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 != timespec64_to_ktime(tmp));
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, u64 offset)
125 u64 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 '%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 u64 timekeeping_get_delta(struct tk_read_base *tkr)
163 struct timekeeper *tk = &tk_core.timekeeper;
164 u64 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, u64 offset)
205 static inline u64 timekeeping_get_delta(struct tk_read_base *tkr)
207 u64 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 ++tk->cs_was_changed_seq;
237 old_clock = tk->tkr_mono.clock;
238 tk->tkr_mono.clock = clock;
239 tk->tkr_mono.read = clock->read;
240 tk->tkr_mono.mask = clock->mask;
241 tk->tkr_mono.cycle_last = tk->tkr_mono.read(clock);
243 tk->tkr_raw.clock = clock;
244 tk->tkr_raw.read = clock->read;
245 tk->tkr_raw.mask = clock->mask;
246 tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last;
248 /* Do the ns -> cycle conversion first, using original mult */
249 tmp = NTP_INTERVAL_LENGTH;
250 tmp <<= clock->shift;
252 tmp += clock->mult/2;
253 do_div(tmp, clock->mult);
257 interval = (u64) tmp;
258 tk->cycle_interval = interval;
260 /* Go back from cycles -> shifted ns */
261 tk->xtime_interval = interval * clock->mult;
262 tk->xtime_remainder = ntpinterval - tk->xtime_interval;
263 tk->raw_interval = (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 u64 timekeeping_delta_to_ns(struct tk_read_base *tkr, u64 delta)
305 nsec = delta * tkr->mult + tkr->xtime_nsec;
308 /* If arch requires, add in get_arch_timeoffset() */
309 return nsec + arch_gettimeoffset();
312 static inline u64 timekeeping_get_ns(struct tk_read_base *tkr)
316 delta = timekeeping_get_delta(tkr);
317 return timekeeping_delta_to_ns(tkr, delta);
320 static inline u64 timekeeping_cycles_to_ns(struct tk_read_base *tkr, u64 cycles)
324 /* calculate the delta since the last update_wall_time */
325 delta = clocksource_delta(cycles, tkr->cycle_last, tkr->mask);
326 return timekeeping_delta_to_ns(tkr, delta);
330 * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
331 * @tkr: Timekeeping readout base from which we take the update
333 * We want to use this from any context including NMI and tracing /
334 * instrumenting the timekeeping code itself.
336 * Employ the latch technique; see @raw_write_seqcount_latch.
338 * So if a NMI hits the update of base[0] then it will use base[1]
339 * which is still consistent. In the worst case this can result is a
340 * slightly wrong timestamp (a few nanoseconds). See
341 * @ktime_get_mono_fast_ns.
343 static void update_fast_timekeeper(struct tk_read_base *tkr, struct tk_fast *tkf)
345 struct tk_read_base *base = tkf->base;
347 /* Force readers off to base[1] */
348 raw_write_seqcount_latch(&tkf->seq);
351 memcpy(base, tkr, sizeof(*base));
353 /* Force readers back to base[0] */
354 raw_write_seqcount_latch(&tkf->seq);
357 memcpy(base + 1, base, sizeof(*base));
361 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
363 * This timestamp is not guaranteed to be monotonic across an update.
364 * The timestamp is calculated by:
366 * now = base_mono + clock_delta * slope
368 * So if the update lowers the slope, readers who are forced to the
369 * not yet updated second array are still using the old steeper slope.
378 * |12345678---> reader order
384 * So reader 6 will observe time going backwards versus reader 5.
386 * While other CPUs are likely to be able observe that, the only way
387 * for a CPU local observation is when an NMI hits in the middle of
388 * the update. Timestamps taken from that NMI context might be ahead
389 * of the following timestamps. Callers need to be aware of that and
392 static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf)
394 struct tk_read_base *tkr;
399 seq = raw_read_seqcount_latch(&tkf->seq);
400 tkr = tkf->base + (seq & 0x01);
401 now = ktime_to_ns(tkr->base);
403 now += timekeeping_delta_to_ns(tkr,
405 tkr->read(tkr->clock),
408 } while (read_seqcount_retry(&tkf->seq, seq));
413 u64 ktime_get_mono_fast_ns(void)
415 return __ktime_get_fast_ns(&tk_fast_mono);
417 EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
419 u64 ktime_get_raw_fast_ns(void)
421 return __ktime_get_fast_ns(&tk_fast_raw);
423 EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns);
426 * ktime_get_boot_fast_ns - NMI safe and fast access to boot clock.
428 * To keep it NMI safe since we're accessing from tracing, we're not using a
429 * separate timekeeper with updates to monotonic clock and boot offset
430 * protected with seqlocks. This has the following minor side effects:
432 * (1) Its possible that a timestamp be taken after the boot offset is updated
433 * but before the timekeeper is updated. If this happens, the new boot offset
434 * is added to the old timekeeping making the clock appear to update slightly
437 * timekeeping_inject_sleeptime64()
438 * __timekeeping_inject_sleeptime(tk, delta);
440 * timekeeping_update(tk, TK_CLEAR_NTP...);
442 * (2) On 32-bit systems, the 64-bit boot offset (tk->offs_boot) may be
443 * partially updated. Since the tk->offs_boot update is a rare event, this
444 * should be a rare occurrence which postprocessing should be able to handle.
446 u64 notrace ktime_get_boot_fast_ns(void)
448 struct timekeeper *tk = &tk_core.timekeeper;
450 return (ktime_get_mono_fast_ns() + ktime_to_ns(tk->offs_boot));
452 EXPORT_SYMBOL_GPL(ktime_get_boot_fast_ns);
454 /* Suspend-time cycles value for halted fast timekeeper. */
455 static u64 cycles_at_suspend;
457 static u64 dummy_clock_read(struct clocksource *cs)
459 return cycles_at_suspend;
463 * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
464 * @tk: Timekeeper to snapshot.
466 * It generally is unsafe to access the clocksource after timekeeping has been
467 * suspended, so take a snapshot of the readout base of @tk and use it as the
468 * fast timekeeper's readout base while suspended. It will return the same
469 * number of cycles every time until timekeeping is resumed at which time the
470 * proper readout base for the fast timekeeper will be restored automatically.
472 static void halt_fast_timekeeper(struct timekeeper *tk)
474 static struct tk_read_base tkr_dummy;
475 struct tk_read_base *tkr = &tk->tkr_mono;
477 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
478 cycles_at_suspend = tkr->read(tkr->clock);
479 tkr_dummy.read = dummy_clock_read;
480 update_fast_timekeeper(&tkr_dummy, &tk_fast_mono);
483 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
484 tkr_dummy.read = dummy_clock_read;
485 update_fast_timekeeper(&tkr_dummy, &tk_fast_raw);
488 #ifdef CONFIG_GENERIC_TIME_VSYSCALL_OLD
490 static inline void update_vsyscall(struct timekeeper *tk)
492 struct timespec xt, wm;
494 xt = timespec64_to_timespec(tk_xtime(tk));
495 wm = timespec64_to_timespec(tk->wall_to_monotonic);
496 update_vsyscall_old(&xt, &wm, tk->tkr_mono.clock, tk->tkr_mono.mult,
497 tk->tkr_mono.cycle_last);
500 static inline void old_vsyscall_fixup(struct timekeeper *tk)
505 * Store only full nanoseconds into xtime_nsec after rounding
506 * it up and add the remainder to the error difference.
