);
637 * ktime_get_raw - Returns the raw monotonic time in ktime_t format
639 ktime_t ktime_get_raw(void)
641 struct timekeeper *tk = &tk_core.timekeeper;
647 seq = read_seqcount_begin(&tk_core.seq);
649 nsecs = timekeeping_get_ns_raw(tk);
651 } while (read_seqcount_retry(&tk_core.seq, seq));
653 return ktime_add_ns(base, nsecs);
655 EXPORT_SYMBOL_GPL(ktime_get_raw);
658 * ktime_get_ts64 - get the monotonic clock in timespec64 format
659 * @ts: pointer to timespec variable
661 * The function calculates the monotonic clock from the realtime
662 * clock and the wall_to_monotonic offset and stores the result
663 * in normalized timespec64 format in the variable pointed to by @ts.
665 void ktime_get_ts64(struct timespec64 *ts)
667 struct timekeeper *tk = &tk_core.timekeeper;
668 struct timespec64 tomono;
672 WARN_ON(timekeeping_suspended);
675 seq = read_seqcount_begin(&tk_core.seq);
676 ts->tv_sec = tk->xtime_sec;
677 nsec = timekeeping_get_ns(&tk->tkr);
678 tomono = tk->wall_to_monotonic;
680 } while (read_seqcount_retry(&tk_core.seq, seq));
682 ts->tv_sec += tomono.tv_sec;
684 timespec64_add_ns(ts, nsec + tomono.tv_nsec);
686 EXPORT_SYMBOL_GPL(ktime_get_ts64);
689 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
691 * Returns the seconds portion of CLOCK_MONOTONIC with a single non
692 * serialized read. tk->ktime_sec is of type 'unsigned long' so this
693 * works on both 32 and 64 bit systems. On 32 bit systems the readout
694 * covers ~136 years of uptime which should be enough to prevent
695 * premature wrap arounds.
697 time64_t ktime_get_seconds(void)
699 struct timekeeper *tk = &tk_core.timekeeper;
701 WARN_ON(timekeeping_suspended);
702 return tk->ktime_sec;
704 EXPORT_SYMBOL_GPL(ktime_get_seconds);
707 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
709 * Returns the wall clock seconds since 1970. This replaces the
710 * get_seconds() interface which is not y2038 safe on 32bit systems.
712 * For 64bit systems the fast access to tk->xtime_sec is preserved. On
713 * 32bit systems the access must be protected with the sequence
714 * counter to provide "atomic" access to the 64bit tk->xtime_sec
717 time64_t ktime_get_real_seconds(void)
719 struct timekeeper *tk = &tk_core.timekeeper;
723 if (IS_ENABLED(CONFIG_64BIT))
724 return tk->xtime_sec;
727 seq = read_seqcount_begin(&tk_core.seq);
728 seconds = tk->xtime_sec;
730 } while (read_seqcount_retry(&tk_core.seq, seq));
734 EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
736 #ifdef CONFIG_NTP_PPS
739 * getnstime_raw_and_real - get day and raw monotonic time in timespec format
740 * @ts_raw: pointer to the timespec to be set to raw monotonic time
741 * @ts_real: pointer to the timespec to be set to the time of day
743 * This function reads both the time of day and raw monotonic time at the
744 * same time atomically and stores the resulting timestamps in timespec
747 void getnstime_raw_and_real(struct timespec *ts_raw, struct timespec *ts_real)
749 struct timekeeper *tk = &tk_core.timekeeper;
751 s64 nsecs_raw, nsecs_real;
753 WARN_ON_ONCE(timekeeping_suspended);
756 seq = read_seqcount_begin(&tk_core.seq);
758 *ts_raw = timespec64_to_timespec(tk->raw_time);
759 ts_real->tv_sec = tk->xtime_sec;
760 ts_real->tv_nsec = 0;
762 nsecs_raw = timekeeping_get_ns_raw(tk);
763 nsecs_real = timekeeping_get_ns(&tk->tkr);
765 } while (read_seqcount_retry(&tk_core.seq, seq));
767 timespec_add_ns(ts_raw, nsecs_raw);
768 timespec_add_ns(ts_real, nsecs_real);
770 EXPORT_SYMBOL(getnstime_raw_and_real);
772 #endif /* CONFIG_NTP_PPS */
775 * do_gettimeofday - Returns the time of day in a timeval
776 * @tv: pointer to the timeval to be set
778 * NOTE: Users should be converted to using getnstimeofday()
780 void do_gettimeofday(struct timeval *tv)
782 struct timespec64 now;
784 getnstimeofday64(&now);
785 tv->tv_sec = now.tv_sec;
786 tv->tv_usec = now.tv_nsec/1000;
788 EXPORT_SYMBOL(do_gettimeofday);
791 * do_settimeofday64 - Sets the time of day.
