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;
62 static struct tk_fast tk_fast_raw ____cacheline_aligned;
64 /* flag for if timekeeping is suspended */
65 int __read_mostly timekeeping_suspended;
67 static inline void tk_normalize_xtime(struct timekeeper *tk)
69 while (tk->tkr_mono.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_mono.shift)) {
70 tk->tkr_mono.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
75 static inline struct timespec64 tk_xtime(struct timekeeper *tk)
79 ts.tv_sec = tk->xtime_sec;
80 ts.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
84 static void tk_set_xtime(struct timekeeper *tk, const struct timespec64 *ts)
86 tk->xtime_sec = ts->tv_sec;
87 tk->tkr_mono.xtime_nsec = (u64)ts->tv_nsec << tk->tkr_mono.shift;
90 static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts)
92 tk->xtime_sec += ts->tv_sec;
93 tk->tkr_mono.xtime_nsec += (u64)ts->tv_nsec << tk->tkr_mono.shift;
94 tk_normalize_xtime(tk);
97 static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm)
99 struct timespec64 tmp;
102 * Verify consistency of: offset_real = -wall_to_monotonic
103 * before modifying anything
105 set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec,
106 -tk->wall_to_monotonic.tv_nsec);
107 WARN_ON_ONCE(tk->offs_real.tv64 != timespec64_to_ktime(tmp).tv64);
108 tk->wall_to_monotonic = wtm;
109 set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
110 tk->offs_real = timespec64_to_ktime(tmp);
111 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0));
114 static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta)
116 tk->offs_boot = ktime_add(tk->offs_boot, delta);
119 #ifdef CONFIG_DEBUG_TIMEKEEPING
120 #define WARNING_FREQ (HZ*300) /* 5 minute rate-limiting */
122 * These simple flag variables are managed
123 * without locks, which is racy, but ok since
124 * we don't really care about being super
125 * precise about how many events were seen,
126 * just that a problem was observed.
128 static int timekeeping_underflow_seen;
129 static int timekeeping_overflow_seen;
131 /* last_warning is only modified under the timekeeping lock */
132 static long timekeeping_last_warning;
134 static void timekeeping_check_update(struct timekeeper *tk, cycle_t offset)
137 cycle_t max_cycles = tk->tkr_mono.clock->max_cycles;
138 const char *name = tk->tkr_mono.clock->name;
140 if (offset > max_cycles) {
141 printk_deferred("WARNING: timekeeping: Cycle offset (%lld) is larger than allowed by the '%s' clock's max_cycles value (%lld): time overflow danger\n",
142 offset, name, max_cycles);
143 printk_deferred(" timekeeping: Your kernel is sick, but tries to cope by capping time updates\n");
145 if (offset > (max_cycles >> 1)) {
146 printk_deferred("INFO: timekeeping: Cycle offset (%lld) is larger than the the '%s' clock's 50%% safety margin (%lld)\n",
147 offset, name, max_cycles >> 1);
148 printk_deferred(" timekeeping: Your kernel is still fine, but is feeling a bit nervous\n");
152 if (timekeeping_underflow_seen) {
153 if (jiffies - timekeeping_last_warning > WARNING_FREQ) {
154 printk_deferred("WARNING: Underflow in clocksource '%s' observed, time update ignored.\n", name);
155 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n");
156 printk_deferred(" Your kernel is probably still fine.\n");
157 timekeeping_last_warning = jiffies;
159 timekeeping_underflow_seen = 0;
162 if (timekeeping_overflow_seen) {
163 if (jiffies - timekeeping_last_warning > WARNING_FREQ) {
164 printk_deferred("WARNING: Overflow in clocksource '%s' observed, time update capped.\n", name);
165 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n");
166 printk_deferred(" Your kernel is probably still fine.\n");
167 timekeeping_last_warning = jiffies;
169 timekeeping_overflow_seen = 0;
173 static inline cycle_t timekeeping_get_delta(struct tk_read_base *tkr)
175 cycle_t now, last, mask, max, delta;
179 * Since we're called holding a seqlock, the data may shift
180 * under us while we're doing the calculation. This can cause
181 * false positives, since we'd note a problem but throw the
182 * results away. So nest another seqlock here to atomically
183 * grab the points we are checking with.
186 seq = read_seqcount_begin(&tk_core.seq);
187 now = tkr->read(tkr->clock);
188 last = tkr->cycle_last;
190 max = tkr->clock->max_cycles;
191 } while (read_seqcount_retry(&tk_core.seq, seq));
193 delta = clocksource_delta(now, last, mask);
196 * Try to catch underflows by checking if we are seeing small
197 * mask-relative negative values.
199 if (unlikely((~delta & mask) < (mask >> 3))) {
200 timekeeping_underflow_seen = 1;
204 /* Cap delta value to the max_cycles values to avoid mult overflows */
205 if (unlikely(delta > max)) {
206 timekeeping_overflow_seen = 1;
207 delta = tkr->clock->max_cycles;
213 static inline void timekeeping_check_update(struct timekeeper *tk, cycle_t offset)
216 static inline cycle_t timekeeping_get_delta(struct tk_read_base *tkr)
218 cycle_t cycle_now, delta;
220 /* read clocksource */
221 cycle_now = tkr->read(tkr->clock);
223 /* calculate the delta since the last update_wall_time */
224 delta = clocksource_delta(cycle_now, tkr->cycle_last, tkr->mask);
231 * tk_setup_internals - Set up internals to use clocksource clock.
233 * @tk: The target timekeeper to setup.
234 * @clock: Pointer to clocksource.
236 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
237 * pair and interval request.
239 * Unless you're the timekeeping code, you should not be using this!
241 static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
244 u64 tmp, ntpinterval;
245 struct clocksource *old_clock;
247 old_clock = tk->tkr_mono.clock;
248 tk->tkr_mono.clock = clock;
249 tk->tkr_mono.read = clock->read;
250 tk->tkr_mono.mask = clock->mask;
251 tk->tkr_mono.cycle_last = tk->tkr_mono.read(clock);
253 tk->tkr_raw.clock = clock;
254 tk->tkr_raw.read = clock->read;
255 tk->tkr_raw.mask = clock->mask;
256 tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last;
258 /* Do the ns -> cycle conversion first, using original mult */
259 tmp = NTP_INTERVAL_LENGTH;
260 tmp <<= clock->shift;
262 tmp += clock->mult/2;
263 do_div(tmp, clock->mult);
267 interval = (cycle_t) tmp;
268 tk->cycle_interval = interval;
270 /* Go back from cycles -> shifted ns */
271 tk->xtime_interval = (u64) interval * clock->mult;
272 tk->xtime_remainder = ntpinterval - tk->xtime_interval;
274 ((u64) interval * clock->mult) >> clock->shift;
276 /* if changing clocks, convert xtime_nsec shift units */
278 int shift_change = clock->shift - old_clock->shift;
279 if (shift_change < 0)
280 tk->tkr_mono.xtime_nsec >>= -shift_change;
282 tk->tkr_mono.xtime_nsec <<= shift_change;
284 tk->tkr_raw.xtime_nsec = 0;
286 tk->tkr_mono.shift = clock->shift;
287 tk->tkr_raw.shift = clock->shift;
290 tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
291 tk->ntp_tick = ntpinterval << tk->ntp_error_shift;
294 * The timekeeper keeps its own mult values for the currently
295 * active clocksource. These value will be adjusted via NTP
296 * to counteract clock drifting.