507 * XXX - This is necessary to avoid small 1ns inconsistnecies caused
508 * by truncating the remainder in vsyscalls. However, it causes
509 * additional work to be done in timekeeping_adjust(). Once
510 * the vsyscall implementations are converted to use xtime_nsec
511 * (shifted nanoseconds), and CONFIG_GENERIC_TIME_VSYSCALL_OLD
512 * users are removed, this can be killed.
514 remainder = tk->tkr_mono.xtime_nsec & ((1ULL << tk->tkr_mono.shift) - 1);
515 if (remainder != 0) {
516 tk->tkr_mono.xtime_nsec -= remainder;
517 tk->tkr_mono.xtime_nsec += 1ULL << tk->tkr_mono.shift;
518 tk->ntp_error += remainder << tk->ntp_error_shift;
519 tk->ntp_error -= (1ULL << tk->tkr_mono.shift) << tk->ntp_error_shift;
523 #define old_vsyscall_fixup(tk)
526 static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
528 static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
530 raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
534 * pvclock_gtod_register_notifier - register a pvclock timedata update listener
536 int pvclock_gtod_register_notifier(struct notifier_block *nb)
538 struct timekeeper *tk = &tk_core.timekeeper;
542 raw_spin_lock_irqsave(&timekeeper_lock, flags);
543 ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
544 update_pvclock_gtod(tk, true);
545 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
549 EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
552 * pvclock_gtod_unregister_notifier - unregister a pvclock
553 * timedata update listener
555 int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
560 raw_spin_lock_irqsave(&timekeeper_lock, flags);
561 ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
562 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
566 EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
569 * tk_update_leap_state - helper to update the next_leap_ktime
571 static inline void tk_update_leap_state(struct timekeeper *tk)
573 tk->next_leap_ktime = ntp_get_next_leap();
574 if (tk->next_leap_ktime != KTIME_MAX)
575 /* Convert to monotonic time */
576 tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real);
580 * Update the ktime_t based scalar nsec members of the timekeeper
582 static inline void tk_update_ktime_data(struct timekeeper *tk)
588 * The xtime based monotonic readout is:
589 * nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
590 * The ktime based monotonic readout is:
591 * nsec = base_mono + now();
592 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
594 seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
595 nsec = (u32) tk->wall_to_monotonic.tv_nsec;
596 tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
598 /* Update the monotonic raw base */
599 tk->tkr_raw.base = timespec64_to_ktime(tk->raw_time);
602 * The sum of the nanoseconds portions of xtime and
603 * wall_to_monotonic can be greater/equal one second. Take
604 * this into account before updating tk->ktime_sec.
606 nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
607 if (nsec >= NSEC_PER_SEC)
609 tk->ktime_sec = seconds;
612 /* must hold timekeeper_lock */
613 static void timekeeping_update(struct timekeeper *tk, unsigned int action)
615 if (action & TK_CLEAR_NTP) {
620 tk_update_leap_state(tk);
621 tk_update_ktime_data(tk);
624 update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
626 update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono);
627 update_fast_timekeeper(&tk->tkr_raw, &tk_fast_raw);
629 if (action & TK_CLOCK_WAS_SET)
630 tk->clock_was_set_seq++;
632 * The mirroring of the data to the shadow-timekeeper needs
633 * to happen last here to ensure we don't over-write the
634 * timekeeper structure on the next update with stale data
636 if (action & TK_MIRROR)
637 memcpy(&shadow_timekeeper, &tk_core.timekeeper,
638 sizeof(tk_core.timekeeper));
642 * timekeeping_forward_now - update clock to the current time
644 * Forward the current clock to update its state since the last call to
645 * update_wall_time(). This is useful before significant clock changes,
646 * as it avoids having to deal with this time offset explicitly.
648 static void timekeeping_forward_now(struct timekeeper *tk)
650 struct clocksource *clock = tk->tkr_mono.clock;
651 u64 cycle_now, delta;
654 cycle_now = tk->tkr_mono.read(clock);
655 delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
656 tk->tkr_mono.cycle_last = cycle_now;
657 tk->tkr_raw.cycle_last = cycle_now;
659 tk->tkr_mono.xtime_nsec += delta * tk->tkr_mono.mult;
661 /* If arch requires, add in get_arch_timeoffset() */
662 tk->tkr_mono.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr_mono.shift;
664 tk_normalize_xtime(tk);
666 nsec = clocksource_cyc2ns(delta, tk->tkr_raw.mult, tk->tkr_raw.shift);
667 timespec64_add_ns(&tk->raw_time, nsec);
671 * __getnstimeofday64 - Returns the time of day in a timespec64.
672 * @ts: pointer to the timespec to be set
674 * Updates the time of day in the timespec.
675 * Returns 0 on success, or -ve when suspended (timespec will be undefined).
677 int __getnstimeofday64(struct timespec64 *ts)
679 struct timekeeper *tk = &tk_core.timekeeper;
684 seq = read_seqcount_begin(&tk_core.seq);
686 ts->tv_sec = tk->xtime_sec;
687 nsecs = timekeeping_get_ns(&tk->tkr_mono);
689 } while (read_seqcount_retry(&tk_core.seq, seq));
692 timespec64_add_ns(ts, nsecs);
695 * Do not bail out early, in case there were callers still using
696 * the value, even in the face of the WARN_ON.
698 if (unlikely(timekeeping_suspended))
702 EXPORT_SYMBOL(__getnstimeofday64);
705 * getnstimeofday64 - Returns the time of day in a timespec64.
706 * @ts: pointer to the timespec64 to be set
708 * Returns the time of day in a timespec64 (WARN if suspended).
710 void getnstimeofday64(struct timespec64 *ts)
712 WARN_ON(__getnstimeofday64(ts));
714 EXPORT_SYMBOL(getnstimeofday64);
716 ktime_t ktime_get(void)
718 struct timekeeper *tk = &tk_core.timekeeper;
723 WARN_ON(timekeeping_suspended);
726 seq = read_seqcount_begin(&tk_core.seq);
727 base = tk->tkr_mono.base;
728 nsecs = timekeeping_get_ns(&tk->tkr_mono);
730 } while (read_seqcount_retry(&tk_core.seq, seq));
732 return ktime_add_ns(base, nsecs);
734 EXPORT_SYMBOL_GPL(ktime_get);
736 u32 ktime_get_resolution_ns(void)
738 struct timekeeper *tk = &tk_core.timekeeper;
742 WARN_ON(timekeeping_suspended);
745 seq = read_seqcount_begin(&tk_core.seq);
746 nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift;
747 } while (read_seqcount_retry(&tk_core.seq, seq));
751 EXPORT_SYMBOL_GPL(ktime_get_resolution_ns);
753 static ktime_t *offsets[TK_OFFS_MAX] = {
754 [TK_OFFS_REAL] = &tk_core.timekeeper.offs_real,
755 [TK_OFFS_BOOT] = &tk_core.timekeeper.offs_boot,
756 [TK_OFFS_TAI] = &tk_core.timekeeper.offs_tai,
759 ktime_t ktime_get_with_offset(enum tk_offsets offs)
761 struct timekeeper *tk = &tk_core.timekeeper;
763 ktime_t base, *offset = offsets[offs];
766 WARN_ON(timekeeping_suspended);
769 seq = read_seqcount_begin(&tk_core.seq);
770 base = ktime_add(tk->tkr_mono.base, *offset);
771 nsecs = timekeeping_get_ns(&tk->tkr_mono);
773 } while (read_seqcount_retry(&tk_core.seq, seq));
775 return ktime_add_ns(base, nsecs);
778 EXPORT_SYMBOL_GPL(ktime_get_with_offset);
781 * ktime_mono_to_any() - convert mononotic time to any other time
782 * @tmono: time to convert.
