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 static void timekeeping_check_update(struct timekeeper *tk, cycle_t offset)
125 cycle_t max_cycles = tk->tkr_mono.clock->max_cycles;
126 const char *name = tk->tkr_mono.clock->name;
128 if (offset > max_cycles) {
129 printk_deferred("WARNING: timekeeping: Cycle offset (%lld) is larger than allowed by the '%s' clock's max_cycles value (%lld): time overflow danger\n",
130 offset, name, max_cycles);
131 printk_deferred(" timekeeping: Your kernel is sick, but tries to cope by capping time updates\n");
133 if (offset > (max_cycles >> 1)) {
134 printk_deferred("INFO: timekeeping: Cycle offset (%lld) is larger than the the '%s' clock's 50%% safety margin (%lld)\n",
135 offset, name, max_cycles >> 1);
136 printk_deferred(" timekeeping: Your kernel is still fine, but is feeling a bit nervous\n");
140 if (tk->underflow_seen) {
141 if (jiffies - tk->last_warning > WARNING_FREQ) {
142 printk_deferred("WARNING: Underflow in clocksource '%s' observed, time update ignored.\n", name);
143 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n");
144 printk_deferred(" Your kernel is probably still fine.\n");
145 tk->last_warning = jiffies;
147 tk->underflow_seen = 0;
150 if (tk->overflow_seen) {
151 if (jiffies - tk->last_warning > WARNING_FREQ) {
152 printk_deferred("WARNING: Overflow in clocksource '%s' observed, time update capped.\n", name);
153 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n");
154 printk_deferred(" Your kernel is probably still fine.\n");
155 tk->last_warning = jiffies;
157 tk->overflow_seen = 0;
161 static inline cycle_t timekeeping_get_delta(struct tk_read_base *tkr)
163 struct timekeeper *tk = &tk_core.timekeeper;
164 cycle_t now, last, mask, max, delta;
168 * Since we're called holding a seqlock, the data may shift
169 * under us while we're doing the calculation. This can cause
170 * false positives, since we'd note a problem but throw the
171 * results away. So nest another seqlock here to atomically
172 * grab the points we are checking with.
175 seq = read_seqcount_begin(&tk_core.seq);
176 now = tkr->read(tkr->clock);
177 last = tkr->cycle_last;
179 max = tkr->clock->max_cycles;
180 } while (read_seqcount_retry(&tk_core.seq, seq));
182 delta = clocksource_delta(now, last, mask);
185 * Try to catch underflows by checking if we are seeing small
186 * mask-relative negative values.
188 if (unlikely((~delta & mask) < (mask >> 3))) {
189 tk->underflow_seen = 1;
193 /* Cap delta value to the max_cycles values to avoid mult overflows */
194 if (unlikely(delta > max)) {
195 tk->overflow_seen = 1;
196 delta = tkr->clock->max_cycles;
202 static inline void timekeeping_check_update(struct timekeeper *tk, cycle_t offset)
205 static inline cycle_t timekeeping_get_delta(struct tk_read_base *tkr)
207 cycle_t cycle_now, delta;
209 /* read clocksource */
210 cycle_now = tkr->read(tkr->clock);
212 /* calculate the delta since the last update_wall_time */
213 delta = clocksource_delta(cycle_now, tkr->cycle_last, tkr->mask);
220 * tk_setup_internals - Set up internals to use clocksource clock.
222 * @tk: The target timekeeper to setup.
223 * @clock: Pointer to clocksource.
225 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
226 * pair and interval request.
228 * Unless you're the timekeeping code, you should not be using this!
230 static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
233 u64 tmp, ntpinterval;
234 struct clocksource *old_clock;
236 old_clock = tk->tkr_mono.clock;
237 tk->tkr_mono.clock = clock;
238 tk->tkr_mono.read = clock->read;
239 tk->tkr_mono.mask = clock->mask;
240 tk->tkr_mono.cycle_last = tk->tkr_mono.read(clock);
242 tk->tkr_raw.clock = clock;
243 tk->tkr_raw.read = clock->read;
244 tk->tkr_raw.mask = clock->mask;
245 tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last;
247 /* Do the ns -> cycle conversion first, using original mult */
248 tmp = NTP_INTERVAL_LENGTH;
249 tmp <<= clock->shift;
251 tmp += clock->mult/2;
252 do_div(tmp, clock->mult);
256 interval = (cycle_t) tmp;
257 tk->cycle_interval = interval;
259 /* Go back from cycles -> shifted ns */
260 tk->xtime_interval = (u64) interval * clock->mult;
261 tk->xtime_remainder = ntpinterval - tk->xtime_interval;
263 ((u64) interval * clock->mult) >> clock->shift;
265 /* if changing clocks, convert xtime_nsec shift units */
267 int shift_change = clock->shift - old_clock->shift;
268 if (shift_change < 0)
269 tk->tkr_mono.xtime_nsec >>= -shift_change;
271 tk->tkr_mono.xtime_nsec <<= shift_change;
273 tk->tkr_raw.xtime_nsec = 0;
275 tk->tkr_mono.shift = clock->shift;
276 tk->tkr_raw.shift = clock->shift;
279 tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
280 tk->ntp_tick = ntpinterval << tk->ntp_error_shift;
283 * The timekeeper keeps its own mult values for the currently
284 * active clocksource. These value will be adjusted via NTP
285 * to counteract clock drifting.
287 tk->tkr_mono.mult = clock->mult;
288 tk->tkr_raw.mult = clock->mult;
289 tk->ntp_err_mult = 0;
292 /* Timekeeper helper functions. */
294 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
295 static u32 default_arch_gettimeoffset(void) { return 0; }
296 u32 (*arch_gettimeoffset)(void) = default_arch_gettimeoffset;
298 static inline u32 arch_gettimeoffset(void) { return 0; }
301 static inline s64 timekeeping_get_ns(struct tk_read_base *tkr)
306 delta = timekeeping_get_delta(tkr);
308 nsec = delta * tkr->mult + tkr->xtime_nsec;
311 /* If arch requires, add in get_arch_timeoffset() */
312 return nsec + arch_gettimeoffset();
316 * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
317 * @tkr: Timekeeping readout base from which we take the update
319 * We want to use this from any context including NMI and tracing /
320 * instrumenting the timekeeping code itself.
322 * So we handle this differently than the other timekeeping accessor
323 * functions which retry when the sequence count has changed. The
326 * smp_wmb(); <- Ensure that the last base[1] update is visible
328 * smp_wmb(); <- Ensure that the seqcount update is visible
329 * update(tkf->base[0], tkr);
330 * smp_wmb(); <- Ensure that the base[0] update is visible
332 * smp_wmb(); <- Ensure that the seqcount update is visible
333 * update(tkf->base[1], tkr);
335 * The reader side does:
341 * now = now(tkf->base[idx]);
343 * } while (seq != tkf->seq)
345 * As long as we update base[0] readers are forced off to
346 * base[1]. Once base[0] is updated readers are redirected to base[0]
347 * and the base[1] update takes place.