298 tk->tkr_mono.mult = clock->mult;
299 tk->tkr_raw.mult = clock->mult;
300 tk->ntp_err_mult = 0;
303 /* Timekeeper helper functions. */
305 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
306 static u32 default_arch_gettimeoffset(void) { return 0; }
307 u32 (*arch_gettimeoffset)(void) = default_arch_gettimeoffset;
309 static inline u32 arch_gettimeoffset(void) { return 0; }
312 static inline s64 timekeeping_get_ns(struct tk_read_base *tkr)
317 delta = timekeeping_get_delta(tkr);
319 nsec = delta * tkr->mult + tkr->xtime_nsec;
322 /* If arch requires, add in get_arch_timeoffset() */
323 return nsec + arch_gettimeoffset();
327 * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
328 * @tkr: Timekeeping readout base from which we take the update
330 * We want to use this from any context including NMI and tracing /
331 * instrumenting the timekeeping code itself.
333 * So we handle this differently than the other timekeeping accessor
334 * functions which retry when the sequence count has changed. The
337 * smp_wmb(); <- Ensure that the last base[1] update is visible
339 * smp_wmb(); <- Ensure that the seqcount update is visible
340 * update(tkf->base[0], tkr);
341 * smp_wmb(); <- Ensure that the base[0] update is visible
343 * smp_wmb(); <- Ensure that the seqcount update is visible
344 * update(tkf->base[1], tkr);
346 * The reader side does:
352 * now = now(tkf->base[idx]);
354 * } while (seq != tkf->seq)
356 * As long as we update base[0] readers are forced off to
357 * base[1]. Once base[0] is updated readers are redirected to base[0]
358 * and the base[1] update takes place.
360 * So if a NMI hits the update of base[0] then it will use base[1]
361 * which is still consistent. In the worst case this can result is a
362 * slightly wrong timestamp (a few nanoseconds). See
363 * @ktime_get_mono_fast_ns.
365 static void update_fast_timekeeper(struct tk_read_base *tkr, struct tk_fast *tkf)
367 struct tk_read_base *base = tkf->base;
369 /* Force readers off to base[1] */
370 raw_write_seqcount_latch(&tkf->seq);
373 memcpy(base, tkr, sizeof(*base));
375 /* Force readers back to base[0] */
376 raw_write_seqcount_latch(&tkf->seq);
379 memcpy(base + 1, base, sizeof(*base));
383 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
385 * This timestamp is not guaranteed to be monotonic across an update.
386 * The timestamp is calculated by:
388 * now = base_mono + clock_delta * slope
390 * So if the update lowers the slope, readers who are forced to the
391 * not yet updated second array are still using the old steeper slope.
400 * |12345678---> reader order
406 * So reader 6 will observe time going backwards versus reader 5.
408 * While other CPUs are likely to be able observe that, the only way
409 * for a CPU local observation is when an NMI hits in the middle of
410 * the update. Timestamps taken from that NMI context might be ahead
411 * of the following timestamps. Callers need to be aware of that and
414 static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf)
416 struct tk_read_base *tkr;
421 seq = raw_read_seqcount(&tkf->seq);
422 tkr = tkf->base + (seq & 0x01);
423 now = ktime_to_ns(tkr->base) + timekeeping_get_ns(tkr);
424 } while (read_seqcount_retry(&tkf->seq, seq));
429 u64 ktime_get_mono_fast_ns(void)
431 return __ktime_get_fast_ns(&tk_fast_mono);
433 EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
435 u64 ktime_get_raw_fast_ns(void)
437 return __ktime_get_fast_ns(&tk_fast_raw);
439 EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns);
441 /* Suspend-time cycles value for halted fast timekeeper. */
442 static cycle_t cycles_at_suspend;
444 static cycle_t dummy_clock_read(struct clocksource *cs)
446 return cycles_at_suspend;
450 * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
451 * @tk: Timekeeper to snapshot.
453 * It generally is unsafe to access the clocksource after timekeeping has been
454 * suspended, so take a snapshot of the readout base of @tk and use it as the
455 * fast timekeeper's readout base while suspended. It will return the same
456 * number of cycles every time until timekeeping is resumed at which time the
457 * proper readout base for the fast timekeeper will be restored automatically.
459 static void halt_fast_timekeeper(struct timekeeper *tk)
461 static struct tk_read_base tkr_dummy;
462 struct tk_read_base *tkr = &tk->tkr_mono;
464 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
465 cycles_at_suspend = tkr->read(tkr->clock);
466 tkr_dummy.read = dummy_clock_read;
467 update_fast_timekeeper(&tkr_dummy, &tk_fast_mono);
470 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
471 tkr_dummy.read = dummy_clock_read;
472 update_fast_timekeeper(&tkr_dummy, &tk_fast_raw);
475 #ifdef CONFIG_GENERIC_TIME_VSYSCALL_OLD
477 static inline void update_vsyscall(struct timekeeper *tk)
479 struct timespec xt, wm;
481 xt = timespec64_to_timespec(tk_xtime(tk));
482 wm = timespec64_to_timespec(tk->wall_to_monotonic);
483 update_vsyscall_old(&xt, &wm, tk->tkr_mono.clock, tk->tkr_mono.mult,
484 tk->tkr_mono.cycle_last);
487 static inline void old_vsyscall_fixup(struct timekeeper *tk)
492 * Store only full nanoseconds into xtime_nsec after rounding
493 * it up and add the remainder to the error difference.
494 * XXX - This is necessary to avoid small 1ns inconsistnecies caused
495 * by truncating the remainder in vsyscalls. However, it causes
496 * additional work to be done in timekeeping_adjust(). Once
497 * the vsyscall implementations are converted to use xtime_nsec
498 * (shifted nanoseconds), and CONFIG_GENERIC_TIME_VSYSCALL_OLD
499 * users are removed, this can be killed.