783 * @offs: which offset to use
785 ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
787 ktime_t *offset = offsets[offs];
792 seq = read_seqcount_begin(&tk_core.seq);
793 tconv = ktime_add(tmono, *offset);
794 } while (read_seqcount_retry(&tk_core.seq, seq));
798 EXPORT_SYMBOL_GPL(ktime_mono_to_any);
801 * ktime_get_raw - Returns the raw monotonic time in ktime_t format
803 ktime_t ktime_get_raw(void)
805 struct timekeeper *tk = &tk_core.timekeeper;
811 seq = read_seqcount_begin(&tk_core.seq);
812 base = tk->tkr_raw.base;
813 nsecs = timekeeping_get_ns(&tk->tkr_raw);
815 } while (read_seqcount_retry(&tk_core.seq, seq));
817 return ktime_add_ns(base, nsecs);
819 EXPORT_SYMBOL_GPL(ktime_get_raw);
822 * ktime_get_ts64 - get the monotonic clock in timespec64 format
823 * @ts: pointer to timespec variable
825 * The function calculates the monotonic clock from the realtime
826 * clock and the wall_to_monotonic offset and stores the result
827 * in normalized timespec64 format in the variable pointed to by @ts.
829 void ktime_get_ts64(struct timespec64 *ts)
831 struct timekeeper *tk = &tk_core.timekeeper;
832 struct timespec64 tomono;
836 WARN_ON(timekeeping_suspended);
839 seq = read_seqcount_begin(&tk_core.seq);
840 ts->tv_sec = tk->xtime_sec;
841 nsec = timekeeping_get_ns(&tk->tkr_mono);
842 tomono = tk->wall_to_monotonic;
844 } while (read_seqcount_retry(&tk_core.seq, seq));
846 ts->tv_sec += tomono.tv_sec;
848 timespec64_add_ns(ts, nsec + tomono.tv_nsec);
850 EXPORT_SYMBOL_GPL(ktime_get_ts64);
853 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
855 * Returns the seconds portion of CLOCK_MONOTONIC with a single non
856 * serialized read. tk->ktime_sec is of type 'unsigned long' so this
857 * works on both 32 and 64 bit systems. On 32 bit systems the readout
858 * covers ~136 years of uptime which should be enough to prevent
859 * premature wrap arounds.
861 time64_t ktime_get_seconds(void)
863 struct timekeeper *tk = &tk_core.timekeeper;
865 WARN_ON(timekeeping_suspended);
866 return tk->ktime_sec;
868 EXPORT_SYMBOL_GPL(ktime_get_seconds);
871 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
873 * Returns the wall clock seconds since 1970. This replaces the
874 * get_seconds() interface which is not y2038 safe on 32bit systems.
876 * For 64bit systems the fast access to tk->xtime_sec is preserved. On
877 * 32bit systems the access must be protected with the sequence
878 * counter to provide "atomic" access to the 64bit tk->xtime_sec
881 time64_t ktime_get_real_seconds(void)
883 struct timekeeper *tk = &tk_core.timekeeper;
887 if (IS_ENABLED(CONFIG_64BIT))
888 return tk->xtime_sec;
891 seq = read_seqcount_begin(&tk_core.seq);
892 seconds = tk->xtime_sec;
894 } while (read_seqcount_retry(&tk_core.seq, seq));
898 EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
901 * __ktime_get_real_seconds - The same as ktime_get_real_seconds
902 * but without the sequence counter protect. This internal function
903 * is called just when timekeeping lock is already held.
905 time64_t __ktime_get_real_seconds(void)
907 struct timekeeper *tk = &tk_core.timekeeper;
909 return tk->xtime_sec;
913 * ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter
914 * @systime_snapshot: pointer to struct receiving the system time snapshot
916 void ktime_get_snapshot(struct system_time_snapshot *systime_snapshot)
918 struct timekeeper *tk = &tk_core.timekeeper;
926 WARN_ON_ONCE(timekeeping_suspended);
929 seq = read_seqcount_begin(&tk_core.seq);
931 now = tk->tkr_mono.read(tk->tkr_mono.clock);
932 systime_snapshot->cs_was_changed_seq = tk->cs_was_changed_seq;
933 systime_snapshot->clock_was_set_seq = tk->clock_was_set_seq;
934 base_real = ktime_add(tk->tkr_mono.base,
935 tk_core.timekeeper.offs_real);
936 base_raw = tk->tkr_raw.base;
937 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, now);
938 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw, now);
939 } while (read_seqcount_retry(&tk_core.seq, seq));
941 systime_snapshot->cycles = now;
942 systime_snapshot->real = ktime_add_ns(base_real, nsec_real);
943 systime_snapshot->raw = ktime_add_ns(base_raw, nsec_raw);
945 EXPORT_SYMBOL_GPL(ktime_get_snapshot);
947 /* Scale base by mult/div checking for overflow */
948 static int scale64_check_overflow(u64 mult, u64 div, u64 *base)
952 tmp = div64_u64_rem(*base, div, &rem);
954 if (((int)sizeof(u64)*8 - fls64(mult) < fls64(tmp)) ||
955 ((int)sizeof(u64)*8 - fls64(mult) < fls64(rem)))
966 * adjust_historical_crosststamp - adjust crosstimestamp previous to current interval
967 * @history: Snapshot representing start of history
968 * @partial_history_cycles: Cycle offset into history (fractional part)
969 * @total_history_cycles: Total history length in cycles
970 * @discontinuity: True indicates clock was set on history period
971 * @ts: Cross timestamp that should be adjusted using
972 * partial/total ratio
974 * Helper function used by get_device_system_crosststamp() to correct the
975 * crosstimestamp corresponding to the start of the current interval to the
976 * system counter value (timestamp point) provided by the driver. The
977 * total_history_* quantities are the total history starting at the provided
978 * reference point and ending at the start of the current interval. The cycle
979 * count between the driver timestamp point and the start of the current
980 * interval is partial_history_cycles.
982 static int adjust_historical_crosststamp(struct system_time_snapshot *history,
983 u64 partial_history_cycles,
984 u64 total_history_cycles,
986 struct system_device_crosststamp *ts)
988 struct timekeeper *tk = &tk_core.timekeeper;
989 u64 corr_raw, corr_real;
993 if (total_history_cycles == 0 || partial_history_cycles == 0)
996 /* Interpolate shortest distance from beginning or end of history */
997 interp_forward = partial_history_cycles > total_history_cycles/2 ?
999 partial_history_cycles = interp_forward ?