792 * @ts: pointer to the timespec64 variable containing the new time
794 * Sets the time of day to the new time and update NTP and notify hrtimers
796 int do_settimeofday64(const struct timespec64 *ts)
798 struct timekeeper *tk = &tk_core.timekeeper;
799 struct timespec64 ts_delta, xt;
802 if (!timespec64_valid_strict(ts))
805 raw_spin_lock_irqsave(&timekeeper_lock, flags);
806 write_seqcount_begin(&tk_core.seq);
808 timekeeping_forward_now(tk);
811 ts_delta.tv_sec = ts->tv_sec - xt.tv_sec;
812 ts_delta.tv_nsec = ts->tv_nsec - xt.tv_nsec;
814 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
816 tk_set_xtime(tk, ts);
818 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
820 write_seqcount_end(&tk_core.seq);
821 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
823 /* signal hrtimers about time change */
828 EXPORT_SYMBOL(do_settimeofday64);
831 * timekeeping_inject_offset - Adds or subtracts from the current time.
832 * @tv: pointer to the timespec variable containing the offset
834 * Adds or subtracts an offset value from the current time.
836 int timekeeping_inject_offset(struct timespec *ts)
838 struct timekeeper *tk = &tk_core.timekeeper;
840 struct timespec64 ts64, tmp;
843 if ((unsigned long)ts->tv_nsec >= NSEC_PER_SEC)
846 ts64 = timespec_to_timespec64(*ts);
848 raw_spin_lock_irqsave(&timekeeper_lock, flags);
849 write_seqcount_begin(&tk_core.seq);
851 timekeeping_forward_now(tk);
853 /* Make sure the proposed value is valid */
854 tmp = timespec64_add(tk_xtime(tk), ts64);
855 if (!timespec64_valid_strict(&tmp)) {
860 tk_xtime_add(tk, &ts64);
861 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts64));
863 error: /* even if we error out, we forwarded the time, so call update */
864 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
866 write_seqcount_end(&tk_core.seq);
867 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
869 /* signal hrtimers about time change */
874 EXPORT_SYMBOL(timekeeping_inject_offset);
878 * timekeeping_get_tai_offset - Returns current TAI offset from UTC
881 s32 timekeeping_get_tai_offset(void)
883 struct timekeeper *tk = &tk_core.timekeeper;
888 seq = read_seqcount_begin(&tk_core.seq);
889 ret = tk->tai_offset;
890 } while (read_seqcount_retry(&tk_core.seq, seq));
896 * __timekeeping_set_tai_offset - Lock free worker function
899 static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
901 tk->tai_offset = tai_offset;
902 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
906 * timekeeping_set_tai_offset - Sets the current TAI offset from UTC
909 void timekeeping_set_tai_offset(s32 tai_offset)
911 struct timekeeper *tk = &tk_core.timekeeper;
914 raw_spin_lock_irqsave(&timekeeper_lock, flags);
915 write_seqcount_begin(&tk_core.seq);
916 __timekeeping_set_tai_offset(tk, tai_offset);
917 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
918 write_seqcount_end(&tk_core.seq);
919 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
924 * change_clocksource - Swaps clocksources if a new one is available
926 * Accumulates current time interval and initializes new clocksource
928 static int change_clocksource(void *data)
930 struct timekeeper *tk = &tk_core.timekeeper;
931 struct clocksource *new, *old;
934 new = (struct clocksource *) data;
936 raw_spin_lock_irqsave(&timekeeper_lock, flags);
937 write_seqcount_begin(&tk_core.seq);
939 timekeeping_forward_now(tk);
941 * If the cs is in module, get a module reference. Succeeds
942 * for built-in code (owner == NULL) as well.
944 if (try_module_get(new->owner)) {
945 if (!new->enable || new->enable(new) == 0) {
947 tk_setup_internals(tk, new);
950 module_put(old->owner);
952 module_put(new->owner);
955 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
957 write_seqcount_end(&tk_core.seq);
958 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
964 * timekeeping_notify - Install a new clock source
965 * @clock: pointer to the clock source
967 * This function is called from clocksource.c after a new, better clock
968 * source has been registered. The caller holds the clocksource_mutex.