501 remainder = tk->tkr_mono.xtime_nsec & ((1ULL << tk->tkr_mono.shift) - 1);
502 tk->tkr_mono.xtime_nsec -= remainder;
503 tk->tkr_mono.xtime_nsec += 1ULL << tk->tkr_mono.shift;
504 tk->ntp_error += remainder << tk->ntp_error_shift;
505 tk->ntp_error -= (1ULL << tk->tkr_mono.shift) << tk->ntp_error_shift;
508 #define old_vsyscall_fixup(tk)
511 static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
513 static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
515 raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
519 * pvclock_gtod_register_notifier - register a pvclock timedata update listener
521 int pvclock_gtod_register_notifier(struct notifier_block *nb)
523 struct timekeeper *tk = &tk_core.timekeeper;
527 raw_spin_lock_irqsave(&timekeeper_lock, flags);
528 ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
529 update_pvclock_gtod(tk, true);
530 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
534 EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
537 * pvclock_gtod_unregister_notifier - unregister a pvclock
538 * timedata update listener
540 int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
545 raw_spin_lock_irqsave(&timekeeper_lock, flags);
546 ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
547 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
551 EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
554 * Update the ktime_t based scalar nsec members of the timekeeper
556 static inline void tk_update_ktime_data(struct timekeeper *tk)
562 * The xtime based monotonic readout is:
563 * nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
564 * The ktime based monotonic readout is:
565 * nsec = base_mono + now();
566 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
568 seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
569 nsec = (u32) tk->wall_to_monotonic.tv_nsec;
570 tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
572 /* Update the monotonic raw base */
573 tk->tkr_raw.base = timespec64_to_ktime(tk->raw_time);
576 * The sum of the nanoseconds portions of xtime and
577 * wall_to_monotonic can be greater/equal one second. Take
578 * this into account before updating tk->ktime_sec.
580 nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
581 if (nsec >= NSEC_PER_SEC)
583 tk->ktime_sec = seconds;
586 /* must hold timekeeper_lock */
587 static void timekeeping_update(struct timekeeper *tk, unsigned int action)
589 if (action & TK_CLEAR_NTP) {
594 tk_update_ktime_data(tk);
597 update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
599 if (action & TK_MIRROR)
600 memcpy(&shadow_timekeeper, &tk_core.timekeeper,
601 sizeof(tk_core.timekeeper));
603 update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono);
604 update_fast_timekeeper(&tk->tkr_raw, &tk_fast_raw);
606 if (action & TK_CLOCK_WAS_SET)
607 tk->clock_was_set_seq++;
611 * timekeeping_forward_now - update clock to the current time
613 * Forward the current clock to update its state since the last call to
614 * update_wall_time(). This is useful before significant clock changes,
615 * as it avoids having to deal with this time offset explicitly.
617 static void timekeeping_forward_now(struct timekeeper *tk)
619 struct clocksource *clock = tk->tkr_mono.clock;
620 cycle_t cycle_now, delta;
623 cycle_now = tk->tkr_mono.read(clock);
624 delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
625 tk->tkr_mono.cycle_last = cycle_now;
626 tk->tkr_raw.cycle_last = cycle_now;
628 tk->tkr_mono.xtime_nsec += delta * tk->tkr_mono.mult;
630 /* If arch requires, add in get_arch_timeoffset() */
631 tk->tkr_mono.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr_mono.shift;
633 tk_normalize_xtime(tk);
635 nsec = clocksource_cyc2ns(delta, tk->tkr_raw.mult, tk->tkr_raw.shift);
636 timespec64_add_ns(&tk->raw_time, nsec);
640 * __getnstimeofday64 - Returns the time of day in a timespec64.
641 * @ts: pointer to the timespec to be set
643 * Updates the time of day in the timespec.
644 * Returns 0 on success, or -ve when suspended (timespec will be undefined).
646 int __getnstimeofday64(struct timespec64 *ts)
648 struct timekeeper *tk = &tk_core.timekeeper;
653 seq = read_seqcount_begin(&tk_core.seq);
655 ts->tv_sec = tk->xtime_sec;
656 nsecs = timekeeping_get_ns(&tk->tkr_mono);
658 } while (read_seqcount_retry(&tk_core.seq, seq));
661 timespec64_add_ns(ts, nsecs);
664 * Do not bail out early, in case there were callers still using
665 * the value, even in the face of the WARN_ON.
667 if (unlikely(timekeeping_suspended))
671 EXPORT_SYMBOL(__getnstimeofday64);
674 * getnstimeofday64 - Returns the time of day in a timespec64.
675 * @ts: pointer to the timespec64 to be set
677 * Returns the time of day in a timespec64 (WARN if suspended).
679 void getnstimeofday64(struct timespec64 *ts)
681 WARN_ON(__getnstimeofday64(ts));
683 EXPORT_SYMBOL(getnstimeofday64);
685 ktime_t ktime_get(void)
687 struct timekeeper *tk = &tk_core.timekeeper;
692 WARN_ON(timekeeping_suspended);
695 seq = read_seqcount_begin(&tk_core.seq);
696 base = tk->tkr_mono.base;
697 nsecs = timekeeping_get_ns(&tk->tkr_mono);
699 } while (read_seqcount_retry(&tk_core.seq, seq));
701 return ktime_add_ns(base, nsecs);
703 EXPORT_SYMBOL_GPL(ktime_get);
705 static ktime_t *offsets[TK_OFFS_MAX] = {
706 [TK_OFFS_REAL] = &tk_core.timekeeper.offs_real,
707 [TK_OFFS_BOOT] = &tk_core.timekeeper.offs_boot,
708 [TK_OFFS_TAI] = &tk_core.timekeeper.offs_tai,
711 ktime_t ktime_get_with_offset(enum tk_offsets offs)
713 struct timekeeper *tk = &tk_core.timekeeper;
715 ktime_t base, *offset = offsets[offs];
718 WARN_ON(timekeeping_suspended);
721 seq = read_seqcount_begin(&tk_core.seq);
722 base = ktime_add(tk->tkr_mono.base, *offset);
723 nsecs = timekeeping_get_ns(&tk->tkr_mono);
725 } while (read_seqcount_retry(&tk_core.seq, seq));
727 return ktime_add_ns(base, nsecs);
730 EXPORT_SYMBOL_GPL(ktime_get_with_offset);
733 * ktime_mono_to_any() - convert mononotic time to any other time
734 * @tmono: time to convert.
735 * @offs: which offset to use
737 ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
739 ktime_t *offset = offsets[offs];
744 seq = read_seqcount_begin(&tk_core.seq);
745 tconv = ktime_add(tmono, *offset);
746 } while (read_seqcount_retry(&tk_core.seq, seq));
750 EXPORT_SYMBOL_GPL(ktime_mono_to_any);
753 * ktime_get_raw - Returns the raw monotonic time in ktime_t format
755 ktime_t ktime_get_raw(void)
757 struct timekeeper *tk = &tk_core.timekeeper;
763 seq = read_seqcount_begin(&tk_core.seq);
764 base = tk->tkr_raw.base;
765 nsecs = timekeeping_get_ns(&tk->tkr_raw);
767 } while (read_seqcount_retry(&tk_core.seq, seq));
769 return ktime_add_ns(base, nsecs);
771 EXPORT_SYMBOL_GPL(ktime_get_raw);
774 * ktime_get_ts64 - get the monotonic clock in timespec64 format
775 * @ts: pointer to timespec variable
777 * The function calculates the monotonic clock from the realtime
778 * clock and the wall_to_monotonic offset and stores the result
779 * in normalized timespec64 format in the variable pointed to by @ts.