1000 total_history_cycles - partial_history_cycles :
1001 partial_history_cycles;
1004 * Scale the monotonic raw time delta by:
1005 * partial_history_cycles / total_history_cycles
1007 corr_raw = (u64)ktime_to_ns(
1008 ktime_sub(ts->sys_monoraw, history->raw));
1009 ret = scale64_check_overflow(partial_history_cycles,
1010 total_history_cycles, &corr_raw);
1015 * If there is a discontinuity in the history, scale monotonic raw
1017 * mult(real)/mult(raw) yielding the realtime correction
1018 * Otherwise, calculate the realtime correction similar to monotonic
1021 if (discontinuity) {
1022 corr_real = mul_u64_u32_div
1023 (corr_raw, tk->tkr_mono.mult, tk->tkr_raw.mult);
1025 corr_real = (u64)ktime_to_ns(
1026 ktime_sub(ts->sys_realtime, history->real));
1027 ret = scale64_check_overflow(partial_history_cycles,
1028 total_history_cycles, &corr_real);
1033 /* Fixup monotonic raw and real time time values */
1034 if (interp_forward) {
1035 ts->sys_monoraw = ktime_add_ns(history->raw, corr_raw);
1036 ts->sys_realtime = ktime_add_ns(history->real, corr_real);
1038 ts->sys_monoraw = ktime_sub_ns(ts->sys_monoraw, corr_raw);
1039 ts->sys_realtime = ktime_sub_ns(ts->sys_realtime, corr_real);
1046 * cycle_between - true if test occurs chronologically between before and after
1048 static bool cycle_between(u64 before, u64 test, u64 after)
1050 if (test > before && test < after)
1052 if (test < before && before > after)
1058 * get_device_system_crosststamp - Synchronously capture system/device timestamp
1059 * @get_time_fn: Callback to get simultaneous device time and
1060 * system counter from the device driver
1061 * @ctx: Context passed to get_time_fn()
1062 * @history_begin: Historical reference point used to interpolate system
1063 * time when counter provided by the driver is before the current interval
1064 * @xtstamp: Receives simultaneously captured system and device time
1066 * Reads a timestamp from a device and correlates it to system time
1068 int get_device_system_crosststamp(int (*get_time_fn)
1069 (ktime_t *device_time,
1070 struct system_counterval_t *sys_counterval,
1073 struct system_time_snapshot *history_begin,
1074 struct system_device_crosststamp *xtstamp)
1076 struct system_counterval_t system_counterval;
1077 struct timekeeper *tk = &tk_core.timekeeper;
1078 u64 cycles, now, interval_start;
1079 unsigned int clock_was_set_seq = 0;
1080 ktime_t base_real, base_raw;
1081 u64 nsec_real, nsec_raw;
1082 u8 cs_was_changed_seq;
1088 seq = read_seqcount_begin(&tk_core.seq);
1090 * Try to synchronously capture device time and a system
1091 * counter value calling back into the device driver
1093 ret = get_time_fn(&xtstamp->device, &system_counterval, ctx);
1098 * Verify that the clocksource associated with the captured
1099 * system counter value is the same as the currently installed
1100 * timekeeper clocksource
1102 if (tk->tkr_mono.clock != system_counterval.cs)
1104 cycles = system_counterval.cycles;
1107 * Check whether the system counter value provided by the
1108 * device driver is on the current timekeeping interval.
1110 now = tk->tkr_mono.read(tk->tkr_mono.clock);
1111 interval_start = tk->tkr_mono.cycle_last;
1112 if (!cycle_between(interval_start, cycles, now)) {
1113 clock_was_set_seq = tk->clock_was_set_seq;
1114 cs_was_changed_seq = tk->cs_was_changed_seq;
1115 cycles = interval_start;
1121 base_real = ktime_add(tk->tkr_mono.base,
1122 tk_core.timekeeper.offs_real);
1123 base_raw = tk->tkr_raw.base;
1125 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono,
1126 system_counterval.cycles);
1127 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw,
1128 system_counterval.cycles);
1129 } while (read_seqcount_retry(&tk_core.seq, seq));
1131 xtstamp->sys_realtime = ktime_add_ns(base_real, nsec_real);
1132 xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw);
1135 * Interpolate if necessary, adjusting back from the start of the
1139 u64 partial_history_cycles, total_history_cycles;
1143 * Check that the counter value occurs after the provided
1144 * history reference and that the history doesn't cross a
1145 * clocksource change
1147 if (!history_begin ||
1148 !cycle_between(history_begin->cycles,
1149 system_counterval.cycles, cycles) ||
1150 history_begin->cs_was_changed_seq != cs_was_changed_seq)
1152 partial_history_cycles = cycles - system_counterval.cycles;
1153 total_history_cycles = cycles - history_begin->cycles;
1155 history_begin->clock_was_set_seq != clock_was_set_seq;
1157 ret = adjust_historical_crosststamp(history_begin,
1158 partial_history_cycles,
1159 total_history_cycles,
1160 discontinuity, xtstamp);
1167 EXPORT_SYMBOL_GPL(get_device_system_crosststamp);
1170 * do_gettimeofday - Returns the time of day in a timeval
1171 * @tv: pointer to the timeval to be set
1173 * NOTE: Users should be converted to using getnstimeofday()
1175 void do_gettimeofday(struct timeval *tv)
1177 struct timespec64 now;
1179 getnstimeofday64(&now);
1180 tv->tv_sec = now.tv_sec;
1181 tv->tv_usec = now.tv_nsec/1000;
1183 EXPORT_SYMBOL(do_gettimeofday);
1186 * do_settimeofday64 - Sets the time of day.
1187 * @ts: pointer to the timespec64 variable containing the new time
1189 * Sets the time of day to the new time and update NTP and notify hrtimers
1191 int do_settimeofday64(const struct timespec64 *ts)
1193 struct timekeeper *tk = &tk_core.timekeeper;
1194 struct timespec64 ts_delta, xt;
1195 unsigned long flags;
1198 if (!timespec64_valid_strict(ts))
1201 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1202 write_seqcount_begin(&tk_core.seq);
1204 timekeeping_forward_now(tk);
1207 ts_delta.tv_sec = ts->tv_sec - xt.tv_sec;
1208 ts_delta.tv_nsec = ts->tv_nsec - xt.tv_nsec;
1210 if (timespec64_compare(&tk->wall_to_monotonic, &ts_delta) > 0) {
1215 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
1217 tk_set_xtime(tk, ts);
1219 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1221 write_seqcount_end(&tk_core.seq);
1222 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1224 /* signal hrtimers about time change */
1229 EXPORT_SYMBOL(do_settimeofday64);
1232 * timekeeping_inject_offset - Adds or subtracts from the current time.
1233 * @tv: pointer to the timespec variable containing the offset
1235 * Adds or subtracts an offset value from the current time.