970 int timekeeping_notify(struct clocksource *clock)
972 struct timekeeper *tk = &tk_core.timekeeper;
974 if (tk->tkr.clock == clock)
976 stop_machine(change_clocksource, clock, NULL);
978 return tk->tkr.clock == clock ? 0 : -1;
982 * getrawmonotonic64 - Returns the raw monotonic time in a timespec
983 * @ts: pointer to the timespec64 to be set
985 * Returns the raw monotonic time (completely un-modified by ntp)
987 void getrawmonotonic64(struct timespec64 *ts)
989 struct timekeeper *tk = &tk_core.timekeeper;
990 struct timespec64 ts64;
995 seq = read_seqcount_begin(&tk_core.seq);
996 nsecs = timekeeping_get_ns_raw(tk);
999 } while (read_seqcount_retry(&tk_core.seq, seq));
1001 timespec64_add_ns(&ts64, nsecs);
1004 EXPORT_SYMBOL(getrawmonotonic64);
1008 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1010 int timekeeping_valid_for_hres(void)
1012 struct timekeeper *tk = &tk_core.timekeeper;
1017 seq = read_seqcount_begin(&tk_core.seq);
1019 ret = tk->tkr.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1021 } while (read_seqcount_retry(&tk_core.seq, seq));
1027 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1029 u64 timekeeping_max_deferment(void)
1031 struct timekeeper *tk = &tk_core.timekeeper;
1036 seq = read_seqcount_begin(&tk_core.seq);
1038 ret = tk->tkr.clock->max_idle_ns;
1040 } while (read_seqcount_retry(&tk_core.seq, seq));
1046 * read_persistent_clock - Return time from the persistent clock.
1048 * Weak dummy function for arches that do not yet support it.
1049 * Reads the time from the battery backed persistent clock.
1050 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1052 * XXX - Do be sure to remove it once all arches implement it.
1054 void __weak read_persistent_clock(struct timespec *ts)
1061 * read_boot_clock - Return time of the system start.
1063 * Weak dummy function for arches that do not yet support it.
1064 * Function to read the exact time the system has been started.
1065 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1067 * XXX - Do be sure to remove it once all arches implement it.
1069 void __weak read_boot_clock(struct timespec *ts)
1076 * timekeeping_init - Initializes the clocksource and common timekeeping values
1078 void __init timekeeping_init(void)
1080 struct timekeeper *tk = &tk_core.timekeeper;
1081 struct clocksource *clock;
1082 unsigned long flags;
1083 struct timespec64 now, boot, tmp;
1086 read_persistent_clock(&ts);
1087 now = timespec_to_timespec64(ts);
1088 if (!timespec64_valid_strict(&now)) {
1089 pr_warn("WARNING: Persistent clock returned invalid value!\n"
1090 " Check your CMOS/BIOS settings.\n");
1093 } else if (now.tv_sec || now.tv_nsec)
1094 persistent_clock_exist = true;
1096 read_boot_clock(&ts);
1097 boot = timespec_to_timespec64(ts);
1098 if (!timespec64_valid_strict(&boot)) {
1099 pr_warn("WARNING: Boot clock returned invalid value!\n"
1100 " Check your CMOS/BIOS settings.\n");
1105 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1106 write_seqcount_begin(&tk_core.seq);
1109 clock = clocksource_default_clock();
1111 clock->enable(clock);
1112 tk_setup_internals(tk, clock);
1114 tk_set_xtime(tk, &now);
1115 tk->raw_time.tv_sec = 0;
1116 tk->raw_time.tv_nsec = 0;
1117 tk->base_raw.tv64 = 0;
1118 if (boot.tv_sec == 0 && boot.tv_nsec == 0)
1119 boot = tk_xtime(tk);
1121 set_normalized_timespec64(&tmp, -boot.tv_sec, -boot.tv_nsec);
1122 tk_set_wall_to_mono(tk, tmp);
1124 timekeeping_update(tk, TK_MIRROR);
1126 write_seqcount_end(&tk_core.seq);
1127 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1130 /* time in seconds when suspend began */
1131 static struct timespec64 timekeeping_suspend_time;
1134 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1135 * @delta: pointer to a timespec delta value
1137 * Takes a timespec offset measuring a suspend interval and properly
1138 * adds the sleep offset to the timekeeping variables.
1140 static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1141 struct timespec64 *delta)
1143 if (!timespec64_valid_strict(delta)) {
1144 printk_deferred(KERN_WARNING
1145 "__timekeeping_inject_sleeptime: Invalid "
1146 "sleep delta value!\n");
1149 tk_xtime_add(tk, delta);
1150 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1151 tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1152 tk_debug_account_sleep_time(delta);
1156 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1157 * @delta: pointer to a timespec64 delta value
1159 * This hook is for architectures that cannot support read_persistent_clock
1160 * because their RTC/persistent clock is only accessible when irqs are enabled.