349 * So if a NMI hits the update of base[0] then it will use base[1]
350 * which is still consistent. In the worst case this can result is a
351 * slightly wrong timestamp (a few nanoseconds). See
352 * @ktime_get_mono_fast_ns.
354 static void update_fast_timekeeper(struct tk_read_base *tkr, struct tk_fast *tkf)
356 struct tk_read_base *base = tkf->base;
358 /* Force readers off to base[1] */
359 raw_write_seqcount_latch(&tkf->seq);
362 memcpy(base, tkr, sizeof(*base));
364 /* Force readers back to base[0] */
365 raw_write_seqcount_latch(&tkf->seq);
368 memcpy(base + 1, base, sizeof(*base));
372 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
374 * This timestamp is not guaranteed to be monotonic across an update.
375 * The timestamp is calculated by:
377 * now = base_mono + clock_delta * slope
379 * So if the update lowers the slope, readers who are forced to the
380 * not yet updated second array are still using the old steeper slope.
389 * |12345678---> reader order
395 * So reader 6 will observe time going backwards versus reader 5.
397 * While other CPUs are likely to be able observe that, the only way
398 * for a CPU local observation is when an NMI hits in the middle of
399 * the update. Timestamps taken from that NMI context might be ahead
400 * of the following timestamps. Callers need to be aware of that and
403 static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf)
405 struct tk_read_base *tkr;
410 seq = raw_read_seqcount(&tkf->seq);
411 tkr = tkf->base + (seq & 0x01);
412 now = ktime_to_ns(tkr->base) + timekeeping_get_ns(tkr);
413 } while (read_seqcount_retry(&tkf->seq, seq));
418 u64 ktime_get_mono_fast_ns(void)
420 return __ktime_get_fast_ns(&tk_fast_mono);
422 EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
424 u64 ktime_get_raw_fast_ns(void)
426 return __ktime_get_fast_ns(&tk_fast_raw);
428 EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns);
430 /* Suspend-time cycles value for halted fast timekeeper. */
431 static cycle_t cycles_at_suspend;
433 static cycle_t dummy_clock_read(struct clocksource *cs)
435 return cycles_at_suspend;
439 * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
440 * @tk: Timekeeper to snapshot.
442 * It generally is unsafe to access the clocksource after timekeeping has been
443 * suspended, so take a snapshot of the readout base of @tk and use it as the
444 * fast timekeeper's readout base while suspended. It will return the same
445 * number of cycles every time until timekeeping is resumed at which time the
446 * proper readout base for the fast timekeeper will be restored automatically.
448 static void halt_fast_timekeeper(struct timekeeper *tk)
450 static struct tk_read_base tkr_dummy;
451 struct tk_read_base *tkr = &tk->tkr_mono;
453 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
454 cycles_at_suspend = tkr->read(tkr->clock);
455 tkr_dummy.read = dummy_clock_read;
456 update_fast_timekeeper(&tkr_dummy, &tk_fast_mono);
459 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
460 tkr_dummy.read = dummy_clock_read;
461 update_fast_timekeeper(&tkr_dummy, &tk_fast_raw);
464 #ifdef CONFIG_GENERIC_TIME_VSYSCALL_OLD
466 static inline void update_vsyscall(struct timekeeper *tk)
468 struct timespec xt, wm;
470 xt = timespec64_to_timespec(tk_xtime(tk));
471 wm = timespec64_to_timespec(tk->wall_to_monotonic);
472 update_vsyscall_old(&xt, &wm, tk->tkr_mono.clock, tk->tkr_mono.mult,
473 tk->tkr_mono.cycle_last);
476 static inline void old_vsyscall_fixup(struct timekeeper *tk)
481 * Store only full nanoseconds into xtime_nsec after rounding
482 * it up and add the remainder to the error difference.
483 * XXX - This is necessary to avoid small 1ns inconsistnecies caused
484 * by truncating the remainder in vsyscalls. However, it causes
485 * additional work to be done in timekeeping_adjust(). Once
486 * the vsyscall implementations are converted to use xtime_nsec
487 * (shifted nanoseconds), and CONFIG_GENERIC_TIME_VSYSCALL_OLD
488 * users are removed, this can be killed.
490 remainder = tk->tkr_mono.xtime_nsec & ((1ULL << tk->tkr_mono.shift) - 1);
491 tk->tkr_mono.xtime_nsec -= remainder;
492 tk->tkr_mono.xtime_nsec += 1ULL << tk->tkr_mono.shift;
493 tk->ntp_error += remainder << tk->ntp_error_shift;
494 tk->ntp_error -= (1ULL << tk->tkr_mono.shift) << tk->ntp_error_shift;
497 #define old_vsyscall_fixup(tk)
500 static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
502 static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
504 raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
508 * pvclock_gtod_register_notifier - register a pvclock timedata update listener
510 int pvclock_gtod_register_notifier(struct notifier_block *nb)
512 struct timekeeper *tk = &tk_core.timekeeper;
516 raw_spin_lock_irqsave(&timekeeper_lock, flags);
517 ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
518 update_pvclock_gtod(tk, true);
519 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
523 EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
526 * pvclock_gtod_unregister_notifier - unregister a pvclock
527 * timedata update listener
529 int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
534 raw_spin_lock_irqsave(&timekeeper_lock, flags);
535 ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
536 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
540 EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
543 * Update the ktime_t based scalar nsec members of the timekeeper
545 static inline void tk_update_ktime_data(struct timekeeper *tk)
551 * The xtime based monotonic readout is:
552 * nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
553 * The ktime based monotonic readout is:
554 * nsec = base_mono + now();
555 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
557 seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
558 nsec = (u32) tk->wall_to_monotonic.tv_nsec;
559 tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
561 /* Update the monotonic raw base */
562 tk->tkr_raw.base = timespec64_to_ktime(tk->raw_time);
565 * The sum of the nanoseconds portions of xtime and
566 * wall_to_monotonic can be greater/equal one second. Take
567 * this into account before updating tk->ktime_sec.
569 nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
570 if (nsec >= NSEC_PER_SEC)
572 tk->ktime_sec = seconds;
575 /* must hold timekeeper_lock */
576 static void timekeeping_update(struct timekeeper *tk, unsigned int action)
578 if (action & TK_CLEAR_NTP) {
583 tk_update_ktime_data(tk);
586 update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
588 update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono);
589 update_fast_timekeeper(&tk->tkr_raw, &tk_fast_raw);
591 if (action & TK_CLOCK_WAS_SET)
592 tk->clock_was_set_seq++;
594 * The mirroring of the data to the shadow-timekeeper needs
595 * to happen last here to ensure we don't over-write the
596 * timekeeper structure on the next update with stale data
598 if (action & TK_MIRROR)
599 memcpy(&shadow_timekeeper, &tk_core.timekeeper,
600 sizeof(tk_core.timekeeper));
604 * timekeeping_forward_now - update clock to the current time
606 * Forward the current clock to update its state since the last call to
607 * update_wall_time(). This is useful before significant clock changes,
608 * as it avoids having to deal with this time offset explicitly.