1237 int timekeeping_inject_offset(struct timespec *ts)
1239 struct timekeeper *tk = &tk_core.timekeeper;
1240 unsigned long flags;
1241 struct timespec64 ts64, tmp;
1244 if (!timespec_inject_offset_valid(ts))
1247 ts64 = timespec_to_timespec64(*ts);
1249 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1250 write_seqcount_begin(&tk_core.seq);
1252 timekeeping_forward_now(tk);
1254 /* Make sure the proposed value is valid */
1255 tmp = timespec64_add(tk_xtime(tk), ts64);
1256 if (timespec64_compare(&tk->wall_to_monotonic, &ts64) > 0 ||
1257 !timespec64_valid_strict(&tmp)) {
1262 tk_xtime_add(tk, &ts64);
1263 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts64));
1265 error: /* even if we error out, we forwarded the time, so call update */
1266 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1268 write_seqcount_end(&tk_core.seq);
1269 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1271 /* signal hrtimers about time change */
1276 EXPORT_SYMBOL(timekeeping_inject_offset);
1279 * __timekeeping_set_tai_offset - Sets the TAI offset from UTC and monotonic
1282 static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1284 tk->tai_offset = tai_offset;
1285 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1289 * change_clocksource - Swaps clocksources if a new one is available
1291 * Accumulates current time interval and initializes new clocksource
1293 static int change_clocksource(void *data)
1295 struct timekeeper *tk = &tk_core.timekeeper;
1296 struct clocksource *new, *old;
1297 unsigned long flags;
1299 new = (struct clocksource *) data;
1301 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1302 write_seqcount_begin(&tk_core.seq);
1304 timekeeping_forward_now(tk);
1306 * If the cs is in module, get a module reference. Succeeds
1307 * for built-in code (owner == NULL) as well.
1309 if (try_module_get(new->owner)) {
1310 if (!new->enable || new->enable(new) == 0) {
1311 old = tk->tkr_mono.clock;
1312 tk_setup_internals(tk, new);
1315 module_put(old->owner);
1317 module_put(new->owner);
1320 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1322 write_seqcount_end(&tk_core.seq);
1323 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1329 * timekeeping_notify - Install a new clock source
1330 * @clock: pointer to the clock source
1332 * This function is called from clocksource.c after a new, better clock
1333 * source has been registered. The caller holds the clocksource_mutex.
1335 int timekeeping_notify(struct clocksource *clock)
1337 struct timekeeper *tk = &tk_core.timekeeper;
1339 if (tk->tkr_mono.clock == clock)
1341 stop_machine(change_clocksource, clock, NULL);
1342 tick_clock_notify();
1343 return tk->tkr_mono.clock == clock ? 0 : -1;
1347 * getrawmonotonic64 - Returns the raw monotonic time in a timespec
1348 * @ts: pointer to the timespec64 to be set
1350 * Returns the raw monotonic time (completely un-modified by ntp)
1352 void getrawmonotonic64(struct timespec64 *ts)
1354 struct timekeeper *tk = &tk_core.timekeeper;
1355 struct timespec64 ts64;
1360 seq = read_seqcount_begin(&tk_core.seq);
1361 nsecs = timekeeping_get_ns(&tk->tkr_raw);
1362 ts64 = tk->raw_time;
1364 } while (read_seqcount_retry(&tk_core.seq, seq));
1366 timespec64_add_ns(&ts64, nsecs);
1369 EXPORT_SYMBOL(getrawmonotonic64);
1373 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1375 int timekeeping_valid_for_hres(void)
1377 struct timekeeper *tk = &tk_core.timekeeper;
1382 seq = read_seqcount_begin(&tk_core.seq);
1384 ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1386 } while (read_seqcount_retry(&tk_core.seq, seq));
1392 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1394 u64 timekeeping_max_deferment(void)
1396 struct timekeeper *tk = &tk_core.timekeeper;
1401 seq = read_seqcount_begin(&tk_core.seq);
1403 ret = tk->tkr_mono.clock->max_idle_ns;
1405 } while (read_seqcount_retry(&tk_core.seq, seq));
1411 * read_persistent_clock - Return time from the persistent clock.
1413 * Weak dummy function for arches that do not yet support it.
1414 * Reads the time from the battery backed persistent clock.
1415 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1417 * XXX - Do be sure to remove it once all arches implement it.
1419 void __weak read_persistent_clock(struct timespec *ts)
1425 void __weak read_persistent_clock64(struct timespec64 *ts64)
1429 read_persistent_clock(&ts);
1430 *ts64 = timespec_to_timespec64(ts);
1434 * read_boot_clock64 - Return time of the system start.
1436 * Weak dummy function for arches that do not yet support it.
1437 * Function to read the exact time the system has been started.
1438 * Returns a timespec64 with tv_sec=0 and tv_nsec=0 if unsupported.
1440 * XXX - Do be sure to remove it once all arches implement it.
1442 void __weak read_boot_clock64(struct timespec64 *ts)
1448 /* Flag for if timekeeping_resume() has injected sleeptime */
1449 static bool sleeptime_injected;
1451 /* Flag for if there is a persistent clock on this platform */
1452 static bool persistent_clock_exists;
1455 * timekeeping_init - Initializes the clocksource and common timekeeping values
1457 void __init timekeeping_init(void)
1459 struct timekeeper *tk = &tk_core.timekeeper;
1460 struct clocksource *clock;
1461 unsigned long flags;
1462 struct timespec64 now, boot, tmp;
1464 read_persistent_clock64(&now);
1465 if (!timespec64_valid_strict(&now)) {
1466 pr_warn("WARNING: Persistent clock returned invalid value!\n"
1467 " Check your CMOS/BIOS settings.\n");
1470 } else if (now.tv_sec || now.tv_nsec)
1471 persistent_clock_exists = true;
1473 read_boot_clock64(&boot);
1474 if (!timespec64_valid_strict(&boot)) {
1475 pr_warn("WARNING: Boot clock returned invalid value!\n"
1476 " Check your CMOS/BIOS settings.\n");
1481 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1482 write_seqcount_begin(&tk_core.seq);
1485 clock = clocksource_default_clock();
1487 clock->enable(clock);
1488 tk_setup_internals(tk, clock);
1490 tk_set_xtime(tk, &now);
1491 tk->raw_time.tv_sec = 0;
1492 tk->raw_time.tv_nsec = 0;
1493 if (boot.tv_sec == 0 && boot.tv_nsec == 0)
1494 boot = tk_xtime(tk);
1496 set_normalized_timespec64(&tmp, -boot.tv_sec, -boot.tv_nsec);
1497 tk_set_wall_to_mono(tk, tmp);
1499 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1501 write_seqcount_end(&tk_core.seq);
1502 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1505 /* time in seconds when suspend began for persistent clock */
1506 static struct timespec64 timekeeping_suspend_time;
1509 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1510 * @delta: pointer to a timespec delta value
1512 * Takes a timespec offset measuring a suspend interval and properly
1513 * adds the sleep offset to the timekeeping variables.
1515 static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1516 struct timespec64 *delta)
1518 if (!timespec64_valid_strict(delta)) {
1519 printk_deferred(KERN_WARNING
1520 "__timekeeping_inject_sleeptime: Invalid "
1521 "sleep delta value!\n");
1524 tk_xtime_add(tk, delta);
1525 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1526 tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1527 tk_debug_account_sleep_time(delta);
1530 #if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1532 * We have three kinds of time sources to use for sleep time
1533 * injection, the preference order is:
1534 * 1) non-stop clocksource
1535 * 2) persistent clock (ie: RTC accessible when irqs are off)
1538 * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1539 * If system has neither 1) nor 2), 3) will be used finally.
1542 * If timekeeping has injected sleeptime via either 1) or 2),
1543 * 3) becomes needless, so in this case we don't need to call
1544 * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1547 bool timekeeping_rtc_skipresume(void)
1549 return sleeptime_injected;
1553 * 1) can be determined whether to use or not only when doing
1554 * timekeeping_resume() which is invoked after rtc_suspend(),
1555 * so we can't skip rtc_suspend() surely if system has 1).