1162 * This function should only be called by rtc_resume(), and allows
1163 * a suspend offset to be injected into the timekeeping values.
1165 void timekeeping_inject_sleeptime64(struct timespec64 *delta)
1167 struct timekeeper *tk = &tk_core.timekeeper;
1168 unsigned long flags;
1171 * Make sure we don't set the clock twice, as timekeeping_resume()
1174 if (has_persistent_clock())
1177 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1178 write_seqcount_begin(&tk_core.seq);
1180 timekeeping_forward_now(tk);
1182 __timekeeping_inject_sleeptime(tk, delta);
1184 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1186 write_seqcount_end(&tk_core.seq);
1187 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1189 /* signal hrtimers about time change */
1194 * timekeeping_resume - Resumes the generic timekeeping subsystem.
1196 * This is for the generic clocksource timekeeping.
1197 * xtime/wall_to_monotonic/jiffies/etc are
1198 * still managed by arch specific suspend/resume code.
1200 void timekeeping_resume(void)
1202 struct timekeeper *tk = &tk_core.timekeeper;
1203 struct clocksource *clock = tk->tkr.clock;
1204 unsigned long flags;
1205 struct timespec64 ts_new, ts_delta;
1206 struct timespec tmp;
1207 cycle_t cycle_now, cycle_delta;
1208 bool suspendtime_found = false;
1210 read_persistent_clock(&tmp);
1211 ts_new = timespec_to_timespec64(tmp);
1213 clockevents_resume();
1214 clocksource_resume();
1216 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1217 write_seqcount_begin(&tk_core.seq);
1220 * After system resumes, we need to calculate the suspended time and
1221 * compensate it for the OS time. There are 3 sources that could be
1222 * used: Nonstop clocksource during suspend, persistent clock and rtc
1225 * One specific platform may have 1 or 2 or all of them, and the
1226 * preference will be:
1227 * suspend-nonstop clocksource -> persistent clock -> rtc
1228 * The less preferred source will only be tried if there is no better
1229 * usable source. The rtc part is handled separately in rtc core code.
1231 cycle_now = tk->tkr.read(clock);
1232 if ((clock->flags & CLOCK_SOURCE_SUSPEND_NONSTOP) &&
1233 cycle_now > tk->tkr.cycle_last) {
1234 u64 num, max = ULLONG_MAX;
1235 u32 mult = clock->mult;
1236 u32 shift = clock->shift;
1239 cycle_delta = clocksource_delta(cycle_now, tk->tkr.cycle_last,
1243 * "cycle_delta * mutl" may cause 64 bits overflow, if the
1244 * suspended time is too long. In that case we need do the
1245 * 64 bits math carefully
1248 if (cycle_delta > max) {
1249 num = div64_u64(cycle_delta, max);
1250 nsec = (((u64) max * mult) >> shift) * num;
1251 cycle_delta -= num * max;
1253 nsec += ((u64) cycle_delta * mult) >> shift;
1255 ts_delta = ns_to_timespec64(nsec);
1256 suspendtime_found = true;
1257 } else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1258 ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1259 suspendtime_found = true;
1262 if (suspendtime_found)
1263 __timekeeping_inject_sleeptime(tk, &ts_delta);
1265 /* Re-base the last cycle value */
1266 tk->tkr.cycle_last = cycle_now;
1268 timekeeping_suspended = 0;
1269 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1270 write_seqcount_end(&tk_core.seq);
1271 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1273 touch_softlockup_watchdog();
1275 clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
1277 /* Resume hrtimers */
1281 int timekeeping_suspend(void)
1283 struct timekeeper *tk = &tk_core.timekeeper;
1284 unsigned long flags;
1285 struct timespec64 delta, delta_delta;
1286 static struct timespec64 old_delta;
1287 struct timespec tmp;
1289 read_persistent_clock(&tmp);
1290 timekeeping_suspend_time = timespec_to_timespec64(tmp);
1293 * On some systems the persistent_clock can not be detected at
1294 * timekeeping_init by its return value, so if we see a valid
1295 * value returned, update the persistent_clock_exists flag.
1297 if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1298 persistent_clock_exist = true;
1300 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1301 write_seqcount_begin(&tk_core.seq);
1302 timekeeping_forward_now(tk);
1303 timekeeping_suspended = 1;
1306 * To avoid drift caused by repeated suspend/resumes,
1307 * which each can add ~1 second drift error,
1308 * try to compensate so the difference in system time
1309 * and persistent_clock time stays close to constant.