610 static void timekeeping_forward_now(struct timekeeper *tk)
612 struct clocksource *clock = tk->tkr_mono.clock;
613 cycle_t cycle_now, delta;
616 cycle_now = tk->tkr_mono.read(clock);
617 delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
618 tk->tkr_mono.cycle_last = cycle_now;
619 tk->tkr_raw.cycle_last = cycle_now;
621 tk->tkr_mono.xtime_nsec += delta * tk->tkr_mono.mult;
623 /* If arch requires, add in get_arch_timeoffset() */
624 tk->tkr_mono.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr_mono.shift;
626 tk_normalize_xtime(tk);
628 nsec = clocksource_cyc2ns(delta, tk->tkr_raw.mult, tk->tkr_raw.shift);
629 timespec64_add_ns(&tk->raw_time, nsec);
633 * __getnstimeofday64 - Returns the time of day in a timespec64.
634 * @ts: pointer to the timespec to be set
636 * Updates the time of day in the timespec.
637 * Returns 0 on success, or -ve when suspended (timespec will be undefined).
639 int __getnstimeofday64(struct timespec64 *ts)
641 struct timekeeper *tk = &tk_core.timekeeper;
646 seq = read_seqcount_begin(&tk_core.seq);
648 ts->tv_sec = tk->xtime_sec;
649 nsecs = timekeeping_get_ns(&tk->tkr_mono);
651 } while (read_seqcount_retry(&tk_core.seq, seq));
654 timespec64_add_ns(ts, nsecs);
657 * Do not bail out early, in case there were callers still using
658 * the value, even in the face of the WARN_ON.
660 if (unlikely(timekeeping_suspended))
664 EXPORT_SYMBOL(__getnstimeofday64);
667 * getnstimeofday64 - Returns the time of day in a timespec64.
668 * @ts: pointer to the timespec64 to be set
670 * Returns the time of day in a timespec64 (WARN if suspended).
672 void getnstimeofday64(struct timespec64 *ts)
674 WARN_ON(__getnstimeofday64(ts));
676 EXPORT_SYMBOL(getnstimeofday64);
678 ktime_t ktime_get(void)
680 struct timekeeper *tk = &tk_core.timekeeper;
685 WARN_ON(timekeeping_suspended);
688 seq = read_seqcount_begin(&tk_core.seq);
689 base = tk->tkr_mono.base;
690 nsecs = timekeeping_get_ns(&tk->tkr_mono);
692 } while (read_seqcount_retry(&tk_core.seq, seq));
694 return ktime_add_ns(base, nsecs);
696 EXPORT_SYMBOL_GPL(ktime_get);
698 u32 ktime_get_resolution_ns(void)
700 struct timekeeper *tk = &tk_core.timekeeper;
704 WARN_ON(timekeeping_suspended);
707 seq = read_seqcount_begin(&tk_core.seq);
708 nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift;
709 } while (read_seqcount_retry(&tk_core.seq, seq));
713 EXPORT_SYMBOL_GPL(ktime_get_resolution_ns);
715 static ktime_t *offsets[TK_OFFS_MAX] = {
716 [TK_OFFS_REAL] = &tk_core.timekeeper.offs_real,
717 [TK_OFFS_BOOT] = &tk_core.timekeeper.offs_boot,
718 [TK_OFFS_TAI] = &tk_core.timekeeper.offs_tai,
721 ktime_t ktime_get_with_offset(enum tk_offsets offs)
723 struct timekeeper *tk = &tk_core.timekeeper;
725 ktime_t base, *offset = offsets[offs];
728 WARN_ON(timekeeping_suspended);
731 seq = read_seqcount_begin(&tk_core.seq);
732 base = ktime_add(tk->tkr_mono.base, *offset);
733 nsecs = timekeeping_get_ns(&tk->tkr_mono);
735 } while (read_seqcount_retry(&tk_core.seq, seq));
737 return ktime_add_ns(base, nsecs);
740 EXPORT_SYMBOL_GPL(ktime_get_with_offset);
743 * ktime_mono_to_any() - convert mononotic time to any other time
744 * @tmono: time to convert.
745 * @offs: which offset to use
747 ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
749 ktime_t *offset = offsets[offs];
754 seq = read_seqcount_begin(&tk_core.seq);
755 tconv = ktime_add(tmono, *offset);
756 } while (read_seqcount_retry(&tk_core.seq, seq));
760 EXPORT_SYMBOL_GPL(ktime_mono_to_any);
763 * ktime_get_raw - Returns the raw monotonic time in ktime_t format
765 ktime_t ktime_get_raw(void)
767 struct timekeeper *tk = &tk_core.timekeeper;
773 seq = read_seqcount_begin(&tk_core.seq);
774 base = tk->tkr_raw.base;
775 nsecs = timekeeping_get_ns(&tk->tkr_raw);
777 } while (read_seqcount_retry(&tk_core.seq, seq));
779 return ktime_add_ns(base, nsecs);
781 EXPORT_SYMBOL_GPL(ktime_get_raw);
784 * ktime_get_ts64 - get the monotonic clock in timespec64 format
785 * @ts: pointer to timespec variable
787 * The function calculates the monotonic clock from the realtime
788 * clock and the wall_to_monotonic offset and stores the result
789 * in normalized timespec64 format in the variable pointed to by @ts.
791 void ktime_get_ts64(struct timespec64 *ts)
793 struct timekeeper *tk = &tk_core.timekeeper;
794 struct timespec64 tomono;
798 WARN_ON(timekeeping_suspended);
801 seq = read_seqcount_begin(&tk_core.seq);
802 ts->tv_sec = tk->xtime_sec;
803 nsec = timekeeping_get_ns(&tk->tkr_mono);
804 tomono = tk->wall_to_monotonic;
806 } while (read_seqcount_retry(&tk_core.seq, seq));
808 ts->tv_sec += tomono.tv_sec;
810 timespec64_add_ns(ts, nsec + tomono.tv_nsec);
812 EXPORT_SYMBOL_GPL(ktime_get_ts64);
815 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
817 * Returns the seconds portion of CLOCK_MONOTONIC with a single non
818 * serialized read. tk->ktime_sec is of type 'unsigned long' so this
819 * works on both 32 and 64 bit systems. On 32 bit systems the readout
820 * covers ~136 years of uptime which should be enough to prevent
821 * premature wrap arounds.