1557 * But if system has 2), 2) will definitely be used, so in this
1558 * case we don't need to call rtc_suspend(), and this is what
1559 * timekeeping_rtc_skipsuspend() means.
1561 bool timekeeping_rtc_skipsuspend(void)
1563 return persistent_clock_exists;
1567 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1568 * @delta: pointer to a timespec64 delta value
1570 * This hook is for architectures that cannot support read_persistent_clock64
1571 * because their RTC/persistent clock is only accessible when irqs are enabled.
1572 * and also don't have an effective nonstop clocksource.
1574 * This function should only be called by rtc_resume(), and allows
1575 * a suspend offset to be injected into the timekeeping values.
1577 void timekeeping_inject_sleeptime64(struct timespec64 *delta)
1579 struct timekeeper *tk = &tk_core.timekeeper;
1580 unsigned long flags;
1582 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1583 write_seqcount_begin(&tk_core.seq);
1585 timekeeping_forward_now(tk);
1587 __timekeeping_inject_sleeptime(tk, delta);
1589 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1591 write_seqcount_end(&tk_core.seq);
1592 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1594 /* signal hrtimers about time change */
1600 * timekeeping_resume - Resumes the generic timekeeping subsystem.
1602 void timekeeping_resume(void)
1604 struct timekeeper *tk = &tk_core.timekeeper;
1605 struct clocksource *clock = tk->tkr_mono.clock;
1606 unsigned long flags;
1607 struct timespec64 ts_new, ts_delta;
1610 sleeptime_injected = false;
1611 read_persistent_clock64(&ts_new);
1613 clockevents_resume();
1614 clocksource_resume();
1616 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1617 write_seqcount_begin(&tk_core.seq);
1620 * After system resumes, we need to calculate the suspended time and
1621 * compensate it for the OS time. There are 3 sources that could be
1622 * used: Nonstop clocksource during suspend, persistent clock and rtc
1625 * One specific platform may have 1 or 2 or all of them, and the
1626 * preference will be:
1627 * suspend-nonstop clocksource -> persistent clock -> rtc
1628 * The less preferred source will only be tried if there is no better
1629 * usable source. The rtc part is handled separately in rtc core code.
1631 cycle_now = tk->tkr_mono.read(clock);
1632 if ((clock->flags & CLOCK_SOURCE_SUSPEND_NONSTOP) &&
1633 cycle_now > tk->tkr_mono.cycle_last) {
1634 u64 nsec, cyc_delta;
1636 cyc_delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last,
1638 nsec = mul_u64_u32_shr(cyc_delta, clock->mult, clock->shift);
1639 ts_delta = ns_to_timespec64(nsec);
1640 sleeptime_injected = true;
1641 } else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1642 ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1643 sleeptime_injected = true;
1646 if (sleeptime_injected)
1647 __timekeeping_inject_sleeptime(tk, &ts_delta);
1649 /* Re-base the last cycle value */
1650 tk->tkr_mono.cycle_last = cycle_now;
1651 tk->tkr_raw.cycle_last = cycle_now;
1654 timekeeping_suspended = 0;
1655 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1656 write_seqcount_end(&tk_core.seq);
1657 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1659 touch_softlockup_watchdog();
1665 int timekeeping_suspend(void)
1667 struct timekeeper *tk = &tk_core.timekeeper;
1668 unsigned long flags;
1669 struct timespec64 delta, delta_delta;
1670 static struct timespec64 old_delta;
1672 read_persistent_clock64(&timekeeping_suspend_time);
1675 * On some systems the persistent_clock can not be detected at
1676 * timekeeping_init by its return value, so if we see a valid
1677 * value returned, update the persistent_clock_exists flag.
1679 if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1680 persistent_clock_exists = true;
1682 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1683 write_seqcount_begin(&tk_core.seq);
1684 timekeeping_forward_now(tk);
1685 timekeeping_suspended = 1;
1687 if (persistent_clock_exists) {
1689 * To avoid drift caused by repeated suspend/resumes,
1690 * which each can add ~1 second drift error,
1691 * try to compensate so the difference in system time
1692 * and persistent_clock time stays close to constant.
1694 delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1695 delta_delta = timespec64_sub(delta, old_delta);
1696 if (abs(delta_delta.tv_sec) >= 2) {
1698 * if delta_delta is too large, assume time correction
1699 * has occurred and set old_delta to the current delta.
1703 /* Otherwise try to adjust old_system to compensate */
1704 timekeeping_suspend_time =
1705 timespec64_add(timekeeping_suspend_time, delta_delta);
1709 timekeeping_update(tk, TK_MIRROR);
1710 halt_fast_timekeeper(tk);
1711 write_seqcount_end(&tk_core.seq);
1712 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1715 clocksource_suspend();
1716 clockevents_suspend();
1721 /* sysfs resume/suspend bits for timekeeping */
1722 static struct syscore_ops timekeeping_syscore_ops = {
1723 .resume = timekeeping_resume,
1724 .suspend = timekeeping_suspend,
1727 static int __init timekeeping_init_ops(void)
1729 register_syscore_ops(&timekeeping_syscore_ops);
1732 device_initcall(timekeeping_init_ops);
1735 * Apply a multiplier adjustment to the timekeeper
1737 static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1742 s64 interval = tk->cycle_interval;
1746 mult_adj = -mult_adj;
1747 interval = -interval;
1750 mult_adj <<= adj_scale;
1751 interval <<= adj_scale;
1752 offset <<= adj_scale;
1755 * So the following can be confusing.
1757 * To keep things simple, lets assume mult_adj == 1 for now.
1759 * When mult_adj != 1, remember that the interval and offset values
1760 * have been appropriately scaled so the math is the same.
1762 * The basic idea here is that we're increasing the multiplier
1763 * by one, this causes the xtime_interval to be incremented by
1764 * one cycle_interval. This is because:
1765 * xtime_interval = cycle_interval * mult
1766 * So if mult is being incremented by one:
1767 * xtime_interval = cycle_interval * (mult + 1)
1769 * xtime_interval = (cycle_interval * mult) + cycle_interval
1770 * Which can be shortened to:
1771 * xtime_interval += cycle_interval
1773 * So offset stores the non-accumulated cycles. Thus the current
1774 * time (in shifted nanoseconds) is:
1775 * now = (offset * adj) + xtime_nsec
1776 * Now, even though we're adjusting the clock frequency, we have
1777 * to keep time consistent. In other words, we can't jump back
1778 * in time, and we also want to avoid jumping forward in time.
1780 * So given the same offset value, we need the time to be the same
1781 * both before and after the freq adjustment.
1782 * now = (offset * adj_1) + xtime_nsec_1
1783 * now = (offset * adj_2) + xtime_nsec_2
1785 * (offset * adj_1) + xtime_nsec_1 =
1786 * (offset * adj_2) + xtime_nsec_2
1790 * (offset * adj_1) + xtime_nsec_1 =
1791 * (offset * (adj_1+1)) + xtime_nsec_2
1792 * (offset * adj_1) + xtime_nsec_1 =
1793 * (offset * adj_1) + offset + xtime_nsec_2
1794 * Canceling the sides:
1795 * xtime_nsec_1 = offset + xtime_nsec_2
1797 * xtime_nsec_2 = xtime_nsec_1 - offset
1798 * Which simplfies to:
1799 * xtime_nsec -= offset
1801 * XXX - TODO: Doc ntp_error calculation.