781 void ktime_get_ts64(struct timespec64 *ts)
783 struct timekeeper *tk = &tk_core.timekeeper;
784 struct timespec64 tomono;
788 WARN_ON(timekeeping_suspended);
791 seq = read_seqcount_begin(&tk_core.seq);
792 ts->tv_sec = tk->xtime_sec;
793 nsec = timekeeping_get_ns(&tk->tkr_mono);
794 tomono = tk->wall_to_monotonic;
796 } while (read_seqcount_retry(&tk_core.seq, seq));
798 ts->tv_sec += tomono.tv_sec;
800 timespec64_add_ns(ts, nsec + tomono.tv_nsec);
802 EXPORT_SYMBOL_GPL(ktime_get_ts64);
805 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
807 * Returns the seconds portion of CLOCK_MONOTONIC with a single non
808 * serialized read. tk->ktime_sec is of type 'unsigned long' so this
809 * works on both 32 and 64 bit systems. On 32 bit systems the readout
810 * covers ~136 years of uptime which should be enough to prevent
811 * premature wrap arounds.
813 time64_t ktime_get_seconds(void)
815 struct timekeeper *tk = &tk_core.timekeeper;
817 WARN_ON(timekeeping_suspended);
818 return tk->ktime_sec;
820 EXPORT_SYMBOL_GPL(ktime_get_seconds);
823 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
825 * Returns the wall clock seconds since 1970. This replaces the
826 * get_seconds() interface which is not y2038 safe on 32bit systems.
828 * For 64bit systems the fast access to tk->xtime_sec is preserved. On
829 * 32bit systems the access must be protected with the sequence
830 * counter to provide "atomic" access to the 64bit tk->xtime_sec
833 time64_t ktime_get_real_seconds(void)
835 struct timekeeper *tk = &tk_core.timekeeper;
839 if (IS_ENABLED(CONFIG_64BIT))
840 return tk->xtime_sec;
843 seq = read_seqcount_begin(&tk_core.seq);
844 seconds = tk->xtime_sec;
846 } while (read_seqcount_retry(&tk_core.seq, seq));
850 EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
852 #ifdef CONFIG_NTP_PPS
855 * getnstime_raw_and_real - get day and raw monotonic time in timespec format
856 * @ts_raw: pointer to the timespec to be set to raw monotonic time
857 * @ts_real: pointer to the timespec to be set to the time of day
859 * This function reads both the time of day and raw monotonic time at the
860 * same time atomically and stores the resulting timestamps in timespec
863 void getnstime_raw_and_real(struct timespec *ts_raw, struct timespec *ts_real)
865 struct timekeeper *tk = &tk_core.timekeeper;
867 s64 nsecs_raw, nsecs_real;
869 WARN_ON_ONCE(timekeeping_suspended);
872 seq = read_seqcount_begin(&tk_core.seq);
874 *ts_raw = timespec64_to_timespec(tk->raw_time);
875 ts_real->tv_sec = tk->xtime_sec;
876 ts_real->tv_nsec = 0;
878 nsecs_raw = timekeeping_get_ns(&tk->tkr_raw);
879 nsecs_real = timekeeping_get_ns(&tk->tkr_mono);
881 } while (read_seqcount_retry(&tk_core.seq, seq));
883 timespec_add_ns(ts_raw, nsecs_raw);
884 timespec_add_ns(ts_real, nsecs_real);
886 EXPORT_SYMBOL(getnstime_raw_and_real);
888 #endif /* CONFIG_NTP_PPS */
891 * do_gettimeofday - Returns the time of day in a timeval
892 * @tv: pointer to the timeval to be set
894 * NOTE: Users should be converted to using getnstimeofday()
896 void do_gettimeofday(struct timeval *tv)
898 struct timespec64 now;
900 getnstimeofday64(&now);
901 tv->tv_sec = now.tv_sec;
902 tv->tv_usec = now.tv_nsec/1000;
904 EXPORT_SYMBOL(do_gettimeofday);
907 * do_settimeofday64 - Sets the time of day.
908 * @ts: pointer to the timespec64 variable containing the new time
910 * Sets the time of day to the new time and update NTP and notify hrtimers
912 int do_settimeofday64(const struct timespec64 *ts)
914 struct timekeeper *tk = &tk_core.timekeeper;
915 struct timespec64 ts_delta, xt;
918 if (!timespec64_valid_strict(ts))
921 raw_spin_lock_irqsave(&timekeeper_lock, flags);
922 write_seqcount_begin(&tk_core.seq);
924 timekeeping_forward_now(tk);
927 ts_delta.tv_sec = ts->tv_sec - xt.tv_sec;
928 ts_delta.tv_nsec = ts->tv_nsec - xt.tv_nsec;
930 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
932 tk_set_xtime(tk, ts);
934 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
936 write_seqcount_end(&tk_core.seq);
937 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
939 /* signal hrtimers about time change */
944 EXPORT_SYMBOL(do_settimeofday64);
947 * timekeeping_inject_offset - Adds or subtracts from the current time.
948 * @tv: pointer to the timespec variable containing the offset
950 * Adds or subtracts an offset value from the current time.
952 int timekeeping_inject_offset(struct timespec *ts)
954 struct timekeeper *tk = &tk_core.timekeeper;
956 struct timespec64 ts64, tmp;
959 if ((unsigned long)ts->tv_nsec >= NSEC_PER_SEC)
962 ts64 = timespec_to_timespec64(*ts);
964 raw_spin_lock_irqsave(&timekeeper_lock, flags);
965 write_seqcount_begin(&tk_core.seq);
967 timekeeping_forward_now(tk);
969 /* Make sure the proposed value is valid */
970 tmp = timespec64_add(tk_xtime(tk), ts64);
971 if (!timespec64_valid_strict(&tmp)) {
976 tk_xtime_add(tk, &ts64);
977 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts64));
979 error: /* even if we error out, we forwarded the time, so call update */
980 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
982 write_seqcount_end(&tk_core.seq);
983 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
985 /* signal hrtimers about time change */
990 EXPORT_SYMBOL(timekeeping_inject_offset);
994 * timekeeping_get_tai_offset - Returns current TAI offset from UTC
997 s32 timekeeping_get_tai_offset(void)
999 struct timekeeper *tk = &tk_core.timekeeper;
1004 seq = read_seqcount_begin(&tk_core.seq);
1005 ret = tk->tai_offset;
1006 } while (read_seqcount_retry(&tk_core.seq, seq));
1012 * __timekeeping_set_tai_offset - Lock free worker function
1015 static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1017 tk->tai_offset = tai_offset;
1018 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1022 * timekeeping_set_tai_offset - Sets the current TAI offset from UTC
1025 void timekeeping_set_tai_offset(s32 tai_offset)
1027 struct timekeeper *tk = &tk_core.timekeeper;
1028 unsigned long flags;
1030 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1031 write_seqcount_begin(&tk_core.seq);
1032 __timekeeping_set_tai_offset(tk, tai_offset);
1033 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1034 write_seqcount_end(&tk_core.seq);
1035 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1040 * change_clocksource - Swaps clocksources if a new one is available
1042 * Accumulates current time interval and initializes new clocksource
1044 static int change_clocksource(void *data)
1046 struct timekeeper *tk = &tk_core.timekeeper;
1047 struct clocksource *new, *old;
1048 unsigned long flags;
1050 new = (struct clocksource *) data;
1052 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1053 write_seqcount_begin(&tk_core.seq);
1055 timekeeping_forward_now(tk);
1057 * If the cs is in module, get a module reference. Succeeds
1058 * for built-in code (owner == NULL) as well.