823 time64_t ktime_get_seconds(void)
825 struct timekeeper *tk = &tk_core.timekeeper;
827 WARN_ON(timekeeping_suspended);
828 return tk->ktime_sec;
830 EXPORT_SYMBOL_GPL(ktime_get_seconds);
833 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
835 * Returns the wall clock seconds since 1970. This replaces the
836 * get_seconds() interface which is not y2038 safe on 32bit systems.
838 * For 64bit systems the fast access to tk->xtime_sec is preserved. On
839 * 32bit systems the access must be protected with the sequence
840 * counter to provide "atomic" access to the 64bit tk->xtime_sec
843 time64_t ktime_get_real_seconds(void)
845 struct timekeeper *tk = &tk_core.timekeeper;
849 if (IS_ENABLED(CONFIG_64BIT))
850 return tk->xtime_sec;
853 seq = read_seqcount_begin(&tk_core.seq);
854 seconds = tk->xtime_sec;
856 } while (read_seqcount_retry(&tk_core.seq, seq));
860 EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
862 #ifdef CONFIG_NTP_PPS
865 * getnstime_raw_and_real - get day and raw monotonic time in timespec format
866 * @ts_raw: pointer to the timespec to be set to raw monotonic time
867 * @ts_real: pointer to the timespec to be set to the time of day
869 * This function reads both the time of day and raw monotonic time at the
870 * same time atomically and stores the resulting timestamps in timespec
873 void getnstime_raw_and_real(struct timespec *ts_raw, struct timespec *ts_real)
875 struct timekeeper *tk = &tk_core.timekeeper;
877 s64 nsecs_raw, nsecs_real;
879 WARN_ON_ONCE(timekeeping_suspended);
882 seq = read_seqcount_begin(&tk_core.seq);
884 *ts_raw = timespec64_to_timespec(tk->raw_time);
885 ts_real->tv_sec = tk->xtime_sec;
886 ts_real->tv_nsec = 0;
888 nsecs_raw = timekeeping_get_ns(&tk->tkr_raw);
889 nsecs_real = timekeeping_get_ns(&tk->tkr_mono);
891 } while (read_seqcount_retry(&tk_core.seq, seq));
893 timespec_add_ns(ts_raw, nsecs_raw);
894 timespec_add_ns(ts_real, nsecs_real);
896 EXPORT_SYMBOL(getnstime_raw_and_real);
898 #endif /* CONFIG_NTP_PPS */
901 * do_gettimeofday - Returns the time of day in a timeval
902 * @tv: pointer to the timeval to be set
904 * NOTE: Users should be converted to using getnstimeofday()
906 void do_gettimeofday(struct timeval *tv)
908 struct timespec64 now;
910 getnstimeofday64(&now);
911 tv->tv_sec = now.tv_sec;
912 tv->tv_usec = now.tv_nsec/1000;
914 EXPORT_SYMBOL(do_gettimeofday);
917 * do_settimeofday64 - Sets the time of day.
918 * @ts: pointer to the timespec64 variable containing the new time
920 * Sets the time of day to the new time and update NTP and notify hrtimers
922 int do_settimeofday64(const struct timespec64 *ts)
924 struct timekeeper *tk = &tk_core.timekeeper;
925 struct timespec64 ts_delta, xt;
928 if (!timespec64_valid_strict(ts))
931 raw_spin_lock_irqsave(&timekeeper_lock, flags);
932 write_seqcount_begin(&tk_core.seq);
934 timekeeping_forward_now(tk);
937 ts_delta.tv_sec = ts->tv_sec - xt.tv_sec;
938 ts_delta.tv_nsec = ts->tv_nsec - xt.tv_nsec;
940 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
942 tk_set_xtime(tk, ts);
944 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
946 write_seqcount_end(&tk_core.seq);
947 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
949 /* signal hrtimers about time change */
954 EXPORT_SYMBOL(do_settimeofday64);
957 * timekeeping_inject_offset - Adds or subtracts from the current time.
958 * @tv: pointer to the timespec variable containing the offset
960 * Adds or subtracts an offset value from the current time.
962 int timekeeping_inject_offset(struct timespec *ts)
964 struct timekeeper *tk = &tk_core.timekeeper;
966 struct timespec64 ts64, tmp;
969 if ((unsigned long)ts->tv_nsec >= NSEC_PER_SEC)
972 ts64 = timespec_to_timespec64(*ts);
974 raw_spin_lock_irqsave(&timekeeper_lock, flags);
975 write_seqcount_begin(&tk_core.seq);
977 timekeeping_forward_now(tk);
979 /* Make sure the proposed value is valid */
980 tmp = timespec64_add(tk_xtime(tk), ts64);
981 if (!timespec64_valid_strict(&tmp)) {
986 tk_xtime_add(tk, &ts64);
987 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts64));
989 error: /* even if we error out, we forwarded the time, so call update */
990 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
992 write_seqcount_end(&tk_core.seq);
993 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
995 /* signal hrtimers about time change */
1000 EXPORT_SYMBOL(timekeeping_inject_offset);
1004 * timekeeping_get_tai_offset - Returns current TAI offset from UTC
1007 s32 timekeeping_get_tai_offset(void)
1009 struct timekeeper *tk = &tk_core.timekeeper;
1014 seq = read_seqcount_begin(&tk_core.seq);
1015 ret = tk->tai_offset;
1016 } while (read_seqcount_retry(&tk_core.seq, seq));
1022 * __timekeeping_set_tai_offset - Lock free worker function
1025 static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1027 tk->tai_offset = tai_offset;
1028 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1032 * timekeeping_set_tai_offset - Sets the current TAI offset from UTC
1035 void timekeeping_set_tai_offset(s32 tai_offset)
1037 struct timekeeper *tk = &tk_core.timekeeper;
1038 unsigned long flags;
1040 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1041 write_seqcount_begin(&tk_core.seq);
1042 __timekeeping_set_tai_offset(tk, tai_offset);
1043 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1044 write_seqcount_end(&tk_core.seq);
1045 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1050 * change_clocksource - Swaps clocksources if a new one is available
1052 * Accumulates current time interval and initializes new clocksource
1054 static int change_clocksource(void *data)
1056 struct timekeeper *tk = &tk_core.timekeeper;
1057 struct clocksource *new, *old;
1058 unsigned long flags;
1060 new = (struct clocksource *) data;
1062 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1063 write_seqcount_begin(&tk_core.seq);
1065 timekeeping_forward_now(tk);
1067 * If the cs is in module, get a module reference. Succeeds
1068 * for built-in code (owner == NULL) as well.
1070 if (try_module_get(new->owner)) {
1071 if (!new->enable || new->enable(new) == 0) {
1072 old = tk->tkr_mono.clock;
1073 tk_setup_internals(tk, new);
1076 module_put(old->owner);
1078 module_put(new->owner);
1081 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1083 write_seqcount_end(&tk_core.seq);
1084 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1090 * timekeeping_notify - Install a new clock source
1091 * @clock: pointer to the clock source
1093 * This function is called from clocksource.c after a new, better clock
1094 * source has been registered. The caller holds the clocksource_mutex.