1803 if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
1804 /* NTP adjustment caused clocksource mult overflow */
1809 tk->tkr_mono.mult += mult_adj;
1810 tk->xtime_interval += interval;
1811 tk->tkr_mono.xtime_nsec -= offset;
1812 tk->ntp_error -= (interval - offset) << tk->ntp_error_shift;
1816 * Calculate the multiplier adjustment needed to match the frequency
1819 static __always_inline void timekeeping_freqadjust(struct timekeeper *tk,
1822 s64 interval = tk->cycle_interval;
1823 s64 xinterval = tk->xtime_interval;
1824 u32 base = tk->tkr_mono.clock->mult;
1825 u32 max = tk->tkr_mono.clock->maxadj;
1826 u32 cur_adj = tk->tkr_mono.mult;
1831 /* Remove any current error adj from freq calculation */
1832 if (tk->ntp_err_mult)
1833 xinterval -= tk->cycle_interval;
1835 tk->ntp_tick = ntp_tick_length();
1837 /* Calculate current error per tick */
1838 tick_error = ntp_tick_length() >> tk->ntp_error_shift;
1839 tick_error -= (xinterval + tk->xtime_remainder);
1841 /* Don't worry about correcting it if its small */
1842 if (likely((tick_error >= 0) && (tick_error <= interval)))
1845 /* preserve the direction of correction */
1846 negative = (tick_error < 0);
1848 /* If any adjustment would pass the max, just return */
1849 if (negative && (cur_adj - 1) <= (base - max))
1851 if (!negative && (cur_adj + 1) >= (base + max))
1854 * Sort out the magnitude of the correction, but
1855 * avoid making so large a correction that we go
1856 * over the max adjustment.
1859 tick_error = abs(tick_error);
1860 while (tick_error > interval) {
1861 u32 adj = 1 << (adj_scale + 1);
1863 /* Check if adjustment gets us within 1 unit from the max */
1864 if (negative && (cur_adj - adj) <= (base - max))
1866 if (!negative && (cur_adj + adj) >= (base + max))
1873 /* scale the corrections */
1874 timekeeping_apply_adjustment(tk, offset, negative, adj_scale);
1878 * Adjust the timekeeper's multiplier to the correct frequency
1879 * and also to reduce the accumulated error value.
1881 static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
1883 /* Correct for the current frequency error */
1884 timekeeping_freqadjust(tk, offset);
1886 /* Next make a small adjustment to fix any cumulative error */
1887 if (!tk->ntp_err_mult && (tk->ntp_error > 0)) {
1888 tk->ntp_err_mult = 1;
1889 timekeeping_apply_adjustment(tk, offset, 0, 0);
1890 } else if (tk->ntp_err_mult && (tk->ntp_error <= 0)) {
1891 /* Undo any existing error adjustment */
1892 timekeeping_apply_adjustment(tk, offset, 1, 0);
1893 tk->ntp_err_mult = 0;
1896 if (unlikely(tk->tkr_mono.clock->maxadj &&
1897 (abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
1898 > tk->tkr_mono.clock->maxadj))) {
1899 printk_once(KERN_WARNING
1900 "Adjusting %s more than 11%% (%ld vs %ld)\n",
1901 tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
1902 (long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
1906 * It may be possible that when we entered this function, xtime_nsec
1907 * was very small. Further, if we're slightly speeding the clocksource
1908 * in the code above, its possible the required corrective factor to
1909 * xtime_nsec could cause it to underflow.
1911 * Now, since we already accumulated the second, cannot simply roll
1912 * the accumulated second back, since the NTP subsystem has been
1913 * notified via second_overflow. So instead we push xtime_nsec forward
1914 * by the amount we underflowed, and add that amount into the error.
1916 * We'll correct this error next time through this function, when
1917 * xtime_nsec is not as small.
1919 if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
1920 s64 neg = -(s64)tk->tkr_mono.xtime_nsec;
1921 tk->tkr_mono.xtime_nsec = 0;
1922 tk->ntp_error += neg << tk->ntp_error_shift;
1927 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
1929 * Helper function that accumulates the nsecs greater than a second
1930 * from the xtime_nsec field to the xtime_secs field.
1931 * It also calls into the NTP code to handle leapsecond processing.
1934 static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
1936 u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
1937 unsigned int clock_set = 0;
1939 while (tk->tkr_mono.xtime_nsec >= nsecps) {
1942 tk->tkr_mono.xtime_nsec -= nsecps;
1945 /* Figure out if its a leap sec and apply if needed */
1946 leap = second_overflow(tk->xtime_sec);
1947 if (unlikely(leap)) {
1948 struct timespec64 ts;
1950 tk->xtime_sec += leap;
1954 tk_set_wall_to_mono(tk,
1955 timespec64_sub(tk->wall_to_monotonic, ts));
1957 __timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
1959 clock_set = TK_CLOCK_WAS_SET;
1966 * logarithmic_accumulation - shifted accumulation of cycles
1968 * This functions accumulates a shifted interval of cycles into
1969 * into a shifted interval nanoseconds. Allows for O(log) accumulation
1972 * Returns the unconsumed cycles.
1974 static u64 logarithmic_accumulation(struct timekeeper *tk, u64 offset,
1975 u32 shift, unsigned int *clock_set)
1977 u64 interval = tk->cycle_interval << shift;
1980 /* If the offset is smaller than a shifted interval, do nothing */
1981 if (offset < interval)
1984 /* Accumulate one shifted interval */
1986 tk->tkr_mono.cycle_last += interval;
1987 tk->tkr_raw.cycle_last += interval;
1989 tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
1990 *clock_set |= accumulate_nsecs_to_secs(tk);
1992 /* Accumulate raw time */
1993 raw_nsecs = (u64)tk->raw_interval << shift;
1994 raw_nsecs += tk->raw_time.tv_nsec;
1995 if (raw_nsecs >= NSEC_PER_SEC) {
1996 u64 raw_secs = raw_nsecs;
1997 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
1998 tk->raw_time.tv_sec += raw_secs;
2000 tk->raw_time.tv_nsec = raw_nsecs;
2002 /* Accumulate error between NTP and clock interval */
2003 tk->ntp_error += tk->ntp_tick << shift;
2004 tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
2005 (tk->ntp_error_shift + shift);
2011 * update_wall_time - Uses the current clocksource to increment the wall time
2014 void update_wall_time(void)
2016 struct timekeeper *real_tk = &tk_core.timekeeper;
2017 struct timekeeper *tk = &shadow_timekeeper;
2019 int shift = 0, maxshift;
2020 unsigned int clock_set = 0;
2021 unsigned long flags;
2023 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2025 /* Make sure we're fully resumed: */
2026 if (unlikely(timekeeping_suspended))
2029 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
2030 offset = real_tk->cycle_interval;
2032 offset = clocksource_delta(tk->tkr_mono.read(tk->tkr_mono.clock),
2033 tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
2036 /* Check if there's really nothing to do */
2037 if (offset < real_tk->cycle_interval)
2040 /* Do some additional sanity checking */
2041 timekeeping_check_update(real_tk, offset);
2044 * With NO_HZ we may have to accumulate many cycle_intervals
2045 * (think "ticks") worth of time at once. To do this efficiently,
2046 * we calculate the largest doubling multiple of cycle_intervals
2047 * that is smaller than the offset. We then accumulate that
2048 * chunk in one go, and then try to consume the next smaller
2051 shift = ilog2(offset) - ilog2(tk->cycle_interval);
2052 shift = max(0, shift);
2053 /* Bound shift to one less than what overflows tick_length */
2054 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
2055 shift = min(shift, maxshift);
2056 while (offset >= tk->cycle_interval) {
2057 offset = logarithmic_accumulation(tk, offset, shift,
2059 if (offset < tk->cycle_interval<<shift)
2063 /* correct the clock when NTP error is too big */
2064 timekeeping_adjust(tk, offset);
2067 * XXX This can be killed once everyone converts
2068 * to the new update_vsyscall.