1311 delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1312 delta_delta = timespec64_sub(delta, old_delta);
1313 if (abs(delta_delta.tv_sec) >= 2) {
1315 * if delta_delta is too large, assume time correction
1316 * has occured and set old_delta to the current delta.
1320 /* Otherwise try to adjust old_system to compensate */
1321 timekeeping_suspend_time =
1322 timespec64_add(timekeeping_suspend_time, delta_delta);
1325 timekeeping_update(tk, TK_MIRROR);
1326 halt_fast_timekeeper(tk);
1327 write_seqcount_end(&tk_core.seq);
1328 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1330 clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
1331 clocksource_suspend();
1332 clockevents_suspend();
1337 /* sysfs resume/suspend bits for timekeeping */
1338 static struct syscore_ops timekeeping_syscore_ops = {
1339 .resume = timekeeping_resume,
1340 .suspend = timekeeping_suspend,
1343 static int __init timekeeping_init_ops(void)
1345 register_syscore_ops(&timekeeping_syscore_ops);
1348 device_initcall(timekeeping_init_ops);
1351 * Apply a multiplier adjustment to the timekeeper
1353 static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1358 s64 interval = tk->cycle_interval;
1362 mult_adj = -mult_adj;
1363 interval = -interval;
1366 mult_adj <<= adj_scale;
1367 interval <<= adj_scale;
1368 offset <<= adj_scale;
1371 * So the following can be confusing.
1373 * To keep things simple, lets assume mult_adj == 1 for now.
1375 * When mult_adj != 1, remember that the interval and offset values
1376 * have been appropriately scaled so the math is the same.
1378 * The basic idea here is that we're increasing the multiplier
1379 * by one, this causes the xtime_interval to be incremented by
1380 * one cycle_interval. This is because:
1381 * xtime_interval = cycle_interval * mult
1382 * So if mult is being incremented by one:
1383 * xtime_interval = cycle_interval * (mult + 1)
1385 * xtime_interval = (cycle_interval * mult) + cycle_interval
1386 * Which can be shortened to:
1387 * xtime_interval += cycle_interval
1389 * So offset stores the non-accumulated cycles. Thus the current
1390 * time (in shifted nanoseconds) is:
1391 * now = (offset * adj) + xtime_nsec
1392 * Now, even though we're adjusting the clock frequency, we have
1393 * to keep time consistent. In other words, we can't jump back
1394 * in time, and we also want to avoid jumping forward in time.
1396 * So given the same offset value, we need the time to be the same
1397 * both before and after the freq adjustment.
1398 * now = (offset * adj_1) + xtime_nsec_1
1399 * now = (offset * adj_2) + xtime_nsec_2
1401 * (offset * adj_1) + xtime_nsec_1 =
1402 * (offset * adj_2) + xtime_nsec_2
1406 * (offset * adj_1) + xtime_nsec_1 =
1407 * (offset * (adj_1+1)) + xtime_nsec_2
1408 * (offset * adj_1) + xtime_nsec_1 =
1409 * (offset * adj_1) + offset + xtime_nsec_2
1410 * Canceling the sides:
1411 * xtime_nsec_1 = offset + xtime_nsec_2
1413 * xtime_nsec_2 = xtime_nsec_1 - offset
1414 * Which simplfies to:
1415 * xtime_nsec -= offset
1417 * XXX - TODO: Doc ntp_error calculation.
1419 if ((mult_adj > 0) && (tk->tkr.mult + mult_adj < mult_adj)) {
1420 /* NTP adjustment caused clocksource mult overflow */
1425 tk->tkr.mult += mult_adj;
1426 tk->xtime_interval += interval;
1427 tk->tkr.xtime_nsec -= offset;
1428 tk->ntp_error -= (interval - offset) << tk->ntp_error_shift;
1432 * Calculate the multiplier adjustment needed to match the frequency
1435 static __always_inline void timekeeping_freqadjust(struct timekeeper *tk,
1438 s64 interval = tk->cycle_interval;
1439 s64 xinterval = tk->xtime_interval;
1444 /* Remove any current error adj from freq calculation */
1445 if (tk->ntp_err_mult)
1446 xinterval -= tk->cycle_interval;
1448 tk->ntp_tick = ntp_tick_length();
1450 /* Calculate current error per tick */
1451 tick_error = ntp_tick_length() >> tk->ntp_error_shift;
1452 tick_error -= (xinterval + tk->xtime_remainder);
1454 /* Don't worry about correcting it if its small */
1455 if (likely((tick_error >= 0) && (tick_error <= interval)))
1458 /* preserve the direction of correction */
1459 negative = (tick_error < 0);
1461 /* Sort out the magnitude of the correction */
1462 tick_error = abs(tick_error);
1463 for (adj = 0; tick_error > interval; adj++)
1466 /* scale the corrections */
1467 timekeeping_apply_adjustment(tk, offset, negative, adj);
1471 * Adjust the timekeeper's multiplier to the correct frequency
1472 * and also to reduce the accumulated error value.