1096 int timekeeping_notify(struct clocksource *clock)
1098 struct timekeeper *tk = &tk_core.timekeeper;
1100 if (tk->tkr_mono.clock == clock)
1102 stop_machine(change_clocksource, clock, NULL);
1103 tick_clock_notify();
1104 return tk->tkr_mono.clock == clock ? 0 : -1;
1108 * getrawmonotonic64 - Returns the raw monotonic time in a timespec
1109 * @ts: pointer to the timespec64 to be set
1111 * Returns the raw monotonic time (completely un-modified by ntp)
1113 void getrawmonotonic64(struct timespec64 *ts)
1115 struct timekeeper *tk = &tk_core.timekeeper;
1116 struct timespec64 ts64;
1121 seq = read_seqcount_begin(&tk_core.seq);
1122 nsecs = timekeeping_get_ns(&tk->tkr_raw);
1123 ts64 = tk->raw_time;
1125 } while (read_seqcount_retry(&tk_core.seq, seq));
1127 timespec64_add_ns(&ts64, nsecs);
1130 EXPORT_SYMBOL(getrawmonotonic64);
1134 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1136 int timekeeping_valid_for_hres(void)
1138 struct timekeeper *tk = &tk_core.timekeeper;
1143 seq = read_seqcount_begin(&tk_core.seq);
1145 ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1147 } while (read_seqcount_retry(&tk_core.seq, seq));
1153 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1155 u64 timekeeping_max_deferment(void)
1157 struct timekeeper *tk = &tk_core.timekeeper;
1162 seq = read_seqcount_begin(&tk_core.seq);
1164 ret = tk->tkr_mono.clock->max_idle_ns;
1166 } while (read_seqcount_retry(&tk_core.seq, seq));
1172 * read_persistent_clock - Return time from the persistent clock.
1174 * Weak dummy function for arches that do not yet support it.
1175 * Reads the time from the battery backed persistent clock.
1176 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1178 * XXX - Do be sure to remove it once all arches implement it.
1180 void __weak read_persistent_clock(struct timespec *ts)
1186 void __weak read_persistent_clock64(struct timespec64 *ts64)
1190 read_persistent_clock(&ts);
1191 *ts64 = timespec_to_timespec64(ts);
1195 * read_boot_clock64 - Return time of the system start.
1197 * Weak dummy function for arches that do not yet support it.
1198 * Function to read the exact time the system has been started.
1199 * Returns a timespec64 with tv_sec=0 and tv_nsec=0 if unsupported.
1201 * XXX - Do be sure to remove it once all arches implement it.
1203 void __weak read_boot_clock64(struct timespec64 *ts)
1209 /* Flag for if timekeeping_resume() has injected sleeptime */
1210 static bool sleeptime_injected;
1212 /* Flag for if there is a persistent clock on this platform */
1213 static bool persistent_clock_exists;
1216 * timekeeping_init - Initializes the clocksource and common timekeeping values
1218 void __init timekeeping_init(void)
1220 struct timekeeper *tk = &tk_core.timekeeper;
1221 struct clocksource *clock;
1222 unsigned long flags;
1223 struct timespec64 now, boot, tmp;
1225 read_persistent_clock64(&now);
1226 if (!timespec64_valid_strict(&now)) {
1227 pr_warn("WARNING: Persistent clock returned invalid value!\n"
1228 " Check your CMOS/BIOS settings.\n");
1231 } else if (now.tv_sec || now.tv_nsec)
1232 persistent_clock_exists = true;
1234 read_boot_clock64(&boot);
1235 if (!timespec64_valid_strict(&boot)) {
1236 pr_warn("WARNING: Boot clock returned invalid value!\n"
1237 " Check your CMOS/BIOS settings.\n");
1242 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1243 write_seqcount_begin(&tk_core.seq);
1246 clock = clocksource_default_clock();
1248 clock->enable(clock);
1249 tk_setup_internals(tk, clock);
1251 tk_set_xtime(tk, &now);
1252 tk->raw_time.tv_sec = 0;
1253 tk->raw_time.tv_nsec = 0;
1254 if (boot.tv_sec == 0 && boot.tv_nsec == 0)
1255 boot = tk_xtime(tk);
1257 set_normalized_timespec64(&tmp, -boot.tv_sec, -boot.tv_nsec);
1258 tk_set_wall_to_mono(tk, tmp);
1260 timekeeping_update(tk, TK_MIRROR);
1262 write_seqcount_end(&tk_core.seq);
1263 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1266 /* time in seconds when suspend began for persistent clock */
1267 static struct timespec64 timekeeping_suspend_time;
1270 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1271 * @delta: pointer to a timespec delta value
1273 * Takes a timespec offset measuring a suspend interval and properly
1274 * adds the sleep offset to the timekeeping variables.
1276 static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1277 struct timespec64 *delta)
1279 if (!timespec64_valid_strict(delta)) {
1280 printk_deferred(KERN_WARNING
1281 "__timekeeping_inject_sleeptime: Invalid "
1282 "sleep delta value!\n");
1285 tk_xtime_add(tk, delta);
1286 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1287 tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1288 tk_debug_account_sleep_time(delta);
1291 #if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1293 * We have three kinds of time sources to use for sleep time
1294 * injection, the preference order is:
1295 * 1) non-stop clocksource
1296 * 2) persistent clock (ie: RTC accessible when irqs are off)
1299 * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1300 * If system has neither 1) nor 2), 3) will be used finally.
1303 * If timekeeping has injected sleeptime via either 1) or 2),
1304 * 3) becomes needless, so in this case we don't need to call
1305 * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1308 bool timekeeping_rtc_skipresume(void)
1310 return sleeptime_injected;
1314 * 1) can be determined whether to use or not only when doing
1315 * timekeeping_resume() which is invoked after rtc_suspend(),
1316 * so we can't skip rtc_suspend() surely if system has 1).
1318 * But if system has 2), 2) will definitely be used, so in this
1319 * case we don't need to call rtc_suspend(), and this is what
1320 * timekeeping_rtc_skipsuspend() means.
1322 bool timekeeping_rtc_skipsuspend(void)
1324 return persistent_clock_exists;
1328 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1329 * @delta: pointer to a timespec64 delta value
1331 * This hook is for architectures that cannot support read_persistent_clock64
1332 * because their RTC/persistent clock is only accessible when irqs are enabled.
1333 * and also don't have an effective nonstop clocksource.
1335 * This function should only be called by rtc_resume(), and allows
1336 * a suspend offset to be injected into the timekeeping values.