2070 old_vsyscall_fixup(tk);
2073 * Finally, make sure that after the rounding
2074 * xtime_nsec isn't larger than NSEC_PER_SEC
2076 clock_set |= accumulate_nsecs_to_secs(tk);
2078 write_seqcount_begin(&tk_core.seq);
2080 * Update the real timekeeper.
2082 * We could avoid this memcpy by switching pointers, but that
2083 * requires changes to all other timekeeper usage sites as
2084 * well, i.e. move the timekeeper pointer getter into the
2085 * spinlocked/seqcount protected sections. And we trade this
2086 * memcpy under the tk_core.seq against one before we start
2089 timekeeping_update(tk, clock_set);
2090 memcpy(real_tk, tk, sizeof(*tk));
2091 /* The memcpy must come last. Do not put anything here! */
2092 write_seqcount_end(&tk_core.seq);
2094 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2096 /* Have to call _delayed version, since in irq context*/
2097 clock_was_set_delayed();
2101 * getboottime64 - Return the real time of system boot.
2102 * @ts: pointer to the timespec64 to be set
2104 * Returns the wall-time of boot in a timespec64.
2106 * This is based on the wall_to_monotonic offset and the total suspend
2107 * time. Calls to settimeofday will affect the value returned (which
2108 * basically means that however wrong your real time clock is at boot time,
2109 * you get the right time here).
2111 void getboottime64(struct timespec64 *ts)
2113 struct timekeeper *tk = &tk_core.timekeeper;
2114 ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
2116 *ts = ktime_to_timespec64(t);
2118 EXPORT_SYMBOL_GPL(getboottime64);
2120 unsigned long get_seconds(void)
2122 struct timekeeper *tk = &tk_core.timekeeper;
2124 return tk->xtime_sec;
2126 EXPORT_SYMBOL(get_seconds);
2128 struct timespec __current_kernel_time(void)
2130 struct timekeeper *tk = &tk_core.timekeeper;
2132 return timespec64_to_timespec(tk_xtime(tk));
2135 struct timespec64 current_kernel_time64(void)
2137 struct timekeeper *tk = &tk_core.timekeeper;
2138 struct timespec64 now;
2142 seq = read_seqcount_begin(&tk_core.seq);
2145 } while (read_seqcount_retry(&tk_core.seq, seq));
2149 EXPORT_SYMBOL(current_kernel_time64);
2151 struct timespec64 get_monotonic_coarse64(void)
2153 struct timekeeper *tk = &tk_core.timekeeper;
2154 struct timespec64 now, mono;
2158 seq = read_seqcount_begin(&tk_core.seq);
2161 mono = tk->wall_to_monotonic;
2162 } while (read_seqcount_retry(&tk_core.seq, seq));
2164 set_normalized_timespec64(&now, now.tv_sec + mono.tv_sec,
2165 now.tv_nsec + mono.tv_nsec);
2169 EXPORT_SYMBOL(get_monotonic_coarse64);
2172 * Must hold jiffies_lock
2174 void do_timer(unsigned long ticks)
2176 jiffies_64 += ticks;
2177 calc_global_load(ticks);
2181 * ktime_get_update_offsets_now - hrtimer helper
2182 * @cwsseq: pointer to check and store the clock was set sequence number
2183 * @offs_real: pointer to storage for monotonic -> realtime offset
2184 * @offs_boot: pointer to storage for monotonic -> boottime offset
2185 * @offs_tai: pointer to storage for monotonic -> clock tai offset
2187 * Returns current monotonic time and updates the offsets if the
2188 * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
2191 * Called from hrtimer_interrupt() or retrigger_next_event()
2193 ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real,
2194 ktime_t *offs_boot, ktime_t *offs_tai)
2196 struct timekeeper *tk = &tk_core.timekeeper;
2202 seq = read_seqcount_begin(&tk_core.seq);
2204 base = tk->tkr_mono.base;
2205 nsecs = timekeeping_get_ns(&tk->tkr_mono);
2206 base = ktime_add_ns(base, nsecs);
2208 if (*cwsseq != tk->clock_was_set_seq) {
2209 *cwsseq = tk->clock_was_set_seq;
2210 *offs_real = tk->offs_real;
2211 *offs_boot = tk->offs_boot;
2212 *offs_tai = tk->offs_tai;
2215 /* Handle leapsecond insertion adjustments */
2216 if (unlikely(base >= tk->next_leap_ktime))
2217 *offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0));
2219 } while (read_seqcount_retry(&tk_core.seq, seq));
2225 * do_adjtimex() - Accessor function to NTP __do_adjtimex function
2227 int do_adjtimex(struct timex *txc)
2229 struct timekeeper *tk = &tk_core.timekeeper;
2230 unsigned long flags;
2231 struct timespec64 ts;
2235 /* Validate the data before disabling interrupts */
2236 ret = ntp_validate_timex(txc);
2240 if (txc->modes & ADJ_SETOFFSET) {
2241 struct timespec delta;
2242 delta.tv_sec = txc->time.tv_sec;
2243 delta.tv_nsec = txc->time.tv_usec;
2244 if (!(txc->modes & ADJ_NANO))
2245 delta.tv_nsec *= 1000;
2246 ret = timekeeping_inject_offset(&delta);
2251 getnstimeofday64(&ts);
2253 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2254 write_seqcount_begin(&tk_core.seq);
2256 orig_tai = tai = tk->tai_offset;
2257 ret = __do_adjtimex(txc, &ts, &tai);
2259 if (tai != orig_tai) {
2260 __timekeeping_set_tai_offset(tk, tai);
2261 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
2263 tk_update_leap_state(tk);
2265 write_seqcount_end(&tk_core.seq);
2266 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2268 if (tai != orig_tai)
2271 ntp_notify_cmos_timer();
2276 #ifdef CONFIG_NTP_PPS
2278 * hardpps() - Accessor function to NTP __hardpps function
2280 void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts)
2282 unsigned long flags;
2284 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2285 write_seqcount_begin(&tk_core.seq);
2287 __hardpps(phase_ts, raw_ts);
2289 write_seqcount_end(&tk_core.seq);
2290 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2292 EXPORT_SYMBOL(hardpps);
2296 * xtime_update() - advances the timekeeping infrastructure
2297 * @ticks: number of ticks, that have elapsed since the last call.
2299 * Must be called with interrupts disabled.
2301 void xtime_update(unsigned long ticks)
2303 write_seqlock(&jiffies_lock);
2305 write_sequnlock(&jiffies_lock);