1474 static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
1476 /* Correct for the current frequency error */
1477 timekeeping_freqadjust(tk, offset);
1479 /* Next make a small adjustment to fix any cumulative error */
1480 if (!tk->ntp_err_mult && (tk->ntp_error > 0)) {
1481 tk->ntp_err_mult = 1;
1482 timekeeping_apply_adjustment(tk, offset, 0, 0);
1483 } else if (tk->ntp_err_mult && (tk->ntp_error <= 0)) {
1484 /* Undo any existing error adjustment */
1485 timekeeping_apply_adjustment(tk, offset, 1, 0);
1486 tk->ntp_err_mult = 0;
1489 if (unlikely(tk->tkr.clock->maxadj &&
1490 (abs(tk->tkr.mult - tk->tkr.clock->mult)
1491 > tk->tkr.clock->maxadj))) {
1492 printk_once(KERN_WARNING
1493 "Adjusting %s more than 11%% (%ld vs %ld)\n",
1494 tk->tkr.clock->name, (long)tk->tkr.mult,
1495 (long)tk->tkr.clock->mult + tk->tkr.clock->maxadj);
1499 * It may be possible that when we entered this function, xtime_nsec
1500 * was very small. Further, if we're slightly speeding the clocksource
1501 * in the code above, its possible the required corrective factor to
1502 * xtime_nsec could cause it to underflow.
1504 * Now, since we already accumulated the second, cannot simply roll
1505 * the accumulated second back, since the NTP subsystem has been
1506 * notified via second_overflow. So instead we push xtime_nsec forward
1507 * by the amount we underflowed, and add that amount into the error.
1509 * We'll correct this error next time through this function, when
1510 * xtime_nsec is not as small.
1512 if (unlikely((s64)tk->tkr.xtime_nsec < 0)) {
1513 s64 neg = -(s64)tk->tkr.xtime_nsec;
1514 tk->tkr.xtime_nsec = 0;
1515 tk->ntp_error += neg << tk->ntp_error_shift;
1520 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
1522 * Helper function that accumulates a the nsecs greater then a second
1523 * from the xtime_nsec field to the xtime_secs field.
1524 * It also calls into the NTP code to handle leapsecond processing.
1527 static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
1529 u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr.shift;
1530 unsigned int clock_set = 0;
1532 while (tk->tkr.xtime_nsec >= nsecps) {
1535 tk->tkr.xtime_nsec -= nsecps;
1538 /* Figure out if its a leap sec and apply if needed */
1539 leap = second_overflow(tk->xtime_sec);
1540 if (unlikely(leap)) {
1541 struct timespec64 ts;
1543 tk->xtime_sec += leap;
1547 tk_set_wall_to_mono(tk,
1548 timespec64_sub(tk->wall_to_monotonic, ts));
1550 __timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
1552 clock_set = TK_CLOCK_WAS_SET;
1559 * logarithmic_accumulation - shifted accumulation of cycles
1561 * This functions accumulates a shifted interval of cycles into
1562 * into a shifted interval nanoseconds. Allows for O(log) accumulation
1565 * Returns the unconsumed cycles.