1060 if (try_module_get(new->owner)) {
1061 if (!new->enable || new->enable(new) == 0) {
1062 old = tk->tkr_mono.clock;
1063 tk_setup_internals(tk, new);
1066 module_put(old->owner);
1068 module_put(new->owner);
1071 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1073 write_seqcount_end(&tk_core.seq);
1074 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1080 * timekeeping_notify - Install a new clock source
1081 * @clock: pointer to the clock source
1083 * This function is called from clocksource.c after a new, better clock
1084 * source has been registered. The caller holds the clocksource_mutex.
1086 int timekeeping_notify(struct clocksource *clock)
1088 struct timekeeper *tk = &tk_core.timekeeper;
1090 if (tk->tkr_mono.clock == clock)
1092 stop_machine(change_clocksource, clock, NULL);
1093 tick_clock_notify();
1094 return tk->tkr_mono.clock == clock ? 0 : -1;
1098 * getrawmonotonic64 - Returns the raw monotonic time in a timespec
1099 * @ts: pointer to the timespec64 to be set
1101 * Returns the raw monotonic time (completely un-modified by ntp)
1103 void getrawmonotonic64(struct timespec64 *ts)
1105 struct timekeeper *tk = &tk_core.timekeeper;
1106 struct timespec64 ts64;
1111 seq = read_seqcount_begin(&tk_core.seq);
1112 nsecs = timekeeping_get_ns(&tk->tkr_raw);
1113 ts64 = tk->raw_time;
1115 } while (read_seqcount_retry(&tk_core.seq, seq));
1117 timespec64_add_ns(&ts64, nsecs);
1120 EXPORT_SYMBOL(getrawmonotonic64);
1124 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1126 int timekeeping_valid_for_hres(void)
1128 struct timekeeper *tk = &tk_core.timekeeper;
1133 seq = read_seqcount_begin(&tk_core.seq);
1135 ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1137 } while (read_seqcount_retry(&tk_core.seq, seq));
1143 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1145 u64 timekeeping_max_deferment(void)
1147 struct timekeeper *tk = &tk_core.timekeeper;
1152 seq = read_seqcount_begin(&tk_core.seq);
1154 ret = tk->tkr_mono.clock->max_idle_ns;
1156 } while (read_seqcount_retry(&tk_core.seq, seq));
1162 * read_persistent_clock - Return time from the persistent clock.
1164 * Weak dummy function for arches that do not yet support it.
1165 * Reads the time from the battery backed persistent clock.
1166 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1168 * XXX - Do be sure to remove it once all arches implement it.
1170 void __weak read_persistent_clock(struct timespec *ts)
1176 void __weak read_persistent_clock64(struct timespec64 *ts64)
1180 read_persistent_clock(&ts);
1181 *ts64 = timespec_to_timespec64(ts);
1185 * read_boot_clock - Return time of the system start.
1187 * Weak dummy function for arches that do not yet support it.
1188 * Function to read the exact time the system has been started.
1189 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1191 * XXX - Do be sure to remove it once all arches implement it.
1193 void __weak read_boot_clock(struct timespec *ts)
1199 void __weak read_boot_clock64(struct timespec64 *ts64)
1203 read_boot_clock(&ts);
1204 *ts64 = timespec_to_timespec64(ts);
1207 /* Flag for if timekeeping_resume() has injected sleeptime */
1208 static bool sleeptime_injected;
1210 /* Flag for if there is a persistent clock on this platform */
1211 static bool persistent_clock_exists;
1214 * timekeeping_init - Initializes the clocksource and common timekeeping values
1216 void __init timekeeping_init(void)
1218 struct timekeeper *tk = &tk_core.timekeeper;
1219 struct clocksource *clock;
1220 unsigned long flags;
1221 struct timespec64 now, boot, tmp;
1223 read_persistent_clock64(&now);
1224 if (!timespec64_valid_strict(&now)) {
1225 pr_warn("WARNING: Persistent clock returned invalid value!\n"
1226 " Check your CMOS/BIOS settings.\n");
1229 } else if (now.tv_sec || now.tv_nsec)
1230 persistent_clock_exists = true;
1232 read_boot_clock64(&boot);
1233 if (!timespec64_valid_strict(&boot)) {
1234 pr_warn("WARNING: Boot clock returned invalid value!\n"
1235 " Check your CMOS/BIOS settings.\n");
1240 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1241 write_seqcount_begin(&tk_core.seq);
1244 clock = clocksource_default_clock();
1246 clock->enable(clock);
1247 tk_setup_internals(tk, clock);
1249 tk_set_xtime(tk, &now);
1250 tk->raw_time.tv_sec = 0;
1251 tk->raw_time.tv_nsec = 0;
1252 if (boot.tv_sec == 0 && boot.tv_nsec == 0)
1253 boot = tk_xtime(tk);
1255 set_normalized_timespec64(&tmp, -boot.tv_sec, -boot.tv_nsec);
1256 tk_set_wall_to_mono(tk, tmp);
1258 timekeeping_update(tk, TK_MIRROR);
1260 write_seqcount_end(&tk_core.seq);
1261 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1264 /* time in seconds when suspend began for persistent clock */
1265 static struct timespec64 timekeeping_suspend_time;
1268 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1269 * @delta: pointer to a timespec delta value
1271 * Takes a timespec offset measuring a suspend interval and properly
1272 * adds the sleep offset to the timekeeping variables.
1274 static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1275 struct timespec64 *delta)
1277 if (!timespec64_valid_strict(delta)) {
1278 printk_deferred(KERN_WARNING
1279 "__timekeeping_inject_sleeptime: Invalid "
1280 "sleep delta value!\n");
1283 tk_xtime_add(tk, delta);
1284 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1285 tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1286 tk_debug_account_sleep_time(delta);
1289 #if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1291 * We have three kinds of time sources to use for sleep time
1292 * injection, the preference order is:
1293 * 1) non-stop clocksource
1294 * 2) persistent clock (ie: RTC accessible when irqs are off)
1297 * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1298 * If system has neither 1) nor 2), 3) will be used finally.
1301 * If timekeeping has injected sleeptime via either 1) or 2),
1302 * 3) becomes needless, so in this case we don't need to call
1303 * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1306 bool timekeeping_rtc_skipresume(void)
1308 return sleeptime_injected;
1312 * 1) can be determined whether to use or not only when doing
1313 * timekeeping_resume() which is invoked after rtc_suspend(),
1314 * so we can't skip rtc_suspend() surely if system has 1).
1316 * But if system has 2), 2) will definitely be used, so in this
1317 * case we don't need to call rtc_suspend(), and this is what
1318 * timekeeping_rtc_skipsuspend() means.