1338 void timekeeping_inject_sleeptime64(struct timespec64 *delta)
1340 struct timekeeper *tk = &tk_core.timekeeper;
1341 unsigned long flags;
1343 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1344 write_seqcount_begin(&tk_core.seq);
1346 timekeeping_forward_now(tk);
1348 __timekeeping_inject_sleeptime(tk, delta);
1350 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1352 write_seqcount_end(&tk_core.seq);
1353 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1355 /* signal hrtimers about time change */
1361 * timekeeping_resume - Resumes the generic timekeeping subsystem.
1363 void timekeeping_resume(void)
1365 struct timekeeper *tk = &tk_core.timekeeper;
1366 struct clocksource *clock = tk->tkr_mono.clock;
1367 unsigned long flags;
1368 struct timespec64 ts_new, ts_delta;
1369 cycle_t cycle_now, cycle_delta;
1371 sleeptime_injected = false;
1372 read_persistent_clock64(&ts_new);
1374 clockevents_resume();
1375 clocksource_resume();
1377 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1378 write_seqcount_begin(&tk_core.seq);
1381 * After system resumes, we need to calculate the suspended time and
1382 * compensate it for the OS time. There are 3 sources that could be
1383 * used: Nonstop clocksource during suspend, persistent clock and rtc
1386 * One specific platform may have 1 or 2 or all of them, and the
1387 * preference will be:
1388 * suspend-nonstop clocksource -> persistent clock -> rtc
1389 * The less preferred source will only be tried if there is no better
1390 * usable source. The rtc part is handled separately in rtc core code.
1392 cycle_now = tk->tkr_mono.read(clock);
1393 if ((clock->flags & CLOCK_SOURCE_SUSPEND_NONSTOP) &&
1394 cycle_now > tk->tkr_mono.cycle_last) {
1395 u64 num, max = ULLONG_MAX;
1396 u32 mult = clock->mult;
1397 u32 shift = clock->shift;
1400 cycle_delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last,
1404 * "cycle_delta * mutl" may cause 64 bits overflow, if the
1405 * suspended time is too long. In that case we need do the
1406 * 64 bits math carefully
1409 if (cycle_delta > max) {
1410 num = div64_u64(cycle_delta, max);
1411 nsec = (((u64) max * mult) >> shift) * num;
1412 cycle_delta -= num * max;
1414 nsec += ((u64) cycle_delta * mult) >> shift;
1416 ts_delta = ns_to_timespec64(nsec);
1417 sleeptime_injected = true;
1418 } else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1419 ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1420 sleeptime_injected = true;
1423 if (sleeptime_injected)
1424 __timekeeping_inject_sleeptime(tk, &ts_delta);
1426 /* Re-base the last cycle value */
1427 tk->tkr_mono.cycle_last = cycle_now;
1428 tk->tkr_raw.cycle_last = cycle_now;
1431 timekeeping_suspended = 0;
1432 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1433 write_seqcount_end(&tk_core.seq);
1434 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1436 touch_softlockup_watchdog();
1442 int timekeeping_suspend(void)
1444 struct timekeeper *tk = &tk_core.timekeeper;
1445 unsigned long flags;
1446 struct timespec64 delta, delta_delta;
1447 static struct timespec64 old_delta;
1449 read_persistent_clock64(&timekeeping_suspend_time);
1452 * On some systems the persistent_clock can not be detected at
1453 * timekeeping_init by its return value, so if we see a valid
1454 * value returned, update the persistent_clock_exists flag.
1456 if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1457 persistent_clock_exists = true;
1459 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1460 write_seqcount_begin(&tk_core.seq);
1461 timekeeping_forward_now(tk);
1462 timekeeping_suspended = 1;
1464 if (persistent_clock_exists) {
1466 * To avoid drift caused by repeated suspend/resumes,
1467 * which each can add ~1 second drift error,
1468 * try to compensate so the difference in system time
1469 * and persistent_clock time stays close to constant.
1471 delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1472 delta_delta = timespec64_sub(delta, old_delta);
1473 if (abs(delta_delta.tv_sec) >= 2) {
1475 * if delta_delta is too large, assume time correction
1476 * has occurred and set old_delta to the current delta.
1480 /* Otherwise try to adjust old_system to compensate */
1481 timekeeping_suspend_time =
1482 timespec64_add(timekeeping_suspend_time, delta_delta);
1486 timekeeping_update(tk, TK_MIRROR);
1487 halt_fast_timekeeper(tk);
1488 write_seqcount_end(&tk_core.seq);
1489 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1492 clocksource_suspend();
1493 clockevents_suspend();
1498 /* sysfs resume/suspend bits for timekeeping */
1499 static struct syscore_ops timekeeping_syscore_ops = {
1500 .resume = timekeeping_resume,
1501 .suspend = timekeeping_suspend,
1504 static int __init timekeeping_init_ops(void)
1506 register_syscore_ops(&timekeeping_syscore_ops);
1509 device_initcall(timekeeping_init_ops);
1512 * Apply a multiplier adjustment to the timekeeper
1514 static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1519 s64 interval = tk->cycle_interval;
1523 mult_adj = -mult_adj;
1524 interval = -interval;
1527 mult_adj <<= adj_scale;
1528 interval <<= adj_scale;
1529 offset <<= adj_scale;
1532 * So the following can be confusing.
1534 * To keep things simple, lets assume mult_adj == 1 for now.
1536 * When mult_adj != 1, remember that the interval and offset values
1537 * have been appropriately scaled so the math is the same.
1539 * The basic idea here is that we're increasing the multiplier
1540 * by one, this causes the xtime_interval to be incremented by
1541 * one cycle_interval. This is because:
1542 * xtime_interval = cycle_interval * mult
1543 * So if mult is being incremented by one:
1544 * xtime_interval = cycle_interval * (mult + 1)
1546 * xtime_interval = (cycle_interval * mult) + cycle_interval
1547 * Which can be shortened to:
1548 * xtime_interval += cycle_interval
1550 * So offset stores the non-accumulated cycles. Thus the current
1551 * time (in shifted nanoseconds) is:
1552 * now = (offset * adj) + xtime_nsec
1553 * Now, even though we're adjusting the clock frequency, we have
1554 * to keep time consistent. In other words, we can't jump back
1555 * in time, and we also want to avoid jumping forward in time.
1557 * So given the same offset value, we need the time to be the same
1558 * both before and after the freq adjustment.
1559 * now = (offset * adj_1) + xtime_nsec_1
1560 * now = (offset * adj_2) + xtime_nsec_2
1562 * (offset * adj_1) + xtime_nsec_1 =
1563 * (offset * adj_2) + xtime_nsec_2
1567 * (offset * adj_1) + xtime_nsec_1 =
1568 * (offset * (adj_1+1)) + xtime_nsec_2
1569 * (offset * adj_1) + xtime_nsec_1 =
1570 * (offset * adj_1) + offset + xtime_nsec_2
1571 * Canceling the sides:
1572 * xtime_nsec_1 = offset + xtime_nsec_2
1574 * xtime_nsec_2 = xtime_nsec_1 - offset
1575 * Which simplfies to:
1576 * xtime_nsec -= offset
1578 * XXX - TODO: Doc ntp_error calculation.