1567 static cycle_t logarithmic_accumulation(struct timekeeper *tk, cycle_t offset,
1569 unsigned int *clock_set)
1571 cycle_t interval = tk->cycle_interval << shift;
1574 /* If the offset is smaller then a shifted interval, do nothing */
1575 if (offset < interval)
1578 /* Accumulate one shifted interval */
1580 tk->tkr.cycle_last += interval;
1582 tk->tkr.xtime_nsec += tk->xtime_interval << shift;
1583 *clock_set |= accumulate_nsecs_to_secs(tk);
1585 /* Accumulate raw time */
1586 raw_nsecs = (u64)tk->raw_interval << shift;
1587 raw_nsecs += tk->raw_time.tv_nsec;
1588 if (raw_nsecs >= NSEC_PER_SEC) {
1589 u64 raw_secs = raw_nsecs;
1590 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
1591 tk->raw_time.tv_sec += raw_secs;
1593 tk->raw_time.tv_nsec = raw_nsecs;
1595 /* Accumulate error between NTP and clock interval */
1596 tk->ntp_error += tk->ntp_tick << shift;
1597 tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
1598 (tk->ntp_error_shift + shift);
1604 * update_wall_time - Uses the current clocksource to increment the wall time
1607 void update_wall_time(void)
1609 struct timekeeper *real_tk = &tk_core.timekeeper;
1610 struct timekeeper *tk = &shadow_timekeeper;
1612 int shift = 0, maxshift;
1613 unsigned int clock_set = 0;
1614 unsigned long flags;
1616 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1618 /* Make sure we're fully resumed: */
1619 if (unlikely(timekeeping_suspended))
1622 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
1623 offset = real_tk->cycle_interval;
1625 offset = clocksource_delta(tk->tkr.read(tk->tkr.clock),
1626 tk->tkr.cycle_last, tk->tkr.mask);
1629 /* Check if there's really nothing to do */
1630 if (offset < real_tk->cycle_interval)
1634 * With NO_HZ we may have to accumulate many cycle_intervals
1635 * (think "ticks") worth of time at once. To do this efficiently,
1636 * we calculate the largest doubling multiple of cycle_intervals
1637 * that is smaller than the offset. We then accumulate that
1638 * chunk in one go, and then try to consume the next smaller
1641 shift = ilog2(offset) - ilog2(tk->cycle_interval);
1642 shift = max(0, shift);
1643 /* Bound shift to one less than what overflows tick_length */
1644 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
1645 shift = min(shift, maxshift);
1646 while (offset >= tk->cycle_interval) {
1647 offset = logarithmic_accumulation(tk, offset, shift,
1649 if (offset < tk->cycle_interval<<shift)
1653 /* correct the clock when NTP error is too big */
1654 timekeeping_adjust(tk, offset);
1657 * XXX This can be killed once everyone converts
1658 * to the new update_vsyscall.
1660 old_vsyscall_fixup(tk);
1663 * Finally, make sure that after the rounding
1664 * xtime_nsec isn't larger than NSEC_PER_SEC
1666 clock_set |= accumulate_nsecs_to_secs(tk);
1668 write_seqcount_begin(&tk_core.seq);
1670 * Update the real timekeeper.
1672 * We could avoid this memcpy by switching pointers, but that
1673 * requires changes to all other timekeeper usage sites as
1674 * well, i.e. move the timekeeper pointer getter into the
1675 * spinlocked/seqcount protected sections. And we trade this
1676 * memcpy under the tk_core.seq against one before we start
1679 memcpy(real_tk, tk, sizeof(*tk));
1680 timekeeping_update(real_tk, clock_set);
1681 write_seqcount_end(&tk_core.seq);
1683 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1685 /* Have to call _delayed version, since in irq context*/
1686 clock_was_set_delayed();
1690 * getboottime64 - Return the real time of system boot.
1691 * @ts: pointer to the timespec64 to be set
1693 * Returns the wall-time of boot in a timespec64.
1695 * This is based on the wall_to_monotonic offset and the total suspend
1696 * time. Calls to settimeofday will affect the value returned (which
1697 * basically means that however wrong your real time clock is at boot time,
1698 * you get the right time here).