1320 bool timekeeping_rtc_skipsuspend(void)
1322 return persistent_clock_exists;
1326 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1327 * @delta: pointer to a timespec64 delta value
1329 * This hook is for architectures that cannot support read_persistent_clock64
1330 * because their RTC/persistent clock is only accessible when irqs are enabled.
1331 * and also don't have an effective nonstop clocksource.
1333 * This function should only be called by rtc_resume(), and allows
1334 * a suspend offset to be injected into the timekeeping values.
1336 void timekeeping_inject_sleeptime64(struct timespec64 *delta)
1338 struct timekeeper *tk = &tk_core.timekeeper;
1339 unsigned long flags;
1341 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1342 write_seqcount_begin(&tk_core.seq);
1344 timekeeping_forward_now(tk);
1346 __timekeeping_inject_sleeptime(tk, delta);
1348 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1350 write_seqcount_end(&tk_core.seq);
1351 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1353 /* signal hrtimers about time change */
1359 * timekeeping_resume - Resumes the generic timekeeping subsystem.
1361 void timekeeping_resume(void)
1363 struct timekeeper *tk = &tk_core.timekeeper;
1364 struct clocksource *clock = tk->tkr_mono.clock;
1365 unsigned long flags;
1366 struct timespec64 ts_new, ts_delta;
1367 cycle_t cycle_now, cycle_delta;
1369 sleeptime_injected = false;
1370 read_persistent_clock64(&ts_new);
1372 clockevents_resume();
1373 clocksource_resume();
1375 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1376 write_seqcount_begin(&tk_core.seq);
1379 * After system resumes, we need to calculate the suspended time and
1380 * compensate it for the OS time. There are 3 sources that could be
1381 * used: Nonstop clocksource during suspend, persistent clock and rtc
1384 * One specific platform may have 1 or 2 or all of them, and the
1385 * preference will be:
1386 * suspend-nonstop clocksource -> persistent clock -> rtc
1387 * The less preferred source will only be tried if there is no better
1388 * usable source. The rtc part is handled separately in rtc core code.
1390 cycle_now = tk->tkr_mono.read(clock);
1391 if ((clock->flags & CLOCK_SOURCE_SUSPEND_NONSTOP) &&
1392 cycle_now > tk->tkr_mono.cycle_last) {
1393 u64 num, max = ULLONG_MAX;
1394 u32 mult = clock->mult;
1395 u32 shift = clock->shift;
1398 cycle_delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last,
1402 * "cycle_delta * mutl" may cause 64 bits overflow, if the
1403 * suspended time is too long. In that case we need do the
1404 * 64 bits math carefully
1407 if (cycle_delta > max) {
1408 num = div64_u64(cycle_delta, max);
1409 nsec = (((u64) max * mult) >> shift) * num;
1410 cycle_delta -= num * max;
1412 nsec += ((u64) cycle_delta * mult) >> shift;
1414 ts_delta = ns_to_timespec64(nsec);
1415 sleeptime_injected = true;
1416 } else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1417 ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1418 sleeptime_injected = true;
1421 if (sleeptime_injected)
1422 __timekeeping_inject_sleeptime(tk, &ts_delta);
1424 /* Re-base the last cycle value */
1425 tk->tkr_mono.cycle_last = cycle_now;
1426 tk->tkr_raw.cycle_last = cycle_now;
1429 timekeeping_suspended = 0;
1430 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1431 write_seqcount_end(&tk_core.seq);
1432 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1434 touch_softlockup_watchdog();
1440 int timekeeping_suspend(void)
1442 struct timekeeper *tk = &tk_core.timekeeper;
1443 unsigned long flags;
1444 struct timespec64 delta, delta_delta;
1445 static struct timespec64 old_delta;
1447 read_persistent_clock64(&timekeeping_suspend_time);
1450 * On some systems the persistent_clock can not be detected at
1451 * timekeeping_init by its return value, so if we see a valid
1452 * value returned, update the persistent_clock_exists flag.
1454 if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1455 persistent_clock_exists = true;
1457 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1458 write_seqcount_begin(&tk_core.seq);
1459 timekeeping_forward_now(tk);
1460 timekeeping_suspended = 1;
1462 if (persistent_clock_exists) {
1464 * To avoid drift caused by repeated suspend/resumes,
1465 * which each can add ~1 second drift error,
1466 * try to compensate so the difference in system time
1467 * and persistent_clock time stays close to constant.
1469 delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1470 delta_delta = timespec64_sub(delta, old_delta);
1471 if (abs(delta_delta.tv_sec) >= 2) {
1473 * if delta_delta is too large, assume time correction
1474 * has occurred and set old_delta to the current delta.
1478 /* Otherwise try to adjust old_system to compensate */
1479 timekeeping_suspend_time =
1480 timespec64_add(timekeeping_suspend_time, delta_delta);
1484 timekeeping_update(tk, TK_MIRROR);
1485 halt_fast_timekeeper(tk);
1486 write_seqcount_end(&tk_core.seq);
1487 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1490 clocksource_suspend();
1491 clockevents_suspend();
1496 /* sysfs resume/suspend bits for timekeeping */
1497 static struct syscore_ops timekeeping_syscore_ops = {
1498 .resume = timekeeping_resume,
1499 .suspend = timekeeping_suspend,
1502 static int __init timekeeping_init_ops(void)
1504 register_syscore_ops(&timekeeping_syscore_ops);
1507 device_initcall(timekeeping_init_ops);
1510 * Apply a multiplier adjustment to the timekeeper
1512 static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1517 s64 interval = tk->cycle_interval;
1521 mult_adj = -mult_adj;
1522 interval = -interval;
1525 mult_adj <<= adj_scale;
1526 interval <<= adj_scale;
1527 offset <<= adj_scale;
1530 * So the following can be confusing.
1532 * To keep things simple, lets assume mult_adj == 1 for now.
1534 * When mult_adj != 1, remember that the interval and offset values
1535 * have been appropriately scaled so the math is the same.
1537 * The basic idea here is that we're increasing the multiplier
1538 * by one, this causes the xtime_interval to be incremented by
1539 * one cycle_interval. This is because:
1540 * xtime_interval = cycle_interval * mult
1541 * So if mult is being incremented by one:
1542 * xtime_interval = cycle_interval * (mult + 1)
1544 * xtime_interval = (cycle_interval * mult) + cycle_interval
1545 * Which can be shortened to:
1546 * xtime_interval += cycle_interval
1548 * So offset stores the non-accumulated cycles. Thus the current
1549 * time (in shifted nanoseconds) is:
1550 * now = (offset * adj) + xtime_nsec
1551 * Now, even though we're adjusting the clock frequency, we have
1552 * to keep time consistent. In other words, we can't jump back
1553 * in time, and we also want to avoid jumping forward in time.
1555 * So given the same offset value, we need the time to be the same
1556 * both before and after the freq adjustment.