1580 if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
1581 /* NTP adjustment caused clocksource mult overflow */
1586 tk->tkr_mono.mult += mult_adj;
1587 tk->xtime_interval += interval;
1588 tk->tkr_mono.xtime_nsec -= offset;
1589 tk->ntp_error -= (interval - offset) << tk->ntp_error_shift;
1593 * Calculate the multiplier adjustment needed to match the frequency
1596 static __always_inline void timekeeping_freqadjust(struct timekeeper *tk,
1599 s64 interval = tk->cycle_interval;
1600 s64 xinterval = tk->xtime_interval;
1605 /* Remove any current error adj from freq calculation */
1606 if (tk->ntp_err_mult)
1607 xinterval -= tk->cycle_interval;
1609 tk->ntp_tick = ntp_tick_length();
1611 /* Calculate current error per tick */
1612 tick_error = ntp_tick_length() >> tk->ntp_error_shift;
1613 tick_error -= (xinterval + tk->xtime_remainder);
1615 /* Don't worry about correcting it if its small */
1616 if (likely((tick_error >= 0) && (tick_error <= interval)))
1619 /* preserve the direction of correction */
1620 negative = (tick_error < 0);
1622 /* Sort out the magnitude of the correction */
1623 tick_error = abs(tick_error);
1624 for (adj = 0; tick_error > interval; adj++)
1627 /* scale the corrections */
1628 timekeeping_apply_adjustment(tk, offset, negative, adj);
1632 * Adjust the timekeeper's multiplier to the correct frequency
1633 * and also to reduce the accumulated error value.
1635 static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
1637 /* Correct for the current frequency error */
1638 timekeeping_freqadjust(tk, offset);
1640 /* Next make a small adjustment to fix any cumulative error */
1641 if (!tk->ntp_err_mult && (tk->ntp_error > 0)) {
1642 tk->ntp_err_mult = 1;
1643 timekeeping_apply_adjustment(tk, offset, 0, 0);
1644 } else if (tk->ntp_err_mult && (tk->ntp_error <= 0)) {
1645 /* Undo any existing error adjustment */
1646 timekeeping_apply_adjustment(tk, offset, 1, 0);
1647 tk->ntp_err_mult = 0;
1650 if (unlikely(tk->tkr_mono.clock->maxadj &&
1651 (abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
1652 > tk->tkr_mono.clock->maxadj))) {
1653 printk_once(KERN_WARNING
1654 "Adjusting %s more than 11%% (%ld vs %ld)\n",
1655 tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
1656 (long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
1660 * It may be possible that when we entered this function, xtime_nsec
1661 * was very small. Further, if we're slightly speeding the clocksource
1662 * in the code above, its possible the required corrective factor to
1663 * xtime_nsec could cause it to underflow.
1665 * Now, since we already accumulated the second, cannot simply roll
1666 * the accumulated second back, since the NTP subsystem has been
1667 * notified via second_overflow. So instead we push xtime_nsec forward
1668 * by the amount we underflowed, and add that amount into the error.
1670 * We'll correct this error next time through this function, when
1671 * xtime_nsec is not as small.
1673 if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
1674 s64 neg = -(s64)tk->tkr_mono.xtime_nsec;
1675 tk->tkr_mono.xtime_nsec = 0;
1676 tk->ntp_error += neg << tk->ntp_error_shift;
1681 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
1683 * Helper function that accumulates a the nsecs greater then a second
1684 * from the xtime_nsec field to the xtime_secs field.
1685 * It also calls into the NTP code to handle leapsecond processing.
1688 static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
1690 u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
1691 unsigned int clock_set = 0;
1693 while (tk->tkr_mono.xtime_nsec >= nsecps) {
1696 tk->tkr_mono.xtime_nsec -= nsecps;
1699 /* Figure out if its a leap sec and apply if needed */
1700 leap = second_overflow(tk->xtime_sec);
1701 if (unlikely(leap)) {
1702 struct timespec64 ts;
1704 tk->xtime_sec += leap;
1708 tk_set_wall_to_mono(tk,
1709 timespec64_sub(tk->wall_to_monotonic, ts));
1711 __timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
1713 clock_set = TK_CLOCK_WAS_SET;
1720 * logarithmic_accumulation - shifted accumulation of cycles
1722 * This functions accumulates a shifted interval of cycles into
1723 * into a shifted interval nanoseconds. Allows for O(log) accumulation
1726 * Returns the unconsumed cycles.
1728 static cycle_t logarithmic_accumulation(struct timekeeper *tk, cycle_t offset,
1730 unsigned int *clock_set)
1732 cycle_t interval = tk->cycle_interval << shift;
1735 /* If the offset is smaller then a shifted interval, do nothing */
1736 if (offset < interval)
1739 /* Accumulate one shifted interval */
1741 tk->tkr_mono.cycle_last += interval;
1742 tk->tkr_raw.cycle_last += interval;
1744 tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
1745 *clock_set |= accumulate_nsecs_to_secs(tk);
1747 /* Accumulate raw time */
1748 raw_nsecs = (u64)tk->raw_interval << shift;
1749 raw_nsecs += tk->raw_time.tv_nsec;
1750 if (raw_nsecs >= NSEC_PER_SEC) {
1751 u64 raw_secs = raw_nsecs;
1752 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
1753 tk->raw_time.tv_sec += raw_secs;
1755 tk->raw_time.tv_nsec = raw_nsecs;
1757 /* Accumulate error between NTP and clock interval */
1758 tk->ntp_error += tk->ntp_tick << shift;
1759 tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
1760 (tk->ntp_error_shift + shift);
1766 * update_wall_time - Uses the current clocksource to increment the wall time
1769 void update_wall_time(void)
1771 struct timekeeper *real_tk = &tk_core.timekeeper;
1772 struct timekeeper *tk = &shadow_timekeeper;
1774 int shift = 0, maxshift;
1775 unsigned int clock_set = 0;
1776 unsigned long flags;
1778 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1780 /* Make sure we're fully resumed: */
1781 if (unlikely(timekeeping_suspended))
1784 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
1785 offset = real_tk->cycle_interval;
1787 offset = clocksource_delta(tk->tkr_mono.read(tk->tkr_mono.clock),
1788 tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
1791 /* Check if there's really nothing to do */
1792 if (offset < real_tk->cycle_interval)
1795 /* Do some additional sanity checking */
1796 timekeeping_check_update(real_tk, offset);
1799 * With NO_HZ we may have to accumulate many cycle_intervals
1800 * (think "ticks") worth of time at once. To do this efficiently,
1801 * we calculate the largest doubling multiple of cycle_intervals
1802 * that is smaller than the offset. We then accumulate that
1803 * chunk in one go, and then try to consume the next smaller
1806 shift = ilog2(offset) - ilog2(tk->cycle_interval);
1807 shift = max(0, shift);
1808 /* Bound shift to one less than what overflows tick_length */
1809 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
1810 shift = min(shift, maxshift);
1811 while (offset >= tk->cycle_interval) {
1812 offset = logarithmic_accumulation(tk, offset, shift,
1814 if (offset < tk->cycle_interval<<shift)
1818 /* correct the clock when NTP error is too big */
1819 timekeeping_adjust(tk, offset);
1822 * XXX This can be killed once everyone converts
1823 * to the new update_vsyscall.