1700 void getboottime64(struct timespec64 *ts)
1702 struct timekeeper *tk = &tk_core.timekeeper;
1703 ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
1705 *ts = ktime_to_timespec64(t);
1707 EXPORT_SYMBOL_GPL(getboottime64);
1709 unsigned long get_seconds(void)
1711 struct timekeeper *tk = &tk_core.timekeeper;
1713 return tk->xtime_sec;
1715 EXPORT_SYMBOL(get_seconds);
1717 struct timespec __current_kernel_time(void)
1719 struct timekeeper *tk = &tk_core.timekeeper;
1721 return timespec64_to_timespec(tk_xtime(tk));
1724 struct timespec current_kernel_time(void)
1726 struct timekeeper *tk = &tk_core.timekeeper;
1727 struct timespec64 now;
1731 seq = read_seqcount_begin(&tk_core.seq);
1734 } while (read_seqcount_retry(&tk_core.seq, seq));
1736 return timespec64_to_timespec(now);
1738 EXPORT_SYMBOL(current_kernel_time);
1740 struct timespec64 get_monotonic_coarse64(void)
1742 struct timekeeper *tk = &tk_core.timekeeper;
1743 struct timespec64 now, mono;
1747 seq = read_seqcount_begin(&tk_core.seq);
1750 mono = tk->wall_to_monotonic;
1751 } while (read_seqcount_retry(&tk_core.seq, seq));
1753 set_normalized_timespec64(&now, now.tv_sec + mono.tv_sec,
1754 now.tv_nsec + mono.tv_nsec);
1760 * Must hold jiffies_lock
1762 void do_timer(unsigned long ticks)
1764 jiffies_64 += ticks;
1765 calc_global_load(ticks);
1769 * ktime_get_update_offsets_tick - hrtimer helper
1770 * @offs_real: pointer to storage for monotonic -> realtime offset
1771 * @offs_boot: pointer to storage for monotonic -> boottime offset
1772 * @offs_tai: pointer to storage for monotonic -> clock tai offset
1774 * Returns monotonic time at last tick and various offsets
1776 ktime_t ktime_get_update_offsets_tick(ktime_t *offs_real, ktime_t *offs_boot,
1779 struct timekeeper *tk = &tk_core.timekeeper;
1785 seq = read_seqcount_begin(&tk_core.seq);
1787 base = tk->tkr.base_mono;
1788 nsecs = tk->tkr.xtime_nsec >> tk->tkr.shift;
1790 *offs_real = tk->offs_real;
1791 *offs_boot = tk->offs_boot;
1792 *offs_tai = tk->offs_tai;
1793 } while (read_seqcount_retry(&tk_core.seq, seq));
1795 return ktime_add_ns(base, nsecs);
1798 #ifdef CONFIG_HIGH_RES_TIMERS
1800 * ktime_get_update_offsets_now - hrtimer helper
1801 * @offs_real: pointer to storage for monotonic -> realtime offset
1802 * @offs_boot: pointer to storage for monotonic -> boottime offset
1803 * @offs_tai: pointer to storage for monotonic -> clock tai offset
1805 * Returns current monotonic time and updates the offsets
1806 * Called from hrtimer_interrupt() or retrigger_next_event()
1808 ktime_t ktime_get_update_offsets_now(ktime_t *offs_real, ktime_t *offs_boot,
1811 struct timekeeper *tk = &tk_core.timekeeper;
1817 seq = read_seqcount_begin(&tk_core.seq);
1819 base = tk->tkr.base_mono;
1820 nsecs = timekeeping_get_ns(&tk->tkr);
1822 *offs_real = tk->offs_real;
1823 *offs_boot = tk->offs_boot;
1824 *offs_tai = tk->offs_tai;
1825 } while (read_seqcount_retry(&tk_core.seq, seq));
1827 return ktime_add_ns(base, nsecs);
1832 * do_adjtimex() - Accessor function to NTP __do_adjtimex function
1834 int do_adjtimex(struct timex *txc)
1836 struct timekeeper *tk = &tk_core.timekeeper;
1837 unsigned long flags;
1838 struct timespec64 ts;
1842 /* Validate the data before disabling interrupts */
1843 ret = ntp_validate_timex(txc);
1847 if (txc->modes & ADJ_SETOFFSET) {
1848 struct timespec delta;
1849 delta.tv_sec = txc->time.tv_sec;
1850 delta.tv_nsec = txc->time.tv_usec;
1851 if (!(txc->modes & ADJ_NANO))
1852 delta.tv_nsec *= 1000;
1853 ret = timekeeping_inject_offset(&delta);
1858 getnstimeofday64(&ts);
1860 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1861 write_seqcount_begin(&tk_core.seq);
1863 orig_tai = tai = tk->tai_offset;
1864 ret = __do_adjtimex(txc, &ts, &tai);
1866 if (tai != orig_tai) {
1867 __timekeeping_set_tai_offset(tk, tai);
1868 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1870 write_seqcount_end(&tk_core.seq);
1871 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1873 if (tai != orig_tai)
1876 ntp_notify_cmos_timer();
1881 #ifdef CONFIG_NTP_PPS
1883 * hardpps() - Accessor function to NTP __hardpps function
1885 void hardpps(const struct timespec *phase_ts, const struct timespec *raw_ts)
1887 unsigned long flags;
1889 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1890 write_seqcount_begin(&tk_core.seq);
1892 __hardpps(phase_ts, raw_ts);
1894 write_seqcount_end(&tk_core.seq);
1895 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1897 EXPORT_SYMBOL(hardpps);
1901 * xtime_update() - advances the timekeeping infrastructure
1902 * @ticks: number of ticks, that have elapsed since the last call.
1904 * Must be called with interrupts disabled.
1906 void xtime_update(unsigned long ticks)
1908 write_seqlock(&jiffies_lock);
1910 write_sequnlock(&jiffies_lock);