1557 * now = (offset * adj_1) + xtime_nsec_1
1558 * now = (offset * adj_2) + xtime_nsec_2
1560 * (offset * adj_1) + xtime_nsec_1 =
1561 * (offset * adj_2) + xtime_nsec_2
1565 * (offset * adj_1) + xtime_nsec_1 =
1566 * (offset * (adj_1+1)) + xtime_nsec_2
1567 * (offset * adj_1) + xtime_nsec_1 =
1568 * (offset * adj_1) + offset + xtime_nsec_2
1569 * Canceling the sides:
1570 * xtime_nsec_1 = offset + xtime_nsec_2
1572 * xtime_nsec_2 = xtime_nsec_1 - offset
1573 * Which simplfies to:
1574 * xtime_nsec -= offset
1576 * XXX - TODO: Doc ntp_error calculation.
1578 if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
1579 /* NTP adjustment caused clocksource mult overflow */
1584 tk->tkr_mono.mult += mult_adj;
1585 tk->xtime_interval += interval;
1586 tk->tkr_mono.xtime_nsec -= offset;
1587 tk->ntp_error -= (interval - offset) << tk->ntp_error_shift;
1591 * Calculate the multiplier adjustment needed to match the frequency
1594 static __always_inline void timekeeping_freqadjust(struct timekeeper *tk,
1597 s64 interval = tk->cycle_interval;
1598 s64 xinterval = tk->xtime_interval;
1603 /* Remove any current error adj from freq calculation */
1604 if (tk->ntp_err_mult)
1605 xinterval -= tk->cycle_interval;
1607 tk->ntp_tick = ntp_tick_length();
1609 /* Calculate current error per tick */
1610 tick_error = ntp_tick_length() >> tk->ntp_error_shift;
1611 tick_error -= (xinterval + tk->xtime_remainder);
1613 /* Don't worry about correcting it if its small */
1614 if (likely((tick_error >= 0) && (tick_error <= interval)))
1617 /* preserve the direction of correction */
1618 negative = (tick_error < 0);
1620 /* Sort out the magnitude of the correction */
1621 tick_error = abs(tick_error);
1622 for (adj = 0; tick_error > interval; adj++)
1625 /* scale the corrections */
1626 timekeeping_apply_adjustment(tk, offset, negative, adj);
1630 * Adjust the timekeeper's multiplier to the correct frequency
1631 * and also to reduce the accumulated error value.
1633 static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
1635 /* Correct for the current frequency error */
1636 timekeeping_freqadjust(tk, offset);
1638 /* Next make a small adjustment to fix any cumulative error */
1639 if (!tk->ntp_err_mult && (tk->ntp_error > 0)) {
1640 tk->ntp_err_mult = 1;
1641 timekeeping_apply_adjustment(tk, offset, 0, 0);
1642 } else if (tk->ntp_err_mult && (tk->ntp_error <= 0)) {
1643 /* Undo any existing error adjustment */
1644 timekeeping_apply_adjustment(tk, offset, 1, 0);
1645 tk->ntp_err_mult = 0;
1648 if (unlikely(tk->tkr_mono.clock->maxadj &&
1649 (abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
1650 > tk->tkr_mono.clock->maxadj))) {
1651 printk_once(KERN_WARNING
1652 "Adjusting %s more than 11%% (%ld vs %ld)\n",
1653 tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
1654 (long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
1658 * It may be possible that when we entered this function, xtime_nsec
1659 * was very small. Further, if we're slightly speeding the clocksource
1660 * in the code above, its possible the required corrective factor to
1661 * xtime_nsec could cause it to underflow.
1663 * Now, since we already accumulated the second, cannot simply roll
1664 * the accumulated second back, since the NTP subsystem has been
1665 * notified via second_overflow. So instead we push xtime_nsec forward
1666 * by the amount we underflowed, and add that amount into the error.
1668 * We'll correct this error next time through this function, when
1669 * xtime_nsec is not as small.
1671 if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
1672 s64 neg = -(s64)tk->tkr_mono.xtime_nsec;
1673 tk->tkr_mono.xtime_nsec = 0;
1674 tk->ntp_error += neg << tk->ntp_error_shift;
1679 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
1681 * Helper function that accumulates a the nsecs greater then a second
1682 * from the xtime_nsec field to the xtime_secs field.
1683 * It also calls into the NTP code to handle leapsecond processing.
1686 static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
1688 u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
1689 unsigned int clock_set = 0;
1691 while (tk->tkr_mono.xtime_nsec >= nsecps) {
1694 tk->tkr_mono.xtime_nsec -= nsecps;
1697 /* Figure out if its a leap sec and apply if needed */
1698 leap = second_overflow(tk->xtime_sec);
1699 if (unlikely(leap)) {
1700 struct timespec64 ts;
1702 tk->xtime_sec += leap;
1706 tk_set_wall_to_mono(tk,
1707 timespec64_sub(tk->wall_to_monotonic, ts));
1709 __timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
1711 clock_set = TK_CLOCK_WAS_SET;
1718 * logarithmic_accumulation - shifted accumulation of cycles
1720 * This functions accumulates a shifted interval of cycles into
1721 * into a shifted interval nanoseconds. Allows for O(log) accumulation
1724 * Returns the unconsumed cycles.
1726 static cycle_t logarithmic_accumulation(struct timekeeper *tk, cycle_t offset,
1728 unsigned int *clock_set)
1730 cycle_t interval = tk->cycle_interval << shift;
1733 /* If the offset is smaller then a shifted interval, do nothing */
1734 if (offset < interval)
1737 /* Accumulate one shifted interval */
1739 tk->tkr_mono.cycle_last += interval;
1740 tk->tkr_raw.cycle_last += interval;
1742 tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
1743 *clock_set |= accumulate_nsecs_to_secs(tk);
1745 /* Accumulate raw time */
1746 raw_nsecs = (u64)tk->raw_interval << shift;
1747 raw_nsecs += tk->raw_time.tv_nsec;
1748 if (raw_nsecs >= NSEC_PER_SEC) {
1749 u64 raw_secs = raw_nsecs;
1750 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
1751 tk->raw_time.tv_sec += raw_secs;
1753 tk->raw_time.tv_nsec = raw_nsecs;
1755 /* Accumulate error between NTP and clock interval */
1756 tk->ntp_error += tk->ntp_tick << shift;
1757 tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
1758 (tk->ntp_error_shift + shift);
1764 * update_wall_time - Uses the current clocksource to increment the wall time
1767 void update_wall_time(void)
1769 struct timekeeper *real_tk = &tk_core.timekeeper;
1770 struct timekeeper *tk = &shadow_timekeeper;
1772 int shift = 0, maxshift;
1773 unsigned int clock_set = 0;
1774 unsigned long flags;
1776 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1778 /* Make sure we're fully resumed: */
1779 if (unlikely(timekeeping_suspended))
1782 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
1783 offset = real_tk->cycle_interval;
1785 offset = clocksource_delta(tk->tkr_mono.read(tk->tkr_mono.clock),
1786 tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
1789 /* Check if there's really nothing to do */
1790 if (offset < real_tk->cycle_interval)
1793 /* Do some additional sanity checking */
1794 timekeeping_check_update(real_tk, offset);
1797 * With NO_HZ we may have to accumulate many cycle_intervals
1798 * (think "ticks") worth of time at once. To do this efficiently,
1799 * we calculate the largest doubling multiple of cycle_intervals
1800 * that is smaller than the offset. We then accumulate that
1801 * chunk in one go, and then try to consume the next smaller
1804 shift = ilog2(offset) - ilog2(tk->cycle_interval);
1805 shift = max(0, shift);
1806 /* Bound shift to one less than what overflows tick_length */
1807 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
1808 shift = min(shift, maxshift);
1809 while (offset >= tk->cycle_interval) {
1810 offset = logarithmic_accumulation(tk, offset, shift,
1812 if (offset < tk->cycle_interval<<shift)
1816 /* correct the clock when NTP error is too big */
1817 timekeeping_adjust(tk, offset);
1820 * XXX This can be killed once everyone converts
1821 * to the new update_vsyscall.