1825 old_vsyscall_fixup(tk);
1828 * Finally, make sure that after the rounding
1829 * xtime_nsec isn't larger than NSEC_PER_SEC
1831 clock_set |= accumulate_nsecs_to_secs(tk);
1833 write_seqcount_begin(&tk_core.seq);
1835 * Update the real timekeeper.
1837 * We could avoid this memcpy by switching pointers, but that
1838 * requires changes to all other timekeeper usage sites as
1839 * well, i.e. move the timekeeper pointer getter into the
1840 * spinlocked/seqcount protected sections. And we trade this
1841 * memcpy under the tk_core.seq against one before we start
1844 memcpy(real_tk, tk, sizeof(*tk));
1845 timekeeping_update(real_tk, clock_set);
1846 write_seqcount_end(&tk_core.seq);
1848 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1850 /* Have to call _delayed version, since in irq context*/
1851 clock_was_set_delayed();
1855 * getboottime64 - Return the real time of system boot.
1856 * @ts: pointer to the timespec64 to be set
1858 * Returns the wall-time of boot in a timespec64.
1860 * This is based on the wall_to_monotonic offset and the total suspend
1861 * time. Calls to settimeofday will affect the value returned (which
1862 * basically means that however wrong your real time clock is at boot time,
1863 * you get the right time here).
1865 void getboottime64(struct timespec64 *ts)
1867 struct timekeeper *tk = &tk_core.timekeeper;
1868 ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
1870 *ts = ktime_to_timespec64(t);
1872 EXPORT_SYMBOL_GPL(getboottime64);
1874 unsigned long get_seconds(void)
1876 struct timekeeper *tk = &tk_core.timekeeper;
1878 return tk->xtime_sec;
1880 EXPORT_SYMBOL(get_seconds);
1882 struct timespec __current_kernel_time(void)
1884 struct timekeeper *tk = &tk_core.timekeeper;
1886 return timespec64_to_timespec(tk_xtime(tk));
1889 struct timespec current_kernel_time(void)
1891 struct timekeeper *tk = &tk_core.timekeeper;
1892 struct timespec64 now;
1896 seq = read_seqcount_begin(&tk_core.seq);
1899 } while (read_seqcount_retry(&tk_core.seq, seq));
1901 return timespec64_to_timespec(now);
1903 EXPORT_SYMBOL(current_kernel_time);
1905 struct timespec64 get_monotonic_coarse64(void)
1907 struct timekeeper *tk = &tk_core.timekeeper;
1908 struct timespec64 now, mono;
1912 seq = read_seqcount_begin(&tk_core.seq);
1915 mono = tk->wall_to_monotonic;
1916 } while (read_seqcount_retry(&tk_core.seq, seq));
1918 set_normalized_timespec64(&now, now.tv_sec + mono.tv_sec,
1919 now.tv_nsec + mono.tv_nsec);
1925 * Must hold jiffies_lock
1927 void do_timer(unsigned long ticks)
1929 jiffies_64 += ticks;
1930 calc_global_load(ticks);
1934 * ktime_get_update_offsets_now - hrtimer helper
1935 * @cwsseq: pointer to check and store the clock was set sequence number
1936 * @offs_real: pointer to storage for monotonic -> realtime offset
1937 * @offs_boot: pointer to storage for monotonic -> boottime offset
1938 * @offs_tai: pointer to storage for monotonic -> clock tai offset
1940 * Returns current monotonic time and updates the offsets if the
1941 * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
1944 * Called from hrtimer_interrupt() or retrigger_next_event()
1946 ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real,
1947 ktime_t *offs_boot, ktime_t *offs_tai)
1949 struct timekeeper *tk = &tk_core.timekeeper;
1955 seq = read_seqcount_begin(&tk_core.seq);
1957 base = tk->tkr_mono.base;
1958 nsecs = timekeeping_get_ns(&tk->tkr_mono);
1959 if (*cwsseq != tk->clock_was_set_seq) {
1960 *cwsseq = tk->clock_was_set_seq;
1961 *offs_real = tk->offs_real;
1962 *offs_boot = tk->offs_boot;
1963 *offs_tai = tk->offs_tai;
1965 } while (read_seqcount_retry(&tk_core.seq, seq));
1967 return ktime_add_ns(base, nsecs);
1971 * do_adjtimex() - Accessor function to NTP __do_adjtimex function
1973 int do_adjtimex(struct timex *txc)
1975 struct timekeeper *tk = &tk_core.timekeeper;
1976 unsigned long flags;
1977 struct timespec64 ts;
1981 /* Validate the data before disabling interrupts */
1982 ret = ntp_validate_timex(txc);
1986 if (txc->modes & ADJ_SETOFFSET) {
1987 struct timespec delta;
1988 delta.tv_sec = txc->time.tv_sec;
1989 delta.tv_nsec = txc->time.tv_usec;
1990 if (!(txc->modes & ADJ_NANO))
1991 delta.tv_nsec *= 1000;
1992 ret = timekeeping_inject_offset(&delta);
1997 getnstimeofday64(&ts);
1999 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2000 write_seqcount_begin(&tk_core.seq);
2002 orig_tai = tai = tk->tai_offset;
2003 ret = __do_adjtimex(txc, &ts, &tai);
2005 if (tai != orig_tai) {
2006 __timekeeping_set_tai_offset(tk, tai);
2007 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
2009 write_seqcount_end(&tk_core.seq);
2010 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2012 if (tai != orig_tai)
2015 ntp_notify_cmos_timer();
2020 #ifdef CONFIG_NTP_PPS
2022 * hardpps() - Accessor function to NTP __hardpps function
2024 void hardpps(const struct timespec *phase_ts, const struct timespec *raw_ts)
2026 unsigned long flags;
2028 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2029 write_seqcount_begin(&tk_core.seq);
2031 __hardpps(phase_ts, raw_ts);
2033 write_seqcount_end(&tk_core.seq);
2034 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2036 EXPORT_SYMBOL(hardpps);
2040 * xtime_update() - advances the timekeeping infrastructure
2041 * @ticks: number of ticks, that have elapsed since the last call.
2043 * Must be called with interrupts disabled.
2045 void xtime_update(unsigned long ticks)
2047 write_seqlock(&jiffies_lock);
2049 write_sequnlock(&jiffies_lock);