1823 old_vsyscall_fixup(tk);
1826 * Finally, make sure that after the rounding
1827 * xtime_nsec isn't larger than NSEC_PER_SEC
1829 clock_set |= accumulate_nsecs_to_secs(tk);
1831 write_seqcount_begin(&tk_core.seq);
1833 * Update the real timekeeper.
1835 * We could avoid this memcpy by switching pointers, but that
1836 * requires changes to all other timekeeper usage sites as
1837 * well, i.e. move the timekeeper pointer getter into the
1838 * spinlocked/seqcount protected sections. And we trade this
1839 * memcpy under the tk_core.seq against one before we start
1842 memcpy(real_tk, tk, sizeof(*tk));
1843 timekeeping_update(real_tk, clock_set);
1844 write_seqcount_end(&tk_core.seq);
1846 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1848 /* Have to call _delayed version, since in irq context*/
1849 clock_was_set_delayed();
1853 * getboottime64 - Return the real time of system boot.
1854 * @ts: pointer to the timespec64 to be set
1856 * Returns the wall-time of boot in a timespec64.
1858 * This is based on the wall_to_monotonic offset and the total suspend
1859 * time. Calls to settimeofday will affect the value returned (which
1860 * basically means that however wrong your real time clock is at boot time,
1861 * you get the right time here).
1863 void getboottime64(struct timespec64 *ts)
1865 struct timekeeper *tk = &tk_core.timekeeper;
1866 ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
1868 *ts = ktime_to_timespec64(t);
1870 EXPORT_SYMBOL_GPL(getboottime64);
1872 unsigned long get_seconds(void)
1874 struct timekeeper *tk = &tk_core.timekeeper;
1876 return tk->xtime_sec;
1878 EXPORT_SYMBOL(get_seconds);
1880 struct timespec __current_kernel_time(void)
1882 struct timekeeper *tk = &tk_core.timekeeper;
1884 return timespec64_to_timespec(tk_xtime(tk));
1887 struct timespec current_kernel_time(void)
1889 struct timekeeper *tk = &tk_core.timekeeper;
1890 struct timespec64 now;
1894 seq = read_seqcount_begin(&tk_core.seq);
1897 } while (read_seqcount_retry(&tk_core.seq, seq));
1899 return timespec64_to_timespec(now);
1901 EXPORT_SYMBOL(current_kernel_time);
1903 struct timespec64 get_monotonic_coarse64(void)
1905 struct timekeeper *tk = &tk_core.timekeeper;
1906 struct timespec64 now, mono;
1910 seq = read_seqcount_begin(&tk_core.seq);
1913 mono = tk->wall_to_monotonic;
1914 } while (read_seqcount_retry(&tk_core.seq, seq));
1916 set_normalized_timespec64(&now, now.tv_sec + mono.tv_sec,
1917 now.tv_nsec + mono.tv_nsec);
1923 * Must hold jiffies_lock
1925 void do_timer(unsigned long ticks)
1927 jiffies_64 += ticks;
1928 calc_global_load(ticks);
1932 * ktime_get_update_offsets_now - hrtimer helper
1933 * @cwsseq: pointer to check and store the clock was set sequence number
1934 * @offs_real: pointer to storage for monotonic -> realtime offset
1935 * @offs_boot: pointer to storage for monotonic -> boottime offset
1936 * @offs_tai: pointer to storage for monotonic -> clock tai offset
1938 * Returns current monotonic time and updates the offsets if the
1939 * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
1942 * Called from hrtimer_interrupt() or retrigger_next_event()
1944 ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real,
1945 ktime_t *offs_boot, ktime_t *offs_tai)
1947 struct timekeeper *tk = &tk_core.timekeeper;
1953 seq = read_seqcount_begin(&tk_core.seq);
1955 base = tk->tkr_mono.base;
1956 nsecs = timekeeping_get_ns(&tk->tkr_mono);
1957 if (*cwsseq != tk->clock_was_set_seq) {
1958 *cwsseq = tk->clock_was_set_seq;
1959 *offs_real = tk->offs_real;
1960 *offs_boot = tk->offs_boot;
1961 *offs_tai = tk->offs_tai;
1963 } while (read_seqcount_retry(&tk_core.seq, seq));
1965 return ktime_add_ns(base, nsecs);
1969 * do_adjtimex() - Accessor function to NTP __do_adjtimex function
1971 int do_adjtimex(struct timex *txc)
1973 struct timekeeper *tk = &tk_core.timekeeper;
1974 unsigned long flags;
1975 struct timespec64 ts;
1979 /* Validate the data before disabling interrupts */
1980 ret = ntp_validate_timex(txc);
1984 if (txc->modes & ADJ_SETOFFSET) {
1985 struct timespec delta;
1986 delta.tv_sec = txc->time.tv_sec;
1987 delta.tv_nsec = txc->time.tv_usec;
1988 if (!(txc->modes & ADJ_NANO))
1989 delta.tv_nsec *= 1000;
1990 ret = timekeeping_inject_offset(&delta);
1995 getnstimeofday64(&ts);
1997 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1998 write_seqcount_begin(&tk_core.seq);
2000 orig_tai = tai = tk->tai_offset;
2001 ret = __do_adjtimex(txc, &ts, &tai);
2003 if (tai != orig_tai) {
2004 __timekeeping_set_tai_offset(tk, tai);
2005 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
2007 write_seqcount_end(&tk_core.seq);
2008 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2010 if (tai != orig_tai)
2013 ntp_notify_cmos_timer();
2018 #ifdef CONFIG_NTP_PPS
2020 * hardpps() - Accessor function to NTP __hardpps function
2022 void hardpps(const struct timespec *phase_ts, const struct timespec *raw_ts)
2024 unsigned long flags;
2026 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2027 write_seqcount_begin(&tk_core.seq);
2029 __hardpps(phase_ts, raw_ts);
2031 write_seqcount_end(&tk_core.seq);
2032 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2034 EXPORT_SYMBOL(hardpps);
2038 * xtime_update() - advances the timekeeping infrastructure
2039 * @ticks: number of ticks, that have elapsed since the last call.
2041 * Must be called with interrupts disabled.
2043 void xtime_update(unsigned long ticks)
2045 write_seqlock(&jiffies_lock);
2047 write_sequnlock(&jiffies_lock);